mirror of https://github.com/alibaba/MNN.git
95 lines
3.7 KiB
C++
95 lines
3.7 KiB
C++
//
|
|
// VulkanScale.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2019/01/31.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "VulkanScale.hpp"
|
|
#include "Macro.h"
|
|
#include "TensorUtils.hpp"
|
|
|
|
namespace MNN {
|
|
|
|
struct gpuScaleParam {
|
|
ivec4 imgSize;
|
|
int channel;
|
|
float eps;
|
|
};
|
|
|
|
VulkanScale::VulkanScale(const Op* op, Backend* bn) : VulkanBasicExecution(bn) {
|
|
const auto scale = op->main_as_Scale();
|
|
const int channels = scale->scaleData()->size();
|
|
|
|
std::vector<VkDescriptorType> types{VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
|
|
VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER};
|
|
|
|
auto extra = static_cast<VulkanBackend*>(bn);
|
|
|
|
mScalePipeline = extra->getPipeline("glsl_scale_comp", /*glsl_scale_comp, glsl_scale_comp_len,*/ types);
|
|
mScaleParam = std::make_shared<VulkanBuffer>(extra->getMemoryPool(), false, sizeof(gpuScaleParam), nullptr,
|
|
VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
|
|
mScaleBuffer = std::make_shared<VulkanBuffer>(extra->getMemoryPool(), false, sizeof(float) * channels,
|
|
scale->scaleData()->data(), VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
|
|
mBiasBuffer = std::make_shared<VulkanBuffer>(extra->getMemoryPool(), false, sizeof(float) * channels,
|
|
scale->biasData()->data(), VK_BUFFER_USAGE_STORAGE_BUFFER_BIT);
|
|
mSampler = extra->getCommonSampler();
|
|
}
|
|
|
|
VulkanScale::~VulkanScale() {
|
|
}
|
|
|
|
ErrorCode VulkanScale::onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) {
|
|
auto input = inputs[0];
|
|
auto output = outputs[0];
|
|
|
|
MNN_ASSERT(MNN_DATA_FORMAT_NC4HW4 == TensorUtils::getDescribe(input)->dimensionFormat);
|
|
|
|
auto scaleP = reinterpret_cast<gpuScaleParam*>(mScaleParam->map());
|
|
::memset(scaleP, 0, sizeof(gpuScaleParam));
|
|
|
|
const int channelDiv4 = UP_DIV(input->channel(), 4);
|
|
|
|
scaleP->channel = channelDiv4;
|
|
scaleP->imgSize[0] = input->width();
|
|
scaleP->imgSize[1] = input->height();
|
|
scaleP->imgSize[2] = channelDiv4 * input->batch();
|
|
scaleP->eps = 0.0;
|
|
mScaleParam->flush(true, 0, sizeof(gpuScaleParam));
|
|
mScaleParam->unmap();
|
|
|
|
mDescriptorSet.reset(mScalePipeline->createSet());
|
|
mDescriptorSet->writeImage(reinterpret_cast<VkImageView>(output->deviceId()), mSampler->get(),
|
|
VK_IMAGE_LAYOUT_GENERAL, 0);
|
|
mDescriptorSet->writeImage(reinterpret_cast<VkImageView>(input->deviceId()), mSampler->get(),
|
|
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
|
|
mDescriptorSet->writeBuffer(mScaleBuffer->buffer(), 2, mScaleBuffer->size());
|
|
mDescriptorSet->writeBuffer(mBiasBuffer->buffer(), 3, mBiasBuffer->size());
|
|
mDescriptorSet->writeBuffer(mScaleParam->buffer(), 4, mScaleParam->size());
|
|
mScalePipeline->bind(cmdBuffer->get(), mDescriptorSet->get());
|
|
|
|
cmdBuffer->barrierSource(reinterpret_cast<VkBuffer>(input->deviceId()), 0, input->size());
|
|
|
|
vkCmdDispatch(cmdBuffer->get(), UP_DIV(input->width(), 16), UP_DIV(input->height(), 16),
|
|
channelDiv4 * input->batch());
|
|
|
|
return NO_ERROR;
|
|
}
|
|
|
|
class VulkanScaleCreator : public VulkanBackend::Creator {
|
|
public:
|
|
virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const MNN::Op* op, Backend* bn) const override {
|
|
return new VulkanScale(op, bn);
|
|
}
|
|
};
|
|
|
|
static bool gResistor = []() {
|
|
VulkanBackend::addCreator(OpType_Scale, new VulkanScaleCreator);
|
|
return true;
|
|
}();
|
|
|
|
} // namespace MNN
|