* 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 mean_3510_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/reduce/mean/mean_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/reduce/mean.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_MEAN_MEAN_C310_IMPL_H__
#endif
#ifndef IMPL_REDUCE_MEAN_C310_IMPL_H
#define IMPL_REDUCE_MEAN_C310_IMPL_H
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
#include "include/adv_api/reduce/mean_utils.h"
#include "../../common/common.h"
namespace AscendC {
namespace Internal {
template <typename T, typename accType>
struct GetConvType {
using type = T;
};
template <>
struct GetConvType<half, float> {
using type = float;
};
template <typename T, typename ConvType>
__simd_callee__ inline void LoadSrcData(
Reg::RegTensor<ConvType>& srcReg, __ubuf__ T* src, uint16_t index, uint32_t offset, Reg::MaskReg& mask)
{
Reg::RegTensor<T> srcTmpReg;
if constexpr (std::is_same<T, half>::value && std::is_same<ConvType, float>::value) {
Reg::LoadAlign<T, Reg::LoadDist::DIST_UNPACK_B16>(srcTmpReg, src + index * offset);
Reg::Cast<float, T, castTraitB16ToB32>(srcReg, srcTmpReg, mask);
} else {
Reg::LoadAlign(srcReg, src + index * offset);
}
}
template <typename T, typename U, typename ConvType>
__simd_vf__ inline void MeanForOneRepeatTime(
__ubuf__ T* dstUb, __ubuf__ U* srcUb, const MeanParams meanParams, uint32_t calCount, uint32_t offset)
{
uint32_t count;
ConvType scalarValue = static_cast<ConvType>(1.0f / meanParams.n);
Reg::MaskReg mask;
Reg::UnalignReg uregOut;
Reg::RegTensor<ConvType> srcTmpReg, dstTmpReg;
Reg::RegTensor<T> dstReg;
for (int i = 0; i < meanParams.outter; i++) {
count = calCount;
mask = Reg::UpdateMask<ConvType>(count);
LoadSrcData(srcTmpReg, srcUb, i, offset, mask);
Reg::ReduceSum(dstTmpReg, srcTmpReg, mask);
Reg::Muls(dstTmpReg, dstTmpReg, scalarValue, mask);
if constexpr (sizeof(T) == sizeof(half) && sizeof(ConvType) == sizeof(float)) {
Reg::Cast<T, float, castTraitB32ToB16>(dstReg, dstTmpReg, mask);
Reg::Pack<uint16_t, uint32_t, Reg::HighLowPart::LOWEST>(
(Reg::RegTensor<uint16_t>&)dstReg, (Reg::RegTensor<uint32_t>&)dstReg);
} else {
dstReg = dstTmpReg;
}
Reg::StoreUnAlign<T, Reg::PostLiteral::POST_MODE_UPDATE>(dstUb, dstReg, uregOut, 1);
}
Reg::StoreUnAlignPost(dstUb, uregOut, 0);
}
template <typename T, typename ConvType, bool isFirstRepeat>
__simd_vf__ inline void ReduceSumNextN(
__ubuf__ ConvType* dstUb, __ubuf__ T* srcUb, const MeanParams meanParams, uint32_t calCount, uint32_t repeatTimes,
uint32_t offset)
{
uint32_t count;
Reg::MaskReg mask;
Reg::UnalignReg uregIn;
Reg::RegTensor<ConvType> srcReg, dstReg;
constexpr int32_t eleCountPerVL = GetVecLen() / sizeof(ConvType);
for (uint16_t i = 0; i < meanParams.outter; i++) {
count = calCount;
auto dstTmpUb = dstUb + i * offset;
for (uint16_t j = 0; j < repeatTimes; j++) {
mask = Reg::UpdateMask<ConvType>(count);
if constexpr (isFirstRepeat) {
LoadSrcData(srcReg, srcUb + i * meanParams.inner, j, eleCountPerVL, mask);
} else {
LoadSrcData(srcReg, srcUb + i * offset, j, eleCountPerVL, mask);
}
Reg::ReduceSum(dstReg, srcReg, mask);
Reg::StoreUnAlign<ConvType, Reg::PostLiteral::POST_MODE_UPDATE>(dstTmpUb, dstReg, uregIn, 1);
}
Reg::StoreUnAlignPost(dstTmpUb, uregIn, 0);
}
Reg::LocalMemBar<Reg::MemType::VEC_STORE, Reg::MemType::VEC_LOAD>();
}
}
template <typename T, typename accType, bool isReuseSource, bool isBasicBlock, int32_t reduceDim>
__aicore__ inline void MeanCheckParams(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const MeanParams& meanParams)
{
static_assert(SupportType<T, half, float>(), "current data type is not supported on current device!");
CheckTensorPos<T>(dstTensor, Hardware::UB, "dstTensor", "VECIN / VECCALC / VECOUT", "Mean");
CheckTensorPos<T>(srcTensor, Hardware::UB, "srcTensor", "VECIN / VECCALC / VECOUT", "Mean");
CheckTensorPos<uint8_t>(sharedTmpBuffer, Hardware::UB, "sharedTmpBuffer", "VECIN / VECCALC / VECOUT", "Mean");
constexpr uint32_t meanInnerAlignLen = 32;
ASCENDC_ASSERT((1 <= meanParams.n) && (meanParams.n <= meanParams.inner), {
KERNEL_LOG(KERNEL_ERROR, "The value of n must be greater than or equal to 1 and less than or equal to inner.");
});
ASCENDC_ASSERT((meanParams.inner * sizeof(T) % meanInnerAlignLen == 0), {
KERNEL_LOG(KERNEL_ERROR, "The value of inner * sizeof(T) must be an integer multiple of 32.");
});
}
template <
typename T, typename accType = T, bool isReuseSource = false, bool isBasicBlock = false, int32_t reduceDim = -1>
__aicore__ inline void MeanImpl(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const MeanParams& meanParams)
{
if ASCEND_IS_AIC {
return;
}
MeanCheckParams<T, accType, isReuseSource, isBasicBlock, reduceDim>(
dstTensor, srcTensor, sharedTmpBuffer, meanParams);
using ConvType = typename Internal::GetConvType<T, accType>::type;
__ubuf__ T* dstUb = (__ubuf__ T*)dstTensor.GetPhyAddr();
__ubuf__ T* srcUb = (__ubuf__ T*)srcTensor.GetPhyAddr();
__ubuf__ ConvType* sharedTmpBufferUb = (__ubuf__ ConvType*)sharedTmpBuffer.GetPhyAddr();
constexpr int32_t eleCountPerVL = GetVecLen() / sizeof(ConvType);
uint16_t repeatTimes = CeilDivision(meanParams.n, eleCountPerVL);
uint32_t loopRepeatTimes;
uint32_t calCount = meanParams.n;
uint32_t totalCnt = 1;
uint32_t dataSize = repeatTimes;
uint32_t offset = AlignUp(CeilDivision(meanParams.inner, eleCountPerVL), 32);
while (dataSize > 1) {
++totalCnt;
dataSize = CeilDivision(dataSize, eleCountPerVL);
}
if (repeatTimes == 1) {
Internal::MeanForOneRepeatTime<T, T, ConvType>(dstUb, srcUb, meanParams, meanParams.n, meanParams.inner);
return;
}
Internal::ReduceSumNextN<T, ConvType, true>(sharedTmpBufferUb, srcUb, meanParams, calCount, repeatTimes, offset);
--totalCnt;
loopRepeatTimes = repeatTimes;
while (totalCnt != 0) {
calCount = loopRepeatTimes;
loopRepeatTimes = CeilDivision(loopRepeatTimes, eleCountPerVL);
if (totalCnt == 1) {
Internal::MeanForOneRepeatTime<T, ConvType, ConvType>(
dstUb, sharedTmpBufferUb, meanParams, calCount, offset);
} else {
Internal::ReduceSumNextN<ConvType, ConvType, false>(
sharedTmpBufferUb, sharedTmpBufferUb, meanParams, calCount, loopRepeatTimes, offset);
}
--totalCnt;
}
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_MEAN_MEAN_C310_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_MEAN_MEAN_C310_IMPL_H__
#endif
