// On input, A and ipiv is LU factorization of A. On output, A is overwritten.
// Requires m == n.
// Uses init_matrix() to re-generate original A as needed.
// Generates random RHS b and solves Ax=b.
// Returns residual, |Ax - b| / (n |A| |x|).
float get_residual(
magma_opts &opts,
magma_int_t m, magma_int_t n,
magmaFloatComplex *A, magma_int_t lda,
magma_int_t *ipiv )
{
if ( m != n ) {
printf( "\nERROR: residual check defined only for square matrices\n" );
return -1;
}
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magma_int_t ione = 1;
// this seed should be DIFFERENT than used in init_matrix
// (else x is column of A, so residual can be exactly zero)
magma_int_t ISEED[4] = {0,0,0,2};
magma_int_t info = 0;
magmaFloatComplex *x, *b;
// initialize RHS
TESTING_MALLOC_CPU( x, magmaFloatComplex, n );
TESTING_MALLOC_CPU( b, magmaFloatComplex, n );
lapackf77_clarnv( &ione, ISEED, &n, b );
blasf77_ccopy( &n, b, &ione, x, &ione );
// solve Ax = b
lapackf77_cgetrs( "Notrans", &n, &ione, A, &lda, ipiv, x, &n, &info );
if (info != 0) {
printf("lapackf77_cgetrs returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
// reset to original A
init_matrix( opts, m, n, A, lda );
// compute r = Ax - b, saved in b
blasf77_cgemv( "Notrans", &m, &n, &c_one, A, &lda, x, &ione, &c_neg_one, b, &ione );
// compute residual |Ax - b| / (n*|A|*|x|)
float norm_x, norm_A, norm_r, work[1];
norm_A = lapackf77_clange( "F", &m, &n, A, &lda, work );
norm_r = lapackf77_clange( "F", &n, &ione, b, &n, work );
norm_x = lapackf77_clange( "F", &n, &ione, x, &n, work );
//printf( "r=\n" ); magma_cprint( 1, n, b, 1 );
TESTING_FREE_CPU( x );
TESTING_FREE_CPU( b );
//printf( "r=%.2e, A=%.2e, x=%.2e, n=%d\n", norm_r, norm_A, norm_x, n );
return norm_r / (n * norm_A * norm_x);
}
// 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,
magmaFloatComplex *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;
magmaFloatComplex alpha = MAGMA_C_ONE;
magmaFloatComplex beta = MAGMA_C_ZERO;
magmaFloatComplex *A, *L, *U;
float work[1], matnorm, residual;
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( L, magmaFloatComplex, M*min_mn );
TESTING_MALLOC_CPU( U, magmaFloatComplex, min_mn*N );
memset( L, 0, M*min_mn*sizeof(magmaFloatComplex) );
memset( U, 0, min_mn*N*sizeof(magmaFloatComplex) );
// set to original A
init_matrix( opts, M, N, A, lda );
lapackf77_claswp( &N, A, &lda, &ione, &min_mn, ipiv, &ione);
// copy LU to L and U, and set diagonal to 1
lapackf77_clacpy( MagmaLowerStr, &M, &min_mn, LU, &lda, L, &M );
lapackf77_clacpy( MagmaUpperStr, &min_mn, &N, LU, &lda, U, &min_mn );
for (j=0; j < min_mn; j++)
L[j+j*M] = MAGMA_C_MAKE( 1., 0. );
matnorm = lapackf77_clange("f", &M, &N, A, &lda, work);
blasf77_cgemm("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_C_SUB( LU[i+j*lda], A[i+j*lda] );
}
}
residual = lapackf77_clange("f", &M, &N, LU, &lda, work);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( L );
TESTING_FREE_CPU( U );
return residual / (matnorm * N);
}
double get_LU_error(magma_int_t M, magma_int_t N,
magmaDoubleComplex *A, magma_int_t lda,
magmaDoubleComplex *LU, magma_int_t *IPIV)
{
magma_int_t min_mn = min(M,N);
magma_int_t ione = 1;
magma_int_t i, j;
magmaDoubleComplex alpha = MAGMA_Z_ONE;
magmaDoubleComplex beta = MAGMA_Z_ZERO;
magmaDoubleComplex *L, *U;
double work[1], matnorm, residual;
TESTING_MALLOC_CPU( L, magmaDoubleComplex, M*min_mn );
TESTING_MALLOC_CPU( U, magmaDoubleComplex, min_mn*N );
memset( L, 0, M*min_mn*sizeof(magmaDoubleComplex) );
memset( U, 0, min_mn*N*sizeof(magmaDoubleComplex) );
lapackf77_zlaswp( &N, A, &lda, &ione, &min_mn, IPIV, &ione);
lapackf77_zlacpy( MagmaLowerStr, &M, &min_mn, LU, &lda, L, &M );
lapackf77_zlacpy( MagmaUpperStr, &min_mn, &N, LU, &lda, U, &min_mn );
for(j=0; j<min_mn; j++)
L[j+j*M] = MAGMA_Z_MAKE( 1., 0. );
matnorm = lapackf77_zlange("f", &M, &N, A, &lda, work);
blasf77_zgemm("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_Z_SUB( LU[i+j*lda], A[i+j*lda] );
}
}
residual = lapackf77_zlange("f", &M, &N, LU, &lda, work);
TESTING_FREE_CPU( L );
TESTING_FREE_CPU( U );
return residual / (matnorm * N);
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cprint
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magmaFloatComplex *hA, *dA;
//magma_int_t ione = 1;
//magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t M, N, lda, ldda; //size
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = M;
ldda = ((M + 31)/32)*32;
//size = lda*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( hA, magmaFloatComplex, lda *N );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, ldda*N );
//lapackf77_clarnv( &ione, ISEED, &size, hA );
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < M; ++i ) {
hA[i + j*lda] = MAGMA_C_MAKE( i + j*0.01, 0. );
}
}
magma_csetmatrix( M, N, hA, lda, dA, ldda );
printf( "A=" );
magma_cprint( M, N, hA, lda );
printf( "dA=" );
magma_cprint_gpu( M, N, dA, ldda );
TESTING_FREE_CPU( hA );
TESTING_FREE_DEV( dA );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cposv_gpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
float error, Rnorm, Anorm, Xnorm, *work;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_B, *h_X;
magmaFloatComplex_ptr d_A, d_B;
magma_int_t N, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf(" N NRHS CPU Gflop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = ldb = N;
ldda = ((N+31)/32)*32;
lddb = ldda;
gflops = ( FLOPS_CPOTRF( N ) + FLOPS_CPOTRS( N, opts.nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, ldb*opts.nrhs );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, ldb*opts.nrhs );
TESTING_MALLOC_CPU( work, float, N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaFloatComplex, lddb*opts.nrhs );
/* ====================================================================
Initialize the matrix
=================================================================== */
sizeA = lda*N;
sizeB = ldb*opts.nrhs;
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeB, h_B );
magma_cmake_hpd( N, h_A, lda );
magma_csetmatrix( N, N, h_A, N, d_A, ldda );
magma_csetmatrix( N, opts.nrhs, h_B, N, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_cposv_gpu( opts.uplo, N, opts.nrhs, d_A, ldda, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cpotrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Residual
=================================================================== */
magma_cgetmatrix( N, opts.nrhs, d_B, lddb, h_X, ldb );
Anorm = lapackf77_clange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_clange("I", &N, &opts.nrhs, h_X, &ldb, work);
blasf77_cgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &opts.nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb );
Rnorm = lapackf77_clange("I", &N, &opts.nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cposv( lapack_uplo_const(opts.uplo), &N, &opts.nrhs, h_A, &lda, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cposv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) opts.nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) opts.nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgemm_batched
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cpu_perf, cpu_time;
float magma_error, magma_err, Ynorm, work[1];
magma_int_t M, N, Xm, Ym, lda, ldda;
magma_int_t sizeA, sizeX, sizeY;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_int_t batchCount;
magmaFloatComplex *h_A, *h_X, *h_Y, *h_Ymagma;
magmaFloatComplex *d_A, *d_X, *d_Y;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex alpha = MAGMA_C_MAKE( 0.29, -0.86 );
magmaFloatComplex beta = MAGMA_C_MAKE( -0.48, 0.38 );
magmaFloatComplex **A_array = NULL;
magmaFloatComplex **X_array = NULL;
magmaFloatComplex **Y_array = NULL;
magma_opts opts;
parse_opts( argc, argv, &opts );
batchCount = opts.batchcount;
opts.lapack |= opts.check;
//float tol = opts.tolerance * lapackf77_slamch("E");
printf("trans = %s\n", lapack_trans_const(opts.transA) );
printf("BatchCount M N MAGMA Gflop/s (ms) CPU Gflop/s (ms) MAGMA 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];
lda = ((M+31)/32)*32;
gflops = FLOPS_CGEMV( M, N ) / 1e9 * batchCount;
if ( opts.transA == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
sizeA = lda*N*batchCount;
sizeX = incx*Xm*batchCount;
sizeY = incy*Ym*batchCount;
ldda = ((lda+31)/32)*32;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, sizeA );
TESTING_MALLOC_CPU( h_X, magmaFloatComplex, sizeX );
TESTING_MALLOC_CPU( h_Y, magmaFloatComplex, sizeY );
TESTING_MALLOC_CPU( h_Ymagma, magmaFloatComplex, sizeY );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N*batchCount );
TESTING_MALLOC_DEV( d_X, magmaFloatComplex, sizeX );
TESTING_MALLOC_DEV( d_Y, magmaFloatComplex, sizeY );
magma_malloc((void**)&A_array, batchCount * sizeof(*A_array));
magma_malloc((void**)&X_array, batchCount * sizeof(*X_array));
magma_malloc((void**)&Y_array, batchCount * sizeof(*Y_array));
/* Initialize the matrices */
lapackf77_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_clarnv( &ione, ISEED, &sizeX, h_X );
lapackf77_clarnv( &ione, ISEED, &sizeY, h_Y );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_csetmatrix( M, N*batchCount, h_A, lda, d_A, ldda );
magma_csetvector( Xm*batchCount, h_X, incx, d_X, incx );
magma_csetvector( Ym*batchCount, h_Y, incy, d_Y, incy );
cset_pointer(A_array, d_A, ldda, 0, 0, ldda*N, batchCount, magma_stream);
cset_pointer(X_array, d_X, 1, 0, 0, incx*Xm, batchCount, magma_stream);
cset_pointer(Y_array, d_Y, 1, 0, 0, incy*Ym, batchCount, magma_stream);
magma_time = magma_sync_wtime( NULL );
magmablas_cgemv_batched(opts.transA, M, N,
alpha, A_array, ldda,
X_array, incx,
beta, Y_array, incy, batchCount, magma_stream);
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_cgetvector( Ym*batchCount, d_Y, incy, h_Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
for(int i=0; i<batchCount; i++)
{
blasf77_cgemv(
lapack_trans_const(opts.transA),
&M, &N,
&alpha, h_A + i*lda*N, &lda,
h_X + i*Xm, &incx,
&beta, h_Y + i*Ym, &incy );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
magma_error = 0.0;
for(int s=0; s<batchCount; s++)
{
Ynorm = lapackf77_clange( "M", &M, &ione, h_Y + s*Ym, &incy, work );
blasf77_caxpy( &Ym, &c_neg_one, h_Y + s*Ym, &ione, h_Ymagma + s*Ym, &ione );
magma_err = lapackf77_clange( "M", &M, &ione, h_Ymagma + s*Ym, &incy, work ) / Ynorm;
if ( isnan(magma_err) || isinf(magma_err) ) {
magma_error = magma_err;
break;
}
magma_error = max(fabs(magma_err), magma_error);
}
printf("%10d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e \n",
(int) batchCount, (int) M, (int) N,
magma_perf, 1000.*magma_time,
cpu_perf, 1000.*cpu_time,
magma_error);
}
else {
printf("%10d %5d %5d %7.2f (%7.2f) --- ( --- ) ---\n",
(int) batchCount, (int) M, (int) N,
magma_perf, 1000.*magma_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_Y );
TESTING_FREE_CPU( h_Ymagma );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_X );
TESTING_FREE_DEV( d_Y );
TESTING_FREE_DEV( A_array );
TESTING_FREE_DEV( X_array );
TESTING_FREE_DEV( Y_array );
fflush( stdout);
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgetri_batched
*/
int main( int argc, char** argv)
{
TESTING_INIT();
// constants
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A, *h_Ainv, *h_R, *work;
magmaDoubleComplex_ptr d_A, d_invA;
magmaDoubleComplex_ptr *dA_array;
magmaDoubleComplex_ptr *dinvA_array;
magma_int_t **dipiv_array;
magma_int_t *dinfo_array;
magma_int_t *ipiv, *cpu_info;
magma_int_t *d_ipiv, *d_info;
magma_int_t N, n2, lda, ldda, info, info1, info2, lwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex tmp;
double error, rwork[1];
magma_int_t columns;
magma_int_t status = 0;
magma_opts opts( MagmaOptsBatched );
opts.parse_opts( argc, argv );
magma_int_t batchCount = opts.batchcount;
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% batchCount N CPU Gflop/s (ms) GPU Gflop/s (ms) ||I - A*A^{-1}||_1 / (N*cond(A))\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N * batchCount;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
// This is the correct flops but since this getri_batched is based on
// 2 trsm = getrs and to know the real flops I am using the getrs one
//gflops = (FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRI( N ))/ 1e9 * batchCount;
gflops = (FLOPS_ZGETRF( N, N ) + FLOPS_ZGETRS( N, N ))/ 1e9 * batchCount;
// query for workspace size
lwork = -1;
lapackf77_zgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0) {
printf("lapackf77_zgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = magma_int_t( MAGMA_Z_REAL( tmp ));
TESTING_MALLOC_CPU( cpu_info, magma_int_t, batchCount );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N * batchCount );
TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork*batchCount );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_Ainv, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N * batchCount );
TESTING_MALLOC_DEV( d_invA, magmaDoubleComplex, ldda*N * batchCount );
TESTING_MALLOC_DEV( d_ipiv, magma_int_t, N * batchCount );
TESTING_MALLOC_DEV( d_info, magma_int_t, batchCount );
TESTING_MALLOC_DEV( dA_array, magmaDoubleComplex*, batchCount );
TESTING_MALLOC_DEV( dinvA_array, magmaDoubleComplex*, batchCount );
TESTING_MALLOC_DEV( dinfo_array, magma_int_t, batchCount );
TESTING_MALLOC_DEV( dipiv_array, magma_int_t*, batchCount );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
columns = N * batchCount;
lapackf77_zlacpy( MagmaFullStr, &N, &columns, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaFullStr, &N, &columns, h_A, &lda, h_Ainv, &lda );
magma_zsetmatrix( N, columns, h_R, lda, d_A, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zset_pointer( dA_array, d_A, ldda, 0, 0, ldda * N, batchCount, opts.queue );
magma_zset_pointer( dinvA_array, d_invA, ldda, 0, 0, ldda * N, batchCount, opts.queue );
magma_iset_pointer( dipiv_array, d_ipiv, 1, 0, 0, N, batchCount, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
info1 = magma_zgetrf_batched( N, N, dA_array, ldda, dipiv_array, dinfo_array, batchCount, opts.queue);
info2 = magma_zgetri_outofplace_batched( N, dA_array, ldda, dipiv_array, dinvA_array, ldda, dinfo_array, batchCount, opts.queue);
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
// check correctness of results throught "dinfo_magma" and correctness of argument throught "info"
magma_getvector( batchCount, sizeof(magma_int_t), dinfo_array, 1, cpu_info, 1, opts.queue );
for (magma_int_t i=0; i < batchCount; i++)
{
if (cpu_info[i] != 0 ) {
printf("magma_zgetrf_batched matrix %d returned error %d\n", (int) i, (int)cpu_info[i] );
}
}
if (info1 != 0) printf("magma_zgetrf_batched returned argument error %d: %s.\n", (int) info1, magma_strerror( info1 ));
if (info2 != 0) printf("magma_zgetri_batched returned argument error %d: %s.\n", (int) info2, magma_strerror( info2 ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
#if !defined (BATCHED_DISABLE_PARCPU) && defined(_OPENMP)
magma_int_t nthreads = magma_get_lapack_numthreads();
magma_set_lapack_numthreads(1);
magma_set_omp_numthreads(nthreads);
#pragma omp parallel for schedule(dynamic)
#endif
for (int i=0; i < batchCount; i++)
{
magma_int_t locinfo;
lapackf77_zgetrf(&N, &N, h_Ainv + i*lda*N, &lda, ipiv + i*N, &locinfo);
if (locinfo != 0) {
printf("lapackf77_zgetrf returned error %d: %s.\n",
(int) locinfo, magma_strerror( locinfo ));
}
lapackf77_zgetri(&N, h_Ainv + i*lda*N, &lda, ipiv + i*N, work + i*lwork, &lwork, &locinfo );
if (locinfo != 0) {
printf("lapackf77_zgetri returned error %d: %s.\n",
(int) locinfo, magma_strerror( locinfo ));
}
}
#if !defined (BATCHED_DISABLE_PARCPU) && defined(_OPENMP)
magma_set_lapack_numthreads(nthreads);
#endif
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
printf("%10d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) batchCount, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000. );
}
else {
printf("%10d %5d --- ( --- ) %7.2f (%7.2f)",
(int) batchCount, (int) N, gpu_perf, gpu_time*1000. );
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.check ) {
magma_igetvector( N*batchCount, d_ipiv, 1, ipiv, 1, opts.queue );
magma_zgetmatrix( N, N*batchCount, d_invA, ldda, h_Ainv, lda, opts.queue );
error = 0;
for (magma_int_t i=0; i < batchCount; i++)
{
for (magma_int_t k=0; k < N; k++) {
if (ipiv[i*N+k] < 1 || ipiv[i*N+k] > N )
{
printf("error for matrix %d ipiv @ %d = %d\n", (int) i, (int) k, (int) ipiv[i*N+k]);
error = -1;
}
}
if (error == -1) {
break;
}
// compute 1-norm condition number estimate, following LAPACK's zget03
double normA, normAinv, rcond, err;
normA = lapackf77_zlange( "1", &N, &N, h_A + i*lda*N, &lda, rwork );
normAinv = lapackf77_zlange( "1", &N, &N, h_Ainv + i*lda*N, &lda, rwork );
if ( normA <= 0 || normAinv <= 0 ) {
rcond = 0;
err = 1 / (tol/opts.tolerance); // == 1/eps
}
else {
rcond = (1 / normA) / normAinv;
// R = I
// R -= A*A^{-1}
// err = ||I - A*A^{-1}|| / ( N ||A||*||A^{-1}|| ) = ||R|| * rcond / N, using 1-norm
lapackf77_zlaset( "full", &N, &N, &c_zero, &c_one, h_R + i*lda*N, &lda );
blasf77_zgemm( "no", "no", &N, &N, &N, &c_neg_one,
h_A + i*lda*N, &lda,
h_Ainv + i*lda*N, &lda, &c_one,
h_R + i*lda*N, &lda );
err = lapackf77_zlange( "1", &N, &N, h_R + i*lda*N, &lda, rwork );
err = err * rcond / N;
}
if ( isnan(err) || isinf(err) ) {
error = err;
break;
}
error = max( err, error );
}
bool okay = (error < tol);
status += ! okay;
printf(" %8.2e %s\n", error, (okay ? "ok" : "failed") );
}
else {
printf("\n");
}
TESTING_FREE_CPU( cpu_info );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_Ainv );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_invA );
TESTING_FREE_DEV( d_ipiv );
TESTING_FREE_DEV( d_info );
TESTING_FREE_DEV( dA_array );
TESTING_FREE_DEV( dinvA_array );
TESTING_FREE_DEV( dinfo_array );
TESTING_FREE_DEV( dipiv_array );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgetrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
float error;
magmaFloatComplex *h_A, *h_R;
magmaFloatComplex *d_A;
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, info, min_mn;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf(" M N CPU GFlop/s (ms) GPU GFlop/s (ms) ||PA-LU||/(||A||*N)\n");
printf("=========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = FLOPS_CGETRF( M, N ) / 1e9;
if ( N > 512 ) {
printf( "%5d %5d skipping because cgetf2 does not support N > 512\n", (int) M, (int) N );
continue;
}
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
magma_csetmatrix( M, N, h_R, lda, d_A, ldda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cgetrf(&M, &N, h_A, &lda, ipiv, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_cgetf2_gpu( M, N, d_A, ldda, ipiv, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgetf2_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) M, (int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000. );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) M, (int) N, gpu_perf, gpu_time*1000. );
}
if ( opts.check ) {
magma_cgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_LU_error( M, N, h_R, lda, h_A, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
}
else {
printf(" --- \n");
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cgehrd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaFloatComplex *h_A, *h_R, *h_Q, *h_work, *tau, *twork;
magmaFloatComplex_ptr dT;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
#endif
float eps, result[2];
magma_int_t N, n2, lda, nb, lwork, ltwork, info;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
eps = lapackf77_slamch( "E" );
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) |A-QHQ'|/N|A| |I-QQ'|/N\n");
printf("=========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
nb = magma_get_cgehrd_nb(N);
/* We suppose the magma nb is bigger than lapack nb */
lwork = N*nb;
gflops = FLOPS_CGEHRD( N ) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, N );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, lwork );
TESTING_MALLOC_DEV( dT, magmaFloatComplex, nb*N );
/* Initialize the matrices */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_cgehrd( N, ione, N, h_R, lda, tau, h_work, lwork, dT, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cgehrd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.check ) {
ltwork = 2*(N*N);
TESTING_MALLOC_PIN( h_Q, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( twork, magmaFloatComplex, ltwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_CPU( rwork, float, N );
#endif
lapackf77_clacpy(MagmaUpperLowerStr, &N, &N, h_R, &lda, h_Q, &lda);
for( int j = 0; j < N-1; ++j )
for( int i = j+2; i < N; ++i )
h_R[i+j*lda] = MAGMA_C_ZERO;
magma_cunghr(N, ione, N, h_Q, lda, tau, dT, nb, &info);
if (info != 0) {
printf("magma_cunghr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
exit(1);
}
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_chst01(&N, &ione, &N,
h_A, &lda, h_R, &lda,
h_Q, &lda, twork, <work, rwork, result);
#else
lapackf77_chst01(&N, &ione, &N,
h_A, &lda, h_R, &lda,
h_Q, &lda, twork, <work, result);
#endif
TESTING_FREE_PIN( h_Q );
TESTING_FREE_CPU( twork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_CPU( rwork );
#endif
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_cgehrd(&N, &ione, &N, h_R, &lda, tau, h_work, &lwork, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cgehrd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Print performance and error.
=================================================================== */
if ( opts.lapack ) {
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d --- ( --- ) %7.2f (%7.2f)",
(int) N, gpu_perf, gpu_time );
}
if ( opts.check ) {
printf(" %8.2e %8.2e %s\n",
result[0]*eps, result[1]*eps,
( ( (result[0]*eps < tol) && (result[1]*eps < tol) ) ? "ok" : "failed") );
status += ! (result[0]*eps < tol);
status += ! (result[1]*eps < tol);
}
else {
printf(" --- ---\n");
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( tau );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_DEV( dT );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ctrsm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf=0, cpu_time=0;
float cublas_error, normA, normx, normr, work[1];
magma_int_t N, info;
magma_int_t sizeA;
magma_int_t lda, ldda;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t *ipiv;
magmaFloatComplex *h_A, *h_b, *h_x, *h_xcublas;
magmaFloatComplex *d_A, *d_x;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf("uplo = %c, transA = %c, diag = %c\n", opts.uplo, opts.transA, opts.diag );
printf(" N CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("============================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
gflops = FLOPS_CTRSM(opts.side, N, 1) / 1e9;
lda = N;
ldda = ((lda+31)/32)*32;
sizeA = lda*N;
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, lda*N );
TESTING_MALLOC_CPU( h_b, magmaFloatComplex, N );
TESTING_MALLOC_CPU( h_x, magmaFloatComplex, N );
TESTING_MALLOC_CPU( h_xcublas, magmaFloatComplex, N );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_x, magmaFloatComplex, 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_clarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_cgetrf( &N, &N, h_A, &lda, ipiv, &info );
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < j; ++i ) {
*h_A(i,j) = *h_A(j,i);
}
}
lapackf77_clarnv( &ione, ISEED, &N, h_b );
blasf77_ccopy( &N, h_b, &ione, h_x, &ione );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
magma_csetvector( N, h_x, 1, d_x, 1 );
cublas_time = magma_sync_wtime( NULL );
cublasCtrsv( opts.uplo, opts.transA, opts.diag,
N,
d_A, ldda,
d_x, 1 );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_cgetvector( N, d_x, 1, h_xcublas, 1 );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_ctrsv( &opts.uplo, &opts.transA, &opts.diag,
&N,
h_A, &lda,
h_x, &ione );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
// ||b - Ax|| / (||A||*||x||)
// error for CUBLAS
normA = lapackf77_clange( "F", &N, &N, h_A, &lda, work );
normx = lapackf77_clange( "F", &N, &ione, h_xcublas, &ione, work );
blasf77_ctrmv( &opts.uplo, &opts.transA, &opts.diag,
&N,
h_A, &lda,
h_xcublas, &ione );
blasf77_caxpy( &N, &c_neg_one, h_b, &ione, h_xcublas, &ione );
normr = lapackf77_clange( "F", &N, &ione, h_xcublas, &N, work );
cublas_error = normr / (normA*normx);
if ( opts.lapack ) {
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e\n",
(int) N,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error );
}
else {
printf("%5d %7.2f (%7.2f) --- ( --- ) %8.2e\n",
(int) N,
cublas_perf, 1000.*cublas_time,
cublas_error );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_b );
TESTING_FREE_CPU( h_x );
TESTING_FREE_CPU( h_xcublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_x );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing clarfb_gpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t M, N, K, size, ldc, ldv, ldt, ldw, nv;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float error, work[1];
magma_int_t status = 0;
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans [] = { MagmaConjTrans, MagmaNoTrans };
magma_direct_t direct[] = { MagmaForward, MagmaBackward };
magma_storev_t storev[] = { MagmaColumnwise, MagmaRowwise };
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
printf(" M N K storev side direct trans ||R||_F / ||HC||_F\n");
printf("========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
M = opts.msize[itest];
N = opts.nsize[itest];
K = opts.ksize[itest];
if ( M < K || N < K || K <= 0 ) {
printf( "%5d %5d %5d skipping because clarfb requires M >= K, N >= K, K >= 0\n",
(int) M, (int) N, (int) K );
continue;
}
for( int istor = 0; istor < 2; ++istor ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int idir = 0; idir < 2; ++idir ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
ldc = ((M+31)/32)*32;
ldt = ((K+31)/32)*32;
ldw = (side[iside] == MagmaLeft ? N : M);
// (ldv, nv) get swapped later if rowwise
ldv = (side[iside] == MagmaLeft ? M : N);
nv = K;
// Allocate memory for matrices
magmaFloatComplex *C, *R, *V, *T, *W;
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*N );
TESTING_MALLOC_CPU( V, magmaFloatComplex, ldv*K );
TESTING_MALLOC_CPU( T, magmaFloatComplex, ldt*K );
TESTING_MALLOC_CPU( W, magmaFloatComplex, ldw*K );
magmaFloatComplex_ptr dC, dV, dT, dW;
TESTING_MALLOC_DEV( dC, magmaFloatComplex, ldc*N );
TESTING_MALLOC_DEV( dV, magmaFloatComplex, ldv*K );
TESTING_MALLOC_DEV( dT, magmaFloatComplex, ldt*K );
TESTING_MALLOC_DEV( dW, magmaFloatComplex, ldw*K );
// C is M x N.
size = ldc*N;
lapackf77_clarnv( &ione, ISEED, &size, C );
//printf( "C=" ); magma_cprint( M, N, C, ldc );
// V is ldv x nv. See larfb docs for description.
// if column-wise and left, M x K
// if column-wise and right, N x K
// if row-wise and left, K x M
// if row-wise and right, K x N
size = ldv*nv;
lapackf77_clarnv( &ione, ISEED, &size, V );
if ( storev[istor] == MagmaColumnwise ) {
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaUpperStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( MagmaLowerStr, &K, &K, &c_zero, &c_one, &V[(ldv-K)], &ldv );
}
}
else {
// rowwise, swap V's dimensions
std::swap( ldv, nv );
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaLowerStr, &K, &K, &c_zero, &c_one, V, &ldv );
}
else {
lapackf77_claset( MagmaUpperStr, &K, &K, &c_zero, &c_one, &V[(nv-K)*ldv], &ldv );
}
}
//printf( "# ldv %d, nv %d\n", ldv, nv );
//printf( "V=" ); magma_cprint( ldv, nv, V, ldv );
// T is K x K, upper triangular for forward, and lower triangular for backward
magma_int_t k1 = K-1;
size = ldt*K;
lapackf77_clarnv( &ione, ISEED, &size, T );
if ( direct[idir] == MagmaForward ) {
lapackf77_claset( MagmaLowerStr, &k1, &k1, &c_zero, &c_zero, &T[1], &ldt );
}
else {
lapackf77_claset( MagmaUpperStr, &k1, &k1, &c_zero, &c_zero, &T[1*ldt], &ldt );
}
//printf( "T=" ); magma_cprint( K, K, T, ldt );
magma_csetmatrix( M, N, C, ldc, dC, ldc );
magma_csetmatrix( ldv, nv, V, ldv, dV, ldv );
magma_csetmatrix( K, K, T, ldt, dT, ldt );
lapackf77_clarfb( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
lapack_direct_const( direct[idir] ), lapack_storev_const( storev[istor] ),
&M, &N, &K,
V, &ldv, T, &ldt, C, &ldc, W, &ldw );
//printf( "HC=" ); magma_cprint( M, N, C, ldc );
magma_clarfb_gpu( side[iside], trans[itran], direct[idir], storev[istor],
M, N, K,
dV, ldv, dT, ldt, dC, ldc, dW, ldw );
magma_cgetmatrix( M, N, dC, ldc, R, ldc );
//printf( "dHC=" ); magma_cprint( M, N, R, ldc );
// compute relative error |HC_magma - HC_lapack| / |HC_lapack|
error = lapackf77_clange( "Fro", &M, &N, C, &ldc, work );
size = ldc*N;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "Fro", &M, &N, R, &ldc, work ) / error;
printf( "%5d %5d %5d %c %c %c %c %8.2e %s\n",
(int) M, (int) N, (int) K,
lapacke_storev_const(storev[istor]), lapacke_side_const(side[iside]),
lapacke_direct_const(direct[idir]), lapacke_trans_const(trans[itran]),
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( V );
TESTING_FREE_CPU( T );
TESTING_FREE_CPU( W );
TESTING_FREE_DEV( dC );
TESTING_FREE_DEV( dV );
TESTING_FREE_DEV( dT );
TESTING_FREE_DEV( dW );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}}}}
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhesv
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magmaDoubleComplex *h_A, *h_B, *h_X, *work, temp;
real_Double_t gflops, gpu_perf, gpu_time = 0.0, cpu_perf=0, cpu_time=0;
double error, error_lapack = 0.0;
magma_int_t *ipiv;
magma_int_t N, n2, lda, ldb, sizeB, lwork, info;
magma_int_t status = 0, ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) |Ax-b|/(N*|A|*|x|)\n");
printf("%%========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
ldb = N;
lda = N;
n2 = lda*N;
sizeB = ldb*opts.nrhs;
gflops = ( FLOPS_ZPOTRF( N ) + FLOPS_ZPOTRS( N, opts.nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_PIN( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_B, magmaDoubleComplex, sizeB );
TESTING_MALLOC_PIN( h_X, magmaDoubleComplex, sizeB );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
lwork = -1;
lapackf77_zhesv(lapack_uplo_const(opts.uplo), &N, &opts.nrhs,
h_A, &lda, ipiv, h_X, &ldb, &temp, &lwork, &info);
lwork = (int)MAGMA_Z_REAL(temp);
TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork );
init_matrix( N, N, h_A, lda );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_zlacpy( MagmaFullStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_zhesv(lapack_uplo_const(opts.uplo), &N, &opts.nrhs,
h_A, &lda, ipiv, h_X, &ldb, work, &lwork, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_zhesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
error_lapack = get_residual( opts.uplo, N, opts.nrhs, h_A, lda, ipiv, h_X, ldb, h_B, ldb );
TESTING_FREE_CPU( work );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
init_matrix( N, N, h_A, lda );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_zlacpy( MagmaFullStr, &N, &opts.nrhs, h_B, &ldb, h_X, &ldb );
magma_setdevice(0);
gpu_time = magma_wtime();
magma_zhesv( opts.uplo, N, opts.nrhs, h_A, lda, ipiv, h_X, ldb, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_zhesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) N, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) N, (int) N, gpu_perf, gpu_time );
}
if ( opts.check == 0 ) {
printf(" --- \n");
} else {
error = get_residual( opts.uplo, N, opts.nrhs, h_A, lda, ipiv, h_X, ldb, h_B, ldb );
printf(" %8.2e %s", error, (error < tol ? "ok" : "failed"));
if (opts.lapack)
printf(" (lapack rel.res. = %8.2e)", error_lapack);
printf("\n");
status += ! (error < tol);
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_PIN( h_X );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssygvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
real_Double_t mgpu_time;
float *h_A, *h_Ainit, *h_B, *h_Binit, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
#endif
float *w1, result=0;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: ngpu = %d, itype = %d, jobz = %s, range = %s, uplo = %s, opts.check = %d, fraction = %6.4f\n",
(int) opts.ngpu, (int) opts.itype,
lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M ngpu MGPU Time (sec)\n");
printf("====================================\n");
magma_int_t threads = magma_get_parallel_numthreads();
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = magma_sbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_sbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
//magma_int_t NB = 96;//magma_bulge_get_nb(N);
//magma_int_t sizvblg = magma_sbulge_get_lq2(N, threads);
//magma_int_t siz = max(sizvblg,n2)+2*(N*NB+N)+24*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_PIN( h_A, float, n2 );
TESTING_MALLOC_PIN( h_B, float, n2 );
TESTING_MALLOC_PIN( h_work, float, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, float, lrwork);
#endif
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork);
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slarnv( &ione, ISEED, &n2, h_B );
magma_smake_hpd( N, h_B, N );
magma_smake_symmetric( N, h_A, N );
if ( opts.warmup || opts.check ) {
TESTING_MALLOC_CPU( h_Ainit, float, n2 );
TESTING_MALLOC_CPU( h_Binit, float, n2 );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_Ainit, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_Binit, &N );
}
magma_int_t m1 = 0;
float vl = 0;
float vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == MagmaRangeI) {
il = 1;
iu = (int) (opts.fraction*N);
}
if ( opts.warmup ) {
// ==================================================================
// Warmup using MAGMA. I prefer to use smalltest to warmup A-
// ==================================================================
magma_ssygvdx_2stage_m(opts.ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N );
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
mgpu_time = magma_wtime();
magma_ssygvdx_2stage_m(opts.ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
mgpu_time = magma_wtime() - mgpu_time;
if ( opts.check ) {
// ===================================================================
// Check the results following the LAPACK's [zc]hegvdx routine.
// A x = lambda B x is solved
// and the following 3 tests computed:
// (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
// | A B Z - Z D | / ( |A||Z| N ) (itype = 2)
// | B A Z - Z D | / ( |A||Z| N ) (itype = 3)
// ===================================================================
#if defined(PRECISION_d) || defined(PRECISION_s)
float *rwork = h_work + N*N;
#endif
result = 1.;
result /= lapackf77_slansy("1", lapack_uplo_const(opts.uplo), &N, h_Ainit, &N, rwork);
result /= lapackf77_slange("1", &N , &m1, h_A, &N, rwork);
if (opts.itype == 1) {
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N);
result *= lapackf77_slange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_slange("1", &N, &m1, h_A, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_slange("1", &N, &m1, h_A, &N, rwork)/N;
}
}
// ===================================================================
// Print execution time
// ===================================================================
printf("%5d %5d %4d %7.2f\n",
(int) N, (int) m1, (int) opts.ngpu, mgpu_time);
if ( opts.check ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if (opts.itype==1) {
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==2) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==3) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
printf("\n");
status += ! (result < tol);
}
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( iwork );
if ( opts.warmup || opts.check ) {
TESTING_FREE_CPU( h_Ainit );
TESTING_FREE_CPU( h_Binit );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE_MGPU();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zposv_batched
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
double err = 0.0, Rnorm, Anorm, Xnorm, *work;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_B, *h_X;
magmaDoubleComplex_ptr d_A, d_B;
magma_int_t *dinfo_array;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_int_t batchCount = 1;
magmaDoubleComplex **dA_array = NULL;
magmaDoubleComplex **dB_array = NULL;
magma_queue_t queue = magma_stream;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
batchCount = opts.batchcount ;
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("BatchCount N NRHS CPU GFlop/s (sec) GPU GFlop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
ldda = ((N+31)/32)*32;
lddb = ldda;
gflops = ( FLOPS_ZPOTRF( N) + FLOPS_ZPOTRS( N, nrhs ) ) / 1e9 * batchCount;
sizeA = lda*N*batchCount;
sizeB = ldb*nrhs*batchCount;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, sizeB );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, sizeB );
TESTING_MALLOC_CPU( work, double, N);
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N*batchCount );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*nrhs*batchCount );
TESTING_MALLOC_DEV( dinfo_array, magma_int_t, batchCount );
magma_malloc((void**)&dA_array, batchCount * sizeof(*dA_array));
magma_malloc((void**)&dB_array, batchCount * sizeof(*dB_array));
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
for(int i=0; i<batchCount; i++)
{
magma_zmake_hpd( N, h_A + i * lda * N, lda );// need modification
}
magma_zsetmatrix( N, N*batchCount, h_A, lda, d_A, ldda );
magma_zsetmatrix( N, nrhs*batchCount, h_B, ldb, d_B, lddb );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
zset_pointer(dA_array, d_A, ldda, 0, 0, ldda*N, batchCount, queue);
zset_pointer(dB_array, d_B, lddb, 0, 0, lddb*nrhs, batchCount, queue);
gpu_time = magma_wtime();
info = magma_zposv_batched(opts.uplo, N, nrhs, dA_array, ldda, dB_array, lddb, dinfo_array, batchCount, queue);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
// check correctness of results throught "dinfo_magma" and correctness of argument throught "info"
magma_int_t *cpu_info = (magma_int_t*) malloc(batchCount*sizeof(magma_int_t));
magma_getvector( batchCount, sizeof(magma_int_t), dinfo_array, 1, cpu_info, 1);
for(int i=0; i<batchCount; i++)
{
if(cpu_info[i] != 0 ){
printf("magma_zposv_batched matrix %d returned internal error %d\n",i, (int)cpu_info[i] );
}
}
if (info != 0)
printf("magma_zposv_batched returned argument error %d: %s.\n", (int) info, magma_strerror( info ));
//=====================================================================
// Residual
//=====================================================================
magma_zgetmatrix( N, nrhs*batchCount, d_B, lddb, h_X, ldb );
for(magma_int_t s=0; s<batchCount; s++)
{
Anorm = lapackf77_zlange("I", &N, &N, h_A + s * lda * N, &lda, work);
Xnorm = lapackf77_zlange("I", &N, &nrhs, h_X + s * ldb * nrhs, &ldb, work);
blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A + s * lda * N, &lda,
h_X + s * ldb * nrhs, &ldb,
&c_neg_one, h_B + s * ldb * nrhs, &ldb);
Rnorm = lapackf77_zlange("I", &N, &nrhs, h_B + s * ldb * nrhs, &ldb, work);
double error = Rnorm/(N*Anorm*Xnorm);
if ( isnan(error) || isinf(error) ) {
err = error;
break;
}
err = max(err, error);
}
status += ! (err < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
for(magma_int_t s=0; s<batchCount; s++)
{
lapackf77_zposv( lapack_uplo_const(opts.uplo), &N, &nrhs, h_A + s * lda * N, &lda, h_B + s * ldb * nrhs, &ldb, &info );
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zposv returned err %d: %s.\n",
(int) info, magma_strerror( info ));
printf( "%10d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int)batchCount, (int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
err, (err < tol ? "ok" : "failed"));
}
else {
printf( "%10d %5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int)batchCount, (int) N, (int) nrhs, gpu_perf, gpu_time,
err, (err < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( dinfo_array );
magma_free(dA_array);
magma_free(dB_array);
free(cpu_info);
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing chegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
/* Constants */
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
const magma_int_t ione = 1;
/* Local variables */
real_Double_t gpu_time;
magmaFloatComplex *h_A, *h_R, *h_B, *h_S, *h_work;
#ifdef COMPLEX
float *rwork;
magma_int_t lrwork;
#endif
float *w1, *w2, result[2]={0,0};
magma_int_t *iwork;
magma_int_t N, n2, info, lda, lwork, liwork;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
// pass ngpu = -1 to test multi-GPU code using 1 gpu
magma_int_t abs_ngpu = abs( opts.ngpu );
printf("%% itype = %d, jobz = %s, range = %s, uplo = %s, fraction = %6.4f, ngpu = %d\n",
int(opts.itype), lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
opts.fraction, int(abs_ngpu) );
if (opts.itype == 1) {
printf("%% N M GPU Time (sec) |AZ-BZD| |D - D_magma|\n");
}
else if (opts.itype == 2) {
printf("%% N M GPU Time (sec) |ABZ-ZD| |D - D_magma|\n");
}
else if (opts.itype == 3) {
printf("%% N M GPU Time (sec) |BAZ-ZD| |D - D_magma|\n");
}
printf("%%======================================================\n");
magma_int_t threads = magma_get_parallel_numthreads();
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
// TODO: test vl-vu range
magma_int_t m1 = 0;
float vl = 0;
float vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (opts.fraction == 0) {
il = max( 1, magma_int_t(0.1*N) );
iu = max( 1, magma_int_t(0.3*N) );
}
else {
il = 1;
iu = max( 1, magma_int_t(opts.fraction*N) );
}
magma_cheevdx_getworksize(N, threads, (opts.jobz == MagmaVec),
&lwork,
#ifdef COMPLEX
&lrwork,
#endif
&liwork);
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaFloatComplex, n2 );
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( w2, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaFloatComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaFloatComplex, max( lwork, N*N )); // check needs N*N
#ifdef COMPLEX
TESTING_MALLOC_PIN( rwork, float, lrwork);
#endif
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
lapackf77_clarnv( &ione, ISEED, &n2, h_B );
magma_cmake_hpd( N, h_B, lda );
magma_cmake_hermitian( N, h_A, lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &lda, h_S, &lda );
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
gpu_time = magma_wtime();
if (opts.ngpu == 1) {
magma_chegvdx_2stage( opts.itype, opts.jobz, range, opts.uplo,
N, h_R, lda, h_S, lda, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
else {
magma_chegvdx_2stage_m( abs_ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_R, lda, h_S, lda, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#ifdef COMPLEX
rwork, lrwork,
#endif
iwork, liwork,
&info );
}
gpu_time = magma_wtime() - gpu_time;
if (info != 0) {
printf("magma_chegvdx_2stage returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvdx routine.
A x = lambda B x is solved
and the following 3 tests computed:
(1) | A Z - B Z D | / ( |A| |Z| N ) (itype = 1)
| A B Z - Z D | / ( |A| |Z| N ) (itype = 2)
| B A Z - Z D | / ( |A| |Z| N ) (itype = 3)
(2) | D(with V, magma) - D(w/o V, lapack) | / | D |
=================================================================== */
#ifdef REAL
float *rwork = h_work + N*N;
#endif
if ( opts.jobz != MagmaNoVec ) {
result[0] = 1.;
result[0] /= safe_lapackf77_clanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, rwork);
result[0] /= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork);
if (opts.itype == 1) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int i=0; i < m1; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N);
result[0] *= lapackf77_clange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int i=0; i < m1; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for (int i=0; i < m1; ++i)
blasf77_csscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_chemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_clange("1", &N, &m1, h_R, &lda, rwork)/N;
}
}
lapackf77_clacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_clacpy( MagmaFullStr, &N, &N, h_B, &lda, h_S, &lda );
lapackf77_chegvd( &opts.itype, "N", lapack_uplo_const(opts.uplo), &N,
h_R, &lda, h_S, &lda, w2,
h_work, &lwork,
#ifdef COMPLEX
rwork, &lrwork,
#endif
iwork, &liwork,
&info );
if (info != 0) {
printf("lapackf77_chegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
float maxw=0, diff=0;
for (int j=0; j < m1; j++) {
maxw = max(maxw, fabs(w1[j]));
maxw = max(maxw, fabs(w2[j]));
diff = max(diff, fabs(w1[j] - w2[j]));
}
result[1] = diff / (m1*maxw);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %9.4f ",
(int) N, (int) m1, gpu_time);
if ( opts.check ) {
bool okay = (result[1] < tolulp);
if ( opts.jobz != MagmaNoVec ) {
okay = okay && (result[0] < tol);
printf(" %8.2e", result[0] );
}
else {
printf(" --- ");
}
printf(" %8.2e %s\n", result[1], (okay ? "ok" : "failed"));
status += ! okay;
}
else {
printf(" ---\n");
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
#ifdef COMPLEX
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zcgeqrsv
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time, gpu_perfd, gpu_perfs;
double error, gpu_error, cpu_error, Anorm, work[1];
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_A2, *h_B, *h_X, *h_R;
magmaDoubleComplex_ptr d_A, d_B, d_X, d_T;
magmaFloatComplex *d_SA, *d_SB;
magmaDoubleComplex *h_workd, *tau, tmp[1];
magmaFloatComplex *h_works;
magma_int_t lda, ldb, lhwork, lworkgpu;
magma_int_t ldda, lddb, lddx;
magma_int_t M, N, nrhs, qrsv_iters, info, size, min_mn, max_mn, nb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
printf("Epsilon(double): %8.6e\n"
"Epsilon(single): %8.6e\n\n",
lapackf77_dlamch("Epsilon"), lapackf77_slamch("Epsilon") );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
printf(" CPU Gflop/s GPU Gflop/s |b-Ax|| / (N||A||) ||dx-x||/(N||A||)\n");
printf(" M N NRHS double double single mixed Iter CPU GPU \n");
printf("=============================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
if ( M < N ) {
printf( "%5d %5d %5d skipping because M < N is not yet supported.\n", (int) M, (int) N, (int) nrhs );
continue;
}
min_mn = min(M, N);
max_mn = max(M, N);
lda = M;
ldb = max_mn;
ldda = ((M+31)/32) * 32;
lddb = ((max_mn+31)/32)*32;
lddx = ((N+31)/32) * 32;
nb = max( magma_get_zgeqrf_nb( M ), magma_get_cgeqrf_nb( M ) );
gflops = (FLOPS_ZGEQRF( M, N ) + FLOPS_ZGEQRS( M, N, nrhs )) / 1e9;
lworkgpu = (M - N + nb)*(nrhs + nb) + nrhs*nb;
// query for workspace size
lhwork = -1;
lapackf77_zgels( MagmaNoTransStr, &M, &N, &nrhs,
NULL, &lda, NULL, &ldb, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_Z_REAL( tmp[0] );
lhwork = max( lhwork, lworkgpu );
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_A2, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_R, magmaDoubleComplex, ldb*nrhs );
TESTING_MALLOC_CPU( h_workd, magmaDoubleComplex, lhwork );
h_works = (magmaFloatComplex*)h_workd;
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*nrhs );
TESTING_MALLOC_DEV( d_X, magmaDoubleComplex, lddx*nrhs );
TESTING_MALLOC_DEV( d_T, magmaDoubleComplex, ( 2*min_mn + (N+31)/32*32 )*nb );
/* Initialize the matrices */
size = lda*N;
lapackf77_zlarnv( &ione, ISEED, &size, h_A );
lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_A2, &lda );
// make random RHS
size = ldb*nrhs;
lapackf77_zlarnv( &ione, ISEED, &size, h_B );
lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_R, &ldb );
magma_zsetmatrix( M, N, h_A, lda, d_A, ldda );
magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
//=====================================================================
// Mixed Precision Iterative Refinement - GPU
//=====================================================================
gpu_time = magma_wtime();
magma_zcgeqrsv_gpu( M, N, nrhs,
d_A, ldda, d_B, lddb,
d_X, lddx, &qrsv_iters, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zcgeqrsv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// compute the residual
magma_zgetmatrix( N, nrhs, d_X, lddx, h_X, ldb );
blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A, &lda,
h_X, &ldb,
&c_one, h_R, &ldb);
Anorm = lapackf77_zlange("f", &M, &N, h_A, &lda, work);
//=====================================================================
// Double Precision Solve
//=====================================================================
magma_zsetmatrix( M, N, h_A, lda, d_A, ldda );
magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
gpu_time = magma_wtime();
magma_zgels_gpu( MagmaNoTrans, M, N, nrhs, d_A, ldda,
d_B, lddb, h_workd, lworkgpu, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perfd = gflops / gpu_time;
//=====================================================================
// Single Precision Solve
//=====================================================================
magma_zsetmatrix( M, N, h_A, lda, d_A, ldda );
magma_zsetmatrix( M, nrhs, h_B, ldb, d_B, lddb );
/* The allocation of d_SA and d_SB is done here to avoid
* to double the memory used on GPU with zcgeqrsv */
TESTING_MALLOC_DEV( d_SA, magmaFloatComplex, ldda*N );
TESTING_MALLOC_DEV( d_SB, magmaFloatComplex, lddb*nrhs );
magmablas_zlag2c( M, N, d_A, ldda, d_SA, ldda, &info );
magmablas_zlag2c( N, nrhs, d_B, lddb, d_SB, lddb, &info );
gpu_time = magma_wtime();
magma_cgels_gpu( MagmaNoTrans, M, N, nrhs, d_SA, ldda,
d_SB, lddb, h_works, lhwork, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perfs = gflops / gpu_time;
TESTING_FREE_DEV( d_SA );
TESTING_FREE_DEV( d_SB );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
lapackf77_zlacpy( MagmaUpperLowerStr, &M, &nrhs, h_B, &ldb, h_X, &ldb );
cpu_time = magma_wtime();
lapackf77_zgels( MagmaNoTransStr, &M, &N, &nrhs,
h_A, &lda, h_X, &ldb, h_workd, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zgels returned error %d: %s.\n",
(int) info, magma_strerror( info ));
blasf77_zgemm( MagmaNoTransStr, MagmaNoTransStr, &M, &nrhs, &N,
&c_neg_one, h_A2, &lda,
h_X, &ldb,
&c_one, h_B, &ldb );
cpu_error = lapackf77_zlange("f", &M, &nrhs, h_B, &ldb, work) / (min_mn*Anorm);
gpu_error = lapackf77_zlange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
// error relative to LAPACK
size = M*nrhs;
blasf77_zaxpy( &size, &c_neg_one, h_B, &ione, h_R, &ione );
error = lapackf77_zlange("f", &M, &nrhs, h_R, &ldb, work) / (min_mn*Anorm);
printf("%5d %5d %5d %7.2f %7.2f %7.2f %7.2f %4d %8.2e %8.2e %8.2e %s\n",
(int) M, (int) N, (int) nrhs,
cpu_perf, gpu_perfd, gpu_perfs, gpu_perf,
(int) qrsv_iters,
cpu_error, gpu_error, error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_A2 );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( h_R );
TESTING_FREE_CPU( h_workd );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_X );
TESTING_FREE_DEV( d_T );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cunmbr
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, dwork[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t ione = 1;
magma_int_t m, n, k, mi, ni, mm, nn, nq, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
magmaFloatComplex *C, *R, *A, *work, *tau, *tauq, *taup;
float *d, *e;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_vect_t vect [] = { MagmaQ, MagmaP };
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { Magma_ConjTrans, MagmaNoTrans };
printf(" M N K vect side trans CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||QC||_F\n");
printf("===============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int ivect = 0; ivect < 2; ++ivect ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_cgebrd_nb( m );
ldc = m;
// A is nq x k (vect=Q) or k x nq (vect=P)
// where nq=m (left) or nq=n (right)
nq = (side[iside] == MagmaLeft ? m : n );
mm = (vect[ivect] == MagmaQ ? nq : k );
nn = (vect[ivect] == MagmaQ ? k : nq);
lda = mm;
// MBR calls either MQR or MLQ in various ways
if ( vect[ivect] == MagmaQ ) {
if ( nq >= k ) {
gflops = FLOPS_CUNMQR( m, n, k, side[iside] ) / 1e9;
}
else {
if ( side[iside] == MagmaLeft ) {
mi = m - 1;
ni = n;
}
else {
mi = m;
ni = n - 1;
}
gflops = FLOPS_CUNMQR( mi, ni, nq-1, side[iside] ) / 1e9;
}
}
else {
if ( nq > k ) {
gflops = FLOPS_CUNMLQ( m, n, k, side[iside] ) / 1e9;
}
else {
if ( side[iside] == MagmaLeft ) {
mi = m - 1;
ni = n;
}
else {
mi = m;
ni = n - 1;
}
gflops = FLOPS_CUNMLQ( mi, ni, nq-1, side[iside] ) / 1e9;
}
}
// workspace for gebrd is (mm + nn)*nb
// workspace for unmbr is m*nb or n*nb, depending on side
lwork_max = max( (mm + nn)*nb, max( m*nb, n*nb ));
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*nn );
TESTING_MALLOC_CPU( work, magmaFloatComplex, lwork_max );
TESTING_MALLOC_CPU( d, float, min(mm,nn) );
TESTING_MALLOC_CPU( e, float, min(mm,nn) );
TESTING_MALLOC_CPU( tauq, magmaFloatComplex, min(mm,nn) );
TESTING_MALLOC_CPU( taup, magmaFloatComplex, min(mm,nn) );
// C is full, m x n
size = ldc*n;
lapackf77_clarnv( &ione, ISEED, &size, C );
lapackf77_clacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*nn;
lapackf77_clarnv( &ione, ISEED, &size, A );
// compute BRD factorization to get Householder vectors in A, tauq, taup
//lapackf77_cgebrd( &mm, &nn, A, &lda, d, e, tauq, taup, work, &lwork_max, &info );
magma_cgebrd( mm, nn, A, lda, d, e, tauq, taup, work, lwork_max, &info );
if (info != 0)
printf("magma_cgebrd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( vect[ivect] == MagmaQ ) {
tau = tauq;
} else {
tau = taup;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cunmbr( lapack_vect_const( vect[ivect] ),
lapack_side_const( side[iside] ),
lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, work, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_cunmbr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_cunmbr( vect[ivect], side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, work, lwork, &info );
if (info != 0)
printf("magma_cunmbr (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = (magma_int_t) MAGMA_C_REAL( work[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("optimal lwork %d > lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_cunmbr( vect[ivect], side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, work, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_cunmbr returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
error = lapackf77_clange( "Fro", &m, &n, C, &ldc, dwork );
size = ldc*n;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
error = lapackf77_clange( "Fro", &m, &n, R, &ldc, dwork ) / error;
printf( "%5d %5d %5d %c %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_vect_const( vect[ivect] ),
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( d );
TESTING_FREE_CPU( e );
TESTING_FREE_CPU( taup );
TESTING_FREE_CPU( tauq );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}}} // end ivect, iside, itran
printf( "\n" );
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgeqlf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *h_A, *h_R, *tau, *h_work, tmp[1];
magma_int_t M, N, n2, lda, ldda, lwork, info, min_mn, nb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = 2. * opts.tolerance * lapackf77_dlamch("E");
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\n");
printf("=======================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
nb = magma_get_zgeqrf_nb(M);
gflops = FLOPS_ZGEQLF( M, N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_zgeqlf(&M, &N, NULL, &M, NULL, tmp, &lwork, &info);
lwork = (magma_int_t)MAGMA_Z_REAL( tmp[0] );
lwork = max( lwork, N*nb );
lwork = max( lwork, 2*nb*nb);
TESTING_MALLOC_CPU( tau, magmaDoubleComplex, min_mn );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_work, magmaDoubleComplex, lwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgeqlf( M, N, h_R, lda, tau, h_work, lwork, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zgeqlf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_zgeqlf(&M, &N, h_A, &lda, tau, h_work, &lwork, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapack_zgeqlf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = lapackf77_zlange("f", &M, &N, h_A, &lda, work);
blasf77_zaxpy(&n2, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_zlange("f", &M, &N, h_R, &lda, work) / error;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgetrf
*/
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time, error;
magmaDoubleComplex *h_A, *h_R;
magmaDoubleComplex_ptr d_A, dwork;
magma_int_t N = 0, n2, lda, ldda;
magma_int_t size[10] = { 1024, 2048, 3072, 4032, 5184, 5600, 5600, 5600, 5600, 5600 };
magma_int_t ntest = 10;
magma_int_t i, info;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0, 0, 0, 1};
magmaDoubleComplex *work;
magmaDoubleComplex tmp;
double rwork[1];
magma_int_t *ipiv;
magma_int_t lwork, ldwork;
double A_norm, R_norm;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
}
if (N>0) size[0] = size[ntest-1] = N;
else exit(1);
}
else {
printf("\nUsage: \n");
printf(" testing_zgetri_gpu -N %d\n\n", 1024);
}
/* query for Lapack workspace size */
N = size[ntest-1];
lda = N;
work = &tmp;
lwork = -1;
lapackf77_zgetri( &N, h_A, &lda, ipiv, work, &lwork, &info );
if (info != 0)
printf("lapackf77_zgetri returned error %d\n", (int) info);
lwork = int( MAGMA_Z_REAL( *work ));
/* query for Magma workspace size */
ldwork = N * magma_get_zgetri_nb( N );
/* Initialize */
magma_queue_t queue;
magma_device_t device[ MagmaMaxGPUs ];
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device[0], &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
/* Allocate memory */
n2 = N * N;
ldda = ((N+31)/32) * 32;
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
printf(" N CPU GFlop/s GPU GFlop/s ||R||_F / ||A||_F\n");
printf("========================================================\n");
for( i=0; i < ntest; i++ ){
N = size[i];
lda = N;
n2 = lda*N;
gflops = FLOPS_ZGETRI( (double)N ) / 1e9;
ldda = ((N+31)/32)*32;
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
A_norm = lapackf77_zlange( "f", &N, &N, h_A, &lda, rwork );
/* Factor the matrix. Both MAGMA and LAPACK will use this factor. */
magma_zsetmatrix( N, N, h_A, 0, lda, d_A, 0, ldda, queue );
magma_zgetrf_gpu( N, N, d_A, 0, ldda, ipiv, &info, queue );
magma_zgetmatrix( N, N, d_A, 0, ldda, h_A, 0, lda, queue );
// check for exact singularity
//h_A[ 10 + 10*lda ] = MAGMA_Z_MAKE( 0.0, 0.0 );
//magma_zsetmatrix( N, N, h_A, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
//warm-up
magma_zgetri_gpu( N, d_A, 0, ldda, ipiv, dwork, 0, ldwork, &info, queue );
magma_zsetmatrix( N, N, h_A, 0, lda, d_A, 0, ldda, queue );
gpu_time = magma_wtime();
magma_zgetri_gpu( N, d_A, 0, ldda, ipiv, dwork, 0, ldwork, &info, queue );
gpu_time = magma_wtime()-gpu_time;
if (info != 0)
printf("magma_zgetri_gpu returned error %d\n", (int) info);
gpu_perf = gflops / gpu_time;
magma_zgetmatrix( N, N, d_A, 0, ldda, h_R, 0, lda, queue );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_zgetri( &N, h_A, &lda, ipiv, work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_zgetri returned error %d\n", (int) info);
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
blasf77_zaxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
R_norm = lapackf77_zlange( "f", &N, &N, h_R, &lda, rwork );
printf( "%5d %6.2f %6.2f %e\n",
(int) N, cpu_perf, gpu_perf, R_norm / A_norm );
if (argc != 1)
break;
}
/* Memory clean up */
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dwork );
/* Shutdown */
magma_queue_destroy( queue );
magma_finalize();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing magma_zhemm_mgpu
*/
int main( int argc, char** argv)
{
TESTING_INIT();
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex calpha = MAGMA_Z_MAKE( 3.456, 5.678 );
magmaDoubleComplex cbeta = MAGMA_Z_MAKE( 1.234, 2.456 );
real_Double_t gflops, gpu_perf=0., cpu_perf=0., gpu_time=0., cpu_time=0.;
real_Double_t gpu_perf2=0., gpu_time2=0.;
double error=0., errorbis=0., work[1];
magmaDoubleComplex *hA, *hX, *hB, *hR;
magmaDoubleComplex_ptr dA[MagmaMaxGPUs], dX[MagmaMaxGPUs], dB[MagmaMaxGPUs], dwork[MagmaMaxGPUs], hwork[MagmaMaxGPUs+1];
magmaDoubleComplex_ptr dA2;
magma_int_t M, N, size, lda, ldda, msize, nb, nstream;
magma_int_t ione = 1;
magma_int_t iseed[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
// default values
nb = (opts.nb > 0 ? opts.nb : 64);
nstream = (opts.nstream > 0 ? opts.nstream : 2);
magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2];
magma_int_t nbcmplx = 0;
magma_buildconnection_mgpu(gnode, &nbcmplx, opts.ngpu);
printf("Initializing communication pattern... GPU-ncmplx %d\n\n", (int) nbcmplx);
for (int i=0; i < nbcmplx; ++i) {
int myngpu = gnode[i][MagmaMaxGPUs];
printf("cmplx %d has %d gpu ", i, myngpu);
for(int j=0; j < myngpu; ++j)
printf(" %d", (int) gnode[i][j]);
printf("\n");
}
magma_int_t nbevents = 2;
magma_queue_t streams[MagmaMaxGPUs][20];
magma_event_t redevents[MagmaMaxGPUs][20];
magma_event_t redevents2[MagmaMaxGPUs][MagmaMaxGPUs*MagmaMaxGPUs+10];
for( int d = 0; d < opts.ngpu; ++d ) {
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[d][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
cudaEventCreateWithFlags(&redevents[d][i], cudaEventDisableTiming);
cudaEventCreateWithFlags(&redevents2[d][i], cudaEventDisableTiming);
}
}
printf( "nb %d, ngpu %d, nstream %d version %d\n", (int) nb, (int) opts.ngpu, (int) nstream, (int) opts.version );
printf(" M N nb offset CPU GFlop/s (sec) GPU GFlop/s (sec) CUBLAS hemm (sec) ||R|| / ||A||*||X||\n");
printf("=========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
M = opts.msize[itest];
N = opts.nsize[itest];
for( int offset = 0; offset < N; offset += min(N,nb) ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
msize = M - offset;
lda = M;
ldda = ((M + 31)/32)*32;
size = lda*M;
gflops = FLOPS_ZHEMM( MagmaLeft, (double)msize, (double)N ) / 1e9;
magma_int_t dworksiz = ldda*N*3;
magma_int_t hworksiz = lda*N;
TESTING_MALLOC_CPU( hA, magmaDoubleComplex, lda*M );
TESTING_MALLOC_CPU( hX, magmaDoubleComplex, lda*N );
TESTING_MALLOC_CPU( hB, magmaDoubleComplex, lda*N );
TESTING_MALLOC_PIN( hR, magmaDoubleComplex, lda*N );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_int_t mlocal = ((M / nb) / opts.ngpu + 1) * nb;
magma_setdevice( d );
TESTING_MALLOC_DEV( dA[d], magmaDoubleComplex, ldda*mlocal );
TESTING_MALLOC_DEV( dX[d], magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dB[d], magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dwork[d], magmaDoubleComplex, dworksiz );
TESTING_MALLOC_PIN( hwork[d], magmaDoubleComplex, hworksiz );
}
TESTING_MALLOC_PIN( hwork[opts.ngpu], magmaDoubleComplex, lda*N );
if ( opts.check ) {
magma_setdevice( 0 );
TESTING_MALLOC_DEV( dA2, magmaDoubleComplex, ldda*M );
}
lapackf77_zlarnv( &ione, iseed, &size, hA );
magma_zmake_hermitian( M, hA, lda );
size = lda*N;
lapackf77_zlarnv( &ione, iseed, &size, hX );
lapackf77_zlarnv( &ione, iseed, &size, hB );
lapackf77_zlacpy( "Full", &M, &N, hB, &lda, hR, &lda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_zsetmatrix_1D_col_bcyclic( M, M, hA, lda, dA, ldda, opts.ngpu, nb );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
//magmablasSetKernelStream( streams[ d ][ 0 ] );
magma_zsetmatrix( M, N, hX, lda, dX[d], ldda );
//if (d == 0) magma_zsetmatrix( M, N, hB, lda, dB[d], ldda ); // this is wrong coz when offset != 0 the gpu who do the beta*C may be not 0 so this should be related to stdev(starting device who own i=0 first col)
magma_zsetmatrix( M, N, hB, lda, dB[d], ldda );
}
//memset(hR, 0, lda*N*sizeof(magmaDoubleComplex));
trace_init( 1, opts.ngpu, nstream, (magma_queue_t*) streams );
//magma_int_t offset = 0; //nb;
gpu_time = magma_sync_wtime(0);
magmablas_zhemm_mgpu_com(
MagmaLeft, MagmaLower, msize, N,
calpha, dA, ldda, offset,
dX, ldda,
cbeta, dB, ldda, dwork, dworksiz, hR, lda, hwork, hworksiz,
opts.ngpu, nb, streams, nstream, redevents2, nbevents, gnode, nbcmplx);
gpu_time = magma_sync_wtime(0) - gpu_time;
gpu_perf = gflops / gpu_time;
#ifdef TRACING
char buf[80];
snprintf( buf, sizeof(buf), "zhemm-m%d-n%d-nb%d-stream%d-ngpu%d-run%d.svg",
(int) M, (int) N, (int) nb, (int) nstream, (int) opts.ngpu, (int) iter );
trace_finalize( buf, "trace.css" );
#endif
/* ====================================================================
Performs operation using CUBLAS
=================================================================== */
if ( opts.check && iter == 0 ) {
magma_setdevice( 0 );
magmablasSetKernelStream( 0 );
magma_zsetmatrix( M, M, hA, lda, dA2, ldda );
magma_zsetmatrix( M, N, hX, lda, dX[0], ldda );
magma_zsetmatrix( M, N, hB, lda, dwork[0], ldda );
gpu_time2 = magma_sync_wtime(0);
magma_zhemm(
MagmaLeft, MagmaLower, msize, N,
calpha, dA2+offset*ldda+offset, ldda,
dX[0], ldda,
cbeta, dwork[0], ldda );
gpu_time2 = magma_sync_wtime(0) - gpu_time2;
gpu_perf2 = gflops / gpu_time2;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.check ) {
// store ||A||*||X||
errorbis = lapackf77_zlange("fro", &msize, &msize, hA+offset*lda+offset, &lda, work );
errorbis *= lapackf77_zlange("fro", &msize, &N, hX, &lda, work );
//printf( "A =" ); magma_zprint( M, M, hA, lda );
//printf( "X =" ); magma_zprint( M, N, hX, lda );
//printf( "B =" ); magma_zprint( M, N, hB, lda );
cpu_time = magma_wtime();
blasf77_zhemm( "Left", "Lower", &msize, &N,
&calpha, hA+offset*lda+offset, &lda,
hX, &lda,
&cbeta, hB, &lda );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/*
trace_file = fopen("AJETE/C", "w");
for (int j = 0; j < N; j++)
for (int i = 0; i < siz; i++)
fprintf(trace_file, "%10d%10d%40.30e\n", i+1, j+1, hB[j*lda+i]);
fclose(trace_file);
*/
magma_int_t firstprint=0;
for(magma_int_t dev=0; dev < opts.ngpu; ++dev) {
magma_setdevice( dev );
magma_zgetmatrix( M, N, dB[dev], ldda, hR, lda );
// compute relative error ||R||/||A||*||X||, where R := B_magma - B_lapack = R - B
size = lda*N;
blasf77_zaxpy( &size, &c_neg_one, hB, &ione, hR, &ione );
error = lapackf77_zlange("fro", &msize, &N, hR, &lda, work) / errorbis;
//printf( "R =" ); magma_zprint( M, N, hR, lda );
if (firstprint == 0) {
printf( "%5d %5d %5d %5d %7.1f (%7.4f) %7.1f (%7.4f) %7.1f (%7.4f) %8.2e %s\n",
(int) M, (int) N, (int) nb, (int) offset,
cpu_perf, cpu_time,
gpu_perf, gpu_time,
gpu_perf2, gpu_time2,
error, (error < tol ? "ok" : "failed") );
}
else {
printf( "%89s %8.2e %s\n", " ",
error, (error < tol ? "ok" : "failed") );
}
status += ! (error < tol);
firstprint =1;
}
} else {
printf( "%5d %5d %5d %5d --- ( --- ) %7.1f (%7.4f) --- ( --- ) ---\n",
(int) M, (int) N, (int) nb, (int) offset,
gpu_perf, gpu_time );
}
TESTING_FREE_CPU( hA );
TESTING_FREE_CPU( hX );
TESTING_FREE_CPU( hB );
TESTING_FREE_PIN( hR );
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
TESTING_FREE_DEV( dA[d] );
TESTING_FREE_DEV( dX[d] );
TESTING_FREE_DEV( dB[d] );
TESTING_FREE_DEV( dwork[d] );
TESTING_FREE_PIN( hwork[d] );
}
TESTING_FREE_PIN( hwork[opts.ngpu] );
if ( opts.check ) {
magma_setdevice( 0 );
TESTING_FREE_DEV( dA2 );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
} // offset
printf( "\n" );
}
for( int d = 0; d < opts.ngpu; ++d ) {
magma_setdevice( d );
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[d][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
magma_event_destroy( redevents[d][i] );
magma_event_destroy( redevents2[d][i] );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zlanhe
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A;
double *h_work;
magmaDoubleComplex_ptr d_A;
magmaDouble_ptr d_work;
magma_int_t i, j, N, n2, lda, ldda;
magma_int_t idist = 3; // normal distribution (otherwise max norm is always ~ 1)
magma_int_t ISEED[4] = {0,0,0,1};
double error, norm_magma, norm_lapack;
magma_int_t status = 0;
magma_int_t lapack_nan_fail = 0;
magma_int_t lapack_inf_fail = 0;
bool mkl_warning = false;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tol2;
magma_uplo_t uplo[] = { MagmaLower, MagmaUpper };
magma_norm_t norm[] = { MagmaInfNorm, MagmaOneNorm, MagmaMaxNorm, MagmaFrobeniusNorm };
// Double-Complex inf-norm not supported on Tesla (CUDA arch 1.x)
#if defined(PRECISION_z)
magma_int_t arch = magma_getdevice_arch();
if ( arch < 200 ) {
printf("!!!! NOTE: Double-Complex %s and %s norm are not supported\n"
"!!!! on CUDA architecture %d; requires arch >= 200.\n"
"!!!! It should report \"parameter number 1 had an illegal value\" below.\n\n",
MagmaInfNormStr, MagmaOneNormStr, (int) arch );
for( int inorm = 0; inorm < 2; ++inorm ) {
for( int iuplo = 0; iuplo < 2; ++iuplo ) {
printf( "Testing that magmablas_zlanhe( %s, %s, ... ) returns -1 error...\n",
lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ));
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], 1, NULL, 1, NULL, 1 );
if ( norm_magma != -1 ) {
printf( "expected magmablas_zlanhe to return -1 error, but got %f\n", norm_magma );
status = 1;
}
}}
printf( "...return values %s\n\n", (status == 0 ? "ok" : "failed") );
}
#endif
#ifdef MAGMA_WITH_MKL
// MKL 11.1 has bug in multi-threaded zlanhe; use single thread to work around.
// MKL 11.2 corrects it for inf, one, max norm.
// MKL 11.2 still segfaults for Frobenius norm, which is not tested here
// because MAGMA doesn't implement Frobenius norm yet.
MKLVersion mkl_version;
mkl_get_version( &mkl_version );
magma_int_t la_threads = magma_get_lapack_numthreads();
bool mkl_single_thread = (mkl_version.MajorVersion <= 11 && mkl_version.MinorVersion < 2);
if ( mkl_single_thread ) {
printf( "\nNote: using single thread to work around MKL zlanhe bug.\n\n" );
}
#endif
printf("%% N norm uplo CPU GByte/s (ms) GPU GByte/s (ms) error nan inf\n");
printf("%%=================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int inorm = 0; inorm < 3; ++inorm ) { /* < 4 for Frobenius */
for( int iuplo = 0; iuplo < 2; ++iuplo ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = magma_roundup( N, opts.align );
// read upper or lower triangle
gbytes = 0.5*(N+1)*N*sizeof(magmaDoubleComplex) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_work, double, N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( d_work, double, N );
/* Initialize the matrix */
lapackf77_zlarnv( &idist, ISEED, &n2, h_A );
magma_zsetmatrix( N, N, h_A, lda, d_A, ldda );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gbytes / gpu_time;
if (norm_magma == -1) {
printf( "%5d %4c skipped because %s norm isn't supported\n",
(int) N, lapacke_norm_const( norm[inorm] ), lapack_norm_const( norm[inorm] ));
goto cleanup;
}
else if (norm_magma < 0) {
printf("magmablas_zlanhe returned error %f: %s.\n",
norm_magma, magma_strerror( (int) norm_magma ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
#ifdef MAGMA_WITH_MKL
if ( mkl_single_thread ) {
// work around MKL bug in multi-threaded zlanhe
magma_set_lapack_numthreads( 1 );
}
#endif
cpu_time = magma_wtime();
norm_lapack = lapackf77_zlanhe(
lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ),
&N, h_A, &lda, h_work );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
if (norm_lapack < 0) {
printf("lapackf77_zlanhe returned error %f: %s.\n",
norm_lapack, magma_strerror( (int) norm_lapack ));
}
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = fabs( norm_magma - norm_lapack ) / norm_lapack;
tol2 = tol;
if ( norm[inorm] == MagmaMaxNorm ) {
// max-norm depends on only one element, so for Real precisions,
// MAGMA and LAPACK should exactly agree (tol2 = 0),
// while Complex precisions incur roundoff in cuCabs.
#ifdef REAL
tol2 = 0;
#endif
}
bool okay; okay = (error <= tol2);
status += ! okay;
mkl_warning |= ! okay;
/* ====================================================================
Check for NAN and INF propagation
=================================================================== */
#define h_A(i_, j_) (h_A + (i_) + (j_)*lda)
#define d_A(i_, j_) (d_A + (i_) + (j_)*ldda)
i = rand() % N;
j = rand() % N;
magma_int_t tmp;
if ( uplo[iuplo] == MagmaLower && i < j ) {
tmp = i;
i = j;
j = tmp;
}
else if ( uplo[iuplo] == MagmaUpper && i > j ) {
tmp = i;
i = j;
j = tmp;
}
*h_A(i,j) = MAGMA_Z_NAN;
magma_zsetvector( 1, h_A(i,j), 1, d_A(i,j), 1 );
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N );
norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ),
&N, h_A, &lda, h_work );
bool nan_okay; nan_okay = isnan(norm_magma);
bool la_nan_okay; la_nan_okay = isnan(norm_lapack);
lapack_nan_fail += ! la_nan_okay;
status += ! nan_okay;
*h_A(i,j) = MAGMA_Z_INF;
magma_zsetvector( 1, h_A(i,j), 1, d_A(i,j), 1 );
norm_magma = magmablas_zlanhe( norm[inorm], uplo[iuplo], N, d_A, ldda, d_work, N );
norm_lapack = lapackf77_zlanhe( lapack_norm_const( norm[inorm] ),
lapack_uplo_const( uplo[iuplo] ),
&N, h_A, &lda, h_work );
bool inf_okay; inf_okay = isinf(norm_magma);
bool la_inf_okay; la_inf_okay = isinf(norm_lapack);
lapack_inf_fail += ! la_inf_okay;
status += ! inf_okay;
#ifdef MAGMA_WITH_MKL
if ( mkl_single_thread ) {
// end single thread to work around MKL bug
magma_set_lapack_numthreads( la_threads );
}
#endif
printf("%5d %4c %4c %7.2f (%7.2f) %7.2f (%7.2f) %#9.3g %-6s %6s%1s %6s%1s\n",
(int) N,
lapacke_norm_const( norm[inorm] ),
lapacke_uplo_const( uplo[iuplo] ),
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
error,
(okay ? "ok" : "failed"),
(nan_okay ? "ok" : "failed"), (la_nan_okay ? " " : "*"),
(inf_okay ? "ok" : "failed"), (la_inf_okay ? " " : "*"));
cleanup:
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_work );
fflush( stdout );
} // end iter
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iuplo, inorm
printf( "\n" );
}
// don't print "failed" here because then run_tests.py thinks MAGMA failed
if ( lapack_nan_fail ) {
printf( "* Warning: LAPACK did not pass NAN propagation test; upgrade to LAPACK version >= 3.4.2 (Sep. 2012)\n" );
}
if ( lapack_inf_fail ) {
printf( "* Warning: LAPACK did not pass INF propagation test\n" );
}
if ( mkl_warning ) {
printf("* MKL (e.g., 11.1) has a bug in zlanhe with multiple threads;\n"
" corrected in 11.2 for one, inf, max norms, but still in Frobenius norm.\n"
" Try again with MKL_NUM_THREADS=1.\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zher2k
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
double cublas_error, Cnorm, work[1];
magma_int_t N, K;
magma_int_t Ak, An, Bk, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex *h_A, *h_B, *h_C, *h_Ccublas;
magmaDoubleComplex *d_A, *d_B, *d_C;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
double beta = MAGMA_D_MAKE( -0.48, 0.38 );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("If running lapack (option --lapack), CUBLAS error is computed\n"
"relative to CPU BLAS result.\n\n");
printf("uplo = %s, transA = %s\n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA) );
printf(" N K CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("==================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.msize[itest];
K = opts.ksize[itest];
gflops = FLOPS_ZHER2K(K, N) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = An = N;
Ak = K;
ldb = Bn = N;
Bk = K;
} else {
lda = An = K;
Ak = N;
ldb = Bn = K;
Bk = N;
}
ldc = N;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*Ak;
sizeB = ldb*Ak;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*Bk );
TESTING_MALLOC_CPU( h_C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Ccublas, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*Bk );
TESTING_MALLOC_DEV( d_C, magmaDoubleComplex, lddc*N );
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_zlarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_zsetmatrix( Bn, Bk, h_B, ldb, d_B, lddb );
magma_zsetmatrix( N, N, h_C, ldc, d_C, lddc );
cublas_time = magma_sync_wtime( NULL );
cublasZher2k( handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA), N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_zher2k( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_zlange( "M", &N, &N, h_C, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_zlange( "M", &N, &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Ccublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
// ----------------------------------------
int main( int argc, char** argv )
{
TESTING_INIT();
//real_Double_t t_m, t_c, t_f;
magma_int_t ione = 1;
magmaDoubleComplex *A, *B;
double error_cblas, error_fblas, error_inline;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t i, j, k, m, n, size, maxn, ld;
// complex x for magma, cblas, fortran, inline blas respectively
magmaDoubleComplex x2_m, x2_c, x2_f, x2_i;
// real x for magma, cblas, fortran, inline blas respectively
double x_m, x_c, x_f, x_i;
MAGMA_UNUSED( x_c );
MAGMA_UNUSED( x_f );
MAGMA_UNUSED( x2_c );
MAGMA_UNUSED( x2_f );
MAGMA_UNUSED( x2_m );
magma_opts opts;
opts.parse_opts( argc, argv );
opts.tolerance = max( 100., opts.tolerance );
double tol = opts.tolerance * lapackf77_dlamch("E");
gTol = tol;
magma_int_t inc[] = { -2, -1, 1, 2 }; //{ 1 }; //{ -1, 1 };
magma_int_t ninc = sizeof(inc)/sizeof(*inc);
magma_int_t maxinc = 0;
for( i=0; i < ninc; ++i ) {
maxinc = max( maxinc, abs(inc[i]) );
}
printf( "!! Calling these CBLAS and Fortran BLAS sometimes crashes (segfaults), which !!\n"
"!! is why we use wrappers. It does not necesarily indicate a bug in MAGMA. !!\n"
"!! If MAGMA_WITH_MKL or __APPLE__ are defined, known failures are skipped. !!\n"
"\n" );
// tell user about disabled functions
#ifndef HAVE_CBLAS
printf( "n/a: HAVE_CBLAS not defined, so no cblas functions tested.\n\n" );
#endif
#if defined(MAGMA_WITH_MKL)
printf( "n/a: cblas_zdotc, cblas_zdotu, blasf77_zdotc, and blasf77_zdotu are disabled with MKL, due to segfaults.\n\n" );
#endif
#if defined(__APPLE__)
printf( "n/a: blasf77_zdotc and blasf77_zdotu are disabled on MacOS, due to segfaults.\n\n" );
#endif
printf( "%% Error w.r.t. Error w.r.t. Error w.r.t.\n"
"%% M N K incx incy Function CBLAS Fortran BLAS inline\n"
"%%====================================================================================\n" );
for( int itest = 0; itest < opts.ntest; ++itest ) {
if ( itest > 0 ) {
printf( "%%----------------------------------------------------------------------\n" );
}
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
// allocate matrices
// over-allocate so they can be any combination of
// {m,n,k} * {abs(incx), abs(incy)} by
// {m,n,k} * {abs(incx), abs(incy)}
maxn = max( max( m, n ), k ) * maxinc;
ld = max( 1, maxn );
size = ld*maxn;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, size );
TESTING_MALLOC_CPU( B, magmaDoubleComplex, size );
// initialize matrices
lapackf77_zlarnv( &ione, ISEED, &size, A );
lapackf77_zlarnv( &ione, ISEED, &size, B );
// ----- test DZASUM
for( int iincx = 0; iincx < ninc; ++iincx ) {
magma_int_t incx = inc[iincx];
for( int iincy = 0; iincy < ninc; ++iincy ) {
magma_int_t incy = inc[iincy];
// get one-norm of column j of A
if ( incx > 0 && incx == incy ) { // positive, no incy
error_cblas = 0;
error_fblas = 0;
error_inline = 0;
for( j=0; j < k; ++j ) {
x_m = magma_cblas_dzasum( m, A(0,j), incx );
#ifdef HAVE_CBLAS
x_c = cblas_dzasum( m, A(0,j), incx );
error_cblas = max( error_cblas, fabs(x_m - x_c) / fabs(m*x_c) );
#else
x_c = 0;
error_cblas = SKIPPED_FLAG;
#endif
x_f = blasf77_dzasum( &m, A(0,j), &incx );
error_fblas = max( error_fblas, fabs(x_m - x_f) / fabs(m*x_f) );
// inline implementation
x_i = 0;
for( i=0; i < m; ++i ) {
x_i += MAGMA_Z_ABS1( *A(i*incx,j) ); // |real(Aij)| + |imag(Aij)|
}
error_inline = max( error_inline, fabs(x_m - x_i) / fabs(m*x_i) );
//printf( "dzasum xm %.8e, xc %.8e, xf %.8e, xi %.8e\n", x_m, x_c, x_f, x_i );
}
output( "dzasum", m, n, k, incx, incy, error_cblas, error_fblas, error_inline );
}
}
}
printf( "\n" );
// ----- test DZNRM2
// get two-norm of column j of A
for( int iincx = 0; iincx < ninc; ++iincx ) {
magma_int_t incx = inc[iincx];
for( int iincy = 0; iincy < ninc; ++iincy ) {
magma_int_t incy = inc[iincy];
if ( incx > 0 && incx == incy ) { // positive, no incy
error_cblas = 0;
error_fblas = 0;
error_inline = 0;
for( j=0; j < k; ++j ) {
x_m = magma_cblas_dznrm2( m, A(0,j), incx );
#ifdef HAVE_CBLAS
x_c = cblas_dznrm2( m, A(0,j), incx );
error_cblas = max( error_cblas, fabs(x_m - x_c) / fabs(m*x_c) );
#else
x_c = 0;
error_cblas = SKIPPED_FLAG;
#endif
x_f = blasf77_dznrm2( &m, A(0,j), &incx );
error_fblas = max( error_fblas, fabs(x_m - x_f) / fabs(m*x_f) );
// inline implementation (poor -- doesn't scale)
x_i = 0;
for( i=0; i < m; ++i ) {
x_i += real( *A(i*incx,j) ) * real( *A(i*incx,j) )
+ imag( *A(i*incx,j) ) * imag( *A(i*incx,j) );
// same: real( conj( *A(i*incx,j) ) * *A(i*incx,j) );
}
x_i = sqrt( x_i );
error_inline = max( error_inline, fabs(x_m - x_i) / fabs(m*x_i) );
//printf( "dznrm2 xm %.8e, xc %.8e, xf %.8e, xi %.8e\n", x_m, x_c, x_f, x_i );
}
output( "dznrm2", m, n, k, incx, incy, error_cblas, error_fblas, error_inline );
}
}
}
printf( "\n" );
// ----- test ZDOTC
// dot columns, Aj^H Bj
for( int iincx = 0; iincx < ninc; ++iincx ) {
magma_int_t incx = inc[iincx];
for( int iincy = 0; iincy < ninc; ++iincy ) {
magma_int_t incy = inc[iincy];
error_cblas = 0;
error_fblas = 0;
error_inline = 0;
for( j=0; j < k; ++j ) {
// MAGMA implementation, not just wrapper
x2_m = magma_cblas_zdotc( m, A(0,j), incx, B(0,j), incy );
// crashes with MKL 11.1.2, ILP64
#if defined(HAVE_CBLAS) && ! defined(MAGMA_WITH_MKL)
#ifdef COMPLEX
cblas_zdotc_sub( m, A(0,j), incx, B(0,j), incy, &x2_c );
#else
x2_c = cblas_zdotc( m, A(0,j), incx, B(0,j), incy );
#endif
error_cblas = max( error_cblas, fabs(x2_m - x2_c) / fabs(m*x2_c) );
#else
x2_c = MAGMA_Z_ZERO;
error_cblas = SKIPPED_FLAG;
#endif
// crashes with MKL 11.2.3 and MacOS 10.9
#if (! defined(COMPLEX) || ! defined(MAGMA_WITH_MKL)) && ! defined(__APPLE__)
x2_f = blasf77_zdotc( &m, A(0,j), &incx, B(0,j), &incy );
error_fblas = max( error_fblas, fabs(x2_m - x2_f) / fabs(m*x2_f) );
#else
x2_f = MAGMA_Z_ZERO;
error_fblas = SKIPPED_FLAG;
#endif
// inline implementation
x2_i = MAGMA_Z_ZERO;
magma_int_t A_offset = (incx > 0 ? 0 : (-n + 1)*incx);
magma_int_t B_offset = (incy > 0 ? 0 : (-n + 1)*incy);
for( i=0; i < m; ++i ) {
x2_i += conj( *A(A_offset + i*incx,j) ) * *B(B_offset + i*incy,j);
}
error_inline = max( error_inline, fabs(x2_m - x2_i) / fabs(m*x2_i) );
//printf( "zdotc xm %.8e + %.8ei, xc %.8e + %.8ei, xf %.8e + %.8ei, xi %.8e + %.8ei\n",
// real(x2_m), imag(x2_m),
// real(x2_c), imag(x2_c),
// real(x2_f), imag(x2_f),
// real(x2_i), imag(x2_i) );
}
output( "zdotc", m, n, k, incx, incy, error_cblas, error_fblas, error_inline );
}
}
printf( "\n" );
// ----- test ZDOTU
// dot columns, Aj^T * Bj
for( int iincx = 0; iincx < ninc; ++iincx ) {
magma_int_t incx = inc[iincx];
for( int iincy = 0; iincy < ninc; ++iincy ) {
magma_int_t incy = inc[iincy];
error_cblas = 0;
error_fblas = 0;
error_inline = 0;
for( j=0; j < k; ++j ) {
// MAGMA implementation, not just wrapper
x2_m = magma_cblas_zdotu( m, A(0,j), incx, B(0,j), incy );
// crashes with MKL 11.1.2, ILP64
#if defined(HAVE_CBLAS) && ! defined(MAGMA_WITH_MKL)
#ifdef COMPLEX
cblas_zdotu_sub( m, A(0,j), incx, B(0,j), incy, &x2_c );
#else
x2_c = cblas_zdotu( m, A(0,j), incx, B(0,j), incy );
#endif
error_cblas = max( error_cblas, fabs(x2_m - x2_c) / fabs(m*x2_c) );
#else
x2_c = MAGMA_Z_ZERO;
error_cblas = SKIPPED_FLAG;
#endif
// crashes with MKL 11.2.3 and MacOS 10.9
#if (! defined(COMPLEX) || ! defined(MAGMA_WITH_MKL)) && ! defined(__APPLE__)
x2_f = blasf77_zdotu( &m, A(0,j), &incx, B(0,j), &incy );
error_fblas = max( error_fblas, fabs(x2_m - x2_f) / fabs(m*x2_f) );
#else
x2_f = MAGMA_Z_ZERO;
error_fblas = SKIPPED_FLAG;
#endif
// inline implementation
x2_i = MAGMA_Z_ZERO;
magma_int_t A_offset = (incx > 0 ? 0 : (-n + 1)*incx);
magma_int_t B_offset = (incy > 0 ? 0 : (-n + 1)*incy);
for( i=0; i < m; ++i ) {
x2_i += *A(A_offset + i*incx,j) * *B(B_offset + i*incy,j);
}
error_inline = max( error_inline, fabs(x2_m - x2_i) / fabs(m*x2_i) );
//printf( "zdotu xm %.8e + %.8ei, xc %.8e + %.8ei, xf %.8e + %.8ei, xi %.8e + %.8ei\n",
// real(x2_m), imag(x2_m),
// real(x2_c), imag(x2_c),
// real(x2_f), imag(x2_f),
// real(x2_i), imag(x2_i) );
}
output( "zdotu", m, n, k, incx, incy, error_cblas, error_fblas, error_inline );
}
}
// cleanup
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( B );
fflush( stdout );
} // itest, incx, incy
opts.cleanup();
TESTING_FINALIZE();
return gStatus;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgetrf
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error;
double *h_A;
magmaDouble_ptr d_A;
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, info, min_mn;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
if ( opts.check == 2 ) {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) |Ax-b|/(N*|A|*|x|)\n");
}
else {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) |PA-LU|/(N*|A|)\n");
}
printf("=========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
gflops = FLOPS_DGETRF( M, N ) / 1e9;
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
init_matrix( M, N, h_A, lda );
cpu_time = magma_wtime();
lapackf77_dgetrf(&M, &N, h_A, &lda, ipiv, &info);
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
init_matrix( M, N, h_A, lda );
magma_dsetmatrix( M, N, h_A, lda, d_A, ldda );
gpu_time = magma_wtime();
magma_dgetrf_gpu( M, N, d_A, ldda, ipiv, &info);
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) M, (int) N, gpu_perf, gpu_time );
}
if ( opts.check == 2 ) {
magma_dgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_residual( M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else if ( opts.check ) {
magma_dgetmatrix( M, N, d_A, ldda, h_A, lda );
error = get_LU_error( M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf(" --- \n");
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgetrf
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
magmaDoubleComplex *h_A, *h_R, *work;
magmaDoubleComplex_ptr d_A, dwork;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t N, n2, lda, ldda, info, lwork, ldwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex tmp;
double error, rwork[1];
magma_int_t *ipiv;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
// need looser bound (3000*eps instead of 30*eps) for tests
// TODO: should compute ||I - A*A^{-1}|| / (n*||A||*||A^{-1}||)
opts.tolerance = max( 3000., opts.tolerance );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / (N*||A||_F)\n");
printf("=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
ldwork = N * magma_get_zgetri_nb( N );
gflops = FLOPS_ZGETRI( N ) / 1e9;
// query for workspace size
lwork = -1;
lapackf77_zgetri( &N, NULL, &lda, NULL, &tmp, &lwork, &info );
if (info != 0)
printf("lapackf77_zgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lwork = int( MAGMA_Z_REAL( tmp ));
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_CPU( work, magmaDoubleComplex, lwork );
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
error = lapackf77_zlange( "f", &N, &N, h_A, &lda, rwork ); // norm(A)
/* Factor the matrix. Both MAGMA and LAPACK will use this factor. */
magma_zsetmatrix( N, N, h_A, lda, d_A, 0, ldda, opts.queue );
magma_zgetrf_gpu( N, N, d_A, 0, ldda, ipiv, opts.queue, &info );
magma_zgetmatrix( N, N, d_A, 0, ldda, h_A, lda, opts.queue );
if ( info != 0 )
printf("magma_zgetrf_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
// check for exact singularity
//h_A[ 10 + 10*lda ] = MAGMA_Z_MAKE( 0.0, 0.0 );
//magma_zsetmatrix( N, N, h_A, lda, d_A, 0, ldda, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zgetri_gpu( N, d_A, 0, ldda, ipiv, dwork, 0, ldwork, opts.queues2, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_zgetri_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_zgetmatrix( N, N, d_A, 0, ldda, h_R, lda, opts.queue );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zgetri( &N, h_A, &lda, ipiv, work, &lwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_zgetri returned error %d: %s.\n",
(int) info, magma_strerror( info ));
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
blasf77_zaxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
error = lapackf77_zlange( "f", &N, &N, h_R, &lda, rwork ) / (N*error);
printf( "%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf( "%5d --- ( --- ) %7.2f (%7.2f) ---\n",
(int) N, gpu_perf, gpu_time );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time;
magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
double *rwork;
magma_int_t lrwork;
#endif
/* Matrix size */
double *w1, *w2, result[2]={0,0};
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %s, range = %s, uplo = %s, opts.check = %d, fraction = %6.4f\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M GPU Time (sec)\n");
printf("============================\n");
magma_int_t threads = magma_get_parallel_numthreads();
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = magma_zbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_zbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, double, lrwork);
#endif
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
magma_zmake_hpd( N, h_B, N );
magma_zmake_hermitian( N, h_A, N );
magma_int_t m1 = 0;
double vl = 0;
double vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == MagmaRangeI) {
il = 1;
iu = (int) (opts.fraction*N);
}
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if (opts.warmup) {
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx_2stage(opts.itype, opts.jobz, range, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_zhegvdx_2stage(opts.itype, opts.jobz, range, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
gpu_time = magma_wtime() - gpu_time;
if ( opts.check && opts.jobz != MagmaNoVec ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvdx routine.
A x = lambda B x is solved
and the following 3 tests computed:
(1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
| A B Z - Z D | / ( |A||Z| N ) (itype = 2)
| B A Z - Z D | / ( |A||Z| N ) (itype = 3)
(2) | S(with V) - S(w/o V) | / | S |
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
double *rwork = h_work + N*N;
#endif
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork);
result[0] /= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_int_t m2 = m1;
lapackf77_zhegvd(&opts.itype, "N", lapack_uplo_const(opts.uplo), &N,
h_R, &N, h_S, &N, w2,
h_work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
double maxw=0, diff=0;
for(int j=0; j<m2; j++) {
maxw = max(maxw, fabs(w1[j]));
maxw = max(maxw, fabs(w2[j]));
diff = max(diff, fabs(w1[j] - w2[j]));
}
result[1] = diff / (m2*maxw);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %7.2f\n",
(int) N, (int) m1, gpu_time);
if ( opts.check && opts.jobz != MagmaNoVec ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if (opts.itype==1) {
printf(" | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
else if (opts.itype==2) {
printf(" | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
else if (opts.itype==3) {
printf(" | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
printf( " | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed"));
status += ! (result[0] < tol && result[1] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE();
return status;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
double magma_error, dev_error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t M, N, Xm, Ym, lda, ldda, sizeA, sizeX, sizeY;
magma_int_t incx = 1;
magma_int_t incy = 1;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ydev, *Ymagma;
magmaDoubleComplex_ptr dA, dX, dY;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% trans = %s\n", lapack_trans_const(opts.transA) );
#ifdef HAVE_CUBLAS
printf("%% M N MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
printf("%% M N %s Gflop/s (ms) CPU Gflop/s (ms) %s error\n",
g_platform_str, g_platform_str );
#endif
printf("%%==================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = M;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
gflops = FLOPS_ZGEMV( M, N ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
sizeA = lda*N;
sizeX = incx*Xm;
sizeY = incy*Ym;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ydev, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( M, N, A, lda, dA, ldda, opts.queue );
magma_zsetvector( Xm, X, incx, dX, incx, opts.queue );
magma_zsetvector( Ym, Y, incy, dY, incy, opts.queue );
dev_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasZgemv( opts.handle, cublas_trans_const(opts.transA),
M, N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
#else
magma_zgemv( opts.transA, M, N,
alpha, dA, ldda,
dX, incx,
beta, dY, incy );
#endif
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
dev_perf = gflops / dev_time;
magma_zgetvector( Ym, dY, incy, Ydev, incy, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetvector( Ym, Y, incy, dY, incy, opts.queue );
magma_time = magma_sync_wtime( opts.queue );
magmablas_zgemv( opts.transA, M, N, alpha, dA, ldda, dX, incx, beta, dY, incy, opts.queue );
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( Ym, dY, incy, Ymagma, incy, opts.queue );
#endif
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zgemv( lapack_trans_const(opts.transA), &M, &N,
&alpha, A, &lda,
X, &incx,
&beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
double Anorm = lapackf77_zlange( "F", &M, &N, A, &lda, work );
double Xnorm = lapackf77_zlange( "F", &Xm, &ione, X, &Xm, work );
blasf77_zaxpy( &Ym, &c_neg_one, Y, &incy, Ydev, &incy );
dev_error = lapackf77_zlange( "F", &Ym, &ione, Ydev, &Ym, work ) / (Anorm * Xnorm);
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &Ym, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "F", &Ym, &ione, Ymagma, &Ym, work ) / (Anorm * Xnorm);
bool okay = (magma_error < tol) && (dev_error < tol);
status += ! okay;
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,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(okay ? "ok" : "failed"));
#else
bool okay = (dev_error < tol);
status += ! okay;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(okay ? "ok" : "failed"));
#endif
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ydev );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgesv_gpu
*/
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_perf, cpu_time, gpu_perf, gpu_time;
double error, Rnorm, Anorm, Xnorm, *work;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_B, *h_X;
magmaDouble_ptr d_A, d_B;
magma_int_t *ipiv;
magma_int_t N, nrhs, lda, ldb, ldda, lddb, info, sizeA, sizeB;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
nrhs = opts.nrhs;
printf("%% N NRHS CPU Gflop/s (sec) GPU Gflop/s (sec) ||B - AX|| / N*||A||*||X||\n");
printf("%%===============================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldb = lda;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
lddb = ldda;
gflops = ( FLOPS_DGETRF( N, N ) + FLOPS_DGETRS( N, nrhs ) ) / 1e9;
TESTING_MALLOC_CPU( h_A, double, lda*N );
TESTING_MALLOC_CPU( h_B, double, ldb*nrhs );
TESTING_MALLOC_CPU( h_X, double, ldb*nrhs );
TESTING_MALLOC_CPU( work, double, N );
TESTING_MALLOC_CPU( ipiv, magma_int_t, N );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
TESTING_MALLOC_DEV( d_B, double, lddb*nrhs );
/* Initialize the matrices */
sizeA = lda*N;
sizeB = ldb*nrhs;
lapackf77_dlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B );
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda, opts.queue );
magma_dsetmatrix( N, nrhs, h_B, ldb, d_B, lddb, opts.queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_dgesv_gpu( N, nrhs, d_A, ldda, ipiv, d_B, lddb, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dgesv_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
//=====================================================================
// Residual
//=====================================================================
magma_dgetmatrix( N, nrhs, d_B, lddb, h_X, ldb, opts.queue );
Anorm = lapackf77_dlange("I", &N, &N, h_A, &lda, work);
Xnorm = lapackf77_dlange("I", &N, &nrhs, h_X, &ldb, work);
blasf77_dgemm( MagmaNoTransStr, MagmaNoTransStr, &N, &nrhs, &N,
&c_one, h_A, &lda,
h_X, &ldb,
&c_neg_one, h_B, &ldb);
Rnorm = lapackf77_dlange("I", &N, &nrhs, h_B, &ldb, work);
error = Rnorm/(N*Anorm*Xnorm);
status += ! (error < tol);
/* ====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dgesv( &N, &nrhs, h_A, &lda, ipiv, h_B, &ldb, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_dgesv returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) nrhs, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_X );
TESTING_FREE_CPU( work );
TESTING_FREE_CPU( ipiv );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cunmlq
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf, cpu_time;
float Cnorm, error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t ione = 1;
magma_int_t mm, m, n, k, size, info;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t nb, ldc, lda, lwork, lwork_max;
magmaFloatComplex *C, *R, *A, *W, *tau;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
// need slightly looser bound (60*eps instead of 30*eps) for some tests
opts.tolerance = max( 60., opts.tolerance );
float tol = opts.tolerance * lapackf77_slamch("E");
// test all combinations of input parameters
magma_side_t side [] = { MagmaLeft, MagmaRight };
magma_trans_t trans[] = { Magma_ConjTrans, MagmaNoTrans };
printf("%% M N K side trans CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F / ||QC||_F\n");
printf("%%==============================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iside = 0; iside < 2; ++iside ) {
for( int itran = 0; itran < 2; ++itran ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
nb = magma_get_cgelqf_nb( m, n );
ldc = m;
// A is k x m (left) or k x n (right)
mm = (side[iside] == MagmaLeft ? m : n);
lda = k;
gflops = FLOPS_CUNMLQ( m, n, k, side[iside] ) / 1e9;
if ( side[iside] == MagmaLeft && m < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=left and m < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
if ( side[iside] == MagmaRight && n < k ) {
printf( "%5d %5d %5d %4c %5c skipping because side=right and n < k\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ) );
continue;
}
// need at least 2*nb*nb for gelqf
lwork_max = max( max( m*nb, n*nb ), 2*nb*nb );
// this rounds it up slightly if needed to agree with lwork query
lwork_max = int( real( magma_cmake_lwork( lwork_max )));
TESTING_MALLOC_CPU( C, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( R, magmaFloatComplex, ldc*n );
TESTING_MALLOC_CPU( A, magmaFloatComplex, lda*mm );
TESTING_MALLOC_CPU( W, magmaFloatComplex, lwork_max );
TESTING_MALLOC_CPU( tau, magmaFloatComplex, k );
// C is full, m x n
size = ldc*n;
lapackf77_clarnv( &ione, ISEED, &size, C );
lapackf77_clacpy( "Full", &m, &n, C, &ldc, R, &ldc );
size = lda*mm;
lapackf77_clarnv( &ione, ISEED, &size, A );
// compute LQ factorization to get Householder vectors in A, tau
magma_cgelqf( k, mm, A, lda, tau, W, lwork_max, &info );
if (info != 0) {
printf("magma_cgelqf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_cunmlq( lapack_side_const( side[iside] ), lapack_trans_const( trans[itran] ),
&m, &n, &k,
A, &lda, tau, C, &ldc, W, &lwork_max, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapackf77_cunmlq returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
// query for workspace size
lwork = -1;
magma_cunmlq( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
if (info != 0) {
printf("magma_cunmlq (lwork query) returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
lwork = (magma_int_t) MAGMA_C_REAL( W[0] );
if ( lwork < 0 || lwork > lwork_max ) {
printf("Warning: optimal lwork %d > allocated lwork_max %d\n", (int) lwork, (int) lwork_max );
lwork = lwork_max;
}
gpu_time = magma_wtime();
magma_cunmlq( side[iside], trans[itran],
m, n, k,
A, lda, tau, R, ldc, W, lwork, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_cunmlq returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* =====================================================================
compute relative error |QC_magma - QC_lapack| / |QC_lapack|
=================================================================== */
size = ldc*n;
blasf77_caxpy( &size, &c_neg_one, C, &ione, R, &ione );
Cnorm = lapackf77_clange( "Fro", &m, &n, C, &ldc, work );
error = lapackf77_clange( "Fro", &m, &n, R, &ldc, work ) / (magma_ssqrt(m*n) * Cnorm);
printf( "%5d %5d %5d %4c %5c %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) m, (int) n, (int) k,
lapacke_side_const( side[iside] ),
lapacke_trans_const( trans[itran] ),
cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
TESTING_FREE_CPU( C );
TESTING_FREE_CPU( R );
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( W );
TESTING_FREE_CPU( tau );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}} // end iside, itran
printf( "\n" );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsyevd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
double *h_A, *h_R, *h_work;
double *w1, *w2;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork, lda, aux_iwork[1];
magma_int_t izero = 0;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double result[3], eps, aux_work[1];
eps = lapackf77_dlamch( "E" );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: jobz = %s, uplo = %s\n",
lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo));
printf(" N CPU Time (sec) GPU Time (sec)\n");
printf("=======================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
lda = N;
// query for workspace sizes
magma_dsyevd( opts.jobz, opts.uplo,
N, NULL, lda, NULL,
aux_work, -1,
aux_iwork, -1,
opts.queue, &info );
lwork = (magma_int_t) aux_work[0];
liwork = aux_iwork[0];
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, double, N*lda );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, double, N*lda );
TESTING_MALLOC_PIN( h_work, double, lwork );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* warm up run */
if ( opts.warmup ) {
magma_dsyevd( opts.jobz, opts.uplo,
N, h_R, lda, w1,
h_work, lwork,
iwork, liwork,
opts.queue, &info );
if (info != 0)
printf("magma_dsyevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_dsyevd( opts.jobz, opts.uplo,
N, h_R, lda, w1,
h_work, lwork,
iwork, liwork,
opts.queue, &info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_dsyevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zcds]drvst routine.
A is factored as A = U S U' and the following 3 tests computed:
(1) | A - U S U' | / ( |A| N )
(2) | I - U'U | / ( N )
(3) | S(with U) - S(w/o U) | / | S |
=================================================================== */
double temp1, temp2;
// tau=NULL is unused since itype=1
lapackf77_dsyt21( &ione, lapack_uplo_const(opts.uplo), &N, &izero,
h_A, &lda,
w1, h_work,
h_R, &lda,
h_R, &lda,
NULL, h_work, &result[0] );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
magma_dsyevd( MagmaNoVec, opts.uplo,
N, h_R, lda, w2,
h_work, lwork,
iwork, liwork,
opts.queue, &info );
if (info != 0)
printf("magma_dsyevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for( int j=0; j<N; j++ ) {
temp1 = max(temp1, fabs(w1[j]));
temp1 = max(temp1, fabs(w2[j]));
temp2 = max(temp2, fabs(w1[j]-w2[j]));
}
result[2] = temp2 / (((double)N)*temp1);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dsyevd( lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_A, &lda, w2,
h_work, &lwork,
iwork, &liwork,
&info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_dsyevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf("%5d %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
/* =====================================================================
Print execution time
=================================================================== */
if ( opts.check ) {
printf("Testing the factorization A = U S U' for correctness:\n");
printf("(1) | A - U S U' | / (|A| N) = %8.2e %s\n", result[0]*eps, (result[0]*eps < tol ? "ok" : "failed") );
printf("(2) | I - U'U | / N = %8.2e %s\n", result[1]*eps, (result[1]*eps < tol ? "ok" : "failed") );
printf("(3) | S(w/ U) - S(w/o U) | / |S| = %8.2e %s\n\n", result[2] , (result[2] < tolulp ? "ok" : "failed") );
status += ! (result[0]*eps < tol && result[1]*eps < tol && result[2] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
