Upgraded to the latest version of Eigen that speeds up full reductions on fp16

by about 3 orders of magnitude as well as some partial reductions by 30% when using cuda 7.5 or above
Change: 122191448

eigen.BUILD

@@ -1,6 +1,6 @@
 package(default_visibility = ["//visibility:public"])
 
-archive_dir = "eigen-eigen-50812b426b7c"
+archive_dir = "eigen-eigen-aaca054ed24d"
 
 cc_library(
     name = "eigen",

tensorflow/contrib/cmake/external/eigen.cmake

@@ -7,7 +7,7 @@
 
 include (ExternalProject)
 
-set(eigen_archive_hash "50812b426b7c")
+set(eigen_archive_hash "aaca054ed24d")
 
 set(eigen_INCLUDE_DIRS
     ${CMAKE_CURRENT_BINARY_DIR}
@@ -16,7 +16,7 @@ set(eigen_INCLUDE_DIRS
     ${tensorflow_source_dir}/third_party/eigen3
 )
 set(eigen_URL https://bitbucket.org/eigen/eigen/get/${eigen_archive_hash}.tar.gz)
-set(eigen_HASH SHA256=fa95e425c379c2c7b8a49d9ef7bd0c5a8369171c987affd6dbae5de8a8911c1a)
+set(eigen_HASH SHA256=4abff4b7ba03316856aeece6de99abb74a1dfed453fdab85eed4da2b5fd2fc59)
 set(eigen_BUILD ${CMAKE_CURRENT_BINARY_DIR}/eigen/src/eigen)
 set(eigen_INSTALL ${CMAKE_CURRENT_BINARY_DIR}/eigen/install)
 

tensorflow/core/common_runtime/gpu/gpu_device.cc

@@ -122,14 +122,19 @@ class EigenAllocator : public ::Eigen::Allocator {
 #else
 class EigenCudaStreamDevice : public ::Eigen::StreamInterface {
  public:
-  EigenCudaStreamDevice() { Eigen::initializeDeviceProp(); }
+  EigenCudaStreamDevice() : scratch_(nullptr) { Eigen::initializeDeviceProp(); }
-
+  ~EigenCudaStreamDevice() {
+    if (scratch_) {
+      deallocate(scratch_);
+    }
+  }
   void Reinitialize(OpKernelContext* context, const cudaStream_t* cuda_stream,
                     int gpu_id, ::tensorflow::Allocator* alloc) {
     if (LogMemory::IsEnabled()) {
       operation_ = context->op_kernel().name() + "/EigenAllocator";
       step_id_ = context->step_id();
     }
+    assert(!scratch_);
     stream_ = cuda_stream;
     allocator_ = alloc;
     device_prop_ = &Eigen::m_deviceProperties[gpu_id];
@@ -163,6 +168,15 @@ class EigenCudaStreamDevice : public ::Eigen::StreamInterface {
     CHECK_EQ(err, cudaSuccess);
   }
 
+  // Return a pointer to a per stream scratchpad of 1024 bytes residing
+  // in global memory.
+  void* scratchpad() const {
+    if (scratch_ == nullptr) {
+      scratch_ = allocate(1024);
+    }
+    return scratch_;
+  }
+
  private:
   struct AsyncFreeData {
     AsyncFreeData(::tensorflow::Allocator* a, void* p, const string& o,
@@ -190,6 +204,7 @@ class EigenCudaStreamDevice : public ::Eigen::StreamInterface {
   const cudaStream_t* stream_;          // Not owned.
   const cudaDeviceProp* device_prop_;   // Not owned.
   ::tensorflow::Allocator* allocator_;  // Not owned.
+  mutable void* scratch_;
 
   TF_DISALLOW_COPY_AND_ASSIGN(EigenCudaStreamDevice);
 };

tensorflow/python/kernel_tests/xent_op_test.py

@@ -41,8 +41,8 @@ class XentTest(tf.test.TestCase):
       loss = tf.nn.softmax_cross_entropy_with_logits(np_features, np_labels)
       backprop = loss.op.outputs[1]
       tf_loss, tf_backprop = sess.run([loss, backprop])
-    self.assertAllClose(np_loss, tf_loss)
+    self.assertAllCloseAccordingToType(np_loss, tf_loss)
-    self.assertAllClose(np_backprop, tf_backprop)
+    self.assertAllCloseAccordingToType(np_backprop, tf_backprop)
 
   def _testAll(self, features, labels):
     self._testXent(features, labels, use_gpu=False)

tensorflow/workspace.bzl

@@ -13,8 +13,8 @@ def tf_workspace(path_prefix = "", tf_repo_name = ""):
 
   native.new_http_archive(
     name = "eigen_archive",
-    url = "https://bitbucket.org/eigen/eigen/get/50812b426b7c.tar.gz",
+    url = "https://bitbucket.org/eigen/eigen/get/aaca054ed24d.tar.gz",
-    sha256 = "fa95e425c379c2c7b8a49d9ef7bd0c5a8369171c987affd6dbae5de8a8911c1a",
+    sha256 = "4abff4b7ba03316856aeece6de99abb74a1dfed453fdab85eed4da2b5fd2fc59",
     build_file = path_prefix + "eigen.BUILD",
   )
 

third_party/eigen3/Eigen/Cholesky

@@ -1 +1 @@
-#include "eigen-eigen-50812b426b7c/Eigen/Cholesky"
+#include "eigen-eigen-aaca054ed24d/Eigen/Cholesky"

third_party/eigen3/Eigen/Core

@@ -1 +1 @@
-#include "eigen-eigen-50812b426b7c/Eigen/Core"
+#include "eigen-eigen-aaca054ed24d/Eigen/Core"

third_party/eigen3/Eigen/Eigenvalues

@@ -1 +1 @@
-#include "eigen-eigen-50812b426b7c/Eigen/Eigenvalues"
+#include "eigen-eigen-aaca054ed24d/Eigen/Eigenvalues"

third_party/eigen3/Eigen/LU

@@ -1 +1 @@
-#include "eigen-eigen-50812b426b7c/Eigen/LU"
+#include "eigen-eigen-aaca054ed24d/Eigen/LU"

third_party/eigen3/Eigen/QR

@@ -1 +1 @@
-#include "eigen-eigen-50812b426b7c/Eigen/QR"
+#include "eigen-eigen-aaca054ed24d/Eigen/QR"

third_party/eigen3/unsupported/Eigen/CXX11/Tensor

@@ -1 +1 @@
-#include "eigen-eigen-50812b426b7c/unsupported/Eigen/CXX11/Tensor"
+#include "eigen-eigen-aaca054ed24d/unsupported/Eigen/CXX11/Tensor"