* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This file is a part of the CANN Open Software.
* Licensed under CANN Open Software License Agreement Version 1.0 (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.
*/
#ifndef K_MAX_SHAPE_DIM
#define K_MAX_SHAPE_DIM 0
#endif
#include <iostream>
#include <vector>
#include "helper.hpp"
#include "golden.hpp"
#include "catlass/catlass.hpp"
#include "catlass/arch/arch.hpp"
#include "catlass/gemm/dispatch_policy.hpp"
#include "catlass/gemv/kernel/kernel_gemv_aiv.hpp"
#include "catlass/gemv/block/block_gemv.hpp"
#include "catlass/gemm/gemm_type.hpp"
#include "catlass/layout/layout.hpp"
#include "catlass/gemv/tile/tile_copy.hpp"
#include "catlass/gemv/tile/tile_vmad.hpp"
#include "catlass/gemv/tile/tile_vmuls.hpp"
#include "catlass/gemv/device/device_gemv.hpp"
#include "catlass/status.hpp"
using namespace Catlass;
using ScalarType = float;
struct Options{
const std::string HELPER = "17_gemv_aiv m n [device_id]";
uint32_t M = 32;
uint32_t N = 32;
uint32_t deviceId{0};
Options() = default;
GemvCoord problemShape{M, N};
int Parse(int argc, const char **argv)
{
enum ArgsIndex {
M_INDEX = 1,
N_INDEX,
DEVICE_ID_INDEX,
ARGS_MAX
};
if (argc > ARGS_MAX || argc <= N_INDEX)
{
std::cerr << HELPER << std::endl;
return -1;
}
problemShape.m() = std::atoi(argv[M_INDEX]);
problemShape.n() = std::atoi(argv[N_INDEX]);
if (argc == ARGS_MAX)
{
deviceId = std::atoi(argv[DEVICE_ID_INDEX]);
}
return 0;
}
};
uint32_t getSplictNum(bool trans, uint32_t M, uint32_t N, uint32_t M1, uint32_t N1, uint32_t maxSplict)
{
uint32_t CORENUM = 20;
uint32_t splitNum = 1;
uint32_t maxOccupancy = 0;
uint32_t blockNum = (M - 1) / M1 + 1;
if (!trans)
{
splitNum = 1;
}
else{
uint32_t splitNum1 = 1, splitNum2 = 1;
for (uint32_t i = 1; i <= maxSplict; i += 1)
{
uint32_t occupancy = (i * blockNum) % (CORENUM * 2);
if (!occupancy)
occupancy = (CORENUM * 2);
if (occupancy > maxOccupancy)
{
maxOccupancy = occupancy;
splitNum1 = i;
}
}
maxOccupancy = 0;
for (uint32_t i = 1; i <= maxSplict; i <<= 1)
{
uint32_t occupancy = (i * blockNum) % (CORENUM * 2);
if (!occupancy)
occupancy = (CORENUM * 2);
if (occupancy > maxOccupancy)
{
maxOccupancy = occupancy;
splitNum2 = i;
}
}
splitNum = (splitNum1 - splitNum2) > 4 ? splitNum1 : splitNum2;
}
return splitNum;
}
template <class Adapter>
void RunAdapter(Adapter gemv_op, typename Adapter::Arguments args, aclrtStream stream,
uint32_t aicCoreNum)
{
size_t sizeWorkspace = gemv_op.GetWorkspaceSize(args);
uint8_t *deviceWorkspace = nullptr;
if (sizeWorkspace > 0) {
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceWorkspace), sizeWorkspace, ACL_MEM_MALLOC_HUGE_FIRST));
}
gemv_op.Initialize(args, deviceWorkspace);
gemv_op(stream, aicCoreNum);
ACL_CHECK(aclrtSynchronizeStream(stream));
if (sizeWorkspace > 0) {
ACL_CHECK(aclrtFree(deviceWorkspace));
}
}
template<class ElementRandom>
void FillRandomScalarData(ElementRandom &scalarData, ElementRandom low, ElementRandom high)
{
scalarData = static_cast<ElementRandom>(low + (static_cast<ElementRandom>(rand()) / static_cast<ElementRandom>(RAND_MAX)) * (high - low));
}
void Run(Options options){
aclrtStream stream{nullptr};
ACL_CHECK(aclInit(nullptr));
ACL_CHECK(aclrtSetDevice(options.deviceId));
ACL_CHECK(aclrtCreateStream(&stream));
uint32_t m = options.problemShape.m();
uint32_t n = options.problemShape.n();
using UBTileShape = GemvShape<32,512>;
uint32_t maxSplict = 20;
uint32_t const split = getSplictNum(false, m, n, UBTileShape::M, UBTileShape::N, maxSplict);
size_t lenA = static_cast<size_t>(m) * n;
size_t lenX = static_cast<size_t>(n) * 1;
size_t lenY = static_cast<size_t>(m) * 1;
size_t scalarLen = 1;
size_t sizeA = lenA * sizeof(float);
size_t sizeX = lenX * sizeof(float);
size_t sizeY = lenY * sizeof(float);
using LayoutA = layout::RowMajor;
using LayoutX = layout::VectorLayout;
using LayoutY = layout::VectorLayout;
LayoutA layoutA{m, n};
LayoutX layoutX{n};
LayoutY layoutY{m};
ScalarType alpha{0};
ScalarType beta{0};
FillRandomScalarData(alpha, -1.0f, 1.0f);
FillRandomScalarData(beta, -1.0f, 1.0f);
std::vector<float> hostA(lenA);
std::vector<float> hostX(lenX);
std::vector<float> hostY(lenY);
golden::FillRandomData(hostA, -1.0f, 1.0f);
golden::FillRandomData(hostX, -1.0f, 1.0f);
golden::FillRandomData(hostY, -1.0f, 1.0f);
uint8_t *deviceA{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceA), sizeA, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceA, sizeA, hostA.data(), sizeA, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *deviceX{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceX), sizeX, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceX, sizeX, hostX.data(), sizeX, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *deviceY{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceY), sizeY, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceY, sizeY, hostY.data(), sizeY, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *deviceZ{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceZ), sizeY, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceZ, sizeY, hostY.data(), sizeY, ACL_MEMCPY_HOST_TO_DEVICE));
auto aicCoreNum = platform_ascendc::PlatformAscendCManager::GetInstance()->GetCoreNumAiv();
using ArchTag = Arch::AtlasA2;
using DispatchPolicy = Gemm::GemvAtlasA2;
using AType = Gemm::GemmType<float, LayoutA>;
using XType = Gemm::GemmType<float, LayoutX>;
using YType = Gemm::GemmType<float, LayoutY>;
using BiasType = void;
using TileCopy = Gemv::Tile::TileCopyGemvAiv<typename DispatchPolicy::ArchTag, AType, XType, YType, BiasType>;
using TileVmad = Gemv::Tile::TileVmad<typename DispatchPolicy::ArchTag, AType, XType, YType, BiasType>;
using TileVmuls = Gemv::Tile::TileVmuls<typename DispatchPolicy::ArchTag, XType>;
using GemvBlock = Gemv::Block::BlockGemv<DispatchPolicy, UBTileShape, AType, XType, YType, BiasType, TileCopy, TileVmad, TileVmuls>;
using BlockEpilogue = void;
using GemvKernel = Gemv::Kernel::KernelGemvAiv<GemvBlock, BlockEpilogue>;
typename GemvKernel::Arguments arguments{options.problemShape, deviceA, deviceX, deviceY, deviceZ, alpha, beta, split};
using GemvAdapter = Gemv::Device::DeviceGemv<GemvKernel>;
GemvAdapter gemv_op;
gemv_op.CanImplement(arguments);
RunAdapter(gemv_op, arguments, stream, aicCoreNum);
std::vector<float> hostRes(lenY);
ACL_CHECK(aclrtMemcpy(hostRes.data(), sizeY, deviceZ, sizeY, ACL_MEMCPY_DEVICE_TO_HOST));
std::vector<float> hostGolden(lenY);
golden::ComputeGemv(options.problemShape, alpha, beta, hostA, layoutA, hostX, layoutX, hostY, layoutY, hostGolden, layoutY);
std::vector<uint64_t> errorIndices = golden::CompareData(hostRes, hostGolden, m);
if (errorIndices.empty()) {
std::cout << "Compare success." << std::endl;
} else {
std::cerr << "Compare failed. Error count: " << errorIndices.size() << std::endl;
}
ACL_CHECK(aclrtFree(deviceA));
ACL_CHECK(aclrtFree(deviceX));
ACL_CHECK(aclrtFree(deviceY));
ACL_CHECK(aclrtFree(deviceZ));
ACL_CHECK(aclrtDestroyStream(stream));
ACL_CHECK(aclrtResetDevice(options.deviceId));
ACL_CHECK(aclFinalize());
}
int main(int argc, const char **argv){
Options options;
if(options.Parse(argc, argv) != 0){
return -1;
}
Run(options);
return 0;
}