extern "C" magma_int_t
magma_zpidr_merge(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_PIDRMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
solver_par->init_res = 0.0;
solver_par->final_res = 0.0;
solver_par->iter_res = 0.0;
solver_par->runtime = 0.0;
// constants
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magmaDoubleComplex c_n_one = MAGMA_Z_NEG_ONE;
// internal user parameters
const magma_int_t smoothing = 1; // 0 = disable, 1 = enable
const double angle = 0.7; // [0-1]
// local variables
magma_int_t iseed[4] = {0, 0, 0, 1};
magma_int_t dof;
magma_int_t s;
magma_int_t distr;
magma_int_t k, i, sk;
magma_int_t innerflag;
magma_int_t ldd;
double residual;
double nrm;
double nrmb;
double nrmr;
double nrmt;
double rho;
magmaDoubleComplex om;
magmaDoubleComplex gamma;
magmaDoubleComplex fk;
// matrices and vectors
magma_z_matrix dxs = {Magma_CSR};
magma_z_matrix dr = {Magma_CSR}, drs = {Magma_CSR};
magma_z_matrix dP = {Magma_CSR}, dP1 = {Magma_CSR};
magma_z_matrix dG = {Magma_CSR}, dGcol = {Magma_CSR};
magma_z_matrix dU = {Magma_CSR};
magma_z_matrix dM = {Magma_CSR}, hMdiag = {Magma_CSR};
magma_z_matrix df = {Magma_CSR};
magma_z_matrix dt = {Magma_CSR}, dtt = {Magma_CSR};
magma_z_matrix dc = {Magma_CSR};
magma_z_matrix dv = {Magma_CSR};
magma_z_matrix dlu = {Magma_CSR};
magma_z_matrix dskp = {Magma_CSR}, hskp = {Magma_CSR};
magma_z_matrix dalpha = {Magma_CSR}, halpha = {Magma_CSR};
magma_z_matrix dbeta = {Magma_CSR}, hbeta = {Magma_CSR};
magmaDoubleComplex *d1 = NULL, *d2 = NULL;
// chronometry
real_Double_t tempo1, tempo2;
// initial s space
// TODO: add option for 's' (shadow space number)
// Hack: uses '--restart' option as the shadow space number.
// This is not a good idea because the default value of restart option is used to detect
// if the user provided a custom restart. This means that if the default restart value
// is changed then the code will think it was the user (unless the default value is
// also updated in the 'if' statement below.
s = 1;
if ( solver_par->restart != 50 ) {
if ( solver_par->restart > A.num_cols ) {
s = A.num_cols;
} else {
s = solver_par->restart;
}
}
solver_par->restart = s;
// set max iterations
solver_par->maxiter = min( 2 * A.num_cols, solver_par->maxiter );
// check if matrix A is square
if ( A.num_rows != A.num_cols ) {
//printf("Matrix A is not square.\n");
info = MAGMA_ERR_NOT_SUPPORTED;
goto cleanup;
}
// |b|
nrmb = magma_dznrm2( b.num_rows, b.dval, 1, queue );
if ( nrmb == 0.0 ) {
magma_zscal( x->num_rows, MAGMA_Z_ZERO, x->dval, 1, queue );
info = MAGMA_SUCCESS;
goto cleanup;
}
// t = 0
// make t twice as large to contain both, dt and dr
ldd = magma_roundup( b.num_rows, 32 );
CHECK( magma_zvinit( &dt, Magma_DEV, ldd, 2, c_zero, queue ));
dt.num_rows = b.num_rows;
dt.num_cols = 1;
dt.nnz = dt.num_rows;
// redirect the dr.dval to the second part of dt
CHECK( magma_zvinit( &dr, Magma_DEV, b.num_rows, 1, c_zero, queue ));
magma_free( dr.dval );
dr.dval = dt.dval + ldd;
// r = b - A x
CHECK( magma_zresidualvec( A, b, *x, &dr, &nrmr, queue ));
// |r|
solver_par->init_res = nrmr;
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nrmr;
}
// check if initial is guess good enough
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
// P = randn(n, s)
// P = ortho(P)
//---------------------------------------
// P = 0.0
CHECK( magma_zvinit( &dP, Magma_CPU, A.num_cols, s, c_zero, queue ));
// P = randn(n, s)
distr = 3; // 1 = unif (0,1), 2 = unif (-1,1), 3 = normal (0,1)
dof = dP.num_rows * dP.num_cols;
lapackf77_zlarnv( &distr, iseed, &dof, dP.val );
// transfer P to device
CHECK( magma_zmtransfer( dP, &dP1, Magma_CPU, Magma_DEV, queue ));
magma_zmfree( &dP, queue );
// P = ortho(P1)
if ( dP1.num_cols > 1 ) {
// P = magma_zqr(P1), QR factorization
CHECK( magma_zqr( dP1.num_rows, dP1.num_cols, dP1, dP1.ld, &dP, NULL, queue ));
} else {
// P = P1 / |P1|
nrm = magma_dznrm2( dof, dP1.dval, 1, queue );
nrm = 1.0 / nrm;
magma_zdscal( dof, nrm, dP1.dval, 1, queue );
CHECK( magma_zmtransfer( dP1, &dP, Magma_DEV, Magma_DEV, queue ));
}
magma_zmfree( &dP1, queue );
//---------------------------------------
// allocate memory for the scalar products
CHECK( magma_zvinit( &hskp, Magma_CPU, 4, 1, c_zero, queue ));
CHECK( magma_zvinit( &dskp, Magma_DEV, 4, 1, c_zero, queue ));
CHECK( magma_zvinit( &halpha, Magma_CPU, s, 1, c_zero, queue ));
CHECK( magma_zvinit( &dalpha, Magma_DEV, s, 1, c_zero, queue ));
CHECK( magma_zvinit( &hbeta, Magma_CPU, s, 1, c_zero, queue ));
CHECK( magma_zvinit( &dbeta, Magma_DEV, s, 1, c_zero, queue ));
// workspace for merged dot product
CHECK( magma_zmalloc( &d1, max(2, s) * b.num_rows ));
CHECK( magma_zmalloc( &d2, max(2, s) * b.num_rows ));
// smoothing enabled
if ( smoothing > 0 ) {
// set smoothing solution vector
CHECK( magma_zmtransfer( *x, &dxs, Magma_DEV, Magma_DEV, queue ));
// tt = 0
// make tt twice as large to contain both, dtt and drs
ldd = magma_roundup( b.num_rows, 32 );
CHECK( magma_zvinit( &dtt, Magma_DEV, ldd, 2, c_zero, queue ));
dtt.num_rows = dr.num_rows;
dtt.num_cols = 1;
dtt.nnz = dtt.num_rows;
// redirect the drs.dval to the second part of dtt
CHECK( magma_zvinit( &drs, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
magma_free( drs.dval );
drs.dval = dtt.dval + ldd;
// set smoothing residual vector
magma_zcopyvector( dr.num_rows, dr.dval, 1, drs.dval, 1, queue );
}
// G(n,s) = 0
if ( s > 1 ) {
ldd = magma_roundup( A.num_rows, 32 );
CHECK( magma_zvinit( &dG, Magma_DEV, ldd, s, c_zero, queue ));
dG.num_rows = A.num_rows;
} else {
CHECK( magma_zvinit( &dG, Magma_DEV, A.num_rows, s, c_zero, queue ));
}
// dGcol represents a single column of dG, array pointer is set inside loop
CHECK( magma_zvinit( &dGcol, Magma_DEV, dG.num_rows, 1, c_zero, queue ));
magma_free( dGcol.dval );
// U(n,s) = 0
if ( s > 1 ) {
ldd = magma_roundup( A.num_cols, 32 );
CHECK( magma_zvinit( &dU, Magma_DEV, ldd, s, c_zero, queue ));
dU.num_rows = A.num_cols;
} else {
CHECK( magma_zvinit( &dU, Magma_DEV, A.num_cols, s, c_zero, queue ));
}
// M(s,s) = I
CHECK( magma_zvinit( &dM, Magma_DEV, s, s, c_zero, queue ));
CHECK( magma_zvinit( &hMdiag, Magma_CPU, s, 1, c_zero, queue ));
magmablas_zlaset( MagmaFull, dM.num_rows, dM.num_cols, c_zero, c_one, dM.dval, dM.ld, queue );
// f = 0
CHECK( magma_zvinit( &df, Magma_DEV, dP.num_cols, 1, c_zero, queue ));
// c = 0
CHECK( magma_zvinit( &dc, Magma_DEV, dM.num_cols, 1, c_zero, queue ));
// v = 0
CHECK( magma_zvinit( &dv, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
// lu = 0
CHECK( magma_zvinit( &dlu, Magma_DEV, dr.num_rows, 1, c_zero, queue ));
//--------------START TIME---------------
// chronometry
tempo1 = magma_sync_wtime( queue );
if ( solver_par->verbose > 0 ) {
solver_par->timing[0] = 0.0;
}
om = MAGMA_Z_ONE;
innerflag = 0;
// start iteration
do
{
solver_par->numiter++;
// new RHS for small systems
// f = P' r
magma_zgemvmdot_shfl( dP.num_rows, dP.num_cols, dP.dval, dr.dval, d1, d2, df.dval, queue );
// shadow space loop
for ( k = 0; k < s; ++k ) {
sk = s - k;
// c(k:s) = M(k:s,k:s) \ f(k:s)
magma_zcopyvector( sk, &df.dval[k], 1, &dc.dval[k], 1, queue );
magma_ztrsv( MagmaLower, MagmaNoTrans, MagmaNonUnit, sk, &dM.dval[k*dM.ld+k], dM.ld, &dc.dval[k], 1, queue );
// v = r - G(:,k:s) c(k:s)
magma_zcopyvector( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
magmablas_zgemv( MagmaNoTrans, dG.num_rows, sk, c_n_one, &dG.dval[k*dG.ld], dG.ld, &dc.dval[k], 1, c_one, dv.dval, 1, queue );
// preconditioning operation
// v = L \ v;
// v = U \ v;
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, dv, &dlu, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, dlu, &dv, precond_par, queue ));
// U(:,k) = om * v + U(:,k:s) c(k:s)
magmablas_zgemv( MagmaNoTrans, dU.num_rows, sk, c_one, &dU.dval[k*dU.ld], dU.ld, &dc.dval[k], 1, om, dv.dval, 1, queue );
magma_zcopyvector( dU.num_rows, dv.dval, 1, &dU.dval[k*dU.ld], 1, queue );
// G(:,k) = A U(:,k)
dGcol.dval = dG.dval + k * dG.ld;
CHECK( magma_z_spmv( c_one, A, dv, c_zero, dGcol, queue ));
solver_par->spmv_count++;
// bi-orthogonalize the new basis vectors
for ( i = 0; i < k; ++i ) {
// alpha = P(:,i)' G(:,k)
halpha.val[i] = magma_zdotc( dP.num_rows, &dP.dval[i*dP.ld], 1, &dG.dval[k*dG.ld], 1, queue );
// alpha = alpha / M(i,i)
halpha.val[i] = halpha.val[i] / hMdiag.val[i];
// G(:,k) = G(:,k) - alpha * G(:,i)
magma_zaxpy( dG.num_rows, -halpha.val[i], &dG.dval[i*dG.ld], 1, &dG.dval[k*dG.ld], 1, queue );
}
// non-first s iteration
if ( k > 0 ) {
// U update outside of loop using GEMV
// U(:,k) = U(:,k) - U(:,1:k) * alpha(1:k)
magma_zsetvector( k, halpha.val, 1, dalpha.dval, 1, queue );
magmablas_zgemv( MagmaNoTrans, dU.num_rows, k, c_n_one, dU.dval, dU.ld, dalpha.dval, 1, c_one, &dU.dval[k*dU.ld], 1, queue );
}
// new column of M = P'G, first k-1 entries are zero
// M(k:s,k) = P(:,k:s)' G(:,k)
magma_zgemvmdot_shfl( dP.num_rows, sk, &dP.dval[k*dP.ld], &dG.dval[k*dG.ld], d1, d2, &dM.dval[k*dM.ld+k], queue );
magma_zgetvector( 1, &dM.dval[k*dM.ld+k], 1, &hMdiag.val[k], 1, queue );
// check M(k,k) == 0
if ( MAGMA_Z_EQUAL(hMdiag.val[k], MAGMA_Z_ZERO) ) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// beta = f(k) / M(k,k)
magma_zgetvector( 1, &df.dval[k], 1, &fk, 1, queue );
hbeta.val[k] = fk / hMdiag.val[k];
// check for nan
if ( magma_z_isnan( hbeta.val[k] ) || magma_z_isinf( hbeta.val[k] )) {
innerflag = 1;
info = MAGMA_DIVERGENCE;
break;
}
// r = r - beta * G(:,k)
magma_zaxpy( dr.num_rows, -hbeta.val[k], &dG.dval[k*dG.ld], 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// |r|
nrmr = magma_dznrm2( dr.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// x = x + beta * U(:,k)
magma_zaxpy( x->num_rows, hbeta.val[k], &dU.dval[k*dU.ld], 1, x->dval, 1, queue );
// smoothing operation
//---------------------------------------
// t = rs - r
magma_zidr_smoothing_1( drs.num_rows, drs.num_cols, drs.dval, dr.dval, dtt.dval, queue );
// t't
// t'rs
CHECK( magma_zgemvmdot_shfl( dt.ld, 2, dtt.dval, dtt.dval, d1, d2, &dskp.dval[2], queue ));
magma_zgetvector( 2, &dskp.dval[2], 1, &hskp.val[2], 1, queue );
// gamma = (t' * rs) / (t' * t)
gamma = hskp.val[3] / hskp.val[2];
// rs = rs - gamma * (rs - r)
magma_zaxpy( drs.num_rows, -gamma, dtt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_zidr_smoothing_2( dxs.num_rows, dxs.num_cols, -gamma, x->dval, dxs.dval, queue );
// |rs|
nrmr = magma_dznrm2( drs.num_rows, drs.dval, 1, queue );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter) % solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)nrmr;
solver_par->timing[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)tempo2 - tempo1;
}
}
// check convergence or iteration limit
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
s = k + 1; // for the x-update outside the loop
innerflag = 2;
info = MAGMA_SUCCESS;
break;
}
// non-last s iteration
if ( (k + 1) < s ) {
// f(k+1:s) = f(k+1:s) - beta * M(k+1:s,k)
magma_zaxpy( sk-1, -hbeta.val[k], &dM.dval[k*dM.ld+(k+1)], 1, &df.dval[k+1], 1, queue );
}
}
// smoothing disabled
if ( smoothing <= 0 && innerflag != 1 ) {
// update solution approximation x
// x = x + U(:,1:s) * beta(1:s)
magma_zsetvector( s, hbeta.val, 1, dbeta.dval, 1, queue );
magmablas_zgemv( MagmaNoTrans, dU.num_rows, s, c_one, dU.dval, dU.ld, dbeta.dval, 1, c_one, x->dval, 1, queue );
}
// check convergence or iteration limit or invalid result of inner loop
if ( innerflag > 0 ) {
break;
}
// v = r
magma_zcopy( dr.num_rows, dr.dval, 1, dv.dval, 1, queue );
// preconditioning operation
// v = L \ v;
// v = U \ v;
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, dv, &dlu, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, dlu, &dv, precond_par, queue ));
// t = A v
CHECK( magma_z_spmv( c_one, A, dv, c_zero, dt, queue ));
solver_par->spmv_count++;
// computation of a new omega
//---------------------------------------
// t't
// t'r
CHECK( magma_zgemvmdot_shfl( dt.ld, 2, dt.dval, dt.dval, d1, d2, dskp.dval, queue ));
magma_zgetvector( 2, dskp.dval, 1, hskp.val, 1, queue );
// |t|
nrmt = magma_dsqrt( MAGMA_Z_REAL(hskp.val[0]) );
// rho = abs((t' * r) / (|t| * |r|))
rho = MAGMA_D_ABS( MAGMA_Z_REAL(hskp.val[1]) / (nrmt * nrmr) );
// om = (t' * r) / (|t| * |t|)
om = hskp.val[1] / hskp.val[0];
if ( rho < angle ) {
om = (om * angle) / rho;
}
//---------------------------------------
if ( MAGMA_Z_EQUAL(om, MAGMA_Z_ZERO) ) {
info = MAGMA_DIVERGENCE;
break;
}
// update approximation vector
// x = x + om * v
magma_zaxpy( x->num_rows, om, dv.dval, 1, x->dval, 1, queue );
// update residual vector
// r = r - om * t
magma_zaxpy( dr.num_rows, -om, dt.dval, 1, dr.dval, 1, queue );
// smoothing disabled
if ( smoothing <= 0 ) {
// residual norm
nrmr = magma_dznrm2( dr.num_rows, dr.dval, 1, queue );
// smoothing enabled
} else {
// smoothing operation
//---------------------------------------
// t = rs - r
magma_zidr_smoothing_1( drs.num_rows, drs.num_cols, drs.dval, dr.dval, dtt.dval, queue );
// t't
// t'rs
CHECK( magma_zgemvmdot_shfl( dt.ld, 2, dtt.dval, dtt.dval, d1, d2, &dskp.dval[2], queue ));
magma_zgetvector( 2, &dskp.dval[2], 1, &hskp.val[2], 1, queue );
// gamma = (t' * rs) / (t' * t)
gamma = hskp.val[3] / hskp.val[2];
// rs = rs - gamma * (rs - r)
magma_zaxpy( drs.num_rows, -gamma, dtt.dval, 1, drs.dval, 1, queue );
// xs = xs - gamma * (xs - x)
magma_zidr_smoothing_2( dxs.num_rows, dxs.num_cols, -gamma, x->dval, dxs.dval, queue );
// |rs|
nrmr = magma_dznrm2( drs.num_rows, drs.dval, 1, queue );
//---------------------------------------
}
// store current timing and residual
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter) % solver_par->verbose == 0 ) {
solver_par->res_vec[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)nrmr;
solver_par->timing[(solver_par->numiter) / solver_par->verbose]
= (real_Double_t)tempo2 - tempo1;
}
}
// check convergence
if ( nrmr <= solver_par->atol ||
nrmr/nrmb <= solver_par->rtol ) {
info = MAGMA_SUCCESS;
break;
}
}
while ( solver_par->numiter + 1 <= solver_par->maxiter );
// smoothing enabled
if ( smoothing > 0 ) {
// x = xs
magma_zcopyvector( x->num_rows, dxs.dval, 1, x->dval, 1, queue );
// r = rs
magma_zcopyvector( dr.num_rows, drs.dval, 1, dr.dval, 1, queue );
}
// get last iteration timing
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t)tempo2 - tempo1;
//--------------STOP TIME----------------
// get final stats
solver_par->iter_res = nrmr;
CHECK( magma_zresidualvec( A, b, *x, &dr, &residual, queue ));
solver_par->final_res = residual;
// set solver conclusion
if ( info != MAGMA_SUCCESS && info != MAGMA_DIVERGENCE ) {
if ( solver_par->init_res > solver_par->final_res ) {
info = MAGMA_SLOW_CONVERGENCE;
}
}
cleanup:
// free resources
// smoothing enabled
if ( smoothing > 0 ) {
drs.dval = NULL; // needed because its pointer is redirected to dtt
magma_zmfree( &dxs, queue );
magma_zmfree( &drs, queue );
magma_zmfree( &dtt, queue );
}
dr.dval = NULL; // needed because its pointer is redirected to dt
dGcol.dval = NULL; // needed because its pointer is redirected to dG
magma_zmfree( &dr, queue );
magma_zmfree( &dP, queue );
magma_zmfree( &dP1, queue );
magma_zmfree( &dG, queue );
magma_zmfree( &dGcol, queue );
magma_zmfree( &dU, queue );
magma_zmfree( &dM, queue );
magma_zmfree( &hMdiag, queue );
magma_zmfree( &df, queue );
magma_zmfree( &dt, queue );
magma_zmfree( &dc, queue );
magma_zmfree( &dv, queue );
magma_zmfree( &dlu, queue );
magma_zmfree( &dskp, queue );
magma_zmfree( &dalpha, queue );
magma_zmfree( &dbeta, queue );
magma_zmfree( &hskp, queue );
magma_zmfree( &halpha, queue );
magma_zmfree( &hbeta, queue );
magma_free( d1 );
magma_free( d2 );
solver_par->info = info;
return info;
/* magma_zpidr_merge */
}
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;
}
int main(int argc, char **argv)
{
TESTING_CUDA_INIT();
magma_timestr_t start, end;
double flops, magma_perf, cuda_perf, error, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
cuDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
FILE *fp ;
magma_int_t i, lda, Xm, Ym;
magma_int_t M, M0 = 0;
magma_int_t N, N0 = 0;
magma_int_t szeA, szeX, szeY;
magma_int_t istart = 64;
magma_int_t iend = 10240;
magma_int_t incx = 1;
magma_int_t incy = 1;
char trans = MagmaNoTrans;
cuDoubleComplex alpha = MAGMA_Z_MAKE(1., 0.); // MAGMA_Z_MAKE( 1.5, -2.3 );
cuDoubleComplex beta = MAGMA_Z_MAKE(0., 0.); // MAGMA_Z_MAKE( -0.6, 0.8 );
cuDoubleComplex *A, *X, *Y, *Ycublas, *Ymagma;
cuDoubleComplex *dA, *dX, *dY;
if (argc != 1){
for(i=1; i<argc; i++){
if ( strcmp("-n", argv[i]) == 0 ){
N0 = atoi(argv[++i]);
}
else if ( strcmp("-m", argv[i]) == 0 ){
M0 = atoi(argv[++i]);
}
else if (strcmp("-N", argv[i])==0){
trans = MagmaNoTrans;
}
else if (strcmp("-T", argv[i])==0){
trans = MagmaTrans;
}
#if defined(PRECISION_z) || defined(PRECISION_c)
else if (strcmp("-C", argv[i])==0){
trans = MagmaConjTrans;
}
#endif
}
}
if ( (M0 != 0) && (N0 != 0) )
iend = istart + 1;
M = N = iend;
if ( M0 != 0 ) M = M0;
if ( N0 != 0 ) N = N0;
if( trans == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
lda = ((M+31)/32)*32;
szeA = lda*N;
szeX = incx*Xm;
szeY = incy*Ym;
TESTING_MALLOC( A, cuDoubleComplex, szeA );
TESTING_MALLOC( X, cuDoubleComplex, szeX );
TESTING_MALLOC( Y, cuDoubleComplex, szeY );
TESTING_MALLOC( Ycublas, cuDoubleComplex, szeY );
TESTING_MALLOC( Ymagma, cuDoubleComplex, szeY );
TESTING_DEVALLOC( dA, cuDoubleComplex, szeA );
TESTING_DEVALLOC( dX, cuDoubleComplex, szeX );
TESTING_DEVALLOC( dY, cuDoubleComplex, szeY );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &szeA, A );
lapackf77_zlarnv( &ione, ISEED, &szeX, X );
lapackf77_zlarnv( &ione, ISEED, &szeY, Y );
fp = fopen ("results_zgemv.txt", "w") ;
if( fp == NULL ){ printf("Couldn't open output file\n"); exit(1);}
printf("\nUsage: \n");
printf(" testing_zgemv [-N|T|C] [-m %d] [-n %d]\n\n", 1024, 1024);
printf( " m n CUBLAS,Gflop/s MAGMABLAS Gflop/s \"error\"\n"
"==============================================================\n");
fprintf(fp, " m n CUBLAS,Gflop/s MAGMABLAS Gflop/s \"error\"\n"
"==============================================================\n");
for( i=istart; i < iend; i = (int)((i+1)*1.1) )
{
M = N = i;
if ( M0 != 0 ) M = M0;
if ( N0 != 0 ) N = N0;
if( trans == MagmaNoTrans ) {
Xm = N;
Ym = M;
} else {
Xm = M;
Ym = N;
}
lda = ((M+31)/32)*32;
flops = FLOPS( (double)M, (double)N ) / 1000000;
printf( "%5d %5d ", (int) M, (int) N );
fprintf( fp, "%5d %5d ", (int) M, (int) N );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_zsetmatrix( M, N, A, lda, dA, lda );
magma_zsetvector( Xm, X, incx, dX, incx );
magma_zsetvector( Ym, Y, incy, dY, incy );
/*
* Cublas Version
*/
start = get_current_time();
cublasZgemv( trans, M, N, alpha, dA, lda, dX, incx, beta, dY, incy );
end = get_current_time();
magma_zgetvector( Ym, dY, incy, Ycublas, incy );
cuda_perf = flops / GetTimerValue(start, end);
printf( "%11.2f", cuda_perf );
fprintf(fp, "%11.2f", cuda_perf );
/*
* Magma Version
*/
magma_zsetvector( Ym, Y, incy, dY, incy );
start = get_current_time();
magmablas_zgemv( trans, M, N, alpha, dA, lda, dX, incx, beta, dY, incy );
end = get_current_time();
magma_zgetvector( Ym, dY, incx, Ymagma, incx );
magma_perf = flops / GetTimerValue(start, end);
printf( "%11.2f", magma_perf );
fprintf(fp, "%11.2f", magma_perf );
/* =====================================================================
Computing the Difference Cublas VS Magma
=================================================================== */
blasf77_zaxpy( &Ym, &c_neg_one, Ymagma, &incy, Ycublas, &incy);
error = lapackf77_zlange( "M", &Ym, &ione, Ycublas, &Ym, work );
#if 0
printf( "\t\t %8.6e", error / (double)Ym );
fprintf( fp, "\t\t %8.6e", error / (double)Ym );
/*
* Blas comparaison
*/
{
char *blastrans = MagmaNoTransStr;
if ( trans == MagmaConjTrans )
blastrans = MagmaConjTransStr;
else if ( trans == MagmaTrans )
blastrans = MagmaTransStr;
blasf77_zcopy( &Ym, Y, &incy, Ycublas, &incy );
blasf77_zgemv( blastrans, &M, &N,
&alpha, A, &lda,
X, &incx,
&beta, Ycublas, &incy );
blasf77_zaxpy( &Ym, &c_neg_one, Ymagma, &incy, Ycublas, &incy);
error = lapackf77_zlange( "M", &Ym, &ione, Ycublas, &Ym, work );
}
#endif
printf( "\t\t %8.6e\n", error / (double)Ym );
fprintf( fp, "\t\t %8.6e\n", error / (double)Ym );
}
/* Free Memory */
TESTING_FREE( A );
TESTING_FREE( X );
TESTING_FREE( Y );
TESTING_FREE( Ycublas );
TESTING_FREE( Ymagma );
TESTING_DEVFREE( dA );
TESTING_DEVFREE( dX );
TESTING_DEVFREE( dY );
/* Free device */
TESTING_CUDA_FINALIZE();
return EXIT_SUCCESS;
}
extern "C" magma_int_t
magma_z_spmv(
magmaDoubleComplex alpha,
magma_z_sparse_matrix A,
magma_z_vector x,
magmaDoubleComplex beta,
magma_z_vector y,
magma_queue_t queue )
{
// set queue for old dense routines
magma_queue_t orig_queue;
magmablasGetKernelStream( &orig_queue );
if ( A.memory_location != x.memory_location ||
x.memory_location != y.memory_location ) {
printf("error: linear algebra objects are not located in same memory!\n");
printf("memory locations are: %d %d %d\n",
A.memory_location, x.memory_location, y.memory_location );
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_INVALID_PTR;
}
// DEV case
if ( A.memory_location == Magma_DEV ) {
if ( A.num_cols == x.num_rows && x.num_cols == 1 ) {
if ( A.storage_type == Magma_CSR
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ) {
//printf("using CSR kernel for SpMV: ");
//magma_zgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
// A.dval, A.drow, A.dcol, x.dval, beta, y.dval );
//printf("done.\n");
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
cusparseZcsrmv( cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_cols, A.nnz, &alpha, descr,
A.dval, A.drow, A.dcol, x.dval, &beta, y.dval );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
}
else if ( A.storage_type == Magma_ELL ) {
//printf("using ELLPACKT kernel for SpMV: ");
magma_zgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, x.dval, beta,
y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_ELLPACKT ) {
//printf("using ELL kernel for SpMV: ");
magma_zgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, x.dval, beta,
y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_ELLRT ) {
//printf("using ELLRT kernel for SpMV: ");
magma_zgeellrtmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.dval, A.dcol, A.drow, x.dval,
beta, y.dval, A.alignment, A.blocksize, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_SELLP ) {
//printf("using SELLP kernel for SpMV: ");
magma_zgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x.dval, beta, y.dval, queue );
//printf("done.\n");
}
else if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_zgemv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
A.dval, A.num_rows, x.dval, 1, beta, y.dval,
1 );
//printf("done.\n");
}
else if ( A.storage_type == Magma_SPMVFUNCTION ) {
//printf("using DENSE kernel for SpMV: ");
magma_zcustomspmv( alpha, x, beta, y, queue );
//printf("done.\n");
}
/* else if ( A.storage_type == Magma_BCSR ) {
//printf("using CUSPARSE BCSR kernel for SpMV: ");
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
// end CUSPARSE context //
cusparseDirection_t dirA = CUSPARSE_DIRECTION_ROW;
int mb = (A.num_rows + A.blocksize-1)/A.blocksize;
int nb = (A.num_cols + A.blocksize-1)/A.blocksize;
cusparseZbsrmv( cusparseHandle, dirA,
CUSPARSE_OPERATION_NON_TRANSPOSE, mb, nb, A.numblocks,
&alpha, descr, A.dval, A.drow, A.dcol, A.blocksize, x.dval,
&beta, y.dval );
//printf("done.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
else if ( A.num_cols < x.num_rows || x.num_cols > 1 ) {
magma_int_t num_vecs = x.num_rows / A.num_cols * x.num_cols;
if ( A.storage_type == Magma_CSR ) {
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseSetStream( cusparseHandle, queue );
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
if ( x.major == MagmaColMajor) {
cusparseZcsrmm(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, num_vecs, A.num_cols, A.nnz,
&alpha, descr, A.dval, A.drow, A.dcol,
x.dval, A.num_cols, &beta, y.dval, A.num_cols);
} else if ( x.major == MagmaRowMajor) {
cusparseZcsrmm2(cusparseHandle,
CUSPARSE_OPERATION_NON_TRANSPOSE,
CUSPARSE_OPERATION_TRANSPOSE,
A.num_rows, num_vecs, A.num_cols, A.nnz,
&alpha, descr, A.dval, A.drow, A.dcol,
x.dval, A.num_cols, &beta, y.dval, A.num_cols);
}
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
}
else if ( A.storage_type == Magma_ELL ) {
if ( x.major == MagmaColMajor) {
magma_zmgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
}
else if ( x.major == MagmaRowMajor) {
// transpose first to col major
magma_z_vector x2;
magma_zvtranspose( x, &x2, queue );
magma_zmgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
magma_z_vfree(&x2, queue );
}
}
else if ( A.storage_type == Magma_ELLPACKT ) {
if ( x.major == MagmaColMajor) {
magma_zmgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
}
else if ( x.major == MagmaRowMajor) {
// transpose first to col major
magma_z_vector x2;
magma_zvtranspose( x, &x2, queue );
magma_zmgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.dval, A.dcol, x.dval,
beta, y.dval, queue );
magma_z_vfree(&x2, queue );
}
} else if ( A.storage_type == Magma_SELLP ) {
if ( x.major == MagmaRowMajor) {
magma_zmgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x.dval, beta, y.dval, queue );
}
else if ( x.major == MagmaColMajor) {
// transpose first to row major
magma_z_vector x2;
magma_zvtranspose( x, &x2, queue );
magma_zmgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.dval, A.dcol, A.drow, x2.dval, beta, y.dval, queue );
magma_z_vfree(&x2, queue );
}
}/*
if ( A.storage_type == Magma_DENSE ) {
//printf("using DENSE kernel for SpMV: ");
magmablas_zmgemv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.dval, A.num_rows, x.dval, 1,
beta, y.dval, 1 );
//printf("done.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
}
}
// CPU case missing!
else {
printf("error: CPU not yet supported.\n");
magmablasSetKernelStream( orig_queue );
return MAGMA_ERR_NOT_SUPPORTED;
}
magmablasSetKernelStream( orig_queue );
return MAGMA_SUCCESS;
}
magma_int_t
magma_z_spmv( magmaDoubleComplex alpha, magma_z_sparse_matrix A,
magma_z_vector x, magmaDoubleComplex beta, magma_z_vector y )
{
if( A.memory_location != x.memory_location ||
x.memory_location != y.memory_location ){
printf("error: linear algebra objects are not located in same memory!\n");
printf("memory locations are: %d %d %d\n",
A.memory_location, x.memory_location, y.memory_location );
return MAGMA_ERR_INVALID_PTR;
}
// DEV case
if( A.memory_location == Magma_DEV ){
if( A.num_cols == x.num_rows ){
if( A.storage_type == Magma_CSR
|| A.storage_type == Magma_CSRL
|| A.storage_type == Magma_CSRU ){
//printf("using CSR kernel for SpMV: ");
//magma_zgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
// A.val, A.row, A.col, x.val, beta, y.val );
//printf("done.\n");
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);
cusparseZcsrmv( cusparseHandle,CUSPARSE_OPERATION_NON_TRANSPOSE,
A.num_rows, A.num_cols, A.nnz, &alpha, descr,
A.val, A.row, A.col, x.val, &beta, y.val );
cusparseDestroyMatDescr( descr );
cusparseDestroy( cusparseHandle );
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELLPACK ){
//printf("using ELLPACK kernel for SpMV: ");
magma_zgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.val, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELL ){
//printf("using ELL kernel for SpMV: ");
magma_zgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.val, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELLRT ){
//printf("using ELLRT kernel for SpMV: ");
magma_zgeellrtmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.max_nnz_row, alpha, A.val, A.col, A.row, x.val,
beta, y.val, A.alignment, A.blocksize );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_SELLC ){
//printf("using SELLC kernel for SpMV: ");
magma_zgesellcmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.val, A.col, A.row, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_SELLP ){
//printf("using SELLP kernel for SpMV: ");
magma_zgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
A.blocksize, A.numblocks, A.alignment,
alpha, A.val, A.col, A.row, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_DENSE ){
//printf("using DENSE kernel for SpMV: ");
magmablas_zgemv( MagmaNoTrans, A.num_rows, A.num_cols, alpha,
A.val, A.num_rows, x.val, 1, beta, y.val, 1 );
//printf("done.\n");
return MAGMA_SUCCESS;
}
/* else if( A.storage_type == Magma_BCSR ){
//printf("using CUSPARSE BCSR kernel for SpMV: ");
// CUSPARSE context //
cusparseHandle_t cusparseHandle = 0;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&cusparseHandle);
cusparseMatDescr_t descr = 0;
cusparseStatus = cusparseCreateMatDescr(&descr);
// end CUSPARSE context //
cusparseDirection_t dirA = CUSPARSE_DIRECTION_ROW;
int mb = (A.num_rows + A.blocksize-1)/A.blocksize;
int nb = (A.num_cols + A.blocksize-1)/A.blocksize;
cusparseZbsrmv( cusparseHandle, dirA,
CUSPARSE_OPERATION_NON_TRANSPOSE, mb, nb, A.numblocks,
&alpha, descr, A.val, A.row, A.col, A.blocksize, x.val,
&beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
return MAGMA_ERR_NOT_SUPPORTED;
}
}
else if( A.num_cols < x.num_rows ){
magma_int_t num_vecs = x.num_rows / A.num_cols;
if( A.storage_type == Magma_CSR ){
//printf("using CSR kernel for SpMV: ");
magma_zmgecsrmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.val, A.row, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELLPACK ){
//printf("using ELLPACK kernel for SpMV: ");
magma_zmgeellmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.val, A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}
else if( A.storage_type == Magma_ELL ){
//printf("using ELL kernel for SpMV: ");
magma_zmgeelltmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.max_nnz_row, alpha, A.val,
A.col, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}else if( A.storage_type == Magma_SELLP ){
//printf("using SELLP kernel for SpMV: ");
magma_zmgesellpmv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, A.blocksize, A.numblocks, A.alignment,
alpha, A.val, A.col, A.row, x.val, beta, y.val );
//printf("done.\n");
return MAGMA_SUCCESS;
}/*
if( A.storage_type == Magma_DENSE ){
//printf("using DENSE kernel for SpMV: ");
magmablas_zmgemv( MagmaNoTrans, A.num_rows, A.num_cols,
num_vecs, alpha, A.val, A.num_rows, x.val, 1,
beta, y.val, 1 );
//printf("done.\n");
return MAGMA_SUCCESS;
}*/
else {
printf("error: format not supported.\n");
return MAGMA_ERR_NOT_SUPPORTED;
}
}
}
// CPU case missing!
else{
printf("error: CPU not yet supported.\n");
return MAGMA_ERR_NOT_SUPPORTED;
}
return MAGMA_SUCCESS;
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time;
double magma_error, cublas_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, 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, *Ycublas, *Ymagma;
magmaDoubleComplex *dA, *dX, *dY;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("trans = %s\n", lapack_trans_const(opts.transA) );
printf(" M N MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n");
printf("===================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
lda = ((M+31)/32)*32;
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( Ycublas, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, sizeA );
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, lda );
magma_zsetvector( Xm, X, incx, dX, incx );
magma_zsetvector( Ym, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasZgemv( handle, cublas_trans_const(opts.transA),
M, N, &alpha, dA, lda, dX, incx, &beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( Ym, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetvector( Ym, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
magmablas_zgemv( opts.transA, M, N, alpha, dA, lda, dX, incx, beta, dY, incy );
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( Ym, dY, incy, Ymagma, incy );
/* =====================================================================
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
=================================================================== */
blasf77_zaxpy( &Ym, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &Ym, &ione, Ymagma, &Ym, work ) / Ym;
blasf77_zaxpy( &Ym, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "M", &Ym, &ione, Ycublas, &Ym, work ) / Ym;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error,
(magma_error < tol && cublas_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && cublas_error < tol);
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- testing zdot
*/
int main( int argc, char** argv)
{
TESTING_INIT();
printf("#================================================================================================================================================\n");
printf("\n");
printf(" | runtime | GFLOPS\n");
printf("#n num_vecs | CUDOT CUGEMV MAGMAGEMV MDOT MDGM | CUDOT CUGEMV MAGMAGEMV MDOT MDGM \n");
printf("#------------------------------------------------------------------------------------------------------------------------------------------------\n");
printf("\n");
for( magma_int_t num_vecs=5; num_vecs<6; num_vecs+=1 ){
for( magma_int_t n=10000; n<100000001; n=n+10000 ){
magma_z_sparse_matrix A, B, C, D, E, F, G, H, I, J, K, Z;
magma_z_vector a,b,c,x, y, z, skp;
int iters = 10;
double computations = (2.* n * iters * num_vecs);
magmaDoubleComplex one = MAGMA_Z_MAKE(1.0, 0.0);
magmaDoubleComplex zero = MAGMA_Z_MAKE(0.0, 0.0);
magmaDoubleComplex alpha;
#define ENABLE_TIMER
#ifdef ENABLE_TIMER
double mdot1, mdot2, mdgm1, mdgm2, magmagemv1, magmagemv2, cugemv1, cugemv2, cudot1, cudot2;
double mdot_time, mdgm_time, magmagemv_time, cugemv_time, cudot_time;
#endif
magma_z_vinit( &a, Magma_DEV, n*num_vecs, one );
magma_z_vinit( &b, Magma_DEV, num_vecs, one );
int min_ten = min(num_vecs, 15);
magma_z_vinit( &x, Magma_DEV, min_ten*n, one );
magma_z_vinit( &y, Magma_DEV, min_ten*n, one );
magma_z_vinit( &skp, Magma_DEV, num_vecs, zero );
// warm up
magma_zgemvmdot( n, num_vecs, a.val, b.val, x.val, y.val, skp.val );
// CUDOT
#ifdef ENABLE_TIMER
magma_device_sync(); cudot1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
for( int l=0; l<num_vecs; l++)
alpha = magma_zdotc(n, a.val, 1, b.val, 1);
}
#ifdef ENABLE_TIMER
magma_device_sync(); cudot2=magma_wtime();
cudot_time=cudot2-cudot1;
#endif
// CUGeMV
#ifdef ENABLE_TIMER
magma_device_sync(); cugemv1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
magma_zgemv(MagmaTrans, n, num_vecs, one, a.val, n, b.val, 1, zero, skp.val, 1);
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); cugemv2=magma_wtime();
cugemv_time=cugemv2-cugemv1;
#endif
// MAGMAGeMV
#ifdef ENABLE_TIMER
magma_device_sync(); magmagemv1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
magmablas_zgemv(MagmaTrans, n, num_vecs, one, a.val, n, b.val, 1, zero, skp.val, 1);
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); magmagemv2=magma_wtime();
magmagemv_time=magmagemv2-magmagemv1;
#endif
// MDOT
#ifdef ENABLE_TIMER
magma_device_sync(); mdot1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
//magma_zmdotc( n, num_vecs, a.val, b.val, x.val, y.val, skp.val );
magma_zmdotc( n, 2, a.val, b.val, x.val, y.val, skp.val );
magma_zmdotc( n, 2, a.val, b.val, x.val, y.val, skp.val );
magma_zmdotc( n, 1, a.val, b.val, x.val, y.val, skp.val );
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); mdot2=magma_wtime();
mdot_time=mdot2-mdot1;
#endif
// MDGM
#ifdef ENABLE_TIMER
magma_device_sync(); mdgm1=magma_wtime();
#endif
for( int h=0; h<iters; h++){
magma_zgemvmdot( n, num_vecs, a.val, b.val, x.val, y.val, skp.val );
//h++;
}
#ifdef ENABLE_TIMER
magma_device_sync(); mdgm2=magma_wtime();
mdgm_time=mdgm2-mdgm1;
#endif
//magma_zprint_gpu(num_vecs,1,skp.val,num_vecs);
//Chronometry
#ifdef ENABLE_TIMER
printf("%d %d %e %e %e %e %e %e %e %e %e %e\n",
n, num_vecs,
cudot_time/iters,
(cugemv_time)/iters,
(magmagemv_time)/iters,
(mdot_time)/iters,
(mdgm_time)/iters,
(double)(computations)/(cudot_time*(1.e+09)),
(double)(computations)/(cugemv_time*(1.e+09)),
(double)(computations)/(magmagemv_time*(1.e+09)),
(double)(computations)/(mdot_time*(1.e+09)),
(double)(computations)/(mdgm_time*(1.e+09)) );
#endif
magma_z_vfree(&a);
magma_z_vfree(&b);
magma_z_vfree(&x);
magma_z_vfree(&y);
magma_z_vfree(&skp);
}
// }
printf("#================================================================================================================================================\n");
printf("\n");
printf("\n");
}
TESTING_FINALIZE();
return 0;
}
