[ARM] Added ARM32 ACL kernels calls

3.5 squash list:
[FORK][FIX] ARM changes for oneDNN 3.5 upgrade

.github/automation/build.sh

@@ -79,7 +79,7 @@ CMAKE_OPTIONS="${CMAKE_OPTIONS}
 # NOTE: only for AArch64 builds.
 if [ ! -z ${ACL_DIR} ]; then
  export ACL_ROOT_DIR=$ACL_DIR
- CMAKE_OPTIONS="${CMAKE_OPTIONS} -DDNNL_AARCH64_USE_ACL=ON"
+ CMAKE_OPTIONS="${CMAKE_OPTIONS} -DDNNL_USE_ACL=ON"
  echo "Info: Building with Arm Compute Library backend for Aarch64..."
 fi
 

cmake/ACL.cmake

@@ -25,7 +25,7 @@ if(NOT DNNL_TARGET_ARCH MATCHES "^(AARCH64|ARM)$")
  return()
 endif()
 
-if(NOT DNNL_AARCH64_USE_ACL)
+if(NOT DNNL_USE_ACL)
  return()
 endif()
 
@@ -67,7 +67,7 @@ if(ACL_FOUND)
  message(STATUS "Arm Compute Library: ${ACL_LIBRARIES}")
  message(STATUS "Arm Compute Library headers: ${ACL_INCLUDE_DIRS}")
 
- add_definitions(-DDNNL_AARCH64_USE_ACL)
+ add_definitions(-DDNNL_USE_ACL)
  set(CMAKE_CXX_STANDARD 14)
  set(CMAKE_CXX_EXTENSIONS "OFF")
 endif()

cmake/options.cmake

@@ -388,7 +388,7 @@ set(DNNL_BLAS_VENDOR "NONE" CACHE STRING
 # AArch64 optimizations with Arm Compute Library
 # ==============================================
 
-option(DNNL_AARCH64_USE_ACL "Enables use of AArch64 optimised functions
+option(DNNL_USE_ACL "Enables use of ARM optimised functions
  from Arm Compute Library.
  This is only supported on AArch64 builds and assumes there is a
  functioning Compute Library build available at the location specified by the

doc/build/build.md

@@ -117,7 +117,7 @@ make -j
 ~~~sh
 export ACL_ROOT_DIR=<path/to/Compute Library>
 cmake .. \
- -DDNNL_AARCH64_USE_ACL=ON \
+ -DDNNL_USE_ACL=ON \
  <extra build options>
 ~~~
 

src/cpu/CMakeLists.txt

@@ -130,6 +130,9 @@ endif()
 if (DNNL_TARGET_ARCH STREQUAL "AARCH64")
  add_subdirectory(aarch64)
 endif()
+if (DNNL_USE_ACL)
+ add_subdirectory(acl)
+endif()
 if (DNNL_TARGET_ARCH STREQUAL "PPC64")
  add_subdirectory(ppc64)
 endif()

src/cpu/aarch64/CMakeLists.txt

@@ -27,14 +27,6 @@ file(GLOB XBYAK_AARCH64_FILES
 
 list(REMOVE_ITEM SOURCES ${XBYAK_AARCH64_FILES})
 
-if(NOT DNNL_AARCH64_USE_ACL)
- file(GLOB_RECURSE ACL_FILES
- ${CMAKE_CURRENT_SOURCE_DIR}/acl_*.[ch]
- ${CMAKE_CURRENT_SOURCE_DIR}/acl_*.[ch]pp
- )
- list(REMOVE_ITEM SOURCES ${ACL_FILES})
-endif()
-
 # If the runtime is not THREADPOOL remove threadpool_scheduler sources.
 if(NOT DNNL_CPU_RUNTIME STREQUAL "THREADPOOL")
  list(APPEND ACL_THREADPOOL_FILES

src/cpu/aarch64/acl_reorder.cpp

@@ -19,7 +19,7 @@
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 status_t acl_reorder_fwd_t::execute_forward(const exec_ctx_t &ctx) const {
  // Lock here is needed because resource_mapper does not support
@@ -46,7 +46,7 @@ status_t acl_reorder_fwd_t::execute_forward(const exec_ctx_t &ctx) const {
  return status::success;
 }
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
-} // namespace dnnl
+} // namespace dnnl

src/cpu/aarch64/acl_reorder.hpp

@@ -13,19 +13,19 @@
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *******************************************************************************/
-#ifndef CPU_AARCH64_ACL_REORDER_HPP
-#define CPU_AARCH64_ACL_REORDER_HPP
+#ifndef CPU_ACL_REORDER_HPP
+#define CPU_ACL_REORDER_HPP
 
 #include "arm_compute/core/Types.h"
 #include "common/utils.hpp"
-#include "cpu/aarch64/acl_utils.hpp"
+#include "cpu/acl/acl_utils.hpp"
 #include "cpu/aarch64/cpu_isa_traits.hpp"
 #include "cpu/reorder/cpu_reorder_pd.hpp"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 struct acl_reorder_obj_t {
  arm_compute::NEReorderLayer reorder;
@@ -131,7 +131,7 @@ struct acl_reorder_fwd_t : public primitive_t {
  if (dst_tag == format_tag::BA4b4a || dst_tag == format_tag::Acdb4a
  || dst_tag == format_tag::Ab4a) {
  _pd->app_.dst_wf = arm_compute::WeightFormat::OHWIo4;
- } else if (mayiuse(sve_256)
+ } else if (aarch64::mayiuse(aarch64::sve_256)
  && (dst_tag == format_tag::BA8b4a
  || dst_tag == format_tag::Acdb8a
  || dst_tag == format_tag::Ab8a)) {
@@ -239,9 +239,9 @@ struct acl_reorder_fwd_t : public primitive_t {
 
 }; // acl_reorder_fwd_t
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl
 
-#endif // CPU_AARCH64_ACL_REORDER_HPP
+#endif // CPU_ACL_REORDER_HPP

src/cpu/acl/CMakeLists.txt

@@ -0,0 +1,36 @@
+#*******************************************************************************
+# Copyright 2020-2022 Arm Ltd. and affiliates
+# Copyright 2020-2021 FUJITSU LIMITED
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#*******************************************************************************
+
+file(GLOB_RECURSE SOURCES
+ ${CMAKE_CURRENT_SOURCE_DIR}/*.[ch]
+ ${CMAKE_CURRENT_SOURCE_DIR}/*.[ch]pp
+ )
+
+# If the runtime is not THREADPOOL remove threadpool_scheduler sources.
+if(NOT DNNL_CPU_RUNTIME STREQUAL "THREADPOOL")
+ list(APPEND ACL_THREADPOOL_FILES
+ ${CMAKE_CURRENT_SOURCE_DIR}/acl_threadpool_scheduler.cpp
+ ${CMAKE_CURRENT_SOURCE_DIR}/acl_threadpool_scheduler.hpp
+ )
+ list(REMOVE_ITEM SOURCES ${ACL_THREADPOOL_FILES})
+endif()
+
+set(OBJ_LIB ${DNNL_LIBRARY_NAME}_cpu_acl)
+add_library(${OBJ_LIB} OBJECT ${SOURCES})
+set_property(GLOBAL APPEND PROPERTY DNNL_LIB_DEPS
+ $<TARGET_OBJECTS:${OBJ_LIB}>)
+enable_conditional_compilation4(${OBJ_LIB})

src/cpu/aarch64/acl_batch_normalization.cpp → src/cpu/acl/acl_batch_normalization.cpp

@@ -14,12 +14,12 @@
 * limitations under the License.
 *******************************************************************************/
 
-#include "cpu/aarch64/acl_batch_normalization.hpp"
+#include "cpu/acl/acl_batch_normalization.hpp"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 status_t acl_batch_normalization_fwd_t::execute_forward(
  const exec_ctx_t &ctx) const {
@@ -72,7 +72,7 @@ status_t acl_batch_normalization_fwd_t::execute_forward(
  return status::success;
 }
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl

src/cpu/aarch64/acl_batch_normalization.hpp → src/cpu/acl/acl_batch_normalization.hpp

@@ -14,18 +14,18 @@
 * limitations under the License.
 *******************************************************************************/
 
-#ifndef CPU_AARCH64_ACL_BATCH_NORMALIZATION_HPP
-#define CPU_AARCH64_ACL_BATCH_NORMALIZATION_HPP
+#ifndef CPU_ACL_BATCH_NORMALIZATION_HPP
+#define CPU_ACL_BATCH_NORMALIZATION_HPP
 
 #include "cpu/cpu_batch_normalization_pd.hpp"
 
-#include "cpu/aarch64/acl_post_ops.hpp"
-#include "cpu/aarch64/acl_utils.hpp"
+#include "cpu/acl/acl_post_ops.hpp"
+#include "cpu/acl/acl_utils.hpp"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 struct acl_batch_normalization_obj_t {
  arm_compute::NEBatchNormalizationLayer bnorm;
@@ -274,7 +274,7 @@ struct acl_batch_normalization_fwd_t : public primitive_t {
  const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
 }; // acl_batch_normalization_fwd_t
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl

src/cpu/aarch64/acl_benchmark_scheduler.cpp → src/cpu/acl/acl_benchmark_scheduler.cpp

@@ -14,13 +14,13 @@
 * limitations under the License.
 *******************************************************************************/
 
-#include "cpu/aarch64/acl_benchmark_scheduler.hpp"
+#include "cpu/acl/acl_benchmark_scheduler.hpp"
 #include "common/verbose.hpp"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 using namespace arm_compute;
 
 BenchmarkScheduler::BenchmarkScheduler(IScheduler &real_scheduler)
@@ -72,7 +72,7 @@ void BenchmarkScheduler::run_workloads(std::vector<Workload> &workloads) {
  ARM_COMPUTE_ERROR("Can't be reached");
 }
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl

src/cpu/aarch64/acl_benchmark_scheduler.hpp → src/cpu/acl/acl_benchmark_scheduler.hpp

@@ -13,16 +13,16 @@
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *******************************************************************************/
-#ifndef CPU_AARCH64_ACL_BENCHMARK_SCHEDULER_HPP
-#define CPU_AARCH64_ACL_BENCHMARK_SCHEDULER_HPP
+#ifndef CPU_ACL_BENCHMARK_SCHEDULER_HPP
+#define CPU_ACL_BENCHMARK_SCHEDULER_HPP
 
 #include "arm_compute/core/CPP/ICPPKernel.h"
 #include "arm_compute/runtime/IScheduler.h"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 // BenchmarkScheduler implement's ACL IScheduler interface and acts as an interceptor scheduler
 // when DNNL_VERBOSE=profile,profile_externals. It intercepts calls made by the actual scheduler used by ACL and adds
 // timers to benchmark execution time of ACL kernels and store kernel information.
@@ -52,9 +52,9 @@ class BenchmarkScheduler final : public arm_compute::IScheduler {
  IScheduler &_real_scheduler;
 };
 
-#endif // CPU_AARCH64_ACL_BENCHMARK_SCHEDULER_HPP
+#endif // CPU_ACL_BENCHMARK_SCHEDULER_HPP
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl

src/cpu/aarch64/acl_binary.cpp → src/cpu/acl/acl_binary.cpp

@@ -14,12 +14,12 @@
 * limitations under the License.
 *******************************************************************************/
 
-#include "cpu/aarch64/acl_binary.hpp"
+#include "cpu/acl/acl_binary.hpp"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 status_t acl_binary_t::execute_forward(const exec_ctx_t &ctx, const void *src0,
  const void *src1, void *dst) const {
@@ -55,7 +55,7 @@ status_t acl_binary_t::execute_forward(const exec_ctx_t &ctx) const {
  return execute_forward(ctx, src0, src1, dst);
 }
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl

src/cpu/aarch64/acl_binary.hpp → src/cpu/acl/acl_binary.hpp

@@ -14,17 +14,17 @@
 * limitations under the License.
 *******************************************************************************/
 
-#ifndef CPU_AARCH64_ACL_BINARY_HPP
-#define CPU_AARCH64_ACL_BINARY_HPP
+#ifndef CPU_ACL_BINARY_HPP
+#define CPU_ACL_BINARY_HPP
 
 #include "cpu/cpu_binary_pd.hpp"
 
-#include "cpu/aarch64/acl_utils.hpp"
+#include "cpu/acl/acl_utils.hpp"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 struct acl_binary_obj_t {
  std::unique_ptr<arm_compute::IFunction> binary_op;
@@ -270,7 +270,7 @@ struct acl_binary_t : public primitive_t {
 
 }; // acl_binary_t
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl

src/cpu/aarch64/acl_convolution_utils.cpp → src/cpu/acl/acl_convolution_utils.cpp

@@ -16,15 +16,15 @@
 
 #include <cstdint>
 
-#include "cpu/aarch64/acl_convolution_utils.hpp"
+#include "cpu/acl/acl_convolution_utils.hpp"
 #include "common/convolution_pd.hpp"
 #include "common/utils.hpp"
 #include "oneapi/dnnl/dnnl.h"
 
 namespace dnnl {
 namespace impl {
 namespace cpu {
-namespace aarch64 {
+namespace acl {
 
 namespace acl_convolution_utils {
 
@@ -418,7 +418,7 @@ status_t init_conf_depthwise(acl_conv_conf_t &acp, memory_desc_t &src_md,
 
 } // namespace acl_convolution_utils
 
-} // namespace aarch64
+} // namespace acl
 } // namespace cpu
 } // namespace impl
 } // namespace dnnl