xref: /aosp_15_r20/external/tensorflow/tensorflow/lite/delegates/gpu/cl/buffer.h (revision b6fb3261f9314811a0f4371741dbb8839866f948)
1 /* Copyright 2019 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_LITE_DELEGATES_GPU_CL_BUFFER_H_
17 #define TENSORFLOW_LITE_DELEGATES_GPU_CL_BUFFER_H_
18 
19 #include "absl/strings/str_cat.h"
20 #include "absl/types/span.h"
21 #include "tensorflow/lite/delegates/gpu/cl/cl_command_queue.h"
22 #include "tensorflow/lite/delegates/gpu/cl/cl_context.h"
23 #include "tensorflow/lite/delegates/gpu/cl/gpu_object.h"
24 #include "tensorflow/lite/delegates/gpu/cl/opencl_wrapper.h"
25 #include "tensorflow/lite/delegates/gpu/cl/util.h"
26 #include "tensorflow/lite/delegates/gpu/common/status.h"
27 #include "tensorflow/lite/delegates/gpu/common/task/buffer_desc.h"
28 
29 namespace tflite {
30 namespace gpu {
31 namespace cl {
32 
33 // Buffer represent linear GPU data storage with arbitrary data format.
34 // Buffer is moveable but not copyable.
35 class Buffer : public GPUObject {
36  public:
Buffer()37   Buffer() {}  // just for using Buffer as a class members
38   Buffer(cl_mem buffer, size_t size_in_bytes, bool is_sub_buffer = false);
39 
40   // Move only
41   Buffer(Buffer&& buffer);
42   Buffer& operator=(Buffer&& buffer);
43   Buffer(const Buffer&) = delete;
44   Buffer& operator=(const Buffer&) = delete;
45 
~Buffer()46   ~Buffer() override { Release(); }
47 
48   // for profiling and memory statistics
GetMemorySizeInBytes()49   uint64_t GetMemorySizeInBytes() const { return size_; }
50 
GetMemoryPtr()51   cl_mem GetMemoryPtr() const { return buffer_; }
52 
IsSubBuffer()53   bool IsSubBuffer() const { return is_sub_buffer_; }
54 
55   // Writes data to a buffer. Data should point to a region that
56   // has exact size in bytes as size_in_bytes(constructor parameter).
57   template <typename T>
58   absl::Status WriteData(CLCommandQueue* queue, const absl::Span<T> data);
59 
60   // Reads data from Buffer into CPU memory.
61   template <typename T>
62   absl::Status ReadData(CLCommandQueue* queue, std::vector<T>* result) const;
63 
64   absl::Status GetGPUResources(const GPUObjectDescriptor* obj_ptr,
65                                GPUResourcesWithValue* resources) const override;
66 
67   absl::Status CreateFromBufferDescriptor(const BufferDescriptor& desc,
68                                           CLContext* context);
69 
70  private:
71   void Release();
72 
73   cl_mem buffer_ = nullptr;
74   size_t size_ = 0;
75   bool is_sub_buffer_ = false;
76 };
77 
78 absl::Status CreateReadOnlyBuffer(size_t size_in_bytes, CLContext* context,
79                                   Buffer* result);
80 
81 absl::Status CreateReadOnlyBuffer(size_t size_in_bytes, const void* data,
82                                   CLContext* context, Buffer* result);
83 
84 absl::Status CreateReadWriteBuffer(size_t size_in_bytes, CLContext* context,
85                                    Buffer* result);
86 
87 absl::Status CreateReadWriteSubBuffer(const Buffer& parent,
88                                       size_t origin_in_bytes,
89                                       size_t size_in_bytes, CLContext* context,
90                                       Buffer* result);
91 
92 template <typename T>
WriteData(CLCommandQueue * queue,const absl::Span<T> data)93 absl::Status Buffer::WriteData(CLCommandQueue* queue,
94                                const absl::Span<T> data) {
95   if (size_ != sizeof(T) * data.size()) {
96     return absl::InvalidArgumentError(
97         "absl::Span<T> data size is different from buffer allocated size.");
98   }
99   RETURN_IF_ERROR(queue->EnqueueWriteBuffer(buffer_, size_, data.data()));
100   return absl::OkStatus();
101 }
102 
103 template <typename T>
ReadData(CLCommandQueue * queue,std::vector<T> * result)104 absl::Status Buffer::ReadData(CLCommandQueue* queue,
105                               std::vector<T>* result) const {
106   if (size_ % sizeof(T) != 0) {
107     return absl::UnknownError("Wrong element size(typename T is not correct?");
108   }
109 
110   const int elements_count = size_ / sizeof(T);
111   result->resize(elements_count);
112 
113   return queue->EnqueueReadBuffer(buffer_, size_, result->data());
114 }
115 
116 }  // namespace cl
117 }  // namespace gpu
118 }  // namespace tflite
119 
120 #endif  // TENSORFLOW_LITE_DELEGATES_GPU_CL_BUFFER_H_
121