* 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.
*/
#ifndef K_MAX_SHAPE_DIM
#define K_MAX_SHAPE_DIM 0
#endif
#include "catlass/gemm/kernel/matmul_full_dequant.hpp"
#include "catlass/arch/arch.hpp"
#include "catlass/catlass.hpp"
#include "catlass/epilogue/block/block_epilogue.hpp"
#include "catlass/epilogue/block/block_epilogue_dequant.hpp"
#include "catlass/epilogue/dispatch_policy.hpp"
#include "catlass/gemm/block/block_mmad.hpp"
#include "catlass/gemm/block/block_scheduler_aswt.hpp"
#include "catlass/gemm/device/device_gemm.hpp"
#include "catlass/gemm/dispatch_policy.hpp"
#include "catlass/gemm/gemm_type.hpp"
#include "catlass/layout/layout.hpp"
#include "catlass/status.hpp"
#include "tla/layout.hpp"
#include "golden.hpp"
#include "helper.hpp"
using namespace Catlass;
using namespace tla;
using Options = GemmOptions;
using QuantMode = Epilogue::Block::QuantMode;
struct MatmulShape{
uint32_t m;
uint32_t n;
uint32_t k;
};
static void Run(const Options &options, QuantMode x1QuantMode, QuantMode x2QuantMode, bool hasQuantBias)
{
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();
uint32_t k = options.problemShape.k();
using ElementA = int8_t;
using ElementB = int8_t;
using ElementC = int32_t;
using ElementD = half;
using ElementX1 = float;
using ElementX2 = float;
using ElementBias = float;
using LayoutTagA = layout::RowMajor;
using LayoutTagB = layout::RowMajor;
using LayoutTagC = layout::RowMajor;
LayoutTagA tagA = LayoutTagA::MakeLayout<ElementA>(m, k);
LayoutTagB tagB = LayoutTagB::MakeLayout<ElementB>(k, n);
LayoutTagC tagC = LayoutTagC::MakeLayout<ElementC>(m, n);
size_t lenA = tagA.Capacity();
size_t lenB = tagB.Capacity();
size_t lenC = tagC.Capacity();
size_t lenBias = static_cast<size_t>(n);
size_t sizeA = lenA * sizeof(ElementA);
size_t sizeB = lenB * sizeof(ElementB);
size_t sizeC = lenC * sizeof(fp16_t);
size_t sizeBias = lenBias * sizeof(ElementBias);
size_t goldenSize = lenC * sizeof(float);
size_t sizeWorkspace;
uint8_t *hostA;
ACL_CHECK(aclrtMallocHost((void **)(&hostA), sizeA));
ReadFile("./input/x1.bin", hostA, sizeA);
uint8_t *deviceA{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceA), sizeA, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceA, sizeA, hostA, sizeA, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *hostB;
ACL_CHECK(aclrtMallocHost((void **)(&hostB), sizeB));
ReadFile("./input/x2.bin", hostB, sizeB);
uint8_t *deviceB{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceB), sizeB, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceB, sizeB, hostB, sizeB, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *deviceC{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceC), sizeC, ACL_MEM_MALLOC_HUGE_FIRST));
uint8_t *x1ScaleHost = nullptr;
uint8_t *x1ScaleDevice = nullptr;
size_t x1ScaleFileSize = 0;
if (x1QuantMode == QuantMode::PERTENSOR_MODE) {
x1ScaleFileSize = sizeof(float);
} else if (x1QuantMode == QuantMode::PERTOKEN_MODE) {
x1ScaleFileSize = m * sizeof(float);
}
if (x1QuantMode != QuantMode::DEFAULT) {
ACL_CHECK(aclrtMallocHost((void **)(&x1ScaleHost), x1ScaleFileSize));
ACL_CHECK(aclrtMalloc((void **)&x1ScaleDevice, x1ScaleFileSize, ACL_MEM_MALLOC_HUGE_FIRST));
ReadFile("./input/x1_scale.bin", x1ScaleHost, x1ScaleFileSize);
ACL_CHECK(aclrtMemcpy(x1ScaleDevice, x1ScaleFileSize, x1ScaleHost, x1ScaleFileSize, ACL_MEMCPY_HOST_TO_DEVICE));
}
uint8_t *x2ScaleHost = nullptr;
uint8_t *x2ScaleDevice = nullptr;
size_t x2ScaleFileSize = 0;
if (x2QuantMode == QuantMode::PERTENSOR_MODE) {
x2ScaleFileSize = sizeof(float);
} else if (x2QuantMode == QuantMode::PERCHANNEL_MODE) {
x2ScaleFileSize = n * sizeof(float);
}
if (x2QuantMode != QuantMode::DEFAULT) {
ACL_CHECK(aclrtMallocHost((void **)(&x2ScaleHost), x2ScaleFileSize));
ACL_CHECK(aclrtMalloc((void **)&x2ScaleDevice, x2ScaleFileSize, ACL_MEM_MALLOC_HUGE_FIRST));
ReadFile("./input/x2_scale.bin", x2ScaleHost, x2ScaleFileSize);
ACL_CHECK(aclrtMemcpy(x2ScaleDevice, x2ScaleFileSize, x2ScaleHost, x2ScaleFileSize, ACL_MEMCPY_HOST_TO_DEVICE));
}
uint8_t *quantBiasHost = nullptr;
uint8_t *quantBiasDevice = nullptr;
if (hasQuantBias) {
size_t quantBiasFileSize = n * sizeof(float);
ACL_CHECK(aclrtMallocHost((void **)(&quantBiasHost), quantBiasFileSize));
ACL_CHECK(aclrtMalloc((void **)&quantBiasDevice, quantBiasFileSize, ACL_MEM_MALLOC_HUGE_FIRST));
ReadFile("./input/bias.bin", quantBiasHost, quantBiasFileSize);
ACL_CHECK(aclrtMemcpy(quantBiasDevice, quantBiasFileSize, quantBiasHost, quantBiasFileSize, ACL_MEMCPY_HOST_TO_DEVICE));
}
uint8_t *deviceWorkspace{nullptr};
auto aicCoreNum = platform_ascendc::PlatformAscendCManager::GetInstance()->GetCoreNumAic();
using ArchTag = Arch::Ascend950;
constexpr bool enableUnitFlag = true;
using DispatchPolicy = Gemm::MmadPingpong<ArchTag, enableUnitFlag>;
using L1TileShape = Shape<Int<128>, Int<256>, Int<256>>;
using L0TileShape = Shape<Int<128>, Int<256>, Int<128>>;
auto layoutA = tla::MakeLayout<ElementA, LayoutTagA>(m, k);
auto layoutB = tla::MakeLayout<ElementB, LayoutTagB>(k, n);
auto layoutC = tla::MakeLayout<ElementC, LayoutTagC>(m, n);
using ProblemShape = MatmulShape;
MatmulShape shape = {m, n, k};
using TileCopy = Gemm::Tile::PackedTileCopyTlaToUB<
ArchTag, ElementA, LayoutTagA, ElementB, LayoutTagB, ElementC, LayoutTagC, ElementBias,
Gemm::Tile::CopyL0CToUBMode::SPLIT_M>;
using BlockMmad = Gemm::Block::BlockMmadTla<
DispatchPolicy, L1TileShape, L0TileShape, ElementA, ElementB, ElementC, ElementBias, TileCopy>;
using EpilogueDispatchPolicy = Epilogue::BlockEpilogueDequant;
using BlockEpilogue = Epilogue::Block::BlockEpilogue<EpilogueDispatchPolicy, L0TileShape, ElementD, ElementC, ElementX1, ElementX2, ElementBias>;
uint32_t taskNum = CeilDiv(options.problemShape.m(), tla::get<0>(L1TileShape{})) *
CeilDiv(options.problemShape.n(), tla::get<1>(L1TileShape{}));
uint32_t aicCoreUsed = min(aicCoreNum, taskNum);
using BlockScheduler = typename Gemm::Block::BlockSchedulerAswt<L1TileShape, L0TileShape>;
using MatmulKernel = Gemm::Kernel::KernelMatmulDequant<ProblemShape, BlockMmad, BlockEpilogue, BlockScheduler>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
using Arguments = typename MatmulKernel::Arguments;
Arguments arguments = {
shape,
{deviceA, deviceB, deviceC},
{
deviceC, x2ScaleDevice, x1ScaleDevice, quantBiasDevice,
{
tla::get<0>(L1TileShape{}),
tla::get<1>(L1TileShape{}),
x1QuantMode,
x2QuantMode,
AscendC::DT_FLOAT,
quantBiasDevice != nullptr
}
}
};
MatmulAdapter matmulOp;
matmulOp.CanImplement(arguments);
sizeWorkspace = matmulOp.GetWorkspaceSize(arguments);
if (sizeWorkspace > 0) {
ACL_CHECK(
aclrtMalloc(reinterpret_cast<void **>(&deviceWorkspace), sizeWorkspace, ACL_MEM_MALLOC_HUGE_FIRST)
);
}
matmulOp.Initialize(arguments, deviceWorkspace);
matmulOp(stream, aicCoreUsed);
ACL_CHECK(aclrtSynchronizeStream(stream));
std::vector<fp16_t> hostC(lenC);
ACL_CHECK(aclrtMemcpy(hostC.data(), sizeC, deviceC, sizeC, ACL_MEMCPY_DEVICE_TO_HOST));
std::vector<float> hostGolden(lenC);
std::string expected_path = "./output/golden_o.bin";
ReadFile(expected_path, hostGolden.data(), goldenSize);
std::vector<uint64_t> errorIndices = golden::CompareData(hostC, hostGolden, k);
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(deviceB));
ACL_CHECK(aclrtFree(deviceC));
if (x1QuantMode != QuantMode::DEFAULT) {
ACL_CHECK(aclrtFree(x1ScaleDevice));
}
if (x2QuantMode != QuantMode::DEFAULT) {
ACL_CHECK(aclrtFree(x2ScaleDevice));
}
if (hasQuantBias) {
ACL_CHECK(aclrtFree(quantBiasDevice));
}
if (sizeWorkspace > 0) {
ACL_CHECK(aclrtFree(deviceWorkspace));
}
ACL_CHECK(aclrtDestroyStream(stream));
ACL_CHECK(aclrtResetDevice(options.deviceId));
ACL_CHECK(aclFinalize());
}
QuantMode StringToQuantMode(const std::string& s)
{
if (s == "per_tensor") {
return QuantMode::PERTENSOR_MODE;
} else if (s == "per_channel") {
return QuantMode::PERCHANNEL_MODE;
} else if (s == "per_token") {
return QuantMode::PERTOKEN_MODE;
} else if (s == "default") {
return QuantMode::DEFAULT;
} else {
throw std::invalid_argument("Invalid quant mode string: " + s);
}
}
struct ArgsParams {
uint32_t m;
uint32_t n;
uint32_t k;
QuantMode x1QuantMode;
QuantMode x2QuantMode;
bool enableBias = false;
};
bool IsValidQuantMode(
QuantMode x1QuantMode,
QuantMode x2QuantMode,
bool enableBias
)
{
using TupleType = std::tuple<QuantMode, QuantMode, bool>;
static const std::vector<TupleType> validPairs = {
{ QuantMode::PERTOKEN_MODE, QuantMode::PERTENSOR_MODE, false },
{ QuantMode::PERTOKEN_MODE, QuantMode::PERCHANNEL_MODE, false },
{ QuantMode::PERTENSOR_MODE, QuantMode::PERCHANNEL_MODE, false },
{ QuantMode::DEFAULT, QuantMode::PERCHANNEL_MODE, false },
{ QuantMode::PERTOKEN_MODE, QuantMode::PERTENSOR_MODE, true },
{ QuantMode::PERTOKEN_MODE, QuantMode::PERCHANNEL_MODE, true },
{ QuantMode::DEFAULT, QuantMode::PERTENSOR_MODE, true },
{ QuantMode::DEFAULT, QuantMode::PERCHANNEL_MODE, true }
};
const auto target = std::make_tuple(x1QuantMode, x2QuantMode, enableBias);
return std::find(validPairs.begin(), validPairs.end(), target) != validPairs.end();
}
ArgsParams ParseArguments(int32_t argc, const char* argv[])
{
ArgsParams params;
constexpr static int32_t minArgs = 5;
if (argc < minArgs) {
throw std::invalid_argument("Insufficient arguments provided");
}
try {
params.m = std::stoi(argv[1]);
params.n = std::stoi(argv[2]);
params.k = std::stoi(argv[3]);
params.x1QuantMode = StringToQuantMode(argv[4]);
params.x2QuantMode = StringToQuantMode(argv[5]);
if (argc >= 7) {
if (std::string(argv[6]) == "has_bias") {
params.enableBias = true;
} else {
throw std::invalid_argument("Invalid argument for bias flag: " + std::string(argv[6]));
}
}
} catch (const std::exception& e) {
throw std::invalid_argument(std::string("Error parsing arguments: ") + e.what());
}
return params;
}
int main(int argc, const char **argv)
{
for (int32_t i = 0;i < argc;i++) {
std::cout << "arg[" << i << "]: " << argv[i] << std::endl;
}
ArgsParams params = ParseArguments(argc, argv);
std::cout \
<< "m: " << params.m << std::endl
<< "n: " << params.n << std::endl
<< "k: " << params.k << std::endl
<< "x1QuantMode: " << static_cast<int>(params.x1QuantMode) << std::endl
<< "x2QuantMode: " << static_cast<int>(params.x2QuantMode) << std::endl
<< "enableBias: " << (params.enableBias ? "true" : "false") << std::endl;
if (!IsValidQuantMode(params.x1QuantMode, params.x2QuantMode, params.enableBias)) {
std::cerr << "Invalid combination of quantization modes and bias flag." << std::endl;
return -1;
}
Options options;
options.problemShape = {params.m, params.n, params.k};
options.deviceId = 0;
Run(options,
params.x1QuantMode,
params.x2QuantMode,
params.enableBias);
return 0;
}