* 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.
*/
#include "aclnn/opdev/common_types.h"
#include "aclnn/opdev/fp16_t.h"
#include "blas_common.h"
#include "blasplan/include/blasplan/BlasGemmBatchedPlan.h"
#include "hcgemm_batched.h"
#include "cgemm_batched.h"
using namespace AsdSip;
using namespace Mki;
constexpr int TWO = 2;
constexpr int THREE = 3;
constexpr int64_t PADNUM = 128;
namespace AsdSip {
static int64_t GetTrans(const asdBlasOperation_t trans)
{
int64_t gTrans = -1;
if (trans == asdBlasOperation_t::ASDBLAS_OP_N) {
gTrans = 0;
} else if (trans == asdBlasOperation_t::ASDBLAS_OP_T) {
gTrans = 1;
} else if (trans == asdBlasOperation_t::ASDBLAS_OP_C) {
gTrans = TWO;
}
return gTrans;
}
static AspbStatus checkStorage(aclTensor *tensor, const char *tensorName)
{
int64_t *storageDims = nullptr;
uint64_t storageDimsNum = 0;
CHECK_STATUS_WITH_ACL_RETURN(
aclGetStorageShape(tensor, &storageDims, &storageDimsNum), "asdBlasCgemm: aclGetStorageShape");
if (*storageDims <= 0) {
delete[] storageDims;
storageDims = nullptr;
ASDSIP_ELOG(ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH)
<< "blas asdBlasCgemm get wrong " << tensorName << " tensor num.";
return ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH;
}
delete[] storageDims;
return ErrorType::ACL_SUCCESS;
}
static AspbStatus checkDtype(aclTensor &tensor, op::DataType dtype, const char *tensorName)
{
op::DataType dataType = op::DataType::DT_UNDEFINED;
dataType = tensor.GetDataType();
std::string errMsg{"blas asdBlasCgemm get wrong "};
errMsg = errMsg + tensorName + " tensor dtype.";
ASDSIP_ECHECK(dataType == dtype, errMsg, ErrorType::ACL_ERROR_UNSUPPORTED_DATA_TYPE);
return ErrorType::ACL_SUCCESS;
}
AspbStatus asdBlasHCgemmBatched(asdBlasHandle handle, asdBlasOperation_t transa, asdBlasOperation_t transb,
const int64_t m, const int64_t n, const int64_t k, const std::complex<op::fp16_t> &alpha, aclTensor *A,
const int64_t lda, aclTensor *B, const int64_t ldb, const std::complex<op::fp16_t> &beta, aclTensor *C,
const int64_t ldc, const int64_t batchCount)
{
std::lock_guard<std::mutex> lock(blas_mtx);
ASDSIP_ECHECK(BlasPlanCache::doesPlanExist(handle),
"blas HCgemmBatched get cached plan failed.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
AspbStatus status;
if ((status = checkStorage(A, "A")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkStorage(B, "B")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkStorage(C, "C")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkDtype(*A, op::DataType::DT_COMPLEX32, "A")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkDtype(*B, op::DataType::DT_COMPLEX32, "B")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkDtype(*C, op::DataType::DT_COMPLEX32, "C")) != ErrorType::ACL_SUCCESS) {
return status;
}
ASDSIP_ECHECK(m > 0 && k > 0 && n > 0,
"blas CgemmBatched only supports m > 0 && k > 0 && n > 0.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(m <= 32 && k <= 32 && n <= 32,
"blas CgemmBatched only supports m <= 32 && k <= 32 && n <= 32.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
int64_t gTransA = GetTrans(transa);
ASDSIP_ECHECK(gTransA == 0, "blas HCgemmBatched only supports transa == 0.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
int64_t gTransB = GetTrans(transb);
ASDSIP_ECHECK(gTransB == 0, "blas HCgemmBatched only supports transb == 0.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK((alpha == std::complex<op::fp16_t>{1.0f, 0.0f}),
"blas HCgemmBatched only supports alpha == {1.0, 0.0}.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK((beta == std::complex<op::fp16_t>{0.0f, 0.0f}),
"blas HCgemmBatched only supports beta == {0.0, 0.0}.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(lda == k, "blas HCgemmBatched only supports lda == k.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(ldb == n, "blas HCgemmBatched only supports ldb == n.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(ldc == n, "blas HCgemmBatched only supports ldc == n.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
try {
AsdSip::BlasHCgemmBatchedPlan &plan =
dynamic_cast<AsdSip::BlasHCgemmBatchedPlan &>(BlasPlanCache::getPlan(handle));
ASDSIP_ECHECK(plan.IsInitialized(), "HCgemmBatched plan init Error!.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
OpDesc opDesc;
opDesc.opName = "HCgemmBatchedOperation";
AsdSip::OpParam::HCgemmBatched param;
param = {m, k, n, batchCount};
opDesc.specificParam = param;
SVector<aclTensor *> inTensors;
inTensors.push_back(A);
inTensors.push_back(B);
inTensors.push_back(plan.gatherOffsets);
SVector<aclTensor *> outTensors;
outTensors.push_back(C);
ASDSIP_LOG(DEBUG) << "OpDesc: " << opDesc.opName << "; OpDesc info: " << param.ToString();
Status HCgemmBatchedStatus = RunAsdOpsV2(plan.GetStream(), opDesc, inTensors, outTensors, plan.GetWorkspace());
ASDSIP_ECHECK(HCgemmBatchedStatus.Ok(), HCgemmBatchedStatus.Message(), ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_LOG(INFO) << "Execute asdBlasHCgemmBatched success.";
return ErrorType::ACL_SUCCESS;
} catch (std::bad_cast &e) {
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "HCgemmBatched Error: " << e.what();
return ErrorType::ACL_ERROR_INTERNAL_ERROR;
}
}
AspbStatus asdBlasMakeHCgemmBatchedPlan(asdBlasHandle handle)
{
std::lock_guard<std::mutex> lock(blas_mtx);
ASDSIP_ECHECK(handle != nullptr, "blas MakeHCgemmBatchedPlan failed.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
AsdSip::BlasHCgemmBatchedPlan *plan = nullptr;
try {
plan = new AsdSip::BlasHCgemmBatchedPlan();
BlasPlanCache::MakePlan(handle, plan);
} catch (const std::exception &e) {
if (plan != nullptr) {
delete plan;
}
delete static_cast<int *>(handle);
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "Make HCgemmBatched Plan failed: " << e.what();
throw std::runtime_error("Make HCgemmBatched Plan failed.");
}
if (plan->CreateTensor() != ErrorType::ACL_SUCCESS) {
plan->MarkFailed();
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "Fail to create blas asdBlasMakeHCgemmBatchedPlan mask tensor.";
return ErrorType::ACL_ERROR_INTERNAL_ERROR;
}
plan->MarkInitialized();
return ErrorType::ACL_SUCCESS;
}
AspbStatus asdBlasCgemmBatched(asdBlasHandle handle, asdBlasOperation_t transa, asdBlasOperation_t transb,
const int64_t m, const int64_t n, const int64_t k, const std::complex<float> &alpha, aclTensor *A,
const int64_t lda, aclTensor *B, const int64_t ldb, const std::complex<float> &beta, aclTensor *C,
const int64_t ldc, const int64_t batchCount)
{
std::lock_guard<std::mutex> lock(blas_mtx);
ASDSIP_ECHECK(BlasPlanCache::doesPlanExist(handle),
"blas CgemmBatched get cached plan failed.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
AspbStatus status;
if ((status = checkStorage(A, "A")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkStorage(B, "B")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkStorage(C, "C")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkDtype(*A, op::DataType::DT_COMPLEX64, "A")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkDtype(*B, op::DataType::DT_COMPLEX64, "B")) != ErrorType::ACL_SUCCESS) {
return status;
}
if ((status = checkDtype(*C, op::DataType::DT_COMPLEX64, "C")) != ErrorType::ACL_SUCCESS) {
return status;
}
ASDSIP_ECHECK(m > 0 && k > 0 && n > 0,
"blas CgemmBatched only supports m > 0 && k > 0 && n > 0.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(m <= 32 && k <= 32 && n <= 32,
"blas CgemmBatched only supports m <= 32 && k <= 32 && n <= 32.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
int64_t gTransA = GetTrans(transa);
ASDSIP_ECHECK(gTransA == 0, "blas CgemmBatched only supports transa == 0.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
int64_t gTransB = GetTrans(transb);
ASDSIP_ECHECK(gTransB == 0, "blas CgemmBatched only supports transb == 0.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK((alpha == std::complex<float>{1.0f, 0.0f}),
"blas CgemmBatched only supports alpha == {1.0, 0.0}.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK((beta == std::complex<float>{0.0f, 0.0f}),
"blas CgemmBatched only supports beta == {0.0, 0.0}.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(lda == k, "blas CgemmBatched only supports lda == k.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(ldb == n, "blas CgemmBatched only supports ldb == n.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(ldc == n, "blas CgemmBatched only supports ldc == n.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
try {
AsdSip::BlasCgemmBatchedPlan &plan =
dynamic_cast<AsdSip::BlasCgemmBatchedPlan &>(BlasPlanCache::getPlan(handle));
ASDSIP_ECHECK(plan.IsInitialized(), "CgemmBatched plan init Error!.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
OpDesc opDesc;
opDesc.opName = "CgemmBatchedOperation";
AsdSip::OpParam::CgemmBatched param;
param = {m, k, n, batchCount};
opDesc.specificParam = param;
SVector<aclTensor *> inTensors;
inTensors.push_back(A);
inTensors.push_back(B);
inTensors.push_back(plan.gatherOffsets);
SVector<aclTensor *> outTensors;
outTensors.push_back(C);
ASDSIP_LOG(DEBUG) << "OpDesc: " << opDesc.opName << "; OpDesc info: " << param.ToString();
Status CgemmBatchedStatus = RunAsdOpsV2(plan.GetStream(), opDesc, inTensors, outTensors, plan.GetWorkspace());
ASDSIP_ECHECK(CgemmBatchedStatus.Ok(), CgemmBatchedStatus.Message(), ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_LOG(INFO) << "Execute asdBlasCgemmBatched success.";
return ErrorType::ACL_SUCCESS;
} catch (std::bad_cast &e) {
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "CgemmBatched Error: " << e.what();
return ErrorType::ACL_ERROR_INTERNAL_ERROR;
}
}
AspbStatus asdBlasMakeCgemmBatchedPlan(asdBlasHandle handle)
{
std::lock_guard<std::mutex> lock(blas_mtx);
ASDSIP_ECHECK(handle != nullptr, "blas MakeCgemmBatchedPlan failed.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
AsdSip::BlasCgemmBatchedPlan *plan = nullptr;
try {
plan = new AsdSip::BlasCgemmBatchedPlan();
BlasPlanCache::MakePlan(handle, plan);
} catch (const std::exception &e) {
if (plan != nullptr) {
delete plan;
}
delete static_cast<int *>(handle);
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "Make CgemmBatched Plan failed: " << e.what();
throw std::runtime_error("Make CgemmBatched Plan failed.");
}
if (plan->CreateTensor() != ErrorType::ACL_SUCCESS) {
plan->MarkFailed();
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "Fail to create blas asdBlasMakeCgemmBatchedPlan mask tensor.";
return ErrorType::ACL_ERROR_INTERNAL_ERROR;
}
plan->MarkInitialized();
return ErrorType::ACL_SUCCESS;
}
}
