Remove MKLML build config (#5559)

Remove MKLML build config
2026-07-24 19:43:35 +00:00 · 2020-10-21 13:11:25 -07:00 · 2020-10-21 13:11:25 -07:00 · 5802fe1699
commit 5802fe1699
parent 82c7a9756e
27 changed files with 18 additions and 513 deletions
--- a/cmake/CMakeLists.txt
+++ b/cmake/CMakeLists.txt
@ -62,7 +62,6 @@ option(onnxruntime_USE_EIGEN_FOR_BLAS "Use eign for blas" ON)
 option(onnxruntime_USE_NNAPI_BUILTIN "Build with builtin NNAPI lib for Android NNAPI support" OFF)
 option(onnxruntime_USE_RKNPU "Build with RKNPU support" OFF)
 option(onnxruntime_USE_DNNL "Build with DNNL support" OFF)
-option(onnxruntime_USE_MKLML "Build the default cpu provider with MKL-ML binary dependency" OFF)
 option(onnxruntime_USE_FEATURIZERS "Build ML Featurizers support" OFF)
 option(onnxruntime_USE_NGRAPH "Build with nGraph support" OFF)
 option(onnxruntime_USE_OPENBLAS "Use openblas" OFF)
@ -198,12 +197,7 @@ if(onnxruntime_USE_OPENMP)
  if (OPENMP_FOUND)
    set (CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
    set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
-    include_directories(${OpenMP_CXX_INCLUDE_DIR})
-    # MKLML and NGraph depend on their own OpenMP library that may be different with the compiler's.
-    # Disable the options to build mklml/NGraph and OpenMP together.
-    if((WIN32 OR APPLE) AND onnxruntime_USE_MKLML)
-      message(FATAL_ERROR "Please use only one of onnxruntime_USE_MKLML, onnxruntime_USE_OPENMP")
-    endif()
+    include_directories(${OpenMP_CXX_INCLUDE_DIR})    
    if(onnxruntime_USE_NGRAPH)
      message(FATAL_ERROR "Please use only one of onnxruntime_USE_NGRAPH, onnxruntime_USE_OPENMP")
    endif()
@ -696,8 +690,7 @@ if (onnxruntime_USE_ARMNN)
  list(APPEND onnxruntime_EXTERNAL_LIBRARIES armnn pthread arm_compute_core arm_compute arm_compute_graph)
 endif()

-# MKLML
-if (onnxruntime_USE_DNNL OR onnxruntime_USE_MKLML)
+if (onnxruntime_USE_DNNL)
  include(dnnl)
 endif()

@ -713,15 +706,6 @@ if (onnxruntime_USE_TVM)
  if (onnxruntime_USE_OPENMP)
    set(USE_OPENMP "gnu")
  endif()
-  if (onnxruntime_USE_MKLML)
-    set(USE_OPENMP "intel")
-    # make sure MKLML in ORT is used by TVM
-    if (WIN32)
-      set(OMP_LIBRARY ${MKLML_LIB_DIR}/${IOMP5MD_IMPORT_LIB})
-    else()
-      set(OMP_LIBRARY ${MKLML_LIB_DIR}/${IOMP5MD_SHARED_LIB})
-    endif()
-  endif()

  add_subdirectory(${PROJECT_SOURCE_DIR}/external/tvm EXCLUDE_FROM_ALL)
  set_target_properties(tvm PROPERTIES FOLDER "External/tvm")
@ -729,13 +713,6 @@ if (onnxruntime_USE_TVM)
  set_target_properties(tvm_runtime PROPERTIES FOLDER "External/tvm")
  set_target_properties(nnvm_compiler PROPERTIES FOLDER "External/tvm")

-  if (onnxruntime_USE_MKLML)
-    add_dependencies(tvm project_mklml)
-    add_dependencies(tvm_topi project_mklml)
-    add_dependencies(tvm_runtime project_mklml)
-    add_dependencies(nnvm_compiler project_mklml)
-  endif()
-
  set(TVM_INCLUDES ${PROJECT_SOURCE_DIR}/external/tvm/include
    ${PROJECT_SOURCE_DIR}/external/tvm/3rdparty/dmlc-core/include
    ${PROJECT_SOURCE_DIR}/external/tvm/3rdparty/dlpack/include
@ -784,22 +761,6 @@ endif()
 set_target_properties(onnx PROPERTIES FOLDER "External/ONNX")
 set_target_properties(onnx_proto PROPERTIES FOLDER "External/ONNX")

-if (onnxruntime_USE_MKLML)
-  add_definitions(-DUSE_MKLML=1 -DUSE_MKLML_FOR_BLAS=1)
-  if (WIN32 OR APPLE)
-    list(APPEND onnxruntime_EXTERNAL_LIBRARIES mklml)
-  else()
-    if(onnxruntime_USE_OPENMP)
-      list(APPEND onnxruntime_EXTERNAL_LIBRARIES mklml_gnu)
-    else()
-      list(APPEND onnxruntime_EXTERNAL_LIBRARIES mklml_intel)
-    endif()
-  endif()
-  list(APPEND onnxruntime_EXTERNAL_DEPENDENCIES project_mklml)
-  include_directories(${MKLML_INCLUDE_DIR})
-  link_directories(${MKLML_LIB_DIR})
-endif()
-
 if (onnxruntime_USE_NGRAPH)
  if (NOT onnxruntime_USE_FULL_PROTOBUF)
    message(FATAL_ERROR "Please set onnxruntime_USE_FULL_PROTOBUF=ON for nGraph execution provider.")
--- a/cmake/external/dnnl.cmake
+++ b/cmake/external/dnnl.cmake
@ -1,64 +1,25 @@
 include (ExternalProject)

 set(DNNL_URL https://github.com/intel/mkl-dnn.git)
-# If DNNL_TAG is updated, check if MKLML_VERSION and platform.cmake.patch need to be updated.
+# If DNNL_TAG is updated, check if platform.cmake.patch need to be updated.
 set(DNNL_TAG v1.1.1)
-set(MKLML_VERSION 2019.0.5.20190502)

 if(WIN32)
  set(MKLML_OS_VERSION_STR "win")
  set(MKLML_FILE_EXTENSION "zip")
  set(DNNL_SHARED_LIB dnnl.dll)
-  set(DNNL_IMPORT_LIB dnnl.lib)
-  if(onnxruntime_USE_MKLML)
-    # Windows-only updated MKLML binary which contains fix for thread cleanup hang.
-    set(MKLML_VERSION 2020.0.20190813)
-    set(MKLML_SHARED_LIB mklml.dll)
-    set(MKLML_IMPORT_LIB mklml.lib)
-    set(IOMP5MD_SHARED_LIB libiomp5md.dll)
-    set(IOMP5MD_IMPORT_LIB libiomp5md.lib)
-  endif()
+  set(DNNL_IMPORT_LIB dnnl.lib)  
 else()
  set(MKLML_FILE_EXTENSION "tgz")
  if (APPLE)
    set(DNNL_SHARED_LIB libdnnl.1.dylib)
-    set(MKLML_OS_VERSION_STR "mac")
-    if(onnxruntime_USE_MKLML)
-      set(MKLML_SHARED_LIB libmklml.dylib)
-      set(IOMP5MD_SHARED_LIB libiomp5.dylib)
-    endif()
+    set(MKLML_OS_VERSION_STR "mac")    
  else()
    set(DNNL_SHARED_LIB libdnnl.so.1)
-    set(MKLML_OS_VERSION_STR "lnx")
-    if(onnxruntime_USE_MKLML)
-      if(onnxruntime_USE_OPENMP)
-        set(MKLML_SHARED_LIB libmklml_gnu.so)
-      else()
-        set(MKLML_SHARED_LIB libmklml_intel.so)
-        set(IOMP5MD_SHARED_LIB libiomp5.so)
-      endif()
-    endif()
+    set(MKLML_OS_VERSION_STR "lnx")    
  endif()  
 endif()

-if (onnxruntime_USE_MKLML)
-  set(MKLDNN_VERSION_SHORT v0.20)
-  set(MKLML_URL https://github.com/intel/mkl-dnn/releases/download/${MKLDNN_VERSION_SHORT}/mklml_${MKLML_OS_VERSION_STR}_${MKLML_VERSION}.${MKLML_FILE_EXTENSION})
-
-  ExternalProject_Add(project_mklml
-    PREFIX mklml
-    URL ${MKLML_URL}
-    CONFIGURE_COMMAND ""
-    BUILD_COMMAND ""
-    UPDATE_COMMAND ""
-    INSTALL_COMMAND ""  )
-
-  set(MKML_DIR ${CMAKE_CURRENT_BINARY_DIR}/mklml/src/project_mklml)
-  set(MKLML_INCLUDE_DIR "${MKML_DIR}/include")
-  set(MKLML_LIB_DIR "${MKML_DIR}/lib")
-  link_directories(${MKLML_LIB_DIR})
-endif()
-
 if (onnxruntime_USE_DNNL)
  set(DNNL_SOURCE ${CMAKE_CURRENT_BINARY_DIR}/dnnl/src/dnnl/src)
  set(DNNL_INSTALL ${CMAKE_CURRENT_BINARY_DIR}/dnnl/install)
@ -88,7 +49,4 @@ if (onnxruntime_USE_DNNL)
    CMAKE_ARGS -DDNNL_BUILD_TESTS=OFF -DDNNL_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} -DCMAKE_INSTALL_PREFIX=${DNNL_INSTALL}
  )
  link_directories(${DNNL_LIB_DIR})
-  #if (onnxruntime_USE_MKLML)
-  #  add_dependencies(project_dnnl project_mklml)
-  #endif()
 endif()
--- a/cmake/onnxruntime_python.cmake
+++ b/cmake/onnxruntime_python.cmake
@ -378,15 +378,6 @@ if (onnxruntime_USE_TVM)
  )
 endif()

-if (onnxruntime_USE_MKLML)
-  add_custom_command(
-    TARGET onnxruntime_pybind11_state POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E copy
-        ${MKLML_LIB_DIR}/${MKLML_SHARED_LIB} ${MKLML_LIB_DIR}/${IOMP5MD_SHARED_LIB}
-        $<TARGET_FILE_DIR:${test_data_target}>/onnxruntime/capi/
-  )
-endif()
-
 if (onnxruntime_USE_NUPHAR)
  file(GLOB onnxruntime_python_nuphar_python_srcs CONFIGURE_DEPENDS
    "${ONNXRUNTIME_ROOT}/core/providers/nuphar/scripts/*"
--- a/cmake/onnxruntime_unittests.cmake
+++ b/cmake/onnxruntime_unittests.cmake
@ -645,14 +645,6 @@ if (onnxruntime_USE_DNNL)
    COMMAND ${CMAKE_COMMAND} -E copy ${DNNL_DLL_PATH} $<TARGET_FILE_DIR:${test_data_target}>
    )
 endif()
-if (onnxruntime_USE_MKLML)
-  add_custom_command(
-    TARGET ${test_data_target} POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E copy
-    ${MKLML_LIB_DIR}/${MKLML_SHARED_LIB} ${MKLML_LIB_DIR}/${IOMP5MD_SHARED_LIB}
-    $<TARGET_FILE_DIR:${test_data_target}>
-  )
-endif()
 if(WIN32)
  if (onnxruntime_USE_NGRAPH)
    add_custom_command(
--- a/onnxruntime/contrib_ops/cpu/bert/attention.cc
+++ b/onnxruntime/contrib_ops/cpu/bert/attention.cc
@ -21,9 +21,7 @@ class Attention : public OpKernel, public AttentionCPUBase {
  explicit Attention(const OpKernelInfo& info);

  Status Compute(OpKernelContext* context) const override;
-#if !defined(USE_MKLML_FOR_BLAS)
  Status PrePack(const Tensor& tensor, int input_idx, bool& is_packed) override;
-#endif

 private:
  BufferUniquePtr packed_weights_;
@ -178,7 +176,6 @@ template <typename T>
 Attention<T>::Attention(const OpKernelInfo& info) : OpKernel(info), AttentionCPUBase(info) {
 }

-#if !defined(USE_MKLML_FOR_BLAS)

 template <typename T>
 Status Attention<T>::PrePack(const Tensor& weights, int input_idx, bool& is_packed) {
@ -225,8 +222,6 @@ Status Attention<T>::PrePack(const Tensor& weights, int input_idx, bool& is_pack
  return Status::OK();
 }

-#endif
-
 template <typename T>
 Status Attention<T>::Compute(OpKernelContext* context) const {
  const Tensor* input = context->Input<Tensor>(0);
--- a/onnxruntime/contrib_ops/cpu/cdist.cc
+++ b/onnxruntime/contrib_ops/cpu/cdist.cc
@ -56,8 +56,8 @@ static void CalculateSqeuclidean(const Tensor& a, const Tensor& b, Tensor& c, co
  // in Xij and Yjk are very similar, so subtracting can be problematic.
  // Due to that we calculate -2*sum_k(Xik*Yjk) using GEMM, add sum_k(Xik**2) next, and add sum_k(Yjk**2) last.

-// use MLAS on 64-bit (no 32-bit dgemm), or MKL on 32-bit or 64-bit
-#if defined(_M_AMD64) || defined(__x86_64__) || defined(USE_MKLML_FOR_BLAS)
+// use MLAS on 64-bit (no 32-bit dgemm)
+#if defined(_M_AMD64) || defined(__x86_64__)
  // Use GEMM of A and B^T with -2 as alpha to calculate -2*sum_k(Xik*Yjk)
  math::Gemm<T>(CBLAS_TRANSPOSE::CblasNoTrans, CBLAS_TRANSPOSE::CblasTrans,
                m, n, k,
--- a/onnxruntime/core/providers/cpu/math/matmul.cc
+++ b/onnxruntime/core/providers/cpu/math/matmul.cc
@ -125,7 +125,6 @@ Status MatMul<T>::Compute(OpKernelContext* ctx) const {
  return Status::OK();
 }

-#if !defined(USE_MKLML_FOR_BLAS)
 Status MatMul<float>::PrePack(const Tensor& tensor, int input_idx, bool& is_packed) {
  is_packed = false;

@ -162,7 +161,6 @@ Status MatMul<float>::PrePack(const Tensor& tensor, int input_idx, bool& is_pack
  }
  return Status::OK();
 }
-#endif

 Status MatMul<float>::Compute(OpKernelContext* ctx) const {
  concurrency::ThreadPool* thread_pool = ctx->GetOperatorThreadPool();
@ -190,7 +188,6 @@ Status MatMul<float>::Compute(OpKernelContext* ctx) const {
  // TODO: replace it with GemmBatch for performance, it's OK for now as GemmBatch unrolls as well
  size_t max_len = helper.OutputOffsets().size();
  for (size_t i = 0; i < max_len; i++) {
-#if !defined(USE_MKLML_FOR_BLAS)
    if (packed_b_) {
      MlasGemm(
          trans_a ? CblasTrans : CblasNoTrans,
@ -207,7 +204,6 @@ Status MatMul<float>::Compute(OpKernelContext* ctx) const {
          thread_pool);
      continue;
    }
-#endif
    math::Gemm<float, concurrency::ThreadPool>(
        trans_a ? CblasTrans : CblasNoTrans,
        trans_b ? CblasTrans : CblasNoTrans,
--- a/onnxruntime/core/providers/cpu/math/matmul.h
+++ b/onnxruntime/core/providers/cpu/math/matmul.h
@ -24,9 +24,7 @@ class MatMul<float> final : public OpKernel {
    info.GetAttrOrDefault<float>("alpha", &alpha_attr_, 1.0);
  }

-#if !defined(USE_MKLML_FOR_BLAS)
  Status PrePack(const Tensor& tensor, int input_idx, bool& is_packed) override;
-#endif

  Status Compute(OpKernelContext* context) const override;

--- a/onnxruntime/core/providers/cpu/rnn/deep_cpu_lstm.cc
+++ b/onnxruntime/core/providers/cpu/rnn/deep_cpu_lstm.cc
@ -327,7 +327,6 @@ Status DeepCpuLstmOp::TryPackWeights(const Tensor& weights, PackedWeights& packe
  return Status::OK();
 }

-#if !defined(USE_MKLML_FOR_BLAS)
 Status DeepCpuLstmOp::PrePack(const Tensor& tensor, int input_idx, bool& is_packed) {
  is_packed = false;

@ -341,7 +340,6 @@ Status DeepCpuLstmOp::PrePack(const Tensor& tensor, int input_idx, bool& is_pack

  return Status::OK();
 }
-#endif

 Status DeepCpuLstmOp::Compute(OpKernelContext* context) const {
  const Tensor& X = *context->Input<Tensor>(0);  // inputs. [seq_length, batch_size, input_size]
--- a/onnxruntime/core/providers/cpu/rnn/deep_cpu_lstm.h
+++ b/onnxruntime/core/providers/cpu/rnn/deep_cpu_lstm.h
@ -50,9 +50,7 @@ class DeepCpuLstmOp final : public OpKernel {
                                                     activation_func_betas);
  }

-#if !defined(USE_MKLML_FOR_BLAS)
  Status PrePack(const Tensor& tensor, int input_idx, bool& is_packed) override;
-#endif
  Status Compute(OpKernelContext* context) const override;

  ~DeepCpuLstmOp() override = default;
--- a/onnxruntime/core/providers/nuphar/common/nuphar_settings.cc
+++ b/onnxruntime/core/providers/nuphar/common/nuphar_settings.cc
@ -58,7 +58,7 @@ void SetDefaultOptions(std::map<std::string, std::string>& options) {
  options.insert(std::make_pair(cache_so_name_opt, cache_so_name_default));

  std::string parallel_min_workloads_opt(kNupharParallelMinWorkloads);
-#if defined(_OPENMP) || defined(USE_MKLML)
+#if defined(_OPENMP)
  // a rough estimate of workloads based on static dimensions for each thread, when using parallel schedule
  // user may change it to 0 to turn it off,
  // or use OMP_NUM_THREADS to control TVM thread pool similar to control MKL
--- a/onnxruntime/core/providers/nuphar/extern/igemv_mkl.cc
+++ b/onnxruntime/core/providers/nuphar/extern/igemv_mkl.cc
@ -1,36 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#include "igemv_mkl.h"
-
-namespace onnxruntime {
-#ifdef NUPHAR_USE_MKL
-void MKLIntGemvS16S16S32R(
-    int16_t* matrixA,
-    int16_t* matrixB,
-    int M,
-    int N,
-    int K,
-    int32_t* output) {
-  MKL_INT32 co = 0;
-  cblas_gemm_s16s16s32(CBLAS_LAYOUT::CblasColMajor, CBLAS_TRANSPOSE::CblasTrans, CBLAS_TRANSPOSE::CblasNoTrans, CBLAS_OFFSET::CblasFixOffset,
-                       M, N, K,
-                       1, matrixA, K,
-                       0, matrixB, K, 0, 0, output, M, &co);
-}
-void MKLIntGemvS8U8S32R(
-    int8_t* matrixA,
-    uint8_t* matrixB,
-    int M,
-    int N,
-    int K,
-    int32_t* output) {
-  MKL_INT32 co = 0;
-  cblas_gemm_s8u8s32(CBLAS_LAYOUT::CblasColMajor, CBLAS_TRANSPOSE::CblasTrans, CBLAS_TRANSPOSE::CblasNoTrans, CBLAS_OFFSET::CblasFixOffset,
-                     M, N, K,
-                     1, matrixA, K,
-                     0, matrixB, K, 0, 0, output, M, &co);
-}
-#endif
-
-}  // namespace onnxruntime
--- a/onnxruntime/core/providers/nuphar/extern/igemv_mkl.h
+++ b/onnxruntime/core/providers/nuphar/extern/igemv_mkl.h
@ -1,30 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#pragma once
-#include <stdint.h>
-
-#ifdef NUPHAR_USE_MKL
-// Need to build with USE_MKLML
-#include <mkl_cblas.h>
-#endif  // NUPHAR_USE_MKL
-
-namespace onnxruntime {
-#ifdef NUPHAR_USE_MKL
-void MKLIntGemvS16S16S32R(
-    int16_t* matrixA,
-    int16_t* matrixB,
-    int M,
-    int N,
-    int K,
-    int32_t* output);
-
-void MKLIntGemvS8U8S32R(
-    int8_t* matrixA,
-    uint8_t* matrixB,
-    int M,
-    int N,
-    int K,
-    int32_t* output);
-#endif
-}  // namespace onnxruntime
--- a/onnxruntime/core/providers/nuphar/mti_x86/quantize/imatmul16_extern.cc
+++ b/onnxruntime/core/providers/nuphar/mti_x86/quantize/imatmul16_extern.cc
@ -5,7 +5,6 @@

 #include "core/common/common.h"
 #include "core/codegen/mti/mti_tvm_utils.h"
-#include "core/providers/nuphar/extern/igemv_mkl.h"
 #include "core/providers/nuphar/extern/igemv_avx2.h"
 #include <topi/detail/extern.h>

--- a/onnxruntime/core/providers/nuphar/mti_x86/quantize/imatmul_extern.cc
+++ b/onnxruntime/core/providers/nuphar/mti_x86/quantize/imatmul_extern.cc
@ -5,7 +5,6 @@

 #include "core/common/common.h"
 #include "core/codegen/mti/mti_tvm_utils.h"
-#include "core/providers/nuphar/extern/igemv_mkl.h"
 #include "core/providers/nuphar/extern/igemv_avx2.h"
 #include <topi/detail/extern.h>

--- a/onnxruntime/core/util/math.h
+++ b/onnxruntime/core/util/math.h
@ -20,16 +20,11 @@
 // still keep it simple, so all platforms would be able to support it fairly
 // easily.

-#ifdef USE_MKLML_FOR_BLAS
-// when USE_MKLML is defined, use MKLML cblas for GEMM
-#include "mkl_cblas.h"
-#define CBLAS_ENUM_DEFINED_H
-#else
+
 // We include the cblas header here so that we can obtain the macros from cblas.
 extern "C" {
 #include "core/framework/cblas.h"
 }
-#endif

 #include "core/common/common.h"
 #include "core/framework/tensor.h"
--- a/onnxruntime/core/util/math_cpu.cc
+++ b/onnxruntime/core/util/math_cpu.cc
@ -59,8 +59,6 @@ EIGEN_MATMUL_FUNCTION(uint64_t)
 // will delegate the Caffe math functions that are BLAS-related to either the
 // CBLAS call or the Eigen implementation.
 ////////////////////////////////////////////////////////////////////////////////
-// when USE_MKLML is defined, use cblas APIs for MKLML
-#if !defined(USE_MKLML_FOR_BLAS)

 // Caffe2 gemm provides a simpler interface to the gemm functions, with the
 // limitation that the data has to be contiguous in memory.
@ -212,79 +210,6 @@ template void Gemv<double, CPUMathUtil>(const CBLAS_TRANSPOSE TransA, int M, int
 SPECIALIZED_AXPY(float)
 #undef SPECIALIZED_AXPY

-#else
-
-template <>
-void Gemm<float, ThreadPool>(const CBLAS_TRANSPOSE TransA, const CBLAS_TRANSPOSE TransB, const int64_t M,
-                             const int64_t N, const int64_t K, float alpha, const float* A, const float* B, float beta,
-                             float* C, ThreadPool* /*context*/) {
-  int lda = gsl::narrow_cast<int>((TransA == CblasNoTrans) ? K : M);
-  int ldb = gsl::narrow_cast<int>((TransB == CblasNoTrans) ? N : K);
-  cblas_sgemm(CblasRowMajor, TransA, TransB,
-              gsl::narrow_cast<int>(M),
-              gsl::narrow_cast<int>(N),
-              gsl::narrow_cast<int>(K),
-              alpha, A, lda, B, ldb,
-              beta, C, gsl::narrow_cast<int>(N));
-}
-
-template <>
-void Gemm<double, ThreadPool>(const CBLAS_TRANSPOSE TransA, const CBLAS_TRANSPOSE TransB, const int64_t M,
-                              const int64_t N, const int64_t K, double alpha, const double* A, const double* B, double beta,
-                              double* C, ThreadPool* /*context*/) {
-  int lda = gsl::narrow_cast<int>((TransA == CblasNoTrans) ? K : M);
-  int ldb = gsl::narrow_cast<int>((TransB == CblasNoTrans) ? N : K);
-  cblas_dgemm(CblasRowMajor, TransA, TransB,
-              gsl::narrow_cast<int>(M),
-              gsl::narrow_cast<int>(N),
-              gsl::narrow_cast<int>(K),
-              alpha, A, lda, B, ldb,
-              beta, C, gsl::narrow_cast<int>(N));
-}
-
-template <>
-void MatMul<float>(int M, int N, int K, const float* A, const float* B, float* C, ThreadPool*) {
-  cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K, 1, A, K, B, N, 0, C, N);
-}
-
-template <>
-void MatMul<double>(int M, int N, int K, const double* A, const double* B, double* C, ThreadPool*) {
-  cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, M, N, K, 1, A, K, B, N, 0, C, N);
-}
-
-template <>
-void GemmEx<float, ThreadPool>(const CBLAS_TRANSPOSE TransA, const CBLAS_TRANSPOSE TransB, int M, int N, int K,
-                               float alpha, const float* A, int lda, const float* B, int ldb, float beta, float* C,
-                               int ldc, ThreadPool* /*context*/) {
-  cblas_sgemm(CblasRowMajor, TransA, TransB, M, N, K, alpha, A, lda, B, ldb,
-              beta, C, ldc);
-}
-
-template <>
-void Gemv<float, CPUMathUtil>(const CBLAS_TRANSPOSE TransA, int M, int N, float alpha, const float* A, const float* x,
-                              float beta, float* y, CPUMathUtil* /*context*/) {
-  cblas_sgemv(CblasRowMajor, TransA, M, N, alpha, A, N, x, 1, beta, y, 1);
-}
-
-template <>
-void Gemv<double, CPUMathUtil>(const CBLAS_TRANSPOSE TransA, int M, int N, float alpha, const double* A, const double* x,
-                               float beta, double* y, CPUMathUtil* /*context*/) {
-  cblas_dgemv(CblasRowMajor, TransA, M, N, alpha, A, N, x, 1, beta, y, 1);
-}
-
-#define CAFFE2_SPECIALIZED_AXPY(T, prefix)                                           \
-  template <>                                                                        \
-  void Axpy<T, CPUMathUtil>(int N, const T alpha, const T* x, T* y, CPUMathUtil*) {  \
-    cblas_##prefix##axpy(N, alpha, x, 1, y, 1);                                      \
-  }                                                                                  \
-  template <>                                                                        \
-  void Axpy<T, CPUMathUtil>(int N, const T* alpha, const T* x, T* y, CPUMathUtil*) { \
-    cblas_##prefix##axpy(N, *alpha, x, 1, y, 1);                                     \
-  }
-CAFFE2_SPECIALIZED_AXPY(float, s)
-#undef CAFFE2_SPECIALIZED_AXPY
-
-#endif

 #define DELEGATE_SIMPLE_UNARY_FUNCTION(T, Funcname, expr)                  \
  template <>                                                              \
--- a/onnxruntime/python/onnxruntime_pybind_state.cc
+++ b/onnxruntime/python/onnxruntime_pybind_state.cc
@ -50,12 +50,6 @@ struct OrtStatus {
 #define BACKEND_DNNL ""
 #endif

-#if USE_MKLML
-#define BACKEND_MKLML "-MKL-ML"
-#else
-#define BACKEND_MKLML ""
-#endif
-
 #if USE_NGRAPH
 #define BACKEND_NGRAPH "-NGRAPH"
 #include "core/providers/ngraph/ngraph_execution_provider.h"
@ -129,7 +123,7 @@ struct OrtStatus {
 #define BACKEND_DML ""
 #endif

-#define BACKEND_DEVICE BACKEND_PROC BACKEND_DNNL BACKEND_MKLML BACKEND_NGRAPH BACKEND_OPENVINO BACKEND_NUPHAR BACKEND_OPENBLAS BACKEND_MIGRAPHX BACKEND_ACL BACKEND_ARMNN BACKEND_DML
+#define BACKEND_DEVICE BACKEND_PROC BACKEND_DNNL BACKEND_NGRAPH BACKEND_OPENVINO BACKEND_NUPHAR BACKEND_OPENBLAS BACKEND_MIGRAPHX BACKEND_ACL BACKEND_ARMNN BACKEND_DML
 #include "core/session/onnxruntime_cxx_api.h"
 #include "core/providers/providers.h"
 #include "core/providers/cpu/cpu_execution_provider.h"
--- a/tools/ci_build/build.py
+++ b/tools/ci_build/build.py
@ -260,8 +260,6 @@ def parse_arguments():
        "--use_openblas", action='store_true', help="Build with OpenBLAS.")
    parser.add_argument(
        "--use_dnnl", action='store_true', help="Build with DNNL.")
-    parser.add_argument(
-        "--use_mklml", action='store_true', help="Build with MKLML.")
    parser.add_argument(
        "--use_featurizers", action='store_true',
        help="Build with ML Featurizer support.")
@ -619,13 +617,12 @@ def generate_build_tree(cmake_path, source_dir, build_dir, cuda_home, cudnn_home
            "OFF" if args.use_openblas else "ON"),
        "-Donnxruntime_USE_OPENBLAS=" + ("ON" if args.use_openblas else "OFF"),
        "-Donnxruntime_USE_DNNL=" + ("ON" if args.use_dnnl else "OFF"),
-        "-Donnxruntime_USE_MKLML=" + ("ON" if args.use_mklml else "OFF"),
        "-Donnxruntime_USE_NGRAPH=" + ("ON" if args.use_ngraph else "OFF"),
        "-Donnxruntime_USE_NNAPI_BUILTIN=" + ("ON" if args.use_nnapi else "OFF"),
        "-Donnxruntime_USE_RKNPU=" + ("ON" if args.use_rknpu else "OFF"),
        "-Donnxruntime_USE_OPENMP=" + (
            "ON" if args.use_openmp and not (
-                args.use_nnapi or (args.use_mklml and (is_macOS() or is_windows())) or args.use_ngraph or
+                args.use_nnapi or args.use_ngraph or
                args.android or (args.ios and is_macOS())
                or args.use_rknpu)
            else "OFF"),
@ -1308,8 +1305,6 @@ def run_onnxruntime_tests(args, source_dir, ctest_path, build_dir, configs):
                build_dir, config, "external", "tvm", config))
        if args.use_tensorrt:
            dll_path_list.append(os.path.join(args.tensorrt_home, 'lib'))
-        if args.use_mklml:
-            dll_path_list.append(os.path.join(build_dir, config, "mklml", "src", "project_mklml", "lib"))
        if not is_windows():
            # A workaround for making libonnxruntime_providers_shared.so loadable.
            dll_path_list.append(os.path.join(build_dir, config))
@ -1509,7 +1504,7 @@ def derive_linux_build_property():
        return "/p:IsLinuxBuild=\"true\""


-def build_nuget_package(source_dir, build_dir, configs, use_cuda, use_openvino, use_tensorrt, use_dnnl, use_mklml):
+def build_nuget_package(source_dir, build_dir, configs, use_cuda, use_openvino, use_tensorrt, use_dnnl):
    if not (is_windows() or is_linux()):
        raise BuildError(
            'Currently csharp builds and nuget package creation is only supportted '
@ -1532,8 +1527,6 @@ def build_nuget_package(source_dir, build_dir, configs, use_cuda, use_openvino,
        package_name = "/p:OrtPackageId=\"Microsoft.ML.OnnxRuntime.DNNL\""
    elif use_cuda:
        package_name = "/p:OrtPackageId=\"Microsoft.ML.OnnxRuntime.Gpu\""
-    elif use_mklml:
-        package_name = "/p:OrtPackageId=\"Microsoft.ML.OnnxRuntime.MKLML\""
    else:
        pass

@ -1925,8 +1918,7 @@ def main():
                args.use_cuda,
                args.use_openvino,
                args.use_tensorrt,
-                args.use_dnnl,
-                args.use_mklml
+                args.use_dnnl
            )

    if args.test and args.build_nuget:
--- a/tools/ci_build/github/azure-pipelines/linux-ci-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/linux-ci-pipeline.yml
@ -39,7 +39,7 @@ jobs:
              --use_openmp \
              --enable_onnx_tests \
              --enable_symbolic_shape_infer_tests \
-              --use_mklml --enable_pybind --build_java --build_nodejs
+              --enable_pybind --build_java --build_nodejs
      workingDirectory: $(Build.SourcesDirectory)
  - task: Docker@2
    displayName: logout
--- a/tools/ci_build/github/azure-pipelines/linux-gpu-cuda-11-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/linux-gpu-cuda-11-pipeline.yml
@ -38,7 +38,7 @@ jobs:
              --build_wheel \
              --use_openmp \
              --enable_onnx_tests --use_cuda --cuda_version=11.0 --cuda_home=/usr/local/cuda-11.0 --cudnn_home=/usr/local/cuda-11.0 \
-              --use_mklml --enable_pybind --build_java --build_nodejs \
+              --enable_pybind --build_java --build_nodejs \
              --cmake_extra_defines PYTHON_INCLUDE_DIR=/opt/python/cp37-cp37m/include/python3.7m PYTHON_LIBRARY=/usr/lib64/librt.so CMAKE_CUDA_ARCHITECTURES=52
      workingDirectory: $(Build.SourcesDirectory)
  - task: Docker@2
--- a/tools/ci_build/github/azure-pipelines/nuget/cpu-mklml-esrp-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/nuget/cpu-mklml-esrp-pipeline.yml
@ -1,16 +0,0 @@
-schedules:
- cron: "0 10 * * *"
-  displayName: Daily Build
-  branches:
-    include:
-    - master
-  always: true
-
-variables:
-  PackageName: 'Microsoft.ML.OnnxRuntime.MKLML'
-
-jobs: 
- template: templates/cpu-mklml.yml
-  parameters:
-    AgentPool : 'Win-CPU-2019'
-    DoEsrp: 'true'
--- a/tools/ci_build/github/azure-pipelines/nuget/cpu-mklml-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/nuget/cpu-mklml-pipeline.yml
@ -1,8 +0,0 @@
-variables:
-  PackageName: 'Microsoft.ML.OnnxRuntime.MKLML'
-
-jobs: 
- template: templates/cpu-mklml.yml
-  parameters:
-    AgentPool : 'Win-CPU-2019'
-    DoEsrp: 'false'
--- a/tools/ci_build/github/azure-pipelines/nuget/templates/cpu-mklml.yml
+++ b/tools/ci_build/github/azure-pipelines/nuget/templates/cpu-mklml.yml
@ -1,190 +0,0 @@
-parameters:
-  DoEsrp: 'false'
-
-jobs: 
- template: ../../templates/win-ci-2019.yml
-  parameters:
-    AgentPool : 'Win-CPU-2019'
-    JobName: 'Windows_CI_Dev'
-    BuildCommand:  '--build_dir $(Build.BinariesDirectory) --skip_submodule_sync --use_mklml --use_winml --enable_wcos --build_shared_lib --enable_onnx_tests  --cmake_generator "Visual Studio 16 2019"'
-    BuildArch: 'x64'
-    EnvSetupScript: 'setup_env.bat'
-    sln_platform: 'x64'
-    DoDebugBuild: 'false'
-    DoNugetPack : 'true'
-    DoCompliance: 'false'
-    DoEsrp: ${{ parameters.DoEsrp }}
-    OrtPackageId: 'Microsoft.ML.OnnxRuntime.MKLML'
-    NuPackScript: |
-     msbuild $(Build.SourcesDirectory)\csharp\OnnxRuntime.CSharp.proj /p:Configuration=RelWithDebInfo /t:CreatePackage /p:OrtPackageId=Microsoft.ML.OnnxRuntime.MKLML
-     copy $(Build.SourcesDirectory)\csharp\src\Microsoft.ML.OnnxRuntime\bin\RelWithDebInfo\*.nupkg $(Build.ArtifactStagingDirectory)
-     copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\*.nupkg $(Build.ArtifactStagingDirectory)
-     mkdir $(Build.ArtifactStagingDirectory)\testdata
-     copy $(Build.BinariesDirectory)\RelWithDebInfo\RelWithDebInfo\custom_op_library.* $(Build.ArtifactStagingDirectory)\testdata
-
- job: 'Linux_CI_Dev'
-  workspace:
-    clean: all
-  timeoutInMinutes: 120
-  pool: $(AgentPoolLinux)
-  steps:
-    - template: ../../templates/set-version-number-variables-step.yml
-    - template: ../../templates/linux-set-variables-and-download.yml
-    - task: Docker@2
-      displayName: login
-      inputs:
-        containerRegistry: onnxruntimeregistry
-        command: login
-        addPipelineData: false
-    - task: CmdLine@2
-      inputs:
-        script: |
-          mkdir -p $HOME/.onnx
-          docker run --rm --volume /data/onnx:/data/onnx:ro --volume $(Build.SourcesDirectory):/onnxruntime_src --volume $(Build.BinariesDirectory):/build --volume /data/models:/build/models:ro --volume $HOME/.onnx:/home/onnxruntimedev/.onnx -e NIGHTLY_BUILD onnxruntimeregistry.azurecr.io/internal/azureml/onnxruntimecentoscpubuild:ch9m /bin/bash -c "python3 /onnxruntime_src/tools/ci_build/build.py --build_dir /build --config Release --skip_submodule_sync --parallel --build_shared_lib --use_openmp --enable_onnx_tests --use_mklml && cd /build/Release && make install DESTDIR=/build/linux-x64"
-        workingDirectory: $(Build.SourcesDirectory)
-    - task: Docker@2
-      displayName: logout
-      inputs:
-        containerRegistry: onnxruntimeregistry
-        command: logout
-        addPipelineData: false
-    - script: |
-       set -e -x
-       mv $(Build.BinariesDirectory)/linux-x64/usr/local/lib64 $(Build.BinariesDirectory)/linux-x64/linux-x64
-       cp $(Build.BinariesDirectory)/Release/mklml/src/project_mklml/lib/libmklml_gnu.so $(Build.BinariesDirectory)/linux-x64/linux-x64
-       ldd $(Build.BinariesDirectory)/linux-x64/linux-x64/libonnxruntime.so
-       cd $(Build.BinariesDirectory)/linux-x64
-       zip -r linux-x64.zip linux-x64
-       cp $(Build.BinariesDirectory)/linux-x64/linux*.zip $(Build.ArtifactStagingDirectory)
-       mkdir $(Build.ArtifactStagingDirectory)/testdata
-       cp $(Build.BinariesDirectory)/Release/libcustom_op_library.so* $(Build.ArtifactStagingDirectory)/testdata 
-       ls -al $(Build.ArtifactStagingDirectory)
-      displayName: 'Create Artifacts'
-    - task: PublishPipelineArtifact@0
-      displayName: 'Publish Pipeline Artifact'
-      inputs:
-        artifactName: 'drop-linux'
-        targetPath: '$(Build.ArtifactStagingDirectory)'
-    - template: ../../templates/component-governance-component-detection-steps.yml
-      parameters :
-        condition : 'succeeded'
-    - template: ../../templates/clean-agent-build-directory-step.yml
-
- template: ../../templates/mac-ci.yml
-  parameters:
-    AgentPool : $(AgentPoolMacOS)
-    JobName: 'MacOS_CI_Dev'
-    BuildCommand: 'python3 $(Build.SourcesDirectory)/tools/ci_build/build.py --build_dir $(Build.BinariesDirectory) --skip_submodule_sync --parallel --build_shared_lib --use_mklml --config RelWithDebInfo'
-    DoNugetPack : 'true'
-    NuPackScript: |
-     set -e -x
-     mkdir $(Build.BinariesDirectory)/osx-x64
-     find $(Build.BinariesDirectory)
-     cp $(Build.BinariesDirectory)/RelWithDebInfo/libonnxruntime.dylib $(Build.BinariesDirectory)/osx-x64/
-     cp $(Build.BinariesDirectory)/RelWithDebInfo/mklml/src/project_mklml/lib/libmklml.dylib $(Build.BinariesDirectory)/osx-x64/
-     cp $(Build.BinariesDirectory)/RelWithDebInfo/mklml/src/project_mklml/lib/libiomp5.dylib $(Build.BinariesDirectory)/osx-x64/
-     dsymutil $(Build.BinariesDirectory)/osx-x64/libonnxruntime.dylib -o $(Build.BinariesDirectory)/osx-x64/libonnxruntime.dylib.dSYM
-     strip -S -x $(Build.BinariesDirectory)/osx-x64/libonnxruntime.dylib
-     find $(Build.BinariesDirectory)/osx-x64 -ls
-     install_name_tool -change "@rpath/libmklml.dylib" "@loader_path/libmklml.dylib" $(Build.BinariesDirectory)/osx-x64/libonnxruntime.dylib
-     install_name_tool -change "@rpath/libiomp5.dylib" "@loader_path/libiomp5.dylib" $(Build.BinariesDirectory)/osx-x64/libmklml.dylib
-     otool -L $(Build.BinariesDirectory)/osx-x64/libonnxruntime.dylib
-     otool -L $(Build.BinariesDirectory)/osx-x64/libmklml.dylib
-     otool -L $(Build.BinariesDirectory)/osx-x64/libiomp5.dylib
-     cwd=`pwd`
-     cd $(Build.BinariesDirectory)
-     zip -r osx-x64.zip osx-x64
-     cp $(Build.BinariesDirectory)/osx-x64.zip $(Build.ArtifactStagingDirectory)
-     mkdir $(Build.ArtifactStagingDirectory)/testdata
-     cp $(Build.BinariesDirectory)/RelWithDebInfo/libcustom_op_library.dylib $(Build.ArtifactStagingDirectory)/testdata
-     cd $cwd
-
- job: NuGet_Packaging
-  workspace:
-    clean: all
-  pool: 'Win-CPU-2019'
-  dependsOn:
-  - Windows_CI_Dev
-  - Linux_CI_Dev
-  - MacOS_CI_Dev
-  condition: succeeded()
-  steps:  
-  - task: DownloadPipelineArtifact@0
-    displayName: 'Download Pipeline Artifact - NuGet'
-    inputs:
-      artifactName: 'drop-nuget'
-      targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
-    continueOnError: true
-
-  - task: DownloadPipelineArtifact@0
-    displayName: 'Download Pipeline Artifact - Linux'
-    inputs:
-      artifactName: 'drop-linux'
-      targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
-    continueOnError: true
-
-  - task: DownloadPipelineArtifact@0
-    displayName: 'Download Pipeline Artifact - MacOS'
-    inputs:
-      artifactName: 'drop-osx'
-      targetPath: '$(Build.BinariesDirectory)/nuget-artifact'
-    continueOnError: true
-
-  - script: |
-     pushd $(Build.BinariesDirectory)\nuget-artifact
-     dir
-     powershell -Command "Invoke-WebRequest http://stahlworks.com/dev/unzip.exe -OutFile unzip.exe"
-     powershell -Command "Invoke-WebRequest http://stahlworks.com/dev/zip.exe -OutFile zip.exe"
-     set PATH=%CD%;%PATH%
-     SETLOCAL EnableDelayedExpansion   
-     FOR /R %%i IN (*.nupkg) do (
-         set filename=%%~ni
-         IF NOT "!filename:~25,7!"=="Managed" (     
-             rename %%~ni.nupkg %%~ni.zip
-             unzip %%~ni.zip -d %%~ni
-             del /Q %%~ni.zip
-             unzip linux-x64.zip -d linux-x64
-             mkdir %%~ni\runtimes\linux-x64
-             mkdir %%~ni\runtimes\linux-x64\native
-             move linux-x64\linux-x64\libonnxruntime.so %%~ni\runtimes\linux-x64\native\libonnxruntime.so
-             move linux-x64\linux-x64\libmklml_gnu.so %%~ni\runtimes\linux-x64\native\libmklml_gnu.so
-             unzip osx-x64.zip -d osx-x64
-             dir osx-x64 /s
-             mkdir %%~ni\runtimes\osx-x64
-             mkdir %%~ni\runtimes\osx-x64\native
-             move osx-x64\osx-x64\libonnxruntime.dylib %%~ni\runtimes\osx-x64\native\libonnxruntime.dylib
-             move osx-x64\osx-x64\libonnxruntime.dylib.dSYM %%~ni\runtimes\osx-x64\native\libonnxruntime.dylib.dSYM
-             move osx-x64\osx-x64\libmklml.dylib %%~ni\runtimes\osx-x64\native\libmklml.dylib
-             move osx-x64\osx-x64\libiomp5.dylib %%~ni\runtimes\osx-x64\native\libiomp5.dylib
-             pushd %%~ni
-             zip -r ..\%%~ni.zip .
-             popd
-             move %%~ni.zip %%~ni.nupkg
-         )
-     )
-     popd
-     copy $(Build.BinariesDirectory)\nuget-artifact\*.nupkg $(Build.ArtifactStagingDirectory)
-    displayName: 'Bundle Native NuGet and other binaries'
-
-  - template: ../../templates/esrp_nuget.yml
-    parameters:
-      DisplayName: 'ESRP - sign NuGet package'
-      FolderPath: '$(Build.ArtifactStagingDirectory)'
-      DoEsrp: ${{ parameters.DoEsrp }}
-
-  - template: ../../templates/validate-nuget.yml
-    parameters:
-      NugetPath: '$(Build.ArtifactStagingDirectory)'
-      NugetPackage: 'Microsoft.ML.OnnxRuntime.MKLML*nupkg'
-      PlatformsSupported: 'win-x64,linux-x64,osx-x64'
-      VerifyNugetSigning: ${{ parameters.DoEsrp }}
-
-  - task: PublishPipelineArtifact@0
-    displayName: 'Publish Pipeline NuGet Artifact'
-    inputs:
-      artifactName: 'drop-signed-nuget'
-      targetPath: '$(Build.ArtifactStagingDirectory)'
-
- template: test_all_os.yml
-  parameters:
-    Skipx86Tests: 'true'
--- a/tools/ci_build/github/azure-pipelines/orttraining-linux-ci-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/orttraining-linux-ci-pipeline.yml
@ -38,7 +38,7 @@ jobs:
              --build_wheel \
              --use_openmp \
              --enable_onnx_tests \
-              --use_mklml --enable_pybind --build_java --build_nodejs --enable_training \
+              --enable_pybind --build_java --build_nodejs --enable_training \
              --cmake_extra_defines PYTHON_INCLUDE_DIR=/opt/python/cp37-cp37m/include/python3.7m PYTHON_LIBRARY=/usr/lib64/librt.so
      workingDirectory: $(Build.SourcesDirectory)
  - task: Docker@2
--- a/tools/ci_build/github/azure-pipelines/templates/linux-ci.yml
+++ b/tools/ci_build/github/azure-pipelines/templates/linux-ci.yml
@ -2,7 +2,7 @@ parameters:
  AgentPool : 'Linux-CPU'
  JobName : 'Linux_CI_Dev'
  SubmoduleCheckoutMode: ''
-  BuildCommand: 'tools/ci_build/github/linux/run_dockerbuild.sh -o ubuntu16.04 -d cpu -r $(Build.BinariesDirectory) -x "--use_mklml --use_tvm --build_wheel"'
+  BuildCommand: 'tools/ci_build/github/linux/run_dockerbuild.sh -o ubuntu16.04 -d cpu -r $(Build.BinariesDirectory) -x "--use_tvm --build_wheel"'
  DoNodejsPack: 'false'
  DoNugetPack: 'false'
  NuPackScript: ''
--- a/tools/ci_build/github/azure-pipelines/win-mklml-ci-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/win-mklml-ci-pipeline.yml
@ -1,6 +0,0 @@
-jobs:
- template: templates/win-ci.yml
-  parameters:
-    JobName: 'Windows_CI_Dev'
-    BuildCommand:  '$(Build.SourcesDirectory)\tools\ci_build\build.py --build_dir $(Build.BinariesDirectory) --skip_submodule_sync --cmake_path $(Build.BinariesDirectory)\cmake\bin\cmake.exe --ctest_path $(Build.BinariesDirectory)\cmake\bin\ctest.exe --use_mklml --build_shared_lib --build_csharp --enable_onnx_tests'
-