* 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 broadcast_v220_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message("impl/adv_api/detail/pad/broadcast/broadcast_v220_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/pad/broadcast.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_PAD_BROADCAST_BROADCAST_V220_IMPL_H__
#endif
#ifndef IMPL_PAD_BROADCAST_BROADCAST_V220_IMPL_H
#define IMPL_PAD_BROADCAST_BROADCAST_V220_IMPL_H
#include "kernel_basic_intf.h"
#include "kernel_tensor.h"
namespace AscendC {
constexpr uint32_t BRCB_ONE_SIZE = 8;
constexpr uint32_t BRCB_HALF_MAX_REPEATE_TIMES = 254;
constexpr uint32_t BRCB_FLOAT_MAX_REPEATE_TIMES = 255;
constexpr uint8_t GATHER_MASK_PATTERN = 7;
template <typename T>
__aicore__ inline void BrcbToOneBlock(const LocalTensor<T> &srcLocal, const uint32_t firstDim,
uint32_t oneBlockElementNum, LocalTensor<T> &brcbOneBlockTempBuffer)
{
const uint32_t brcbRepeatTime = (firstDim + BRCB_ONE_SIZE - 1) / BRCB_ONE_SIZE;
uint32_t brcbMaxRepeatTimes = BRCB_HALF_MAX_REPEATE_TIMES;
if constexpr (sizeof(T) == sizeof(float)) {
brcbMaxRepeatTimes = BRCB_FLOAT_MAX_REPEATE_TIMES;
}
const uint32_t brcbCount = brcbRepeatTime / brcbMaxRepeatTimes;
const uint32_t tailBrcbRepeatTime = brcbRepeatTime % brcbMaxRepeatTimes;
uint32_t brcbSrcOffset = 0;
uint32_t brcbOneBlockTempBufferOffset = 0;
for (uint32_t i = 0; i < brcbCount; i++) {
Brcb(brcbOneBlockTempBuffer[brcbOneBlockTempBufferOffset],
srcLocal[brcbSrcOffset],
brcbMaxRepeatTimes,
{1, DEFAULT_REPEAT_STRIDE});
brcbOneBlockTempBufferOffset += brcbMaxRepeatTimes * oneBlockElementNum * BRCB_ONE_SIZE;
brcbSrcOffset += brcbMaxRepeatTimes * BRCB_ONE_SIZE;
}
if (tailBrcbRepeatTime != 0) {
Brcb(brcbOneBlockTempBuffer[brcbOneBlockTempBufferOffset],
srcLocal[brcbSrcOffset],
tailBrcbRepeatTime,
{1, DEFAULT_REPEAT_STRIDE});
}
PipeBarrier<PIPE_V>();
}
template <typename T, bool isReuseSource>
__aicore__ inline void TwoDimBroadCastLastDimAlign220(const LocalTensor<T> &dstLocal, const LocalTensor<T> &srcLocal,
LocalTensor<T> &tmpBuffer, const uint32_t firstDim, const uint32_t numBlocks)
{
constexpr uint32_t oneBlockElementNum = ONE_BLK_SIZE / sizeof(T);
BrcbToOneBlock(srcLocal, firstDim, oneBlockElementNum, tmpBuffer);
SetVectorMask<T, MaskMode::COUNTER>(numBlocks);
const CopyRepeatParams copyRepeatParams = {1, 0, (uint16_t)(numBlocks / oneBlockElementNum), 1};
uint32_t CopyCounts = firstDim / MAX_REPEAT_TIMES;
uint32_t dstOffset = 0;
uint32_t brcbOneBlockTempBufferOffset = 0;
for (uint32_t i = 0; i < CopyCounts; i++) {
Copy<T, false>(dstLocal[dstOffset],
tmpBuffer[brcbOneBlockTempBufferOffset],
MASK_PLACEHOLDER,
MAX_REPEAT_TIMES,
copyRepeatParams);
dstOffset += MAX_REPEAT_TIMES * numBlocks;
brcbOneBlockTempBufferOffset += MAX_REPEAT_TIMES * oneBlockElementNum;
}
uint32_t tailsCopyRepeatTimes = firstDim % MAX_REPEAT_TIMES;
if (tailsCopyRepeatTimes != 0) {
Copy<T, false>(dstLocal[dstOffset],
tmpBuffer[brcbOneBlockTempBufferOffset],
MASK_PLACEHOLDER,
tailsCopyRepeatTimes,
copyRepeatParams);
}
PipeBarrier<PIPE_V>();
}
template <typename T, bool isReuseSource>
__aicore__ inline void TwoDimBroadCastLastDimNotAlign220(const LocalTensor<T> &dstLocal, const LocalTensor<T> &srcLocal,
LocalTensor<T> &tmpBuffer, const uint32_t firstDim, const uint32_t numBlocks)
{
constexpr uint32_t oneBlockElementNum = ONE_BLK_SIZE / sizeof(T);
BrcbToOneBlock(srcLocal, firstDim, oneBlockElementNum, tmpBuffer);
const uint32_t dstRepeatSize = (numBlocks + oneBlockElementNum - 1) / oneBlockElementNum;
const uint32_t numBlocksAlign = dstRepeatSize * oneBlockElementNum;
SetVectorMask<T, MaskMode::COUNTER>(numBlocksAlign);
const CopyRepeatParams copyRepeatParams = {1, 0, (uint16_t)dstRepeatSize, 1};
uint32_t CopyCounts = firstDim / MAX_REPEAT_TIMES;
uint32_t dstOffset = 0;
uint32_t brcbOneBlockTempBufferOffset = 0;
auto copyTempBuffer = tmpBuffer[firstDim * oneBlockElementNum];
for (uint32_t i = 0; i < CopyCounts; i++) {
Copy<T, false>(copyTempBuffer[dstOffset],
tmpBuffer[brcbOneBlockTempBufferOffset],
MASK_PLACEHOLDER,
MAX_REPEAT_TIMES,
copyRepeatParams);
dstOffset += MAX_REPEAT_TIMES * numBlocksAlign;
brcbOneBlockTempBufferOffset += MAX_REPEAT_TIMES * oneBlockElementNum;
}
uint32_t tailsCopyRepeatTimes = firstDim % MAX_REPEAT_TIMES;
if (tailsCopyRepeatTimes != 0) {
Copy<T, false>(copyTempBuffer[dstOffset],
tmpBuffer[brcbOneBlockTempBufferOffset],
MASK_PLACEHOLDER,
tailsCopyRepeatTimes,
copyRepeatParams);
}
PipeBarrier<PIPE_V>();
const GatherMaskParams gatherMaskParams = {
1, (uint16_t)firstDim, (uint16_t)dstRepeatSize, 0};
uint64_t rsvdCnt = 0;
GatherMask(dstLocal, copyTempBuffer, GATHER_MASK_PATTERN, true, numBlocks, gatherMaskParams, rsvdCnt);
SetMaskCount();
PipeBarrier<PIPE_V>();
}
template <typename T>
__aicore__ inline void GetAlignLoopNumbers(const uint32_t firstDim, const uint32_t numBlocks,
const uint32_t tmpBufferSize, uint32_t &oneRepeatSize, uint32_t &rangeM, uint32_t &tailM)
{
constexpr uint32_t oneBlockElementNum = ONE_BLK_SIZE / sizeof(T);
constexpr uint32_t minBrcbTempBufferSize = oneBlockElementNum * oneBlockElementNum;
constexpr uint32_t minTmpBufferSize = minBrcbTempBufferSize;
ASCENDC_ASSERT((tmpBufferSize >= minTmpBufferSize), {
KERNEL_LOG(KERNEL_ERROR,
"tmpBufferSize can't smaller than minTmpBufferSize, tmpBufferSize is %u, minTmpBufferSize is %u!",
tmpBufferSize,
minTmpBufferSize);
});
oneRepeatSize = tmpBufferSize / minTmpBufferSize * oneBlockElementNum;
rangeM = firstDim / oneRepeatSize;
tailM = firstDim - oneRepeatSize * rangeM;
}
template <typename T>
__aicore__ inline void GetNotAlignLoopNumbers(const uint32_t firstDim, const uint32_t numBlocks,
const uint32_t tmpBufferSize, uint32_t &oneRepeatSize, uint32_t &rangeM, uint32_t &tailM)
{
constexpr uint32_t oneBlockElementNum = ONE_BLK_SIZE / sizeof(T);
constexpr uint32_t minBrcbTempBufferSize = oneBlockElementNum * oneBlockElementNum;
const uint32_t dstRepeatSize = (numBlocks + oneBlockElementNum - 1) / oneBlockElementNum;
const uint32_t numBlocksAlign = dstRepeatSize * oneBlockElementNum;
const uint32_t minCopyTempBufferSize = oneBlockElementNum * numBlocksAlign;
const uint32_t minTmpBufferSize = minBrcbTempBufferSize + minCopyTempBufferSize;
ASCENDC_ASSERT((tmpBufferSize >= minTmpBufferSize), {
KERNEL_LOG(KERNEL_ERROR,
"tmpBufferSize can't smaller than minTmpBufferSize, tmpBufferSize is %u, minTmpBufferSize is %u!",
tmpBufferSize,
minTmpBufferSize);
});
oneRepeatSize = tmpBufferSize / minTmpBufferSize * oneBlockElementNum;
rangeM = firstDim / oneRepeatSize;
tailM = firstDim - oneRepeatSize * rangeM;
}
template <typename T, int32_t dim, int32_t axis, bool isReuseSource = false>
__aicore__ inline void TwoDimBroadCastLastDim(const LocalTensor<T> &dstLocal, const LocalTensor<T> &srcLocal,
const uint32_t dstShape[dim], const uint32_t srcShape[dim], LocalTensor<T> &tmpBuffer)
{
const auto firstDim = dstShape[0];
const auto numBlocks = dstShape[axis];
uint32_t oneRepeatSize = 0;
uint32_t rangeM = 0;
uint32_t tailM = 0;
uint32_t dstLocalOffset = 0;
uint32_t srcLocalOffset = 0;
if (numBlocks * sizeof(T) % ONE_BLK_SIZE == 0) {
GetAlignLoopNumbers<T>(firstDim, numBlocks, tmpBuffer.GetSize(), oneRepeatSize, rangeM, tailM);
for (uint32_t i = 0; i < rangeM; i++) {
TwoDimBroadCastLastDimAlign220<T, isReuseSource>(
dstLocal[dstLocalOffset], srcLocal[srcLocalOffset], tmpBuffer, oneRepeatSize, numBlocks);
dstLocalOffset += oneRepeatSize * numBlocks;
srcLocalOffset += oneRepeatSize;
}
if (tailM != 0) {
TwoDimBroadCastLastDimAlign220<T, isReuseSource>(
dstLocal[dstLocalOffset], srcLocal[srcLocalOffset], tmpBuffer, tailM, numBlocks);
}
} else {
GetNotAlignLoopNumbers<T>(firstDim, numBlocks, tmpBuffer.GetSize(), oneRepeatSize, rangeM, tailM);
for (uint32_t i = 0; i < rangeM; i++) {
TwoDimBroadCastLastDimNotAlign220<T, isReuseSource>(
dstLocal[dstLocalOffset], srcLocal[srcLocalOffset], tmpBuffer, oneRepeatSize, numBlocks);
dstLocalOffset += oneRepeatSize * numBlocks;
srcLocalOffset += oneRepeatSize;
}
if (tailM != 0) {
TwoDimBroadCastLastDimNotAlign220<T, isReuseSource>(
dstLocal[dstLocalOffset], srcLocal[srcLocalOffset], tmpBuffer, tailM, numBlocks);
}
}
}
template <typename T>
__aicore__ inline void NoBroad(const LocalTensor<T> &dstLocal, const LocalTensor<T> &srcLocal, const uint32_t size)
{
SetVectorMask<T, MaskMode::COUNTER>(size);
Copy<T, false>(dstLocal, srcLocal, MASK_PLACEHOLDER, 1, {1, 1, DEFAULT_REPEAT_STRIDE, DEFAULT_REPEAT_STRIDE});
PipeBarrier<PIPE_V>();
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_PAD_BROADCAST_BROADCAST_V220_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_PAD_BROADCAST_BROADCAST_V220_IMPL_H__
#endif
