/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
/**
@deprecated
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]
dA REAL array, dimension (LDDA,N), on the GPU.
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]
ldda INTEGER
The leading dimension of the array A. LDDA >= 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]
dauxv REAL array, dimension (NB), on the GPU
Auxiliary vector.
@param[in,out]
dF REAL array, dimension (LDDF,NB), on the GPU
Matrix F' = L*Y'*A.
@param[in]
lddf INTEGER
The leading dimension of the array F. LDDF >= max(1,N).
@ingroup magma_sgeqp3_aux
********************************************************************/
extern "C" magma_int_t
magma_slaqps_gpu(
magma_int_t m, magma_int_t n, magma_int_t offset,
magma_int_t nb, magma_int_t *kb,
magmaFloat_ptr dA, magma_int_t ldda,
magma_int_t *jpvt, float *tau,
float *vn1, float *vn2,
magmaFloat_ptr dauxv,
magmaFloat_ptr dF, magma_int_t lddf)
{
#define dA(i, j) (dA + (i) + (j)*(ldda))
#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 z__1;
magma_int_t k, rk;
magmaFloat_ptr dAks;
float tauk = MAGMA_S_ZERO;
magma_int_t pvt;
float tol3z;
magma_int_t itemp;
float lsticc;
magmaFloat_ptr dlsticcs;
magma_smalloc( &dlsticcs, 1+256*(n+255)/256 );
tol3z = magma_ssqrt( lapackf77_slamch("Epsilon"));
lsticc = 0;
k = 0;
magma_smalloc( &dAks, nb );
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
while( k < nb && lsticc == 0 ) {
rk = offset + k;
/* Determine ith pivot column and swap if necessary */
// subtract 1 from Fortran/CUBLAS isamax; pvt, k are 0-based.
pvt = k + magma_isamax( n-k, &vn1[k], ione, queue ) - 1;
if (pvt != k) {
/* F gets swapped so F must be sent at the end to GPU */
i__1 = k;
magmablas_sswap( m, dA(0, pvt), ione, dA(0, k), ione, queue );
magmablas_sswap( i__1, dF(pvt, 0), lddf, dF(k, 0), lddf, queue );
itemp = jpvt[pvt];
jpvt[pvt] = jpvt[k];
jpvt[k] = itemp;
magma_sswap( 2, &vn1[pvt], n+offset, &vn1[k], n+offset, queue );
}
/* 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) {
//#define RIGHT_UPDATE
#ifdef RIGHT_UPDATE
i__1 = m - offset - nb;
i__2 = k;
magma_sgemv( MagmaNoTrans, i__1, i__2,
c_neg_one, A(offset+nb, 0), lda,
F(k, 0), ldf,
c_one, A(offset+nb, k), ione, queue );
#else
i__1 = m - rk;
i__2 = k;
magma_sgemv( MagmaNoTrans, i__1, i__2,
c_neg_one, dA(rk, 0), ldda,
dF(k, 0), lddf,
c_one, dA(rk, k), ione, queue );
#endif
}
/* Generate elementary reflector H(k). */
magma_slarfg_gpu( m-rk, dA(rk, k), dA(rk + 1, k), &tau[k], &vn1[k], &dAks[k], queue );
/* needed to avoid the race condition */
if (k == 0) magma_ssetvector( 1, &c_one, 1, dA(rk, k), 1, queue );
else magma_scopymatrix( 1, 1, dA(offset, 0), 1, dA(rk, k), 1, queue );
/* 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 || k > 0) magma_sgetvector( 1, &tau[k], 1, &tauk, 1, queue );
if (k < n-1) {
i__1 = m - rk;
i__2 = n - k - 1;
/* Multiply on GPU */
magma_sgemv( MagmaConjTrans, m-rk, n-k-1,
tauk, dA( rk, k+1 ), ldda,
dA( rk, k ), 1,
c_zero, dF( k+1, k ), 1, queue );
}
/* 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).
F(1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K)
:= tau(K)(A(RK:M,K+1:N)' - F(1:N,1:K-1)*A(RK:M,1:K-1)') A(RK:M,K)
so, F is (updated A)*V */
if (k > 0) {
z__1 = MAGMA_S_NEGATE( tauk );
#ifdef RIGHT_UPDATE
i__1 = m - offset - nb;
i__2 = k;
magma_sgemv( MagmaConjTrans, i__1, i__2,
z__1, dA(offset+nb, 0), lda,
dA(offset+nb, k), ione,
c_zero, dauxv, ione, queue );
i__1 = k;
magma_sgemv( MagmaNoTrans, n-k-1, i__1,
c_one, F(k+1,0), ldf,
dauxv, ione,
c_one, F(k+1,k), ione, queue );
#else
i__1 = m - rk;
i__2 = k;
magma_sgemv( MagmaConjTrans, i__1, i__2,
z__1, dA(rk, 0), ldda,
dA(rk, k), ione,
c_zero, dauxv, ione, queue );
/* I think we only need stricly lower-triangular part :) */
magma_sgemv( MagmaNoTrans, n-k-1, i__2,
c_one, dF(k+1,0), lddf,
dauxv, ione,
c_one, dF(k+1,k), ione, queue );
#endif
}
/* 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;
#ifdef RIGHT_UPDATE
/* right-looking update of rows, */
magma_sgemm( MagmaNoTrans, MagmaConjTrans, nb-k, i__1, ione,
c_neg_one, dA(rk, k ), ldda,
dF(k+1, k ), lddf,
c_one, dA(rk, k+1), ldda, queue );
#else
/* left-looking update of rows, *
* since F=A'v with original A, so no right-looking */
magma_sgemm( MagmaNoTrans, MagmaConjTrans, ione, i__1, i__2,
c_neg_one, dA(rk, 0 ), ldda,
dF(k+1,0 ), lddf,
c_one, dA(rk, k+1), ldda, queue );
#endif
}
/* Update partial column norms. */
if (rk < min(m, n+offset)-1 ) {
magmablas_snrm2_row_check_adjust( n-k-1, tol3z, &vn1[k+1], &vn2[k+1],
dA(rk,k+1), ldda, dlsticcs, queue );
//magma_device_sync();
magma_sgetvector( 1, &dlsticcs[0], 1, &lsticc, 1, queue );
}
++k;
}
magma_scopymatrix( 1, k, dAks, 1, dA(offset, 0), ldda+1, queue );
// 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;
magma_sgemm( MagmaNoTrans, MagmaConjTrans, i__1, i__2, *kb,
c_neg_one, dA(rk+1, 0 ), ldda,
dF(*kb, 0 ), lddf,
c_one, dA(rk+1, *kb), ldda, queue );
}
/* Recomputation of difficult columns. */
if ( lsticc > 0 ) {
// printf( " -- recompute dnorms --\n" );
magmablas_snrm2_check( m-rk-1, n-*kb, dA(rk+1,*kb), ldda,
&vn1[*kb], dlsticcs, queue );
magma_scopymatrix( n-*kb, 1, &vn1[*kb], *kb, &vn2[*kb], *kb, queue );
}
magma_free( dAks );
magma_free( dlsticcs );
magma_queue_destroy( queue );
return MAGMA_SUCCESS;
} /* magma_slaqps */
/* ////////////////////////////////////////////////////////////////////////////
-- 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;
}
