/**
* 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 mmad_unitflag_disable.asc
* \brief
*/
#include "acl/acl.h"
#include "kernel_operator.h"
#include "data_utils.h"
// half type, cube block: [16, 16]
constexpr uint32_t CUBE_BLOCK = 16;
constexpr uint32_t CUBE_BLOCK_SIZE = 16 * 16;
constexpr uint32_t M = 128;
constexpr uint32_t K = 512;
constexpr uint32_t N = 256;
constexpr uint32_t kRound = 8;
constexpr static uint16_t LIMIT_MNSIZE = 10;
constexpr static uint16_t ALIGN_NUM = 16;
class KernelMmad {
public:
__aicore__ inline KernelMmad()
{
aSingleSize = m * k;
bSingleSize = k * n;
cSingleSize = m * n;
aSize = kRound * aSingleSize;
bSize = kRound * bSingleSize;
cSize = kRound * cSingleSize;
}
__aicore__ inline void Init(GM_ADDR a, GM_ADDR b, GM_ADDR c, AscendC::TPipe* pipeIn)
{
// set cube only
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIC_ONLY);
pipe = pipeIn;
aGM.SetGlobalBuffer((__gm__ half *)a);
bGM.SetGlobalBuffer((__gm__ half *)b);
cGM.SetGlobalBuffer((__gm__ float *)c);
pipe->InitBuffer(inQueueA1, 1, aSingleSize * sizeof(half));
pipe->InitBuffer(inQueueA2, 1, aSingleSize * sizeof(half));
pipe->InitBuffer(inQueueB1, 1, bSingleSize * sizeof(half));
pipe->InitBuffer(inQueueB2, 1, bSingleSize * sizeof(half));
pipe->InitBuffer(outQueueCO1, 1, cSingleSize * sizeof(float));
}
__aicore__ inline void Process()
{
uint32_t chunkSize = kRound;
AscendC::LocalTensor<float> c = outQueueCO1.AllocTensor<float>();
for (uint32_t kIndex = 0; kIndex < chunkSize; kIndex++) {
CopyIn(kIndex);
SplitA(kIndex);
SplitB(kIndex);
Compute(kIndex, c);
}
CopyOut();
}
private:
__aicore__ inline uint32_t CeilCubeBlock(uint32_t len)
{
return (len + CUBE_BLOCK - 1) / CUBE_BLOCK;
}
__aicore__ inline void CopyIn(uint32_t kIndex)
{
AscendC::LocalTensor<half> a1Local = inQueueA1.AllocTensor<half>();
AscendC::LocalTensor<half> b1Local = inQueueB1.AllocTensor<half>();
AscendC::Nd2NzParams nd2nzA1Params;
// 传输ND矩阵的数目
nd2nzA1Params.ndNum = 1;
nd2nzA1Params.nValue = m;
nd2nzA1Params.dValue = k;
// 源操作数相邻ND矩阵起始地址间的偏移，单位为元素。
nd2nzA1Params.srcNdMatrixStride = 0;
nd2nzA1Params.srcDValue = k;
// 目的NZ矩阵中，来自源操作数同一行的多行数据相邻行起始地址间的偏移，取值范围：dstNzC0Stride∈[1, 16384]，单位：C0_SIZE（32B）。
nd2nzA1Params.dstNzC0Stride = CeilCubeBlock(m) * CUBE_BLOCK;
// 目的NZ矩阵中，Z型矩阵相邻行起始地址之间的偏移。单位：C0_SIZE（32B）。
nd2nzA1Params.dstNzNStride = 1;
// 目的NZ矩阵中，相邻NZ矩阵起始地址间的偏移，单位为元素。
nd2nzA1Params.dstNzMatrixStride = 0;
AscendC::DataCopy(a1Local, aGM[kIndex * aSingleSize], nd2nzA1Params);
AscendC::Nd2NzParams nd2nzB1Params;
nd2nzB1Params.ndNum = 1;
nd2nzB1Params.nValue = k;
nd2nzB1Params.dValue = n;
nd2nzB1Params.srcNdMatrixStride = 0;
nd2nzB1Params.srcDValue = n;
nd2nzB1Params.dstNzC0Stride = CeilCubeBlock(k) * CUBE_BLOCK;
nd2nzB1Params.dstNzNStride = 1;
nd2nzB1Params.dstNzMatrixStride = 0;
AscendC::DataCopy(b1Local, bGM[kIndex * bSingleSize], nd2nzB1Params);
inQueueA1.EnQue(a1Local);
inQueueB1.EnQue(b1Local);
}
__aicore__ inline void SplitA(uint32_t kIndex)
{
AscendC::LocalTensor<half> a1Local = inQueueA1.DeQue<half>();
AscendC::LocalTensor<half> a = inQueueA2.AllocTensor<half>();
uint32_t dstOffset = CeilCubeBlock(k) * CUBE_BLOCK_SIZE;
uint32_t srcOffset = CUBE_BLOCK_SIZE;
// Nz -> Zz
AscendC::LoadData2DParams loadDataParams;
loadDataParams.repeatTimes = CeilCubeBlock(k);
loadDataParams.srcStride = CeilCubeBlock(m);
loadDataParams.dstGap = 0;
loadDataParams.ifTranspose = false;
for (int i = 0; i < CeilCubeBlock(m); ++i) {
AscendC::LoadData(a[i * dstOffset], a1Local[i * srcOffset], loadDataParams);
}
inQueueA2.EnQue<half>(a);
inQueueA1.FreeTensor(a1Local);
}
__aicore__ inline void SplitB(uint32_t kIndex)
{
AscendC::LocalTensor<half> b1Local = inQueueB1.DeQue<half>();
AscendC::LocalTensor<half> b = inQueueB2.AllocTensor<half>();
uint32_t dstOffset = CeilCubeBlock(n) * CUBE_BLOCK_SIZE;
uint32_t srcOffset = CUBE_BLOCK_SIZE;
// Nz -> Zn
AscendC::LoadData2DParams loadDataParams;
loadDataParams.repeatTimes = CeilCubeBlock(n);
loadDataParams.srcStride = CeilCubeBlock(k);
loadDataParams.dstGap = 0;
loadDataParams.ifTranspose = true;
for (int i = 0; i < CeilCubeBlock(k); ++i) {
AscendC::LoadData(b[i * dstOffset], b1Local[i * srcOffset], loadDataParams);
}
inQueueB2.EnQue<half>(b);
inQueueB1.FreeTensor(b1Local);
}
__aicore__ inline void Compute(uint32_t kIndex, AscendC::LocalTensor<float>& c)
{
AscendC::LocalTensor<half> a = inQueueA2.DeQue<half>();
AscendC::LocalTensor<half> b = inQueueB2.DeQue<half>();
AscendC::MmadParams mmadParams;
mmadParams.m = m;
mmadParams.n = n;
mmadParams.k = k;
if (kIndex == 0) {
mmadParams.cmatrixInitVal = true;
} else {
mmadParams.cmatrixInitVal = false;
}
AscendC::Mmad(c, a, b, mmadParams);
if ((m / ALIGN_NUM) * (n / ALIGN_NUM) < LIMIT_MNSIZE) {
AscendC::PipeBarrier<PIPE_M>();
}
if (kIndex == kRound - 1) {
outQueueCO1.EnQue<float>(c);
}
inQueueA2.FreeTensor(a);
inQueueB2.FreeTensor(b);
}
__aicore__ inline void CopyOut()
{
AscendC::LocalTensor<float> c = outQueueCO1.DeQue<float>();
// LOC-->GM,随路NZ2ND
AscendC::FixpipeParamsV220 fixpipeParams;
fixpipeParams.nSize = n;
fixpipeParams.mSize = m;
fixpipeParams.srcStride = m;
fixpipeParams.dstStride = n;
fixpipeParams.ndNum = 1;
fixpipeParams.srcNdStride = 0;
fixpipeParams.dstNdStride = 0;
AscendC::Fixpipe(cGM, c, fixpipeParams);
outQueueCO1.FreeTensor(c);
}
private:
AscendC::TPipe* pipe;
AscendC::TQue<AscendC::TPosition::A1, 1> inQueueA1;
AscendC::TQue<AscendC::TPosition::A2, 1> inQueueA2;
AscendC::TQue<AscendC::TPosition::B1, 1> inQueueB1;
AscendC::TQue<AscendC::TPosition::B2, 1> inQueueB2;
AscendC::TQue<AscendC::TPosition::CO1, 1> outQueueCO1;
AscendC::GlobalTensor<half> aGM;
AscendC::GlobalTensor<half> bGM;
AscendC::GlobalTensor<float> cGM;
// 注意沿着K轴切分
uint16_t m = M, k = K / kRound, n = N;
uint16_t aSize, bSize, cSize;
uint16_t aSingleSize, bSingleSize, cSingleSize;
};
extern "C" __global__ __aicore__ void chunk_mmad_custom(GM_ADDR a, GM_ADDR b, GM_ADDR c)
{
AscendC::TPipe pipe;
KernelMmad op;
op.Init(a, b, c, &pipe);
op.Process();
}
int32_t main(int32_t argc, char *argv[])
{
size_t aFileSize = M * K * sizeof(int16_t);
size_t bFileSize = K * N * sizeof(int16_t);
size_t cFileSize = M * N * sizeof(float);
uint32_t numBlocks = 1;
aclInit(nullptr);
int32_t deviceId = 0;
aclrtSetDevice(deviceId);
aclrtStream stream = nullptr;
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 *cHost;
uint8_t *cDevice;
aclrtMallocHost((void **)(&cHost), cFileSize);
aclrtMalloc((void **)&cDevice, cFileSize, ACL_MEM_MALLOC_HUGE_FIRST);
chunk_mmad_custom<<<numBlocks, nullptr, stream>>>(aDevice, bDevice, cDevice);
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);
aclrtFree(cDevice);
aclrtFreeHost(cHost);
aclrtDestroyStream(stream);
aclrtResetDevice(deviceId);
aclFinalize();
return 0;
}
