* 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 kernel_operator_gemm_base_impl.h
* \brief
*/
#if !defined(__ASCENDC_INCLUDE_INTERNAL_HEADERS__)
#pragma message("impl/basic_api/kernel_operator_gemm_base_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 \"basic_api/kernel_operator_gemm_intf.h\"\" and use public functions or variables defined in interface headers files.")
#define __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#define __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_GEMM_BASE_IMPL_H__
#endif
#ifndef ASCENDC_MODULE_OPERATOR_GEMM_BASE_IMPL_H
#define ASCENDC_MODULE_OPERATOR_GEMM_BASE_IMPL_H
#if ASCENDC_CPU_DEBUG
#include <unordered_set>
#endif
#include "kernel_tensor.h"
#include "kernel_operator_mm_base_impl.h"
#include "kernel_struct_conv2d.h"
#include "kernel_struct_mm.h"
#include "kernel_tpipe.h"
#include "kernel_operator_block_sync_intf.h"
namespace AscendC {
#if ASCENDC_CPU_DEBUG
const std::unordered_set<std::string> MATMUL_SUPPORT_TYPE { "s8s8s32", "f16f16f32", "f16f16f16" };
template <typename T> __aicore__ inline std::string GetTypeStr(const LocalTensor<T>& input)
{
if (std::is_same<PrimT<T>, uint8_t>::value) {
return "u8";
} else if (std::is_same<PrimT<T>, int8_t>::value) {
return "s8";
} else if (std::is_same<PrimT<T>, half>::value) {
return "f16";
} else if (std::is_same<PrimT<T>, float>::value) {
return "f32";
} else if (std::is_same<PrimT<T>, int32_t>::value) {
return "s32";
} else {
return "None";
}
}
__aicore__ inline bool CheckRange(std::pair<uint32_t, uint32_t>& range, const uint32_t num)
{
if (num < range.first || num > range.second) {
return false;
} else {
return true;
}
}
template <typename T, typename U, typename S>
__aicore__ inline bool CheckOverflow(const LocalTensor<T>& dst, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, const uint32_t m, const uint32_t k, const uint32_t n, GemmTiling& tilling)
{
uint32_t roundM = DivCeil(m, tilling.blockSize) * tilling.blockSize;
uint32_t roundN = DivCeil(n, tilling.blockSize) * tilling.blockSize;
uint32_t roundK = DivCeil(k, tilling.c0Size) * tilling.c0Size;
uint32_t needElementLoc = roundM * roundN * sizeof(uint32_t);
if (needElementLoc > TOTAL_L0C_SIZE) {
return false;
}
uint32_t needElementL0a = roundM * roundK * sizeof(PrimT<U>);
uint32_t needElementL0b = roundN * roundK * sizeof(PrimT<S>);
if ((needElementL0b + needElementL0a) > TOTAL_L1_SIZE) {
return false;
}
return true;
}
template <typename T, typename U, typename S>
__aicore__ inline bool CheckParams(const LocalTensor<T>& dst, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, const uint32_t m, const uint32_t k, const uint32_t n, GemmTiling& tilling)
{
if (tilling.c0Size != 16 && tilling.c0Size != 32) {
return false;
}
const Hardware src0Scope = GetPhyType((TPosition)src0.GetPosition());
const Hardware src1Scope = GetPhyType((TPosition)src1.GetPosition());
const Hardware dstScope = GetPhyType((TPosition)dst.GetPosition());
if (src0Scope != Hardware::L1 || src1Scope != Hardware::L1) {
return false;
}
if (dstScope != Hardware::UB && dstScope != Hardware::L0C) {
return false;
}
std::string dtypeStr = GetTypeStr(src0) + GetTypeStr(src1) + GetTypeStr(dst);
if (MATMUL_SUPPORT_TYPE.find(dtypeStr) == MATMUL_SUPPORT_TYPE.end()) {
return false;
}
std::pair<uint32_t, uint32_t> mRange(1, 4096);
std::pair<uint32_t, uint32_t> nRange = mRange;
std::pair<uint32_t, uint32_t> kRange;
if (std::is_same<U, half>::value) {
kRange = std::make_pair(1, 32768);
} else {
kRange = std::make_pair(1, 16384);
}
if (!CheckRange(mRange, m) || !CheckRange(nRange, n) || !CheckRange(kRange, k)) {
return false;
}
if (!CheckOverflow(dst, src0, src1, m, k, n, tilling)) {
return false;
}
return true;
}
#endif
__aicore__ inline void CalculateGemmTiling(GemmTiling& tilling)
{
tilling.mIterNum = 1;
tilling.nIterNum = 1;
tilling.kIterNum = DivCeil(tilling.kBlockNum, tilling.kTileBlock);
tilling.mTileBlock = DivCeil(tilling.mBlockNum, tilling.mIterNum);
tilling.nTileBlock = DivCeil(tilling.nBlockNum, tilling.nIterNum);
tilling.kTailBlock = tilling.kBlockNum - (tilling.kIterNum - 1) * tilling.kTileBlock;
tilling.mTailBlock = tilling.mBlockNum - (tilling.mIterNum - 1) * tilling.mTileBlock;
tilling.nTailBlock = tilling.nBlockNum - (tilling.nIterNum - 1) * tilling.nTileBlock;
tilling.kHasTail = tilling.kTailBlock != tilling.kTileBlock;
tilling.kHasTailEle = tilling.roundK != tilling.kNum;
tilling.kTailEle = tilling.kNum % (tilling.kTileBlock * tilling.c0Size);
if (tilling.mNum != tilling.mTileBlock * tilling.blockSize) {
tilling.mHasTail = true;
} else {
tilling.mHasTail = false;
}
tilling.nHasTail = tilling.nTileBlock != tilling.nTailBlock;
}
template <typename T>
__aicore__ inline void LoadL0B(uint32_t kBlocks, uint32_t nBlocks, GemmTiling tilling, uint32_t i, uint32_t j,
const LocalTensor<T>& src1, const LocalTensor<T>& l0b)
{
if (tilling.nIterNum == 1) {
uint32_t wSize = tilling.blockSize * tilling.c0Size;
uint32_t wIdx = (i * tilling.kTileBlock * tilling.nBlockNum + j * tilling.nTileBlock) * wSize;
LoadData2DParams params;
params.startIndex = 0;
params.repeatTimes = kBlocks * nBlocks;
params.srcStride = 1;
LoadDataImpl(l0b, src1[wIdx], params);
} else {
for (size_t index = 0; index < kBlocks; ++index) {
uint32_t wSize = j * tilling.nTileBlock * tilling.blockSize * tilling.c0Size;
uint32_t wIdx =
(i * tilling.kTileBlock + index) * tilling.nBlockNum * tilling.blockSize * tilling.c0Size + wSize;
uint32_t l0bIdx = index * nBlocks * tilling.blockSize * tilling.c0Size;
LoadData2DParams params;
params.startIndex = 0;
params.repeatTimes = nBlocks;
params.srcStride = 1;
LoadDataImpl(l0b[l0bIdx], src1[wIdx], params);
}
}
}
template <typename T>
__aicore__ inline void LoadL0A(uint32_t kBlocks, uint32_t mBlocks, GemmTiling tilling, uint32_t i, uint32_t t,
const LocalTensor<T>& src0, const LocalTensor<T>& l0a)
{
if (kBlocks == 1) {
uint32_t l1aSize = i * tilling.kTileBlock * tilling.mBlockNum * tilling.blockSize * tilling.c0Size;
uint32_t l1aOffset = t * tilling.mTileBlock * tilling.blockSize * tilling.c0Size + l1aSize;
LoadData2DParams params;
params.startIndex = 0;
params.repeatTimes = mBlocks;
params.srcStride = 1;
LoadDataImpl(l0a, src0[l1aOffset], params);
} else {
for (size_t index = 0; index < mBlocks; index++) {
uint32_t l0aOffset = index * kBlocks * tilling.blockSize * tilling.c0Size;
uint32_t l1aOffset = (t * tilling.mTileBlock + index) * tilling.blockSize * tilling.c0Size +
i * tilling.kTileBlock * tilling.mBlockNum * tilling.blockSize * tilling.c0Size;
LoadData2DParams params;
params.startIndex = 0;
params.repeatTimes = kBlocks;
params.srcStride = tilling.mBlockNum;
LoadDataImpl(l0a[l0aOffset], src0[l1aOffset], params);
}
}
}
template <typename T, typename U, typename S>
__aicore__ inline void MmadFunc(const LocalTensor<U>& l0a, const LocalTensor<S>& l0b,
const LocalTensor<T>& l0c, int32_t initValue, GemmTiling tilling, size_t i)
{
MmadParams mmadParams;
mmadParams.m = tilling.mTileBlock * tilling.blockSize;
mmadParams.n = tilling.nTileBlock * tilling.blockSize;
mmadParams.isBias = 1;
if (tilling.kIterNum == 1) {
mmadParams.k = tilling.kNum;
mmadParams.isBias = initValue;
} else if (initValue == 1 && tilling.kHasTailEle) {
if (i == tilling.kIterNum - 1) {
mmadParams.k = tilling.kTailEle;
} else {
mmadParams.k = tilling.kTileBlock * tilling.c0Size;
}
} else if (initValue != 1 && tilling.kHasTailEle) {
if (i == 0) {
mmadParams.k = tilling.kTileBlock * tilling.c0Size;
mmadParams.isBias = 0;
} else if (i == tilling.kIterNum - 1) {
mmadParams.k = tilling.kTailEle;
} else {
mmadParams.k = tilling.kTileBlock * tilling.c0Size;
}
} else if (initValue == 1 && !tilling.kHasTailEle) {
if (i == tilling.kIterNum - 1) {
mmadParams.k = tilling.kTailBlock * tilling.c0Size;
} else {
mmadParams.k = tilling.kTileBlock * tilling.c0Size;
}
} else {
if (i == 0) {
mmadParams.k = tilling.kTileBlock * tilling.c0Size;
mmadParams.isBias = 0;
} else if (i == tilling.kIterNum - 1) {
mmadParams.k = tilling.kTailBlock * tilling.c0Size;
} else {
mmadParams.k = tilling.kTileBlock * tilling.c0Size;
}
}
MmadImpl(l0c, l0a, l0b, mmadParams);
}
template <typename T, typename U>
__aicore__ inline void GetPingPongBuffer(LocalTensor<T>& l0aPing, LocalTensor<T>& l0aPong,
LocalTensor<U>& l0bPing, LocalTensor<U>& l0bPong)
{
TBuffAddr tbufaPing;
tbufaPing.logicPos = static_cast<uint8_t>(TPosition::A2);
l0aPing.SetAddr(tbufaPing);
l0aPing.InitBuffer(0, TOTAL_L0A_SIZE / 2 / sizeof(PrimT<T>));
TBuffAddr tbufaPong;
tbufaPong.logicPos = static_cast<uint8_t>(TPosition::A2);
l0aPong.SetAddr(tbufaPong);
l0aPong.InitBuffer(TOTAL_L0A_SIZE / 2, TOTAL_L0A_SIZE / 2 / sizeof(PrimT<T>));
TBuffAddr tbufbPing;
tbufbPing.logicPos = static_cast<uint8_t>(TPosition::B2);
l0bPing.SetAddr(tbufbPing);
l0bPing.InitBuffer(0, TOTAL_L0B_SIZE / 2 / sizeof(PrimT<U>));
TBuffAddr tbufbPong;
tbufbPong.logicPos = static_cast<uint8_t>(TPosition::B2);
l0bPong.SetAddr(tbufbPong);
l0bPong.InitBuffer(TOTAL_L0B_SIZE / 2, TOTAL_L0B_SIZE / 2 / sizeof(PrimT<U>));
return;
}
template <typename T, typename U>
__aicore__ inline void GetSingleThreadBuffer(LocalTensor<T>& l0a, LocalTensor<U>& l0b)
{
TBuffAddr tbufa;
tbufa.logicPos = static_cast<uint8_t>(TPosition::A2);
l0a.SetAddr(tbufa);
l0a.InitBuffer(0, TOTAL_L0A_SIZE / sizeof(PrimT<T>));
TBuffAddr tbufb;
tbufb.logicPos = static_cast<uint8_t>(TPosition::B2);
l0b.SetAddr(tbufb);
l0b.InitBuffer(0, TOTAL_L0B_SIZE / sizeof(PrimT<U>));
return;
}
template <typename T, typename U, typename S>
__aicore__ inline void GemmExecNmNopingpong(const LocalTensor<T>& l0c, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, GemmTiling tilling, const int32_t initValue)
{
LocalTensor<U> l0a;
LocalTensor<S> l0b;
GetSingleThreadBuffer(l0a, l0b);
event_t eventIdMToMte1 = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::M_MTE1));
SetFlag<HardEvent::M_MTE1>(eventIdMToMte1);
for (size_t indexK = 0; indexK < tilling.kIterNum; indexK++) {
uint32_t kBlocks = tilling.kTileBlock;
if (indexK == tilling.kIterNum - 1) {
kBlocks = tilling.kTailBlock;
}
WaitFlag<HardEvent::M_MTE1>(eventIdMToMte1);
for (size_t indexN = 0; indexN < tilling.nIterNum; indexN++) {
LoadL0B(kBlocks, tilling.nTileBlock, tilling, indexK, indexN, src1, l0b);
for (size_t indexM = 0; indexM < tilling.mIterNum; indexM++) {
LoadL0A(kBlocks, tilling.mTileBlock, tilling, indexK, indexM, src0, l0a);
event_t eventIdMte1ToM = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE1_M));
SetFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
WaitFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
PipeBarrier<PIPE_M>();
MmadFunc(l0a, l0b, l0c, initValue, tilling, indexK);
}
}
SetFlag<HardEvent::M_MTE1>(eventIdMToMte1);
}
WaitFlag<HardEvent::M_MTE1>(eventIdMToMte1);
}
template <typename T, typename U, typename S>
__aicore__ inline void GemmExecNmPingPong(const LocalTensor<T>& l0c, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, GemmTiling tilling, const int32_t initValue)
{
uint32_t ping = 1;
LocalTensor<U> l0aPing;
LocalTensor<U> l0aPong;
LocalTensor<S> l0bPing;
LocalTensor<S> l0bPong;
GetPingPongBuffer(l0aPing, l0aPong, l0bPing, l0bPong);
event_t eventId0 = static_cast<event_t>(GetTPipePtr()->AllocEventID<HardEvent::M_MTE1>());
event_t eventId1 = static_cast<event_t>(GetTPipePtr()->AllocEventID<HardEvent::M_MTE1>());
SetFlag<HardEvent::M_MTE1>(eventId0);
SetFlag<HardEvent::M_MTE1>(eventId1);
for (size_t i = 0; i < tilling.kIterNum; i++) {
uint32_t kBlocks = tilling.kTileBlock;
if (i == tilling.kIterNum - 1) {
kBlocks = tilling.kTailBlock;
}
if (ping == 1) {
WaitFlag<HardEvent::M_MTE1>(eventId0);
for (size_t indexN = 0; indexN < tilling.nIterNum; indexN++) {
LoadL0B(kBlocks, tilling.nTileBlock, tilling, i, indexN, src1, l0bPing);
for (size_t indexM = 0; indexM < tilling.mIterNum; indexM++) {
LoadL0A(kBlocks, tilling.mTileBlock, tilling, i, indexM, src0, l0aPing);
event_t eventIdMte1ToM = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE1_M));
SetFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
WaitFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
PipeBarrier<PIPE_M>();
MmadFunc(l0aPing, l0bPing, l0c, initValue, tilling, i);
}
}
SetFlag<HardEvent::M_MTE1>(eventId0);
} else {
WaitFlag<HardEvent::M_MTE1>(eventId1);
for (size_t indexN = 0; indexN < tilling.nIterNum; indexN++) {
LoadL0B(kBlocks, tilling.nTileBlock, tilling, i, indexN, src1, l0bPong);
for (size_t indexM = 0; indexM < tilling.mIterNum; indexM++) {
LoadL0A(kBlocks, tilling.mTileBlock, tilling, i, indexM, src0, l0aPong);
event_t eventIdMte1ToM = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE1_M));
SetFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
WaitFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
PipeBarrier<PIPE_M>();
MmadFunc(l0aPong, l0bPong, l0c, initValue, tilling, i);
}
}
SetFlag<HardEvent::M_MTE1>(eventId1);
}
ping = 1 - ping;
}
#if __NPU_ARCH__ == 2201
WaitFlag<HardEvent::M_MTE1>(eventId0);
GetTPipePtr()->ReleaseEventID<HardEvent::M_MTE1>(eventId0);
WaitFlag<HardEvent::M_MTE1>(eventId1);
GetTPipePtr()->ReleaseEventID<HardEvent::M_MTE1>(eventId1);
#else
WaitFlag<HardEvent::M_MTE1>(eventId0);
WaitFlag<HardEvent::M_MTE1>(eventId1);
#endif
}
template <typename T, typename U, typename S>
__aicore__ inline void GemmExecNm(const LocalTensor<T>& l0c, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, GemmTiling tilling, const int32_t initValue)
{
uint32_t needL0Asize = tilling.roundM * tilling.dtypeSize * tilling.c0Size * tilling.kTileBlock * 2;
uint32_t needL0Bsize = tilling.roundN * tilling.dtypeSize * tilling.c0Size * tilling.kTileBlock * 2;
if (needL0Asize > TOTAL_L0A_SIZE || needL0Bsize > TOTAL_L0B_SIZE) {
GemmExecNmNopingpong(l0c, src0, src1, tilling, initValue);
return;
}
GemmExecNmPingPong(l0c, src0, src1, tilling, initValue);
}
template <typename T, typename U, typename S>
__aicore__ inline void GemmExecMnNopingpong(const LocalTensor<T>& l0c, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, GemmTiling tilling, const int32_t initValue)
{
LocalTensor<S> l0b;
LocalTensor<U> l0a;
GetSingleThreadBuffer(l0a, l0b);
event_t eventIdMToMte1 = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::M_MTE1));
SetFlag<HardEvent::M_MTE1>(eventIdMToMte1);
for (size_t indexK = 0; indexK < tilling.kIterNum; indexK++) {
uint32_t kBlocks = tilling.kTileBlock;
if (indexK == tilling.kIterNum - 1) {
kBlocks = tilling.kTailBlock;
}
WaitFlag<HardEvent::M_MTE1>(eventIdMToMte1);
for (size_t indexM = 0; indexM < tilling.mIterNum; indexM++) {
LoadL0A(kBlocks, tilling.mTileBlock, tilling, indexK, indexM, src0, l0a);
for (size_t indexN = 0; indexN < tilling.nIterNum; indexN++) {
LoadL0B(kBlocks, tilling.nTileBlock, tilling, indexK, indexN, src1, l0b);
event_t eventIdMte1ToM = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE1_M));
SetFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
WaitFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
PipeBarrier<PIPE_M>();
MmadFunc(l0a, l0b, l0c, initValue, tilling, indexK);
}
}
SetFlag<HardEvent::M_MTE1>(eventIdMToMte1);
}
WaitFlag<HardEvent::M_MTE1>(eventIdMToMte1);
}
template <typename T, typename U, typename S>
__aicore__ inline void GemmExecMnPingPong(const LocalTensor<T>& l0c, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, GemmTiling tilling, const int32_t initValue)
{
uint32_t ping = 1;
LocalTensor<U> l0aPing;
LocalTensor<U> l0aPong;
LocalTensor<S> l0bPing;
LocalTensor<S> l0bPong;
GetPingPongBuffer(l0aPing, l0aPong, l0bPing, l0bPong);
event_t eventId0 = static_cast<event_t>(GetTPipePtr()->AllocEventID<HardEvent::M_MTE1>());
event_t eventId1 = static_cast<event_t>(GetTPipePtr()->AllocEventID<HardEvent::M_MTE1>());
SetFlag<HardEvent::M_MTE1>(eventId0);
SetFlag<HardEvent::M_MTE1>(eventId1);
for (size_t i = 0; i < tilling.kIterNum; i++) {
uint32_t kBlocks = tilling.kTileBlock;
if (i == tilling.kIterNum - 1) {
kBlocks = tilling.kTailBlock;
}
if (ping == 1) {
WaitFlag<HardEvent::M_MTE1>(eventId0);
for (size_t indexM = 0; indexM < tilling.mIterNum; indexM++) {
LoadL0A(kBlocks, tilling.mTileBlock, tilling, i, indexM, src0, l0aPing);
for (size_t indexN = 0; indexN < tilling.nIterNum; indexN++) {
LoadL0B(kBlocks, tilling.nTileBlock, tilling, i, indexN, src1, l0bPing);
event_t eventIdMte1ToM = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE1_M));
SetFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
WaitFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
PipeBarrier<PIPE_M>();
MmadFunc(l0aPing, l0bPing, l0c, initValue, tilling, i);
}
}
SetFlag<HardEvent::M_MTE1>(eventId0);
} else {
WaitFlag<HardEvent::M_MTE1>(eventId1);
for (size_t indexM = 0; indexM < tilling.mIterNum; indexM++) {
LoadL0A(kBlocks, tilling.mTileBlock, tilling, i, indexM, src0, l0aPong);
for (size_t indexN = 0; indexN < tilling.nIterNum; indexN++) {
LoadL0B(kBlocks, tilling.nTileBlock, tilling, i, indexN, src1, l0bPong);
event_t eventIdMte1ToM = static_cast<event_t>(GetTPipePtr()->FetchEventID(HardEvent::MTE1_M));
SetFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
WaitFlag<HardEvent::MTE1_M>(eventIdMte1ToM);
PipeBarrier<PIPE_M>();
MmadFunc(l0aPong, l0bPong, l0c, initValue, tilling, i);
}
}
SetFlag<HardEvent::M_MTE1>(eventId1);
}
ping = 1 - ping;
}
WaitFlag<HardEvent::M_MTE1>(eventId0);
GetTPipePtr()->ReleaseEventID<HardEvent::M_MTE1>(eventId0);
WaitFlag<HardEvent::M_MTE1>(eventId1);
GetTPipePtr()->ReleaseEventID<HardEvent::M_MTE1>(eventId1);
}
template <typename T, typename U, typename S>
__aicore__ inline void GemmExecMn(const LocalTensor<T>& l0c, const LocalTensor<U>& src0,
const LocalTensor<S>& src1, GemmTiling tilling, const int32_t initValue)
{
uint32_t needL0Bsize = tilling.roundN * tilling.dtypeSize * tilling.c0Size * tilling.kTileBlock * 2;
uint32_t needL0Asize = tilling.roundM * tilling.dtypeSize * tilling.c0Size * tilling.kTileBlock * 2;
if (needL0Asize > TOTAL_L0A_SIZE || needL0Bsize > TOTAL_L0B_SIZE) {
GemmExecMnNopingpong(l0c, src0, src1, tilling, initValue);
return;
}
GemmExecMnPingPong(l0c, src0, src1, tilling, initValue);
}
}
#endif
#if defined(__UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_GEMM_BASE_IMPL_H__)
#undef __ASCENDC_INCLUDE_INTERNAL_HEADERS__
#undef __UNDEF_ASCENDC_INCLUDE_INTERNAL_HEADERS_KERNEL_OPERATOR_GEMM_BASE_IMPL_H__
#endif