Merge pull request #19 from tensorflow/master

downstream merge

Author: tg-at-google

.bazelrc

@@ -28,6 +28,7 @@
 #
 # Other build options:
 #     short_logs:       Only log errors during build, skip warnings.
+#     verbose_logs:     Show all compiler warnings during build.
 #     monolithic:       Build all TF C++ code into a single shared object.
 #     dynamic_kernels:  Try to link all kernels dynamically (experimental).
 #     libc++:           Link against libc++ instead of stdlibc++
@@ -172,6 +173,11 @@ build:using_cuda --define=using_cuda=true
 build:using_cuda --action_env TF_NEED_CUDA=1
 build:using_cuda --crosstool_top=@local_config_cuda//crosstool:toolchain
 
+# Enable the mlir generated GPU kernels only for cuda builds.
+build --define=tensorflow_enable_mlir_generated_gpu_kernels=0
+# This is a more specific option, so it takes precedence over the line above for cuda builds.
+build:using_cuda --define=tensorflow_enable_mlir_generated_gpu_kernels=1
+
 # This config refers to building CUDA op kernels with nvcc.
 build:cuda --config=using_cuda
 build:cuda --define=using_cuda_nvcc=true
@@ -331,6 +337,8 @@ build:windows --distinct_host_configuration=false
 
 # Suppress all warning messages.
 build:short_logs --output_filter=DONT_MATCH_ANYTHING
+build:verbose_logs --output_filter=
+build --config=short_logs
 
 # Instruction set optimizations
 # TODO(gunan): Create a feature in toolchains for avx/avx2 to
@@ -547,6 +555,7 @@ try-import %workspace%/.bazelrc.user
 # Here are bazelrc configs for release builds
 build:release_common --config=opt
 build:release_common --config=v2
+build:release_common --distinct_host_configuration=false
 build:release_common --action_env TF_CONFIGURE_IOS="0"
 
 build:release_cpu_linux --config=release_common
@@ -564,9 +573,10 @@ build:release_gpu_common --config=tensorrt
 build:release_gpu_common --action_env CUDA_TOOLKIT_PATH="/usr/local/cuda-10.1"
 build:release_gpu_common --action_env=TF_CUDA_VERSION="10"
 build:release_gpu_common --action_env=TF_CUDNN_VERSION="7"
+build:release_gpu_common --action_env=TF_NEED_TENSORRT="1"
 build:release_gpu_common --action_env=TF_CUDA_COMPUTE_CAPABILITIES="sm_35,sm_37,sm_52,sm_60,sm_61,compute_70"
 build:release_gpu_common --action_env=TENSORRT_INSTALL_PATH="/usr/local/tensorrt"
-build:release_gpu_common --action_env=LD_LIBRARY_PATH="/usr/local/tensorrt/lib"
+build:release_gpu_common --action_env=LD_LIBRARY_PATH="/usr/local/cuda:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/tensorrt/lib"
 build:release_gpu_common --action_env=GCC_HOST_COMPILER_PATH="/usr/bin/gcc-5"
 
 

RELEASE.md

@@ -125,6 +125,14 @@ TF_CAPI_EXPORT extern void TF_DeleteBuffer(TF_Buffer*);
 
 TF_CAPI_EXPORT extern TF_Buffer TF_GetBuffer(TF_Buffer* buffer);
 
+// --------------------------------------------------------------------------
+// Used to return strings across the C API. The caller does not take ownership
+// of the underlying data pointer and is not responsible for freeing it.
+typedef struct TF_StringView {
+  const char* data;
+  size_t len;
+} TF_StringView;
+
 // --------------------------------------------------------------------------
 // TF_SessionOptions holds options that can be passed during session creation.
 typedef struct TF_SessionOptions TF_SessionOptions;

tensorflow/c/c_api.h

@@ -262,6 +262,7 @@ cc_library(
     ],
     deps = [
         "//tensorflow/core:protos_all_cc",
+        "//tensorflow/core/platform:refcount",
     ],
 )
 
@@ -549,6 +550,7 @@ tf_cuda_cc_test(
     args = ["--heap_check=local"],
     extra_copts = tfe_xla_copts(),
     tags = [
+        "no_oss",  # b/162361408
         "no_windows",
         "noasan",  # leaks gRPC server instances
     ],

tensorflow/c/eager/BUILD

@@ -18,11 +18,12 @@ limitations under the License.
 #include <memory>
 
 #include "tensorflow/core/framework/types.pb.h"
+#include "tensorflow/core/platform/refcount.h"
 namespace tensorflow {
 
 // Abstract interface to a Tensor handle in either tracing or immediate
 // execution mode.
-class AbstractTensorHandle {
+class AbstractTensorHandle : public core::RefCounted {
  protected:
   enum AbstractTensorHandleKind { kGraph, kMlir, kEager, kTfrt };
   explicit AbstractTensorHandle(AbstractTensorHandleKind kind) : kind_(kind) {}
@@ -34,14 +35,6 @@ class AbstractTensorHandle {
 
   AbstractTensorHandleKind getKind() const { return kind_; }
 
-  // Release any underlying resources, including the interface object.
-  //
-  // WARNING: The destructor of this class is marked as protected to disallow
-  // clients from directly destroying this object since it may manage it's own
-  // lifetime through ref counting. Thus this must be allocated on the heap and
-  // clients MUST call Release() in order to destroy an instance of this class.
-  virtual void Release() = 0;
-
  private:
   const AbstractTensorHandleKind kind_;
 };
@@ -50,7 +43,7 @@ namespace internal {
 struct AbstractTensorHandleDeleter {
   void operator()(AbstractTensorHandle* p) const {
     if (p != nullptr) {
-      p->Release();
+      p->Unref();
     }
   }
 };

tensorflow/c/eager/abstract_tensor_handle.h

@@ -1072,11 +1072,13 @@ TFE_TensorHandle* TFE_NewTensorHandleFromDeviceMemory(
   status->status = context->FindDeviceFromName(device_name, &device);
   tensorflow::CustomDevice* custom_device = nullptr;
   if (!status->status.ok()) {
-    status->status =
-        context->FindCustomDeviceFromName(device_name, &custom_device);
-    if (!status->status.ok()) {
+    if (!context->FindCustomDeviceFromName(device_name, &custom_device)) {
       deallocator(data, len, deallocator_arg);
+      status->status =
+          tensorflow::errors::InvalidArgument(device_name, " unknown device.");
       return nullptr;
+    } else {
+      status->status = tensorflow::Status::OK();
     }
   }
   std::vector<tensorflow::int64> dimvec(num_dims);

tensorflow/c/eager/c_api.cc

@@ -26,14 +26,13 @@ limitations under the License.
 #include "tensorflow/compiler/jit/xla_device.h"
 #endif  // TENSORFLOW_EAGER_USE_XLA
 
-using tensorflow::int64;
 using tensorflow::string;
 
 namespace {
 
-std::vector<int64> TensorShapeAsVector(const tensorflow::TensorHandle& handle,
-                                       tensorflow::Status* status) {
-  std::vector<int64> shape;
+std::vector<tensorflow::int64> TensorShapeAsVector(
+    const tensorflow::TensorHandle& handle, tensorflow::Status* status) {
+  std::vector<tensorflow::int64> shape;
   int rank = -1;
   *status = handle.NumDims(&rank);
   if (!status->ok()) {
@@ -79,7 +78,7 @@ TF_CAPI_EXPORT extern TFE_TensorDebugInfo* TFE_TensorHandleTensorDebugInfo(
       return nullptr;
     }
     if (VLOG_IS_ON(3)) {
-      std::vector<int64> shape_to_log =
+      std::vector<tensorflow::int64> shape_to_log =
           TensorShapeAsVector(*handle, &status->status);
       if (!status->status.ok()) {
         // Ignore the status here as we are simply logging.
@@ -128,14 +127,14 @@ TF_CAPI_EXPORT extern TFE_TensorDebugInfo* TFE_TensorHandleTensorDebugInfo(
     }
 
     int rank = padded_shape.dimensions_size();
-    std::vector<int64> dev_dims;
+    std::vector<tensorflow::int64> dev_dims;
     dev_dims.reserve(rank);
     if (rank == 1) {
       // Rank 1 tensors might not have padded_shape.layout.minor_to_major set,
       dev_dims.push_back(padded_shape.dimensions(0));
     } else {
       for (int i = rank - 1; i >= 0; --i) {
-        int64 dim_index = padded_shape.layout().minor_to_major(i);
+        tensorflow::int64 dim_index = padded_shape.layout().minor_to_major(i);
         dev_dims.push_back(padded_shape.dimensions(dim_index));
       }
     }
@@ -146,7 +145,8 @@ TF_CAPI_EXPORT extern TFE_TensorDebugInfo* TFE_TensorHandleTensorDebugInfo(
 
   // If the tensor is not an XLA tensor, the device shape is
   // the same as regular tensor shape.
-  std::vector<int64> dev_dims = TensorShapeAsVector(*handle, &status->status);
+  std::vector<tensorflow::int64> dev_dims =
+      TensorShapeAsVector(*handle, &status->status);
   if (!status->status.ok()) {
     return nullptr;
   }

tensorflow/c/eager/c_api_debug.cc

@@ -147,7 +147,7 @@ TF_AbstractOp* TF_NewAbstractOp(TF_ExecutionContext* c) {
 
 void TF_DeleteAbstractOp(TF_AbstractOp* op) { unwrap(op)->Release(); }
 
-void TF_DeleteAbstractTensor(TF_AbstractTensor* t) { unwrap(t)->Release(); }
+void TF_DeleteAbstractTensor(TF_AbstractTensor* t) { unwrap(t)->Unref(); }
 
 TF_OutputList* TF_NewOutputList() { return wrap(new OutputList); }
 void TF_DeleteOutputList(TF_OutputList* o) { delete unwrap(o); }

tensorflow/c/eager/c_api_unified_experimental.cc

@@ -49,7 +49,6 @@ class GraphTensor : public TracingTensorHandle {
  public:
   explicit GraphTensor(TF_Output output)
       : TracingTensorHandle(kGraph), output_(output) {}
-  void Release() override { delete this; }
 
   tensorflow::DataType DataType() const override {
     return static_cast<tensorflow::DataType>(TF_OperationOutputType(output_));

tensorflow/c/eager/c_api_unified_experimental_graph.cc

@@ -51,25 +51,14 @@ int64 ToId(AbstractTensorHandle* t) {
 
 TapeTensor::TapeTensor(AbstractTensorHandle* handle, AbstractContext* ctx)
     : handle_(handle), ctx_(ctx) {
-  // TODO(b/160888114): Make AbstractTensorHandle RefCounted. Right now we rely
-  // on the client to keep this tensor live for the duration of the gradient
-  // computation.
-  // handle_->Ref();
+  handle_->Ref();
 }
 TapeTensor::TapeTensor(const TapeTensor& other) {
   handle_ = other.handle_;
-  // TODO(b/160888114): Make AbstractTensorHandle RefCounted. Right now we rely
-  // on the client to keep this tensor live for the duration of the gradient
-  // computation.
-  // handle_->Ref();
+  handle_->Ref();
   ctx_ = other.ctx_;
 }
-TapeTensor::~TapeTensor() {
-  // TODO(b/160888114): Make AbstractTensorHandle RefCounted. Right now we rely
-  // on the client to keep this tensor live for the duration of the gradient
-  // computation.
-  // handle_->Unref();
-}
+TapeTensor::~TapeTensor() { handle_->Unref(); }
 
 tensorflow::int64 TapeTensor::GetID() const { return ToId(handle_); }
 
@@ -192,7 +181,7 @@ TapeTensor TapeVSpace::TapeTensorFromGradient(AbstractTensorHandle* g) const {
 void TapeVSpace::MarkAsResult(AbstractTensorHandle* gradient) const {}
 
 void TapeVSpace::DeleteGradient(AbstractTensorHandle* gradient) const {
-  gradient->Release();
+  gradient->Unref();
 }
 
 // Helper functions which delegate to `AbstractOperation`, update