MNN/source/backend/vulkan/image/execution/VulkanImageConverter.cpp

149 lines
6.2 KiB
C++

//
// VulkanImageConverter.cpp
// MNN
//
// Created by MNN on 2019/01/31.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "VulkanImageConverter.hpp"
#include "core/Macro.h"
#include "core/TensorUtils.hpp"
#include "VulkanBackend.hpp"
namespace MNN {
VulkanImageConverter::VulkanImageConverter(const VulkanBackend* bn) {
mBackend = bn;
mSampler = bn->getCommonSampler();
mConst.reset(
new VulkanBuffer(bn->getMemoryPool(), false, 8 * sizeof(int), nullptr, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT));
}
void VulkanImageConverter::_setUpPipeline(MNN_DATA_FORMAT sourceFormat, MNN_DATA_FORMAT destFormat, TYPE type,
halide_type_t datatype) {
if (nullptr != mPipeline && sourceFormat == mCurrentSource && destFormat == mCurrentDest && mConvertImage == type) {
return;
}
mCurrentDest = destFormat;
mCurrentSource = sourceFormat;
mConvertImage = type;
std::vector<VkDescriptorType> types{VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER};
std::string name;
if (type == BUFFER_TO_IMAGE) {
types[0] = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE;
if (sourceFormat == MNN_DATA_FORMAT_NC4HW4) {
name = "glsl_nc4hw4toimage_comp";
} else {
name = "glsl_nchwToimage_comp";
}
} else {
if (destFormat == MNN_DATA_FORMAT_NC4HW4) {
name = "glsl_imageTonc4hw4_comp";
} else {
name = "glsl_imageTonchw_comp";
}
}
// FUNC_PRINT_ALL(name.c_str(), s);
mPipeline = mBackend->getPipeline(name, types);
MNN_ASSERT(nullptr != mPipeline);
}
void VulkanImageConverter::encodeBufferToTensor(VkBuffer srcBuffer, const Tensor* destTensor, const int bufferSize,
VkDeviceSize bufferOffset, MNN_DATA_FORMAT srcBufferFormat,
const VulkanCommandPool::Buffer* cmdBuffer) {
auto destFormat = TensorUtils::getDescribe(destTensor)->dimensionFormat;
auto sourceFormat = srcBufferFormat;
cmdBuffer->barrierSource(srcBuffer, 0, bufferSize);
auto tensor = destTensor;
_setUpPipeline(sourceFormat, destFormat, BUFFER_TO_IMAGE, tensor->buffer().type);
auto vkTensor = (VulkanTensor*)(destTensor->deviceId());
for (int i=0; i<vkTensor->imageSize(); ++i) {
vkTensor->image(i)->barrierWrite(cmdBuffer->get());
}
_encodeImageBufferConvert(tensor, srcBuffer, bufferSize, bufferOffset, cmdBuffer, VK_IMAGE_LAYOUT_GENERAL, srcBufferFormat);
}
void VulkanImageConverter::_encodeImageBufferConvert(const Tensor* tensor, VkBuffer destBuffer, const int bufferSize, VkDeviceSize bufferOffset, const VulkanCommandPool::Buffer* cmdBuffer, VkImageLayout layout, MNN_DATA_FORMAT bufferFormat) {
auto dims = (int*)mConst->map();// W, H, C, N
auto nhwc = VulkanTensor::tensorShapeFormat(tensor);
dims[0] = nhwc[2];
dims[1] = nhwc[1];
dims[2] = nhwc[3];
dims[3] = nhwc[0];
// Set stride for W, H, C, N
if (bufferFormat == MNN_DATA_FORMAT_NHWC) {
dims[4] = nhwc[3];
dims[5] = nhwc[3] * nhwc[2];
dims[6] = 1;
dims[7] = nhwc[3] * nhwc[2] * nhwc[1];
} else {
dims[4] = 1;
dims[5] = nhwc[2];
dims[6] = nhwc[2] * nhwc[1];
dims[7] = nhwc[3] * nhwc[2] * nhwc[1] ;
}
mConst->unmap();
auto vkTensor = reinterpret_cast<VulkanTensor*>(tensor->deviceId());
auto& mBlocks = vkTensor->blocks();
auto& limits = mBackend->proty().limits;
int wUnit = limits.maxImageDimension2D;
int hUnit = limits.maxImageDimension2D;
struct OffsetBuffer {
int offset[4]; // Offset w, h, c, n
int size[4];//w, h, c, w*h*c
};
mSet.resize(vkTensor->imageSize());
mOffset.resize(vkTensor->imageSize());
for (int y=0; y<mBlocks[1]; ++y) {
auto ySta = y * hUnit;
for (int x=0; x<mBlocks[0]; ++x) {
auto xSta = x * wUnit;
OffsetBuffer offset;
offset.offset[0] = xSta;
offset.offset[1] = ySta;
auto index = y*mBlocks[0] + x;
auto image = vkTensor->image(index);
offset.size[0] = image->width();
offset.size[1] = image->height();
offset.size[2] = 0;
offset.size[3] = image->width() * image->height();
mOffset[index].reset(new VulkanBuffer(mBackend->getMemoryPool(), false, sizeof(offset), &offset, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT));
mSet[index].reset(mPipeline->createSet());
mSet[index]->writeImage(image->view(), mSampler->get(), layout, 0);
mSet[index]->writeBuffer(destBuffer, 1, bufferSize, bufferOffset);
mSet[index]->writeBuffer(mConst->buffer(), 2, mConst->size());
mSet[index]->writeBuffer(mOffset[index]->buffer(), 3, mOffset[index]->size());
mPipeline->bind(cmdBuffer->get(), mSet[index]->get());
vkCmdDispatch(cmdBuffer->get(), UP_DIV(offset.size[3], 256), 1, 1);
}
}
}
void VulkanImageConverter::encodeTensorToBuffer(const Tensor* srcTensor, VkBuffer destBuffer, const int bufferSize,
VkDeviceSize bufferOffset, MNN_DATA_FORMAT destBufferFormat,
const VulkanCommandPool::Buffer* cmdBuffer) {
auto sourceFormat = TensorUtils::getDescribe(srcTensor)->dimensionFormat;
auto destFormat = destBufferFormat;
auto vkTensor = (VulkanTensor*)(srcTensor->deviceId());
auto tensor = srcTensor;
_setUpPipeline(sourceFormat, destFormat, IMAGE_TO_BUFFER, tensor->buffer().type);
for (int i=0; i<vkTensor->imageSize(); ++i) {
vkTensor->image(i)->barrierRead(cmdBuffer->get());
}
_encodeImageBufferConvert(tensor, destBuffer, bufferSize, bufferOffset, cmdBuffer,
VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, destBufferFormat);
}
MNN_DATA_FORMAT VulkanImageConverter::getTensorLinearFormat(const Tensor* tensor) {
auto format = TensorUtils::getDescribe(tensor)->dimensionFormat;
if (MNN_DATA_FORMAT_NC4HW4 == format) {
return MNN_DATA_FORMAT_NCHW;
}
return format;
}
} // namespace MNN