* 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/strided_batched_matmul_tla.hpp"
#include "catlass/arch/arch.hpp"
#include "catlass/catlass.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/layout/layout.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;
struct BatchedMatmulTlaOptions {
const std::string HELPER =
"problem_count m n k [device_id] [lda ldb ldc] [strideA strideB strideC] [layoutA layoutB]\n"
" layoutA/layoutB: row | col (case-insensitive)\n"
" Note: C is row in this example (ldc>=n, strideC based on row).";
Catlass::GemmCoord problemShape{128, 128, 128};
uint32_t problemCount{1};
int32_t deviceId{0};
int64_t lda{-1};
int64_t ldb{-1};
int64_t ldc{-1};
int64_t strideA{-1};
int64_t strideB{-1};
int64_t strideC{-1};
enum class MatrixLayout { RowMajor, ColumnMajor };
MatrixLayout layoutA{MatrixLayout::RowMajor};
MatrixLayout layoutB{MatrixLayout::RowMajor};
BatchedMatmulTlaOptions() = default;
static bool IsLayoutToken(const std::string &s) {
if (s.empty()) {
return false;
}
std::string t;
t.resize(s.size());
std::transform(s.begin(), s.end(), t.begin(), [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
return (t == "row" || t == "col");
}
static MatrixLayout ParseLayoutToken(const std::string &s) {
std::string t;
t.resize(s.size());
std::transform(s.begin(), s.end(), t.begin(), [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
if (t == "row") {
return MatrixLayout::RowMajor;
}
return MatrixLayout::ColumnMajor;
}
int Parse(int argc, const char **argv) {
int argcEffective = argc;
if (argc >= 7) {
std::string maybeA(argv[argc - 2]);
std::string maybeB(argv[argc - 1]);
if (IsLayoutToken(maybeA) && IsLayoutToken(maybeB)) {
layoutA = ParseLayoutToken(maybeA);
layoutB = ParseLayoutToken(maybeB);
argcEffective -= 2;
}
}
if (!(argcEffective == 5 || argcEffective == 6 || argcEffective == 9 || argcEffective == 12)) {
std::cerr << TOSTRING(CATLASS_EXAMPLE_NAME) << " " << HELPER << std::endl;
return -1;
}
problemCount = std::atoi(argv[1]);
problemShape.m() = std::atoi(argv[2]);
problemShape.n() = std::atoi(argv[3]);
problemShape.k() = std::atoi(argv[4]);
if (argcEffective >= 6) {
deviceId = std::atoi(argv[5]);
}
uint32_t m = problemShape.m();
uint32_t n = problemShape.n();
uint32_t k = problemShape.k();
lda = (layoutA == MatrixLayout::RowMajor) ? static_cast<int64_t>(k) : static_cast<int64_t>(m);
ldb = (layoutB == MatrixLayout::RowMajor) ? static_cast<int64_t>(n) : static_cast<int64_t>(k);
ldc = static_cast<int64_t>(n);
strideA = (layoutA == MatrixLayout::RowMajor) ? static_cast<int64_t>(m) * lda : static_cast<int64_t>(k) * lda;
strideB = (layoutB == MatrixLayout::RowMajor) ? static_cast<int64_t>(k) * ldb : static_cast<int64_t>(n) * ldb;
strideC = static_cast<int64_t>(m) * ldc;
if (argcEffective >= 9) {
lda = std::atoll(argv[6]);
ldb = std::atoll(argv[7]);
ldc = std::atoll(argv[8]);
strideA = (layoutA == MatrixLayout::RowMajor) ? static_cast<int64_t>(m) * lda : static_cast<int64_t>(k) * lda;
strideB = (layoutB == MatrixLayout::RowMajor) ? static_cast<int64_t>(k) * ldb : static_cast<int64_t>(n) * ldb;
strideC = static_cast<int64_t>(m) * ldc;
}
if (argcEffective == 12) {
strideA = std::atoll(argv[9]);
strideB = std::atoll(argv[10]);
strideC = std::atoll(argv[11]);
}
int64_t minLda = (layoutA == MatrixLayout::RowMajor) ? static_cast<int64_t>(k) : static_cast<int64_t>(m);
int64_t minLdb = (layoutB == MatrixLayout::RowMajor) ? static_cast<int64_t>(n) : static_cast<int64_t>(k);
if (lda < minLda || ldb < minLdb || ldc < static_cast<int64_t>(n)) {
std::cerr << "Invalid leading dimensions: require lda>=" << minLda
<< ", ldb>=" << minLdb
<< ", ldc>=" << n << "." << std::endl;
return -1;
}
int64_t minMatA = (layoutA == MatrixLayout::RowMajor)
? (static_cast<int64_t>(m - 1) * lda + static_cast<int64_t>(k))
: (static_cast<int64_t>(k - 1) * lda + static_cast<int64_t>(m));
int64_t minMatB = (layoutB == MatrixLayout::RowMajor)
? (static_cast<int64_t>(k - 1) * ldb + static_cast<int64_t>(n))
: (static_cast<int64_t>(n - 1) * ldb + static_cast<int64_t>(k));
int64_t minMatC = static_cast<int64_t>(m - 1) * ldc + static_cast<int64_t>(n);
if (strideA < minMatA || strideB < minMatB || strideC < minMatC) {
std::cerr << "Invalid batch strides: require strideA/strideB/strideC large enough for one matrix."
<< std::endl;
return -1;
}
return 0;
}
};
using Options = BatchedMatmulTlaOptions;
template <typename LayoutTagA>
static auto MakeTlaLayoutA(uint32_t batchCount, uint32_t m, uint32_t k, int64_t strideA, int64_t lda) {
if constexpr (std::is_same_v<LayoutTagA, layout::RowMajor>) {
return tla::MakeLayout(
tla::MakeShape(batchCount, m, k),
tla::MakeStride(strideA, lda, tla::Int<1>{})
);
} else {
return tla::MakeLayout(
tla::MakeShape(batchCount, m, k),
tla::MakeStride(strideA, tla::Int<1>{}, lda)
);
}
}
template <typename LayoutTagB>
static auto MakeTlaLayoutB(uint32_t batchCount, uint32_t k, uint32_t n, int64_t strideB, int64_t ldb) {
if constexpr (std::is_same_v<LayoutTagB, layout::RowMajor>) {
return tla::MakeLayout(
tla::MakeShape(batchCount, k, n),
tla::MakeStride(strideB, ldb, tla::Int<1>{})
);
} else {
return tla::MakeLayout(
tla::MakeShape(batchCount, k, n),
tla::MakeStride(strideB, tla::Int<1>{}, ldb)
);
}
}
template <typename LayoutTagA, typename LayoutTagB>
static void RunWithLayouts(const Options &options) {
aclrtStream stream{nullptr};
ACL_CHECK(aclInit(nullptr));
ACL_CHECK(aclrtSetDevice(options.deviceId));
ACL_CHECK(aclrtCreateStream(&stream));
uint32_t batchCount = options.problemCount;
uint32_t m = options.problemShape.m();
uint32_t n = options.problemShape.n();
uint32_t k = options.problemShape.k();
using ElementA = half;
using ElementB = half;
using ElementC = half;
using HostElementA = fp16_t;
using HostElementB = fp16_t;
using HostElementC = fp16_t;
using LayoutTagC = layout::RowMajor;
LayoutTagA tagA{m, k, options.lda};
LayoutTagB tagB{k, n, options.ldb};
LayoutTagC tagC{m, n, options.ldc};
int64_t capA = (static_cast<int64_t>(batchCount) - 1) * options.strideA +
static_cast<int64_t>(tagA.GetOffset(MakeCoord(m - 1, k - 1))) + 1;
int64_t capB = (static_cast<int64_t>(batchCount) - 1) * options.strideB +
static_cast<int64_t>(tagB.GetOffset(MakeCoord(k - 1, n - 1))) + 1;
int64_t capC = (static_cast<int64_t>(batchCount) - 1) * options.strideC +
static_cast<int64_t>(tagC.GetOffset(MakeCoord(m - 1, n - 1))) + 1;
size_t lenA = static_cast<size_t>(capA);
size_t lenB = static_cast<size_t>(capB);
size_t lenC = static_cast<size_t>(capC);
size_t sizeA = lenA * sizeof(ElementA);
size_t sizeB = lenB * sizeof(ElementB);
size_t sizeC = lenC * sizeof(ElementC);
std::vector<HostElementA> hostA(lenA, 1.0f);
golden::FillRandomData<HostElementA>(hostA, -5.0f, 5.0f);
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));
std::vector<HostElementB> hostB(lenB, 1.0f);
golden::FillRandomData<HostElementB>(hostB, -5.0f, 5.0f);
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));
std::vector<HostElementC> hostC(lenC);
uint8_t *deviceC{nullptr};
ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceC), sizeC, ACL_MEM_MALLOC_HUGE_FIRST));
auto aicCoreNum = platform_ascendc::PlatformAscendCManager::GetInstance()->GetCoreNumAic();
using ArchTag = Arch::AtlasA2;
using DispatchPolicy = Gemm::MmadPingpongTlaV2<ArchTag, true>;
using L1TileShape = Shape<_128, _256, _256>;
using L0TileShape = Shape<_128, _256, _64>;
using TileCopy =
Gemm::Tile::PackedTileCopyTla<ArchTag, ElementA, LayoutTagA, ElementB, LayoutTagB, ElementC, LayoutTagC>;
using BlockMmad = Gemm::Block::BlockMmadTla<
DispatchPolicy, L1TileShape, L0TileShape, ElementA, ElementB, ElementC, void, TileCopy>;
using BlockEpilogue = void;
auto layoutA = MakeTlaLayoutA<LayoutTagA>(batchCount, m, k, options.strideA, options.lda);
auto layoutB = MakeTlaLayoutB<LayoutTagB>(batchCount, k, n, options.strideB, options.ldb);
auto layoutC = tla::MakeLayout(
tla::MakeShape(batchCount, m, n),
tla::MakeStride(options.strideC, options.ldc, tla::Int<1>{})
);
if (options.problemShape.m() > options.problemShape.n()) {
using BlockScheduler = typename Gemm::Block::GemmIdentityBlockSwizzle<3, 0>;
using MatmulKernel = Gemm::Kernel::StridedBatchedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
typename MatmulKernel::Arguments arguments{
batchCount, options.problemShape,
deviceA, layoutA,
deviceB, layoutB,
deviceC, layoutC
};
MatmulAdapter matmulOp;
uint8_t *deviceWorkspace{nullptr};
matmulOp.CanImplement(arguments);
matmulOp.Initialize(arguments, deviceWorkspace);
matmulOp(stream, aicCoreNum);
ACL_CHECK(aclrtSynchronizeStream(stream));
} else {
using BlockScheduler = typename Gemm::Block::GemmIdentityBlockSwizzle<3, 1>;
using MatmulKernel = Gemm::Kernel::StridedBatchedMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler>;
using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;
typename MatmulKernel::Arguments arguments{
batchCount, options.problemShape,
deviceA, layoutA,
deviceB, layoutB,
deviceC, layoutC
};
MatmulAdapter matmulOp;
uint8_t *deviceWorkspace{nullptr};
matmulOp.CanImplement(arguments);
matmulOp.Initialize(arguments, deviceWorkspace);
matmulOp(stream, aicCoreNum);
ACL_CHECK(aclrtSynchronizeStream(stream));
}
ACL_CHECK(aclrtMemcpy(hostC.data(), sizeC, deviceC, sizeC, ACL_MEMCPY_DEVICE_TO_HOST));
size_t packedLenC = static_cast<size_t>(batchCount) * m * n;
std::vector<HostElementC> packedC(packedLenC);
std::vector<float> packedGolden(packedLenC);
for (uint32_t b = 0; b < batchCount; ++b) {
size_t basePacked = static_cast<size_t>(b) * m * n;
size_t baseA = static_cast<size_t>(b) * static_cast<size_t>(options.strideA);
size_t baseB = static_cast<size_t>(b) * static_cast<size_t>(options.strideB);
size_t baseC = static_cast<size_t>(b) * static_cast<size_t>(options.strideC);
for (uint32_t i = 0; i < m; ++i) {
for (uint32_t j = 0; j < n; ++j) {
size_t idxPacked = basePacked + static_cast<size_t>(i) * n + j;
size_t offC = baseC + static_cast<size_t>(tagC.GetOffset(MakeCoord(i, j)));
packedC[idxPacked] = hostC[offC];
float acc = 0.0f;
for (uint32_t kk = 0; kk < k; ++kk) {
size_t offA = baseA + static_cast<size_t>(tagA.GetOffset(MakeCoord(i, kk)));
size_t offB = baseB + static_cast<size_t>(tagB.GetOffset(MakeCoord(kk, j)));
acc += static_cast<float>(hostA[offA]) * static_cast<float>(hostB[offB]);
}
packedGolden[idxPacked] = acc;
}
}
}
std::vector<uint64_t> errorIndices = golden::CompareData(packedC, packedGolden, 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));
ACL_CHECK(aclrtDestroyStream(stream));
ACL_CHECK(aclrtResetDevice(options.deviceId));
ACL_CHECK(aclFinalize());
}
static void Run(const Options &options) {
using ML = Options::MatrixLayout;
if (options.layoutA == ML::RowMajor && options.layoutB == ML::RowMajor) {
RunWithLayouts<layout::RowMajor, layout::RowMajor>(options);
} else if (options.layoutA == ML::RowMajor && options.layoutB == ML::ColumnMajor) {
RunWithLayouts<layout::RowMajor, layout::ColumnMajor>(options);
} else if (options.layoutA == ML::ColumnMajor && options.layoutB == ML::RowMajor) {
RunWithLayouts<layout::ColumnMajor, layout::RowMajor>(options);
} else {
RunWithLayouts<layout::ColumnMajor, layout::ColumnMajor>(options);
}
}
int main(int argc, const char **argv) {
Options options;
if (options.Parse(argc, argv) != 0) {
return -1;
}
Run(options);
return 0;
}