[INTEL MKL] Remove unnecessary MKL build macros and bug fixes

tensorflow/core/common_runtime/mkl_cpu_allocator.h

@@ -29,10 +29,6 @@ limitations under the License.
 #include "tensorflow/core/platform/mem.h"
 #include "tensorflow/core/platform/numa.h"
 
-#ifndef INTEL_MKL_DNN_ONLY
-#include "i_malloc.h"
-#endif
-
 #ifdef _WIN32
 typedef unsigned int uint;
 #endif
@@ -186,14 +182,6 @@ class MklCPUAllocator : public Allocator {
         new MklSmallSizeAllocator(sub_allocator_, max_mem_bytes, kName);
     large_size_allocator_ =
         new BFCAllocator(sub_allocator_, max_mem_bytes, kAllowGrowth, kName);
-#ifndef INTEL_MKL_DNN_ONLY
-    // For redirecting all allocations from MKL to this allocator
-    // From: http://software.intel.com/en-us/node/528565
-    i_malloc = MallocHook;
-    i_calloc = CallocHook;
-    i_realloc = ReallocHook;
-    i_free = FreeHook;
-#endif
     return Status::OK();
   }
 

tensorflow/core/grappler/optimizers/auto_mixed_precision.cc

@@ -1986,7 +1986,7 @@ Status AutoMixedPrecision::Optimize(Cluster* cluster, const GrapplerItem& item,
     return errors::InvalidArgument("cluster == nullptr");
   }
 
-#if !defined(INTEL_MKL) || !defined(ENABLE_INTEL_MKL_BFLOAT16)
+#if !defined(INTEL_MKL)
   if (mode_ == AutoMixedPrecisionMode::MKL) {
     return errors::Unimplemented(
         "The auto_mixed_precision_mkl optimizer cannot be used since "

tensorflow/core/grappler/optimizers/auto_mixed_precision_test.cc

@@ -13,8 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM || \
-    (INTEL_MKL && defined(ENABLE_INTEL_MKL_BFLOAT16))
+#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM || INTEL_MKL 
 
 #include "tensorflow/core/grappler/optimizers/auto_mixed_precision.h"
 
@@ -1178,7 +1177,6 @@ TEST_F(AutoMixedPrecisionTest, TanhOp) {
 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
 
 #if INTEL_MKL
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
 
 class AutoMixedPrecisionMklTest : public GrapplerTest {
  protected:
@@ -1354,12 +1352,10 @@ TEST_F(AutoMixedPrecisionMklTest, TensorListSetGet) {
   }
 }
 
-#endif  // ENABLE_INTEL_MKL_BFLOAT16
 #endif  // INTEL_MKL
 
 }  // namespace
 }  // namespace grappler
 }  // namespace tensorflow
 
-#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM || (INTEL_MKL &&
-        // defined(ENABLE_INTEL_MKL_BFLOAT16))
+#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM || INTEL_MKL

tensorflow/core/grappler/optimizers/remapper.cc

@@ -230,15 +230,9 @@ bool IsCpuCompatibleDataType(const NodeDef* contraction,
                              const string& type_attr = "T") {
   DataType dtype = GetDataTypeFromAttr(*contraction, type_attr);
 #if defined(INTEL_MKL)
-#if defined(ENABLE_INTEL_MKL_BFLOAT16)
   if (IsConv2D(*contraction) || IsDepthwiseConv2dNative(*contraction) ||
       IsMatMul(*contraction)) {
     return dtype == DT_FLOAT || dtype == DT_BFLOAT16;
-#else
-  if (IsConv2D(*contraction) || IsDepthwiseConv2dNative(*contraction) ||
-      IsMatMul(*contraction)) {
-    return dtype == DT_FLOAT;
-#endif  // ENABLE_INTEL_MKL_BFLOAT16
 #else
   if (IsConv2D(*contraction)) {
     return dtype == DT_FLOAT || dtype == DT_DOUBLE;
@@ -677,14 +671,9 @@ bool FindContractionWithBiasAddAndAdd(const RemapperContext& ctx,
   const auto* node_def = node_view.node();
   if (!IsAddN(*node_def) && !IsAddWithNoBroadcast(ctx, *node_def)) return false;
 
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
   // MKL AddN ops only support float and bfloat16 data types.
   if (!HasDataType(node_def, DT_FLOAT) && !HasDataType(node_def, DT_BFLOAT16))
     return false;
-#else
-  // MKL AddN ops only support float data type.
-  if (!HasDataType(node_def, DT_FLOAT)) return false;
-#endif  // ENABLE_INTEL_MKL_BFLOAT16
 
   ContractionWithBiasAdd base;
   matched->port_id = 0;
@@ -730,14 +719,9 @@ bool FindContractionWithBiasAndAddActivation(
   // Currently, Contraction + Bias + Add + Tanh pattern is not supported
   if (IsTanh(*node_def)) return false;
 
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
   // MKL activation op only supports float and bfloat16 data types.
   if (!HasDataType(node_def, DT_FLOAT) && !HasDataType(node_def, DT_BFLOAT16))
     return false;
-#else
-  // MKL activation op only supports float data type.
-  if (!HasDataType(node_def, DT_FLOAT)) return false;
-#endif  // ENABLE_INTEL_MKL_BFLOAT16
 
   // And input to activation must match ContractionWithBiasAddAndAdd pattern.
   if (node_view->NumRegularFanins() < 1) return false;
@@ -843,15 +827,15 @@ bool FindFusedBatchNormEx(const RemapperContext& ctx, int node_index,
     const auto* fused_batch_norm_node_def = fused_batch_norm.node();
     if (!IsFusedBatchNorm(*fused_batch_norm_node_def)) return false;
 
-#ifndef ENABLE_MKLDNN_V1
+#ifndef INTEL_MKL
     // We fuse FusedBatchNorm on GPU or MKL CPU.
     if (!NodeIsOnGpu(fused_batch_norm_node_def)) return false;
 #else
     if (DisableMKL()) return false;
 #endif
 
     DataType t_dtype = GetDataTypeFromAttr(*fused_batch_norm_node_def, "T");
-#ifndef ENABLE_MKLDNN_V1
+#ifndef INTEL_MKL
     if (t_dtype != DT_FLOAT && t_dtype != DT_HALF) return false;
 #else
     if (t_dtype != DT_FLOAT && t_dtype != DT_BFLOAT16) return false;
@@ -919,7 +903,7 @@ bool FindFusedBatchNormEx(const RemapperContext& ctx, int node_index,
   if (IsAdd(*relu_fanin_0_node_def)) {
     // Currently no CPU implementation for "FusedBatchNorm + SideInput +
     // <Activation>""
-#ifdef ENABLE_MKLDNN_V1
+#ifdef INTEL_MKL
     return false;
 #endif
 
@@ -1022,7 +1006,7 @@ void CopyFusedBatchNormAttributes(const NodeDef& fused_batch_norm,
   if (fused_batch_norm.op() != "FusedBatchNorm") {
     SetAttrValue(src_attr.at("U"), &(*attr)["U"]);
   } else {
-#ifndef ENABLE_MKLDNN_V1
+#ifndef INTEL_MKL
     SetAttrValue(src_attr.at("T"), &(*attr)["U"]);
 #else
     SetAttrValue(DT_FLOAT, &(*attr)["U"]);

tensorflow/core/grappler/optimizers/remapper_test.cc

@@ -461,10 +461,10 @@ class RemapperFuseMatMulWithBiasTest : public RemapperTest {
 TEST_F(RemapperFuseMatMulWithBiasTest, F32) { RunTest<DT_FLOAT>(); }
 
 TEST_F(RemapperFuseMatMulWithBiasTest, Bf16) {
-#if !defined(INTEL_MKL) || !defined(ENABLE_INTEL_MKL_BFLOAT16)
+#if !defined(INTEL_MKL)
   GTEST_SKIP() << "Intel MKL with bfloat16 support is not enabled, skipping "
                   "FuseMatMulWithBias with bfloat16.";
-#endif  // !defined(INTEL_MKL) || !defined(ENABLE_INTEL_MKL_BFLOAT16)
+#endif  // !defined(INTEL_MKL)
   RunTest<DT_BFLOAT16>();  // NOLINT
 }
 
@@ -742,10 +742,10 @@ TEST_F(RemapperFuseMatMulWithBiasAndActivationTest, F32) {
 }
 
 TEST_F(RemapperFuseMatMulWithBiasAndActivationTest, Bf16) {
-#if !defined(INTEL_MKL) || !defined(ENABLE_INTEL_MKL_BFLOAT16)
+#if !defined(INTEL_MKL)
   GTEST_SKIP() << "Intel MKL with bfloat16 support is not enabled, skipping "
                   "FuseMatMulWithBiasAndActivation with bfloat16.";
-#endif  // !defined(INTEL_MKL) || !defined(ENABLE_INTEL_MKL_BFLOAT16)
+#endif  // !defined(INTEL_MKL)
   RunTest<DT_BFLOAT16>();  // NOLINT
 }
 

tensorflow/core/kernels/mkl/mkl_conv_ops_test.cc

@@ -26,10 +26,7 @@ limitations under the License.
 #include "tensorflow/core/platform/test.h"
 #include "tensorflow/core/platform/test_benchmark.h"
 #include "tensorflow/core/public/session.h"
-
-#if defined(INTEL_MKL_DNN_ONLY)
 #include "tensorflow/core/util/mkl_util.h"
-#endif
 
 // TODO(ezhulenev): Add numerical tests that will compare results of default
 // (aka Eigen) convolutions with MKL convolutions.
@@ -123,7 +120,6 @@ static Graph* DefaultConv2D(const Conv2DDimensions& dims) {
   return graph;
 }
 
-#if defined(INTEL_MKL_DNN_ONLY)
 static Graph* MklConv2D(const Conv2DDimensions& dims) {
   auto* graph = new Graph(OpRegistry::Global());
 
@@ -150,7 +146,6 @@ static Graph* MklConv2D(const Conv2DDimensions& dims) {
 
   return graph;
 }
-#endif
 
 static Graph* DefaultConv2DBwdInput(const Conv2DDimensions& dims) {
   auto* graph = new Graph(OpRegistry::Global());
@@ -179,7 +174,6 @@ static Graph* DefaultConv2DBwdInput(const Conv2DDimensions& dims) {
   return graph;
 }
 
-#if defined(INTEL_MKL_DNN_ONLY)
 static Graph* MklConv2DBwdInput(const Conv2DDimensions& dims) {
   auto* graph = new Graph(OpRegistry::Global());
 
@@ -213,7 +207,6 @@ static Graph* MklConv2DBwdInput(const Conv2DDimensions& dims) {
 
   return graph;
 }
-#endif
 
 static Graph* DefaultConv2DBwdFilter(const Conv2DDimensions& dims) {
   auto* graph = new Graph(OpRegistry::Global());
@@ -243,7 +236,6 @@ static Graph* DefaultConv2DBwdFilter(const Conv2DDimensions& dims) {
   return graph;
 }
 
-#if defined(INTEL_MKL_DNN_ONLY)
 static Graph* MklConv2DBwdFilter(const Conv2DDimensions& dims) {
   Graph* graph = new Graph(OpRegistry::Global());
 
@@ -278,7 +270,6 @@ static Graph* MklConv2DBwdFilter(const Conv2DDimensions& dims) {
 
   return graph;
 }
-#endif
 
 // Macro arguments names: --------------------------------------------------- //
 //    N: batch size
@@ -311,14 +302,9 @@ static Graph* MklConv2DBwdFilter(const Conv2DDimensions& dims) {
   }                                                                      \
   BENCHMARK(BM_NAME(Conv2D_##kind, type, N, H, W, C, FC, FH, FW))
 
-#if defined(INTEL_MKL_DNN_ONLY)
 #define BM_Conv2D(N, H, W, C, FC, FH, FW, type, LABEL)      \
   BM_Conv2DT(Default, N, H, W, C, FC, FH, FW, type, LABEL); \
   BM_Conv2DT(Mkl, N, H, W, C, FC, FH, FW, type, LABEL);
-#else
-#define BM_Conv2D(N, H, W, C, FC, FH, FW, type, LABEL) \
-  BM_Conv2DT(Default, N, H, W, C, FC, FH, FW, type, LABEL);
-#endif
 
 #define BM_Conv2DBwdInputT(kind, N, H, W, C, FC, FH, FW, type, LABEL)         \
   static void BM_NAME(Conv2DBwdInput_##kind, type, N, H, W, C, FC, FH,        \
@@ -334,14 +320,9 @@ static Graph* MklConv2DBwdFilter(const Conv2DDimensions& dims) {
   }                                                                           \
   BENCHMARK(BM_NAME(Conv2DBwdInput_##kind, type, N, H, W, C, FC, FH, FW))
 
-#if defined(INTEL_MKL_DNN_ONLY)
 #define BM_Conv2DBwdInput(N, H, W, C, FC, FH, FW, type, LABEL)      \
   BM_Conv2DBwdInputT(Default, N, H, W, C, FC, FH, FW, type, LABEL); \
   BM_Conv2DBwdInputT(Mkl, N, H, W, C, FC, FH, FW, type, LABEL);
-#else
-#define BM_Conv2DBwdInput(N, H, W, C, FC, FH, FW, type, LABEL) \
-  BM_Conv2DBwdInputT(Default, N, H, W, C, FC, FH, FW, type, LABEL);
-#endif
 
 #define BM_Conv2DBwdFilterT(kind, N, H, W, C, FC, FH, FW, type, LABEL)         \
   static void BM_NAME(Conv2DBwdFilter_##kind, type, N, H, W, C, FC, FH,        \
@@ -357,14 +338,9 @@ static Graph* MklConv2DBwdFilter(const Conv2DDimensions& dims) {
   }                                                                            \
   BENCHMARK(BM_NAME(Conv2DBwdFilter_##kind, type, N, H, W, C, FC, FH, FW))
 
-#if defined(INTEL_MKL_DNN_ONLY)
 #define BM_Conv2DBwdFilter(N, H, W, C, FC, FH, FW, type, LABEL)      \
   BM_Conv2DBwdFilterT(Default, N, H, W, C, FC, FH, FW, type, LABEL); \
   BM_Conv2DBwdFilterT(Mkl, N, H, W, C, FC, FH, FW, type, LABEL);
-#else
-#define BM_Conv2DBwdFilter(N, H, W, C, FC, FH, FW, type, LABEL) \
-  BM_Conv2DBwdFilterT(Default, N, H, W, C, FC, FH, FW, type, LABEL);
-#endif
 
 // ImageNet Convolutions ---------------------------------------------------- //
 

tensorflow/core/kernels/mkl/mkl_fused_ops_test.cc

@@ -888,11 +888,7 @@ TYPED_TEST_P(FusedPadConvOpTest, PaddingConvTestNchw) { this->Run("NCHW"); }
 REGISTER_TYPED_TEST_SUITE_P(FusedPadConvOpTest, PaddingConvTest,
                             PaddingConvTestNchw);
 
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
 using FusedPadConvDataTypes = ::testing::Types<float, bfloat16>;
-#else
-using FusedPadConvDataTypes = ::testing::Types<float>;
-#endif
 INSTANTIATE_TYPED_TEST_SUITE_P(Test, FusedPadConvOpTest, FusedPadConvDataTypes);
 
 class FilterCacheTest : public OpsTestBase {

tensorflow/core/kernels/mkl/mkl_matmul_op.cc

@@ -167,7 +167,6 @@ class MklMatMulOp : public OpKernel {
 #endif  // ENABLE_MKLDNN_THREADPOOL
   }
 
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
   void MklBlasGemm(OpKernelContext* ctx, bool transa, bool transb, const int m,
                    const int n, const int k, const bfloat16* a, const int lda,
                    const bfloat16* b, const int ldb, bfloat16* c,
@@ -181,7 +180,6 @@ class MklMatMulOp : public OpKernel {
     dnnl_gemm<bfloat16>(ftrans[index_transa], ftrans[index_transb], m, n, k,
                         alpha, a, lda, b, ldb, beta, c, ldc, ctx);
   }
-#endif  // ENABLE_INTEL_MKL_BFLOAT16
 };
 
 #define REGISTER_CPU(T)                                   \
@@ -196,9 +194,7 @@ class MklMatMulOp : public OpKernel {
 // TODO(inteltf) Consider template specialization when adding/removing
 // additional types
 TF_CALL_float(REGISTER_CPU);
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
 TF_CALL_bfloat16(REGISTER_CPU);
-#endif  // ENABLE_INTEL_MKL_BFLOAT16
 #endif  // ENABLE_MKL
 }  // namespace tensorflow
 #endif  // INTEL_MKL

tensorflow/core/kernels/mkl/mkl_matmul_ops_common.h

@@ -51,12 +51,11 @@ struct MklDnnMatMulFwdParams {
   };
   std::vector<PostOpParam> post_op_params;
 
-  MklDnnMatMulFwdParams(
-      memory::dims src_dims, memory::dims weight_dims, memory::dims bias_dims,
-      memory::dims dst_dims,
-      memory::format_tag src_format = memory::format_tag::any,
-      memory::format_tag weight_format = memory::format_tag::any,
-      memory::format_tag dst_format = memory::format_tag::any)
+  MklDnnMatMulFwdParams(memory::dims src_dims, memory::dims weight_dims,
+                        memory::dims bias_dims, memory::dims dst_dims,
+                        memory::format_tag src_format = memory::format_tag::any,
+                        memory::format_tag weight_format = memory::format_tag::any,
+                        memory::format_tag dst_format = memory::format_tag::any)
       : src_dims(src_dims),
         weight_dims(weight_dims),
         bias_dims(bias_dims),

tensorflow/core/kernels/mkl/mkl_tmp_bf16_ops.cc

@@ -19,8 +19,6 @@ limitations under the License.
 #include "tensorflow/core/framework/register_types.h"
 #include "tensorflow/core/kernels/no_op.h"
 
-#ifdef ENABLE_INTEL_MKL_BFLOAT16
-
 namespace tensorflow {
 
 // This file contains temporary registrations for some of the Eigen CPU backend
@@ -62,5 +60,3 @@ TF_CALL_bfloat16(REGISTER_CPU);
 #undef REGISTER_CPU
 
 }  // namespace tensorflow
-
-#endif  // ENABLE_INTEL_MKL_BFLOAT16