* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file swiglu_3510_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/activation/swiglu/swiglu_3510_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use \"#include \"adv_api/activation/swiglu.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SWIGLU_C310_IMPL_H__
#endif
#ifndef IMPL_ACTIVATION_SWIGLU_SWIGLU_IMPL_C310_H
#define IMPL_ACTIVATION_SWIGLU_SWIGLU_IMPL_C310_H
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
#include "../../common/check.h"
#include "../../common/common.h"
namespace AscendC {
namespace Internal {
template <typename T>
__simd_vf__ inline void SwiGLUImplVF(
__ubuf__ T* dst, __ubuf__ T* src0, __ubuf__ T* src1, const float scalarValue, uint32_t count,
const uint16_t repeatTimes)
{
Reg::RegTensor<float> srcVreg0;
Reg::RegTensor<float> srcVreg1;
Reg::RegTensor<half> srcOrigin0;
Reg::RegTensor<half> srcOrigin1;
Reg::RegTensor<float> tmpReg0;
Reg::RegTensor<float> tmpReg1;
Reg::RegTensor<float> dstVreg;
Reg::MaskReg mask;
constexpr uint32_t oneRepElm = static_cast<uint32_t>(GetVecLen() / sizeof(float));
for (uint16_t i = 0; i < repeatTimes; ++i) {
mask = Reg::UpdateMask<float>(count);
if constexpr (sizeof(T) == sizeof(half)) {
Reg::LoadAlign<half, Reg::LoadDist::DIST_UNPACK_B16>(srcOrigin0, src0 + i * oneRepElm);
Reg::Cast<float, half, castTraitB16ToB32>(srcVreg0, srcOrigin0, mask);
Reg::LoadAlign<half, Reg::LoadDist::DIST_UNPACK_B16>(srcOrigin1, src1 + i * oneRepElm);
Reg::Cast<float, half, castTraitB16ToB32>(srcVreg1, srcOrigin1, mask);
} else {
Reg::LoadAlign(srcVreg0, src0 + i * oneRepElm);
Reg::LoadAlign(srcVreg1, src1 + i * oneRepElm);
}
Reg::Muls(tmpReg0, srcVreg1, scalarValue, mask);
Reg::Exp(tmpReg1, tmpReg0, mask);
Reg::Adds(tmpReg1, tmpReg1, 1.0f, mask);
Reg::Div(tmpReg1, srcVreg1, tmpReg1, mask);
Reg::Mul(dstVreg, srcVreg0, tmpReg1, mask);
if constexpr (sizeof(T) == sizeof(half)) {
Reg::Cast<half, float, castTraitB32ToB16>((Reg::RegTensor<half>&)dstVreg, dstVreg, mask);
Reg::StoreAlign<half, Reg::StoreDist::DIST_PACK_B32>(
dst + i * oneRepElm, (Reg::RegTensor<half>&)dstVreg, mask);
} else {
Reg::StoreAlign(dst + i * oneRepElm, dstVreg, mask);
}
}
}
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void SwiGLUImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1,
const float scalarValue, const uint32_t count)
{
if ASCEND_IS_AIC {
return;
}
static_assert(SupportType<T, half, float>(), "SwiGLU only support half/float data type on current device!");
ASCENDC_ASSERT((srcTensor0.GetSize() == srcTensor1.GetSize()), {
KERNEL_LOG(KERNEL_ERROR, "Input params.GetSize must be equal with each other!");
});
CheckTensorPosition(dstTensor, "dstTensor", "VECIN, VECOUT, VECCALC");
CheckTensorPosition(srcTensor0, "srcTensor0", "VECIN, VECOUT, VECCALC");
CheckTensorPosition(srcTensor1, "srcTensor1", "VECIN, VECOUT, VECCALC");
CheckCalCount(count, "count", dstTensor, "dstTensor", "SwiGLU");
CheckCalCount(count, "count", srcTensor0, "srcTensor0", "SwiGLU");
CheckCalCount(count, "count", srcTensor1, "srcTensor1", "SwiGLU");
constexpr uint32_t oneRepElm = static_cast<uint32_t>(GetVecLen() / sizeof(float));
uint16_t repeatTimes = static_cast<uint16_t>(CeilDivision(count, oneRepElm));
const float scalar = static_cast<float>(static_cast<float>(-1.0) * static_cast<float>(scalarValue));
Internal::SwiGLUImplVF<T>(
(__ubuf__ T*)dstTensor.GetPhyAddr(), (__ubuf__ T*)srcTensor0.GetPhyAddr(), (__ubuf__ T*)srcTensor1.GetPhyAddr(),
scalar, count, repeatTimes);
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void SwiGLUImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1,
const float scalarValue, const LocalTensor<uint8_t>& sharedTmpBuffer, const uint32_t count)
{
CheckTensorPosition(sharedTmpBuffer, "sharedTmpBuffer", "VECIN, VECOUT, VECCALC");
SwiGLUImpl<T, isReuseSource>(dstTensor, srcTensor0, srcTensor1, scalarValue, count);
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void SwiGLUImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1,
const float scalarValue, const LocalTensor<uint8_t>& sharedTmpBuffer)
{
CheckTensorPosition(sharedTmpBuffer, "sharedTmpBuffer", "VECIN, VECOUT, VECCALC");
SwiGLUImpl<T, isReuseSource>(dstTensor, srcTensor0, srcTensor1, scalarValue, srcTensor0.GetSize());
}
template <typename T, bool isReuseSource = false>
__aicore__ inline void SwiGLUImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor0, const LocalTensor<T>& srcTensor1,
const float scalarValue)
{
SwiGLUImpl<T, isReuseSource>(dstTensor, srcTensor0, srcTensor1, scalarValue, srcTensor0.GetSize());
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SWIGLU_C310_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_SWIGLU_C310_IMPL_H__
#endif
