Minimal support for running OpsTest on GPU, using CUDA unified memory.

PiperOrigin-RevId: 169897567

2 files changed, 78 insertions, 0 deletions

diff --git a/tensorflow/core/common_runtime/gpu/gpu_managed_allocator.cc b/tensorflow/core/common_runtime/gpu/gpu_managed_allocator.cc
new file mode 100644
index 0000000000..41f7249d89
--- /dev/null
+++ b/tensorflow/core/common_runtime/gpu/gpu_managed_allocator.cc
@@ -0,0 +1,42 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/core/common_runtime/gpu/gpu_managed_allocator.h"
+
+#ifdef GOOGLE_CUDA
+#include "cuda/include/cuda.h"
+#endif  // GOOGLE_CUDA
+
+namespace tensorflow {
+
+void* GpuManagedAllocator::AllocateRaw(size_t alignment, size_t num_bytes) {
+#ifdef GOOGLE_CUDA
+  CUdeviceptr ptr = 0;
+  CHECK_EQ(cuMemAllocManaged(&ptr, num_bytes, CU_MEM_ATTACH_GLOBAL),
+           CUDA_SUCCESS);
+  CHECK(!(ptr & (alignment - 1)));
+  return reinterpret_cast<void*>(ptr);
+#else
+  return nullptr;
+#endif
+}
+
+void GpuManagedAllocator::DeallocateRaw(void* ptr) {
+#ifdef GOOGLE_CUDA
+  CHECK_EQ(cuMemFree(reinterpret_cast<CUdeviceptr>(ptr)), CUDA_SUCCESS);
+#endif
+}
+
+}  // namespace tensorflow
diff --git a/tensorflow/core/common_runtime/gpu/gpu_managed_allocator.h b/tensorflow/core/common_runtime/gpu/gpu_managed_allocator.h
new file mode 100644
index 0000000000..006b2ca448
--- /dev/null
+++ b/tensorflow/core/common_runtime/gpu/gpu_managed_allocator.h
@@ -0,0 +1,36 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef THIRD_PARTY_TENSORFLOW_CORE_COMMON_RUNTIME_GPU_GPU_MANAGED_ALLOCATOR_H_
+#define THIRD_PARTY_TENSORFLOW_CORE_COMMON_RUNTIME_GPU_GPU_MANAGED_ALLOCATOR_H_
+
+#include "tensorflow/core/framework/allocator.h"
+
+namespace tensorflow {
+
+// An allocator for CUDA unified memory. Memory allocated with this allocator
+// can be accessed from both host and device. CUDA transparently migrates dirty
+// pages, which can be slow. Therefore, this allocator is intended for
+// convenience in functional tests only.
+class GpuManagedAllocator : public Allocator {
+ public:
+  string Name() override { return "GpuManagedAllocator"; }
+  void* AllocateRaw(size_t alignment, size_t num_bytes) override;
+  void DeallocateRaw(void* ptr) override;
+};
+
+}  // namespace tensorflow
+
+#endif  // THIRD_PARTY_TENSORFLOW_CORE_COMMON_RUNTIME_GPU_GPU_MANAGED_ALLOCATOR_H_