wehub-resource-sync
61 lines
1.9 KiB
C++
61 lines
1.9 KiB
C++
//
|
|
// SeqLen2SpatialExecution.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2023/09/18.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
#ifdef MNN_SUPPORT_TRANSFORMER_FUSE
|
|
|
|
#include "SeqLen2SpatialExecution.hpp"
|
|
#include "core/TensorUtils.hpp"
|
|
|
|
namespace MNN {
|
|
namespace CUDA {
|
|
|
|
SeqLen2SpatialExecution::SeqLen2SpatialExecution(Backend* backend) : Execution(backend) {
|
|
// Nothing todo
|
|
}
|
|
ErrorCode SeqLen2SpatialExecution::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
|
|
// Nothing todo
|
|
return NO_ERROR;
|
|
}
|
|
|
|
ErrorCode SeqLen2SpatialExecution::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("start SeqLen2SpatialExecution onExecute...");
|
|
#endif
|
|
auto runtime = static_cast<CUDABackend*>(backend())->getCUDARuntime();
|
|
|
|
int32_t const BS = inputs[1]->length(0);
|
|
int32_t const HW = inputs[1]->length(1);
|
|
int32_t const C = inputs[1]->length(2);
|
|
int32_t const gridSize = BS * HW;
|
|
|
|
bool isHalf = static_cast<CUDABackend*>(backend())->useFp16();
|
|
void *input0 = (void *)inputs[1]->deviceId();
|
|
void *input1 = (void *)inputs[2]->deviceId();
|
|
void *input2 = (void *)inputs[3]->deviceId();
|
|
void *output = (void *)outputs[0]->deviceId();
|
|
|
|
launchSeqLen2SpatialKernel(input0, input1, input2, output, isHalf, gridSize, C);
|
|
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("end SeqLen2SpatialExecution onExecute...");
|
|
#endif
|
|
return NO_ERROR;
|
|
}
|
|
|
|
|
|
class SeqLen2SpatialCreator : public CUDABackend::Creator {
|
|
public:
|
|
virtual Execution* onCreate(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
|
|
const MNN::Op* op, Backend* backend) const override {
|
|
return new SeqLen2SpatialExecution(backend);
|
|
}
|
|
};
|
|
|
|
CUDACreatorRegister<SeqLen2SpatialCreator> __SeqLen2SpatialExecution(OpType_SeqLen2Spatial);
|
|
} // namespace CUDA
|
|
} // namespace MNN
|
|
#endif |