1 //
2 // VulkanRelu.cpp
3 // MNN
4 //
5 // Created by MNN on 2019/01/31.
6 // Copyright © 2018, Alibaba Group Holding Limited
7 //
8
9 #include "VulkanRelu.hpp"
10 #include "core/Macro.h"
11 #include "core/TensorUtils.hpp"
12 namespace MNN {
13
14 struct GpuReluParam {
15 ivec4 imgSize;
16 vec4 slope;
17 };
18
19 //--------------------------relu--------------------------//
VulkanRelu(Backend * bn,const Op * op)20 VulkanRelu::VulkanRelu(Backend *bn, const Op* op) : VulkanBasicExecution(bn) {
21 auto vulkanBn = static_cast<VulkanBackend *>(bn);
22 mGpuReluParam.reset(new VulkanBuffer(vulkanBn->getMemoryPool(), false, sizeof(GpuReluParam), nullptr,
23 VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT));
24 if (op->type() == OpType_ReLU6) {
25 float minv = 0.0f;
26 float maxv = 6.0f;
27 if (nullptr != op->main_as_Relu6()) {
28 minv = op->main_as_Relu6()->minValue();
29 maxv = op->main_as_Relu6()->maxValue();
30 }
31 mSlope[0] = minv;
32 mSlope[1] = maxv;
33 mReluPipeline = vulkanBn->getPipeline("glsl_relu6_comp", {VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER});
34 } else {
35 if (op->type() == OpType_ReLU) {
36 mSlope[0] = op->main_as_Relu()->slope();
37 mSlope[1] = op->main_as_Relu()->slope();
38 mSlope[2] = op->main_as_Relu()->slope();
39 mSlope[3] = op->main_as_Relu()->slope();
40 } else {
41 // PRELU
42 auto slope = op->main_as_PRelu()->slope()->data()[0];
43 mSlope[0] = slope;
44 mSlope[1] = slope;
45 mSlope[2] = slope;
46 mSlope[3] = slope;
47 }
48
49 mReluPipeline = vulkanBn->getPipeline("glsl_relu_comp", {VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER});
50 }
51 }
52
~VulkanRelu()53 VulkanRelu::~VulkanRelu() {
54 }
55
onEncode(const std::vector<Tensor * > & inputs,const std::vector<Tensor * > & outputs,const VulkanCommandPool::Buffer * cmdBuffer)56 ErrorCode VulkanRelu::onEncode(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
57 const VulkanCommandPool::Buffer *cmdBuffer) {
58 auto input = inputs[0];
59 auto output = outputs[0];
60
61 auto vkBn = (VulkanBackend *)backend();
62
63 auto inputTensor = reinterpret_cast<VulkanTensor*>(input->deviceId());
64 auto outputTensor = reinterpret_cast<VulkanTensor*>(output->deviceId());
65 auto reluParam = reinterpret_cast<GpuReluParam *>(mGpuReluParam->map());
66 ::memset(reluParam, 0, sizeof(GpuReluParam));
67 reluParam->imgSize[0] = inputTensor->image()->width();
68 reluParam->imgSize[1] = inputTensor->image()->height();
69 reluParam->imgSize[2] = inputTensor->image()->depth();
70 reluParam->imgSize[3] = 0;
71 for (int i=0; i<4; ++i) {
72 reluParam->slope[i] = mSlope[i];
73 }
74 mGpuReluParam->unmap();
75 mDescriptorSet.reset(mReluPipeline->createSet());
76 mDescriptorSet->writeImage(outputTensor->image()->view(), vkBn->getCommonSampler()->get(),
77 VK_IMAGE_LAYOUT_GENERAL, 0);
78 mDescriptorSet->writeImage(inputTensor->image()->view(), vkBn->getCommonSampler()->get(),
79 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
80 mDescriptorSet->writeBuffer(mGpuReluParam->buffer(), 2, mGpuReluParam->size());
81 mReluPipeline->bind(cmdBuffer->get(), mDescriptorSet->get());
82 vkCmdDispatch(cmdBuffer->get(), UP_DIV(inputTensor->image()->width(), 16), UP_DIV(inputTensor->image()->height(), 16), 1);
83 return NO_ERROR;
84 }
85 //--------------------------Prelu--------------------------//
VulkanPrelu(Backend * bn,const Op * op)86 VulkanPrelu::VulkanPrelu(Backend *bn, const Op *op) : VulkanBasicExecution(bn) {
87 std::vector<VkDescriptorType> types{VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
88 VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER};
89 auto vulkanBn = static_cast<VulkanBackend *>(bn);
90 mPreluPipeline = vulkanBn->getPipeline("glsl_preluWithChannel_comp",
91 /*glsl_preluWithChannel_comp, glsl_preluWithChannel_comp_len,*/ types);
92 const auto prelu = op->main_as_PRelu();
93 mGpuPreluParam.reset(new VulkanBuffer(vulkanBn->getMemoryPool(), false, sizeof(GpuReluParam), nullptr,
94 VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT));
95 int count = ALIGN_UP4(prelu->slope()->size());
96
97 mSlope.reset(new VulkanImage(vulkanBn->getMemoryPool(), false, std::vector<int>{count / 4, 1}));
98 {
99 std::shared_ptr<VulkanBuffer> slopeBuffer(new VulkanBuffer(
100 vulkanBn->getMemoryPool(), false, sizeof(float) * count, nullptr, VK_BUFFER_USAGE_STORAGE_BUFFER_BIT));
101 auto slope = slopeBuffer->map();
102 ::memset(slope, 0, count * sizeof(float));
103 ::memcpy(slope, prelu->slope()->data(), prelu->slope()->size() * sizeof(float));
104 slopeBuffer->unmap();
105 vulkanBn->copyBufferToImage(slopeBuffer.get(), mSlope.get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
106 }
107 }
108
~VulkanPrelu()109 VulkanPrelu::~VulkanPrelu() {
110 }
111
onEncode(const std::vector<Tensor * > & inputs,const std::vector<Tensor * > & outputs,const VulkanCommandPool::Buffer * cmdBuffer)112 ErrorCode VulkanPrelu::onEncode(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
113 const VulkanCommandPool::Buffer *cmdBuffer) {
114 auto input = inputs[0];
115 auto output = outputs[0];
116
117 auto preluParam = reinterpret_cast<GpuReluParam *>(mGpuPreluParam->map());
118 ::memset(preluParam, 0, sizeof(GpuReluParam));
119 auto vkBn = static_cast<VulkanBackend *>(backend());
120
121 const int channelDiv4 = UP_DIV(input->channel(), 4);
122 preluParam->imgSize[0] = input->width();
123 preluParam->imgSize[1] = input->height();
124 preluParam->imgSize[2] = channelDiv4;
125 preluParam->imgSize[3] = 0;
126 mGpuPreluParam->flush(true, 0, sizeof(GpuReluParam));
127 mGpuPreluParam->unmap();
128
129 auto vkBackend = (VulkanBackend*)backend();
130 auto vkOutput = (VulkanTensor*)output->deviceId();
131 auto vkInput = (VulkanTensor*)input->deviceId();
132 cmdBuffer->barrierImageIfNeeded(vkOutput->image(), VK_IMAGE_LAYOUT_GENERAL);
133 cmdBuffer->barrierImageIfNeeded(vkInput->image(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
134 cmdBuffer->barrierImageIfNeeded(mSlope.get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
135
136 mDescriptorSet.reset(mPreluPipeline->createSet());
137 mDescriptorSet->writeImage(((VulkanTensor*)output->deviceId())->image()->view(), vkBn->getCommonSampler()->get(),
138 VK_IMAGE_LAYOUT_GENERAL, 0);
139 mDescriptorSet->writeImage(((VulkanTensor*)input->deviceId())->image()->view(), vkBn->getCommonSampler()->get(),
140 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
141 mDescriptorSet->writeImage((mSlope->view()), vkBn->getCommonSampler()->get(),
142 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 2);
143 mDescriptorSet->writeBuffer(mGpuPreluParam->buffer(), 3, mGpuPreluParam->size());
144
145 mPreluPipeline->bind(cmdBuffer->get(), mDescriptorSet->get());
146
147 vkCmdDispatch(cmdBuffer->get(), UP_DIV(input->width(), 16), UP_DIV(input->height(), 16), channelDiv4 * input->batch());
148 return NO_ERROR;
149 }
150
151 class VulkanReluCreator : public VulkanBackend::Creator {
152 public:
onCreate(const std::vector<Tensor * > & inputs,const std::vector<Tensor * > & outputs,const MNN::Op * op,Backend * bn) const153 virtual VulkanBasicExecution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor*>& outputs, const MNN::Op *op, Backend *bn) const override {
154 auto type = op->type();
155 if (OpType_ReLU6 == type) {
156 return new VulkanRelu(bn, op);
157 }
158 if (OpType_ReLU == type) {
159 return new VulkanRelu(bn, op);
160 } else if (1 == op->main_as_PRelu()->slopeCount()) {
161 return new VulkanRelu(bn, op);
162 } else {
163 return new VulkanPrelu(bn, op);
164 }
165 return nullptr;
166 }
167 };
168
__anon764c0b250102() 169 static bool gr = []() {
170 VulkanBackend::addCreator(OpType_ReLU, new VulkanReluCreator);
171 VulkanBackend::addCreator(OpType_PReLU, new VulkanReluCreator);
172 VulkanBackend::addCreator(OpType_ReLU6, new VulkanReluCreator);
173 return true;
174 }();
175
176 } // namespace MNN
177