* Copyright (c) 2026 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.
*/
#ifndef OP_KERNEL_SPLIT_V_GENERAL_H_
#define OP_KERNEL_SPLIT_V_GENERAL_H_
#include "kernel_operator.h"
#include "kernel_operator_list_tensor_intf.h"
#include "kernel_tiling/kernel_tiling.h"
#include "split_v_tiling_data.h"
#include "split_v_kernel_common.h"
using namespace AscendC;
template <typename T>
class SplitVGeneral : private SplitVTileCopyState<T> {
SPLIT_V_DECLARE_TILE_COPY_API(SplitVGeneral);
private:
static constexpr uint32_t BUFFER_NUM = 2;
static constexpr uint32_t BLOCK_SIZE = 32;
using Base = SplitVTileCopyState<T>;
SPLIT_V_USE_TILE_COPY_STATE(Base);
uint64_t midTileTotal;
};
template <typename T>
__aicore__ inline void SplitVGeneral<T>::Init(GM_ADDR input, GM_ADDR output, SplitVTilingData* tiling, TPipe* pipe)
{
blockIdx = GetBlockIdx();
if (blockIdx >= GetBlockNum()) {
return;
}
LoadTileCopyTiling(output, tiling);
midTileTotal = midLength * tileNum;
LoadTileCopyLoop(tiling);
InitTileCopyBuffer(input, outerLength * midLength * innerLength, pipe);
}
template <typename T>
__aicore__ inline void SplitVGeneral<T>::Process()
{
const uint64_t firstTaskIdx = loopOffset;
const uint64_t firstOuterIdx = firstTaskIdx / midTileTotal;
const uint64_t firstMidIdx = (firstTaskIdx % midTileTotal) / tileNum;
int64_t splitIdx = 0;
uint64_t splitStart = 0;
uint64_t splitLength = static_cast<uint64_t>(sizeSplits[splitIdx]);
while (firstMidIdx >= splitStart + splitLength) {
splitStart += splitLength;
++splitIdx;
splitLength = static_cast<uint64_t>(sizeSplits[splitIdx]);
}
__gm__ T* dstPtr = outList.template GetDataPtr<__gm__ T>(splitIdx);
outputGm.SetGlobalBuffer(dstPtr, outerLength * splitLength * innerLength);
uint64_t prevOuterIdx = firstOuterIdx;
for (uint64_t i = 0; i < loopNum; ++i) {
const uint64_t taskIdx = loopOffset + i;
const uint64_t outerIdx = taskIdx / midTileTotal;
const uint64_t midIdx = (taskIdx % midTileTotal) / tileNum;
const uint32_t tileIdx = static_cast<uint32_t>(taskIdx % tileNum);
if (outerIdx != prevOuterIdx) {
splitStart = 0;
splitIdx = 0;
splitLength = static_cast<uint64_t>(sizeSplits[0]);
dstPtr = outList.template GetDataPtr<__gm__ T>(0);
outputGm.SetGlobalBuffer(dstPtr, outerLength * splitLength * innerLength);
prevOuterIdx = outerIdx;
}
while (midIdx >= splitStart + splitLength) {
splitStart += splitLength;
++splitIdx;
splitLength = static_cast<uint64_t>(sizeSplits[splitIdx]);
dstPtr = outList.template GetDataPtr<__gm__ T>(splitIdx);
outputGm.SetGlobalBuffer(dstPtr, outerLength * splitLength * innerLength);
}
const uint64_t srcOff = outerIdx * midLength * innerLength + midIdx * innerLength +
static_cast<uint64_t>(tileIdx) * tileLength;
const uint64_t dstOff = outerIdx * splitLength * innerLength + (midIdx - splitStart) * innerLength +
static_cast<uint64_t>(tileIdx) * tileLength;
const uint32_t length = (tileIdx == tileNum - 1) ? lastTileLength : tileLength;
const uint32_t alignedLength = (length + alignedNum - 1) / alignedNum * alignedNum;
CopyIn(srcOff, length, alignedLength);
CopyOut(dstOff, length, alignedLength);
}
}
template <typename T>
__aicore__ inline void SplitVGeneral<T>::CopyIn(uint64_t offset, uint32_t length, uint32_t alignedLength)
{
SplitVQueueCopyIn(queue, inputGm, offset, length, alignedLength);
}
template <typename T>
__aicore__ inline void SplitVGeneral<T>::CopyOut(uint64_t offset, uint32_t length, uint32_t alignedLength)
{
SplitVQueueCopyOut(queue, outputGm, offset, length, alignedLength);
}
#endif