Update on "[Executorch][BE] Rename sdpa_with_kv_cache.py to custom_ops.py"

Because now we have more than sdpa_with_kv_cache in it

Differential Revision: [D66269486](https://our.internmc.facebook.com/intern/diff/D66269486/)

[ghstack-poisoned]

Author: kimishpatel

.github/scripts/extract_benchmark_results.py

@@ -451,7 +451,7 @@ def main() -> None:
             continue
 
         output_dir = os.path.join(args.output_dir, schema)
-        os.mkdir(output_dir)
+        os.makedirs(output_dir, exist_ok=True)
 
         output_file = os.path.basename(args.artifacts)
         with open(f"{output_dir}/{output_file}", "w") as f:

CMakeLists.txt

@@ -200,8 +200,6 @@ option(EXECUTORCH_BUILD_EXTENSION_TENSOR "Build the Tensor extension" OFF)
 
 option(EXECUTORCH_BUILD_EXTENSION_TRAINING "Build the training extension" OFF)
 
-option(EXECUTORCH_BUILD_GTESTS "Build googletest based test binaries" OFF)
-
 option(EXECUTORCH_BUILD_MPS "Build the MPS backend" OFF)
 
 option(EXECUTORCH_BUILD_NEURON "Build the backends/mediatek directory" OFF)
@@ -216,6 +214,8 @@ option(EXECUTORCH_BUILD_KERNELS_QUANTIZED "Build the quantized kernels" OFF)
 
 option(EXECUTORCH_BUILD_DEVTOOLS "Build the ExecuTorch Developer Tools")
 
+option(EXECUTORCH_BUILD_TESTS "Build CMake-based unit tests" OFF)
+
 option(EXECUTORCH_NNLIB_OPT "Build Cadence backend Hifi nnlib kernel" OFF)
 
 option(EXECUTORCH_CADENCE_CPU_RUNNER "Build Cadence backend CPU runner" OFF)
@@ -330,6 +330,10 @@ if(EXECUTORCH_BUILD_PTHREADPOOL)
   )
 endif()
 
+if(EXECUTORCH_BUILD_TESTS)
+  include(CTest)
+endif()
+
 if(NOT PYTHON_EXECUTABLE)
   resolve_python_executable()
 endif()
@@ -625,7 +629,7 @@ cmake_dependent_option(
 )
 
 # Add googletest if any test targets should be built
-if(EXECUTORCH_BUILD_GTESTS)
+if(BUILD_TESTING)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/googletest)
 endif()
 
@@ -829,5 +833,7 @@ if(EXECUTORCH_BUILD_VULKAN)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/vulkan)
 endif()
 
+include(Test.cmake)
+
 # Print all summary
 executorch_print_configuration_summary()

Test.cmake

@@ -0,0 +1,29 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+#
+# A helper CMake file to trigger C++ unit tests.
+#
+
+if(BUILD_TESTING)
+  # This contains the list of tests which are always built
+  add_subdirectory(extension/evalue_util/test)
+  add_subdirectory(extension/kernel_util/test)
+  add_subdirectory(extension/memory_allocator/test)
+  add_subdirectory(extension/parallel/test)
+  add_subdirectory(extension/pytree/test)
+  add_subdirectory(kernels/portable/cpu/util/test)
+  add_subdirectory(kernels/prim_ops/test)
+  add_subdirectory(kernels/test)
+  add_subdirectory(runtime/core/exec_aten/testing_util/test)
+  add_subdirectory(runtime/core/exec_aten/util/test)
+  add_subdirectory(runtime/core/portable_type/test)
+  add_subdirectory(runtime/core/test)
+  add_subdirectory(runtime/executor/test)
+  add_subdirectory(runtime/kernel/test)
+  add_subdirectory(runtime/platform/test)
+  add_subdirectory(test/utils)
+endif()

backends/arm/_passes/arm_pass_utils.py

@@ -156,7 +156,7 @@ def get_node_arg(args: list | dict, key: int | str | type, default_value=None):
                 f"Out of bounds index {key} for getting value in args (of size {len(args)})"
             )
     elif isinstance(key, str):
-        return args.get(key, default_value)
+        return args.get(key, default_value)  # pyre-ignore[16]
     elif isclass(key):
         for arg in args:
             if isinstance(arg, key):

backends/arm/_passes/keep_dims_false_to_squeeze_pass.py

@@ -64,12 +64,17 @@ def call(self, graph_module: torch.fx.GraphModule):
                 continue
 
             sum_node = cast(torch.fx.Node, node)
-            keep_dim = get_node_arg(sum_node.args, keep_dim_index, False)
+            keep_dim = get_node_arg(
+                # pyre-ignore[6]
+                sum_node.args,
+                keep_dim_index,
+                False,
+            )
 
             if keep_dim:
                 continue
 
-            dim_list = get_node_arg(sum_node.args, 1, [0])
+            dim_list = get_node_arg(sum_node.args, 1, [0])  # pyre-ignore[6]
 
             # Add keep_dim = True arg to sum node.
             set_node_arg(sum_node, 2, True)

backends/arm/operator_support/to_copy_support.py

@@ -33,7 +33,9 @@ class ToCopySupported(SupportedTOSAOperatorCheck):
 
     @staticmethod
     def _merge_supported_types(
-        dtypes1: SupportedTypeDict, dtypes2: SupportedTypeDict
+        # pyre-ignore[11]
+        dtypes1: SupportedTypeDict,
+        dtypes2: SupportedTypeDict,
     ) -> SupportedTypeDict:
         merged_dtypes = dtypes1
         for k, v in dtypes2.items():
@@ -92,7 +94,7 @@ def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool
         if input_dtype not in supported_dtypes:
             logger.info(
                 f"Input dtype {input_val.dtype} is not supported in "
-                f"{node.target.name()}."
+                f"{node.target.name()}."  # pyre-ignore[16]
             )
             return False
 
@@ -102,7 +104,7 @@ def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool
         if output_val.dtype not in supported_dtypes[input_dtype]:
             logger.info(
                 f"Output dtype {output_val.dtype} is not supported in "
-                f"{node.target.name()} for input dtype {input_dtype}. "
+                f"{node.target.name()} for input dtype {input_dtype}. "  # pyre-ignore[16]
                 f"Supported output types: "
                 f"{''.join(str(t) for t in supported_dtypes[input_dtype])}"
             )
@@ -113,7 +115,7 @@ def is_node_supported(self, node: fx.Node, tosa_spec: TosaSpecification) -> bool
             if node.kwargs["memory_format"] in (torch.preserve_format,):
                 logger.info(
                     f"Argument 'memory_format' is not supported for "
-                    f"{node.target.name()} right now."
+                    f"{node.target.name()} right now."  # pyre-ignore[16]
                 )
                 return False
 

backends/cadence/CMakeLists.txt

@@ -23,7 +23,6 @@ include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 
 # Let files say "include <executorch/path/to/header.h>".
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)
-set(TARGET_DIR reference)
 
 if(EXECUTORCH_CADENCE_CPU_RUNNER)
   include(${EXECUTORCH_ROOT}/build/Codegen.cmake)
@@ -61,6 +60,9 @@ if(EXECUTORCH_CADENCE_CPU_RUNNER)
                                       ${_common_include_directories}
   )
 
+  set(TARGET_DIR reference)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)
+
   target_link_libraries(
   cadence_runner
   executorch

backends/cadence/aot/functions.yaml

@@ -142,6 +142,41 @@
     - arg_meta: null
       kernel_name: torch::executor::where_out
 
+- op: transpose_copy.int_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::transpose_copy_int_out
+
+- op: eq.Scalar_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::eq_scalar_out
+
+- op: logical_not.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::logical_not_out
+
+- op: any.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::any_out
+
+- op: native_group_norm.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::native_group_norm_out
+
+- op: sum.IntList_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::sum_dim_out
+
+- op: select_copy.int_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::select_copy_int_out
+
 # custom ops
 - func: cadence::quantize_per_tensor.out(Tensor input, float scale, int zero_point, int quant_min, int quant_max, ScalarType dtype, *, Tensor(a!) out) -> Tensor(a!)
   variants: function
@@ -183,3 +218,18 @@
   kernels:
     - arg_meta: null
       kernel_name: impl::reference::quantized_matmul_out
+
+- func: cadence::quantized_linear.per_tensor_out(Tensor src, Tensor weight, Tensor bias, SymInt src_zero_point, SymInt weight_zero_point, SymInt out_multiplier, SymInt out_shift, SymInt out_zero_point, Tensor? offset, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_linear_per_tensor_out
+
+- func: cadence::im2row.out(Tensor input, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride, Tensor in_zero_point, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::im2row_out
+
+- func: cadence::quantized_conv.per_tensor_out(Tensor input, Tensor weight, Tensor bias, int[] stride, SymInt[] padding, int[] dilation, int groups, int input_zero_point, int weight_zero_point, float bias_scale, float out_scale, int out_zero_point, int out_multiplier, int out_shift, bool channel_last=False, *, Tensor(a!) out) -> Tensor(a!)
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::reference::quantized_conv_per_tensor_out

backends/cadence/reference/operators/CMakeLists.txt

@@ -55,6 +55,16 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_expand_copy.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_gelu.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_empty.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_transpose_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_eq.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_logical_not.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_any.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_native_group_norm.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sum.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_select_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/dtype_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/normalization_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/select_copy_util.cpp"
 )
 add_library(aten_ops_cadence ${_aten_ops__srcs})
 target_link_libraries(aten_ops_cadence PUBLIC executorch)
@@ -78,6 +88,7 @@ add_library(
   "quantize_per_tensor.cpp"
   "dequantize_per_tensor.cpp"
   "quantized_matmul_out.cpp"
+  "im2row_out.cpp"
 )
 target_include_directories(
   custom_ops PUBLIC ${ROOT_DIR}/.. ${CMAKE_BINARY_DIR}