fixed CPU kernel, unary_ufunc_realhbbf16_to_floathbf16, cpu flow (pyt…

…orch#5991)

Summary:
kernel of cpu was not included.
unary_ufunc_realhbbf16_to_floathbf16 was recently renamed
cpu cmake was broken due to hifi change. need to move cpu cadence runner to 1 level up to include cadence kernel and ops.

Pull Request resolved: pytorch#5991

Reviewed By: mcremon-meta

Differential Revision: D64056710

Pulled By: zonglinpeng

fbshipit-source-id: 91875312f657b78588a4e9212e02b57bf6408cad

CMakeLists.txt

@@ -201,6 +201,10 @@ option(EXECUTORCH_BUILD_KERNELS_QUANTIZED "Build the quantized kernels" OFF)
 
 option(EXECUTORCH_BUILD_DEVTOOLS "Build the ExecuTorch Developer Tools")
 
+option(EXECUTORCH_NNLIB_OPT "Build Cadence backend Hifi nnlib kernel" OFF)
+
+option(EXECUTORCH_CADENCE_CPU_RUNNER "Build Cadence backend CPU runner" OFF)
+
 option(EXECUTORCH_BUILD_SIZE_TEST "Build the size test" OFF)
 
 option(EXECUTORCH_BUILD_XNNPACK "Build the XNNPACK backend" OFF)

backends/cadence/CMakeLists.txt

@@ -25,6 +25,54 @@ include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 set(TARGET_DIR reference)
 
+if(EXECUTORCH_CADENCE_CPU_RUNNER)
+  include(${EXECUTORCH_ROOT}/build/Codegen.cmake)
+
+  if(NOT PYTHON_EXECUTABLE)
+  resolve_python_executable()
+  endif()
+
+  set(_common_compile_options -Wno-deprecated-declarations -fPIC)
+
+  # 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
+      ${EXECUTORCH_ROOT}/examples/devtools/example_runner/example_runner.cpp
+  )
+  target_compile_options(executorch INTERFACE -DET_EVENT_TRACER_ENABLED)
+
+  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}
+                                      ${_common_include_directories}
+  )
+
+  target_link_libraries(
+  cadence_runner
+  executorch
+  gflags
+  etdump
+  extension_data_loader
+  bundled_program
+  cadence_ops_lib
+  flatccrt
+  )
+endif()
+
 if(EXECUTORCH_NNLIB_OPT)
   set(TARGET_DIR hifi)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/${TARGET_DIR}/third-party/nnlib)

backends/cadence/cadence_runner/build_cadence_runner.sh → backends/cadence/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}"
@@ -25,15 +25,7 @@ main() {
     -DCMAKE_BUILD_TYPE=Release \
     -DEXECUTORCH_BUILD_DEVTOOLS=ON \
     -DEXECUTORCH_ENABLE_EVENT_TRACER=ON \
-    -DPYTHON_EXECUTABLE=python3 \
-    -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON \
-    -DEXECUTORCH_BUILD_HOST_TARGETS=ON \
-    -DEXECUTORCH_BUILD_EXECUTOR_RUNNER=OFF \
-    -DEXECUTORCH_BUILD_PTHREADPOOL=OFF \
-    -DEXECUTORCH_BUILD_CPUINFO=OFF \
-    -DEXECUTORCH_ENABLE_LOGGING=ON \
-    -DEXECUTORCH_NNLIB_OPT=OFF \
-    -Bcmake-out
+    -Bcmake-out .
   cmake --build cmake-out --target install --config Release -j16
 
   local example_dir=backends/cadence
@@ -42,6 +34,7 @@ main() {
   rm -rf ${build_dir}
   cmake -DCMAKE_PREFIX_PATH="${cmake_prefix_path}" \
     -DCMAKE_BUILD_TYPE=Release \
+    -DEXECUTORCH_CADENCE_CPU_RUNNER=ON \
     -B"${build_dir}" \
     "${example_dir}"
   cmake --build "${build_dir}" --config Release -j16

backends/cadence/reference/kernels/kernels.cpp

@@ -6,6 +6,7 @@
  * LICENSE file in the root directory of this source tree.
  */
 
+#include <executorch/backends/cadence/reference/kernels/kernels.h>
 #include <math.h>
 #include <algorithm>
 #include <cstring>

backends/cadence/reference/operators/CMakeLists.txt

@@ -33,7 +33,7 @@ set(_aten_ops__srcs
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/util/reduce_util.cpp"
     "${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"
+    "${EXECUTORCH_ROOT}/kernels/portable/cpu/pattern/unary_ufunc_realhbbf16_to_floathbf16.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_bmm.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_cat.cpp"
     "${EXECUTORCH_ROOT}/kernels/portable/cpu/op_clone.cpp"

backends/cadence/runtime/executor.py

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