extern "C" magma_int_t
magma_cpotrf3_mgpu(magma_int_t num_gpus, char uplo, magma_int_t m, magma_int_t n,
magma_int_t off_i, magma_int_t off_j, magma_int_t nb,
magmaFloatComplex *d_lA[], magma_int_t ldda,
magmaFloatComplex *d_lP[], magma_int_t lddp,
magmaFloatComplex *a, magma_int_t lda, magma_int_t h,
magma_queue_t stream[][3], magma_event_t event[][5],
magma_int_t *info )
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
CPOTRF computes the Cholesky factorization of a complex Hermitian
positive definite matrix dA.
Auxiliary subroutine for cpotrf2_ooc. It is multiple gpu interface to compute
Cholesky of a "rectangular" matrix.
The factorization has the form
dA = U**H * U, if UPLO = 'U', or
dA = L * L**H, if UPLO = 'L',
where U is an upper triangular matrix and L is lower triangular.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
=========
UPLO (input) CHARACTER*1
= 'U': Upper triangle of dA is stored;
= 'L': Lower triangle of dA is stored.
N (input) INTEGER
The order of the matrix dA. N >= 0.
dA (input/output) COMPLEX array on the GPU, dimension (LDDA,N)
On entry, the Hermitian matrix dA. If UPLO = 'U', the leading
N-by-N upper triangular part of dA contains the upper
triangular part of the matrix dA, and the strictly lower
triangular part of dA is not referenced. If UPLO = 'L', the
leading N-by-N lower triangular part of dA contains the lower
triangular part of the matrix dA, and the strictly upper
triangular part of dA is not referenced.
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization dA = U**H * U or dA = L * L**H.
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
> 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
===================================================================== */
magma_int_t j, jb, nb0, nb2, d, dd, id, j_local, j_local2, buf;
char uplo_[2] = {uplo, 0};
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float d_one = 1.0;
float d_neg_one = -1.0;
int upper = lapackf77_lsame(uplo_, "U");
magmaFloatComplex *dlpanel;
magma_int_t n_local[MagmaMaxGPUs], ldpanel;
const magma_int_t stream1 = 0, stream2 = 1, stream3 = 2;
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
/* used by ctrsm_work */
int trsm_nb = 128;
int trsm_n = trsm_nb*((nb+trsm_nb-1)/trsm_nb);
magmaFloatComplex *d_dinvA[MagmaMaxGPUs];
magmaFloatComplex *d_x[MagmaMaxGPUs];
#define dinvA(d,j) &(d_dinvA[(d)][(j)*trsm_nb*trsm_n])
#define dx(d,j) &(d_x[(d)][(j)*nb*m])
/*
* Allocate device memory for the inversed diagonal blocks, size=N*BLOCK_SIZE
*/
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
if ( (MAGMA_SUCCESS != magma_cmalloc( &d_dinvA[d], 2*trsm_nb*trsm_n )) ||
(MAGMA_SUCCESS != magma_cmalloc( &d_x[d], 2*nb*(upper ? n : m) )) ) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
magma_setdevice(0);
#endif
*info = 0;
if ( (! upper) && (! lapackf77_lsame(uplo_, "L")) ) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (!upper && num_gpus*ldda < max(1,n)) {
*info = -4;
} else if (upper && ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* initialization */
for( d=0; d<num_gpus; d++ ) {
/* local-n and local-ld */
if (upper) {
n_local[d] = (n/(nb*num_gpus))*nb;
if (d < (n/nb)%num_gpus)
n_local[d] += nb;
else if (d == (n/nb)%num_gpus)
n_local[d] += n%nb;
} else {
n_local[d] = (m/(nb*num_gpus))*nb;
if (d < (m/nb)%num_gpus)
n_local[d] += nb;
else if (d == (m/nb)%num_gpus)
n_local[d] += m%nb;
}
}
/* == initialize the trace */
trace_init( 1, num_gpus, 3, (CUstream_st**)stream );
if (upper)
{
/* ---------------------------------------------- */
/* Upper-triangular case */
/* > Compute the Cholesky factorization A = U'*U. */
/* ---------------------------------------------- */
for (j=0; j<m; j+=nb) {
/* Set the GPU number that holds the current panel */
id = (j/nb)%num_gpus;
buf = (j/nb)%num_gpus; // right now, we have num_gpu buffers, so id and buf are the same..
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (m-j));
/* Update the current diagonal block on stream1 */
magma_setdevice(id);
if( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
trace_gpu_start( id, stream1, "syrk", "syrk" );
magma_cherk(MagmaUpper, MagmaConjTrans, jb, j,
d_neg_one, dlA(id, 0, nb*j_local), ldda,
d_one, dlA(id, j, nb*j_local), ldda);
trace_gpu_end( id, stream1 );
}
/* send the diagonal to cpu on stream1 */
trace_gpu_start( id, stream1, "comm", "D to CPU" );
magma_cgetmatrix_async( jb, jb,
dlA(id, j, nb*j_local), ldda,
Aup(j,j), lda,
stream[id][stream1] );
trace_gpu_end( id, stream1 );
/* update off-diagonal blocks in the panel */
if( j > 0 ) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
j_local2 = j_local+1;
if( d > id ) j_local2 --;
nb0 = nb*j_local2; // number of local columns in the panel, while jb is panel-size (number of rows)
if( n_local[d] > nb0 ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
if( d == id ) {
dlpanel = dlA(d,0,nb*j_local);
ldpanel = ldda;
// the GPU owns the row from start, and no need of synch.
//magma_queue_wait_event( stream[d][stream2], event[d][0] ); // rows arrived at gpu
} else {
dlpanel = dlP(d,nb,0,buf);
ldpanel = lddp;
magma_queue_wait_event( stream[d][stream2], event[d][0] ); // rows arrived at gpu
}
trace_gpu_start( d, stream2, "gemm", "gemm" );
magma_cgemm(MagmaConjTrans, MagmaNoTrans,
jb, n_local[d]-nb0, j,
c_neg_one, dlpanel, ldpanel,
dlA(d, 0, nb0), ldda,
c_one, dlA(d, j, nb0), ldda);
trace_gpu_end( d, stream2 );
magma_event_record( event[d][2], stream[d][stream2] );
}
d = (d+1)%num_gpus;
}
}
/* wait for panel and factorize it on cpu */
magma_setdevice(id);
magma_queue_sync( stream[id][stream1] );
trace_cpu_start( 0, "getrf", "getrf" );
lapackf77_cpotrf(MagmaUpperStr, &jb, Aup(j,j), &lda, info);
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus on stream1 */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
if( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d,0,0,buf);
ldpanel = lddp;
}
magma_setdevice(d);
trace_gpu_start( d, stream1, "comm", "comm" );
magma_csetmatrix_async( jb, jb,
Aup(j,j), lda,
dlpanel, ldpanel,
stream[d][stream1] );
trace_gpu_end( d, stream1 );
magma_event_record( event[d][1], stream[d][stream1] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
trace_gpu_start( id, stream1, "comm", "comm" );
magma_csetmatrix_async( jb, jb,
Aup(j,j), lda,
dlA(id, j, nb*j_local), ldda,
stream[id][stream1] );
trace_gpu_end( id, stream1 );
}
/* panel-factorize the off-diagonal */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
/* next column */
j_local2 = j_local+1;
if( d > id ) j_local2--;
if( d == id ) {
dlpanel = dlA(d,j,nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d,0,0,buf);
ldpanel = lddp;
}
nb2 = n_local[d] - j_local2*nb;
magma_setdevice(d);
if( j+jb < m && d == (j/nb+1)%num_gpus ) {
/* owns the next column, look-ahead next block on stream1 */
nb0 = min(nb, nb2);
magmablasSetKernelStream(stream[d][stream1]);
magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for gemm update
trace_gpu_start( d, stream1, "trsm", "trsm" );
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
magmablas_claset( MagmaUpperLower, trsm_nb, trsm_n, dinvA(d,0),trsm_nb );
magmablas_claset( MagmaUpperLower, nb0,jb, dx(d,0),nb0 );
magmablas_ctrsm_work( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ctrsm( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
magma_event_record( event[d][4], stream[d][stream1] );
trace_gpu_end( d, stream1 );
} else if( nb2 > 0 ) {
/* update all the blocks on stream2 */
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // wait for cholesky factor
trace_gpu_start( d, stream2, "trsm", "trsm" );
magmablasSetKernelStream(stream[d][stream2]);
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
magmablas_claset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,0),trsm_nb );
magmablas_claset( MagmaUpperLower, nb2,jb, dx(d,0),nb2 );
magmablas_ctrsm_work( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ctrsm( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
trace_gpu_end( d, stream2 );
}
d = (d+1)%num_gpus;
} /* end of for */
/* ========================================================== */
if( j+jb < m ) {
d = (j/nb+1)%num_gpus;
/* next column */
j_local2 = j_local+1;
if( d > id ) j_local2--;
nb0 = min(nb, n_local[d]-nb*j_local2 );
/* even on 1 gpu, off-diagonals are copied to cpu (synchronize at the end). *
* so we have the Cholesky factor, but only diagonal submatrix of the big panel, *
* on cpu at the end. */
int d2, buf2;
magma_setdevice(d);
/* lookahead done */
magma_queue_wait_event( stream[d][stream3], event[d][4] );
trace_gpu_start( d, stream3, "comm", "row to CPU" );
magma_cgetmatrix_async( (j+jb), nb0,
dlA(d, 0, nb*j_local2), ldda,
Aup(0,j+jb), lda,
stream[d][stream3] );
trace_gpu_end( d, stream3 );
magma_event_record( event[d][3], stream[d][stream3] );
/* needed on pluto */
//magma_queue_sync( stream[d][stream3] );
/* broadcast rows to gpus on stream2 */
buf2 = ((j+jb)/nb)%num_gpus;
for( d2=0; d2<num_gpus; d2++ ) {
if( d2 != d )
{
magma_setdevice(d2);
trace_gpu_start( d2, stream3, "comm", "row to GPUs" );
magma_queue_wait_event( stream[d2][stream3], event[d][3] ); // rows arrived at cpu on stream3
magma_csetmatrix_async( j+jb, nb0,
Aup(0,j+jb), lda,
dlP(d2,nb,0,buf2), lddp,
stream[d2][stream3] );
trace_gpu_end( d2, stream3 );
magma_event_record( event[d2][0], stream[d2][stream3] );
}
}
/* =========================== */
/* update the remaining blocks */
nb2 = n_local[d]-(nb*j_local2 + nb0);
if( nb2 > 0 ) {
if( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d,0,0,buf);
ldpanel = lddp;
}
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
trace_gpu_start( d, stream2, "trsm", "trsm" );
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
int flag = 0;
if (flag == 0) {
magma_queue_wait_event( stream[d][stream2], event[d][4] ); // lookahead -> diagonal inversion
} else {
magmablas_claset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,flag),trsm_nb );
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // panel received
}
magmablas_claset( MagmaUpperLower, nb2,jb, dx(d,1),nb2 );
magmablas_ctrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda,
flag, dinvA(d,flag), dx(d,1) );
#else
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // wait for cholesky factor
magma_ctrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda);
#endif
trace_gpu_end( d, stream2 );
}
}
} /* end of ctrsm */
} /* end of for j=1, .., n */
} else {
/* ---------------------------------------------- */
/* Lower-triangular case */
/* > Compute the Cholesky factorization A = L*L'. */
/* ---------------------------------------------- */
for (j=0; j<n; j+=nb) {
/* Set the GPU number that holds the current panel */
id = (j/nb)%num_gpus;
buf = (j/nb)%num_gpus;
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (n-j));
/* Update the current diagonal block on stream1 */
magma_setdevice(id);
if( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
magma_cherk(MagmaLower, MagmaNoTrans, jb, j,
d_neg_one, dlA(id, nb*j_local, 0), ldda,
d_one, dlA(id, nb*j_local, j), ldda);
}
/* send the diagonal to cpu on stream1 */
magma_cgetmatrix_async( jb, jb,
dlA(id, nb*j_local, j), ldda,
Alo(j,j), lda,
stream[id][stream1] );
/* update off-diagonal blocks of the panel */
if( j > 0 ) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
j_local2 = j_local+1;
if( d > id ) j_local2 --;
nb0 = nb*j_local2;
if( nb0 < n_local[d] ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
if( d == id ) {
dlpanel = dlA(d, nb*j_local, 0);
ldpanel = ldda;
} else {
dlpanel = dlPT(d,0,nb,buf);
ldpanel = nb;
magma_queue_wait_event( stream[d][stream2], event[d][0] ); // rows arrived at gpu
}
magma_cgemm( MagmaNoTrans, MagmaConjTrans,
n_local[d]-nb0, jb, j,
c_neg_one, dlA(d, nb0, 0), ldda,
dlpanel, ldpanel,
c_one, dlA(d, nb0, j), ldda);
magma_event_record( event[d][2], stream[d][stream2] );
}
d = (d+1)%num_gpus;
}
}
/* wait for the panel and factorized it on cpu */
magma_setdevice(id);
magma_queue_sync( stream[id][stream1] );
lapackf77_cpotrf(MagmaLowerStr, &jb, Alo(j,j), &lda, info);
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus on stream1 */
if ( (j+jb) < m) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
if( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
magma_setdevice(d);
magma_csetmatrix_async( jb, jb,
Alo(j,j), lda,
dlpanel, ldpanel,
stream[d][stream1] );
magma_event_record( event[d][1], stream[d][stream1] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
magma_csetmatrix_async( jb, jb,
Alo(j,j), lda,
dlA(id, nb*j_local, j), ldda,
stream[id][stream1] );
}
/* panel factorize the off-diagonal */
if ( (j+jb) < m) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
/* next column */
j_local2 = j_local+1;
if( d > id ) j_local2--;
if( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
nb2 = n_local[d] - j_local2*nb;
nb0 = min(nb, nb2);
magma_setdevice(d);
if( j+nb < n && d == (j/nb+1)%num_gpus ) { /* owns next column, look-ahead next block on stream1 */
if ( j > 0 ) magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for gemm update
magmablasSetKernelStream(stream[d][stream1]);
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
magmablas_claset( MagmaUpperLower, trsm_nb, trsm_n, dinvA(d,0),trsm_nb );
magmablas_claset( MagmaUpperLower, nb0,jb, dx(d,0),nb0 );
magmablas_ctrsm_work( MagmaRight, MagmaLower,
MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ctrsm( MagmaRight, MagmaLower,
MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
magma_event_record( event[d][4], stream[d][stream1] );
} else if( nb2 > 0 ) { /* other gpus updating all the blocks on stream2 */
/* update the entire column */
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // wait for the cholesky factor
magmablasSetKernelStream(stream[d][stream2]);
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
magmablas_claset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,0),trsm_nb );
magmablas_claset( MagmaUpperLower, nb2,jb, dx(d,0),nb2 );
magmablas_ctrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ctrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
}
d = (d+1)%num_gpus;
} /* end for d */
/* ========================================================== */
if( j+jb < n ) {
d = (j/nb+1)%num_gpus;
/* next column */
j_local2 = j_local+1;
if( d > id ) j_local2--;
nb0 = min(nb, n_local[d]-nb*j_local2 );
/* even on 1 gpu, we copy off-diagonal to cpu (but don't synchronize). */
/* so we have the Cholesky factor on cpu at the end. */
int d2, buf2;
//#define CPOTRF_DEVICE_TO_DEVICE
#ifdef CPOTRF_DEVICE_TO_DEVICE
// lookahead done
/* broadcast the rows to gpus */
buf2 = ((j+jb)/nb)%num_gpus;
for( d2=0; d2<num_gpus; d2++ ) {
magma_setdevice(d2);
magma_queue_wait_event( stream[d2][stream3], event[d][4] );
if( d2 != d ) {
magma_ccopymatrix_async( nb0, j+jb,
dlPT(d2,0,nb,buf2), nb, // first nbxnb reserved for diagonal block
dlA(d, nb*j_local2, 0), ldda,
stream[d2][stream3] );
magma_event_record( event[d2][0], stream[d2][stream3] );
} else {
magma_cgetmatrix_async( nb0, j+jb,
dlA(d, nb*j_local2, 0), ldda,
Alo(j+jb,0), lda,
stream[d][stream3] );
}
}
#else
// lookahead done
magma_setdevice(d);
magma_queue_wait_event( stream[d][stream3], event[d][4] );
magma_cgetmatrix_async( nb0, j+jb,
dlA(d, nb*j_local2, 0), ldda,
Alo(j+jb,0), lda,
stream[d][stream3] );
magma_event_record( event[d][3], stream[d][stream3] );
/* syn on rows on CPU, seem to be needed on Pluto */
//magma_queue_sync( stream[d][stream3] );
/* broadcast the rows to gpus */
buf2 = ((j+jb)/nb)%num_gpus;
for( d2=0; d2<num_gpus; d2++ ) {
if( d2 != d )
{
magma_setdevice(d2);
magma_queue_wait_event( stream[d2][stream3], event[d][3] ); // getmatrix done
magma_csetmatrix_async( nb0, j+jb,
Alo(j+jb,0), lda,
dlPT(d2,0,nb,buf2), nb, // first nbxnb reserved for diagonal block
stream[d2][stream3] );
magma_event_record( event[d2][0], stream[d2][stream3] );
}
}
#endif
/* =================================== */
/* updates remaining blocks on stream2 */
nb2 = n_local[d] - (j_local2*nb + nb0);
if( nb2 > 0 ) {
if( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d,0,0,buf);
ldpanel = nb;
}
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
/* update the remaining blocks in the column */
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
int flag = 0;
if (flag == 0) {
magma_queue_wait_event( stream[d][stream2], event[d][4] ); // lookahead -> diagonal inversion
} else {
magmablas_claset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,flag),trsm_nb );
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // panel received
}
magmablas_claset( MagmaUpperLower, nb2,jb, dx(d,1),nb2 );
magmablas_ctrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda,
flag, dinvA(d,flag), dx(d,1) );
#else
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // panel received
magma_ctrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda);
#endif
}
}
}
}
} /* end of else not upper */
/* == finalize the trace == */
trace_finalize( "cpotrf.svg","trace.css" );
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j<3; j++ ) {
magma_queue_sync( stream[d][j] );
}
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(CTRSM_WORK)
magma_free( d_dinvA[d] );
magma_free( d_x[d] );
#endif
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
return *info;
} /* magma_cpotrf_mgpu */
/**
Purpose
-------
CPOTRF computes the Cholesky factorization of a complex Hermitian
positive definite matrix dA.
The factorization has the form
dA = U**H * U, if UPLO = MagmaUpper, or
dA = L * L**H, if UPLO = MagmaLower,
where U is an upper triangular matrix and L is lower triangular.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of dA is stored;
- = MagmaLower: Lower triangle of dA is stored.
@param[in]
n INTEGER
The order of the matrix dA. N >= 0.
@param[in,out]
dA COMPLEX array on the GPU, dimension (LDDA,N)
On entry, the Hermitian matrix dA. If UPLO = MagmaUpper, the leading
N-by-N upper triangular part of dA contains the upper
triangular part of the matrix dA, and the strictly lower
triangular part of dA is not referenced. If UPLO = MagmaLower, the
leading N-by-N lower triangular part of dA contains the lower
triangular part of the matrix dA, and the strictly upper
triangular part of dA is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization dA = U**H * U or dA = L * L**H.
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
- > 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_cposv_comp
********************************************************************/
extern "C" magma_int_t
magma_cpotrf2_mgpu(int num_gpus, magma_uplo_t uplo, magma_int_t m, magma_int_t n,
magma_int_t off_i, magma_int_t off_j, magma_int_t nb,
magmaFloatComplex **d_lA, magma_int_t ldda,
magmaFloatComplex **d_lP, magma_int_t lddp,
magmaFloatComplex *A, magma_int_t lda, magma_int_t h,
magma_queue_t stream[][3], magma_event_t event[][5],
magma_int_t *info )
{
#define Alo(i, j) (A + ((j)+off_j)*lda + (nb*(((i)/nb)%h)+off_i))
#define Aup(i, j) (A + (nb*(((j)/nb)%h)+off_j)*lda + (i+off_i))
#define dlA(id, i, j) (d_lA[(id)] + (j)*ldda + (i))
#define dlP(id, i, j, k) (d_lP[(id)] + (k)*nb*lddp + (j)*lddp + (i))
#define dlPT(id, i, j, k) (d_lP[(id)] + (k)*nb*lddp + (j)*nb + (i))
magma_int_t j, jb, nb0, nb2, dd, d, id, j_local, j_local2, buf;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
float d_one = 1.0;
float d_neg_one = -1.0;
int upper = (uplo == MagmaUpper);
magmaFloatComplex *dlpanel;
//magma_event_t event0[MagmaMaxGPUs], // syrk
// event1[MagmaMaxGPUs], // send off-diagonal
// event2[MagmaMaxGPUs], // send diagonal
// event3[MagmaMaxGPUs]; // trsm
magma_int_t n_local[MagmaMaxGPUs], ldpanel;
int stream0 = 0, stream1 = 1;
#ifdef CTRSM_WORK
magmaFloatComplex *d_dinvA[MagmaMaxGPUs][2], *d_x[MagmaMaxGPUs][2]; /* used by ctrsm_work */
#endif
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (!upper && num_gpus*ldda < max(1,n)) {
*info = -4;
} else if (upper && ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
for( d=0; d < num_gpus; d++ ) {
/* local-n and local-ld */
if (upper) {
n_local[d] = ((n/nb)/num_gpus)*nb;
if (d < (n/nb)%num_gpus)
n_local[d] += nb;
else if (d == (n/nb)%num_gpus)
n_local[d] += n%nb;
} else {
n_local[d] = ((m/nb)/num_gpus)*nb;
if (d < (m/nb)%num_gpus)
n_local[d] += nb;
else if (d == (m/nb)%num_gpus)
n_local[d] += m%nb;
}
//magma_setdevice(d);
//magma_event_create( &event0[d] );
//magma_event_create( &event1[d] );
//magma_event_create( &event2[d] );
//magma_event_create( &event3[d] );
}
magma_setdevice(0);
/* == initialize the trace */
trace_init( 1, num_gpus, 3, (magma_queue_t*)stream );
/* Use blocked code. */
if (upper) {
/* ---------------------------------------------- */
/* Upper-triangular case */
/* > Compute the Cholesky factorization A = U'*U. */
/* ---------------------------------------------- */
#if defined(PRECISION_d) && defined(CTRSM_WORK)
/* invert the diagonals
* Allocate device memory for the inversed diagonal blocks, size=m*NB
*/
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j < 2; j++ ) {
magma_cmalloc( &d_dinvA[d][j], nb*nb );
magma_cmalloc( &d_x[d][j], n*nb );
cudaMemset(d_dinvA[d][j], 0, nb*nb*sizeof(magmaFloatComplex));
cudaMemset(d_x[d][j], 0, n*nb*sizeof(magmaFloatComplex));
}
}
magma_setdevice(0);
#endif
for (j=0; j < m; j += nb) {
/* Set the GPU number that holds the current panel */
id = (j/nb)%num_gpus;
buf = (j/nb)%num_gpus;
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (m-j));
if ( j > 0 ) {
/* needed on pluto... */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] ); // wait for the column on CPU
/* broadcast off-diagonal column to all gpus */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
if ( d != id ) {
magma_setdevice(d);
/* wait for it on CPU */
magma_queue_wait_event( stream[d][stream0], event[id][1] );
/* send it to GPU */
trace_gpu_start( d, stream0, "comm", "rows to GPUs" );
magma_csetmatrix_async( j, jb,
Aup(0,j), lda,
dlP(d,jb,0,buf), lddp,
stream[d][stream0] );
trace_gpu_end( d, stream0 );
magma_event_record( event[d][1], stream[d][stream0] );
}
d = (d+1)%num_gpus;
}
}
/* Update the current diagonal block */
magma_setdevice(id);
if ( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
trace_gpu_start( id, stream1, "syrk", "syrk" );
magma_cherk(MagmaUpper, MagmaConjTrans, jb, j,
d_neg_one, dlA(id, 0, nb*j_local), ldda,
d_one, dlA(id, j, nb*j_local), ldda);
trace_gpu_end( id, stream1 );
magma_event_record( event[id][0], stream[id][stream1] );
}
/* send the diagonal to cpu */
magma_queue_wait_event( stream[id][stream0], event[id][0] ); // wait for syrk
trace_gpu_start( id, stream0, "comm", "D to CPU" );
magma_cgetmatrix_async( jb, jb,
dlA(id, j, nb*j_local), ldda,
Aup(j,j), lda,
stream[id][stream0] );
trace_gpu_end( id, stream0 );
if ( j > 0 ) {
/* Compute the local block column of the panel. */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
j_local2 = j_local+1;
if ( d > id ) j_local2 --;
nb0 = nb*j_local2;
if ( n_local[d] > nb0 ) {
/* wait for the off-diagonal */
if ( d != id ) {
//magma_queue_sync( stream[id][3] );
dlpanel = dlP(d, jb, 0, buf);
ldpanel = lddp;
/* wait for the offdiagonal column */
magma_queue_wait_event( stream[d][stream1], event[d][1] );
} else {
dlpanel = dlA(d, 0, nb*j_local);
ldpanel = ldda;
}
/* update the panel */
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
trace_gpu_start( d, stream1, "gemm", "gemm" );
magma_cgemm(MagmaConjTrans, MagmaNoTrans,
jb, n_local[d]-nb0, j,
c_neg_one, dlpanel, ldpanel,
dlA(d, 0, nb0), ldda,
c_one, dlA(d, j, nb0), ldda);
trace_gpu_end( d, stream1 );
}
d = (d+1)%num_gpus;
}
}
/* factor the diagonal */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] ); // wait for the diagonal
trace_cpu_start( 0, "getrf", "getrf" );
lapackf77_cpotrf(MagmaUpperStr, &jb, Aup(j,j), &lda, info);
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
if ( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d, 0, 0, buf);
ldpanel = lddp;
}
trace_gpu_start( d, stream0, "comm", "D to GPUs" );
magma_csetmatrix_async( jb, jb,
Aup(j,j), lda,
dlpanel, ldpanel,
stream[d][stream0] );
trace_gpu_end( d, stream0 );
magma_event_record( event[d][2], stream[d][stream0] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
trace_gpu_start( id, stream0, "comm", "D to GPUs" );
magma_csetmatrix_async( jb, jb,
Aup(j,j), lda,
dlA(id, j, nb*j_local), ldda,
stream[id][stream0] );
trace_gpu_end( id, stream0 );
}
/* panel-factorize the off-diagonal */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
/* next column */
j_local2 = j_local+1;
if ( d > id ) j_local2--;
if ( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d, 0, 0, buf);
ldpanel = lddp;
}
nb2 = n_local[d]-nb*j_local2;
nb0 = min(nb, nb2 );
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for the diagonal
if ( j+jb < m && d == (j/nb+1)%num_gpus ) {
/* owns the next column, look-ahead the column */
trace_gpu_start( d, stream1, "trsm", "trsm" );
#if defined(PRECISION_d) && defined(CTRSM_WORK)
magmablas_ctrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d][0], d_x[d][0] );
/*nb2 = n_local[d] - j_local2*nb;
magmablas_ctrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d], d_x[d] ); */
#else
/*nb2 = n_local[d] - j_local2*nb;
magma_ctrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldda,
dlA(d, j, nb*j_local2), ldda);
*/
magma_ctrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
trace_gpu_end( d, stream1 );
magma_event_record( event[d][3], stream[d][stream1] );
/* send the column to cpu */
if ( j+jb < m ) {
trace_gpu_start( d, stream0, "comm", "rows to CPU" );
magma_queue_wait_event( stream[d][stream0], event[d][3] ); // wait for lookahead
magma_cgetmatrix_async( (j+jb), nb0,
dlA(d, 0, nb*j_local2), ldda,
Aup(0,j+jb), lda,
stream[d][stream0] );
trace_gpu_end( d, stream0 );
magma_event_record( event[d][1], stream[d][stream0] );
}
/* update the remaining blocks */
nb2 = nb2 - nb0;
#if defined(PRECISION_d) && defined(CTRSM_WORK)
magmablas_ctrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_ctrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda);
#endif
} else if ( nb2 > 0 ) {
/* update the entire trailing matrix */
trace_gpu_start( d, stream1, "trsm", "trsm" );
#if defined(PRECISION_d) && defined(CTRSM_WORK)
magmablas_ctrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_ctrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
trace_gpu_end( d, stream1 );
}
d = (d+1)%num_gpus;
}
} /* end of ctrsm */
} /* end of for j=1, .., n */
} else {
/* -------------------------------------------- */
/* Lower-triangular case */
/* Compute the Cholesky factorization A = L*L'. */
/* -------------------------------------------- */
#if defined(PRECISION_d) && defined(CTRSM_WORK)
/*
* Allocate device memory for the inversed diagonal blocks, size=N*BLOCK_SIZE
*/
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j < 2; j++ ) {
magma_cmalloc( &d_dinvA[d][j], nb*nb );
magma_cmalloc( &d_x[d][j], nb*m );
cudaMemset(d_dinvA[d][j], 0, nb*nb*sizeof(magmaFloatComplex));
cudaMemset(d_x[d][j], 0, nb* m*sizeof(magmaFloatComplex));
}
}
magma_setdevice(0);
#endif
for (j=0; j < n; j += nb) {
/* Set the GPU number that holds the current panel */
id = (j/nb)%num_gpus;
buf = (j/nb)%num_gpus;
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (n-j));
if ( j > 0 ) {
/* needed on pluto... */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] ); // wait for the column on CPU
/* broadcast offdiagonal row to all gpus */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
if ( d != id ) {
magma_setdevice(d);
/* wait for it on CPU */
magma_queue_wait_event( stream[d][stream0], event[id][1] );
/* send it to GPU */
magma_csetmatrix_async( jb, j,
Alo(j,0), lda,
dlPT(d,0,jb,buf), nb,
stream[d][stream0] );
magma_event_record( event[d][1], stream[d][stream0] );
}
d = (d+1)%num_gpus;
}
}
/* Update the current diagonal block */
magma_setdevice(id);
if ( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
magma_cherk(MagmaLower, MagmaNoTrans, jb, j,
d_neg_one, dlA(id, nb*j_local, 0), ldda,
d_one, dlA(id, nb*j_local, j), ldda);
magma_event_record( event[id][0], stream[id][stream1] );
}
/* send the diagonal to cpu */
magma_queue_wait_event( stream[id][stream0], event[id][0] ); // wait for syrk
magma_cgetmatrix_async( jb, jb,
dlA(id, nb*j_local, j), ldda,
Alo(j,j), lda,
stream[id][stream0] );
/* update the offdiagonal blocks */
if ( j > 0 ) {
/* compute the block-rows of the panel */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
j_local2 = j_local+1;
if ( d > id ) j_local2 --;
nb0 = nb*j_local2;
if ( nb0 < n_local[d] ) {
if ( d != id ) {
dlpanel = dlPT(d, 0, jb, buf);
ldpanel = nb;
/* wait for offdiagonal row */
magma_queue_wait_event( stream[d][stream1], event[d][1] );
} else {
dlpanel = dlA(d, nb*j_local, 0);
ldpanel = ldda;
}
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
magma_cgemm( MagmaNoTrans, MagmaConjTrans,
n_local[d]-nb0, jb, j,
c_neg_one, dlA(d, nb0, 0), ldda,
dlpanel, ldpanel,
c_one, dlA(d, nb0, j), ldda);
}
d = (d+1)%num_gpus;
}
}
/* factor the diagonal */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] );
lapackf77_cpotrf(MagmaLowerStr, &jb, Alo(j,j), &lda, info);
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus */
if ( (j+jb) < m ) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
if ( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
magma_csetmatrix_async( jb, jb,
Alo(j,j), lda,
dlpanel, ldpanel,
stream[d][stream0] );
magma_event_record( event[d][2], stream[d][stream0] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
magma_csetmatrix_async( jb, jb,
Alo(j,j), lda,
dlA(id, nb*j_local, j), ldda,
stream[id][stream0] );
}
/* factorize off-diagonal blocks */
if ( (j+jb) < m ) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
/* next column */
j_local2 = j_local+1;
if ( d > id ) j_local2--;
if ( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
nb2 = n_local[d] - j_local2*nb;
nb0 = min(nb, nb2 );
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for the diagonal
if ( j+jb < n && d == (j/nb+1)%num_gpus ) {
/* owns the next column, look-ahead the column */
#if defined(PRECISION_d) && defined(CTRSM_WORK)
magmablas_ctrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
d_dinvA[d][0], d_x[d][0]);
#else
magma_ctrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
magma_event_record( event[d][3], stream[d][stream1] );
/* send the column to cpu */
if ( j+jb < n ) {
magma_queue_wait_event( stream[d][stream0], event[d][3] ); // wait for lookahead
magma_cgetmatrix_async( nb0, j+jb,
dlA(d, nb*j_local2, 0), ldda,
Alo(j+jb,0), lda,
stream[d][stream0] );
magma_event_record( event[d][1], stream[d][stream0] );
}
/* update the remaining blocks */
nb2 = nb2 - nb0;
#if defined(PRECISION_d) && defined(CTRSM_WORK)
magmablas_ctrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_ctrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda);
#endif
} else if ( nb2 > 0 ) {
/* update the entire trailing matrix */
#if defined(PRECISION_d) && defined(CTRSM_WORK)
magmablas_ctrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_ctrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
}
d = (d+1)%num_gpus;
}
}
}
} /* end of else not upper */
/* == finalize the trace == */
trace_finalize( "cpotrf.svg", "trace.css" );
/* clean up */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
//magma_event_destroy( event0[d] );
//magma_event_destroy( event1[d] );
//magma_event_destroy( event2[d] );
//magma_event_destroy( event3[d] );
}
magma_setdevice(0);
return *info;
} /* magma_cpotrf_mgpu */
