wehub-resource-sync
79 lines
3.0 KiB
C++
79 lines
3.0 KiB
C++
#ifndef VulkanConv1x1Coop_hpp
|
|
#define VulkanConv1x1Coop_hpp
|
|
|
|
#include "VulkanBasicExecution.hpp"
|
|
#include "VulkanConvolution.hpp"
|
|
#include "core/ConvolutionCommon.hpp"
|
|
|
|
namespace MNN {
|
|
|
|
class VulkanConv1x1Coop : public VulkanBasicExecution {
|
|
public:
|
|
VulkanConv1x1Coop(VulkanBackend* backend, const Convolution2DCommon* convOption, const float* weightPtr, const float* biasPtr, int ci, int co, VulkanDevice::CoopMatInfo coopMatInfo, std::shared_ptr<ConvolutionCommon::Int8Common> quantInfo = nullptr);
|
|
virtual ~VulkanConv1x1Coop();
|
|
virtual ErrorCode onEncode(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const VulkanCommandPool::Buffer* cmdBuffer) override;
|
|
virtual bool onClone(Backend* bn, const Op* op, VulkanBasicExecution** dst) override;
|
|
|
|
private:
|
|
VulkanConv1x1Coop(VulkanBackend* backend, const Convolution2DCommon* convOption, int ci, int co,
|
|
uint32_t coopM, uint32_t coopN, uint32_t coopK, uint32_t subgroupSize,
|
|
std::shared_ptr<ConvolutionCommon::Int8Common> quantInfo,
|
|
bool initStaticResource);
|
|
|
|
bool _init(const float* weightPtr, const float* biasPtr, bool initStaticResource);
|
|
|
|
const Convolution2DCommon* mCommon;
|
|
int mCi;
|
|
int mCo;
|
|
bool mIsQuant = false;
|
|
std::shared_ptr<ConvolutionCommon::Int8Common> mQuantCommon;
|
|
uint32_t mBlockSize = 0; // quant block size along K
|
|
uint32_t mPadK = 0;
|
|
uint32_t mPadN = 0;
|
|
bool mQuantConverted = false;
|
|
|
|
// Weight and Bias Buffer
|
|
std::shared_ptr<VulkanBuffer> mWeightBuffer; // float/half weight (float path) or dequantized coop weight (quant path)
|
|
std::shared_ptr<VulkanBuffer> mBiasBuffer;
|
|
// Quant resources (static)
|
|
std::shared_ptr<VulkanBuffer> mQuantWeightBuffer; // int8/int4 packed in [padN, padK]
|
|
std::shared_ptr<VulkanBuffer> mQuantMetaBuffer; // [scale, offset] per (N, K-block)
|
|
|
|
// Pipelines
|
|
const VulkanPipeline* mPackPipeline;
|
|
const VulkanPipeline* mMatMulPipeline;
|
|
const VulkanPipeline* mUnpackPipeline;
|
|
const VulkanPipeline* mPrefillDequantPipeline = nullptr;
|
|
const VulkanPipeline* mDecodePipeline = nullptr;
|
|
|
|
// Descriptor Sets
|
|
std::shared_ptr<VulkanLayout::DescriptorSet> mPackSet;
|
|
std::shared_ptr<VulkanLayout::DescriptorSet> mMatMulSet;
|
|
std::shared_ptr<VulkanLayout::DescriptorSet> mUnpackSet;
|
|
std::shared_ptr<VulkanLayout::DescriptorSet> mPrefillDequantSet;
|
|
std::shared_ptr<VulkanLayout::DescriptorSet> mDecodeSet;
|
|
|
|
// Constant Buffers
|
|
std::shared_ptr<VulkanBuffer> mPackConst;
|
|
std::shared_ptr<VulkanBuffer> mMatMulConst;
|
|
std::shared_ptr<VulkanBuffer> mUnpackConst;
|
|
|
|
// Temp Tensors
|
|
std::shared_ptr<Tensor> mTempInput;
|
|
std::shared_ptr<Tensor> mTempOutput;
|
|
std::shared_ptr<Tensor> mTempWeight;
|
|
|
|
// COOP Shape
|
|
uint32_t COOP_M = 64;
|
|
uint32_t COOP_N = 64;
|
|
uint32_t COOP_K = 16;
|
|
|
|
// subgroupSize
|
|
uint32_t mSubgroupSize = 64;
|
|
};
|
|
|
|
} // namespace MNN
|
|
|
|
#endif
|