1 //
2 //  PoolBufExecution.cpp
3 //  MNN
4 //
5 //  Created by MNN on 2019/02/28.
6 //  Copyright © 2018, Alibaba Group Holding Limited
7 //
8 
9 #ifndef MNN_OPENCL_BUFFER_CLOSED
10 
11 #include "backend/opencl/execution/buffer/PoolBufExecution.hpp"
12 #include "core/Macro.h"
13 #include "core/TensorUtils.hpp"
14 #include "backend/opencl/core/OpenCLBackend.hpp"
15 
16 namespace MNN {
17 namespace OpenCL {
18 
PoolBufExecution(const std::vector<Tensor * > & inputs,const MNN::Op * op,Backend * backend)19 PoolBufExecution::PoolBufExecution(const std::vector<Tensor *> &inputs, const MNN::Op *op, Backend *backend)
20     : Execution(backend) {
21     mOpenCLBackend = static_cast<OpenCLBackend *>(backend);
22     mPoolParams    = op->main_as_Pool();
23     mPoolType      = mPoolParams->type();
24 
25     mStrides[0] = mPoolParams->strideY();
26     mStrides[1] = mPoolParams->strideX();
27     mKernels[0] = mPoolParams->kernelY();
28     mKernels[1] = mPoolParams->kernelX();
29 
30     mPaddings[0] = mPoolParams->padY() * 2;
31     mPaddings[1] = mPoolParams->padX() * 2;
32     mPadType     = mPoolParams->padType();
33     if (mPadType == PoolPadType_VALID) {
34         mPaddings[0] = 0;
35         mPaddings[1] = 0;
36     }
37 }
38 
onResize(const std::vector<Tensor * > & inputs,const std::vector<Tensor * > & outputs)39 ErrorCode PoolBufExecution::onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
40 #ifdef LOG_VERBOSE
41     MNN_PRINT("start PoolBufExecution onResize !\n");
42 #endif
43     auto input  = inputs[0];
44     auto output = outputs[0];
45 
46     if (mPoolParams->isGlobal()) {
47         std::vector<int> inputShape = tensorShapeFormat(inputs[0]);
48         mKernels                    = {inputShape.at(1), inputShape.at(2)};
49         mStrides                    = {inputShape.at(1), inputShape.at(2)};
50         mPaddings                   = {0, 0};
51     }
52 
53     if (mPadType == PoolPadType_SAME) {
54         int padNeededHeight = std::max(0, (output->height() - 1) * mStrides[0] + mKernels[0] - input->height());
55         int padNeededWidth  = std::max(0, (output->width() - 1) * mStrides[1] + mKernels[1] - input->width());
56 
57         mPaddings[0] = padNeededHeight;
58         mPaddings[1] = padNeededWidth;
59     }
60 
61     MNN_ASSERT(mDilations[0] == 1 && mDilations[1] == 1);
62 
63     std::vector<int> inputShape  = tensorShapeFormat(input);
64     std::vector<int> outputShape = tensorShapeFormat(output);
65 
66     const int batch        = outputShape.at(0);
67     const int outputHeight = outputShape.at(1);
68     const int outputWidth  = outputShape.at(2);
69     const int channels     = outputShape.at(3);
70 
71     const int inputHeight = inputShape.at(1);
72     const int inputWidth  = inputShape.at(2);
73     int channelBlocks = (channels + 3) / 4;
74 
75     std::set<std::string> buildOptions;
76     std::string kernelName = "pooling";
77     auto runtime           = mOpenCLBackend->getOpenCLRuntime();
78 
79     if (mPoolType == PoolType_AVEPOOL) {
80         buildOptions.emplace("-DPOOL_AVG");
81     }
82 
83     mKernel           = runtime->buildKernel("pooling_buf", kernelName, buildOptions);
84     mMaxWorkGroupSize = static_cast<uint32_t>(runtime->getMaxWorkGroupSize(mKernel));
85 
86     mGlobalWorkSize = {
87         static_cast<uint32_t>(outputWidth),
88         static_cast<uint32_t>(batch * outputHeight),
89         static_cast<uint32_t>(channelBlocks),
90     };
91 
92     int inputImageShape[2] = {inputHeight, inputWidth};
93     int outputImageShape[2] = {outputHeight, outputWidth};
94     int paddingShape[2]    = {mPaddings[0] / 2, mPaddings[1] / 2};
95     int strideShape[2]     = {mStrides[0], mStrides[1]};
96     int kernelShape[2]     = {mKernels[0], mKernels[1]};
97 
98     uint32_t idx   = 0;
99     mKernel.setArg(idx++, mGlobalWorkSize[0]);
100     mKernel.setArg(idx++, mGlobalWorkSize[1]);
101     mKernel.setArg(idx++, mGlobalWorkSize[2]);
102     mKernel.setArg(idx++, openCLBuffer(input));
103     mKernel.setArg(idx++, sizeof(inputImageShape), inputImageShape);
104     mKernel.setArg(idx++, sizeof(outputImageShape), outputImageShape);
105     mKernel.setArg(idx++, sizeof(paddingShape), paddingShape);
106     mKernel.setArg(idx++, sizeof(strideShape), strideShape);
107     mKernel.setArg(idx++, sizeof(kernelShape), kernelShape);
108     mKernel.setArg(idx++, openCLBuffer(output));
109     mKernel.setArg(idx++, channelBlocks);
110 
111     std::string kernelNameTune = "pooling_buf";
112     mLocalWorkSize =
113     localWS3DDefault(mGlobalWorkSize, mMaxWorkGroupSize, mOpenCLBackend->getOpenCLRuntime(), kernelNameTune, mKernel).first;
114 
115 #ifdef LOG_VERBOSE
116     MNN_PRINT("end PoolBufExecution onResize !\n");
117 #endif
118     return NO_ERROR;
119 }
120 
onExecute(const std::vector<Tensor * > & inputs,const std::vector<Tensor * > & outputs)121 ErrorCode PoolBufExecution::onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
122 #ifdef LOG_VERBOSE
123     MNN_PRINT("start PoolBufExecution onExecute !\n");
124 #endif
125 
126 #ifdef ENABLE_OPENCL_TIME_PROFILER
127     cl::Event event;
128     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize,
129                        mOpenCLBackend->getOpenCLRuntime(), &event);
130 
131     int costTime = (int)mOpenCLBackend->getOpenCLRuntime()->getCostTime(&event);
132     MNN_PRINT("kernel cost:%d    us Pooling\n",costTime);
133 #else
134     run3DKernelDefault(mKernel, mGlobalWorkSize, mLocalWorkSize,
135                        mOpenCLBackend->getOpenCLRuntime());
136 #endif
137 
138 #ifdef LOG_VERBOSE
139     MNN_PRINT("end PoolBufExecution onExecute !\n");
140 #endif
141     return NO_ERROR;
142 }
143 
144 OpenCLCreatorRegister<TypedCreator<PoolBufExecution>> __PoolBuf_op(OpType_Pooling, BUFFER);
145 } // namespace OpenCL
146 } // namespace MNN
147 #endif /* MNN_OPENCL_BUFFER_CLOSED */
148