CPU: remove oneDNN

Author: atafra

.github/workflows/internal.ci.yml

@@ -270,17 +270,6 @@ jobs:
       artifact-out: build-rockylinux8-gcc-minimal
       artifact-path: build install deps
 
-  build-rockylinux8-gcc-onednn:
-    secrets: inherit
-    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/docker.yml@main
-    with:
-      project: oidn
-      image: oidn/rockylinux:8.8
-      dpcpp-version: intel-llvm/v6.0.1-rk
-      cmd: scripts/build.py install --compiler gcc -D OIDN_DEVICE_CPU_ONEDNN=ON
-      artifact-out: build-rockylinux8-gcc-onednn
-      artifact-path: build install deps
-
   build-rockylinux8-gta-aotjit:
     secrets: inherit
     uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/docker.yml@main
@@ -754,20 +743,6 @@ jobs:
       artifact-on-failure: true
       cmd: scripts/test.py --device cpu --log test.log
 
-  test-rockylinux8-gcc-onednn-avx2:
-    needs: build-rockylinux8-gcc-onednn
-    secrets: inherit
-    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/docker.yml@main
-    with:
-      project: oidn
-      runs-on: '[ "Linux", "avx2" ]'
-      image: oidn/rockylinux:8.8
-      artifact-in: build-rockylinux8-gcc-onednn
-      artifact-out: test-rockylinux8-gcc-onednn-avx2
-      artifact-path: test.log
-      artifact-on-failure: true
-      cmd: scripts/test.py --device cpu --log test.log
-
   test-ubuntu2204-clang-asan:
     needs: build-ubuntu2204-clang-asan
     secrets: inherit
@@ -1433,16 +1408,6 @@ jobs:
       runs-on: '[ "Windows", "build" ]'
       cmd: python scripts/build.py --compiler msvc17 -D OIDN_LIBRARY_NAME=OIDN -D OIDN_API_NAMESPACE=myoidn
 
-  build-windows-msvc17-onednn:
-    secrets: inherit
-    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/windows.yml@main
-    with:
-      project: oidn
-      runs-on: '[ "Windows", "build" ]'
-      artifact-out: build-windows-msvc17-onednn
-      artifact-path: build install
-      cmd: python scripts/build.py install --compiler msvc17 -D OIDN_DEVICE_CPU_ONEDNN=ON
-
   build-windows-msvc17-asan:
     secrets: inherit
     uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/windows.yml@main
@@ -1580,19 +1545,6 @@ jobs:
       artifact-on-failure: true
       cmd: python scripts/test.py --device cpu --log test.log
 
-  test-windows-msvc17-onednn-avx512:
-    needs: build-windows-msvc17-onednn
-    secrets: inherit
-    uses: intel-innersource/libraries.devops.renderkit.workflows/.github/workflows/windows.yml@main
-    with:
-      project: oidn
-      runs-on: '[ "Windows", "avx512", "NAS" ]'
-      artifact-in: build-windows-msvc17-onednn
-      artifact-out: test-windows-msvc17-onednn-avx512
-      artifact-path: test.log
-      artifact-on-failure: true
-      cmd: python scripts/test.py --device cpu --minimal --log test.log
-
   # FIXME: ASAN not on the path?
   # test-windows-msvc17-asan-avx2:
   #   needs: build-windows-msvc17-asan

.gitmodules

@@ -1,6 +1,3 @@
-[submodule "mkl-dnn"]
-	path = external/mkl-dnn
-	url = ../mkl-dnn.git
 [submodule "weights"]
 	path = weights
 	url = ../oidn-weights.git

CMakeLists.txt

@@ -134,7 +134,6 @@ install(
     ${PROJECT_SOURCE_DIR}/CHANGELOG.md
     ${PROJECT_SOURCE_DIR}/third-party-programs.txt
     ${PROJECT_SOURCE_DIR}/third-party-programs-DPCPP.txt
-    ${PROJECT_SOURCE_DIR}/third-party-programs-oneDNN.txt
     ${PROJECT_SOURCE_DIR}/third-party-programs-oneTBB.txt
   DESTINATION ${CMAKE_INSTALL_DOCDIR}
   COMPONENT lib

cmake/oidn_dnnl.cmake

@@ -16,15 +16,7 @@ if(NOT TBB_FOUND)
   find_package(TBB ${OIDN_TBB_MIN_VERSION} REQUIRED tbb)
 endif()
 
-if(OIDN_ARCH STREQUAL "X64")
-  # oneDNN
-  option(OIDN_DEVICE_CPU_ONEDNN "Enable oneDNN for CPU device." OFF)
-  mark_as_advanced(OIDN_DEVICE_CPU_ONEDNN)
-  if(OIDN_DEVICE_CPU_ONEDNN)
-    set(OIDN_DNNL ON)
-    include(oidn_dnnl)
-  endif()
-elseif(APPLE AND OIDN_ARCH STREQUAL "ARM64")
+if(APPLE AND OIDN_ARCH STREQUAL "ARM64")
   # BNNS
   option(OIDN_DEVICE_CPU_BNNS "Enable BNNS for CPU device." ON)
   mark_as_advanced(OIDN_DEVICE_CPU_BNNS)
@@ -88,18 +80,7 @@ set(OIDN_ISPC_SOURCES_COMMON
   vec.isph
 )
 
-if(OIDN_DNNL)
-  list(APPEND OIDN_CPU_SOURCES
-    dnnl/dnnl_conv.h
-    dnnl/dnnl_conv.cpp
-    dnnl/dnnl_common.h
-    dnnl/dnnl_common.cpp
-    dnnl/dnnl_engine.h
-    dnnl/dnnl_engine.cpp
-    dnnl/dnnl_tensor.h
-    dnnl/dnnl_tensor.cpp
-  )
-elseif(OIDN_BNNS)
+if(OIDN_BNNS)
   list(APPEND OIDN_CPU_SOURCES
     bnns/bnns_common.h
     bnns/bnns_common.cpp
@@ -139,11 +120,7 @@ if(OIDN_LIBRARY_VERSIONED)
   set_target_properties(OpenImageDenoise_device_cpu PROPERTIES VERSION ${PROJECT_VERSION})
 endif()
 
-if(OIDN_DNNL)
-  target_compile_definitions(OpenImageDenoise_device_cpu PRIVATE OIDN_DNNL)
-  ispc_add_definitions(-DOIDN_DNNL)
-  target_link_libraries(OpenImageDenoise_device_cpu PRIVATE dnnl)
-elseif(OIDN_BNNS)
+if(OIDN_BNNS)
   target_compile_definitions(OpenImageDenoise_device_cpu PRIVATE OIDN_BNNS)
   ispc_add_definitions(-DOIDN_BNNS)
   target_link_libraries(OpenImageDenoise_device_cpu PRIVATE "-framework Accelerate")

devices/cpu/CMakeLists.txt

@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 
 #include "cpu_common.h"
-#if !defined(OIDN_DNNL) && !defined(OIDN_BNNS)
+#if !defined(OIDN_BNNS)
   #include "cpu_conv_ispc.h"
   #if defined(OIDN_ARCH_X64)
     #include "cpu_conv_amx_ispc.h"
@@ -89,7 +89,7 @@ OIDN_NAMESPACE_BEGIN
   }
 #endif
 
-#if !defined(OIDN_DNNL) && !defined(OIDN_BNNS)
+#if !defined(OIDN_BNNS)
   Tensor::operator ispc::TensorAccessor1D()
   {
     if (layout != TensorLayout::x)

devices/cpu/cpu_common.cpp

@@ -4,9 +4,7 @@
 #include "cpu_device.h"
 #include "platform_ispc.h"
 
-#if defined(OIDN_DNNL)
-  #include "dnnl/dnnl_engine.h"
-#elif defined(OIDN_BNNS)
+#if defined(OIDN_BNNS)
   #include "bnns/bnns_engine.h"
 #else
   #include "cpu_engine.h"
@@ -151,22 +149,7 @@ OIDN_NAMESPACE_BEGIN
 
     std::unique_ptr<CPUEngine> engine;
 
-  #if defined(OIDN_DNNL)
-    if (arch == CPUArch::AVX512)
-    {
-      tensorLayout = TensorLayout::Chw16c;
-      weightLayout = TensorLayout::OIhw16i16o;
-      tensorBlockC = 16;
-    }
-    else
-    {
-      tensorLayout = TensorLayout::Chw8c;
-      weightLayout = TensorLayout::OIhw8i8o;
-      tensorBlockC = 8;
-    }
-
-    engine.reset(new DNNLEngine(this, numThreads));
-  #elif defined(OIDN_BNNS)
+  #if defined(OIDN_BNNS)
     tensorLayout = TensorLayout::chw;
     weightLayout = TensorLayout::oihw;
     tensorBlockC = 1;

devices/cpu/cpu_device.cpp

@@ -33,7 +33,6 @@ OIDN_NAMESPACE_BEGIN
   class CPUDevice final : public Device
   {
     friend class CPUEngine;
-    friend class DNNLEngine;
 
   public:
     static std::vector<Ref<PhysicalDevice>> getPhysicalDevices();
@@ -44,9 +43,7 @@ OIDN_NAMESPACE_BEGIN
 
     DeviceType getType() const override { return DeviceType::CPU; }
 
-  #if !defined(OIDN_DNNL)
     bool needWeightAndBiasOnDevice() const override { return false; } // no need to copy
-  #endif
     Storage getPtrStorage(const void* ptr) override;
 
     int getInt(const std::string& name) override;

devices/cpu/cpu_device.h

@@ -3,7 +3,7 @@
 
 #include "core/conv.h"
 #include "cpu_engine.h"
-#if !defined(OIDN_DNNL) && !defined(OIDN_BNNS)
+#if !defined(OIDN_BNNS)
   #include "cpu_conv.h"
   #if defined(OIDN_ARCH_X64)
     #include "cpu_conv_amx.h"
@@ -75,7 +75,7 @@ OIDN_NAMESPACE_BEGIN
       return postOp == PostOp::None;
   }
 
-#if !defined(OIDN_DNNL) && !defined(OIDN_BNNS)
+#if !defined(OIDN_BNNS)
   Ref<Conv> CPUEngine::newConv(const ConvDesc& desc)
   {
   #if defined(OIDN_ARCH_X64)