* 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_common_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message( \
"impl/adv_api/detail/reduce/mean/mean_common_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_COMMON_IMPL_H__
#endif
#ifndef LIB_REDUCE_MEAN_MEAN_COMMON_IMPL_H
#define LIB_REDUCE_MEAN_MEAN_COMMON_IMPL_H
#include "kernel_tensor.h"
#include "kernel_basic_intf.h"
#include "include/adv_api/reduce/mean_utils.h"
#ifdef ASCENDC_CPU_DEBUG
#include "../../api_check/kernel_check/reduce/mean/mean_check.h"
#endif
#include "../../api_check/kernel_api_check.h"
namespace AscendC {
constexpr uint32_t HALF_NUM_PER = 128;
constexpr uint32_t FLOAT_NUM_PER = 64;
__aicore__ inline void MeanCast(
const LocalTensor<half>& dstTensor, const LocalTensor<half>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const MeanParams& meanParams)
{
uint32_t elementNumPerRep = FLOAT_NUM_PER;
uint32_t repeatTimes = (meanParams.n + elementNumPerRep - 1) / elementNumPerRep;
const UnaryRepeatParams unaryParams;
float scalarValue = static_cast<float>(1) / static_cast<float>(static_cast<int32_t>(meanParams.n));
LocalTensor<float> TmpTensor = sharedTmpBuffer.ReinterpretCast<float>();
LocalTensor<half> castTensor = sharedTmpBuffer.ReinterpretCast<half>();
SetMaskCount();
for (uint32_t row = 0; row < meanParams.outter; ++row) {
SetVectorMask<half>(0, meanParams.n);
Cast<float, half, false>(
TmpTensor, srcTensor[row * meanParams.inner], RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
{1, 1, DEFAULT_REPEAT_STRIDE, HALF_DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
RepeatReduceSum<float, false>(
TmpTensor[meanParams.inner], TmpTensor, 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE,
DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
uint32_t reduceNums = repeatTimes;
while (reduceNums > 1) {
SetVectorMask<half>(0, reduceNums);
reduceNums = (reduceNums + elementNumPerRep - 1) / elementNumPerRep;
RepeatReduceSum<float, false>(
TmpTensor[meanParams.inner], TmpTensor[meanParams.inner], 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE,
DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
}
SetVectorMask<half>(0, 1);
Muls<float, false>(
TmpTensor[meanParams.inner], TmpTensor[meanParams.inner], scalarValue, MASK_PLACEHOLDER, 1, unaryParams);
PipeBarrier<PIPE_V>();
Cast<half, float, false>(
castTensor, TmpTensor[meanParams.inner], RoundMode::CAST_NONE, MASK_PLACEHOLDER, 1,
{1, 1, HALF_DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
RepeatReduceSum<half, false>(
dstTensor[row], castTensor, 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE,
DEFAULT_REPEAT_STRIDE);
}
SetMaskNorm();
ResetMask();
}
template <typename T>
__aicore__ inline void MeanForOneRepeatTime(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const MeanParams& meanParams, T scalarValue)
{
SetVectorMask<T>(0, meanParams.n);
for (uint32_t row = 0; row < meanParams.outter; ++row) {
RepeatReduceSum<T, false>(
dstTensor[row], srcTensor[row * meanParams.inner], 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE,
DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
}
PipeBarrier<PIPE_V>();
SetVectorMask<T>(0, meanParams.outter);
const UnaryRepeatParams unaryParams;
Muls<T, false>(dstTensor, dstTensor, scalarValue, MASK_PLACEHOLDER, 1, unaryParams);
SetMaskNorm();
ResetMask();
}
template <typename T, typename accType, bool isReuseSource>
__aicore__ inline void MeanCommon(
const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor, const LocalTensor<uint8_t>& sharedTmpBuffer,
const MeanParams& meanParams)
{
uint32_t elementNumPerRep = FLOAT_NUM_PER;
if constexpr (sizeof(T) == sizeof(half)) {
elementNumPerRep = HALF_NUM_PER;
}
uint32_t repeatTimes = (meanParams.n + elementNumPerRep - 1) / elementNumPerRep;
T scalarValue = static_cast<T>(static_cast<float>(1) / static_cast<float>(static_cast<int32_t>(meanParams.n)));
SetMaskCount();
if (repeatTimes == 1) {
return MeanForOneRepeatTime(dstTensor, srcTensor, meanParams, scalarValue);
}
const UnaryRepeatParams unaryParams;
LocalTensor<T> TmpTensor = sharedTmpBuffer.ReinterpretCast<T>();
for (uint32_t row = 0; row < meanParams.outter; ++row) {
uint32_t reduceNums = repeatTimes;
SetVectorMask<T>(0, meanParams.n);
RepeatReduceSum<T, false>(
TmpTensor, srcTensor[row * meanParams.inner], 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE,
DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
PipeBarrier<PIPE_V>();
while (reduceNums > 1) {
SetVectorMask<T>(0, reduceNums);
reduceNums = (reduceNums + elementNumPerRep - 1) / elementNumPerRep;
if (reduceNums == 1) {
RepeatReduceSum<T, false>(
dstTensor[row], TmpTensor, 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE,
DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
} else {
RepeatReduceSum<T, false>(
TmpTensor, TmpTensor, 1, MASK_PLACEHOLDER, DEFAULT_BLK_STRIDE, DEFAULT_BLK_STRIDE,
DEFAULT_BLK_STRIDE, DEFAULT_REPEAT_STRIDE);
}
PipeBarrier<PIPE_V>();
}
}
SetVectorMask<T>(0, meanParams.outter);
Muls<T, false>(dstTensor, dstTensor, scalarValue, MASK_PLACEHOLDER, 1, unaryParams);
SetMaskNorm();
}
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)
{
uint32_t elementNumPerRep = FLOAT_NUM_PER;
if constexpr (sizeof(T) == sizeof(half) && sizeof(accType) == sizeof(float)) {
uint32_t repeatTimes = (meanParams.n + elementNumPerRep - 1) / elementNumPerRep;
uint32_t finalWorkSize =
meanParams.inner * sizeof(float) + (repeatTimes + ONE_BLK_SIZE - 1) / ONE_BLK_SIZE * ONE_BLK_SIZE;
CHECK_FUNC_HIGHLEVEL_API(
Mean, (T, accType, isReuseSource, isBasicBlock, reduceDim),
(dstTensor, srcTensor, sharedTmpBuffer, meanParams, finalWorkSize));
MeanCast(dstTensor, srcTensor, sharedTmpBuffer, meanParams);
} else {
if constexpr (sizeof(T) == sizeof(half)) {
elementNumPerRep = HALF_NUM_PER;
}
uint32_t repeatTimes = (meanParams.n + elementNumPerRep - 1) / elementNumPerRep;
uint32_t finalWorkSize = (repeatTimes + ONE_BLK_SIZE - 1) / ONE_BLK_SIZE * ONE_BLK_SIZE;
CHECK_FUNC_HIGHLEVEL_API(
Mean, (T, accType, isReuseSource, isBasicBlock, reduceDim),
(dstTensor, srcTensor, sharedTmpBuffer, meanParams, finalWorkSize));
MeanCommon<T, accType, isReuseSource>(dstTensor, srcTensor, sharedTmpBuffer, meanParams);
}
}
#pragma end_pipe
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_MEAN_MEAN_COMMON_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_REDUCE_MEAN_MEAN_COMMON_IMPL_H__
#endif
