* 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 "cgemv_batched.h"
#include "blas_common.h"
#include "blasplan/include/blasplan/BlasCgemvBatchedPlan.h"
using namespace AsdSip;
using namespace Mki;
static constexpr uint32_t DIM_0 = 0;
static constexpr uint32_t DIM_1 = 1;
static constexpr uint32_t DIM_2 = 2;
static constexpr uint32_t DIM_3 = 3;
static constexpr uint32_t MAX_ELENUM_INLINE = 32;
namespace AsdSip {
struct CgemvBatchedImplParam {
asdBlasOperation_t trans;
asdDataType_t dtype;
int64_t m;
int64_t n;
int64_t batchCount;
aclTensor *A;
aclTensor *x;
aclTensor *y;
};
AspbStatus asdBlasCgemvBatchedShapeCheck(CgemvBatchedImplParam implParam)
{
ASDSIP_ECHECK(implParam.m > 0 && implParam.n > 0 && implParam.m <= MAX_ELENUM_INLINE &&
implParam.n <= MAX_ELENUM_INLINE && implParam.batchCount > 0,
"blas asdBlasCgemvBatched get m <= 0 || n <= 0 || m > 32 || n > 32 || batchCount <= 0.",
ErrorType::ACL_ERROR_INVALID_PARAM);
int64_t *xstorageDims = nullptr;
uint64_t xstorageDimsNum = 0;
CHECK_STATUS_WITH_ACL_RETURN(
aclGetStorageShape(implParam.x, &xstorageDims, &xstorageDimsNum), "asdBlasCgemvBatched: aclGetStorageShape");
if (xstorageDimsNum > DIM_3 || *xstorageDims <= 0) {
delete[] xstorageDims;
xstorageDims = nullptr;
ASDSIP_ELOG(ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH) << "blas asdBlasCgemvBatched get wrong x tensor num.";
return ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH;
}
if (xstorageDims != nullptr) {
delete[] xstorageDims;
xstorageDims = nullptr;
}
int64_t *ystorageDims = nullptr;
uint64_t ystorageDimsNum = 0;
CHECK_STATUS_WITH_ACL_RETURN(
aclGetStorageShape(implParam.y, &ystorageDims, &ystorageDimsNum), "asdBlasCgemvBatched: aclGetStorageShape");
if (ystorageDimsNum > DIM_3 || *ystorageDims <= 0) {
delete[] ystorageDims;
ystorageDims = nullptr;
ASDSIP_ELOG(ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH) << "blas asdBlasCgemvBatched get wrong y tensor num.";
return ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH;
}
if (ystorageDims != nullptr) {
delete[] ystorageDims;
ystorageDims = nullptr;
}
int64_t *AstorageDims = nullptr;
uint64_t AstorageDimsNum = 0;
CHECK_STATUS_WITH_ACL_RETURN(
aclGetStorageShape(implParam.A, &AstorageDims, &AstorageDimsNum), "asdBlasCgemvBatched: aclGetStorageShape");
if (AstorageDimsNum != DIM_3 || AstorageDims[DIM_0] != implParam.batchCount || AstorageDims[DIM_1] != implParam.m ||
AstorageDims[DIM_2] != implParam.n) {
delete[] AstorageDims;
AstorageDims = nullptr;
ASDSIP_ELOG(ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH) << "blas asdBlasCgemvBatched get wrong A tensor num.";
return ErrorType::ACL_ERROR_OP_INPUT_NOT_MATCH;
}
if (AstorageDims != nullptr) {
delete[] AstorageDims;
AstorageDims = nullptr;
}
return ErrorType::ACL_SUCCESS;
}
AspbStatus asdBlasCgemvBatchedImpl(asdBlasHandle handle, CgemvBatchedImplParam implParam)
{
ASDSIP_ECHECK(BlasPlanCache::doesPlanExist(handle),
"blas CgemvBatched get cached plan failed.",
ErrorType::ACL_ERROR_INTERNAL_ERROR);
AsdSip::AspbStatus shapeCheckStatus = asdBlasCgemvBatchedShapeCheck(implParam);
if (shapeCheckStatus != AsdSip::ErrorType::ACL_SUCCESS) {
return shapeCheckStatus;
}
try {
AsdSip::BlasCgemvBatchedPlan &plan =
dynamic_cast<AsdSip::BlasCgemvBatchedPlan &>(BlasPlanCache::getPlan(handle));
ASDSIP_ECHECK(plan.IsInitialized(), "CgemvBatched plan init Error!.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
Status status;
OpDesc opDesc;
opDesc.opName = "CgemvBatchedOperation";
AsdSip::OpParam::CgemvBatched param;
int64_t transType = static_cast<int64_t>(implParam.trans);
int64_t dataType = static_cast<int64_t>(implParam.dtype);
param = {transType, dataType, implParam.m, implParam.n, implParam.batchCount, plan.GetMaxMatNum()};
opDesc.specificParam = param;
ASDSIP_LOG(DEBUG) << "OpDesc: " << opDesc.opName << "; OpDesc info: " << param.ToString();
SVector<aclTensor *> inTensors{implParam.A, implParam.x, plan.GetAclMaskTensor()};
SVector<aclTensor *> outTensors{implParam.y};
status = RunAsdOpsV2(plan.GetStream(), opDesc, inTensors, outTensors, plan.GetWorkspace());
ASDSIP_ECHECK(status.Ok(), status.Message(), ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_LOG(INFO) << "Execute asdBlasCgemvBatched success.";
return ErrorType::ACL_SUCCESS;
} catch (std::bad_cast &e) {
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "CgemvBatched Error: " << e.what();
return ErrorType::ACL_ERROR_INTERNAL_ERROR;
}
}
AspbStatus asdBlasHCgemvBatched(asdBlasHandle handle, asdBlasOperation_t trans, const int64_t m, const int64_t n,
const std::complex<op::fp16_t> &alpha, aclTensor *A, const int64_t lda, aclTensor *x, const int64_t incx,
const std::complex<op::fp16_t> &beta, aclTensor *y, const int64_t incy, const int64_t batchCount)
{
(void)alpha;
(void)beta;
std::lock_guard<std::mutex> lock(blas_mtx);
if (lda <= 0) {
ASDSIP_LOG(INFO) << "blas asdBlasHCgemvBatched get lda <= 0.";
}
if (incx <= 0) {
ASDSIP_LOG(INFO) << "blas asdBlasHCgemvBatched get incx <= 0.";
}
if (incy <= 0) {
ASDSIP_LOG(INFO) << "blas asdBlasHCgemvBatched get incy <= 0.";
}
return asdBlasCgemvBatchedImpl(handle, {trans, asdDataType_t::ASD_C_32F, m, n, batchCount, A, x, y});
}
AspbStatus asdBlasCgemvBatched(asdBlasHandle handle, asdBlasOperation_t trans, const int64_t m, const int64_t n,
const std::complex<float> &alpha, aclTensor *A, const int64_t lda, aclTensor *x, const int64_t incx,
const std::complex<float> &beta, aclTensor *y, const int64_t incy, const int64_t batchCount)
{
(void)alpha;
(void)beta;
std::lock_guard<std::mutex> lock(blas_mtx);
aclDataType dataType = aclDataType::ACL_DT_UNDEFINED;
CHECK_STATUS_WITH_ACL_RETURN(aclGetDataType(x, &dataType), "asdBlasCgemvBatched: aclGetDataType");
ASDSIP_ECHECK(dataType == aclDataType::ACL_COMPLEX64,
"blas asdBlasCgemvBatched get wrong x tensor dtype.",
ErrorType::ACL_ERROR_UNSUPPORTED_DATA_TYPE);
CHECK_STATUS_WITH_ACL_RETURN(aclGetDataType(y, &dataType), "asdBlasCgemvBatched: aclGetDataType");
ASDSIP_ECHECK(dataType == aclDataType::ACL_COMPLEX64,
"blas asdBlasCgemvBatched get wrong y tensor dtype.",
ErrorType::ACL_ERROR_UNSUPPORTED_DATA_TYPE);
CHECK_STATUS_WITH_ACL_RETURN(aclGetDataType(A, &dataType), "asdBlasCgemvBatched: aclGetDataType");
ASDSIP_ECHECK(dataType == aclDataType::ACL_COMPLEX64,
"blas asdBlasCgemvBatched get wrong A tensor dtype.",
ErrorType::ACL_ERROR_UNSUPPORTED_DATA_TYPE);
if (lda <= 0) {
ASDSIP_LOG(INFO) << "blas asdBlasCgemvBatched get lda <= 0.";
}
if (incx <= 0) {
ASDSIP_LOG(INFO) << "blas asdBlasCgemvBatched get incx <= 0.";
}
if (incy <= 0) {
ASDSIP_LOG(INFO) << "blas asdBlasCgemvBatched get incy <= 0.";
}
return asdBlasCgemvBatchedImpl(handle, {trans, asdDataType_t::ASD_C_64F, m, n, batchCount, A, x, y});
}
AspbStatus asdBlasMakeCgemvBatchedPlanImpl(
asdBlasHandle handle, asdBlasOperation_t trans, asdDataType_t dtype, const int64_t m)
{
ASDSIP_ECHECK(handle != nullptr, "blas CgemvBatched Make Plan failed.", ErrorType::ACL_ERROR_INTERNAL_ERROR);
ASDSIP_ECHECK(m > 0, "blas asdBlasMakeCgemvBatchedPlan get m <= 0.", ErrorType::ACL_ERROR_INVALID_PARAM);
AsdSip::BlasCgemvBatchedPlan *plan = nullptr;
try {
plan = new AsdSip::BlasCgemvBatchedPlan(trans, dtype, m);
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 CgemvBatched Plan failed: " << e.what();
throw std::runtime_error("Make CgemvBatched Plan failed.");
}
if (plan->CreateTensor() != ErrorType::ACL_SUCCESS) {
plan->MarkFailed();
ASDSIP_ELOG(ErrorType::ACL_ERROR_INTERNAL_ERROR) << "Fail to create blas CgemvBatchedPlan mask tensor.";
return ErrorType::ACL_ERROR_INTERNAL_ERROR;
}
plan->MarkInitialized();
return ErrorType::ACL_SUCCESS;
}
AspbStatus asdBlasMakeHCgemvBatchedPlan(asdBlasHandle handle, asdBlasOperation_t trans, const int64_t m)
{
std::lock_guard<std::mutex> lock(blas_mtx);
return asdBlasMakeCgemvBatchedPlanImpl(handle, trans, asdDataType_t::ASD_C_32F, m);
}
AspbStatus asdBlasMakeCgemvBatchedPlan(asdBlasHandle handle, asdBlasOperation_t trans, const int64_t m)
{
std::lock_guard<std::mutex> lock(blas_mtx);
return asdBlasMakeCgemvBatchedPlanImpl(handle, trans, asdDataType_t::ASD_C_64F, m);
}
}
