* 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 diag_v2_base.h
* \brief
*/
#ifndef DIAG_V2_BASE_H
#define DIAG_V2_BASE_H
#include "diag_v2_assist.h"
namespace DiagV2 {
using namespace AscendC;
__aicore__ inline constexpr bool IsDataCopyPadSupport()
{
#if __CCE_AICORE__ == 220 || __CCE_AICORE__ == 310 || (defined(__NPU_ARCH__) && __NPU_ARCH__ == 3003)
return true;
#else
return false;
#endif
}
enum DiagV2TilingKey
{
ASSIST_SIZE_1 = 2101,
ASSIST_SIZE_2 = 2102,
ASSIST_SIZE_4 = 2103,
ASSIST_SIZE_8 = 2104,
ASSIST_SIZE_16 = 2105,
SCALAR_SIZE_1 = 2401,
SCALAR_SIZE_2 = 2402,
SCALAR_SIZE_4 = 2403,
SCALAR_SIZE_8 = 2404
};
template <typename T>
class DiagV2Base
{
public:
__aicore__ inline DiagV2Base(){};
protected:
template <typename T1, typename T2>
__aicore__ inline T1 CeilDiv(T1 a, T2 b)
{
if (b == 0) {
return 0;
}
return (a + b - 1) / b;
};
template <typename T1, typename T2>
__aicore__ inline void LocalTensor2NewTensor(LocalTensor<T1>& tensor_new, const LocalTensor<T2>& tensor_old)
{
tensor_new = tensor_old.template ReinterpretCast<T1>();
};
__aicore__ inline void ParseTilingData(const DiagV2TilingData* tilingData, DiagV2TilingData& m_tilingData);
__aicore__ inline void CopyIn(
LocalTensor<T>& xLocal, const GlobalTensor<T>& xGm, const DiagV2TilingData& m_tilingData, const int64_t& index,
const int64_t& dataCount);
__aicore__ inline void CopyInNotAlign(
LocalTensor<T>& xLocal, const GlobalTensor<T>& xGm, const DiagV2TilingData& m_tilingData, const int64_t& index,
const int64_t& dataCount);
__aicore__ inline void CopyInWithPad(
LocalTensor<T>& xLocal, const GlobalTensor<T>& xGm, const DiagV2TilingData& m_tilingData, const int64_t& index,
const int64_t& dataCount);
private:
constexpr static int32_t blockSize = 32;
constexpr static int32_t maxBlockStride = 65535;
int32_t perBlockNum = blockSize / sizeof(T);
};
template <typename T>
__aicore__ inline void DiagV2Base<T>::ParseTilingData(
const DiagV2TilingData* tilingData, DiagV2TilingData& m_tilingData)
{
m_tilingData.xWidth = tilingData->xWidth;
m_tilingData.xHeight = tilingData->xHeight;
m_tilingData.gmOffset = tilingData->gmOffset;
m_tilingData.numOut = tilingData->numOut;
m_tilingData.realCoreNum = tilingData->realCoreNum;
m_tilingData.numPerCore = tilingData->numPerCore;
m_tilingData.tailNum = tilingData->tailNum;
m_tilingData.tilingKey = tilingData->tilingKey;
m_tilingData.matrixRowLength = tilingData->matrixRowLength;
}
template <typename T>
__aicore__ inline void DiagV2Base<T>::CopyInWithPad(
LocalTensor<T>& xLocal, const GlobalTensor<T>& xGm, const DiagV2TilingData& m_tilingData, const int64_t& index,
const int64_t& dataCount)
{
int64_t blockOffset =
(GetBlockIdx() * m_tilingData.numPerCore + index * m_tilingData.matrixRowLength) * (m_tilingData.xWidth + 1) +
m_tilingData.gmOffset;
for (int64_t idx = 0; idx < dataCount; idx++) {
DataCopyExtParams copyParams = {1, 0, 0, 0, 0};
copyParams.blockLen = dataCount * sizeof(T);
DataCopyPadExtParams<T> padParams = {false, 0, 0, 0};
DataCopyPad(
xLocal[idx * m_tilingData.matrixRowLength], xGm[blockOffset + idx * m_tilingData.xWidth], copyParams,
padParams);
}
}
template <typename T>
__aicore__ inline void DiagV2Base<T>::CopyInNotAlign(
LocalTensor<T>& xLocal, const GlobalTensor<T>& xGm, const DiagV2TilingData& m_tilingData, const int64_t& index,
const int64_t& dataCount)
{
int64_t loopTailAlign = this->CeilDiv(dataCount, perBlockNum) * perBlockNum;
int64_t blockOffset =
(GetBlockIdx() * m_tilingData.numPerCore + index * m_tilingData.matrixRowLength - (loopTailAlign - dataCount)) *
(m_tilingData.xWidth + 1) +
m_tilingData.gmOffset;
for (int64_t idx = 0; idx < loopTailAlign; idx++) {
DataCopy(
xLocal[idx * m_tilingData.matrixRowLength], xGm[blockOffset + idx * m_tilingData.xWidth], loopTailAlign);
}
}
template <typename T>
__aicore__ inline void DiagV2Base<T>::CopyIn(
LocalTensor<T>& xLocal, const GlobalTensor<T>& xGm, const DiagV2TilingData& m_tilingData, const int64_t& index,
const int64_t& dataCount)
{
int64_t blockOffset =
(GetBlockIdx() * m_tilingData.numPerCore + index * m_tilingData.matrixRowLength) * (m_tilingData.xWidth + 1) +
m_tilingData.gmOffset;
int64_t loopTailAlign = this->CeilDiv(dataCount, perBlockNum) * perBlockNum;
if ((m_tilingData.xWidth - loopTailAlign) % perBlockNum == 0 &&
(m_tilingData.xWidth - loopTailAlign) / perBlockNum <= maxBlockStride) {
int64_t srcStride = (m_tilingData.xWidth - loopTailAlign) / perBlockNum;
int64_t dstStride = (m_tilingData.matrixRowLength - loopTailAlign) / perBlockNum;
DataCopyParams intriParams;
intriParams.blockCount = dataCount;
intriParams.blockLen = loopTailAlign / perBlockNum;
intriParams.srcStride = srcStride;
intriParams.dstStride = dstStride;
DataCopy(xLocal, xGm[blockOffset], intriParams);
} else {
for (int64_t idx = 0; idx < dataCount; idx++) {
DataCopy(
xLocal[idx * m_tilingData.matrixRowLength], xGm[blockOffset + idx * m_tilingData.xWidth],
loopTailAlign);
}
}
}
}
#endif