* 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.
*/
* \file kernel_matmul_streamk.h
* \brief
*/
#pragma once
#include "kernel_universal.h"
#define ASCENDC_CUBE_ONLY
#if ASC_DEVKIT_MAJOR >= 9
#include "kernel_basic_intf.h"
#else
#include "kernel_operator.h"
#endif
#include "lib/matmul_intf.h"
#include "blaze/epilogue/block/block_epilogue_matmul_streamk.h"
#include "blaze/gemm/block/block_mmad_matmul_streamk.h"
#include "blaze/gemm/utils/common_utils.h"
#include "tensor_api/tensor.h"
namespace Blaze {
namespace Gemm {
namespace Kernel {
template <class ProblemShape_, class BlockMmad_, class BlockEpilogue_, class BlockScheduler_>
class GemmUniversal<
ProblemShape_, BlockMmad_, BlockEpilogue_, BlockScheduler_,
AscendC::Std::enable_if_t<
AscendC::Std::is_same_v<KernelMultiBlockStreamK, typename BlockMmad_::DispatchPolicy::ScheduleType> &&
AscendC::Std::is_same_v<KernelMultiBlockStreamK, typename BlockEpilogue_::DispatchPolicy::ScheduleType>>> {
public:
__aicore__ inline GemmUniversal()
{}
__aicore__ inline ~GemmUniversal()
{}
using BlockMmadOp = BlockMmad_;
using ProblemShape = ProblemShape_;
using BlockScheduler = BlockScheduler_;
using BlockEpilogue = BlockEpilogue_;
using BlockMmadParams = typename BlockMmadOp::GmParams;
using BlockEpilogueParams = typename BlockEpilogue::Params;
using BlockSchedulerParams = typename BlockScheduler::Params;
using AType = typename BlockMmadOp::AType;
using BType = typename BlockMmadOp::BType;
using CType = typename BlockMmadOp::CType;
using BiasType = typename BlockMmadOp::BiasType;
using LayoutA = typename BlockMmadOp::LayoutA;
using LayoutB = typename BlockMmadOp::LayoutB;
using LayoutC = typename BlockMmadOp::LayoutC;
using LayoutBias = typename BlockMmadOp::LayoutBias;
using TupleShape = AscendC::Te::Shape<int64_t, int64_t, int64_t, int64_t>;
__gm__ AType* aGmAddr_;
__gm__ BType* bGmAddr_;
__gm__ CType* cGmAddr_;
__gm__ BiasType* biasGmAddr_;
__gm__ float* workspaceGmAddr_;
using MakeLayoutA = AscendC::Te::FrameLayoutFormat<LayoutA, AscendC::Te::LayoutTraitDefault<AType>>;
using MakeLayoutB = AscendC::Te::FrameLayoutFormat<LayoutB, AscendC::Te::LayoutTraitDefault<BType>>;
using MakeLayoutC = AscendC::Te::FrameLayoutFormat<LayoutC, AscendC::Te::LayoutTraitDefault<CType>>;
using MakeLayoutBias = AscendC::Te::FrameLayoutFormat<LayoutBias, AscendC::Te::LayoutTraitDefault<BiasType>>;
int64_t m_ = 0;
int64_t n_ = 0;
int64_t k_ = 0;
int64_t usedCoreNum_ = 0;
TupleShape problemShape_{};
bool isBias_ = false;
constexpr static uint16_t NUM_TWO = 2;
constexpr static uint16_t AIC_SYNC_AIV_MODE_4 = 4;
constexpr static uint16_t AIV_SYNC_AIC_FLAG = 6;
constexpr static uint16_t AIC_SYNC_AIV_FLAG = 8;
constexpr static uint16_t FLAG_ID_MAX = 16;
constexpr static uint16_t BLOCK_BASE_M = 256;
constexpr static uint16_t BLOCK_BASE_N = 256;
struct Params {
ProblemShape problemShape;
BlockMmadParams mmadParams;
BlockEpilogueParams epilogueParams;
BlockSchedulerParams schParams;
Params() = default;
};
__aicore__ inline void Init(Params const& params)
{
problemShape_ = params.problemShape;
BlockMmadParams blockMmadParams_ = params.mmadParams;
BlockEpilogueParams blockEpilogueParams_ = params.epilogueParams;
m_ = Get<MNK_M>(problemShape_);
n_ = Get<MNK_N>(problemShape_);
k_ = Get<MNK_K>(problemShape_);
usedCoreNum_ = params.schParams.usedCoreNum;
aGmAddr_ = reinterpret_cast<__gm__ AType*>(blockMmadParams_.aGmAddr);
bGmAddr_ = reinterpret_cast<__gm__ BType*>(blockMmadParams_.bGmAddr);
cGmAddr_ = reinterpret_cast<__gm__ CType*>(blockMmadParams_.cGmAddr);
workspaceGmAddr_ = reinterpret_cast<__gm__ float*>(blockMmadParams_.workspaceGmAddr);
if (blockMmadParams_.biasGmAddr != nullptr) {
isBias_ = true;
biasGmAddr_ = reinterpret_cast<__gm__ BiasType*>(blockMmadParams_.biasGmAddr);
}
}
__aicore__ inline void operator()(Params const& params)
{
Init(params);
if (usedCoreNum_ <= 0) {
return;
}
BlockScheduler bs(params.problemShape, params.schParams);
TupleShape tileL1 = bs.GetTileL1Shape();
int64_t mL1 = Get<MNK_M>(tileL1);
int64_t nL1 = Get<MNK_N>(tileL1);
int64_t kL1 = Get<MNK_K>(tileL1);
int64_t mTileNum = Get<MNK_M>(bs.GetMNKTileNum());
int64_t nTileNum = Get<MNK_N>(bs.GetMNKTileNum());
int64_t skKTileNum = Get<MNK_K>(bs.GetMNKTileNum());
int64_t tileNum = bs.GetTotalTileNum();
if ASCEND_IS_AIC {
BlockMmadOp blockMmadOp;
int64_t curBlockIdx = AscendC::GetBlockIdx();
TupleShape tileL0 = bs.GetTileL0Shape();
int64_t isHf32 = bs.GetHf32Flag();
if (curBlockIdx >= bs.GetBlockNum(usedCoreNum_)) {
CrossCoreSetFlag<AIC_SYNC_AIV_MODE_4, PIPE_FIX>(AIC_SYNC_AIV_FLAG);
CrossCoreSetFlag<AIC_SYNC_AIV_MODE_4, PIPE_FIX>(AIC_SYNC_AIV_FLAG + FLAG_ID_MAX);
return;
}
if (isHf32) {
AscendC::SetHF32Mode(1);
AscendC::SetHF32TransMode(1);
}
SetMMLayoutTransform(true);
blockMmadOp.Init(problemShape_, tileL1, tileL0, isBias_);
int64_t tailSKTotalTileNum = static_cast<int64_t>(((mTileNum * nTileNum) % usedCoreNum_) * skKTileNum);
auto layoutA = MakeLayoutA{}(m_, k_);
auto layoutB = MakeLayoutB{}(k_, n_);
auto layoutC = MakeLayoutC{}(m_, n_);
auto layoutBias = MakeLayoutBias{}(1L, n_);
auto gmA = AscendC::Te::MakeTensor(AscendC::Te::MakeMemPtr<AscendC::Te::Location::GM>(aGmAddr_), layoutA);
auto gmB = AscendC::Te::MakeTensor(AscendC::Te::MakeMemPtr<AscendC::Te::Location::GM>(bGmAddr_), layoutB);
auto gmC = AscendC::Te::MakeTensor(AscendC::Te::MakeMemPtr<AscendC::Te::Location::GM>(cGmAddr_), layoutC);
auto gmBias =
AscendC::Te::MakeTensor(AscendC::Te::MakeMemPtr<AscendC::Te::Location::GM>(biasGmAddr_), layoutBias);
for (int64_t tileIdx = curBlockIdx; tileIdx < tileNum; tileIdx += usedCoreNum_) {
int64_t tmpTileIdx = tileIdx;
if (!bs.CheckIsSkScene(0)) {
if (tileIdx % usedCoreNum_ < tailSKTotalTileNum &&
(CeilDiv(tileIdx + 1, usedCoreNum_) == (CeilDiv(tileNum, usedCoreNum_) - 1))) {
tmpTileIdx = tileIdx + usedCoreNum_;
} else if (
tileIdx % usedCoreNum_ < tailSKTotalTileNum &&
(CeilDiv(tileIdx + 1, usedCoreNum_) == CeilDiv(tileNum, usedCoreNum_))) {
tmpTileIdx = tileIdx - usedCoreNum_;
}
}
auto singleCoreShape = bs.GetSingleCoreShape(tmpTileIdx);
auto singleCoreCoord = bs.GetSingleCoreCoord(tmpTileIdx);
int64_t kSingleCore = bs.GetCurKSingleCore(tmpTileIdx);
int64_t offsetWorkspace =
(((tmpTileIdx % usedCoreNum_) / skKTileNum) * skKTileNum + Get<MNK_K>(singleCoreCoord)) *
BLOCK_BASE_M * BLOCK_BASE_N;
auto workspaceStrideColumn0 = BlockMmadOp::DispatchPolicy::fixpOpti_ == MatMulL0C2Out::ND_FIXPIPE_1_2 ?
CeilAlign(Get<MNK_N>(singleCoreShape), BLOCK_BYTE_SIZE) :
Get<MNK_N>(singleCoreShape);
auto workspaceShape = AscendC::Te::MakeShape(
AscendC::Te::MakeShape(Std::Int<1>{}, Get<MNK_M>(singleCoreShape)),
AscendC::Te::MakeShape(Std::Int<1>{}, Get<MNK_N>(singleCoreShape)));
auto workspaceStride = AscendC::Te::MakeStride(
AscendC::Te::MakeStride(Std::Int<0>{}, workspaceStrideColumn0),
AscendC::Te::MakeStride(Std::Int<0>{}, Std::Int<1>{}));
auto layoutWorkspace =
AscendC::Te::MakePatternLayout<AscendC::Te::NDExtLayoutPtn, AscendC::Te::LayoutTraitDefault<float>>(
workspaceShape, workspaceStride);
auto gmWorkSpace = AscendC::Te::MakeTensor(
AscendC::Te::MakeMemPtr<AscendC::Te::Location::GM>(workspaceGmAddr_ + offsetWorkspace),
layoutWorkspace);
auto gmBlockA = gmA.Slice(
AscendC::Te::MakeCoord(
Get<MNK_M>(singleCoreCoord) * mL1,
Get<MNK_K>(singleCoreCoord) * kSingleCore),
AscendC::Te::MakeShape(Get<MNK_M>(singleCoreShape), Get<MNK_K>(singleCoreShape)));
auto gmBlockB = gmB.Slice(
AscendC::Te::MakeCoord(
Get<MNK_K>(singleCoreCoord) * kSingleCore, Get<MNK_N>(singleCoreCoord) * nL1),
AscendC::Te::MakeShape(Get<MNK_K>(singleCoreShape), Get<MNK_N>(singleCoreShape)));
auto gmBlockC = gmC.Slice(
AscendC::Te::MakeCoord(Get<MNK_M>(singleCoreCoord) * mL1, Get<MNK_N>(singleCoreCoord) * nL1),
AscendC::Te::MakeShape(Get<MNK_M>(singleCoreShape), Get<MNK_N>(singleCoreShape)));
auto gmBlockBias = gmBias.Slice(
AscendC::Te::MakeCoord(0L, Get<MNK_N>(singleCoreCoord) * nL1),
AscendC::Te::MakeShape(1L, Get<MNK_N>(singleCoreShape)));
blockMmadOp(
gmBlockC, gmBlockA, gmBlockB, gmBlockBias, gmWorkSpace, singleCoreShape,
Get<MNK_K>(singleCoreCoord), bs.CheckIsSkScene(tmpTileIdx));
if (tmpTileIdx + usedCoreNum_ >= tileNum) {
CrossCoreSetFlag<AIC_SYNC_AIV_MODE_4, PIPE_FIX>(AIC_SYNC_AIV_FLAG);
CrossCoreSetFlag<AIC_SYNC_AIV_MODE_4, PIPE_FIX>(AIC_SYNC_AIV_FLAG + FLAG_ID_MAX);
}
}
SetMMLayoutTransform(false);
if (isHf32) {
AscendC::SetHF32Mode(0);
}
}
if ASCEND_IS_AIV {
uint64_t lastLoopTotalCnt = (mTileNum * nTileNum % usedCoreNum_) * skKTileNum;
uint64_t curBlockIdxInAiv = AscendC::GetBlockIdx();
if (curBlockIdxInAiv >= lastLoopTotalCnt * AscendC::GetTaskRation()) {
CrossCoreWaitFlag<AIC_SYNC_AIV_MODE_4, PIPE_MTE3>(AIC_SYNC_AIV_FLAG);
SyncAll();
return;
}
CrossCoreWaitFlag<AIC_SYNC_AIV_MODE_4, PIPE_MTE3>(AIC_SYNC_AIV_FLAG);
SyncAll();
BlockEpilogue epilogueOp;
epilogueOp.Init(
params.epilogueParams, problemShape_, tileL1, bs.GetMNKTileNum(), usedCoreNum_, bs.CheckIsSkScene(0));
epilogueOp();
}
}
};
}
}
}