Hello，I tried to call cusparseSpGEMM and I followed this example and wrote a code using openacc (Problems in the use of cusparseSpGEMM in CUDA Fortran).
The following error occurred: Segmentation fault (core dumped)
I don’t know how to solve it. Looking forward to your answers. Thank you very much!
Here is my code:
The computational environment is V100 16GB with CUDA 12.0
program spgemm_text
use openacc
use cusparse
implicit none
type(cusparseHandle) :: handle
type(cusparseSpMatDescr) :: A_matrix,B_matrix,C_matrix
type(cusparseSpGEMMDescr) :: SpGEMMDescr
integer(kind=8):: istat,i
integer(kind=8):: A_num_rows = 4
integer(kind=8):: A_num_cols = 4
integer(kind=8):: A_nnz = 9
integer(kind=8):: B_num_rows = 4
integer(kind=8):: B_num_cols = 4
integer(kind=8):: B_nnz = 8
integer(kind=8):: C_num_rows = 4
integer(kind=8):: C_num_cols = 4
integer(kind=8):: C_nnz
integer(kind=8), dimension(:), allocatable :: A_csrOffsets ,A_columns
integer(kind=8), dimension(:), allocatable :: B_csrOffsets ,B_columns
integer(kind=8), dimension(:), allocatable :: C_csrOffsets ,C_columns
real(kind=8), dimension(:), allocatable :: A_values, B_values, C_values
real(kind=8)::alpha=1.0,beta=0.0
integer(kind=8) :: bufferSize1
integer(1), pointer :: buffer1(:)
integer(kind=8) :: bufferSize2
integer(1), pointer :: buffer2(:)
ALLOCATE(A_csrOffsets(A_num_rows + 1), A_columns(A_nnz), A_values(A_nnz))
ALLOCATE(B_csrOffsets(B_num_rows + 1), B_columns(B_nnz), B_values(B_nnz))
A_csrOffsets = [1, 4, 5, 8, 10]
A_columns = [1, 3, 4, 2, 1, 3, 4, 2, 4]
A_values = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0]
B_csrOffsets = [1, 3, 5, 8, 10]
B_columns = [1, 4, 2, 4, 1, 2, 3, 2]
B_values = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0]
C_columns=0
C_csrOffsets=0
C_values=0.0
!$acc data copyin(A_csrOffsets, B_csrOffsets, A_columns, B_columns, A_values, B_values, C_csrOffsets, C_columns, C_values)
istat = cusparseCreate(handle)
istat=cusparseSpGEMM_createDescr(SpGEMMDescr)
istat=cusparseCreateCsr(A_matrix, A_num_rows, A_num_cols, A_nnz, A_csrOffsets, A_columns,A_values,&
CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ONE, CUDA_R_64F)
istat=cusparseCreateCsr(B_matrix, B_num_rows, B_num_cols, B_nnz, B_csrOffsets, B_columns,B_values,&
CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ONE, CUDA_R_64F)
istat=cusparseCreateCsr(C_matrix, C_num_rows, C_num_cols, 0,&
null(),null(),null(),&
!C_csrOffsets, C_columns,C_values,&
CUSPARSE_INDEX_32I, CUSPARSE_INDEX_32I, CUSPARSE_INDEX_BASE_ONE, CUDA_R_64F)
if(istat/=CUSPARSE_STATUS_SUCCESS) print *, 'cusparseCreateCsr error: ', istat
nullify(buffer1)
istat=cusparseSpGEMM_workEstimation(handle,&
CUSPARSE_OPERATION_NON_TRANSPOSE,CUSPARSE_OPERATION_NON_TRANSPOSE,&
alpha,A_matrix,B_matrix,beta,C_matrix,&
CUDA_R_64F,CUSPARSE_SPGEMM_DEFAULT,&
SpGEMMDescr,bufferSize1,buffer1)
!SpGEMMDescr,bufferSize1,null())
if(istat/=CUSPARSE_STATUS_SUCCESS) print *, 'cusparseSpGEMM_workEstimation error: ', istat
ALLOCATE(buffer1(bufferSize1))
!$acc data create(buffer1)
istat=cusparseSpGEMM_workEstimation(handle,&
CUSPARSE_OPERATION_NON_TRANSPOSE,CUSPARSE_OPERATION_NON_TRANSPOSE,&
alpha,A_matrix,B_matrix,beta,C_matrix,&
CUDA_R_64F,CUSPARSE_SPGEMM_DEFAULT,&
SpGEMMDescr,bufferSize1,buffer1)
if(istat/=CUSPARSE_STATUS_SUCCESS) print *, 'cusparseSpGEMM_workEstimation error: ', istat
nullify(buffer2)
istat=cusparseSpGEMM_compute(handle,&
CUSPARSE_OPERATION_NON_TRANSPOSE,CUSPARSE_OPERATION_NON_TRANSPOSE,&
alpha,A_matrix,B_matrix,beta,C_matrix,&
CUDA_R_64F,CUSPARSE_SPGEMM_DEFAULT,&
SpGEMMDescr,bufferSize2,buffer2)
!SpGEMMDescr,bufferSize2,null())
ALLOCATE(buffer2(bufferSize2))
!$acc data create(buffer2)
istat=cusparseSpGEMM_compute(handle,&
CUSPARSE_OPERATION_NON_TRANSPOSE,CUSPARSE_OPERATION_NON_TRANSPOSE,&
alpha,A_matrix,B_matrix,beta,C_matrix,&
CUDA_R_64F,CUSPARSE_SPGEMM_DEFAULT,&
SpGEMMDescr,bufferSize2,buffer2)
if(istat/=CUSPARSE_STATUS_SUCCESS) print *, 'cusparseSpGEMM_workEstimation error: ', istat
istat=cusparseSpMatGetSize(C_matrix,C_num_rows,C_num_cols,C_nnz)
allocate(C_csrOffsets(C_num_rows+1),C_columns(C_nnz),C_values(C_nnz))
istat=cusparseCsrSetPointers(C_matrix, C_csrOffsets, C_columns, C_values)
istat=cusparseSpGEMM_copy(handle,&
CUSPARSE_OPERATION_NON_TRANSPOSE,CUSPARSE_OPERATION_NON_TRANSPOSE,&
alpha,A_matrix,B_matrix,beta,C_matrix,&
CUDA_R_64F,CUSPARSE_SPGEMM_DEFAULT,SpGEMMDescr)
deallocate(buffer1(bufferSize1))
deallocate(buffer2(bufferSize2))
istat=cusparseSpGEMM_destroyDescr(SpGEMMDescr)
istat=cusparseDestroySpMat(A_matrix)
istat=cusparseDestroySpMat(B_matrix)
istat=cusparseDestroySpMat(C_matrix)
istat=cusparseDestroy(handle)
!$acc update host(C_csrOffsets,C_columns,C_values)
print*,'C_csrOffsets:',C_csrOffsets
print*,'C_columns:',C_columns
print*,'C_values:',C_values
!$acc end data
!$acc end data
!$acc end data
end program spgemm_text