int main( int argc, char** argv)
{
real_Double_t gflops, magma_perf, magma_time, clblas_perf, clblas_time, cpu_perf, cpu_time;
float magma_error, clblas_error, work[1];
magma_trans_t transA = MagmaNoTrans;
magma_trans_t transB = MagmaNoTrans;
magma_int_t istart = 1024;
magma_int_t iend = 6240;
magma_int_t M, M0 = 0;
magma_int_t N, N0 = 0;
magma_int_t K, K0 = 0;
magma_int_t i;
magma_int_t Am, An, Bm, Bn;
magma_int_t szeA, szeB, szeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float *h_A, *h_B, *h_C, *h_C2, *h_C3;
magmaFloat_ptr d_A, d_B, d_C;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 0.29, -0.86 );
float beta = MAGMA_S_MAKE( -0.48, 0.38 );
int lapack = getenv("MAGMA_RUN_LAPACK") != NULL;
int count = 1;
printf("\nUsage: testing_sgemm [-NN|NT|TN|TT|NC|CN|TC|CT|CC] -M m -N n -K k -count c -l\n"
" -l or setting $MAGMA_RUN_LAPACK runs CPU BLAS,\n"
" and computes both MAGMA and CLBLAS error using CPU BLAS result.\n"
" Else, MAGMA error is computed using CLBLAS result.\n\n");
for( int i = 1; i < argc; ++i ) {
if ( strcmp("-N", argv[i]) == 0 && i+1 < argc ){
N0 = atoi(argv[++i]);
}
else if ( strcmp("-M", argv[i]) == 0 && i+1 < argc ){
M0 = atoi(argv[++i]);
}
else if ( strcmp("-K", argv[i]) == 0 && i+1 < argc ){
K0 = atoi(argv[++i]);
}
else if (strcmp("-NN", argv[i])==0){
transA = transB = MagmaNoTrans;
}
else if (strcmp("-TT", argv[i])==0){
transA = transB = MagmaTrans;
}
else if (strcmp("-NT", argv[i])==0){
transA = MagmaNoTrans;
transB = MagmaTrans;
}
else if (strcmp("-TN", argv[i])==0){
transA = MagmaTrans;
transB = MagmaNoTrans;
}
else if (strcmp("-NC", argv[i])==0){
transA = MagmaNoTrans;
transB = MagmaConjTrans;
}
else if (strcmp("-TC", argv[i])==0){
transA = MagmaTrans;
transB = MagmaConjTrans;
}
else if (strcmp("-CN", argv[i])==0){
transA = MagmaConjTrans;
transB = MagmaNoTrans;
}
else if (strcmp("-CT", argv[i])==0){
transA = MagmaConjTrans;
transB = MagmaTrans;
}
else if (strcmp("-CC", argv[i])==0){
transA = transB = MagmaConjTrans;
}
else if (strcmp("-l", argv[i])==0) {
lapack = true;
}
else if ( strcmp("-count", argv[i]) == 0 && i+1 < argc ){
count = atoi(argv[++i]);
}
else {
printf( "invalid argument: %s\n", argv[i] );
exit(1);
}
}
if ( (M0 != 0) && (N0 != 0) && (K0 != 0) )
iend = istart + 1;
M = N = K = iend;
if ( M0 != 0 ) M = M0;
if ( N0 != 0 ) N = N0;
if ( K0 != 0 ) K = K0;
if( transA == MagmaNoTrans ) {
Am = M;
An = K;
} else {
Am = K;
An = M;
}
if( transB == MagmaNoTrans ) {
Bm = K;
Bn = N;
} else {
Bm = N;
Bn = K;
}
/* Initialize */
magma_queue_t queue;
magma_device_t device[ MagmaMaxGPUs ];
magma_int_t num = 0;
magma_int_t err;
magma_init();
err = magma_getdevices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_getdevices failed: %d\n", (int) err );
exit(-1);
}
err = magma_queue_create( device[0], &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", (int) err );
exit(-1);
}
lda = ldc = M;
ldb = Bm;
ldda = ((M+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ldda;
K += 32;
M += 32;
N += 32;
TESTING_MALLOC_CPU( h_A, float, lda*K );
TESTING_MALLOC_CPU( h_B, float, ldb*Bn );
TESTING_MALLOC_CPU( h_C, float, ldc*N );
TESTING_MALLOC_CPU( h_C2, float, ldc*N );
TESTING_MALLOC_CPU( h_C3, float, ldc*N );
TESTING_MALLOC_DEV( d_A, float, ldda*K );
TESTING_MALLOC_DEV( d_B, float, lddb*Bn );
TESTING_MALLOC_DEV( d_C, float, lddc*N );
printf("Testing transA = %c transB = %c\n", *lapack_const(transA), *lapack_const(transB));
printf(" M N K MAGMA Gflop/s (sec) CLBLAS Gflop/s (sec) CPU Gflop/s (sec) MAGMA error CLBLAS error\n");
printf("===========================================================================================================\n");
for( i=istart; i<iend; i = (int)(i*1.25) ) {
for( int cnt = 0; cnt < count; ++cnt ) {
M = N = K = i;
if ( M0 != 0 ) M = M0;
if ( N0 != 0 ) N = N0;
if ( K0 != 0 ) K = K0;
if( transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if( transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
gflops = FLOPS_SGEMM( M, N, K ) / 1e9;
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
szeA = lda * An;
szeB = ldb * Bn;
szeC = ldc * N;
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &szeA, h_A );
lapackf77_slarnv( &ione, ISEED, &szeB, h_B );
lapackf77_slarnv( &ione, ISEED, &szeC, h_C );
/* =====================================================================
Performs operation using MAGMA-BLAS
=================================================================== */
magma_ssetmatrix( Am, An, h_A, lda, d_A, 0, ldda, queue );
magma_ssetmatrix( Bm, Bn, h_B, ldb, d_B, 0, lddb, queue );
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magmablas_sgemm_reduce( M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magma_queue_sync(queue);
magma_time = magma_wtime();
magmablas_sgemm_reduce( M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_queue_sync(queue);
magma_time = magma_wtime() - magma_time;
magma_perf = gflops / magma_time;
magma_sgetmatrix( M, N, d_C, 0, lddc, h_C2, ldc, queue );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magma_sgemm( transA, transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magma_queue_sync(queue);
clblas_time = magma_wtime();
magma_sgemm( transA, transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_queue_sync(queue);
clblas_time = magma_wtime() - clblas_time;
clblas_perf = gflops / clblas_time;
magma_sgetmatrix( M, N, d_C, 0, lddc, h_C3, ldc, queue );
/* =====================================================================
Performs operation using BLAS
=================================================================== */
if ( lapack ) {
cpu_time = magma_wtime();
blasf77_sgemm( lapack_const(transA), lapack_const(transB), &M, &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Error Computation and Performance Compariosn
=================================================================== */
if ( lapack ) {
// compare both magma & clblas to lapack
blasf77_saxpy(&szeC, &c_neg_one, h_C, &ione, h_C2, &ione);
magma_error = lapackf77_slange("M", &M, &N, h_C2, &ldc, work);
blasf77_saxpy(&szeC, &c_neg_one, h_C, &ione, h_C3, &ione);
clblas_error = lapackf77_slange("M", &M, &N, h_C3, &ldc, work);
printf("%5d %5d %5d %7.2f (%7.4f) %7.2f (%7.4f) %7.2f (%7.4f) %8.2e %8.2e\n",
(int) M, (int) N, (int) K,
magma_perf, magma_time, clblas_perf, clblas_time, cpu_perf, cpu_time,
magma_error, clblas_error );
}
else {
// compare magma to clblas
blasf77_saxpy(&szeC, &c_neg_one, h_C3, &ione, h_C2, &ione);
magma_error = lapackf77_slange("M", &M, &N, h_C2, &ldc, work);
printf("%5d %5d %5d %7.2f (%7.4f) %7.2f (%7.4f) --- ( --- ) %8.2e ---\n",
(int) M, (int) N, (int) K,
magma_perf, magma_time, clblas_perf, clblas_time,
magma_error );
}
}
if ( count > 1 ) {
printf( "\n" );
}
}
/* Memory clean up */
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_C2 );
TESTING_FREE_CPU( h_C3 );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
magma_queue_destroy( queue );
magma_finalize();
}
/**
Purpose
-------
SPOTRF computes the Cholesky factorization of a real symmetric
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]
d_lA REAL array of pointers on the GPU, dimension (ngpu)
On entry, the symmetric matrix dA distributed over GPUs
(dl_A[d] points to the local matrix on the d-th GPU).
It is distributed in 1D block column or row cyclic (with the
block size of nb) if UPLO = MagmaUpper or MagmaLower, respectively.
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_sposv_comp
********************************************************************/
extern "C" magma_int_t
magma_spotrf_mgpu_right(
magma_int_t ngpu,
magma_uplo_t uplo, magma_int_t n,
magmaFloat_ptr d_lA[], magma_int_t ldda,
magma_int_t *info )
{
#define dlA(id, i, j) (d_lA[(id)] + (j) * ldda + (i))
#define dlP(id, i, j) (d_lP[(id)] + (j) * ldda + (i))
#define panel(j) (panel + (j))
#define tmppanel(j) (tmppanel + (j))
#define tmpprevpanel(j) (tmpprevpanel + (j))
#define STREAM_ID(i) (nqueue > 1 ? 1+((i)/nb)%(nqueue-1) : 0)
float z_one = MAGMA_S_MAKE( 1.0, 0.0 );
float mz_one = MAGMA_S_MAKE( -1.0, 0.0 );
float one = 1.0;
float m_one = -1.0;
const char* uplo_ = lapack_uplo_const( uplo );
magma_int_t j, nb, d, id, j_local, blkid, crosspoint, prevtrsmrows=0, nqueue = 5;
float *panel, *tmppanel0, *tmppanel1, *tmppanel, *tmpprevpanel;
float *d_lP[MagmaMaxGPUs], *dlpanel, *dlpanels[MagmaMaxGPUs];
magma_int_t rows, trsmrows, igpu, n_local[MagmaMaxGPUs], ldpanel;
magma_queue_t queues[MagmaMaxGPUs][10];
*info = 0;
if ( uplo != MagmaUpper && uplo != MagmaLower ) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
nb = magma_get_spotrf_nb(n);
ldpanel = ldda;
magma_setdevice(0);
if (MAGMA_SUCCESS != magma_smalloc_pinned( &panel, 2 * nb * ldpanel )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
tmppanel0 = panel;
tmppanel1 = tmppanel0 + nb * ldpanel;
if ((nb <= 1) || (nb >= n)) {
// Use unblocked code.
magma_sgetmatrix( n, n, dlA(0, 0, 0), ldda, panel, ldpanel);
lapackf77_spotrf( uplo_, &n, panel, &ldpanel, info);
magma_ssetmatrix( n, n, panel, ldpanel, dlA(0, 0, 0), ldda );
} else {
for( d = 0; d < ngpu; d++ ) {
// local-n and local-ld
n_local[d] = ((n / nb) / ngpu) * nb;
if (d < (n / nb) % ngpu)
n_local[d] += nb;
else if (d == (n / nb) % ngpu)
n_local[d] += n % nb;
magma_setdevice(d);
magma_device_sync();
if (MAGMA_SUCCESS != magma_smalloc( &d_lP[d], nb * ldda )) {
for( j = 0; j < d; j++ ) {
magma_setdevice(j);
magma_free( d_lP[d] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
for( j=0; j < nqueue; j++ ) {
magma_queue_create( &queues[d][j] );
}
}
//#define ENABLE_TIMER
#if defined (ENABLE_TIMER)
real_Double_t therk[4], tmtc, tcchol, tctrsm, tctm, tmnp, tcnp;
real_Double_t ttot_herk[4] = {0,0,0,0}, ttot_mtc = 0, ttot_cchol = 0, ttot_ctrsm = 0, ttot_ctm = 0, ttot_mnp = 0, ttot_cnp = 0;
printf("\n\n %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s\n",
"j", "nb", "row", "mtc", "CPU_np", "panel", "ctrsm", "CH+TRSM", "CPU", "dsyrk[0]", "dsyrk[1]", "dsyrk[2]", "dsyrk[3]", "ctm P", "gpu_np");
printf(" ====================================================================================================\n");
#endif
// Use blocked code.
if (uplo == MagmaUpper) {
printf( " === not supported, yet ===\n" );
} else {
blkid = -1;
if (ngpu == 4)
crosspoint = n;
else if (ngpu == 3)
crosspoint = n;
else if (ngpu == 2)
crosspoint = 20160;
else
crosspoint = 0;
crosspoint = 0; //n; //n -- > gpu always does next panel, 0 --> cpu always does next panel
crosspoint = n;
#if defined (ENABLE_TIMER)
real_Double_t tget = magma_wtime(), tset = 0.0, ttot = 0.0;
#endif
if ( n > nb ) {
// send first panel to cpu
magma_setdevice(0);
tmppanel = tmppanel0;
magma_sgetmatrix_async(n, nb,
dlA(0, 0, 0), ldda,
tmppanel(0), ldpanel,
queues[0][0] );
}
#if defined (ENABLE_TIMER)
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_device_sync();
}
tget = magma_wtime()-tget;
#endif
// Compute the Cholesky factorization A = L*L'
for (j = 0; (j + nb) < n; j += nb) {
#if defined (ENABLE_TIMER)
therk[0] = therk[1] = therk[2] = therk[3] = tmtc = tcchol = tctrsm = tctm = tmnp = tcnp = 0.0;
#endif
blkid += 1;
tmppanel = (blkid % 2 == 0) ? tmppanel0 : tmppanel1;
// Set the gpu number that holds the current panel
id = (j / nb) % ngpu;
magma_setdevice(id);
// Set the local index where the current panel is
j_local = j / (nb * ngpu) * nb;
rows = n - j;
// Wait for the panel on cpu
magma_queue_sync( queues[id][0] );
if (j > 0 && prevtrsmrows > crosspoint) {
#if defined (ENABLE_TIMER)
tcnp = magma_wtime();
#endif
tmpprevpanel = ((blkid - 1) % 2) == 0 ? tmppanel0 : tmppanel1;
blasf77_sgemm( MagmaNoTransStr, MagmaConjTransStr,
&rows, &nb, &nb,
&mz_one, tmpprevpanel(j), &ldpanel,
tmpprevpanel(j), &ldpanel,
&z_one, tmppanel(j), &ldpanel );
#if defined (ENABLE_TIMER)
tcnp = magma_wtime() - tcnp;
ttot_cnp += tcnp;
#endif
}
#if defined (ENABLE_TIMER)
tcchol = magma_wtime();
#endif
lapackf77_spotrf(MagmaLowerStr, &nb, tmppanel(j), &ldpanel, info);
if (*info != 0) {
*info = *info + j;
break;
}
#if defined (ENABLE_TIMER)
tcchol = magma_wtime() - tcchol;
ttot_cchol += tcchol;
tctrsm = magma_wtime();
#endif
trsmrows = rows - nb;
if (trsmrows > 0) {
blasf77_strsm(MagmaRightStr, MagmaLowerStr, MagmaConjTransStr, MagmaNonUnitStr,
&trsmrows, &nb,
&z_one, tmppanel(j), &ldpanel,
tmppanel(j + nb), &ldpanel);
}
#if defined (ENABLE_TIMER)
tctrsm = magma_wtime() - tctrsm;
ttot_ctrsm += tctrsm;
tctm = magma_wtime();
#endif
d = (id + 1) % ngpu;
// send current panel to gpus
for (igpu = 0; igpu < ngpu; igpu++, d = (d + 1) % ngpu ) {
magma_int_t myrows = 0;
magma_int_t row_offset = 0;
if ( d == id ) {
dlpanel = dlA(d, j, j_local);
myrows = rows;
row_offset = 0;
} else {
dlpanel = dlP(d, 0, 0);
myrows = trsmrows;
row_offset = nb;
}
if (myrows > 0) {
magma_setdevice(d);
magma_ssetmatrix_async(myrows, nb,
tmppanel(j + row_offset), ldpanel,
dlpanel, ldda, queues[d][0] );
}
}
/* make sure panel is on GPUs */
d = (id + 1) % ngpu;
for (igpu = 0; igpu < ngpu; igpu++, d = (d + 1) % ngpu ) {
magma_setdevice(d);
magma_queue_sync( queues[d][0] );
}
#if defined (ENABLE_TIMER)
tctm = magma_wtime() - tctm;
ttot_ctm += tctm;
#endif
if ( (j + nb) < n) {
magma_int_t offset = 0;
magma_int_t row_offset = 0;
if (j + nb + nb < n) {
d = (id + 1) % ngpu;
magma_setdevice(d);
magma_int_t j_local2 = (j + nb) / (nb * ngpu) * nb;
if (trsmrows <= crosspoint) {
#if defined (ENABLE_TIMER)
tmnp = magma_wtime();
#endif
// do gemm on look ahead panel
if ( d == id ) {
dlpanel = dlA(d, j + nb, j_local);
} else {
dlpanel = dlP(d, 0, 0);
}
magmablasSetKernelStream( queues[d][STREAM_ID(j_local2)] );
#define SSYRK_ON_DIAG
#ifdef SSYRK_ON_DIAG
magma_ssyrk( MagmaLower, MagmaNoTrans,
nb, nb,
m_one, dlpanel, ldda,
one, dlA(d, j + nb, j_local2), ldda);
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
trsmrows-nb, nb, nb,
mz_one, dlpanel+nb, ldda,
dlpanel, ldda,
z_one, dlA(d, j + nb +nb, j_local2), ldda);
#else
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
trsmrows, nb, nb,
mz_one, dlpanel, ldda,
dlpanel, ldda,
z_one, dlA(d, j + nb, j_local2), ldda);
#endif
#if defined (ENABLE_TIMER)
magma_device_sync();
tmnp = magma_wtime() - tmnp;
ttot_mnp += tmnp;
#endif
}
// send next panel to cpu
magma_queue_sync( queues[d][STREAM_ID(j_local2)] ); // make sure lookahead is done
tmppanel = ((blkid+1) % 2 == 0) ? tmppanel0 : tmppanel1;
magma_sgetmatrix_async(rows-nb, nb,
dlA(d, j+nb, j_local2), ldda,
tmppanel(j+nb), ldpanel,
queues[d][0] );
tmppanel = (blkid % 2 == 0) ? tmppanel0 : tmppanel1;
offset = j + nb + nb;
row_offset = nb;
} else {
offset = j + nb;
row_offset = 0;
}
if (n - offset > 0) {
// syrk on multiple gpu
for (d = 0; d < ngpu; d++ ) {
if ( d == id ) {
dlpanels[d] = dlA(d, j + nb + row_offset, j_local);
} else {
dlpanels[d] = dlP(d, row_offset, 0);
}
}
#if defined (ENABLE_TIMER)
for( d=0; d < ngpu; d++ ) therk[d] = magma_wtime();
#endif
//magmablasSetKernelStream( queues[d] );
//magma_ssyrk(MagmaLower, MagmaNoTrans, n - offset, nb,
// m_one, dlpanel, ldda,
// one, &d_lA[d][offset + offset*ldda], ldda );
#ifdef SSYRK_ON_DIAG
magma_ssyrk_mgpu
#else
magma_ssyrk_mgpu2
#endif
(ngpu, MagmaLower, MagmaNoTrans,
nb, n - offset, nb,
m_one, dlpanels, ldda, 0,
one, d_lA, ldda, offset,
nqueue, queues );
#if defined (ENABLE_TIMER)
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_device_sync();
therk[d] = magma_wtime() - therk[d];
ttot_herk[d] += therk[d];
}
#endif
}
prevtrsmrows = trsmrows;
#if defined (ENABLE_TIMER)
ttot += (tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp);
printf("%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(%d) %10.3lf\n",
j, nb, rows, tmtc,
tcnp, // gemm
tcchol, // potrf
tctrsm, // trsm
(tcchol + tctrsm),
(tmtc+tcnp+tcchol+tctrsm),
therk[0], therk[1], therk[2], therk[3], // syrk
tctm, // copy panel to GPU
tmnp, // lookahead on GPU
(id + 1) % ngpu,
(tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp));
fflush(0);
#endif
}
}
for( d = 0; d < ngpu; d++ ) {
magma_setdevice(d);
for( id=0; id < nqueue; id++ ) {
magma_queue_sync( queues[d][id] );
}
}
#if defined (ENABLE_TIMER)
printf("\n%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(-) %10.3lf\n",
n, n, 0, ttot_mtc,
ttot_cnp, // gemm
ttot_cchol, // potrf
ttot_ctrsm, // trsm
(ttot_cchol + ttot_ctrsm),
(ttot_mtc+ttot_cnp+ttot_cchol+ttot_ctrsm),
ttot_herk[0], ttot_herk[1], ttot_herk[2], ttot_herk[3], // syrk
ttot_ctm, // copy panel to GPU
ttot_mnp, // lookahead on GPU
(ttot_cnp+ttot_cchol+ttot_ctrsm+ttot_herk[0]+ttot_herk[1]+ttot_herk[2]+ttot_ctm+ttot_mnp));
printf("%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(-) %10.3lf (ratio)\n",
n, n, 0, ttot_mtc/ttot,
ttot_cnp/ttot, // gemm
ttot_cchol/ttot, // potrf
ttot_ctrsm/ttot, // trsm
(ttot_cchol + ttot_ctrsm)/ttot,
(ttot_mtc+ttot_cnp+ttot_cchol+ttot_ctrsm)/ttot,
ttot_herk[0]/ttot, ttot_herk[1]/ttot, ttot_herk[2]/ttot, ttot_herk[3]/ttot, // syrk
ttot_ctm/ttot, // copy panel to GPU
ttot_mnp/ttot, // lookahead on GPU
(ttot_cnp+ttot_cchol+ttot_ctrsm+ttot_herk[0]+ttot_herk[1]+ttot_herk[2]+ttot_ctm+ttot_mnp)/ttot);
#endif
// cholesky for the last block
if (j < n && *info == 0) {
rows = n - j;
id = (j / nb) % ngpu;
// Set the local index where the current panel is
j_local = j / (nb * ngpu) * nb;
magma_setdevice(id);
#if defined (ENABLE_TIMER)
tset = magma_wtime();
#endif
magma_sgetmatrix(rows, rows, dlA(id, j, j_local), ldda, panel(j), ldpanel);
lapackf77_spotrf(MagmaLowerStr, &rows, panel(j), &ldpanel, info);
magma_ssetmatrix(rows, rows, panel(j), ldpanel, dlA(id, j, j_local), ldda);
#if defined (ENABLE_TIMER)
tset = magma_wtime() - tset;
#endif
}
#if defined (ENABLE_TIMER)
printf( " matrix_get,set: %10.3lf %10.3lf -> %10.3lf\n",tget,tset,ttot+tget+tset );
#endif
} // end of else not upper
// clean up
for( d = 0; d < ngpu; d++ ) {
magma_setdevice(d);
for( j=0; j < nqueue; j++ ) {
magma_queue_destroy( queues[d][j] );
}
magma_free( d_lP[d] );
}
} // end of not lapack
// free workspace
magma_free_pinned( panel );
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_spotrf_mgpu_right */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgemm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
float magma_error, dev_error, Cnorm, work[1];
magma_int_t M, N, K;
magma_int_t Am, An, Bm, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
float *h_A, *h_B, *h_C, *h_Cmagma, *h_Cdev;
magmaFloat_ptr d_A, d_B, d_C;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 0.29, -0.86 );
float beta = MAGMA_S_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
#ifdef HAVE_CUBLAS
// for CUDA, we can check MAGMA vs. CUBLAS, without running LAPACK
printf("If running lapack (option --lapack), MAGMA and %s error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to %s result.\n\n",
g_platform_str, g_platform_str );
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
// for others, we need LAPACK for check
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K %s Gflop/s (ms) CPU Gflop/s (ms) %s error\n",
g_platform_str, g_platform_str );
#endif
printf("=========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
K = opts.ksize[itest];
gflops = FLOPS_SGEMM( M, N, K ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if ( opts.transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*An;
sizeB = ldb*Bn;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, float, lda*An );
TESTING_MALLOC_CPU( h_B, float, ldb*Bn );
TESTING_MALLOC_CPU( h_C, float, ldc*N );
TESTING_MALLOC_CPU( h_Cmagma, float, ldc*N );
TESTING_MALLOC_CPU( h_Cdev, float, ldc*N );
TESTING_MALLOC_DEV( d_A, float, ldda*An );
TESTING_MALLOC_DEV( d_B, float, lddb*Bn );
TESTING_MALLOC_DEV( d_C, float, lddc*N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_slarnv( &ione, ISEED, &sizeC, h_C );
magma_ssetmatrix( Am, An, h_A, lda, d_A, 0, ldda, opts.queue );
magma_ssetmatrix( Bm, Bn, h_B, ldb, d_B, 0, lddb, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_ssetmatrix( M, N, h_C, ldc, d_C, lddc );
magma_time = magma_sync_wtime( NULL );
magmablas_sgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, ldda,
d_B, lddb,
beta, d_C, lddc );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_sgetmatrix( M, N, d_C, lddc, h_Cmagma, ldc );
#endif
/* =====================================================================
Performs operation using CUBLAS / clBLAS / Xeon Phi MKL
=================================================================== */
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, opts.queue );
#ifdef HAVE_CUBLAS
dev_time = magma_sync_wtime( NULL );
cublasSgemm( opts.handle, cublas_trans_const(opts.transA), cublas_trans_const(opts.transB), M, N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
dev_time = magma_sync_wtime( NULL ) - dev_time;
#else
dev_time = magma_sync_wtime( opts.queue );
magma_sgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, opts.queue );
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
#endif
dev_perf = gflops / dev_time;
magma_sgetmatrix( M, N, d_C, 0, lddc, h_Cdev, ldc, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_sgemm( lapack_trans_const(opts.transA), lapack_trans_const(opts.transB), &M, &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & dev, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_slange( "F", &M, &N, h_C, &ldc, work );
blasf77_saxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cdev, &ione );
dev_error = lapackf77_slange( "F", &M, &N, h_Cdev, &ldc, work ) / Cnorm;
#ifdef HAVE_CUBLAS
blasf77_saxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cmagma, &ione );
magma_error = lapackf77_slange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(magma_error < tol && dev_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && dev_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(dev_error < tol ? "ok" : "failed"));
status += ! (dev_error < tol);
#endif
}
else {
#ifdef HAVE_CUBLAS
// compute relative error for magma, relative to dev (currently only with CUDA)
Cnorm = lapackf77_slange( "F", &M, &N, h_Cdev, &ldc, work );
blasf77_saxpy( &sizeC, &c_neg_one, h_Cdev, &ione, h_Cmagma, &ione );
magma_error = lapackf77_slange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e --- %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
magma_error,
(magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) --- ( --- ) ---\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time );
#endif
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Cmagma );
TESTING_FREE_CPU( h_Cdev );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgegqr
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float e1, e2, e3, e4, e5, *work;
float c_neg_one = MAGMA_S_NEG_ONE;
float c_one = MAGMA_S_ONE;
float c_zero = MAGMA_S_ZERO;
float *h_A, *h_R, *tau, *dtau, *h_work, *h_rwork, tmp[1];
float *d_A, *dwork;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn;
magma_int_t ione = 1, ldwork;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
// versions 1...4 are valid
if (opts.version < 1 || opts.version > 4) {
printf("Unknown version %d; exiting\n", opts.version );
return -1;
}
float tol, eps = lapackf77_slamch("E");
tol = 10* opts.tolerance * eps;
printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||I-Q'Q||_F / M ||I-Q'Q||_I / M ||A-Q R||_I\n");
printf(" MAGMA / LAPACK MAGMA / LAPACK\n");
printf("==========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if (N > 128) {
printf("%5d %5d skipping because sgegqr requires N <= 128\n",
(int) M, (int) N);
continue;
}
if (M < N) {
printf("%5d %5d skipping because sgegqr requires M >= N\n",
(int) M, (int) N);
continue;
}
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = FLOPS_SGEQRF( M, N ) / 1e9 + FLOPS_SORGQR( M, N, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgeqrf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] );
lwork = max(lwork, 3*N*N);
ldwork = N*N;
if (opts.version == 2) {
ldwork = 3*N*N + min_mn;
}
TESTING_MALLOC_PIN( tau, float, min_mn );
TESTING_MALLOC_PIN( h_work, float, lwork );
TESTING_MALLOC_PIN(h_rwork, float, lwork );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_CPU( h_R, float, n2 );
TESTING_MALLOC_CPU( work, float, M );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( dtau, float, min_mn );
TESTING_MALLOC_DEV( dwork, float, ldwork );
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_ssetmatrix( M, N, h_R, lda, d_A, ldda );
// warmup
magma_sgegqr_gpu( 1, M, N, d_A, ldda, dwork, h_work, &info );
magma_ssetmatrix( M, N, h_R, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_sync_wtime( 0 );
magma_sgegqr_gpu( opts.version, M, N, d_A, ldda, dwork, h_rwork, &info );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sgegqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_sgetmatrix( M, N, d_A, ldda, h_R, M );
// Regenerate R
// blasf77_sgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_A, &M, &c_zero, h_rwork, &N);
// magma_sprint(N, N, h_work, N);
blasf77_strmm("r", "u", "n", "n", &M, &N, &c_one, h_rwork, &N, h_R, &M);
blasf77_saxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
e5 = lapackf77_slange("i", &M, &N, h_R, &M, work) /
lapackf77_slange("i", &M, &N, h_A, &lda, work);
magma_sgetmatrix( M, N, d_A, ldda, h_R, M );
if ( opts.lapack ) {
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
/* Orthogonalize on the CPU */
lapackf77_sgeqrf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
lapackf77_sorgqr(&M, &N, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sorgqr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
blasf77_sgemm("t", "n", &N, &N, &M, &c_one, h_R, &M, h_R, &M, &c_zero, h_work, &N);
for(int ii = 0; ii < N*N; ii += N+1 ) {
h_work[ii] = MAGMA_S_SUB(h_work[ii], c_one);
}
e1 = lapackf77_slange("f", &N, &N, h_work, &N, work) / N;
e3 = lapackf77_slange("i", &N, &N, h_work, &N, work) / N;
blasf77_sgemm("t", "n", &N, &N, &M, &c_one, h_A, &M, h_A, &M, &c_zero, h_work, &N);
for(int ii = 0; ii < N*N; ii += N+1 ) {
h_work[ii] = MAGMA_S_SUB(h_work[ii], c_one);
}
e2 = lapackf77_slange("f", &N, &N, h_work, &N, work) / N;
e4 = lapackf77_slange("i", &N, &N, h_work, &N, work) / N;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e / %8.2e %8.2e / %8.2e %8.2e %s\n",
(int) M, (int) N, cpu_perf, 1000.*cpu_time, gpu_perf, 1000.*gpu_time,
e1, e2, e3, e4, e5,
(e1 < tol ? "ok" : "failed"));
status += ! (e1 < tol);
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, 1000.*gpu_time );
}
TESTING_FREE_PIN( tau );
TESTING_FREE_PIN( h_work );
TESTING_FREE_PIN( h_rwork );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_CPU( work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dtau );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
SLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to SLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code makes very mild assumptions about floating point
arithmetic. It will work on machines with a guard digit in
add/subtract, or on those binary machines without guard digits
which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
---------
@param[in]
k INTEGER
The number of terms in the rational function to be solved by
SLAED4. K >= 0.
@param[in]
n INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N > K).
@param[in]
n1 INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
@param[out]
d REAL array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
@param[out]
Q REAL array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
@param[in]
ldq INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
@param[in]
rho REAL
The value of the parameter in the rank one update equation.
RHO >= 0 required.
@param[in,out]
dlamda REAL array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
@param[in]
Q2 REAL array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
TODO what is LDQ2?
@param[in]
indx INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see SLAED2).
The rows of the eigenvectors found by SLAED4 must be likewise
permuted before the matrix multiply can take place.
@param[in]
ctot INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
@param[in,out]
w REAL array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
@param
s (workspace) REAL array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
@param[out]
indxq INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
@param
dwork (workspace) REAL array, dimension (3*N*N/2+3*N)
@param[in]
range magma_range_t
- = MagmaRangeAll: all eigenvalues will be found.
- = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU]
will be found.
- = MagmaRangeI: the IL-th through IU-th eigenvalues will be found.
TODO verify range, vl, vu, il, iu -- copied from slaex1.
@param[in]
vl REAL
@param[in]
vu REAL
if RANGE=MagmaRangeV, the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeI.
@param[in]
il INTEGER
@param[in]
iu INTEGER
if RANGE=MagmaRangeI, the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeV.
@param[out]
info INTEGER
- = 0: successful exit.
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: if INFO = 1, an eigenvalue did not converge
Further Details
---------------
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
@ingroup magma_ssyev_aux
********************************************************************/
extern "C" magma_int_t
magma_slaex3(magma_int_t k, magma_int_t n, magma_int_t n1, float* d,
float* Q, magma_int_t ldq, float rho,
float* dlamda, float* Q2, magma_int_t* indx,
magma_int_t* ctot, float* w, float* s, magma_int_t* indxq,
float* dwork,
magma_range_t range, float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t* info )
{
#define Q(i_,j_) (Q + (i_) + (j_)*ldq)
float d_one = 1.;
float d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
magma_int_t iil, iiu, rk;
float* dq2= dwork;
float* ds = dq2 + n*(n/2+1);
float* dq = ds + n*(n/2+1);
magma_int_t lddq = n/2 + 1;
magma_int_t i, iq2, j, n12, n2, n23, tmp, lq2;
float temp;
magma_int_t alleig, valeig, indeig;
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
*info = 0;
if (k < 0)
*info=-1;
else if (n < k)
*info=-2;
else if (ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if (*info != 0) {
magma_xerbla(__func__, -(*info));
return *info;
}
// Quick return if possible
if (k == 0)
return *info;
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
lq2 = iq2 + n2 * n23;
magma_ssetvector_async( lq2, Q2, 1, dq2, 1, NULL );
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for (i = ib; i < ie; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = ib; j < ie; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0) {
#pragma omp critical (info)
*info=iinfo;
break;
}
}
#pragma omp barrier
if (*info == 0) {
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2) {
#pragma omp single
{
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for (j = 0; j < k; ++j) {
magma_int_t i_tmp = min(j, ie);
for (i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for (i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if (id < tot) {
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else {
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for (j = ib; j < ie; ++j) {
for (i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s+id*k, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return *info;
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
for (i = 0; i < k; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = 0; j < k; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0)
*info=iinfo;
}
if (*info != 0)
return *info;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2) {
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &k, Q, &tmp, w, &ione);
for (j = 0; j < k; ++j) {
for (i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for (i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for (j = iil-1; j < iiu; ++j) {
for (i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#endif //_OPENMP
// Compute the updated eigenvectors.
timer_start( time );
magma_queue_sync( NULL );
if (rk != 0) {
if ( n23 != 0 ) {
if (rk < magma_get_slaed3_k()) {
lapackf77_slacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_sgemm("N", "N", &n2, &rk, &n23, &d_one, &Q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
} else {
magma_ssetmatrix( n23, rk, Q(ctot[0],iil-1), ldq, ds, n23 );
magma_sgemm( MagmaNoTrans, MagmaNoTrans, n2, rk, n23, d_one, &dq2[iq2], n2, ds, n23, d_zero, dq, lddq);
magma_sgetmatrix( n2, rk, dq, lddq, Q(n1,iil-1), ldq );
}
} else
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 != 0 ) {
if (rk < magma_get_slaed3_k()) {
lapackf77_slacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_sgemm("N", "N", &n1, &rk, &n12, &d_one, Q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
} else {
magma_ssetmatrix( n12, rk, Q(0,iil-1), ldq, ds, n12 );
magma_sgemm( MagmaNoTrans, MagmaNoTrans, n1, rk, n12, d_one, dq2, n1, ds, n12, d_zero, dq, lddq);
magma_sgetmatrix( n1, rk, dq, lddq, Q(0,iil-1), ldq );
}
} else
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
timer_stop( time );
timer_printf( "gemms = %6.2f\n", time );
return *info;
} /* magma_slaex3 */
extern "C" magma_int_t
magma_sgetrf_nopiv(magma_int_t *m, magma_int_t *n, float *a,
magma_int_t *lda, magma_int_t *info)
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
SGETRF_NOPIV computes an LU factorization of a general M-by-N
matrix A without pivoting.
The factorization has the form
A = L * U
where L is lower triangular with unit diagonal elements (lower
trapezoidal if m > n), and U is upper triangular (upper
trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
A (input/output) COMPLEX*16 array, dimension (LDA,N)
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,M).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
float c_one = MAGMA_S_ONE;
magma_int_t a_dim1, a_offset, min_mn, i__3, i__4;
float z__1;
magma_int_t j, jb, nb, iinfo;
a_dim1 = *lda;
a_offset = 1 + a_dim1;
a -= a_offset;
/* Function Body */
*info = 0;
if (*m < 0) {
*info = -1;
} else if (*n < 0) {
*info = -2;
} else if (*lda < max(1,*m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (*m == 0 || *n == 0) {
return *info;
}
/* Determine the block size for this environment. */
nb = 128;
min_mn = min(*m,*n);
if (nb <= 1 || nb >= min_mn)
{
/* Use unblocked code. */
magma_sgetf2_nopiv(m, n, &a[a_offset], lda, info);
}
else
{
/* Use blocked code. */
for (j = 1; j <= min_mn; j += nb)
{
/* Computing MIN */
i__3 = min_mn - j + 1;
jb = min(i__3,nb);
/* Factor diagonal and subdiagonal blocks and test for exact
singularity. */
i__3 = *m - j + 1;
//magma_sgetf2_nopiv(&i__3, &jb, &a[j + j * a_dim1], lda, &iinfo);
i__3 -= jb;
magma_sgetf2_nopiv(&jb, &jb, &a[j + j * a_dim1], lda, &iinfo);
blasf77_strsm("R", "U", "N", "N", &i__3, &jb, &c_one,
&a[j + j * a_dim1], lda,
&a[j + jb + j * a_dim1], lda);
/* Adjust INFO */
if (*info == 0 && iinfo > 0)
*info = iinfo + j - 1;
if (j + jb <= *n)
{
/* Compute block row of U. */
i__3 = *n - j - jb + 1;
blasf77_strsm("Left", "Lower", "No transpose", "Unit", &jb, &i__3,
&c_one, &a[j + j * a_dim1], lda, &a[j + (j+jb)*a_dim1], lda);
if (j + jb <= *m)
{
/* Update trailing submatrix. */
i__3 = *m - j - jb + 1;
i__4 = *n - j - jb + 1;
z__1 = MAGMA_S_NEG_ONE;
blasf77_sgemm("No transpose", "No transpose", &i__3, &i__4, &jb,
&z__1, &a[j + jb + j * a_dim1], lda,
&a[j + (j + jb) * a_dim1], lda, &c_one,
&a[j + jb + (j + jb) * a_dim1], lda);
}
}
}
}
return *info;
} /* magma_sgetrf_nopiv */
/**
Purpose
-------
SLAQPS computes a step of QR factorization with column pivoting
of a real M-by-N matrix A by using Blas-3. It tries to factorize
NB columns from A starting from the row OFFSET+1, and updates all
of the matrix with Blas-3 xGEMM.
In some cases, due to catastrophic cancellations, it cannot
factorize NB columns. Hence, the actual number of factorized
columns is returned in KB.
Block A(1:OFFSET,1:N) is accordingly pivoted, but not factorized.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0
@param[in]
offset INTEGER
The number of rows of A that have been factorized in
previous steps.
@param[in]
nb INTEGER
The number of columns to factorize.
@param[out]
kb INTEGER
The number of columns actually factorized.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, block A(OFFSET+1:M,1:KB) is the triangular
factor obtained and block A(1:OFFSET,1:N) has been
accordingly pivoted, but no factorized.
The rest of the matrix, block A(OFFSET+1:M,KB+1:N) has
been updated.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[in,out]
jpvt INTEGER array, dimension (N)
JPVT(I) = K <==> Column K of the full matrix A has been
permuted into position I in AP.
@param[out]
tau REAL array, dimension (KB)
The scalar factors of the elementary reflectors.
@param[in,out]
vn1 REAL array, dimension (N)
The vector with the partial column norms.
@param[in,out]
vn2 REAL array, dimension (N)
The vector with the exact column norms.
@param[in,out]
auxv REAL array, dimension (NB)
Auxiliar vector.
@param[in,out]
F REAL array, dimension (LDF,NB)
Matrix F' = L*Y'*A.
@param[in]
ldf INTEGER
The leading dimension of the array F. LDF >= max(1,N).
@ingroup magma_sgeqp3_aux
********************************************************************/
extern "C" magma_int_t
magma_slaqps(magma_int_t m, magma_int_t n, magma_int_t offset,
magma_int_t nb, magma_int_t *kb,
float *A, magma_int_t lda,
float *dA, magma_int_t ldda,
magma_int_t *jpvt, float *tau, float *vn1, float *vn2,
float *auxv,
float *F, magma_int_t ldf,
float *dF, magma_int_t lddf)
{
#define A(i, j) (A + (i) + (j)*(lda ))
#define dA(i, j) (dA + (i) + (j)*(ldda))
#define F(i, j) (F + (i) + (j)*(ldf ))
#define dF(i, j) (dF + (i) + (j)*(lddf))
float c_zero = MAGMA_S_MAKE( 0.,0.);
float c_one = MAGMA_S_MAKE( 1.,0.);
float c_neg_one = MAGMA_S_MAKE(-1.,0.);
magma_int_t ione = 1;
magma_int_t i__1, i__2;
float d__1;
float z__1;
magma_int_t j, k, rk;
float Akk;
magma_int_t pvt;
float temp, temp2, tol3z;
magma_int_t itemp;
magma_int_t lsticc;
magma_int_t lastrk;
lastrk = min( m, n + offset );
tol3z = magma_ssqrt( lapackf77_slamch("Epsilon"));
magma_queue_t stream;
magma_queue_create( &stream );
lsticc = 0;
k = 0;
while( k < nb && lsticc == 0 ) {
rk = offset + k;
/* Determine ith pivot column and swap if necessary */
// subtract 1 from Fortran isamax; pvt, k are 0-based.
i__1 = n-k;
pvt = k + blasf77_isamax( &i__1, &vn1[k], &ione ) - 1;
if (pvt != k) {
if (pvt >= nb) {
/* 1. Start copy from GPU */
magma_sgetmatrix_async( m - offset - nb, 1,
dA(offset + nb, pvt), ldda,
A (offset + nb, pvt), lda, stream );
}
/* F gets swapped so F must be sent at the end to GPU */
i__1 = k;
blasf77_sswap( &i__1, F(pvt,0), &ldf, F(k,0), &ldf );
itemp = jpvt[pvt];
jpvt[pvt] = jpvt[k];
jpvt[k] = itemp;
vn1[pvt] = vn1[k];
vn2[pvt] = vn2[k];
if (pvt < nb) {
/* no need of transfer if pivot is within the panel */
blasf77_sswap( &m, A(0, pvt), &ione, A(0, k), &ione );
}
else {
/* 1. Finish copy from GPU */
magma_queue_sync( stream );
/* 2. Swap as usual on CPU */
blasf77_sswap(&m, A(0, pvt), &ione, A(0, k), &ione);
/* 3. Restore the GPU */
magma_ssetmatrix_async( m - offset - nb, 1,
A (offset + nb, pvt), lda,
dA(offset + nb, pvt), ldda, stream);
}
}
/* Apply previous Householder reflectors to column K:
A(RK:M,K) := A(RK:M,K) - A(RK:M,1:K-1)*F(K,1:K-1)'.
Optimization: multiply with beta=0; wait for vector and subtract */
if (k > 0) {
#if defined(PRECISION_c) || defined(PRECISION_z)
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_S_CNJG( *F(k,j) );
}
#endif
i__1 = m - rk;
i__2 = k;
blasf77_sgemv( MagmaNoTransStr, &i__1, &i__2,
&c_neg_one, A(rk, 0), &lda,
F(k, 0), &ldf,
&c_one, A(rk, k), &ione );
#if defined(PRECISION_c) || defined(PRECISION_z)
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_S_CNJG( *F(k,j) );
}
#endif
}
/* Generate elementary reflector H(k). */
if (rk < m-1) {
i__1 = m - rk;
lapackf77_slarfg( &i__1, A(rk, k), A(rk + 1, k), &ione, &tau[k] );
} else {
lapackf77_slarfg( &ione, A(rk, k), A(rk, k), &ione, &tau[k] );
}
Akk = *A(rk, k);
*A(rk, k) = c_one;
/* Compute Kth column of F:
Compute F(K+1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) on the GPU */
if (k < n-1) {
i__1 = m - rk;
i__2 = n - k - 1;
/* Send the vector to the GPU */
magma_ssetmatrix( i__1, 1, A(rk, k), lda, dA(rk,k), ldda );
/* Multiply on GPU */
// was CALL SGEMV( 'Conjugate transpose', M-RK+1, N-K,
// TAU( K ), A( RK, K+1 ), LDA,
// A( RK, K ), 1,
// CZERO, F( K+1, K ), 1 )
magma_int_t i__3 = nb-k-1;
magma_int_t i__4 = i__2 - i__3;
magma_int_t i__5 = nb-k;
magma_sgemv( MagmaConjTrans, i__1 - i__5, i__2 - i__3,
tau[k], dA(rk +i__5, k+1+i__3), ldda,
dA(rk +i__5, k ), ione,
c_zero, dF(k+1+i__3, k ), ione );
magma_sgetmatrix_async( i__2-i__3, 1,
dF(k + 1 +i__3, k), i__2,
F (k + 1 +i__3, k), i__2, stream );
blasf77_sgemv( MagmaConjTransStr, &i__1, &i__3,
&tau[k], A(rk, k+1), &lda,
A(rk, k ), &ione,
&c_zero, F(k+1, k ), &ione );
magma_queue_sync( stream );
blasf77_sgemv( MagmaConjTransStr, &i__5, &i__4,
&tau[k], A(rk, k+1+i__3), &lda,
A(rk, k ), &ione,
&c_one, F(k+1+i__3, k ), &ione );
}
/* Padding F(1:K,K) with zeros. */
for (j = 0; j < k; ++j) {
*F(j, k) = c_zero;
}
/* Incremental updating of F:
F(1:N,K) := F(1:N,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K). */
if (k > 0) {
i__1 = m - rk;
i__2 = k;
z__1 = MAGMA_S_NEGATE( tau[k] );
blasf77_sgemv( MagmaConjTransStr, &i__1, &i__2,
&z__1, A(rk, 0), &lda,
A(rk, k), &ione,
&c_zero, auxv, &ione );
i__1 = k;
blasf77_sgemv( MagmaNoTransStr, &n, &i__1,
&c_one, F(0,0), &ldf,
auxv, &ione,
&c_one, F(0,k), &ione );
}
/* Optimization: On the last iteration start sending F back to the GPU */
/* Update the current row of A:
A(RK,K+1:N) := A(RK,K+1:N) - A(RK,1:K)*F(K+1:N,1:K)'. */
if (k < n-1) {
i__1 = n - k - 1;
i__2 = k + 1;
blasf77_sgemm( MagmaNoTransStr, MagmaConjTransStr, &ione, &i__1, &i__2,
&c_neg_one, A(rk, 0 ), &lda,
F(k+1,0 ), &ldf,
&c_one, A(rk, k+1), &lda );
}
/* Update partial column norms. */
if (rk < lastrk) {
for (j = k + 1; j < n; ++j) {
if (vn1[j] != 0.) {
/* NOTE: The following 4 lines follow from the analysis in
Lapack Working Note 176. */
temp = MAGMA_S_ABS( *A(rk,j) ) / vn1[j];
temp = max( 0., ((1. + temp) * (1. - temp)) );
d__1 = vn1[j] / vn2[j];
temp2 = temp * (d__1 * d__1);
if (temp2 <= tol3z) {
vn2[j] = (float) lsticc;
lsticc = j;
} else {
vn1[j] *= magma_ssqrt(temp);
}
}
}
}
*A(rk, k) = Akk;
++k;
}
// leave k as the last column done
--k;
*kb = k + 1;
rk = offset + *kb - 1;
/* Apply the block reflector to the rest of the matrix:
A(OFFSET+KB+1:M,KB+1:N) := A(OFFSET+KB+1:M,KB+1:N) - A(OFFSET+KB+1:M,1:KB)*F(KB+1:N,1:KB)' */
if (*kb < min(n, m - offset)) {
i__1 = m - rk - 1;
i__2 = n - *kb;
/* Send F to the GPU */
magma_ssetmatrix( i__2, *kb,
F (*kb, 0), ldf,
dF(*kb, 0), i__2 );
magma_sgemm( MagmaNoTrans, MagmaConjTrans, i__1, i__2, *kb,
c_neg_one, dA(rk+1, 0 ), ldda,
dF(*kb, 0 ), i__2,
c_one, dA(rk+1, *kb), ldda );
}
/* Recomputation of difficult columns. */
while( lsticc > 0 ) {
itemp = (magma_int_t)(vn2[lsticc] >= 0. ? floor(vn2[lsticc] + .5) : -floor(.5 - vn2[lsticc]));
i__1 = m - rk - 1;
if (lsticc <= nb)
vn1[lsticc] = magma_cblas_snrm2( i__1, A(rk+1,lsticc), ione );
else {
/* Where is the data, CPU or GPU ? */
float r1, r2;
r1 = magma_cblas_snrm2( nb-k, A(rk+1,lsticc), ione );
r2 = magma_snrm2(m-offset-nb, dA(offset + nb + 1, lsticc), ione);
//vn1[lsticc] = magma_snrm2(i__1, dA(rk + 1, lsticc), ione);
vn1[lsticc] = magma_ssqrt(r1*r1 + r2*r2);
}
/* NOTE: The computation of VN1( LSTICC ) relies on the fact that
SNRM2 does not fail on vectors with norm below the value of SQRT(SLAMCH('S')) */
vn2[lsticc] = vn1[lsticc];
lsticc = itemp;
}
magma_queue_destroy( stream );
return MAGMA_SUCCESS;
} /* magma_slaqps */
extern "C" magma_int_t
magma_slaex3(magma_int_t k, magma_int_t n, magma_int_t n1, float* d,
float* q, magma_int_t ldq, float rho,
float* dlamda, float* q2, magma_int_t* indx,
magma_int_t* ctot, float* w, float* s, magma_int_t* indxq,
float* dwork,
char range, float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t* info )
{
/*
Purpose
=======
SLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to SLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code makes very mild assumptions about floating point
arithmetic. It will work on machines with a guard digit in
add/subtract, or on those binary machines without guard digits
which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
=========
K (input) INTEGER
The number of terms in the rational function to be solved by
SLAED4. K >= 0.
N (input) INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N>K).
N1 (input) INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
D (output) REAL array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
Q (output) REAL array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
LDQ (input) INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
RHO (input) REAL
The value of the parameter in the rank one update equation.
RHO >= 0 required.
DLAMDA (input/output) REAL array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
Q2 (input) REAL array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
INDX (input) INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see SLAED2).
The rows of the eigenvectors found by SLAED4 must be likewise
permuted before the matrix multiply can take place.
CTOT (input) INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
W (input/output) REAL array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
S (workspace) REAL array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
INDXQ (output) INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
DWORK (device workspace) REAL array, dimension (3*N*N/2+3*N)
INFO (output) INTEGER
= 0: successful exit.
< 0: if INFO = -i, the i-th argument had an illegal value.
> 0: if INFO = 1, an eigenvalue did not converge
Further Details
===============
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
===================================================================== */
float d_one = 1.;
float d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
char range_[] = {range, 0};
magma_int_t iil, iiu, rk;
float* dq2= dwork;
float* ds = dq2 + n*(n/2+1);
float* dq = ds + n*(n/2+1);
magma_int_t lddq = n/2 + 1;
magma_int_t i,iq2,j,n12,n2,n23,tmp,lq2;
float temp;
magma_int_t alleig, valeig, indeig;
alleig = lapackf77_lsame(range_, "A");
valeig = lapackf77_lsame(range_, "V");
indeig = lapackf77_lsame(range_, "I");
*info = 0;
if(k < 0)
*info=-1;
else if(n < k)
*info=-2;
else if(ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if(*info != 0){
magma_xerbla(__func__, -(*info));
return MAGMA_ERR_ILLEGAL_VALUE;
}
// Quick return if possible
if(k == 0)
return MAGMA_SUCCESS;
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
lq2 = iq2 + n2 * n23;
magma_ssetvector_async( lq2, q2, 1, dq2, 1, NULL );
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
magma_timestr_t start, end;
start = get_current_time();
#endif
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for(i = ib; i < ie; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for(j = ib; j < ie; ++j){
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if(iinfo != 0){
#pragma omp critical (info)
*info=iinfo;
break;
}
}
#pragma omp barrier
if(*info == 0){
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione , &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2){
#pragma omp single
{
for(j = 0; j < k; ++j){
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if(k != 1){
// Compute updated W.
blasf77_scopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for(j = 0; j < k; ++j){
magma_int_t i_tmp = min(j, ie);
for(i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for(i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for(i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if(id < tot){
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else{
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for(j = ib; j < ie; ++j){
for(i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = cblas_snrm2( k, s+id*k, 1);
for(i = 0; i < k; ++i){
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return MAGMA_SUCCESS; //??????
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
end = get_current_time();
printf("eigenvalues/vector D+zzT = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
magma_timestr_t start, end;
start = get_current_time();
#endif
for(i = 0; i < k; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for(j = 0; j < k; ++j){
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if(iinfo != 0)
*info=iinfo;
}
if(*info != 0)
return MAGMA_SUCCESS;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione , &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2){
for(j = 0; j < k; ++j){
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if(k != 1){
// Compute updated W.
blasf77_scopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &k, q, &tmp, w, &ione);
for(j = 0; j < k; ++j){
for(i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for(i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for(i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for(j = iil-1; j < iiu; ++j){
for(i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = cblas_snrm2( k, s, 1);
for(i = 0; i < k; ++i){
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
end = get_current_time();
printf("eigenvalues/vector D+zzT = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
#endif //_OPENMP
// Compute the updated eigenvectors.
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
start = get_current_time();
#endif
magma_queue_sync( NULL );
if (rk != 0){
if( n23 != 0 ){
if (rk < magma_get_slaed3_k()){
lapackf77_slacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_sgemm("N", "N", &n2, &rk, &n23, &d_one, &q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
} else {
magma_ssetmatrix( n23, rk, Q(ctot[0],iil-1), ldq, ds, n23 );
magma_sgemm('N', 'N', n2, rk, n23, d_one, &dq2[iq2], n2, ds, n23, d_zero, dq, lddq);
magma_sgetmatrix( n2, rk, dq, lddq, Q(n1,iil-1), ldq );
}
} else
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if( n12 != 0 ) {
if (rk < magma_get_slaed3_k()){
lapackf77_slacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_sgemm("N", "N", &n1, &rk, &n12, &d_one, q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
} else {
magma_ssetmatrix( n12, rk, Q(0,iil-1), ldq, ds, n12 );
magma_sgemm('N', 'N', n1, rk, n12, d_one, dq2, n1, ds, n12, d_zero, dq, lddq);
magma_sgetmatrix( n1, rk, dq, lddq, Q(0,iil-1), ldq );
}
} else
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
#ifdef ENABLE_TIMER_DIVIDE_AND_CONQUER
end = get_current_time();
printf("gemms = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
return MAGMA_SUCCESS;
} /*magma_slaed3*/
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgels
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float gpu_error, cpu_error, error, Anorm, work[1];
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *h_work, tmp[1];
float *d_A, *d_B;
magma_int_t M, N, size, nrhs, lda, ldb, ldda, lddb, min_mn, max_mn, nb, info;
magma_int_t lworkgpu, lhwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t status = 0;
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf(" ||b-Ax|| / (N||A||) ||dx-x||/(N||A||)\n");
printf(" M N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) CPU GPU \n");
printf("===================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if ( M < N ) {
printf( "%5d %5d %5d skipping because M < N is not yet supported.\n", (int) M, (int) N, (int) nrhs );
continue;
}
min_mn = min(M, N);
max_mn = max(M, N);
lda = M;
ldb = max_mn;
ldda = ((M+31)/32)*32;
lddb = ((max_mn+31)/32)*32;
nb = magma_get_sgeqrf_nb(M);
gflops = (FLOPS_SGEQRF( M, N ) + FLOPS_SGEQRS( M, N, nrhs )) / 1e9;
lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb;
// query for workspace size
lhwork = -1;
lapackf77_sgels( MagmaNoTransStr, &M, &N, &nrhs,
NULL, &lda, NULL, &ldb, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_S_REAL( tmp[0] );
lhwork = max( lhwork, lworkgpu );
TESTING_MALLOC_CPU( tau, float, min_mn );
TESTING_MALLOC_CPU( h_A, float, lda*N );
TESTING_MALLOC_CPU( h_A2, float, lda*N );
TESTING_MALLOC_CPU( h_B, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_R, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_work, float, lhwork );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( d_B, float, lddb*nrhs );
/* Initialize the matrices */
size = lda*N;
lapackf77_slarnv( &ione, ISEED, &size, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda );
// make random RHS
size = ldb*nrhs;
lapackf77_slarnv( &ione, ISEED, &size, h_B );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
// make consistent RHS
//size = N*nrhs;
//lapackf77_slarnv( &ione, ISEED, &size, h_X );
//blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
// &c_one, h_A, &lda,
// h_X, &ldb,
// &c_zero, h_B, &ldb );
//lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_ssetmatrix( M, N, h_A, lda, d_A, ldda );
magma_ssetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
gpu_time = magma_wtime();
magma_sgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, ldda,
d_B, lddb, h_work, lworkgpu, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sgels_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute the residual
magma_sgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb );
Anorm = lapackf77_slange("f", &M, &N, h_A, &lda, work);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_sgels( MagmaNoTransStr, &M, &N, &nrhs,
h_A, &lda, h_X, &ldb, h_work, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sgels returned error %d: %s.\n",
(int) info, magma_strerror( info ));
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb );
cpu_error = lapackf77_slange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm);
gpu_error = lapackf77_slange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
// error relative to LAPACK
size = M*nrhs;
blasf77_saxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error = lapackf77_slange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e",
(int) M, (int) N, (int) nrhs,
cpu_perf, cpu_time, gpu_perf, gpu_time, cpu_error, gpu_error, error );
if ( M == N ) {
printf( " %s\n", (gpu_error < tol && error < tol ? "ok" : "failed"));
status += ! (gpu_error < tol && error < tol);
}
else {
printf( " %s\n", (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_A2 );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_R );
TESTING_FREE_CPU( h_work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgesv_gpu
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
float error, Rnorm, Anorm, Xnorm, *work;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_B, *h_X;
magmaFloat_ptr d_A, d_B;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf(" N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
ldda = ((N+31)/32)*32;
lddb = ldda;
gflops = ( FLOPS_SGETRF( N, N ) + FLOPS_SGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, float, lda*N );
TESTING_MALLOC_CPU( h_B, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, float, ldb*nrhs );
TESTING_MALLOC_CPU( work, float, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( d_B, float, lddb*nrhs );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &sizeB, h_B );
magma_ssetmatrix( N, N, h_A, lda, d_A, ldda );
magma_ssetmatrix( N, nrhs, h_B, ldb, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgesv_gpu( N, nrhs, d_A, ldda, ipiv, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_sgesv_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
magma_sgetmatrix( N, nrhs, d_B, lddb, h_X, ldb );
Anorm = lapackf77_slange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_slange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_slange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgesv( &N, &nrhs, h_A, &lda, ipiv, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_sgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
extern "C" magma_int_t
magma_slaqps(magma_int_t m, magma_int_t n, magma_int_t offset,
magma_int_t nb, magma_int_t *kb,
float *A, magma_int_t lda,
float *dA, magma_int_t ldda,
magma_int_t *jpvt, float *tau, float *vn1, float *vn2,
float *auxv,
float *F, magma_int_t ldf,
float *dF, magma_int_t lddf)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
SLAQPS computes a step of QR factorization with column pivoting
of a real M-by-N matrix A by using Blas-3. It tries to factorize
NB columns from A starting from the row OFFSET+1, and updates all
of the matrix with Blas-3 xGEMM.
In some cases, due to catastrophic cancellations, it cannot
factorize NB columns. Hence, the actual number of factorized
columns is returned in KB.
Block A(1:OFFSET,1:N) is accordingly pivoted, but not factorized.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0
OFFSET (input) INTEGER
The number of rows of A that have been factorized in
previous steps.
NB (input) INTEGER
The number of columns to factorize.
KB (output) INTEGER
The number of columns actually factorized.
A (input/output) REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, block A(OFFSET+1:M,1:KB) is the triangular
factor obtained and block A(1:OFFSET,1:N) has been
accordingly pivoted, but no factorized.
The rest of the matrix, block A(OFFSET+1:M,KB+1:N) has
been updated.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,M).
JPVT (input/output) INTEGER array, dimension (N)
JPVT(I) = K <==> Column K of the full matrix A has been
permuted into position I in AP.
TAU (output) REAL array, dimension (KB)
The scalar factors of the elementary reflectors.
VN1 (input/output) DOUBLE PRECISION array, dimension (N)
The vector with the partial column norms.
VN2 (input/output) DOUBLE PRECISION array, dimension (N)
The vector with the exact column norms.
AUXV (input/output) REAL array, dimension (NB)
Auxiliar vector.
F (input/output) REAL array, dimension (LDF,NB)
Matrix F' = L*Y'*A.
LDF (input) INTEGER
The leading dimension of the array F. LDF >= max(1,N).
===================================================================== */
#define A(i, j) (A + (i) + (j)*(lda ))
#define dA(i, j) (dA + (i) + (j)*(ldda))
#define F(i, j) (F + (i) + (j)*(ldf ))
#define dF(i, j) (dF + (i) + (j)*(lddf))
float c_zero = MAGMA_S_MAKE( 0.,0.);
float c_one = MAGMA_S_MAKE( 1.,0.);
float c_neg_one = MAGMA_S_MAKE(-1.,0.);
magma_int_t ione = 1;
magma_int_t i__1, i__2;
float d__1;
float z__1;
magma_int_t j, k, rk;
float Akk;
magma_int_t pvt;
float temp, temp2, tol3z;
magma_int_t itemp;
magma_int_t lsticc;
magma_int_t lastrk;
lastrk = min( m, n + offset );
tol3z = magma_ssqrt( lapackf77_slamch("Epsilon"));
magma_queue_t stream;
magma_queue_create( &stream );
lsticc = 0;
k = 0;
while( k < nb && lsticc == 0 ) {
rk = offset + k;
/* Determine ith pivot column and swap if necessary */
// Fortran: pvt, k, isamax are all 1-based; subtract 1 from k.
// C: pvt, k, isamax are all 0-based; don't subtract 1.
pvt = k + cblas_isamax( n-k, &vn1[k], ione );
if (pvt != k) {
if (pvt >= nb) {
/* 1. Start copy from GPU */
magma_sgetmatrix_async( m - offset - nb, 1,
dA(offset + nb, pvt), ldda,
A (offset + nb, pvt), lda, stream );
}
/* F gets swapped so F must be sent at the end to GPU */
i__1 = k;
blasf77_sswap( &i__1, F(pvt,0), &ldf, F(k,0), &ldf );
itemp = jpvt[pvt];
jpvt[pvt] = jpvt[k];
jpvt[k] = itemp;
vn1[pvt] = vn1[k];
vn2[pvt] = vn2[k];
if (pvt < nb){
/* no need of transfer if pivot is within the panel */
blasf77_sswap( &m, A(0, pvt), &ione, A(0, k), &ione );
}
else {
/* 1. Finish copy from GPU */
magma_queue_sync( stream );
/* 2. Swap as usual on CPU */
blasf77_sswap(&m, A(0, pvt), &ione, A(0, k), &ione);
/* 3. Restore the GPU */
magma_ssetmatrix_async( m - offset - nb, 1,
A (offset + nb, pvt), lda,
dA(offset + nb, pvt), ldda, stream);
}
}
/* Apply previous Householder reflectors to column K:
A(RK:M,K) := A(RK:M,K) - A(RK:M,1:K-1)*F(K,1:K-1)'.
Optimization: multiply with beta=0; wait for vector and subtract */
if (k > 0) {
#if defined(PRECISION_c) || defined(PRECISION_z)
for (j = 0; j < k; ++j){
*F(k,j) = MAGMA_S_CNJG( *F(k,j) );
}
#endif
i__1 = m - rk;
i__2 = k;
blasf77_sgemv( MagmaNoTransStr, &i__1, &i__2,
&c_neg_one, A(rk, 0), &lda,
F(k, 0), &ldf,
&c_one, A(rk, k), &ione );
#if defined(PRECISION_c) || defined(PRECISION_z)
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_S_CNJG( *F(k,j) );
}
#endif
}
/* Generate elementary reflector H(k). */
if (rk < m-1) {
i__1 = m - rk;
lapackf77_slarfg( &i__1, A(rk, k), A(rk + 1, k), &ione, &tau[k] );
} else {
lapackf77_slarfg( &ione, A(rk, k), A(rk, k), &ione, &tau[k] );
}
Akk = *A(rk, k);
*A(rk, k) = c_one;
/* Compute Kth column of F:
Compute F(K+1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) on the GPU */
if (k < n-1) {
i__1 = m - rk;
i__2 = n - k - 1;
/* Send the vector to the GPU */
magma_ssetmatrix( i__1, 1, A(rk, k), lda, dA(rk,k), ldda );
/* Multiply on GPU */
// was CALL SGEMV( 'Conjugate transpose', M-RK+1, N-K,
// TAU( K ), A( RK, K+1 ), LDA,
// A( RK, K ), 1,
// CZERO, F( K+1, K ), 1 )
magma_int_t i__3 = nb-k-1;
magma_int_t i__4 = i__2 - i__3;
magma_int_t i__5 = nb-k;
magma_sgemv( MagmaTrans, i__1 - i__5, i__2 - i__3,
tau[k], dA(rk +i__5, k+1+i__3), ldda,
dA(rk +i__5, k ), ione,
c_zero, dF(k+1+i__3, k ), ione );
magma_sgetmatrix_async( i__2-i__3, 1,
dF(k + 1 +i__3, k), i__2,
F (k + 1 +i__3, k), i__2, stream );
blasf77_sgemv( MagmaTransStr, &i__1, &i__3,
&tau[k], A(rk, k+1), &lda,
A(rk, k ), &ione,
&c_zero, F(k+1, k ), &ione );
magma_queue_sync( stream );
blasf77_sgemv( MagmaTransStr, &i__5, &i__4,
&tau[k], A(rk, k+1+i__3), &lda,
A(rk, k ), &ione,
&c_one, F(k+1+i__3, k ), &ione );
}
/* Padding F(1:K,K) with zeros. */
for (j = 0; j < k; ++j) {
*F(j, k) = c_zero;
}
/* Incremental updating of F:
F(1:N,K) := F(1:N,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K). */
if (k > 0) {
i__1 = m - rk;
i__2 = k;
z__1 = MAGMA_S_NEGATE( tau[k] );
blasf77_sgemv( MagmaTransStr, &i__1, &i__2,
&z__1, A(rk, 0), &lda,
A(rk, k), &ione,
&c_zero, auxv, &ione );
i__1 = k;
blasf77_sgemv( MagmaNoTransStr, &n, &i__1,
&c_one, F(0,0), &ldf,
auxv, &ione,
&c_one, F(0,k), &ione );
}
/* Optimization: On the last iteration start sending F back to the GPU */
/* Update the current row of A:
A(RK,K+1:N) := A(RK,K+1:N) - A(RK,1:K)*F(K+1:N,1:K)'. */
if (k < n-1) {
i__1 = n - k - 1;
i__2 = k + 1;
blasf77_sgemm( MagmaNoTransStr, MagmaTransStr, &ione, &i__1, &i__2,
&c_neg_one, A(rk, 0 ), &lda,
F(k+1,0 ), &ldf,
&c_one, A(rk, k+1), &lda );
}
/* Update partial column norms. */
if (rk < lastrk) {
for (j = k + 1; j < n; ++j) {
if (vn1[j] != 0.) {
/* NOTE: The following 4 lines follow from the analysis in
Lapack Working Note 176. */
temp = MAGMA_S_ABS( *A(rk,j) ) / vn1[j];
temp = max( 0., ((1. + temp) * (1. - temp)) );
d__1 = vn1[j] / vn2[j];
temp2 = temp * (d__1 * d__1);
if (temp2 <= tol3z) {
vn2[j] = (float) lsticc;
lsticc = j;
} else {
vn1[j] *= magma_ssqrt(temp);
}
}
}
}
*A(rk, k) = Akk;
++k;
}
// leave k as the last column done
--k;
*kb = k + 1;
rk = offset + *kb - 1;
/* Apply the block reflector to the rest of the matrix:
A(OFFSET+KB+1:M,KB+1:N) := A(OFFSET+KB+1:M,KB+1:N) - A(OFFSET+KB+1:M,1:KB)*F(KB+1:N,1:KB)' */
if (*kb < min(n, m - offset)) {
i__1 = m - rk - 1;
i__2 = n - *kb;
/* Send F to the GPU */
magma_ssetmatrix( i__2, *kb,
F (*kb, 0), ldf,
dF(*kb, 0), i__2 );
magma_sgemm( MagmaNoTrans, MagmaTrans, i__1, i__2, *kb,
c_neg_one, dA(rk+1, 0 ), ldda,
dF(*kb, 0 ), i__2,
c_one, dA(rk+1, *kb), ldda );
}
/* Recomputation of difficult columns. */
while( lsticc > 0 ) {
itemp = (magma_int_t)(vn2[lsticc] >= 0. ? floor(vn2[lsticc] + .5) : -floor(.5 - vn2[lsticc]));
i__1 = m - rk - 1;
if (lsticc <= nb)
vn1[lsticc] = cblas_snrm2(i__1, A(rk + 1, lsticc), ione);
else {
/* Where is the data, CPU or GPU ? */
float r1, r2;
r1 = cblas_snrm2(nb-k, A(rk + 1, lsticc), ione);
r2 = magma_snrm2(m-offset-nb, dA(offset + nb + 1, lsticc), ione);
//vn1[lsticc] = magma_snrm2(i__1, dA(rk + 1, lsticc), ione);
vn1[lsticc] = magma_ssqrt(r1*r1+r2*r2);
}
/* NOTE: The computation of VN1( LSTICC ) relies on the fact that
SNRM2 does not fail on vectors with norm below the value of SQRT(SLAMCH('S')) */
vn2[lsticc] = vn1[lsticc];
lsticc = itemp;
}
magma_queue_destroy( stream );
return MAGMA_SUCCESS;
} /* magma_slaqps */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeqrs_gpu
*/
int main( int argc, char** argv)
{
//#if defined(PRECISION_s)
/* Initialize */
magma_queue_t queue;
magma_device_t device[ MagmaMaxGPUs ];
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device[0], &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float matnorm, work[1];
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_A2, *h_B, *h_X, *h_R, *tau, *hwork, tmp[1];
magmaFloat_ptr d_A, d_B;
/* Matrix size */
magma_int_t M = 0, N = 0, n2;
magma_int_t lda, ldb, ldda, lddb, lworkgpu, lhwork;
magma_int_t size[7] = {1024,2048,3072,4032,5184,6016,7000};
magma_int_t i, info, min_mn, nb, l1, l2;
magma_int_t ione = 1;
magma_int_t nrhs = 3;
magma_int_t ISEED[4] = {0,0,0,1};
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
else if (strcmp("-M", argv[i])==0)
M = atoi(argv[++i]);
else if (strcmp("-nrhs", argv[i])==0)
nrhs = atoi(argv[++i]);
}
if (N>0 && M>0 && M >= N)
printf(" testing_sgeqrs_gpu -nrhs %d -M %d -N %d\n\n", nrhs, M, N);
else
{
printf("\nUsage: \n");
printf(" testing_sgeqrs_gpu -nrhs %d -M %d -N %d\n\n", nrhs, M, N);
printf(" M has to be >= N, exit.\n");
exit(1);
}
}
else {
printf("\nUsage: \n");
printf(" testing_sgeqrs_gpu -nrhs %d -M %d -N %d\n\n", nrhs, 1024, 1024);
M = N = size[6];
}
ldda = ((M+31)/32)*32;
lddb = ldda;
n2 = M * N;
min_mn = min(M, N);
nb = magma_get_sgeqrf_nb(M);
lda = ldb = M;
lworkgpu = (M-N + nb)*(nrhs+2*nb);
/* Allocate host memory for the matrix */
TESTING_MALLOC_PIN( tau, float, min_mn );
TESTING_MALLOC_PIN( h_A, float, lda*N );
TESTING_MALLOC_PIN( h_A2, float, lda*N );
TESTING_MALLOC_PIN( h_B, float, ldb*nrhs );
TESTING_MALLOC_PIN( h_X, float, ldb*nrhs );
TESTING_MALLOC_PIN( h_R, float, ldb*nrhs );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( d_B, float, lddb*nrhs );
/*
* Get size for host workspace
*/
lhwork = -1;
lapackf77_sgeqrf(&M, &N, h_A, &M, tau, tmp, &lhwork, &info);
l1 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
lhwork = -1;
lapackf77_sormqr( MagmaLeftStr, MagmaTransStr,
&M, &nrhs, &min_mn, h_A, &lda, tau,
h_X, &ldb, tmp, &lhwork, &info);
l2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
lhwork = max( max( l1, l2 ), lworkgpu );
TESTING_MALLOC_PIN( hwork, float, lhwork );
printf("\n");
printf(" ||b-Ax|| / (N||A||)\n");
printf(" M N CPU GFlop/s GPU GFlop/s CPU GPU \n");
printf("============================================================\n");
for(i=0; i<7; i++){
if (argc == 1){
M = N = size[i];
}
min_mn= min(M, N);
ldb = lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = (FLOPS_GEQRF( (float)M, (float)N )
+ FLOPS_GEQRS( (float)M, (float)N, (float)nrhs )) / 1e9;
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda );
n2 = M*nrhs;
lapackf77_slarnv( &ione, ISEED, &n2, h_B );
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* Warm up to measure the performance */
magma_ssetmatrix( M, N, h_A, 0, lda, d_A, 0, ldda, queue );
magma_ssetmatrix( M, nrhs, h_B, 0, ldb, d_B, 0, lddb, queue );
magma_sgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, 0, ldda,
d_B, 0, lddb, hwork, lworkgpu, &info, queue);
magma_ssetmatrix( M, N, h_A, 0, lda, d_A, 0, ldda, queue );
magma_ssetmatrix( M, nrhs, h_B, 0, ldb, d_B, 0, lddb, queue );
gpu_time = magma_wtime();
magma_sgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, 0, ldda,
d_B, 0, lddb, hwork, lworkgpu, &info, queue);
gpu_time = magma_wtime() - gpu_time;
if (info < 0)
printf("Argument %d of magma_sgels had an illegal value.\n", -info);
gpu_perf = gflops / gpu_time;
// Get the solution in h_X
magma_sgetmatrix( N, nrhs, d_B, 0, lddb, h_X, 0, ldb, queue );
// compute the residual
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb);
matnorm = lapackf77_slange("f", &M, &N, h_A, &lda, work);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_slacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_sgels( MagmaNoTransStr, &M, &N, &nrhs,
h_A, &lda, h_X, &ldb, hwork, &lhwork, &info);
cpu_time = magma_wtime()-cpu_time;
cpu_perf = gflops / cpu_time;
if (info < 0)
printf("Argument %d of lapackf77_sgels had an illegal value.\n", -info);
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb);
printf("%5d %5d %6.1f %6.1f %7.2e %7.2e\n",
M, N, cpu_perf, gpu_perf,
lapackf77_slange("f", &M, &nrhs, h_B, &M, work)/(min_mn*matnorm),
lapackf77_slange("f", &M, &nrhs, h_R, &M, work)/(min_mn*matnorm) );
if (argc != 1)
break;
}
/* Memory clean up */
TESTING_FREE_PIN( tau );
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_A2 );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_X );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( hwork );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
/* Shutdown */
magma_queue_destroy( queue );
magma_finalize();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgesv
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
float error, lerror, Rnorm, Anorm, Xnorm, *work;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *h_A, *h_LU, *h_B, *h_B0, *h_X;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
nrhs = opts.nrhs;
printf("%% ngpu %d\n", (int) opts.ngpu );
if (opts.lapack) {
printf("%% N NRHS CPU Gflop/s (sec) GPU Gflop/s (sec) ||B - AX|| / N*||A||*||X|| ||B - AX|| / N*||A||*||X||_CPU\n");
printf("%%================================================================================================================\n");
} else {
printf("%% N NRHS CPU Gflop/s (sec) GPU Gflop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("%%===============================================================================\n");
}
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
gflops = ( FLOPS_SGETRF( N, N ) + FLOPS_SGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, float, lda*N );
TESTING_MALLOC_CPU( h_LU, float, lda*N );
TESTING_MALLOC_CPU( h_B0, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_B, float, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, float, ldb*nrhs );
TESTING_MALLOC_CPU( work, float, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &sizeB, h_B );
// copy A to LU and B to X; save A and B for residual
lapackf77_slacpy( "F", &N, &N, h_A, &lda, h_LU, &lda );
lapackf77_slacpy( "F", &N, &nrhs, h_B, &ldb, h_X, &ldb );
lapackf77_slacpy( "F", &N, &nrhs, h_B, &ldb, h_B0, &ldb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgesv( N, nrhs, h_LU, lda, ipiv, h_X, ldb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Residual
//=====================================================================
Anorm = lapackf77_slange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_slange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_slange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
bool okay = (error < tol);
status += ! okay;
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
lapackf77_slacpy( "F", &N, &N, h_A, &lda, h_LU, &lda );
lapackf77_slacpy( "F", &N, &nrhs, h_B0, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_sgesv( &N, &nrhs, h_LU, &lda, ipiv, h_X, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//Anorm = lapackf77_slange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_slange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_sgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B0, &ldb);
Rnorm = lapackf77_slange("I", &N, &nrhs, h_B0, &ldb, work);
lerror = Rnorm/(N*Anorm*Xnorm);
bool lokay = (lerror < tol);
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %-6s %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (okay ? "ok" : "failed"),
lerror, (lokay ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (okay ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_LU );
TESTING_FREE_CPU( h_B0 );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgetri
*/
int main( int argc, char** argv )
{
TESTING_INIT();
// constants
const float c_zero = MAGMA_S_ZERO;
const float c_one = MAGMA_S_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float *h_A, *h_Ainv, *h_R, *work;
magmaFloat_ptr d_A, dwork;
magma_int_t N, n2, lda, ldda, info, lwork, ldwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float tmp;
float error, rwork[1];
magma_int_t *ipiv;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% N CPU Gflop/s (sec) GPU Gflop/s (sec) ||I - A*A^{-1}||_1 / (N*cond(A))\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
ldwork = N * magma_get_sgetri_nb( N );
gflops = FLOPS_SGETRI( N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0) {
printf("lapackf77_sgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = magma_int_t( MAGMA_S_REAL( tmp ));
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( work, float, lwork );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_CPU( h_Ainv, float, n2 );
TESTING_MALLOC_CPU( h_R, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
TESTING_MALLOC_DEV( dwork, float, ldwork );
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
/* Factor the matrix. Both MAGMA and LAPACK will use this factor. */
magma_ssetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
magma_sgetrf_gpu( N, N, d_A, ldda, ipiv, &info );
magma_sgetmatrix( N, N, d_A, ldda, h_Ainv, lda, opts.queue );
if (info != 0) {
printf("magma_sgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
// check for exact singularity
//h_Ainv[ 10 + 10*lda ] = MAGMA_S_MAKE( 0.0, 0.0 );
//magma_ssetmatrix( N, N, h_Ainv, lda, d_A, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_sgetri_gpu( N, d_A, ldda, ipiv, dwork, ldwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgetri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgetri( &N, h_Ainv, &lda, ipiv, work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf( "%5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf( "%5d --- ( --- ) %7.2f (%7.2f)",
(int) N, gpu_perf, gpu_time );
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
magma_sgetmatrix( N, N, d_A, ldda, h_Ainv, lda, opts.queue );
// compute 1-norm condition number estimate, following LAPACK's zget03
float normA, normAinv, rcond;
normA = lapackf77_slange( "1", &N, &N, h_A, &lda, rwork );
normAinv = lapackf77_slange( "1", &N, &N, h_Ainv, &lda, rwork );
if ( normA <= 0 || normAinv <= 0 ) {
rcond = 0;
error = 1 / (tol/opts.tolerance); // == 1/eps
}
else {
rcond = (1 / normA) / normAinv;
// R = I
// R -= A*A^{-1}
// err = ||I - A*A^{-1}|| / ( N ||A||*||A^{-1}|| ) = ||R|| * rcond / N, using 1-norm
lapackf77_slaset( "full", &N, &N, &c_zero, &c_one, h_R, &lda );
blasf77_sgemm( "no", "no", &N, &N, &N,
&c_neg_one, h_A, &lda,
h_Ainv, &lda,
&c_one, h_R, &lda );
error = lapackf77_slange( "1", &N, &N, h_R, &lda, rwork );
error = error * rcond / N;
}
bool okay = (error < tol);
status += ! okay;
printf( " %8.2e %s\n",
error, (okay ? "ok" : "failed"));
}
else {
printf( "\n" );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_Ainv );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgeqrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
const float d_neg_one = MAGMA_D_NEG_ONE;
const float d_one = MAGMA_D_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
const float c_one = MAGMA_S_ONE;
const float c_zero = MAGMA_S_ZERO;
const magma_int_t ione = 1;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float Anorm, error=0, error2=0;
float *h_A, *h_R, *tau, *h_work, tmp[1];
magmaFloat_ptr d_A, dT;
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb, size;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
opts.parse_opts( argc, argv );
magma_int_t status = 0;
float tol = opts.tolerance * lapackf77_slamch("E");
// version 3 can do either check
if (opts.check == 1 && opts.version == 1) {
opts.check = 2;
printf( "%% version 1 requires check 2 (solve A*x=b)\n" );
}
if (opts.check == 2 && opts.version == 2) {
opts.check = 1;
printf( "%% version 2 requires check 1 (R - Q^H*A)\n" );
}
printf( "%% version %d\n", (int) opts.version );
if ( opts.check == 1 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |R - Q^H*A| |I - Q^H*Q|\n");
printf("%%==============================================================================\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |b - A*x|\n");
printf("%%===============================================================\n");
}
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min( M, N );
lda = M;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
nb = magma_get_sgeqrf_nb( M, N );
gflops = FLOPS_SGEQRF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_sgeqrf( &M, &N, NULL, &M, NULL, tmp, &lwork, &info );
lwork = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( tau, float, min_mn );
TESTING_MALLOC_CPU( h_A, float, n2 );
TESTING_MALLOC_CPU( h_work, float, lwork );
TESTING_MALLOC_PIN( h_R, float, n2 );
TESTING_MALLOC_DEV( d_A, float, ldda*N );
if ( opts.version == 1 || opts.version == 3 ) {
size = (2*min(M, N) + magma_roundup( N, 32 ) )*nb;
TESTING_MALLOC_DEV( dT, float, size );
magmablas_slaset( MagmaFull, size, 1, c_zero, c_zero, dT, size );
}
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
magma_ssetmatrix( M, N, h_R, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
nb = magma_get_sgeqrf_nb( M, N );
gpu_time = magma_wtime();
if ( opts.version == 1 ) {
// stores dT, V blocks have zeros, R blocks inverted & stored in dT
magma_sgeqrf_gpu( M, N, d_A, ldda, tau, dT, &info );
}
else if ( opts.version == 2 ) {
// LAPACK complaint arguments
magma_sgeqrf2_gpu( M, N, d_A, ldda, tau, &info );
}
#ifdef HAVE_CUBLAS
else if ( opts.version == 3 ) {
// stores dT, V blocks have zeros, R blocks stored in dT
magma_sgeqrf3_gpu( M, N, d_A, ldda, tau, dT, &info );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
return -1;
}
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
if ( opts.check == 1 && (opts.version == 2 || opts.version == 3) ) {
if ( opts.version == 3 ) {
// copy diagonal blocks of R back to A
for( int i=0; i < min_mn-nb; i += nb ) {
magma_int_t ib = min( min_mn-i, nb );
magmablas_slacpy( MagmaUpper, ib, ib, &dT[min_mn*nb + i*nb], nb, &d_A[ i + i*ldda ], ldda );
}
}
/* =====================================================================
Check the result, following zqrt01 except using the reduced Q.
This works for any M,N (square, tall, wide).
Only for version 2, which has LAPACK complaint output.
Or for version 3, after restoring diagonal blocks of A above.
=================================================================== */
magma_sgetmatrix( M, N, d_A, ldda, h_R, lda );
magma_int_t ldq = M;
magma_int_t ldr = min_mn;
float *Q, *R;
float *work;
TESTING_MALLOC_CPU( Q, float, ldq*min_mn ); // M by K
TESTING_MALLOC_CPU( R, float, ldr*N ); // K by N
TESTING_MALLOC_CPU( work, float, min_mn );
// generate M by K matrix Q, where K = min(M,N)
lapackf77_slacpy( "Lower", &M, &min_mn, h_R, &lda, Q, &ldq );
lapackf77_sorgqr( &M, &min_mn, &min_mn, Q, &ldq, tau, h_work, &lwork, &info );
assert( info == 0 );
// copy K by N matrix R
lapackf77_slaset( "Lower", &min_mn, &N, &c_zero, &c_zero, R, &ldr );
lapackf77_slacpy( "Upper", &min_mn, &N, h_R, &lda, R, &ldr );
// error = || R - Q^H*A || / (N * ||A||)
blasf77_sgemm( "Conj", "NoTrans", &min_mn, &N, &M,
&c_neg_one, Q, &ldq, h_A, &lda, &c_one, R, &ldr );
Anorm = lapackf77_slange( "1", &M, &N, h_A, &lda, work );
error = lapackf77_slange( "1", &min_mn, &N, R, &ldr, work );
if ( N > 0 && Anorm > 0 )
error /= (N*Anorm);
// set R = I (K by K identity), then R = I - Q^H*Q
// error = || I - Q^H*Q || / N
lapackf77_slaset( "Upper", &min_mn, &min_mn, &c_zero, &c_one, R, &ldr );
blasf77_ssyrk( "Upper", "Conj", &min_mn, &M, &d_neg_one, Q, &ldq, &d_one, R, &ldr );
error2 = safe_lapackf77_slansy( "1", "Upper", &min_mn, R, &ldr, work );
if ( N > 0 )
error2 /= N;
TESTING_FREE_CPU( Q ); Q = NULL;
TESTING_FREE_CPU( R ); R = NULL;
TESTING_FREE_CPU( work ); work = NULL;
}
else if ( opts.check == 2 && M >= N && (opts.version == 1 || opts.version == 3) ) {
/* =====================================================================
Check the result by solving consistent linear system, A*x = b.
Only for versions 1 & 3 with M >= N.
=================================================================== */
magma_int_t lwork2;
float *x, *b, *hwork;
magmaFloat_ptr d_B;
// initialize RHS, b = A*random
TESTING_MALLOC_CPU( x, float, N );
TESTING_MALLOC_CPU( b, float, M );
lapackf77_slarnv( &ione, ISEED, &N, x );
blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_zero, b, &ione );
// copy to GPU
TESTING_MALLOC_DEV( d_B, float, M );
magma_ssetvector( M, b, 1, d_B, 1 );
if ( opts.version == 1 ) {
// allocate hwork
magma_sgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, float, lwork2 );
// solve linear system
magma_sgeqrs_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork2, &info );
if (info != 0) {
printf("magma_sgeqrs returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
TESTING_FREE_CPU( hwork );
}
#ifdef HAVE_CUBLAS
else if ( opts.version == 3 ) {
// allocate hwork
magma_sgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, tmp, -1, &info );
lwork2 = (magma_int_t)MAGMA_S_REAL( tmp[0] );
TESTING_MALLOC_CPU( hwork, float, lwork2 );
// solve linear system
magma_sgeqrs3_gpu( M, N, 1,
d_A, ldda, tau, dT,
d_B, M, hwork, lwork2, &info );
if (info != 0) {
printf("magma_sgeqrs3 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
TESTING_FREE_CPU( hwork );
}
#endif
else {
printf( "Unknown version %d\n", (int) opts.version );
return -1;
}
magma_sgetvector( N, d_B, 1, x, 1 );
// compute r = Ax - b, saved in b
blasf77_sgemv( "Notrans", &M, &N, &c_one, h_A, &lda, x, &ione, &c_neg_one, b, &ione );
// compute residual |Ax - b| / (max(m,n)*|A|*|x|)
float norm_x, norm_A, norm_r, work[1];
norm_A = lapackf77_slange( "F", &M, &N, h_A, &lda, work );
norm_r = lapackf77_slange( "F", &M, &ione, b, &M, work );
norm_x = lapackf77_slange( "F", &N, &ione, x, &N, work );
TESTING_FREE_CPU( x );
TESTING_FREE_CPU( b );
TESTING_FREE_DEV( d_B );
error = norm_r / (max(M,N) * norm_A * norm_x);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_sgeqrf( &M, &N, h_A, &lda, tau, h_work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_sgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
}
/* =====================================================================
Print performance and error.
=================================================================== */
printf("%5d %5d ", (int) M, (int) N );
if ( opts.lapack ) {
printf( "%7.2f (%7.2f)", cpu_perf, cpu_time );
}
else {
printf(" --- ( --- )" );
}
printf( " %7.2f (%7.2f) ", gpu_perf, gpu_time );
if ( opts.check == 1 ) {
bool okay = (error < tol && error2 < tol);
status += ! okay;
printf( "%11.2e %11.2e %s\n", error, error2, (okay ? "ok" : "failed") );
}
else if ( opts.check == 2 ) {
if ( M >= N ) {
bool okay = (error < tol);
status += ! okay;
printf( "%10.2e %s\n", error, (okay ? "ok" : "failed") );
}
else {
printf( "(error check only for M >= N)\n" );
}
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
if ( opts.version == 1 || opts.version == 3 ) {
TESTING_FREE_DEV( dT );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
SGETRF_NOPIV computes an LU factorization of a general M-by-N
matrix A without pivoting.
The factorization has the form
A = L * U
where L is lower triangular with unit diagonal elements (lower
trapezoidal if m > n), and U is upper triangular (upper
trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
- > 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
@ingroup magma_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrf_nopiv(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *info)
{
#define A(i_,j_) (A + (i_) + (j_)*lda)
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t min_mn, i__3, i__4;
magma_int_t j, jb, nb, iinfo;
A -= 1 + lda;
/* Function Body */
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0) {
return *info;
}
/* Determine the block size for this environment. */
nb = 128;
min_mn = min(m,n);
if (nb <= 1 || nb >= min_mn) {
/* Use unblocked code. */
magma_sgetf2_nopiv( m, n, A(1,1), lda, info );
}
else {
/* Use blocked code. */
for (j = 1; j <= min_mn; j += nb) {
jb = min( min_mn - j + 1, nb );
/* Factor diagonal and subdiagonal blocks and test for exact
singularity. */
i__3 = m - j + 1;
//magma_sgetf2_nopiv( i__3, jb, A(j,j), lda, &iinfo );
i__3 -= jb;
magma_sgetf2_nopiv( jb, jb, A(j,j), lda, &iinfo );
blasf77_strsm( "R", "U", "N", "N", &i__3, &jb, &c_one,
A(j,j), &lda,
A(j+jb,j), &lda );
/* Adjust INFO */
if (*info == 0 && iinfo > 0)
*info = iinfo + j - 1;
if (j + jb <= n) {
/* Compute block row of U. */
i__3 = n - j - jb + 1;
blasf77_strsm( "Left", "Lower", "No transpose", "Unit",
&jb, &i__3, &c_one,
A(j,j), &lda,
A(j,j+jb), &lda );
if (j + jb <= m) {
/* Update trailing submatrix. */
i__3 = m - j - jb + 1;
i__4 = n - j - jb + 1;
blasf77_sgemm( "No transpose", "No transpose",
&i__3, &i__4, &jb, &c_neg_one,
A(j+jb,j), &lda,
A(j,j+jb), &lda, &c_one,
A(j+jb,j+jb), &lda );
}
}
}
}
return *info;
} /* magma_sgetrf_nopiv */
/**
Purpose
-------
SLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to SLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code makes very mild assumptions about floating point
arithmetic. It will work on machines with a guard digit in
add/subtract, or on those binary machines without guard digits
which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
---------
@param[in]
ngpu INTEGER
Number of GPUs to use. ngpu > 0.
@param[in]
k INTEGER
The number of terms in the rational function to be solved by
SLAED4. K >= 0.
@param[in]
n INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N > K).
@param[in]
n1 INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
@param[out]
d REAL array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
@param[out]
Q REAL array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
@param[in]
ldq INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
@param[in]
rho REAL
The value of the parameter in the rank one update equation.
RHO >= 0 required.
@param[in,out]
dlamda REAL array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
@param[in]
Q2 REAL array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
@param[in]
indx INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see SLAED2).
The rows of the eigenvectors found by SLAED4 must be likewise
permuted before the matrix multiply can take place.
@param[in]
ctot INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
@param[in,out]
w REAL array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
@param
s (workspace) REAL array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
@param[out]
indxq INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
@param
dwork (devices workspaces) REAL array of arrays,
dimension NRGPU.
if NRGPU = 1 the dimension of the first workspace
should be (3*N*N/2+3*N)
otherwise the NRGPU workspaces should have the size
ceil((N-N1) * (N-N1) / floor(ngpu/2)) +
NB * ((N-N1) + (N-N1) / floor(ngpu/2))
@param
queues (device queues) magma_queue_t array,
dimension (MagmaMaxGPUs,2)
@param[in]
range magma_range_t
- = MagmaRangeAll: all eigenvalues will be found.
- = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU]
will be found.
- = MagmaRangeI: the IL-th through IU-th eigenvalues will be found.
TODO verify range, vl, vu, il, iu -- copied from slaex1.
@param[in]
vl REAL
@param[in]
vu REAL
if RANGE=MagmaRangeV, the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeI.
@param[in]
il INTEGER
@param[in]
iu INTEGER
if RANGE=MagmaRangeI, the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeV.
@param[out]
info INTEGER
- = 0: successful exit.
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: if INFO = 1, an eigenvalue did not converge
Further Details
---------------
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
@ingroup magma_ssyev_aux
********************************************************************/
extern "C" magma_int_t
magma_slaex3_m(
magma_int_t ngpu,
magma_int_t k, magma_int_t n, magma_int_t n1, float *d,
float *Q, magma_int_t ldq, float rho,
float *dlamda, float *Q2, magma_int_t *indx,
magma_int_t *ctot, float *w, float *s, magma_int_t *indxq,
magmaFloat_ptr dwork[],
magma_queue_t queues[MagmaMaxGPUs][2],
magma_range_t range, float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t *info )
{
#define Q(i_,j_) (Q + (i_) + (j_)*ldq)
#define dQ2(id) (dwork[id])
#define dS(id, ii) (dwork[id] + n2*n2_loc + (ii)*(n2*nb))
#define dQ(id, ii) (dwork[id] + n2*n2_loc + 2*(n2*nb) + (ii)*(n2_loc*nb))
if (ngpu == 1) {
magma_setdevice(0);
magma_slaex3(k, n, n1, d, Q, ldq, rho,
dlamda, Q2, indx, ctot, w, s, indxq,
*dwork, range, vl, vu, il, iu, info );
return *info;
}
float d_one = 1.;
float d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
magma_int_t iil, iiu, rk;
magma_int_t n1_loc, n2_loc, ib, nb, ib2, igpu;
magma_int_t ni_loc[MagmaMaxGPUs];
magma_int_t i, ind, iq2, j, n12, n2, n23, tmp;
float temp;
magma_int_t alleig, valeig, indeig;
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
*info = 0;
if (k < 0)
*info=-1;
else if (n < k)
*info=-2;
else if (ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if (*info != 0) {
magma_xerbla(__func__, -(*info));
return *info;
}
// Quick return if possible
if (k == 0)
return *info;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
//#define CHECK_CPU
#ifdef CHECK_CPU
float *hwS[2][MagmaMaxGPUs], *hwQ[2][MagmaMaxGPUs], *hwQ2[MagmaMaxGPUs];
#define hQ2(id) (hwQ2[id])
#define hS(id, ii) (hwS[ii][id])
#define hQ(id, ii) (hwQ[ii][id])
#endif
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
//lq2 = iq2 + n2 * n23;
n1_loc = (n1-1) / (ngpu/2) + 1;
n2_loc = (n2-1) / (ngpu/2) + 1;
nb = magma_get_slaex3_m_nb();
if (n1 >= magma_get_slaex3_m_k()) {
#ifdef CHECK_CPU
for (igpu = 0; igpu < ngpu; ++igpu) {
magma_smalloc_pinned( &(hwS[0][igpu]), n2*nb );
magma_smalloc_pinned( &(hwS[1][igpu]), n2*nb );
magma_smalloc_pinned( &(hwQ2[igpu]), n2*n2_loc );
magma_smalloc_pinned( &(hwQ[0][igpu]), n2_loc*nb );
magma_smalloc_pinned( &(hwQ[1][igpu]), n2_loc*nb );
}
#endif
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
ni_loc[igpu] = min(n1_loc, n1 - igpu/2 * n1_loc);
#ifdef CHECK_CPU
lapackf77_slacpy("A", &ni_loc[igpu], &n12, Q2+n1_loc*(igpu/2), &n1, hQ2(igpu), &n1_loc);
#endif
magma_setdevice(igpu);
magma_ssetmatrix_async( ni_loc[igpu], n12,
Q2+n1_loc*(igpu/2), n1,
dQ2(igpu), n1_loc, queues[igpu][0] );
ni_loc[igpu+1] = min(n2_loc, n2 - igpu/2 * n2_loc);
#ifdef CHECK_CPU
lapackf77_slacpy("A", &ni_loc[igpu+1], &n23, Q2+iq2+n2_loc*(igpu/2), &n2, hQ2(igpu+1), &n2_loc);
#endif
magma_setdevice(igpu+1);
magma_ssetmatrix_async( ni_loc[igpu+1], n23,
Q2+iq2+n2_loc*(igpu/2), n2,
dQ2(igpu+1), n2_loc, queues[igpu+1][0] );
}
}
//
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for (i = ib; i < ie; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = ib; j < ie; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0) {
#pragma omp critical (info)
*info = iinfo;
break;
}
}
#pragma omp barrier
if (*info == 0) {
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2) {
#pragma omp single
{
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for (j = 0; j < k; ++j) {
magma_int_t i_tmp = min(j, ie);
for (i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for (i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if (id < tot) {
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else {
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for (j = ib; j < ie; ++j) {
for (i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s+id*k, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return *info;
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
for (i = 0; i < k; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = 0; j < k; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0)
*info=iinfo;
}
if (*info != 0)
return *info;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2) {
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &k, Q, &tmp, w, &ione);
for (j = 0; j < k; ++j) {
for (i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for (i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for (j = iil-1; j < iiu; ++j) {
for (i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#endif //_OPENMP
// Compute the updated eigenvectors.
timer_start( time );
if (rk > 0) {
if (n1 < magma_get_slaex3_m_k()) {
// stay on the CPU
if ( n23 != 0 ) {
lapackf77_slacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_sgemm("N", "N", &n2, &rk, &n23, &d_one, &Q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
}
else
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 != 0 ) {
lapackf77_slacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_sgemm("N", "N", &n1, &rk, &n12, &d_one, Q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
}
else
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
else {
//use the gpus
ib = min(nb, rk);
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_ssetmatrix_async( n23, ib,
Q(ctot[0],iil-1), ldq,
dS(igpu+1,0), n23, queues[igpu+1][0] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_ssetmatrix_async( n12, ib,
Q(0,iil-1), ldq,
dS(igpu,0), n12, queues[igpu][0] );
}
}
for (i = 0; i < rk; i += nb) {
ib = min(nb, rk - i);
ind = (i/nb)%2;
if (i+nb < rk) {
ib2 = min(nb, rk - i - nb);
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_ssetmatrix_async( n23, ib2,
Q(ctot[0],iil-1+i+nb), ldq,
dS(igpu+1,(ind+1)%2), n23, queues[igpu+1][(ind+1)%2] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_ssetmatrix_async( n12, ib2,
Q(0,iil-1+i+nb), ldq,
dS(igpu,(ind+1)%2), n12, queues[igpu][(ind+1)%2] );
}
}
}
// Ensure that the data is copied on gpu since we will overwrite it.
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
#ifdef CHECK_CPU
lapackf77_slacpy("A", &n23, &ib, Q(ctot[0],iil-1+i), &ldq, hS(igpu+1,ind), &n23);
#endif
magma_setdevice(igpu+1);
magma_queue_sync( queues[igpu+1][ind] );
}
if (n12 != 0) {
#ifdef CHECK_CPU
lapackf77_slacpy("A", &n12, &ib, Q(0,iil-1+i), &ldq, hS(igpu,ind), &n12);
#endif
magma_setdevice(igpu);
magma_queue_sync( queues[igpu][ind] );
}
}
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
#ifdef CHECK_CPU
blasf77_sgemm("N", "N", &ni_loc[igpu+1], &ib, &n23, &d_one, hQ2(igpu+1), &n2_loc,
hS(igpu+1,ind), &n23, &d_zero, hQ(igpu+1, ind), &n2_loc);
#endif
magma_setdevice(igpu+1);
magmablasSetKernelStream(queues[igpu+1][ind]);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, ni_loc[igpu+1], ib, n23, d_one, dQ2(igpu+1), n2_loc,
dS(igpu+1, ind), n23, d_zero, dQ(igpu+1, ind), n2_loc);
#ifdef CHECK_CPU
printf("norm Q %d: %f\n", igpu+1, cpu_gpu_sdiff(ni_loc[igpu+1], ib, hQ(igpu+1, ind), n2_loc, dQ(igpu+1, ind), n2_loc));
#endif
}
if (n12 != 0) {
#ifdef CHECK_CPU
blasf77_sgemm("N", "N", &ni_loc[igpu], &ib, &n12, &d_one, hQ2(igpu), &n1_loc,
hS(igpu,ind%2), &n12, &d_zero, hQ(igpu, ind%2), &n1_loc);
#endif
magma_setdevice(igpu);
magmablasSetKernelStream(queues[igpu][ind]);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, ni_loc[igpu], ib, n12, d_one, dQ2(igpu), n1_loc,
dS(igpu, ind), n12, d_zero, dQ(igpu, ind), n1_loc);
#ifdef CHECK_CPU
printf("norm Q %d: %f\n", igpu, cpu_gpu_sdiff(ni_loc[igpu], ib, hQ(igpu, ind), n1_loc, dQ(igpu, ind), n1_loc));
#endif
}
}
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_sgetmatrix( ni_loc[igpu+1], ib, dQ(igpu+1, ind), n2_loc,
Q(n1+n2_loc*(igpu/2),iil-1+i), ldq );
// magma_sgetmatrix_async( ni_loc[igpu+1], ib, dQ(igpu+1, ind), n2_loc,
// Q(n1+n2_loc*(igpu/2),iil-1+i), ldq, queues[igpu+1][ind] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_sgetmatrix( ni_loc[igpu], ib, dQ(igpu, ind), n1_loc,
Q(n1_loc*(igpu/2),iil-1+i), ldq );
// magma_sgetmatrix_async( ni_loc[igpu], ib, dQ(igpu, ind), n1_loc,
// Q(n1_loc*(igpu/2),iil-1+i), ldq, queues[igpu][ind] );
}
}
}
for (igpu = 0; igpu < ngpu; ++igpu) {
#ifdef CHECK_CPU
magma_free_pinned( hwS[1][igpu] );
magma_free_pinned( hwS[0][igpu] );
magma_free_pinned( hwQ2[igpu] );
magma_free_pinned( hwQ[1][igpu] );
magma_free_pinned( hwQ[0][igpu] );
#endif
magma_setdevice(igpu);
magma_queue_sync( queues[igpu][0] );
magma_queue_sync( queues[igpu][1] );
}
if ( n23 == 0 )
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 == 0 )
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
}
timer_stop( time );
timer_printf( "gemms = %6.2f\n", time );
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_slaed3_m */
magma_int_t magma_strevc3(
magma_side_t side, magma_vec_t howmany,
magma_int_t *select, // logical in fortran
magma_int_t n,
float *T, magma_int_t ldt,
float *VL, magma_int_t ldvl,
float *VR, magma_int_t ldvr,
magma_int_t mm, magma_int_t *mout,
float *work, magma_int_t lwork,
#ifdef COMPLEX
float *rwork,
#endif
magma_int_t *info )
{
#define T(i,j) (T + (i) + (j)*ldt)
#define VL(i,j) (VL + (i) + (j)*ldvl)
#define VR(i,j) (VR + (i) + (j)*ldvr)
#define X(i,j) (X + (i)-1 + ((j)-1)*2) // still as 1-based indices
#define work(i,j) (work + (i) + (j)*n)
// constants
const magma_int_t ione = 1;
const float c_zero = 0;
const float c_one = 1;
const magma_int_t nbmin = 16, nbmax = 256;
// .. Local Scalars ..
magma_int_t allv, bothv, leftv, over, pair, rightv, somev;
magma_int_t i, ierr, ii, ip, is, j, k, ki, ki2,
iv, n2, nb, nb2, version;
float emax, remax;
// .. Local Arrays ..
// since iv is a 1-based index, allocate one extra here
magma_int_t iscomplex[ nbmax+1 ];
// Decode and test the input parameters
bothv = (side == MagmaBothSides);
rightv = (side == MagmaRight) || bothv;
leftv = (side == MagmaLeft ) || bothv;
allv = (howmany == MagmaAllVec);
over = (howmany == MagmaBacktransVec);
somev = (howmany == MagmaSomeVec);
*info = 0;
if ( ! rightv && ! leftv )
*info = -1;
else if ( ! allv && ! over && ! somev )
*info = -2;
else if ( n < 0 )
*info = -4;
else if ( ldt < max( 1, n ) )
*info = -6;
else if ( ldvl < 1 || ( leftv && ldvl < n ) )
*info = -8;
else if ( ldvr < 1 || ( rightv && ldvr < n ) )
*info = -10;
else if ( lwork < max( 1, 3*n ) )
*info = -14;
else {
// Set mout to the number of columns required to store the selected
// eigenvectors, standardize the array select if necessary, and
// test mm.
if ( somev ) {
*mout = 0;
pair = false;
for( j=0; j < n; ++j ) {
if ( pair ) {
pair = false;
select[j] = false;
}
else {
if ( j < n-1 ) {
if ( *T(j+1,j) == c_zero ) {
if ( select[j] ) {
*mout += 1;
}
}
else {
pair = true;
if ( select[j] || select[j+1] ) {
select[j] = true;
*mout += 2;
}
}
}
else if ( select[n-1] ) {
*mout += 1;
}
}
}
}
else {
*mout = n;
}
if ( mm < *mout ) {
*info = -11;
}
}
if ( *info != 0 ) {
magma_xerbla( __func__, -(*info) );
return *info;
}
// Quick return if possible.
if ( n == 0 ) {
return *info;
}
// Use blocked version (2) if sufficient workspace.
// Requires 1 vector for 1-norms, and 2*nb vectors for x and Q*x.
// Zero-out the workspace to avoid potential NaN propagation.
nb = 2;
if ( lwork >= n + 2*n*nbmin ) {
version = 2;
nb = (lwork - n) / (2*n);
nb = min( nb, nbmax );
nb2 = 1 + 2*nb;
lapackf77_slaset( "F", &n, &nb2, &c_zero, &c_zero, work, &n );
}
else {
version = 1;
}
// Compute 1-norm of each column of strictly upper triangular
// part of T to control overflow in triangular solver.
*work(0,0) = c_zero;
for( j=1; j < n; ++j ) {
*work(j,0) = c_zero;
for( i=0; i < j; ++i ) {
*work(j,0) += fabsf( *T(i,j) );
}
}
magma_timer_t time_total=0, time_trsv=0, time_gemm=0, time_gemv=0, time_trsv_sum=0, time_gemm_sum=0, time_gemv_sum=0;
timer_start( time_total );
// Index ip is used to specify the real or complex eigenvalue:
// ip = 0, real eigenvalue (wr),
// = 1, first of conjugate complex pair: (wr,wi)
// = -1, second of conjugate complex pair: (wr,wi)
// iscomplex array stores ip for each column in current block.
if ( rightv ) {
// ============================================================
// Compute right eigenvectors.
// iv is index of column in current block (1-based).
// For complex right vector, uses iv-1 for real part and iv for complex part.
// Non-blocked version always uses iv=2;
// blocked version starts with iv=nb, goes down to 1 or 2.
// (Note the "0-th" column is used for 1-norms computed above.)
iv = 2;
if ( version == 2 ) {
iv = nb;
}
timer_start( time_trsv );
ip = 0;
is = *mout - 1;
for( ki=n-1; ki >= 0; --ki ) {
if ( ip == -1 ) {
// previous iteration (ki+1) was second of conjugate pair,
// so this ki is first of conjugate pair; skip to end of loop
ip = 1;
continue;
}
else if ( ki == 0 ) {
// last column, so this ki must be real eigenvalue
ip = 0;
}
else if ( *T(ki,ki-1) == c_zero ) {
// zero on sub-diagonal, so this ki is real eigenvalue
ip = 0;
}
else {
// non-zero on sub-diagonal, so this ki is second of conjugate pair
ip = -1;
}
if ( somev ) {
if ( ip == 0 ) {
if ( ! select[ki] ) {
continue;
}
}
else {
if ( ! select[ki-1] ) {
continue;
}
}
}
if ( ip == 0 ) {
// ------------------------------------------------------------
// Real right eigenvector
// Solve upper quasi-triangular system:
// [ T(0:ki-1,0:ki-1) - wr ]*X = -T(0:ki-1,ki)
magma_slaqtrsd( MagmaNoTrans, ki+1, T(0,0), ldt,
work(0,iv), n, work(0,0), &ierr );
// Copy the vector x or Q*x to VR and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VR and normalize.
n2 = ki+1;
blasf77_scopy( &n2, work(0,iv), &ione, VR(0,is), &ione );
ii = blasf77_isamax( &n2, VR(0,is), &ione ) - 1; // subtract 1; ii is 0-based
remax = c_one / fabsf( *VR(ii,is) );
blasf77_sscal( &n2, &remax, VR(0,is), &ione );
for( k=ki + 1; k < n; ++k ) {
*VR(k,is) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemv );
if ( ki > 0 ) {
n2 = ki;
blasf77_sgemv( "n", &n, &n2, &c_one,
VR, &ldvr,
work(0, iv), &ione,
work(ki,iv), VR(0,ki), &ione );
}
time_gemv_sum += timer_stop( time_gemv );
ii = blasf77_isamax( &n, VR(0,ki), &ione ) - 1; // subtract 1; ii is 0-based
remax = c_one / fabsf( *VR(ii,ki) );
blasf77_sscal( &n, &remax, VR(0,ki), &ione );
timer_start( time_trsv );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out below vector
for( k=ki + 1; k < n; ++k ) {
*work(k,iv) = c_zero;
}
iscomplex[ iv ] = ip;
// back-transform and normalization is done below
}
} // end real eigenvector
else {
// ------------------------------------------------------------
// Complex right eigenvector
// Solve upper quasi-triangular system:
// [ T(0:ki-2,0:ki-2) - (wr+i*wi) ]*x = u
magma_slaqtrsd( MagmaNoTrans, ki+1, T(0,0), ldt,
work(0,iv-1), n, work(0,0), &ierr );
// Copy the vector x or Q*x to VR and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VR and normalize.
n2 = ki+1;
blasf77_scopy( &n2, work(0,iv-1), &ione, VR(0,is-1), &ione );
blasf77_scopy( &n2, work(0,iv ), &ione, VR(0,is ), &ione );
emax = c_zero;
for( k=0; k <= ki; ++k ) {
emax = max( emax, fabsf(*VR(k,is-1)) + fabsf(*VR(k,is)) );
}
remax = c_one / emax;
blasf77_sscal( &n2, &remax, VR(0,is-1), &ione );
blasf77_sscal( &n2, &remax, VR(0,is ), &ione );
for( k=ki + 1; k < n; ++k ) {
*VR(k,is-1) = c_zero;
*VR(k,is ) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemv );
if ( ki > 1 ) {
n2 = ki-1;
blasf77_sgemv( "n", &n, &n2, &c_one,
VR, &ldvr,
work(0, iv-1), &ione,
work(ki-1,iv-1), VR(0,ki-1), &ione );
blasf77_sgemv( "n", &n, &n2, &c_one,
VR, &ldvr,
work(0, iv), &ione,
work(ki,iv), VR(0,ki), &ione );
}
else {
blasf77_sscal( &n, work(ki-1,iv-1), VR(0,ki-1), &ione );
blasf77_sscal( &n, work(ki, iv ), VR(0,ki ), &ione );
}
time_gemv_sum += timer_stop( time_gemv );
emax = c_zero;
for( k=0; k < n; ++k ) {
emax = max( emax, fabsf(*VR(k,ki-1)) + fabsf(*VR(k,ki)) );
}
remax = c_one / emax;
blasf77_sscal( &n, &remax, VR(0,ki-1), &ione );
blasf77_sscal( &n, &remax, VR(0,ki ), &ione );
timer_start( time_trsv );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out below vector
for( k=ki + 1; k < n; ++k ) {
*work(k,iv-1) = c_zero;
*work(k,iv ) = c_zero;
}
iscomplex[ iv-1 ] = -ip;
iscomplex[ iv ] = ip;
iv -= 1;
// back-transform and normalization is done below
}
} // end real or complex vector
if ( version == 2 ) {
// ------------------------------------------------------------
// Blocked version of back-transform
// For complex case, ki2 includes both vectors (ki-1 and ki)
if ( ip == 0 ) {
ki2 = ki;
}
else {
ki2 = ki - 1;
}
// Columns iv:nb of work are valid vectors.
// When the number of vectors stored reaches nb-1 or nb,
// or if this was last vector, do the GEMM
if ( (iv <= 2) || (ki2 == 0) ) {
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemm );
nb2 = nb-iv+1;
n2 = ki2+nb-iv+1;
blasf77_sgemm( "n", "n", &n, &nb2, &n2, &c_one,
VR, &ldvr,
work(0,iv), &n,
&c_zero,
work(0,nb+iv), &n );
time_gemm_sum += timer_stop( time_gemm );
// normalize vectors
// TODO if somev, should copy vectors individually to correct location.
for( k=iv; k <= nb; ++k ) {
if ( iscomplex[k] == 0 ) {
// real eigenvector
ii = blasf77_isamax( &n, work(0,nb+k), &ione ) - 1; // subtract 1; ii is 0-based
remax = c_one / fabsf( *work(ii,nb+k) );
}
else if ( iscomplex[k] == 1 ) {
// first eigenvector of conjugate pair
emax = c_zero;
for( ii=0; ii < n; ++ii ) {
emax = max( emax, fabsf( *work(ii,nb+k ) )
+ fabsf( *work(ii,nb+k+1) ) );
}
remax = c_one / emax;
// else if iscomplex[k] == -1
// second eigenvector of conjugate pair
// reuse same remax as previous k
}
blasf77_sscal( &n, &remax, work(0,nb+k), &ione );
}
nb2 = nb-iv+1;
lapackf77_slacpy( "F", &n, &nb2,
work(0,nb+iv), &n,
VR(0,ki2), &ldvr );
iv = nb;
timer_start( time_trsv );
}
else {
iv -= 1;
}
} // end blocked back-transform
is -= 1;
if ( ip != 0 ) {
is -= 1;
}
}
}
timer_stop( time_trsv );
timer_stop( time_total );
timer_printf( "trevc trsv %.4f, gemm %.4f, gemv %.4f, total %.4f\n",
time_trsv_sum, time_gemm_sum, time_gemv_sum, time_total );
if ( leftv ) {
// ============================================================
// Compute left eigenvectors.
// iv is index of column in current block (1-based).
// For complex left vector, uses iv for real part and iv+1 for complex part.
// Non-blocked version always uses iv=1;
// blocked version starts with iv=1, goes up to nb-1 or nb.
// (Note the "0-th" column is used for 1-norms computed above.)
iv = 1;
ip = 0;
is = 0;
for( ki=0; ki < n; ++ki ) {
if ( ip == 1 ) {
// previous iteration (ki-1) was first of conjugate pair,
// so this ki is second of conjugate pair; skip to end of loop
ip = -1;
continue;
}
else if ( ki == n-1 ) {
// last column, so this ki must be real eigenvalue
ip = 0;
}
else if ( *T(ki+1,ki) == c_zero ) {
// zero on sub-diagonal, so this ki is real eigenvalue
ip = 0;
}
else {
// non-zero on sub-diagonal, so this ki is first of conjugate pair
ip = 1;
}
if ( somev ) {
if ( ! select[ki] ) {
continue;
}
}
if ( ip == 0 ) {
// ------------------------------------------------------------
// Real left eigenvector
// Solve transposed quasi-triangular system:
// [ T(ki+1:n,ki+1:n) - wr ]**T * X = -T(ki+1:n,ki)
magma_slaqtrsd( MagmaTrans, n-ki, T(ki,ki), ldt,
work(ki,iv), n, work(ki,0), &ierr );
// Copy the vector x or Q*x to VL and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VL and normalize.
n2 = n-ki;
blasf77_scopy( &n2, work(ki,iv), &ione, VL(ki,is), &ione );
ii = blasf77_isamax( &n2, VL(ki,is), &ione ) + ki - 1; // subtract 1; ii is 0-based
remax = c_one / fabsf( *VL(ii,is) );
blasf77_sscal( &n2, &remax, VL(ki,is), &ione );
for( k=0; k < ki; ++k ) {
*VL(k,is) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
if ( ki < n-1 ) {
n2 = n-ki-1;
blasf77_sgemv( "n", &n, &n2, &c_one,
VL(0,ki+1), &ldvl,
work(ki+1,iv), &ione,
work(ki, iv), VL(0,ki), &ione );
}
ii = blasf77_isamax( &n, VL(0,ki), &ione ) - 1; // subtract 1; ii is 0-based
remax = c_one / fabsf( *VL(ii,ki) );
blasf77_sscal( &n, &remax, VL(0,ki), &ione );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out above vector
// could go from (ki+1)-NV+1 to ki
for( k=0; k < ki; ++k ) {
*work(k,iv) = c_zero;
}
iscomplex[ iv ] = ip;
// back-transform and normalization is done below
}
} // end real eigenvector
else {
// ------------------------------------------------------------
// Complex left eigenvector
// Solve transposed quasi-triangular system:
// [ T(ki+2:n,ki+2:n)**T - (wr-i*wi) ]*X = V
magma_slaqtrsd( MagmaTrans, n-ki, T(ki,ki), ldt,
work(ki,iv), n, work(ki,0), &ierr );
// Copy the vector x or Q*x to VL and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VL and normalize.
n2 = n-ki;
blasf77_scopy( &n2, work(ki,iv ), &ione, VL(ki,is ), &ione );
blasf77_scopy( &n2, work(ki,iv+1), &ione, VL(ki,is+1), &ione );
emax = c_zero;
for( k=ki; k < n; ++k ) {
emax = max( emax, fabsf(*VL(k,is))+ fabsf(*VL(k,is+1)) );
}
remax = c_one / emax;
blasf77_sscal( &n2, &remax, VL(ki,is ), &ione );
blasf77_sscal( &n2, &remax, VL(ki,is+1), &ione );
for( k=0; k < ki; ++k ) {
*VL(k,is ) = c_zero;
*VL(k,is+1) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
if ( ki < n-2 ) {
n2 = n-ki-2;
blasf77_sgemv( "n", &n, &n2, &c_one,
VL(0,ki+2), &ldvl,
work(ki+2,iv), &ione,
work(ki, iv), VL(0,ki), &ione );
blasf77_sgemv( "n", &n, &n2, &c_one,
VL(0,ki+2), &ldvl,
work(ki+2,iv+1), &ione,
work(ki+1,iv+1), VL(0,ki+1), &ione );
}
else {
blasf77_sscal( &n, work(ki, iv ), VL(0, ki ), &ione );
blasf77_sscal( &n, work(ki+1,iv+1), VL(0, ki+1), &ione );
}
emax = c_zero;
for( k=0; k < n; ++k ) {
emax = max( emax, fabsf(*VL(k,ki))+ fabsf(*VL(k,ki+1)) );
}
remax = c_one / emax;
blasf77_sscal( &n, &remax, VL(0,ki ), &ione );
blasf77_sscal( &n, &remax, VL(0,ki+1), &ione );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out above vector
// could go from (ki+1)-NV+1 to ki
for( k=0; k < ki; ++k ) {
*work(k,iv ) = c_zero;
*work(k,iv+1) = c_zero;
}
iscomplex[ iv ] = ip;
iscomplex[ iv+1 ] = -ip;
iv += 1;
// back-transform and normalization is done below
}
} // end real or complex eigenvector
if ( version == 2 ) {
// -------------------------------------------------
// Blocked version of back-transform
// For complex case, (ki2+1) includes both vectors (ki+1) and (ki+2)
if ( ip == 0 ) {
ki2 = ki;
}
else {
ki2 = ki + 1;
}
// Columns 1:iv of work are valid vectors.
// When the number of vectors stored reaches nb-1 or nb,
// or if this was last vector, do the GEMM
if ( (iv >= nb-1) || (ki2 == n-1) ) {
n2 = n-(ki2+1)+iv;
blasf77_sgemm( "n", "n", &n, &iv, &n2, &c_one,
VL(0,ki2-iv+1), &ldvl,
work(ki2-iv+1,1), &n,
&c_zero,
work(0,nb+1), &n );
// normalize vectors
for( k=1; k <= iv; ++k ) {
if ( iscomplex[k] == 0 ) {
// real eigenvector
ii = blasf77_isamax( &n, work(0,nb+k), &ione ) - 1; // subtract 1; ii is 0-based
remax = c_one / fabsf( *work(ii,nb+k) );
}
else if ( iscomplex[k] == 1) {
// first eigenvector of conjugate pair
emax = c_zero;
for( ii=0; ii < n; ++ii ) {
emax = max( emax, fabsf( *work(ii,nb+k ) )
+ fabsf( *work(ii,nb+k+1) ) );
}
remax = c_one / emax;
// else if iscomplex[k] == -1
// second eigenvector of conjugate pair
// reuse same remax as previous k
}
blasf77_sscal( &n, &remax, work(0,nb+k), &ione );
}
lapackf77_slacpy( "F", &n, &iv,
work(0,nb+1), &n,
VL(0,ki2-iv+1), &ldvl );
iv = 1;
}
else {
iv += 1;
}
} // blocked back-transform
is += 1;
if ( ip != 0 ) {
is += 1;
}
}
}
return *info;
} // end of STREVC3
