* 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/quant_optimized_matmul_tla.hpp"
#include "catlass/arch/arch.hpp"
#include "catlass/catlass.hpp"
#include "catlass/epilogue/block/block_epilogue.hpp"
#include "catlass/epilogue/dispatch_policy.hpp"
#include "catlass/epilogue/tile/tile_broadcast_mul.hpp"
#include "catlass/epilogue/tile/tile_broadcast_one_blk.hpp"
#include "catlass/epilogue/tile/tile_swizzle.hpp"
#include "catlass/gemm/block/block_mmad.hpp"
#include "catlass/gemm/block/block_swizzle.hpp"
#include "catlass/gemm/device/device_gemm.hpp"
#include "catlass/gemm/dispatch_policy.hpp"
#include "catlass/gemm/gemm_type.hpp"
#include "catlass/status.hpp"
#include "tla/layout.hpp"
#include "tla/tensor.hpp"
#include "golden.hpp"
#include "helper.hpp"
using namespace Catlass;
using namespace tla;
template <class LayoutTag>
auto GetPaddingLayout(LayoutTag layout, uint32_t blockRows, uint32_t blockCols) {
if constexpr (std::is_same_v<LayoutTag, layout::RowMajor>) {
auto shape = MakeShape(
MakeShape(blockRows, CeilDiv(layout.shape(0), blockRows)),
MakeShape(blockCols, CeilDiv(layout.shape(1), blockCols))
);
auto stride = MakeStride(
MakeStride(
static_cast<int64_t>(blockCols), static_cast<int64_t>(blockRows) * RoundUp(layout.shape(1), blockCols)
),
MakeStride(Int<1>{}, static_cast<int64_t>(blockRows) * blockCols)
);
return MakeLayout(shape, stride);
} else {
auto shape = MakeShape(
MakeShape(blockRows, CeilDiv(layout.shape(0), blockRows)),
MakeShape(blockCols, CeilDiv(layout.shape(1), blockCols))
);
auto stride = MakeStride(
MakeStride(Int<1>{}, static_cast<int64_t>(blockRows) * blockCols),
MakeStride(
static_cast<int64_t>(blockRows), RoundUp(layout.shape(0), blockRows) * static_cast<int64_t>(blockCols)
)
);
return MakeLayout(shape, stride);
}
}
using Options = GemmOptions;
template <class LayoutTag>
size_t GetWorkspaceLen(LayoutTag layout, size_t blockRows, size_t blockCols) {
return RoundUp(static_cast<size_t>(layout.shape(0)), blockRows)
* RoundUp(static_cast<size_t>(layout.shape(1)), blockCols);
}
static void Run(const 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();
uint32_t k = options.problemShape.k();
size_t lenA = static_cast<size_t>(m) * k;
size_t lenB = static_cast<size_t>(k) * n;
size_t lenScale = static_cast<size_t>(n);
size_t lenPerTokenScale = static_cast<size_t>(m);
size_t lenD = static_cast<size_t>(m) * n;
size_t sizeA = lenA * sizeof(int8_t);
size_t sizeB = lenB * sizeof(int8_t);
size_t sizeScale = lenScale * sizeof(float);
size_t sizePerTokenScale = lenPerTokenScale * sizeof(float);
size_t sizeD = lenD * sizeof(bfloat16);
size_t sizeWorkspace;
using ElementA = int8_t;
using ElementB = int8_t;
using ElementC = int32_t;
using ElementScale = float;
using ElementPerTokenScale = float;
using ElementD = bfloat16_t;
using ArchTag = Arch::AtlasA2;
const uint32_t align = 256;
using LayoutTagA = layout::RowMajor;
using LayoutTagB = layout::ColumnMajor;
using LayoutTagC = layout::RowMajor;
using LayoutTagScale = layout::RowMajor;
using LayoutTagPerTokenScale = LayoutTagScale;
using LayoutTagD = LayoutTagC;
LayoutTagA tagA = LayoutTagA::MakeLayout<ElementA>(m, k);
LayoutTagB tagB = LayoutTagB::MakeLayout<ElementB>(k, n);
LayoutTagC tagC = LayoutTagC::MakeLayout<ElementC>(m, n);
LayoutTagScale tagScale = LayoutTagScale::MakeLayout<ElementScale>(1, n);
LayoutTagPerTokenScale tagPerTokenScale = LayoutTagPerTokenScale::MakeLayout<ElementPerTokenScale>(1, m);
LayoutTagD tagD = LayoutTagD::MakeLayout<ElementD>(m, n);
auto layoutA = MakeLayoutFromTag(tagA);
auto layoutB = MakeLayoutFromTag(tagB);
auto layoutC = MakeLayoutFromTag(tagC);
auto layoutScale = MakeLayoutFromTag(tagScale);
auto layoutPerTokenScale = MakeLayoutFromTag(tagPerTokenScale);
auto layoutD = MakeLayoutFromTag(tagC);
using TensorA =
Tensor<AscendC::GlobalTensor<ElementA>, decltype(layoutA), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TensorB =
Tensor<AscendC::GlobalTensor<ElementB>, decltype(layoutB), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TensorC =
Tensor<AscendC::GlobalTensor<ElementC>, decltype(layoutC), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
bool isNeedPaddingA = IsNeedPadding(tagA, align);
bool isNeedPaddingB = IsNeedPadding(tagB, align);
using L1TileShape = std::conditional_t<
std::is_same_v<LayoutTagA, layout::ColumnMajor> && std::is_same_v<LayoutTagB, layout::ColumnMajor>,
Shape<_256, _128, _512>, Shape<_128, _256, _512>>;
using L0TileShape = std::conditional_t<
std::is_same_v<LayoutTagA, layout::ColumnMajor> && std::is_same_v<LayoutTagB, layout::ColumnMajor>,
Shape<_256, _128, _128>, Shape<_128, _256, _128>>;
using BlockScheduler30 = typename Gemm::Block::GemmIdentityBlockSwizzle<3, 0>;
using BlockScheduler31 = typename Gemm::Block::GemmIdentityBlockSwizzle<3, 1>;
size_t sizeWA = GetWorkspaceLen(tagA, get<0>(L1TileShape{}), get<2>(L1TileShape{})) * sizeof(int8_t);
size_t sizeWB = GetWorkspaceLen(tagB, get<2>(L1TileShape{}), get<1>(L1TileShape{})) * sizeof(int8_t);
constexpr const uint32_t computeLengthA = 96 * 1024 / sizeof(ElementA);
using PaddingA = Catlass::Gemm::Kernel::PaddingMatrixBlockND<ArchTag, TensorA, computeLengthA>;
constexpr const uint32_t computeLengthB = 96 * 1024 / sizeof(ElementB);
using PaddingB = Catlass::Gemm::Kernel::PaddingMatrixBlockND<ArchTag, TensorB, computeLengthB>;
std::vector<int8_t> hostA(lenA);
std::vector<int8_t> hostB(lenB);
std::vector<float> hostScale(lenScale);
std::vector<float> hostPerTokenScale(lenPerTokenScale);
golden::FillRandomData(hostA, -5, 5);
golden::FillRandomData(hostB, -5, 5);
golden::FillRandomData(hostScale, 0.0, 1.0);
golden::FillRandomData(hostPerTokenScale, 0.0, 1.0);
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 *deviceB{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceB), sizeB, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceB, sizeB, hostB.data(), sizeB, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *deviceScale{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceScale), sizeScale, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(deviceScale, sizeScale, hostScale.data(), sizeScale, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *devicePerTokenScale{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&devicePerTokenScale), sizePerTokenScale, ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMemcpy(
devicePerTokenScale, sizePerTokenScale, hostPerTokenScale.data(), sizePerTokenScale, ACL_MEMCPY_HOST_TO_DEVICE));
uint8_t *deviceD{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceD), sizeD, ACL_MEM_MALLOC_HUGE_FIRST));
uint8_t *deviceWA{nullptr};
if (isNeedPaddingA) {
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceWA), sizeWA, ACL_MEM_MALLOC_HUGE_FIRST));
} else {
deviceWA = deviceA;
}
uint8_t *deviceWB{nullptr};
if (isNeedPaddingB) {
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceWB), sizeWB, ACL_MEM_MALLOC_HUGE_FIRST));
} else {
deviceWB = deviceB;
}
uint64_t fftsAddr{0};
uint32_t fftsLen{0};
RT_CHECK(rtGetC2cCtrlAddr(&fftsAddr, &fftsLen));
auto aicCoreNum = platform_ascendc::PlatformAscendCManager::GetInstance()->GetCoreNumAic();
constexpr uint32_t workspaceStages = 2;
using ArchTag = Arch::AtlasA2;
constexpr bool enableUnitFlag = false;
constexpr bool enableShuffleK = true;
using DispatchPolicy = Gemm::MmadAtlasA2Preload<enableUnitFlag, enableShuffleK>;
constexpr uint32_t ubStages = 2;
using EpilogueDispatchPolicy = Epilogue::EpilogueAtlasA2PerTokenDequantTla<ubStages>;
using ElementCompute = float;
using EpilogueTileShape = MatrixShape<32, 256>;
using TileRowBroadcastMul = Epilogue::Tile::TileRowBroadcastMulTla<ArchTag, ElementCompute, EpilogueTileShape>;
using TileBroadcastOneBlk =
Epilogue::Tile::TileBroadcastOneBlkTla<ArchTag, ElementCompute, EpilogueTileShape::ROW>;
using TileOneBlkColumnBroadcastMul =
Epilogue::Tile::TileOneBlkColumnBroadcastMulTla<ArchTag, ElementCompute, EpilogueTileShape>;
using EpilogueTileCopy = Epilogue::Tile::TileCopyDequantTla<
ArchTag, ElementC, LayoutTagC, ElementScale, LayoutTagScale, ElementPerTokenScale, LayoutTagPerTokenScale,
ElementD, LayoutTagD>;
using EpilogueTileScheduler = Epilogue::Tile::EpilogueHorizontalTileSwizzle;
using BlockEpilogue = Epilogue::Block::BlockEpilogue<
EpilogueDispatchPolicy, ElementC, ElementScale, ElementPerTokenScale, ElementD, TileRowBroadcastMul, TileBroadcastOneBlk,
TileOneBlkColumnBroadcastMul, EpilogueTileCopy, EpilogueTileScheduler>;
if (!isNeedPaddingA && !isNeedPaddingB) {
auto layoutWA = MakeLayout(layoutA.shape(), layoutA.stride());
auto layoutWB = MakeLayout(layoutB.shape(), layoutB.stride());
using TensorWA = Tensor<
AscendC::GlobalTensor<ElementA>, decltype(layoutWA), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TensorWB = Tensor<
AscendC::GlobalTensor<ElementB>, decltype(layoutWB), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TileCopy = Gemm::Tile::PaddingPackedTileCopyTla<
ArchTag, TensorWA, LayoutTagA, TensorWB, LayoutTagB, TensorC, LayoutTagC, void, void, false, false>;
using BlockMmad = Gemm::Block::BlockMmadTla<
DispatchPolicy, L1TileShape, L0TileShape, TensorWA, TensorWB, TensorC, void, TileCopy>;
if (options.problemShape.m() > options.problemShape.n()) {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler30, void, void, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
} else {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler31, void, void, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
}
} else if (!isNeedPaddingA && isNeedPaddingB) {
auto layoutWA = MakeLayout(layoutA.shape(), layoutA.stride());
auto layoutWB = GetPaddingLayout(tagB, get<2>(L1TileShape{}), get<1>(L1TileShape{}));
using TensorWA = Tensor<
AscendC::GlobalTensor<ElementA>, decltype(layoutWA), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TensorWB = Tensor<
AscendC::GlobalTensor<ElementB>, decltype(layoutWB), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TileCopy = Gemm::Tile::PaddingPackedTileCopyTla<
ArchTag, TensorWA, LayoutTagA, TensorWB, LayoutTagB, TensorC, LayoutTagC, void, void, false, true>;
using BlockMmad = Gemm::Block::BlockMmadTla<
DispatchPolicy, L1TileShape, L0TileShape, TensorWA, TensorWB, TensorC, void, TileCopy>;
if (options.problemShape.m() > options.problemShape.n()) {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler30, void, PaddingB, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
} else {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler31, void, PaddingB, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
}
} else if (isNeedPaddingA && !isNeedPaddingB) {
auto layoutWA = GetPaddingLayout(tagA, get<0>(L1TileShape{}), get<2>(L1TileShape{}));
auto layoutWB = MakeLayout(layoutB.shape(), layoutB.stride());
using TensorWA = Tensor<
AscendC::GlobalTensor<ElementA>, decltype(layoutWA), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TensorWB = Tensor<
AscendC::GlobalTensor<ElementB>, decltype(layoutWB), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TileCopy = Gemm::Tile::PaddingPackedTileCopyTla<
ArchTag, TensorWA, LayoutTagA, TensorWB, LayoutTagB, TensorC, LayoutTagC, void, void, true, false>;
using BlockMmad = Gemm::Block::BlockMmadTla<
DispatchPolicy, L1TileShape, L0TileShape, TensorWA, TensorWB, TensorC, void, TileCopy>;
if (options.problemShape.m() > options.problemShape.n()) {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler30, PaddingA, void, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
} else {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler31, PaddingA, void, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
}
} else {
auto layoutWA = GetPaddingLayout(tagA, get<0>(L1TileShape{}), get<2>(L1TileShape{}));
auto layoutWB = GetPaddingLayout(tagB, get<2>(L1TileShape{}), get<1>(L1TileShape{}));
using TensorWA = Tensor<
AscendC::GlobalTensor<ElementA>, decltype(layoutWA), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TensorWB = Tensor<
AscendC::GlobalTensor<ElementB>, decltype(layoutWB), tla::Coord<tla::_0, tla::_0>, AscendC::TPosition::GM>;
using TileCopy = Gemm::Tile::PaddingPackedTileCopyTla<
ArchTag, TensorWA, LayoutTagA, TensorWB, LayoutTagB, TensorC, LayoutTagC, void, void, true, true>;
using BlockMmad = Gemm::Block::BlockMmadTla<
DispatchPolicy, L1TileShape, L0TileShape, TensorWA, TensorWB, TensorC, void, TileCopy>;
if (options.problemShape.m() > options.problemShape.n()) {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler30, PaddingA, PaddingB, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
} else {
using MatmulKernel =
Gemm::Kernel::QuantOptimizedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler31, PaddingA, PaddingB, workspaceStages>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
MatmulKernel::Arguments arguments{
options.problemShape, aicCoreNum,
{deviceA, layoutA, deviceB, layoutB, deviceWA, layoutWA, deviceWB, layoutWB},
{deviceScale, layoutScale, devicePerTokenScale, layoutPerTokenScale, deviceD, layoutD}
};
MatmulAdapter matmulOp;
RunAdapter(matmulOp, arguments, stream, aicCoreNum, fftsAddr);
}
}
ACL_CHECK(aclrtSynchronizeStream(stream));
std::vector<bfloat16> hostD(lenD);
ACL_CHECK(aclrtMemcpy(hostD.data(), sizeD, deviceD, sizeD, ACL_MEMCPY_DEVICE_TO_HOST));
std::vector<float> hostGolden(lenD);
golden::QuantMatmul(
options.problemShape, hostA, tagA, hostB, tagB, hostScale, tagScale,
hostPerTokenScale, tagPerTokenScale, hostGolden, tagD
);
std::vector<uint64_t> errorIndices = golden::CompareData(hostD, 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(deviceD));
if (isNeedPaddingA) {
ACL_CHECK(aclrtFree(deviceWA));
}
if (isNeedPaddingB) {
ACL_CHECK(aclrtFree(deviceWB));
}
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;
}