/**
* 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 matmul.asc
* \brief
*/
#include "data_utils.h"
#include "kernel_tiling/kernel_tiling.h"
#include "tiling/platform/platform_ascendc.h"
#include "tiling/tiling_api.h"
#include "acl/acl.h"
#include "kernel_operator.h"
#ifdef ENABLE_CUBE_ONLY
#define ASCENDC_CUBE_ONLY
#endif
#include "lib/matmul_intf.h"
constexpr uint32_t M = 128;
constexpr uint32_t N = 30720;
constexpr uint32_t K = 64;
constexpr bool isTransA = false;
constexpr bool isTransB = false;
constexpr bool isBias = false;
/**
* @brief Copy tiling data to TCubeTiling ptr from tiling gm addr.
* @param tiling: TCubeTiling ptr which needs to copy tiling data.
* @param tilingGM: tiling gm addr.
* @retval None
*/
__aicore__ inline void CopyTiling(TCubeTiling* tiling, GM_ADDR tilingGM)
{
uint32_t* ptr = reinterpret_cast<uint32_t*>(tiling);
auto tiling32 = reinterpret_cast<__gm__ uint32_t*>(tilingGM);
for (uint32_t i = 0; i < sizeof(TCubeTiling) / sizeof(uint32_t); i++, ptr++) { *ptr = *(tiling32 + i); }
return;
}
__aicore__ inline constexpr MatmulConfig GetCustomNormCFG()
{
auto mmCfg = CFG_NORM;
// disable unitflag for performance comparison
mmCfg.enUnitFlag = false;
return mmCfg;
}
__aicore__ inline constexpr MatmulConfig GetCustomMDLCFG()
{
auto mmCfg = CFG_MDL;
#ifdef ENABLE_MDL_UNITFLAG
mmCfg.enUnitFlag = true;
#endif
return mmCfg;
}
template <typename AType, typename BType, typename CType, typename BiasType>
class MatmulKernel {
public:
__aicore__ inline MatmulKernel(){};
/**
* @brief Initialization before process.
* @param a: A matrix gm addr.
* @param b: B matrix gm addr.
* @param bias: Bias matrix gm addr.
* @param c: C matrix gm addr.
* @param tiling: Matmul tiling struct.
* @retval None
*/
__aicore__ inline void Init(GM_ADDR a, GM_ADDR b, GM_ADDR bias, GM_ADDR c, const TCubeTiling& tiling);
/**
* @brief Process matrix calculation.
* @param pipe: TPipe object.
* @retval None
*/
__aicore__ inline void Process(AscendC::TPipe* pipe);
using A_TYPE = AscendC::MatmulType<AscendC::TPosition::GM, CubeFormat::ND, AType>;
using B_TYPE = AscendC::MatmulType<AscendC::TPosition::GM, CubeFormat::ND, BType>;
using C_TYPE = AscendC::MatmulType<AscendC::TPosition::GM, CubeFormat::ND, CType>;
using BIAS_TYPE = AscendC::MatmulType<AscendC::TPosition::GM, CubeFormat::ND, BiasType>;
#ifdef ENABLE_MDL
constexpr static MatmulConfig CUSTOM_CFG_MDL = GetCustomMDLCFG();
AscendC::Matmul<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE, CUSTOM_CFG_MDL> matmulObj;
#else
constexpr static MatmulConfig CUSTOM_CFG_NORM = GetCustomNormCFG();
AscendC::Matmul<A_TYPE, B_TYPE, C_TYPE, BIAS_TYPE, CUSTOM_CFG_NORM> matmulObj;
#endif
private:
/**
* @brief Calculate the gm offset based on the blockIdx.
* @param offsetA: Gm offset of A matrix.
* @param offsetB: Gm offset of B matrix.
* @param offsetC: Gm offset of C matrix.
* @param offsetBias: Gm offset of Bias matrix.
* @retval None
*/
__aicore__ inline void CalcOffset(int32_t& offsetA, int32_t& offsetB, int32_t& offsetC, int32_t& offsetBias);
AscendC::GlobalTensor<AType> aGlobal;
AscendC::GlobalTensor<BType> bGlobal;
AscendC::GlobalTensor<CType> cGlobal;
AscendC::GlobalTensor<BiasType> biasGlobal;
TCubeTiling tiling;
int32_t mCoreIndex;
int32_t nCoreIndex;
};
template <typename AType, typename BType, typename CType, typename BiasType>
__aicore__ inline void MatmulKernel<AType, BType, CType, BiasType>::Init(GM_ADDR a, GM_ADDR b, GM_ADDR bias, GM_ADDR c,
const TCubeTiling& tiling)
{
this->tiling = tiling;
if (AscendC::GetBlockIdx() >= tiling.usedCoreNum) {
return;
}
aGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ AType*>(a), tiling.M * tiling.Ka);
bGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ BType*>(b), tiling.Kb * tiling.N);
cGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ CType*>(c), tiling.M * tiling.N);
biasGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ BiasType*>(bias), tiling.N);
int32_t offsetA = 0;
int32_t offsetB = 0;
int32_t offsetC = 0;
int32_t offsetBias = 0;
CalcOffset(offsetA, offsetB, offsetC, offsetBias);
aGlobal = aGlobal[offsetA];
bGlobal = bGlobal[offsetB];
cGlobal = cGlobal[offsetC];
biasGlobal = biasGlobal[offsetBias];
}
template <typename AType, typename BType, typename CType, typename BiasType>
__aicore__ inline void MatmulKernel<AType, BType, CType, BiasType>::Process(AscendC::TPipe* pipe)
{
if (AscendC::GetBlockIdx() >= tiling.usedCoreNum) {
return;
}
// process with tail block
int tailM = tiling.M - mCoreIndex * tiling.singleCoreM;
tailM = tailM < tiling.singleCoreM ? (tailM > 0 ? tailM : tiling.singleCoreM) : tiling.singleCoreM;
int tailN = tiling.N - nCoreIndex * tiling.singleCoreN;
tailN = tailN < tiling.singleCoreN ? (tailN > 0 ? tailN : tiling.singleCoreN) : tiling.singleCoreN;
matmulObj.SetSingleShape(tailM, tailN, tiling.Ka);
matmulObj.SetTensorA(aGlobal, isTransA);
matmulObj.SetTensorB(bGlobal, isTransB);
if (tiling.isBias) {
matmulObj.SetBias(biasGlobal);
}
matmulObj.IterateAll(cGlobal);
matmulObj.End();
}
template <typename AType, typename BType, typename CType, typename BiasType>
__aicore__ inline void MatmulKernel<AType, BType, CType, BiasType>::CalcOffset(int32_t& offsetA, int32_t& offsetB,
int32_t& offsetC, int32_t& offsetBias)
{
const TCubeTiling& tiling = this->tiling;
int32_t blockIdx = AscendC::GetBlockIdx();
if (blockIdx >= tiling.usedCoreNum) {
return;
}
auto mSingleBlocks = (tiling.M + tiling.singleCoreM - 1) / tiling.singleCoreM; // split M into mSingleBlocks cores
mCoreIndex = blockIdx % mSingleBlocks;
nCoreIndex = blockIdx / mSingleBlocks;
if (isTransA) {
offsetA = mCoreIndex * tiling.singleCoreM;
} else {
offsetA = mCoreIndex * tiling.Ka * tiling.singleCoreM;
}
if (isTransB) {
offsetB = nCoreIndex * tiling.Kb * tiling.singleCoreN;
} else {
offsetB = nCoreIndex * tiling.singleCoreN;
}
offsetC = mCoreIndex * tiling.N * tiling.singleCoreM + nCoreIndex * tiling.singleCoreN;
offsetBias = nCoreIndex * tiling.singleCoreN;
}
/**
* @brief matmul kernel function entry
* @param a: A matrix gm addr.
* @param b: B matrix gm addr.
* @param bias: bias matrix gm addr.
* @param c: C matrix gm addr.
* @param workspace: Temporary gm space addr required by matmul calc.
* @param tilingGm: Tiling data addr.
* @retval None
*/
__global__ __aicore__ void matmul_custom(GM_ADDR a, GM_ADDR b, GM_ADDR bias, GM_ADDR c,
__kfc_workspace__ GM_ADDR workspace, GM_ADDR tilingGm)
{
#ifdef ENABLE_CUBE_ONLY
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIC_ONLY);
#endif
TCubeTiling tiling;
CopyTiling(&tiling, tilingGm);
MatmulKernel<half, half, float, float> matmulKernel;
AscendC::TPipe pipe;
matmulKernel.Init(a, b, bias, c, tiling);
REGIST_MATMUL_OBJ(&pipe, GetSysWorkSpacePtr(), matmulKernel.matmulObj, &tiling);
matmulKernel.Process(&pipe);
}
void GenerateTiling(platform_ascendc::PlatformAscendC* ascendcPlatform, uint8_t* tilingBuf)
{
optiling::TCubeTiling tilingData;
matmul_tiling::MultiCoreMatmulTiling tilingApi(*ascendcPlatform);
#ifdef ENABLE_CUBE_ONLY
tilingApi.SetDim(ascendcPlatform->GetCoreNumAic());
#else
tilingApi.SetDim(ascendcPlatform->GetCoreNumAiv());
#endif
tilingApi.SetAType(matmul_tiling::TPosition::GM, matmul_tiling::CubeFormat::ND, matmul_tiling::DataType::DT_FLOAT16,
isTransA);
tilingApi.SetBType(matmul_tiling::TPosition::GM, matmul_tiling::CubeFormat::ND, matmul_tiling::DataType::DT_FLOAT16,
isTransB);
tilingApi.SetCType(matmul_tiling::TPosition::GM, matmul_tiling::CubeFormat::ND, matmul_tiling::DataType::DT_FLOAT);
tilingApi.SetBiasType(matmul_tiling::TPosition::GM, matmul_tiling::CubeFormat::ND,
matmul_tiling::DataType::DT_FLOAT);
tilingApi.SetOrgShape(M, N, K);
tilingApi.SetShape(M, N, K);
tilingApi.EnableBias(isBias);
tilingApi.SetBufferSpace(-1, -1, -1);
int64_t res = tilingApi.GetTiling(tilingData); // Get matmul tiling data.
if (res == -1) {
std::cout << "gen tiling failed" << std::endl;
}
uint32_t tcubeTilingSize = tilingData.GetDataSize();
tilingData.SaveToBuffer(tilingBuf, tcubeTilingSize);
}
int32_t main(int32_t argc, char* argv[])
{
auto ascendcPlatform = platform_ascendc::PlatformAscendCManager::GetInstance();
size_t aFileSize = static_cast<size_t>(M * K) * sizeof(uint16_t); // uint16_t represent half
size_t bFileSize = static_cast<size_t>(K * N) * sizeof(uint16_t); // uint16_t represent half
size_t cFileSize = static_cast<size_t>(M * N) * sizeof(float);
size_t biasFileSize = static_cast<size_t>(1 * N) * sizeof(float);
size_t userWorkspaceSize = 0;
size_t systemWorkspaceSize = static_cast<size_t>(ascendcPlatform->GetLibApiWorkSpaceSize());
size_t workspaceSize = userWorkspaceSize + systemWorkspaceSize;
// matmul TCubeTiling
size_t tilingFileSize = sizeof(TCubeTiling);
uint8_t* tilingBuf = (uint8_t*)malloc(tilingFileSize);
GenerateTiling(ascendcPlatform, tilingBuf);
uint32_t numBlocks = reinterpret_cast<TCubeTiling*>(tilingBuf)->usedCoreNum;
int32_t deviceId = 0;
aclrtStream stream = nullptr;
aclrtContext context;
aclInit(nullptr);
aclrtSetDevice(deviceId);
aclrtCreateContext(&context, deviceId);
aclrtCreateStream(&stream);
uint8_t* aHost;
uint8_t* aDevice;
aclrtMallocHost((void**)(&aHost), aFileSize);
aclrtMalloc((void**)&aDevice, aFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
ReadFile("./input/x1_gm.bin", aFileSize, aHost, aFileSize);
aclrtMemcpy(aDevice, aFileSize, aHost, aFileSize, ACL_MEMCPY_HOST_TO_DEVICE);
uint8_t* bHost;
uint8_t* bDevice;
aclrtMallocHost((void**)(&bHost), bFileSize);
aclrtMalloc((void**)&bDevice, bFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
ReadFile("./input/x2_gm.bin", bFileSize, bHost, bFileSize);
aclrtMemcpy(bDevice, bFileSize, bHost, bFileSize, ACL_MEMCPY_HOST_TO_DEVICE);
uint8_t* biasHost = nullptr;
uint8_t* biasDevice = nullptr;
if (isBias) {
aclrtMallocHost((void**)(&biasHost), biasFileSize);
aclrtMalloc((void**)&biasDevice, biasFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
ReadFile("../input/bias_gm.bin", biasFileSize, biasHost, biasFileSize);
aclrtMemcpy(biasDevice, biasFileSize, biasHost, biasFileSize, ACL_MEMCPY_HOST_TO_DEVICE);
}
uint8_t* cHost;
uint8_t* cDevice;
aclrtMallocHost((void**)(&cHost), cFileSize);
aclrtMalloc((void**)&cDevice, cFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
uint8_t* workspaceDevice;
aclrtMalloc((void**)&workspaceDevice, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST);
uint8_t* tilingHost;
uint8_t* tilingDevice;
aclrtMallocHost((void**)(&tilingHost), tilingFileSize);
aclrtMalloc((void**)&tilingDevice, tilingFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
aclrtMemcpy(tilingHost, tilingFileSize, tilingBuf, tilingFileSize, ACL_MEMCPY_HOST_TO_HOST);
aclrtMemcpy(tilingDevice, tilingFileSize, tilingHost, tilingFileSize, ACL_MEMCPY_HOST_TO_DEVICE);
matmul_custom<<<numBlocks, nullptr, stream>>>(aDevice, bDevice, biasDevice, cDevice, workspaceDevice, tilingDevice);
aclrtSynchronizeStream(stream);
aclrtMemcpy(cHost, cFileSize, cDevice, cFileSize, ACL_MEMCPY_DEVICE_TO_HOST);
WriteFile("./output/output.bin", cHost, cFileSize);
aclrtFree(aDevice);
aclrtFreeHost(aHost);
aclrtFree(bDevice);
aclrtFreeHost(bHost);
if (isBias) {
aclrtFree(biasDevice);
aclrtFreeHost(biasHost);
}
aclrtFree(workspaceDevice);
aclrtFree(tilingDevice);
aclrtFreeHost(tilingHost);
aclrtFree(cDevice);
aclrtFreeHost(cHost);
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
free(tilingBuf);
return 0;
}
