6 files changed, 27 insertions, 24 deletions
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index 640f270323..102419a264 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -524,11 +524,12 @@ port::Status CudnnSupport::Init() {
                                    ToString(status))};
 }
 
-port::StatusOr<std::tuple<int, int, int>> CudnnSupport::GetVersion() {
+port::StatusOr<perftools::gputools::dnn::VersionInfo>
+CudnnSupport::GetVersion() {
   CudnnVersion version;
   TF_RETURN_IF_ERROR(GetLoadedCudnnVersion(&version));
-  return std::make_tuple(version.major_version, version.minor_version,
-                         version.patch_level);
+  return perftools::gputools::dnn::VersionInfo(
+      version.major_version, version.minor_version, version.patch_level);
 }
 
 // Turns a BatchDescriptor structure into a cudnn tensor handle within a scope.
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.h b/tensorflow/stream_executor/cuda/cuda_dnn.h
index e6d12bfef9..5ded7cf154 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.h
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.h
@@ -45,7 +45,7 @@ class CudnnSupport : public dnn::DnnSupport {
   ~CudnnSupport() override;
 
   port::Status Init() override;
-  port::StatusOr<std::tuple<int, int, int>> GetVersion() override;
+  port::StatusOr<perftools::gputools::dnn::VersionInfo> GetVersion() override;
 
   port::StatusOr<std::unique_ptr<dnn::RnnDescriptor>> createRnnDescriptor(
       int num_layers, int hidden_size, int input_size,
diff --git a/tensorflow/stream_executor/cuda/cuda_driver.cc b/tensorflow/stream_executor/cuda/cuda_driver.cc
index fedf4f53b8..71cab145b9 100644
--- a/tensorflow/stream_executor/cuda/cuda_driver.cc
+++ b/tensorflow/stream_executor/cuda/cuda_driver.cc
@@ -37,14 +37,6 @@ limitations under the License.
 #include "tensorflow/stream_executor/platform/port.h"
 #include "tensorflow/stream_executor/lib/inlined_vector.h"
 
-#if defined(PLATFORM_WINDOWS)
-// TODO: in windows ARRAYSIZE is defined in winnt.h but including it
-//  here creates a conflict with cuda.h - for now define it here.
-#define ARRAYSIZE(a) \
-  ((sizeof(a) / sizeof(*(a))) / \
-  static_cast<size_t>(!(sizeof(a) % sizeof(*(a)))))
-#endif
-
 bool FLAGS_gpuexec_cuda_driver_inject_init_error = false;
 bool FLAGS_gpuexec_cuda_sync_around_driver_calls = false;
 bool FLAGS_gpuexec_cuda_device_0_only = false;
@@ -719,15 +711,15 @@ CUDADriver::ContextGetSharedMemConfig(CudaContext* context) {
         port::bit_cast<void *>(uintptr_t(info_log_buffer_bytes)),
         port::bit_cast<void *>(info_log_buffer.data()),
         port::bit_cast<void *>(uintptr_t(log_verbose))};
-    CHECK(ARRAYSIZE(options) == ARRAYSIZE(option_values));
+    CHECK(TF_ARRAYSIZE(options) == TF_ARRAYSIZE(option_values));
 
     CUresult res;
     {
       // TODO(leary) Need to see if NVIDIA can expunge the leakiness in their
       // module loading: see http://b/13248943
 
-      res = cuModuleLoadDataEx(module, ptx_data, ARRAYSIZE(options), options,
-                               option_values);
+      res = cuModuleLoadDataEx(module, ptx_data, TF_ARRAYSIZE(options),
+                               options, option_values);
     }
 
     // The PTX JIT mutates the values in the option values array to reflect the
diff --git a/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc b/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc
index 9700daca89..7c87d33d21 100644
--- a/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc
+++ b/tensorflow/stream_executor/cuda/cuda_gpu_executor.cc
@@ -1126,7 +1126,7 @@ DeviceDescription *CUDAExecutor::PopulateDeviceDescription() const {
     builder.set_name(device_name);
   }
 
-  for (size_t i = 0; i < ARRAYSIZE(kAllUnqueryableDeviceParams); i++) {
+  for (size_t i = 0; i < TF_ARRAYSIZE(kAllUnqueryableDeviceParams); i++) {
     const auto &params = kAllUnqueryableDeviceParams[i];
     if (params.cc_major == cc_major_ && params.cc_minor == cc_minor_) {
       builder.set_blocks_per_core_limit(params.blocks_per_core_limit);
diff --git a/tensorflow/stream_executor/dnn.h b/tensorflow/stream_executor/dnn.h
index 8e202d115a..39f21d8b10 100644
--- a/tensorflow/stream_executor/dnn.h
+++ b/tensorflow/stream_executor/dnn.h
@@ -875,6 +875,22 @@ enum class ElementwiseOperation { kAdd, kMultiply };
 
 string ElementwiseOperationString(ElementwiseOperation op);
 
+// A simple class representing the version of the backing library, to 
+// workaround the "too perfect forwarding" issue in gcc6+ compilers. 
+// See PR#16309 and issue #18402 for links discussing the issue.
+class VersionInfo {
+ public:
+  VersionInfo(int major = 0, int minor = 0, int patch = 0)
+      : major_(major), minor_(minor), patch_(patch) {}
+  int major_version() { return major_; }
+  int minor_version() { return minor_; }
+  int patch() { return patch_; }
+ private:
+  int major_;
+  int minor_;
+  int patch_;
+};
+
 // Suite of operations typically used for implementing Deep/Convolutional Neural
 // Nets. Note: A false return value of an operation indicates the
 // implementation is not available.
@@ -885,8 +901,8 @@ class DnnSupport {
 
   virtual port::Status Init() = 0;
 
-  // Gets the version of the backing library, as a {major, minor, patch} tuple.
-  virtual port::StatusOr<std::tuple<int, int, int>> GetVersion() {
+  // Gets the version of the backing library, as a VersionInfo object.
+  virtual port::StatusOr<VersionInfo> GetVersion() {
     return port::UnimplementedError(
         "DnnSupport::GetVersion not implemented on this platform.");
   }
diff --git a/tensorflow/stream_executor/platform/port.h b/tensorflow/stream_executor/platform/port.h
index 259cf380d6..57ad965ef1 100644
--- a/tensorflow/stream_executor/platform/port.h
+++ b/tensorflow/stream_executor/platform/port.h
@@ -38,12 +38,6 @@ using tensorflow::uint64;
 using std::string;
 #endif
 
-#if !defined(COMPILER_MSVC)
-#define ARRAYSIZE(a)              \
-    ((sizeof(a) / sizeof(*(a))) / \
-    static_cast<size_t>(!(sizeof(a) % sizeof(*(a)))))
-#endif
-
 using tensorflow::LinkerInitialized;
 using tensorflow::LINKER_INITIALIZED;