diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 00000000..9a949a0a --- /dev/null +++ b/.gitattributes @@ -0,0 +1 @@ +third_party/onnxruntime/lib/** filter=lfs diff=lfs merge=lfs -text diff --git a/MODULE.bazel b/MODULE.bazel index 0fec2bc3..db9aaee4 100644 --- a/MODULE.bazel +++ b/MODULE.bazel @@ -77,21 +77,3 @@ http_archive( strip_prefix = "OpenCL-Headers-2023.04.17", url = "https://github.com/KhronosGroup/OpenCL-Headers/archive/refs/tags/v2023.04.17.tar.gz", ) - -new_git_repository( - name = "onnx_runtime", - build_file = "//third_party:onnxruntime.BUILD", - # branch= "rel-1.20.1", - commit = "5c1b7ccbff7e5141c1da7a9d963d660e5741c319", - patches = ["//third_party/patch:onnx.patch"], - recursive_init_submodules = True, - remote = "https://github.com/microsoft/onnxruntime", -) - -new_git_repository( - name = "onnx_runtime_extensions", - build_file = "//third_party:onnxruntime_extensions.BUILD", - commit = "81e7799c69044c745239202085eb0a98f102937b", - patches = ["//third_party/patch:onnx_ext.patch"], - remote = "https://github.com/microsoft/onnxruntime-extensions", -) diff --git a/kernels/onnxRT/BUILD b/kernels/onnxRT/BUILD index 654b51ed..deb847d4 100644 --- a/kernels/onnxRT/BUILD +++ b/kernels/onnxRT/BUILD @@ -14,7 +14,9 @@ cc_library( ], deps = [ "//:corevx", - "@onnx_runtime//:onnxruntime_lib", + "//third_party:onnxruntime", ], + linkstatic = True, + alwayslink = True, visibility = ["//visibility:public"] ) \ No newline at end of file diff --git a/kernels/onnxRT/ort_runner.hpp b/kernels/onnxRT/ort_runner.hpp index d9302b29..679013e2 100644 --- a/kernels/onnxRT/ort_runner.hpp +++ b/kernels/onnxRT/ort_runner.hpp @@ -13,8 +13,8 @@ #include #include -#include #include +#include /** * @brief Onnx Runtime Model Runner Object diff --git a/targets/onnxRT/BUILD b/targets/onnxRT/BUILD index cac3d127..e2280eea 100644 --- a/targets/onnxRT/BUILD +++ b/targets/onnxRT/BUILD @@ -1,4 +1,3 @@ - cc_library( name = "onnxRT", srcs = glob([ @@ -15,19 +14,21 @@ cc_library( "//:corevx", "//kernels/onnxRT:ort_kernels", ], + linkstatic = True, + alwayslink = True, visibility = ["//visibility:public"] ) cc_shared_library( - name = "openvx-onnxRT", - deps = [ + name="openvx-onnxRT", + deps=[ ":onnxRT", ], - visibility = ["//visibility:public"] + visibility=["//visibility:public"] ) cc_import( - name = "imported_openvx_onnxRT", - shared_library = ":openvx-onnxRT", - visibility = ["//visibility:public"] + name="imported_openvx_onnxRT", + shared_library=":openvx-onnxRT", + visibility=["//visibility:public"] ) diff --git a/tests/BUILD b/tests/BUILD index 0b407797..79c64e47 100644 --- a/tests/BUILD +++ b/tests/BUILD @@ -10,10 +10,14 @@ cc_binary( ], deps = [ "//:corevx", + "//targets/ai_server:imported_openvx_ai_server", "//targets/c_model:imported_openvx_c_model", "//targets/debug:imported_openvx_debug", "//targets/extras:imported_openvx_extras", + "//targets/liteRT:imported_openvx_liteRT", "//targets/opencl:imported_openvx_opencl", + "//targets/onnxRT:imported_openvx_onnxRT", + "//targets/executorch:imported_openvx_torch", "//tests/xyz:imported_xyz", ], linkopts = select({ diff --git a/tests/integration_test/BUILD b/tests/integration_test/BUILD index 64dd6cc8..4df07631 100644 --- a/tests/integration_test/BUILD +++ b/tests/integration_test/BUILD @@ -20,30 +20,30 @@ cc_test( size = "small" ) -# cc_test( -# name = "test_ort", -# srcs = [ -# "test_ort.cpp" -# ], -# deps = [ -# "//:corevx", -# "@googletest//:gtest_main", -# "//targets/c_model:imported_openvx_c_model", -# "//targets/debug:imported_openvx_debug", -# "//targets/extras:imported_openvx_extras", -# "//targets/opencl:imported_openvx_opencl", -# "//targets/onnxRT:imported_openvx_onnxRT", -# ], -# linkopts = select({ -# "@platforms//os:linux": ["-Wl,-rpath,$ORIGIN"], -# "@platforms//os:macos": ["-Wl,-rpath,@executable_path"], -# "//conditions:default": [], -# }), -# data = [ -# "//tests/raw:models", -# ], -# size = "small" -# ) +cc_test( + name="test_ort", + srcs=[ + "test_ort.cpp" + ], + deps=[ + "//:corevx", + "@googletest//:gtest_main", + "//targets/c_model:imported_openvx_c_model", + "//targets/debug:imported_openvx_debug", + "//targets/extras:imported_openvx_extras", + "//targets/opencl:imported_openvx_opencl", + "//targets/onnxRT:imported_openvx_onnxRT", + ], + linkopts=select({ + "@platforms//os:linux": ["-Wl,-rpath,$ORIGIN"], + "@platforms//os:macos": ["-Wl,-rpath,@executable_path"], + "//conditions:default": [], + }), + data=[ + "//tests/raw:models", + ], + size="small" +) cc_test( name = "test_tflite", diff --git a/third_party/BUILD b/third_party/BUILD index 3907bada..62378748 100644 --- a/third_party/BUILD +++ b/third_party/BUILD @@ -2,6 +2,35 @@ BUILD file for build defs for third party deps """ +# ONNX Runtime +cc_import( + name="onnxruntime-providers", + shared_library=select({ + "@platforms//os:linux": "onnxruntime/lib/libonnxruntime_providers_shared.so", + "//conditions:default": None, + }), + visibility=["//visibility:public"], +) + +cc_import( + name="onnxruntime-prebuilt", + shared_library=select({ + "@platforms//os:macos": "onnxruntime/lib/libonnxruntime.1.22.0.dylib", + "@platforms//os:linux": "onnxruntime/lib/libonnxruntime.so.1", + "//conditions:default": None, + }), + deps=[":onnxruntime-providers"], + visibility=["//visibility:public"], +) + +cc_library( + name="onnxruntime", + hdrs=glob(["onnxruntime/include/**/*.h"]), + includes=["onnxruntime/include"], + deps=[":onnxruntime-prebuilt"], + visibility=["//visibility:public"], +) + # TensorFlow Lite cc_import( name = "tflite", diff --git a/third_party/onnxruntime.BUILD b/third_party/onnxruntime.BUILD deleted file mode 100644 index a3ab289c..00000000 --- a/third_party/onnxruntime.BUILD +++ /dev/null @@ -1,380 +0,0 @@ -load("@rules_foreign_cc//foreign_cc:defs.bzl", "cmake") -load("@cuda_env//:cuda_home.bzl", "CUDA_HOME") -load("@cuda_env//:cudnn_home.bzl", "CUDNN_HOME") - -filegroup( - name = "all_srcs", - srcs = glob(["**"], exclude = ["**/foo.onnx"]), - visibility = ["//visibility:public"], -) - -cc_library( - name = "hdrs", - hdrs = glob(["include/onnxruntime/**/*.h"]), - strip_include_prefix = "include/onnxruntime", - visibility = ["//visibility:public"], -) - -config_setting( - name = "with_cuda", - define_values = { "use_cuda": "" } -) - -__POSTFIX = """ -mkdir -p $INSTALLDIR/lib -mkdir -p $INSTALLDIR/lib/_deps -mkdir -p $INSTALLDIR/lib/_deps/onnx-build -mkdir -p $INSTALLDIR/lib/_deps/re2-build -mkdir -p $INSTALLDIR/lib/_deps/abseil_cpp-build -mkdir -p $INSTALLDIR/lib/_deps/protobuf-build -mkdir -p $INSTALLDIR/lib/_deps/abseil_cpp-build/absl/container -mkdir -p $INSTALLDIR/lib/_deps/abseil_cpp-build/absl/hash -mkdir -p $INSTALLDIR/lib/_deps/pytorch_cpuinfo-build -mkdir -p $INSTALLDIR/lib/_deps/pytorch_cpuinfo-build/deps -mkdir -p $INSTALLDIR/lib/_deps/pytorch_clog-build -mkdir -p $INSTALLDIR/lib/_deps/google_nsync-build -mkdir -p $INSTALLDIR/lib/_deps/opencv-build -mkdir -p $INSTALLDIR/lib/_deps/opencv-build/lib -mkdir -p $INSTALLDIR/lib/_deps/opencv-build/3rdparty -mkdir -p $INSTALLDIR/lib/_deps/opencv-build/3rdparty/lib -cp $BUILD_TMPDIR/_deps/onnx-build/libonnx.a $INSTALLDIR/lib/_deps/onnx-build -cp $BUILD_TMPDIR/_deps/onnx-build/libonnx_proto.a $INSTALLDIR/lib/_deps/onnx-build -cp $BUILD_TMPDIR/_deps/re2-build/libre2.a $INSTALLDIR/lib/_deps/re2-build -cp -r $BUILD_TMPDIR/_deps/abseil_cpp-build/. $INSTALLDIR/lib/_deps/abseil_cpp-build -cp $BUILD_TMPDIR/_deps/google_nsync-build/libnsync_cpp.a $INSTALLDIR/lib/_deps/google_nsync-build -cp $BUILD_TMPDIR/_deps/pytorch_clog-build/libclog.a $INSTALLDIR/lib/_deps/pytorch_clog-build -cp $BUILD_TMPDIR/_deps/pytorch_cpuinfo-build/libcpuinfo.a $INSTALLDIR/lib/_deps/pytorch_cpuinfo-build -cp $BUILD_TMPDIR/_deps/protobuf-build/libprotobuf.a $INSTALLDIR/lib/_deps/protobuf-build -# cp $BUILD_TMPDIR/_deps/opencv-build/lib/libopencv_imgcodecs.a $INSTALLDIR/lib/_deps/opencv-build/lib -# cp $BUILD_TMPDIR/_deps/opencv-build/lib/libopencv_imgproc.a $INSTALLDIR/lib/_deps/opencv-build/lib -# cp $BUILD_TMPDIR/_deps/opencv-build/lib/libopencv_core.a $INSTALLDIR/lib/_deps/opencv-build/lib -# cp $BUILD_TMPDIR/_deps/opencv-build/3rdparty/lib/liblibjpeg-turbo.a $INSTALLDIR/lib/_deps/opencv-build/3rdparty/lib -# cp $BUILD_TMPDIR/_deps/opencv-build/3rdparty/lib/liblibpng.a $INSTALLDIR/lib/_deps/opencv-build/3rdparty/lib -# cp $BUILD_TMPDIR/lib/libnoexcep_operators.a $INSTALLDIR/lib -# cp $BUILD_TMPDIR/lib/libocos_operators.a $INSTALLDIR/lib -# cp $BUILD_TMPDIR/lib/libortcustomops.a $INSTALLDIR/lib -""" - -__POSTFIX_WITH_CUDA = __POSTFIX + """ -cp $BUILD_TMPDIR/libonnxruntime_providers_shared.so $INSTALLDIR/../../../ -cp $BUILD_TMPDIR/libonnxruntime_providers_cuda.so $INSTALLDIR/../../../ -""" - -__ONNXRUNTIME_WITHOUT_CUDA = { -'onnxruntime_RUN_ONNX_TESTS':'OFF', -'onnxruntime_GENERATE_TEST_REPORTS':'OFF', -'onnxruntime_USE_MIMALLOC':'OFF', -'onnxruntime_ENABLE_PYTHON':'OFF', -'onnxruntime_BUILD_CSHARP':'OFF', -'onnxruntime_BUILD_JAVA':'OFF', -'onnxruntime_BUILD_NODEJS':'OFF', -'onnxruntime_BUILD_OBJC':'OFF', -'onnxruntime_BUILD_SHARED_LIB':'OFF', -'onnxruntime_BUILD_APPLE_FRAMEWORK':'OFF', -'onnxruntime_USE_DNNL':'OFF', -'onnxruntime_USE_NNAPI_BUILTIN':'OFF', -'onnxruntime_USE_RKNPU':'OFF', -'onnxruntime_USE_LLVM':'OFF', -'onnxruntime_ENABLE_MICROSOFT_INTERNAL':'OFF', -'onnxruntime_USE_VITISAI':'OFF', -'onnxruntime_USE_TENSORRT':'OFF', -'onnxruntime_SKIP_AND_PERFORM_FILTERED_TENSORRT_TESTS':'OFF', -'onnxruntime_USE_TENSORRT_BUILTIN_PARSER':'OFF', -'onnxruntime_TENSORRT_PLACEHOLDER_BUILDER':'OFF', -'onnxruntime_USE_TVM':'OFF', -'onnxruntime_TVM_CUDA_RUNTIME':'OFF', -'onnxruntime_TVM_USE_HASH':'OFF', -'onnxruntime_USE_MIGRAPHX':'OFF', -'onnxruntime_CROSS_COMPILING':'OFF', -'onnxruntime_DISABLE_CONTRIB_OPS':'OFF', -'onnxruntime_DISABLE_ML_OPS':'OFF', -'onnxruntime_DISABLE_RTTI':'OFF', -'onnxruntime_DISABLE_EXCEPTIONS':'OFF', -'onnxruntime_MINIMAL_BUILD':'OFF', -'onnxruntime_EXTENDED_MINIMAL_BUILD':'OFF', -'onnxruntime_MINIMAL_BUILD_CUSTOM_OPS':'OFF', -'onnxruntime_REDUCED_OPS_BUILD':'OFF', -'onnxruntime_ENABLE_LANGUAGE_INTEROP_OPS':'OFF', -'onnxruntime_USE_DML':'OFF', -'onnxruntime_USE_WINML':'OFF', -'onnxruntime_BUILD_MS_EXPERIMENTAL_OPS':'OFF', -'onnxruntime_USE_TELEMETRY':'OFF', -'onnxruntime_ENABLE_LTO':'OFF', -'onnxruntime_USE_ACL':'OFF', -'onnxruntime_USE_ACL_1902':'OFF', -'onnxruntime_USE_ACL_1905':'OFF', -'onnxruntime_USE_ACL_1908':'OFF', -'onnxruntime_USE_ACL_2002':'OFF', -'onnxruntime_USE_ARMNN':'OFF', -'onnxruntime_ARMNN_RELU_USE_CPU':'ON', -'onnxruntime_ARMNN_BN_USE_CPU':'ON', -'onnxruntime_ENABLE_NVTX_PROFILE':'OFF', -'onnxruntime_ENABLE_TRAINING':'OFF', -'onnxruntime_ENABLE_TRAINING_OPS':'OFF', -'onnxruntime_ENABLE_TRAINING_APIS':'OFF', -'onnxruntime_ENABLE_CPU_FP16_OPS':'OFF', -'onnxruntime_USE_NCCL':'OFF', -'onnxruntime_BUILD_BENCHMARKS':'OFF', -'onnxruntime_USE_ROCM':'OFF', -'Onnxruntime_GCOV_COVERAGE':'OFF', -'onnxruntime_USE_MPI':'OFF', -'onnxruntime_ENABLE_MEMORY_PROFILE':'OFF', -'onnxruntime_ENABLE_CUDA_LINE_NUMBER_INFO':'OFF', -'onnxruntime_BUILD_WEBASSEMBLY':'OFF', -'onnxruntime_BUILD_WEBASSEMBLY_STATIC_LIB':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_EXCEPTION_CATCHING':'ON', -'onnxruntime_ENABLE_WEBASSEMBLY_API_EXCEPTION_CATCHING':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_EXCEPTION_THROWING':'ON', -'onnxruntime_ENABLE_WEBASSEMBLY_THREADS':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_DEBUG_INFO':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_PROFILING':'OFF', -'onnxruntime_ENABLE_EAGER_MODE':'OFF', -'onnxruntime_ENABLE_LAZY_TENSOR':'OFF', -'onnxruntime_ENABLE_EXTERNAL_CUSTOM_OP_SCHEMAS':'OFF', -'onnxruntime_ENABLE_CUDA_PROFILING':'OFF', -'onnxruntime_ENABLE_ROCM_PROFILING':'OFF', -'onnxruntime_USE_XNNPACK':'OFF', -'onnxruntime_USE_CANN':'OFF', -'CMAKE_TLS_VERIFY':'ON', -'FETCHCONTENT_QUIET':'OFF', -'onnxruntime_PYBIND_EXPORT_OPSCHEMA':'OFF', -'onnxruntime_ENABLE_MEMLEAK_CHECKER':'OFF', -'CMAKE_BUILD_TYPE':'Release', -'onnxruntime_USE_EXTENSIONS': 'OFF', -'onnxruntime_EXTENSIONS_PATH': '$EXT_BUILD_ROOT/external/onnx_runtime/cmake/external/onnxruntime-extensions', -'onnxruntime_USE_FULL_PROTOBUF': 'ON', -'Protobuf_USE_STATIC_LIBS': 'ON', -'OCOS_ENABLE_BLINGFIRE': 'OFF' -} - -__ONNXRUNTIME_WITH_CUDA = { -'onnxruntime_RUN_ONNX_TESTS':'OFF', -'onnxruntime_GENERATE_TEST_REPORTS':'OFF', -'onnxruntime_USE_MIMALLOC':'OFF', -'onnxruntime_ENABLE_PYTHON':'OFF', -'onnxruntime_BUILD_CSHARP':'OFF', -'onnxruntime_BUILD_JAVA':'OFF', -'onnxruntime_BUILD_NODEJS':'OFF', -'onnxruntime_BUILD_OBJC':'OFF', -'onnxruntime_BUILD_SHARED_LIB':'OFF', -'onnxruntime_BUILD_APPLE_FRAMEWORK':'OFF', -'onnxruntime_USE_DNNL':'OFF', -'onnxruntime_USE_NNAPI_BUILTIN':'OFF', -'onnxruntime_USE_RKNPU':'OFF', -'onnxruntime_USE_LLVM':'OFF', -'onnxruntime_ENABLE_MICROSOFT_INTERNAL':'OFF', -'onnxruntime_USE_VITISAI':'OFF', -'onnxruntime_USE_TENSORRT':'OFF', -'onnxruntime_SKIP_AND_PERFORM_FILTERED_TENSORRT_TESTS':'OFF', -'onnxruntime_USE_TENSORRT_BUILTIN_PARSER':'OFF', -'onnxruntime_TENSORRT_PLACEHOLDER_BUILDER':'OFF', -'onnxruntime_USE_TVM':'OFF', -'onnxruntime_TVM_CUDA_RUNTIME':'OFF', -'onnxruntime_TVM_USE_HASH':'OFF', -'onnxruntime_USE_MIGRAPHX':'OFF', -'onnxruntime_CROSS_COMPILING':'OFF', -'onnxruntime_DISABLE_CONTRIB_OPS':'OFF', -'onnxruntime_DISABLE_ML_OPS':'OFF', -'onnxruntime_DISABLE_RTTI':'OFF', -'onnxruntime_DISABLE_EXCEPTIONS':'OFF', -'onnxruntime_MINIMAL_BUILD':'OFF', -'onnxruntime_EXTENDED_MINIMAL_BUILD':'OFF', -'onnxruntime_MINIMAL_BUILD_CUSTOM_OPS':'OFF', -'onnxruntime_REDUCED_OPS_BUILD':'OFF', -'onnxruntime_ENABLE_LANGUAGE_INTEROP_OPS':'OFF', -'onnxruntime_USE_DML':'OFF', -'onnxruntime_USE_WINML':'OFF', -'onnxruntime_BUILD_MS_EXPERIMENTAL_OPS':'OFF', -'onnxruntime_USE_TELEMETRY':'OFF', -'onnxruntime_ENABLE_LTO':'OFF', -'onnxruntime_USE_ACL':'OFF', -'onnxruntime_USE_ACL_1902':'OFF', -'onnxruntime_USE_ACL_1905':'OFF', -'onnxruntime_USE_ACL_1908':'OFF', -'onnxruntime_USE_ACL_2002':'OFF', -'onnxruntime_USE_ARMNN':'OFF', -'onnxruntime_ARMNN_RELU_USE_CPU':'ON', -'onnxruntime_ARMNN_BN_USE_CPU':'ON', -'onnxruntime_ENABLE_NVTX_PROFILE':'OFF', -'onnxruntime_ENABLE_TRAINING':'OFF', -'onnxruntime_ENABLE_TRAINING_OPS':'OFF', -'onnxruntime_ENABLE_TRAINING_APIS':'OFF', -'onnxruntime_ENABLE_CPU_FP16_OPS':'OFF', -'onnxruntime_USE_NCCL':'OFF', -'onnxruntime_BUILD_BENCHMARKS':'OFF', -'onnxruntime_USE_ROCM':'OFF', -'Onnxruntime_GCOV_COVERAGE':'OFF', -'onnxruntime_USE_MPI':'OFF', -'onnxruntime_ENABLE_MEMORY_PROFILE':'OFF', -'onnxruntime_ENABLE_CUDA_LINE_NUMBER_INFO':'OFF', -'onnxruntime_BUILD_WEBASSEMBLY':'OFF', -'onnxruntime_BUILD_WEBASSEMBLY_STATIC_LIB':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_EXCEPTION_CATCHING':'ON', -'onnxruntime_ENABLE_WEBASSEMBLY_API_EXCEPTION_CATCHING':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_EXCEPTION_THROWING':'ON', -'onnxruntime_ENABLE_WEBASSEMBLY_THREADS':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_DEBUG_INFO':'OFF', -'onnxruntime_ENABLE_WEBASSEMBLY_PROFILING':'OFF', -'onnxruntime_ENABLE_EAGER_MODE':'OFF', -'onnxruntime_ENABLE_LAZY_TENSOR':'OFF', -'onnxruntime_ENABLE_EXTERNAL_CUSTOM_OP_SCHEMAS':'OFF', -'onnxruntime_ENABLE_CUDA_PROFILING':'OFF', -'onnxruntime_ENABLE_ROCM_PROFILING':'OFF', -'onnxruntime_USE_XNNPACK':'OFF', -'onnxruntime_USE_CANN':'OFF', -'CMAKE_TLS_VERIFY':'ON', -'FETCHCONTENT_QUIET':'OFF', -'onnxruntime_PYBIND_EXPORT_OPSCHEMA':'OFF', -'onnxruntime_ENABLE_MEMLEAK_CHECKER':'OFF', -'CMAKE_BUILD_TYPE':'Release', -'onnxruntime_USE_CUDA':'ON', -'onnxruntime_USE_CUDNN':'ON', -'onnxruntime_USE_EXTENSIONS': 'OFF', -'onnxruntime_EXTENSIONS_PATH': '$EXT_BUILD_ROOT/external/onnx_runtime/cmake/external/onnxruntime-extensions', -'onnxruntime_CUDA_HOME': CUDA_HOME, -'onnxruntime_CUDNN_HOME': CUDNN_HOME, -'CMAKE_CUDA_COMPILER': CUDA_HOME + "/bin/nvcc", -'onnxruntime_USE_FULL_PROTOBUF': 'ON', -'Protobuf_USE_STATIC_LIBS': 'ON', -'OCOS_ENABLE_BLINGFIRE': 'OFF' -} - -cmake( - name = "onnxruntime", - lib_source = "//:all_srcs", - cache_entries = select({ - ":with_cuda": __ONNXRUNTIME_WITH_CUDA, - "//conditions:default": __ONNXRUNTIME_WITHOUT_CUDA - }), - working_directory="cmake", - generate_args=["-GNinja"], - build_args= [ - "--config Release", - ], - tags=[ - "requires-network", - # "no-sandbox" - ], - features=[ - "-default_compile_flags", - "-fno-canonical-system-headers", - "-Wno-builtin-macro-redefined", - "-Wno-deprecated-declarations" - ], - out_static_libs=[ - "libonnxruntime_common.a", - "libonnxruntime_flatbuffers.a", - "libonnxruntime_framework.a", - "libonnxruntime_graph.a", - "libonnxruntime_lora.a", - "libonnxruntime_mlas.a", - "libonnxruntime_optimizer.a", - "libonnxruntime_session.a", - "libonnxruntime_providers.a", - "libonnxruntime_util.a", - # "libortcustomops.a", - # "libocos_operators.a", - # "libnoexcep_operators.a", - "_deps/onnx-build/libonnx.a", - "_deps/onnx-build/libonnx_proto.a", - "_deps/re2-build/libre2.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_cord.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_string_view.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_cord_internal.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_cordz_info.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_strings.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_str_format_internal.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_strings_internal.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_cordz_handle.a", - "_deps/abseil_cpp-build/absl/strings/libabsl_cordz_functions.a", - "_deps/abseil_cpp-build/absl/types/libabsl_bad_optional_access.a", - "_deps/abseil_cpp-build/absl/types/libabsl_bad_any_cast_impl.a", - "_deps/abseil_cpp-build/absl/types/libabsl_bad_variant_access.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_commandlineflag.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_usage_internal.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_private_handle_accessor.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_usage.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_program_name.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_parse.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_config.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_commandlineflag_internal.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_marshalling.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_reflection.a", - "_deps/abseil_cpp-build/absl/flags/libabsl_flags_internal.a", - "_deps/abseil_cpp-build/absl/synchronization/libabsl_kernel_timeout_internal.a", - "_deps/abseil_cpp-build/absl/synchronization/libabsl_graphcycles_internal.a", - "_deps/abseil_cpp-build/absl/synchronization/libabsl_synchronization.a", - "_deps/abseil_cpp-build/absl/hash/libabsl_city.a", - "_deps/abseil_cpp-build/absl/hash/libabsl_hash.a", - "_deps/abseil_cpp-build/absl/hash/libabsl_low_level_hash.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_failure_signal_handler.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_utf8_for_code_point.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_decode_rust_punycode.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_debugging_internal.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_symbolize.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_stacktrace.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_demangle_rust.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_demangle_internal.a", - "_deps/abseil_cpp-build/absl/debugging/libabsl_examine_stack.a", - "_deps/abseil_cpp-build/absl/crc/libabsl_crc_cord_state.a", - "_deps/abseil_cpp-build/absl/crc/libabsl_crc32c.a", - "_deps/abseil_cpp-build/absl/crc/libabsl_crc_internal.a", - "_deps/abseil_cpp-build/absl/crc/libabsl_crc_cpu_detect.a", - "_deps/abseil_cpp-build/absl/time/libabsl_time_zone.a", - "_deps/abseil_cpp-build/absl/time/libabsl_time.a", - "_deps/abseil_cpp-build/absl/time/libabsl_civil_time.a", - "_deps/abseil_cpp-build/absl/container/libabsl_raw_hash_set.a", - "_deps/abseil_cpp-build/absl/container/libabsl_hashtablez_sampler.a", - "_deps/abseil_cpp-build/absl/numeric/libabsl_int128.a", - "_deps/abseil_cpp-build/absl/profiling/libabsl_exponential_biased.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_conditions.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_nullguard.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_log_sink_set.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_format.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_fnmatch.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_globals.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_sink.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_check_op.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_message.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_globals.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_internal_proto.a", - "_deps/abseil_cpp-build/absl/log/libabsl_log_entry.a", - "_deps/abseil_cpp-build/absl/log/libabsl_vlog_config_internal.a", - "_deps/abseil_cpp-build/absl/base/libabsl_spinlock_wait.a", - "_deps/abseil_cpp-build/absl/base/libabsl_log_severity.a", - "_deps/abseil_cpp-build/absl/base/libabsl_raw_logging_internal.a", - "_deps/abseil_cpp-build/absl/base/libabsl_base.a", - "_deps/abseil_cpp-build/absl/base/libabsl_throw_delegate.a", - "_deps/abseil_cpp-build/absl/base/libabsl_malloc_internal.a", - "_deps/abseil_cpp-build/absl/base/libabsl_strerror.a", - "_deps/google_nsync-build/libnsync_cpp.a", - "_deps/protobuf-build/libprotobuf.a", - "_deps/pytorch_cpuinfo-build/libcpuinfo.a", - "_deps/pytorch_clog-build/libclog.a", - # "_deps/opencv-build/lib/libopencv_imgcodecs.a", - # "_deps/opencv-build/lib/libopencv_imgproc.a", - # "_deps/opencv-build/lib/libopencv_core.a", - # "_deps/opencv-build/3rdparty/lib/liblibjpeg-turbo.a", - # "_deps/opencv-build/3rdparty/lib/liblibpng.a" - ], - postfix_script= select({ - ":with_cuda": __POSTFIX_WITH_CUDA, - "//conditions:default": __POSTFIX - }), -) - -cc_library( - name = "onnxruntime_lib", - linkopts = select({ - "@platforms//os:linux": ["-static-libstdc++", "-static-libgcc"], - "//conditions:default": [], - }), - deps = [ - ":hdrs", - ":onnxruntime", - # "@onnx_runtime_extensions//:operators", - ], - includes= ["onnxruntime/include/onnxruntime"], - alwayslink = True, - visibility = ["//visibility:public"] -) \ No newline at end of file diff --git a/third_party/onnxruntime/include/core/providers/custom_op_context.h b/third_party/onnxruntime/include/core/providers/custom_op_context.h new file mode 100644 index 00000000..b10126da --- /dev/null +++ b/third_party/onnxruntime/include/core/providers/custom_op_context.h @@ -0,0 +1,10 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#pragma once + +// CustomOpContext defines an interface allowing a custom op to access ep-specific resources. +struct CustomOpContext { + CustomOpContext() = default; + virtual ~CustomOpContext() {}; +}; \ No newline at end of file diff --git a/third_party/onnxruntime/include/core/providers/resource.h b/third_party/onnxruntime/include/core/providers/resource.h new file mode 100644 index 00000000..bd123e1c --- /dev/null +++ b/third_party/onnxruntime/include/core/providers/resource.h @@ -0,0 +1,14 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#pragma once + +enum ResourceOffset { + cpu_resource_offset = 0, + cuda_resource_offset = 10000, + dml_resource_offset = 20000, + rocm_resource_offset = 30000, + // offsets for other ort eps + custom_ep_resource_offset = 10000000, + // offsets for customized eps +}; \ No newline at end of file diff --git a/third_party/onnxruntime/include/coreml_provider_factory.h b/third_party/onnxruntime/include/coreml_provider_factory.h new file mode 100644 index 00000000..351eafc2 --- /dev/null +++ b/third_party/onnxruntime/include/coreml_provider_factory.h @@ -0,0 +1,90 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. +#pragma once + +#include "onnxruntime_c_api.h" + +// COREMLFlags are bool options we want to set for CoreML EP +// This enum is defined as bit flags, and cannot have negative value +// To generate an uint32_t coreml_flags for using with OrtSessionOptionsAppendExecutionProvider_CoreML below, +// uint32_t coreml_flags = 0; +// coreml_flags |= COREML_FLAG_USE_CPU_ONLY; +enum COREMLFlags { + COREML_FLAG_USE_NONE = 0x000, + + // Using CPU only in CoreML EP, this may decrease the perf but will provide + // reference output value without precision loss, which is useful for validation + COREML_FLAG_USE_CPU_ONLY = 0x001, + + // Enable CoreML EP on subgraph + COREML_FLAG_ENABLE_ON_SUBGRAPH = 0x002, + + // By default CoreML Execution provider will be enabled for all compatible Apple devices + // Enable this option will only enable CoreML EP for Apple devices with ANE (Apple Neural Engine) + // Please note, enable this option does not guarantee the entire model to be executed using ANE only + COREML_FLAG_ONLY_ENABLE_DEVICE_WITH_ANE = 0x004, + + // Only allow CoreML EP to take nodes with inputs with static shapes. By default it will also allow inputs with + // dynamic shapes. However, the performance may be negatively impacted if inputs have dynamic shapes. + COREML_FLAG_ONLY_ALLOW_STATIC_INPUT_SHAPES = 0x008, + + // Create an MLProgram. By default it will create a NeuralNetwork model. Requires Core ML 5 or later. + COREML_FLAG_CREATE_MLPROGRAM = 0x010, + + // https://developer.apple.com/documentation/coreml/mlcomputeunits?language=objc + // there are four compute units: + // MLComputeUnitsCPUAndNeuralEngine|MLComputeUnitsCPUAndGPU|MLComputeUnitsCPUOnly|MLComputeUnitsAll + // different CU will have different performance and power consumption + COREML_FLAG_USE_CPU_AND_GPU = 0x020, + // Keep COREML_FLAG_LAST at the end of the enum definition + // And assign the last COREMLFlag to it + COREML_FLAG_LAST = COREML_FLAG_USE_CPU_AND_GPU, +}; + +// MLComputeUnits can be one of the following values: +// 'MLComputeUnitsCPUAndNeuralEngine|MLComputeUnitsCPUAndGPU|MLComputeUnitsCPUOnly|MLComputeUnitsAll' +// these values are intended to be used with Ort::SessionOptions::AppendExecutionProvider (C++ API) +// and SessionOptionsAppendExecutionProvider (C API). For the old API, use COREMLFlags instead. +static const char* const kCoremlProviderOption_MLComputeUnits = "MLComputeUnits"; +static const char* const kCoremlProviderOption_ModelFormat = "ModelFormat"; +// same as COREML_FLAG_ONLY_ALLOW_STATIC_INPUT_SHAPES +static const char* const kCoremlProviderOption_RequireStaticInputShapes = "RequireStaticInputShapes"; +static const char* const kCoremlProviderOption_EnableOnSubgraphs = "EnableOnSubgraphs"; +// provided by https://developer.apple.com/documentation/coreml/mloptimizationhints-swift.struct/specializationstrategy-swift.property +// Core ML segments the model’s compute graph and specializes each segment for the target compute device. +// This process can affect the model loading time and the prediction latency. +// Use this option to tailor the specialization strategy for your model. +static const char* const kCoremlProviderOption_SpecializationStrategy = "SpecializationStrategy"; +// Profile the Core ML MLComputePlan. +// This logs the hardware each operator is dispatched to and the estimated execution time. +// Intended for developer usage but provide useful diagnostic information if performance is not as expected. +static const char* const kCoremlProviderOption_ProfileComputePlan = "ProfileComputePlan"; +// please refer to https://developer.apple.com/documentation/coreml/mlmodelconfiguration/allowlowprecisionaccumulationongpu +static const char* const kCoremlProviderOption_AllowLowPrecisionAccumulationOnGPU = "AllowLowPrecisionAccumulationOnGPU"; +// Specify the directory to cache any CoreML models created from the ONNX model in. +// CoreML EP will convert onnx subgraph to CoreML model and save to disk. +// If this path is not specified, the model will be saved to a temp directory and deleted after the session is closed. +// otherwise, the model will be saved to the specified path and User should manage to delete the model. + +// we do NOT detect if the onnx model has changed and no longer matches the cached model. +// the user should carefully manage the cache if modifying/replacing a model. +// The cache key is generated by +// 1. User provided key in metadata_props if found (preferred) +// 2. Hash of the model url the inference session was created with +// 3. Hash of the input/output names of the model +// Please find out how to set metadata_props in the onnxruntime API documentation. https://onnxruntime.ai/docs/execution-providers/CoreML-ExecutionProvider.html#configuration-options +static const char* const kCoremlProviderOption_ModelCacheDirectory = "ModelCacheDirectory"; + +// User provided cache-key in metadata_props. +static const char* const kCOREML_CACHE_KEY = "COREML_CACHE_KEY"; + +#ifdef __cplusplus +extern "C" { +#endif + +ORT_EXPORT ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CoreML, + _In_ OrtSessionOptions* options, uint32_t coreml_flags); + +#ifdef __cplusplus +} +#endif diff --git a/third_party/onnxruntime/include/cpu_provider_factory.h b/third_party/onnxruntime/include/cpu_provider_factory.h new file mode 100644 index 00000000..29267869 --- /dev/null +++ b/third_party/onnxruntime/include/cpu_provider_factory.h @@ -0,0 +1,19 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#include "onnxruntime_c_api.h" + +#ifdef __cplusplus +extern "C" { +#endif + +/** + * \param use_arena zero: false. non-zero: true. + */ +ORT_EXPORT +ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CPU, _In_ OrtSessionOptions* options, int use_arena) +ORT_ALL_ARGS_NONNULL; + +#ifdef __cplusplus +} +#endif diff --git a/third_party/onnxruntime/include/onnxruntime_c_api.h b/third_party/onnxruntime/include/onnxruntime_c_api.h new file mode 100644 index 00000000..b01b7e48 --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_c_api.h @@ -0,0 +1,6195 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +// See docs\c_cxx\README.md on generating the Doxygen documentation from this file + +/** \mainpage ONNX Runtime + * + * ONNX Runtime is a high-performance inference and training graph execution engine for deep learning models. + * + * ONNX Runtime's C, C++ APIs offer an easy to use interface to onboard and execute onnx models. + * - \subpage c_cpp_api "Core C, C++ APIs" + * - \subpage training_c_cpp_api "Training C, C++ APIs for on-device training" + * + * \page c_cpp_api Core C, C++ APIs + *

C

+ * + * ::OrtApi - Click here to go to the structure with all C API functions. + * + *

C++

+ * + * ::Ort - Click here to go to the namespace holding all of the C++ wrapper classes + * + * It is a set of header only wrapper classes around the C API. The goal is to turn the C style return value error codes into C++ exceptions, and to + * automate memory management through standard C++ RAII principles. + * + * \addtogroup Global + * ONNX Runtime C API + * @{ + */ + +#pragma once +#include +#include +#include +#include + +/** \brief The API version defined in this header + * + * This value is used by some API functions to behave as this version of the header expects. + */ +#define ORT_API_VERSION 22 + +#ifdef __cplusplus +extern "C" { +#endif + +//! @} +// SAL2 Definitions +#ifndef _MSC_VER +#define _In_ +#define _In_z_ +#define _In_opt_ +#define _In_opt_z_ +#define _Out_ +#define _Outptr_ +#define _Out_opt_ +#define _Inout_ +#define _Inout_opt_ +#define _Frees_ptr_opt_ +#define _Ret_maybenull_ +#define _Ret_notnull_ +#define _Check_return_ +#define _Outptr_result_maybenull_ +#define _In_reads_(X) +#define _Inout_updates_(X) +#define _Out_writes_(X) +#define _Inout_updates_all_(X) +#define _Out_writes_bytes_all_(X) +#define _Out_writes_all_(X) +#define _Success_(X) +#define _Outptr_result_buffer_maybenull_(X) +#define ORT_ALL_ARGS_NONNULL __attribute__((nonnull)) +#else +#include +#define ORT_ALL_ARGS_NONNULL +#endif + +#ifdef _WIN32 +// Define ORT_DLL_IMPORT if your program is dynamically linked to Ort. +// dllexport is not used, we use a .def file. +#ifdef ORT_DLL_IMPORT +#define ORT_EXPORT __declspec(dllimport) +#else +#define ORT_EXPORT +#endif +#define ORT_API_CALL _stdcall +#define ORT_MUST_USE_RESULT +#define ORTCHAR_T wchar_t +#else +// To make symbols visible on macOS/iOS +#ifdef __APPLE__ +#define ORT_EXPORT __attribute__((visibility("default"))) +#else +#define ORT_EXPORT +#endif +#define ORT_API_CALL +#define ORT_MUST_USE_RESULT __attribute__((warn_unused_result)) +#define ORTCHAR_T char +#endif + +/// ORTCHAR_T, ORT_TSTR are reserved specifically for path handling. +/// All other strings are UTF-8 encoded, use char and std::string +#ifndef ORT_TSTR +#ifdef _WIN32 +#define ORT_TSTR(X) L##X +// When X is a macro, L##X is not defined. In this case, we need to use ORT_TSTR_ON_MACRO. +#define ORT_TSTR_ON_MACRO(X) L"" X +#else +#define ORT_TSTR(X) X +#define ORT_TSTR_ON_MACRO(X) X +#endif +#endif + +// On Windows, ORT_FILE is a wchar_t version of the __FILE__ macro. +// Otherwise, ORT_FILE is equivalent to __FILE__. +#ifndef ORT_FILE +#define ORT_FILE_INTERNAL(x) ORT_TSTR(x) +#define ORT_FILE ORT_FILE_INTERNAL(__FILE__) +#endif + +// Any pointer marked with _In_ or _Out_, cannot be NULL. + +// Windows users should use unicode paths when possible to bypass the MAX_PATH limitation +// Every pointer marked with _In_ or _Out_, cannot be NULL. Caller should ensure that. +// for ReleaseXXX(...) functions, they can accept NULL pointer. + +#ifdef __cplusplus +// For any compiler with C++11 support, MSVC 2015 and greater, or Clang version supporting noexcept. +// Such complex condition is needed because compilers set __cplusplus value differently. +#ifndef __has_feature +#define __has_feature(x) 0 +#endif +#if ((__cplusplus >= 201103L) || (_MSC_VER >= 1900) || (defined(__has_feature) && __has_feature(cxx_noexcept))) +#define NO_EXCEPTION noexcept +#else +#define NO_EXCEPTION throw() +#endif +#else +#define NO_EXCEPTION +#endif + +// __VA_ARGS__ on Windows and Linux are different +#define ORT_API(RETURN_TYPE, NAME, ...) RETURN_TYPE ORT_API_CALL NAME(__VA_ARGS__) NO_EXCEPTION + +#define ORT_API_STATUS(NAME, ...) \ + _Success_(return == 0) _Check_return_ _Ret_maybenull_ OrtStatusPtr ORT_API_CALL NAME(__VA_ARGS__) \ + NO_EXCEPTION ORT_MUST_USE_RESULT + +// XXX: Unfortunately, SAL annotations are known to not work with function pointers +#define ORT_API2_STATUS(NAME, ...) \ + _Check_return_ _Ret_maybenull_ OrtStatusPtr(ORT_API_CALL* NAME)(__VA_ARGS__) NO_EXCEPTION ORT_MUST_USE_RESULT + +// Used in *.cc files. Almost as same as ORT_API_STATUS, except without ORT_MUST_USE_RESULT and ORT_EXPORT +#define ORT_API_STATUS_IMPL(NAME, ...) \ + _Success_(return == 0) _Check_return_ _Ret_maybenull_ OrtStatusPtr ORT_API_CALL NAME(__VA_ARGS__) NO_EXCEPTION + +#define ORT_CLASS_RELEASE(X) void(ORT_API_CALL * Release##X)(_Frees_ptr_opt_ Ort##X * input) + +#ifdef __DOXYGEN__ +#undef ORT_API_STATUS +#define ORT_API_STATUS(NAME, ...) OrtStatus* NAME(__VA_ARGS__) +#undef ORT_API2_STATUS +#define ORT_API2_STATUS(NAME, ...) OrtStatus* NAME(__VA_ARGS__) +#undef ORT_CLASS_RELEASE +#define ORT_CLASS_RELEASE(X) void Release##X(Ort##X* input) +#undef NO_EXCEPTION +#define NO_EXCEPTION +#endif +/** \addtogroup Global + * ONNX Runtime C API + * @{ + */ + +/** Copied from TensorProto::DataType + * Currently, Ort doesn't support complex64, complex128 + */ +typedef enum ONNXTensorElementDataType { + ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED, + ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, // maps to c type float + ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8, // maps to c type uint8_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8, // maps to c type int8_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16, // maps to c type uint16_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16, // maps to c type int16_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32, // maps to c type int32_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64, // maps to c type int64_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING, // maps to c++ type std::string + ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL, + ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16, + ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE, // maps to c type double + ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32, // maps to c type uint32_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64, // maps to c type uint64_t + ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX64, // complex with float32 real and imaginary components + ONNX_TENSOR_ELEMENT_DATA_TYPE_COMPLEX128, // complex with float64 real and imaginary components + ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16, // Non-IEEE floating-point format based on IEEE754 single-precision + // float 8 types were introduced in onnx 1.14, see https://onnx.ai/onnx/technical/float8.html + ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FN, // Non-IEEE floating-point format based on IEEE754 single-precision + ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FNUZ, // Non-IEEE floating-point format based on IEEE754 single-precision + ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2, // Non-IEEE floating-point format based on IEEE754 single-precision + ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ, // Non-IEEE floating-point format based on IEEE754 single-precision + // Int4 types were introduced in ONNX 1.16. See https://onnx.ai/onnx/technical/int4.html + ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT4, // maps to a pair of packed uint4 values (size == 1 byte) + ONNX_TENSOR_ELEMENT_DATA_TYPE_INT4 // maps to a pair of packed int4 values (size == 1 byte) +} ONNXTensorElementDataType; + +// Synced with onnx TypeProto oneof +typedef enum ONNXType { + ONNX_TYPE_UNKNOWN, + ONNX_TYPE_TENSOR, + ONNX_TYPE_SEQUENCE, + ONNX_TYPE_MAP, + ONNX_TYPE_OPAQUE, + ONNX_TYPE_SPARSETENSOR, + ONNX_TYPE_OPTIONAL +} ONNXType; + +// These types are synced with internal +// SparseFormatFlags +typedef enum OrtSparseFormat { + ORT_SPARSE_UNDEFINED = 0, + ORT_SPARSE_COO = 0x1, + ORT_SPARSE_CSRC = 0x2, + ORT_SPARSE_BLOCK_SPARSE = 0x4 +} OrtSparseFormat; + +// Enum allows to query sparse tensor indices +enum OrtSparseIndicesFormat { + ORT_SPARSE_COO_INDICES, + ORT_SPARSE_CSR_INNER_INDICES, + ORT_SPARSE_CSR_OUTER_INDICES, + ORT_SPARSE_BLOCK_SPARSE_INDICES +}; + +/** \brief Logging severity levels + * + * In typical API usage, specifying a logging severity level specifies the minimum severity of log messages to show. + */ +typedef enum OrtLoggingLevel { + ORT_LOGGING_LEVEL_VERBOSE, ///< Verbose informational messages (least severe). + ORT_LOGGING_LEVEL_INFO, ///< Informational messages. + ORT_LOGGING_LEVEL_WARNING, ///< Warning messages. + ORT_LOGGING_LEVEL_ERROR, ///< Error messages. + ORT_LOGGING_LEVEL_FATAL, ///< Fatal error messages (most severe). +} OrtLoggingLevel; + +typedef enum OrtErrorCode { + ORT_OK, + ORT_FAIL, + ORT_INVALID_ARGUMENT, + ORT_NO_SUCHFILE, + ORT_NO_MODEL, + ORT_ENGINE_ERROR, + ORT_RUNTIME_EXCEPTION, + ORT_INVALID_PROTOBUF, + ORT_MODEL_LOADED, + ORT_NOT_IMPLEMENTED, + ORT_INVALID_GRAPH, + ORT_EP_FAIL, + ORT_MODEL_LOAD_CANCELED, + ORT_MODEL_REQUIRES_COMPILATION, +} OrtErrorCode; + +typedef enum OrtOpAttrType { + ORT_OP_ATTR_UNDEFINED = 0, + ORT_OP_ATTR_INT, + ORT_OP_ATTR_INTS, + ORT_OP_ATTR_FLOAT, + ORT_OP_ATTR_FLOATS, + ORT_OP_ATTR_STRING, + ORT_OP_ATTR_STRINGS, +} OrtOpAttrType; + +//! @} +#define ORT_RUNTIME_CLASS(X) \ + struct Ort##X; \ + typedef struct Ort##X Ort##X + +/** \addtogroup Global + * ONNX Runtime C API + * @{ + */ +// The actual types defined have an Ort prefix +ORT_RUNTIME_CLASS(Env); +ORT_RUNTIME_CLASS(Status); // nullptr for Status* indicates success +ORT_RUNTIME_CLASS(MemoryInfo); +ORT_RUNTIME_CLASS(IoBinding); +ORT_RUNTIME_CLASS(Session); // Don't call ReleaseSession from Dllmain (because session owns a thread pool) +ORT_RUNTIME_CLASS(Value); +ORT_RUNTIME_CLASS(RunOptions); +ORT_RUNTIME_CLASS(TypeInfo); +ORT_RUNTIME_CLASS(TensorTypeAndShapeInfo); +ORT_RUNTIME_CLASS(MapTypeInfo); +ORT_RUNTIME_CLASS(SequenceTypeInfo); +ORT_RUNTIME_CLASS(OptionalTypeInfo); +ORT_RUNTIME_CLASS(SessionOptions); +ORT_RUNTIME_CLASS(CustomOpDomain); +ORT_RUNTIME_CLASS(ModelMetadata); +ORT_RUNTIME_CLASS(ThreadPoolParams); +ORT_RUNTIME_CLASS(ThreadingOptions); +ORT_RUNTIME_CLASS(ArenaCfg); +ORT_RUNTIME_CLASS(PrepackedWeightsContainer); +ORT_RUNTIME_CLASS(TensorRTProviderOptionsV2); +ORT_RUNTIME_CLASS(NvTensorRtRtxProviderOptions); +ORT_RUNTIME_CLASS(CUDAProviderOptionsV2); +ORT_RUNTIME_CLASS(CANNProviderOptions); +ORT_RUNTIME_CLASS(DnnlProviderOptions); +ORT_RUNTIME_CLASS(Op); +ORT_RUNTIME_CLASS(OpAttr); +ORT_RUNTIME_CLASS(Logger); +ORT_RUNTIME_CLASS(ShapeInferContext); +ORT_RUNTIME_CLASS(LoraAdapter); +ORT_RUNTIME_CLASS(ValueInfo); +ORT_RUNTIME_CLASS(Node); +ORT_RUNTIME_CLASS(Graph); +ORT_RUNTIME_CLASS(Model); +ORT_RUNTIME_CLASS(ModelCompilationOptions); +ORT_RUNTIME_CLASS(HardwareDevice); +ORT_RUNTIME_CLASS(EpDevice); +ORT_RUNTIME_CLASS(KeyValuePairs); + +#ifdef _MSC_VER +typedef _Return_type_success_(return == 0) OrtStatus* OrtStatusPtr; +#else +typedef OrtStatus* OrtStatusPtr; +#endif + +/** \brief Memory allocation interface + * + * Structure of function pointers that defines a memory allocator. This can be created and filled in by the user for custom allocators. + * + * When an allocator is passed to any function, be sure that the allocator object is not destroyed until the last allocated object using it is freed. + */ +typedef struct OrtAllocator { + uint32_t version; ///< Must be initialized to ORT_API_VERSION + void*(ORT_API_CALL* Alloc)(struct OrtAllocator* this_, size_t size); ///< Returns a pointer to an allocated block of `size` bytes + void(ORT_API_CALL* Free)(struct OrtAllocator* this_, void* p); ///< Free a block of memory previously allocated with OrtAllocator::Alloc + const struct OrtMemoryInfo*(ORT_API_CALL* Info)(const struct OrtAllocator* this_); ///< Return a pointer to an ::OrtMemoryInfo that describes this allocator + /** + * @brief Optional allocation function to use for memory allocations made during session initialization. + * Use this function if you want to separate allocations made by ORT during Run() calls from + * those made during session initialization. This allows for separate memory management strategies for these allocations. + */ + void*(ORT_API_CALL* Reserve)(struct OrtAllocator* this_, size_t size); ///< Returns a pointer to an allocated block of `size` bytes +} OrtAllocator; + +typedef void(ORT_API_CALL* OrtLoggingFunction)( + void* param, OrtLoggingLevel severity, const char* category, const char* logid, const char* code_location, + const char* message); + +/** \brief Graph optimization level + * + * Refer to https://www.onnxruntime.ai/docs/performance/graph-optimizations.html#graph-optimization-levels + * for an in-depth understanding of the Graph Optimization Levels. + */ +typedef enum GraphOptimizationLevel { + ORT_DISABLE_ALL = 0, + ORT_ENABLE_BASIC = 1, + ORT_ENABLE_EXTENDED = 2, + ORT_ENABLE_ALL = 99 +} GraphOptimizationLevel; + +typedef enum ExecutionMode { + ORT_SEQUENTIAL = 0, + ORT_PARALLEL = 1, +} ExecutionMode; + +/** \brief Language projection identifiers + * /see OrtApi::SetLanguageProjection + */ +typedef enum OrtLanguageProjection { + ORT_PROJECTION_C = 0, + ORT_PROJECTION_CPLUSPLUS = 1, + ORT_PROJECTION_CSHARP = 2, + ORT_PROJECTION_PYTHON = 3, + ORT_PROJECTION_JAVA = 4, + ORT_PROJECTION_WINML = 5, + ORT_PROJECTION_NODEJS = 6, +} OrtLanguageProjection; + +struct OrtKernelInfo; +typedef struct OrtKernelInfo OrtKernelInfo; +struct OrtKernelContext; +typedef struct OrtKernelContext OrtKernelContext; +struct OrtCustomOp; +typedef struct OrtCustomOp OrtCustomOp; + +typedef enum OrtAllocatorType { + OrtInvalidAllocator = -1, + OrtDeviceAllocator = 0, + OrtArenaAllocator = 1 +} OrtAllocatorType; + +/** \brief Memory types for allocated memory, execution provider specific types should be extended in each provider. + */ +// Whenever this struct is updated, please also update the MakeKey function in onnxruntime / core / framework / execution_provider.cc +typedef enum OrtMemType { + OrtMemTypeCPUInput = -2, ///< Any CPU memory used by non-CPU execution provider + OrtMemTypeCPUOutput = -1, ///< CPU accessible memory outputted by non-CPU execution provider, i.e. CUDA_PINNED + OrtMemTypeCPU = OrtMemTypeCPUOutput, ///< Temporary CPU accessible memory allocated by non-CPU execution provider, i.e. CUDA_PINNED + OrtMemTypeDefault = 0, ///< The default allocator for execution provider +} OrtMemType; + +/** \brief This mimics OrtDevice type constants so they can be returned in the API + */ +typedef enum OrtMemoryInfoDeviceType { + OrtMemoryInfoDeviceType_CPU = 0, + OrtMemoryInfoDeviceType_GPU = 1, + OrtMemoryInfoDeviceType_FPGA = 2 +} OrtMemoryInfoDeviceType; + +typedef enum OrtHardwareDeviceType { + OrtHardwareDeviceType_CPU, + OrtHardwareDeviceType_GPU, + OrtHardwareDeviceType_NPU +} OrtHardwareDeviceType; + +/** \brief These are the default EP selection policies used by ORT when doing automatic EP selection. + */ +typedef enum OrtExecutionProviderDevicePolicy { + OrtExecutionProviderDevicePolicy_DEFAULT, + OrtExecutionProviderDevicePolicy_PREFER_CPU, + OrtExecutionProviderDevicePolicy_PREFER_NPU, + OrtExecutionProviderDevicePolicy_PREFER_GPU, + OrtExecutionProviderDevicePolicy_MAX_PERFORMANCE, + OrtExecutionProviderDevicePolicy_MAX_EFFICIENCY, + OrtExecutionProviderDevicePolicy_MIN_OVERALL_POWER, +} OrtExecutionProviderDevicePolicy; + +/** \brief Delegate to allow providing custom OrtEpDevice selection logic + * + * This delegate is called by the EP selection code to allow the user to provide custom device selection logic. + * The user can use this to select OrtEpDevice instances from the list of available devices. + * + * \param ep_devices The list of available devices. + * \param num_devices The number of available devices. + * \param model_metadata The model metadata. + * \param runtime_metadata The runtime metadata. May be nullptr. + * \param selected Pre-allocated array to populate with selected OrtEpDevice pointers from ep_devices. + * \param max_ep_devices The maximum number of devices that can be selected in the pre-allocated array. + Currently the maximum is 8. + * \param num_ep_devices The number of selected devices. + * \param state Opaque pointer. Required to use the delegate from other languages like C# and python. + * + * \return OrtStatus* Selection status. Return nullptr on success. + * Use CreateStatus to provide error info. Use ORT_FAIL as the error code. + * ORT will release the OrtStatus* if not null. + */ +typedef OrtStatus* (*EpSelectionDelegate)(_In_ const OrtEpDevice** ep_devices, + _In_ size_t num_devices, + _In_ const OrtKeyValuePairs* model_metadata, + _In_opt_ const OrtKeyValuePairs* runtime_metadata, + _Inout_ const OrtEpDevice** selected, + _In_ size_t max_selected, + _Out_ size_t* num_selected, + _In_ void* state); + +/** \brief Algorithm to use for cuDNN Convolution Op + */ +typedef enum OrtCudnnConvAlgoSearch { + OrtCudnnConvAlgoSearchExhaustive, // expensive exhaustive benchmarking using cudnnFindConvolutionForwardAlgorithmEx + OrtCudnnConvAlgoSearchHeuristic, // lightweight heuristic based search using cudnnGetConvolutionForwardAlgorithm_v7 + OrtCudnnConvAlgoSearchDefault, // default algorithm using CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM +} OrtCudnnConvAlgoSearch; + +/** \brief CUDA Provider Options + * + * \see OrtApi::SessionOptionsAppendExecutionProvider_CUDA + */ +typedef struct OrtCUDAProviderOptions { +#ifdef __cplusplus + OrtCUDAProviderOptions() + : device_id{}, + cudnn_conv_algo_search{OrtCudnnConvAlgoSearchExhaustive}, + gpu_mem_limit{SIZE_MAX}, + arena_extend_strategy{}, + do_copy_in_default_stream{1}, + has_user_compute_stream{}, + user_compute_stream{}, + default_memory_arena_cfg{}, + tunable_op_enable{false}, + tunable_op_tuning_enable{false}, + tunable_op_max_tuning_duration_ms{} {} +#endif + + /** \brief CUDA device Id + * Defaults to 0. + */ + int device_id; + + /** \brief CUDA Convolution algorithm search configuration. + * See enum OrtCudnnConvAlgoSearch for more details. + * Defaults to OrtCudnnConvAlgoSearchExhaustive. + */ + OrtCudnnConvAlgoSearch cudnn_conv_algo_search; + + /** \brief CUDA memory limit (To use all possible memory pass in maximum size_t) + * Defaults to SIZE_MAX. + * \note If a ::OrtArenaCfg has been applied, it will override this field + */ + size_t gpu_mem_limit; + + /** \brief Strategy used to grow the memory arena + * 0 = kNextPowerOfTwo
+ * 1 = kSameAsRequested
+ * Defaults to 0. + * \note If a ::OrtArenaCfg has been applied, it will override this field + */ + int arena_extend_strategy; + + /** \brief Flag indicating if copying needs to take place on the same stream as the compute stream in the CUDA EP + * 0 = Use separate streams for copying and compute. + * 1 = Use the same stream for copying and compute. + * Defaults to 1. + * WARNING: Setting this to 0 may result in data races for some models. + * Please see issue #4829 for more details. + */ + int do_copy_in_default_stream; + + /** \brief Flag indicating if there is a user provided compute stream + * Defaults to 0. + */ + int has_user_compute_stream; + + /** \brief User provided compute stream. + * If provided, please set `has_user_compute_stream` to 1. + */ + void* user_compute_stream; + + /** \brief CUDA memory arena configuration parameters + */ + OrtArenaCfg* default_memory_arena_cfg; + + /** \brief Enable TunableOp for using. + * Set it to 1/0 to enable/disable TunableOp. Otherwise, it is disabled by default. + * This option can be overridden by environment variable ORT_CUDA_TUNABLE_OP_ENABLE. + */ + int tunable_op_enable; + + /** \brief Enable TunableOp for tuning. + * Set it to 1/0 to enable/disable TunableOp tuning. Otherwise, it is disabled by default. + * This option can be overridden by environment variable ORT_CUDA_TUNABLE_OP_TUNING_ENABLE. + */ + int tunable_op_tuning_enable; + + /** \brief Max tuning duration time limit for each instance of TunableOp. + * Defaults to 0 to disable the limit. + */ + int tunable_op_max_tuning_duration_ms; + +} OrtCUDAProviderOptions; + +/** \brief ROCM Provider Options + * + * \see OrtApi::SessionOptionsAppendExecutionProvider_ROCM + */ +typedef struct OrtROCMProviderOptions { +#ifdef __cplusplus + OrtROCMProviderOptions() + : device_id{}, + miopen_conv_exhaustive_search{0}, + gpu_mem_limit{SIZE_MAX}, + arena_extend_strategy{}, + do_copy_in_default_stream{1}, + has_user_compute_stream{}, + user_compute_stream{}, + default_memory_arena_cfg{}, + enable_hip_graph{false}, + tunable_op_enable{false}, + tunable_op_tuning_enable{false}, + tunable_op_max_tuning_duration_ms{} {} +#endif + + /** \brief ROCM device Id + * Defaults to 0. + */ + int device_id; + + /** \brief ROCM MIOpen Convolution algorithm exaustive search option. + * Defaults to 0 (false). + */ + int miopen_conv_exhaustive_search; + + /** \brief ROCM memory limit (To use all possible memory pass in maximum size_t) + * Defaults to SIZE_MAX. + * \note If a ::OrtArenaCfg has been applied, it will override this field + */ + size_t gpu_mem_limit; + + /** \brief Strategy used to grow the memory arena + * 0 = kNextPowerOfTwo
+ * 1 = kSameAsRequested
+ * Defaults to 0. + * \note If a ::OrtArenaCfg has been applied, it will override this field + */ + int arena_extend_strategy; + + /** \brief Flag indicating if copying needs to take place on the same stream as the compute stream in the ROCM EP + * 0 = Use separate streams for copying and compute. + * 1 = Use the same stream for copying and compute. + * Defaults to 1. + * WARNING: Setting this to 0 may result in data races for some models. + * Please see issue #4829 for more details. + */ + int do_copy_in_default_stream; + + /** \brief Flag indicating if there is a user provided compute stream + * Defaults to 0. + */ + int has_user_compute_stream; + + /** \brief User provided compute stream. + * If provided, please set `has_user_compute_stream` to 1. + */ + void* user_compute_stream; + + /** \brief ROCM memory arena configuration parameters + */ + OrtArenaCfg* default_memory_arena_cfg; + + int enable_hip_graph; + + /** \brief Enable TunableOp for using. + * Set it to 1/0 to enable/disable TunableOp. Otherwise, it is disabled by default. + * This option can be overridden by environment variable ORT_ROCM_TUNABLE_OP_ENABLE. + */ + int tunable_op_enable; + + /** \brief Enable TunableOp for tuning. + * Set it to 1/0 to enable/disable TunableOp tuning. Otherwise, it is disabled by default. + * This option can be overridden by environment variable ORT_ROCM_TUNABLE_OP_TUNING_ENABLE. + */ + int tunable_op_tuning_enable; + + /** \brief Max tuning duration time limit for each instance of TunableOp. + * Defaults to 0 to disable the limit. + */ + int tunable_op_max_tuning_duration_ms; + +} OrtROCMProviderOptions; + +/** \brief TensorRT Provider Options + * + * \see OrtApi::SessionOptionsAppendExecutionProvider_TensorRT + */ +typedef struct OrtTensorRTProviderOptions { + int device_id; ///< CUDA device id (0 = default device) + int has_user_compute_stream; // indicator of user specified CUDA compute stream. + void* user_compute_stream; // user specified CUDA compute stream. + int trt_max_partition_iterations; // maximum iterations for TensorRT parser to get capability + int trt_min_subgraph_size; // minimum size of TensorRT subgraphs + size_t trt_max_workspace_size; // maximum workspace size for TensorRT. + int trt_fp16_enable; // enable TensorRT FP16 precision. Default 0 = false, nonzero = true + int trt_int8_enable; // enable TensorRT INT8 precision. Default 0 = false, nonzero = true + const char* trt_int8_calibration_table_name; // TensorRT INT8 calibration table name. + int trt_int8_use_native_calibration_table; // use native TensorRT generated calibration table. Default 0 = false, nonzero = true + int trt_dla_enable; // enable DLA. Default 0 = false, nonzero = true + int trt_dla_core; // DLA core number. Default 0 + int trt_dump_subgraphs; // dump TRT subgraph. Default 0 = false, nonzero = true + int trt_engine_cache_enable; // enable engine caching. Default 0 = false, nonzero = true + const char* trt_engine_cache_path; // specify engine cache path + int trt_engine_decryption_enable; // enable engine decryption. Default 0 = false, nonzero = true + const char* trt_engine_decryption_lib_path; // specify engine decryption library path + int trt_force_sequential_engine_build; // force building TensorRT engine sequentially. Default 0 = false, nonzero = true + // This is the legacy struct and don't add new fields here. + // For new field that can be represented by string, please add it in include/onnxruntime/core/providers/tensorrt/tensorrt_provider_options.h + // For non-string field, need to create a new separate api to handle it. +} OrtTensorRTProviderOptions; + +/** \brief MIGraphX Provider Options + * + * \see OrtApi::SessionOptionsAppendExecutionProvider_MIGraphX + */ +typedef struct OrtMIGraphXProviderOptions { + int device_id; // hip device id. + int migraphx_fp16_enable; // MIGraphX FP16 precision. Default 0 = false, nonzero = true + int migraphx_int8_enable; // MIGraphX INT8 precision. Default 0 = false, nonzero = true + int migraphx_use_native_calibration_table; // MIGraphx INT8 cal table. Default 0 = false, noznero = true + const char* migraphx_int8_calibration_table_name; // MIGraphx INT8 calibration table name + int migraphx_save_compiled_model; // migraphx save compiled model. Default 0 = false, noznero = true + const char* migraphx_save_model_path; // migraphx model path name + int migraphx_load_compiled_model; // migraphx int8 cal table. Default 0 = false, noznero = true + const char* migraphx_load_model_path; // migraphx model path name + bool migraphx_exhaustive_tune; // migraphx tuned compile Default = false +} OrtMIGraphXProviderOptions; + +/** \brief OpenVINO Provider Options + * \brief This Struct is frozen since ORT 1.13.0. Its maintained part of Legacy API for compatibility. + * \brief For latest OpenVINO Provider Options update to the ProviderOptions map. + * \brief Latest OpenVINO Provider Options are listed in the + * \htmlonly + * onnxruntime document. + * \endhtmlonly + * \see OrtApi::SessionOptionsAppendExecutionProvider() + */ +typedef struct OrtOpenVINOProviderOptions { +#ifdef __cplusplus + OrtOpenVINOProviderOptions() : device_type{}, + enable_npu_fast_compile{}, + device_id{}, + num_of_threads{}, + cache_dir{}, + context{}, + enable_opencl_throttling{}, + enable_dynamic_shapes{} {} +#endif + /** \brief Device type string + * + * Valid settings are one of: "CPU_FP32", "CPU_FP16", "GPU_FP32", "GPU_FP16" + */ + const char* device_type; + unsigned char enable_npu_fast_compile; ///< 0 = disabled, nonzero = enabled + const char* device_id; + size_t num_of_threads; ///< 0 = Use default number of threads + const char* cache_dir; // path is set to empty by default + void* context; + unsigned char enable_opencl_throttling; ///< 0 = disabled, nonzero = enabled + unsigned char enable_dynamic_shapes; ///< 0 = disabled, nonzero = enabled +} OrtOpenVINOProviderOptions; + +struct OrtApi; +typedef struct OrtApi OrtApi; + +struct OrtTrainingApi; +typedef struct OrtTrainingApi OrtTrainingApi; + +struct OrtModelEditorApi; +typedef struct OrtModelEditorApi OrtModelEditorApi; + +struct OrtCompileApi; +typedef struct OrtCompileApi OrtCompileApi; + +struct OrtEpApi; +typedef struct OrtEpApi OrtEpApi; + +/** \brief The helper interface to get the right version of OrtApi + * + * Get a pointer to this structure through ::OrtGetApiBase + */ +struct OrtApiBase { + /** \brief Get a pointer to the requested version of the ::OrtApi + * + * \param[in] version Must be ::ORT_API_VERSION + * \return The ::OrtApi for the version requested, nullptr will be returned if this version is unsupported, for example when using a runtime + * older than the version created with this header file. + * + * One can call GetVersionString() to get the version of the Onnxruntime library for logging + * and error reporting purposes. + */ + const OrtApi*(ORT_API_CALL* GetApi)(uint32_t version)NO_EXCEPTION; + + /** \brief Returns a null terminated string of the version of the Onnxruntime library (eg: "1.8.1") + * + * \return UTF-8 encoded version string. Do not deallocate the returned buffer. + */ + const char*(ORT_API_CALL* GetVersionString)(void)NO_EXCEPTION; +}; + +typedef struct OrtApiBase OrtApiBase; + +/** \brief The Onnxruntime library's entry point to access the C API + * + * Call this to get the a pointer to an ::OrtApiBase + */ +ORT_EXPORT const OrtApiBase* ORT_API_CALL OrtGetApiBase(void) NO_EXCEPTION; + +/** \brief Thread work loop function + * + * Onnxruntime will provide the working loop on custom thread creation + * Argument is an onnxruntime built-in type which will be provided when thread pool calls OrtCustomCreateThreadFn + */ +typedef void (*OrtThreadWorkerFn)(void* ort_worker_fn_param); + +typedef const struct OrtCustomHandleType { + char __place_holder; +}* OrtCustomThreadHandle; + +/** \brief Ort custom thread creation function + * + * The function should return a thread handle to be used in onnxruntime thread pools + * Onnxruntime will throw exception on return value of nullptr or 0, indicating that the function failed to create a thread + */ +typedef OrtCustomThreadHandle (*OrtCustomCreateThreadFn)(void* ort_custom_thread_creation_options, OrtThreadWorkerFn ort_thread_worker_fn, void* ort_worker_fn_param); + +/** \brief Custom thread join function + * + * Onnxruntime thread pool destructor will call the function to join a custom thread. + * Argument ort_custom_thread_handle is the value returned by OrtCustomCreateThreadFn + */ +typedef void (*OrtCustomJoinThreadFn)(OrtCustomThreadHandle ort_custom_thread_handle); + +typedef OrtStatus*(ORT_API_CALL* RegisterCustomOpsFn)(OrtSessionOptions* options, const OrtApiBase* api); + +/** \brief Callback function for RunAsync + * + * \param[in] user_data User specific data that passed back to the callback + * \param[out] outputs On succeed, outputs host inference results, on error, the value will be nullptr + * \param[out] num_outputs Number of outputs, on error, the value will be zero + * \param[out] status On error, status will provide details + */ +typedef void (*RunAsyncCallbackFn)(void* user_data, OrtValue** outputs, size_t num_outputs, OrtStatusPtr status); + +/** \brief The C API + * + * All C API functions are defined inside this structure as pointers to functions. + * Call OrtApiBase::GetApi to get a pointer to it + * + * \nosubgrouping + */ +struct OrtApi { + /// \name OrtStatus + /// @{ + + /** + * \brief Create an OrtStatus from a null terminated string + * + * \param[in] code + * \param[in] msg A null-terminated string. Its contents will be copied. + * \return A new OrtStatus object, must be destroyed with OrtApi::ReleaseStatus + */ + OrtStatus*(ORT_API_CALL* CreateStatus)(OrtErrorCode code, _In_ const char* msg)NO_EXCEPTION ORT_ALL_ARGS_NONNULL; + + /** \brief Get OrtErrorCode from OrtStatus + * + * \param[in] status + * \return OrtErrorCode that \p status was created with + */ + OrtErrorCode(ORT_API_CALL* GetErrorCode)(_In_ const OrtStatus* status) NO_EXCEPTION ORT_ALL_ARGS_NONNULL; + + /** \brief Get error string from OrtStatus + * + * \param[in] status + * \return The error message inside the `status`. Do not free the returned value. + */ + const char*(ORT_API_CALL* GetErrorMessage)(_In_ const OrtStatus* status)NO_EXCEPTION ORT_ALL_ARGS_NONNULL; + + /// @} + /// \name OrtEnv + /// @{ + + /** \brief Create an OrtEnv + * + * \note Invoking this function will return the same instance of the environment as that returned by a previous call + * to another env creation function; all arguments to this function will be ignored. + * \param[in] log_severity_level The log severity level. + * \param[in] logid The log identifier. + * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateEnv, OrtLoggingLevel log_severity_level, _In_ const char* logid, _Outptr_ OrtEnv** out); + + /** \brief Create an OrtEnv + * + * \note Invoking this function will return the same instance of the environment as that returned by a previous call + * to another env creation function; all arguments to this function will be ignored. If you want to provide your + * own logging function, consider setting it using the SetUserLoggingFunction API instead. + * \param[in] logging_function A pointer to a logging function. + * \param[in] logger_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to + * `logging_function`. This parameter is optional. + * \param[in] log_severity_level The log severity level. + * \param[in] logid The log identifier. + * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateEnvWithCustomLogger, _In_ OrtLoggingFunction logging_function, _In_opt_ void* logger_param, + _In_ OrtLoggingLevel log_severity_level, _In_ const char* logid, _Outptr_ OrtEnv** out); + + /** \brief Enable Telemetry + * + * \note Telemetry events are on by default since they are lightweight + * \param[in] env + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(EnableTelemetryEvents, _In_ const OrtEnv* env); + /** \brief Disable Telemetry + * + * \see OrtApi::EnableTelemetryEvents + * \param[in] env + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(DisableTelemetryEvents, _In_ const OrtEnv* env); + + /// @} + /// \name OrtSession + /// @{ + + /** \brief Create an OrtSession from a model file + * + * \param[in] env + * \param[in] model_path + * \param[in] options + * \param[out] out Returned newly created OrtSession. Must be freed with OrtApi::ReleaseSession + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + // TODO: document the path separator convention? '/' vs '\' + // TODO: should specify the access characteristics of model_path. Is this read only during the + // execution of CreateSession, or does the OrtSession retain a handle to the file/directory + // and continue to access throughout the OrtSession lifetime? + // What sort of access is needed to model_path : read or read/write? + ORT_API2_STATUS(CreateSession, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path, + _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out); + + /** \brief Create an OrtSession from memory + * + * \param[in] env + * \param[in] model_data + * \param[in] model_data_length + * \param[in] options + * \param[out] out Returned newly created OrtSession. Must be freed with OrtApi::ReleaseSession + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateSessionFromArray, _In_ const OrtEnv* env, + _In_ const void* model_data, size_t model_data_length, + _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out); + + /** \brief Run the model in an ::OrtSession + * + * Will not return until the model run has completed. Multiple threads might be used to run the model based on + * the options in the ::OrtSession and settings used when creating the ::OrtEnv + * + * \param[in] session + * \param[in] run_options If nullptr, will use a default ::OrtRunOptions + * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names + * \param[in] inputs Array of ::OrtValue%s of the input values + * \param[in] input_len Number of elements in the input_names and inputs arrays + * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names + * \param[in] output_names_len Number of elements in the output_names and outputs array + * \param[out] outputs Array of ::OrtValue%s that the outputs are stored in. This can also be + * an array of nullptr values, in this case ::OrtValue objects will be allocated and pointers + * to them will be set into the `outputs` array. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(Run, _Inout_ OrtSession* session, _In_opt_ const OrtRunOptions* run_options, + _In_reads_(input_len) const char* const* input_names, + _In_reads_(input_len) const OrtValue* const* inputs, size_t input_len, + _In_reads_(output_names_len) const char* const* output_names, size_t output_names_len, + _Inout_updates_all_(output_names_len) OrtValue** outputs); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Create an ::OrtSessionOptions object + * + * To use additional providers, you must build ORT with the extra providers enabled. Then call one of these + * functions to enable them in the session:
+ * OrtSessionOptionsAppendExecutionProvider_CPU
+ * OrtSessionOptionsAppendExecutionProvider_CUDA
+ * OrtSessionOptionsAppendExecutionProvider_(remaining providers...)
+ * The order they are called indicates the preference order as well. In other words call this method + * on your most preferred execution provider first followed by the less preferred ones. + * If none are called Ort will use its internal CPU execution provider. + * + * \param[out] options The newly created OrtSessionOptions. Must be freed with OrtApi::ReleaseSessionOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateSessionOptions, _Outptr_ OrtSessionOptions** options); + + /** \brief Set filepath to save optimized model after graph level transformations + * + * \param[in] options + * \param[in] optimized_model_filepath + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetOptimizedModelFilePath, _Inout_ OrtSessionOptions* options, + _In_ const ORTCHAR_T* optimized_model_filepath); + + /** \brief Create a copy of an existing ::OrtSessionOptions + * + * \param[in] in_options OrtSessionOptions to copy + * \param[out] out_options Returned newly created ::OrtSessionOptions. Must be freed with OrtApi::ReleaseSessionOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CloneSessionOptions, _In_ const OrtSessionOptions* in_options, + _Outptr_ OrtSessionOptions** out_options); + + /** \brief Set execution mode + * + * Controls whether you want to execute operators in your graph sequentially or in parallel. Usually when the model + * has many branches, setting this option to ExecutionMode.ORT_PARALLEL will give you better performance. + * See [docs/ONNX_Runtime_Perf_Tuning.md] for more details. + * + * \param[in] options + * \param[in] execution_mode + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetSessionExecutionMode, _Inout_ OrtSessionOptions* options, ExecutionMode execution_mode); + + /** \brief Enable profiling for a session + * + * \param[in] options + * \param[in] profile_file_prefix + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(EnableProfiling, _Inout_ OrtSessionOptions* options, _In_ const ORTCHAR_T* profile_file_prefix); + + /** \brief Disable profiling for a session + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(DisableProfiling, _Inout_ OrtSessionOptions* options); + + /** \brief Enable the memory pattern optimization + * + * The idea is if the input shapes are the same, we could trace the internal memory allocation + * and generate a memory pattern for future request. So next time we could just do one allocation + * with a big chunk for all the internal memory allocation. + * \note Memory pattern optimization is only available when Sequential Execution mode is enabled (see OrtApi::SetSessionExecutionMode) + * + * \see OrtApi::DisableMemPattern + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(EnableMemPattern, _Inout_ OrtSessionOptions* options); + + /** \brief Disable the memory pattern optimization + * + * \see OrtApi::EnableMemPattern + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(DisableMemPattern, _Inout_ OrtSessionOptions* options); + + /** \brief Enable the memory arena on CPU + * + * Arena may pre-allocate memory for future usage. + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(EnableCpuMemArena, _Inout_ OrtSessionOptions* options); + + /** \brief Disable the memory arena on CPU + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(DisableCpuMemArena, _Inout_ OrtSessionOptions* options); + + /** \brief Set session log id + * + * \param[in] options + * \param[in] logid The log identifier. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetSessionLogId, _Inout_ OrtSessionOptions* options, const char* logid); + + /** \brief Set session log verbosity level + * + * Applies to session load, initialization, etc + * + * \param[in] options + * \param[in] session_log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetSessionLogVerbosityLevel, _Inout_ OrtSessionOptions* options, int session_log_verbosity_level); + + /** \brief Set session log severity level + * + * \param[in] options + * \param[in] session_log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values). + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetSessionLogSeverityLevel, _Inout_ OrtSessionOptions* options, int session_log_severity_level); + + /** \brief Set the optimization level to apply when loading a graph + * + * Please see https://onnxruntime.ai/docs/performance/model-optimizations/graph-optimizations.html for an in-depth explanation + * \param[in,out] options The session options object + * \param[in] graph_optimization_level The optimization level + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetSessionGraphOptimizationLevel, _Inout_ OrtSessionOptions* options, + GraphOptimizationLevel graph_optimization_level); + + /** \brief Sets the number of threads used to parallelize the execution within nodes + * + * When running a single node operation, ex. add, this sets the maximum number of threads to use. + * + * \note If built with OpenMP, this has no effect on the number of threads used. In this case + * use the OpenMP env variables to configure the number of intra op num threads. + * + * \param[in] options + * \param[in] intra_op_num_threads Number of threads to use
+ * A value of 0 will use the default number of threads
+ * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetIntraOpNumThreads, _Inout_ OrtSessionOptions* options, int intra_op_num_threads); + + /** \brief Sets the number of threads used to parallelize the execution of the graph + * + * If nodes can be run in parallel, this sets the maximum number of threads to use to run them in parallel. + * + * \note If sequential execution is enabled this value is ignored, it acts as if it was set to 1. + * + * \param[in] options + * \param[in] inter_op_num_threads Number of threads to use
+ * A value of 0 will use the default number of threads
+ * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetInterOpNumThreads, _Inout_ OrtSessionOptions* options, int inter_op_num_threads); + + /// @} + /// \name OrtCustomOpDomain + /// @{ + + /** \brief Create a custom op domain + * + * \param[in] domain + * \param[out] out Newly created domain. Must be freed with OrtApi::ReleaseCustomOpDomain + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateCustomOpDomain, _In_ const char* domain, _Outptr_ OrtCustomOpDomain** out); + + /** \brief Add a custom op to a custom op domain + * + * \note The OrtCustomOp* pointer must remain valid until the ::OrtCustomOpDomain using it is released + * + * \param[in] custom_op_domain + * \param[in] op + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CustomOpDomain_Add, _Inout_ OrtCustomOpDomain* custom_op_domain, _In_ const OrtCustomOp* op); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Add custom op domain to a session options + * + * \note The OrtCustomOpDomain* must not be deleted until all sessions using it are released + * + * \param[in] options + * \param[in] custom_op_domain + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(AddCustomOpDomain, _Inout_ OrtSessionOptions* options, _In_ OrtCustomOpDomain* custom_op_domain); + + /** \deprecated Use OrtApi::RegisterCustomOpsLibrary_V2. + * + * Registers custom ops from a shared library. + * + * Loads a shared library (dll on windows, so on linux, etc) named 'library_path' and looks for this entry point: + * OrtStatus* RegisterCustomOps(OrtSessionOptions * options, const OrtApiBase* api); + * It then passes in the provided session options to this function along with the api base. + * The handle to the loaded library is returned in library_handle. It can be freed by the caller after all sessions using the passed in + * session options are destroyed, or if an error occurs and it is non null. + * + * \param[in] options + * \param[in] library_path + * \param[out] library_handle OS specific handle to the loaded library (Use FreeLibrary on Windows, dlclose on Linux, etc.. to unload) + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RegisterCustomOpsLibrary, _Inout_ OrtSessionOptions* options, _In_ const char* library_path, _Outptr_ void** library_handle); + + /// @} + /// \name OrtSession + /// @{ + + /** \brief Get input count for a session + * + * This number must also match the number of inputs passed to OrtApi::Run + * + * \see OrtApi::SessionGetInputTypeInfo, OrtApi::SessionGetInputName, OrtApi::Session + * + * \param[in] session + * \param[out] out Number of inputs + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetInputCount, _In_ const OrtSession* session, _Out_ size_t* out); + + /** \brief Get output count for a session + * + * This number must also match the number of outputs returned by OrtApi::Run + * + * \see OrtApi::SessionGetOutputTypeInfo, OrtApi::SessionGetOutputName, OrtApi::Session + * + * \param[in] session + * \param[out] out Number of outputs + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetOutputCount, _In_ const OrtSession* session, _Out_ size_t* out); + + /** \brief Get overridable initializer count + * + * \see OrtApi::SessionGetOverridableInitializerTypeInfo, OrtApi::SessionGetOverridableInitializerName + * + * \param[in] session + * \param[in] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetOverridableInitializerCount, _In_ const OrtSession* session, _Out_ size_t* out); + + /** \brief Get input type information + * + * \param[in] session + * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetInputCount returns (exclusive) + * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetInputTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info); + + /** \brief Get output type information + * + * \param[in] session + * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOutputCount returns (exclusive) + * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetOutputTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info); + + /** \brief Get overridable initializer type information + * + * \param[in] session + * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOverridableInitializerCount returns (exclusive) + * \param[out] type_info Must be freed with OrtApi::ReleaseTypeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetOverridableInitializerTypeInfo, _In_ const OrtSession* session, size_t index, _Outptr_ OrtTypeInfo** type_info); + + /** \brief Get input name + * + * \param[in] session + * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetInputCount returns (exclusive) + * \param[in] allocator + * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetInputName, _In_ const OrtSession* session, size_t index, _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /** \brief Get output name + * + * \param[in] session + * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOutputCount returns (exclusive) + * \param[in] allocator + * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetOutputName, _In_ const OrtSession* session, size_t index, _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /** \brief Get overridable initializer name + * + * \param[in] session + * \param[in] index Must be between 0 (inclusive) and what OrtApi::SessionGetOverridableInitializerCount returns (exclusive) + * \param[in] allocator + * \param[out] value Set to a null terminated UTF-8 encoded string allocated using `allocator`. Must be freed using `allocator`. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetOverridableInitializerName, _In_ const OrtSession* session, size_t index, + _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /// @} + /// \name OrtRunOptions + /// @{ + + /** \brief Create an OrtRunOptions + * + * \param[out] out Returned newly created ::OrtRunOptions. Must be freed with OrtApi::ReleaseRunOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateRunOptions, _Outptr_ OrtRunOptions** out); + + /** \brief Set per-run log verbosity level + * + * \see OrtApi::RunOptionsGetRunLogVerbosityLevel + * + * \param[in] options + * \param[in] log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RunOptionsSetRunLogVerbosityLevel, _Inout_ OrtRunOptions* options, int log_verbosity_level); + + /** \brief Set per-run log severity level + * + * \see OrtApi::RunOptionsGetRunLogSeverityLevel + * + * \param[in] options + * \param[in] log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values). + */ + ORT_API2_STATUS(RunOptionsSetRunLogSeverityLevel, _Inout_ OrtRunOptions* options, int log_severity_level); + + /** \brief Set per-run tag + * + * This is used in a per-run log identifier. + * + * \see OrtApi::RunOptionsGetRunTag + * + * \param[in] options + * \param[in] run_tag The run tag. + */ + ORT_API2_STATUS(RunOptionsSetRunTag, _Inout_ OrtRunOptions* options, _In_ const char* run_tag); + + /** \brief Get per-run log verbosity level + * + * \see OrtApi::RunOptionsSetRunLogVerbosityLevel + * + * \param[in] options + * \param[out] log_verbosity_level \snippet{doc} snippets.dox Log Verbosity Level + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RunOptionsGetRunLogVerbosityLevel, _In_ const OrtRunOptions* options, + _Out_ int* log_verbosity_level); + + /** \brief Get per-run log severity level + * + * \see OrtApi::RunOptionsSetRunLogSeverityLevel + * + * \param[in] options + * \param[out] log_severity_level The log severity level (refer to ::OrtLoggingLevel for possible values). + */ + ORT_API2_STATUS(RunOptionsGetRunLogSeverityLevel, _In_ const OrtRunOptions* options, _Out_ int* log_severity_level); + + /** \brief Get per-run tag + * + * This is used in a per-run log identifier. + * + * \see OrtApi::RunOptionsSetRunTag + * + * \param[in] options + * \param[out] run_tag The run tag. + * Do not free this value, it is owned by `options`. It will be invalidated if the run tag + * changes (i.e., with OrtApi::RunOptionsSetRunTag) or `options` is freed. + */ + ORT_API2_STATUS(RunOptionsGetRunTag, _In_ const OrtRunOptions* options, _Out_ const char** run_tag); + + /** \brief Set terminate flag + * + * If a currently executing session needs to be force terminated, this can be called from another thread to force it to fail with an error. + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RunOptionsSetTerminate, _Inout_ OrtRunOptions* options); + + /** \brief Clears the terminate flag + * + * Used so the OrtRunOptions instance can be used in a new OrtApi::Run call without it instantly terminating + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RunOptionsUnsetTerminate, _Inout_ OrtRunOptions* options); + + /// @} + /// \name OrtValue + /// @{ + + /** \brief Create a tensor + * + * Create a tensor using a supplied ::OrtAllocator + * + * \param[in] allocator + * \param[in] shape Pointer to the tensor shape dimensions. + * \param[in] shape_len The number of tensor shape dimensions. + * \param[in] type + * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateTensorAsOrtValue, _Inout_ OrtAllocator* allocator, _In_ const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type, _Outptr_ OrtValue** out); + + /** \brief Create a tensor backed by a user supplied buffer + * + * Create a tensor with user's buffer. You can fill the buffer either before calling this function or after. + * p_data is owned by caller. ReleaseValue won't release p_data. + * + * If you wish to transfer ownership of p_data to ORT use CreateTensorWithDataAndDeleterAsOrtValue. + * + * \param[in] info Memory description of where the p_data buffer resides (CPU vs GPU etc). + * \param[in] p_data Pointer to the data buffer. + * \param[in] p_data_len The number of bytes in the data buffer. + * \param[in] shape Pointer to the tensor shape dimensions. + * \param[in] shape_len The number of tensor shape dimensions. + * \param[in] type The data type. + * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateTensorWithDataAsOrtValue, _In_ const OrtMemoryInfo* info, _Inout_ void* p_data, + size_t p_data_len, _In_ const int64_t* shape, size_t shape_len, ONNXTensorElementDataType type, + _Outptr_ OrtValue** out); + + /** \brief Return if an ::OrtValue is a tensor type + * + * \param[in] value A tensor type (string tensors are not supported) + * \param[out] out Set to 1 iff ::OrtValue is a tensor, 0 otherwise + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(IsTensor, _In_ const OrtValue* value, _Out_ int* out); + + /** \brief Get a pointer to the raw data inside a tensor + * + * Used to read/write/modify the internal tensor data directly. + * \note The returned pointer is valid until the \p value is destroyed. + * + * \param[in] value A tensor type (string tensors are not supported) + * \param[out] out Filled in with a pointer to the internal storage + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTensorMutableData, _In_ OrtValue* value, _Outptr_ void** out); + + /** \brief Set all strings at once in a string tensor + * + * \param[in,out] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING + * \param[in] s An array of strings. Each string in this array must be null terminated. + * \param[in] s_len Count of strings in s (Must match the size of \p value's tensor shape) + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(FillStringTensor, _Inout_ OrtValue* value, _In_ const char* const* s, size_t s_len); + + /** \brief Get total byte length for all strings in a string tensor + * + * Typically used with OrtApi::GetStringTensorContent + * + * \param[in] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING + * \param[out] len Total byte length of all strings (does not include trailing nulls) + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetStringTensorDataLength, _In_ const OrtValue* value, _Out_ size_t* len); + + /** \brief Get all strings from a string tensor + * + * An example of the results:
+ * Given \p value is a string tensor with the strings { "This" "is" "a" "test" }
+ * \p s must have a size of 11 bytes
+ * \p offsets must have 4 elements
+ * After the call, these values will be filled in:
+ * \p s will contain "Thisisatest"
+ * \p offsets will contain { 0, 4, 6, 7 }
+ * The length of the last string is just s_len - offsets[last] + * + * \param[in] value A tensor of type ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING + * \param[in] s Buffer to sequentially write all tensor strings to. Each string is NOT null-terminated. + * \param[in] s_len Number of bytes of buffer pointed to by \p s (Get it from OrtApi::GetStringTensorDataLength) + * \param[out] offsets Array of start offsets into the strings written to \p s + * \param[in] offsets_len Number of elements in offsets + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetStringTensorContent, _In_ const OrtValue* value, _Out_writes_bytes_all_(s_len) void* s, + size_t s_len, _Out_writes_all_(offsets_len) size_t* offsets, size_t offsets_len); + + /// @} + /// \name OrtTypeInfo + /// @{ + + /** \brief Get ::OrtTensorTypeAndShapeInfo from an ::OrtTypeInfo + * + * \param[in] type_info + * \param[out] out Do not free this value, it will be valid until type_info is freed. + * If type_info does not represent tensor, this value will be set to nullptr. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CastTypeInfoToTensorInfo, _In_ const OrtTypeInfo* type_info, + _Outptr_result_maybenull_ const OrtTensorTypeAndShapeInfo** out); + + /** \brief Get ::ONNXType from ::OrtTypeInfo + * + * \param[in] type_info + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetOnnxTypeFromTypeInfo, _In_ const OrtTypeInfo* type_info, _Out_ enum ONNXType* out); + + /// @} + /// \name OrtTensorTypeAndShapeInfo + /// @{ + + /** \brief Create an ::OrtTensorTypeAndShapeInfo object + * + * \param[out] out Returns newly created ::OrtTensorTypeAndShapeInfo. Must be freed with OrtApi::ReleaseTensorTypeAndShapeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateTensorTypeAndShapeInfo, _Outptr_ OrtTensorTypeAndShapeInfo** out); + + /** \brief Set element type in ::OrtTensorTypeAndShapeInfo + * + * \param[in] info + * \param[in] type + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetTensorElementType, _Inout_ OrtTensorTypeAndShapeInfo* info, enum ONNXTensorElementDataType type); + + /** \brief Set shape information in ::OrtTensorTypeAndShapeInfo + * + * \param[in] info + * \param[in] dim_values Array with `dim_count` elements. Can contain negative values. + * \param[in] dim_count Number of elements in `dim_values` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetDimensions, OrtTensorTypeAndShapeInfo* info, _In_ const int64_t* dim_values, size_t dim_count); + + /** \brief Get element type in ::OrtTensorTypeAndShapeInfo + * + * \see OrtApi::SetTensorElementType + * + * \param[in] info + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTensorElementType, _In_ const OrtTensorTypeAndShapeInfo* info, + _Out_ enum ONNXTensorElementDataType* out); + + /** \brief Get dimension count in ::OrtTensorTypeAndShapeInfo + * + * \see OrtApi::GetDimensions + * + * \param[in] info + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetDimensionsCount, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ size_t* out); + + /** \brief Get dimensions in ::OrtTensorTypeAndShapeInfo + * + * \param[in] info + * \param[out] dim_values Array with `dim_values_length` elements. On return, filled with the dimensions stored in the ::OrtTensorTypeAndShapeInfo + * \param[in] dim_values_length Number of elements in `dim_values`. Use OrtApi::GetDimensionsCount to get this value + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetDimensions, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ int64_t* dim_values, + size_t dim_values_length); + + /** \brief Get symbolic dimension names in ::OrtTensorTypeAndShapeInfo + * + * \param[in] info + * \param[in] dim_params Array with `dim_params_length` elements. On return filled with pointers to null terminated strings of the dimension names + * \param[in] dim_params_length Number of elements in `dim_params`. Use OrtApi::GetDimensionsCount to get this value + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSymbolicDimensions, _In_ const OrtTensorTypeAndShapeInfo* info, + _Out_writes_all_(dim_params_length) const char* dim_params[], size_t dim_params_length); + + /** \brief Get total number of elements in a tensor shape from an ::OrtTensorTypeAndShapeInfo + * + * Return the number of elements specified by the tensor shape (all dimensions multiplied by each other). + * For 0 dimensions, 1 is returned. If any dimension is less than 0, the result is always -1. + * + * Examples:
+ * [] = 1
+ * [1,3,4] = 12
+ * [2,0,4] = 0
+ * [-1,3,4] = -1
+ * + * \param[in] info + * \param[out] out Number of elements + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTensorShapeElementCount, _In_ const OrtTensorTypeAndShapeInfo* info, _Out_ size_t* out); + + /// @} + /// \name OrtValue + /// @{ + + /** \brief Get type and shape information from a tensor ::OrtValue + * + * \param[in] value Must be a tensor (not a map/sequence/etc) or will return failure + * \param[out] out Newly created ::OrtTensorTypeAndShapeInfo. Must be freed with OrtApi::ReleaseTensorTypeAndShapeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTensorTypeAndShape, _In_ const OrtValue* value, _Outptr_ OrtTensorTypeAndShapeInfo** out); + + /** \brief Get type information of an OrtValue + * + * \param[in] value + * \param[out] out Newly created ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTypeInfo, _In_ const OrtValue* value, _Outptr_result_maybenull_ OrtTypeInfo** out); + + /** \brief Get ONNXType of an ::OrtValue + * + * \param[in] value + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetValueType, _In_ const OrtValue* value, _Out_ enum ONNXType* out); + + /// @} + /// \name OrtMemoryInfo + /// @{ + + /** \brief Create an ::OrtMemoryInfo + * + * \param[in] name + * \param[in] type + * \param[in] id + * \param[in] mem_type + * \param[out] out Newly created ::OrtMemoryInfo. Must be freed with OrtAPi::ReleaseMemoryInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateMemoryInfo, _In_ const char* name, enum OrtAllocatorType type, int id, + enum OrtMemType mem_type, _Outptr_ OrtMemoryInfo** out); + + /** \brief Create an ::OrtMemoryInfo for CPU memory + * + * Special case version of OrtApi::CreateMemoryInfo for CPU based memory. Same as using OrtApi::CreateMemoryInfo with name = "Cpu" and id = 0. + * + * \param[in] type + * \param[in] mem_type + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateCpuMemoryInfo, enum OrtAllocatorType type, enum OrtMemType mem_type, + _Outptr_ OrtMemoryInfo** out); + + /** \brief Compare ::OrtMemoryInfo objects for equality + * + * Compares all settings of each ::OrtMemoryInfo for equality + * + * \param[in] info1 + * \param[in] info2 + * \param[out] out Set to 0 if equal, -1 if not equal + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CompareMemoryInfo, _In_ const OrtMemoryInfo* info1, _In_ const OrtMemoryInfo* info2, _Out_ int* out); + + /** \brief Get name from ::OrtMemoryInfo + * + * \param[in] ptr + * \param[out] out Writes null terminated string to this pointer. Do NOT free the returned pointer. It is valid for the lifetime of the ::OrtMemoryInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(MemoryInfoGetName, _In_ const OrtMemoryInfo* ptr, _Out_ const char** out); + + /** \brief Get the id from ::OrtMemoryInfo + */ + ORT_API2_STATUS(MemoryInfoGetId, _In_ const OrtMemoryInfo* ptr, _Out_ int* out); + + /** \brief Get the ::OrtMemType from ::OrtMemoryInfo + */ + ORT_API2_STATUS(MemoryInfoGetMemType, _In_ const OrtMemoryInfo* ptr, _Out_ OrtMemType* out); + + /** \brief Get the ::OrtAllocatorType from ::OrtMemoryInfo + */ + ORT_API2_STATUS(MemoryInfoGetType, _In_ const OrtMemoryInfo* ptr, _Out_ OrtAllocatorType* out); + + /// @} + /// \name OrtAllocator + /// @{ + + /// \brief Calls OrtAllocator::Alloc function + ORT_API2_STATUS(AllocatorAlloc, _Inout_ OrtAllocator* ort_allocator, size_t size, _Outptr_ void** out); + /// \brief Calls OrtAllocator::Free function + ORT_API2_STATUS(AllocatorFree, _Inout_ OrtAllocator* ort_allocator, void* p); + /// \brief Calls OrtAllocator::Info function + ORT_API2_STATUS(AllocatorGetInfo, _In_ const OrtAllocator* ort_allocator, _Outptr_ const struct OrtMemoryInfo** out); + + /** \brief Get the default allocator + * + * The default allocator is a CPU based, non-arena. Always returns the same pointer to the same default allocator. + * + * \param[out] out Returned value should NOT be freed + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetAllocatorWithDefaultOptions, _Outptr_ OrtAllocator** out); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Override session symbolic dimensions + * + * Override symbolic dimensions (by specific denotation strings) with actual values if known at session initialization time to enable + * optimizations that can take advantage of fixed values (such as memory planning, etc) + * + * \param[in] options + * \param[in] dim_denotation + * \param[in] dim_value + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(AddFreeDimensionOverride, _Inout_ OrtSessionOptions* options, _In_ const char* dim_denotation, + _In_ int64_t dim_value); + + /// @} + /// \name OrtValue + /// @{ + + /* Internal information (not seen in Doxygen) + * + * APIs to support non-tensor types - map and sequence. + * Currently only the following types are supported + * Note: the following types should be kept in sync with data_types.h + * Map types + * ========= + * std::map + * std::map + * std::map + * std::map + * std::map + * std::map + * std::map + * std::map + * + * Sequence types + * ============== + * std::vector + * std::vector + * std::vector + * std::vector + * std::vector> + * std::vector + */ + + /** \brief Get non tensor data from an ::OrtValue + * + * If `value` is of type ONNX_TYPE_MAP, you need to retrieve the keys and values + * separately. Use index=0 to retrieve keys and index=1 to retrieve values. + * If `value` is of type ONNX_TYPE_SEQUENCE, use index to retrieve the index'th element + * of the sequence. + * + * \param[in] value + * \param[in] index See above for usage based on `value` type + * \param[in] allocator Allocator used to allocate ::OrtValue + * \param[out] out Created ::OrtValue that holds the element requested. Must be freed with OrtApi::ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetValue, _In_ const OrtValue* value, int index, _Inout_ OrtAllocator* allocator, + _Outptr_ OrtValue** out); + + /** \brief Get non tensor value count from an ::OrtValue + * + * If `value` is of type ONNX_TYPE_MAP 2 will always be returned. For ONNX_TYPE_SEQUENCE + * the number of elements in the sequence will be returned + * + * \param[in] value + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetValueCount, _In_ const OrtValue* value, _Out_ size_t* out); + + /** \brief Create a map or sequence ::OrtValue + * + * To construct a map (ONNX_TYPE_MAP), use num_values = 2 and `in` should be an array of 2 ::OrtValue%s + * representing keys and values.
+ * + * To construct a sequence (ONNX_TYPE_SEQUENCE), use num_values = N where N is the number of the elements in the + * sequence. 'in' should be an array of N ::OrtValue%s. + * + * \param[in] in See above for details + * \param[in] num_values + * \param[in] value_type Must be either ONNX_TYPE_MAP or ONNX_TYPE_SEQUENCE + * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateValue, _In_reads_(num_values) const OrtValue* const* in, size_t num_values, + enum ONNXType value_type, _Outptr_ OrtValue** out); + + /** \brief Create an opaque (custom user defined type) ::OrtValue + * + * Constructs an ::OrtValue that contains a value of non-standard type created for + * experiments or while awaiting standardization. ::OrtValue in this case would contain + * an internal representation of the Opaque type. Opaque types are distinguished from + * each other by two strings 1) domain and 2) type name. The combination of the two + * must be unique, so the type representation is properly identified internally. The combination + * must be properly registered from within ORT at both compile/run time or by another API. + * + * To construct the ::OrtValue pass domain and type names, also a pointer to a data container + * the type of which must be known to both ORT and the client program. That data container may or may + * not match the internal representation of the Opaque type. The sizeof(data_container) is passed for + * verification purposes. + * + * \param[in] domain_name Null terminated string of the domain name + * \param[in] type_name Null terminated string of the type name + * \param[in] data_container User pointer Data to populate ::OrtValue + * \param[in] data_container_size Size in bytes of what `data_container` points to + * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateOpaqueValue, _In_z_ const char* domain_name, _In_z_ const char* type_name, + _In_ const void* data_container, size_t data_container_size, _Outptr_ OrtValue** out); + + /** \brief Get internal data from an opaque (custom user defined type) ::OrtValue + * + * Copies internal data from an opaque value into a user provided buffer + * + * \see OrtApi::CreateOpaqueValue + * + * \param[in] domain_name Null terminated string of the domain name + * \param[in] type_name Null terminated string of the type name + * \param[in] in The opaque ::OrtValue + * \param[out] data_container Buffer to copy data into + * \param[out] data_container_size Size in bytes of the buffer pointed to by data_container. Must match the size of the internal buffer. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetOpaqueValue, _In_ const char* domain_name, _In_ const char* type_name, _In_ const OrtValue* in, + _Out_ void* data_container, size_t data_container_size); + + /// @} + /// \name OrtKernelInfo + /// Custom operator APIs. + /// @{ + + /** \brief Get a float stored as an attribute in the graph node + * + * \param[in] info ::OrtKernelInfo instance + * \param[in] name Null terminated string of the name of the attribute + * \param[out] out Pointer to memory where the attribute will be stored + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelInfoGetAttribute_float, _In_ const OrtKernelInfo* info, _In_ const char* name, + _Out_ float* out); + + /** \brief Fetch a 64-bit int stored as an attribute in the graph node + * + * \param[in] info ::OrtKernelInfo instance + * \param[in] name Null terminated string of the name of the attribute + * \param[out] out Pointer to memory where the attribute will be stored + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelInfoGetAttribute_int64, _In_ const OrtKernelInfo* info, _In_ const char* name, + _Out_ int64_t* out); + + /** \brief Fetch a string stored as an attribute in the graph node + * + * If `out` is nullptr, the value of `size` is set to the true size of the string + * attribute, and a success status is returned. + * + * If the `size` parameter is greater than or equal to the actual string attribute's size, + * the value of `size` is set to the true size of the string attribute, the provided memory + * is filled with the attribute's contents, and a success status is returned. + * + * If the `size` parameter is less than the actual string attribute's size and `out` + * is not nullptr, the value of `size` is set to the true size of the string attribute + * and a failure status is returned.) + * + * \param[in] info ::OrtKernelInfo instance + * \param[in] name Null terminated string of the name of the attribute + * \param[out] out Pointer to memory where the attribute will be stored + * \param[in,out] size See above comments for details + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelInfoGetAttribute_string, _In_ const OrtKernelInfo* info, _In_ const char* name, _Out_ char* out, + _Inout_ size_t* size); + + /// @} + /// \name OrtKernelContext + /// Custom operator APIs. + /// @{ + + /** \brief Used for custom operators, get the input count of a kernel + * + * \see ::OrtCustomOp + */ + ORT_API2_STATUS(KernelContext_GetInputCount, _In_ const OrtKernelContext* context, _Out_ size_t* out); + + /** \brief Used for custom operators, get the output count of a kernel + * + * \see ::OrtCustomOp + */ + ORT_API2_STATUS(KernelContext_GetOutputCount, _In_ const OrtKernelContext* context, _Out_ size_t* out); + + /** \brief Used for custom operators, get an input of a kernel + * + * The function attempts fetches the input of the kernel. If the input is optional + * and not present, the function returns success and out is set to nullptr. + * + * \param[in] context ::OrtKernelContext instance + * \param[in] index See KernelContext_GetInputCount for boundaries check. + * \param[out] out OrtValue if the input is present otherwise is set nullptr + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelContext_GetInput, _In_ const OrtKernelContext* context, _In_ size_t index, + _Out_ const OrtValue** out); + + /** \brief Used for custom operators, get an output of a kernel + * + * The function attempts fetches the output of the kernel. If the output is optional + * and not present, the function returns success and out is set to nullptr. + * + * \param[in] context ::OrtKernelContext instance + * \param[in] index See KernelContext_GetOutputCount for boundaries check. + * \param[in] dim_values output dimensions + * \param[in] dim_count number of dimensions + * \param[out] out a ptr to OrtValue to output otherwise set to nullptr + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelContext_GetOutput, _Inout_ OrtKernelContext* context, _In_ size_t index, + _In_ const int64_t* dim_values, size_t dim_count, _Outptr_ OrtValue** out); + + /// @} + /// \name OrtEnv + /// @{ + ORT_CLASS_RELEASE(Env); + /// @} + /// \name OrtStatus + /// @{ + ORT_CLASS_RELEASE(Status); + /// @} + /// \name OrtMemoryInfo + /// @{ + ORT_CLASS_RELEASE(MemoryInfo); + /// @} + /// \name OrtSession + /// @{ + ORT_CLASS_RELEASE(Session); // Don't call ReleaseSession from Dllmain (because session owns a thread pool) + /// @} + /// \name OrtValue + /// @{ + ORT_CLASS_RELEASE(Value); + /// @} + /// \name OrtRunOptions + /// @{ + ORT_CLASS_RELEASE(RunOptions); + /// @} + /// \name OrtTypeInfo + /// @{ + ORT_CLASS_RELEASE(TypeInfo); + /// @} + /// \name OrtTensorTypeAndShapeInfo + /// @{ + ORT_CLASS_RELEASE(TensorTypeAndShapeInfo); + /// @} + /// \name OrtSessionOptions + /// @{ + ORT_CLASS_RELEASE(SessionOptions); + /// @} + /// \name OrtCustomOpDomain + /// @{ + ORT_CLASS_RELEASE(CustomOpDomain); + + /// @} + /// \name OrtTypeInfo + /// @{ + + /** \brief Get denotation from type information + * + * Augments ::OrtTypeInfo to return denotations on the type. + * + * This is used by WinML to determine if an input/output is intended to be an Image or a Tensor. + * + * \param[in] type_info + * \param[out] denotation Pointer to the null terminated denotation string is written to this pointer. This pointer is valid until the object is destroyed or the name is changed, do not free. + * \param[out] len Length in bytes of the string returned in `denotation` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetDenotationFromTypeInfo, _In_ const OrtTypeInfo* type_info, _Out_ const char** const denotation, + _Out_ size_t* len); + + /** \brief Get detailed map information from an ::OrtTypeInfo + * + * This augments ::OrtTypeInfo to return an ::OrtMapTypeInfo when the type is a map. + * The OrtMapTypeInfo has additional information about the map's key type and value type. + * + * This is used by WinML to support model reflection APIs. + * + * \param[out] type_info + * \param[out] out A pointer to the ::OrtMapTypeInfo. Do not free this value. If type_info + * does not contain a map, this value will be set to nullptr. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CastTypeInfoToMapTypeInfo, _In_ const OrtTypeInfo* type_info, + _Outptr_result_maybenull_ const OrtMapTypeInfo** out); + + /** \brief Cast ::OrtTypeInfo to an ::OrtSequenceTypeInfo + * + * This api augments ::OrtTypeInfo to return an ::OrtSequenceTypeInfo when the type is a sequence. + * The ::OrtSequenceTypeInfo has additional information about the sequence's element type. + * + * This is used by WinML to support model reflection APIs. + * + * \param[in] type_info + * \param[out] out A pointer to the OrtSequenceTypeInfo. Do not free this value. If type_info + * doesn not contain a sequence, this value will be set to nullptr. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CastTypeInfoToSequenceTypeInfo, _In_ const OrtTypeInfo* type_info, + _Outptr_result_maybenull_ const OrtSequenceTypeInfo** out); + + /// @} + /// \name OrtMapTypeInfo + /// @{ + + /** \brief Get key type from an ::OrtMapTypeInfo + * + * Key types are restricted to being scalar types. + * + * This is used by WinML to support model reflection APIs. + * + * \param[in] map_type_info + * \param[out] out + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetMapKeyType, _In_ const OrtMapTypeInfo* map_type_info, _Out_ enum ONNXTensorElementDataType* out); + + /** \brief Get the value type from an ::OrtMapTypeInfo + * + * \param[in] map_type_info + * \param[out] type_info A copy of the OrtTypeInfo for the map value type. + * The user must free this value with ReleaseTypeInfo. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetMapValueType, _In_ const OrtMapTypeInfo* map_type_info, _Outptr_ OrtTypeInfo** type_info); + + /// @} + /// \name OrtSequenceTypeInfo + /// @{ + + /** \brief Get element type from an ::OrtSequenceTypeInfo + * + * This is used by WinML to support model reflection APIs. + * + * \param[in] sequence_type_info + * \param[out] type_info A copy of the OrtTypeInfo for the sequence element type. + * The user must free this value with ReleaseTypeInfo. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSequenceElementType, _In_ const OrtSequenceTypeInfo* sequence_type_info, + _Outptr_ OrtTypeInfo** type_info); + + /// @} + /// \name OrtMapTypeInfo + /// @{ + ORT_CLASS_RELEASE(MapTypeInfo); + /// @} + /// \name OrtSequenceTypeInfo + /// @{ + ORT_CLASS_RELEASE(SequenceTypeInfo); + + /// @} + /// \name OrtSession + /// @{ + + /** \brief End profiling and return filename of the profile data + * + * Profiling is turned on through OrtApi::EnableProfiling + * + * \param[in] session + * \param[in] allocator + * \param[out] out Null terminated string of the filename, allocated using `allocator`. Must be freed using `allocator` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionEndProfiling, _In_ OrtSession* session, _Inout_ OrtAllocator* allocator, _Outptr_ char** out); + + /** \brief Get ::OrtModelMetadata from an ::OrtSession + * + * \param[in] session + * \param[out] out Newly created ::OrtModelMetadata. Must be freed using OrtApi::ReleaseModelMetadata + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetModelMetadata, _In_ const OrtSession* session, _Outptr_ OrtModelMetadata** out); + + /// @} + /// \name OrtModelMetadata + /// @{ + + /** \brief Get `producer name` from an ::OrtModelMetadata + * + * \param[in] model_metadata + * \param[in] allocator + * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetProducerName, _In_ const OrtModelMetadata* model_metadata, + _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /** \brief Get `graph name` from an ::OrtModelMetadata + * + * \param[in] model_metadata + * \param[in] allocator + * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetGraphName, _In_ const OrtModelMetadata* model_metadata, + _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /** \brief Get `domain` from an ::OrtModelMetadata + * + * \param[in] model_metadata + * \param[in] allocator + * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetDomain, _In_ const OrtModelMetadata* model_metadata, _Inout_ OrtAllocator* allocator, + _Outptr_ char** value); + + /** \brief Get `description` from an ::OrtModelMetadata + * + * \param[in] model_metadata + * \param[in] allocator + * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetDescription, _In_ const OrtModelMetadata* model_metadata, + _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /** \brief Return data for a key in the custom metadata map in an ::OrtModelMetadata + * + * \param[in] model_metadata + * \param[in] allocator + * \param[in] key Null terminated string + * \param[out] value Set to a null terminated string allocated using `allocator`. Must be freed using `allocator` + * `value` will be set to nullptr if the given key is not found in the custom metadata map. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataLookupCustomMetadataMap, _In_ const OrtModelMetadata* model_metadata, + _Inout_ OrtAllocator* allocator, _In_ const char* key, _Outptr_result_maybenull_ char** value); + + /** \brief Get version number from an ::OrtModelMetadata + * + * \param[in] model_metadata + * \param[out] value Set to the version number + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetVersion, _In_ const OrtModelMetadata* model_metadata, _Out_ int64_t* value); + + ORT_CLASS_RELEASE(ModelMetadata); + + /// @} + /// \name OrtEnv + /// @{ + + /** \brief Create an OrtEnv + * + * Create an environment with global threadpools that will be shared across sessions. + * Use this in conjunction with OrtApi::DisablePerSessionThreads or else the session will use + * its own thread pools. + * + * \param[in] log_severity_level The log severity level. + * \param[in] logid The log identifier. + * \param[in] tp_options + * \param[out] out Returned newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateEnvWithGlobalThreadPools, OrtLoggingLevel log_severity_level, _In_ const char* logid, + _In_ const OrtThreadingOptions* tp_options, _Outptr_ OrtEnv** out); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Use global thread pool on a session + * + * Disable using per session thread pool and use the shared global threadpool. + * This should be used in conjunction with OrtApi::CreateEnvWithGlobalThreadPools. + * + * \param[in] options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(DisablePerSessionThreads, _Inout_ OrtSessionOptions* options); + + /// @} + /// \name OrtThreadingOptions + /// @{ + + /** \brief Create an ::OrtThreadingOptions + * + * \param[out] out Newly created ::OrtThreadingOptions. Must be freed with OrtApi::ReleaseThreadingOptions + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateThreadingOptions, _Outptr_ OrtThreadingOptions** out); + + ORT_CLASS_RELEASE(ThreadingOptions); + + /// @} + /// \name OrtModelMetadata + /// @{ + + /** + * + * \param[in] model_metadata + * \param[in] allocator + * \param[out] keys Array of null terminated strings (array count = num_keys) allocated using `allocator`. + * The strings and the pointer array must be freed using `allocator` + * `keys` will be set to nullptr if the custom metadata map is empty. + * \param[out] num_keys Set to the number of elements in the `keys` array + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetCustomMetadataMapKeys, _In_ const OrtModelMetadata* model_metadata, + _Inout_ OrtAllocator* allocator, _Outptr_result_buffer_maybenull_(*num_keys) char*** keys, _Out_ int64_t* num_keys); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** + * + * Override symbolic dimensions (by specific name strings) with actual values + * if known at session initialization time to enable optimizations that can + * take advantage of fixed values (such as memory planning, etc) + * + */ + ORT_API2_STATUS(AddFreeDimensionOverrideByName, + _Inout_ OrtSessionOptions* options, _In_ const char* dim_name, + _In_ int64_t dim_value); + + /// @} + /// \name Misc + /// @{ + + /** \brief Get the names of all available providers + * + * \note The providers in the list are not guaranteed to be usable. They may fail to load due to missing system dependencies. + * For example, if the CUDA/cuDNN libraries are not installed, the CUDA provider will report an error when it is added to the session options. + * + * \param[out] out_ptr Set to a pointer to an array of null terminated strings of the available providers. The entries and the + * array itself must be freed using OrtApi::ReleaseAvailableProviders + * \param[out] provider_length Set to the number of entries in the `out_ptr` array + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetAvailableProviders, _Outptr_ char*** out_ptr, _Out_ int* provider_length); + + /** \brief Release data from OrtApi::GetAvailableProviders. This API will never fail + * so you can rely on it in a noexcept code. + * + * \param[in] ptr The `out_ptr` result from OrtApi::GetAvailableProviders. + * \param[in] providers_length The `provider_length` result from OrtApi::GetAvailableProviders + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ReleaseAvailableProviders, _In_ char** ptr, + _In_ int providers_length); + + /// @} + /// \name OrtValue + /// @{ + + /** \brief Get the length of a single string in a string tensor + * + * \param[in] value A string tensor + * \param[in] index Index of the string in the tensor + * \param[out] out Set to number of bytes of the string element + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetStringTensorElementLength, _In_ const OrtValue* value, size_t index, _Out_ size_t* out); + + /** \brief Get a single string from a string tensor + * + * \param[in] value A string tensor + * \param[in] s_len Number of bytes in the `s` buffer. Must match the value returned by OrtApi::GetStringTensorElementLength. + * \param[in] index Index of the string in the tensor + * \param[out] s The string element contents in UTF-8 encoding. The string is NOT null-terminated. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetStringTensorElement, _In_ const OrtValue* value, size_t s_len, size_t index, _Out_writes_bytes_all_(s_len) void* s); + + /** \brief Set a single string in a string tensor + * + * \param[in] value A string tensor + * \param[in] s A null terminated UTF-8 encoded string + * \param[in] index Index of the string in the tensor to set + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(FillStringTensorElement, _Inout_ OrtValue* value, _In_ const char* s, size_t index); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Set a session configuration entry as a pair of strings + * + * If a configuration with same key exists, this will overwrite the configuration with the given config_value. + * + * The config_key and the format of config_value are defined in onnxruntime_session_options_config_keys.h + * + * \param[in] options + * \param[in] config_key A null terminated string representation of the config key + * \param[in] config_value A null terminated string representation of the config value + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(AddSessionConfigEntry, _Inout_ OrtSessionOptions* options, + _In_z_ const char* config_key, _In_z_ const char* config_value); + + /// @} + /// \name OrtAllocator + /// @{ + + /** \brief Create an allocator for an ::OrtSession following an ::OrtMemoryInfo + * + * \param[in] session + * \param[in] mem_info valid ::OrtMemoryInfo instance + * \param[out] out Newly created ::OrtAllocator. Must be freed with OrtApi::ReleaseAllocator + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateAllocator, _In_ const OrtSession* session, _In_ const OrtMemoryInfo* mem_info, + _Outptr_ OrtAllocator** out); + + /** \brief Release an ::OrtAllocator obtained from OrtApi::CreateAllocator + */ + ORT_CLASS_RELEASE(Allocator); + + /// @} + /// \name OrtSession + /// @{ + + /** \brief Run a model using Io Bindings for the inputs & outputs + * + * \see OrtApi::Run + * + * \param[in] session + * \param[in] run_options + * \param[in] binding_ptr + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RunWithBinding, _Inout_ OrtSession* session, _In_ const OrtRunOptions* run_options, _In_ const OrtIoBinding* binding_ptr); + + /** \brief Create an ::OrtIoBinding instance + * + * An IoBinding object allows one to bind pre-allocated ::OrtValue%s to input names. + * Thus if you want to use a raw on device buffer as input or output you can avoid + * extra copy during runtime. + * + * \param[in] session + * \param[out] out Newly created ::OrtIoBinding. Must be freed with OrtApi::ReleaseIoBinding + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateIoBinding, _Inout_ OrtSession* session, _Outptr_ OrtIoBinding** out); + + /// @} + /// \name OrtIoBinding + /// @{ + + /** \brief Release an ::OrtIoBinding obtained from OrtApi::CreateIoBinding + */ + ORT_CLASS_RELEASE(IoBinding); + + /** \brief Bind an ::OrtValue to an ::OrtIoBinding input + * + * When using OrtApi::RunWithBinding this value is used for the named input + * + * \param[in] binding_ptr + * \param[in] name Name for the model input + * \param[in] val_ptr ::OrtValue of Tensor type. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(BindInput, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtValue* val_ptr); + + /** \brief Bind an ::OrtValue to an ::OrtIoBinding output + * + * When using OrtApi::RunWithBinding this value is used for the named output + * + * \param[in] binding_ptr + * \param[in] name Null terminated string of the model output name + * \param[in] val_ptr ::OrtValue of Tensor type. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(BindOutput, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtValue* val_ptr); + + /** \brief Bind an ::OrtIoBinding output to a device + * + * Binds the ::OrtValue to a device which is specified by ::OrtMemoryInfo. + * You can either create an instance of ::OrtMemoryInfo with a device id or obtain one from the allocator that you have created/are using + * This is useful when one or more outputs have dynamic shapes and, it is hard to pre-allocate and bind a chunk of + * memory within ::OrtValue ahead of time. + * + * \see OrtApi::RunWithBinding + * + * \param[in] binding_ptr + * \param[in] name Null terminated string of the device name + * \param[in] mem_info_ptr + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(BindOutputToDevice, _Inout_ OrtIoBinding* binding_ptr, _In_ const char* name, _In_ const OrtMemoryInfo* mem_info_ptr); + + /** \brief Get the names of an ::OrtIoBinding's outputs + * + * Returns the names of the outputs in the order they were bound. This is useful after running the model + * with bound outputs because the returned names are in order in which output ::OrtValue are returned. This is useful if + * the order of outputs and their names is not known. + * + * \param[in] binding_ptr + * \param[in] allocator Allocator used to allocate continuous buffers for output strings and lengths. + * \param[out] buffer Returns an array of non-null terminated UTF-8 strings. The number of strings stored is returned in the count parameter. + * This buffer is allocated using `allocator` and must be freed using it. + * \param[out] lengths Returns an array of `count` lengths of the strings returned in `buffer` + * This buffer is allocated using `allocator` and must be freed using it. + * \param[out] count Number of strings returned. If `binding_ptr` has no bound outputs, zero is returned, + * no memory allocation is performed and buffer and lengths are set to nullptr. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetBoundOutputNames, _In_ const OrtIoBinding* binding_ptr, _In_ OrtAllocator* allocator, + _Out_ char** buffer, _Out_writes_all_(count) size_t** lengths, _Out_ size_t* count); + + /** \brief Get the output ::OrtValue objects from an ::OrtIoBinding + * + * Returns an array of pointers to individually allocated ::OrtValue%s that contain results of a model execution with OrtApi::RunWithBinding + * The array contains the same number of ::OrtValue%s and they are in the same order as they were bound with OrtApi::BindOutput + * or OrtApi::BindOutputToDevice. + * + * The returned ::OrtValue%s must be released using OrtApi::ReleaseValue after they are no longer needed. + * The array is allocated using the specified instance of the allocator and must be freed using the same allocator after + * all the ::OrtValue%s contained therein are individually released. + * + * \param[in] binding_ptr + * \param[in] allocator Allocator used to allocate output array + * \param[out] output Set to the allocated array of allocated ::OrtValue outputs. Set to nullptr if there are 0 outputs. + * \param[out] output_count Set to number of ::OrtValue%s returned + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetBoundOutputValues, _In_ const OrtIoBinding* binding_ptr, _In_ OrtAllocator* allocator, + _Out_writes_all_(output_count) OrtValue*** output, _Out_ size_t* output_count); + + /** \brief Clears any previously set Inputs for an ::OrtIoBinding + */ + void(ORT_API_CALL* ClearBoundInputs)(_Inout_ OrtIoBinding* binding_ptr) NO_EXCEPTION ORT_ALL_ARGS_NONNULL; + + /** \brief Clears any previously set Outputs for an ::OrtIoBinding + */ + void(ORT_API_CALL* ClearBoundOutputs)(_Inout_ OrtIoBinding* binding_ptr) NO_EXCEPTION ORT_ALL_ARGS_NONNULL; + + /// @} + /// \name OrtValue + /// @{ + + /** \brief Direct memory access to a specified tensor element + * + * For example, given a tensor with shape of [3,224,224], a pointer to the element at location [2,150,128] can be retrieved + * + * This function only works for numeric type tensors (No strings, etc). + * This is a no-copy method whose returned pointer is valid until the passed in ::OrtValue is free'd. + * + * \param[in] value + * \param[in] location_values Pointer to an array of index values that specify an element's location relative to its shape + * \param[in] location_values_count Number of elements in location_values. Must match the number of elements in the tensor's shape. + * \param[out] out Set to a pointer to the element specified + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(TensorAt, _Inout_ OrtValue* value, const int64_t* location_values, size_t location_values_count, _Outptr_ void** out); + + /// @} + /// \name OrtEnv + /// @{ + + /** \brief Create an allocator and register it with the ::OrtEnv + * + * Enables sharing the allocator between multiple sessions that use the same env instance. + * Lifetime of the created allocator will be valid for the duration of the environment. + * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered. + * + * See https://onnxruntime.ai/docs/get-started/with-c.html for details. + * + * \param[in] env ::OrtEnv instance + * \param[in] mem_info + * \param[in] arena_cfg Pass nullptr for defaults + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateAndRegisterAllocator, _Inout_ OrtEnv* env, _In_ const OrtMemoryInfo* mem_info, + _In_ const OrtArenaCfg* arena_cfg); + + /** \brief Set language projection + * + * Set the language projection for collecting telemetry data when Env is created. + * + * The default is ORT_PROJECTION_C, which means it will classify the language not in the list to C also. + * + * \param[in] ort_env + * \param[in] projection + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetLanguageProjection, _In_ const OrtEnv* ort_env, _In_ OrtLanguageProjection projection); + + /// @} + /// \name OrtSession + /// @{ + + /** \brief Return the time that profiling was started + * + * \note The timer precision varies per platform. On Windows and MacOS, the precision will be ~100ns + * + * \param[in] session + * \param[out] out nanoseconds of profiling's start time + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionGetProfilingStartTimeNs, _In_ const OrtSession* session, _Outptr_ uint64_t* out); + + /// @} + /// \name OrtThreadingOptions + /// @{ + + /** \brief Set global intra-op thread count + * + * This configures the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools + * + * \param[in] tp_options + * \param[in] intra_op_num_threads Number of threads, special values:
+ * 0 = Use default thread count
+ * 1 = The invoking thread will be used; no threads will be created in the thread pool. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalIntraOpNumThreads, _Inout_ OrtThreadingOptions* tp_options, int intra_op_num_threads); + + /** \brief Set global inter-op thread count + * + * This configures the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools + * + * \param[in] tp_options + * \param[in] inter_op_num_threads Number of threads, special values:
+ * 0 = Use default thread count
+ * 1 = The invoking thread will be used; no threads will be created in the thread pool. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalInterOpNumThreads, _Inout_ OrtThreadingOptions* tp_options, int inter_op_num_threads); + + /** \brief Set global spin control options + * + * This will configure the global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools. + * Allow spinning of thread pools when their queues are empty. This will set the value for both + * inter_op and intra_op threadpools. + * + * \param[in] tp_options + * \param[in] allow_spinning Valid values are 0 or 1.
+ * 0 = It won't spin (recommended if CPU usage is high)
+ * 1 = Threadpool will spin to wait for queue to become non-empty + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalSpinControl, _Inout_ OrtThreadingOptions* tp_options, int allow_spinning); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Add a pre-allocated initializer to a session + * + * If a model contains an initializer with a name that is same as the name passed to this call, + * ORT will use this initializer instance instead of deserializing one from the model file. This + * is useful when you want to share the same initializer across sessions. + * + * \param[in] options + * \param[in] name Null terminated string of the initializer name + * \param[in] val ::OrtValue containing the initializer. Its lifetime and the underlying initializer buffer must be + * managed by the user (created using the OrtApi::CreateTensorWithDataAsOrtValue) and it must outlive the session object + * to which it is added. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(AddInitializer, _Inout_ OrtSessionOptions* options, _In_z_ const char* name, + _In_ const OrtValue* val); + + /// @} + /// \name OrtEnv + /// @{ + + /** + * Create a custom environment with global threadpools and logger that will be shared across sessions. + * Use this in conjunction with OrtApi::DisablePerSessionThreads or else the session will use + * its own thread pools. + * + * \param[in] logging_function A pointer to a logging function. + * \param[in] logger_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to + * `logging_function`. + * \param[in] log_severity_level The log severity level. + * \param[in] logid The log identifier. + * \param[in] tp_options + * \param[out] out Newly created OrtEnv. Must be freed with OrtApi::ReleaseEnv + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateEnvWithCustomLoggerAndGlobalThreadPools, OrtLoggingFunction logging_function, _In_opt_ void* logger_param, OrtLoggingLevel log_severity_level, + _In_ const char* logid, _In_ const struct OrtThreadingOptions* tp_options, _Outptr_ OrtEnv** out); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Append CUDA provider to session options + * + * If CUDA is not available (due to a non CUDA enabled build, or if CUDA is not installed on the system), this function will return failure. + * + * \param[in] options + * \param[in] cuda_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CUDA, + _In_ OrtSessionOptions* options, _In_ const OrtCUDAProviderOptions* cuda_options); + + /** \brief Append ROCM execution provider to the session options + * + * If ROCM is not available (due to a non ROCM enabled build, or if ROCM is not installed on the system), this function will return failure. + * + * \param[in] options + * \param[in] rocm_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_ROCM, + _In_ OrtSessionOptions* options, _In_ const OrtROCMProviderOptions* rocm_options); + + /** \brief Append OpenVINO execution provider to the session options + * + * If OpenVINO is not available (due to a non OpenVINO enabled build, or if OpenVINO is not installed on the system), this function will fail. + * + * \param[in] options + * \param[in] provider_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_OpenVINO, + _In_ OrtSessionOptions* options, _In_ const OrtOpenVINOProviderOptions* provider_options); + + /// @} + /// \name OrtThreadingOptions + /// @{ + + /** \brief Set threading flush-to-zero and denormal-as-zero + * + * Sets global thread pool options to be used in the call to OrtApi::CreateEnvWithGlobalThreadPools. + * Flush-to-zero and denormal-as-zero are applied to threads in both intra and inter global thread pool. + * \note This option is not needed if the models used have no denormals. Having no denormals is recommended as this option may hurt model accuracy. + * + * \param[in] tp_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalDenormalAsZero, _Inout_ OrtThreadingOptions* tp_options); + + /// @} + /// \name OrtArenaCfg + /// @{ + + /** \deprecated Use OrtApi::CreateArenaCfgV2 + * + * This will create the configuration of an arena that can eventually be used to define an arena based allocator's behavior + * + * \param[in] max_mem Use 0 to allow ORT to choose the default + * \param[in] arena_extend_strategy Use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested + * \param[in] initial_chunk_size_bytes Use -1 to allow ORT to choose the default + * \param[in] max_dead_bytes_per_chunk Use -1 to allow ORT to choose the default + * \param[in] out A pointer to an OrtArenaCfg instance + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateArenaCfg, _In_ size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, + int max_dead_bytes_per_chunk, _Outptr_ OrtArenaCfg** out); + + ORT_CLASS_RELEASE(ArenaCfg); + + /// @} + /// \name OrtModelMetadata + /// @{ + + /** + * Use this to obtain the description of the graph present in the model + * (doc_string field of the GraphProto message within the ModelProto message). + * If it doesn't exist, an empty string will be returned. + * + * \param[in] model_metadata An instance of ::OrtModelMetadata + * \param[in] allocator Allocator used to allocate the string that will be returned back + * \param[out] value Set to a null terminated string allocated using `allocator`. The caller is responsible for freeing it using `allocator` + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(ModelMetadataGetGraphDescription, _In_ const OrtModelMetadata* model_metadata, + _Inout_ OrtAllocator* allocator, _Outptr_ char** value); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Append TensorRT provider to session options + * + * If TensorRT is not available (due to a non TensorRT enabled build, or if TensorRT is not installed on the system), this function will return failure. + * + * \param[in] options + * \param[in] tensorrt_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_TensorRT, + _In_ OrtSessionOptions* options, _In_ const OrtTensorRTProviderOptions* tensorrt_options); + + /// @} + /// \name Misc + /// @{ + + /** \brief Set current GPU device ID + * + * Set the current device id of the GPU execution provider (CUDA/tensorrt/rocm). The device id should be less + * than the total number of devices available. This is only useful when multiple-GPUs are installed and it is + * required to restrict execution to a single GPU. + * + * \param[in] device_id + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetCurrentGpuDeviceId, _In_ int device_id); + + /** \brief Get current GPU device ID + * + * Get the current device id of the GPU execution provider (CUDA/tensorrt/rocm). + * + * \see OrtApi::SetCurrentGpuDeviceId + * + * \param[out] device_id + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetCurrentGpuDeviceId, _In_ int* device_id); + + /// @} + /// \name OrtKernelInfo + /// Custom operator APIs. + /// @{ + + /** \brief Fetch an array of int64_t values stored as an attribute in the graph node + * + * + * If `out` is nullptr, the value of `size` is set to the true size of the attribute + * array's size, and a success status is returned. + * + * If the `size` parameter is greater than or equal to the actual attribute array's size, + * the value of `size` is set to the true size of the attribute array's size, + * the provided memory is filled with the attribute's contents, + * and a success status is returned. + * + * If the `size` parameter is less than the actual attribute array's size and `out` + * is not nullptr, the value of `size` is set to the true size of the attribute array's size + * and a failure status is returned.) + * + * \param[in] info instance + * \param[in] name name of the attribute to be parsed + * \param[out] out pointer to memory where the attribute's contents are to be stored + * \param[in, out] size actual size of attribute array + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelInfoGetAttributeArray_float, _In_ const OrtKernelInfo* info, _In_ const char* name, + _Out_ float* out, _Inout_ size_t* size); + + /** \brief Fetch an array of int64_t values stored as an attribute in the graph node + * + * If `out` is nullptr, the value of `size` is set to the true size of the attribute + * array's size, and a success status is returned. + * + * If the `size` parameter is greater than or equal to the actual attribute array's size, + * the value of `size` is set to the true size of the attribute array's size, + * the provided memory is filled with the attribute's contents, + * and a success status is returned. + * + * If the `size` parameter is less than the actual attribute array's size and `out` + * is not nullptr, the value of `size` is set to the true size of the attribute array's size + * and a failure status is returned.) + * + * \param[in] info instance + * \param[in] name name of the attribute to be parsed + * \param[out] out pointer to memory where the attribute's contents are to be stored + * \param[in, out] size actual size of attribute array + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelInfoGetAttributeArray_int64, _In_ const OrtKernelInfo* info, _In_ const char* name, + _Out_ int64_t* out, _Inout_ size_t* size); + + /// @} + /// \name OrtArenaCfg + /// @{ + + /** \brief Create an ::OrtArenaCfg + * + * Create the configuration of an arena that can eventually be used to define an arena based allocator's behavior. + * + * Supported keys are (See https://onnxruntime.ai/docs/get-started/with-c.html for details on what the + * following parameters mean and how to choose these values.): + * "max_mem": Maximum memory that can be allocated by the arena based allocator. + * Use 0 for ORT to pick the best value. Default is 0. + * "arena_extend_strategy": 0 = kNextPowerOfTwo, 1 = kSameAsRequested. + * Use -1 to allow ORT to choose the default. + * "initial_chunk_size_bytes": (Possible) Size of the first allocation in the arena. + * Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default. + * Ultimately, the first allocation size is determined by the allocation memory request. + * "max_dead_bytes_per_chunk": Threshold of unused memory in an allocated chunk of arena memory after + * crossing which the current chunk is chunked into 2. + * "initial_growth_chunk_size_bytes": (Possible) Size of the second allocation in the arena. + * Only relevant if arena strategy is `kNextPowerOfTwo`. Use -1 to allow ORT to choose the default. + * "max_power_of_two_extend_bytes": The maximum enxtend size if arena strategy is `kNextPowerOfTwo`. + * It is not an allocation limit, it is only a limit for extension when requested byte is less than the limit. + * When requested bytes is more than the limit, allocator will still return as requested. + * Use -1 to allow ORT to choose the default 1GB for max_power_of_two_extend_bytes. + * Ultimately, the allocation size is determined by the allocation memory request. + * Further allocation sizes are governed by the arena extend strategy. + * + * \param[in] arena_config_keys Keys to configure the arena + * \param[in] arena_config_values Values to configure the arena + * \param[in] num_keys Number of keys in `arena_config_keys` and `arena_config_values` + * \param[out] out Newly created ::OrtArenaCfg. Must be freed with OrtApi::ReleaseArenaCfg + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateArenaCfgV2, _In_reads_(num_keys) const char* const* arena_config_keys, + _In_reads_(num_keys) const size_t* arena_config_values, _In_ size_t num_keys, + _Outptr_ OrtArenaCfg** out); + + /// @} + /// \name OrtRunOptions + /// @{ + + /** \brief Set a single run configuration entry as a pair of strings + * + * If a configuration with same key exists, this will overwrite the configuration with the given config_value + * + * The config_key and the format of config_value are defined in onnxruntime_run_options_config_keys.h + * + * \param[in] options + * \param[in] config_key A null terminated string representation of the config key + * \param[in] config_value A null terminated string representation of the config value + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(AddRunConfigEntry, _Inout_ OrtRunOptions* options, + _In_z_ const char* config_key, _In_z_ const char* config_value); + + /// @} + /// \name OrtPrepackedWeightsContainer + /// @{ + + /** \brief Create an ::OrtPrepackedWeightsContainer + * + * This container will hold pre-packed buffers of shared initializers for sharing between sessions + * (i.e.) if there are shared initializers that can be shared between sessions, the pre-packed buffers + * of these (if any) may possibly be shared to provide memory footprint savings. Pass this container + * to sessions that you would like to share pre-packed buffers of shared initializers at session + * creation time. + * + * \param[out] out Newly created ::OrtPrepackedWeightsContainer. Must be freed with OrtApi::ReleasePrepackedWeightsContainer + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreatePrepackedWeightsContainer, _Outptr_ OrtPrepackedWeightsContainer** out); + + /** \brief Release OrtPrepackedWeightsContainer instance + * + * \note instance must not be released until the sessions using it are released + */ + ORT_CLASS_RELEASE(PrepackedWeightsContainer); + + /// @} + /// \name OrtSession + /// @{ + + /** \brief Create session with prepacked weights container + * + * Same functionality offered by OrtApi::CreateSession except that a container that contains + * pre-packed weights' buffers is written into/read from by the created session. + * This is useful when used in conjunction with OrtApi::AddInitializer which injects + * shared initializer info into sessions. Wherever possible, the pre-packed versions of these + * shared initializers are cached in this container so that multiple sessions can just re-use + * these instead of duplicating these in memory. + * + * \param[in] env OrtEnv instance instance + * \param[in] model_path Null terminated string of the path (wchar on Windows, char otherwise) + * \param[in] options + * \param[in] prepacked_weights_container + * \param[out] out Newly created ::OrtSession. Must be freed with OrtApi::ReleaseSession + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateSessionWithPrepackedWeightsContainer, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path, + _In_ const OrtSessionOptions* options, + _Inout_ OrtPrepackedWeightsContainer* prepacked_weights_container, + _Outptr_ OrtSession** out); + + /** \brief Create session from memory with prepacked weights container + * + * Same functionality offered by OrtApi::CreateSessionFromArray except that a container that contains + * pre-packed weights' buffers is written into/read from by the created session. + * This is useful when used in conjunction with OrtApi::AddInitializer which injects + * shared initializer info into sessions. Wherever possible, the pre-packed versions of these + * shared initializers are cached in this container so that multiple sessions can just re-use + * these instead of duplicating these in memory. + * + * \param[in] env + * \param[in] model_data Array of bytes holding the model + * \param[in] model_data_length Number of bytes in `model_data_model` + * \param[in] options + * \param[in] prepacked_weights_container + * \param[out] out Newly created ::OrtSession. Must be freed with OrtApi::ReleaseSession + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateSessionFromArrayWithPrepackedWeightsContainer, _In_ const OrtEnv* env, + _In_ const void* model_data, size_t model_data_length, + _In_ const OrtSessionOptions* options, + _Inout_ OrtPrepackedWeightsContainer* prepacked_weights_container, + _Outptr_ OrtSession** out); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Append TensorRT execution provider to the session options + * + * If TensorRT is not available (due to a non TensorRT enabled build), this function will return failure. + * + * This is slightly different from OrtApi::SessionOptionsAppendExecutionProvider_TensorRT, it takes an + * ::OrtTensorRTProviderOptions which is publicly defined. This takes an opaque ::OrtTensorRTProviderOptionsV2 + * which must be created with OrtApi::CreateTensorRTProviderOptions. + * + * For OrtApi::SessionOptionsAppendExecutionProvider_TensorRT, the user needs to instantiate ::OrtTensorRTProviderOptions + * as well as allocate/release buffers for some members of ::OrtTensorRTProviderOptions. + * Here, OrtApi::CreateTensorRTProviderOptions and Ortapi::ReleaseTensorRTProviderOptions will do the memory management for you. + * + * \param[in] options + * \param[in] tensorrt_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_TensorRT_V2, + _In_ OrtSessionOptions* options, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options); + + /// @} + /// \name OrtTensorRTProviderOptionsV2 + /// @{ + + /** \brief Create an OrtTensorRTProviderOptionsV2 + * + * \param[out] out Newly created ::OrtTensorRTProviderOptionsV2. Must be released with OrtApi::ReleaseTensorRTProviderOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateTensorRTProviderOptions, _Outptr_ OrtTensorRTProviderOptionsV2** out); + + /** \brief Set options in a TensorRT Execution Provider. + * + * Please refer to https://onnxruntime.ai/docs/execution-providers/TensorRT-ExecutionProvider.html#cc + * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtTensorRTProviderOptionsV2 + * and value should be its related range. Recreates the options and only sets the supplied values. + * + * For example, key="trt_max_workspace_size" and value="2147483648" + * + * \param[in] tensorrt_options + * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys + * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values + * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(UpdateTensorRTProviderOptions, _Inout_ OrtTensorRTProviderOptionsV2* tensorrt_options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** \brief Get serialized TensorRT provider options string. + * + * For example, "trt_max_workspace_size=2147483648;trt_max_partition_iterations=10;trt_int8_enable=1;......" + * + * \param tensorrt_options - OrtTensorRTProviderOptionsV2 instance + * \param allocator - a ptr to an instance of OrtAllocator obtained with OrtApi::CreateAllocator or OrtApi::GetAllocatorWithDefaultOptions + * the specified allocator will be used to allocate continuous buffers for output strings and lengths. + * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTensorRTProviderOptionsAsString, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr); + + /** \brief Release an ::OrtTensorRTProviderOptionsV2 + * + * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does + */ + void(ORT_API_CALL* ReleaseTensorRTProviderOptions)(_Frees_ptr_opt_ OrtTensorRTProviderOptionsV2* input); + + /// @} + /// \name OrtSessionOptions + /// @{ + + /** \brief Enable custom operators + * + * See onnxruntime-extensions: https://github.com/microsoft/onnxruntime-extensions.git + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(EnableOrtCustomOps, _Inout_ OrtSessionOptions* options); + + /// @} + /// \name OrtAllocator + /// @{ + + /** \brief Register a custom allocator + * + * Enables sharing between multiple sessions that use the same env instance. + * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered. + * + * The behavior of this is exactly the same as OrtApi::CreateAndRegisterAllocator except + * instead of ORT creating an allocator based on provided info, in this case + * ORT uses the user-provided custom allocator. + * See https://onnxruntime.ai/docs/get-started/with-c.html for details. + * + * \param[in] env + * \param[in] allocator User provided allocator + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(RegisterAllocator, _Inout_ OrtEnv* env, _In_ OrtAllocator* allocator); + + /** \brief Unregister a custom allocator + * + * It is an error if you provide an ::OrtMemoryInfo not corresponding to any + * registered allocators for sharing. + * + * \param[in] env + * \param[in] mem_info + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(UnregisterAllocator, _Inout_ OrtEnv* env, + _In_ const OrtMemoryInfo* mem_info); + + /// @} + /// \name OrtValue + /// @{ + + /** \brief Sets *out to 1 iff an ::OrtValue is a SparseTensor, and 0 otherwise + * + * \param[in] value existing ::OrtValue + * \param[out] out unless an error occurs, contains 1 iff the value contains an instance + * of sparse tensor or 0 otherwise. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(IsSparseTensor, _In_ const OrtValue* value, _Out_ int* out); + + /** \brief Create an ::OrtValue with a sparse tensor that is empty. + * + * Use FillSparseTensor() functions to populate sparse tensor with non-zero values and + * format specific indices data. + * Use ReleaseValue to destroy the sparse tensor, this will also release the buffer inside the output value + * if any was allocated. + * \param[in,out] allocator allocator to use when performing an allocation. Allocation will be performed + * by FillSparseTensor() APIs. The lifespan of the allocator instance must eclipse the lifespan + * this sparse tensor instance as the same allocator will be used to free memory. + * \param[in] dense_shape shape of the original dense tensor + * \param[in] dense_shape_len number of shape dimensions being passed + * \param[in] type must be one of TENSOR_ELEMENT_DATA_TYPE_xxxx + * \param[out] out Should be freed by calling ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateSparseTensorAsOrtValue, _Inout_ OrtAllocator* allocator, _In_ const int64_t* dense_shape, + size_t dense_shape_len, ONNXTensorElementDataType type, _Outptr_ OrtValue** out); + + /** + * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue. + * This will allocate required memory and copy the supplied NNZ values and COO indices into that memory allocation. + * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue. + * + * \param[in,out] ort_value ::OrtValue to populate with data + * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified + * at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed. + * String data is assumed to be on CPU and will only be copied into a CPU allocated buffer. + * \param[in] values_shape pointer to values shape array + * \param[in] values_shape_len length of the values_shape + * \param[in] values pointer to an array of values. For strings, pass const char**. + * \param[in] indices_data pointer to a location of COO indices + * \param[in] indices_num number of COO indices + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(FillSparseTensorCoo, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info, + _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values, + _In_ const int64_t* indices_data, size_t indices_num); + + /** + * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue. + * This will allocate required memory and copy the supplied NNZ values and CSR indices into that memory allocation. + * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue. + * + * \param[in,out] ort_value ::OrtValue to populate with data + * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified + * at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed. + * String data is assumed to be on CPU and will only be copied into a CPU allocated buffer. + * \param[in] values_shape pointer to values shape array + * \param[in] values_shape_len length of the values_shape + * \param[in] values - pointer to an array of values. For strings, pass const char**. + * \param[in] inner_indices_data pointer to a location of CSR inner indices + * \param[in] inner_indices_num number of CSR inner indices + * \param[in] outer_indices_data pointer to a location of CSR outer indices + * \param[in] outer_indices_num number of CSR outer indices + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(FillSparseTensorCsr, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info, + _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values, + _In_ const int64_t* inner_indices_data, size_t inner_indices_num, + _In_ const int64_t* outer_indices_data, size_t outer_indices_num); + + /** + * This fills populates an empty tensor that was created using OrtApi::CreateSparseTensorAsOrtValue. + * This will allocate required memory and copy the supplied NNZ values and BlockSparse indices into that memory allocation. + * Memory allocation is performed using the allocator that was specified with OrtApi::CreateSparseTensorAsOrtValue. + * + * \param[in,out] ort_value ::OrtValue to populate with data + * \param[in] data_mem_info serves to identify the location of the data to be copied. If the allocator specified + * at the creation time has memory info that is not the same as mem_info argument to this function a X-device copy will be performed. + * String data is assumed to be on CPU and will only be copied into a CPU allocated buffer. + * \param[in] values_shape + * \param[in] values_shape_len + * \param[in] values structure with values information + * \param[in] indices_shape_data pointer to a location of indices shape + * \param[in] indices_shape_len length of the block sparse indices shape + * \param[in] indices_data pointer to a location of indices data. Shape will determine the length of the indices data. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(FillSparseTensorBlockSparse, _Inout_ OrtValue* ort_value, _In_ const OrtMemoryInfo* data_mem_info, + _In_ const int64_t* values_shape, size_t values_shape_len, _In_ const void* values, + _In_ const int64_t* indices_shape_data, size_t indices_shape_len, + _In_ const int32_t* indices_data); + + /** + * Create an ::OrtValue with a sparse tensor. This is the first step. + * Next, use UseIndices() functions to supply sparse tensor with + * format specific indices data and set its sparse format to a specific enum value. + * This will not perform memory allocations. It will + * use supplied user buffer which should outlive the created sparse tensor. + * Use OrtApi::ReleaseValue to destroy the sparse tensor. It would not release the supplied values buffer. + * This function can not be used to map strings from the user allocated memory. Strings must always be copied + * and have UTF-8 encoding. Therefore, use OrtApi::CreateSparseTensorAsOrtValue above and then fill it with data + * using appropriate Make*() function. + * + * \param[in] info memory info where sparse values reside. + * \param[in,out] p_data pointer to a user allocated buffer with values. To create a full sparse tensor with no non-zero + * values, pass nullptr + * \param[in] dense_shape shape of the original dense tensor + * \param[in] dense_shape_len number of shape dimensions being passed + * \param[in] values_shape shape of the values data. To create a fully sparse tensor with no non-zero values, + * pass {0} shape. + * \param[in] values_shape_len number of values shape dimensions + * \param[in] type must be one of TENSOR_ELEMENT_DATA_TYPE_xxxx + * \param[out] out Should be freed by calling ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(CreateSparseTensorWithValuesAsOrtValue, _In_ const OrtMemoryInfo* info, _Inout_ void* p_data, + _In_ const int64_t* dense_shape, size_t dense_shape_len, + _In_ const int64_t* values_shape, size_t values_shape_len, + ONNXTensorElementDataType type, _Outptr_ OrtValue** out); + + /** + * This assigns Coo format indices to the SparseTensor that was created by + * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to + * ORT_SPARSE_COO. This will not allocate any additional memory for data. The life span of + * indices_data buffer should eclipse the life span of this ::OrtValue. + * + * \param[in,out] ort_value ::OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue + * \param[in,out] indices_data pointer to a user pre-allocated buffer or nullptr for fully sparse tensors. + * \param[in] indices_num number of COO indices. Should either be 0 for fully sparse tensors, be equal + * to the number of nnz values specified to OrtApi::CreateSparseTensorWithValuesAsOrtValue for 1-D {nnz} indices or + * be twice as number of nnz values for a 2-D indices {nnz, 2} + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(UseCooIndices, _Inout_ OrtValue* ort_value, _Inout_ int64_t* indices_data, size_t indices_num); + + /** + * The assigns CSR format indices to the SparseTensor that was created by + * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to + * ORT_SPARSE_CSRC. This will not allocate any additional memory for data. The life spans of + * inner_data and outer_data buffers should eclipse the life span of this ::OrtValue. + * + * \param[in,out] ort_value ::OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue + * \param[in,out] inner_data pointer to a user pre-allocated buffer or nullptr for fully sparse tensors. + * \param[in] inner_num number of inner CSR indices. Should either be 0 for fully sparse tensors or be equal + * to the number of nnz values specified to OrtApi::CreateSparseTensorWithValuesAsOrtValue. + * \param[in,out] outer_data pointer to user pre-allocated buffer or nullptr for fully sparse tensors. + * \param[in] outer_num number of CSR outer indices. Should either be 0 for fully sparse tensors or + * equal to rows + 1 of the dense shape. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(UseCsrIndices, _Inout_ OrtValue* ort_value, _Inout_ int64_t* inner_data, size_t inner_num, + _Inout_ int64_t* outer_data, size_t outer_num); + + /** + * The assigns BlockSparse format indices to the SparseTensor that was created by + * OrtApi::CreateSparseTensorWithValuesAsOrtValue above. It also sets OrtSparseFormat to + * ORT_SPARSE_BLOCK_SPARSE. This will not allocate any additional memory for data. The life span of + * indices_data buffer must eclipse the lifespan of this ::OrtValue. + * + * \param[in,out] ort_value OrtValue instance constructed with OrtApi::CreateSparseTensorWithValuesAsOrtValue + * \param[in] indices_shape pointer to indices shape. Use {0} for fully sparse tensors + * \param[in] indices_shape_len length of the indices shape + * \param[in,out] indices_data pointer to user pre-allocated buffer or nullptr for fully sparse tensors. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(UseBlockSparseIndices, _Inout_ OrtValue* ort_value, const int64_t* indices_shape, size_t indices_shape_len, _Inout_ int32_t* indices_data); + + /** \brief Returns sparse tensor format enum iff a given ort value contains an instance of sparse tensor. + * + * \param[in] ort_value ::OrtValue that contains an instance of sparse tensor + * \param[out] out pointer to out parameter + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSparseTensorFormat, _In_ const OrtValue* ort_value, _Out_ enum OrtSparseFormat* out); + + /** \brief Returns data type and shape of sparse tensor values (nnz) iff ::OrtValue contains a SparseTensor. + * + * \param[in] ort_value An ::OrtValue that contains a fully constructed sparse tensor + * \param[out] out Must be freed by OrtApi::ReleaseTensorTypeAndShapeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSparseTensorValuesTypeAndShape, _In_ const OrtValue* ort_value, _Outptr_ OrtTensorTypeAndShapeInfo** out); + + /** \brief Returns numeric data for sparse tensor values (nnz). For string values use GetStringTensor*(). + * + * \param[in] ort_value an instance of ::OrtValue containing sparse tensor + * \param[out] out returns a pointer to values data. Do not attempt to free this ptr. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSparseTensorValues, _In_ const OrtValue* ort_value, _Outptr_ const void** out); + + /** \brief Returns data type, shape for the type of indices specified by indices_format. + * + * \param[in] ort_value ::OrtValue containing sparse tensor. + * \param[in] indices_format One of the indices formats. It is an error to request a format that the sparse + * tensor does not contain. + * \param[out] out an instance of ::OrtTensorTypeAndShapeInfo. Must be freed by OrtApi::ReleaseTensorTypeAndShapeInfo + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSparseTensorIndicesTypeShape, _In_ const OrtValue* ort_value, enum OrtSparseIndicesFormat indices_format, _Outptr_ OrtTensorTypeAndShapeInfo** out); + + /** \brief Returns indices data for the type of the indices specified by indices_format + * + * \param[in] ort_value ::OrtValue containing sparse tensor. + * \param[in] indices_format One of the indices formats. It is an error to request a format that the sparse tensor does not contain. + * \param[out] num_indices Pointer to where the number of indices entries is returned + * \param[out] indices Returned pointer to the indices data. Do not free the returned pointer as it refers to internal data owned by the ::OrtValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetSparseTensorIndices, _In_ const OrtValue* ort_value, enum OrtSparseIndicesFormat indices_format, _Out_ size_t* num_indices, _Outptr_ const void** indices); + /// @} + /// \name OrtSessionOptions + /// @{ + + /** + * \brief Sets out to 1 iff an optional type OrtValue has an element, 0 otherwise (OrtValue is None) + * Use this API to find if the optional type OrtValue is None or not. + * If the optional type OrtValue is not None, use the OrtValue just like any other OrtValue. + * For example, if you get an OrtValue that corresponds to Optional(tensor) and + * if HasValue() returns true, use it as tensor and so on. + + * \param[in] value Input OrtValue. + * \param[out] out indicating if the input OrtValue contains data (1) or if it is a None (0) + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(HasValue, _In_ const OrtValue* value, _Out_ int* out); + + /// @} + /// \name OrtKernelContext + /// Custom operator APIs. + /// @{ + + /** \brief Used for custom operators, gets the GPU compute stream to use to launch the custom a GPU kernel + * \see ::OrtCustomOp + * \param[in] context OrtKernelContext instance + * \param[out] out Returns pointer to a GPU compute stream that can be used to launch the custom GPU kernel. + * If retrieving the GPU compute stream is not relevant (GPU not enabled in the build, kernel partitioned to + * some other EP), then a nullptr is returned as the output param. + * Do not free or mutate the returned pointer as it refers to internal data owned by the underlying session. + * Only use it for custom kernel launching. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelContext_GetGPUComputeStream, _In_ const OrtKernelContext* context, _Outptr_ void** out); + + /// @} + /// \name GetTensorMemoryInfo + /// @{ + /** \brief Returns a pointer to the ::OrtMemoryInfo of a Tensor + * \param[in] value ::OrtValue containing tensor. + * \param[out] mem_info ::OrtMemoryInfo of the tensor. Do NOT free the returned pointer. It is valid for the lifetime of the ::OrtValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetTensorMemoryInfo, _In_ const OrtValue* value, _Out_ const OrtMemoryInfo** mem_info); + + /// @} + /// \name GetExecutionProviderApi + /// @{ + /** \brief Get a pointer to the requested version of the Execution Provider specific + * API extensions to the OrtApi + * \param[in] provider_name The name of the execution provider name. Currently only the following + * values are supported: "DML". + * \param[in] version Must be ::ORT_API_VERSION. + * \param[out] provider_api A void pointer containing a reference to the execution provider versioned api structure. + * For example, the provider_api pointer can be cast to the OrtDmlApi* when the provider_name is "DML". + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetExecutionProviderApi, _In_ const char* provider_name, _In_ uint32_t version, _Outptr_ const void** provider_api); + + /// @} + + /// \name SessionOptions + /// @{ + /** \brief Set custom thread creation function + * + * \param[in] options Session options + * \param[in] ort_custom_create_thread_fn Custom thread creation function + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsSetCustomCreateThreadFn, _Inout_ OrtSessionOptions* options, _In_ OrtCustomCreateThreadFn ort_custom_create_thread_fn); + + /** \brief Set creation options for custom thread + * + * \param[in] options Session options + * \param[in] ort_custom_thread_creation_options Custom thread creation options (can be nullptr) + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsSetCustomThreadCreationOptions, _Inout_ OrtSessionOptions* options, _In_ void* ort_custom_thread_creation_options); + + /** \brief Set custom thread join function + * + * \param[in] options Session options + * \param[in] ort_custom_join_thread_fn Custom join thread function, must not be nullptr when ort_custom_create_thread_fn is set + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SessionOptionsSetCustomJoinThreadFn, _Inout_ OrtSessionOptions* options, _In_ OrtCustomJoinThreadFn ort_custom_join_thread_fn); + /// @} + + /// \name OrtThreadingOptions + /// @{ + /** \brief Set custom thread creation function for global thread pools + * + * \param[inout] tp_options + * \param[in] ort_custom_create_thread_fn Custom thread creation function + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalCustomCreateThreadFn, _Inout_ OrtThreadingOptions* tp_options, _In_ OrtCustomCreateThreadFn ort_custom_create_thread_fn); + + /** \brief Set custom thread creation options for global thread pools + * + * \param[inout] tp_options + * \param[in] ort_custom_thread_creation_options Custom thread creation options (can be nullptr) + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalCustomThreadCreationOptions, _Inout_ OrtThreadingOptions* tp_options, _In_ void* ort_custom_thread_creation_options); + + /** \brief Set custom thread join function for global thread pools + * + * \param[inout] tp_options + * \param[in] ort_custom_join_thread_fn Custom thread join function, must not be nullptr when global ort_custom_create_thread_fn is set + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SetGlobalCustomJoinThreadFn, _Inout_ OrtThreadingOptions* tp_options, _In_ OrtCustomJoinThreadFn ort_custom_join_thread_fn); + /// @} + + /** \brief Synchronize bound inputs. The call may be necessary for some providers, such as cuda, + * in case the system that allocated bound memory operated on a different stream. However, the + * operation is provider specific and could be a no-op. + * + * \param[inout] binding_ptr + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SynchronizeBoundInputs, _Inout_ OrtIoBinding* binding_ptr); + + /** \brief Synchronize bound outputs. The call may be necessary for some providers, such as cuda, + * in case the system that allocated bound memory operated on a different stream. However, the + * operation is provider specific and could be a no-op. + * + * \param[inout] binding_ptr + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(SynchronizeBoundOutputs, _Inout_ OrtIoBinding* binding_ptr); + + /// \name OrtSessionOptions + /// @{ + + /** \brief Append CUDA execution provider to the session options + * + * If CUDA is not available (due to a non CUDA enabled build), this function will return failure. + * + * This is slightly different from OrtApi::SessionOptionsAppendExecutionProvider_CUDA, it takes an + * ::OrtCUDAProviderOptions which is publicly defined. This takes an opaque ::OrtCUDAProviderOptionsV2 + * which must be created with OrtApi::CreateCUDAProviderOptions. + * + * For OrtApi::SessionOptionsAppendExecutionProvider_CUDA, the user needs to instantiate ::OrtCUDAProviderOptions + * as well as allocate/release buffers for some members of ::OrtCUDAProviderOptions. + * Here, OrtApi::CreateCUDAProviderOptions and Ortapi::ReleaseCUDAProviderOptions will do the memory management for you. + * + * \param[in] options + * \param[in] cuda_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.11. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CUDA_V2, + _In_ OrtSessionOptions* options, _In_ const OrtCUDAProviderOptionsV2* cuda_options); + + /// @} + /// \name OrtCUDAProviderOptionsV2 + /// @{ + + /** \brief Create an OrtCUDAProviderOptionsV2 + * + * \param[out] out Newly created ::OrtCUDAProviderOptionsV2. Must be released with OrtApi::ReleaseCudaProviderOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.11. + */ + ORT_API2_STATUS(CreateCUDAProviderOptions, _Outptr_ OrtCUDAProviderOptionsV2** out); + + /** \brief Set options in a CUDA Execution Provider. + * + * Please refer to https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#configuration-options + * to know the available keys and values. Key should be in null terminated string format of the member of ::OrtCUDAProviderOptionsV2 + * and value should be its related range. Recreates the options and only sets the supplied values. + * + * For example, key="device_id" and value="0" + * + * \param[in] cuda_options + * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys + * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values + * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.11. + */ + ORT_API2_STATUS(UpdateCUDAProviderOptions, _Inout_ OrtCUDAProviderOptionsV2* cuda_options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** + * Get serialized CUDA provider options string. + * + * For example, "device_id=0;arena_extend_strategy=0;......" + * + * \param cuda_options - OrtCUDAProviderOptionsV2 instance + * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions() + * the specified allocator will be used to allocate continuous buffers for output strings and lengths. + * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.11. + */ + ORT_API2_STATUS(GetCUDAProviderOptionsAsString, _In_ const OrtCUDAProviderOptionsV2* cuda_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr); + + /** \brief Release an ::OrtCUDAProviderOptionsV2 + * + * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does + * + * \since Version 1.11. + */ + void(ORT_API_CALL* ReleaseCUDAProviderOptions)(_Frees_ptr_opt_ OrtCUDAProviderOptionsV2* input); + + /// @} + + /** \brief Append MIGraphX provider to session options + * + * If MIGraphX is not available (due to a non MIGraphX enabled build, or if MIGraphX is not installed on the system), this function will return failure. + * + * \param[in] options + * \param[in] migraphx_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.11. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_MIGraphX, + _In_ OrtSessionOptions* options, _In_ const OrtMIGraphXProviderOptions* migraphx_options); + + /** \brief Replace initialized Tensors with external data with the data provided in initializers. + * + * The function will find the initialized TensorProtos with external data in the graph with the provided names and + * replace them with the provided tensors. The API verifies that the TensorProto being replaced + * has an external data reference and has the same name, dimensions and data type as its replacement. The replacement + * will occur before any of the optimizations take place. The data will be copied into the graph + * since TensorProto can't refer to the user provided buffers. + * + * Once the model has been loaded, the OrtValue(s) added to SessionOptions instance will be removed + * from the internal SessionOptions copy to save memory, the user provided buffers can then be deallocated + * and the SessionOptions instance that refers to them can be destroyed. + * + * \param[in] options + * \param[in] initializer_names Array of null terminated UTF-8 encoded strings of the initializers names. + * \param[in] initializers Array of ::OrtValue type + * \param[in] num_initializers Number of elements in the initializer_names and initializers + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.12. + */ + ORT_API2_STATUS(AddExternalInitializers, _In_ OrtSessionOptions* options, + _In_reads_(num_initializers) const char* const* initializer_names, + _In_reads_(num_initializers) const OrtValue* const* initializers, size_t num_initializers); + + /** \brief: Create attribute of onnxruntime operator + * + * \param[in] name Name of the attribute + * \param[in] data Data content of the attribute + * \param[in] len Number of bytes stored in data + * \param[in] type Data type + * \param[out] op_attr Attribute that has been created, which must be released by OrtApi::ReleaseOpAttr + * + * \since Version 1.12. + */ + ORT_API2_STATUS(CreateOpAttr, + _In_ const char* name, + _In_ const void* data, + _In_ int len, + _In_ OrtOpAttrType type, + _Outptr_ OrtOpAttr** op_attr); + + /* \brief: Release op attribute + * + * \param[in] opAttr Attribute created by OrtApi::CreateOpAttr + * + * \since Version 1.12. + */ + ORT_CLASS_RELEASE(OpAttr); + + /** \brief: Create onnxruntime native operator + * + * \param[in] info Kernel info + * \param[in] op_name Operator name + * \param[in] domain Operator domain + * \param[in] version Operator opset version + * \param[in] type_constraint_names Name of the type contraints, such as "T" or "T1" + * \param[in] type_constraint_values Type of each contraints + * \param[in] type_constraint_count Number of contraints + * \param[in] attr_values Attributes used to initialize the operator + * \param[in] attr_count Number of the attributes + * \param[in] input_count Number of inputs + * \param[in] output_count Number of outputs + * \param[out] ort_op Operator that has been created + * + * \since Version 1.12. + */ + ORT_API2_STATUS(CreateOp, + _In_ const OrtKernelInfo* info, + _In_z_ const char* op_name, + _In_z_ const char* domain, + int version, + _In_reads_(type_constraint_count) const char** type_constraint_names, + _In_reads_(type_constraint_count) const ONNXTensorElementDataType* type_constraint_values, + int type_constraint_count, + _In_reads_(attr_count) const OrtOpAttr* const* attr_values, + int attr_count, + int input_count, + int output_count, + _Outptr_ OrtOp** ort_op); + + /** \brief: Invoke the operator created by OrtApi::CreateOp + * The inputs must follow the order as specified in onnx specification + * + * \param[in] context Kernel context + * \param[in] ort_op Operator that has been created + * \param[in] input_values Array of inputs + * \param[in] input_count Number of inputs + * \param[in] output_values Array of outputs + * \param[in] output_count Number of outputs + * + * \since Version 1.12. + */ + ORT_API2_STATUS(InvokeOp, + _In_ const OrtKernelContext* context, + _In_ const OrtOp* ort_op, + _In_ const OrtValue* const* input_values, + _In_ int input_count, + _Inout_ OrtValue* const* output_values, + _In_ int output_count); + + /* \brief: Release an onnxruntime operator + * + * \param[in] Op Operator created by OrtApi::CreateOp + * + * \since Version 1.12. + */ + ORT_CLASS_RELEASE(Op); + + /** \brief: Append execution provider to the session options. + * \param[in] options + * \param[in] provider_name - provider to add. + * \param[in] provider_options_keys - keys to configure the provider options + * \param[in] provider_options_values - values to configure the provider options + * \param[in] num_keys - number of keys passed in + * + * Currently supported provider names: + * QNNExecutionProvider (or QNN) + * OpenVINOExecutionProvider (or OpenVINO) + * XnnpackExecutionProvider (or XNNPACK) + * WebNNExecutionProvider (or WEBNN) + * WebGpuExecutionProvider (or WebGPU) + * AzureExecutionProvider (or AZURE) + * JsExecutionProvider (or JS) + * VitisAIExecutionProvider (or VitisAI) + * CoreMLExecutionProvider (or CoreML) + * + * Note: If an execution provider has a dedicated SessionOptionsAppendExecutionProvider_ function + * that should be used to add it. + * + * QNN supported keys: + * "backend_type": Type of QNN backend. Specifies a backend path that is the associated QNN backend library file + * name. E.g., given backend type "htp", on Windows, the backend path would be "QnnHtp.dll", and on other + * platforms, it would be "libQnnHtp.so". Mutually exclusive with "backend_path". + * Available options: + * -# "cpu" + * -# "gpu" + * -# "htp": Default. + * -# "saver" + * "backend_path": File path to QNN backend library. Mutually exclusive with "backend_type". + * "profiling_level": QNN profiling level. + * Available options: + * -# "off": Default. + * -# "basic" + * -# "detailed" + * "profiling_file_path": QNN profiling file path if ETW not enabled. + * "rpc_control_latency": QNN RPC control latency. + * "vtcm_mb": QNN VTCM size in MB. default to 0(not set). + * "htp_performance_mode": QNN performance mode. + * Available options: + * -# "burst" + * -# "balanced" + * -# "default": Default. + * -# "high_performance" + * -# "high_power_saver" + * -# "low_balanced" + * -# "extreme_power_saver" + * -# "low_power_saver" + * -# "power_saver" + * -# "sustained_high_performance" + * "qnn_saver_path": File path to the QNN Saver backend library. If specified, QNN Saver will be enabled and will + * dump QNN API calls to disk for replay/debugging. QNN Saver produces incorrect model inference results and + * may alter model/EP partitioning. Use only for debugging. + * "qnn_context_priority": QNN context priority. + * Available options: + * -# "low" + * -# "normal": Default. + * -# "normal_high" + * -# "high" + * "htp_graph_finalization_optimization_mode": Set the optimization mode for graph finalization on the HTP backend. + * Available options: + * -# "0": Default. + * -# "1": Faster preparation time, less optimal graph. + * -# "2": Longer preparation time, more optimal graph. + * -# "3": Longest preparation time, most likely even more optimal graph. See QNN SDK documentation for specific + * details. + * "soc_model": The SoC model number. Refer to the QNN SDK documentation for valid values. + * Defaults to "0" (unknown). + * "htp_arch": The minimum HTP architecture the driver will use to select compatible QNN operators. + * Available options: + * -# "0": Default (none). + * -# "68" + * -# "69" + * -# "73" + * -# "75" + * "device_id": The ID of the device to use when setting 'htp_arch'. Defaults to "0" (for single device). + * "enable_htp_fp16_precision": Used for float32 model for HTP backend. + * Enable the float32 model to be inferenced with fp16 precision. Otherwise, it will be fp32 precision. + * -# "0": With fp32 precision. + * -# "1": Default. With fp16 precision. + * "offload_graph_io_quantization": Offload graph input quantization and graph output dequantization to another + * execution provider (typically CPU EP). + * -# "0": Disabled. QNN EP will handle quantization and dequantization of graph I/O. + * -# "1": Enabled. This is the default value. + * "enable_htp_spill_fill_buffer": Enable HTP spill fill buffer setting. The flag is used while generating context + * binary. + * -# "0": Default. Disabled. + * -# "1": Enabled. + * "enable_htp_shared_memory_allocator": Enable the QNN HTP shared memory allocator. Requires libcdsprpc.so/dll to + * be available. + * -# "0": Default. Disabled. + * -# "1": Enabled. + * "dump_json_qnn_graph": Set to "1" to dump QNN graphs generated by QNN EP as JSON files. Each graph partition + * assigned to QNN EP is dumped to a separate file. + * "json_qnn_graph_dir": Directory in which to dump QNN JSON graphs. If not specified, QNN graphs are dumped in the + * program's current working directory. Ignored if "dump_json_qnn_graph" is not set. + * + * XNNPACK supported keys: + * "intra_op_num_threads": number of thread-pool size to use for XNNPACK execution provider. + * default value is 0, which means to use the session thread-pool size. + * + * \since Version 1.12. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider, _In_ OrtSessionOptions* options, + _In_ const char* provider_name, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /* \brief: Get a copy of kernel info + * + * \param[in] info Kernel info + * \param[out] info_copy Copy of kernel info + * + * \since Version 1.12. + */ + ORT_API2_STATUS(CopyKernelInfo, + _In_ const OrtKernelInfo* info, + _Outptr_ OrtKernelInfo** info_copy); + + /* \brief: Release kernel info + * + * \param[in] KernelInfo A copy of kernel info returned by CopyKernelInfo + * + * \since Version 1.12. + */ + ORT_CLASS_RELEASE(KernelInfo); + + /// \name Ort Training + /// @{ + /** \brief Gets the Training C Api struct + * + * Call this function to access the ::OrtTrainingApi structure that holds pointers to functions that enable + * training with onnxruntime. + * \note A NULL pointer will be returned and no error message will be printed if the training api + * is not supported with this build. A NULL pointer will be returned and an error message will be + * printed if the provided version is unsupported, for example when using a runtime older than the + * version created with this header file. + * + * \param[in] version Must be ::ORT_API_VERSION + * \return The ::OrtTrainingApi struct for the version requested. + * + * \since Version 1.13 + */ + const OrtTrainingApi*(ORT_API_CALL* GetTrainingApi)(uint32_t version)NO_EXCEPTION; + + /// @} + + /** \brief Append CANN provider to session options + * + * If CANN is not available (due to a non CANN enabled build, or if CANN is not installed on the system), this function will return failure. + * + * \param[in] options + * \param[in] cann_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.13. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_CANN, + _In_ OrtSessionOptions* options, _In_ const OrtCANNProviderOptions* cann_options); + + /** \brief Create an OrtCANNProviderOptions + * + * \param[out] out created ::OrtCANNProviderOptions. Must be released with OrtApi::ReleaseCANNProviderOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.13. + */ + ORT_API2_STATUS(CreateCANNProviderOptions, _Outptr_ OrtCANNProviderOptions** out); + + /** \brief Set options in a CANN Execution Provider. + * + * \param[in] cann_options + * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys + * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values + * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.13. + */ + ORT_API2_STATUS(UpdateCANNProviderOptions, _Inout_ OrtCANNProviderOptions* cann_options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** \brief Get serialized CANN provider options string. + * + * \param[in] cann_options OrtCANNProviderOptions instance + * \param[in] allocator a ptr to an instance of OrtAllocator obtained with CreateAllocator() + * or GetAllocatorWithDefaultOptions(), the specified allocator will be used to allocate + * continuous buffers for output strings and lengths. + * \param[out] ptr is a UTF-8 null terminated string allocated using 'allocator'. + * The caller is responsible for using the same allocator to free it. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.13. + */ + ORT_API2_STATUS(GetCANNProviderOptionsAsString, _In_ const OrtCANNProviderOptions* cann_options, + _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr); + + /** \brief Release an OrtCANNProviderOptions + * + * \param[in] input The pointer of OrtCANNProviderOptions which will been deleted + * + * \since Version 1.13. + */ + void(ORT_API_CALL* ReleaseCANNProviderOptions)(_Frees_ptr_opt_ OrtCANNProviderOptions* input); + + /* \brief Get OrtDevice type from MemoryInfo + * + * \since Version 1.14 + */ + void(ORT_API_CALL* MemoryInfoGetDeviceType)(_In_ const OrtMemoryInfo* ptr, _Out_ OrtMemoryInfoDeviceType* out); + + /* \brief Update the OrtEnv instance with custom log severity level + * + * \param[in] ort_env The OrtEnv instance being used + * \param[in] log_severity_level The log severity level. + * + * \since Version 1.14. + */ + ORT_API2_STATUS(UpdateEnvWithCustomLogLevel, _In_ OrtEnv* ort_env, OrtLoggingLevel log_severity_level); + + /* \brief Set affinities for intra op threads + * + * Affinity string follows format: + * logical_processor_id,logical_processor_id;logical_processor_id,logical_processor_id + * Semicolon isolates configurations among threads, while comma split processors where ith thread expected to attach to. + * e.g. 1,2,3;4,5 + * specifies affinities for two threads, with the 1st thread attach to the 1st, 2nd, and 3rd processor, and 2nd thread to the 4th and 5th. + * To ease the configuration, an "interval" is also allowed: + * e.g. 1-8;8-16;17-24 + * orders that the 1st thread runs on first eight processors, 2nd thread runs on next eight processors, and so forth. + * Note: + * 1. Once set, the number of thread affinities must equal to intra_op_num_threads - 1, + * ort does not set affinity on the main thread which is started and managed by the calling app; + * 2. For windows, ort will infer the group id from a logical processor id, for example, assuming there are two groups with each has 64 logical processors, + * an id of 64 will be inferred as the last processor of the 1st group, while 65 will be interpreted as the 1st processor of the second group. + * Hence 64-65 is an invalid configuration, because a windows thread cannot be attached to processors across group boundary. + * + * \since Version 1.14 + */ + ORT_API2_STATUS(SetGlobalIntraOpThreadAffinity, _Inout_ OrtThreadingOptions* tp_options, const char* affinity_string); + + /** \brief Register custom ops from a shared library. + * + * Loads a shared library (.dll on windows, .so on linux, etc) named 'library_name' and looks for this entry point: + * OrtStatus* RegisterCustomOps(OrtSessionOptions * options, const OrtApiBase* api); + * It then passes in the provided session options to this function along with the api base. + * + * The handle to the loaded library is automatically released by ORT when the last OrtSession that references the + * library handle is released. If no OrtSession is created, then the library handle is released when the provided + * OrtSessionOptions is released. + * + * \param[in] options The session options. + * \param[in] library_name The name of the shared library to load and register. Refer to OS-specific dynamic library + * loading utilities (e.g., LoadLibraryEx on Windows or dlopen on Linux/MacOS) for information + * on the format of library names and search paths. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(RegisterCustomOpsLibrary_V2, _Inout_ OrtSessionOptions* options, _In_ const ORTCHAR_T* library_name); + + /** \brief Register custom ops by calling a RegisterCustomOpsFn function. + * + * Searches for registration_func_name and if found calls it. + * + * The library containing the function must either be linked against or previously loaded by the executable. + * + * If you want ONNX Runtime to load the library and manage its lifetime, use RegisterCustomOpsLibrary_V2. + * + * RegisterCustomOpsUsingFunction can be used in scenarios where it may not be possible for ONNX Runtime to load + * the library from a path. e.g. mobile platforms where the library must be linked into the app. + * + * The registration function must have the signature of RegisterCustomOpsFn: + * OrtStatus* (*fn)(OrtSessionOptions* options, const OrtApiBase* api); + * + * See https://onnxruntime.ai/docs/reference/operators/add-custom-op.html for details on how the registration + * function should be implemented. + * + * \param[in] options OrtSessionOptions that is passed through as the first argument in the call to the + * registration function. + * \param[in] registration_func_name Name of registration function to use. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(RegisterCustomOpsUsingFunction, _Inout_ OrtSessionOptions* options, + _In_ const char* registration_func_name); + + /// \name OrtKernelInfo + /// Custom operator APIs. + /// @{ + + /** \brief Get the number of inputs from ::OrtKernelInfo. + * + * Used in the CreateKernel callback of an OrtCustomOp to query the number of inputs + * during kernel/session creation. + * + * \param[in] info Instance of ::OrtKernelInfo. + * \param[out] out Pointer to variable assigned with the result on success. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(KernelInfo_GetInputCount, _In_ const OrtKernelInfo* info, _Out_ size_t* out); + + /** \brief Get the number of outputs from ::OrtKernelInfo. + * + * Used in the CreateKernel callback of an OrtCustomOp to query the number of outputs + * during kernel/session creation. + * + * \param[in] info Instance of ::OrtKernelInfo. + * \param[out] out Pointer to variable assigned with the result on success. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(KernelInfo_GetOutputCount, _In_ const OrtKernelInfo* info, _Out_ size_t* out); + + /** \brief Get the name of a ::OrtKernelInfo's input. + * + * Used in the CreateKernel callback of an OrtCustomOp to query an input's name + * during kernel/session creation. + * + * If `out` is nullptr, the value of `size` is set to the size of the name + * string (including null-terminator), and a success status is returned. + * + * If the `size` parameter is greater than or equal to the name string's size, + * the value of `size` is set to the true size of the string (including null-terminator), + * the provided memory is filled with the string's contents, and a success status is returned. + * + * If the `size` parameter is less than the actual string's size and `out` + * is not nullptr, the value of `size` is set to the true size of the string + * and a failure status is returned. + * + * \param[in] info An instance of ::OrtKernelInfo. + * \param[in] index The index of the input name to get. Returns a failure status if out-of-bounds. + * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the input's name. + * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(KernelInfo_GetInputName, _In_ const OrtKernelInfo* info, size_t index, _Out_ char* out, + _Inout_ size_t* size); + + /** \brief Get the name of a ::OrtKernelInfo's output. + * + * Used in the CreateKernel callback of an OrtCustomOp to query an output's name + * during kernel/session creation. + * + * If `out` is nullptr, the value of `size` is set to the size of the name + * string (including null-terminator), and a success status is returned. + * + * If the `size` parameter is greater than or equal to the name string's size, + * the value of `size` is set to the true size of the string (including null-terminator), + * the provided memory is filled with the string's contents, and a success status is returned. + * + * If the `size` parameter is less than the actual string's size and `out` + * is not nullptr, the value of `size` is set to the true size of the string + * and a failure status is returned. + * + * \param[in] info An instance of ::OrtKernelInfo. + * \param[in] index The index of the output name to get. Returns a failure status if out-of-bounds. + * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the output's + * name. + * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(KernelInfo_GetOutputName, _In_ const OrtKernelInfo* info, size_t index, _Out_ char* out, + _Inout_ size_t* size); + + /** \brief Get the type information for a ::OrtKernelInfo's input. + * + * Used in the CreateKernel callback of an OrtCustomOp to query the shape and type information + * of an input during kernel/session creation. + * + * \param[in] info An instance of ::OrtKernelInfo. + * \param[in] index Which input to get the type information for + * \param[out] type_info Pointer set to the resulting ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(KernelInfo_GetInputTypeInfo, _In_ const OrtKernelInfo* info, size_t index, + _Outptr_ OrtTypeInfo** type_info); + + /** \brief Get the type information for a ::OrtKernelInfo's output. + * + * Used in the CreateKernel callback of an OrtCustomOp to query the shape and type information + * of an output during kernel/session creation. + * + * \param[in] info An instance of ::OrtKernelInfo. + * \param[in] index Which input to get the type information for + * \param[out] type_info Pointer set to the resulting ::OrtTypeInfo. Must be freed with OrtApi::ReleaseTypeInfo. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(KernelInfo_GetOutputTypeInfo, _In_ const OrtKernelInfo* info, size_t index, + _Outptr_ OrtTypeInfo** type_info); + + /** \brief Get a ::OrtValue tensor stored as an attribute in the graph node. + * + * Used in the CreateKernel callback of an OrtCustomOp to get a tensor attribute. + * + * \param[in] info ::OrtKernelInfo instance. + * \param[in] name UTF-8 null-terminated string representing the attribute's name. + * \param[in] allocator Allocator used to allocate the internal tensor state. + * \param[out] out Returns newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue, + * which will also free internal tensor state allocated with the provided allocator. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(KernelInfoGetAttribute_tensor, _In_ const OrtKernelInfo* info, _In_z_ const char* name, + _Inout_ OrtAllocator* allocator, _Outptr_ OrtValue** out); + + /// @} + /// \name OrtSessionOptions + /// Custom operator APIs + /// @{ + + /** \brief Checks if the given session configuration entry exists. + * + * The config_key formats are defined in onnxruntime_session_options_config_keys.h + * + * Can be used in a custom operator library to check for session configuration entries + * that target one or more custom operators in the library. Example: The config entry + * custom_op.myop.some_key targets a custom op named "myop". + * + * \param[in] options The ::OrtSessionOptions instance. + * \param[in] config_key A null-terminated UTF-8 string representation of the configuration key. + * \param[out] out Pointer set to 1 if the entry exists and 0 otherwise. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(HasSessionConfigEntry, _In_ const OrtSessionOptions* options, + _In_z_ const char* config_key, _Out_ int* out); + + /** \brief Get a session configuration value. + * + * Returns a failure status if the configuration key does not exist. + * The config_key and the format of config_value are defined in onnxruntime_session_options_config_keys.h + * + * If `config_value` is nullptr, the value of `size` is set to the true size of the string + * value (including null-terminator), and a success status is returned. + * + * If the `size` parameter is greater than or equal to the actual string value's size, + * the value of `size` is set to the true size of the string value, the provided memory + * is filled with the value's contents, and a success status is returned. + * + * If the `size` parameter is less than the actual string value's size and `config_value` + * is not nullptr, the value of `size` is set to the true size of the string value + * and a failure status is returned. + * + * Can be used in a custom operator library to get session configuration entries + * that target one or more custom operators in the library. Example: The config entry + * custom_op.myop.some_key targets a custom op named "myop". + * + * \param[in] options The session options. + * \param[in] config_key A null-terminated UTF-8 string representation of the config key. + * \param[in] config_value Pointer to memory where the null-terminated UTF-8 string value will be stored. + * \param[in,out] size Pointer to the size of the `config_value` buffer. See above comments for details. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.14 + */ + ORT_API2_STATUS(GetSessionConfigEntry, _In_ const OrtSessionOptions* options, + _In_z_ const char* config_key, _Out_ char* config_value, _Inout_ size_t* size); + + /// @} + + /** \brief Append dnnl provider to session options + * + * If oneDNN is not available, this function will return failure. + * + * \param[in] options + * \param[in] dnnl_options + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_Dnnl, + _In_ OrtSessionOptions* options, _In_ const OrtDnnlProviderOptions* dnnl_options); + + /** \brief Create an OrtDnnlProviderOptions + * + * \param[out] out Newly created ::OrtDnnlProviderOptions. Must be released with OrtApi::ReleaseDnnlProviderOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(CreateDnnlProviderOptions, _Outptr_ OrtDnnlProviderOptions** out); + + /** \brief Set options in a oneDNN Execution Provider. + * + * Key should be in null terminated string format of the member of ::OrtDnnlProviderOptions + * and value should be its related range. + * + * For example, key="use_arena" and value="1" + * + * \param[in] dnnl_options + * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys + * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values + * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(UpdateDnnlProviderOptions, _Inout_ OrtDnnlProviderOptions* dnnl_options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** + * Get serialized oneDNN provider options string. + * + * For example, "use_arena=1;......" + * + * \param dnnl_options - OrtDnnlProviderOptions instance + * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions() + * the specified allocator will be used to allocate continuous buffers for output strings and lengths. + * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(GetDnnlProviderOptionsAsString, _In_ const OrtDnnlProviderOptions* dnnl_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr); + + /** \brief Release an ::OrtDnnlProviderOptions + * + * \since Version 1.15. + */ + void(ORT_API_CALL* ReleaseDnnlProviderOptions)(_Frees_ptr_opt_ OrtDnnlProviderOptions* input); + + /// \name OrtKernelInfo + /// Custom operator APIs. + /// @{ + + /** \brief Get the graph node name from ::OrtKernelInfo. + * + * If `out` is nullptr, the value of `size` is set to the size of the name + * string (including null-terminator), and a success status is returned. + * + * If the `size` parameter is greater than or equal to the name string's size, + * the value of `size` is set to the true size of the string (including null-terminator), + * the provided memory is filled with the string's contents, and a success status is returned. + * + * If the `size` parameter is less than the actual string's size and `out` + * is not nullptr, the value of `size` is set to the true size of the string + * and a failure status is returned. + * + * Can be used in a custom operator's CreateKernel callback to get the name of the operator's node name in the graph. + * + * \param[in] info An instance of ::OrtKernelInfo. + * \param[out] out Memory location into which to write the UTF-8 null-terminated string representing the name. + * \param[in,out] size Pointer to the size of the `out` buffer. See above comments for details. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.15 + */ + ORT_API2_STATUS(KernelInfo_GetNodeName, _In_ const OrtKernelInfo* info, _Out_ char* out, _Inout_ size_t* size); + + /** \brief Get the session logger from ::OrtKernelInfo. + * + * Used in the CreateKernel callback of an OrtCustomOp to get a logger that can be used to log + * messages. + * + * \param[in] info An instance of ::OrtKernelInfo. + * \param[out] logger Pointer set to the session's ::OrtLogger. Owned by ONNX Runtime, so do not free. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.15 + */ + ORT_API2_STATUS(KernelInfo_GetLogger, _In_ const OrtKernelInfo* info, _Outptr_ const OrtLogger** logger); + + /// @} + /// \name OrtKernelContext + /// Custom operator APIs. + /// @{ + + /** \brief Get the runtime logger from ::OrtKernelContext. + * + * Used in the KernelCompute callback of an OrtCustomOp to get a logger that can be used to log + * messages during inference. + * + * \param[in] context An instance of ::OrtKernelContext. + * \param[out] logger Pointer set to the kernel context's ::OrtLogger. Owned by ONNX Runtime, so do not free. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.15 + */ + ORT_API2_STATUS(KernelContext_GetLogger, _In_ const OrtKernelContext* context, _Outptr_ const OrtLogger** logger); + + /// @} + /// \name OrtLogger + /// Custom operator APIs. + /// @{ + + /** \brief Logs a message at the given severity level using the provided ::OrtLogger. + * + * Only messages with a severity level equal or greater than the ::OrtLogger's logging severity level + * are logged. Use OrtApi::Logger_GetLoggingSeverityLevel to get the ::OrtLogger's logging severity + * level. + * + * Can be used in custom operators to log messages with the logger retrieved via OrtApi::KernelInfo_GetLogger. + * + * \param[in] logger The ::OrtLogger instance. + * \param[in] log_severity_level The message's severity level. + * \param[in] message The message to log. + * \param[in] file_path The filepath of the file in which the message is logged. Usually the value of ORT_FILE. + * \param[in] line_number The file line number in which the message is logged. Usually the value of __LINE__. + * \param[in] func_name The name of the function in which the message is logged. Usually the value of __FUNCTION__. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.15 + */ + ORT_API2_STATUS(Logger_LogMessage, _In_ const OrtLogger* logger, OrtLoggingLevel log_severity_level, + _In_z_ const char* message, _In_z_ const ORTCHAR_T* file_path, int line_number, + _In_z_ const char* func_name); + + /** \brief Get the logging severity level of the ::OrtLogger. + * + * Can be used in a custom operator to get the logging serverity level of the ::OrtLogger associated with + * the ::OrtKernelInfo. + * + * \param[in] logger The ::OrtLogger instance. + * \param[out] out Pointer to variable assigned with the logging severity level on success. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.15 + */ + ORT_API2_STATUS(Logger_GetLoggingSeverityLevel, _In_ const OrtLogger* logger, _Out_ OrtLoggingLevel* out); + + /// @} + + /** \brief Get a ::OrtValue tensor stored as a constant initializer in the graph node. + * + * Used in the CreateKernel callback of an OrtCustomOp to get a tensor value. + * + * \param[in] info ::OrtKernelInfo instance. + * \param[in] index The node index. + * \param[out] is_constant Is it a constant node input or not. + * \param[out] out The OrtValue tensor value. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(KernelInfoGetConstantInput_tensor, _In_ const OrtKernelInfo* info, size_t index, _Out_ int* is_constant, _Outptr_ const OrtValue** out); + + /** \brief Get Optional Type information from an ::OrtTypeInfo + * + * This augments ::OrtTypeInfo to return an ::OrtOptionalTypeInfo when the type is optional. + * The OrtOptionalTypeInfo also has a nested ::OrtTypeInfo that describes the type of the optional value. + * ::OrtOptionalTypeInfo type can only appear within model metadata to describe inputs/outputs. + * The actual OrtValues that are supplied in place of optional type inputs should contain + * specific type that is described by ::OrtOptionalTypeInfo. + * + * So the picture: ::OrtTypeInfo -> ::OrtOptionalTypeInfo -> ::OrtTypeInfo (describes the type that can be supplied + * in place of the optional type when creating the actual ::OrtValue). + * + * \param[in] type_info + * \param[out] out A pointer to the ::OrtOptionalTypeInfo. Do not free this value, + * it is owned by OrtTypeInfo instance. When the type_info does not represent + * optional type, nullptr is returned in out. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(CastTypeInfoToOptionalTypeInfo, _In_ const OrtTypeInfo* type_info, + _Outptr_result_maybenull_ const OrtOptionalTypeInfo** out); + + /** \brief Get OrtTypeInfo for the allowed contained type from an ::OrtOptionalTypeInfo. + * + * This augments ::OrtOptionalTypeInfo to return an ::OrtTypeInfo for the contained type. + * The OrtOptionalTypeInfo has a nested ::OrtTypeInfo that describes the type of the optional value. + * ::OrtOptionalTypeInfo type can only appear within model metadata to describe inputs/outputs. + * The actual OrtValues that are supplied in place of optional type inputs should contain + * specific type that is described by the returned ::OrtTypeInfo. + * + * \param[in] optional_type_info + * \param[out] out A copy of ::OrtTypeInfo for what the optional value could be. + * The user must free this value with ReleaseTypeInfo. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(GetOptionalContainedTypeInfo, _In_ const OrtOptionalTypeInfo* optional_type_info, + _Outptr_ OrtTypeInfo** out); + + /** \brief Set a single string in a string tensor + * Do not zero terminate the string data. + * + * \param[in] value A string tensor + * \param[in] index - flat index of the element + * \param[in] length_in_bytes length of the buffer in utf-8 bytes (without the null terminator) + * \param[inout] buffer - address of return value + * + * \snippet{doc} snippets.dox OrtStatus Return Value + */ + ORT_API2_STATUS(GetResizedStringTensorElementBuffer, _Inout_ OrtValue* value, _In_ size_t index, _In_ size_t length_in_bytes, _Inout_ char** buffer); + + /** \brief Get Allocator from KernelContext for a specific memoryInfo. Please use C API ReleaseAllocator to release out object + * + * \param[in] context OrtKernelContext instance + * \param[in] mem_info OrtMemoryInfo instance + * \param[out] out A pointer to OrtAllocator. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.15. + */ + ORT_API2_STATUS(KernelContext_GetAllocator, _In_ const OrtKernelContext* context, _In_ const OrtMemoryInfo* mem_info, _Outptr_ OrtAllocator** out); + + /** \brief Returns a null terminated string of the build info including git info and cxx flags + * + * \return UTF-8 encoded version string. Do not deallocate the returned buffer. + * + * \since Version 1.15. + */ + const char*(ORT_API_CALL* GetBuildInfoString)(void); + + /// \name OrtROCMProviderOptions + /// @{ + + /** \brief Create an OrtROCMProviderOptions + * + * \param[out] out Newly created ::OrtROCMProviderOptions. Must be released with OrtApi::ReleaseROCMProviderOptions + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.16. + */ + ORT_API2_STATUS(CreateROCMProviderOptions, _Outptr_ OrtROCMProviderOptions** out); + + /** \brief Set options in a ROCm Execution Provider. + * + * Please refer to https://onnxruntime.ai/docs/execution-providers/ROCm-ExecutionProvider.html + * to know the available keys and values. Key should be in null terminated string format of the member of + * ::OrtROCMProviderOptions and value should be its related range. + * + * For example, key="device_id" and value="0" + * + * \param[in] rocm_options + * \param[in] provider_options_keys Array of UTF-8 null-terminated string for provider options keys + * \param[in] provider_options_values Array of UTF-8 null-terminated string for provider options values + * \param[in] num_keys Number of elements in the `provider_option_keys` and `provider_options_values` arrays + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.16. + */ + ORT_API2_STATUS(UpdateROCMProviderOptions, _Inout_ OrtROCMProviderOptions* rocm_options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** + * Get serialized ROCm provider options string. + * + * For example, "device_id=0;arena_extend_strategy=0;......" + * + * \param rocm_options - OrtROCMProviderOptions instance + * \param allocator - a ptr to an instance of OrtAllocator obtained with CreateAllocator() or GetAllocatorWithDefaultOptions() + * the specified allocator will be used to allocate continuous buffers for output strings and lengths. + * \param ptr - is a UTF-8 null terminated string allocated using 'allocator'. The caller is responsible for using the same allocator to free it. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.16. + */ + ORT_API2_STATUS(GetROCMProviderOptionsAsString, _In_ const OrtROCMProviderOptions* rocm_options, _Inout_ OrtAllocator* allocator, _Outptr_ char** ptr); + + /** \brief Release an ::OrtROCMProviderOptions + * + * \note This is an exception in the naming convention of other Release* functions, as the name of the method does not have the V2 suffix, but the type does + * + * \since Version 1.16. + */ + void(ORT_API_CALL* ReleaseROCMProviderOptions)(_Frees_ptr_opt_ OrtROCMProviderOptions* input); + + /** \brief Create an allocator with specific type and register it with the ::OrtEnv + * This API enhance CreateAndRegisterAllocator that it can create an allocator with specific type, not just CPU allocator + * Enables sharing the allocator between multiple sessions that use the same env instance. + * Lifetime of the created allocator will be valid for the duration of the environment. + * Returns an error if an allocator with the same ::OrtMemoryInfo is already registered. + * \param[in] env OrtEnv instance + * \param[in] provider_type ExecutionProvider type + * \param[in] mem_info OrtMemoryInfo instance + * \param[in] arena_cfg Arena configuration + * \param[in] provider_options_keys key of the provider options map + * \param[in] provider_options_values value of the provider options map + * \param[in] num_keys Length of the provider options map + */ + ORT_API2_STATUS(CreateAndRegisterAllocatorV2, _Inout_ OrtEnv* env, _In_ const char* provider_type, _In_ const OrtMemoryInfo* mem_info, _In_ const OrtArenaCfg* arena_cfg, + _In_reads_(num_keys) const char* const* provider_options_keys, _In_reads_(num_keys) const char* const* provider_options_values, _In_ size_t num_keys); + + /** \brief Run the model asynchronously in a thread owned by intra op thread pool + * + * \param[in] session + * \param[in] run_options If nullptr, will use a default ::OrtRunOptions + * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names + * \param[in] input Array of ::OrtValue%s of the input values + * \param[in] input_len Number of elements in the input_names and inputs arrays + * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names + * \param[in] output_names_len Number of elements in the output_names and outputs array + * \param[out] output OrtValue* array of size output_names_len. + * On calling RunAsync, output[i] could either be a null or a pointer to a preallocated OrtValue. + * Later, the output array will be passed to run_async_callback with all null(s) filled with valid + * OrtValue pointer(s) allocated by onnxruntime. + * NOTE: it is customer's duty to finally release the output array and each of its member, + * regardless of whether the member (OrtValue*) is allocated by onnxruntime or preallocated by the customer. + * \param[in] run_async_callback Callback function on model run completion + * \param[in] user_data User data that pass back to run_async_callback + */ + ORT_API2_STATUS(RunAsync, _Inout_ OrtSession* session, _In_opt_ const OrtRunOptions* run_options, + _In_reads_(input_len) const char* const* input_names, + _In_reads_(input_len) const OrtValue* const* input, size_t input_len, + _In_reads_(output_names_len) const char* const* output_names, size_t output_names_len, + _Inout_updates_all_(output_names_len) OrtValue** output, + _In_ RunAsyncCallbackFn run_async_callback, _In_opt_ void* user_data); + + /** + * Update TensorRT EP provider option where its data type is pointer, for example 'user_compute_stream'. + * If the data type of the provider option can be represented by string please use UpdateTensorRTProviderOptions. + * + * Note: It's caller's responsibility to properly manage the lifetime of the instance pointed by this pointer. + * + * \param tensorrt_options - OrtTensorRTProviderOptionsV2 instance + * \param key - Name of the provider option + * \param value - A pointer to the instance that will be assigned to this provider option + * + * \since Version 1.16. + */ + ORT_API2_STATUS(UpdateTensorRTProviderOptionsWithValue, _Inout_ OrtTensorRTProviderOptionsV2* tensorrt_options, _In_ const char* key, _In_ void* value); + + /** + * Get TensorRT EP provider option where its data type is pointer. + * If the data type of the provider option can be represented by string please use GetTensorRTProviderOptionsAsString. + * + * \param tensorrt_options - OrtTensorRTProviderOptionsV2 instance + * \param key - Name of the provider option + * \param ptr - A pointer to the instance that is kept by the provider option + * + * \since Version 1.16. + */ + ORT_API2_STATUS(GetTensorRTProviderOptionsByName, _In_ const OrtTensorRTProviderOptionsV2* tensorrt_options, _In_ const char* key, _Outptr_ void** ptr); + + /** + * Update CUDA EP provider option where its data type is pointer, for example 'user_compute_stream'. + * If the data type of the provider option can be represented by string please use UpdateCUDAProviderOptions. + * + * Note: It's caller's responsibility to properly manage the lifetime of the instance pointed by this pointer. + * + * \param cuda_options - OrtCUDAProviderOptionsV2 instance + * \param key - Name of the provider option + * \param value - A pointer to the instance that will be assigned to this provider option + * + * \since Version 1.16. + */ + ORT_API2_STATUS(UpdateCUDAProviderOptionsWithValue, _Inout_ OrtCUDAProviderOptionsV2* cuda_options, _In_ const char* key, _In_ void* value); + + /** + * Get CUDA EP provider option where its data type is pointer. + * If the data type of the provider option can be represented by string please use GetCUDAProviderOptionsAsString. + * + * \param cuda_options - OrtCUDAProviderOptionsV2 instance + * \param key - Name of the provider option + * \param ptr - A pointer to the instance that is kept by the provider option + * + * \since Version 1.16. + */ + ORT_API2_STATUS(GetCUDAProviderOptionsByName, _In_ const OrtCUDAProviderOptionsV2* cuda_options, _In_ const char* key, _Outptr_ void** ptr); + + /** + * Get a EP resource. + * E.g. a cuda stream or a cublas handle + * + * \param context - Kernel context + * \param resource_version - Version of the resource + * \param resource_id - Type of resource + * \param resource - A pointer to returned resource + * + * \since Version 1.16. + */ + ORT_API2_STATUS(KernelContext_GetResource, _In_ const OrtKernelContext* context, _In_ int resource_version, + _In_ int resource_id, _Outptr_ void** resource); + + /** \brief Set user logging function + * + * By default the logger created by the CreateEnv* functions is used to create the session logger as well. + * This function allows a user to override this default session logger with a logger of their own choosing. This way + * the user doesn't have to create a separate environment with a custom logger. This addresses the problem when + * the user already created an env but now wants to use a different logger for a specific session (for debugging or + * other reasons). + * + * \param[in] options + * \param[in] user_logging_function A pointer to a logging function. + * \param[in] user_logging_param A pointer to arbitrary data passed as the ::OrtLoggingFunction `param` parameter to + * `user_logging_function`. This parameter is optional. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.17. + */ + ORT_API2_STATUS(SetUserLoggingFunction, _Inout_ OrtSessionOptions* options, + _In_ OrtLoggingFunction user_logging_function, _In_opt_ void* user_logging_param); + + /** + * Get number of input from OrtShapeInferContext + * + * \param[in] context + * \param[out] out The number of inputs + * + * \since Version 1.17. + */ + ORT_API2_STATUS(ShapeInferContext_GetInputCount, _In_ const OrtShapeInferContext* context, _Out_ size_t* out); + + /** + * Get type and shape info of an input + * + * \param[in] context + * \param[in] index The index of the input + * \param[out] info Type shape info of the input + * + * \since Version 1.17. + */ + ORT_API2_STATUS(ShapeInferContext_GetInputTypeShape, _In_ const OrtShapeInferContext* context, _In_ size_t index, _Outptr_ OrtTensorTypeAndShapeInfo** info); + + /** + * Get attribute from OrtShapeInferContext. Note that OrtShapeInferContext is a per-node context, one could only read attribute from current node. + * + * \param[in] context + * \param[in] attr_name Name of the attribute + * \param[out] attr Handle of the attribute fetched + * + * \since Version 1.17. + */ + ORT_API2_STATUS(ShapeInferContext_GetAttribute, _In_ const OrtShapeInferContext* context, _In_ const char* attr_name, _Outptr_ const OrtOpAttr** attr); + + /** + * Set type and shape info of an output + * + * \param[in] context + * \param[in] index The index of the output + * \param[out] info Type shape info of the output + * + * \since Version 1.17. + */ + ORT_API2_STATUS(ShapeInferContext_SetOutputTypeShape, _In_ const OrtShapeInferContext* context, _In_ size_t index, _In_ const OrtTensorTypeAndShapeInfo* info); + + /** + * Set symbolic shape to type shape info + * + * \param[in] info Type shape info + * \param[in] dim_params Symbolic strings + * \param[in] dim_params_length Number of strings + * + * \since Version 1.17. + */ + ORT_API2_STATUS(SetSymbolicDimensions, _In_ OrtTensorTypeAndShapeInfo* info, _In_ const char* dim_params[], _In_ size_t dim_params_length); + + /** + * Read contents of an attribute to data + * + * \param[in] op_attr + * \param[in] type Attribute type + * \param[out] data Memory address to save raw content of the attribute + * \param[in] len Number of bytes allowed to store in data + * \param[out] out Number of bytes required to save the data when the call failed, or the real number of bytes saved to data on success + * + * \since Version 1.17. + */ + ORT_API2_STATUS(ReadOpAttr, _In_ const OrtOpAttr* op_attr, _In_ OrtOpAttrType type, _Inout_ void* data, _In_ size_t len, _Out_ size_t* out); + + /** \brief Set whether to use deterministic compute. + * + * Default is false. If set to true, this will enable deterministic compute for GPU kernels where possible. + * Note that this most likely will have a performance cost. + * + * \param[in] options + * \param[in] value + * + * \since Version 1.17. + */ + ORT_API2_STATUS(SetDeterministicCompute, _Inout_ OrtSessionOptions* options, bool value); + + /** + * Run fn in parallel + * + * \param[in] context + * \param[in] fn Function accepting usr_data and an integer as iterator + * \param[in] total The number of times fn is to be invoked + * \param[in] num_batch Number of batches by which the "total" is to be divided in maximum. When zero, there is no limit + * \param[in] usr_data User data to be passed back to fn + * + * \since Version 1.17. + */ + ORT_API2_STATUS(KernelContext_ParallelFor, _In_ const OrtKernelContext* context, _In_ void (*fn)(void*, size_t), _In_ size_t total, _In_ size_t num_batch, _In_ void* usr_data); + + /** \brief Append OpenVINO execution provider to the session options + * + * If OpenVINO is not available (due to a non OpenVINO enabled build, or if OpenVINO is not installed on the system), this function will fail. + * + * \param[in] options + * \param[in] provider_options_keys + * \param[in] provider_options_values + * \param[in] num_keys + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.17. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_OpenVINO_V2, + _In_ OrtSessionOptions* options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** \brief Append VitisAI provider to session options + * + * If VitisAI is not available (due to a non VitisAI enabled build, or if VitisAI is not installed on the system), this function will return failure. + * + * \param[in] options + * \param[in] provider_options_keys + * \param[in] provider_options_values + * \param[in] num_keys + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.18. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_VitisAI, + _In_ OrtSessionOptions* options, + _In_reads_(num_keys) const char* const* provider_options_keys, + _In_reads_(num_keys) const char* const* provider_options_values, + _In_ size_t num_keys); + + /** \brief Get scratch buffer from the corresponding allocator under the sepcific OrtMemoryInfo object. + * NOTE: callers are responsible to release this scratch buffer from the corresponding allocator + * \param[in] context OrtKernelContext instance + * \param[in] mem_info OrtMemoryInfo instance + * \param[in] count_or_bytes How many bytes is this scratch buffer + * \param[out] out A pointer to the scrach buffer + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.18. + */ + ORT_API2_STATUS(KernelContext_GetScratchBuffer, _In_ const OrtKernelContext* context, _In_ const OrtMemoryInfo* mem_info, _In_ size_t count_or_bytes, _Outptr_ void** out); + + /** \brief Get allocator from KernelInfo for a specific memory type. Please use C API ReleaseAllocator to release out object + * + * \param[in] info OrtKernelInfo instance + * \param[in] mem_type OrtMemType object + * \param[out] out A pointer to OrtAllocator + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.18. + */ + ORT_API2_STATUS(KernelInfoGetAllocator, _In_ const OrtKernelInfo* info, _In_ OrtMemType mem_type, _Outptr_ OrtAllocator** out); + + /** \brief Replace initialized Tensors with external data with the provided files in memory + * + * The function will find the initialized TensorProtos with external data in the graph with the provided + * external file names and the file content in memory. The API gets the external file name, offset, data length + * from TensorProto, and locate the tensor data from the file in memory buffer. + * It creates a Tensor to replace the existing Tensor in graph. The replacement + * will occur before any of the optimizations take place. The data will be copied into the graph + * since TensorProto can't refer to the user provided buffers. + * + * \param[in] options + * \param[in] external_initializer_file_names Array of null terminated UTF-8 encoded strings of the file names + * which holds the external initializers. + * \param[in] external_initializer_file_buffer_array Array of pointers to the buffer of the file content. + * The buffer can be freed after session creation. + * \param[in] external_initializer_file_lengths Array of size_t to indicate the length of file content + * \param[in] num_external_initializer_files Number of external files + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.18. + */ + ORT_API2_STATUS(AddExternalInitializersFromFilesInMemory, _In_ OrtSessionOptions* options, + _In_reads_(num_external_initializer_files) const ORTCHAR_T* const* external_initializer_file_names, + _In_reads_(num_external_initializer_files) char* const* external_initializer_file_buffer_array, + _In_reads_(num_external_initializer_files) const size_t* external_initializer_file_lengths, + size_t num_external_initializer_files); + + /** \brief Create an OrtLoraAdapter + * + * The function attempts to locate file specified by adapter_file_path, read it and create an OrtLoraAdapter + * instance. The adapter_file_path should be a valid path to a file that contains a valid Lora Adapter + * format. The function attempts to validate the format at load time. The file will always be memory mapped, unless + * the platform does not support memory mapping, in which case the file will be read into memory. + * + * \param[in] adapter_file_path adapter file path. + * \param[in] allocator optional pointer to a device allocator. If specified + * data is copied to the device at some point before Run() is invoked. If nullptr, data stays on CPU. + * The data would still be copied to device if required by the model at inference time. + * \param[out] out A pointer to a newly created OrtLoraAdapter instance. Must be released with + * OrtApi::ReleaseLoraAdapter. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.20. + */ + ORT_API2_STATUS(CreateLoraAdapter, const ORTCHAR_T* adapter_file_path, _In_ OrtAllocator* allocator, + _Outptr_ OrtLoraAdapter** out); + + /** \brief Create an OrtLoraAdapter + * + * The function copies the bytes from the array and creates an OrtLoraAdapter instance. + * + * + * \param[in] bytes pointer to a valid Lora Adapter format buffer. + * \param[in] num_bytes length of bytes buffer. + * \param[in] allocator optional pointer to a device allocator. If specified + * data is copied to the device at some point before Run() is invoked. If nullptr, data stays on CPU. + * The data would still be copied to device if required by the model at inference time. + * \param[out] out A pointer to a newly created OrtLoraAdapter instance. Must be released with + * OrtApi::ReleaseLoraAdapter. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.20. + */ + ORT_API2_STATUS(CreateLoraAdapterFromArray, _In_ const void* bytes, size_t num_bytes, _In_ OrtAllocator* allocator, + _Outptr_ OrtLoraAdapter** out); + + /** \brief Release an ::OrtLoraAdapter obtained from OrtApi::CreateLoraAdapter + */ + ORT_CLASS_RELEASE(LoraAdapter); + + /** \brief Add the Lora Adapter to the list of active adapters. + * + * The function adds the Lora Adapter to the list of active adapters. The Lora Adapter must be created with + * OrtApi::CreateLoraAdapter or FromArray. The Lora Adapter will be used by the session to run the model. + * The instance of the OrtRunOptions can then be used to customize the Run() calls. + * More than one OrtLoraAdapter can be active at the same time. Lora Parameters that belong to different + * Lora adapters that will be active at the same time must not overlap. + * This setting does not affect RunWithBinding. + * + * \param[in] options OrtRunOptions instance + * \param[in] adapter OrtLoraAdapter instance + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.20. + */ + ORT_API2_STATUS(RunOptionsAddActiveLoraAdapter, _Inout_ OrtRunOptions* options, _In_ const OrtLoraAdapter* adapter); + + /// @} + /// \name OrtEpDynamicOptions + /// @{ + + /** \brief Set DynamicOptions for EPs (Execution Providers) + * + * Valid options can be found in `include\onnxruntime\core\session\onnxruntime_session_options_config_keys.h` + * Look for `kOrtEpDynamicOptions` + * + * \param[in] sess OrtSession + * \param[in] keys Array of null terminated UTF8 encoded strings of EP dynamic option keys + * \param[in] values Array of null terminated UTF8 encoded string of EP dynamic option values + * \param[in] kv_len Number of elements in the keys and values arrays + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.20. + */ + ORT_API2_STATUS(SetEpDynamicOptions, _Inout_ OrtSession* sess, _In_reads_(kv_len) const char* const* keys, + _In_reads_(kv_len) const char* const* values, _In_ size_t kv_len); + + /** \brief Release an OrtValueInfo instance if it was not added to an OrtGraph. + * \since Version 1.22. + */ + ORT_CLASS_RELEASE(ValueInfo); + + /** \brief Release an OrtNode if it was not added to an OrtGraph. + * \since Version 1.22. + */ + ORT_CLASS_RELEASE(Node); + + /** \brief Release an OrtGraph. + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.22. + */ + ORT_CLASS_RELEASE(Graph); + + /** \brief Release an OrtModel. + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.22. + */ + ORT_CLASS_RELEASE(Model); + + /** \brief Get the value name from an OrtValueInfo instance. + * \param[in] value_info The OrtValueInfo instance. + * \param[out] name The name of the OrtValueInfo + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.22. + */ + ORT_API2_STATUS(GetValueInfoName, _In_ const OrtValueInfo* value_info, _Out_ const char** name); + + /** \brief Get the type information from an OrtValueInfo instance. + * \param[in] value_info The OrtValueInfo instance. + * \param[out] type_info The type info of the OrtValueInfo + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.22. + */ + ORT_API2_STATUS(GetValueInfoTypeInfo, _In_ const OrtValueInfo* value_info, _Outptr_ const OrtTypeInfo** type_info); + + /** \brief Get the Model Editor API instance + * + * Get the Model Editor API instance to create a new model or augment an existing model. + * + * \return Model Editor API struct + * + * \since Version 1.22. + */ + const OrtModelEditorApi*(ORT_API_CALL* GetModelEditorApi)(); + + /** \brief Create an OrtValue for a Tensor that uses pre-existing memory. + * + * ORT will take ownership of the memory and free it using the provided deleter when no longer in use. + * + * \param[in] deleter OrtAllocator instance that will be used to free the memory. + * Only the OrtAllocator:Info and OrtAllocator::Release functions are required. + * The OrtMemoryInfo returned by OrtAllocator::Info must match the location of p_data. + * \param[in] p_data Pointer to the memory that will be used by the Tensor. ORT will take ownership of the memory. + * \param[in] p_data_len Length of the memory in bytes. + * \param[in] shape Dimensions of the Tensor. All values should be > 0. + * \param[in] shape_len Number of dimensions in the shape array. + * \param[in] type Data type of the Tensor. + * \param[out] out Newly created ::OrtValue. Must be freed with OrtApi::ReleaseValue + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateTensorWithDataAndDeleterAsOrtValue, _In_ OrtAllocator* deleter, + _In_ void* p_data, size_t p_data_len, + _In_ const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type, + _Outptr_ OrtValue** out); + + /** \brief sets load cancellation flag to abort session loading process. + * + * \param[in] options instance that was passed to the session at creation time. + * \param[in] cancel setting this to true after model loading process was initiated will + * attempt to cancel the loading process. If cancellation is successful, CreateSession() + * CreateSessionFromArray() or any other session creation API that take session options as an + * argument will return an OrtStatus indicating that session loading was canceled at user request, + * error code ORT_MODEL_LOAD_CANCELED. + * The APIs above would not return any valid Session instance. This is the best case effort and the result + * is not guaranteed. The session may have already been created and initialized + * before the cancellation request was issued. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(SessionOptionsSetLoadCancellationFlag, _Inout_ OrtSessionOptions* options, + _In_ bool cancel); + + /** \brief Get the Compile API instance. + * + * Get the Compile API instance to compile ONNX models. Execution providers that support compilation fuse a subgraph + * into an EPContext node that wraps a provider-specific binary representation of the subgraph. + * For more details about the EPContext design, refer to: + * \htmlonly + * EPContext design document. + * \endhtmlonly + * + * \return Compile API struct instance. + * + * \since Version 1.22. + */ + const OrtCompileApi*(ORT_API_CALL* GetCompileApi)(); + + // + // OrtKeyValuePairs + // + + /** \brief Create an OrtKeyValuePairs instance. + * + * \param[out] out A pointer to a newly created OrtKeyValuePairs instance. + * + * \note Must be released by calling ReleaseKeyValuePairs. + * + * \since Version 1.22. + */ + void(ORT_API_CALL* CreateKeyValuePairs)(_Outptr_ OrtKeyValuePairs** out); + + /** \brief Add a key-value pair to the OrtKeyValuePairs instance. + * + * \param[in] kvps OrtKeyValuePairs instance. + * \param[in] key Key to be added. + * \param[in] value Value to be added. + * + * \note The `key` and `value` are copied internally. + * + * \since Version 1.22. + */ + + void(ORT_API_CALL* AddKeyValuePair)(_In_ OrtKeyValuePairs* kvps, _In_ const char* key, _In_ const char* value); + + /** \brief Get the value associated with a key in the OrtKeyValuePairs instance. + * + * \param[in] kvps OrtKeyValuePairs instance. + * \param[in] key Key to be searched. + * + * \return The value associated with the key, or nullptr if the key does not exist. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* GetKeyValue)(_In_ const OrtKeyValuePairs* kvps, _In_ const char* key); + + /** \brief Get all the key-value pairs from the OrtKeyValuePairs instance. + * + * \param[in] kvps OrtKeyValuePairs instance. + * \param[out] keys Array of keys from `kvps`. + * \param[out] values Array of values from `kvps`. + * \param[out] num_entries Number of entries in `keys` and `values`. + * + * \since Version 1.22. + */ + void(ORT_API_CALL* GetKeyValuePairs)(_In_ const OrtKeyValuePairs* kvps, + _Outptr_ const char* const** keys, _Outptr_ const char* const** values, + _Out_ size_t* num_entries); + + /** \brief Remove a key-value pair from the OrtKeyValuePairs instance. + * + * \param[in] kvps OrtKeyValuePairs instance. + * \param[in] key Key to be removed. No error if not found. + * + * \since Version 1.22. + */ + void(ORT_API_CALL* RemoveKeyValuePair)(_In_ OrtKeyValuePairs* kvps, _In_ const char* key); + + /** \brief Release an OrtKeyValuePairs instance. + * + * \param[in] input OrtKeyValuePairs instance to be released. + * + * \since Version 1.22. + */ + ORT_CLASS_RELEASE(KeyValuePairs); + + /** \brief Register an execution provider library with ORT. + * + * The library must export 'CreateEpFactories' and 'ReleaseEpFactory' functions. + * See OrtEpApi for more details. + * + * \param[in] env The OrtEnv instance to register the library in. + * \param[in] registration_name The name to register the execution provider library under. + * \param[in] path The path to the execution provider library. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(RegisterExecutionProviderLibrary, _In_ OrtEnv* env, _In_ const char* registration_name, + _In_ const ORTCHAR_T* path); + + /** \brief Unregister an execution provider library with ORT. + * + * ORT will call ReleaseEpFactory for all factories created by the library, and unload the library. + * + * You MUST ensure there are no Session instances using execution providers created by the library + * before calling this function. + * + * \param[in] env The OrtEnv instance to unregister the library from. + * \param[in] registration_name The name the execution provider library was registered under. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(UnregisterExecutionProviderLibrary, _In_ OrtEnv* env, _In_ const char* registration_name); + + /** \brief Get the list of available OrtEpDevice instances. + * + * Each OrtEpDevice instance contains details of the execution provider and the device it will use. + * + * \param[in] env The OrtEnv instance to query. + * \param[out] ep_devices The OrtEpDevice instances that the execution provider will use. + * \param[out] num_ep_devices The number of OrtEpDevice instances returned. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(GetEpDevices, _In_ const OrtEnv* env, + _Outptr_ const OrtEpDevice* const** ep_devices, _Out_ size_t* num_ep_devices); + + /** \brief Append the execution provider that is responsible for the selected OrtEpDevice instances + * to the session options. + * + * \param[in] session_options Session options to add execution provider to. + * \param[in] env Environment that execution providers were registered with. + * \param[in] ep_devices One or more OrtEpDevice instances to create an execution provider for. + * Obtain from GetEpDevices. All OrtEpDevice instances must be from the same execution + * provider. It is only necessary to provide multiple OrtEpDevices if you want to use the + * same execution provider for multiple devices. + * e.g. the EP is capable of running on GPU and NPU. + * \param[in] num_ep_devices Number of OrtEpDevice instances. + * \param[in] ep_option_keys Optional keys to configure the execution provider. + * \param[in] ep_option_vals Optional values to configure the execution provider. + * \param[in] num_ep_options Number of execution provide options to add. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(SessionOptionsAppendExecutionProvider_V2, _In_ OrtSessionOptions* session_options, + _In_ OrtEnv* env, + _In_reads_(num_ep_devices) const OrtEpDevice* const* ep_devices, _In_ size_t num_ep_devices, + _In_reads_(num_op_options) const char* const* ep_option_keys, + _In_reads_(num_op_options) const char* const* ep_option_vals, + size_t num_ep_options); + + /** \brief Set the execution provider selection policy for the session. + * + * Allows users to specify a device selection policy for automatic execution provider (EP) selection. + * If custom selection is required please use SessionOptionsSetEpSelectionPolicyDelegate instead. + * + * \param[in] session_options The OrtSessionOptions instance. + * \param[in] policy The device selection policy to use (see OrtExecutionProviderDevicePolicy). + * + * \since Version 1.22 + */ + ORT_API2_STATUS(SessionOptionsSetEpSelectionPolicy, _In_ OrtSessionOptions* session_options, + _In_ OrtExecutionProviderDevicePolicy policy); + + /** \brief Set the execution provider selection policy delegate for the session. + * + * Allows users to provide a custom device selection policy for automatic execution provider (EP) selection. + * + * \param[in] session_options The OrtSessionOptions instance. + * \param[in] delegate Delegate callback for custom selection. + * \param[in] delegate_state Optional state that will be passed to the delegate callback. nullptr if not required. + * + * \since Version 1.22 + */ + ORT_API2_STATUS(SessionOptionsSetEpSelectionPolicyDelegate, _In_ OrtSessionOptions* session_options, + _In_ EpSelectionDelegate delegate, + _In_opt_ void* delegate_state); + + /** \brief Get the hardware device type. + * + * \param[in] device The OrtHardwareDevice instance to query. + * \return The hardware device type. + * + * \since Version 1.22. + */ + OrtHardwareDeviceType(ORT_API_CALL* HardwareDevice_Type)(_In_ const OrtHardwareDevice* device); + + /** \brief Get the hardware device's vendor identifier. + * + * \param[in] device The OrtHardwareDevice instance to query. + * \return The hardware device vendor identifier. + * + * \since Version 1.22. + */ + uint32_t(ORT_API_CALL* HardwareDevice_VendorId)(_In_ const OrtHardwareDevice* device); + + /** \brief Get the hardware device's vendor name. + * + * \param[in] device The OrtHardwareDevice instance to query. + * \return The hardware device's vendor name. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* HardwareDevice_Vendor)(_In_ const OrtHardwareDevice* device); + + /** \brief Get the hardware device's unique identifier. + * + * \param[in] device The OrtHardwareDevice instance to query. + * \return The device id. + * + * \note This is not a unique identifier. It identifies the hardware type when combined with vendor id. + * \since Version 1.22. + */ + uint32_t(ORT_API_CALL* HardwareDevice_DeviceId)(_In_ const OrtHardwareDevice* device); + + /** \brief Get hardware device metadata. + * + * \param[in] device The OrtHardwareDevice instance to query. + * \return An OrtKeyValuePairs instance containing the metadata for the device. + * Note: ORT owns the instance so the user must not call ReleaseKeyValuePairs with it. + * + * \since Version 1.22. + */ + const OrtKeyValuePairs*(ORT_API_CALL* HardwareDevice_Metadata)(_In_ const OrtHardwareDevice* device); + + /** \brief Get the execution provider name. + * + * \param[in] ep_device The OrtEpDevice instance to query. + * \return The execution provider name. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* EpDevice_EpName)(_In_ const OrtEpDevice* ep_device); + + /** \brief Get the execution provider's vendor name. + * + * \param[in] ep_device The OrtEpDevice instance to query. + * \return The execution provider's vendor name. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* EpDevice_EpVendor)(_In_ const OrtEpDevice* ep_device); + + /** \brief Get the metadata for the OrtEpDevice. + * + * \param[in] ep_device The OrtEpDevice instance to query. + * \return An OrtKeyValuePairs instance containing the metadata for the device. + * + * \since Version 1.22. + */ + const OrtKeyValuePairs*(ORT_API_CALL* EpDevice_EpMetadata)(_In_ const OrtEpDevice* ep_device); + + /** \brief Get the execution provider options for the OrtEpDevice. + * + * \param[in] ep_device The OrtEpDevice instance to query. + * \return An OrtKeyValuePairs instance containing the execution provider options for the device. + * + * \since Version 1.22. + */ + const OrtKeyValuePairs*(ORT_API_CALL* EpDevice_EpOptions)(_In_ const OrtEpDevice* ep_device); + + /** \brief Get the OrtHardwareDevice instance for the OrtEpDevice. + * + * \param[in] ep_device The OrtEpDevice instance to query. + * \return The OrtHardwareDevice instance for the device. + * + * \since Version 1.22. + */ + const OrtHardwareDevice*(ORT_API_CALL* EpDevice_Device)(_In_ const OrtEpDevice* ep_device); + + /** \brief Get the OrtEpApi instance for implementing an execution provider. + * + * \since Version 1.22. + */ + const OrtEpApi*(ORT_API_CALL* GetEpApi)(); +}; + +/* + * Steps to use a custom op: + * 1 Create an OrtCustomOpDomain with the domain name used by the custom ops + * 2 Create an OrtCustomOp structure for each op and add them to the domain + * 3 Call OrtAddCustomOpDomain to add the custom domain of ops to the session options + */ + +// Specifies some characteristics of inputs/outputs of custom ops: +// Specify if the inputs/outputs are one of: +// 1) Non-optional (input/output must be present in the node) +// 2) Optional (input/output may be absent in the node) +// 3) Variadic: A variadic input or output specifies N (i.e., the minimum arity) or more operands. +// Only the last input or output of a custom op may be marked as variadic. +// The homogeneity of the variadic input or output determines whether all operands must be of the same +// tensor element type. +typedef enum OrtCustomOpInputOutputCharacteristic { + INPUT_OUTPUT_REQUIRED = 0, + INPUT_OUTPUT_OPTIONAL, + INPUT_OUTPUT_VARIADIC, +} OrtCustomOpInputOutputCharacteristic; + +/* + * The OrtCustomOp structure defines a custom op's schema and its kernel callbacks. The callbacks are filled in by + * the implementor of the custom op. + */ +struct OrtCustomOp { + uint32_t version; // Must be initialized to ORT_API_VERSION + + // This callback creates the kernel, which is a user defined + // parameter that is passed to the Kernel* callbacks below. It is + // recommended to use CreateKernelV2 which allows for a safe error + // propagation by returning an OrtStatusPtr. + void*(ORT_API_CALL* CreateKernel)(_In_ const struct OrtCustomOp* op, _In_ const OrtApi* api, + _In_ const OrtKernelInfo* info); + + // Returns the name of the op + const char*(ORT_API_CALL* GetName)(_In_ const struct OrtCustomOp* op); + + // Returns the type of the execution provider, return nullptr to use CPU execution provider + const char*(ORT_API_CALL* GetExecutionProviderType)(_In_ const struct OrtCustomOp* op); + + // Returns the count and types of the input & output tensors + ONNXTensorElementDataType(ORT_API_CALL* GetInputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index); + size_t(ORT_API_CALL* GetInputTypeCount)(_In_ const struct OrtCustomOp* op); + ONNXTensorElementDataType(ORT_API_CALL* GetOutputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index); + size_t(ORT_API_CALL* GetOutputTypeCount)(_In_ const struct OrtCustomOp* op); + + // Perform a computation step. It is recommended to use + // KernelComputeV2 which allows for a safe error propagation by + // returning an OrtStatusPtr. + void(ORT_API_CALL* KernelCompute)(_In_ void* op_kernel, _In_ OrtKernelContext* context); + void(ORT_API_CALL* KernelDestroy)(_In_ void* op_kernel); + + // Returns the characteristics of the input & output tensors + OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetInputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index); + OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetOutputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index); + + // Returns the memory type of the input tensors. This API allows the custom op + // to place the inputs on specific devices. By default, it returns + // OrtMemTypeDefault, which means the input is placed on the default device for + // the execution provider. If the inputs need to be with different memory tyeps, + // this function can be overridden to return the specific memory types. + OrtMemType(ORT_API_CALL* GetInputMemoryType)(_In_ const struct OrtCustomOp* op, _In_ size_t index); + + // Returns the minimum number of input arguments expected for the variadic input. + // Applicable only for custom ops that have a variadic input. + int(ORT_API_CALL* GetVariadicInputMinArity)(_In_ const struct OrtCustomOp* op); + + // Returns true (non-zero) if all arguments of a variadic input have to be of the same type (homogeneous), + // and false (zero) otherwise. + // Applicable only for custom ops that have a variadic input. + int(ORT_API_CALL* GetVariadicInputHomogeneity)(_In_ const struct OrtCustomOp* op); + + // Returns the minimum number of output values expected for the variadic output. + // Applicable only for custom ops that have a variadic output. + int(ORT_API_CALL* GetVariadicOutputMinArity)(_In_ const struct OrtCustomOp* op); + + // Returns true (non-zero) if all outputs values of a variadic output have to be of the same type (homogeneous), + // and false (zero) otherwise. + // Applicable only for custom ops that have a variadic output. + int(ORT_API_CALL* GetVariadicOutputHomogeneity)(_In_ const struct OrtCustomOp* op); + + // Create the kernel state which is passed to each compute call. + OrtStatusPtr(ORT_API_CALL* CreateKernelV2)(_In_ const struct OrtCustomOp* op, _In_ const OrtApi* api, + _In_ const OrtKernelInfo* info, + _Out_ void** kernel); + + // Perform the computation step. + OrtStatusPtr(ORT_API_CALL* KernelComputeV2)(_In_ void* op_kernel, _In_ OrtKernelContext* context); + + OrtStatusPtr(ORT_API_CALL* InferOutputShapeFn)(_In_ const struct OrtCustomOp* op, _In_ OrtShapeInferContext*); + + // Get start range + int(ORT_API_CALL* GetStartVersion)(_In_ const struct OrtCustomOp* op); + int(ORT_API_CALL* GetEndVersion)(_In_ const struct OrtCustomOp* op); + + // Get the inplace_map that defines which output can reuse which input + // Callers will provide 2 raw int* and pass in their address, this function will fill these 2 arrays + // when return, output (*output_index)[i] may reuse the input (*input_index[i]). + // The return value is the size of these 2 arrays. + // Callers are responsible to delete these 2 arrays after use by calling OrtCustomOp::ReleaseMayInplace(). + size_t(ORT_API_CALL* GetMayInplace)(_Out_ int** input_index, _Out_ int** output_index); + + // Release the pointer input_index and output_index allocated from GetMayInplace() function. + // If GetMayInplace() is defined, this function MUST be defined as well. + void(ORT_API_CALL* ReleaseMayInplace)(_Frees_ptr_opt_ int* input_index, _Frees_ptr_opt_ int* output_index); + + // Same as GetMayInplace() and ReleaseMayInplace() + size_t(ORT_API_CALL* GetAliasMap)(_Out_ int** input_index, _Out_ int** output_index); + void(ORT_API_CALL* ReleaseAliasMap)(_Frees_ptr_opt_ int* input_index, _Frees_ptr_opt_ int* output_index); +}; + +/** + * ORT Model Editor API + */ + +/** + * \brief The OrtModelEditorApi struct provides functions to create or edit an ONNX model. + * + * See onnxruntime/test/shared_lib/test_model_editor_api.cc for example usage. + * + * \since Version 1.22. + */ +struct OrtModelEditorApi { + // Model building/editing requires a full build. We return nullptr from GetModelEditorApi if this is a minimal + // build, so it doesn't matter if there are no function pointers in this struct as a user will never get an + // OrtModelEditorApi instance. We do however need a dummy field to avoid empty struct warning. +#if defined(ORT_MINIMAL_BUILD) + const bool not_defined_in_this_build; +#else + /** \brief Create an OrtTypeInfo instance for a Tensor. + * + * Create an OrtTypeInfo instance for a Tensor to use as graph inputs/outputs with the Model Editor API. + * + * User can release `tensor_info` after creating the OrtTypeInfo. + * + * \param[in] tensor_info Tensor type and shape information. + * \param[out] type_info TypeInfo instance for the tensor. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateTensorTypeInfo, _In_ const OrtTensorTypeAndShapeInfo* tensor_info, + _Outptr_ OrtTypeInfo** type_info); + + /** \brief Create an OrtTypeInfo instance for a SparseTensor. + * + * Create an OrtTypeInfo instance for a SparseTensor to use as graph inputs/outputs with the Model Editor API. + * + * User can release `tensor_info` after creating the OrtTypeInfo. + * + * \param[in] tensor_info SparseTensor type and shape information. + * \param[out] type_info TypeInfo instance for the tensor. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateSparseTensorTypeInfo, _In_ const OrtTensorTypeAndShapeInfo* tensor_info, + _Outptr_ OrtTypeInfo** type_info); + + /** \brief Create an OrtTypeInfo instance for a Map. + * + * Create an OrtTypeInfo instance for a Map to use as graph inputs/outputs with the Model Editor API. + * + * User can release `map_value_type` after creating the OrtTypeInfo. + * + * \param[in] map_key_type Key type for the map. + * \param[in] map_value_type Value type for the map. + * \param[out] type_info TypeInfo instance for the map. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateMapTypeInfo, ONNXTensorElementDataType map_key_type, _In_ const OrtTypeInfo* map_value_type, + _Outptr_ OrtTypeInfo** type_info); + + /** \brief Create an OrtTypeInfo instance for a Sequence. + * + * Create an OrtTypeInfo instance for a Sequence to use as graph inputs/outputs with the Model Editor API. + * + * User can release `sequence_type` after creating the OrtTypeInfo. + * + * \param[in] sequence_type Sequence type and shape information. + * \param[out] type_info TypeInfo instance for the sequence. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateSequenceTypeInfo, _In_ const OrtTypeInfo* sequence_type, _Outptr_ OrtTypeInfo** type_info); + + /** \brief Create an OrtTypeInfo instance for an Optional. + * + * Create an OrtTypeInfo instance for an Optional to use as graph inputs/outputs with the Model Editor API. + * + * User can release `contained_type` after creating the OrtTypeInfo. + * + * \param[in] contained_type Tensor type and shape information. + * \param[out] type_info TypeInfo instance for the tensor. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateOptionalTypeInfo, _In_ const OrtTypeInfo* contained_type, _Outptr_ OrtTypeInfo** type_info); + + /** \brief Create an OrtValueInfo for use as an OrtGraph input or output. + * + * \param[in] name The name of the input or output. + * \param[in] type_info The type information for the input or output. The provided value is copied. + * \param[out] value_info The OrtValueInfo instance. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateValueInfo, _In_ const char* name, _In_ const OrtTypeInfo* type_info, + _Outptr_ OrtValueInfo** value_info); + + /** \brief Create an OrtNode to add to an OrtGraph. + * + * Create an OrtNode. + * + * Create attributes with CreateOpAttr. OrtOpAttr instances are copied. + * + * \param[in] operator_name The name of the operator. + * \param[in] domain_name The domain of the operator. Use an empty string for ONNX operators. + * \param[in] node_name The name of the node. + * \param[in] input_names The names of the inputs. + * \param[in] input_names_len The number of input names. + * \param[in] output_names The names of the outputs. + * \param[in] output_names_len The number of output names. + * \param[in] attributes The optional attributes of the node. + * \param[in] attribs_len The number of attributes. May be zero. + * \param[out] node The OrtNode instance. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateNode, _In_ const char* operator_name, _In_ const char* domain_name, _In_ const char* node_name, + _In_reads_(input_names_len) const char* const* input_names, size_t input_names_len, + _In_reads_(output_names_len) const char* const* output_names, size_t output_names_len, + _In_reads_(attribs_len) _In_opt_ OrtOpAttr** attributes, _In_ size_t attribs_len, + _Outptr_ OrtNode** node); + + /** \brief Create an OrtGraph + * \snippet{doc} snippets.dox OrtStatus Return Value + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateGraph, _Outptr_ OrtGraph** graph); + + /** \brief Set the inputs for the OrtGraph. + * + * Set the graph inputs. This will replace any existing inputs with the new values. + * The OrtGraph takes ownership of the OrtValueInfo instances and you should NOT call ReleaseOrtValueInfo. + * + * \param[in] graph The OrtGraph instance to update. + * \param[in] inputs The input OrtValueInfo instances. + * \param[in] inputs_len The number of input OrtValueInfo instances. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(SetGraphInputs, _Inout_ OrtGraph* graph, + _In_reads_(inputs_len) _In_ OrtValueInfo** inputs, _In_ size_t inputs_len); + + /** \brief Set the outputs for the OrtGraph. + * + * Set the graph outputs. This will replace any existing outputs with the new values. + * The OrtGraph takes ownership of the OrtValueInfo instances provided and you should NOT call ReleaseOrtValueInfo. + * + * \param[in] graph The OrtGraph instance to update. + * \param[in] outputs The output OrtValueInfo instances. + * \param[in] outputs_len The number of output OrtValueInfo instances. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(SetGraphOutputs, _Inout_ OrtGraph* graph, + _In_reads_(outputs_len) _In_ OrtValueInfo** outputs, _In_ size_t outputs_len); + + /** \brief Add an initializer to the OrtGraph + * + * ORT will take ownership of the OrtValue and you should NOT call ReleaseOrtValue. + * + * Two options: + * + * Allocated memory: + * Use CreateTensorAsOrtValue (allocates memory) and populate the tensor with the data. + * Set `data_is_external` to false. + * + * Pre-existing memory: + * Use CreateTensorWithDataAsOrtValue or CreateTensorWithDataAndDeleterAsOrtValue to create an OrtValue + * with a tensor that contains a pointer to the existing data. + * Set `data_is_external` to true. + * + * The pointer must remain valid for the duration of the inference session. + * If using CreateTensorWithDataAsOrtValue you are responsible for freeing the memory after the inference session + * is released. + * If using CreateTensorWithDataAndDeleterAsOrtValue, ORT will free the memory using the provided deleter as + * soon as the OrtValue is no longer in use. + * + * NOTE: A tensor containing pre-existing memory MUST have 128 bytes of data or more. + * For smaller tensors use CreateTensorAsOrtValue. + * + * ONNX shape inferencing does not support external data. An initializer involved in shape inferencing is + * typically small (a single value or limited by the rank of a tensor) and uses less than 128 bytes of + * memory, so this limit acts as a simple catch-all rule to avoid issues. + * e.g. Reshape's `shape`, Clip's `min` and `max`, various ops `axes`. + * + * \param[in] graph The OrtGraph instance to update. + * \param[in] name The value name for the initializer. + * \param[in] tensor The OrtValue instance containing the tensor data. + * \param[in] data_is_external Set to true if the data is external and should not be copied. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(AddInitializerToGraph, _Inout_ OrtGraph* graph, _In_ const char* name, _In_ OrtValue* tensor, + bool data_is_external); + + /** \brief Add an OrtNode to an OrtGraph + * + * Add the node to the graph. The OrtGraph will take ownership of OrtNode and you should NOT call ReleaseOrtNode. + * + * \param[in] graph The OrtGraph instance to update. + * \param[in] node The OrtNode instance to add to the graph. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(AddNodeToGraph, _Inout_ OrtGraph* graph, _In_ OrtNode* node); + + /** \brief Create an OrtModel. + * + * Create an OrtModel. + * + * This can be used to build a new model, or to augment an existing model. + * + * \param[in] domain_names The domain names for the model. + * If augmenting an existing model add additional domains if needed. + * \param[in] opset_versions The opset versions for the model. + * If augmenting an existing model add additional opset versions if needed. + * \param[in] opset_entries_len The number of domain_names and opset_versions entries. + * Domain and opset entries should be 1:1 + * \param[out] model The OrtModel instance. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateModel, + _In_reads_(opset_entries_len) const char* const* domain_names, + _In_reads_(opset_entries_len) const int* opset_versions, + size_t opset_entries_len, + _Outptr_ OrtModel** model); + + /** \brief Add an OrtGraph to an OrtModel. + * + * Add the graph to a model. This should be called once when creating a new model. + * + * The OrtModel takes ownership of the OrtGraph and you should NOT call ReleaseOrtGraph. + * + * \param[in] model The OrtModel instance to update. + * \param[in] graph The OrtGraph instance to add to the model. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(AddGraphToModel, _Inout_ OrtModel* model, _In_ OrtGraph* graph); + + /** \brief Create an OrtSession using the OrtModel. + * + * Create an inference session using the OrtModel instance. + * The OrtModel should have been populated with an OrtGraph containing nodes and initializers, and SetGraphInputs + * and SetGraphOutputs must have been called. + * This will validate the model, run optimizers, and prepare the session for inferencing. + * + * ReleaseOrtModel must be called to free the OrtModel after session creation. + * + * \param[in] env The OrtEnv instance. + * \param[in] model The OrtModel instance. + * \param[in] options The OrtSessionOptions instance. + * \param[out] out The OrtSession instance. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateSessionFromModel, _In_ const OrtEnv* env, _In_ const OrtModel* model, + _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out); + + /** \brief Create an OrtSession to augment an existing model. + * + * Create an OrtSession with an existing model that will be augmented with additional nodes and initializers. + * Nodes can be added before or after the existing nodes in the model. ONNX Runtime will connect the nodes when the + * model is finalized. + * + * To add nodes and initializers to the existing model, first create an OrtModel using CreateModel. + * Add nodes and initializers to the OrtModel using AddNodeToGraph and AddInitializerToGraph. + * Graph inputs/outputs should be updated with SetGraphInputs and SetGraphOutputs as needed to reflect changes made + * by the new nodes. The list of graph inputs/outputs should be for the overall model and not just the new nodes. + * + * Add the new information from the OrtModel to the original model using ApplyModelToSession, and prepare the + * session for inferencing by calling FinalizeModelEditorSession. + * + * \param{in} env The OrtEnv instance. + * \param{in} model_path The path to the existing ONNX model to augment. + * \param{in} options The OrtSessionOptions instance. + * \param{out} out The created OrtSession instance. + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateModelEditorSession, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path, + _In_ const OrtSessionOptions* options, + _Outptr_ OrtSession** out); + + /** \brief Create an OrtSession to augment an existing model. + * + * Create an OrtSession with an existing model that will be augmented with additional nodes and initializers. + * Nodes can be added before or after the existing nodes in the model. ONNX Runtime will connect the nodes when the + * model is finalized. + * + * To add nodes and initializers to the existing model, first create an OrtModel using CreateModel. + * Add nodes and initializers to the OrtModel using AddNodeToGraph and AddInitializerToGraph. + * Graph inputs/outputs should be updated with SetGraphInputs and SetGraphOutputs as needed to reflect changes made + * by the new nodes. The list of graph inputs/outputs should be for the overall model and not just the new nodes. + * + * Add the new information from the OrtModel to the original model using ApplyModelToSession, and prepare the + * session for inferencing by calling FinalizeModelEditorSession. + * + * \param{in} env The OrtEnv instance. + * \param{in} model_data The model data for the existing model to augment. + * \param{in} model_data_length The length of the model data. + * \param{in} options The OrtSessionOptions instance. + * \param{out} out The created OrtSession instance. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateModelEditorSessionFromArray, _In_ const OrtEnv* env, + _In_ const void* model_data, size_t model_data_length, + _In_ const OrtSessionOptions* options, + _Outptr_ OrtSession** out); + + /** \brief Query the session for the opset version of a domain. + * + * When using the Model Editor API to augment a model, any new nodes must conform to the opset version of the + * original model. To do that the user must be able to discover that opset version. + * Returns an error if the domain is not used in the model. + * + * \param[in] session OrtSession to query + * \param[in] domain Domain to query. The ONNX domain is an empty string. + * \param[out] opset The opset version of the domain. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(SessionGetOpsetForDomain, _In_ const OrtSession* session, _In_ const char* domain, _Out_ int* opset); + + /** \brief Apply changes to augment the ONNX model in a session created using CreateModelEditorSession[FromArray] + * + * Adds new nodes and updates graph inputs/outputs using `model` to augment the original ONNX model in the session. + * All changes will be validated. + * Call FinalizeModelEditorSession to prepare the session for inferencing. + * + * Existing input/outputs will only be updated if the OrtGraph inputs/outputs are set in the OrtModel. + * i.e. you don't need to call SetGraphInputs/SetGraphOutputs if they are unchanged. + * + * ReleaseOrtModel must be called to free the OrtModel after it is applied to the session. + * + * \param[in] session OrtSession to update. Session must have been created using CreateModelEditorSession[FromArray]. + * \param[in] model OrtModel containing new nodes, new initializers, and updated graph input and/or output info. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ApplyModelToModelEditorSession, _Inout_ OrtSession* session, _In_ OrtModel* model); + + /** \brief Finalize the Model Editor session that was created using CreateModelEditorSession[FromArray]. + * + * Finalize the Model Editor session that augmented an ONNX model by adding new nodes. + * This will run optimizers and prepare the session for inferencing. + * + * \param[in] session OrtSession to finalize. Session must have been created using CreateModelEditorSession[FromArray]. + * \param[in] options OrtSessionOptions to use for the session. + * \param[in] prepacked_weights_container Optional OrtPrepackedWeightsContainer to use for the session. + Set to nullptr if not used. + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(FinalizeModelEditorSession, _Inout_ OrtSession* session, _In_ const OrtSessionOptions* options, + _In_opt_ OrtPrepackedWeightsContainer* prepacked_weights_container); +#endif // !defined(ORT_MINIMAL_BUILD) +}; + +/** + * ORT Compile API + */ + +/** + * \brief The OrtCompileApi struct provides functions to compile ONNX models. + * + * Execution providers that support compilation fuse a subgraph into an EPContext node that wraps a provider-specific + * binary representation of the subgraph. + * For more details about the EPContext design, refer to: + * \htmlonly + * EPContext design document. + * \endhtmlonly + * + * Example (error handling not shown): + * OrtStatus* status = NULL; + * OrtCompileApi* compile_api = ort_api->GetCompileApi(); + * OrtModelCompilationOptions* compile_options = NULL; + * + * status = compile_api->CreateModelCompilationOptionsFromSessionOptions(env, session_options, &compile_options); + * status = compile_api->ModelCompilationOptions_SetInputModelPath(compile_options, ORT_TSTR("model.onnx")); + * status = compile_api->ModelCompilationOptions_SetOutputModelPath(compile_options, ORT_TSTR("model.compiled.onnx")); + * status = compile_api->CompileModel(env, compile_options); + * compile_api->ReleaseModelCompilationOptions(compile_options); + * + * \since Version 1.22. + */ +struct OrtCompileApi { + /// @} + /// \name OrtModelCompilationOptions + /// @{ + ORT_CLASS_RELEASE(ModelCompilationOptions); + + /** \brief Creates an OrtModelCompilationOptions object from an existing OrtSessionOptions object. + * + * An OrtModelCompilationOptions object contains the settings used to generate a compiled ONNX model. + * The OrtSessionOptions object has the execution providers with which the model will be compiled. + * + * ReleaseOrtModelCompilationsOptions must be called to free the OrtModelCompilationOptions after calling + * CompileModel. + * + * \param[in] env OrtEnv object. + * \param[in] session_options The OrtSessionOptions instance from which to create the OrtModelCompilationOptions. + * \param[out] out The created OrtModelCompilationOptions instance. + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateModelCompilationOptionsFromSessionOptions, _In_ const OrtEnv* env, + _In_ const OrtSessionOptions* session_options, _Outptr_ OrtModelCompilationOptions** out); + + /** \brief Sets the file path to the input ONNX model to compile. + * + * The input model's location (e.g., file path or memory buffer) must be set with either + * ModelCompilationOptions_SetInputModelPath or ModelCompilationOptions_SetInputModelFromBuffer. + * + * \param[in] model_compile_options The OrtModelCompilationOptions instance. + * \param[in] input_model_path Null terminated string of the path (wchar on Windows, char otherwise). + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ModelCompilationOptions_SetInputModelPath, _In_ OrtModelCompilationOptions* model_compile_options, + _In_ const ORTCHAR_T* input_model_path); + + /** \brief Sets the buffer that stores the bytes of the loaded ONNX model to compile. + * + * The input model's location (e.g., file path or memory buffer) must be set with either + * ModelCompilationOptions_SetInputModelPath or ModelCompilationOptions_SetInputModelFromBuffer. + * + * \param[in] model_compile_options The OrtModelCompilationOptions instance. + * \param[in] input_model_data Buffer containing the loaded ONNX model bytes. + * \param[in] input_model_data_size The number of bytes in the `input_model_data` buffer. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ModelCompilationOptions_SetInputModelFromBuffer, + _In_ OrtModelCompilationOptions* model_compile_options, + _In_ const void* input_model_data, + size_t input_model_data_size); + + /** \brief Sets the file path for the output ONNX model generated by CompileModel. + * + * The output model's location (e.g., file path or memory buffer) can be set with either + * ModelCompilationOptions_SetOutputModelPath or ModelCompilationOptions_SetOutputModelBuffer. + * + * If the output model's location is not set, ONNX Runtime will generate an output file with a path based on + * the input model's file path. Examples: + * /Path/my_model.onnx -> /Path/my_model_ctx.onnx + * /Path/my_model -> /Path/my_model_ctx.onnx + * + * \param[in] model_compile_options The OrtModelCompilationOptions instance. + * \param[in] output_model_path Null terminated string of the path (wchar on Windows, char otherwise). + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ModelCompilationOptions_SetOutputModelPath, _In_ OrtModelCompilationOptions* model_compile_options, + _In_ const ORTCHAR_T* output_model_path); + + /** \brief Optionally sets the file that should store external initializers for the compiled ONNX model. + * If not set, initializers are stored within the model. + * + * Only initializers for nodes that were not compiled are stored in the external initializers file. + * Compiled nodes contain their initializer data within the `ep_cache_context` attribute of EPContext nodes. + * Refer to ModelCompilationOptions_SetEpContextEmbedMode. + * + * \param[in] model_compile_options The OrtModelCompilationOptions instance. + * \param[in] external_initializers_file_path Null terminated string of the path to the file. + * \param[in] external_initializers_size_threshold Initializers larger than this threshold are stored in the file. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ModelCompilationOptions_SetOutputModelExternalInitializersFile, + _In_ OrtModelCompilationOptions* model_compile_options, + _In_ const ORTCHAR_T* external_initializers_file_path, + size_t external_initializers_size_threshold); + + /** \brief Configures model compilation to store the output compiled ONNX model in a buffer. + * + * The caller passes an OrtAllocator that ONNX Runtime uses to allocate memory for the buffer. + * + * The output model's location (e.g., file path or memory buffer) can be set with either + * ModelCompilationOptions_SetOutputModelPath or ModelCompilationOptions_SetOutputModelBuffer. + * + * If the output model's location is not set, ONNX Runtime will generate an output file with a path based on + * the input model's file path. Examples: + * /Path/my_model.onnx -> /Path/my_model_ctx.onnx + * /Path/my_model -> /Path/my_model_ctx.onnx + * + * \param[in] model_compile_options The OrtModelCompilationOptions instance. + * \param[in] allocator The allocator used to allocate the buffer for the compiled model. + * \param[out] output_model_buffer_ptr Pointer to the buffer that stores the compiled model. + * \param[out] output_model_buffer_size_ptr Pointer set to the size of output model in bytes. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ModelCompilationOptions_SetOutputModelBuffer, + _In_ OrtModelCompilationOptions* model_compile_options, + _Inout_ OrtAllocator* allocator, + _Outptr_ void** output_model_buffer_ptr, + _Out_ size_t* output_model_buffer_size_ptr); + + /** \brief Enables or disables the embedding of EPContext binary data into the `ep_cache_context` attribute + * of EPContext nodes. Defaults to false. + * + * If enabled, the `ep_cache_context` attribute of EPContext nodes will store the context binary data, which may + * include weights for compiled subgraphs. + * + * If disabled, the `ep_cache_context` attribute of EPContext nodes will contain the path to the file containing the + * context binary data. The path is set by the execution provider creating the EPContext node. + * + * More details relate to EPContext design refers to: + * \htmlonly + * EPContext design document. + * \endhtmlonly + * + * \param[in] model_compile_options The OrtModelCompilationOptions instance. + * \param[in] embed_ep_context_in_model True to embed EPContext binary data into the EPContext node + * `ep_cache_context` attributes. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(ModelCompilationOptions_SetEpContextEmbedMode, _In_ OrtModelCompilationOptions* model_compile_options, + bool embed_ep_context_in_model); + + /** \brief Compiles an input ONNX model with the given compilation options. + * + * \param[in] env OrtEnv object. + * \param[in] model_options The compilation options that defines compilation options for a model. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CompileModel, _In_ const OrtEnv* env, _In_ const OrtModelCompilationOptions* model_options); +}; + +ORT_RUNTIME_CLASS(Ep); +ORT_RUNTIME_CLASS(EpFactory); + +struct OrtEpApi { + /** \brief Create an OrtEpDevice for the EP and an OrtHardwareDevice. + * \param[in] ep_factory Execution provider factory that is creating the instance. + * \param[in] hardware_device Hardware device that the EP can utilize. + * \param[in] ep_metadata Optional OrtKeyValuePairs instance for execution provider metadata that may be used + * during execution provider selection and passed to CreateEp. + * ep_device will copy this instance and the user should call ReleaseKeyValuePairs. + * \param[in] ep_options Optional OrtKeyValuePairs instance for execution provider options that will be added + * to the Session configuration options if the execution provider is selected. + * ep_device will copy this instance and the user should call ReleaseKeyValuePairs. + * \param ep_device OrtExecutionDevice that is created. + * + * \since Version 1.22. + */ + ORT_API2_STATUS(CreateEpDevice, _In_ OrtEpFactory* ep_factory, + _In_ const OrtHardwareDevice* hardware_device, + _In_opt_ const OrtKeyValuePairs* ep_metadata, + _In_opt_ const OrtKeyValuePairs* ep_options, + _Out_ OrtEpDevice** ep_device); + + ORT_CLASS_RELEASE(EpDevice); +}; + +/** + * \brief The OrtEp struct provides functions to implement for an execution provider. + * \since Version 1.22. + */ +struct OrtEp { + /** \brief The ONNX Runtime version the execution provider was compiled with. + * + * Implementation should set to ORT_API_VERSION. + * ORT will use this to ensure it does not call functions that were not available when the library was compiled. + * + * \since Version 1.22. + */ + uint32_t ort_version_supported; + + /** \brief Get the execution provider name. + * + * \param[in] this_ptr The OrtEp instance. + * \return The execution provider name. + * + * \note Returned string is owned by ORT and valid until UnregisterExecutionProviderLibrary is called. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* GetName)(const OrtEp* this_ptr); + + // OrtStatus* GetCapability(OrtEp* ep, const OrtGraph* graph, + // size_t* num_supported_subgraphs, + // OrtIndexedSubgraph** supported_subgraphs, OrtAllocator* allocator); + + // OrtStatus* Compile(OrtEp* ep, const OrtGraph** graphs, OrtNode** fused_graph_nodes, + // size_t count, OrtNodeComputeInfo* node_compute_infos); + + // TODO: Implement OrtEpApi and the complete OrtEp interface as the next step. +}; + +/** \brief The function signature that ORT will call to create OrtEpFactory instances. + * + * This must be available in a function called 'CreateEpFactories' in the execution provider library. + * + * \param[in] registered_name The name the execution library is registered with by RegisterExecutionProviderLibrary + * \param[in] ort_api_base The OrtApiBase instance that is used by the factory to get the OrtApi instance for the + * version of ORT that the library was compiled against. + * \param[in,out] factories The implementation should create and add OrtEpFactory instances to this + * pre-allocated array. + * i.e. usage is `factories[0] = new MyEpFactory();` + * \param[in] max_factories The maximum number of OrtEpFactory instances that can be added to `factories`. + * Current default is to allow 4 factories. This can be increased in the future if needed. + * \param[out] num_factories The number of OrtEpFactory instances created by the factory and added to `factories`. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ +typedef OrtStatus* (*CreateEpApiFactoriesFn)(_In_ const char* registered_name, _In_ const OrtApiBase* ort_api_base, + _Inout_ OrtEpFactory** factories, _In_ size_t max_factories, + _Out_ size_t* num_factories); + +/** \brief The function signature that ORT will call to release an OrtEpFactory instance. + * + * This must be available in a function called 'ReleaseEpFactory' in the execution provider library. + * + * \param[in] factory The OrtEpFactory instance to release. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version 1.22. + */ +typedef OrtStatus* (*ReleaseEpApiFactoryFn)(_In_ OrtEpFactory* factory); + +/** + * \brief The OrtEpFactory provides functions to create and manage execution providers. + * \since Version 1.22. + */ +struct OrtEpFactory { + /** \brief The ONNX Runtime version the execution provider was compiled with. + * + * Implementation should set to ORT_API_VERSION. + * ORT will use this to ensure it does not call functions that were not available when the library was compiled. + * + * \since Version 1.22. + */ + uint32_t ort_version_supported; + + /** \brief Get the name the of the execution provider that the factory creates. + * + * \param[in] this_ptr The OrtEpFactory instance. + * \return The name of the execution provider the factory creates. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* GetName)(const OrtEpFactory* this_ptr); + + /** \brief Get the name of vendor who owns the execution provider that the factory creates. + * + * \param[in] this_ptr The OrtEpFactory instance. + * \return vendor The vendor name of the execution provider the factory creates. + * + * \since Version 1.22. + */ + const char*(ORT_API_CALL* GetVendor)(const OrtEpFactory* this_ptr); // return EP vendor + + /** \brief Get information from the execution provider if it supports the OrtHardwareDevice. + * + * \param[in] this_ptr The OrtEpFactory instance. + * Non-const as the factory is passed through to the CreateEp call via the OrtEpDevice. + * \param[in] devices The OrtHardwareDevice instances that are available. + * \param[in] num_devices The number of OrtHardwareDevice instances. + * \param[out] ep_devices OrtEpDevice instances for each OrtHardwareDevice that the EP can use. + * The implementation should call OrtEpApi::CreateEpDevice to create, and add the OrtEpDevice + * instances to this pre-allocated array. ORT will take ownership of the values returned. + * i.e. usage is `ep_devices[0] = ;` + * \param[in] max_ep_devices The maximum number of OrtEpDevices that can be added to ep_devices. + * Current default is 8. This can be increased if needed. + * \param[out] num_ep_devices The number of EP devices added to ep_devices. + * \return true if the factory can create an execution provider that uses `device`. + * + * \note ORT will take ownership or ep_metadata and/or ep_options if they are not null. + * + * \since Version 1.22. + */ + OrtStatus*(ORT_API_CALL* GetSupportedDevices)(_In_ OrtEpFactory* this_ptr, + _In_reads_(num_devices) const OrtHardwareDevice* const* devices, + _In_ size_t num_devices, + _Inout_ OrtEpDevice** ep_devices, + _In_ size_t max_ep_devices, + _Out_ size_t* num_ep_devices); + + /** \brief Function to create an OrtEp instance for use in a Session. + * + * ORT will call ReleaseEp to release the instance when it is no longer needed. + * + * \param[in] this_ptr The OrtEpFactory instance. + * \param[in] devices The OrtHardwareDevice instances that the execution provider was selected to use. + * \param[in] ep_metadata_pairs Execution provider metadata that was provided to OrtEpApi::CreateEpDevice, for each + * device. + * \param[in] num_devices The number of devices the execution provider was selected for. + * \param[in] session_options The OrtSessionOptions instance that contains the configuration options for the + * session. This will include ep_options from GetSupportedDevices as well as any + * user provided overrides. + * Execution provider options will have been added with a prefix of 'ep..'. + * The OrtSessionOptions instance will NOT be valid after this call and should not be + * stored for later use. + * \param[in] logger The OrtLogger instance for the session that the execution provider should use for logging. + * \param[out] ep The OrtEp instance created by the factory. + * + * \snippet{doc} snippets.dox OrtStatus Return Value + * + * \since Version . This is a placeholder. + */ + OrtStatus*(ORT_API_CALL* CreateEp)(_In_ OrtEpFactory* this_ptr, + _In_reads_(num_devices) const OrtHardwareDevice* const* devices, + _In_reads_(num_devices) const OrtKeyValuePairs* const* ep_metadata_pairs, + _In_ size_t num_devices, + _In_ const OrtSessionOptions* session_options, + _In_ const OrtLogger* logger, _Outptr_ OrtEp** ep); + + /** \brief Release the OrtEp instance. + * + * \param[in] this_ptr The OrtEpFactory instance. + * \param[in] ep The OrtEp instance to release. + * + * \since Version . This is a placeholder. + */ + void(ORT_API_CALL* ReleaseEp)(OrtEpFactory* this_ptr, struct OrtEp* ep); +}; + +/* + * This is the old way to add the CUDA provider to the session, please use SessionOptionsAppendExecutionProvider_CUDA above to access the latest functionality + * This function always exists, but will only succeed if Onnxruntime was built with CUDA support and the CUDA provider shared library exists + * + * \param device_id CUDA device id, starts from zero. + */ +ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_CUDA, _In_ OrtSessionOptions* options, int device_id); + +/* + * This is the old way to add the ROCm provider to the session, please use + * SessionOptionsAppendExecutionProvider_ROCM above to access the latest functionality + * This function always exists, but will only succeed if Onnxruntime was built with + * HIP support and the ROCm provider shared library exists + * + * \param device_id HIP device id, starts from zero. + */ +ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_ROCM, _In_ OrtSessionOptions* options, int device_id); + +/* + * This is the old way to add the MIGraphX provider to the session, please use + * SessionOptionsAppendExecutionProvider_MIGraphX above to access the latest functionality + * This function always exists, but will only succeed if Onnxruntime was built with + * HIP support and the MIGraphX provider shared library exists + * + * \param device_id HIP device id, starts from zero. + */ +ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_MIGraphX, _In_ OrtSessionOptions* options, int device_id); + +/* + * This is the old way to add the oneDNN provider to the session, please use + * SessionOptionsAppendExecutionProvider_oneDNN above to access the latest functionality + * This function always exists, but will only succeed if Onnxruntime was built with + * oneDNN support and the oneDNN provider shared library exists + * + * \param use_arena zero: false. non-zero: true. + */ +ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_Dnnl, _In_ OrtSessionOptions* options, int use_arena); + +/* + * This is the old way to add the TensorRT provider to the session, please use SessionOptionsAppendExecutionProvider_TensorRT_V2 above to access the latest functionality + * This function always exists, but will only succeed if Onnxruntime was built with TensorRT support and the TensorRT provider shared library exists + * + * \param device_id CUDA device id, starts from zero. + */ +ORT_API_STATUS(OrtSessionOptionsAppendExecutionProvider_Tensorrt, _In_ OrtSessionOptions* options, int device_id); + +#ifdef __cplusplus +} +#endif +/// @} diff --git a/third_party/onnxruntime/include/onnxruntime_cxx_api.h b/third_party/onnxruntime/include/onnxruntime_cxx_api.h new file mode 100644 index 00000000..bc6f381b --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_cxx_api.h @@ -0,0 +1,2865 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +// Summary: The Ort C++ API is a header only wrapper around the Ort C API. +// +// The C++ API simplifies usage by returning values directly instead of error codes, throwing exceptions on errors +// and automatically releasing resources in the destructors. The primary purpose of C++ API is exception safety so +// all the resources follow RAII and do not leak memory. +// +// Each of the C++ wrapper classes holds only a pointer to the C internal object. Treat them like smart pointers. +// To create an empty object, pass 'nullptr' to the constructor (for example, Env e{nullptr};). However, you can't use them +// until you assign an instance that actually holds an underlying object. +// +// For Ort objects only move assignment between objects is allowed, there are no copy constructors. +// Some objects have explicit 'Clone' methods for this purpose. +// +// ConstXXXX types are copyable since they do not own the underlying C object, so you can pass them to functions as arguments +// by value or by reference. ConstXXXX types are restricted to const only interfaces. +// +// UnownedXXXX are similar to ConstXXXX but also allow non-const interfaces. +// +// The lifetime of the corresponding owning object must eclipse the lifetimes of the ConstXXXX/UnownedXXXX types. They exists so you do not +// have to fallback to C types and the API with the usual pitfalls. In general, do not use C API from your C++ code. + +#pragma once +#include "onnxruntime_c_api.h" +#include "onnxruntime_float16.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef ORT_NO_EXCEPTIONS +#include +#endif + +/** \brief All C++ Onnxruntime APIs are defined inside this namespace + * + */ +namespace Ort { + +/** \brief All C++ methods that can fail will throw an exception of this type + * + * If ORT_NO_EXCEPTIONS is defined, then any error will result in a call to abort() + */ +struct Exception : std::exception { + Exception(std::string&& string, OrtErrorCode code) : message_{std::move(string)}, code_{code} {} + + OrtErrorCode GetOrtErrorCode() const { return code_; } + const char* what() const noexcept override { return message_.c_str(); } + + private: + std::string message_; + OrtErrorCode code_; +}; + +#ifdef ORT_NO_EXCEPTIONS +// The #ifndef is for the very special case where the user of this library wants to define their own way of handling errors. +// NOTE: This header expects control flow to not continue after calling ORT_CXX_API_THROW +#ifndef ORT_CXX_API_THROW +#define ORT_CXX_API_THROW(string, code) \ + do { \ + std::cerr << Ort::Exception(string, code) \ + .what() \ + << std::endl; \ + abort(); \ + } while (false) +#endif +#else +#define ORT_CXX_API_THROW(string, code) \ + throw Ort::Exception(string, code) +#endif + +// This is used internally by the C++ API. This class holds the global variable that points to the OrtApi, +// it's in a template so that we can define a global variable in a header and make +// it transparent to the users of the API. +template +struct Global { + static const OrtApi* api_; +}; + +// If macro ORT_API_MANUAL_INIT is defined, no static initialization will be performed. Instead, user must call InitApi() before using it. +template +#ifdef ORT_API_MANUAL_INIT +const OrtApi* Global::api_{}; +inline void InitApi() noexcept { Global::api_ = OrtGetApiBase()->GetApi(ORT_API_VERSION); } + +// Used by custom operator libraries that are not linked to onnxruntime. Sets the global API object, which is +// required by C++ APIs. +// +// Example mycustomop.cc: +// +// #define ORT_API_MANUAL_INIT +// #include +// #undef ORT_API_MANUAL_INIT +// +// OrtStatus* ORT_API_CALL RegisterCustomOps(OrtSessionOptions* options, const OrtApiBase* api_base) { +// Ort::InitApi(api_base->GetApi(ORT_API_VERSION)); +// // ... +// } +// +inline void InitApi(const OrtApi* api) noexcept { Global::api_ = api; } +#else +#if defined(_MSC_VER) && !defined(__clang__) +#pragma warning(push) +// "Global initializer calls a non-constexpr function." Therefore you can't use ORT APIs in the other global initializers. +// Please define ORT_API_MANUAL_INIT if it conerns you. +#pragma warning(disable : 26426) +#endif +const OrtApi* Global::api_ = OrtGetApiBase()->GetApi(ORT_API_VERSION); +#if defined(_MSC_VER) && !defined(__clang__) +#pragma warning(pop) +#endif +#endif + +/// This returns a reference to the ORT C API. +inline const OrtApi& GetApi() noexcept { return *Global::api_; } + +/// +/// This function returns the onnxruntime version string +/// +/// version string major.minor.rev +std::string GetVersionString(); + +/// +/// This function returns the onnxruntime build information: including git branch, +/// git commit id, build type(Debug/Release/RelWithDebInfo) and cmake cpp flags. +/// +/// string +std::string GetBuildInfoString(); + +/// +/// This is a C++ wrapper for OrtApi::GetAvailableProviders() and +/// returns a vector of strings representing the available execution providers. +/// +/// vector of strings +std::vector GetAvailableProviders(); + +/// +/// This returns a reference to the ORT C Model Editor API. Used if building or augmenting a model at runtime. +/// +/// ORT C Model Editor API reference +inline const OrtModelEditorApi& GetModelEditorApi() { + auto* api = GetApi().GetModelEditorApi(); + if (api == nullptr) { + // minimal build + ORT_CXX_API_THROW("Model Editor API is not available in this build", ORT_FAIL); + } + + return *api; +} + +/// +/// This returns a reference to the ORT C Compile API. Used if compiling a model at runtime. +/// +/// ORT C Compile API reference +inline const OrtCompileApi& GetCompileApi() { + auto* api = GetApi().GetCompileApi(); + if (api == nullptr) { + // minimal build + ORT_CXX_API_THROW("Compile API is not available in this build", ORT_FAIL); + } + + return *api; +} + +/// +/// This returns a reference to the ORT C EP API. Used if authoring a plugin execution provider. +/// +/// ORT C EP API reference +inline const OrtEpApi& GetEpApi() { + auto* api = GetApi().GetEpApi(); + if (api == nullptr) { + // minimal build + ORT_CXX_API_THROW("EP API is not available in this build", ORT_FAIL); + } + + return *api; +} + +/** \brief IEEE 754 half-precision floating point data type + * + * \details This struct is used for converting float to float16 and back + * so the user could feed inputs and fetch outputs using these type. + * + * The size of the structure should align with uint16_t and one can freely cast + * uint16_t buffers to/from Ort::Float16_t to feed and retrieve data. + * + * \code{.unparsed} + * // This example demonstrates converion from float to float16 + * constexpr float values[] = {1.f, 2.f, 3.f, 4.f, 5.f}; + * std::vector fp16_values; + * fp16_values.reserve(std::size(values)); + * std::transform(std::begin(values), std::end(values), std::back_inserter(fp16_values), + * [](float value) { return Ort::Float16_t(value); }); + * + * \endcode + */ +struct Float16_t : onnxruntime_float16::Float16Impl { + private: + /// + /// Constructor from a 16-bit representation of a float16 value + /// No conversion is done here. + /// + /// 16-bit representation + constexpr explicit Float16_t(uint16_t v) noexcept { val = v; } + + public: + using Base = onnxruntime_float16::Float16Impl; + + /// + /// Default constructor + /// + Float16_t() = default; + + /// + /// Explicit conversion to uint16_t representation of float16. + /// + /// uint16_t bit representation of float16 + /// new instance of Float16_t + constexpr static Float16_t FromBits(uint16_t v) noexcept { return Float16_t(v); } + + /// + /// __ctor from float. Float is converted into float16 16-bit representation. + /// + /// float value + explicit Float16_t(float v) noexcept { val = Base::ToUint16Impl(v); } + + /// + /// Converts float16 to float + /// + /// float representation of float16 value + float ToFloat() const noexcept { return Base::ToFloatImpl(); } + + /// + /// Checks if the value is negative + /// + /// true if negative + using Base::IsNegative; + + /// + /// Tests if the value is NaN + /// + /// true if NaN + using Base::IsNaN; + + /// + /// Tests if the value is finite + /// + /// true if finite + using Base::IsFinite; + + /// + /// Tests if the value represents positive infinity. + /// + /// true if positive infinity + using Base::IsPositiveInfinity; + + /// + /// Tests if the value represents negative infinity + /// + /// true if negative infinity + using Base::IsNegativeInfinity; + + /// + /// Tests if the value is either positive or negative infinity. + /// + /// True if absolute value is infinity + using Base::IsInfinity; + + /// + /// Tests if the value is NaN or zero. Useful for comparisons. + /// + /// True if NaN or zero. + using Base::IsNaNOrZero; + + /// + /// Tests if the value is normal (not zero, subnormal, infinite, or NaN). + /// + /// True if so + using Base::IsNormal; + + /// + /// Tests if the value is subnormal (denormal). + /// + /// True if so + using Base::IsSubnormal; + + /// + /// Creates an instance that represents absolute value. + /// + /// Absolute value + using Base::Abs; + + /// + /// Creates a new instance with the sign flipped. + /// + /// Flipped sign instance + using Base::Negate; + + /// + /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check + /// for two values by or'ing the private bits together and stripping the sign. They are both zero, + /// and therefore equivalent, if the resulting value is still zero. + /// + /// first value + /// second value + /// True if both arguments represent zero + using Base::AreZero; + + /// + /// User defined conversion operator. Converts Float16_t to float. + /// + explicit operator float() const noexcept { return ToFloat(); } + + using Base::operator==; + using Base::operator!=; + using Base::operator<; +}; + +static_assert(sizeof(Float16_t) == sizeof(uint16_t), "Sizes must match"); + +/** \brief bfloat16 (Brain Floating Point) data type + * + * \details This struct is used for converting float to bfloat16 and back + * so the user could feed inputs and fetch outputs using these type. + * + * The size of the structure should align with uint16_t and one can freely cast + * uint16_t buffers to/from Ort::BFloat16_t to feed and retrieve data. + * + * \code{.unparsed} + * // This example demonstrates converion from float to float16 + * constexpr float values[] = {1.f, 2.f, 3.f, 4.f, 5.f}; + * std::vector bfp16_values; + * bfp16_values.reserve(std::size(values)); + * std::transform(std::begin(values), std::end(values), std::back_inserter(bfp16_values), + * [](float value) { return Ort::BFloat16_t(value); }); + * + * \endcode + */ +struct BFloat16_t : onnxruntime_float16::BFloat16Impl { + private: + /// + /// Constructor from a uint16_t representation of bfloat16 + /// used in FromBits() to escape overload resolution issue with + /// constructor from float. + /// No conversion is done. + /// + /// 16-bit bfloat16 value + constexpr explicit BFloat16_t(uint16_t v) noexcept { val = v; } + + public: + using Base = onnxruntime_float16::BFloat16Impl; + + BFloat16_t() = default; + + /// + /// Explicit conversion to uint16_t representation of bfloat16. + /// + /// uint16_t bit representation of bfloat16 + /// new instance of BFloat16_t + static constexpr BFloat16_t FromBits(uint16_t v) noexcept { return BFloat16_t(v); } + + /// + /// __ctor from float. Float is converted into bfloat16 16-bit representation. + /// + /// float value + explicit BFloat16_t(float v) noexcept { val = Base::ToUint16Impl(v); } + + /// + /// Converts bfloat16 to float + /// + /// float representation of bfloat16 value + float ToFloat() const noexcept { return Base::ToFloatImpl(); } + + /// + /// Checks if the value is negative + /// + /// true if negative + using Base::IsNegative; + + /// + /// Tests if the value is NaN + /// + /// true if NaN + using Base::IsNaN; + + /// + /// Tests if the value is finite + /// + /// true if finite + using Base::IsFinite; + + /// + /// Tests if the value represents positive infinity. + /// + /// true if positive infinity + using Base::IsPositiveInfinity; + + /// + /// Tests if the value represents negative infinity + /// + /// true if negative infinity + using Base::IsNegativeInfinity; + + /// + /// Tests if the value is either positive or negative infinity. + /// + /// True if absolute value is infinity + using Base::IsInfinity; + + /// + /// Tests if the value is NaN or zero. Useful for comparisons. + /// + /// True if NaN or zero. + using Base::IsNaNOrZero; + + /// + /// Tests if the value is normal (not zero, subnormal, infinite, or NaN). + /// + /// True if so + using Base::IsNormal; + + /// + /// Tests if the value is subnormal (denormal). + /// + /// True if so + using Base::IsSubnormal; + + /// + /// Creates an instance that represents absolute value. + /// + /// Absolute value + using Base::Abs; + + /// + /// Creates a new instance with the sign flipped. + /// + /// Flipped sign instance + using Base::Negate; + + /// + /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check + /// for two values by or'ing the private bits together and stripping the sign. They are both zero, + /// and therefore equivalent, if the resulting value is still zero. + /// + /// first value + /// second value + /// True if both arguments represent zero + using Base::AreZero; + + /// + /// User defined conversion operator. Converts BFloat16_t to float. + /// + explicit operator float() const noexcept { return ToFloat(); } + + // We do not have an inherited impl for the below operators + // as the internal class implements them a little differently + bool operator==(const BFloat16_t& rhs) const noexcept; + bool operator!=(const BFloat16_t& rhs) const noexcept { return !(*this == rhs); } + bool operator<(const BFloat16_t& rhs) const noexcept; +}; + +static_assert(sizeof(BFloat16_t) == sizeof(uint16_t), "Sizes must match"); + +/** \brief float8e4m3fn (Float8 Floating Point) data type + * \details It is necessary for type dispatching to make use of C++ API + * The type is implicitly convertible to/from uint8_t. + * See https://onnx.ai/onnx/technical/float8.html for further details. + */ +struct Float8E4M3FN_t { + uint8_t value; + constexpr Float8E4M3FN_t() noexcept : value(0) {} + constexpr Float8E4M3FN_t(uint8_t v) noexcept : value(v) {} + constexpr operator uint8_t() const noexcept { return value; } + // nan values are treated like any other value for operator ==, != + constexpr bool operator==(const Float8E4M3FN_t& rhs) const noexcept { return value == rhs.value; }; + constexpr bool operator!=(const Float8E4M3FN_t& rhs) const noexcept { return value != rhs.value; }; +}; + +static_assert(sizeof(Float8E4M3FN_t) == sizeof(uint8_t), "Sizes must match"); + +/** \brief float8e4m3fnuz (Float8 Floating Point) data type + * \details It is necessary for type dispatching to make use of C++ API + * The type is implicitly convertible to/from uint8_t. + * See https://onnx.ai/onnx/technical/float8.html for further details. + */ +struct Float8E4M3FNUZ_t { + uint8_t value; + constexpr Float8E4M3FNUZ_t() noexcept : value(0) {} + constexpr Float8E4M3FNUZ_t(uint8_t v) noexcept : value(v) {} + constexpr operator uint8_t() const noexcept { return value; } + // nan values are treated like any other value for operator ==, != + constexpr bool operator==(const Float8E4M3FNUZ_t& rhs) const noexcept { return value == rhs.value; }; + constexpr bool operator!=(const Float8E4M3FNUZ_t& rhs) const noexcept { return value != rhs.value; }; +}; + +static_assert(sizeof(Float8E4M3FNUZ_t) == sizeof(uint8_t), "Sizes must match"); + +/** \brief float8e5m2 (Float8 Floating Point) data type + * \details It is necessary for type dispatching to make use of C++ API + * The type is implicitly convertible to/from uint8_t. + * See https://onnx.ai/onnx/technical/float8.html for further details. + */ +struct Float8E5M2_t { + uint8_t value; + constexpr Float8E5M2_t() noexcept : value(0) {} + constexpr Float8E5M2_t(uint8_t v) noexcept : value(v) {} + constexpr operator uint8_t() const noexcept { return value; } + // nan values are treated like any other value for operator ==, != + constexpr bool operator==(const Float8E5M2_t& rhs) const noexcept { return value == rhs.value; }; + constexpr bool operator!=(const Float8E5M2_t& rhs) const noexcept { return value != rhs.value; }; +}; + +static_assert(sizeof(Float8E5M2_t) == sizeof(uint8_t), "Sizes must match"); + +/** \brief float8e5m2fnuz (Float8 Floating Point) data type + * \details It is necessary for type dispatching to make use of C++ API + * The type is implicitly convertible to/from uint8_t. + * See https://onnx.ai/onnx/technical/float8.html for further details. + */ +struct Float8E5M2FNUZ_t { + uint8_t value; + constexpr Float8E5M2FNUZ_t() noexcept : value(0) {} + constexpr Float8E5M2FNUZ_t(uint8_t v) noexcept : value(v) {} + constexpr operator uint8_t() const noexcept { return value; } + // nan values are treated like any other value for operator ==, != + constexpr bool operator==(const Float8E5M2FNUZ_t& rhs) const noexcept { return value == rhs.value; }; + constexpr bool operator!=(const Float8E5M2FNUZ_t& rhs) const noexcept { return value != rhs.value; }; +}; + +static_assert(sizeof(Float8E5M2FNUZ_t) == sizeof(uint8_t), "Sizes must match"); + +namespace detail { +// This is used internally by the C++ API. This macro is to make it easy to generate overloaded methods for all of the various OrtRelease* functions for every Ort* type +// This can't be done in the C API since C doesn't have function overloading. +#define ORT_DEFINE_RELEASE(NAME) \ + inline void OrtRelease(Ort##NAME* ptr) { GetApi().Release##NAME(ptr); } + +#define ORT_DEFINE_RELEASE_FROM_API_STRUCT(NAME, API_GETTER) \ + inline void OrtRelease(Ort##NAME* ptr) { API_GETTER().Release##NAME(ptr); } + +ORT_DEFINE_RELEASE(Allocator); +ORT_DEFINE_RELEASE(MemoryInfo); +ORT_DEFINE_RELEASE(CustomOpDomain); +ORT_DEFINE_RELEASE(ThreadingOptions); +ORT_DEFINE_RELEASE(Env); +ORT_DEFINE_RELEASE(RunOptions); +ORT_DEFINE_RELEASE(LoraAdapter); +ORT_DEFINE_RELEASE(Session); +ORT_DEFINE_RELEASE(SessionOptions); +ORT_DEFINE_RELEASE(TensorTypeAndShapeInfo); +ORT_DEFINE_RELEASE(SequenceTypeInfo); +ORT_DEFINE_RELEASE(MapTypeInfo); +ORT_DEFINE_RELEASE(TypeInfo); +ORT_DEFINE_RELEASE(Value); +ORT_DEFINE_RELEASE(ModelMetadata); +ORT_DEFINE_RELEASE(IoBinding); +ORT_DEFINE_RELEASE(ArenaCfg); +ORT_DEFINE_RELEASE(Status); +ORT_DEFINE_RELEASE(OpAttr); +ORT_DEFINE_RELEASE(Op); +ORT_DEFINE_RELEASE(KernelInfo); +ORT_DEFINE_RELEASE(ValueInfo); +ORT_DEFINE_RELEASE(Node); +ORT_DEFINE_RELEASE(Graph); +ORT_DEFINE_RELEASE(Model); +ORT_DEFINE_RELEASE(KeyValuePairs) +ORT_DEFINE_RELEASE_FROM_API_STRUCT(ModelCompilationOptions, GetCompileApi); +ORT_DEFINE_RELEASE_FROM_API_STRUCT(EpDevice, GetEpApi); + +#undef ORT_DEFINE_RELEASE +#undef ORT_DEFINE_RELEASE_FROM_API_STRUCT + +/** \brief This is a tagging template type. Use it with Base to indicate that the C++ interface object + * has no ownership of the underlying C object. + */ +template +struct Unowned { + using Type = T; +}; + +/** \brief Used internally by the C++ API. C++ wrapper types inherit from this. + * This is a zero cost abstraction to wrap the C API objects and delete them on destruction. + * + * All of the C++ classes + * a) serve as containers for pointers to objects that are created by the underlying C API. + * Their size is just a pointer size, no need to dynamically allocate them. Use them by value. + * b) Each of struct XXXX, XXX instances function as smart pointers to the underlying C API objects. + * they would release objects owned automatically when going out of scope, they are move-only. + * c) ConstXXXX and UnownedXXX structs function as non-owning, copyable containers for the above pointers. + * ConstXXXX allow calling const interfaces only. They give access to objects that are owned by somebody else + * such as Onnxruntime or instances of XXXX classes. + * d) serve convenient interfaces that return C++ objects and further enhance exception and type safety so they can be used + * in C++ code. + * + */ + +/// +/// This is a non-const pointer holder that is move-only. Disposes of the pointer on destruction. +/// +template +struct Base { + using contained_type = T; + + constexpr Base() = default; + constexpr explicit Base(contained_type* p) noexcept : p_{p} {} + ~Base() { + OrtRelease(p_); + } + + Base(const Base&) = delete; + Base& operator=(const Base&) = delete; + + Base(Base&& v) noexcept : p_{v.p_} { v.p_ = nullptr; } + Base& operator=(Base&& v) noexcept { + OrtRelease(p_); + p_ = v.release(); + return *this; + } + + constexpr operator contained_type*() const noexcept { return p_; } + + /// \brief Relinquishes ownership of the contained C object pointer + /// The underlying object is not destroyed + contained_type* release() { + T* p = p_; + p_ = nullptr; + return p; + } + + protected: + contained_type* p_{}; +}; + +// Undefined. For const types use Base> +template +struct Base; + +/// +/// Covers unowned pointers owned by either the ORT +/// or some other instance of CPP wrappers. +/// Used for ConstXXX and UnownedXXXX types that are copyable. +/// Also convenient to wrap raw OrtXX pointers . +/// +/// +template +struct Base> { + using contained_type = typename Unowned::Type; + + constexpr Base() = default; + constexpr explicit Base(contained_type* p) noexcept : p_{p} {} + + ~Base() = default; + + Base(const Base&) = default; + Base& operator=(const Base&) = default; + + Base(Base&& v) noexcept : p_{v.p_} { v.p_ = nullptr; } + Base& operator=(Base&& v) noexcept { + p_ = nullptr; + std::swap(p_, v.p_); + return *this; + } + + constexpr operator contained_type*() const noexcept { return p_; } + + protected: + contained_type* p_{}; +}; + +// Light functor to release memory with OrtAllocator +struct AllocatedFree { + OrtAllocator* allocator_; + explicit AllocatedFree(OrtAllocator* allocator) + : allocator_(allocator) {} + void operator()(void* ptr) const { + if (ptr) allocator_->Free(allocator_, ptr); + } +}; + +} // namespace detail + +struct AllocatorWithDefaultOptions; +struct Env; +struct EpDevice; +struct Graph; +struct Model; +struct Node; +struct ModelMetadata; +struct TypeInfo; +struct Value; +struct ValueInfo; + +/** \brief unique_ptr typedef used to own strings allocated by OrtAllocators + * and release them at the end of the scope. The lifespan of the given allocator + * must eclipse the lifespan of AllocatedStringPtr instance + */ +using AllocatedStringPtr = std::unique_ptr; + +/** \brief The Status that holds ownership of OrtStatus received from C API + * Use it to safely destroy OrtStatus* returned from the C API. Use appropriate + * constructors to construct an instance of a Status object from exceptions. + */ +struct Status : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit Status(std::nullptr_t) noexcept {} ///< Create an empty object, must be assigned a valid one to be used + explicit Status(OrtStatus* status) noexcept; ///< Takes ownership of OrtStatus instance returned from the C API. + explicit Status(const Exception&) noexcept; ///< Creates status instance out of exception + explicit Status(const std::exception&) noexcept; ///< Creates status instance out of exception + Status(const char* message, OrtErrorCode code) noexcept; ///< Creates status instance out of null-terminated string message. + std::string GetErrorMessage() const; + OrtErrorCode GetErrorCode() const; + bool IsOK() const noexcept; ///< Returns true if instance represents an OK (non-error) status. +}; + +/** \brief The ThreadingOptions + * + * The ThreadingOptions used for set global threadpools' options of The Env. + */ +struct ThreadingOptions : detail::Base { + /// \brief Wraps OrtApi::CreateThreadingOptions + ThreadingOptions(); + + /// \brief Wraps OrtApi::SetGlobalIntraOpNumThreads + ThreadingOptions& SetGlobalIntraOpNumThreads(int intra_op_num_threads); + + /// \brief Wraps OrtApi::SetGlobalInterOpNumThreads + ThreadingOptions& SetGlobalInterOpNumThreads(int inter_op_num_threads); + + /// \brief Wraps OrtApi::SetGlobalSpinControl + ThreadingOptions& SetGlobalSpinControl(int allow_spinning); + + /// \brief Wraps OrtApi::SetGlobalDenormalAsZero + ThreadingOptions& SetGlobalDenormalAsZero(); + + /// \brief Wraps OrtApi::SetGlobalCustomCreateThreadFn + ThreadingOptions& SetGlobalCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn); + + /// \brief Wraps OrtApi::SetGlobalCustomThreadCreationOptions + ThreadingOptions& SetGlobalCustomThreadCreationOptions(void* ort_custom_thread_creation_options); + + /// \brief Wraps OrtApi::SetGlobalCustomJoinThreadFn + ThreadingOptions& SetGlobalCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn); +}; + +namespace detail { +template +struct KeyValuePairsImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; + + const char* GetValue(const char* key) const; + + // get the pairs in unordered_map. needs to copy to std::string so the hash works as expected + std::unordered_map GetKeyValuePairs() const; + // get the pairs in two vectors. entries will be 1:1 between keys and values. avoids copying to std::string + void GetKeyValuePairs(std::vector& keys, std::vector& values) const; +}; +} // namespace detail + +// Const object holder that does not own the underlying object +using ConstKeyValuePairs = detail::KeyValuePairsImpl>; + +/** \brief Wrapper around ::OrtKeyValuePair */ +struct KeyValuePairs : detail::KeyValuePairsImpl { + explicit KeyValuePairs(std::nullptr_t) {} ///< No instance is created + /// Take ownership of a pointer created by C API + explicit KeyValuePairs(OrtKeyValuePairs* p) : KeyValuePairsImpl{p} {} + + /// \brief Wraps OrtApi::CreateKeyValuePairs + explicit KeyValuePairs(); + + /// \brief Wraps OrtApi::CreateKeyValuePairs and OrtApi::AddKeyValuePair + explicit KeyValuePairs(const std::unordered_map& kv_pairs); + + /// \brief Wraps OrtApi::AddKeyValuePair + void Add(const char* key, const char* value); + + /// \brief Wraps OrtApi::RemoveKeyValuePair + void Remove(const char* key); + + ConstKeyValuePairs GetConst() const { return ConstKeyValuePairs{this->p_}; } +}; + +namespace detail { +template +struct HardwareDeviceImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; + + OrtHardwareDeviceType Type() const; + uint32_t VendorId() const; + uint32_t DeviceId() const; + const char* Vendor() const; + ConstKeyValuePairs Metadata() const; +}; +} // namespace detail + +/** \brief Wrapper around ::OrtHardwareDevice + * \remarks HardwareDevice is always read-only for API users. + */ +using ConstHardwareDevice = detail::HardwareDeviceImpl>; + +namespace detail { +template +struct EpDeviceImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; + + const char* EpName() const; + const char* EpVendor() const; + ConstKeyValuePairs EpMetadata() const; + ConstKeyValuePairs EpOptions() const; + ConstHardwareDevice Device() const; +}; +} // namespace detail + +/** \brief Wrapper around ::OrtEpDevice + * \remarks EpDevice is always read-only for ORT API users. + */ +using ConstEpDevice = detail::EpDeviceImpl>; + +/** \brief Mutable EpDevice that is created by EpApi users. + */ +struct EpDevice : detail::EpDeviceImpl { + explicit EpDevice(std::nullptr_t) {} ///< No instance is created + explicit EpDevice(OrtEpDevice* p) : EpDeviceImpl{p} {} ///< Take ownership of a pointer created by C API + + /// \brief Wraps OrtEpApi::CreateEpDevice + EpDevice(OrtEpFactory& ep_factory, ConstHardwareDevice& hardware_device, + ConstKeyValuePairs ep_metadata = {}, ConstKeyValuePairs ep_options = {}); +}; + +/** \brief The Env (Environment) + * + * The Env holds the logging state used by all other objects. + * Note: One Env must be created before using any other Onnxruntime functionality + */ +struct Env : detail::Base { + explicit Env(std::nullptr_t) {} ///< Create an empty Env object, must be assigned a valid one to be used + + /// \brief Wraps OrtApi::CreateEnv + Env(OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = ""); + + /// \brief Wraps OrtApi::CreateEnvWithCustomLogger + Env(OrtLoggingLevel logging_level, const char* logid, OrtLoggingFunction logging_function, void* logger_param); + + /// \brief Wraps OrtApi::CreateEnvWithGlobalThreadPools + Env(const OrtThreadingOptions* tp_options, OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = ""); + + /// \brief Wraps OrtApi::CreateEnvWithCustomLoggerAndGlobalThreadPools + Env(const OrtThreadingOptions* tp_options, OrtLoggingFunction logging_function, void* logger_param, + OrtLoggingLevel logging_level = ORT_LOGGING_LEVEL_WARNING, _In_ const char* logid = ""); + + /// \brief C Interop Helper + explicit Env(OrtEnv* p) : Base{p} {} + + Env& EnableTelemetryEvents(); ///< Wraps OrtApi::EnableTelemetryEvents + Env& DisableTelemetryEvents(); ///< Wraps OrtApi::DisableTelemetryEvents + + Env& UpdateEnvWithCustomLogLevel(OrtLoggingLevel log_severity_level); ///< Wraps OrtApi::UpdateEnvWithCustomLogLevel + + Env& CreateAndRegisterAllocator(const OrtMemoryInfo* mem_info, const OrtArenaCfg* arena_cfg); ///< Wraps OrtApi::CreateAndRegisterAllocator + + Env& CreateAndRegisterAllocatorV2(const std::string& provider_type, const OrtMemoryInfo* mem_info, + const std::unordered_map& options, + const OrtArenaCfg* arena_cfg); ///< Wraps OrtApi::CreateAndRegisterAllocatorV2 + + Env& RegisterExecutionProviderLibrary(const char* registration_name, const std::basic_string& path); ///< Wraps OrtApi::RegisterExecutionProviderLibrary + Env& UnregisterExecutionProviderLibrary(const char* registration_name); ///< Wraps OrtApi::UnregisterExecutionProviderLibrary + + std::vector GetEpDevices() const; +}; + +/** \brief Custom Op Domain + * + */ +struct CustomOpDomain : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit CustomOpDomain(std::nullptr_t) {} ///< Create an empty CustomOpDomain object, must be assigned a valid one to be used + + /// \brief Wraps OrtApi::CreateCustomOpDomain + explicit CustomOpDomain(const char* domain); + + // This does not take ownership of the op, simply registers it. + void Add(const OrtCustomOp* op); ///< Wraps CustomOpDomain_Add +}; + +/// \brief LoraAdapter holds a set of Lora Parameters loaded from a single file +struct LoraAdapter : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit LoraAdapter(std::nullptr_t) {} ///< Create an empty LoraAdapter object, must be assigned a valid one to be used + /// \brief Wraps OrtApi::CreateLoraAdapter + /// + /// The function attempts to load the adapter from the specified file + /// \param adapter_path The path to the Lora adapter + /// \param allocator optional pointer to a device allocator. If nullptr, the data stays on CPU. It would still + /// be copied to device if required by the model at inference time. + static LoraAdapter CreateLoraAdapter(const std::basic_string& adapter_path, + OrtAllocator* allocator); + + /// \brief Wraps OrtApi::CreateLoraAdapterFromArray + /// + /// The function attempts to load the adapter from the specified byte array. + /// \param bytes The byte array containing file LoraAdapter format + /// \param num_bytes The number of bytes in the byte array + /// \param allocator optional pointer to a device allocator. If nullptr, the data stays on CPU. It would still + /// be copied to device if required by the model at inference time. + static LoraAdapter CreateLoraAdapterFromArray(const void* bytes, size_t num_bytes, + OrtAllocator* allocator); +}; + +/** \brief RunOptions + * + */ +struct RunOptions : detail::Base { + explicit RunOptions(std::nullptr_t) {} ///< Create an empty RunOptions object, must be assigned a valid one to be used + RunOptions(); ///< Wraps OrtApi::CreateRunOptions + + RunOptions& SetRunLogVerbosityLevel(int); ///< Wraps OrtApi::RunOptionsSetRunLogVerbosityLevel + int GetRunLogVerbosityLevel() const; ///< Wraps OrtApi::RunOptionsGetRunLogVerbosityLevel + + RunOptions& SetRunLogSeverityLevel(int); ///< Wraps OrtApi::RunOptionsSetRunLogSeverityLevel + int GetRunLogSeverityLevel() const; ///< Wraps OrtApi::RunOptionsGetRunLogSeverityLevel + + RunOptions& SetRunTag(const char* run_tag); ///< wraps OrtApi::RunOptionsSetRunTag + const char* GetRunTag() const; ///< Wraps OrtApi::RunOptionsGetRunTag + + RunOptions& AddConfigEntry(const char* config_key, const char* config_value); ///< Wraps OrtApi::AddRunConfigEntry + + /** \brief Terminates all currently executing Session::Run calls that were made using this RunOptions instance + * + * If a currently executing session needs to be force terminated, this can be called from another thread to force it to fail with an error + * Wraps OrtApi::RunOptionsSetTerminate + */ + RunOptions& SetTerminate(); + + /** \brief Clears the terminate flag so this RunOptions instance can be used in a new Session::Run call without it instantly terminating + * + * Wraps OrtApi::RunOptionsUnsetTerminate + */ + RunOptions& UnsetTerminate(); + + /** \brief Add the LoraAdapter to the list of active adapters. + * The setting does not affect RunWithBinding() calls. + * + * Wraps OrtApi::RunOptionsAddActiveLoraAdapter + * \param adapter The LoraAdapter to be used as the active adapter + */ + RunOptions& AddActiveLoraAdapter(const LoraAdapter& adapter); +}; + +namespace detail { +// Utility function that returns a SessionOption config entry key for a specific custom operator. +// Ex: custom_op.[custom_op_name].[config] +std::string MakeCustomOpConfigEntryKey(const char* custom_op_name, const char* config); +} // namespace detail + +/// +/// Class that represents session configuration entries for one or more custom operators. +/// +/// Example: +/// Ort::CustomOpConfigs op_configs; +/// op_configs.AddConfig("my_custom_op", "device_type", "CPU"); +/// +/// Passed to Ort::SessionOptions::RegisterCustomOpsLibrary. +/// +struct CustomOpConfigs { + CustomOpConfigs() = default; + ~CustomOpConfigs() = default; + CustomOpConfigs(const CustomOpConfigs&) = default; + CustomOpConfigs& operator=(const CustomOpConfigs&) = default; + CustomOpConfigs(CustomOpConfigs&& o) = default; + CustomOpConfigs& operator=(CustomOpConfigs&& o) = default; + + /** \brief Adds a session configuration entry/value for a specific custom operator. + * + * \param custom_op_name The name of the custom operator for which to add a configuration entry. + * Must match the name returned by the CustomOp's GetName() method. + * \param config_key The name of the configuration entry. + * \param config_value The value of the configuration entry. + * \return A reference to this object to enable call chaining. + */ + CustomOpConfigs& AddConfig(const char* custom_op_name, const char* config_key, const char* config_value); + + /** \brief Returns a flattened map of custom operator configuration entries and their values. + * + * The keys has been flattened to include both the custom operator name and the configuration entry key name. + * For example, a prior call to AddConfig("my_op", "key", "value") corresponds to the flattened key/value pair + * {"my_op.key", "value"}. + * + * \return An unordered map of flattened configurations. + */ + const std::unordered_map& GetFlattenedConfigs() const; + + private: + std::unordered_map flat_configs_; +}; + +/** \brief Options object used when creating a new Session object + * + * Wraps ::OrtSessionOptions object and methods + */ + +struct SessionOptions; + +namespace detail { +// we separate const-only methods because passing const ptr to non-const methods +// is only discovered when inline methods are compiled which is counter-intuitive +template +struct ConstSessionOptionsImpl : Base { + using B = Base; + using B::B; + + SessionOptions Clone() const; ///< Creates and returns a copy of this SessionOptions object. Wraps OrtApi::CloneSessionOptions + + std::string GetConfigEntry(const char* config_key) const; ///< Wraps OrtApi::GetSessionConfigEntry + bool HasConfigEntry(const char* config_key) const; ///< Wraps OrtApi::HasSessionConfigEntry + std::string GetConfigEntryOrDefault(const char* config_key, const std::string& def) const; +}; + +template +struct SessionOptionsImpl : ConstSessionOptionsImpl { + using B = ConstSessionOptionsImpl; + using B::B; + + SessionOptionsImpl& SetIntraOpNumThreads(int intra_op_num_threads); ///< Wraps OrtApi::SetIntraOpNumThreads + SessionOptionsImpl& SetInterOpNumThreads(int inter_op_num_threads); ///< Wraps OrtApi::SetInterOpNumThreads + SessionOptionsImpl& SetGraphOptimizationLevel(GraphOptimizationLevel graph_optimization_level); ///< Wraps OrtApi::SetSessionGraphOptimizationLevel + SessionOptionsImpl& SetDeterministicCompute(bool value); ///< Wraps OrtApi::SetDeterministicCompute + + SessionOptionsImpl& EnableCpuMemArena(); ///< Wraps OrtApi::EnableCpuMemArena + SessionOptionsImpl& DisableCpuMemArena(); ///< Wraps OrtApi::DisableCpuMemArena + + SessionOptionsImpl& SetOptimizedModelFilePath(const ORTCHAR_T* optimized_model_file); ///< Wraps OrtApi::SetOptimizedModelFilePath + + SessionOptionsImpl& EnableProfiling(const ORTCHAR_T* profile_file_prefix); ///< Wraps OrtApi::EnableProfiling + SessionOptionsImpl& DisableProfiling(); ///< Wraps OrtApi::DisableProfiling + + SessionOptionsImpl& EnableOrtCustomOps(); ///< Wraps OrtApi::EnableOrtCustomOps + + SessionOptionsImpl& EnableMemPattern(); ///< Wraps OrtApi::EnableMemPattern + SessionOptionsImpl& DisableMemPattern(); ///< Wraps OrtApi::DisableMemPattern + + SessionOptionsImpl& SetExecutionMode(ExecutionMode execution_mode); ///< Wraps OrtApi::SetSessionExecutionMode + + SessionOptionsImpl& SetLoadCancellationFlag(bool value); ///< Wraps OrtApi::SessionOptionsSetLoadCancellationFlag + + SessionOptionsImpl& SetLogId(const char* logid); ///< Wraps OrtApi::SetSessionLogId + SessionOptionsImpl& SetLogSeverityLevel(int level); ///< Wraps OrtApi::SetSessionLogSeverityLevel + + SessionOptionsImpl& Add(OrtCustomOpDomain* custom_op_domain); ///< Wraps OrtApi::AddCustomOpDomain + + SessionOptionsImpl& DisablePerSessionThreads(); ///< Wraps OrtApi::DisablePerSessionThreads + + SessionOptionsImpl& AddConfigEntry(const char* config_key, const char* config_value); ///< Wraps OrtApi::AddSessionConfigEntry + + SessionOptionsImpl& AddInitializer(const char* name, const OrtValue* ort_val); ///< Wraps OrtApi::AddInitializer + SessionOptionsImpl& AddExternalInitializers(const std::vector& names, const std::vector& ort_values); ///< Wraps OrtApi::AddExternalInitializers + SessionOptionsImpl& AddExternalInitializersFromFilesInMemory(const std::vector>& external_initializer_file_names, + const std::vector& external_initializer_file_buffer_array, + const std::vector& external_initializer_file_lengths); ///< Wraps OrtApi::AddExternalInitializersFromFilesInMemory + + SessionOptionsImpl& AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA + SessionOptionsImpl& AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_CUDA_V2 + SessionOptionsImpl& AppendExecutionProvider_ROCM(const OrtROCMProviderOptions& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_ROCM + SessionOptionsImpl& AppendExecutionProvider_OpenVINO(const OrtOpenVINOProviderOptions& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO + ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_OpenVINO_V2 + SessionOptionsImpl& AppendExecutionProvider_OpenVINO_V2(const std::unordered_map& provider_options = {}); + SessionOptionsImpl& AppendExecutionProvider_TensorRT(const OrtTensorRTProviderOptions& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_TensorRT + SessionOptionsImpl& AppendExecutionProvider_TensorRT_V2(const OrtTensorRTProviderOptionsV2& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_TensorRT + SessionOptionsImpl& AppendExecutionProvider_MIGraphX(const OrtMIGraphXProviderOptions& provider_options); ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_MIGraphX + /// Wraps OrtApi::SessionOptionsAppendExecutionProvider_CANN + SessionOptionsImpl& AppendExecutionProvider_CANN(const OrtCANNProviderOptions& provider_options); + /// Wraps OrtApi::SessionOptionsAppendExecutionProvider_Dnnl + SessionOptionsImpl& AppendExecutionProvider_Dnnl(const OrtDnnlProviderOptions& provider_options); + /// Wraps OrtApi::SessionOptionsAppendExecutionProvider. Currently supports QNN, SNPE and XNNPACK. + SessionOptionsImpl& AppendExecutionProvider(const std::string& provider_name, + const std::unordered_map& provider_options = {}); + + /// Append EPs that have been registered previously with the OrtEnv. + /// Wraps OrtApi::SessionOptionsAppendExecutionProvider_V2 + SessionOptionsImpl& AppendExecutionProvider_V2(Env& env, const std::vector& ep_devices, + const KeyValuePairs& ep_options); + /// Append EPs that have been registered previously with the OrtEnv. + /// Wraps OrtApi::SessionOptionsAppendExecutionProvider_V2 + SessionOptionsImpl& AppendExecutionProvider_V2(Env& env, const std::vector& ep_devices, + const std::unordered_map& ep_options); + + /// Wraps OrtApi::SessionOptionsSetEpSelectionPolicy + SessionOptionsImpl& SetEpSelectionPolicy(OrtExecutionProviderDevicePolicy policy); + + /// Wraps OrtApi::SessionOptionsSetEpSelectionPolicyDelegate + SessionOptionsImpl& SetEpSelectionPolicy(EpSelectionDelegate delegate, void* state = nullptr); + + SessionOptionsImpl& SetCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn); ///< Wraps OrtApi::SessionOptionsSetCustomCreateThreadFn + SessionOptionsImpl& SetCustomThreadCreationOptions(void* ort_custom_thread_creation_options); ///< Wraps OrtApi::SessionOptionsSetCustomThreadCreationOptions + SessionOptionsImpl& SetCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn); ///< Wraps OrtApi::SessionOptionsSetCustomJoinThreadFn + + ///< Registers the custom operator from the specified shared library via OrtApi::RegisterCustomOpsLibrary_V2. + ///< The custom operator configurations are optional. If provided, custom operator configs are set via + ///< OrtApi::AddSessionConfigEntry. + SessionOptionsImpl& RegisterCustomOpsLibrary(const ORTCHAR_T* library_name, const CustomOpConfigs& custom_op_configs = {}); + + SessionOptionsImpl& RegisterCustomOpsUsingFunction(const char* function_name); ///< Wraps OrtApi::RegisterCustomOpsUsingFunction + + ///< Wraps OrtApi::SessionOptionsAppendExecutionProvider_VitisAI + SessionOptionsImpl& AppendExecutionProvider_VitisAI(const std::unordered_map& provider_options = {}); +}; +} // namespace detail + +using UnownedSessionOptions = detail::SessionOptionsImpl>; +using ConstSessionOptions = detail::ConstSessionOptionsImpl>; + +/** \brief Wrapper around ::OrtSessionOptions + * + */ +struct SessionOptions : detail::SessionOptionsImpl { + explicit SessionOptions(std::nullptr_t) {} ///< Create an empty SessionOptions object, must be assigned a valid one to be used + SessionOptions(); ///< Wraps OrtApi::CreateSessionOptions + explicit SessionOptions(OrtSessionOptions* p) : SessionOptionsImpl{p} {} ///< Used for interop with the C API + UnownedSessionOptions GetUnowned() const { return UnownedSessionOptions{this->p_}; } + ConstSessionOptions GetConst() const { return ConstSessionOptions{this->p_}; } +}; + +/** \brief Options object used when compiling a model. + * + * Wraps ::OrtModelCompilationOptions object and methods + */ +struct ModelCompilationOptions : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit ModelCompilationOptions(std::nullptr_t) {} ///< Create an empty ModelCompilationOptions object, must be assigned a valid one to be used. + + ModelCompilationOptions(const Env& env, const SessionOptions& session_options); ///< Wraps OrtApi::CreateModelCompilationOptionsFromSessionOptions + ModelCompilationOptions(const Env& env, ConstSessionOptions session_options); ///< Wraps OrtApi::CreateModelCompilationOptionsFromSessionOptions + + ModelCompilationOptions& SetInputModelPath(const ORTCHAR_T* input_model_path); ///< Wraps OrtApi::ModelCompilationOptions_SetInputModelPath + ModelCompilationOptions& SetInputModelFromBuffer(const void* input_model_data, + size_t input_model_data_size); ///< Wraps OrtApi::ModelCompilationOptions_SetInputModelFromBuffer + ModelCompilationOptions& SetEpContextEmbedMode(bool embed_ep_context_in_model); ///< Wraps OrtApi::ModelCompilationOptions_SetEpContextEmbedMode + ModelCompilationOptions& SetOutputModelPath(const ORTCHAR_T* output_model_path); ///< Wraps OrtApi::ModelCompilationOptions_SetOutputModelPath + ModelCompilationOptions& SetOutputModelExternalInitializersFile(const ORTCHAR_T* file_path, + size_t initializer_size_threshold); ///< Wraps OrtApi::ModelCompilationOptions_SetOutputModelExternalInitializersFile + ModelCompilationOptions& SetOutputModelBuffer(OrtAllocator* allocator, void** output_model_buffer_ptr, + size_t* output_model_buffer_size_ptr); ///< Wraps OrtApi::ModelCompilationOptions_SetOutputModelBuffer +}; + +/** \brief Compiles an input model to generate a model with EPContext nodes that execute EP-specific kernels. Wraps OrtApi::CompileModels. + * + * \param env: ORT environment object. + * \param model_compilation_options: Compilation options for a model. + * \return A Status indicating success or failure. + */ +Status CompileModel(const Env& env, const ModelCompilationOptions& model_compilation_options); + +/** \brief Wrapper around ::OrtModelMetadata + * + */ +struct ModelMetadata : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit ModelMetadata(std::nullptr_t) {} ///< Create an empty ModelMetadata object, must be assigned a valid one to be used + + /** \brief Returns a copy of the producer name. + * + * \param allocator to allocate memory for the copy of the name returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetProducerNameAllocated(OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataGetProducerName + + /** \brief Returns a copy of the graph name. + * + * \param allocator to allocate memory for the copy of the name returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetGraphNameAllocated(OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataGetGraphName + + /** \brief Returns a copy of the domain name. + * + * \param allocator to allocate memory for the copy of the name returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetDomainAllocated(OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataGetDomain + + /** \brief Returns a copy of the description. + * + * \param allocator to allocate memory for the copy of the string returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetDescriptionAllocated(OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataGetDescription + + /** \brief Returns a copy of the graph description. + * + * \param allocator to allocate memory for the copy of the string returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetGraphDescriptionAllocated(OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataGetGraphDescription + + /** \brief Returns a vector of copies of the custom metadata keys. + * + * \param allocator to allocate memory for the copy of the string returned + * \return a instance std::vector of smart pointers that would deallocate the buffers when out of scope. + * The OrtAllocator instance must be valid at the point of memory release. + */ + std::vector GetCustomMetadataMapKeysAllocated(OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataGetCustomMetadataMapKeys + + /** \brief Looks up a value by a key in the Custom Metadata map + * + * \param key zero terminated string key to lookup + * \param allocator to allocate memory for the copy of the string returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * maybe nullptr if key is not found. + * + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr LookupCustomMetadataMapAllocated(const char* key, OrtAllocator* allocator) const; ///< Wraps OrtApi::ModelMetadataLookupCustomMetadataMap + + int64_t GetVersion() const; ///< Wraps OrtApi::ModelMetadataGetVersion +}; + +struct IoBinding; + +namespace detail { + +// we separate const-only methods because passing const ptr to non-const methods +// is only discovered when inline methods are compiled which is counter-intuitive +template +struct ConstSessionImpl : Base { + using B = Base; + using B::B; + + size_t GetInputCount() const; ///< Returns the number of model inputs + size_t GetOutputCount() const; ///< Returns the number of model outputs + size_t GetOverridableInitializerCount() const; ///< Returns the number of inputs that have defaults that can be overridden + + std::vector GetInputNames() const; + std::vector GetOutputNames() const; + std::vector GetOverridableInitializerNames() const; + + /** \brief Returns a copy of input name at the specified index. + * + * \param index must less than the value returned by GetInputCount() + * \param allocator to allocate memory for the copy of the name returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetInputNameAllocated(size_t index, OrtAllocator* allocator) const; + + /** \brief Returns a copy of output name at then specified index. + * + * \param index must less than the value returned by GetOutputCount() + * \param allocator to allocate memory for the copy of the name returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetOutputNameAllocated(size_t index, OrtAllocator* allocator) const; + + /** \brief Returns a copy of the overridable initializer name at then specified index. + * + * \param index must less than the value returned by GetOverridableInitializerCount() + * \param allocator to allocate memory for the copy of the name returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr GetOverridableInitializerNameAllocated(size_t index, OrtAllocator* allocator) const; ///< Wraps OrtApi::SessionGetOverridableInitializerName + + uint64_t GetProfilingStartTimeNs() const; ///< Wraps OrtApi::SessionGetProfilingStartTimeNs + ModelMetadata GetModelMetadata() const; ///< Wraps OrtApi::SessionGetModelMetadata + + TypeInfo GetInputTypeInfo(size_t index) const; ///< Wraps OrtApi::SessionGetInputTypeInfo + TypeInfo GetOutputTypeInfo(size_t index) const; ///< Wraps OrtApi::SessionGetOutputTypeInfo + TypeInfo GetOverridableInitializerTypeInfo(size_t index) const; ///< Wraps OrtApi::SessionGetOverridableInitializerTypeInfo + + int GetOpset(const std::string& domain) const; ///< Wraps OrtApi::SessionGetOpsetForDomain + + // Will move before checkin if that's the case. + std::vector GetInputs() const; + std::vector GetOutputs() const; +}; + +template +struct SessionImpl : ConstSessionImpl { + using B = ConstSessionImpl; + using B::B; + + /** \brief Run the model returning results in an Ort allocated vector. + * + * Wraps OrtApi::Run + * + * The caller provides a list of inputs and a list of the desired outputs to return. + * + * See the output logs for more information on warnings/errors that occur while processing the model. + * Common errors are.. (TODO) + * + * \param[in] run_options + * \param[in] input_names Array of null terminated strings of length input_count that is the list of input names + * \param[in] input_values Array of Value objects of length input_count that is the list of input values + * \param[in] input_count Number of inputs (the size of the input_names & input_values arrays) + * \param[in] output_names Array of C style strings of length output_count that is the list of output names + * \param[in] output_count Number of outputs (the size of the output_names array) + * \return A std::vector of Value objects that directly maps to the output_names array (eg. output_name[0] is the first entry of the returned vector) + */ + std::vector Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count, + const char* const* output_names, size_t output_count); + + /** \brief Run the model returning results in user provided outputs + * Same as Run(const RunOptions&, const char* const*, const Value*, size_t,const char* const*, size_t) + */ + void Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count, + const char* const* output_names, Value* output_values, size_t output_count); + + void Run(const RunOptions& run_options, const IoBinding&); ///< Wraps OrtApi::RunWithBinding + + /** \brief Run the model asynchronously in a thread owned by intra op thread pool + * + * Wraps OrtApi::RunAsync + * + * \param[in] run_options + * \param[in] input_names Array of null terminated UTF8 encoded strings of the input names + * \param[in] input_values Array of Value objects of length input_count + * \param[in] input_count Number of elements in the input_names and inputs arrays + * \param[in] output_names Array of null terminated UTF8 encoded strings of the output names + * \param[out] output_values Array of provided Values to be filled with outputs. + * On calling RunAsync, output_values[i] could either be initialized by a null pointer or a preallocated OrtValue*. + * Later, on invoking the callback, each output_values[i] of null will be filled with an OrtValue* allocated by onnxruntime. + * Then, an OrtValue** pointer will be casted from output_values, and pass to the callback. + * NOTE: it is customer's duty to finally release output_values and each of its member, + * regardless of whether the member (Ort::Value) is allocated by onnxruntime or preallocated by the customer. + * \param[in] output_count Number of elements in the output_names and outputs array + * \param[in] callback Callback function on model run completion + * \param[in] user_data User data that pass back to the callback + */ + void RunAsync(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count, + const char* const* output_names, Value* output_values, size_t output_count, RunAsyncCallbackFn callback, void* user_data); + + /** \brief End profiling and return a copy of the profiling file name. + * + * \param allocator to allocate memory for the copy of the string returned + * \return a instance of smart pointer that would deallocate the buffer when out of scope. + * The OrtAllocator instances must be valid at the point of memory release. + */ + AllocatedStringPtr EndProfilingAllocated(OrtAllocator* allocator); ///< Wraps OrtApi::SessionEndProfiling + + /** \brief Set DynamicOptions for EPs (Execution Providers) + * + * Wraps OrtApi::SetEpDynamicOptions + * + * Valid options can be found in `include\onnxruntime\core\session\onnxruntime_session_options_config_keys.h` + * Look for `kOrtEpDynamicOptions` + * + * \param[in] keys Array of null terminated UTF8 encoded strings of EP dynamic option keys + * \param[in] values Array of null terminated UTF8 encoded string of EP dynamic option values + * \param[in] kv_len Number of elements in the keys and values arrays + */ + void SetEpDynamicOptions(const char* const* keys, const char* const* values, size_t kv_len); + + void FinalizeModelEditorSession(const Model& model, const SessionOptions& options, + OrtPrepackedWeightsContainer* prepacked_weights_container = nullptr); +}; + +} // namespace detail + +using ConstSession = detail::ConstSessionImpl>; +using UnownedSession = detail::SessionImpl>; + +/** \brief Wrapper around ::OrtSession + * + */ +struct Session : detail::SessionImpl { + /// Create an empty Session object, must be assigned a valid one to be used. Wraps OrtApi::CreateSession + explicit Session(std::nullptr_t) {} + explicit Session(OrtSession* p) : SessionImpl{p} {} ///< C API Interop + + Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options); + + /// Wraps OrtApi::CreateSessionWithPrepackedWeightsContainer + Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options, + OrtPrepackedWeightsContainer* prepacked_weights_container); + + /// Wraps OrtApi::CreateSessionFromArray + Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options); + + /// Wraps OrtApi::CreateSessionFromArrayWithPrepackedWeightsContainer + Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options, + OrtPrepackedWeightsContainer* prepacked_weights_container); + +#if !defined(ORT_MINIMAL_BUILD) + /// Wraps OrtModelEditorApi::CreateSessionFromModel + Session(const Env& env, const Model& model, const SessionOptions& options); + + /// Wraps OrtModelEditorApi::CreateModelEditorSession + static Session CreateModelEditorSession(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options); + + /// Wraps OrtModelEditorApi::CreateModelEditorSession + static Session CreateModelEditorSession(const Env& env, const void* model_data, size_t model_data_length, + const SessionOptions& options); +#endif // !defined(ORT_MINIMAL_BUILD) + + ConstSession GetConst() const { return ConstSession{this->p_}; } + UnownedSession GetUnowned() const { return UnownedSession{this->p_}; } +}; + +namespace detail { +template +struct MemoryInfoImpl : Base { + using B = Base; + using B::B; + + std::string GetAllocatorName() const; + OrtAllocatorType GetAllocatorType() const; + int GetDeviceId() const; + OrtMemoryInfoDeviceType GetDeviceType() const; + OrtMemType GetMemoryType() const; + + template + bool operator==(const MemoryInfoImpl& o) const; +}; +} // namespace detail + +// Const object holder that does not own the underlying object +using ConstMemoryInfo = detail::MemoryInfoImpl>; + +/** \brief Wrapper around ::OrtMemoryInfo + * + */ +struct MemoryInfo : detail::MemoryInfoImpl { + static MemoryInfo CreateCpu(OrtAllocatorType type, OrtMemType mem_type1); + explicit MemoryInfo(std::nullptr_t) {} ///< No instance is created + explicit MemoryInfo(OrtMemoryInfo* p) : MemoryInfoImpl{p} {} ///< Take ownership of a pointer created by C API + MemoryInfo(const char* name, OrtAllocatorType type, int id, OrtMemType mem_type); + ConstMemoryInfo GetConst() const { return ConstMemoryInfo{this->p_}; } +}; + +namespace detail { +template +struct TensorTypeAndShapeInfoImpl : Base { + using B = Base; + using B::B; + + ONNXTensorElementDataType GetElementType() const; ///< Wraps OrtApi::GetTensorElementType + size_t GetElementCount() const; ///< Wraps OrtApi::GetTensorShapeElementCount + + size_t GetDimensionsCount() const; ///< Wraps OrtApi::GetDimensionsCount + + /** \deprecated use GetShape() returning std::vector + * [[deprecated]] + * This interface is unsafe to use + */ + [[deprecated("use GetShape()")]] void GetDimensions(int64_t* values, size_t values_count) const; ///< Wraps OrtApi::GetDimensions + + void GetSymbolicDimensions(const char** values, size_t values_count) const; ///< Wraps OrtApi::GetSymbolicDimensions + std::vector GetSymbolicDimensions() const; + + std::vector GetShape() const; ///< Uses GetDimensionsCount & GetDimensions to return a std::vector of the shape +}; + +} // namespace detail + +using ConstTensorTypeAndShapeInfo = detail::TensorTypeAndShapeInfoImpl>; + +/** \brief Wrapper around ::OrtTensorTypeAndShapeInfo + * + */ +struct TensorTypeAndShapeInfo : detail::TensorTypeAndShapeInfoImpl { + using Base = detail::TensorTypeAndShapeInfoImpl; + using Base::Base; + + /// Create an empty TensorTypeAndShapeInfo object, must be assigned a valid one to be used + explicit TensorTypeAndShapeInfo(std::nullptr_t) {} + /// Used for interop with the C API + explicit TensorTypeAndShapeInfo(OrtTensorTypeAndShapeInfo* p) : TensorTypeAndShapeInfoImpl{p} {} + + // Create a TensorTypeAndShapeInfo object with the specified element type and dimensions + // symbolic_dims are optional, but should be 1:1 with dims. + // The value in symbolic_dims will be used for all entries in dims that are -1. + explicit TensorTypeAndShapeInfo(ONNXTensorElementDataType element_type, + const std::vector& dims, + const std::vector* symbolic_dims = nullptr); + + ConstTensorTypeAndShapeInfo GetConst() const { return ConstTensorTypeAndShapeInfo{this->p_}; } +}; + +namespace detail { +template +struct SequenceTypeInfoImpl : Base { + using B = Base; + using B::B; + TypeInfo GetSequenceElementType() const; ///< Wraps OrtApi::GetSequenceElementType +}; + +} // namespace detail + +using ConstSequenceTypeInfo = detail::SequenceTypeInfoImpl>; + +/** \brief Wrapper around ::OrtSequenceTypeInfo + * + */ +struct SequenceTypeInfo : detail::SequenceTypeInfoImpl { + using Base = detail::SequenceTypeInfoImpl; + using Base::Base; + + explicit SequenceTypeInfo(std::nullptr_t) {} ///< Create an empty SequenceTypeInfo object, must be assigned a valid one to be used + explicit SequenceTypeInfo(OrtSequenceTypeInfo* p) : SequenceTypeInfoImpl{p} {} ///< Used for interop with the C API + ConstSequenceTypeInfo GetConst() const { return ConstSequenceTypeInfo{this->p_}; } +}; + +namespace detail { +template +struct OptionalTypeInfoImpl : Base { + using B = Base; + using B::B; + TypeInfo GetOptionalElementType() const; ///< Wraps OrtApi::CastOptionalTypeToContainedTypeInfo +}; + +} // namespace detail + +// This is always owned by the TypeInfo and can only be obtained from it. +using ConstOptionalTypeInfo = detail::OptionalTypeInfoImpl>; + +namespace detail { +template +struct MapTypeInfoImpl : detail::Base { + using B = Base; + using B::B; + ONNXTensorElementDataType GetMapKeyType() const; ///< Wraps OrtApi::GetMapKeyType + TypeInfo GetMapValueType() const; ///< Wraps OrtApi::GetMapValueType +}; + +} // namespace detail + +using ConstMapTypeInfo = detail::MapTypeInfoImpl>; + +/** \brief Wrapper around ::OrtMapTypeInfo + * + */ +struct MapTypeInfo : detail::MapTypeInfoImpl { + using Base = detail::MapTypeInfoImpl; + using Base::Base; + + explicit MapTypeInfo(std::nullptr_t) {} ///< Create an empty MapTypeInfo object, must be assigned a valid one to be used + explicit MapTypeInfo(OrtMapTypeInfo* p) : MapTypeInfoImpl{p} {} ///< Used for interop with the C API + ConstMapTypeInfo GetConst() const { return ConstMapTypeInfo{this->p_}; } +}; + +namespace detail { +template +struct TypeInfoImpl : detail::Base { + using B = Base; + using B::B; + + ConstTensorTypeAndShapeInfo GetTensorTypeAndShapeInfo() const; ///< Wraps OrtApi::CastTypeInfoToTensorInfo + ConstSequenceTypeInfo GetSequenceTypeInfo() const; ///< Wraps OrtApi::CastTypeInfoToSequenceTypeInfo + ConstMapTypeInfo GetMapTypeInfo() const; ///< Wraps OrtApi::CastTypeInfoToMapTypeInfo + ConstOptionalTypeInfo GetOptionalTypeInfo() const; ///< wraps OrtApi::CastTypeInfoToOptionalTypeInfo + + ONNXType GetONNXType() const; +}; +} // namespace detail + +/// +/// Contains a constant, unowned OrtTypeInfo that can be copied and passed around by value. +/// Provides access to const OrtTypeInfo APIs. +/// +using ConstTypeInfo = detail::TypeInfoImpl>; + +/// +/// Type information that may contain either TensorTypeAndShapeInfo or +/// the information about contained sequence or map depending on the ONNXType. +/// +struct TypeInfo : detail::TypeInfoImpl { + using Base = detail::TypeInfoImpl; + using Base::Base; + + /// Create an empty TypeInfo object, must be assigned a valid one to be used + explicit TypeInfo(std::nullptr_t) {} + explicit TypeInfo(OrtTypeInfo* p) : TypeInfoImpl{p} {} ///< C API Interop + +#if !defined(ORT_MINIMAL_BUILD) + static TypeInfo CreateTensorInfo(ConstTensorTypeAndShapeInfo tensor_info); + static TypeInfo CreateSparseTensorInfo(ConstTensorTypeAndShapeInfo sparse_tensor_info); + static TypeInfo CreateSequenceTypeInfo(ConstTypeInfo sequence_type); + static TypeInfo CreateMapTypeInfo(ONNXTensorElementDataType key_type, ConstTypeInfo value_type); + static TypeInfo CreateOptionalTypeInfo(ConstTypeInfo contained_type); +#endif // !defined(ORT_MINIMAL_BUILD) + + ConstTypeInfo GetConst() const { return ConstTypeInfo{this->p_}; } +}; + +namespace detail { +// This structure is used to feed sparse tensor values +// information for use with FillSparseTensor() API +// if the data type for the sparse tensor values is numeric +// use data.p_data, otherwise, use data.str pointer to feed +// values. data.str is an array of const char* that are zero terminated. +// number of strings in the array must match shape size. +// For fully sparse tensors use shape {0} and set p_data/str +// to nullptr. +struct OrtSparseValuesParam { + const int64_t* values_shape; + size_t values_shape_len; + union { + const void* p_data; + const char** str; + } data; +}; + +// Provides a way to pass shape in a single +// argument +struct Shape { + const int64_t* shape; + size_t shape_len; +}; + +template +struct ConstValueImpl : Base { + using B = Base; + using B::B; + + /// + /// Obtains a pointer to a user defined data for experimental purposes + /// + template + void GetOpaqueData(const char* domain, const char* type_name, R&) const; ///< Wraps OrtApi::GetOpaqueValue + + bool IsTensor() const; ///< Returns true if Value is a tensor, false for other types like map/sequence/etc + bool HasValue() const; /// < Return true if OrtValue contains data and returns false if the OrtValue is a None + + size_t GetCount() const; // If a non tensor, returns 2 for map and N for sequence, where N is the number of elements + Value GetValue(int index, OrtAllocator* allocator) const; + + /// + /// This API returns a full length of string data contained within either a tensor or a sparse Tensor. + /// For sparse tensor it returns a full length of stored non-empty strings (values). The API is useful + /// for allocating necessary memory and calling GetStringTensorContent(). + /// + /// total length of UTF-8 encoded bytes contained. No zero terminators counted. + size_t GetStringTensorDataLength() const; + + /// + /// The API copies all of the UTF-8 encoded string data contained within a tensor or a sparse tensor + /// into a supplied buffer. Use GetStringTensorDataLength() to find out the length of the buffer to allocate. + /// The user must also allocate offsets buffer with the number of entries equal to that of the contained + /// strings. + /// + /// Strings are always assumed to be on CPU, no X-device copy. + /// + /// user allocated buffer + /// length in bytes of the allocated buffer + /// a pointer to the offsets user allocated buffer + /// count of offsets, must be equal to the number of strings contained. + /// that can be obtained from the shape of the tensor or from GetSparseTensorValuesTypeAndShapeInfo() + /// for sparse tensors + void GetStringTensorContent(void* buffer, size_t buffer_length, size_t* offsets, size_t offsets_count) const; + + /// + /// Returns a const typed pointer to the tensor contained data. + /// No type checking is performed, the caller must ensure the type matches the tensor type. + /// + /// + /// const pointer to data, no copies made + template + const R* GetTensorData() const; ///< Wraps OrtApi::GetTensorMutableData /// + + /// + /// Returns a non-typed pointer to a tensor contained data. + /// + /// const pointer to data, no copies made + const void* GetTensorRawData() const; + + /// + /// The API returns type information for data contained in a tensor. For sparse + /// tensors it returns type information for contained non-zero values. + /// It returns dense shape for sparse tensors. + /// + /// TypeInfo + TypeInfo GetTypeInfo() const; + + /// + /// The API returns type information for data contained in a tensor. For sparse + /// tensors it returns type information for contained non-zero values. + /// It returns dense shape for sparse tensors. + /// + /// TensorTypeAndShapeInfo + TensorTypeAndShapeInfo GetTensorTypeAndShapeInfo() const; + + /// + /// This API returns information about the memory allocation used to hold data. + /// + /// Non owning instance of MemoryInfo + ConstMemoryInfo GetTensorMemoryInfo() const; + + /// + /// The API copies UTF-8 encoded bytes for the requested string element + /// contained within a tensor or a sparse tensor into a provided buffer. + /// Use GetStringTensorElementLength() to obtain the length of the buffer to allocate. + /// + /// + /// + /// + void GetStringTensorElement(size_t buffer_length, size_t element_index, void* buffer) const; + + /// + /// Returns string tensor UTF-8 encoded string element. + /// Use of this API is recommended over GetStringTensorElement() that takes void* buffer pointer. + /// + /// + /// std::string + std::string GetStringTensorElement(size_t element_index) const; + + /// + /// The API returns a byte length of UTF-8 encoded string element + /// contained in either a tensor or a spare tensor values. + /// + /// + /// byte length for the specified string element + size_t GetStringTensorElementLength(size_t element_index) const; + +#if !defined(DISABLE_SPARSE_TENSORS) + /// + /// The API returns the sparse data format this OrtValue holds in a sparse tensor. + /// If the sparse tensor was not fully constructed, i.e. Use*() or Fill*() API were not used + /// the value returned is ORT_SPARSE_UNDEFINED. + /// + /// Format enum + OrtSparseFormat GetSparseFormat() const; + + /// + /// The API returns type and shape information for stored non-zero values of the + /// sparse tensor. Use GetSparseTensorValues() to obtain values buffer pointer. + /// + /// TensorTypeAndShapeInfo values information + TensorTypeAndShapeInfo GetSparseTensorValuesTypeAndShapeInfo() const; + + /// + /// The API returns type and shape information for the specified indices. Each supported + /// indices have their own enum values even if a give format has more than one kind of indices. + /// Use GetSparseTensorIndicesData() to obtain pointer to indices buffer. + /// + /// enum requested + /// type and shape information + TensorTypeAndShapeInfo GetSparseTensorIndicesTypeShapeInfo(OrtSparseIndicesFormat format) const; + + /// + /// The API retrieves a pointer to the internal indices buffer. The API merely performs + /// a convenience data type casting on the return type pointer. Make sure you are requesting + /// the right type, use GetSparseTensorIndicesTypeShapeInfo(); + /// + /// type to cast to + /// requested indices kind + /// number of indices entries + /// Pinter to the internal sparse tensor buffer containing indices. Do not free this pointer. + template + const R* GetSparseTensorIndicesData(OrtSparseIndicesFormat indices_format, size_t& num_indices) const; + + /// + /// Returns true if the OrtValue contains a sparse tensor + /// + /// + bool IsSparseTensor() const; + + /// + /// The API returns a pointer to an internal buffer of the sparse tensor + /// containing non-zero values. The API merely does casting. Make sure you + /// are requesting the right data type by calling GetSparseTensorValuesTypeAndShapeInfo() + /// first. + /// + /// numeric data types only. Use GetStringTensor*() to retrieve strings. + /// a pointer to the internal values buffer. Do not free this pointer. + template + const R* GetSparseTensorValues() const; + +#endif +}; + +template +struct ValueImpl : ConstValueImpl { + using B = ConstValueImpl; + using B::B; + + /// + /// Returns a non-const typed pointer to an OrtValue/Tensor contained buffer + /// No type checking is performed, the caller must ensure the type matches the tensor type. + /// + /// non-const pointer to data, no copies made + template + R* GetTensorMutableData(); + + /// + /// Returns a non-typed non-const pointer to a tensor contained data. + /// + /// pointer to data, no copies made + void* GetTensorMutableRawData(); + + /// + // Obtain a reference to an element of data at the location specified + /// by the vector of dims. + /// + /// + /// [in] expressed by a vecotr of dimensions offsets + /// + template + R& At(const std::vector& location); + + /// + /// Set all strings at once in a string tensor + /// + /// [in] An array of strings. Each string in this array must be null terminated. + /// [in] Count of strings in s (Must match the size of \p value's tensor shape) + void FillStringTensor(const char* const* s, size_t s_len); + + /// + /// Set a single string in a string tensor + /// + /// [in] A null terminated UTF-8 encoded string + /// [in] Index of the string in the tensor to set + void FillStringTensorElement(const char* s, size_t index); + + /// + /// Allocate if necessary and obtain a pointer to a UTF-8 + /// encoded string element buffer indexed by the flat element index, + /// of the specified length. + /// + /// This API is for advanced usage. It avoids a need to construct + /// an auxiliary array of string pointers, and allows to write data directly + /// (do not zero terminate). + /// + /// + /// + /// a pointer to a writable buffer + char* GetResizedStringTensorElementBuffer(size_t index, size_t buffer_length); + +#if !defined(DISABLE_SPARSE_TENSORS) + /// + /// Supplies COO format specific indices and marks the contained sparse tensor as being a COO format tensor. + /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user + /// allocated buffers lifespan must eclipse that of the OrtValue. + /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time. + /// + /// pointer to the user allocated buffer with indices. Use nullptr for fully sparse tensors. + /// number of indices entries. Use 0 for fully sparse tensors + void UseCooIndices(int64_t* indices_data, size_t indices_num); + + /// + /// Supplies CSR format specific indices and marks the contained sparse tensor as being a CSR format tensor. + /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user + /// allocated buffers lifespan must eclipse that of the OrtValue. + /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time. + /// + /// pointer to the user allocated buffer with inner indices or nullptr for fully sparse tensors + /// number of csr inner indices or 0 for fully sparse tensors + /// pointer to the user allocated buffer with outer indices or nullptr for fully sparse tensors + /// number of csr outer indices or 0 for fully sparse tensors + void UseCsrIndices(int64_t* inner_data, size_t inner_num, int64_t* outer_data, size_t outer_num); + + /// + /// Supplies BlockSparse format specific indices and marks the contained sparse tensor as being a BlockSparse format tensor. + /// Values are supplied with a CreateSparseTensor() API. The supplied indices are not copied and the user + /// allocated buffers lifespan must eclipse that of the OrtValue. + /// The location of the indices is assumed to be the same as specified by OrtMemoryInfo argument at the creation time. + /// + /// indices shape or a {0} for fully sparse + /// user allocated buffer with indices or nullptr for fully spare tensors + void UseBlockSparseIndices(const Shape& indices_shape, int32_t* indices_data); + + /// + /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API + /// and copy the values and COO indices into it. If data_mem_info specifies that the data is located + /// at difference device than the allocator, a X-device copy will be performed if possible. + /// + /// specified buffer memory description + /// values buffer information. + /// coo indices buffer or nullptr for fully sparse data + /// number of COO indices or 0 for fully sparse data + void FillSparseTensorCoo(const OrtMemoryInfo* data_mem_info, const OrtSparseValuesParam& values_param, + const int64_t* indices_data, size_t indices_num); + + /// + /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API + /// and copy the values and CSR indices into it. If data_mem_info specifies that the data is located + /// at difference device than the allocator, a X-device copy will be performed if possible. + /// + /// specified buffer memory description + /// values buffer information + /// csr inner indices pointer or nullptr for fully sparse tensors + /// number of csr inner indices or 0 for fully sparse tensors + /// pointer to csr indices data or nullptr for fully sparse tensors + /// number of csr outer indices or 0 + void FillSparseTensorCsr(const OrtMemoryInfo* data_mem_info, + const OrtSparseValuesParam& values, + const int64_t* inner_indices_data, size_t inner_indices_num, + const int64_t* outer_indices_data, size_t outer_indices_num); + + /// + /// The API will allocate memory using the allocator instance supplied to the CreateSparseTensor() API + /// and copy the values and BlockSparse indices into it. If data_mem_info specifies that the data is located + /// at difference device than the allocator, a X-device copy will be performed if possible. + /// + /// specified buffer memory description + /// values buffer information + /// indices shape. use {0} for fully sparse tensors + /// pointer to indices data or nullptr for fully sparse tensors + void FillSparseTensorBlockSparse(const OrtMemoryInfo* data_mem_info, + const OrtSparseValuesParam& values, + const Shape& indices_shape, + const int32_t* indices_data); + +#endif +}; + +} // namespace detail + +using ConstValue = detail::ConstValueImpl>; +using UnownedValue = detail::ValueImpl>; + +/** \brief Wrapper around ::OrtValue + * + */ +struct Value : detail::ValueImpl { + using Base = detail::ValueImpl; + using Base::Base; + using OrtSparseValuesParam = detail::OrtSparseValuesParam; + using Shape = detail::Shape; + + explicit Value(std::nullptr_t) {} ///< Create an empty Value object, must be assigned a valid one to be used + Value(Value&&) = default; + Value& operator=(Value&&) = default; + + ConstValue GetConst() const { return ConstValue{this->p_}; } + UnownedValue GetUnowned() const { return UnownedValue{this->p_}; } + + /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAsOrtValue. + * \tparam T The numeric datatype. This API is not suitable for strings. + * \param info Memory description of where the p_data buffer resides (CPU vs GPU etc). + * \param p_data Pointer to the data buffer. + * \param p_data_element_count The number of elements in the data buffer. + * \param shape Pointer to the tensor shape dimensions. + * \param shape_len The number of tensor shape dimensions. + */ + template + static Value CreateTensor(const OrtMemoryInfo* info, T* p_data, size_t p_data_element_count, + const int64_t* shape, size_t shape_len); + + /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAsOrtValue. + * + * \param info Memory description of where the p_data buffer resides (CPU vs GPU etc). + * \param p_data Pointer to the data buffer. + * \param p_data_byte_count The number of bytes in the data buffer. + * \param shape Pointer to the tensor shape dimensions. + * \param shape_len The number of tensor shape dimensions. + * \param type The data type. + */ + static Value CreateTensor(const OrtMemoryInfo* info, void* p_data, size_t p_data_byte_count, + const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type); + + /** \brief Creates a tensor with a user supplied buffer. Wraps OrtApi::CreateTensorWithDataAndDeleterAsOrtValue. + * + * \param deleter OrtAllocator that will be used to free the buffer when no longer required. + * \param p_data Pointer to the data buffer. + * \param p_data_byte_count The number of bytes in the data buffer. + * \param shape Pointer to the tensor shape dimensions. + * \param shape_len The number of tensor shape dimensions. + * \param type The data type. + */ + static Value CreateTensor(OrtAllocator* deleter, void* p_data, size_t p_data_byte_count, + const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type); + + /** \brief Creates an OrtValue with a tensor using a supplied OrtAllocator. Wraps OrtApi::CreateTensorAsOrtValue. + * This overload will allocate the buffer for the tensor according to the supplied shape and data type. + * The allocated buffer will be owned by the returned OrtValue and will be freed when the OrtValue is released. + * The input data would need to be copied into the allocated buffer. + * This API is not suitable for strings. + * + * \tparam T The numeric datatype. This API is not suitable for strings. + * \param allocator The allocator to use. + * \param shape Pointer to the tensor shape dimensions. + * \param shape_len The number of tensor shape dimensions. + */ + template + static Value CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len); + + /** \brief Creates an OrtValue with a tensor using the supplied OrtAllocator. + * Wraps OrtApi::CreateTensorAsOrtValue. + * The allocated buffer will be owned by the returned OrtValue and will be freed when the OrtValue is released. + * The input data would need to be copied into the allocated buffer. + * This API is not suitable for strings. + * + * \param allocator The allocator to use. + * \param shape Pointer to the tensor shape dimensions. + * \param shape_len The number of tensor shape dimensions. + * \param type The data type. + */ + static Value CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type); + + /** \brief Creates an OrtValue with a Map Onnx type representation. + * The API would ref-count the supplied OrtValues and they will be released + * when the returned OrtValue is released. The caller may release keys and values after the call + * returns. + * + * \param keys an OrtValue containing a tensor with primitive data type keys. + * \param values an OrtValue that may contain a tensor. Ort currently supports only primitive data type values. + */ + static Value CreateMap(const Value& keys, const Value& values); ///< Wraps OrtApi::CreateValue + + /** \brief Creates an OrtValue with a Sequence Onnx type representation. + * The API would ref-count the supplied OrtValues and they will be released + * when the returned OrtValue is released. The caller may release the values after the call + * returns. + * + * \param values a vector of OrtValues that must have the same Onnx value type. + */ + static Value CreateSequence(const std::vector& values); ///< Wraps OrtApi::CreateValue + + /** \brief Creates an OrtValue wrapping an Opaque type. + * This is used for experimental support of non-tensor types. + * + * \tparam T - the type of the value. + * \param domain - zero terminated utf-8 string. Domain of the type. + * \param type_name - zero terminated utf-8 string. Name of the type. + * \param value - the value to be wrapped. + */ + template + static Value CreateOpaque(const char* domain, const char* type_name, const T& value); ///< Wraps OrtApi::CreateOpaqueValue + +#if !defined(DISABLE_SPARSE_TENSORS) + /// + /// This is a simple forwarding method to the other overload that helps deducing + /// data type enum value from the type of the buffer. + /// + /// numeric datatype. This API is not suitable for strings. + /// Memory description where the user buffers reside (CPU vs GPU etc) + /// pointer to the user supplied buffer, use nullptr for fully sparse tensors + /// a would be dense shape of the tensor + /// non zero values shape. Use a single 0 shape for fully sparse tensors. + /// + template + static Value CreateSparseTensor(const OrtMemoryInfo* info, T* p_data, const Shape& dense_shape, + const Shape& values_shape); + + /// + /// Creates an OrtValue instance containing SparseTensor. This constructs + /// a sparse tensor that makes use of user allocated buffers. It does not make copies + /// of the user provided data and does not modify it. The lifespan of user provided buffers should + /// eclipse the life span of the resulting OrtValue. This call constructs an instance that only contain + /// a pointer to non-zero values. To fully populate the sparse tensor call UseIndices() API below + /// to supply a sparse format specific indices. + /// This API is not suitable for string data. Use CreateSparseTensor() with allocator specified so strings + /// can be properly copied into the allocated buffer. + /// + /// Memory description where the user buffers reside (CPU vs GPU etc) + /// pointer to the user supplied buffer, use nullptr for fully sparse tensors + /// a would be dense shape of the tensor + /// non zero values shape. Use a single 0 shape for fully sparse tensors. + /// data type + /// Ort::Value instance containing SparseTensor + static Value CreateSparseTensor(const OrtMemoryInfo* info, void* p_data, const Shape& dense_shape, + const Shape& values_shape, ONNXTensorElementDataType type); + + /// + /// This is a simple forwarding method to the below CreateSparseTensor. + /// This helps to specify data type enum in terms of C++ data type. + /// Use CreateSparseTensor + /// + /// numeric data type only. String data enum must be specified explicitly. + /// allocator to use + /// a would be dense shape of the tensor + /// Ort::Value + template + static Value CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape); + + /// + /// Creates an instance of OrtValue containing sparse tensor. The created instance has no data. + /// The data must be supplied by on of the FillSparseTensor() methods that take both non-zero values + /// and indices. The data will be copied into a buffer that would be allocated using the supplied allocator. + /// Use this API to create OrtValues that contain sparse tensors with all supported data types including + /// strings. + /// + /// allocator to use. The allocator lifespan must eclipse that of the resulting OrtValue + /// a would be dense shape of the tensor + /// data type + /// an instance of Ort::Value + static Value CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape, ONNXTensorElementDataType type); + +#endif // !defined(DISABLE_SPARSE_TENSORS) +}; + +/// +/// Represents native memory allocation coming from one of the +/// OrtAllocators registered with OnnxRuntime. +/// Use it to wrap an allocation made by an allocator +/// so it can be automatically released when no longer needed. +/// +struct MemoryAllocation { + MemoryAllocation(OrtAllocator* allocator, void* p, size_t size); + ~MemoryAllocation(); + MemoryAllocation(const MemoryAllocation&) = delete; + MemoryAllocation& operator=(const MemoryAllocation&) = delete; + MemoryAllocation(MemoryAllocation&&) noexcept; + MemoryAllocation& operator=(MemoryAllocation&&) noexcept; + + void* get() { return p_; } + size_t size() const { return size_; } + + private: + OrtAllocator* allocator_; + void* p_; + size_t size_; +}; + +namespace detail { +template +struct AllocatorImpl : Base { + using B = Base; + using B::B; + + void* Alloc(size_t size); + MemoryAllocation GetAllocation(size_t size); + void Free(void* p); + ConstMemoryInfo GetInfo() const; +}; + +} // namespace detail + +/** \brief Wrapper around ::OrtAllocator default instance that is owned by Onnxruntime + * + */ +struct AllocatorWithDefaultOptions : detail::AllocatorImpl> { + explicit AllocatorWithDefaultOptions(std::nullptr_t) {} ///< Convenience to create a class member and then replace with an instance + AllocatorWithDefaultOptions(); +}; + +/** \brief Wrapper around ::OrtAllocator + * + */ +struct Allocator : detail::AllocatorImpl { + explicit Allocator(std::nullptr_t) {} ///< Convenience to create a class member and then replace with an instance + Allocator(const Session& session, const OrtMemoryInfo*); +}; + +using UnownedAllocator = detail::AllocatorImpl>; + +namespace detail { +namespace binding_utils { +// Bring these out of template +std::vector GetOutputNamesHelper(const OrtIoBinding* binding, OrtAllocator*); +std::vector GetOutputValuesHelper(const OrtIoBinding* binding, OrtAllocator*); +} // namespace binding_utils + +template +struct ConstIoBindingImpl : Base { + using B = Base; + using B::B; + + std::vector GetOutputNames() const; + std::vector GetOutputNames(OrtAllocator*) const; + std::vector GetOutputValues() const; + std::vector GetOutputValues(OrtAllocator*) const; +}; + +template +struct IoBindingImpl : ConstIoBindingImpl { + using B = ConstIoBindingImpl; + using B::B; + + void BindInput(const char* name, const Value&); + void BindOutput(const char* name, const Value&); + void BindOutput(const char* name, const OrtMemoryInfo*); + void ClearBoundInputs(); + void ClearBoundOutputs(); + void SynchronizeInputs(); + void SynchronizeOutputs(); +}; + +} // namespace detail + +using ConstIoBinding = detail::ConstIoBindingImpl>; +using UnownedIoBinding = detail::IoBindingImpl>; + +/** \brief Wrapper around ::OrtIoBinding + * + */ +struct IoBinding : detail::IoBindingImpl { + explicit IoBinding(std::nullptr_t) {} ///< Create an empty object for convenience. Sometimes, we want to initialize members later. + explicit IoBinding(Session& session); + ConstIoBinding GetConst() const { return ConstIoBinding{this->p_}; } + UnownedIoBinding GetUnowned() const { return UnownedIoBinding{this->p_}; } +}; + +/*! \struct Ort::ArenaCfg + * \brief it is a structure that represents the configuration of an arena based allocator + * \details Please see docs/C_API.md for details + */ +struct ArenaCfg : detail::Base { + explicit ArenaCfg(std::nullptr_t) {} ///< Create an empty ArenaCfg object, must be assigned a valid one to be used + /** + * Wraps OrtApi::CreateArenaCfg + * \param max_mem - use 0 to allow ORT to choose the default + * \param arena_extend_strategy - use -1 to allow ORT to choose the default, 0 = kNextPowerOfTwo, 1 = kSameAsRequested + * \param initial_chunk_size_bytes - use -1 to allow ORT to choose the default + * \param max_dead_bytes_per_chunk - use -1 to allow ORT to choose the default + * See docs/C_API.md for details on what the following parameters mean and how to choose these values + */ + ArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, int max_dead_bytes_per_chunk); +}; + +// +// Custom OPs (only needed to implement custom OPs) +// + +/// +/// This struct provides life time management for custom op attribute +/// +struct OpAttr : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit OpAttr(std::nullptr_t) {} + OpAttr(const char* name, const void* data, int len, OrtOpAttrType type); +}; + +/** + * Macro that logs a message using the provided logger. Throws an exception if OrtApi::Logger_LogMessage fails. + * Example: ORT_CXX_LOG(logger, ORT_LOGGING_LEVEL_INFO, "Log a message"); + * + * \param logger The Ort::Logger instance to use. Must be a value or reference. + * \param message_severity The logging severity level of the message. + * \param message A null-terminated UTF-8 message to log. + */ +#define ORT_CXX_LOG(logger, message_severity, message) \ + do { \ + if (message_severity >= logger.GetLoggingSeverityLevel()) { \ + Ort::ThrowOnError(logger.LogMessage(message_severity, ORT_FILE, __LINE__, \ + static_cast(__FUNCTION__), message)); \ + } \ + } while (false) + +/** + * Macro that logs a message using the provided logger. Can be used in noexcept code since errors are silently ignored. + * Example: ORT_CXX_LOG_NOEXCEPT(logger, ORT_LOGGING_LEVEL_INFO, "Log a message"); + * + * \param logger The Ort::Logger instance to use. Must be a value or reference. + * \param message_severity The logging severity level of the message. + * \param message A null-terminated UTF-8 message to log. + */ +#define ORT_CXX_LOG_NOEXCEPT(logger, message_severity, message) \ + do { \ + if (message_severity >= logger.GetLoggingSeverityLevel()) { \ + static_cast(logger.LogMessage(message_severity, ORT_FILE, __LINE__, \ + static_cast(__FUNCTION__), message)); \ + } \ + } while (false) + +/** + * Macro that logs a printf-like formatted message using the provided logger. Throws an exception if + * OrtApi::Logger_LogMessage fails or if a formatting error occurs. + * Example: ORT_CXX_LOGF(logger, ORT_LOGGING_LEVEL_INFO, "Log an int: %d", 12); + * + * \param logger The Ort::Logger instance to use. Must be a value or reference. + * \param message_severity The logging severity level of the message. + * \param format A null-terminated UTF-8 format string forwarded to a printf-like function. + * Refer to https://en.cppreference.com/w/cpp/io/c/fprintf for information on valid formats. + * \param ... Zero or more variadic arguments referenced by the format string. + */ +#define ORT_CXX_LOGF(logger, message_severity, /*format,*/...) \ + do { \ + if (message_severity >= logger.GetLoggingSeverityLevel()) { \ + Ort::ThrowOnError(logger.LogFormattedMessage(message_severity, ORT_FILE, __LINE__, \ + static_cast(__FUNCTION__), __VA_ARGS__)); \ + } \ + } while (false) + +/** + * Macro that logs a printf-like formatted message using the provided logger. Can be used in noexcept code since errors + * are silently ignored. + * Example: ORT_CXX_LOGF_NOEXCEPT(logger, ORT_LOGGING_LEVEL_INFO, "Log an int: %d", 12); + * + * \param logger The Ort::Logger instance to use. Must be a value or reference. + * \param message_severity The logging severity level of the message. + * \param format A null-terminated UTF-8 format string forwarded to a printf-like function. + * Refer to https://en.cppreference.com/w/cpp/io/c/fprintf for information on valid formats. + * \param ... Zero or more variadic arguments referenced by the format string. + */ +#define ORT_CXX_LOGF_NOEXCEPT(logger, message_severity, /*format,*/...) \ + do { \ + if (message_severity >= logger.GetLoggingSeverityLevel()) { \ + static_cast(logger.LogFormattedMessage(message_severity, ORT_FILE, __LINE__, \ + static_cast(__FUNCTION__), __VA_ARGS__)); \ + } \ + } while (false) + +/// +/// This class represents an ONNX Runtime logger that can be used to log information with an +/// associated severity level and source code location (file path, line number, function name). +/// +/// A Logger can be obtained from within custom operators by calling Ort::KernelInfo::GetLogger(). +/// Instances of Ort::Logger are the size of two pointers and can be passed by value. +/// +/// Use the ORT_CXX_LOG macros to ensure the source code location is set properly from the callsite +/// and to take advantage of a cached logging severity level that can bypass calls to the underlying C API. +/// +struct Logger { + /** + * Creates an empty Ort::Logger. Must be initialized from a valid Ort::Logger before use. + */ + Logger() = default; + + /** + * Creates an empty Ort::Logger. Must be initialized from a valid Ort::Logger before use. + */ + explicit Logger(std::nullptr_t) {} + + /** + * Creates a logger from an ::OrtLogger instance. Caches the logger's current severity level by calling + * OrtApi::Logger_GetLoggingSeverityLevel. Throws an exception if OrtApi::Logger_GetLoggingSeverityLevel fails. + * + * \param logger The ::OrtLogger to wrap. + */ + explicit Logger(const OrtLogger* logger); + + ~Logger() = default; + + Logger(const Logger&) = default; + Logger& operator=(const Logger&) = default; + + Logger(Logger&& v) noexcept = default; + Logger& operator=(Logger&& v) noexcept = default; + + /** + * Returns the logger's current severity level from the cached member. + * + * \return The current ::OrtLoggingLevel. + */ + OrtLoggingLevel GetLoggingSeverityLevel() const noexcept; + + /** + * Logs the provided message via OrtApi::Logger_LogMessage. Use the ORT_CXX_LOG or ORT_CXX_LOG_NOEXCEPT + * macros to properly set the source code location and to use the cached severity level to potentially bypass + * calls to the underlying C API. + * + * \param log_severity_level The message's logging severity level. + * \param file_path The filepath of the file in which the message is logged. Usually the value of ORT_FILE. + * \param line_number The file line number in which the message is logged. Usually the value of __LINE__. + * \param func_name The name of the function in which the message is logged. Usually the value of __FUNCTION__. + * \param message The message to log. + * \return A Ort::Status value to indicate error or success. + */ + Status LogMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, int line_number, + const char* func_name, const char* message) const noexcept; + + /** + * Logs a printf-like formatted message via OrtApi::Logger_LogMessage. Use the ORT_CXX_LOGF or ORT_CXX_LOGF_NOEXCEPT + * macros to properly set the source code location and to use the cached severity level to potentially bypass + * calls to the underlying C API. Returns an error status if a formatting error occurs. + * + * \param log_severity_level The message's logging severity level. + * \param file_path The filepath of the file in which the message is logged. Usually the value of ORT_FILE. + * \param line_number The file line number in which the message is logged. Usually the value of __LINE__. + * \param func_name The name of the function in which the message is logged. Usually the value of __FUNCTION__. + * \param format A null-terminated UTF-8 format string forwarded to a printf-like function. + * Refer to https://en.cppreference.com/w/cpp/io/c/fprintf for information on valid formats. + * \param args Zero or more variadic arguments referenced by the format string. + * \return A Ort::Status value to indicate error or success. + */ + template + Status LogFormattedMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, int line_number, + const char* func_name, const char* format, Args&&... args) const noexcept; + + private: + const OrtLogger* logger_{}; + OrtLoggingLevel cached_severity_level_{}; +}; + +/// +/// This class wraps a raw pointer OrtKernelContext* that is being passed +/// to the custom kernel Compute() method. Use it to safely access context +/// attributes, input and output parameters with exception safety guarantees. +/// See usage example in onnxruntime/test/testdata/custom_op_library/custom_op_library.cc +/// +struct KernelContext { + explicit KernelContext(OrtKernelContext* context); + size_t GetInputCount() const; + size_t GetOutputCount() const; + // If input is optional and is not present, the method returns an empty ConstValue + // which can be compared to nullptr. + ConstValue GetInput(size_t index) const; + // If output is optional and is not present, the method returns an empty UnownedValue + // which can be compared to nullptr. + UnownedValue GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const; + UnownedValue GetOutput(size_t index, const std::vector& dims) const; + void* GetGPUComputeStream() const; + Logger GetLogger() const; + OrtAllocator* GetAllocator(const OrtMemoryInfo& memory_info) const; + OrtKernelContext* GetOrtKernelContext() const { return ctx_; } + void ParallelFor(void (*fn)(void*, size_t), size_t total, size_t num_batch, void* usr_data) const; + + private: + OrtKernelContext* ctx_; +}; + +struct KernelInfo; + +namespace detail { +namespace attr_utils { +void GetAttr(const OrtKernelInfo* p, const char* name, float&); +void GetAttr(const OrtKernelInfo* p, const char* name, int64_t&); +void GetAttr(const OrtKernelInfo* p, const char* name, std::string&); +void GetAttrs(const OrtKernelInfo* p, const char* name, std::vector&); +void GetAttrs(const OrtKernelInfo* p, const char* name, std::vector&); +} // namespace attr_utils + +template +struct KernelInfoImpl : Base { + using B = Base; + using B::B; + + KernelInfo Copy() const; + + template // R is only implemented for float, int64_t, and string + R GetAttribute(const char* name) const { + R val; + attr_utils::GetAttr(this->p_, name, val); + return val; + } + + template // R is only implemented for std::vector, std::vector + std::vector GetAttributes(const char* name) const { + std::vector result; + attr_utils::GetAttrs(this->p_, name, result); + return result; + } + + Value GetTensorAttribute(const char* name, OrtAllocator* allocator) const; + + size_t GetInputCount() const; + size_t GetOutputCount() const; + + std::string GetInputName(size_t index) const; + std::string GetOutputName(size_t index) const; + + TypeInfo GetInputTypeInfo(size_t index) const; + TypeInfo GetOutputTypeInfo(size_t index) const; + + ConstValue GetTensorConstantInput(size_t index, int* is_constant) const; + + std::string GetNodeName() const; + Logger GetLogger() const; +}; + +} // namespace detail + +using ConstKernelInfo = detail::KernelInfoImpl>; + +/// +/// This struct owns the OrtKernInfo* pointer when a copy is made. +/// For convenient wrapping of OrtKernelInfo* passed to kernel constructor +/// and query attributes, warp the pointer with Ort::Unowned instance +/// so it does not destroy the pointer the kernel does not own. +/// +struct KernelInfo : detail::KernelInfoImpl { + using Base = detail::KernelInfoImpl; + using Base::Base; + explicit KernelInfo(std::nullptr_t) {} ///< Create an empty instance to initialize later + explicit KernelInfo(OrtKernelInfo* info); ///< Take ownership of the instance + ConstKernelInfo GetConst() const { return ConstKernelInfo{this->p_}; } +}; + +/// +/// Create and own custom defined operation. +/// +struct Op : detail::Base { + using Base = detail::Base; + using Base::Base; + + explicit Op(std::nullptr_t) {} ///< Create an empty Operator object, must be assigned a valid one to be used + + explicit Op(OrtOp*); ///< Take ownership of the OrtOp + + static Op Create(const OrtKernelInfo* info, const char* op_name, const char* domain, + int version, const char** type_constraint_names, + const ONNXTensorElementDataType* type_constraint_values, + size_t type_constraint_count, + const OpAttr* attr_values, + size_t attr_count, + size_t input_count, size_t output_count); + + void Invoke(const OrtKernelContext* context, + const Value* input_values, + size_t input_count, + Value* output_values, + size_t output_count); + + // For easier refactoring + void Invoke(const OrtKernelContext* context, + const OrtValue* const* input_values, + size_t input_count, + OrtValue* const* output_values, + size_t output_count); +}; + +/// +/// Provide access to per-node attributes and input shapes, so one could compute and set output shapes. +/// +struct ShapeInferContext { + struct SymbolicInteger { + SymbolicInteger(int64_t i) : i_(i), is_int_(true) {}; + SymbolicInteger(const char* s) : s_(s), is_int_(false) {}; + SymbolicInteger(const SymbolicInteger&) = default; + SymbolicInteger(SymbolicInteger&&) = default; + + SymbolicInteger& operator=(const SymbolicInteger&) = default; + SymbolicInteger& operator=(SymbolicInteger&&) = default; + + bool operator==(const SymbolicInteger& dim) const { + if (is_int_ == dim.is_int_) { + if (is_int_) { + return i_ == dim.i_; + } else { + return std::string{s_} == std::string{dim.s_}; + } + } + return false; + } + + bool IsInt() const { return is_int_; } + int64_t AsInt() const { return i_; } + const char* AsSym() const { return s_; } + + static constexpr int INVALID_INT_DIM = -2; + + private: + union { + int64_t i_; + const char* s_; + }; + bool is_int_; + }; + + using Shape = std::vector; + + ShapeInferContext(const OrtApi* ort_api, OrtShapeInferContext* ctx); + + const Shape& GetInputShape(size_t indice) const { return input_shapes_.at(indice); } + + size_t GetInputCount() const { return input_shapes_.size(); } + + Status SetOutputShape(size_t indice, const Shape& shape, ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT); + + int64_t GetAttrInt(const char* attr_name); + + using Ints = std::vector; + Ints GetAttrInts(const char* attr_name); + + float GetAttrFloat(const char* attr_name); + + using Floats = std::vector; + Floats GetAttrFloats(const char* attr_name); + + std::string GetAttrString(const char* attr_name); + + using Strings = std::vector; + Strings GetAttrStrings(const char* attr_name); + + private: + const OrtOpAttr* GetAttrHdl(const char* attr_name) const; + const OrtApi* ort_api_; + OrtShapeInferContext* ctx_; + std::vector input_shapes_; +}; + +using ShapeInferFn = Ort::Status (*)(Ort::ShapeInferContext&); + +#define MAX_CUSTOM_OP_END_VER (1UL << 31) - 1 + +template +struct CustomOpBase : OrtCustomOp { + CustomOpBase() { + OrtCustomOp::version = ORT_API_VERSION; + OrtCustomOp::GetName = [](const OrtCustomOp* this_) { return static_cast(this_)->GetName(); }; + + OrtCustomOp::GetExecutionProviderType = [](const OrtCustomOp* this_) { return static_cast(this_)->GetExecutionProviderType(); }; + + OrtCustomOp::GetInputTypeCount = [](const OrtCustomOp* this_) { return static_cast(this_)->GetInputTypeCount(); }; + OrtCustomOp::GetInputType = [](const OrtCustomOp* this_, size_t index) { return static_cast(this_)->GetInputType(index); }; + OrtCustomOp::GetInputMemoryType = [](const OrtCustomOp* this_, size_t index) { return static_cast(this_)->GetInputMemoryType(index); }; + + OrtCustomOp::GetOutputTypeCount = [](const OrtCustomOp* this_) { return static_cast(this_)->GetOutputTypeCount(); }; + OrtCustomOp::GetOutputType = [](const OrtCustomOp* this_, size_t index) { return static_cast(this_)->GetOutputType(index); }; + +#if defined(_MSC_VER) && !defined(__clang__) +#pragma warning(push) +#pragma warning(disable : 26409) +#endif + OrtCustomOp::KernelDestroy = [](void* op_kernel) { delete static_cast(op_kernel); }; +#if defined(_MSC_VER) && !defined(__clang__) +#pragma warning(pop) +#endif + OrtCustomOp::GetInputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast(this_)->GetInputCharacteristic(index); }; + OrtCustomOp::GetOutputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast(this_)->GetOutputCharacteristic(index); }; + + OrtCustomOp::GetVariadicInputMinArity = [](const OrtCustomOp* this_) { return static_cast(this_)->GetVariadicInputMinArity(); }; + OrtCustomOp::GetVariadicInputHomogeneity = [](const OrtCustomOp* this_) { return static_cast(static_cast(this_)->GetVariadicInputHomogeneity()); }; + OrtCustomOp::GetVariadicOutputMinArity = [](const OrtCustomOp* this_) { return static_cast(this_)->GetVariadicOutputMinArity(); }; + OrtCustomOp::GetVariadicOutputHomogeneity = [](const OrtCustomOp* this_) { return static_cast(static_cast(this_)->GetVariadicOutputHomogeneity()); }; +#ifdef __cpp_if_constexpr + if constexpr (WithStatus) { +#else + if (WithStatus) { +#endif + OrtCustomOp::CreateKernelV2 = [](const OrtCustomOp* this_, const OrtApi* api, const OrtKernelInfo* info, void** op_kernel) -> OrtStatusPtr { + return static_cast(this_)->CreateKernelV2(*api, info, op_kernel); + }; + OrtCustomOp::KernelComputeV2 = [](void* op_kernel, OrtKernelContext* context) -> OrtStatusPtr { + return static_cast(op_kernel)->ComputeV2(context); + }; + } else { + OrtCustomOp::CreateKernelV2 = nullptr; + OrtCustomOp::KernelComputeV2 = nullptr; + + OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* api, const OrtKernelInfo* info) { return static_cast(this_)->CreateKernel(*api, info); }; + OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) { + static_cast(op_kernel)->Compute(context); + }; + } + + SetShapeInferFn(0); + + OrtCustomOp::GetStartVersion = [](const OrtCustomOp* this_) { + return static_cast(this_)->start_ver_; + }; + + OrtCustomOp::GetEndVersion = [](const OrtCustomOp* this_) { + return static_cast(this_)->end_ver_; + }; + + OrtCustomOp::GetMayInplace = nullptr; + OrtCustomOp::ReleaseMayInplace = nullptr; + OrtCustomOp::GetAliasMap = nullptr; + OrtCustomOp::ReleaseAliasMap = nullptr; + } + + // Default implementation of GetExecutionProviderType that returns nullptr to default to the CPU provider + const char* GetExecutionProviderType() const { return nullptr; } + + // Default implementations of GetInputCharacteristic() and GetOutputCharacteristic() below + // (inputs and outputs are required by default) + OrtCustomOpInputOutputCharacteristic GetInputCharacteristic(size_t /*index*/) const { + return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED; + } + + OrtCustomOpInputOutputCharacteristic GetOutputCharacteristic(size_t /*index*/) const { + return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED; + } + + // Default implemention of GetInputMemoryType() that returns OrtMemTypeDefault + OrtMemType GetInputMemoryType(size_t /*index*/) const { + return OrtMemTypeDefault; + } + + // Default implementation of GetVariadicInputMinArity() returns 1 to specify that a variadic input + // should expect at least 1 argument. + int GetVariadicInputMinArity() const { + return 1; + } + + // Default implementation of GetVariadicInputHomegeneity() returns true to specify that all arguments + // to a variadic input should be of the same type. + bool GetVariadicInputHomogeneity() const { + return true; + } + + // Default implementation of GetVariadicOutputMinArity() returns 1 to specify that a variadic output + // should produce at least 1 output value. + int GetVariadicOutputMinArity() const { + return 1; + } + + // Default implementation of GetVariadicOutputHomegeneity() returns true to specify that all output values + // produced by a variadic output should be of the same type. + bool GetVariadicOutputHomogeneity() const { + return true; + } + + // Declare list of session config entries used by this Custom Op. + // Implement this function in order to get configs from CustomOpBase::GetSessionConfigs(). + // This default implementation returns an empty vector of config entries. + std::vector GetSessionConfigKeys() const { + return std::vector{}; + } + + // Ort::CustomOpBase derived class should provide the following static method with the type/shape inferencing + // implementation if needed: + // static OrtStatusPtr InferOutputShape(Ort::ShapeInferContext& context) + template + decltype(&C::InferOutputShape) SetShapeInferFn(decltype(&C::InferOutputShape)) { + OrtCustomOp::InferOutputShapeFn = [](const OrtCustomOp*, OrtShapeInferContext* ort_ctx) -> OrtStatusPtr { + ShapeInferContext ctx(&GetApi(), ort_ctx); + return C::InferOutputShape(ctx); + }; + return {}; + } + + template + void SetShapeInferFn(...) { + OrtCustomOp::InferOutputShapeFn = {}; + } + + protected: + // Helper function that returns a map of session config entries specified by CustomOpBase::GetSessionConfigKeys. + void GetSessionConfigs(std::unordered_map& out, ConstSessionOptions options) const; + + int start_ver_ = 1; + int end_ver_ = MAX_CUSTOM_OP_END_VER; +}; + +namespace detail { +template +struct ValueInfoImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; + + std::string Name() const; + ConstTypeInfo TypeInfo() const; +}; +} // namespace detail + +// Const object holder that does not own the underlying object +using ConstValueInfo = detail::ValueInfoImpl>; + +/** \brief Wrapper around ::OrtValueInfo + * + */ +struct ValueInfo : detail::ValueInfoImpl { + explicit ValueInfo(std::nullptr_t) {} ///< No instance is created + /// Take ownership of a pointer created by C API + explicit ValueInfo(OrtValueInfo* p) : ValueInfoImpl{p} {} + + // Create ValueInfo for a tensor + explicit ValueInfo(const std::string& name, const ConstTypeInfo& type_info); + + ConstValueInfo GetConst() const { return ConstValueInfo{this->p_}; } +}; + +namespace detail { +template +struct NodeImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; +}; +} // namespace detail + +/** \brief Wrapper around ::OrtNode + * + */ +struct Node : detail::NodeImpl { + explicit Node(std::nullptr_t) {} ///< No instance is created + explicit Node(OrtNode* p) : NodeImpl{p} {} ///< Take ownership of a pointer created by C API + +#if !defined(ORT_MINIMAL_BUILD) + Node(const std::string& operator_name, const std::string& operator_domain, + const std::string& node_name, + const std::vector& input_names, + const std::vector& output_names); + + /// + /// Wraps CreateNode. Node takes ownership of attributes on success and updates the OpAttr in `attributes` to do so. + /// + Node(const std::string& operator_name, const std::string& operator_domain, + const std::string& node_name, + const std::vector& input_names, + const std::vector& output_names, + std::vector& attributes); + + private: + static void Init(const std::string& operator_name, const std::string& operator_domain, + const std::string& node_name, + const std::vector& input_names, + const std::vector& output_names, + std::vector& attributes, + OrtNode*& node); +#endif // !defined(ORT_MINIMAL_BUILD) +}; + +namespace detail { +template +struct GraphImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; + +#if !defined(ORT_MINIMAL_BUILD) + void SetInputs(std::vector& inputs); + void SetOutputs(std::vector& outputs); + void AddInitializer(const std::string& name, Value& initializer, bool data_is_external); // Graph takes ownership of Value + void AddNode(Node& node); // Graph takes ownership of Node +#endif // !defined(ORT_MINIMAL_BUILD) +}; +} // namespace detail + +/** \brief Wrapper around ::OrtGraph + * + */ +struct Graph : detail::GraphImpl { + explicit Graph(std::nullptr_t) {} ///< No instance is created + explicit Graph(OrtGraph* p) : GraphImpl{p} {} ///< Take ownership of a pointer created by C API +#if !defined(ORT_MINIMAL_BUILD) + Graph(); +#endif +}; + +namespace detail { +template +struct ModelImpl : Ort::detail::Base { + using B = Ort::detail::Base; + using B::B; + +#if !defined(ORT_MINIMAL_BUILD) + void AddGraph(Graph& graph); +#endif +}; +} // namespace detail + +// Const object holder that does not own the underlying object +using ConstModel = detail::ModelImpl>; + +/** \brief Wrapper around ::OrtModel + * + */ +struct Model : detail::ModelImpl { + using DomainOpsetPair = std::pair; + + explicit Model(std::nullptr_t) {} ///< No instance is created + explicit Model(OrtModel* p) : ModelImpl{p} {} ///< Take ownership of a pointer created by C API + +#if !defined(ORT_MINIMAL_BUILD) + explicit Model(const std::vector& opsets); +#endif + + ConstModel GetConst() const { return ConstModel{this->p_}; } +}; +} // namespace Ort +#include "onnxruntime_cxx_inline.h" diff --git a/third_party/onnxruntime/include/onnxruntime_cxx_inline.h b/third_party/onnxruntime/include/onnxruntime_cxx_inline.h new file mode 100644 index 00000000..94ad2118 --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_cxx_inline.h @@ -0,0 +1,2778 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +// Do not include this file directly. Please include "onnxruntime_cxx_api.h" instead. +// If interested in trying out features of the new experimental C++ API, include "experimental_onnxruntime_cxx_api.h" instead. +// +// These are the inline implementations of the C++ header APIs. They're in this separate file as to not clutter +// the main C++ file with implementation details. + +#include +#include +#include +#include +#include +#include + +// Convert OrtStatus to Ort::Status and return +// instead of throwing +#define ORT_CXX_RETURN_ON_API_FAIL(expression) \ + { \ + auto ort_status = (expression); \ + if (ort_status) { \ + return Ort::Status(ort_status); \ + } \ + } + +#ifdef __cpp_if_constexpr +#define ORT_CXX_IF_CONSTEXPR if constexpr +#else +#define ORT_CXX_IF_CONSTEXPR if +#endif + +namespace Ort { + +namespace detail { +inline void ThrowStatus(const Status& st) { + std::string error_message = st.GetErrorMessage(); + OrtErrorCode error_code = st.GetErrorCode(); + ORT_CXX_API_THROW(std::move(error_message), error_code); +} +} // namespace detail + +inline void ThrowOnError(OrtStatus* ort_status) { + if (ort_status) { + Ort::Status st(ort_status); + detail::ThrowStatus(st); + } +} + +inline void ThrowOnError(const Status& st) { + if (st) { + detail::ThrowStatus(st); + } +} + +inline Status::Status(OrtStatus* status) noexcept : detail::Base{status} { +} + +inline Status::Status(const std::exception& e) noexcept { + p_ = GetApi().CreateStatus(ORT_FAIL, e.what()); +} + +inline Status::Status(const Exception& e) noexcept { + p_ = GetApi().CreateStatus(e.GetOrtErrorCode(), e.what()); +} + +inline Status::Status(const char* message, OrtErrorCode code) noexcept { + p_ = GetApi().CreateStatus(code, message); +} + +inline std::string Status::GetErrorMessage() const { + std::string message(GetApi().GetErrorMessage(p_)); + return message; +} + +inline OrtErrorCode Status::GetErrorCode() const { + return GetApi().GetErrorCode(p_); +} + +inline bool Status::IsOK() const noexcept { + return (p_ == nullptr); +} + +// This template converts a C++ type into it's ONNXTensorElementDataType +template +struct TypeToTensorType; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL; +}; + +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FN; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FNUZ; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2; +}; +template <> +struct TypeToTensorType { + static constexpr ONNXTensorElementDataType type = ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ; +}; + +inline bool BFloat16_t::operator==(const BFloat16_t& rhs) const noexcept { + if (IsNaN() || rhs.IsNaN()) { + // IEEE defines that NaN is not equal to anything, including itself. + return false; + } + return val == rhs.val; +} + +inline bool BFloat16_t::operator<(const BFloat16_t& rhs) const noexcept { + if (IsNaN() || rhs.IsNaN()) { + // IEEE defines that NaN is unordered with respect to everything, including itself. + return false; + } + + const bool left_is_negative = IsNegative(); + if (left_is_negative != rhs.IsNegative()) { + // When the signs of left and right differ, we know that left is less than right if it is + // the negative value. The exception to this is if both values are zero, in which case IEEE + // says they should be equal, even if the signs differ. + return left_is_negative && !AreZero(*this, rhs); + } + return (val != rhs.val) && ((val < rhs.val) ^ left_is_negative); +} + +inline MemoryAllocation::MemoryAllocation(OrtAllocator* allocator, void* p, size_t size) + : allocator_(allocator), p_(p), size_(size) { +} + +inline MemoryAllocation::~MemoryAllocation() { + if (p_ != nullptr) { + // We do not throw out of destructor + auto ret = GetApi().AllocatorFree(allocator_, p_); + static_cast(ret); + } +} + +inline MemoryAllocation::MemoryAllocation(MemoryAllocation&& o) noexcept : allocator_(nullptr), p_(nullptr), size_(0) { + *this = std::move(o); +} + +inline MemoryAllocation& MemoryAllocation::operator=(MemoryAllocation&& o) noexcept { + OrtAllocator* alloc = nullptr; + void* p = nullptr; + size_t sz = 0; + + // Swap out this + std::swap(alloc, allocator_); + std::swap(p, p_); + std::swap(sz, size_); + + // Swap with incoming + std::swap(allocator_, o.allocator_); + std::swap(p_, o.p_); + std::swap(size_, o.size_); + + // Destroy this instance if needed + MemoryAllocation this_alloc(alloc, p, sz); + return *this; +} + +namespace detail { + +template +inline void* AllocatorImpl::Alloc(size_t size) { + void* out; + ThrowOnError(GetApi().AllocatorAlloc(this->p_, size, &out)); + return out; +} + +template +inline MemoryAllocation AllocatorImpl::GetAllocation(size_t size) { + void* out; + ThrowOnError(GetApi().AllocatorAlloc(this->p_, size, &out)); + MemoryAllocation result(this->p_, out, size); + return result; +} + +template +inline void AllocatorImpl::Free(void* p) { + ThrowOnError(GetApi().AllocatorFree(this->p_, p)); +} + +template +inline ConstMemoryInfo AllocatorImpl::GetInfo() const { + const OrtMemoryInfo* out; + ThrowOnError(GetApi().AllocatorGetInfo(this->p_, &out)); + return ConstMemoryInfo{out}; +} + +} // namespace detail + +inline AllocatorWithDefaultOptions::AllocatorWithDefaultOptions() { + ThrowOnError(GetApi().GetAllocatorWithDefaultOptions(&this->p_)); +} + +inline Allocator::Allocator(const Session& sess, const OrtMemoryInfo* mem_info) { + ThrowOnError(GetApi().CreateAllocator(sess, mem_info, &this->p_)); +} + +namespace detail { + +template +inline std::string MemoryInfoImpl::GetAllocatorName() const { + const char* name = nullptr; + ThrowOnError(GetApi().MemoryInfoGetName(this->p_, &name)); + return std::string(name); +} + +template +inline OrtAllocatorType MemoryInfoImpl::GetAllocatorType() const { + OrtAllocatorType type; + ThrowOnError(GetApi().MemoryInfoGetType(this->p_, &type)); + return type; +} + +template +inline int MemoryInfoImpl::GetDeviceId() const { + int id = 0; + ThrowOnError(GetApi().MemoryInfoGetId(this->p_, &id)); + return id; +} + +template +inline OrtMemoryInfoDeviceType MemoryInfoImpl::GetDeviceType() const { + OrtMemoryInfoDeviceType type; + GetApi().MemoryInfoGetDeviceType(this->p_, &type); + return type; +} + +template +inline OrtMemType MemoryInfoImpl::GetMemoryType() const { + OrtMemType type; + ThrowOnError(GetApi().MemoryInfoGetMemType(this->p_, &type)); + return type; +} + +template +template +inline bool MemoryInfoImpl::operator==(const MemoryInfoImpl& o) const { + int comp_result = 0; + ThrowOnError(Ort::GetApi().CompareMemoryInfo(this->p_, o, &comp_result)); + return comp_result == 0; +} + +} // namespace detail + +inline MemoryInfo MemoryInfo::CreateCpu(OrtAllocatorType type, OrtMemType mem_type) { + OrtMemoryInfo* p; + ThrowOnError(GetApi().CreateCpuMemoryInfo(type, mem_type, &p)); + return MemoryInfo(p); +} + +inline MemoryInfo::MemoryInfo(const char* name, OrtAllocatorType type, int id, OrtMemType mem_type) { + ThrowOnError(GetApi().CreateMemoryInfo(name, type, id, mem_type, &this->p_)); +} + +namespace detail { +template +inline std::vector ConstIoBindingImpl::GetOutputNames() const { + AllocatorWithDefaultOptions allocator; + return binding_utils::GetOutputNamesHelper(this->p_, allocator); +} + +template +inline std::vector ConstIoBindingImpl::GetOutputNames(OrtAllocator* allocator) const { + return binding_utils::GetOutputNamesHelper(this->p_, allocator); +} + +template +inline std::vector ConstIoBindingImpl::GetOutputValues() const { + AllocatorWithDefaultOptions allocator; + return binding_utils::GetOutputValuesHelper(this->p_, allocator); +} + +template +inline std::vector ConstIoBindingImpl::GetOutputValues(OrtAllocator* allocator) const { + return binding_utils::GetOutputValuesHelper(this->p_, allocator); +} + +template +inline void IoBindingImpl::BindInput(const char* name, const Value& value) { + ThrowOnError(GetApi().BindInput(this->p_, name, value)); +} + +template +inline void IoBindingImpl::BindOutput(const char* name, const Value& value) { + ThrowOnError(GetApi().BindOutput(this->p_, name, value)); +} + +template +inline void IoBindingImpl::BindOutput(const char* name, const OrtMemoryInfo* mem_info) { + ThrowOnError(GetApi().BindOutputToDevice(this->p_, name, mem_info)); +} + +template +inline void IoBindingImpl::ClearBoundInputs() { + GetApi().ClearBoundInputs(this->p_); +} + +template +inline void IoBindingImpl::ClearBoundOutputs() { + GetApi().ClearBoundOutputs(this->p_); +} + +template +inline void IoBindingImpl::SynchronizeInputs() { + ThrowOnError(GetApi().SynchronizeBoundInputs(this->p_)); +} + +template +inline void IoBindingImpl::SynchronizeOutputs() { + ThrowOnError(GetApi().SynchronizeBoundOutputs(this->p_)); +} + +namespace binding_utils { +inline std::vector GetOutputNamesHelper(const OrtIoBinding* binding, OrtAllocator* allocator) { + std::vector result; + auto free_fn = detail::AllocatedFree(allocator); + using Ptr = std::unique_ptr; + + char* buffer = nullptr; + size_t* lengths = nullptr; + size_t count = 0; + ThrowOnError(GetApi().GetBoundOutputNames(binding, allocator, &buffer, &lengths, &count)); + + if (count == 0) { + return result; + } + + Ptr buffer_g(buffer, free_fn); + Ptr lengths_g(lengths, free_fn); + + result.reserve(count); + for (size_t i = 0; i < count; ++i) { + auto sz = *lengths; + result.emplace_back(buffer, sz); + buffer += sz; + ++lengths; + } + return result; +} + +inline std::vector GetOutputValuesHelper(const OrtIoBinding* binding, OrtAllocator* allocator) { + std::vector result; + size_t owned = 0; + size_t output_count = 0; + // Lambda to release the buffer when no longer needed and + // make sure that we destroy all instances on exception + auto free_fn = [&owned, &output_count, allocator](OrtValue** buffer) { + if (buffer) { + while (owned < output_count) { + auto* p = buffer + owned++; + GetApi().ReleaseValue(*p); + } + allocator->Free(allocator, buffer); + } + }; + using Ptr = std::unique_ptr; + + OrtValue** output_buffer = nullptr; + ThrowOnError(GetApi().GetBoundOutputValues(binding, allocator, &output_buffer, &output_count)); + if (output_count == 0) { + return result; + } + + Ptr buffer_g(output_buffer, free_fn); + + result.reserve(output_count); + for (size_t i = 0; i < output_count; ++i) { + result.emplace_back(output_buffer[i]); + ++owned; + } + return result; +} + +} // namespace binding_utils +} // namespace detail + +inline IoBinding::IoBinding(Session& session) { + ThrowOnError(GetApi().CreateIoBinding(session, &this->p_)); +} + +inline ArenaCfg::ArenaCfg(size_t max_mem, int arena_extend_strategy, int initial_chunk_size_bytes, int max_dead_bytes_per_chunk) { + ThrowOnError(GetApi().CreateArenaCfg(max_mem, arena_extend_strategy, initial_chunk_size_bytes, max_dead_bytes_per_chunk, &p_)); +} + +inline ThreadingOptions::ThreadingOptions() { + ThrowOnError(GetApi().CreateThreadingOptions(&p_)); +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalIntraOpNumThreads(int intra_op_num_threads) { + ThrowOnError(GetApi().SetGlobalIntraOpNumThreads(p_, intra_op_num_threads)); + return *this; +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalInterOpNumThreads(int inter_op_num_threads) { + ThrowOnError(GetApi().SetGlobalInterOpNumThreads(p_, inter_op_num_threads)); + return *this; +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalSpinControl(int allow_spinning) { + ThrowOnError(GetApi().SetGlobalSpinControl(p_, allow_spinning)); + return *this; +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalDenormalAsZero() { + ThrowOnError(GetApi().SetGlobalDenormalAsZero(p_)); + return *this; +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn) { + ThrowOnError(GetApi().SetGlobalCustomCreateThreadFn(p_, ort_custom_create_thread_fn)); + return *this; +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalCustomThreadCreationOptions(void* ort_custom_thread_creation_options) { + ThrowOnError(GetApi().SetGlobalCustomThreadCreationOptions(p_, ort_custom_thread_creation_options)); + return *this; +} + +inline ThreadingOptions& ThreadingOptions::SetGlobalCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn) { + ThrowOnError(GetApi().SetGlobalCustomJoinThreadFn(p_, ort_custom_join_thread_fn)); + return *this; +} + +namespace detail { +template +inline const char* KeyValuePairsImpl::GetValue(const char* key) const { + return GetApi().GetKeyValue(this->p_, key); +} + +template +inline std::unordered_map KeyValuePairsImpl::GetKeyValuePairs() const { + std::unordered_map out; + + size_t num_pairs = 0; + const char* const* keys = nullptr; + const char* const* values = nullptr; + GetApi().GetKeyValuePairs(this->p_, &keys, &values, &num_pairs); + if (num_pairs > 0) { + out.reserve(num_pairs); + for (size_t i = 0; i < num_pairs; ++i) { + out.emplace(keys[i], values[i]); + } + } + + return out; +} + +template +inline void KeyValuePairsImpl::GetKeyValuePairs(std::vector& keys, + std::vector& values) const { + keys.clear(); + values.clear(); + + size_t num_pairs = 0; + const char* const* keys_ptr = nullptr; + const char* const* values_ptr = nullptr; + GetApi().GetKeyValuePairs(this->p_, &keys_ptr, &values_ptr, &num_pairs); + if (num_pairs > 0) { + keys.resize(num_pairs); + values.resize(num_pairs); + std::copy(keys_ptr, keys_ptr + num_pairs, keys.begin()); + std::copy(values_ptr, values_ptr + num_pairs, values.begin()); + } +} +} // namespace detail + +inline KeyValuePairs::KeyValuePairs() { + GetApi().CreateKeyValuePairs(&p_); +} + +inline KeyValuePairs::KeyValuePairs(const std::unordered_map& kv_pairs) { + GetApi().CreateKeyValuePairs(&p_); + for (const auto& kv : kv_pairs) { + GetApi().AddKeyValuePair(this->p_, kv.first.c_str(), kv.second.c_str()); + } +} + +inline void KeyValuePairs::Add(const char* key, const char* value) { + GetApi().AddKeyValuePair(this->p_, key, value); +} + +inline void KeyValuePairs::Remove(const char* key) { + GetApi().RemoveKeyValuePair(this->p_, key); +} + +namespace detail { +template +inline OrtHardwareDeviceType HardwareDeviceImpl::Type() const { + return GetApi().HardwareDevice_Type(this->p_); +} + +template +inline uint32_t HardwareDeviceImpl::VendorId() const { + return GetApi().HardwareDevice_VendorId(this->p_); +} + +template +inline uint32_t HardwareDeviceImpl::DeviceId() const { + return GetApi().HardwareDevice_DeviceId(this->p_); +} + +template +inline const char* HardwareDeviceImpl::Vendor() const { + return GetApi().HardwareDevice_Vendor(this->p_); +} + +template +inline ConstKeyValuePairs HardwareDeviceImpl::Metadata() const { + return ConstKeyValuePairs{GetApi().HardwareDevice_Metadata(this->p_)}; +} + +template +inline const char* EpDeviceImpl::EpName() const { + return GetApi().EpDevice_EpName(this->p_); +} + +template +inline const char* EpDeviceImpl::EpVendor() const { + return GetApi().EpDevice_EpVendor(this->p_); +} + +template +inline ConstKeyValuePairs EpDeviceImpl::EpMetadata() const { + return ConstKeyValuePairs(GetApi().EpDevice_EpMetadata(this->p_)); +} + +template +inline ConstKeyValuePairs EpDeviceImpl::EpOptions() const { + return ConstKeyValuePairs(GetApi().EpDevice_EpOptions(this->p_)); +} + +template +inline ConstHardwareDevice EpDeviceImpl::Device() const { + return ConstHardwareDevice(GetApi().EpDevice_Device(this->p_)); +} +} // namespace detail + +inline EpDevice::EpDevice(OrtEpFactory& ep_factory, ConstHardwareDevice& hardware_device, + ConstKeyValuePairs ep_metadata, ConstKeyValuePairs ep_options) { + ThrowOnError(GetEpApi().CreateEpDevice(&ep_factory, hardware_device, ep_metadata, ep_options, &p_)); +} + +inline Env::Env(OrtLoggingLevel logging_level, _In_ const char* logid) { + ThrowOnError(GetApi().CreateEnv(logging_level, logid, &p_)); + if (strcmp(logid, "onnxruntime-node") == 0) { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS)); + } else { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS)); + } +} + +inline Env::Env(OrtLoggingLevel logging_level, const char* logid, OrtLoggingFunction logging_function, void* logger_param) { + ThrowOnError(GetApi().CreateEnvWithCustomLogger(logging_function, logger_param, logging_level, logid, &p_)); + if (strcmp(logid, "onnxruntime-node") == 0) { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS)); + } else { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS)); + } +} + +inline Env::Env(const OrtThreadingOptions* tp_options, OrtLoggingLevel logging_level, _In_ const char* logid) { + ThrowOnError(GetApi().CreateEnvWithGlobalThreadPools(logging_level, logid, tp_options, &p_)); + if (strcmp(logid, "onnxruntime-node") == 0) { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS)); + } else { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS)); + } +} + +inline Env::Env(const OrtThreadingOptions* tp_options, OrtLoggingFunction logging_function, void* logger_param, + OrtLoggingLevel logging_level, _In_ const char* logid) { + ThrowOnError(GetApi().CreateEnvWithCustomLoggerAndGlobalThreadPools(logging_function, logger_param, logging_level, logid, tp_options, &p_)); + if (strcmp(logid, "onnxruntime-node") == 0) { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_NODEJS)); + } else { + ThrowOnError(GetApi().SetLanguageProjection(p_, OrtLanguageProjection::ORT_PROJECTION_CPLUSPLUS)); + } +} + +inline Env& Env::EnableTelemetryEvents() { + ThrowOnError(GetApi().EnableTelemetryEvents(p_)); + return *this; +} + +inline Env& Env::DisableTelemetryEvents() { + ThrowOnError(GetApi().DisableTelemetryEvents(p_)); + return *this; +} + +inline Env& Env::UpdateEnvWithCustomLogLevel(OrtLoggingLevel log_severity_level) { + ThrowOnError(GetApi().UpdateEnvWithCustomLogLevel(p_, log_severity_level)); + return *this; +} + +inline Env& Env::CreateAndRegisterAllocator(const OrtMemoryInfo* mem_info, const OrtArenaCfg* arena_cfg) { + ThrowOnError(GetApi().CreateAndRegisterAllocator(p_, mem_info, arena_cfg)); + return *this; +} + +inline Env& Env::CreateAndRegisterAllocatorV2(const std::string& provider_type, const OrtMemoryInfo* mem_info, const std::unordered_map& options, const OrtArenaCfg* arena_cfg) { + std::vector keys, values; + auto num_entries = options.size(); + if (num_entries > 0) { + keys.reserve(num_entries); + values.reserve(num_entries); + for (const auto& entry : options) { + keys.push_back(entry.first.c_str()); + values.push_back(entry.second.c_str()); + } + } + ThrowOnError(GetApi().CreateAndRegisterAllocatorV2(p_, provider_type.c_str(), mem_info, arena_cfg, keys.data(), values.data(), num_entries)); + return *this; +} + +inline Env& Env::RegisterExecutionProviderLibrary(const char* registration_name, + const std::basic_string& path) { + ThrowOnError(GetApi().RegisterExecutionProviderLibrary(p_, registration_name, path.c_str())); + return *this; +} + +inline Env& Env::UnregisterExecutionProviderLibrary(const char* registration_name) { + ThrowOnError(GetApi().UnregisterExecutionProviderLibrary(p_, registration_name)); + return *this; +} + +inline std::vector Env::GetEpDevices() const { + size_t num_devices = 0; + const OrtEpDevice* const* device_ptrs = nullptr; + ThrowOnError(GetApi().GetEpDevices(p_, &device_ptrs, &num_devices)); + + std::vector devices; + if (num_devices > 0) { + devices.reserve(num_devices); + for (size_t i = 0; i < num_devices; ++i) { + devices.emplace_back(device_ptrs[i]); + } + } + + return devices; +} + +inline CustomOpDomain::CustomOpDomain(const char* domain) { + ThrowOnError(GetApi().CreateCustomOpDomain(domain, &p_)); +} + +inline void CustomOpDomain::Add(const OrtCustomOp* op) { + ThrowOnError(GetApi().CustomOpDomain_Add(p_, op)); +} + +inline LoraAdapter LoraAdapter::CreateLoraAdapter(const std::basic_string& adapter_path, + OrtAllocator* allocator) { + OrtLoraAdapter* p; + ThrowOnError(GetApi().CreateLoraAdapter(adapter_path.c_str(), allocator, &p)); + return LoraAdapter{p}; +} + +inline LoraAdapter LoraAdapter::CreateLoraAdapterFromArray(const void* bytes, size_t num_bytes, + OrtAllocator* allocator) { + OrtLoraAdapter* p; + ThrowOnError(GetApi().CreateLoraAdapterFromArray(bytes, num_bytes, allocator, &p)); + return LoraAdapter{p}; +} + +inline RunOptions::RunOptions() { + ThrowOnError(GetApi().CreateRunOptions(&p_)); +} + +inline RunOptions& RunOptions::SetRunLogVerbosityLevel(int level) { + ThrowOnError(GetApi().RunOptionsSetRunLogVerbosityLevel(p_, level)); + return *this; +} + +inline RunOptions& RunOptions::SetRunLogSeverityLevel(int level) { + ThrowOnError(GetApi().RunOptionsSetRunLogSeverityLevel(p_, level)); + return *this; +} + +inline int RunOptions::GetRunLogVerbosityLevel() const { + int out; + ThrowOnError(GetApi().RunOptionsGetRunLogVerbosityLevel(p_, &out)); + return out; +} + +inline int RunOptions::GetRunLogSeverityLevel() const { + int out; + ThrowOnError(GetApi().RunOptionsGetRunLogSeverityLevel(p_, &out)); + return out; +} + +inline RunOptions& RunOptions::SetRunTag(const char* run_tag) { + ThrowOnError(GetApi().RunOptionsSetRunTag(p_, run_tag)); + return *this; +} + +inline const char* RunOptions::GetRunTag() const { + const char* out; + ThrowOnError(GetApi().RunOptionsGetRunTag(p_, &out)); + return out; +} + +inline RunOptions& RunOptions::AddConfigEntry(const char* config_key, const char* config_value) { + ThrowOnError(GetApi().AddRunConfigEntry(p_, config_key, config_value)); + return *this; +} + +inline RunOptions& RunOptions::SetTerminate() { + ThrowOnError(GetApi().RunOptionsSetTerminate(p_)); + return *this; +} + +inline RunOptions& RunOptions::UnsetTerminate() { + ThrowOnError(GetApi().RunOptionsUnsetTerminate(p_)); + return *this; +} + +inline RunOptions& RunOptions::AddActiveLoraAdapter(const LoraAdapter& adapter) { + ThrowOnError(GetApi().RunOptionsAddActiveLoraAdapter(p_, adapter)); + return *this; +} + +inline ModelCompilationOptions::ModelCompilationOptions(const Env& env, const SessionOptions& session_options) { + ThrowOnError(GetCompileApi().CreateModelCompilationOptionsFromSessionOptions(env, session_options, &this->p_)); +} + +inline ModelCompilationOptions::ModelCompilationOptions(const Env& env, ConstSessionOptions session_options) { + ThrowOnError(GetCompileApi().CreateModelCompilationOptionsFromSessionOptions(env, session_options, &this->p_)); +} + +inline Status CompileModel(const Env& env, const ModelCompilationOptions& model_compilation_options) { + return Ort::Status(GetCompileApi().CompileModel(env, model_compilation_options)); +} + +inline ModelCompilationOptions& ModelCompilationOptions::SetInputModelPath( + const ORTCHAR_T* input_model_path) { + Ort::ThrowOnError(GetCompileApi().ModelCompilationOptions_SetInputModelPath(this->p_, input_model_path)); + return *this; +} + +inline ModelCompilationOptions& ModelCompilationOptions::SetInputModelFromBuffer( + const void* input_model_data, size_t input_model_data_size) { + Ort::ThrowOnError(GetCompileApi().ModelCompilationOptions_SetInputModelFromBuffer(this->p_, input_model_data, + input_model_data_size)); + return *this; +} + +inline ModelCompilationOptions& ModelCompilationOptions::SetOutputModelPath( + const ORTCHAR_T* output_model_path) { + Ort::ThrowOnError(GetCompileApi().ModelCompilationOptions_SetOutputModelPath(this->p_, output_model_path)); + return *this; +} + +inline ModelCompilationOptions& ModelCompilationOptions::SetOutputModelExternalInitializersFile( + const ORTCHAR_T* file_path, size_t initializer_size_threshold) { + Ort::ThrowOnError(GetCompileApi().ModelCompilationOptions_SetOutputModelExternalInitializersFile( + this->p_, + file_path, + initializer_size_threshold)); + return *this; +} + +inline ModelCompilationOptions& ModelCompilationOptions::SetOutputModelBuffer( + OrtAllocator* allocator, void** output_model_buffer_ptr, size_t* output_model_buffer_size_ptr) { + Ort::ThrowOnError(GetCompileApi().ModelCompilationOptions_SetOutputModelBuffer(this->p_, allocator, + output_model_buffer_ptr, + output_model_buffer_size_ptr)); + return *this; +} + +inline ModelCompilationOptions& ModelCompilationOptions::SetEpContextEmbedMode( + bool embed_ep_context_in_model) { + Ort::ThrowOnError(GetCompileApi().ModelCompilationOptions_SetEpContextEmbedMode( + this->p_, + embed_ep_context_in_model)); + return *this; +} + +namespace detail { + +template +inline Ort::SessionOptions ConstSessionOptionsImpl::Clone() const { + OrtSessionOptions* out; + ThrowOnError(GetApi().CloneSessionOptions(this->p_, &out)); + return SessionOptions{out}; +} + +template +inline std::string ConstSessionOptionsImpl::GetConfigEntry(const char* config_key) const { + size_t size = 0; + // Feed nullptr for the data buffer to query the true size of the string value + Ort::ThrowOnError(GetApi().GetSessionConfigEntry(this->p_, config_key, nullptr, &size)); + + std::string out; + out.resize(size); + Ort::ThrowOnError(GetApi().GetSessionConfigEntry(this->p_, config_key, &out[0], &size)); + out.resize(size - 1); // remove the terminating character '\0' + + return out; +} + +template +inline bool ConstSessionOptionsImpl::HasConfigEntry(const char* config_key) const { + int out = 0; + Ort::ThrowOnError(GetApi().HasSessionConfigEntry(this->p_, config_key, &out)); + return static_cast(out); +} + +template +inline std::string ConstSessionOptionsImpl::GetConfigEntryOrDefault(const char* config_key, + const std::string& def) const { + if (!this->HasConfigEntry(config_key)) { + return def; + } + + return this->GetConfigEntry(config_key); +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetIntraOpNumThreads(int intra_op_num_threads) { + ThrowOnError(GetApi().SetIntraOpNumThreads(this->p_, intra_op_num_threads)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetInterOpNumThreads(int inter_op_num_threads) { + ThrowOnError(GetApi().SetInterOpNumThreads(this->p_, inter_op_num_threads)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetGraphOptimizationLevel(GraphOptimizationLevel graph_optimization_level) { + ThrowOnError(GetApi().SetSessionGraphOptimizationLevel(this->p_, graph_optimization_level)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetDeterministicCompute(bool value) { + ThrowOnError(GetApi().SetDeterministicCompute(this->p_, value)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetOptimizedModelFilePath(const ORTCHAR_T* optimized_model_filepath) { + ThrowOnError(GetApi().SetOptimizedModelFilePath(this->p_, optimized_model_filepath)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::EnableProfiling(const ORTCHAR_T* profile_file_prefix) { + ThrowOnError(GetApi().EnableProfiling(this->p_, profile_file_prefix)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::DisableProfiling() { + ThrowOnError(GetApi().DisableProfiling(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::EnableOrtCustomOps() { + ThrowOnError(GetApi().EnableOrtCustomOps(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::EnableMemPattern() { + ThrowOnError(GetApi().EnableMemPattern(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::DisableMemPattern() { + ThrowOnError(GetApi().DisableMemPattern(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::EnableCpuMemArena() { + ThrowOnError(GetApi().EnableCpuMemArena(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::DisableCpuMemArena() { + ThrowOnError(GetApi().DisableCpuMemArena(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetExecutionMode(ExecutionMode execution_mode) { + ThrowOnError(GetApi().SetSessionExecutionMode(this->p_, execution_mode)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetLoadCancellationFlag(bool value) { + ThrowOnError(GetApi().SessionOptionsSetLoadCancellationFlag(this->p_, value)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetLogId(const char* logid) { + ThrowOnError(GetApi().SetSessionLogId(this->p_, logid)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetLogSeverityLevel(int level) { + ThrowOnError(GetApi().SetSessionLogSeverityLevel(this->p_, level)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::Add(OrtCustomOpDomain* custom_op_domain) { + ThrowOnError(GetApi().AddCustomOpDomain(this->p_, custom_op_domain)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AddConfigEntry(const char* config_key, const char* config_value) { + ThrowOnError(GetApi().AddSessionConfigEntry(this->p_, config_key, config_value)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AddInitializer(const char* name, const OrtValue* ort_val) { + ThrowOnError(GetApi().AddInitializer(this->p_, name, ort_val)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::DisablePerSessionThreads() { + ThrowOnError(GetApi().DisablePerSessionThreads(this->p_)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AddExternalInitializers(const std::vector& names, + const std::vector& ort_values) { + const size_t inputs_num = names.size(); + if (inputs_num != ort_values.size()) { + ORT_CXX_API_THROW("Expecting names and ort_values to have the same length", ORT_INVALID_ARGUMENT); + } + std::vector names_ptr; + std::vector ort_values_ptrs; + names_ptr.reserve(inputs_num); + ort_values_ptrs.reserve(inputs_num); + for (size_t i = 0; i < inputs_num; ++i) { + names_ptr.push_back(names[i].c_str()); + ort_values_ptrs.push_back(ort_values[i]); + } + ThrowOnError(GetApi().AddExternalInitializers(this->p_, names_ptr.data(), ort_values_ptrs.data(), inputs_num)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AddExternalInitializersFromFilesInMemory(const std::vector>& file_names, + const std::vector& buffer_array, + const std::vector& file_lengths) { + const size_t inputs_num = file_names.size(); + if (inputs_num != buffer_array.size()) { + ORT_CXX_API_THROW("Expecting names and buffer_array to have the same length", ORT_INVALID_ARGUMENT); + } + if (inputs_num != file_lengths.size()) { + ORT_CXX_API_THROW("Expecting names and file_lengths to have the same length", ORT_INVALID_ARGUMENT); + } + std::vector names_ptr; + names_ptr.reserve(inputs_num); + for (size_t i = 0; i < inputs_num; ++i) { + names_ptr.push_back(file_names[i].c_str()); + } + ThrowOnError(GetApi().AddExternalInitializersFromFilesInMemory(this->p_, names_ptr.data(), buffer_array.data(), + file_lengths.data(), inputs_num)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_CUDA(const OrtCUDAProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_CUDA_V2(const OrtCUDAProviderOptionsV2& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CUDA_V2(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_ROCM(const OrtROCMProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_ROCM(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_TensorRT(const OrtTensorRTProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_TensorRT(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_TensorRT_V2(const OrtTensorRTProviderOptionsV2& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_TensorRT_V2(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_MIGraphX(const OrtMIGraphXProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_MIGraphX(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_CANN(const OrtCANNProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_CANN(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_Dnnl(const OrtDnnlProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_Dnnl(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider( + const std::string& provider_name, + const std::unordered_map& provider_options) { + auto num_entries = provider_options.size(); + std::vector keys, values; + if (num_entries > 0) { + keys.reserve(num_entries); + values.reserve(num_entries); + + for (const auto& entry : provider_options) { + keys.push_back(entry.first.c_str()); + values.push_back(entry.second.c_str()); + } + } + + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider(this->p_, provider_name.c_str(), + keys.data(), values.data(), num_entries)); + + return *this; +} + +namespace { +template +void SessionOptionsAppendEP(detail::SessionOptionsImpl& session_options, + Env& env, const std::vector& ep_devices, + const std::vector& ep_options_keys, + const std::vector& ep_options_values) { + std::vector ep_devices_ptrs; + ep_devices_ptrs.reserve(ep_devices.size()); + for (const auto& ep_device : ep_devices) { + ep_devices_ptrs.push_back(ep_device); + } + + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_V2( + session_options, env, ep_devices_ptrs.data(), ep_devices_ptrs.size(), + ep_options_keys.data(), ep_options_values.data(), ep_options_keys.size())); +} +} // namespace + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_V2( + Env& env, const std::vector& ep_devices, const KeyValuePairs& ep_options) { + std::vector ep_options_keys, ep_options_values; + ep_options.GetKeyValuePairs(ep_options_keys, ep_options_values); + + SessionOptionsAppendEP(*this, env, ep_devices, ep_options_keys, ep_options_values); + + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_V2( + Env& env, const std::vector& ep_devices, + const std::unordered_map& ep_options) { + std::vector ep_options_keys, ep_options_values; + ep_options_keys.reserve(ep_options.size()); + ep_options_values.reserve(ep_options.size()); + + for (const auto& [key, value] : ep_options) { + ep_options_keys.push_back(key.c_str()); + ep_options_values.push_back(value.c_str()); + } + + SessionOptionsAppendEP(*this, env, ep_devices, ep_options_keys, ep_options_values); + + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetEpSelectionPolicy(OrtExecutionProviderDevicePolicy policy) { + ThrowOnError(GetApi().SessionOptionsSetEpSelectionPolicy(this->p_, policy)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetEpSelectionPolicy(EpSelectionDelegate delegate, void* state) { + ThrowOnError(GetApi().SessionOptionsSetEpSelectionPolicyDelegate(this->p_, delegate, state)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetCustomCreateThreadFn(OrtCustomCreateThreadFn ort_custom_create_thread_fn) { + ThrowOnError(GetApi().SessionOptionsSetCustomCreateThreadFn(this->p_, ort_custom_create_thread_fn)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetCustomThreadCreationOptions(void* ort_custom_thread_creation_options) { + ThrowOnError(GetApi().SessionOptionsSetCustomThreadCreationOptions(this->p_, ort_custom_thread_creation_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::SetCustomJoinThreadFn(OrtCustomJoinThreadFn ort_custom_join_thread_fn) { + ThrowOnError(GetApi().SessionOptionsSetCustomJoinThreadFn(this->p_, ort_custom_join_thread_fn)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_OpenVINO(const OrtOpenVINOProviderOptions& provider_options) { + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_OpenVINO(this->p_, &provider_options)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_OpenVINO_V2(const std::unordered_map& provider_options) { + auto num_entries = provider_options.size(); + std::vector keys, values; + if (num_entries > 0) { + keys.reserve(num_entries); + values.reserve(num_entries); + + for (const auto& entry : provider_options) { + keys.push_back(entry.first.c_str()); + values.push_back(entry.second.c_str()); + } + } + + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_OpenVINO_V2(this->p_, + keys.data(), values.data(), num_entries)); + + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::AppendExecutionProvider_VitisAI(const std::unordered_map& provider_options) { + auto num_entries = provider_options.size(); + std::vector keys, values; + if (num_entries > 0) { + keys.reserve(num_entries); + values.reserve(num_entries); + + for (const auto& entry : provider_options) { + keys.push_back(entry.first.c_str()); + values.push_back(entry.second.c_str()); + } + } + + ThrowOnError(GetApi().SessionOptionsAppendExecutionProvider_VitisAI(this->p_, keys.data(), values.data(), num_entries)); + + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::RegisterCustomOpsLibrary(const ORTCHAR_T* library_name, + const CustomOpConfigs& custom_op_configs) { + // Add custom op config entries before registering the custom op library. Otherwise, the config entries _may_ be ignored by + // the custom op library. + for (const auto& config_iter : custom_op_configs.GetFlattenedConfigs()) { + AddConfigEntry(config_iter.first.c_str(), config_iter.second.c_str()); + } + + ThrowOnError(GetApi().RegisterCustomOpsLibrary_V2(this->p_, library_name)); + return *this; +} + +template +inline SessionOptionsImpl& SessionOptionsImpl::RegisterCustomOpsUsingFunction(const char* registration_function_name) { + ThrowOnError(GetApi().RegisterCustomOpsUsingFunction(this->p_, registration_function_name)); + return *this; +} + +/// Session +template +inline size_t ConstSessionImpl::GetInputCount() const { + size_t out; + ThrowOnError(GetApi().SessionGetInputCount(this->p_, &out)); + return out; +} + +template +inline size_t ConstSessionImpl::GetOutputCount() const { + size_t out; + ThrowOnError(GetApi().SessionGetOutputCount(this->p_, &out)); + return out; +} + +template +inline size_t ConstSessionImpl::GetOverridableInitializerCount() const { + size_t out; + ThrowOnError(GetApi().SessionGetOverridableInitializerCount(this->p_, &out)); + return out; +} + +template +inline std::vector ConstSessionImpl::GetInputNames() const { + AllocatorWithDefaultOptions allocator; + + auto num_inputs = GetInputCount(); + std::vector input_names; + input_names.reserve(num_inputs); + + for (size_t i = 0; i < num_inputs; ++i) { + char* name = nullptr; + ThrowOnError(GetApi().SessionGetInputName(this->p_, i, allocator, &name)); + input_names.push_back(name); + allocator.Free(name); + } + + return input_names; +} + +template +inline std::vector ConstSessionImpl::GetOutputNames() const { + AllocatorWithDefaultOptions allocator; + + auto num_inputs = GetOutputCount(); + std::vector output_names; + output_names.reserve(num_inputs); + + for (size_t i = 0; i < num_inputs; ++i) { + char* name = nullptr; + ThrowOnError(GetApi().SessionGetOutputName(this->p_, i, allocator, &name)); + output_names.push_back(name); + allocator.Free(name); + } + + return output_names; +} + +template +inline std::vector ConstSessionImpl::GetOverridableInitializerNames() const { + AllocatorWithDefaultOptions allocator; + + auto num_initializers = GetOverridableInitializerCount(); + std::vector initializer_names; + initializer_names.reserve(num_initializers); + + for (size_t i = 0; i < num_initializers; ++i) { + char* name = nullptr; + ThrowOnError(GetApi().SessionGetOverridableInitializerName(this->p_, i, allocator, &name)); + initializer_names.push_back(name); + } + + return initializer_names; +} + +template +inline AllocatedStringPtr ConstSessionImpl::GetInputNameAllocated(size_t index, OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().SessionGetInputName(this->p_, index, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +template +inline AllocatedStringPtr ConstSessionImpl::GetOutputNameAllocated(size_t index, OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().SessionGetOutputName(this->p_, index, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +template +inline AllocatedStringPtr ConstSessionImpl::GetOverridableInitializerNameAllocated(size_t index, OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().SessionGetOverridableInitializerName(this->p_, index, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +template +inline uint64_t ConstSessionImpl::GetProfilingStartTimeNs() const { + uint64_t out; + ThrowOnError(GetApi().SessionGetProfilingStartTimeNs(this->p_, &out)); + return out; +} + +template +inline ModelMetadata ConstSessionImpl::GetModelMetadata() const { + OrtModelMetadata* out; + ThrowOnError(GetApi().SessionGetModelMetadata(this->p_, &out)); + return ModelMetadata{out}; +} + +template +inline TypeInfo ConstSessionImpl::GetInputTypeInfo(size_t index) const { + OrtTypeInfo* out; + ThrowOnError(GetApi().SessionGetInputTypeInfo(this->p_, index, &out)); + return TypeInfo{out}; +} + +template +inline TypeInfo ConstSessionImpl::GetOutputTypeInfo(size_t index) const { + OrtTypeInfo* out; + ThrowOnError(GetApi().SessionGetOutputTypeInfo(this->p_, index, &out)); + return TypeInfo{out}; +} + +template +inline TypeInfo ConstSessionImpl::GetOverridableInitializerTypeInfo(size_t index) const { + OrtTypeInfo* out; + ThrowOnError(GetApi().SessionGetOverridableInitializerTypeInfo(this->p_, index, &out)); + return TypeInfo{out}; +} + +#if !defined(ORT_MINIMAL_BUILD) +template +inline int ConstSessionImpl::GetOpset(const std::string& domain) const { + int opset; + ThrowOnError(GetModelEditorApi().SessionGetOpsetForDomain(this->p_, domain.c_str(), &opset)); + return opset; +} +#endif // !defined(ORT_MINIMAL_BUILD) + +template +std::vector ConstSessionImpl::GetInputs() const { + const std::vector input_names = GetInputNames(); + + std::vector inputs; + inputs.reserve(input_names.size()); + + for (size_t i = 0; i < input_names.size(); ++i) { + auto type_info = GetInputTypeInfo(i); + inputs.emplace_back(ValueInfo{input_names[i], type_info.GetConst()}); + } + + return inputs; +} + +template +std::vector ConstSessionImpl::GetOutputs() const { + const std::vector output_names = GetOutputNames(); + + std::vector outputs; + outputs.reserve(output_names.size()); + + for (size_t i = 0; i < output_names.size(); ++i) { + auto type_info = GetOutputTypeInfo(i); + outputs.emplace_back(ValueInfo{output_names[i], type_info.GetConst()}); + } + + return outputs; +} + +template +inline std::vector SessionImpl::Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count, + const char* const* output_names, size_t output_count) { + std::vector output_values; + output_values.reserve(output_count); + for (size_t i = 0; i < output_count; i++) + output_values.emplace_back(nullptr); + Run(run_options, input_names, input_values, input_count, output_names, output_values.data(), output_count); + return output_values; +} + +template +inline void SessionImpl::Run(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count, + const char* const* output_names, Value* output_values, size_t output_count) { + static_assert(sizeof(Value) == sizeof(OrtValue*), "Value is really just an array of OrtValue* in memory, so we can reinterpret_cast safely"); + auto ort_input_values = reinterpret_cast(input_values); + auto ort_output_values = reinterpret_cast(output_values); + ThrowOnError(GetApi().Run(this->p_, run_options, input_names, ort_input_values, input_count, output_names, output_count, ort_output_values)); +} + +template +inline void SessionImpl::Run(const RunOptions& run_options, const IoBinding& io_binding) { + ThrowOnError(GetApi().RunWithBinding(this->p_, run_options, io_binding)); +} + +template +inline void SessionImpl::RunAsync(const RunOptions& run_options, const char* const* input_names, const Value* input_values, size_t input_count, + const char* const* output_names, Value* output_values, size_t output_count, RunAsyncCallbackFn callback, void* user_data) { + auto ort_input_values = reinterpret_cast(input_values); + auto ort_output_values = reinterpret_cast(output_values); + ThrowOnError(GetApi().RunAsync(this->p_, run_options, input_names, + ort_input_values, input_count, output_names, output_count, + ort_output_values, callback, user_data)); +} + +template +inline AllocatedStringPtr SessionImpl::EndProfilingAllocated(OrtAllocator* allocator) { + char* out = nullptr; + ThrowOnError(GetApi().SessionEndProfiling(this->p_, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +template +inline void SessionImpl::SetEpDynamicOptions(const char* const* keys, const char* const* values, size_t kv_len) { + ThrowOnError(GetApi().SetEpDynamicOptions(this->p_, keys, values, kv_len)); +} + +#if !defined(ORT_MINIMAL_BUILD) +template +inline void SessionImpl::FinalizeModelEditorSession(const Model& model, const SessionOptions& options, + OrtPrepackedWeightsContainer* prepacked_weights_container) { + ThrowOnError(GetModelEditorApi().ApplyModelToModelEditorSession(this->p_, model)); + ThrowOnError(GetModelEditorApi().FinalizeModelEditorSession(this->p_, options, prepacked_weights_container)); +} +#endif // #if !defined(ORT_MINIMAL_BUILD) + +} // namespace detail + +inline SessionOptions::SessionOptions() { + ThrowOnError(GetApi().CreateSessionOptions(&this->p_)); +} + +/// CustomOpConfigs +inline std::string detail::MakeCustomOpConfigEntryKey(const char* custom_op_name, const char* config) { + std::string config_key = "custom_op."; + + config_key += custom_op_name; + config_key += "."; + config_key += config; + + return config_key; +} + +inline CustomOpConfigs& CustomOpConfigs::AddConfig(const char* custom_op_name, const char* config_key, const char* config_value) { + const std::string full_flat_key = detail::MakeCustomOpConfigEntryKey(custom_op_name, config_key); + flat_configs_[full_flat_key] = config_value; + return *this; +} + +inline const std::unordered_map& CustomOpConfigs::GetFlattenedConfigs() const { + return flat_configs_; +} + +inline Session::Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options) { + ThrowOnError(GetApi().CreateSession(env, model_path, options, &this->p_)); +} + +inline Session::Session(const Env& env, const ORTCHAR_T* model_path, const SessionOptions& options, + OrtPrepackedWeightsContainer* prepacked_weights_container) { + ThrowOnError(GetApi().CreateSessionWithPrepackedWeightsContainer(env, model_path, options, prepacked_weights_container, &this->p_)); +} + +inline Session::Session(const Env& env, const void* model_data, size_t model_data_length, const SessionOptions& options) { + ThrowOnError(GetApi().CreateSessionFromArray(env, model_data, model_data_length, options, &this->p_)); +} + +inline Session::Session(const Env& env, const void* model_data, size_t model_data_length, + const SessionOptions& options, OrtPrepackedWeightsContainer* prepacked_weights_container) { + ThrowOnError(GetApi().CreateSessionFromArrayWithPrepackedWeightsContainer(env, model_data, model_data_length, options, + prepacked_weights_container, &this->p_)); +} + +#if !defined(ORT_MINIMAL_BUILD) +inline Session::Session(const Env& env, const Model& model, const SessionOptions& options) { + ThrowOnError(GetModelEditorApi().CreateSessionFromModel(env, model.GetConst(), options, &this->p_)); +} + +// static +inline Session Session::CreateModelEditorSession(const Env& env, const ORTCHAR_T* model_path, + const SessionOptions& options) { + OrtSession* session = nullptr; + ThrowOnError(GetModelEditorApi().CreateModelEditorSession(env, model_path, options, &session)); + return Session(session); +} + +// static +inline Session Session::CreateModelEditorSession(const Env& env, const void* model_data, size_t model_data_length, + const SessionOptions& options) { + OrtSession* session = nullptr; + ThrowOnError(GetModelEditorApi().CreateModelEditorSessionFromArray(env, model_data, model_data_length, options, + &session)); + return Session(session); +} + +void FinalizeModelEditorSession(const Model& model, const SessionOptions& options, + OrtPrepackedWeightsContainer* prepacked_weights_container); +#endif // #if !defined(ORT_MINIMAL_BUILD) + +inline AllocatedStringPtr ModelMetadata::GetProducerNameAllocated(OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().ModelMetadataGetProducerName(p_, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +inline AllocatedStringPtr ModelMetadata::GetGraphNameAllocated(OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().ModelMetadataGetGraphName(p_, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +inline AllocatedStringPtr ModelMetadata::GetDomainAllocated(OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().ModelMetadataGetDomain(p_, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +inline AllocatedStringPtr Ort::ModelMetadata::GetDescriptionAllocated(OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().ModelMetadataGetDescription(p_, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +inline AllocatedStringPtr ModelMetadata::GetGraphDescriptionAllocated(OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().ModelMetadataGetGraphDescription(p_, allocator, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +inline AllocatedStringPtr ModelMetadata::LookupCustomMetadataMapAllocated(const char* key, OrtAllocator* allocator) const { + char* out; + ThrowOnError(GetApi().ModelMetadataLookupCustomMetadataMap(p_, allocator, key, &out)); + return AllocatedStringPtr(out, detail::AllocatedFree(allocator)); +} + +inline std::vector ModelMetadata::GetCustomMetadataMapKeysAllocated(OrtAllocator* allocator) const { + auto deletor = detail::AllocatedFree(allocator); + std::vector result; + + char** out = nullptr; + int64_t num_keys = 0; + ThrowOnError(GetApi().ModelMetadataGetCustomMetadataMapKeys(p_, allocator, &out, &num_keys)); + if (num_keys <= 0) { + return result; + } + + // array of pointers will be freed + std::unique_ptr array_guard(out, deletor); + // reserve may throw + auto strings_deletor = [&deletor, num_keys](char** out) { for(int64_t i = 0; i < num_keys; ++i) deletor(out[i]); }; + std::unique_ptr strings_guard(out, strings_deletor); + result.reserve(static_cast(num_keys)); + strings_guard.release(); + for (int64_t i = 0; i < num_keys; ++i) { + result.push_back(AllocatedStringPtr(out[i], deletor)); + } + + return result; +} + +inline int64_t ModelMetadata::GetVersion() const { + int64_t out; + ThrowOnError(GetApi().ModelMetadataGetVersion(p_, &out)); + return out; +} + +inline TensorTypeAndShapeInfo::TensorTypeAndShapeInfo(ONNXTensorElementDataType element_type, + const std::vector& dims, + const std::vector* symbolic_dims) { + ThrowOnError(GetApi().CreateTensorTypeAndShapeInfo(&p_)); + ThrowOnError(GetApi().SetTensorElementType(p_, element_type)); + ThrowOnError(GetApi().SetDimensions(p_, dims.data(), dims.size())); + + if (symbolic_dims) { + std::vector symbolic_dims_cstr; + symbolic_dims_cstr.reserve(symbolic_dims->size()); + std::transform(symbolic_dims->begin(), symbolic_dims->end(), std::back_inserter(symbolic_dims_cstr), + [](const std::string& s) { return s.c_str(); }); + ThrowOnError(GetApi().SetSymbolicDimensions(p_, symbolic_dims_cstr.data(), symbolic_dims_cstr.size())); + } +} + +#if !defined(ORT_MINIMAL_BUILD) +// static +inline TypeInfo TypeInfo::CreateTensorInfo(ConstTensorTypeAndShapeInfo tensor_type_and_shape_info) { + OrtTypeInfo* output = nullptr; + ThrowOnError(GetModelEditorApi().CreateTensorTypeInfo(tensor_type_and_shape_info, &output)); + return TypeInfo{output}; +} + +// static +inline TypeInfo TypeInfo::CreateSparseTensorInfo(ConstTensorTypeAndShapeInfo sparse_tensor_type_and_shape_info) { + OrtTypeInfo* output = nullptr; + ThrowOnError(GetModelEditorApi().CreateSparseTensorTypeInfo(sparse_tensor_type_and_shape_info, &output)); + return TypeInfo{output}; +} + +// static +inline TypeInfo TypeInfo::CreateSequenceTypeInfo(ConstTypeInfo sequence_type) { + OrtTypeInfo* output; + ThrowOnError(GetModelEditorApi().CreateSequenceTypeInfo(sequence_type, &output)); + return TypeInfo{output}; +} + +// static +inline TypeInfo TypeInfo::CreateMapTypeInfo(ONNXTensorElementDataType key_type, ConstTypeInfo value_type) { + OrtTypeInfo* output; + ThrowOnError(GetModelEditorApi().CreateMapTypeInfo(key_type, value_type, &output)); + return TypeInfo{output}; +} + +// static +inline TypeInfo TypeInfo::CreateOptionalTypeInfo(ConstTypeInfo contained_type) { + OrtTypeInfo* output; + ThrowOnError(GetModelEditorApi().CreateOptionalTypeInfo(contained_type, &output)); + return TypeInfo{output}; +} +#endif // #if !defined(ORT_MINIMAL_BUILD) + +namespace detail { + +template +inline ONNXTensorElementDataType TensorTypeAndShapeInfoImpl::GetElementType() const { + ONNXTensorElementDataType out; + ThrowOnError(GetApi().GetTensorElementType(this->p_, &out)); + return out; +} + +template +inline size_t TensorTypeAndShapeInfoImpl::GetElementCount() const { + size_t out; + ThrowOnError(GetApi().GetTensorShapeElementCount(this->p_, &out)); + return static_cast(out); +} + +template +inline size_t TensorTypeAndShapeInfoImpl::GetDimensionsCount() const { + size_t out; + ThrowOnError(GetApi().GetDimensionsCount(this->p_, &out)); + return out; +} + +template +inline void TensorTypeAndShapeInfoImpl::GetDimensions(int64_t* values, size_t values_count) const { + ThrowOnError(GetApi().GetDimensions(this->p_, values, values_count)); +} + +template +inline void TensorTypeAndShapeInfoImpl::GetSymbolicDimensions(const char** values, size_t values_count) const { + ThrowOnError(GetApi().GetSymbolicDimensions(this->p_, values, values_count)); +} + +template +inline std::vector TensorTypeAndShapeInfoImpl::GetSymbolicDimensions() const { + std::vector out(GetDimensionsCount(), nullptr); + ThrowOnError(GetApi().GetSymbolicDimensions(this->p_, out.data(), out.size())); + return out; +} + +template +inline std::vector TensorTypeAndShapeInfoImpl::GetShape() const { + std::vector out(GetDimensionsCount(), -1); + ThrowOnError(GetApi().GetDimensions(this->p_, out.data(), out.size())); + return out; +} + +template +inline ConstTensorTypeAndShapeInfo TypeInfoImpl::GetTensorTypeAndShapeInfo() const { + const OrtTensorTypeAndShapeInfo* out; + ThrowOnError(GetApi().CastTypeInfoToTensorInfo(this->p_, &out)); + return ConstTensorTypeAndShapeInfo{out}; +} + +template +inline ConstSequenceTypeInfo TypeInfoImpl::GetSequenceTypeInfo() const { + const OrtSequenceTypeInfo* out; + ThrowOnError(GetApi().CastTypeInfoToSequenceTypeInfo(this->p_, &out)); + return ConstSequenceTypeInfo{out}; +} + +template +inline ConstMapTypeInfo TypeInfoImpl::GetMapTypeInfo() const { + const OrtMapTypeInfo* out; + ThrowOnError(GetApi().CastTypeInfoToMapTypeInfo(this->p_, &out)); + return ConstMapTypeInfo{out}; +} + +template +inline ONNXType TypeInfoImpl::GetONNXType() const { + ONNXType out; + ThrowOnError(GetApi().GetOnnxTypeFromTypeInfo(this->p_, &out)); + return out; +} + +template +inline TypeInfo SequenceTypeInfoImpl::GetSequenceElementType() const { + OrtTypeInfo* output; + ThrowOnError(GetApi().GetSequenceElementType(this->p_, &output)); + return TypeInfo{output}; +} + +template +inline TypeInfo OptionalTypeInfoImpl::GetOptionalElementType() const { + OrtTypeInfo* info; + ThrowOnError(GetApi().GetOptionalContainedTypeInfo(this->p_, &info)); + return TypeInfo{info}; +} + +template +inline ONNXTensorElementDataType MapTypeInfoImpl::GetMapKeyType() const { + ONNXTensorElementDataType out; + ThrowOnError(GetApi().GetMapKeyType(this->p_, &out)); + return out; +} + +template +inline TypeInfo MapTypeInfoImpl::GetMapValueType() const { + OrtTypeInfo* output; + ThrowOnError(GetApi().GetMapValueType(this->p_, &output)); + return TypeInfo{output}; +} + +template +inline ConstOptionalTypeInfo TypeInfoImpl::GetOptionalTypeInfo() const { + const OrtOptionalTypeInfo* info; + ThrowOnError(GetApi().CastTypeInfoToOptionalTypeInfo(this->p_, &info)); + return ConstOptionalTypeInfo{info}; +} + +} // namespace detail + +namespace detail { + +template +template +inline void ConstValueImpl::GetOpaqueData(const char* domain, const char* type_name, R& out) const { + ThrowOnError(GetApi().GetOpaqueValue(domain, type_name, this->p_, &out, sizeof(R))); +} + +template +inline bool ConstValueImpl::IsTensor() const { + int out; + ThrowOnError(GetApi().IsTensor(this->p_, &out)); + return out != 0; +} + +template +inline bool ConstValueImpl::HasValue() const { + int out; + ThrowOnError(GetApi().HasValue(this->p_, &out)); + return out != 0; +} + +template +inline size_t ConstValueImpl::GetCount() const { + size_t out; + ThrowOnError(GetApi().GetValueCount(this->p_, &out)); + return out; +} + +template +inline Value ConstValueImpl::GetValue(int index, OrtAllocator* allocator) const { + OrtValue* out; + ThrowOnError(GetApi().GetValue(this->p_, index, allocator, &out)); + return Value{out}; +} + +template +inline size_t ConstValueImpl::GetStringTensorDataLength() const { + size_t out; + ThrowOnError(GetApi().GetStringTensorDataLength(this->p_, &out)); + return out; +} + +template +inline size_t ConstValueImpl::GetStringTensorElementLength(size_t element_index) const { + size_t out; + ThrowOnError(GetApi().GetStringTensorElementLength(this->p_, element_index, &out)); + return out; +} + +template +template +inline const R* ConstValueImpl::GetTensorData() const { + R* out; + ThrowOnError(GetApi().GetTensorMutableData(const_cast(this->p_), (void**)&out)); + return out; +} + +template +inline const void* ConstValueImpl::GetTensorRawData() const { + void* out; + ThrowOnError(GetApi().GetTensorMutableData(const_cast(this->p_), &out)); + return out; +} + +template +inline TypeInfo ConstValueImpl::GetTypeInfo() const { + OrtTypeInfo* output; + ThrowOnError(GetApi().GetTypeInfo(this->p_, &output)); + return TypeInfo{output}; +} + +template +inline TensorTypeAndShapeInfo ConstValueImpl::GetTensorTypeAndShapeInfo() const { + OrtTensorTypeAndShapeInfo* output; + ThrowOnError(GetApi().GetTensorTypeAndShape(this->p_, &output)); + return TensorTypeAndShapeInfo{output}; +} + +template +inline ConstMemoryInfo ConstValueImpl::GetTensorMemoryInfo() const { + const OrtMemoryInfo* mem_info; + ThrowOnError(GetApi().GetTensorMemoryInfo(this->p_, &mem_info)); + return ConstMemoryInfo(mem_info); +} + +template +inline void ConstValueImpl::GetStringTensorElement(size_t buffer_length, size_t element_index, void* buffer) const { + ThrowOnError(GetApi().GetStringTensorElement(this->p_, buffer_length, element_index, buffer)); +} + +template +inline std::string ConstValueImpl::GetStringTensorElement(size_t element_index) const { + size_t buffer_length; + ThrowOnError(GetApi().GetStringTensorElementLength(this->p_, element_index, &buffer_length)); + + std::string s; + s.resize(buffer_length); + ThrowOnError(GetApi().GetStringTensorElement(this->p_, buffer_length, element_index, &s[0])); + return s; +} + +template +inline void ConstValueImpl::GetStringTensorContent(void* buffer, size_t buffer_length, size_t* offsets, size_t offsets_count) const { + ThrowOnError(GetApi().GetStringTensorContent(this->p_, buffer, buffer_length, offsets, offsets_count)); +} + +#if !defined(DISABLE_SPARSE_TENSORS) +template +inline OrtSparseFormat ConstValueImpl::GetSparseFormat() const { + OrtSparseFormat format; + ThrowOnError(GetApi().GetSparseTensorFormat(this->p_, &format)); + return format; +} + +template +inline TensorTypeAndShapeInfo ConstValueImpl::GetSparseTensorValuesTypeAndShapeInfo() const { + OrtTensorTypeAndShapeInfo* output; + ThrowOnError(GetApi().GetSparseTensorValuesTypeAndShape(this->p_, &output)); + return TensorTypeAndShapeInfo{output}; +} + +template +inline TensorTypeAndShapeInfo ConstValueImpl::GetSparseTensorIndicesTypeShapeInfo(OrtSparseIndicesFormat indices_format) const { + OrtTensorTypeAndShapeInfo* output; + ThrowOnError(GetApi().GetSparseTensorIndicesTypeShape(this->p_, indices_format, &output)); + return TensorTypeAndShapeInfo{output}; +} + +template +template +inline const R* ConstValueImpl::GetSparseTensorIndicesData(OrtSparseIndicesFormat indices_format, size_t& num_indices) const { + const void* out; + ThrowOnError(GetApi().GetSparseTensorIndices(this->p_, indices_format, &num_indices, &out)); + return reinterpret_cast(out); +} + +template +inline bool ConstValueImpl::IsSparseTensor() const { + int out; + ThrowOnError(GetApi().IsSparseTensor(this->p_, &out)); + return out != 0; +} + +template +template +inline const R* ConstValueImpl::GetSparseTensorValues() const { + const void* out; + ThrowOnError(GetApi().GetSparseTensorValues(this->p_, &out)); + return reinterpret_cast(out); +} + +#endif + +template +void ValueImpl::FillStringTensor(const char* const* s, size_t s_len) { + ThrowOnError(GetApi().FillStringTensor(this->p_, s, s_len)); +} + +template +void ValueImpl::FillStringTensorElement(const char* s, size_t index) { + ThrowOnError(GetApi().FillStringTensorElement(this->p_, s, index)); +} + +template +inline char* ValueImpl::GetResizedStringTensorElementBuffer(size_t index, size_t buffer_length) { + char* result; + ThrowOnError(GetApi().GetResizedStringTensorElementBuffer(this->p_, index, buffer_length, &result)); + return result; +} + +template +void* ValueImpl::GetTensorMutableRawData() { + void* out; + ThrowOnError(GetApi().GetTensorMutableData(this->p_, &out)); + return out; +} + +template +template +R* ValueImpl::GetTensorMutableData() { + R* out; + ThrowOnError(GetApi().GetTensorMutableData(this->p_, (void**)&out)); + return out; +} + +template +template +R& ValueImpl::At(const std::vector& location) { + static_assert(!std::is_same::value, "this api does not support std::string"); + R* out; + ThrowOnError(GetApi().TensorAt(this->p_, location.data(), location.size(), (void**)&out)); + return *out; +} + +#if !defined(DISABLE_SPARSE_TENSORS) +template +void ValueImpl::UseCooIndices(int64_t* indices_data, size_t indices_num) { + ThrowOnError(GetApi().UseCooIndices(this->p_, indices_data, indices_num)); +} + +template +void ValueImpl::UseCsrIndices(int64_t* inner_data, size_t inner_num, int64_t* outer_data, size_t outer_num) { + ThrowOnError(GetApi().UseCsrIndices(this->p_, inner_data, inner_num, outer_data, outer_num)); +} + +template +void ValueImpl::UseBlockSparseIndices(const Shape& indices_shape, int32_t* indices_data) { + ThrowOnError(GetApi().UseBlockSparseIndices(this->p_, indices_shape.shape, indices_shape.shape_len, indices_data)); +} + +template +void ValueImpl::FillSparseTensorCoo(const OrtMemoryInfo* mem_info, const OrtSparseValuesParam& values_param, + const int64_t* indices_data, size_t indices_num) { + ThrowOnError(GetApi().FillSparseTensorCoo(this->p_, mem_info, values_param.values_shape, + values_param.values_shape_len, values_param.data.p_data, + indices_data, indices_num)); +} + +template +void ValueImpl::FillSparseTensorCsr(const OrtMemoryInfo* data_mem_info, + const OrtSparseValuesParam& values, + const int64_t* inner_indices_data, size_t inner_indices_num, + const int64_t* outer_indices_data, size_t outer_indices_num) { + ThrowOnError(GetApi().FillSparseTensorCsr(this->p_, data_mem_info, values.values_shape, values.values_shape_len, values.data.p_data, + inner_indices_data, inner_indices_num, + outer_indices_data, outer_indices_num)); +} + +template +void ValueImpl::FillSparseTensorBlockSparse(const OrtMemoryInfo* data_mem_info, + const OrtSparseValuesParam& values, + const Shape& indices_shape, + const int32_t* indices_data) { + ThrowOnError(GetApi().FillSparseTensorBlockSparse(this->p_, data_mem_info, values.values_shape, values.values_shape_len, values.data.p_data, + indices_shape.shape, indices_shape.shape_len, + indices_data)); +} + +#endif // !defined(DISABLE_SPARSE_TENSORS) + +} // namespace detail + +template +inline Value Value::CreateTensor(const OrtMemoryInfo* info, T* p_data, size_t p_data_element_count, + const int64_t* shape, size_t shape_len) { + return CreateTensor(info, p_data, p_data_element_count * sizeof(T), shape, shape_len, TypeToTensorType::type); +} + +inline Value Value::CreateTensor(const OrtMemoryInfo* info, void* p_data, size_t p_data_byte_count, + const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type) { + OrtValue* out; + ThrowOnError(GetApi().CreateTensorWithDataAsOrtValue(info, p_data, p_data_byte_count, shape, shape_len, type, &out)); + return Value{out}; +} + +inline Value Value::CreateTensor(OrtAllocator* deleter, void* p_data, size_t p_data_byte_count, + const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type) { + OrtValue* out; + ThrowOnError(GetApi().CreateTensorWithDataAndDeleterAsOrtValue(deleter, p_data, p_data_byte_count, + shape, shape_len, type, &out)); + return Value{out}; +} + +template +inline Value Value::CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len) { + return CreateTensor(allocator, shape, shape_len, TypeToTensorType::type); +} + +inline Value Value::CreateTensor(OrtAllocator* allocator, const int64_t* shape, size_t shape_len, + ONNXTensorElementDataType type) { + OrtValue* out; + ThrowOnError(GetApi().CreateTensorAsOrtValue(allocator, shape, shape_len, type, &out)); + return Value{out}; +} + +#if !defined(DISABLE_SPARSE_TENSORS) + +template +inline Value Value::CreateSparseTensor(const OrtMemoryInfo* info, T* p_data, const Shape& dense_shape, + const Shape& values_shape) { + return CreateSparseTensor(info, p_data, dense_shape, values_shape, TypeToTensorType::type); +} + +inline Value Value::CreateSparseTensor(const OrtMemoryInfo* info, void* p_data, const Shape& dense_shape, + const Shape& values_shape, ONNXTensorElementDataType type) { + OrtValue* out; + ThrowOnError(GetApi().CreateSparseTensorWithValuesAsOrtValue(info, p_data, dense_shape.shape, dense_shape.shape_len, + values_shape.shape, values_shape.shape_len, type, + &out)); + return Value{out}; +} + +template +inline Value Value::CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape) { + return CreateSparseTensor(allocator, dense_shape, TypeToTensorType::type); +} + +inline Value Value::CreateSparseTensor(OrtAllocator* allocator, const Shape& dense_shape, + ONNXTensorElementDataType type) { + OrtValue* out; + ThrowOnError(GetApi().CreateSparseTensorAsOrtValue(allocator, dense_shape.shape, dense_shape.shape_len, type, &out)); + return Value{out}; +} +#endif // !defined(DISABLE_SPARSE_TENSORS) + +inline Value Value::CreateMap(const Value& keys, const Value& values) { + OrtValue* out; + const OrtValue* inputs[2] = {keys, values}; + ThrowOnError(GetApi().CreateValue(inputs, 2, ONNX_TYPE_MAP, &out)); + return Value{out}; +} + +inline Value Value::CreateSequence(const std::vector& values) { + OrtValue* out; + std::vector values_ort{values.data(), values.data() + values.size()}; + ThrowOnError(GetApi().CreateValue(values_ort.data(), values_ort.size(), ONNX_TYPE_SEQUENCE, &out)); + return Value{out}; +} + +template +inline Value Value::CreateOpaque(const char* domain, const char* type_name, const T& data_container) { + OrtValue* out; + ThrowOnError(GetApi().CreateOpaqueValue(domain, type_name, &data_container, sizeof(T), &out)); + return Value{out}; +} + +// +// Custom OP Inlines +// +inline Logger::Logger(const OrtLogger* logger) : logger_(logger) { + Ort::ThrowOnError(GetApi().Logger_GetLoggingSeverityLevel(this->logger_, &this->cached_severity_level_)); +} + +inline OrtLoggingLevel Logger::GetLoggingSeverityLevel() const noexcept { + return cached_severity_level_; +} + +inline Status Logger::LogMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, int line_number, + const char* func_name, const char* message) const noexcept { + OrtStatus* status = GetApi().Logger_LogMessage(logger_, log_severity_level, message, file_path, line_number, + func_name); + return Status{status}; +} + +// Disable warnings about the format string not being a literal (-Wformat-nonliteral and -Wformat-security) +// for gcc and clang. The alternative is to use actual C-style variadic parameters and apply +// __attribute__(format(printf...)), which does not work with variadic templates. +#if defined(__GNUC__) +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wformat-nonliteral" +#pragma GCC diagnostic ignored "-Wformat-security" +#elif defined(__clang__) +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Wformat-nonliteral" +#pragma clang diagnostic ignored "-Wformat-security" +#endif +template +inline Status Logger::LogFormattedMessage(OrtLoggingLevel log_severity_level, const ORTCHAR_T* file_path, + int line_number, const char* func_name, const char* format, + Args&&... args) const noexcept { + int msg_len = std::snprintf(nullptr, 0U, format, std::forward(args)...); + + if (msg_len < 0) { // Formatting error + return Status("Failed to log message due to formatting error", OrtErrorCode::ORT_FAIL); + } + + OrtStatus* status = nullptr; + const size_t buffer_size = static_cast(msg_len) + 1U; + + constexpr size_t kStackBufferSize = 1024; + + if (buffer_size < kStackBufferSize) { + char buffer[kStackBufferSize]; + snprintf(buffer, kStackBufferSize, format, std::forward(args)...); + status = GetApi().Logger_LogMessage(logger_, log_severity_level, buffer, file_path, line_number, func_name); + } else { + // std::make_unique is only supported starting at C++14. +#if (__cplusplus >= 201402L) || (_MSC_VER >= 1900) + auto buffer = std::make_unique(buffer_size); +#else + std::unique_ptr buffer(new char[buffer_size]); +#endif + std::snprintf(buffer.get(), buffer_size, format, std::forward(args)...); + status = GetApi().Logger_LogMessage(logger_, log_severity_level, buffer.get(), file_path, line_number, func_name); + } + + return Status{status}; +} +// Re-enable -Wformat-nonliteral and -Wformat-security +#if defined(__GNUC__) +#pragma GCC diagnostic pop +#elif defined(__clang__) +#pragma clang diagnostic pop +#endif + +inline KernelContext::KernelContext(OrtKernelContext* context) : ctx_(context) { +} + +inline size_t KernelContext::GetInputCount() const { + size_t out = 0; + Ort::ThrowOnError(GetApi().KernelContext_GetInputCount(ctx_, &out)); + return out; +} + +inline size_t KernelContext::GetOutputCount() const { + size_t out = 0; + Ort::ThrowOnError(GetApi().KernelContext_GetOutputCount(ctx_, &out)); + return out; +} + +inline ConstValue KernelContext::GetInput(size_t index) const { + const OrtValue* out = nullptr; + Ort::ThrowOnError(GetApi().KernelContext_GetInput(ctx_, index, &out)); + return ConstValue{out}; +} + +inline UnownedValue KernelContext::GetOutput(size_t index, const int64_t* dim_values, size_t dim_count) const { + OrtValue* out = nullptr; + Ort::ThrowOnError(GetApi().KernelContext_GetOutput(ctx_, index, dim_values, dim_count, &out)); + return UnownedValue(out); +} + +inline UnownedValue KernelContext::GetOutput(size_t index, const std::vector& dims) const { + OrtValue* out = nullptr; + Ort::ThrowOnError(GetApi().KernelContext_GetOutput(ctx_, index, dims.data(), dims.size(), &out)); + return UnownedValue(out); +} + +inline void* KernelContext::GetGPUComputeStream() const { + void* out = nullptr; + Ort::ThrowOnError(GetApi().KernelContext_GetGPUComputeStream(ctx_, &out)); + return out; +} + +inline OrtAllocator* KernelContext::GetAllocator(const OrtMemoryInfo& memory_info) const { + OrtAllocator* out = nullptr; + Ort::ThrowOnError(GetApi().KernelContext_GetAllocator(ctx_, &memory_info, &out)); + return out; +} + +inline Logger KernelContext::GetLogger() const { + const OrtLogger* out = nullptr; + ThrowOnError(GetApi().KernelContext_GetLogger(this->ctx_, &out)); + return Logger{out}; +} + +inline void KernelContext::ParallelFor(void (*fn)(void*, size_t), size_t total, size_t num_batch, void* usr_data) const { + ThrowOnError(GetApi().KernelContext_ParallelFor(ctx_, fn, total, num_batch, usr_data)); +} + +inline OpAttr::OpAttr(const char* name, const void* data, int len, OrtOpAttrType type) { + Ort::ThrowOnError(GetApi().CreateOpAttr(name, data, len, type, &p_)); +} + +namespace detail { +template +inline KernelInfo KernelInfoImpl::Copy() const { + OrtKernelInfo* info_copy = nullptr; + Ort::ThrowOnError(GetApi().CopyKernelInfo(this->p_, &info_copy)); + return KernelInfo{info_copy}; +} + +template +inline size_t KernelInfoImpl::GetInputCount() const { + size_t out = 0; + ThrowOnError(GetApi().KernelInfo_GetInputCount(this->p_, &out)); + return out; +} + +template +inline size_t KernelInfoImpl::GetOutputCount() const { + size_t out = 0; + ThrowOnError(GetApi().KernelInfo_GetOutputCount(this->p_, &out)); + return out; +} + +template +inline std::string KernelInfoImpl::GetInputName(size_t index) const { + size_t size = 0; + + // Feed nullptr for the data buffer to query the true size of the string value + Ort::ThrowOnError(GetApi().KernelInfo_GetInputName(this->p_, index, nullptr, &size)); + + std::string out; + out.resize(size); + Ort::ThrowOnError(GetApi().KernelInfo_GetInputName(this->p_, index, &out[0], &size)); + out.resize(size - 1); // remove the terminating character '\0' + + return out; +} + +template +inline std::string KernelInfoImpl::GetOutputName(size_t index) const { + size_t size = 0; + + // Feed nullptr for the data buffer to query the true size of the string value + Ort::ThrowOnError(GetApi().KernelInfo_GetOutputName(this->p_, index, nullptr, &size)); + + std::string out; + out.resize(size); + Ort::ThrowOnError(GetApi().KernelInfo_GetOutputName(this->p_, index, &out[0], &size)); + out.resize(size - 1); // remove the terminating character '\0' + + return out; +} + +template +inline TypeInfo KernelInfoImpl::GetInputTypeInfo(size_t index) const { + OrtTypeInfo* out = nullptr; + ThrowOnError(GetApi().KernelInfo_GetInputTypeInfo(this->p_, index, &out)); + return TypeInfo{out}; +} + +template +inline TypeInfo KernelInfoImpl::GetOutputTypeInfo(size_t index) const { + OrtTypeInfo* out = nullptr; + ThrowOnError(GetApi().KernelInfo_GetOutputTypeInfo(this->p_, index, &out)); + return TypeInfo{out}; +} + +template +inline Value KernelInfoImpl::GetTensorAttribute(const char* name, OrtAllocator* allocator) const { + OrtValue* out = nullptr; + ThrowOnError(GetApi().KernelInfoGetAttribute_tensor(this->p_, name, allocator, &out)); + return Value{out}; +} + +template +inline ConstValue KernelInfoImpl::GetTensorConstantInput(size_t index, int* is_constant) const { + const OrtValue* out = nullptr; + ThrowOnError(GetApi().KernelInfoGetConstantInput_tensor(this->p_, index, is_constant, &out)); + return ConstValue{out}; +} + +template +inline std::string KernelInfoImpl::GetNodeName() const { + size_t size = 0; + + // Feed nullptr for the data buffer to query the true size of the string value + Ort::ThrowOnError(GetApi().KernelInfo_GetNodeName(this->p_, nullptr, &size)); + + std::string out; + out.resize(size); + Ort::ThrowOnError(GetApi().KernelInfo_GetNodeName(this->p_, &out[0], &size)); + out.resize(size - 1); // remove the terminating character '\0' + + return out; +} + +template +inline Logger KernelInfoImpl::GetLogger() const { + const OrtLogger* out = nullptr; + ThrowOnError(GetApi().KernelInfo_GetLogger(this->p_, &out)); + return Logger{out}; +} + +inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, float& out) { + Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_float(p, name, &out)); +} + +inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, int64_t& out) { + Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_int64(p, name, &out)); +} + +inline void attr_utils::GetAttr(const OrtKernelInfo* p, const char* name, std::string& result) { + size_t size = 0; + // Feed nullptr for the data buffer to query the true size of the string attribute + Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_string(p, name, nullptr, &size)); + + std::string out; + out.resize(size); + Ort::ThrowOnError(GetApi().KernelInfoGetAttribute_string(p, name, &out[0], &size)); + out.resize(size - 1); // remove the terminating character '\0' + out.swap(result); +} + +inline void attr_utils::GetAttrs(const OrtKernelInfo* p, const char* name, std::vector& result) { + size_t size = 0; + // Feed nullptr for the data buffer to query the true size of the attribute + Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_float(p, name, nullptr, &size)); + + std::vector out; + out.resize(size); + Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_float(p, name, out.data(), &size)); + out.swap(result); +} + +inline void attr_utils::GetAttrs(const OrtKernelInfo* p, const char* name, std::vector& result) { + size_t size = 0; + + // Feed nullptr for the data buffer to query the true size of the attribute + Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_int64(p, name, nullptr, &size)); + + std::vector out; + out.resize(size); + Ort::ThrowOnError(GetApi().KernelInfoGetAttributeArray_int64(p, name, out.data(), &size)); + out.swap(result); +} +} // namespace detail + +inline KernelInfo::KernelInfo(OrtKernelInfo* info) : detail::KernelInfoImpl{info} {} + +inline Op::Op(OrtOp* p) : detail::Base(p) {} + +inline Op Op::Create(const OrtKernelInfo* info, const char* op_name, const char* domain, int version, + const char** type_constraint_names, + const ONNXTensorElementDataType* type_constraint_values, + size_t type_constraint_count, + const OpAttr* attr_values, size_t attr_count, + size_t input_count, size_t output_count) { + static_assert(sizeof(OpAttr) == sizeof(OrtOpAttr*), + "OpAttr's is expected to be just an array of OrtOpAttr in memory so we can reinterpret safely"); + auto attr_input_values = reinterpret_cast(attr_values); + OrtOp* op; + Ort::ThrowOnError(GetApi().CreateOp(info, op_name, domain, version, type_constraint_names, type_constraint_values, + static_cast(type_constraint_count), + attr_input_values, + static_cast(attr_count), + static_cast(input_count), + static_cast(output_count), &op)); + return Op{op}; +} + +inline void Op::Invoke(const OrtKernelContext* context, + const Value* input_values, + size_t input_count, + Value* output_values, + size_t output_count) { + static_assert(sizeof(Value) == sizeof(OrtValue*), + "Value is really just an array of OrtValue* in memory, so we can reinterpret_cast safely"); + auto ort_input_values = reinterpret_cast(input_values); + auto ort_output_values = reinterpret_cast(output_values); + Ort::ThrowOnError(GetApi().InvokeOp(context, p_, ort_input_values, static_cast(input_count), + ort_output_values, static_cast(output_count))); +} + +inline void Op::Invoke(const OrtKernelContext* context, + const OrtValue* const* input_values, + size_t input_count, + OrtValue* const* output_values, + size_t output_count) { + Ort::ThrowOnError(GetApi().InvokeOp(context, p_, input_values, static_cast(input_count), + output_values, static_cast(output_count))); +} + +inline std::string GetVersionString() { + return OrtGetApiBase()->GetVersionString(); +} + +inline std::string GetBuildInfoString() { + return GetApi().GetBuildInfoString(); +} + +inline std::vector GetAvailableProviders() { + char** providers; + int len; + + auto release_fn = [&len](char** providers) { + // This should always return nullptr. + ThrowOnError(GetApi().ReleaseAvailableProviders(providers, len)); + }; + + ThrowOnError(GetApi().GetAvailableProviders(&providers, &len)); + std::unique_ptr guard(providers, release_fn); + std::vector available_providers; + available_providers.reserve(static_cast(len)); + for (int i = 0; i < len; ++i) { + available_providers.emplace_back(providers[i]); + } + return available_providers; +} + +template +void CustomOpBase::GetSessionConfigs(std::unordered_map& out, + ConstSessionOptions options) const { + const TOp* derived = static_cast(this); + std::vector keys = derived->GetSessionConfigKeys(); + + out.reserve(keys.size()); + + std::string config_entry_key = detail::MakeCustomOpConfigEntryKey(derived->GetName(), ""); + const size_t prefix_size = config_entry_key.length(); + + for (const auto& key : keys) { + config_entry_key.resize(prefix_size); + config_entry_key.append(key); + out[key] = options.GetConfigEntryOrDefault(config_entry_key.c_str(), ""); + } +} + +inline ShapeInferContext::ShapeInferContext(const OrtApi* ort_api, + OrtShapeInferContext* ctx) : ort_api_(ort_api), ctx_(ctx) { + size_t input_count = 0; + Ort::ThrowOnError(ort_api_->ShapeInferContext_GetInputCount(ctx_, &input_count)); + for (size_t ith_input = 0; ith_input < input_count; ++ith_input) { + OrtTensorTypeAndShapeInfo* info{}; + Ort::ThrowOnError(ort_api_->ShapeInferContext_GetInputTypeShape(ctx, ith_input, &info)); + TensorTypeAndShapeInfo type_shape_info(info); + auto integer_shape = type_shape_info.GetShape(); + std::vector symbolic_shape(integer_shape.size(), {}); + if (!integer_shape.empty()) { + type_shape_info.GetSymbolicDimensions(&symbolic_shape[0], integer_shape.size()); + } + Shape shape; + for (size_t ith = 0; ith < integer_shape.size(); ++ith) { + if (symbolic_shape[ith] && std::string{symbolic_shape[ith]}.size() > 0) { + shape.emplace_back(symbolic_shape[ith]); + } else { + shape.emplace_back(integer_shape[ith]); + } + } + input_shapes_.push_back(std::move(shape)); + type_shape_info.release(); + } +} + +inline Status ShapeInferContext::SetOutputShape(size_t indice, const Shape& shape, ONNXTensorElementDataType type) { + OrtTensorTypeAndShapeInfo* info = {}; + ORT_CXX_RETURN_ON_API_FAIL(ort_api_->CreateTensorTypeAndShapeInfo(&info)); + ORT_CXX_RETURN_ON_API_FAIL(ort_api_->SetTensorElementType(info, type)); + + using InfoPtr = std::unique_ptr>; + + InfoPtr info_ptr(info, [this](OrtTensorTypeAndShapeInfo* obj) { + ort_api_->ReleaseTensorTypeAndShapeInfo(obj); + }); + + std::vector integer_dims; + std::vector symbolic_dims; + + for (const auto dim : shape) { + if (dim.IsInt()) { + integer_dims.push_back(dim.AsInt()); + symbolic_dims.push_back(""); + } else { + if (!dim.AsSym() || std::string{dim.AsSym()}.empty()) { + ORT_CXX_API_THROW("Symbolic dim must not be an empty string", ORT_INVALID_ARGUMENT); + } + integer_dims.push_back(SymbolicInteger::INVALID_INT_DIM); + symbolic_dims.push_back(dim.AsSym()); + } + } + + ORT_CXX_RETURN_ON_API_FAIL(ort_api_->SetDimensions(info, integer_dims.data(), integer_dims.size())); + ORT_CXX_RETURN_ON_API_FAIL(ort_api_->SetSymbolicDimensions(info, symbolic_dims.data(), symbolic_dims.size())); + ORT_CXX_RETURN_ON_API_FAIL(ort_api_->ShapeInferContext_SetOutputTypeShape(ctx_, indice, info)); + return Status{nullptr}; +} + +inline int64_t ShapeInferContext::GetAttrInt(const char* attr_name) { + const auto* attr = GetAttrHdl(attr_name); + int64_t i = {}; + size_t out = {}; + Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INT, &i, sizeof(i), &out)); + return i; +} + +inline ShapeInferContext::Ints ShapeInferContext::GetAttrInts(const char* attr_name) { + const auto* attr = GetAttrHdl(attr_name); + int64_t i = {}; + size_t out = {}; + // first call to get the bytes needed + // 1. A status == nullptr means that ReadOpAttr was successful. A status != nullptr means failure. + // 2. The ReadOpAttr function should normally be called twice: once to get the needed buffer size (returns a status != nullptr), and a second time to actually read the ints (returns status == null on success). + // 3. This code tries a subtle optimization in the first call to ReadOpAttr. It passes in a buffer (&i) of size 1 just in case there is only 1 int. In this case, status == nullptr and we need to return {i}. + auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INTS, &i, sizeof(i), &out); + if (status) { + size_t num_i = out / sizeof(int64_t); + ShapeInferContext::Ints ints(num_i, 0); + Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INTS, ints.data(), out, &out)); + return ints; + } else { + if (out == 0u) { + return {}; + } + return {i}; + } +} + +inline float ShapeInferContext::GetAttrFloat(const char* attr_name) { + const auto* attr = GetAttrHdl(attr_name); + float f = {}; + size_t out = {}; + Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOAT, &f, sizeof(f), &out)); + return f; +} + +inline ShapeInferContext::Floats ShapeInferContext::GetAttrFloats(const char* attr_name) { + const auto* attr = GetAttrHdl(attr_name); + float f = {}; + size_t out = {}; + // first call to get the bytes needed + // 1. A status == nullptr means that ReadOpAttr was successful. A status != nullptr means failure. + // 2. The ReadOpAttr function should normally be called twice: once to get the needed buffer size (returns a status != nullptr), and a second time to actually read the ints (returns status == null on success). + // 3. This code tries a subtle optimization in the first call to ReadOpAttr. It passes in a buffer (&i) of size 1 just in case there is only 1 int. In this case, status == nullptr and we need to return {i}. + auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOATS, &f, sizeof(f), &out); + if (status) { + size_t num_f = out / sizeof(float); + ShapeInferContext::Floats floats(num_f, 0); + Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOATS, floats.data(), out, &out)); + return floats; + } else { + if (out == 0u) { + return {}; + } + return {f}; + } +} + +inline std::string ShapeInferContext::GetAttrString(const char* attr_name) { + const auto* attr = GetAttrHdl(attr_name); + char c = {}; + size_t out = {}; + // first call to get the bytes needed + auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRING, &c, sizeof(char), &out); + if (status) { + std::vector chars(out, '\0'); + Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRING, chars.data(), out, &out)); + return {chars.data()}; + } else { + return {c}; + } +} + +inline ShapeInferContext::Strings ShapeInferContext::GetAttrStrings(const char* attr_name) { + const auto* attr = GetAttrHdl(attr_name); + char c = {}; + size_t out = {}; + // first call to get the bytes needed + // 1. A status == nullptr means that ReadOpAttr was successful. A status != nullptr means failure. + // 2. The ReadOpAttr function should normally be called twice: once to get the needed buffer size (returns a status != nullptr), and a second time to actually read the ints (returns status == null on success). + // 3. This code tries a subtle optimization in the first call to ReadOpAttr. It passes in a buffer (&i) of size 1 just in case there is only 1 int. In this case, status == nullptr and we need to return {i}. + auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRINGS, &c, sizeof(char), &out); + if (status) { + std::vector chars(out, '\0'); + Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRINGS, chars.data(), out, &out)); + ShapeInferContext::Strings strings; + char* char_st = chars.data(); + char* char_ed = char_st + out; + while (char_st < char_ed) { + strings.emplace_back(char_st); + while (*char_st != '\0') { + char_st++; + } + char_st++; + } + return strings; + } else { + if (out == 0u) { + return {}; + } + return {std::string{c}}; + } +} + +inline const OrtOpAttr* ShapeInferContext::GetAttrHdl(const char* attr_name) const { + const OrtOpAttr* attr_hdl = {}; + Ort::ThrowOnError(ort_api_->ShapeInferContext_GetAttribute(ctx_, attr_name, &attr_hdl)); + return attr_hdl; +} + +namespace detail { +inline std::vector StringsToCharPtrs(const std::vector& strings) { + std::vector ptrs; + ptrs.reserve(strings.size()); + std::transform(strings.begin(), strings.end(), std::back_inserter(ptrs), + [](const std::string& s) { return s.c_str(); }); + + return ptrs; +} +} // namespace detail + +#if !defined(ORT_MINIMAL_BUILD) +// static +inline void Node::Init(const std::string& operator_name, const std::string& operator_domain, + const std::string& node_name, + const std::vector& input_names, + const std::vector& output_names, + std::vector& attributes, + OrtNode*& node) { + auto inputs = detail::StringsToCharPtrs(input_names); + auto outputs = detail::StringsToCharPtrs(output_names); + + std::vector attributes_ptrs; + attributes_ptrs.reserve(attributes.size()); + std::transform(attributes.begin(), attributes.end(), std::back_inserter(attributes_ptrs), + [](OpAttr& attr) -> OrtOpAttr* { return attr; }); + + ThrowOnError(GetModelEditorApi().CreateNode(operator_name.c_str(), operator_domain.c_str(), node_name.c_str(), + inputs.data(), inputs.size(), + outputs.data(), outputs.size(), + attributes_ptrs.data(), attributes_ptrs.size(), + &node)); + + // Node now owns the attributes + std::for_each(attributes.begin(), attributes.end(), [](OpAttr& attr) { attr.release(); }); +} + +inline Node::Node(const std::string& operator_name, const std::string& operator_domain, + const std::string& node_name, + const std::vector& input_names, + const std::vector& output_names, + std::vector& attributes) { + Init(operator_name, operator_domain, node_name, input_names, output_names, attributes, p_); +} + +inline Node::Node(const std::string& operator_name, const std::string& operator_domain, + const std::string& node_name, + const std::vector& input_names, + const std::vector& output_names) { + std::vector empty_attributes; + Init(operator_name, operator_domain, node_name, input_names, output_names, empty_attributes, p_); +} + +inline Graph::Graph() { + ThrowOnError(GetModelEditorApi().CreateGraph(&p_)); +} + +inline Model::Model(const std::vector& opsets) { + std::vector domains; + std::vector versions; + domains.reserve(opsets.size()); + versions.reserve(opsets.size()); + + for (const auto& pair : opsets) { + domains.push_back(pair.first.c_str()); + versions.push_back(pair.second); + } + + ThrowOnError(GetModelEditorApi().CreateModel(domains.data(), versions.data(), opsets.size(), &p_)); +} + +inline ValueInfo::ValueInfo(const std::string& name, const ConstTypeInfo& type_info) { + ThrowOnError(GetModelEditorApi().CreateValueInfo(name.c_str(), type_info, &p_)); +} +#endif // !defined(ORT_MINIMAL_BUILD) + +namespace detail { +template <> +inline std::string ValueInfoImpl::Name() const { + const char* name = nullptr; + ThrowOnError(GetApi().GetValueInfoName(this->p_, &name)); + return name; +} + +template <> +inline ConstTypeInfo ValueInfoImpl::TypeInfo() const { + const OrtTypeInfo* type_info = nullptr; + ThrowOnError(GetApi().GetValueInfoTypeInfo(this->p_, &type_info)); + return ConstTypeInfo{type_info}; +} + +#if !defined(ORT_MINIMAL_BUILD) +template <> +inline void GraphImpl::SetInputs(std::vector& inputs) { + std::vector inputs_ptrs; + inputs_ptrs.reserve(inputs.size()); + std::transform(inputs.begin(), inputs.end(), std::back_inserter(inputs_ptrs), + [](ValueInfo& vi) -> OrtValueInfo* { return vi; }); + + ThrowOnError(GetModelEditorApi().SetGraphInputs(p_, inputs_ptrs.data(), inputs_ptrs.size())); + + // Graph now owns the inputs + std::for_each(inputs.begin(), inputs.end(), [](ValueInfo& vi) { vi.release(); }); +} + +template <> +inline void GraphImpl::SetOutputs(std::vector& outputs) { + std::vector outputs_ptrs; + outputs_ptrs.reserve(outputs.size()); + std::transform(outputs.begin(), outputs.end(), std::back_inserter(outputs_ptrs), + [](ValueInfo& vi) -> OrtValueInfo* { return vi; }); + + ThrowOnError(GetModelEditorApi().SetGraphOutputs(p_, outputs_ptrs.data(), outputs_ptrs.size())); + + // Graph now owns the outputs + std::for_each(outputs.begin(), outputs.end(), [](ValueInfo& vi) { vi.release(); }); +} + +template <> +inline void GraphImpl::AddInitializer(const std::string& name, Value& initializer, bool data_is_external) { + // Graph takes ownership of `initializer` + ThrowOnError(GetModelEditorApi().AddInitializerToGraph(p_, name.c_str(), initializer.release(), data_is_external)); +} + +template <> +inline void GraphImpl::AddNode(Node& node) { + // Graph takes ownership of `node` + ThrowOnError(GetModelEditorApi().AddNodeToGraph(p_, node.release())); +} + +template <> +inline void ModelImpl::AddGraph(Graph& graph) { + // Model takes ownership of `graph` + ThrowOnError(GetModelEditorApi().AddGraphToModel(p_, graph.release())); +} +#endif // !defined(ORT_MINIMAL_BUILD) + +} // namespace detail +} // namespace Ort diff --git a/third_party/onnxruntime/include/onnxruntime_float16.h b/third_party/onnxruntime/include/onnxruntime_float16.h new file mode 100644 index 00000000..408d3ccf --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_float16.h @@ -0,0 +1,535 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#pragma once + +#include +#include +#include +#include + +namespace onnxruntime_float16 { + +namespace detail { + +enum class endian { +#if defined(_WIN32) + little = 0, + big = 1, + native = little, +#elif defined(__GNUC__) || defined(__clang__) + little = __ORDER_LITTLE_ENDIAN__, + big = __ORDER_BIG_ENDIAN__, + native = __BYTE_ORDER__, +#else +#error onnxruntime_float16::detail::endian is not implemented in this environment. +#endif +}; + +static_assert( + endian::native == endian::little || endian::native == endian::big, + "Only little-endian or big-endian native byte orders are supported."); + +} // namespace detail + +/// +/// Shared implementation between public and internal classes. CRTP pattern. +/// +template +struct Float16Impl { + protected: + /// + /// Converts from float to uint16_t float16 representation + /// + /// + /// + constexpr static uint16_t ToUint16Impl(float v) noexcept; + + /// + /// Converts float16 to float + /// + /// float representation of float16 value + float ToFloatImpl() const noexcept; + + /// + /// Creates an instance that represents absolute value. + /// + /// Absolute value + uint16_t AbsImpl() const noexcept { + return static_cast(val & ~kSignMask); + } + + /// + /// Creates a new instance with the sign flipped. + /// + /// Flipped sign instance + uint16_t NegateImpl() const noexcept { + return IsNaN() ? val : static_cast(val ^ kSignMask); + } + + public: + // uint16_t special values + static constexpr uint16_t kSignMask = 0x8000U; + static constexpr uint16_t kBiasedExponentMask = 0x7C00U; + static constexpr uint16_t kPositiveInfinityBits = 0x7C00U; + static constexpr uint16_t kNegativeInfinityBits = 0xFC00U; + static constexpr uint16_t kPositiveQNaNBits = 0x7E00U; + static constexpr uint16_t kNegativeQNaNBits = 0xFE00U; + static constexpr uint16_t kMaxValueBits = 0x7BFFU; // Largest normal number + static constexpr uint16_t kOneBits = 0x3C00U; + static constexpr uint16_t kMinusOneBits = 0xBC00U; + + uint16_t val{0}; + + Float16Impl() = default; + + /// + /// Checks if the value is negative + /// + /// true if negative + bool IsNegative() const noexcept { + return static_cast(val) < 0; + } + + /// + /// Tests if the value is NaN + /// + /// true if NaN + bool IsNaN() const noexcept { + return AbsImpl() > kPositiveInfinityBits; + } + + /// + /// Tests if the value is finite + /// + /// true if finite + bool IsFinite() const noexcept { + return AbsImpl() < kPositiveInfinityBits; + } + + /// + /// Tests if the value represents positive infinity. + /// + /// true if positive infinity + bool IsPositiveInfinity() const noexcept { + return val == kPositiveInfinityBits; + } + + /// + /// Tests if the value represents negative infinity + /// + /// true if negative infinity + bool IsNegativeInfinity() const noexcept { + return val == kNegativeInfinityBits; + } + + /// + /// Tests if the value is either positive or negative infinity. + /// + /// True if absolute value is infinity + bool IsInfinity() const noexcept { + return AbsImpl() == kPositiveInfinityBits; + } + + /// + /// Tests if the value is NaN or zero. Useful for comparisons. + /// + /// True if NaN or zero. + bool IsNaNOrZero() const noexcept { + auto abs = AbsImpl(); + return (abs == 0 || abs > kPositiveInfinityBits); + } + + /// + /// Tests if the value is normal (not zero, subnormal, infinite, or NaN). + /// + /// True if so + bool IsNormal() const noexcept { + auto abs = AbsImpl(); + return (abs < kPositiveInfinityBits) // is finite + && (abs != 0) // is not zero + && ((abs & kBiasedExponentMask) != 0); // is not subnormal (has a non-zero exponent) + } + + /// + /// Tests if the value is subnormal (denormal). + /// + /// True if so + bool IsSubnormal() const noexcept { + auto abs = AbsImpl(); + return (abs < kPositiveInfinityBits) // is finite + && (abs != 0) // is not zero + && ((abs & kBiasedExponentMask) == 0); // is subnormal (has a zero exponent) + } + + /// + /// Creates an instance that represents absolute value. + /// + /// Absolute value + Derived Abs() const noexcept { return Derived::FromBits(AbsImpl()); } + + /// + /// Creates a new instance with the sign flipped. + /// + /// Flipped sign instance + Derived Negate() const noexcept { return Derived::FromBits(NegateImpl()); } + + /// + /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check + /// for two values by or'ing the private bits together and stripping the sign. They are both zero, + /// and therefore equivalent, if the resulting value is still zero. + /// + /// first value + /// second value + /// True if both arguments represent zero + static bool AreZero(const Float16Impl& lhs, const Float16Impl& rhs) noexcept { + return static_cast((lhs.val | rhs.val) & ~kSignMask) == 0; + } + + bool operator==(const Float16Impl& rhs) const noexcept { + if (IsNaN() || rhs.IsNaN()) { + // IEEE defines that NaN is not equal to anything, including itself. + return false; + } + return val == rhs.val; + } + + bool operator!=(const Float16Impl& rhs) const noexcept { return !(*this == rhs); } + + bool operator<(const Float16Impl& rhs) const noexcept { + if (IsNaN() || rhs.IsNaN()) { + // IEEE defines that NaN is unordered with respect to everything, including itself. + return false; + } + + const bool left_is_negative = IsNegative(); + if (left_is_negative != rhs.IsNegative()) { + // When the signs of left and right differ, we know that left is less than right if it is + // the negative value. The exception to this is if both values are zero, in which case IEEE + // says they should be equal, even if the signs differ. + return left_is_negative && !AreZero(*this, rhs); + } + return (val != rhs.val) && ((val < rhs.val) ^ left_is_negative); + } +}; + +// The following Float16_t conversions are based on the code from +// Eigen library. + +// The conversion routines are Copyright (c) Fabian Giesen, 2016. +// The original license follows: +// +// Copyright (c) Fabian Giesen, 2016 +// All rights reserved. +// Redistribution and use in source and binary forms, with or without +// modification, are permitted. +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR +// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT +// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, +// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT +// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, +// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY +// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT +// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + +namespace detail { +union float32_bits { + unsigned int u; + float f; +}; +} // namespace detail + +template +inline constexpr uint16_t Float16Impl::ToUint16Impl(float v) noexcept { + detail::float32_bits f{}; + f.f = v; + + constexpr detail::float32_bits f32infty = {255 << 23}; + constexpr detail::float32_bits f16max = {(127 + 16) << 23}; + constexpr detail::float32_bits denorm_magic = {((127 - 15) + (23 - 10) + 1) << 23}; + constexpr unsigned int sign_mask = 0x80000000u; + uint16_t val = static_cast(0x0u); + + unsigned int sign = f.u & sign_mask; + f.u ^= sign; + + // NOTE all the integer compares in this function can be safely + // compiled into signed compares since all operands are below + // 0x80000000. Important if you want fast straight SSE2 code + // (since there's no unsigned PCMPGTD). + + if (f.u >= f16max.u) { // result is Inf or NaN (all exponent bits set) + val = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf + } else { // (De)normalized number or zero + if (f.u < (113 << 23)) { // resulting FP16 is subnormal or zero + // use a magic value to align our 10 mantissa bits at the bottom of + // the float. as long as FP addition is round-to-nearest-even this + // just works. + f.f += denorm_magic.f; + + // and one integer subtract of the bias later, we have our final float! + val = static_cast(f.u - denorm_magic.u); + } else { + unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd + + // update exponent, rounding bias part 1 + // Equivalent to `f.u += ((unsigned int)(15 - 127) << 23) + 0xfff`, but + // without arithmetic overflow. + f.u += 0xc8000fffU; + // rounding bias part 2 + f.u += mant_odd; + // take the bits! + val = static_cast(f.u >> 13); + } + } + + val |= static_cast(sign >> 16); + return val; +} + +template +inline float Float16Impl::ToFloatImpl() const noexcept { + constexpr detail::float32_bits magic = {113 << 23}; + constexpr unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift + detail::float32_bits o{}; + + o.u = (val & 0x7fff) << 13; // exponent/mantissa bits + unsigned int exp = shifted_exp & o.u; // just the exponent + o.u += (127 - 15) << 23; // exponent adjust + + // handle exponent special cases + if (exp == shifted_exp) { // Inf/NaN? + o.u += (128 - 16) << 23; // extra exp adjust + } else if (exp == 0) { // Zero/Denormal? + o.u += 1 << 23; // extra exp adjust + o.f -= magic.f; // re-normalize + } + + // Attempt to workaround the Internal Compiler Error on ARM64 + // for bitwise | operator, including std::bitset +#if (defined _MSC_VER) && (defined _M_ARM || defined _M_ARM64 || defined _M_ARM64EC) + if (IsNegative()) { + return -o.f; + } +#else + // original code: + o.u |= (val & 0x8000U) << 16U; // sign bit +#endif + return o.f; +} + +/// Shared implementation between public and internal classes. CRTP pattern. +template +struct BFloat16Impl { + protected: + /// + /// Converts from float to uint16_t float16 representation + /// + /// + /// + static uint16_t ToUint16Impl(float v) noexcept; + + /// + /// Converts bfloat16 to float + /// + /// float representation of bfloat16 value + float ToFloatImpl() const noexcept; + + /// + /// Creates an instance that represents absolute value. + /// + /// Absolute value + uint16_t AbsImpl() const noexcept { + return static_cast(val & ~kSignMask); + } + + /// + /// Creates a new instance with the sign flipped. + /// + /// Flipped sign instance + uint16_t NegateImpl() const noexcept { + return IsNaN() ? val : static_cast(val ^ kSignMask); + } + + public: + // uint16_t special values + static constexpr uint16_t kSignMask = 0x8000U; + static constexpr uint16_t kBiasedExponentMask = 0x7F80U; + static constexpr uint16_t kPositiveInfinityBits = 0x7F80U; + static constexpr uint16_t kNegativeInfinityBits = 0xFF80U; + static constexpr uint16_t kPositiveQNaNBits = 0x7FC1U; + static constexpr uint16_t kNegativeQNaNBits = 0xFFC1U; + static constexpr uint16_t kMaxValueBits = 0x7F7FU; + static constexpr uint16_t kRoundToNearest = 0x7FFFU; + static constexpr uint16_t kOneBits = 0x3F80U; + static constexpr uint16_t kMinusOneBits = 0xBF80U; + + uint16_t val{0}; + + BFloat16Impl() = default; + + /// + /// Checks if the value is negative + /// + /// true if negative + bool IsNegative() const noexcept { + return static_cast(val) < 0; + } + + /// + /// Tests if the value is NaN + /// + /// true if NaN + bool IsNaN() const noexcept { + return AbsImpl() > kPositiveInfinityBits; + } + + /// + /// Tests if the value is finite + /// + /// true if finite + bool IsFinite() const noexcept { + return AbsImpl() < kPositiveInfinityBits; + } + + /// + /// Tests if the value represents positive infinity. + /// + /// true if positive infinity + bool IsPositiveInfinity() const noexcept { + return val == kPositiveInfinityBits; + } + + /// + /// Tests if the value represents negative infinity + /// + /// true if negative infinity + bool IsNegativeInfinity() const noexcept { + return val == kNegativeInfinityBits; + } + + /// + /// Tests if the value is either positive or negative infinity. + /// + /// True if absolute value is infinity + bool IsInfinity() const noexcept { + return AbsImpl() == kPositiveInfinityBits; + } + + /// + /// Tests if the value is NaN or zero. Useful for comparisons. + /// + /// True if NaN or zero. + bool IsNaNOrZero() const noexcept { + auto abs = AbsImpl(); + return (abs == 0 || abs > kPositiveInfinityBits); + } + + /// + /// Tests if the value is normal (not zero, subnormal, infinite, or NaN). + /// + /// True if so + bool IsNormal() const noexcept { + auto abs = AbsImpl(); + return (abs < kPositiveInfinityBits) // is finite + && (abs != 0) // is not zero + && ((abs & kBiasedExponentMask) != 0); // is not subnormal (has a non-zero exponent) + } + + /// + /// Tests if the value is subnormal (denormal). + /// + /// True if so + bool IsSubnormal() const noexcept { + auto abs = AbsImpl(); + return (abs < kPositiveInfinityBits) // is finite + && (abs != 0) // is not zero + && ((abs & kBiasedExponentMask) == 0); // is subnormal (has a zero exponent) + } + + /// + /// Creates an instance that represents absolute value. + /// + /// Absolute value + Derived Abs() const noexcept { return Derived::FromBits(AbsImpl()); } + + /// + /// Creates a new instance with the sign flipped. + /// + /// Flipped sign instance + Derived Negate() const noexcept { return Derived::FromBits(NegateImpl()); } + + /// + /// IEEE defines that positive and negative zero are equal, this gives us a quick equality check + /// for two values by or'ing the private bits together and stripping the sign. They are both zero, + /// and therefore equivalent, if the resulting value is still zero. + /// + /// first value + /// second value + /// True if both arguments represent zero + static bool AreZero(const BFloat16Impl& lhs, const BFloat16Impl& rhs) noexcept { + // IEEE defines that positive and negative zero are equal, this gives us a quick equality check + // for two values by or'ing the private bits together and stripping the sign. They are both zero, + // and therefore equivalent, if the resulting value is still zero. + return static_cast((lhs.val | rhs.val) & ~kSignMask) == 0; + } +}; + +template +inline uint16_t BFloat16Impl::ToUint16Impl(float v) noexcept { + uint16_t result; + if (std::isnan(v)) { + result = kPositiveQNaNBits; + } else { + auto get_msb_half = [](float fl) { + uint16_t result; +#ifdef __cpp_if_constexpr + if constexpr (detail::endian::native == detail::endian::little) { +#else + if (detail::endian::native == detail::endian::little) { +#endif + std::memcpy(&result, reinterpret_cast(&fl) + sizeof(uint16_t), sizeof(uint16_t)); + } else { + std::memcpy(&result, &fl, sizeof(uint16_t)); + } + return result; + }; + + uint16_t upper_bits = get_msb_half(v); + union { + uint32_t U32; + float F32; + }; + F32 = v; + U32 += (upper_bits & 1) + kRoundToNearest; + result = get_msb_half(F32); + } + return result; +} + +template +inline float BFloat16Impl::ToFloatImpl() const noexcept { + if (IsNaN()) { + return std::numeric_limits::quiet_NaN(); + } + float result; + char* const first = reinterpret_cast(&result); + char* const second = first + sizeof(uint16_t); +#ifdef __cpp_if_constexpr + if constexpr (detail::endian::native == detail::endian::little) { +#else + if (detail::endian::native == detail::endian::little) { +#endif + std::memset(first, 0, sizeof(uint16_t)); + std::memcpy(second, &val, sizeof(uint16_t)); + } else { + std::memcpy(first, &val, sizeof(uint16_t)); + std::memset(second, 0, sizeof(uint16_t)); + } + return result; +} + +} // namespace onnxruntime_float16 diff --git a/third_party/onnxruntime/include/onnxruntime_lite_custom_op.h b/third_party/onnxruntime/include/onnxruntime_lite_custom_op.h new file mode 100644 index 00000000..ce87d8c5 --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_lite_custom_op.h @@ -0,0 +1,1119 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +// Summary +// The header has APIs to save custom op authors the trouble of defining schemas, +// which will be inferred by functions' signature, as long as their argument list has types supported here. +// Input could be: +// 1. Tensor of onnx data types. +// 2. Span of onnx data types. +// 3. Scalar of onnx data types. +// A input could be optional if indicated as std::optional<...>. +// For an output, it must be a tensor of onnx data types. +// Further, the header also has utility for a simple custom struct, where resources could be kept, to be registered as a custom op. +// For concrete examples, please search keyword "LiteCustomOpTest" under "/onnxruntime/test/". +// Note - all APIs in this header are ABI. + +#pragma once +#include "onnxruntime_cxx_api.h" +#include +#include +#include +#include + +namespace Ort { +namespace Custom { + +class ArgBase { + public: + ArgBase(OrtKernelContext* ctx, + size_t indice, + bool is_input) : ctx_(ctx), indice_(indice), is_input_(is_input) {} + virtual ~ArgBase() {}; + + protected: + struct KernelContext ctx_; + size_t indice_; + bool is_input_; +}; + +using ArgPtr = std::unique_ptr; +using ArgPtrs = std::vector; + +class TensorBase : public ArgBase { + public: + TensorBase(OrtKernelContext* ctx, + size_t indice, + bool is_input) : ArgBase(ctx, indice, is_input) {} + + operator bool() const { + return shape_.has_value(); + } + + const std::vector& Shape() const { + if (!shape_.has_value()) { + ORT_CXX_API_THROW("tensor shape is not yet initialized", OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + return shape_.value(); + } + + ONNXTensorElementDataType Type() const { + return type_; + } + + int64_t NumberOfElement() const { + if (shape_.has_value()) { + return std::accumulate(shape_->begin(), shape_->end(), 1LL, std::multiplies()); + } else { + return 0; + } + } + + std::string Shape2Str() const { + if (shape_.has_value()) { + std::string shape_str; + for (const auto& dim : *shape_) { + shape_str.append(std::to_string(dim)); + shape_str.append(", "); + } + return shape_str; + } else { + return "empty"; + } + } + + bool IsCpuTensor() const { + return strcmp("Cpu", mem_type_) == 0; + } + + virtual const void* DataRaw() const = 0; + virtual size_t SizeInBytes() const = 0; + + protected: + std::optional> shape_; + ONNXTensorElementDataType type_ = ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED; + const char* mem_type_ = "Cpu"; +}; + +template +struct Span { + const T* data_ = {}; + size_t size_ = {}; + void Assign(const T* data, size_t size) { + data_ = data; + size_ = size; + } + size_t size() const { return size_; } + T operator[](size_t indice) const { + return data_[indice]; + } + const T* data() const { return data_; } +}; + +template +class Tensor : public TensorBase { + public: + using TT = typename std::remove_reference::type; + Tensor(OrtKernelContext* ctx, size_t indice, bool is_input) : TensorBase(ctx, indice, is_input) { + if (is_input_) { + if (indice >= ctx_.GetInputCount()) { + ORT_CXX_API_THROW("invalid indice for Ort::Custom::Tensor", OrtErrorCode::ORT_INVALID_ARGUMENT); + } + const_value_ = ctx_.GetInput(indice); + auto type_shape_info = const_value_.GetTensorTypeAndShapeInfo(); + shape_ = type_shape_info.GetShape(); + } + } + const TT* Data() const { + return reinterpret_cast(const_value_.GetTensorRawData()); + } + TT* Allocate(const std::vector& shape) { + shape_ = shape; + if (!data_) { + shape_ = shape; + data_ = ctx_.GetOutput(indice_, shape).template GetTensorMutableData(); + } + return data_; + } + static TT GetT() { return (TT)0; } + const Span& AsSpan() { + if (!shape_.has_value() || shape_->size() != 1) { + ORT_CXX_API_THROW("invalid shape while trying to get a span out of Ort::Custom::Tensor", + OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + span_.Assign(Data(), static_cast((*shape_)[0])); + return span_; + } + const T& AsScalar() { + if (!shape_.has_value() || shape_->size() != 1 || (*shape_)[0] != 1) { + ORT_CXX_API_THROW("invalid shape while trying to get a scalar from Ort::Custom::Tensor", + OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + return *Data(); + } + const void* DataRaw() const override { + return reinterpret_cast(Data()); + } + + size_t SizeInBytes() const override { + return sizeof(TT) * static_cast(NumberOfElement()); + } + + private: + ConstValue const_value_; // for input + TT* data_{}; // for output + Span span_; +}; + +template <> +class Tensor : public TensorBase { + public: + using strings = std::vector; + + Tensor(OrtKernelContext* ctx, size_t indice, bool is_input) : TensorBase(ctx, indice, is_input) { + if (is_input_) { + if (indice >= ctx_.GetInputCount()) { + ORT_CXX_API_THROW("invalid indice for Ort::Custom::Tensor", OrtErrorCode::ORT_INVALID_ARGUMENT); + } + auto const_value = ctx_.GetInput(indice); + auto type_shape_info = const_value.GetTensorTypeAndShapeInfo(); + shape_ = type_shape_info.GetShape(); + auto num_chars = const_value.GetStringTensorDataLength(); + // note - there will be copy ... + auto num_strings = static_cast(NumberOfElement()); + if (num_strings) { + std::vector chars(num_chars + 1, '\0'); + std::vector offsets(num_strings); + const_value.GetStringTensorContent(static_cast(chars.data()), num_chars, offsets.data(), offsets.size()); + auto upper_bound = num_strings - 1; + input_strings_.resize(num_strings); + for (size_t i = upper_bound;; --i) { + if (i < upper_bound) { + chars[offsets[i + 1]] = '\0'; + } + input_strings_[i] = chars.data() + offsets[i]; + if (0 == i) { + break; + } + } + } + } + } + const strings& Data() const { + return input_strings_; + } + const void* DataRaw() const override { + if (input_strings_.size() != 1) { + ORT_CXX_API_THROW("DataRaw() only applies to string scalar", ORT_RUNTIME_EXCEPTION); + } + return reinterpret_cast(input_strings_[0].c_str()); + } + size_t SizeInBytes() const override { + if (input_strings_.size() != 1) { + ORT_CXX_API_THROW("SizeInBytes() only applies to string scalar", ORT_RUNTIME_EXCEPTION); + } + return input_strings_[0].size(); + } + void SetStringOutput(const strings& ss, const std::vector& dims) { + shape_ = dims; + std::vector raw; + for (const auto& s : ss) { + raw.push_back(s.data()); + } + auto output = ctx_.GetOutput(indice_, dims.data(), dims.size()); + // note - there will be copy ... + output.FillStringTensor(raw.data(), raw.size()); + } + const Span& AsSpan() { + ORT_CXX_API_THROW("span for TensorT of string not implemented", OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + const std::string& AsScalar() { + if (input_strings_.size() != 1) { + ORT_CXX_API_THROW("invalid shape while trying to get a scalar string from Ort::Custom::Tensor", + OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + return input_strings_[0]; + } + + private: + std::vector input_strings_; // for input +}; + +template <> +class Tensor : public TensorBase { + public: + using strings = std::vector; + using string_views = std::vector; + + Tensor(OrtKernelContext* ctx, size_t indice, bool is_input) : TensorBase(ctx, indice, is_input) { + if (is_input_) { + if (indice >= ctx_.GetInputCount()) { + ORT_CXX_API_THROW("invalid indice for Ort::Custom::Tensor", OrtErrorCode::ORT_INVALID_ARGUMENT); + } + auto const_value = ctx_.GetInput(indice); + auto type_shape_info = const_value.GetTensorTypeAndShapeInfo(); + shape_ = type_shape_info.GetShape(); + auto num_chars = const_value.GetStringTensorDataLength(); + chars_.resize(num_chars + 1, '\0'); + auto num_strings = static_cast(NumberOfElement()); + if (num_strings) { + std::vector offsets(num_strings); + const_value.GetStringTensorContent(static_cast(chars_.data()), num_chars, offsets.data(), offsets.size()); + offsets.push_back(num_chars); + for (size_t i = 0; i < num_strings; ++i) { + input_string_views_.emplace_back(chars_.data() + offsets[i], offsets[i + 1] - offsets[i]); + } + } + } + } + const string_views& Data() const { + return input_string_views_; + } + const void* DataRaw() const override { + if (input_string_views_.size() != 1) { + ORT_CXX_API_THROW("DataRaw() only applies to string scalar", ORT_RUNTIME_EXCEPTION); + } + return reinterpret_cast(input_string_views_[0].data()); + } + size_t SizeInBytes() const override { + if (input_string_views_.size() != 1) { + ORT_CXX_API_THROW("SizeInBytes() only applies to string scalar", ORT_RUNTIME_EXCEPTION); + } + return input_string_views_[0].size(); + } + void SetStringOutput(const strings& ss, const std::vector& dims) { + shape_ = dims; + std::vector raw; + for (const auto& s : ss) { + raw.push_back(s.data()); + } + auto output = ctx_.GetOutput(indice_, dims.data(), dims.size()); + // note - there will be copy ... + output.FillStringTensor(raw.data(), raw.size()); + } + const Span& AsSpan() { + ORT_CXX_API_THROW("span for TensorT of string view not implemented", OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + std::string_view AsScalar() { + if (input_string_views_.size() != 1) { + ORT_CXX_API_THROW("invalid shape while trying to get a scalar string view from Ort::Custom::Tensor", + OrtErrorCode::ORT_RUNTIME_EXCEPTION); + } + return input_string_views_[0]; + } + + private: + std::vector chars_; // for input + std::vector input_string_views_; // for input +}; + +using TensorPtr = std::unique_ptr; +using TensorPtrs = std::vector; + +struct TensorArray : public ArgBase { + TensorArray(OrtKernelContext* ctx, + size_t start_indice, + bool is_input) : ArgBase(ctx, + start_indice, + is_input) { + if (is_input) { + auto input_count = ctx_.GetInputCount(); + for (size_t ith_input = start_indice; ith_input < input_count; ++ith_input) { + auto const_value = ctx_.GetInput(start_indice); + auto type_shape_info = const_value.GetTensorTypeAndShapeInfo(); + auto type = type_shape_info.GetElementType(); + TensorPtr tensor; + switch (type) { + case ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64: + tensor = std::make_unique>(ctx, ith_input, true); + break; + case ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING: + tensor = std::make_unique>(ctx, ith_input, true); + break; + default: + ORT_CXX_API_THROW("unknow input type", ORT_RUNTIME_EXCEPTION); + break; + } + tensors_.emplace_back(tensor.release()); + } // for + } + } + template + T* AllocateOutput(size_t ith_output, const std::vector& shape) { + // ith_output is the indice of output relative to the tensor array + // indice_ + ith_output is the indice relative to context + auto tensor = std::make_unique>(ctx_.GetOrtKernelContext(), indice_ + ith_output, false); + auto raw_output = tensor.get()->Allocate(shape); + tensors_.emplace_back(tensor.release()); + return raw_output; + } + Tensor& AllocateStringTensor(size_t ith_output) { + // ith_output is the indice of output relative to the tensor array + // indice_ + ith_output is the indice relative to context + auto tensor = std::make_unique>(ctx_.GetOrtKernelContext(), indice_ + ith_output, false); + Tensor& output = *tensor; + tensors_.emplace_back(tensor.release()); + return output; + } + size_t Size() const { + return tensors_.size(); + } + const TensorPtr& operator[](size_t ith_input) const { + // ith_input is the indice of output relative to the tensor array + return tensors_.at(ith_input); + } + + private: + TensorPtrs tensors_; +}; + +using Variadic = TensorArray; + +/* +Note: +OrtLiteCustomOp inherits from OrtCustomOp to bridge tween a custom func/struct and ort core. +The lifetime of an OrtLiteCustomOp instance is managed by customer code, not ort, so: +1. DO NOT cast OrtLiteCustomOp to OrtCustomOp and release since there is no virtual destructor in the hierarchy. +2. OrtLiteCustomFunc and OrtLiteCustomStruct, as two sub-structs, can be released in form of OrtLiteCustomOp since all members are kept in the OrtLiteCustomOp, + hence memory could still be recycled properly. +Further, OrtCustomOp is a c struct bearing no v-table, so offspring structs are by design to be of zero virtual functions to maintain cast safety. +*/ +struct OrtLiteCustomOp : public OrtCustomOp { + using ConstOptionalFloatTensor = std::optional&>; + using OptionalFloatTensor = std::optional>; + + // CreateTuple + template + static typename std::enable_if>::type + CreateTuple(OrtKernelContext*, ArgPtrs&, size_t, size_t, const std::string&) { + return std::make_tuple(); + } + + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + std::tuple current = std::tuple{context}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } + + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + std::tuple current = std::tuple{*context}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } + +#ifdef ORT_CUDA_CTX + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + thread_local CudaContext cuda_context; + cuda_context.Init(*context); + std::tuple current = std::tuple{cuda_context}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } +#endif + +#ifdef ORT_ROCM_CTX + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + thread_local RocmContext rocm_context; + rocm_context.Init(*context); + std::tuple current = std::tuple{rocm_context}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } +#endif + + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + args.push_back(std::make_unique(context, ith_input, true)); + std::tuple current = std::tuple{reinterpret_cast(args.back().get())}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } + + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + args.push_back(std::make_unique(context, ith_input, true)); + std::tuple current = std::tuple{reinterpret_cast(*args.back().get())}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } + + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + args.push_back(std::make_unique(context, ith_output, false)); + std::tuple current = std::tuple{reinterpret_cast(args.back().get())}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } + + template + static typename std::enable_if::value, std::tuple>::type + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { + args.push_back(std::make_unique(context, ith_output, false)); + std::tuple current = std::tuple{reinterpret_cast(*args.back().get())}; + auto next = CreateTuple(context, args, num_input, num_output, ep); + return std::tuple_cat(current, next); + } + +#define CREATE_TUPLE_INPUT(data_type) \ + template \ + static typename std::enable_if*>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{reinterpret_cast(args.back().get())}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if&>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{reinterpret_cast(*args.back().get())}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if*>>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if (ith_input < num_input) { \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{reinterpret_cast*>(args.back().get())}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } else { \ + std::tuple current = std::tuple{}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + } \ + template \ + static typename std::enable_if*>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if ("CPUExecutionProvider" != ep) { \ + ORT_CXX_API_THROW("span input could only be applied to CPU EP", OrtErrorCode::ORT_RUNTIME_EXCEPTION); \ + } \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{&reinterpret_cast*>(args.back().get())->AsSpan()}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if&>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if ("CPUExecutionProvider" != ep) { \ + ORT_CXX_API_THROW("span input could only be applied to CPU EP", OrtErrorCode::ORT_RUNTIME_EXCEPTION); \ + } \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{reinterpret_cast*>(args.back().get())->AsSpan()}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if*>>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if (ith_input < num_input) { \ + if ("CPUExecutionProvider" != ep) { \ + ORT_CXX_API_THROW("span input could only be applied to CPU EP", OrtErrorCode::ORT_RUNTIME_EXCEPTION); \ + } \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{&reinterpret_cast*>(args.back().get())->AsSpan()}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } else { \ + std::tuple current = std::tuple{}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + } \ + template \ + static typename std::enable_if::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if ("CPUExecutionProvider" != ep) { \ + ORT_CXX_API_THROW("scalar input could only be applied to CPU EP", OrtErrorCode::ORT_RUNTIME_EXCEPTION); \ + } \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{reinterpret_cast*>(args.back().get())->AsScalar()}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if (ith_input < num_input) { \ + if ("CPUExecutionProvider" != ep) { \ + ORT_CXX_API_THROW("scalar input could only be applied to CPU EP", OrtErrorCode::ORT_RUNTIME_EXCEPTION); \ + } \ + args.push_back(std::make_unique>(context, ith_input, true)); \ + std::tuple current = std::tuple{reinterpret_cast*>(args.back().get())->AsScalar()}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } else { \ + std::tuple current = std::tuple{}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + } +#define CREATE_TUPLE_OUTPUT(data_type) \ + template \ + static typename std::enable_if*>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + args.push_back(std::make_unique>(context, ith_output, false)); \ + std::tuple current = std::tuple{reinterpret_cast(args.back().get())}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if&>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + args.push_back(std::make_unique>(context, ith_output, false)); \ + std::tuple current = std::tuple{reinterpret_cast(*args.back().get())}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + template \ + static typename std::enable_if*>>::value, std::tuple>::type \ + CreateTuple(OrtKernelContext* context, ArgPtrs& args, size_t num_input, size_t num_output, const std::string& ep) { \ + if (ith_output < num_output) { \ + args.push_back(std::make_unique>(context, ith_output, false)); \ + std::tuple current = std::tuple{reinterpret_cast*>(args.back().get())}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } else { \ + std::tuple current = std::tuple{}; \ + auto next = CreateTuple(context, args, num_input, num_output, ep); \ + return std::tuple_cat(current, next); \ + } \ + } +#define CREATE_TUPLE(data_type) \ + CREATE_TUPLE_INPUT(data_type) \ + CREATE_TUPLE_OUTPUT(data_type) + + CREATE_TUPLE(bool) + CREATE_TUPLE(float) + CREATE_TUPLE(Ort::Float16_t) + CREATE_TUPLE(Ort::BFloat16_t) + CREATE_TUPLE(double) + CREATE_TUPLE(int8_t) + CREATE_TUPLE(int16_t) + CREATE_TUPLE(int32_t) + CREATE_TUPLE(int64_t) + CREATE_TUPLE(uint8_t) + CREATE_TUPLE(uint16_t) + CREATE_TUPLE(uint32_t) + CREATE_TUPLE(uint64_t) + CREATE_TUPLE(std::string) + CREATE_TUPLE_INPUT(std::string_view) + CREATE_TUPLE(Ort::Float8E4M3FN_t) + CREATE_TUPLE(Ort::Float8E4M3FNUZ_t) + CREATE_TUPLE(Ort::Float8E5M2_t) + CREATE_TUPLE(Ort::Float8E5M2FNUZ_t) + + // ParseArgs ... + template + static typename std::enable_if<0 == sizeof...(Ts)>::type + ParseArgs(std::vector&, std::vector&) { + } + + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + ParseArgs(input_types, output_types); + } + + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + ParseArgs(input_types, output_types); + } + +#ifdef ORT_CUDA_CTX + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + ParseArgs(input_types, output_types); + } +#endif + +#ifdef ORT_ROCM_CTX + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + ParseArgs(input_types, output_types); + } +#endif + + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + input_types.push_back(ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED); + ParseArgs(input_types, output_types); + } + + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + input_types.push_back(ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED); + ParseArgs(input_types, output_types); + } + + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + output_types.push_back(ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED); + ParseArgs(input_types, output_types); + } + + template + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type + ParseArgs(std::vector& input_types, std::vector& output_types) { + output_types.push_back(ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED); + ParseArgs(input_types, output_types); + } + +#define PARSE_INPUT_BASE(pack_type, onnx_type) \ + template \ + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same::value>::type \ + ParseArgs(std::vector& input_types, std::vector& output_types) { \ + input_types.push_back(onnx_type); \ + ParseArgs(input_types, output_types); \ + } \ + template \ + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same>::value>::type \ + ParseArgs(std::vector& input_types, std::vector& output_types) { \ + input_types.push_back(onnx_type); \ + ParseArgs(input_types, output_types); \ + } \ + template \ + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same>::value>::type \ + ParseArgs(std::vector& input_types, std::vector& output_types) { \ + input_types.push_back(onnx_type); \ + ParseArgs(input_types, output_types); \ + } + +#define PARSE_INPUT(data_type, onnx_type) \ + PARSE_INPUT_BASE(const Custom::Tensor*, onnx_type) \ + PARSE_INPUT_BASE(const Custom::Tensor&, onnx_type) \ + PARSE_INPUT_BASE(const Custom::Span*, onnx_type) \ + PARSE_INPUT_BASE(const Custom::Span&, onnx_type) \ + PARSE_INPUT_BASE(data_type, onnx_type) + +#define PARSE_OUTPUT(data_type, onnx_type) \ + template \ + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same*>::value>::type \ + ParseArgs(std::vector& input_types, std::vector& output_types) { \ + output_types.push_back(onnx_type); \ + ParseArgs(input_types, output_types); \ + } \ + template \ + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same&>::value>::type \ + ParseArgs(std::vector& input_types, std::vector& output_types) { \ + output_types.push_back(onnx_type); \ + ParseArgs(input_types, output_types); \ + } \ + template \ + static typename std::enable_if<0 <= sizeof...(Ts) && std::is_same*>>::value>::type \ + ParseArgs(std::vector& input_types, std::vector& output_types) { \ + output_types.push_back(onnx_type); \ + ParseArgs(input_types, output_types); \ + } + +#define PARSE_ARGS(data_type, onnx_type) \ + PARSE_INPUT(data_type, onnx_type) \ + PARSE_OUTPUT(data_type, onnx_type) + + PARSE_ARGS(bool, ONNX_TENSOR_ELEMENT_DATA_TYPE_BOOL) + PARSE_ARGS(float, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT) + PARSE_ARGS(Ort::Float16_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16) + PARSE_ARGS(Ort::BFloat16_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_BFLOAT16) + PARSE_ARGS(double, ONNX_TENSOR_ELEMENT_DATA_TYPE_DOUBLE) + PARSE_ARGS(int8_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_INT8) + PARSE_ARGS(int16_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_INT16) + PARSE_ARGS(int32_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32) + PARSE_ARGS(int64_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64) + PARSE_ARGS(uint8_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT8) + PARSE_ARGS(uint16_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT16) + PARSE_ARGS(uint32_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT32) + PARSE_ARGS(uint64_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_UINT64) + PARSE_ARGS(std::string, ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING) + PARSE_ARGS(std::string_view, ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING) // todo - remove string_view output + PARSE_ARGS(Ort::Float8E4M3FN_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FN) + PARSE_ARGS(Ort::Float8E4M3FNUZ_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E4M3FNUZ) + PARSE_ARGS(Ort::Float8E5M2_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2) + PARSE_ARGS(Ort::Float8E5M2FNUZ_t, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT8E5M2FNUZ) + + OrtLiteCustomOp(const char* op_name, + const char* execution_provider, + ShapeInferFn shape_infer_fn, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) : op_name_(op_name), + execution_provider_(execution_provider), + shape_infer_fn_(shape_infer_fn), + start_ver_(start_ver), + end_ver_(end_ver) { + OrtCustomOp::version = ORT_API_VERSION; + + OrtCustomOp::GetName = [](const OrtCustomOp* op) { return static_cast(op)->op_name_.c_str(); }; + OrtCustomOp::GetExecutionProviderType = [](const OrtCustomOp* op) { return ((OrtLiteCustomOp*)op)->execution_provider_.c_str(); }; + OrtCustomOp::GetInputMemoryType = [](const OrtCustomOp*, size_t) { return OrtMemTypeDefault; }; + + OrtCustomOp::GetInputTypeCount = [](const OrtCustomOp* op) { + auto self = reinterpret_cast(op); + return self->input_types_.size(); + }; + + OrtCustomOp::GetInputType = [](const OrtCustomOp* op, size_t indice) { + auto self = reinterpret_cast(op); + return self->input_types_[indice]; + }; + + OrtCustomOp::GetOutputTypeCount = [](const OrtCustomOp* op) { + auto self = reinterpret_cast(op); + return self->output_types_.size(); + }; + + OrtCustomOp::GetOutputType = [](const OrtCustomOp* op, size_t indice) { + auto self = reinterpret_cast(op); + return self->output_types_[indice]; + }; + + OrtCustomOp::GetInputCharacteristic = [](const OrtCustomOp* op, size_t indice) { + auto self = reinterpret_cast(op); + return self->input_types_[indice] == ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED ? INPUT_OUTPUT_VARIADIC : INPUT_OUTPUT_OPTIONAL; + }; + + OrtCustomOp::GetOutputCharacteristic = [](const OrtCustomOp* op, size_t indice) { + auto self = reinterpret_cast(op); + return self->output_types_[indice] == ONNX_TENSOR_ELEMENT_DATA_TYPE_UNDEFINED ? INPUT_OUTPUT_VARIADIC : INPUT_OUTPUT_OPTIONAL; + }; + + OrtCustomOp::GetVariadicInputMinArity = [](const OrtCustomOp*) { + return 1; + }; + + OrtCustomOp::GetVariadicInputHomogeneity = [](const OrtCustomOp*) { + return 0; + }; + + OrtCustomOp::GetVariadicOutputMinArity = [](const OrtCustomOp*) { + return 1; + }; + + OrtCustomOp::GetVariadicOutputHomogeneity = [](const OrtCustomOp*) { + return 0; + }; + + OrtCustomOp::GetVariadicInputMinArity = [](const OrtCustomOp*) { return 0; }; + OrtCustomOp::GetVariadicInputHomogeneity = [](const OrtCustomOp*) { return 0; }; + OrtCustomOp::GetVariadicOutputMinArity = [](const OrtCustomOp*) { return 0; }; + OrtCustomOp::GetVariadicOutputHomogeneity = [](const OrtCustomOp*) { return 0; }; + + OrtCustomOp::CreateKernelV2 = {}; + OrtCustomOp::KernelComputeV2 = {}; + OrtCustomOp::KernelCompute = {}; + + OrtCustomOp::InferOutputShapeFn = {}; + + OrtCustomOp::GetStartVersion = [](const OrtCustomOp* op) { + auto self = reinterpret_cast(op); + return self->start_ver_; + }; + + OrtCustomOp::GetEndVersion = [](const OrtCustomOp* op) { + auto self = reinterpret_cast(op); + return self->end_ver_; + }; + + OrtCustomOp::GetMayInplace = {}; + OrtCustomOp::ReleaseMayInplace = {}; + OrtCustomOp::GetAliasMap = {}; + OrtCustomOp::ReleaseAliasMap = {}; + } + + const std::string op_name_; + const std::string execution_provider_; + + std::vector input_types_; + std::vector output_types_; + + ShapeInferFn shape_infer_fn_ = {}; + + int start_ver_ = 1; + int end_ver_ = MAX_CUSTOM_OP_END_VER; + + void* compute_fn_ = {}; + void* compute_fn_return_status_ = {}; +}; + +//////////////////////////// OrtLiteCustomFunc //////////////////////////////// +// The struct is to implement function-as-op. +// E.g. a function might be defined as: +// void Filter(const Ort::Custom::Tensor& floats_in, Ort::Custom::Tensor& floats_out) { ... } +// It could be registered this way: +// Ort::CustomOpDomain v2_domain{"v2"}; +// std::unique_ptr fil_op_ptr{Ort::Custom::CreateLiteCustomOp("Filter", "CPUExecutionProvider", Filter)}; +// v2_domain.Add(fil_op_ptr.get()); +// session_options.Add(v2_domain); +// For the complete example, please search keyword "LiteCustomOpTest" under "/onnxruntime/test/". +template +struct OrtLiteCustomFunc : public OrtLiteCustomOp { + using ComputeFn = void (*)(Args...); + using ComputeFnReturnStatus = Status (*)(Args...); + using MyType = OrtLiteCustomFunc; + + struct Kernel { + size_t num_input_{}; + size_t num_output_{}; + ComputeFn compute_fn_{}; + ComputeFnReturnStatus compute_fn_return_status_{}; + std::string ep_{}; + }; + + OrtLiteCustomFunc(const char* op_name, + const char* execution_provider, + ComputeFn compute_fn, + ShapeInferFn shape_infer_fn = {}, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) : OrtLiteCustomOp(op_name, execution_provider, shape_infer_fn, start_ver, end_ver) { + compute_fn_ = reinterpret_cast(compute_fn); + ParseArgs(input_types_, output_types_); + + OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) { + auto kernel = reinterpret_cast(op_kernel); + std::vector args; + auto t = CreateTuple<0, 0, Args...>(context, args, kernel->num_input_, kernel->num_output_, kernel->ep_); + std::apply([kernel](Args const&... t_args) { kernel->compute_fn_(t_args...); }, t); + }; + + OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* ort_api, const OrtKernelInfo* info) { + auto kernel = std::make_unique(); + auto me = static_cast(this_); + kernel->compute_fn_ = reinterpret_cast(me->compute_fn_); + Ort::ThrowOnError(ort_api->KernelInfo_GetInputCount(info, &kernel->num_input_)); + Ort::ThrowOnError(ort_api->KernelInfo_GetOutputCount(info, &kernel->num_output_)); + auto self = static_cast(this_); + kernel->ep_ = self->execution_provider_; + return reinterpret_cast(kernel.release()); + }; + + OrtCustomOp::KernelDestroy = [](void* op_kernel) { + delete reinterpret_cast(op_kernel); + }; + + if (shape_infer_fn_) { + OrtCustomOp::InferOutputShapeFn = [](const OrtCustomOp* op, OrtShapeInferContext* ort_ctx) -> OrtStatusPtr { + auto shape_info_fn = static_cast(op)->shape_infer_fn_; + ShapeInferContext ctx(&GetApi(), ort_ctx); + return shape_info_fn(ctx); + }; + } + } + + OrtLiteCustomFunc(const char* op_name, + const char* execution_provider, + ComputeFnReturnStatus compute_fn_return_status, + ShapeInferFn shape_infer_fn = {}, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) : OrtLiteCustomOp(op_name, execution_provider, shape_infer_fn, start_ver, end_ver) { + compute_fn_return_status_ = reinterpret_cast(compute_fn_return_status); + ParseArgs(input_types_, output_types_); + + OrtCustomOp::KernelComputeV2 = [](void* op_kernel, OrtKernelContext* context) -> OrtStatusPtr { + auto kernel = reinterpret_cast(op_kernel); + std::vector args; + auto t = CreateTuple<0, 0, Args...>(context, args, kernel->num_input_, kernel->num_output_, kernel->ep_); + return std::apply([kernel](Args const&... t_args) { Status status = kernel->compute_fn_return_status_(t_args...); return status.release(); }, t); + }; + + OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* ort_api, const OrtKernelInfo* info) { + auto kernel = std::make_unique(); + auto me = static_cast(this_); + kernel->compute_fn_return_status_ = reinterpret_cast(me->compute_fn_return_status_); + Ort::ThrowOnError(ort_api->KernelInfo_GetInputCount(info, &kernel->num_input_)); + Ort::ThrowOnError(ort_api->KernelInfo_GetOutputCount(info, &kernel->num_output_)); + auto self = static_cast(this_); + kernel->ep_ = self->execution_provider_; + return reinterpret_cast(kernel.release()); + }; + + OrtCustomOp::KernelDestroy = [](void* op_kernel) { + delete reinterpret_cast(op_kernel); + }; + + if (shape_infer_fn_) { + OrtCustomOp::InferOutputShapeFn = [](const OrtCustomOp* op, OrtShapeInferContext* ort_ctx) -> OrtStatusPtr { + auto shape_info_fn = static_cast(op)->shape_infer_fn_; + ShapeInferContext ctx(&GetApi(), ort_ctx); + return shape_info_fn(ctx); + }; + } + } +}; // struct OrtLiteCustomFunc + +/////////////////////////// OrtLiteCustomStruct /////////////////////////// +// The struct is to implement struct-as-op. +// E.g. a struct might be defined as: +// struct Merge { +// Merge(const OrtApi* ort_api, const OrtKernelInfo* info) {...} +// void Compute(const Ort::Custom::Tensor& strings_in, +// std::string_view string_in, +// Ort::Custom::Tensor* strings_out) {...} +// bool reverse_ = false; +// }; +// It could be registered this way: +// Ort::CustomOpDomain v2_domain{"v2"}; +// std::unique_ptr mrg_op_ptr{Ort::Custom::CreateLiteCustomOp("Merge", "CPUExecutionProvider")}; +// v2_domain.Add(mrg_op_ptr.get()); +// session_options.Add(v2_domain); +// For the complete example, please search keyword "LiteCustomOpTest" under "/onnxruntime/test/". +template +struct OrtLiteCustomStruct : public OrtLiteCustomOp { + template + using CustomComputeFn = void (CustomOp::*)(Args...); + + template + using CustomComputeFnReturnStatus = Status (CustomOp::*)(Args...); + + using MyType = OrtLiteCustomStruct; + + struct Kernel { + size_t num_input_{}; + size_t num_output_{}; + std::unique_ptr custom_op_; + std::string ep_{}; + }; + + OrtLiteCustomStruct(const char* op_name, + const char* execution_provider, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) : OrtLiteCustomOp(op_name, execution_provider, {}, start_ver, end_ver) { + SetCompute(&CustomOp::Compute); + + OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* ort_api, const OrtKernelInfo* info) { + auto kernel = std::make_unique(); + Ort::ThrowOnError(ort_api->KernelInfo_GetInputCount(info, &kernel->num_input_)); + Ort::ThrowOnError(ort_api->KernelInfo_GetOutputCount(info, &kernel->num_output_)); + kernel->custom_op_ = std::make_unique(ort_api, info); + auto self = static_cast(this_); + kernel->ep_ = self->execution_provider_; + return reinterpret_cast(kernel.release()); + }; + + OrtCustomOp::KernelDestroy = [](void* op_kernel) { + delete reinterpret_cast(op_kernel); + }; + + SetShapeInfer(0); + } + + template + void SetCompute(CustomComputeFn) { + ParseArgs(input_types_, output_types_); + OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) { + auto kernel = reinterpret_cast(op_kernel); + ArgPtrs args; + auto t = CreateTuple<0, 0, Args...>(context, args, kernel->num_input_, kernel->num_output_, kernel->ep_); + std::apply([kernel](Args const&... t_args) { kernel->custom_op_->Compute(t_args...); }, t); + }; + } + + template + void SetCompute(CustomComputeFnReturnStatus) { + ParseArgs(input_types_, output_types_); + OrtCustomOp::KernelComputeV2 = [](void* op_kernel, OrtKernelContext* context) -> OrtStatusPtr { + auto kernel = reinterpret_cast(op_kernel); + ArgPtrs args; + auto t = CreateTuple<0, 0, Args...>(context, args, kernel->num_input_, kernel->num_output_, kernel->ep_); + return std::apply([kernel](Args const&... t_args) { Status status = kernel->custom_op_->Compute(t_args...); return status.release(); }, t); + }; + } + + template + decltype(&C::InferOutputShape) SetShapeInfer(decltype(&C::InferOutputShape)) { + OrtCustomOp::InferOutputShapeFn = [](const OrtCustomOp*, OrtShapeInferContext* ort_ctx) -> OrtStatusPtr { + ShapeInferContext ctx(&GetApi(), ort_ctx); + return C::InferOutputShape(ctx); + }; + return {}; + } + + template + void SetShapeInfer(...) { + OrtCustomOp::InferOutputShapeFn = {}; + } +}; // struct OrtLiteCustomStruct + +/////////////////////////// CreateLiteCustomOp //////////////////////////// + +template +OrtLiteCustomOp* CreateLiteCustomOp(const char* op_name, + const char* execution_provider, + void (*custom_compute_fn)(Args...), + Status (*shape_infer_fn)(ShapeInferContext&) = {}, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) { + using LiteOp = OrtLiteCustomFunc; + return std::make_unique(op_name, execution_provider, custom_compute_fn, shape_infer_fn, start_ver, end_ver).release(); +} + +template +OrtLiteCustomOp* CreateLiteCustomOp(const char* op_name, + const char* execution_provider, + Status (*custom_compute_fn_v2)(Args...), + Status (*shape_infer_fn)(ShapeInferContext&) = {}, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) { + using LiteOp = OrtLiteCustomFunc; + return std::make_unique(op_name, execution_provider, custom_compute_fn_v2, shape_infer_fn, start_ver, end_ver).release(); +} + +template +OrtLiteCustomOp* CreateLiteCustomOp(const char* op_name, + const char* execution_provider, + int start_ver = 1, + int end_ver = MAX_CUSTOM_OP_END_VER) { + using LiteOp = OrtLiteCustomStruct; + return std::make_unique(op_name, execution_provider, start_ver, end_ver).release(); +} + +} // namespace Custom +} // namespace Ort diff --git a/third_party/onnxruntime/include/onnxruntime_run_options_config_keys.h b/third_party/onnxruntime/include/onnxruntime_run_options_config_keys.h new file mode 100644 index 00000000..f40ea659 --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_run_options_config_keys.h @@ -0,0 +1,54 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#pragma once + +/* + * This file defines RunOptions Config Keys and format of the Config Values. + * + * The Naming Convention for a RunOptions Config Key, + * "[Area][.[SubArea1].[SubArea2]...].[Keyname]" + * Such as "ep.cuda.use_arena" + * The Config Key cannot be empty + * The maximum length of the Config Key is 128 + * + * The string format of a RunOptions Config Value is defined individually for each Config. + * The maximum length of the Config Value is 1024 + */ + +// Key for enabling shrinkages of user listed device memory arenas. +// Expects a list of semi-colon separated key value pairs separated by colon in the following format: +// "device_0:device_id_0;device_1:device_id_1" +// No white-spaces allowed in the provided list string. +// Currently, the only supported devices are : "cpu", "gpu" (case sensitive). +// If "cpu" is included in the list, DisableCpuMemArena() API must not be called (i.e.) arena for cpu should be enabled. +// Example usage: "cpu:0;gpu:0" (or) "gpu:0" +// By default, the value for this key is empty (i.e.) no memory arenas are shrunk +static const char* const kOrtRunOptionsConfigEnableMemoryArenaShrinkage = "memory.enable_memory_arena_shrinkage"; + +// Set to '1' to not synchronize execution providers with CPU at the end of session run. +// Per default it will be set to '0' +// Taking CUDA EP as an example, it omit triggering cudaStreamSynchronize on the compute stream. +static const char* const kOrtRunOptionsConfigDisableSynchronizeExecutionProviders = "disable_synchronize_execution_providers"; + +// Set HTP performance mode for QNN HTP backend before session run. +// options for HTP performance mode: "burst", "balanced", "default", "high_performance", +// "high_power_saver", "low_balanced", "extreme_power_saver", "low_power_saver", "power_saver", +// "sustained_high_performance". Default to "default". +static const char* const kOrtRunOptionsConfigQnnPerfMode = "qnn.htp_perf_mode"; + +// Set HTP performance mode for QNN HTP backend post session run. +static const char* const kOrtRunOptionsConfigQnnPerfModePostRun = "qnn.htp_perf_mode_post_run"; + +// Set RPC control latency for QNN HTP backend +static const char* const kOrtRunOptionsConfigQnnRpcControlLatency = "qnn.rpc_control_latency"; + +// Set QNN Lora Config File for apply Lora in QNN context binary +static const char* const kOrtRunOptionsConfigQnnLoraConfig = "qnn.lora_config"; + +// Set graph annotation id for CUDA EP. Use with enable_cuda_graph=true. +// The value should be an integer. If the value is not set, the default value is 0 and +// ORT session only captures one cuda graph before another capture is requested. +// If the value is set to -1, cuda graph capture/replay is disabled in that run. +// User are not expected to set the value to 0 as it is reserved for internal use. +static const char* const kOrtRunOptionsConfigCudaGraphAnnotation = "gpu_graph_id"; diff --git a/third_party/onnxruntime/include/onnxruntime_session_options_config_keys.h b/third_party/onnxruntime/include/onnxruntime_session_options_config_keys.h new file mode 100644 index 00000000..5497d7c7 --- /dev/null +++ b/third_party/onnxruntime/include/onnxruntime_session_options_config_keys.h @@ -0,0 +1,366 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#pragma once + +/* + * This file defines SessionOptions Config Keys and format of the Config Values. + * + * The Naming Convention for a SessionOptions Config Key, + * "[Area][.[SubArea1].[SubArea2]...].[Keyname]" + * Such as "ep.cuda.use_arena" + * The Config Key cannot be empty + * The maximum length of the Config Key is 1024 + * + * The string format of a SessionOptions Config Value is defined individually for each Config. + * The maximum length of the Config Value is 2048 + */ + +// Key for disable PrePacking, +// If the config value is set to "1" then the prepacking is disabled, otherwise prepacking is enabled (default value) +static const char* const kOrtSessionOptionsConfigDisablePrepacking = "session.disable_prepacking"; + +// A value of "1" means allocators registered in the env will be used. "0" means the allocators created in the session +// will be used. Use this to override the usage of env allocators on a per session level. +static const char* const kOrtSessionOptionsConfigUseEnvAllocators = "session.use_env_allocators"; + +// Set to 'ORT' (case sensitive) to load an ORT format model. +// If unset, model type will default to ONNX unless inferred from filename ('.ort' == ORT format) or bytes to be ORT +static const char* const kOrtSessionOptionsConfigLoadModelFormat = "session.load_model_format"; + +// Set to 'ORT' (case sensitive) to save optimized model in ORT format when SessionOptions.optimized_model_path is set. +// If unset, format will default to ONNX unless optimized_model_filepath ends in '.ort'. +static const char* const kOrtSessionOptionsConfigSaveModelFormat = "session.save_model_format"; + +// If a value is "1", flush-to-zero and denormal-as-zero are applied. The default is "0". +// When multiple sessions are created, a main thread doesn't override changes from succeeding session options, +// but threads in session thread pools follow option changes. +// When ORT runs with OpenMP, the same rule is applied, i.e. the first session option to flush-to-zero and +// denormal-as-zero is only applied to global OpenMP thread pool, which doesn't support per-session thread pool. +// Note that an alternative way not using this option at runtime is to train and export a model without denormals +// and that's recommended because turning this option on may hurt model accuracy. +static const char* const kOrtSessionOptionsConfigSetDenormalAsZero = "session.set_denormal_as_zero"; + +// It controls to run quantization model in QDQ (QuantizelinearDeQuantizelinear) format or not. +// "0": enable. ORT does fusion logic for QDQ format. +// "1": disable. ORT doesn't do fusion logic for QDQ format. +// Its default value is "0" unless the DirectML execution provider is registered, in which case it defaults to "1". +static const char* const kOrtSessionOptionsDisableQuantQDQ = "session.disable_quant_qdq"; + +// It controls whether to enable Double QDQ remover and Identical Children Consolidation +// "0": not to disable. ORT does remove the middle 2 Nodes from a Q->(QD->Q)->QD pairs +// "1": disable. ORT doesn't remove the middle 2 Nodes from a Q->(QD->Q)->QD pairs +// Its default value is "0" +static const char* const kOrtSessionOptionsDisableDoubleQDQRemover = "session.disable_double_qdq_remover"; + +// If set to "1", enables the removal of QuantizeLinear/DequantizeLinear node pairs once all QDQ handling has been +// completed. e.g. If after all QDQ handling has completed and we have -> FloatOp -> Q -> DQ -> FloatOp -> the +// Q -> DQ could potentially be removed. This will provide a performance benefit by avoiding going from float to +// 8-bit and back to float, but could impact accuracy. The impact on accuracy will be model specific and depend on +// other factors like whether the model was created using Quantization Aware Training or Post Training Quantization. +// As such, it's best to test to determine if enabling this works well for your scenario. +// The default value is "0" +// Available since version 1.11. +static const char* const kOrtSessionOptionsEnableQuantQDQCleanup = "session.enable_quant_qdq_cleanup"; + +// Enable or disable gelu approximation in graph optimization. "0": disable; "1": enable. The default is "0". +// GeluApproximation has side effects which may change the inference results. It is disabled by default due to this. +static const char* const kOrtSessionOptionsEnableGeluApproximation = "optimization.enable_gelu_approximation"; + +// This setting controls whether to enable AheadOfTime function inlining. +// AOT function inlining examines the graph and attempts to inline as many locally defined functions in the model +// as possible with the help of enabled execution providers. +// This can reduce the number of function calls and improve performance because it is done before +// Level1 optimizers and constant folding. However, under some circumstances, when the EPs are not available, +// one can disable the AOT inlining, produce an optimized model and postpone AOT until run time. +// "0": enable; "1": disable. +// Its default value is "0". +static const char* const kOrtSessionOptionsDisableAheadOfTimeFunctionInlining = "session.disable_aot_function_inlining"; + +#ifdef ENABLE_TRAINING +// Specifies a path of the file containing a list of memory optimization configurations. +// The value should be a string indicating the file path of the config file. +// The content of the config file is a JSON struct like this: +// [ +// "Gelu+Cast+:1:0", +// "Dropout+:1:1" +// ] +// Taking the example of "Gelu+Cast+:1:0", +// > "Gelu+Cast+" is the subgraph string, a valid "subgraph string" should be one subgraph representation +// output by ORT graph transformations. +// > "1" is "optimization strategy", valid values: 0 - disabled, 1 - recompute. +// > "0" is "number of subgraph to apply" which is used to control how many subgraphs to apply optimization, +// to avoid "oversaving" the memory. +static const char* const kOrtSessionOptionsMemoryOptimizerApplyConfig = "optimization.memory_optimizer_config"; + +// Specifies the config for detecting subgraphs for memory footprint reduction. +// The value should be a string contains int separated using commas. The default value is "0:0". +static const char* const kOrtSessionOptionsMemoryOptimizerProbeConfig = "optimization.enable_memory_probe_recompute_config"; +#endif + +// This setting if set should contain a comma separated list of optimizers names that should be disabled. +// Optimizers may take time to execute and affect model loading time. If you feel that a specific optimizer +// does not provider runtime benefits, but affects your model loading time you may disable it using this config +// entry. This option is not enabled in ORT_MINIMAL_BUILD build. +// A list of optimizes is available in onnxruntime/core/optimizer/graph_transformer_utils.cc +// +// Default is an empty string which means no optimizers are disabled. +static const char* const kOrtSessionOptionsDisableSpecifiedOptimizers = "optimization.disable_specified_optimizers"; + +// Enable or disable using device allocator for allocating initialized tensor memory. "1": enable; "0": disable. The default is "0". +// Using device allocators means the memory allocation is made using malloc/new. +static const char* const kOrtSessionOptionsUseDeviceAllocatorForInitializers = "session.use_device_allocator_for_initializers"; + +// Configure whether to allow the inter_op/intra_op threads spinning a number of times before blocking +// "0": thread will block if found no job to run +// "1": default, thread will spin a number of times before blocking +static const char* const kOrtSessionOptionsConfigAllowInterOpSpinning = "session.inter_op.allow_spinning"; +static const char* const kOrtSessionOptionsConfigAllowIntraOpSpinning = "session.intra_op.allow_spinning"; + +// Key for using model bytes directly for ORT format +// If a session is created using an input byte array contains the ORT format model data, +// By default we will copy the model bytes at the time of session creation to ensure the model bytes +// buffer is valid. +// Setting this option to "1" will disable copy the model bytes, and use the model bytes directly. The caller +// has to guarantee that the model bytes are valid until the ORT session using the model bytes is destroyed. +static const char* const kOrtSessionOptionsConfigUseORTModelBytesDirectly = "session.use_ort_model_bytes_directly"; + +/// +/// Key for using the ORT format model flatbuffer bytes directly for initializers. +/// This avoids copying the bytes and reduces peak memory usage during model loading and initialization. +/// Requires `session.use_ort_model_bytes_directly` to be true. +/// If set, the flatbuffer bytes provided when creating the InferenceSession MUST remain valid for the entire +/// duration of the InferenceSession. +/// +static const char* const kOrtSessionOptionsConfigUseORTModelBytesForInitializers = + "session.use_ort_model_bytes_for_initializers"; + +// This should only be specified when exporting an ORT format model for use on a different platform. +// If the ORT format model will be used on ARM platforms set to "1". For other platforms set to "0" +// Available since version 1.11. +static const char* const kOrtSessionOptionsQDQIsInt8Allowed = "session.qdqisint8allowed"; + +// x64 SSE4.1/AVX2/AVX512(with no VNNI) has overflow problem with quantizied matrix multiplication with U8S8. +// To avoid this we need to use slower U8U8 matrix multiplication instead. This option, if +// turned on, use slower U8U8 matrix multiplications. Only effective with AVX2 or AVX512 +// platforms. +static const char* const kOrtSessionOptionsAvx2PrecisionMode = "session.x64quantprecision"; + +// Specifies how minimal build graph optimizations are handled in a full build. +// These optimizations are at the extended level or higher. +// Possible values and their effects are: +// "save": Save runtime optimizations when saving an ORT format model. +// "apply": Only apply optimizations available in a minimal build. +// ""/: Apply optimizations available in a full build. +// Available since version 1.11. +static const char* const kOrtSessionOptionsConfigMinimalBuildOptimizations = + "optimization.minimal_build_optimizations"; + +// Note: The options specific to an EP should be specified prior to appending that EP to the session options object in +// order for them to take effect. + +// Specifies a list of stop op types. Nodes of a type in the stop op types and nodes downstream from them will not be +// run by the NNAPI EP. +// The value should be a ","-delimited list of op types. For example, "Add,Sub". +// If not specified, the default set of stop ops is used. To specify an empty stop ops types list and disable stop op +// exclusion, set the value to "". +static const char* const kOrtSessionOptionsConfigNnapiEpPartitioningStopOps = "ep.nnapi.partitioning_stop_ops"; + +// Enabling dynamic block-sizing for multithreading. +// With a positive value, thread pool will split a task of N iterations to blocks of size starting from: +// N / (num_of_threads * dynamic_block_base) +// As execution progresses, the size will decrease according to the diminishing residual of N, +// meaning the task will be distributed in smaller granularity for better parallelism. +// For some models, it helps to reduce the variance of E2E inference latency and boost performance. +// The feature will not function by default, specify any positive integer, e.g. "4", to enable it. +// Available since version 1.11. +static const char* const kOrtSessionOptionsConfigDynamicBlockBase = "session.dynamic_block_base"; + +// This option allows to decrease CPU usage between infrequent +// requests and forces any TP threads spinning stop immediately when the last of +// concurrent Run() call returns. +// Spinning is restarted on the next Run() call. +// Applies only to internal thread-pools +static const char* const kOrtSessionOptionsConfigForceSpinningStop = "session.force_spinning_stop"; + +// "1": all inconsistencies encountered during shape and type inference +// will result in failures. +// "0": in some cases warnings will be logged but processing will continue. The default. +// May be useful to expose bugs in models. +static const char* const kOrtSessionOptionsConfigStrictShapeTypeInference = "session.strict_shape_type_inference"; + +// "1": every model using a more recent opset than the latest released one will fail +// "0": the model may or may not work if onnxruntime cannot find an implementation, this option +// is used for development purpose. +static const char* const kOrtSessionOptionsConfigStrictAllowReleasedOpsetsOnly = "session.allow_released_opsets_only"; + +// The file saves configuration for partitioning node among logic streams +static const char* const kNodePartitionConfigFile = "session.node_partition_config_file"; + +// This Option allows setting affinities for intra op threads. +// Affinity string follows format: +// logical_processor_id,logical_processor_id;logical_processor_id,logical_processor_id +// Semicolon isolates configurations among threads, while comma split processors where ith thread expected to attach to. +// e.g.1,2,3;4,5 +// specifies affinities for two threads, with the 1st thread attach to the 1st, 2nd, and 3rd processor, and 2nd thread to the 4th and 5th. +// To ease the configuration, an "interval" is also allowed: +// e.g. 1-8;8-16;17-24 +// orders that the 1st thread runs on first eight processors, 2nd thread runs on next eight processors, and so forth. +// Note: +// 1. Once set, the number of thread affinities must equal to intra_op_num_threads - 1, since ort does not set affinity on the main thread which +// is started and managed by the calling app; +// 2. For windows, ort will infer the group id from a logical processor id, for example, assuming there are two groups with each has 64 logical processors, +// an id of 64 will be inferred as the last processor of the 1st group, while 65 will be interpreted as the 1st processor of the second group. +// Hence 64-65 is an invalid configuration, because a windows thread cannot be attached to processors across group boundary. +static const char* const kOrtSessionOptionsConfigIntraOpThreadAffinities = "session.intra_op_thread_affinities"; + +// This option will dump out the model to assist debugging any issues with layout transformation, +// and is primarily intended for developer usage. It is only relevant if an execution provider that requests +// NHWC layout is enabled such as NNAPI, XNNPACK or QNN. +// +// Default is off. Set to "1" to enable. +// +// If modified by layout transformation the model will be dumped after these steps: +// 1) insertion of the layout transformation Transpose nodes +// 2) after those are optimized using the transpose optimizer, +// 3) after the L1 transformers are applied to the updated graph. +// The model will be saved to filename post_layout_transform_step_.onnx. +static const char* const kDebugLayoutTransformation = "session.debug_layout_transformation"; + +// Graph nodes that are not supported by the execution providers (EPs) explicitly added to the session are +// assigned (i.e., "fallback") to the CPU EP by default. +// +// This option allows the user to disable the fallback of unsupported graph nodes to the CPU EP. +// If this option is set to "1", session creation will fail if the execution providers other than the CPU EP cannot +// fully support all of the nodes in the graph. +// +// It is invalid to set this option and explicitly add the CPU EP to the session. In this case, session creation +// will also fail with an error. +// +// Option values: +// - "0": CPU EP fallback is not disabled. [DEFAULT] +// - "1": CPU EP fallback is disabled. +static const char* const kOrtSessionOptionsDisableCPUEPFallback = "session.disable_cpu_ep_fallback"; + +// Use this config when serializing a large model after optimization to specify an external initializers file +static const char* const kOrtSessionOptionsOptimizedModelExternalInitializersFileName = + "session.optimized_model_external_initializers_file_name"; + +// Use this config to control the minimum size of the initializer when externalizing it during serialization +static const char* const kOrtSessionOptionsOptimizedModelExternalInitializersMinSizeInBytes = + "session.optimized_model_external_initializers_min_size_in_bytes"; + +// When loading model from memory buffer and the model has external initializers +// Use this config to set the external data file folder path +// All external data files should be in the same folder +static const char* const kOrtSessionOptionsModelExternalInitializersFileFolderPath = + "session.model_external_initializers_file_folder_path"; + +// Use this config when saving pre-packed constant initializers to an external data file. +// This allows you to memory map pre-packed initializers on model load and leave it to +// to the OS the amount of memory consumed by the pre-packed initializers. Otherwise, +// pre-packed data resides on the heap. +// +// - "0": Default is not save pre-packed initializers to a data file. +// - "1": Save pre-packed constant initializers to an external data file. +// Sample usage: sess_options.add_session_config_entry(kOrtSessionOptionsSavePrePackedConstantInitializers, "1") +static const char* const kOrtSessionOptionsSavePrePackedConstantInitializers = + "session.save_external_prepacked_constant_initializers"; + +// Use this config when you want to collect memory stats for each node in the graph. +// The file format is a CSV file with the following columns: +// The file will be created if it does not exist, and will be overwritten if it does. +// +// The content of the file can be used to estimate memory requirements at run time including +// the temporary allocations. This operation is preferably done on a CPU device, as the model may exceed +// device memory limits in constrained environments. When enabling this option, it is important to disable +// memory patterns, as they tend to allocate large blocks to avoid fragmentation and accommodate needs of multiple +// kernels. Memory patterns may make it difficult to allocate on a device with limited memory. +// +// The collected stats then can be used to partition the graph among the devices in a way that only the +// required memory is allocated on each device. +// +// node_name, initializers_memory, dynamic_outputs_sizes, temp_allocations_size +// +// - "full path to file": there is not a default for this option. If the file can not be opened for writing, an error will be returned. +static const char* const kOrtSessionOptionsCollectNodeMemoryStatsToFile = "session.collect_node_memory_stats_to_file"; + +/// This is a composite CSV setting formatted as "memory limit in kb,file name for collected stats" +/// "limit > 0": enables Capacity Aware Partitioning for Cuda EP. `limit` is optional and when absent +/// the provider may attempt to figure out the memory available automatically. +/// The setting with no limit is expected to look like: ",file name for collected stats" +/// The EP will place nodes on device "file name" : +/// this file is expected to be found at the same folder with the model. The file contains +/// pre-recorded stats collected when running with kOrtSessionOptionsCollectNodeMemoryStatsToFile enforce (see above) +static const char* const kOrtSessionOptionsResourceCudaPartitioningSettings = + "session.resource_cuda_partitioning_settings"; + +// Enable EP context feature to dump the partitioned graph which includes the EP context into Onnx file. +// The dumped Onnx model with EP context can be used for future inference to avoid the EP graph partitioning/compile overhead. +// "0": disable. (default) +// "1": enable. +static const char* const kOrtSessionOptionEpContextEnable = "ep.context_enable"; + +// Specify the file path for the Onnx model which has EP context. +// Default to original_file_name_ctx.onnx if not specified +// Folder is not a valid option +static const char* const kOrtSessionOptionEpContextFilePath = "ep.context_file_path"; + +// Flag to specify whether to dump the EP context into the Onnx model. +// "0": dump the EP context into separate file, keep the file name in the Onnx model. (default). +// "1": dump the EP context into the Onnx model. +static const char* const kOrtSessionOptionEpContextEmbedMode = "ep.context_embed_mode"; + +// Specify the EPContext node name prefix to make it unique +// in case user need to merge/connect multiple EPContext nodes in one model +static const char* const kOrtSessionOptionEpContextNodeNamePrefix = "ep.context_node_name_prefix"; + +// Share EP related resources across sessions +static const char* const kOrtSessionOptionShareEpContexts = "ep.share_ep_contexts"; + +// Stop to share EP related resources across sessions from then on +static const char* const kOrtSessionOptionStopShareEpContexts = "ep.stop_share_ep_contexts"; + +// Used only for context model generation. +// This configuration is used when some nodes are partitioned on the CPU EP and those nodes have external initializers. +// When generating the EP context model, the new model should not rely on the old external data file used by the source ONNX model. +// Use this setting when dumping the EP context model with an external initializers file. +// If specified, all initializers will be placed inside the external data file. +// Otherwise, all initializers will be embedded inside the generated ONNX file. +// By default, this option is not set, meaning all initializers will be included within the ONNX file. +static const char* const kOrtSessionOptionsEpContextModelExternalInitializersFileName = + "ep.context_model_external_initializers_file_name"; + +// Gemm fastmath mode provides fp32 gemm acceleration with bfloat16 based matmul. +// Option values: +// - "0": Gemm FastMath mode is not enabled. [DEFAULT] +// - "1": Gemm FastMath mode is enabled. +static const char* const kOrtSessionOptionsMlasGemmFastMathArm64Bfloat16 = "mlas.enable_gemm_fastmath_arm64_bfloat16"; + +// When converting DQ + MatMul -> MatMulNBits, the accuracy level of the MatMulNBits is controlled by this option. +// Refer to MatMulNBits op schema for more details. +// If not provided, default is 4. +static const char* const kOrtSessionOptionsQDQMatMulNBitsAccuracyLevel = "session.qdq_matmulnbits_accuracy_level"; + +// THIS OPTION IS NOT A REGULAR SESSION OPTION SINCE IT CAN BE MODIFIED AT ANY TIME +// Meant to be used with SetEpDynamicOptions +// Specify the type of workload for this session. +// “Default”: OS determines the scheduling priority and processor performance to service this workload. [Default] +// “Efficient”: OS treats this workload is efficiency oriented with low scheduling priority and efficient processor performance. +static const char* const kOrtEpDynamicOptionsWorkloadType = "ep.dynamic.workload_type"; + +// Disables model compilation during session initialization. +// +// If this option is set to "1", inference session creation will fail with error code ORT_MODEL_REQUIRES_COMPILATION +// if compilation is required to run the model on any Execution Provider added to the session. +// Only the following kinds of models are valid when this option is set to "1": +// - Pre-compiled models that have EPContext nodes for the compiling Execution Providers in the session. +// - Non-compiled models that run only on non-compiling Execution Providers, like CPU EP. +// +// See \href https://onnxruntime.ai/docs/execution-providers/EP-Context-Design.html for details about +// compiled models with EPContext nodes. +// +// Option values: +// - "0": EP compile is not disabled. [DEFAULT] +// - "1": EP compile is disabled. +static const char* const kOrtSessionOptionsDisableModelCompile = "session.disable_model_compile"; diff --git a/third_party/onnxruntime/include/provider_options.h b/third_party/onnxruntime/include/provider_options.h new file mode 100644 index 00000000..aab13e80 --- /dev/null +++ b/third_party/onnxruntime/include/provider_options.h @@ -0,0 +1,18 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +#pragma once + +#include +#include +#include + +namespace onnxruntime { + +// data types for execution provider options + +using ProviderOptions = std::unordered_map; +using ProviderOptionsVector = std::vector; +using ProviderOptionsMap = std::unordered_map; + +} // namespace onnxruntime diff --git a/third_party/onnxruntime/include/webgpu_provider_factory.h b/third_party/onnxruntime/include/webgpu_provider_factory.h new file mode 100644 index 00000000..0b45b847 --- /dev/null +++ b/third_party/onnxruntime/include/webgpu_provider_factory.h @@ -0,0 +1,14 @@ +// Copyright (c) Microsoft Corporation. All rights reserved. +// Licensed under the MIT License. + +// Dummy file to provide a signal in the ONNX Runtime C cocoapod as to whether the WebGPU EP was included in the build. +// If it was, this file will be included in the cocoapod, and a test like this can be used: +// +// #if __has_include() +// #define WEBGPU_EP_AVAILABLE 1 +// #else +// #define WEBGPU_EP_AVAILABLE 0 +// #endif + +// The WebGPU EP can be enabled via the generic SessionOptionsAppendExecutionProvider method, so no direct usage of +// the provider factory is required. diff --git a/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib new file mode 100755 index 00000000..4313c296 --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db045368293215c9d22aa7b8c983d688b3ae9ca1da3f64ffbe01ba7df31c3355 +size 70774760 diff --git a/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Info.plist b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Info.plist new file mode 100644 index 00000000..463d6b4a --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Info.plist @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28804d24c39e029589fe6fa0df9cba7573dac77b5af73c8c06ccebb45590a4de +size 656 diff --git a/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/DWARF/libonnxruntime.1.22.0.dylib b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/DWARF/libonnxruntime.1.22.0.dylib new file mode 100644 index 00000000..e7d02e3b --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/DWARF/libonnxruntime.1.22.0.dylib @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd4fdf091af6de95c8f0f9d0c07ec220e50afca3e28ffa48aae6bf0abdafa0bf +size 150291495 diff --git a/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/Relocations/aarch64/libonnxruntime.1.22.0.dylib.yml b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/Relocations/aarch64/libonnxruntime.1.22.0.dylib.yml new file mode 100644 index 00000000..dbb30a1c --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/Relocations/aarch64/libonnxruntime.1.22.0.dylib.yml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fc2fbdbb0303f4b3f8ff39cefb4db91a89cc56d660e5d67c8239d252111af00 +size 2972517 diff --git a/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/Relocations/x86_64/libonnxruntime.1.22.0.dylib.yml b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/Relocations/x86_64/libonnxruntime.1.22.0.dylib.yml new file mode 100644 index 00000000..2d00f894 --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime.1.22.0.dylib.dSYM/Contents/Resources/Relocations/x86_64/libonnxruntime.1.22.0.dylib.yml @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a77ba16f4e21fd0b6c94e8253c4f79fd3320200d7d97f61e85804e9f51f6a78f +size 2143723 diff --git a/third_party/onnxruntime/lib/libonnxruntime.so.1 b/third_party/onnxruntime/lib/libonnxruntime.so.1 new file mode 100755 index 00000000..fda66fd5 --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime.so.1 @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3da6146e14e7b8aaec625dde11d6114c7457c87a5f93d744897da8781e35c673 +size 21042416 diff --git a/third_party/onnxruntime/lib/libonnxruntime_providers_shared.so b/third_party/onnxruntime/lib/libonnxruntime_providers_shared.so new file mode 100755 index 00000000..d1856781 --- /dev/null +++ b/third_party/onnxruntime/lib/libonnxruntime_providers_shared.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4706ea02be3999167f000bb1ff7391c86e95d2d96f78eb33225c14b172a8451 +size 14632 diff --git a/third_party/onnxruntime_extensions.BUILD b/third_party/onnxruntime_extensions.BUILD deleted file mode 100644 index dddc1b9f..00000000 --- a/third_party/onnxruntime_extensions.BUILD +++ /dev/null @@ -1,31 +0,0 @@ -cc_library( - name = "headers", - hdrs = glob(["includes/**"]), - strip_include_prefix = "includes", - deps = [ - "@github_nlohmann_json//:json", - ], - visibility = ["//visibility:public"], -) - -cc_library( - name = "operators", - hdrs = [ - "operators/base64.h", - "operators/ustring.h", - "operators/tokenizer/bert_tokenizer.hpp", - "operators/tokenizer/basic_tokenizer.hpp" - ] + glob([ - "operators/string_utils*", - ]), - srcs = [ - "operators/base64.cc", - "operators/ustring.cc", - "operators/tokenizer/bert_tokenizer.cc", - "operators/tokenizer/basic_tokenizer.cc" - ] + glob([ - "operators/string_utils*", - ]), - strip_include_prefix = "operators", - visibility = ["//visibility:public"], -) \ No newline at end of file diff --git a/third_party/patch/onnx.patch b/third_party/patch/onnx.patch deleted file mode 100644 index 8a7deae3..00000000 --- a/third_party/patch/onnx.patch +++ /dev/null @@ -1,45 +0,0 @@ -diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt -index d90a2a3550..d5278575ae 100644 ---- cmake/CMakeLists.txt -+++ cmake/CMakeLists.txt -@@ -1456,6 +1456,7 @@ if (onnxruntime_USE_CUDA) - set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_53,code=sm_53") # TX1, Nano - set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_62,code=sm_62") # TX2 - set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_72,code=sm_72") # AGX Xavier, NX Xavier -+ set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_75,code=sm_75") # T4 - if (CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 11) - set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -gencode=arch=compute_87,code=sm_87") # AGX Orin, NX Orin - endif() -diff --git a/cmake/deps.txt b/cmake/deps.txt -index 9219f16be0..e1559bd3da 100644 ---- cmake/deps.txt -+++ cmake/deps.txt -@@ -27,7 +27,7 @@ flatbuffers;https://github.com/google/flatbuffers/archive/refs/tags/v23.5.26.zip - fp16;https://github.com/Maratyszcza/FP16/archive/0a92994d729ff76a58f692d3028ca1b64b145d91.zip;b985f6985a05a1c03ff1bb71190f66d8f98a1494 - fxdiv;https://github.com/Maratyszcza/FXdiv/archive/63058eff77e11aa15bf531df5dd34395ec3017c8.zip;a5658f4036402dbca7cebee32be57fb8149811e1 - google_benchmark;https://github.com/google/benchmark/archive/refs/tags/v1.8.5.zip;cd47d3d272faf353600c8cc2fdec2b52d6f69177 --google_nsync;https://github.com/google/nsync/archive/refs/tags/1.26.0.zip;5e7c00ef6bf5b787386fc040067903ec774e2752 -+google_nsync;https://github.com/amikhail48/nsync/archive/refs/tags/1.29.3.zip;1cdfb3b740dadf9a6cc6d6b65976d31f9d9c2900 - googletest;https://github.com/google/googletest/archive/refs/tags/v1.15.0.zip;9d2d0af8d77ac726ea55d44a8fa727ec98311349 - #xnnpack 2024.09.04 - googlexnnpack;https://github.com/google/XNNPACK/archive/309b75c9e56e0a674bf78d59872ce131f814dfb6.zip;39FA5259EAEACE0547284B63D5CEDC4F05553F5A -diff --git a/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx2_int8_blklen32.h b/onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx2_int8_blklen32.h -index af6f52090a..37ae94f1ae 100644 ---- onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx2_int8_blklen32.h -+++ onnxruntime/core/mlas/lib/sqnbitgemm_kernel_avx2_int8_blklen32.h -@@ -6,10 +6,11 @@ - #include "sqnbitgemm.h" - #include "sqnbitgemm_kernel_avx_common.h" - -+#pragma GCC diagnostic push -+#pragma GCC diagnostic ignored "-Warray-bounds" - - MLAS_FORCEINLINE void -@@ -1044,6 +1050,7 @@ MlasQ4Int8TileGemmKernelBlkLen32Avx2( - BiasPtr += BiasPtr != nullptr ? 1 : 0; - SumPtr += 1; - } - } // m - return CountM; - } -+#pragma GCC diagnostic pop diff --git a/third_party/patch/onnx_ext.patch b/third_party/patch/onnx_ext.patch deleted file mode 100644 index ef29f876..00000000 --- a/third_party/patch/onnx_ext.patch +++ /dev/null @@ -1,763 +0,0 @@ -diff --git a/operators/string_tensor.cc b/operators/string_tensor.cc -index 3d49e64..84975f6 100644 ---- operators/string_tensor.cc -+++ operators/string_tensor.cc -@@ -1,6 +1,6 @@ - // Copyright (c) Microsoft Corporation. All rights reserved. - // Licensed under the MIT License. --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_tensor.h" - - void GetTensorMutableDataString(const OrtApi& api, OrtW::CustomOpApi& ort, OrtKernelContext* context, -diff --git a/operators/string_utils.cc b/operators/string_utils_onnx.cc -similarity index 99% -rename from operators/string_utils.cc -rename to operators/string_utils_onnx.cc -index ecb6713..91dbe76 100644 ---- operators/string_utils.cc -+++ operators/string_utils_onnx.cc -@@ -2,7 +2,7 @@ - #include "farmhash.h" - #endif - --#include "string_utils.h" -+#include "string_utils_onnx.h" - - std::vector SplitString(const std::string_view& str, const std::string_view& seps, bool remove_empty_entries) { - std::vector result; -diff --git a/operators/string_utils.h b/operators/string_utils_onnx.h -similarity index 89% -rename from operators/string_utils.h -rename to operators/string_utils_onnx.h -index 5653fbd..6556666 100644 ---- operators/string_utils.h -+++ operators/string_utils_onnx.h -@@ -4,7 +4,6 @@ - #include - #include - #include --#include "ocos.h" - - template - inline void MakeStringInternal(std::ostringstream& ss, const T& t) noexcept { -@@ -23,11 +22,6 @@ inline void MakeStringInternal(std::ostringstream& ss, const std::vector --inline void MakeStringInternal(std::ostringstream& ss, const OrtTensorDimensions& t) noexcept { -- MakeStringInternal(ss, static_cast&>(t)); --} -- - template <> - inline void MakeStringInternal(std::ostringstream& ss, const std::vector& t) noexcept { - ss << "["; -diff --git a/operators/tokenizer/basic_tokenizer.cc b/operators/tokenizer/basic_tokenizer.cc -index 324a774..00eac2b 100644 ---- operators/tokenizer/basic_tokenizer.cc -+++ operators/tokenizer/basic_tokenizer.cc -@@ -1,9 +1,8 @@ - // Copyright (c) Microsoft Corporation. All rights reserved. - // Licensed under the MIT License. - --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "basic_tokenizer.hpp" --#include "string_tensor.h" - #include - #include - #include -@@ -81,52 +80,3 @@ std::vector BasicTokenizer::Tokenize(ustring text) { - push_current_token_and_clear(); - return result; - } -- --KernelBasicTokenizer::KernelBasicTokenizer(const OrtApi& api, const OrtKernelInfo* info) : BaseKernel(api, info) { -- bool do_lower_case = TryToGetAttributeWithDefault("do_lower_case", true); -- bool tokenize_chinese_chars = TryToGetAttributeWithDefault("tokenize_chinese_chars", true); -- bool strip_accents = TryToGetAttributeWithDefault("strip_accents", false); -- bool tokenize_punctuation = TryToGetAttributeWithDefault("tokenize_punctuation", false); -- bool remove_control_chars = TryToGetAttributeWithDefault("remove_control_chars", true); -- -- tokenizer_ = std::make_shared(do_lower_case, tokenize_chinese_chars, strip_accents, tokenize_punctuation, remove_control_chars); --} -- --void KernelBasicTokenizer::Compute(OrtKernelContext* context) { -- // Setup inputs -- const OrtValue* input = ort_.KernelContext_GetInput(context, 0); -- std::vector input_data; -- GetTensorMutableDataString(api_, ort_, context, input, input_data); -- -- OrtTensorDimensions dimensions(ort_, input); -- if (dimensions.size() != 1 && dimensions[0] != 1) { -- ORTX_CXX_API_THROW("[BasicTokenizer]: only support string scalar.", ORT_INVALID_GRAPH); -- } -- -- OrtValue* output = ort_.KernelContext_GetOutput(context, 0, dimensions.data(), dimensions.size()); -- std::vector result = tokenizer_->Tokenize(ustring(input_data[0])); -- -- FillTensorDataString(api_, ort_, context, result, output); --} -- --void* CustomOpBasicTokenizer::CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const { -- return CreateKernelImpl(api, info); --}; -- --const char* CustomOpBasicTokenizer::GetName() const { return "BasicTokenizer"; }; -- --size_t CustomOpBasicTokenizer::GetInputTypeCount() const { -- return 1; --}; -- --ONNXTensorElementDataType CustomOpBasicTokenizer::GetInputType(size_t /*index*/) const { -- return ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING; --}; -- --size_t CustomOpBasicTokenizer::GetOutputTypeCount() const { -- return 1; --}; -- --ONNXTensorElementDataType CustomOpBasicTokenizer::GetOutputType(size_t /*index*/) const { -- return ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING; --}; -diff --git a/operators/tokenizer/basic_tokenizer.hpp b/operators/tokenizer/basic_tokenizer.hpp -index 046499e..9fd6f1a 100644 ---- operators/tokenizer/basic_tokenizer.hpp -+++ operators/tokenizer/basic_tokenizer.hpp -@@ -3,8 +3,7 @@ - - #pragma once - --#include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "ustring.h" - - class BasicTokenizer { -@@ -19,19 +18,3 @@ class BasicTokenizer { - bool tokenize_punctuation_; - bool remove_control_chars_; - }; -- --struct KernelBasicTokenizer : BaseKernel { -- KernelBasicTokenizer(const OrtApi& api, const OrtKernelInfo* info); -- void Compute(OrtKernelContext* context); -- private: -- std::shared_ptr tokenizer_; --}; -- --struct CustomOpBasicTokenizer : OrtW::CustomOpBase { -- void* CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const; -- const char* GetName() const; -- size_t GetInputTypeCount() const; -- ONNXTensorElementDataType GetInputType(size_t index) const; -- size_t GetOutputTypeCount() const; -- ONNXTensorElementDataType GetOutputType(size_t index) const; --}; -diff --git a/operators/tokenizer/bert_tokenizer.cc b/operators/tokenizer/bert_tokenizer.cc -index b860ba6..9f43c5e 100644 ---- operators/tokenizer/bert_tokenizer.cc -+++ operators/tokenizer/bert_tokenizer.cc -@@ -33,7 +33,8 @@ int32_t BertTokenizerVocab::FindTokenId(const ustring& token) { - - auto it = vocab_.find(utf8_token); - if (it == vocab_.end()) { -- ORTX_CXX_API_THROW("[BertTokenizerVocab]: can not find tokens: " + std::string(token), ORT_RUNTIME_EXCEPTION); -+ std::cout << "[BertTokenizerVocab]: can not find tokens: " + std::string(token); -+ return -1; - } - - return it->second; -@@ -276,138 +277,3 @@ TruncateStrategy::TruncateStrategy(std::string_view strategy_name) : strategy_(T - } - } - --KernelBertTokenizer::KernelBertTokenizer(const OrtApi& api, const OrtKernelInfo* info) : BaseKernel(api, info) { -- std::string vocab = ort_.KernelInfoGetAttribute(info, "vocab_file"); -- bool do_lower_case = TryToGetAttributeWithDefault("do_lower_case", true); -- bool do_basic_tokenize = TryToGetAttributeWithDefault("do_basic_tokenize", true); -- std::string unk_token = TryToGetAttributeWithDefault("unk_token", std::string("[UNK]")); -- std::string sep_token = TryToGetAttributeWithDefault("sep_token", std::string("[SEP]")); -- std::string pad_token = TryToGetAttributeWithDefault("pad_token", std::string("[PAD]")); -- std::string cls_token = TryToGetAttributeWithDefault("cls_token", std::string("[CLS]")); -- std::string mask_token = TryToGetAttributeWithDefault("mask_token", std::string("[MASK]")); -- bool tokenize_chinese_chars = TryToGetAttributeWithDefault("tokenize_chinese_chars", true); -- bool strip_accents = TryToGetAttributeWithDefault("strip_accents", false); -- std::string suffix_indicator = TryToGetAttributeWithDefault("suffix_indicator", std::string("##")); -- std::string truncation_strategy_name = TryToGetAttributeWithDefault("truncation_strategy_name", std::string("longest_first")); -- int32_t max_len = static_cast(TryToGetAttributeWithDefault("max_length", int64_t(-1))); -- -- tokenizer_ = std::make_unique( -- vocab, do_lower_case, do_basic_tokenize, ustring(unk_token), -- ustring(sep_token), ustring(pad_token), ustring(cls_token), -- ustring(mask_token), tokenize_chinese_chars, strip_accents, -- ustring(suffix_indicator), max_len, truncation_strategy_name); --} -- --void KernelBertTokenizer::Compute(OrtKernelContext* context) { -- // Setup inputs -- const OrtValue* input = ort_.KernelContext_GetInput(context, 0); -- std::vector input_data; -- GetTensorMutableDataString(api_, ort_, context, input, input_data); -- -- if (input_data.size() != 1 && input_data.size() != 2) { -- ORTX_CXX_API_THROW("[BertTokenizer]: only support one or two query.", ORT_INVALID_GRAPH); -- } -- std::vector input_ids; -- std::vector token_type_ids; -- -- if (input_data.size() == 1) { -- std::vector tokens = tokenizer_->Tokenize(ustring(input_data[0])); -- std::vector encoded = tokenizer_->Encode(tokens); -- tokenizer_->Truncate(encoded); -- input_ids = tokenizer_->AddSpecialToken(encoded); -- token_type_ids = tokenizer_->GenerateTypeId(encoded); -- } else { -- std::vector tokens1 = tokenizer_->Tokenize(ustring(input_data[0])); -- std::vector tokens2 = tokenizer_->Tokenize(ustring(input_data[1])); -- std::vector encoded1 = tokenizer_->Encode(tokens1); -- std::vector encoded2 = tokenizer_->Encode(tokens2); -- input_ids = tokenizer_->AddSpecialToken(encoded1, encoded2); -- token_type_ids = tokenizer_->GenerateTypeId(encoded1, encoded2); -- } -- -- std::vector attention_mask(input_ids.size(), 1); -- -- std::vector output_dim{static_cast(input_ids.size())}; -- -- SetOutput(context, 0, output_dim, input_ids); -- SetOutput(context, 1, output_dim, token_type_ids); -- SetOutput(context, 2, output_dim, attention_mask); --} -- --void* CustomOpBertTokenizer::CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const { -- return CreateKernelImpl(api, info); --} -- --const char* CustomOpBertTokenizer::GetName() const { return "BertTokenizer"; } -- --size_t CustomOpBertTokenizer::GetInputTypeCount() const { -- return 1; --} -- --ONNXTensorElementDataType CustomOpBertTokenizer::GetInputType(size_t /* index */) const { -- return ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING; --} -- --size_t CustomOpBertTokenizer::GetOutputTypeCount() const { -- return 3; --} -- --ONNXTensorElementDataType CustomOpBertTokenizer::GetOutputType(size_t /* index */) const { -- return ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64; --} -- --KernelHfBertTokenizer::KernelHfBertTokenizer(const OrtApi& api, const OrtKernelInfo* info) : KernelBertTokenizer(api, info) {} -- --void KernelHfBertTokenizer::Compute(OrtKernelContext* context) { -- // Setup inputs -- const OrtValue *const input = ort_.KernelContext_GetInput(context, 0); -- std::vector input_data; -- GetTensorMutableDataString(api_, ort_, context, input, input_data); -- -- if (input_data.size() != 2) { -- ORTX_CXX_API_THROW("[HfBertTokenizer]: Support only two input strings.", ORT_INVALID_GRAPH); -- } -- -- std::vector tokens1 = tokenizer_->Tokenize(ustring(input_data[0])); -- std::vector tokens2 = tokenizer_->Tokenize(ustring(input_data[1])); -- std::vector encoded1 = tokenizer_->Encode(tokens1); -- std::vector encoded2 = tokenizer_->Encode(tokens2); -- std::vector input_ids = tokenizer_->AddSpecialToken(encoded1, encoded2); -- std::vector token_type_ids = tokenizer_->GenerateTypeId(encoded1, encoded2); -- std::vector attention_mask(input_ids.size(), 1LL); -- -- const std::vector outer_dims{1LL, static_cast(input_ids.size())}; -- const std::vector inner_dims{1LL}; -- for (int32_t i = 0; i < 3; ++i) { -- OrtValue* const value = ort_.KernelContext_GetOutput(context, i, outer_dims.data(), outer_dims.size()); -- OrtTensorTypeAndShapeInfo *const info = ort_.GetTensorTypeAndShape(value); -- ort_.SetDimensions(info, inner_dims.data(), inner_dims.size()); -- ort_.ReleaseTensorTypeAndShapeInfo(info); -- } -- -- SetOutput(context, 0, outer_dims, input_ids); -- SetOutput(context, 1, outer_dims, attention_mask); -- SetOutput(context, 2, outer_dims, token_type_ids); --} -- --void* CustomOpHfBertTokenizer::CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const { -- return CreateKernelImpl(api, info); --} -- --const char* CustomOpHfBertTokenizer::GetName() const { return "HfBertTokenizer"; } -- --size_t CustomOpHfBertTokenizer::GetInputTypeCount() const { -- return 1; --} -- --ONNXTensorElementDataType CustomOpHfBertTokenizer::GetInputType(size_t /* index */) const { -- return ONNX_TENSOR_ELEMENT_DATA_TYPE_STRING; --} -- --size_t CustomOpHfBertTokenizer::GetOutputTypeCount() const { -- return 3; --} -- --ONNXTensorElementDataType CustomOpHfBertTokenizer::GetOutputType(size_t /* index */) const { -- return ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64; --} -diff --git a/operators/tokenizer/bert_tokenizer.hpp b/operators/tokenizer/bert_tokenizer.hpp -index 6dfcd84..10565e4 100644 ---- operators/tokenizer/bert_tokenizer.hpp -+++ operators/tokenizer/bert_tokenizer.hpp -@@ -3,12 +3,11 @@ - - #pragma once - -+#include - #include - #include --#include "ocos.h" - #include "ustring.h" --#include "string_utils.h" --#include "string_tensor.h" -+#include "string_utils_onnx.h" - #include "basic_tokenizer.hpp" - - class BertTokenizerVocab final { -@@ -89,33 +88,4 @@ class BertTokenizer final { - std::unique_ptr wordpiece_tokenizer_; - }; - --struct KernelBertTokenizer : BaseKernel { -- KernelBertTokenizer(const OrtApi& api, const OrtKernelInfo* info); -- void Compute(OrtKernelContext* context); - -- protected: -- std::unique_ptr tokenizer_; --}; -- --struct CustomOpBertTokenizer : OrtW::CustomOpBase { -- void* CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const; -- const char* GetName() const; -- size_t GetInputTypeCount() const; -- ONNXTensorElementDataType GetInputType(size_t index) const; -- size_t GetOutputTypeCount() const; -- ONNXTensorElementDataType GetOutputType(size_t index) const; --}; -- --struct KernelHfBertTokenizer : KernelBertTokenizer { -- KernelHfBertTokenizer(const OrtApi& api, const OrtKernelInfo* info); -- void Compute(OrtKernelContext* context); --}; -- --struct CustomOpHfBertTokenizer : OrtW::CustomOpBase { -- void* CreateKernel(const OrtApi& api, const OrtKernelInfo* info) const; -- const char* GetName() const; -- size_t GetInputTypeCount() const; -- ONNXTensorElementDataType GetInputType(size_t index) const; -- size_t GetOutputTypeCount() const; -- ONNXTensorElementDataType GetOutputType(size_t index) const; --}; -diff --git a/operators/cv2/imgcodecs/imdecode.hpp b/operators/cv2/imgcodecs/imdecode.hpp -index 846e8a0..7a0a9b1 100644 ---- operators/cv2/imgcodecs/imdecode.hpp -+++ operators/cv2/imgcodecs/imdecode.hpp -@@ -8,7 +8,7 @@ - #include - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - #include - -diff --git a/operators/math/segment_extraction.hpp b/operators/math/segment_extraction.hpp -index 45c4eec..3450f07 100644 ---- operators/math/segment_extraction.hpp -+++ operators/math/segment_extraction.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelSegmentExtraction : BaseKernel { - KernelSegmentExtraction(const OrtApi& api); -diff --git a/operators/math/segment_sum.hpp b/operators/math/segment_sum.hpp -index 9ae20ad..7cf095d 100644 ---- operators/math/segment_sum.hpp -+++ operators/math/segment_sum.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelSegmentSum : BaseKernel { - KernelSegmentSum(const OrtApi& api); -diff --git a/operators/text/masked_fill.hpp b/operators/text/masked_fill.hpp -index 35d408a..37971ae 100644 ---- operators/text/masked_fill.hpp -+++ operators/text/masked_fill.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include - - struct KernelMaskedFill : BaseKernel { -diff --git a/operators/text/op_equal.hpp b/operators/text/op_equal.hpp -index cbe1099..22f24dc 100644 ---- operators/text/op_equal.hpp -+++ operators/text/op_equal.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringEqual : BaseKernel { - KernelStringEqual(const OrtApi& api); -diff --git a/operators/text/op_equal_impl.hpp b/operators/text/op_equal_impl.hpp -index 89ad075..38e1776 100644 ---- operators/text/op_equal_impl.hpp -+++ operators/text/op_equal_impl.hpp -@@ -3,7 +3,7 @@ - #pragma once - #include - #include --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_tensor.h" - - template -diff --git a/operators/text/op_ragged_tensor.cc b/operators/text/op_ragged_tensor.cc -index 6314001..4e9b88d 100644 ---- operators/text/op_ragged_tensor.cc -+++ operators/text/op_ragged_tensor.cc -@@ -1,6 +1,6 @@ - // Copyright (c) Microsoft Corporation. All rights reserved. - // Licensed under the MIT License. --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_tensor.h" - #include "op_ragged_tensor.hpp" - -diff --git a/operators/text/re2_strings/string_regex_replace.hpp b/operators/text/re2_strings/string_regex_replace.hpp -index f12d002..4320953 100644 ---- operators/text/re2_strings/string_regex_replace.hpp -+++ operators/text/re2_strings/string_regex_replace.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringRegexReplace : BaseKernel { - KernelStringRegexReplace(const OrtApi& api, const OrtKernelInfo* info); -diff --git a/operators/text/re2_strings/string_regex_split.hpp b/operators/text/re2_strings/string_regex_split.hpp -index 9900269..959f4db 100644 ---- operators/text/re2_strings/string_regex_split.hpp -+++ operators/text/re2_strings/string_regex_split.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - // See https://github.com/tensorflow/text/blob/master/docs/api_docs/python/text/regex_split_with_offsets.md. - struct KernelStringRegexSplitWithOffsets : BaseKernel { -diff --git a/operators/text/string_concat.hpp b/operators/text/string_concat.hpp -index 597cc53..623f3e9 100644 ---- operators/text/string_concat.hpp -+++ operators/text/string_concat.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringConcat : BaseKernel { - KernelStringConcat(const OrtApi& api); -diff --git a/operators/text/string_ecmaregex_replace.hpp b/operators/text/string_ecmaregex_replace.hpp -index d52a02f..86eb24a 100644 ---- operators/text/string_ecmaregex_replace.hpp -+++ operators/text/string_ecmaregex_replace.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringECMARegexReplace : BaseKernel { - KernelStringECMARegexReplace(const OrtApi& api, const OrtKernelInfo* info); -diff --git a/operators/text/string_ecmaregex_split.hpp b/operators/text/string_ecmaregex_split.hpp -index 78cd599..6bf6029 100644 ---- operators/text/string_ecmaregex_split.hpp -+++ operators/text/string_ecmaregex_split.hpp -@@ -5,7 +5,7 @@ - - #include - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - // See https://github.com/tensorflow/text/blob/master/docs/api_docs/python/text/regex_split_with_offsets.md. - struct KernelStringECMARegexSplitWithOffsets : BaseKernel { -diff --git a/operators/text/string_hash.hpp b/operators/text/string_hash.hpp -index 701066f..05cb024 100644 ---- operators/text/string_hash.hpp -+++ operators/text/string_hash.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringHash : BaseKernel { - KernelStringHash(const OrtApi& api); -diff --git a/operators/text/string_join.hpp b/operators/text/string_join.hpp -index 637f84c..05c90b6 100644 ---- operators/text/string_join.hpp -+++ operators/text/string_join.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringJoin : BaseKernel { - KernelStringJoin(const OrtApi& api); -diff --git a/operators/text/string_length.hpp b/operators/text/string_length.hpp -index 592a3b2..402a9c1 100644 ---- operators/text/string_length.hpp -+++ operators/text/string_length.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringLength : BaseKernel { - KernelStringLength(const OrtApi& api); -diff --git a/operators/text/string_lower.hpp b/operators/text/string_lower.hpp -index a9421a1..312f11a 100644 ---- operators/text/string_lower.hpp -+++ operators/text/string_lower.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringLower : BaseKernel { - KernelStringLower(const OrtApi& api); -diff --git a/operators/text/string_mapping.hpp b/operators/text/string_mapping.hpp -index 1afcb66..0b23ac5 100644 ---- operators/text/string_mapping.hpp -+++ operators/text/string_mapping.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include - - struct KernelStringMapping : BaseKernel { -diff --git a/operators/text/string_split.hpp b/operators/text/string_split.hpp -index f2c2737..be6aad5 100644 ---- operators/text/string_split.hpp -+++ operators/text/string_split.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringSplit : BaseKernel { - KernelStringSplit(const OrtApi& api); -diff --git a/operators/text/string_to_vector.hpp b/operators/text/string_to_vector.hpp -index 3d4d62f..89a1fb2 100644 ---- operators/text/string_to_vector.hpp -+++ operators/text/string_to_vector.hpp -@@ -7,7 +7,7 @@ - #include - #include - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - - class StringToVectorImpl { -diff --git a/operators/text/string_upper.hpp b/operators/text/string_upper.hpp -index eb9d6e2..afe2540 100644 ---- operators/text/string_upper.hpp -+++ operators/text/string_upper.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - struct KernelStringUpper : BaseKernel { - KernelStringUpper(const OrtApi& api); -diff --git a/operators/text/vector_to_string.cc b/operators/text/vector_to_string.cc -index be1eb19..578dc5f 100644 ---- operators/text/vector_to_string.cc -+++ operators/text/vector_to_string.cc -@@ -1,6 +1,6 @@ - #include - #include "farmhash.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "vector_to_string.hpp" - #include "string_tensor.h" - -diff --git a/operators/text/vector_to_string.hpp b/operators/text/vector_to_string.hpp -index d558070..bb78cf0 100644 ---- operators/text/vector_to_string.hpp -+++ operators/text/vector_to_string.hpp -@@ -7,7 +7,7 @@ - #include - #include - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - namespace std { - -diff --git a/operators/tokenizer/blingfire_sentencebreaker.hpp b/operators/tokenizer/blingfire_sentencebreaker.hpp -index 623f72b..d11b7b4 100644 ---- operators/tokenizer/blingfire_sentencebreaker.hpp -+++ operators/tokenizer/blingfire_sentencebreaker.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - extern "C" const int TextToSentencesWithOffsetsWithModel( - const char* pInUtf8Str, int InUtf8StrByteCount, -diff --git a/operators/tokenizer/clip_tokenizer.hpp b/operators/tokenizer/clip_tokenizer.hpp -index aec7874..896d3f7 100644 ---- operators/tokenizer/clip_tokenizer.hpp -+++ operators/tokenizer/clip_tokenizer.hpp -@@ -1,7 +1,7 @@ - #include - #include "ocos.h" - #include "ustring.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - class VocabData; - -diff --git a/operators/tokenizer/sentencepiece_decoder.hpp b/operators/tokenizer/sentencepiece_decoder.hpp -index 8b71346..4f3e340 100644 ---- operators/tokenizer/sentencepiece_decoder.hpp -+++ operators/tokenizer/sentencepiece_decoder.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_tensor.h" - #include "sentencepiece_processor.h" - #include "sentencepiece_model.pb.h" -diff --git a/operators/tokenizer/sentencepiece_tokenizer.hpp b/operators/tokenizer/sentencepiece_tokenizer.hpp -index 97327ab..83c71c1 100644 ---- operators/tokenizer/sentencepiece_tokenizer.hpp -+++ operators/tokenizer/sentencepiece_tokenizer.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "sentencepiece_processor.h" - - -diff --git a/operators/tokenizer/wordpiece_tokenizer.hpp b/operators/tokenizer/wordpiece_tokenizer.hpp -index 5f6eb7e..0f30014 100644 ---- operators/tokenizer/wordpiece_tokenizer.hpp -+++ operators/tokenizer/wordpiece_tokenizer.hpp -@@ -7,7 +7,7 @@ - #include - #include "ocos.h" - #include "ustring.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_tensor.h" - - struct KernelWordpieceTokenizer : BaseKernel { -diff --git a/operators/vision/decode_image.hpp b/operators/vision/decode_image.hpp -index c21a41a..bae62df 100644 ---- operators/vision/decode_image.hpp -+++ operators/vision/decode_image.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - #include - -diff --git a/operators/vision/encode_image.hpp b/operators/vision/encode_image.hpp -index f637eef..373eb71 100644 ---- operators/vision/encode_image.hpp -+++ operators/vision/encode_image.hpp -@@ -4,7 +4,7 @@ - #pragma once - - #include "ocos.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - - #include - -diff --git a/operators/text/string_to_vector.cc b/operators/text/string_to_vector.cc -index e924a9d..434be84 100644 ---- operators/text/string_to_vector.cc -+++ operators/text/string_to_vector.cc -@@ -1,6 +1,6 @@ - #include - #include "farmhash.h" --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_to_vector.hpp" - #include "string_tensor.h" - -diff --git a/pyop/pyfunc.cc b/pyop/pyfunc.cc -index fd9634c..e78b238 100644 ---- pyop/pyfunc.cc -+++ pyop/pyfunc.cc -@@ -16,7 +16,7 @@ - #include - #include - #include --#include "string_utils.h" -+#include "string_utils_onnx.h" - #include "string_tensor.h" - #include "pykernel.h" - diff --git a/tools/BUILD b/tools/BUILD index d4d080f9..c4608db2 100644 --- a/tools/BUILD +++ b/tools/BUILD @@ -9,7 +9,7 @@ cc_binary( "//targets/extras:imported_openvx_extras", "//targets/liteRT:imported_openvx_liteRT", "//targets/opencl:imported_openvx_opencl", - # "//targets/onnxRT:imported_openvx_onnxRT", + "//targets/onnxRT:imported_openvx_onnxRT", "//targets/executorch:imported_openvx_torch", ], linkopts = select({