Device or host arguments to magmablas_zgemm_batched

[Wileam Phan]

Hi everyone,

I'm using MAGMA with Fortran code and OpenACC (with PGI compiler). So far I have only the array of matrices on device. I have the following wrapper to the MAGMA v2 interface (adapted from the example one in fortran subdirectory in the source code):

INTERFACE

  SUBROUTINE magmablas_zgemm_batched( opA, opB, m, n, k, alpha, dptr_A, ldda, dptr_B, lddb, beta, dptr_C, lddc, batchCount, queue) BIND(C, NAME='magmablas_zgemm_batched')

    USE ISO_C_BINDING

    INTEGER(C_INT), VALUE :: opA, opB, m, n, k, ldda, lddb, lddc

    COMPLEX(C_DOUBLE_COMPLEX), VALUE :: alpha, beta

    TYPE(C_PTR), DIMENSION(batchCount) :: dptr_A, dptr_B, dptr_C

    TYPE(C_PTR), VALUE :: queue

  END SUBROUTINE

END INTERFACE

opA = magma_trans_const( 'N' )

opB = magma_trans_const( 'N' )

!$ACC DATA PRESENT(matA, matB, matC)

!$ACC HOST_DATA USE_DEVICE(matA, matB, matC)

DO ibatch = 1, batchCount

  dptr_A(ibatch) = C_LOC( matA(1,1,ibatch) )

  dptr_B(ibatch) = C_LOC( matB(1,1,ibatch) )

  dptr_C(ibatch) = C_LOC( matC(1,1,ibatch) )

END DO

CALL magmablas_zgemm_batched( opA, opB, m, n, k, alpha, dptr_A, ldda, dptr_B, lddb, beta, dptr_C, lddc, batchCount, queue )

!$ACC END HOST_DATA

!$ACC END DATA

If I use managed memory (-ta=tesla:managed) the code executes successfully.

If I don't use managed memory and run it through cuda-gdb, I get CUDA_EXCEPTION_14, Warp illegal address, which suggests that one of the arguments need to be on device, instead of on host?

Thanks,

Wileam Y. Phan

[Ahmad Abdelfattah]

The pointer arrays of all the magma batch routines must be on device memory.

Ahmad

--
You received this message because you are subscribed to the Google Groups "MAGMA User" group.
To unsubscribe from this group and stop receiving emails from it, send an email to magma-user+...@icl.utk.edu.
To view this discussion on the web visit https://groups.google.com/a/icl.utk.edu/d/msgid/magma-user/f3d14c86-152e-4eaa-98c2-1fd12b709d53n%40icl.utk.edu.

[Wileam Phan]

Hi Ahmad,

Thanks, I've now exposed their device pointer addresses using !$ACC HOST_DATA USE_DATA(matA, matB, matC, dptr_A, dptr_B, dptr_C) after filling in the pointer arrays.

What about alpha and beta? Should they reside on host, or on device?

Thanks,
Wileam Y. Phan

[Ahmad Abdelfattah]

alpha and beta are scalars, not arrays. They should be on host memory. 

Ahmad

[Wileam Phan]

Hi Ahmad,

Ah, I see. I guess this is where MAGMA differs from cuBLAS, where alpha and beta can be either on host or device. I wish this would be made clearer in the documentation.

Thanks a lot,
Wileam Y. Phan