MNN/source/backend/vulkan/execution/VulkanMatrixMultier.cpp

91 lines
3.7 KiB
C++

//
// VulkanMatrixMultier.cpp
// MNN
//
// Created by MNN on 2019/01/31.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "VulkanMatrixMultier.hpp"
#include "Macro.h"
namespace MNN {
struct constUniform {
ivec4 outputSize;
int multiLength;
};
VulkanMatrixMultier::~VulkanMatrixMultier() {
}
VulkanMatrixMultier::VulkanMatrixMultier(VulkanBackend* backend, const float* B, int w, int h, int c) {
mBackend = backend;
mWidth = w;
mHeight = h;
mDepth = c;
mConstBuffer = std::make_shared<VulkanBuffer>(backend->getMemoryPool(), false, sizeof(constUniform), nullptr,
VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
{
std::vector<VkDescriptorType> types{VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER};
if (backend->gpuType() == VulkanBackend::ADRENO || backend->gpuType() == VulkanBackend::MALI) {
mPipeline = mBackend->getPipeline("glsl_gemm16x16Half_comp",
/*glsl_gemm16x16Half_comp, glsl_gemm16x16Half_comp_len,*/ types);
} else {
mPipeline =
mBackend->getPipeline("glsl_gemm16x16_comp", /*glsl_gemm16x16_comp, glsl_gemm16x16_comp_len,*/ types);
}
}
mDescriptorSet.reset(mPipeline->createSet());
mKernel = std::make_shared<VulkanImage>(backend->getMemoryPool(), false,
std::vector<int>{ALIGN_UP4(w), UP_DIV(h, 4) * c});
mSampler = mBackend->getCommonSampler();
// Compute mKernel
auto tempBuffer = std::make_shared<VulkanBuffer>(backend->getMemoryPool(), false,
ALIGN_UP4(w) * ALIGN_UP4(h) * c * sizeof(float));
{
auto dest = tempBuffer->map();
::memcpy(dest, B, ALIGN_UP4(w) * ALIGN_UP4(h) * c * sizeof(float));
tempBuffer->unmap();
}
mBackend->copyBufferToImage(tempBuffer.get(), mKernel.get());
}
void VulkanMatrixMultier::prepare(int srcHeight) {
int sw = ALIGN_UP4(mWidth);
int sh = UP_DIV(srcHeight, 4);
mSource = std::make_shared<VulkanImage>(mBackend->getDynamicMemoryPool(), false, std::vector<int>{sw, sh * mDepth});
int ow = sh;
int oh = ALIGN_UP4(mHeight);
mDest = std::make_shared<VulkanImage>(mBackend->getDynamicMemoryPool(), false, std::vector<int>{ow * mDepth, oh});
MNN_ASSERT(nullptr != mSource && nullptr != mDest);
mSource->release();
mDest->release();
mDescriptorSet->writeImage(mDest->view(), mSampler->get(), VK_IMAGE_LAYOUT_GENERAL, 0);
mDescriptorSet->writeImage(mSource->view(), mSampler->get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
mDescriptorSet->writeImage(mKernel->view(), mSampler->get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 2);
{
auto uniform = (constUniform*)mConstBuffer->map();
uniform->outputSize[0] = ow;
uniform->outputSize[1] = oh / 4;
uniform->multiLength = sw / 4;
mConstBuffer->unmap();
}
mDescriptorSet->writeBuffer(mConstBuffer->buffer(), 3, mConstBuffer->size());
mOutputWidth = ow;
mOutputHeight = oh;
}
void VulkanMatrixMultier::compute(const VulkanCommandPool::Buffer* commandBuffer) const {
mPipeline->bind(commandBuffer->get(), mDescriptorSet->get());
commandBuffer->barrierImage(mSource->get(), VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
vkCmdDispatch(commandBuffer->get(), UP_DIV(mOutputWidth, 8), UP_DIV(mOutputHeight / 4, 8), mDepth);
}
} // namespace MNN