wehub-resource-sync
103 lines
3.5 KiB
C++
103 lines
3.5 KiB
C++
//
|
|
// VulkanRange.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2026/06/11.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#include "VulkanRange.hpp"
|
|
#include "core/Macro.h"
|
|
#include "core/TensorUtils.hpp"
|
|
#include "VulkanBackend.hpp"
|
|
|
|
namespace MNN {
|
|
|
|
struct GpuRangeParam {
|
|
ivec4 size;
|
|
};
|
|
|
|
VulkanRange::VulkanRange(Backend* bn) : VulkanBasicExecution(bn) {
|
|
auto vkBn = static_cast<VulkanBackend*>(bn);
|
|
mPipeline = vkBn->getPipeline("glsl_range_comp", {
|
|
VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
|
|
VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
|
|
VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER});
|
|
}
|
|
|
|
VulkanRange::~VulkanRange() {
|
|
}
|
|
|
|
ErrorCode VulkanRange::onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) {
|
|
auto outVT = reinterpret_cast<VulkanTensor*>(outputs[0]->deviceId());
|
|
auto startVT = reinterpret_cast<VulkanTensor*>(inputs[0]->deviceId());
|
|
auto deltaVT = reinterpret_cast<VulkanTensor*>(inputs[2]->deviceId());
|
|
auto imageSize = outVT->imageSize();
|
|
auto vkBn = static_cast<VulkanBackend*>(backend());
|
|
auto sampler = vkBn->getCommonSampler()->get();
|
|
|
|
size_t oldPara = mParams.size();
|
|
if (oldPara != imageSize) {
|
|
mParams.resize(imageSize);
|
|
mDescriptorSets.resize(imageSize);
|
|
for (size_t i = oldPara; i < imageSize; ++i) {
|
|
mParams[i] = std::make_shared<VulkanBuffer>(
|
|
vkBn->getMemoryPool(), false, sizeof(GpuRangeParam),
|
|
nullptr, VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT);
|
|
mDescriptorSets[i].reset(mPipeline->createSet());
|
|
}
|
|
}
|
|
|
|
auto startImg = startVT->image();
|
|
auto deltaImg = deltaVT->image();
|
|
startImg->barrierRead(cmdBuffer->get());
|
|
deltaImg->barrierRead(cmdBuffer->get());
|
|
int pixelOffset = 0;
|
|
for(int i = 0; i < imageSize; ++i){
|
|
auto outImg = outVT->image(i);
|
|
auto total = outImg->width() * outImg->height();
|
|
|
|
auto p = reinterpret_cast<GpuRangeParam*>(mParams[i]->map());
|
|
::memset(p, 0, sizeof(GpuRangeParam));
|
|
p->size[0] = pixelOffset;
|
|
p->size[3] = total;
|
|
mParams[i]->flush(true, 0, sizeof(GpuRangeParam));
|
|
mParams[i]->unmap();
|
|
|
|
auto desSet = mDescriptorSets[i];
|
|
desSet->writeImage(outImg->view(), sampler, VK_IMAGE_LAYOUT_GENERAL, 0);
|
|
desSet->writeImage(startImg->view(), sampler, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 1);
|
|
desSet->writeImage(deltaImg->view(), sampler, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, 2);
|
|
desSet->writeBuffer(mParams[i]->buffer(), 3, mParams[i]->size());
|
|
|
|
outImg->barrierWrite(cmdBuffer->get());
|
|
|
|
mPipeline->bind(cmdBuffer->get(), desSet->get());
|
|
vkCmdDispatch(cmdBuffer->get(), UP_DIV(total, 256), 1, 1);
|
|
|
|
pixelOffset += total;
|
|
}
|
|
|
|
return ErrorCode::NO_ERROR;
|
|
}
|
|
|
|
class VulkanRangeCreator : public VulkanBackend::Creator {
|
|
public:
|
|
virtual VulkanBasicExecution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const MNN::Op* op, Backend* bn) const override {
|
|
if (outputs[0]->getType().code != halide_type_float) {
|
|
return nullptr;
|
|
}
|
|
return new VulkanRange(bn);
|
|
}
|
|
};
|
|
|
|
static bool gResistor = []() {
|
|
VulkanBackend::addCreator(OpType_Range, new VulkanRangeCreator);
|
|
return true;
|
|
}();
|
|
|
|
}// namespace MNN
|