1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 16 #ifndef TENSORFLOW_COMPILER_XLA_STREAM_EXECUTOR_SCRATCH_ALLOCATOR_H_ 17 #define TENSORFLOW_COMPILER_XLA_STREAM_EXECUTOR_SCRATCH_ALLOCATOR_H_ 18 19 #include <memory> 20 #include <utility> 21 22 #include "absl/container/inlined_vector.h" 23 #include "tensorflow/compiler/xla/stream_executor/device_memory.h" 24 #include "tensorflow/compiler/xla/stream_executor/device_memory_allocator.h" 25 #include "tensorflow/compiler/xla/stream_executor/lib/statusor.h" 26 #include "tensorflow/compiler/xla/stream_executor/platform/port.h" 27 #include "tensorflow/compiler/xla/stream_executor/temporary_device_memory.h" 28 29 namespace stream_executor { 30 31 class Stream; 32 33 // Interface for "scratch" allocator for device memory, which deallocates all 34 // buffers it has allocated at destruction. Returned memory pointers are not 35 // owning. 36 // 37 // Used by stream operations (e.g. Stream::ThenConvolveWithScratch) to 38 // optionally request scratch space to speed up the operation. 39 class ScratchAllocator { 40 public: ~ScratchAllocator()41 virtual ~ScratchAllocator() {} 42 43 // Returns a limit of memory this scratch allocator wants to produce, in 44 // bytes. This information may be used to help select an algorithm. 45 // 46 // Returns values < 0 to indicate that there is no recommended limit. 47 virtual int64_t GetMemoryLimitInBytes() = 0; 48 49 // Returns an allocation on byte_size bytes for use in an operation on stream. 50 // 51 // This is a temporary allocation, and the caller is responsible for 52 // deallocating at some known-safe point. See the class comment above. 53 virtual port::StatusOr<DeviceMemory<uint8>> AllocateBytes( 54 int64_t byte_size) = 0; 55 }; 56 57 // Allocates a single temporary memory allocation -- this memory is deallocated 58 // at the next stream synchronization point after this object has gone out of 59 // scope. This satisfies the lifetime and deallocation properties given in the 60 // class comment above. 61 // 62 // Thread-compatible, but not thread-safe (use in scenarios where only one 63 // thread will request the scratch allocation). 64 class OneTimeScratchAllocator : public ScratchAllocator { 65 public: OneTimeScratchAllocator(Stream * stream)66 explicit OneTimeScratchAllocator(Stream* stream) : stream_(stream) {} 67 GetMemoryLimitInBytes()68 int64_t GetMemoryLimitInBytes() override { return -1; } 69 70 port::StatusOr<DeviceMemory<uint8>> AllocateBytes(int64_t byte_size) override; 71 72 private: 73 std::unique_ptr<TemporaryDeviceMemory<uint8>> temporary_; 74 Stream* stream_; 75 76 SE_DISALLOW_COPY_AND_ASSIGN(OneTimeScratchAllocator); 77 }; 78 79 // Can allocate several times -- this memory is deallocated when the scratch 80 // allocator is destroyed. 81 // 82 // Thread-compatible, but not thread-safe (use in scenarios where only one 83 // thread will request the scratch allocation). 84 template <size_t N = 1> 85 class OwningScratchAllocator : public ScratchAllocator { 86 public: OwningScratchAllocator(int device_ordinal,DeviceMemoryAllocator * allocator)87 OwningScratchAllocator(int device_ordinal, DeviceMemoryAllocator* allocator) 88 : device_ordinal_(device_ordinal), allocator_(allocator) {} 89 GetMemoryLimitInBytes()90 int64_t GetMemoryLimitInBytes() override { return -1; } 91 AllocateBytes(int64_t byte_size)92 port::StatusOr<DeviceMemory<uint8>> AllocateBytes( 93 int64_t byte_size) override { 94 TF_ASSIGN_OR_RETURN(OwningDeviceMemory buffer, 95 allocator_->Allocate(device_ordinal_, byte_size, 96 /*retry_on_failure=*/false)); 97 buffers_.push_back(std::move(buffer)); 98 return *buffers_.back(); 99 } 100 101 private: 102 int device_ordinal_; 103 DeviceMemoryAllocator* allocator_; 104 absl::InlinedVector<OwningDeviceMemory, N> buffers_; 105 106 SE_DISALLOW_COPY_AND_ASSIGN(OwningScratchAllocator); 107 }; 108 109 } // namespace stream_executor 110 111 #endif // TENSORFLOW_COMPILER_XLA_STREAM_EXECUTOR_SCRATCH_ALLOCATOR_H_ 112