Milestone 1: Added Fusion G3 NN library with kernels related to add, mul, quantize…

backends/cadence/CMakeLists.txt

@@ -76,7 +76,12 @@ endif()
 if(EXECUTORCH_NNLIB_OPT)
   set(TARGET_DIR hifi)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/third-party/nnlib)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)
+endif()
+
+if(EXECUTORCH_FUSION_G3_OPT)
+  set(TARGET_DIR fusion_g3)
+  add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/third-party/nnlib)
 endif()
 
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/operators)
-add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)

backends/cadence/aot/functions_fusion_g3.yaml

@@ -0,0 +1,119 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This yaml file contains operators that are also defined by the ATen library.
+# For lean mode:
+#   - Codegen'd target `executorch_generated_lib` will be reading all the information
+#     from this file, including operator schema and kernel metadata.
+#   - Selective build target `codegen:executorch_defined_ops` now is selecting all the
+#     operators in this file, by dumping all the op names into `selected_operators.yaml`.
+#
+# See the README.md file in executorch/kernels/portable for a description of the syntax used
+# by this file.
+
+
+# aten ops
+- op: _to_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::to_copy_out
+
+- op: _softmax.out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::softmax_out
+
+- op: add.out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::add_out
+
+- op: add.Scalar_out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::add_scalar_out
+
+- op: bmm.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::bmm_out
+
+- op: cat.out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::cat_out
+
+- op: clone.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::clone_out
+
+- op: div.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::div_out
+
+- op: div.out_mode
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::div_out_mode
+
+- op: embedding.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::embedding_out
+
+- op: full.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::full_out
+
+- op: mul.out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::mul_out
+
+- op: mul.Scalar_out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::mul_scalar_out
+
+- op: permute_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::permute_copy_out
+
+- op: sigmoid.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::sigmoid_out
+
+- op: slice_copy.Tensor_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::slice_copy_Tensor_out
+
+- op: split_with_sizes_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::split_with_sizes_copy_out
+
+- op: sub.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::sub_out
+
+- op: view_copy.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::view_copy_out
+
+- op: where.self_out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::where_out
+
+- op: native_layer_norm.out
+  kernels:
+    - arg_meta: null
+      kernel_name: impl::FusionG3::native_layer_norm_out
+

backends/cadence/fusion_g3/operators/CMakeLists.txt

@@ -0,0 +1,85 @@
+# 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.
+
+cmake_minimum_required(VERSION 3.19)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
+
+include(${EXECUTORCH_ROOT}/build/Utils.cmake)
+include(${EXECUTORCH_ROOT}/build/Codegen.cmake)
+
+if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+endif()
+
+# ATen compliant ops that are needed to run this model.
+set(_aten_ops__srcs
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/activation_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/copy_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/broadcast_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/index_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/kernel_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/matmul_ops_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/reduce_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/repeat_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/slice_util.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_add.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_mul.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_cat.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_softmax.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_native_layer_norm.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_quantize.cpp"
+    "${CMAKE_CURRENT_SOURCE_DIR}/op_dequantize.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_bmm.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_clone.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_div.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_embedding.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_full.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_permute_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sigmoid.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_slice_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_split_with_sizes_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_sub.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_to_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_view_copy.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_where.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/dtype_util.cpp"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/normalization_ops_util.cpp"
+)
+add_library(aten_ops_cadence ${_aten_ops__srcs})
+target_link_libraries(aten_ops_cadence PUBLIC executorch)
+target_link_libraries(aten_ops_cadence PRIVATE xa_nnlib)
+
+# Let files say "include <executorch/path/to/header.h>".
+set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+
+target_include_directories(
+  aten_ops_cadence PUBLIC ${ROOT_DIR}/.. ${CMAKE_BINARY_DIR}
+                          ${_common_include_directories}
+                          ${EXECUTORCH_ROOT}/backends/cadence/fusion_g3/third-party/nnlib/nnlib-FuG3/xa_nnlib/algo/common/include/
+                          ${EXECUTORCH_ROOT}/backends/cadence/fusion_g3/third-party/nnlib/nnlib-FuG3/xa_nnlib/include/nnlib
+                          ${EXECUTORCH_ROOT}/backends/cadence/fusion_g3/third-party/nnlib/nnlib-FuG3/xa_nnlib/include
+                          ${EXECUTORCH_ROOT}/backends/cadence/fusion_g3/third-party/nnlib/nnlib-FuG3/xa_nnlib/algo/kernels/tables/include
+)
+
+# Generate C++ bindings to register kernels into both PyTorch (for AOT) and
+# Executorch (for runtime). Here select all ops in functions.yaml
+gen_selected_ops(
+  LIB_NAME "cadence_ops_lib" OPS_SCHEMA_YAML
+  "${CMAKE_CURRENT_LIST_DIR}/../../aot/functions_fusion_g3.yaml" "" ""
+)
+generate_bindings_for_kernels(
+  LIB_NAME "cadence_ops_lib" OPS_SCHEMA_YAML FUNCTIONS_YAML
+  ${CMAKE_CURRENT_SOURCE_DIR}/../../aot/functions_fusion_g3.yaml
+)
+message("Generated files ${gen_command_sources}")
+
+gen_operators_lib(
+  LIB_NAME "cadence_ops_lib" KERNEL_LIBS DEPS aten_ops_cadence
+)