Remove EIGEN_USE_LIBXSMM support from Tensorflow.

PiperOrigin-RevId: 264222375
tensorflow · Aug 19, 2019 · afe53b5 · afe53b5
1 parent 1454690
commit afe53b5
Show file tree

Hide file tree

Showing 10 changed files with 111 additions and 175 deletions.
diff --git a/tensorflow/core/grappler/clusters/utils.cc b/tensorflow/core/grappler/clusters/utils.cc
@@ -27,7 +27,7 @@ limitations under the License.
 #include "rocm/include/hip/hip_runtime.h"
 #endif
 
-#ifdef EIGEN_USE_LIBXSMM
+#ifdef TENSORFLOW_USE_LIBXSMM
 #include "include/libxsmm.h"
 #endif
 
@@ -67,7 +67,7 @@ DeviceProperties GetLocalCPUInfo() {
 
   (*device.mutable_environment())["eigen"] = strings::StrCat(
       EIGEN_WORLD_VERSION, ".", EIGEN_MAJOR_VERSION, ".", EIGEN_MINOR_VERSION);
-#ifdef EIGEN_USE_LIBXSMM
+#ifdef TENSORFLOW_USE_LIBXSMM
   (*device.mutable_environment())["libxsmm"] = LIBXSMM_VERSION;
 #endif
 

diff --git a/tensorflow/core/grappler/optimizers/remapper.cc b/tensorflow/core/grappler/optimizers/remapper.cc
@@ -64,15 +64,6 @@ constexpr char kIsTraining[] = "is_training";
 
 constexpr int kMissingIndex = -1;
 
-// TODO(b/119765980): Upgrade upstream Eigen to set `m_can_use_xsmm=false` for
-// contractions with non-default contraction output kernels.
-bool EigenSupportsContractionOutputKernel() {
-#if defined(EIGEN_USE_LIBXSMM)
-  return false;
-#endif
-  return true;
-}
-
 struct RemapperContext {
   explicit RemapperContext(GrapplerItem* item, Status* status)
       : nodes_to_preserve(item->NodesToPreserve()),
@@ -353,8 +344,6 @@ inline bool HasAtMostOneFanoutAtPort0(const utils::MutableNodeView& node_view) {
 bool FindContractionWithBias(const RemapperContext& ctx, int node_index,
                              ContractionWithBiasAdd* matched,
                              bool check_device_compatible = true) {
-  if (!EigenSupportsContractionOutputKernel()) return false;
-
   const auto* node_view = ctx.graph_view.GetNode(node_index);
   // Root of the pattern must be a BiasAdd.
   // TODO(lyandy): Forward controls for patterns with control dependencies.
@@ -394,8 +383,6 @@ bool FindContractionWithBias(const RemapperContext& ctx, int node_index,
 bool FindContractionWithBiasAndActivation(
     const RemapperContext& ctx, int node_index,
     ContractionWithBiasAddAndActivation* matched) {
-  if (!EigenSupportsContractionOutputKernel()) return false;
-
   const auto* node_view = ctx.graph_view.GetNode(node_index);
   // Root of the pattern must be an activation node.
   // TODO(lyandy): Forward controls for patterns with control dependencies.
@@ -431,8 +418,6 @@ bool FindContractionWithBiasAndActivation(
 
 bool FindConv2DWithSqueezeAndBias(const RemapperContext& ctx, int node_index,
                                   ContractionWithSqueezeAndBiasAdd* matched) {
-  if (!EigenSupportsContractionOutputKernel()) return false;
-
   const auto* node_view = ctx.graph_view.GetNode(node_index);
   // TODO(lyandy): Forward controls for patterns with control dependencies.
   if (HasControlFaninOrFanout(*node_view)) return false;
@@ -488,8 +473,6 @@ bool FindConv2DWithSqueezeAndBias(const RemapperContext& ctx, int node_index,
 
 bool FindConv2DWithBatchNorm(const RemapperContext& ctx, int node_index,
                              ContractionWithBatchNorm* matched) {
-  if (!EigenSupportsContractionOutputKernel()) return false;
-
   const auto* node_view = ctx.graph_view.GetNode(node_index);
   const auto* node_def = node_view->node();
   // Root of the pattern must be a FusedBatchNorm.
@@ -539,8 +522,6 @@ bool FindConv2DWithBatchNorm(const RemapperContext& ctx, int node_index,
 bool FindConv2DWithBatchNormAndActivation(
     const RemapperContext& ctx, int node_index,
     ContractionWithBatchNormAndActivation* matched) {
-  if (!EigenSupportsContractionOutputKernel()) return false;
-
   const auto* node_view = ctx.graph_view.GetNode(node_index);
   // TODO(lyandy): Forward controls for patterns with control dependencies.
   if (HasControlFaninOrFanout(*node_view)) return false;

diff --git a/tensorflow/core/grappler/optimizers/remapper_test.cc b/tensorflow/core/grappler/optimizers/remapper_test.cc
@@ -37,15 +37,6 @@ class RemapperTest : public GrapplerTest {
     // This is a requirement for fusing FusedBatchNorm + SideInput + Activation.
     setenv("TF_USE_CUDNN_BATCHNORM_SPATIAL_PERSISTENT", "1", 1 /* replace */);
   }
-
-  // TODO(b/119765980): Upgrade upstream Eigen to set `m_can_use_xsmm=false` for
-  // contractions with non-default contraction output kernels.
-  bool EigenSupportsContractionOutputKernel() {
-#if defined(EIGEN_USE_LIBXSMM)
-    return false;
-#endif
-    return true;
-  }
 };
 
 TEST_F(RemapperTest, FusedBatchNorm) {
@@ -336,8 +327,6 @@ TEST_F(RemapperTest, FuseBatchNormWithAddAndRelu) {
 }
 
 TEST_F(RemapperTest, FuseConv2DWithBias) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ::tensorflow::ops::Placeholder;
 
   tensorflow::Scope s = tensorflow::Scope::NewRootScope();
@@ -400,8 +389,6 @@ TEST_F(RemapperTest, FuseConv2DWithBias) {
 }
 
 TEST_F(RemapperTest, FuseMatMulWithBias) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ::tensorflow::ops::Placeholder;
 
   tensorflow::Scope s = tensorflow::Scope::NewRootScope();
@@ -463,8 +450,6 @@ TEST_F(RemapperTest, FuseMatMulWithBias) {
 }
 
 TEST_F(RemapperTest, FuseConv2DWithBiasAndActivation) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ::tensorflow::ops::Placeholder;
 
   for (const string& activation : {"Relu", "Relu6", "Elu"}) {
@@ -545,8 +530,6 @@ TEST_F(RemapperTest, FuseConv2DWithBiasAndActivation) {
 }
 
 TEST_F(RemapperTest, FuseMatMulWithBiasAndActivation) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ::tensorflow::ops::Placeholder;
 
   for (const string& activation : {"Relu", "Relu6", "Elu"}) {
@@ -625,8 +608,6 @@ TEST_F(RemapperTest, FuseMatMulWithBiasAndActivation) {
 }
 
 TEST_F(RemapperTest, FuseConv2DWithBatchNorm) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ops::Placeholder;
 
   tensorflow::Scope s = tensorflow::Scope::NewRootScope();
@@ -705,8 +686,6 @@ TEST_F(RemapperTest, FuseConv2DWithBatchNorm) {
 }
 
 TEST_F(RemapperTest, FuseConv2DWithBatchNormAndActivation) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ops::Placeholder;
 
   for (const string& activation : {"Relu", "Relu6", "Elu"}) {
@@ -802,8 +781,6 @@ TEST_F(RemapperTest, FuseConv2DWithBatchNormAndActivation) {
 }
 
 TEST_F(RemapperTest, FuseConv2DWithSqueezeAndBias) {
-  if (!EigenSupportsContractionOutputKernel()) return;
-
   using ops::Placeholder;
 
   tensorflow::Scope s = tensorflow::Scope::NewRootScope();

diff --git a/tensorflow/core/kernels/BUILD b/tensorflow/core/kernels/BUILD
@@ -3783,10 +3783,7 @@ tf_kernel_library(
     ],
     hdrs = ["matmul_op.h"],
     defines = select({
-        ":xsmm": [
-            "TENSORFLOW_USE_LIBXSMM",
-            "EIGEN_USE_LIBXSMM",
-        ],
+        ":xsmm": ["TENSORFLOW_USE_LIBXSMM"],
         "//conditions:default": [],
     }),
     deps = MATH_DEPS + [
@@ -4209,12 +4206,7 @@ tf_kernel_library(
         "//conditions:default": [],
     }),
     defines = select({
-        ":xsmm_convolutions": [
-            "TENSORFLOW_USE_LIBXSMM_CONVOLUTIONS",
-        ],
-        "//conditions:default": [],
-    }) + select({
-        ":xsmm": ["EIGEN_USE_LIBXSMM"],
+        ":xsmm_convolutions": ["TENSORFLOW_USE_LIBXSMM_CONVOLUTIONS"],
         "//conditions:default": [],
     }) + select({
         ":xsmm_backward_convolutions": ["TENSORFLOW_USE_LIBXSMM_BACKWARD_CONVOLUTIONS"],

diff --git a/tensorflow/core/kernels/conv_ops_fused_double.cc b/tensorflow/core/kernels/conv_ops_fused_double.cc
@@ -23,9 +23,7 @@ namespace tensorflow {
 
 // If we're using the alternative GEMM-based implementation of Conv2D for the
 // CPU implementation, don't register this EigenTensor-based version.
-// TODO(b/119765980): Upgrade upstream Eigen to set `m_can_use_xsmm=false` for
-// contractions with non-default contraction output kernels.
-#if !defined(USE_GEMM_FOR_CONV) && !defined(EIGEN_USE_LIBXSMM)
+#if !defined(USE_GEMM_FOR_CONV)
 TF_CALL_double(REGISTER_FUSED_CPU_CONV2D);
 #endif  // !USE_GEMM_FOR_CONV
 

diff --git a/tensorflow/core/kernels/conv_ops_fused_float.cc b/tensorflow/core/kernels/conv_ops_fused_float.cc
@@ -23,9 +23,7 @@ namespace tensorflow {
 
 // If we're using the alternative GEMM-based implementation of Conv2D for the
 // CPU implementation, don't register this EigenTensor-based version.
-// TODO(b/119765980): Upgrade upstream Eigen to set `m_can_use_xsmm=false` for
-// contractions with non-default contraction output kernels.
-#if !defined(USE_GEMM_FOR_CONV) && !defined(EIGEN_USE_LIBXSMM)
+#if !defined(USE_GEMM_FOR_CONV)
 TF_CALL_float(REGISTER_FUSED_CPU_CONV2D);
 #endif  // !USE_GEMM_FOR_CONV
 

diff --git a/tensorflow/core/kernels/matmul_op_fused.cc b/tensorflow/core/kernels/matmul_op_fused.cc
@@ -189,9 +189,7 @@ class FusedMatMulOp : public OpKernel {
       Name("_FusedMatMul").Device(DEVICE_CPU).TypeConstraint<T>("T"), \
       FusedMatMulOp<CPUDevice, T>);
 
-#ifndef EIGEN_USE_LIBXSMM
 TF_CALL_float(REGISTER_FUSED_CPU_MATMUL);
-#endif  // !EIGEN_USE_LIBXSMM
 
 #undef REGISTER_FUSED_CPU_MATMUL
 

diff --git a/tensorflow/lite/kernels/internal/optimized/eigen_tensor_reduced_instantiations_google.h b/tensorflow/lite/kernels/internal/optimized/eigen_tensor_reduced_instantiations_google.h
@@ -70,10 +70,6 @@ typedef unsigned __int64 uint64_t;
 #include <time.h>
 #endif
 
-// #if defined(EIGEN_USE_LIBXSMM)
-// #include "libxsmm.h"
-// #endif
-
 #ifdef EIGEN_USE_THREADS
 #include "third_party/eigen3/unsupported/Eigen/CXX11/ThreadPool"
 #endif

diff --git a/tensorflow/lite/kernels/internal/optimized/eigen_tensor_reduced_instantiations_oss.h b/tensorflow/lite/kernels/internal/optimized/eigen_tensor_reduced_instantiations_oss.h
@@ -70,10 +70,6 @@ typedef unsigned __int64 uint64_t;
 #include <time.h>
 #endif
 
-// #if defined(EIGEN_USE_LIBXSMM)
-// #include "libxsmm.h"
-// #endif
-
 #ifdef EIGEN_USE_THREADS
 #include "unsupported/Eigen/CXX11/ThreadPool"
 #endif