wehub-resource-sync
75 lines
2.5 KiB
C++
75 lines
2.5 KiB
C++
//
|
|
// SplitGeluExecution.cpp
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2023/09/18.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
#ifdef MNN_SUPPORT_TRANSFORMER_FUSE
|
|
|
|
#include "SplitGeluExecution.hpp"
|
|
#include "core/TensorUtils.hpp"
|
|
|
|
namespace MNN {
|
|
namespace CUDA {
|
|
|
|
SplitGeluExecution::SplitGeluExecution(Backend* backend) : Execution(backend) {
|
|
mFDiv = 1.4140625F;
|
|
mFAdd = 1.F;
|
|
mFMul = 0.5F;
|
|
}
|
|
ErrorCode SplitGeluExecution::onResize(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
|
|
// Nothing todo
|
|
return NO_ERROR;
|
|
}
|
|
|
|
ErrorCode SplitGeluExecution::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("start SplitGeluExecution onExecute...");
|
|
#endif
|
|
auto runtime = static_cast<CUDABackend*>(backend())->getCUDARuntime();
|
|
|
|
int32_t const gridSize = inputs[0]->length(0) * inputs[0]->length(1);
|
|
int32_t const nHalfHiddenSize = inputs[0]->length(2) / 2; // HHS
|
|
if(inputs.size() > 1) {
|
|
MNN_ASSERT(inputs[1]->dimensions() == 1);
|
|
MNN_ASSERT(inputs[1]->length(0) == inputs[0]->length(2));
|
|
}
|
|
|
|
if(static_cast<CUDABackend*>(backend())->useFp16()) {
|
|
auto const input0 = static_cast<half const*>((void *)inputs[0]->deviceId());
|
|
void* input1 = nullptr;
|
|
if(inputs.size() > 1) {
|
|
input1 = (void *)inputs[1]->deviceId();
|
|
}
|
|
auto output = static_cast<half*>((void *)outputs[0]->deviceId());
|
|
launchSplitGeLUKernel<half>(gridSize, nHalfHiddenSize, input0, static_cast<half const*>(input1), output, mFDiv, mFAdd, mFMul);
|
|
} else {
|
|
auto const input0 = static_cast<float const*>((void *)inputs[0]->deviceId());
|
|
void* input1 = nullptr;
|
|
if(inputs.size() > 1) {
|
|
input1 = (void *)inputs[1]->deviceId();
|
|
}
|
|
auto output = static_cast<float*>((void *)outputs[0]->deviceId());
|
|
launchSplitGeLUKernel<float>(gridSize, nHalfHiddenSize, input0, static_cast<float const*>(input1), output, mFDiv, mFAdd, mFMul);
|
|
}
|
|
|
|
#ifdef LOG_VERBOSE
|
|
MNN_PRINT("end SplitGeluExecution onExecute...");
|
|
#endif
|
|
return NO_ERROR;
|
|
}
|
|
|
|
|
|
class SplitGeluCreator : 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 SplitGeluExecution(backend);
|
|
}
|
|
};
|
|
|
|
CUDACreatorRegister<SplitGeluCreator> __SplitGeluExecution(OpType_SplitGeLU);
|
|
} // namespace CUDA
|
|
} // namespace MNN
|
|
#endif |