// task execution code
void SCHED_ztrsm(Quark* quark)
{
magma_int_t UPPER;
magma_int_t M;
magma_int_t N;
cuDoubleComplex *A1;
magma_int_t LDA;
cuDoubleComplex *A2;
cuDoubleComplex one = MAGMA_Z_ONE;
quark_unpack_args_6(quark, UPPER, M, N, A1, LDA, A2);
if (UPPER) {
blasf77_ztrsm(MagmaLeftStr, MagmaUpperStr, MagmaConjTransStr, MagmaNonUnitStr,
&M, &N, &one, A1, &LDA, A2, &LDA);
} else {
blasf77_ztrsm(MagmaRightStr, MagmaLowerStr, MagmaConjTransStr, MagmaNonUnitStr,
&M, &N, &one, A1, &LDA, A2, &LDA);
}
}
/* Task execution code */
static void SCHED_panel_update(Quark* quark)
{
int N;
cuDoubleComplex *A1;
int LDA;
int K2;
int *IPIV;
cuDoubleComplex *A2;
int M;
int K;
cuDoubleComplex *A3;
cuDoubleComplex *A4;
int ione=1;
cuDoubleComplex mone = MAGMA_Z_NEG_ONE;
cuDoubleComplex one = MAGMA_Z_ONE;
quark_unpack_args_10(quark, N, A1, LDA, K2, IPIV, A2, M, K, A3, A4);
lapackf77_zlaswp(&N, A1, &LDA, &ione, &K2, IPIV, &ione);
blasf77_ztrsm("l", "l", "n", "u",
&K2, &N, &one, A2, &LDA, A1, &LDA);
if (M > 0) {
blasf77_zgemm("n","n",
&M, &N, &K, &mone, A3, &LDA, A1, &LDA, &one, A4, &LDA);
}
}
extern "C" magma_int_t
magma_zgeqrs_gpu(magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaDoubleComplex *dA, magma_int_t ldda,
magmaDoubleComplex *tau, magmaDoubleComplex *dT,
magmaDoubleComplex *dB, magma_int_t lddb,
magmaDoubleComplex *hwork, magma_int_t lwork,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
Solves the least squares problem
min || A*X - C ||
using the QR factorization A = Q*R computed by ZGEQRF_GPU.
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. M >= N >= 0.
NRHS (input) INTEGER
The number of columns of the matrix C. NRHS >= 0.
A (input) COMPLEX_16 array on the GPU, dimension (LDDA,N)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,n, as returned by
ZGEQRF_GPU in the first n columns of its array argument A.
LDDA (input) INTEGER
The leading dimension of the array A, LDDA >= M.
TAU (input) COMPLEX_16 array, dimension (N)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by MAGMA_ZGEQRF_GPU.
DB (input/output) COMPLEX_16 array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
DT (input) COMPLEX_16 array that is the output (the 6th argument)
of magma_zgeqrf_gpu of size
2*MIN(M, N)*NB + ((N+31)/32*32 )* MAX(NB, NRHS).
The array starts with a block of size MIN(M,N)*NB that stores
the triangular T matrices used in the QR factorization,
followed by MIN(M,N)*NB block storing the diagonal block
inverses for the R matrix, followed by work space of size
((N+31)/32*32 )* MAX(NB, NRHS).
LDDB (input) INTEGER
The leading dimension of the array DB. LDDB >= M.
HWORK (workspace/output) COMPLEX_16 array, dimension (LWORK)
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The dimension of the array WORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_zgeqrf_nb( M ).
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the WORK array.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
===================================================================== */
#define a_ref(a_1,a_2) (dA+(a_2)*(ldda) + (a_1))
#define d_ref(a_1) (dT+(lddwork+(a_1))*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *dwork;
magma_int_t i, k, lddwork, rows, ib;
magma_int_t ione = 1;
magma_int_t nb = magma_get_zgeqrf_nb(m);
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
int lquery = (lwork == -1);
hwork[0] = MAGMA_Z_MAKE( (double)lwkopt, 0. );
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0 || m < n)
*info = -2;
else if (nrhs < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
else if (lddb < max(1,m))
*info = -9;
else if (lwork < lwkopt && ! lquery)
*info = -11;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = c_one;
return *info;
}
/* B := Q' * B */
magma_zunmqr_gpu( MagmaLeft, MagmaConjTrans,
m, nrhs, n,
a_ref(0,0), ldda, tau,
dB, lddb, hwork, lwork, dT, nb, info );
if ( *info != 0 ) {
return *info;
}
/* Solve R*X = B(1:n,:) */
lddwork= k;
if (nb < k)
dwork = dT+2*lddwork*nb;
else
dwork = dT;
// To do: Why did we have this line originally; seems to be a bug (Stan)?
// dwork = dT;
i = (k-1)/nb * nb;
ib = n-i;
rows = m-i;
// TODO: this assumes that, on exit from magma_zunmqr_gpu, hwork contains
// the last block of A and B (i.e., C in zunmqr). This should be fixed.
// Seems this data should already be on the GPU, so could switch to
// magma_ztrsm and drop the zsetmatrix.
if ( nrhs == 1 ) {
blasf77_ztrsv( MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, hwork, &rows,
hwork+rows*ib, &ione);
} else {
blasf77_ztrsm( MagmaLeftStr, MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, &nrhs,
&c_one, hwork, &rows,
hwork+rows*ib, &rows);
}
// update the solution vector
magma_zsetmatrix( ib, nrhs, hwork+rows*ib, rows, dwork+i, lddwork );
// update c
if (nrhs == 1)
magma_zgemv( MagmaNoTrans, i, ib,
c_neg_one, a_ref(0, i), ldda,
dwork + i, 1,
c_one, dB, 1);
else
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, a_ref(0, i), ldda,
dwork + i, lddwork,
c_one, dB, lddb);
int start = i-nb;
if (nb < k) {
for (i = start; i >=0; i -= nb) {
ib = min(k-i, nb);
rows = m -i;
if (i + ib < n) {
if (nrhs == 1) {
magma_zgemv( MagmaNoTrans, ib, ib,
c_one, d_ref(i), ib,
dB+i, 1,
c_zero, dwork+i, 1);
magma_zgemv( MagmaNoTrans, i, ib,
c_neg_one, a_ref(0, i), ldda,
dwork + i, 1,
c_one, dB, 1);
}
else {
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
ib, nrhs, ib,
c_one, d_ref(i), ib,
dB+i, lddb,
c_zero, dwork+i, lddwork);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, a_ref(0, i), ldda,
dwork + i, lddwork,
c_one, dB, lddb);
}
}
}
}
magma_zcopymatrix( (n), nrhs,
dwork, lddwork,
dB, lddb );
return *info;
}
extern "C" magma_int_t
magma_zgetrf_nopiv(magma_int_t *m, magma_int_t *n, magmaDoubleComplex *a,
magma_int_t *lda, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
ZGETRF_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.
===================================================================== */
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t a_dim1, a_offset, min_mn, i__3, i__4;
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_zgetf2_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_zgetf2_nopiv(&i__3, &jb, &a[j + j * a_dim1], lda, &iinfo);
i__3 -= jb;
magma_zgetf2_nopiv(&jb, &jb, &a[j + j * a_dim1], lda, &iinfo);
blasf77_ztrsm("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_ztrsm("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;
blasf77_zgemm("No transpose", "No transpose", &i__3, &i__4, &jb,
&c_neg_one, &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_zgetrf_nopiv */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ztrsm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf=0, cpu_time=0;
double magma_error, cublas_error, work[1];
magma_int_t M, N, info;
magma_int_t Ak;
magma_int_t sizeA, sizeB;
magma_int_t lda, ldb, ldda, lddb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t *piv;
magma_err_t err;
magmaDoubleComplex *h_A, *h_B, *h_Bcublas, *h_Bmagma, *h_B1, *h_X1, *h_X2, *LU, *LUT;
magmaDoubleComplex *d_A, *d_B;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
magma_opts opts;
parse_opts( argc, argv, &opts );
printf("If running lapack (option --lapack), MAGMA and CUBLAS error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.\n\n"
"side = %c, uplo = %c, transA = %c, diag = %c \n", opts.side, opts.uplo, opts.transA, opts.diag );
printf(" M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n");
printf("==================================================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
gflops = FLOPS_ZTRSM(opts.side, M, N) / 1e9;
if ( opts.side == MagmaLeft ) {
lda = M;
Ak = M;
} else {
lda = N;
Ak = N;
}
ldb = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
sizeA = lda*Ak;
sizeB = ldb*N;
TESTING_MALLOC( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC( LU, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC( LUT, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC( h_B, magmaDoubleComplex, ldb*N );
TESTING_MALLOC( h_B1, magmaDoubleComplex, ldb*N );
TESTING_MALLOC( h_X1, magmaDoubleComplex, ldb*N );
TESTING_MALLOC( h_X2, magmaDoubleComplex, ldb*N );
TESTING_MALLOC( h_Bcublas, magmaDoubleComplex, ldb*N );
TESTING_MALLOC( h_Bmagma, magmaDoubleComplex, ldb*N );
TESTING_DEVALLOC( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_DEVALLOC( d_B, magmaDoubleComplex, lddb*N );
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, LU );
err = magma_malloc_cpu( (void**) &piv, Ak*sizeof(magma_int_t) ); assert( err == 0 );
lapackf77_zgetrf( &Ak, &Ak, LU, &lda, piv, &info );
int i, j;
for(i=0;i<Ak;i++){
for(j=0;j<Ak;j++){
LUT[j+i*lda] = LU[i+j*lda];
}
}
lapackf77_zlacpy(MagmaUpperStr, &Ak, &Ak, LUT, &lda, LU, &lda);
if(opts.uplo == MagmaLower){
lapackf77_zlacpy(MagmaLowerStr, &Ak, &Ak, LU, &lda, h_A, &lda);
}else{
lapackf77_zlacpy(MagmaUpperStr, &Ak, &Ak, LU, &lda, h_A, &lda);
}
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
memcpy(h_B1, h_B, sizeB*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using MAGMA-BLAS
=================================================================== */
magma_zsetmatrix( Ak, Ak, h_A, lda, d_A, ldda );
magma_zsetmatrix( M, N, h_B, ldb, d_B, lddb );
magma_time = magma_sync_wtime( NULL );
magmablas_ztrsm( opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetmatrix( M, N, d_B, lddb, h_Bmagma, ldb );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_zsetmatrix( M, N, h_B, ldb, d_B, lddb );
cublas_time = magma_sync_wtime( NULL );
cublasZtrsm( opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetmatrix( M, N, d_B, lddb, h_Bcublas, ldb );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_ztrsm( &opts.side, &opts.uplo, &opts.transA, &opts.diag,
&M, &N,
&alpha, h_A, &lda,
h_B, &ldb );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
// ||b - Ax|| / (||A||*||x||)
memcpy(h_X1, h_Bmagma, sizeB*sizeof(magmaDoubleComplex));
magmaDoubleComplex alpha2 = MAGMA_Z_DIV( c_one, alpha );
blasf77_ztrmm( &opts.side, &opts.uplo, &opts.transA, &opts.diag,
&M, &N,
&alpha2, h_A, &lda,
h_X1, &ldb );
blasf77_zaxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X1, &ione );
double norm1 = lapackf77_zlange( "M", &M, &N, h_X1, &ldb, work );
double normx = lapackf77_zlange( "M", &M, &N, h_Bmagma, &ldb, work );
double normA = lapackf77_zlange( "M", &Ak, &Ak, h_A, &lda, work );
magma_error = norm1/(normx*normA);
memcpy(h_X2, h_Bcublas, sizeB*sizeof(magmaDoubleComplex));
blasf77_ztrmm( &opts.side, &opts.uplo, &opts.transA, &opts.diag,
&M, &N,
&alpha2, h_A, &lda,
h_X2, &ldb );
blasf77_zaxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X2, &ione );
norm1 = lapackf77_zlange( "M", &M, &N, h_X2, &ldb, work );
normx = lapackf77_zlange( "M", &M, &N, h_Bcublas, &ldb, work );
normA = lapackf77_zlange( "M", &Ak, &Ak, h_A, &lda, work );
cublas_error = norm1/(normx*normA);
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error );
}
else {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e %8.2e\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
magma_error, cublas_error );
}
TESTING_FREE( h_A );
TESTING_FREE( LU );
TESTING_FREE( LUT );
TESTING_FREE( h_B );
TESTING_FREE( h_Bcublas );
TESTING_FREE( h_Bmagma );
TESTING_FREE( h_B1 );
TESTING_FREE( h_X1 );
TESTING_FREE( h_X2 );
TESTING_DEVFREE( d_A );
TESTING_DEVFREE( d_B );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
/**
Purpose
-------
ZPOTRF computes the Cholesky factorization of a complex Hermitian
positive definite matrix dA.
The factorization has the form
dA = U**H * U, if UPLO = MagmaUpper, or
dA = L * L**H, if UPLO = MagmaLower,
where U is an upper triangular matrix and L is lower triangular.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of dA is stored;
- = MagmaLower: Lower triangle of dA is stored.
@param[in]
n INTEGER
The order of the matrix dA. N >= 0.
@param[in,out]
d_lA COMPLEX_16 array of pointers on the GPU, dimension (num_gpus)
On entry, the Hermitian 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_zposv_comp
********************************************************************/
extern "C" magma_int_t
magma_zpotrf_mgpu_right(magma_int_t num_gpus, magma_uplo_t uplo, magma_int_t n,
magmaDoubleComplex **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) (num_streams > 1 ? 1+((i)/nb)%(num_streams-1) : 0)
magmaDoubleComplex z_one = MAGMA_Z_MAKE( 1.0, 0.0 );
magmaDoubleComplex mz_one = MAGMA_Z_MAKE( -1.0, 0.0 );
double one = 1.0;
double m_one = -1.0;
const char* uplo_ = lapack_uplo_const( uplo );
magma_int_t j, nb, d, id, j_local, blkid, crosspoint, prevj, prevtrsmrows, num_streams = 5;
magmaDoubleComplex *panel, *tmppanel0, *tmppanel1, *tmppanel, *tmpprevpanel;
magmaDoubleComplex *d_lP[MagmaMaxGPUs], *dlpanel, *dlpanels[MagmaMaxGPUs];
magma_int_t rows, trsmrows, ngpu, n_local[MagmaMaxGPUs], ldpanel;
magma_queue_t stream[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_zpotrf_nb(n);
ldpanel = ldda;
magma_setdevice(0);
if (MAGMA_SUCCESS != magma_zmalloc_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_zgetmatrix( n, n, dlA(0, 0, 0), ldda, panel, ldpanel);
lapackf77_zpotrf( uplo_, &n, panel, &ldpanel, info);
magma_zsetmatrix( n, n, panel, ldpanel, dlA(0, 0, 0), ldda );
} else {
for( d = 0; d < num_gpus; d++ ) {
// local-n and local-ld
n_local[d] = ((n / nb) / num_gpus) * nb;
if (d < (n / nb) % num_gpus)
n_local[d] += nb;
else if (d == (n / nb) % num_gpus)
n_local[d] += n % nb;
magma_setdevice(d);
magma_device_sync();
if (MAGMA_SUCCESS != magma_zmalloc( &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 < num_streams; j++ ) {
magma_queue_create( &stream[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 (num_gpus == 4)
crosspoint = n;
else if (num_gpus == 3)
crosspoint = n;
else if (num_gpus == 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 = get_time(), tset = 0.0, ttot = 0.0;
#endif
if ( n > nb ) {
// send first panel to cpu
magma_setdevice(0);
tmppanel = tmppanel0;
magma_zgetmatrix_async(n, nb,
dlA(0, 0, 0), ldda,
tmppanel(0), ldpanel,
stream[0][0] );
}
#if defined (ENABLE_TIMER)
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_device_sync();
}
tget = get_time()-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) % num_gpus;
magma_setdevice(id);
// Set the local index where the current panel is
j_local = j / (nb * num_gpus) * nb;
rows = n - j;
// Wait for the panel on cpu
magma_queue_sync( stream[id][0] );
if (j > 0 && prevtrsmrows > crosspoint) {
#if defined (ENABLE_TIMER)
tcnp = get_time();
#endif
tmpprevpanel = ((blkid - 1) % 2) == 0 ? tmppanel0 : tmppanel1;
blasf77_zgemm( MagmaNoTransStr, MagmaConjTransStr,
&rows, &nb, &nb,
&mz_one, tmpprevpanel(j), &ldpanel,
tmpprevpanel(j), &ldpanel,
&z_one, tmppanel(j), &ldpanel );
#if defined (ENABLE_TIMER)
tcnp = get_time() - tcnp;
ttot_cnp += tcnp;
#endif
}
#if defined (ENABLE_TIMER)
tcchol = get_time();
#endif
lapackf77_zpotrf(MagmaLowerStr, &nb, tmppanel(j), &ldpanel, info);
if (*info != 0) {
*info = *info + j;
break;
}
#if defined (ENABLE_TIMER)
tcchol = get_time() - tcchol;
ttot_cchol += tcchol;
tctrsm = get_time();
#endif
trsmrows = rows - nb;
if (trsmrows > 0) {
blasf77_ztrsm(MagmaRightStr, MagmaLowerStr, MagmaConjTransStr, MagmaNonUnitStr,
&trsmrows, &nb,
&z_one, tmppanel(j), &ldpanel,
tmppanel(j + nb), &ldpanel);
}
#if defined (ENABLE_TIMER)
tctrsm = get_time() - tctrsm;
ttot_ctrsm += tctrsm;
tctm = get_time();
#endif
d = (id + 1) % num_gpus;
// send current panel to gpus
for (ngpu = 0; ngpu < num_gpus; ngpu++, d = (d + 1) % num_gpus ) {
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_zsetmatrix_async(myrows, nb,
tmppanel(j + row_offset), ldpanel,
dlpanel, ldda, stream[d][0] );
}
}
/* make sure panel is on GPUs */
d = (id + 1) % num_gpus;
for (ngpu = 0; ngpu < num_gpus; ngpu++, d = (d + 1) % num_gpus ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
}
#if defined (ENABLE_TIMER)
tctm = get_time() - 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) % num_gpus;
magma_setdevice(d);
magma_int_t j_local2 = (j + nb) / (nb * num_gpus) * nb;
if (trsmrows <= crosspoint) {
#if defined (ENABLE_TIMER)
tmnp = get_time();
#endif
// do gemm on look ahead panel
if ( d == id ) {
dlpanel = dlA(d, j + nb, j_local);
} else {
dlpanel = dlP(d, 0, 0);
}
magmablasSetKernelStream(stream[d][STREAM_ID(j_local2)]);
#define ZHERK_ON_DIAG
#ifdef ZHERK_ON_DIAG
magma_zherk( MagmaLower, MagmaNoTrans,
nb, nb,
m_one, dlpanel, ldda,
one, dlA(d, j + nb, j_local2), ldda);
magma_zgemm( 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_zgemm( 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 = get_time() - tmnp;
ttot_mnp += tmnp;
#endif
}
// send next panel to cpu
magma_queue_sync( stream[d][STREAM_ID(j_local2)] ); // make sure lookahead is done
tmppanel = ((blkid+1) % 2 == 0) ? tmppanel0 : tmppanel1;
magma_zgetmatrix_async(rows-nb, nb,
dlA(d, j+nb, j_local2), ldda,
tmppanel(j+nb), ldpanel,
stream[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 < num_gpus; 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 < num_gpus; d++ ) therk[d] = get_time();
#endif
//magmablasSetKernelStream(stream[d]);
//magma_zherk(MagmaLower, MagmaNoTrans, n - offset, nb,
// m_one, dlpanel, ldda,
// one, &d_lA[d][offset + offset*ldda], ldda );
#ifdef ZHERK_ON_DIAG
magma_zherk_mgpu
#else
magma_zherk_mgpu2
#endif
(num_gpus, MagmaLower, MagmaNoTrans,
nb, n - offset, nb,
m_one, dlpanels, ldda, 0,
one, d_lA, ldda, offset,
num_streams, stream );
#if defined (ENABLE_TIMER)
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_device_sync();
therk[d] = get_time() - therk[d];
ttot_herk[d] += therk[d];
}
#endif
}
prevtrsmrows = trsmrows;
prevj = j;
#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) % num_gpus,
(tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp));
fflush(0);
#endif
}
}
for( d = 0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( id=0; id < num_streams; id++ ) {
magma_queue_sync( stream[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) % num_gpus;
// Set the local index where the current panel is
j_local = j / (nb * num_gpus) * nb;
magma_setdevice(id);
#if defined (ENABLE_TIMER)
tset = get_time();
#endif
magma_zgetmatrix(rows, rows, dlA(id, j, j_local), ldda, panel(j), ldpanel);
lapackf77_zpotrf(MagmaLowerStr, &rows, panel(j), &ldpanel, info);
magma_zsetmatrix(rows, rows, panel(j), ldpanel, dlA(id, j, j_local), ldda);
#if defined (ENABLE_TIMER)
tset = get_time() - 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 < num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j < num_streams; j++ ) {
magma_queue_destroy( stream[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_zpotrf_mgpu_right */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ztrsm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time=0, cublas_perf, cublas_time, cpu_perf=0, cpu_time=0;
double magma_error, cublas_error, work[1];
magma_int_t M, N, info;
magma_int_t Ak;
magma_int_t sizeA, sizeB;
magma_int_t lda, ldb, ldda, lddb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t *ipiv;
magmaDoubleComplex *h_A, *h_B, *h_Bcublas, *h_Bmagma, *h_B1, *h_X1, *h_X2;
magmaDoubleComplex *d_A, *d_B;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("side = %s, uplo = %s, transA = %s, diag = %s \n",
lapack_side_const(opts.side), lapack_uplo_const(opts.uplo),
lapack_trans_const(opts.transA), lapack_diag_const(opts.diag) );
printf(" M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\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];
gflops = FLOPS_ZTRSM(opts.side, M, N) / 1e9;
if ( opts.side == MagmaLeft ) {
lda = M;
Ak = M;
} else {
lda = N;
Ak = N;
}
ldb = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
sizeA = lda*Ak;
sizeB = ldb*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_B1, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_X1, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_X2, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_Bcublas, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( h_Bmagma, magmaDoubleComplex, ldb*N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, Ak );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*N );
/* Initialize the matrices */
/* Factor A into LU to get well-conditioned triangular matrix.
* Copy L to U, since L seems okay when used with non-unit diagonal
* (i.e., from U), while U fails when used with unit diagonal. */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zgetrf( &Ak, &Ak, h_A, &lda, ipiv, &info );
for( int j = 0; j < Ak; ++j ) {
for( int i = 0; i < j; ++i ) {
*h_A(i,j) = *h_A(j,i);
}
}
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
memcpy(h_B1, h_B, sizeB*sizeof(magmaDoubleComplex));
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetmatrix( Ak, Ak, h_A, lda, d_A, ldda );
magma_zsetmatrix( M, N, h_B, ldb, d_B, lddb );
magma_time = magma_sync_wtime( NULL );
magmablas_ztrsm( opts.side, opts.uplo, opts.transA, opts.diag,
M, N,
alpha, d_A, ldda,
d_B, lddb );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetmatrix( M, N, d_B, lddb, h_Bmagma, ldb );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( M, N, h_B, ldb, d_B, lddb );
cublas_time = magma_sync_wtime( NULL );
cublasZtrsm( handle, cublas_side_const(opts.side), cublas_uplo_const(opts.uplo),
cublas_trans_const(opts.transA), cublas_diag_const(opts.diag),
M, N,
&alpha, d_A, ldda,
d_B, lddb );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetmatrix( M, N, d_B, lddb, h_Bcublas, ldb );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_ztrsm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha, h_A, &lda,
h_B, &ldb );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
// ||b - Ax|| / (||A||*||x||)
memcpy(h_X1, h_Bmagma, sizeB*sizeof(magmaDoubleComplex));
magmaDoubleComplex alpha2 = MAGMA_Z_DIV( c_one, alpha );
blasf77_ztrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha2, h_A, &lda,
h_X1, &ldb );
blasf77_zaxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X1, &ione );
double norm1 = lapackf77_zlange( "M", &M, &N, h_X1, &ldb, work );
double normx = lapackf77_zlange( "M", &M, &N, h_Bmagma, &ldb, work );
double normA = lapackf77_zlange( "M", &Ak, &Ak, h_A, &lda, work );
magma_error = norm1/(normx*normA);
memcpy(h_X2, h_Bcublas, sizeB*sizeof(magmaDoubleComplex));
blasf77_ztrmm( lapack_side_const(opts.side), lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&M, &N,
&alpha2, h_A, &lda,
h_X2, &ldb );
blasf77_zaxpy( &sizeB, &c_neg_one, h_B1, &ione, h_X2, &ione );
norm1 = lapackf77_zlange( "M", &M, &N, h_X2, &ldb, work );
normx = lapackf77_zlange( "M", &M, &N, h_Bcublas, &ldb, work );
normA = lapackf77_zlange( "M", &Ak, &Ak, h_A, &lda, work );
cublas_error = norm1/(normx*normA);
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_B1 );
TESTING_FREE_CPU( h_X1 );
TESTING_FREE_CPU( h_X2 );
TESTING_FREE_CPU( h_Bcublas );
TESTING_FREE_CPU( h_Bmagma );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
Solves the least squares problem
min || A*X - C ||
using the QR factorization A = Q*R computed by ZGEQRF_GPU.
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. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of columns of the matrix C. NRHS >= 0.
@param[in]
dA COMPLEX_16 array on the GPU, dimension (LDDA,N)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,n, as returned by
ZGEQRF_GPU in the first n columns of its array argument A.
@param[in]
ldda INTEGER
The leading dimension of the array A, LDDA >= M.
@param[in]
tau COMPLEX_16 array, dimension (N)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by MAGMA_ZGEQRF_GPU.
@param[in,out]
dB COMPLEX_16 array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
@param[in]
dT COMPLEX_16 array that is the output (the 6th argument)
of magma_zgeqrf_gpu of size
2*MIN(M, N)*NB + ((N+31)/32*32 )* MAX(NB, NRHS).
The array starts with a block of size MIN(M,N)*NB that stores
the triangular T matrices used in the QR factorization,
followed by MIN(M,N)*NB block storing the diagonal block
inverses for the R matrix, followed by work space of size
((N+31)/32*32 )* MAX(NB, NRHS).
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= M.
@param[out]
hwork (workspace) COMPLEX_16 array, dimension (LWORK)
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_zgeqrf_nb( M ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the WORK array.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_zgels_comp
********************************************************************/
extern "C" magma_int_t
magma_zgeqrs_gpu(
magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaDoubleComplex_ptr dA, magma_int_t ldda,
magmaDoubleComplex *tau,
magmaDoubleComplex_ptr dT,
magmaDoubleComplex_ptr dB, magma_int_t lddb,
magmaDoubleComplex *hwork, magma_int_t lwork,
magma_int_t *info)
{
#define dA(a_1,a_2) (dA + (a_2)*(ldda) + (a_1))
#define dT(a_1) (dT + (lddwork+(a_1))*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex_ptr dwork;
magma_int_t i, k, lddwork, rows, ib;
magma_int_t ione = 1;
magma_int_t nb = magma_get_zgeqrf_nb(m);
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
int lquery = (lwork == -1);
hwork[0] = MAGMA_Z_MAKE( (double)lwkopt, 0. );
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0 || m < n)
*info = -2;
else if (nrhs < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
else if (lddb < max(1,m))
*info = -9;
else if (lwork < lwkopt && ! lquery)
*info = -11;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = c_one;
return *info;
}
/* B := Q' * B */
magma_zunmqr_gpu( MagmaLeft, Magma_ConjTrans,
m, nrhs, n,
dA(0,0), ldda, tau,
dB, lddb, hwork, lwork, dT, nb, info );
if ( *info != 0 ) {
return *info;
}
/* Solve R*X = B(1:n,:) */
lddwork= k;
if (nb < k)
dwork = dT+2*lddwork*nb;
else
dwork = dT;
// To do: Why did we have this line originally; seems to be a bug (Stan)?
// dwork = dT;
i = (k-1)/nb * nb;
ib = n-i;
rows = m-i;
// TODO: this assumes that, on exit from magma_zunmqr_gpu, hwork contains
// the last block of A and B (i.e., C in zunmqr). This should be fixed.
// Seems this data should already be on the GPU, so could switch to
// magma_ztrsm and drop the zsetmatrix.
if ( nrhs == 1 ) {
blasf77_ztrsv( MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, hwork, &rows,
hwork+rows*ib, &ione);
} else {
blasf77_ztrsm( MagmaLeftStr, MagmaUpperStr, MagmaNoTransStr, MagmaNonUnitStr,
&ib, &nrhs,
&c_one, hwork, &rows,
hwork+rows*ib, &rows);
}
// update the solution vector
magma_zsetmatrix( ib, nrhs, hwork+rows*ib, rows, dwork+i, lddwork );
// update c
if (nrhs == 1)
magma_zgemv( MagmaNoTrans, i, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, 1,
c_one, dB, 1);
else
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, lddwork,
c_one, dB, lddb);
int start = i-nb;
if (nb < k) {
for (i = start; i >= 0; i -= nb) {
ib = min(k-i, nb);
rows = m -i;
if (i + ib < n) {
if (nrhs == 1) {
magma_zgemv( MagmaNoTrans, ib, ib,
c_one, dT(i), ib,
dB+i, 1,
c_zero, dwork+i, 1);
magma_zgemv( MagmaNoTrans, i, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, 1,
c_one, dB, 1);
}
else {
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
ib, nrhs, ib,
c_one, dT(i), ib,
dB+i, lddb,
c_zero, dwork+i, lddwork);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
i, nrhs, ib,
c_neg_one, dA(0, i), ldda,
dwork + i, lddwork,
c_one, dB, lddb);
}
}
}
}
magma_zcopymatrix( (n), nrhs,
dwork, lddwork,
dB, lddb );
return *info;
}
