build cadence cpu flow as a stand-alone cmake dependency (#5555)

Summary:
Pull Request resolved: #5555
This PR
- Re-enable cadence e2e flow (quant ops, models) on CPU
- move cadence runner builder to cadence runner
- create executor cmake to build runner separately from hifi
- update to cmake builder script for the cpu runner
- add slice_util to reference ops because its missing
- disabled __attribute__((always_inline)) in reference kernel because of  {F1886927939}

Reviewed By: mcremon-meta

Differential Revision: D63284100

fbshipit-source-id: 10955f2ade74788da10aa441d164977b44e4eec5

Author: zonglinpeng

backends/cadence/cadence_runner/CMakeLists.txt

@@ -0,0 +1,72 @@
+# 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.
+
+# Set the minimum required version of CMake for this project.
+cmake_minimum_required(VERSION 3.10)
+
+if(NOT CMAKE_CXX_STANDARD)
+  set(CMAKE_CXX_STANDARD 17)
+endif()
+
+# Set the project name.
+project(cadence_backend)
+
+# Source root directory for executorch.
+if(NOT EXECUTORCH_ROOT)
+  set(EXECUTORCH_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
+endif()
+
+include(${EXECUTORCH_ROOT}/build/Utils.cmake)
+include(${EXECUTORCH_ROOT}/build/Codegen.cmake)
+
+# Let files say "include <executorch/path/to/header.h>".
+set(_common_include_directories ${EXECUTORCH_ROOT}/..)
+set(TARGET_DIR reference)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+endif()
+
+# Find prebuilt libraries. executorch package should contain portable_ops_lib,
+# etdump, bundled_program.
+find_package(executorch CONFIG REQUIRED)
+target_link_options_shared_lib(executorch)
+target_link_options_shared_lib(portable_ops_lib)
+
+target_include_directories(executorch INTERFACE ${_common_include_directories})
+
+find_package(
+  gflags REQUIRED PATHS ${CMAKE_CURRENT_BINARY_DIR}/../../../third-party
+)
+
+add_executable(cadence_runner cadence_runner.cpp)
+target_compile_options(executorch INTERFACE -DET_EVENT_TRACER_ENABLED)
+
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/operators)
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/kernels)
+
+target_include_directories(
+  etdump INTERFACE ${CMAKE_CURRENT_BINARY_DIR}/../../../devtools/include
+                   ${EXECUTORCH_ROOT}/third-party/flatcc/include
+)
+
+target_include_directories(
+  cadence_runner PUBLIC ${ROOT_DIR}/../.. ${CMAKE_BINARY_DIR}
+                        ${_common_include_directories}
+)
+
+target_link_libraries(
+  cadence_runner
+  executorch
+  gflags
+  etdump
+  extension_data_loader
+  bundled_program
+  cadence_ops_lib
+  flatccrt
+)

backends/cadence/build_cadence_runner.sh renamed to backends/cadence/cadence_runner/build_cadence_runner.sh

@@ -12,7 +12,7 @@ set -euo pipefail
 SCRIPT_DIR="$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )"
 readonly SCRIPT_DIR
 
-readonly EXECUTORCH_ROOT="${SCRIPT_DIR}/../.."
+readonly EXECUTORCH_ROOT="${SCRIPT_DIR}/../../.."
 
 # Allow overriding the number of build jobs. Default to 9.
 export CMAKE_BUILD_PARALLEL_LEVEL="${CMAKE_BUILD_PARALLEL_LEVEL:-9}"
@@ -32,8 +32,9 @@ main() {
     -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
     -DEXECUTORCH_BUILD_CPUINFO=OFF \
     -DEXECUTORCH_ENABLE_LOGGING=ON \
-    -Bcmake-out .
-  cmake --build cmake-out --target install --config Release
+    -DEXECUTORCH_NNLIB_OPT=OFF \
+    -Bcmake-out
+  cmake --build cmake-out --target install --config Release -j16
 
   local example_dir=backends/cadence
   local build_dir="cmake-out/${example_dir}"
@@ -43,7 +44,7 @@ main() {
     -DCMAKE_BUILD_TYPE=Release \
     -B"${build_dir}" \
     "${example_dir}"
-  cmake --build "${build_dir}" --config Release
+  cmake --build "${build_dir}" --config Release -j16
 
   local runner="${PWD}/${build_dir}/cadence_runner"
   if [[ ! -f "${runner}" ]]; then

backends/cadence/reference/kernels/kernels.cpp

@@ -9,6 +9,7 @@
 #include <math.h>
 #include <algorithm>
 #include <cstring>
+#include <limits>
 #include <numeric>
 
 namespace impl {
@@ -17,8 +18,7 @@ namespace kernels {
 
 // Quantize a fp32 value to an int8_t/uint8_t value
 template <typename T>
-__attribute__((always_inline)) T
-quantize(const float x, float scale, int32_t zero_point) {
+T quantize(const float x, float scale, int32_t zero_point) {
   constexpr float min_val = std::numeric_limits<T>::min();
   constexpr float max_val = std::numeric_limits<T>::max();
   float tmp = roundf(x * scale + zero_point);
@@ -40,8 +40,7 @@ void quantize(
 
 // Dequantize an int8_t/uint8_t value to an fp32 value
 template <typename T>
-__attribute__((always_inline)) float
-dequantize(const T x, float scale, int32_t zero_point) {
+float dequantize(const T x, float scale, int32_t zero_point) {
   return scale * (x - zero_point);
 }
 
@@ -60,9 +59,8 @@ void dequantize(
 
 // explicit template instantiation
 
-#define typed_quantize_val(dtype)                         \
-  template __attribute__((always_inline)) dtype quantize( \
-      const float x, float inv_scale, int32_t zero_point);
+#define typed_quantize_val(dtype) \
+  template dtype quantize(const float x, float inv_scale, int32_t zero_point);
 typed_quantize_val(int8_t);
 typed_quantize_val(uint8_t);
 typed_quantize_val(int16_t);
@@ -82,9 +80,8 @@ typed_quantize_vec(int16_t);
 typed_quantize_vec(int32_t);
 #undef typed_quantize_vec
 
-#define typed_dequantize_val(dtype)                         \
-  template __attribute__((always_inline)) float dequantize( \
-      const dtype x, float scale, int32_t zero_point);
+#define typed_dequantize_val(dtype) \
+  template float dequantize(const dtype x, float scale, int32_t zero_point);
 typed_dequantize_val(int8_t);
 typed_dequantize_val(uint8_t);
 typed_dequantize_val(int16_t);

backends/cadence/reference/operators/CMakeLists.txt

@@ -32,6 +32,7 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/matmul_ops_util.cpp"
     "${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"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/pattern/unary_ufunc_realhb_to_floath.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_bmm.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_cat.cpp"

backends/cadence/runtime/executor.py

@@ -106,7 +106,9 @@ def __init__(
         working_dir: str = "",
     ):
         self.working_dir = working_dir
-        self.executor_builder = "./backends/cadence/build_cadence_runner.sh"
+        self.executor_builder = (
+            "./backends/cadence/cadence_runner/build_cadence_runner.sh"
+        )
         self.execute_runner = "./cmake-out/backends/cadence/cadence_runner"
         self.bundled_program_path: str = "CadenceDemoModel.bpte"