Adding GPU kernels such as geadd
View options
We would like to port some kernels to GPU to minimize the CPU-GPU data transfer. I have started with zgeadd: I added the cl_zgeadd_cuda_func definition to its codelet, and I added cuda_zgeadd to call cublasZgeam as follows:
#include "cudablas/include/cudablas.h"
#include "cudablas/include/cudablas_z.h"
int CUDA_zgeadd(MORSE_enum trans,
int m, int n,
cuDoubleComplex *alpha,
const cuDoubleComplex *A, int lda,
cuDoubleComplex *beta,
cuDoubleComplex *B, int ldb,
CUBLAS_STREAM_PARAM)
{
#if !defined(CHAMELEON_USE_CUBLAS_V2)
cublasSetKernelStream( stream );
#endif
//cublasHandle_t handle;
//cublasZgeam(handle,
cublasZgeam(CUBLAS_HANDLE
morse_lapack_const(trans), morse_lapack_const(MorseNoTrans),
m, n,
CUBLAS_VALUE(alpha), A, lda,
CUBLAS_VALUE(beta), B, ldb,
B, ldb);
assert( CUBLAS_STATUS_SUCCESS == cublasGetError() );
return MORSE_SUCCESS;
}I am using spack to install chameleon v0.9.1 with starpu v1.2.1. I got the following errors and warnings:
compilation aborted for /home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_zgeadd.c (code 2)
make[2]: *** [cudablas/compute/CMakeFiles/cudablas.dir/cuda_zgeadd.c.o] Error 2
/home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_dgeadd.c(45): warning #167: argument of type "char" is incompatible with parameter of type "cublasHandle_t"
morse_lapack_const(trans), morse_lapack_const(MorseNoTrans),
^
/home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_dgeadd.c(48): warning #167: argument of type "int" is incompatible with parameter of type "const double *"
CUBLAS_VALUE(alpha), A, lda,
^
/home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_dgeadd.c(49): warning #167: argument of type "int" is incompatible with parameter of type "const double *"
CUBLAS_VALUE(beta), B, ldb,
^
/home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_dgeadd.c(50): warning #810: conversion from "double *" to "int" may lose significant bits
B, ldb);
^
/home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_dgeadd.c(50): warning #167: argument of type "int" is incompatible with parameter of type "double *"
B, ldb);
^
/home/sukkarde/codes/chameleon/spack-build/cudablas/compute/cuda_dgeadd.c(50): error #165: too few arguments in function call
B, ldb);If I add cublasHandle_t handle instead of CUBLAS_HANDLE, then I got a wrong results. Could you please advice how to solve this issue ?
Link items together to show that they're related.