wehub-resource-sync
115 lines
3.3 KiB
C++
115 lines
3.3 KiB
C++
//
|
|
// VulkanTopKV2.cpp
|
|
// MNN
|
|
//
|
|
// Vulkan buffer-mode implementation of TopKV2.
|
|
//
|
|
|
|
#include "VulkanTopKV2.hpp"
|
|
#include "core/Macro.h"
|
|
#include "core/TensorUtils.hpp"
|
|
|
|
namespace MNN {
|
|
|
|
struct TopKV2ConstBuffer {
|
|
int rowSize;
|
|
int k;
|
|
int numRows;
|
|
int pad;
|
|
};
|
|
|
|
VulkanTopKV2::VulkanTopKV2(const Op* op, Backend* bn, int k, Tensor* input) : VulkanBasicExecution(bn) {
|
|
MNN_ASSERT(k > 0);
|
|
MNN_ASSERT(input != nullptr);
|
|
auto vkBn = (VulkanBackend*)backend();
|
|
mK = k;
|
|
mLargest = true;
|
|
auto param = op->main_as_TopKV2();
|
|
if (nullptr != param) {
|
|
mLargest = param->largest();
|
|
}
|
|
|
|
mConstBuffer = vkBn->allocUniform();
|
|
std::vector<VkDescriptorType> types{
|
|
VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
|
|
VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
|
|
VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
|
|
VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER,
|
|
};
|
|
|
|
std::string pKey = "glsl_topkv2_";
|
|
if (mLargest) {
|
|
pKey += "SORT_DESC_";
|
|
}
|
|
if (input->getType().code == halide_type_float && vkBn->useFP16()) {
|
|
pKey += "FP16_";
|
|
}
|
|
pKey += "comp";
|
|
|
|
mPipeline = vkBn->getPipeline(pKey, types);
|
|
mDescriptorSet.reset(mPipeline->createSet());
|
|
}
|
|
|
|
VulkanTopKV2::~VulkanTopKV2() {
|
|
auto vkBn = (VulkanBackend*)backend();
|
|
vkBn->recycleUniform(mConstBuffer);
|
|
}
|
|
|
|
ErrorCode VulkanTopKV2::onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) {
|
|
auto input = inputs[0];
|
|
auto outputValue = outputs[0];
|
|
auto outputIndex = outputs[1];
|
|
|
|
const int rowSize = input->length(input->dimensions() - 1);
|
|
if (rowSize <= 0) {
|
|
return NO_ERROR;
|
|
}
|
|
const int numRows = input->elementSize() / rowSize;
|
|
const int k = mK;
|
|
|
|
auto vkBn = static_cast<VulkanBackend*>(backend());
|
|
|
|
// Set GPU params
|
|
auto topkParam = reinterpret_cast<TopKV2ConstBuffer*>(mConstBuffer->map());
|
|
topkParam->rowSize = rowSize;
|
|
topkParam->k = k;
|
|
topkParam->numRows = numRows;
|
|
topkParam->pad = 0;
|
|
mConstBuffer->unmap();
|
|
|
|
// Bind buffers
|
|
mDescriptorSet->writeBuffer(vkBn->getBuffer(outputValue), 0);
|
|
mDescriptorSet->writeBuffer(vkBn->getBuffer(outputIndex), 1);
|
|
mDescriptorSet->writeBuffer(vkBn->getBuffer(input), 2);
|
|
mDescriptorSet->writeBuffer(mConstBuffer->buffer(), 3, mConstBuffer->size());
|
|
|
|
// Dispatch: x=1 (workgroup handles 128 threads internally), y=numRows, z=1
|
|
mPipeline->bind(cmdBuffer->get(), mDescriptorSet->get());
|
|
vkCmdDispatch(cmdBuffer->get(), 1, numRows, 1);
|
|
|
|
return NO_ERROR;
|
|
}
|
|
|
|
class VulkanTopKV2Creator : public VulkanBackend::Creator {
|
|
public:
|
|
virtual VulkanBasicExecution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs, const MNN::Op* op,
|
|
Backend* backend) const override {
|
|
if (inputs.size() < 2 || outputs.size() != 2) {
|
|
return nullptr;
|
|
}
|
|
if (TensorUtils::getDescribe(inputs[0])->dimensionFormat == MNN_DATA_FORMAT_NC4HW4) {
|
|
return nullptr;
|
|
}
|
|
const int k = inputs[1]->host<int32_t>()[0];
|
|
return new VulkanTopKV2(op, backend, k, inputs[0]);
|
|
}
|
|
};
|
|
|
|
static bool gResistor = []() {
|
|
VulkanBackend::addCreator(OpType_TopKV2, new VulkanTopKV2Creator);
|
|
return true;
|
|
}();
|
|
|
|
} // namespace MNN
|