inline std::ptrdiff_t potrf( Order, const UpLo, const fortran_int_t n,
double* a, const fortran_int_t lda ) {
BOOST_STATIC_ASSERT( (is_same<Order, tag::column_major>::value) );
fortran_int_t info(0);
LAPACK_DPOTRF( &lapack_option< UpLo >::value, &n, a, &lda, &info );
return info;
}
int TEMPLATE2 (CHOLMOD (gpu_lower_potrf))
(
Int nscol2, /* S is nscol2-by-nscol2 */
Int nsrow, /* leading dimension of S */
Int psx, /* S is located at Lx + L_ENTRY*psx */
double *Lx, /* contains S; overwritten with Cholesky factor */
Int *info, /* BLAS info return value */
cholmod_common *Common,
cholmod_gpu_pointers *gpu_p
)
{
double *devPtrA, *devPtrB, *A ;
double alpha, beta ;
cudaError_t cudaStat ;
cublasStatus_t cublasStatus ;
Int j, nsrow2, nb, n, gpu_lda, lda, gpu_ldb ;
int ilda, ijb, iinfo ;
#ifndef NTIMER
double tstart ;
#endif
if (nscol2 * L_ENTRY < CHOLMOD_POTRF_LIMIT)
{
/* too small for the CUDA BLAS; use the CPU instead */
return (0) ;
}
#ifndef NTIMER
tstart = SuiteSparse_time ( ) ;
Common->CHOLMOD_GPU_POTRF_CALLS++ ;
#endif
nsrow2 = nsrow - nscol2 ;
/* ---------------------------------------------------------------------- */
/* heuristic to get the block size depending of the problem size */
/* ---------------------------------------------------------------------- */
nb = 128 ;
if (nscol2 > 4096) nb = 256 ;
if (nscol2 > 8192) nb = 384 ;
n = nscol2 ;
gpu_lda = ((nscol2+31)/32)*32 ;
lda = nsrow ;
A = gpu_p->h_Lx[(Common->ibuffer+CHOLMOD_HOST_SUPERNODE_BUFFERS-1)%
CHOLMOD_HOST_SUPERNODE_BUFFERS];
/* ---------------------------------------------------------------------- */
/* determine the GPU leading dimension of B */
/* ---------------------------------------------------------------------- */
gpu_ldb = 0 ;
if (nsrow2 > 0)
{
gpu_ldb = ((nsrow2+31)/32)*32 ;
}
/* ---------------------------------------------------------------------- */
/* remember where device memory is, to be used by triangular solve later */
/* ---------------------------------------------------------------------- */
devPtrA = gpu_p->d_Lx[0];
devPtrB = gpu_p->d_Lx[1];
/* ---------------------------------------------------------------------- */
/* copy A from device to device */
/* ---------------------------------------------------------------------- */
cudaStat = cudaMemcpy2DAsync ( devPtrA,
gpu_lda * L_ENTRY * sizeof (devPtrA[0]),
gpu_p->d_A[1],
nsrow * L_ENTRY * sizeof (Lx[0]),
nscol2 * L_ENTRY * sizeof (devPtrA[0]),
nscol2,
cudaMemcpyDeviceToDevice,
Common->gpuStream[0] );
if ( cudaStat ) {
ERROR ( CHOLMOD_GPU_PROBLEM, "GPU memcopy device to device");
}
/* ---------------------------------------------------------------------- */
/* copy B in advance, for gpu_triangular_solve */
/* ---------------------------------------------------------------------- */
if (nsrow2 > 0)
{
cudaStat = cudaMemcpy2DAsync (devPtrB,
gpu_ldb * L_ENTRY * sizeof (devPtrB [0]),
gpu_p->d_A[1] + L_ENTRY*nscol2,
nsrow * L_ENTRY * sizeof (Lx [0]),
nsrow2 * L_ENTRY * sizeof (devPtrB [0]),
nscol2,
cudaMemcpyDeviceToDevice,
Common->gpuStream[0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
}
/* ------------------------------------------------------------------ */
/* define the dpotrf stream */
/* ------------------------------------------------------------------ */
cublasStatus = cublasSetStream (Common->cublasHandle,
Common->gpuStream [0]) ;
if (cublasStatus != CUBLAS_STATUS_SUCCESS) {
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS stream") ;
}
/* ---------------------------------------------------------------------- */
/* block Cholesky factorization of S */
/* ---------------------------------------------------------------------- */
for (j = 0 ; j < n ; j += nb)
{
Int jb = nb < (n-j) ? nb : (n-j) ;
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dsyrk */
/* ------------------------------------------------------------------ */
alpha = -1.0 ;
beta = 1.0 ;
#ifdef REAL
cublasStatus = cublasDsyrk (Common->cublasHandle,
CUBLAS_FILL_MODE_LOWER, CUBLAS_OP_N, jb, j,
&alpha, devPtrA + j, gpu_lda,
&beta, devPtrA + j + j*gpu_lda, gpu_lda) ;
#else
cublasStatus = cublasZherk (Common->cublasHandle,
CUBLAS_FILL_MODE_LOWER, CUBLAS_OP_N, jb, j,
&alpha, (cuDoubleComplex*)devPtrA + j,
gpu_lda,
&beta,
(cuDoubleComplex*)devPtrA + j + j*gpu_lda,
gpu_lda) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
/* ------------------------------------------------------------------ */
cudaStat = cudaEventRecord (Common->cublasEventPotrf [0],
Common->gpuStream [0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
cudaStat = cudaStreamWaitEvent (Common->gpuStream [1],
Common->cublasEventPotrf [0], 0) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
/* ------------------------------------------------------------------ */
/* copy back the jb columns on two different streams */
/* ------------------------------------------------------------------ */
cudaStat = cudaMemcpy2DAsync (A + L_ENTRY*(j + j*lda),
lda * L_ENTRY * sizeof (double),
devPtrA + L_ENTRY*(j + j*gpu_lda),
gpu_lda * L_ENTRY * sizeof (double),
L_ENTRY * sizeof (double)*jb,
jb,
cudaMemcpyDeviceToHost,
Common->gpuStream [1]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy from device") ;
}
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dgemm */
/* ------------------------------------------------------------------ */
if ((j+jb) < n)
{
#ifdef REAL
alpha = -1.0 ;
beta = 1.0 ;
cublasStatus = cublasDgemm (Common->cublasHandle,
CUBLAS_OP_N, CUBLAS_OP_T,
(n-j-jb), jb, j,
&alpha,
devPtrA + (j+jb), gpu_lda,
devPtrA + (j) , gpu_lda,
&beta,
devPtrA + (j+jb + j*gpu_lda), gpu_lda) ;
#else
cuDoubleComplex calpha = {-1.0,0.0} ;
cuDoubleComplex cbeta = { 1.0,0.0} ;
cublasStatus = cublasZgemm (Common->cublasHandle,
CUBLAS_OP_N, CUBLAS_OP_C,
(n-j-jb), jb, j,
&calpha,
(cuDoubleComplex*)devPtrA + (j+jb),
gpu_lda,
(cuDoubleComplex*)devPtrA + (j),
gpu_lda,
&cbeta,
(cuDoubleComplex*)devPtrA +
(j+jb + j*gpu_lda),
gpu_lda ) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
}
cudaStat = cudaStreamSynchronize (Common->gpuStream [1]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
/* ------------------------------------------------------------------ */
/* compute the Cholesky factorization of the jbxjb block on the CPU */
/* ------------------------------------------------------------------ */
ilda = (int) lda ;
ijb = jb ;
#ifdef REAL
LAPACK_DPOTRF ("L", &ijb, A + L_ENTRY * (j + j*lda), &ilda, &iinfo) ;
#else
LAPACK_ZPOTRF ("L", &ijb, A + L_ENTRY * (j + j*lda), &ilda, &iinfo) ;
#endif
*info = iinfo ;
if (*info != 0)
{
*info = *info + j ;
break ;
}
/* ------------------------------------------------------------------ */
/* copy the result back to the GPU */
/* ------------------------------------------------------------------ */
cudaStat = cudaMemcpy2DAsync (devPtrA + L_ENTRY*(j + j*gpu_lda),
gpu_lda * L_ENTRY * sizeof (double),
A + L_ENTRY * (j + j*lda),
lda * L_ENTRY * sizeof (double),
L_ENTRY * sizeof (double) * jb,
jb,
cudaMemcpyHostToDevice,
Common->gpuStream [0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dtrsm */
/* ------------------------------------------------------------------ */
if ((j+jb) < n)
{
#ifdef REAL
alpha = 1.0 ;
cublasStatus = cublasDtrsm (Common->cublasHandle,
CUBLAS_SIDE_RIGHT,
CUBLAS_FILL_MODE_LOWER,
CUBLAS_OP_T, CUBLAS_DIAG_NON_UNIT,
(n-j-jb), jb,
&alpha,
devPtrA + (j + j*gpu_lda), gpu_lda,
devPtrA + (j+jb + j*gpu_lda), gpu_lda) ;
#else
cuDoubleComplex calpha = {1.0,0.0};
cublasStatus = cublasZtrsm (Common->cublasHandle,
CUBLAS_SIDE_RIGHT,
CUBLAS_FILL_MODE_LOWER,
CUBLAS_OP_C, CUBLAS_DIAG_NON_UNIT,
(n-j-jb), jb,
&calpha,
(cuDoubleComplex *)devPtrA +
(j + j*gpu_lda),
gpu_lda,
(cuDoubleComplex *)devPtrA +
(j+jb + j*gpu_lda),
gpu_lda) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
/* -------------------------------------------------------------- */
/* Copy factored column back to host. */
/* -------------------------------------------------------------- */
cudaStat = cudaEventRecord (Common->cublasEventPotrf[2],
Common->gpuStream[0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
cudaStat = cudaStreamWaitEvent (Common->gpuStream[1],
Common->cublasEventPotrf[2], 0) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
cudaStat = cudaMemcpy2DAsync (A + L_ENTRY*(j + jb + j * lda),
lda * L_ENTRY * sizeof (double),
devPtrA + L_ENTRY*
(j + jb + j * gpu_lda),
gpu_lda * L_ENTRY * sizeof (double),
L_ENTRY * sizeof (double)*
(n - j - jb), jb,
cudaMemcpyDeviceToHost,
Common->gpuStream[1]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
}
}
#ifndef NTIMER
Common->CHOLMOD_GPU_POTRF_TIME += SuiteSparse_time ( ) - tstart ;
#endif
return (1) ;
}
