extern "C" magma_int_t
magma_sgetrf_piv(magma_int_t m, magma_int_t n, magma_int_t NB,
float *A, magma_int_t lda, magma_int_t *ipiv, magma_int_t *info)
{
magma_int_t I, k1, k2, incx, minmn;
*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;
/* initialize nb */
minmn = min(m,n);
for( I=0; I < minmn-NB; I += NB ) {
k1 = 1+I+NB;
k2 = minmn;
incx = 1;
lapackf77_slaswp(&NB, A(0,I), &lda, &k1, &k2, ipiv, &incx);
}
return *info;
} /* magma_sgetrf_piv */
extern "C" magma_int_t
magma_sgetrf2_piv(magma_int_t m, magma_int_t n, magma_int_t start, magma_int_t end,
float *A, magma_int_t lda, magma_int_t *ipiv, magma_int_t *info)
{
magma_int_t I, k1, k2, nb, incx, minmn;
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0)
return *info;
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* initialize nb */
nb = magma_get_sgetrf_nb(m);
minmn = min( end, min(m,n) );
for( I=start; I < end-nb; I += nb ) {
incx = 1;
k1 = 1+I+nb;
k2 = minmn;
lapackf77_slaswp(&nb, A(0,I), &lda, &k1, &k2, ipiv, &incx);
}
return *info;
} /* magma_sgetrf_piv */
// On input, LU and ipiv is LU factorization of A. On output, LU is overwritten.
// Works for any m, n.
// Uses init_matrix() to re-generate original A as needed.
// Returns error in factorization, |PA - LU| / (n |A|)
// This allocates 3 more matrices to store A, L, and U.
float get_LU_error(
magma_opts &opts,
magma_int_t M, magma_int_t N,
float *LU, magma_int_t lda,
magma_int_t *ipiv)
{
magma_int_t min_mn = min(M,N);
magma_int_t ione = 1;
magma_int_t i, j;
float alpha = MAGMA_S_ONE;
float beta = MAGMA_S_ZERO;
float *A, *L, *U;
float work[1], matnorm, residual;
TESTING_MALLOC_CPU( A, float, lda*N );
TESTING_MALLOC_CPU( L, float, M*min_mn );
TESTING_MALLOC_CPU( U, float, min_mn*N );
memset( L, 0, M*min_mn*sizeof(float) );
memset( U, 0, min_mn*N*sizeof(float) );
// set to original A
init_matrix( opts, M, N, A, lda );
lapackf77_slaswp( &N, A, &lda, &ione, &min_mn, ipiv, &ione);
// copy LU to L and U, and set diagonal to 1
lapackf77_slacpy( MagmaLowerStr, &M, &min_mn, LU, &lda, L, &M );
lapackf77_slacpy( MagmaUpperStr, &min_mn, &N, LU, &lda, U, &min_mn );
for (j=0; j < min_mn; j++)
L[j+j*M] = MAGMA_S_MAKE( 1., 0. );
matnorm = lapackf77_slange("f", &M, &N, A, &lda, work);
blasf77_sgemm("N", "N", &M, &N, &min_mn,
&alpha, L, &M, U, &min_mn, &beta, LU, &lda);
for( j = 0; j < N; j++ ) {
for( i = 0; i < M; i++ ) {
LU[i+j*lda] = MAGMA_S_SUB( LU[i+j*lda], A[i+j*lda] );
}
}
residual = lapackf77_slange("f", &M, &N, LU, &lda, work);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( L );
TESTING_FREE_CPU( U );
return residual / (matnorm * N);
}
float get_LU_error(magma_int_t M, magma_int_t N,
float *A, magma_int_t lda,
float *LU, magma_int_t *IPIV)
{
magma_int_t min_mn = min(M,N);
magma_int_t ione = 1;
magma_int_t i, j;
float alpha = MAGMA_S_ONE;
float beta = MAGMA_S_ZERO;
float *L, *U;
float work[1], matnorm, residual;
TESTING_MALLOC( L, float, M*min_mn);
TESTING_MALLOC( U, float, min_mn*N);
memset( L, 0, M*min_mn*sizeof(float) );
memset( U, 0, min_mn*N*sizeof(float) );
lapackf77_slaswp( &N, A, &lda, &ione, &min_mn, IPIV, &ione);
lapackf77_slacpy( MagmaLowerStr, &M, &min_mn, LU, &lda, L, &M );
lapackf77_slacpy( MagmaUpperStr, &min_mn, &N, LU, &lda, U, &min_mn );
for(j=0; j<min_mn; j++)
L[j+j*M] = MAGMA_S_MAKE( 1., 0. );
matnorm = lapackf77_slange("f", &M, &N, A, &lda, work);
blasf77_sgemm("N", "N", &M, &N, &min_mn,
&alpha, L, &M, U, &min_mn, &beta, LU, &lda);
for( j = 0; j < N; j++ ) {
for( i = 0; i < M; i++ ) {
LU[i+j*lda] = MAGMA_S_SUB( LU[i+j*lda], A[i+j*lda] );
}
}
residual = lapackf77_slange("f", &M, &N, LU, &lda, work);
TESTING_FREE(L);
TESTING_FREE(U);
return residual / (matnorm * N);
}
/**
Purpose
-------
Solves a system of linear equations
A * X = B, A**T * X = B, or A**T * X = B
with a general N-by-N matrix A using the LU factorization computed by SGETRF_GPU.
Arguments
---------
@param[in]
trans magma_trans_t
Specifies the form of the system of equations:
- = MagmaNoTrans: A * X = B (No transpose)
- = MagmaTrans: A**T * X = B (Transpose)
- = MagmaTrans: A**T * X = B (Conjugate transpose)
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in]
nrhs INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
@param[in]
dA REAL array on the GPU, dimension (LDA,N)
The factors L and U from the factorization A = P*L*U as computed
by SGETRF_GPU.
@param[in]
ldda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[in]
ipiv INTEGER array, dimension (N)
The pivot indices from SGETRF; for 1 <= i <= N, row i of the
matrix was interchanged with row IPIV(i).
@param[in,out]
dB REAL array on the GPU, dimension (LDB,NRHS)
On entry, the right hand side matrix B.
On exit, the solution matrix X.
@param[in]
lddb INTEGER
The leading dimension of the array B. LDB >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrs_gpu(magma_trans_t trans, magma_int_t n, magma_int_t nrhs,
float *dA, magma_int_t ldda,
magma_int_t *ipiv,
float *dB, magma_int_t lddb,
magma_int_t *info)
{
float c_one = MAGMA_S_ONE;
float *work = NULL;
int notran = (trans == MagmaNoTrans);
magma_int_t i1, i2, inc;
*info = 0;
if ( (! notran) &&
(trans != MagmaTrans) &&
(trans != MagmaTrans) ) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (nrhs < 0) {
*info = -3;
} else if (ldda < max(1,n)) {
*info = -5;
} else if (lddb < max(1,n)) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (n == 0 || nrhs == 0) {
return *info;
}
magma_smalloc_cpu( &work, n * nrhs );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
i1 = 1;
i2 = n;
if (notran) {
inc = 1;
/* Solve A * X = B. */
magma_sgetmatrix( n, nrhs, dB, lddb, work, n );
lapackf77_slaswp(&nrhs, work, &n, &i1, &i2, ipiv, &inc);
magma_ssetmatrix( n, nrhs, work, n, dB, lddb );
if ( nrhs == 1) {
magma_strsv(MagmaLower, MagmaNoTrans, MagmaUnit, n, dA, ldda, dB, 1 );
magma_strsv(MagmaUpper, MagmaNoTrans, MagmaNonUnit, n, dA, ldda, dB, 1 );
} else {
magma_strsm(MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
magma_strsm(MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
}
} else {
inc = -1;
/* Solve A**T * X = B or A**T * X = B. */
if ( nrhs == 1) {
magma_strsv(MagmaUpper, trans, MagmaNonUnit, n, dA, ldda, dB, 1 );
magma_strsv(MagmaLower, trans, MagmaUnit, n, dA, ldda, dB, 1 );
} else {
magma_strsm(MagmaLeft, MagmaUpper, trans, MagmaNonUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
magma_strsm(MagmaLeft, MagmaLower, trans, MagmaUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
}
magma_sgetmatrix( n, nrhs, dB, lddb, work, n );
lapackf77_slaswp(&nrhs, work, &n, &i1, &i2, ipiv, &inc);
magma_ssetmatrix( n, nrhs, work, n, dB, lddb );
}
magma_free_cpu(work);
return *info;
}
extern "C" magma_int_t
magma_sgetrs_gpu(char trans, magma_int_t n, magma_int_t nrhs,
float *dA, magma_int_t ldda,
magma_int_t *ipiv,
float *dB, magma_int_t lddb,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
Solves a system of linear equations
A * X = B or A' * X = B
with a general N-by-N matrix A using the LU factorization computed by SGETRF_GPU.
Arguments
=========
TRANS (input) CHARACTER*1
Specifies the form of the system of equations:
= 'N': A * X = B (No transpose)
= 'T': A'* X = B (Transpose)
= 'C': A'* X = B (Conjugate transpose = Transpose)
N (input) INTEGER
The order of the matrix A. N >= 0.
NRHS (input) INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
A (input) REAL array on the GPU, dimension (LDA,N)
The factors L and U from the factorization A = P*L*U as computed
by SGETRF_GPU.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,N).
IPIV (input) INTEGER array, dimension (N)
The pivot indices from SGETRF; for 1<=i<=N, row i of the
matrix was interchanged with row IPIV(i).
B (input/output) REAL array on the GPU, dimension (LDB,NRHS)
On entry, the right hand side matrix B.
On exit, the solution matrix X.
LDB (input) INTEGER
The leading dimension of the array B. LDB >= max(1,N).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
HWORK (workspace) REAL array, dimension N*NRHS
===================================================================== */
float c_one = MAGMA_S_ONE;
float *work = NULL;
char trans_[2] = {trans, 0};
int notran = lapackf77_lsame(trans_, "N");
magma_int_t i1, i2, inc;
*info = 0;
if ( (! notran) &&
(! lapackf77_lsame(trans_, "T")) &&
(! lapackf77_lsame(trans_, "C")) ) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (nrhs < 0) {
*info = -3;
} else if (ldda < max(1,n)) {
*info = -5;
} else if (lddb < max(1,n)) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (n == 0 || nrhs == 0) {
return *info;
}
magma_smalloc_cpu( &work, n * nrhs );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
i1 = 1;
i2 = n;
if (notran) {
inc = 1;
/* Solve A * X = B. */
magma_sgetmatrix( n, nrhs, dB, lddb, work, n );
lapackf77_slaswp(&nrhs, work, &n, &i1, &i2, ipiv, &inc);
magma_ssetmatrix( n, nrhs, work, n, dB, lddb );
if ( nrhs == 1) {
magma_strsv(MagmaLower, MagmaNoTrans, MagmaUnit, n, dA, ldda, dB, 1 );
magma_strsv(MagmaUpper, MagmaNoTrans, MagmaNonUnit, n, dA, ldda, dB, 1 );
} else {
magma_strsm(MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
magma_strsm(MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
}
} else {
inc = -1;
/* Solve A' * X = B. */
if ( nrhs == 1) {
magma_strsv(MagmaUpper, trans, MagmaNonUnit, n, dA, ldda, dB, 1 );
magma_strsv(MagmaLower, trans, MagmaUnit, n, dA, ldda, dB, 1 );
} else {
magma_strsm(MagmaLeft, MagmaUpper, trans, MagmaNonUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
magma_strsm(MagmaLeft, MagmaLower, trans, MagmaUnit, n, nrhs, c_one, dA, ldda, dB, lddb );
}
magma_sgetmatrix( n, nrhs, dB, lddb, work, n );
lapackf77_slaswp(&nrhs, work, &n, &i1, &i2, ipiv, &inc);
magma_ssetmatrix( n, nrhs, work, n, dB, lddb );
}
magma_free_cpu(work);
return *info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sswap, sswapblk, spermute, slaswp, slaswpx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
float *h_A1, *h_A2;
float *d_A1, *d_A2;
float *h_R1, *h_R2;
// row-major and column-major performance
real_Double_t row_perf0, col_perf0;
real_Double_t row_perf1, col_perf1;
real_Double_t row_perf2, col_perf2;
real_Double_t row_perf3;
real_Double_t row_perf4;
real_Double_t row_perf5, col_perf5;
real_Double_t row_perf6, col_perf6;
real_Double_t row_perf7;
real_Double_t cpu_perf;
real_Double_t time, gbytes;
magma_int_t N, lda, ldda, nb, j;
magma_int_t ione = 1;
magma_int_t *ipiv, *ipiv2;
magma_int_t *d_ipiv;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_queue_t queue = 0;
printf(" cublasSswap sswap sswapblk slaswp spermute slaswp2 slaswpx scopymatrix CPU (all in )\n");
printf(" N nb row-maj/col-maj row-maj/col-maj row-maj/col-maj row-maj row-maj row-maj row-maj/col-maj row-blk/col-blk slaswp (GByte/s)\n");
printf("==================================================================================================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
// each test is assigned one bit in the check bitmask, bit=1 is failure.
// shift keeps track of which bit is for current test
int shift = 1;
int check = 0;
N = opts.nsize[i];
lda = N;
ldda = ((N+31)/32)*32;
nb = (opts.nb > 0 ? opts.nb : magma_get_sgetrf_nb( N ));
// for each swap, does 2N loads and 2N stores
gbytes = sizeof(float) * 4.*N*nb / 1e9;
TESTING_MALLOC_PIN( h_A1, float, lda*N );
TESTING_MALLOC_PIN( h_A2, float, lda*N );
TESTING_MALLOC_PIN( h_R1, float, lda*N );
TESTING_MALLOC_PIN( h_R2, float, lda*N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, nb );
TESTING_MALLOC_CPU( ipiv2, magma_int_t, nb );
TESTING_MALLOC_DEV( d_ipiv, magma_int_t, nb );
TESTING_MALLOC_DEV( d_A1, float, ldda*N );
TESTING_MALLOC_DEV( d_A2, float, ldda*N );
for( j=0; j < nb; j++ ) {
ipiv[j] = (magma_int_t) ((rand()*1.*N) / (RAND_MAX * 1.)) + 1;
}
/* =====================================================================
* cublasSswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
cublasSswap( N, d_A1+ldda*j, 1, d_A2+ldda*(ipiv[j]-1), 1);
}
}
time = magma_sync_wtime( queue ) - time;
row_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
cublasSswap( N, d_A1+j, ldda, d_A2+ipiv[j]-1, ldda);
}
}
time = magma_sync_wtime( queue ) - time;
col_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* sswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_sswap( N, d_A1+ldda*j, 1, d_A2+ldda*(ipiv[j]-1), 1);
}
}
time = magma_sync_wtime( queue ) - time;
row_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_sswap( N, d_A1+j, ldda, d_A2+ipiv[j]-1, ldda );
}
}
time = magma_sync_wtime( queue ) - time;
col_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* sswapblk, blocked version (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
magmablas_sswapblk( 'R', N, d_A1, ldda, d_A2, ldda, 1, nb, ipiv, 1, 0);
time = magma_sync_wtime( queue ) - time;
row_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
magmablas_sswapblk( 'C', N, d_A1, ldda, d_A2, ldda, 1, nb, ipiv, 1, 0);
time = magma_sync_wtime( queue ) - time;
col_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* spermute_long (1 matrix)
*/
/* Row Major */
memcpy( ipiv2, ipiv, nb*sizeof(magma_int_t) ); // spermute updates ipiv2
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_spermute_long2( N, d_A1, ldda, ipiv2, nb, 0 );
time = magma_sync_wtime( queue ) - time;
row_perf3 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style slaswp (1 matrix)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_slaswp( N, d_A1, ldda, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
row_perf4 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style slaswp (1 matrix) - d_ipiv on GPU
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magma_setvector( nb, sizeof(magma_int_t), ipiv, 1, d_ipiv, 1 );
magmablas_slaswp2( N, d_A1, ldda, 1, nb, d_ipiv );
time = magma_sync_wtime( queue ) - time;
row_perf7 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* LAPACK-style slaswpx (extended for row- and col-major) (1 matrix)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_slaswpx( N, d_A1, ldda, 1, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
row_perf5 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* Col Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_slaswpx( N, d_A1, 1, ldda, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
col_perf5 = gbytes / time;
time = magma_wtime();
lapackf77_slaswp( &N, h_A1, &lda, &ione, &nb, ipiv, &ione);
time = magma_wtime() - time;
cpu_perf = gbytes / time;
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* =====================================================================
* Copy matrix.
*/
time = magma_sync_wtime( queue );
magma_scopymatrix( N, nb, d_A1, ldda, d_A2, ldda );
time = magma_sync_wtime( queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
col_perf6 = 0.5 * gbytes / time;
time = magma_sync_wtime( queue );
magma_scopymatrix( nb, N, d_A1, ldda, d_A2, ldda );
time = magma_sync_wtime( queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
row_perf6 = 0.5 * gbytes / time;
printf("%5d %3d %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c %6.2f%c %6.2f%c %6.2f%c/ %6.2f%c %6.2f / %6.2f %6.2f %10s\n",
(int) N, (int) nb,
row_perf0, ((check & 0x001) != 0 ? '*' : ' '),
col_perf0, ((check & 0x002) != 0 ? '*' : ' '),
row_perf1, ((check & 0x004) != 0 ? '*' : ' '),
col_perf1, ((check & 0x008) != 0 ? '*' : ' '),
row_perf2, ((check & 0x010) != 0 ? '*' : ' '),
col_perf2, ((check & 0x020) != 0 ? '*' : ' '),
row_perf3, ((check & 0x040) != 0 ? '*' : ' '),
row_perf4, ((check & 0x080) != 0 ? '*' : ' '),
row_perf7, ((check & 0x100) != 0 ? '*' : ' '),
row_perf5, ((check & 0x200) != 0 ? '*' : ' '),
col_perf5, ((check & 0x400) != 0 ? '*' : ' '),
row_perf6,
col_perf6,
cpu_perf,
(check == 0 ? "ok" : "* failures") );
TESTING_FREE_PIN( h_A1 );
TESTING_FREE_PIN( h_A2 );
TESTING_FREE_PIN( h_R1 );
TESTING_FREE_PIN( h_R2 );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( ipiv2 );
TESTING_FREE_DEV( d_ipiv );
TESTING_FREE_DEV( d_A1 );
TESTING_FREE_DEV( d_A2 );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
/**
Purpose
-------
SGETRS solves a system of linear equations
A * X = B,
A**T * X = B, or
A**H * X = B
with a general N-by-N matrix A using the LU factorization computed by SGETRF_GPU.
Arguments
---------
@param[in]
trans magma_trans_t
Specifies the form of the system of equations:
- = MagmaNoTrans: A * X = B (No transpose)
- = MagmaTrans: A**T * X = B (Transpose)
- = MagmaConjTrans: A**H * X = B (Conjugate transpose)
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in]
nrhs INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
@param[in]
dA REAL array on the GPU, dimension (LDDA,N)
The factors L and U from the factorization A = P*L*U as computed
by SGETRF_GPU.
@param[in]
ldda INTEGER
The leading dimension of the array A. LDDA >= max(1,N).
@param[in]
ipiv INTEGER array, dimension (N)
The pivot indices from SGETRF; for 1 <= i <= N, row i of the
matrix was interchanged with row IPIV(i).
@param[in,out]
dB REAL array on the GPU, dimension (LDDB,NRHS)
On entry, the right hand side matrix B.
On exit, the solution matrix X.
@param[in]
lddb INTEGER
The leading dimension of the array B. LDDB >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrs_gpu(
magma_trans_t trans, magma_int_t n, magma_int_t nrhs,
magmaFloat_ptr dA, magma_int_t ldda, magma_int_t *ipiv,
magmaFloat_ptr dB, magma_int_t lddb,
magma_int_t *info)
{
// Constants
const float c_one = MAGMA_S_ONE;
// Local variables
float *work = NULL;
bool notran = (trans == MagmaNoTrans);
magma_int_t i1, i2, inc;
*info = 0;
if ( (! notran) &&
(trans != MagmaTrans) &&
(trans != MagmaConjTrans) ) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (nrhs < 0) {
*info = -3;
} else if (ldda < max(1,n)) {
*info = -5;
} else if (lddb < max(1,n)) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (n == 0 || nrhs == 0) {
return *info;
}
magma_smalloc_cpu( &work, n * nrhs );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_queue_t queue = NULL;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
i1 = 1;
i2 = n;
if (notran) {
inc = 1;
/* Solve A * X = B. */
magma_sgetmatrix( n, nrhs, dB, lddb, work, n, queue );
lapackf77_slaswp( &nrhs, work, &n, &i1, &i2, ipiv, &inc );
magma_ssetmatrix( n, nrhs, work, n, dB, lddb, queue );
if ( nrhs == 1) {
magma_strsv( MagmaLower, MagmaNoTrans, MagmaUnit, n, dA, ldda, dB, 1, queue );
magma_strsv( MagmaUpper, MagmaNoTrans, MagmaNonUnit, n, dA, ldda, dB, 1, queue );
} else {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit, n, nrhs, c_one, dA, ldda, dB, lddb, queue );
magma_strsm( MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit, n, nrhs, c_one, dA, ldda, dB, lddb, queue );
}
} else {
inc = -1;
/* Solve A**T * X = B or A**H * X = B. */
if ( nrhs == 1) {
magma_strsv( MagmaUpper, trans, MagmaNonUnit, n, dA, ldda, dB, 1, queue );
magma_strsv( MagmaLower, trans, MagmaUnit, n, dA, ldda, dB, 1, queue );
} else {
magma_strsm( MagmaLeft, MagmaUpper, trans, MagmaNonUnit, n, nrhs, c_one, dA, ldda, dB, lddb, queue );
magma_strsm( MagmaLeft, MagmaLower, trans, MagmaUnit, n, nrhs, c_one, dA, ldda, dB, lddb, queue );
}
magma_sgetmatrix( n, nrhs, dB, lddb, work, n, queue );
lapackf77_slaswp( &nrhs, work, &n, &i1, &i2, ipiv, &inc );
magma_ssetmatrix( n, nrhs, work, n, dB, lddb, queue );
}
magma_queue_destroy( queue );
magma_free_cpu( work );
return *info;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sswap, sswapblk, slaswp, slaswpx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
float *h_A1, *h_A2;
float *h_R1, *h_R2;
magmaFloat_ptr d_A1, d_A2;
// row-major and column-major performance
real_Double_t row_perf0 = MAGMA_D_NAN, col_perf0 = MAGMA_D_NAN;
real_Double_t row_perf1 = MAGMA_D_NAN, col_perf1 = MAGMA_D_NAN;
real_Double_t row_perf2 = MAGMA_D_NAN, col_perf2 = MAGMA_D_NAN;
real_Double_t row_perf4 = MAGMA_D_NAN;
real_Double_t row_perf5 = MAGMA_D_NAN, col_perf5 = MAGMA_D_NAN;
real_Double_t row_perf6 = MAGMA_D_NAN, col_perf6 = MAGMA_D_NAN;
real_Double_t row_perf7 = MAGMA_D_NAN;
real_Double_t cpu_perf = MAGMA_D_NAN;
real_Double_t time, gbytes;
magma_int_t N, lda, ldda, nb, j;
magma_int_t ione = 1;
magma_int_t *ipiv, *ipiv2;
magmaInt_ptr d_ipiv;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_queue_t queue = 0;
printf(" %8s sswap sswap sswapblk slaswp slaswp2 slaswpx scopymatrix CPU (all in )\n", g_platform_str );
printf(" N nb row-maj/col-maj row-maj/col-maj row-maj/col-maj row-maj row-maj row-maj/col-maj row-blk/col-blk slaswp (GByte/s)\n");
printf("=========================================================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
// For an N x N matrix, swap nb rows or nb columns using various methods.
// Each test is assigned one bit in the 'check' bitmask; bit=1 indicates failure.
// The variable 'shift' keeps track of which bit is for current test
int shift = 1;
int check = 0;
N = opts.nsize[itest];
lda = N;
ldda = ((N+31)/32)*32;
nb = (opts.nb > 0 ? opts.nb : magma_get_sgetrf_nb( N ));
nb = min( N, nb );
// each swap does 2N loads and 2N stores, for nb swaps
gbytes = sizeof(float) * 4.*N*nb / 1e9;
TESTING_MALLOC_PIN( h_A1, float, lda*N );
TESTING_MALLOC_PIN( h_A2, float, lda*N );
TESTING_MALLOC_PIN( h_R1, float, lda*N );
TESTING_MALLOC_PIN( h_R2, float, lda*N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, nb );
TESTING_MALLOC_CPU( ipiv2, magma_int_t, nb );
TESTING_MALLOC_DEV( d_ipiv, magma_int_t, nb );
TESTING_MALLOC_DEV( d_A1, float, ldda*N );
TESTING_MALLOC_DEV( d_A2, float, ldda*N );
// getrf always makes ipiv[j] >= j+1, where ipiv is one based and j is zero based
// some implementations (e.g., MacOS dlaswp) assume this
for( j=0; j < nb; j++ ) {
ipiv[j] = (rand() % (N-j)) + j + 1;
assert( ipiv[j] >= j+1 );
assert( ipiv[j] <= N );
}
/* =====================================================================
* cublas / clBLAS / Xeon Phi sswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
#ifdef HAVE_CUBLAS
cublasSswap( opts.handle, N, d_A1+ldda*j, 1, d_A2+ldda*(ipiv[j]-1), 1 );
#else
magma_sswap( N, d_A1, ldda*j, 1, d_A2, ldda*(ipiv[j]-1), 1, opts.queue );
#endif
}
}
time = magma_sync_wtime( queue ) - time;
row_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
#ifdef HAVE_CUBLAS
cublasSswap( opts.handle, N, d_A1+j, ldda, d_A2+ipiv[j]-1, ldda );
#else
magma_sswap( N, d_A1, j, ldda, d_A2, ipiv[j]-1, ldda, opts.queue );
#endif
}
}
time = magma_sync_wtime( queue ) - time;
col_perf0 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* sswap, row-by-row (2 matrices)
*/
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_sswap( N, d_A1+ldda*j, 1, d_A2+ldda*(ipiv[j]-1), 1);
}
}
time = magma_sync_wtime( queue ) - time;
row_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
magmablas_sswap( N, d_A1+j, ldda, d_A2+ipiv[j]-1, ldda );
}
}
time = magma_sync_wtime( queue ) - time;
col_perf1 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* =====================================================================
* sswapblk, blocked version (2 matrices)
*/
#ifdef HAVE_CUBLAS
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
magmablas_sswapblk( MagmaRowMajor, N, d_A1, ldda, d_A2, ldda, 1, nb, ipiv, 1, 0);
time = magma_sync_wtime( queue ) - time;
row_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A2+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
/* Column Major */
init_matrix( N, N, h_A1, lda, 0 );
init_matrix( N, N, h_A2, lda, 100 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
magma_ssetmatrix( N, N, h_A2, lda, d_A2, ldda );
time = magma_sync_wtime( queue );
magmablas_sswapblk( MagmaColMajor, N, d_A1, ldda, d_A2, ldda, 1, nb, ipiv, 1, 0);
time = magma_sync_wtime( queue ) - time;
col_perf2 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+j, &lda, h_A2+(ipiv[j]-1), &lda);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
magma_sgetmatrix( N, N, d_A2, ldda, h_R2, lda );
check += (diff_matrix( N, N, h_A1, lda, h_R1, lda ) ||
diff_matrix( N, N, h_A2, lda, h_R2, lda ))*shift;
shift *= 2;
#endif
/* =====================================================================
* LAPACK-style slaswp (1 matrix)
*/
#ifdef HAVE_CUBLAS
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_slaswp( N, d_A1, ldda, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
row_perf4 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
#endif
/* =====================================================================
* LAPACK-style slaswp (1 matrix) - d_ipiv on GPU
*/
#ifdef HAVE_CUBLAS
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magma_setvector( nb, sizeof(magma_int_t), ipiv, 1, d_ipiv, 1 );
magmablas_slaswp2( N, d_A1, ldda, 1, nb, d_ipiv, 1 );
time = magma_sync_wtime( queue ) - time;
row_perf7 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
#endif
/* =====================================================================
* LAPACK-style slaswpx (extended for row- and col-major) (1 matrix)
*/
#ifdef HAVE_CUBLAS
/* Row Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_slaswpx( N, d_A1, ldda, 1, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
row_perf5 = gbytes / time;
for( j=0; j < nb; j++) {
if ( j != (ipiv[j]-1)) {
blasf77_sswap( &N, h_A1+lda*j, &ione, h_A1+lda*(ipiv[j]-1), &ione);
}
}
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
/* Col Major */
init_matrix( N, N, h_A1, lda, 0 );
magma_ssetmatrix( N, N, h_A1, lda, d_A1, ldda );
time = magma_sync_wtime( queue );
magmablas_slaswpx( N, d_A1, 1, ldda, 1, nb, ipiv, 1);
time = magma_sync_wtime( queue ) - time;
col_perf5 = gbytes / time;
#endif
/* LAPACK swap on CPU for comparison */
time = magma_wtime();
lapackf77_slaswp( &N, h_A1, &lda, &ione, &nb, ipiv, &ione);
time = magma_wtime() - time;
cpu_perf = gbytes / time;
#ifdef HAVE_CUBLAS
magma_sgetmatrix( N, N, d_A1, ldda, h_R1, lda );
check += diff_matrix( N, N, h_A1, lda, h_R1, lda )*shift;
shift *= 2;
#endif
/* =====================================================================
* Copy matrix.
*/
time = magma_sync_wtime( queue );
magma_scopymatrix( N, nb, d_A1, ldda, d_A2, ldda );
time = magma_sync_wtime( queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
col_perf6 = 0.5 * gbytes / time;
time = magma_sync_wtime( queue );
magma_scopymatrix( nb, N, d_A1, ldda, d_A2, ldda );
time = magma_sync_wtime( queue ) - time;
// copy reads 1 matrix and writes 1 matrix, so has half gbytes of swap
row_perf6 = 0.5 * gbytes / time;
printf("%5d %3d %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c/ %6.2f%c %6.2f%c %6.2f%c %6.2f%c/ %6.2f%c %6.2f / %6.2f %6.2f %10s\n",
(int) N, (int) nb,
row_perf0, ((check & 0x001) != 0 ? '*' : ' '),
col_perf0, ((check & 0x002) != 0 ? '*' : ' '),
row_perf1, ((check & 0x004) != 0 ? '*' : ' '),
col_perf1, ((check & 0x008) != 0 ? '*' : ' '),
row_perf2, ((check & 0x010) != 0 ? '*' : ' '),
col_perf2, ((check & 0x020) != 0 ? '*' : ' '),
row_perf4, ((check & 0x040) != 0 ? '*' : ' '),
row_perf7, ((check & 0x080) != 0 ? '*' : ' '),
row_perf5, ((check & 0x100) != 0 ? '*' : ' '),
col_perf5, ((check & 0x200) != 0 ? '*' : ' '),
row_perf6,
col_perf6,
cpu_perf,
(check == 0 ? "ok" : "* failed") );
status += ! (check == 0);
TESTING_FREE_PIN( h_A1 );
TESTING_FREE_PIN( h_A2 );
TESTING_FREE_PIN( h_R1 );
TESTING_FREE_PIN( h_R2 );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( ipiv2 );
TESTING_FREE_DEV( d_ipiv );
TESTING_FREE_DEV( d_A1 );
TESTING_FREE_DEV( d_A2 );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
