inline void gemm( const Order order, const TransA transa, const TransB transb,
const int m, const int n, const int k,
const std::complex<double> alpha, const std::complex<double>* a,
const int lda, const std::complex<double>* b, const int ldb,
const std::complex<double> beta, std::complex<double>* c,
const int ldc ) {
BOOST_STATIC_ASSERT( (is_same<Order, tag::column_major>::value) );
cublasZgemm( blas_option< TransA >::value, blas_option< TransB >::value,
m, n, k, alpha, a, lda, b, ldb, beta, c, ldc );
}
/** Perform matrix-matrix product, \f$ C = \alpha op(A) op(B) + \beta C \f$.
@param[in]
transA Operation op(A) to perform on matrix A.
@param[in]
transB Operation op(B) to perform on matrix B.
@param[in]
m Number of rows of C and op(A). m >= 0.
@param[in]
n Number of columns of C and op(B). n >= 0.
@param[in]
k Number of columns of op(A) and rows of op(B). k >= 0.
@param[in]
alpha Scalar \f$ \alpha \f$
@param[in]
dA COMPLEX_16 array on GPU device.
If transA == MagmaNoTrans, the m-by-k matrix A of dimension (ldda,k), ldda >= max(1,m); \n
otherwise, the k-by-m matrix A of dimension (ldda,m), ldda >= max(1,k).
@param[in]
ldda Leading dimension of dA.
@param[in]
dB COMPLEX_16 array on GPU device.
If transB == MagmaNoTrans, the k-by-n matrix B of dimension (lddb,n), lddb >= max(1,k); \n
otherwise, the n-by-k matrix B of dimension (lddb,k), lddb >= max(1,n).
@param[in]
lddb Leading dimension of dB.
@param[in]
beta Scalar \f$ \beta \f$
@param[in,out]
dC COMPLEX_16 array on GPU device.
The m-by-n matrix C of dimension (lddc,n), lddc >= max(1,m).
@param[in]
lddc Leading dimension of dC.
@ingroup magma_zblas3
*/
extern "C" void
magma_zgemm(
magma_trans_t transA, magma_trans_t transB,
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex alpha,
magmaDoubleComplex_const_ptr dA, magma_int_t ldda,
magmaDoubleComplex_const_ptr dB, magma_int_t lddb,
magmaDoubleComplex beta,
magmaDoubleComplex_ptr dC, magma_int_t lddc )
{
cublasZgemm(
cublas_trans_const( transA ),
cublas_trans_const( transB ),
m, n, k,
alpha, dA, ldda,
dB, lddb,
beta, dC, lddc );
}
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, t1, t2;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_trans_t trans[] = { MagmaNoTrans, MagmaConjTrans, MagmaTrans };
magma_uplo_t uplo [] = { MagmaLower, MagmaUpper };
magma_diag_t diag [] = { MagmaUnit, MagmaNonUnit };
magma_side_t side [] = { MagmaLeft, MagmaRight };
magmaDoubleComplex *A, *B, *C, *C2, *LU;
magmaDoubleComplex *dA, *dB, *dC1, *dC2;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.5, 0.1 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( 0.7, 0.2 );
double dalpha = 0.6;
double dbeta = 0.8;
double work[1], error, total_error;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t m, n, k, size, maxn, ld, info;
magma_int_t *piv;
magma_int_t err;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf( "Compares magma wrapper function to cublas function; all diffs should be exactly 0.\n\n" );
total_error = 0.;
for( int itest = 0; itest < opts.ntest; ++itest ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
printf("=========================================================================\n");
printf( "m=%d, n=%d, k=%d\n", (int) m, (int) n, (int) k );
// allocate matrices
// over-allocate so they can be any combination of {m,n,k} x {m,n,k}.
maxn = max( max( m, n ), k );
ld = max( 1, maxn );
size = ld*maxn;
err = magma_malloc_cpu( (void**) &piv, maxn*sizeof(magma_int_t) ); assert( err == 0 );
err = magma_zmalloc_pinned( &A, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &B, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &C, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &C2, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &LU, size ); assert( err == 0 );
err = magma_zmalloc( &dA, size ); assert( err == 0 );
err = magma_zmalloc( &dB, size ); assert( err == 0 );
err = magma_zmalloc( &dC1, size ); assert( err == 0 );
err = magma_zmalloc( &dC2, size ); assert( err == 0 );
// initialize matrices
size = maxn*maxn;
lapackf77_zlarnv( &ione, ISEED, &size, A );
lapackf77_zlarnv( &ione, ISEED, &size, B );
lapackf77_zlarnv( &ione, ISEED, &size, C );
printf( "========== Level 1 BLAS ==========\n" );
// ----- test ZSWAP
// swap columns 2 and 3 of dA, then copy to C2 and compare with A
if ( n >= 3 ) {
magma_zsetmatrix( m, n, A, ld, dA, ld );
magma_zsetmatrix( m, n, A, ld, dB, ld );
magma_zswap( m, dA(0,1), 1, dA(0,2), 1 );
magma_zswap( m, dB(0,1), 1, dB(0,2), 1 );
// check results, storing diff between magma and cuda calls in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dA, 1, dB, 1 );
magma_zgetmatrix( m, n, dB, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &k, C2, &ld, work );
total_error += error;
printf( "zswap diff %.2g\n", error );
}
else {
printf( "zswap skipped for n < 3\n" );
}
// ----- test IZAMAX
// get argmax of column of A
magma_zsetmatrix( m, k, A, ld, dA, ld );
error = 0;
for( int j = 0; j < k; ++j ) {
magma_int_t i1 = magma_izamax( m, dA(0,j), 1 );
int i2; // NOT magma_int_t, for cublas
cublasIzamax( handle, m, dA(0,j), 1, &i2 );
// todo need sync here?
assert( i1 == i2 );
error += abs( i1 - i2 );
}
total_error += error;
gflops = (double)m * k / 1e9;
printf( "izamax diff %.2g\n", error );
printf( "\n" );
printf( "========== Level 2 BLAS ==========\n" );
// ----- test ZGEMV
// c = alpha*A*b + beta*c, with A m*n; b,c m or n-vectors
// try no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
magma_zsetmatrix( m, n, A, ld, dA, ld );
magma_zsetvector( maxn, B, 1, dB, 1 );
magma_zsetvector( maxn, C, 1, dC1, 1 );
magma_zsetvector( maxn, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_zgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZgemv( handle, cublas_trans_const(trans[ia]),
m, n, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
size = (trans[ia] == MagmaNoTrans ? m : n);
cublasZaxpy( handle, size, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( size, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &size, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZGEMV( m, n ) / 1e9;
printf( "zgemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test ZHEMV
// c = alpha*A*b + beta*c, with A m*m symmetric; b,c m-vectors
// try upper/lower
for( int iu = 0; iu < 2; ++iu ) {
magma_zsetmatrix( m, m, A, ld, dA, ld );
magma_zsetvector( m, B, 1, dB, 1 );
magma_zsetvector( m, C, 1, dC1, 1 );
magma_zsetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_zhemv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZhemv( handle, cublas_uplo_const(uplo[iu]),
m, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHEMV( m ) / 1e9;
printf( "zhemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test ZTRSV
// solve A*c = c, with A m*m triangular; c m-vector
// try upper/lower, no-trans/trans, unit/non-unit diag
// Factor A into LU to get well-conditioned triangles, else solve yields garbage.
// Still can give garbage if solves aren't consistent with LU factors,
// e.g., using unit diag for U, so copy lower triangle to upper triangle.
// Also used for trsm later.
lapackf77_zlacpy( "Full", &maxn, &maxn, A, &ld, LU, &ld );
lapackf77_zgetrf( &maxn, &maxn, LU, &ld, piv, &info );
for( int j = 0; j < maxn; ++j ) {
for( int i = 0; i < j; ++i ) {
*LU(i,j) = *LU(j,i);
}
}
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
magma_zsetmatrix( m, m, LU, ld, dA, ld );
magma_zsetvector( m, C, 1, dC1, 1 );
magma_zsetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_ztrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZtrsv( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
cublas_diag_const(diag[id]), m, dA, ld, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRSM( MagmaLeft, m, 1 ) / 1e9;
printf( "ztrsv( %c, %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), lapacke_diag_const(diag[id]),
error, gflops/t1, gflops/t2 );
}}}
printf( "\n" );
printf( "========== Level 3 BLAS ==========\n" );
// ----- test ZGEMM
// C = alpha*A*B + beta*C, with A m*k or k*m; B k*n or n*k; C m*n
// try combinations of no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
for( int ib = 0; ib < 3; ++ib ) {
bool nta = (trans[ia] == MagmaNoTrans);
bool ntb = (trans[ib] == MagmaNoTrans);
magma_zsetmatrix( (nta ? m : k), (nta ? m : k), A, ld, dA, ld );
magma_zsetmatrix( (ntb ? k : n), (ntb ? n : k), B, ld, dB, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZgemm( handle, cublas_trans_const(trans[ia]), cublas_trans_const(trans[ib]),
m, n, k, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZGEMM( m, n, k ) / 1e9;
printf( "zgemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), lapacke_trans_const(trans[ib]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHEMM
// C = alpha*A*B + beta*C (left) with A m*m symmetric; B,C m*n; or
// C = alpha*B*A + beta*C (right) with A n*n symmetric; B,C m*n
// try left/right, upper/lower
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
magma_zsetmatrix( m, m, A, ld, dA, ld );
magma_zsetmatrix( m, n, B, ld, dB, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zhemm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZhemm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
m, n, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHEMM( side[is], m, n ) / 1e9;
printf( "zhemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_side_const(side[is]), lapacke_uplo_const(uplo[iu]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHERK
// C = alpha*A*A^H + beta*C (no-trans) with A m*k and C m*m symmetric; or
// C = alpha*A^H*A + beta*C (trans) with A k*m and C m*m symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
magma_zsetmatrix( n, k, A, ld, dA, ld );
magma_zsetmatrix( n, n, C, ld, dC1, ld );
magma_zsetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zherk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZherk( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &dalpha, dA, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHERK( k, n ) / 1e9;
printf( "zherk( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHER2K
// C = alpha*A*B^H + ^alpha*B*A^H + beta*C (no-trans) with A,B n*k; C n*n symmetric; or
// C = alpha*A^H*B + ^alpha*B^H*A + beta*C (trans) with A,B k*n; C n*n symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
bool nt = (trans[it] == MagmaNoTrans);
magma_zsetmatrix( (nt ? n : k), (nt ? n : k), A, ld, dA, ld );
magma_zsetmatrix( n, n, C, ld, dC1, ld );
magma_zsetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zher2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZher2k( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &alpha, dA, ld, dB, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHER2K( k, n ) / 1e9;
printf( "zher2k( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZTRMM
// C = alpha*A*C (left) with A m*m triangular; C m*n; or
// C = alpha*C*A (right) with A n*n triangular; C m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_zsetmatrix( (left ? m : n), (left ? m : n), A, ld, dA, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ztrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
// note cublas does trmm out-of-place (i.e., adds output matrix C),
// but allows C=B to do in-place.
t2 = magma_sync_wtime( 0 );
cublasZtrmm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRMM( side[is], m, n ) / 1e9;
printf( "ztrmm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// ----- test ZTRSM
// solve A*X = alpha*B (left) with A m*m triangular; B m*n; or
// solve X*A = alpha*B (right) with A n*n triangular; B m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_zsetmatrix( (left ? m : n), (left ? m : n), LU, ld, dA, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ztrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZtrsm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRSM( side[is], m, n ) / 1e9;
printf( "ztrsm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// cleanup
magma_free_cpu( piv );
magma_free_pinned( A );
magma_free_pinned( B );
magma_free_pinned( C );
magma_free_pinned( C2 );
magma_free_pinned( LU );
magma_free( dA );
magma_free( dB );
magma_free( dC1 );
magma_free( dC2 );
fflush( stdout );
}
if ( total_error != 0. ) {
printf( "total error %.2g -- ought to be 0 -- some test failed (see above).\n",
total_error );
}
else {
printf( "all tests passed\n" );
}
TESTING_FINALIZE();
int status = (total_error != 0.);
return status;
}
extern "C" magma_int_t
magma_zlarfb_gpu( char side, char trans, char direct, char storev,
magma_int_t m, magma_int_t n, magma_int_t k,
cuDoubleComplex *dV, magma_int_t ldv,
cuDoubleComplex *dT, magma_int_t ldt,
cuDoubleComplex *dC, magma_int_t ldc,
cuDoubleComplex *dwork, magma_int_t ldwork)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Univ. of California Berkeley
August 2013
Purpose
=======
ZLARFB applies a complex block reflector H or its transpose H' to a
COMPLEX_16 m by n matrix C, from the left.
Arguments
=========
SIDE (input) CHARACTER
= 'L': apply H or H' from the Left
= 'R': apply H or H' from the Right (Not implemented)
TRANS (input) CHARACTER
= 'N': apply H (No transpose) (Not implemented)
= 'C': apply H' (Conjugate transpose)
DIRECT (input) CHARACTER
Indicates how H is formed from a product of elementary
reflectors
= 'F': H = H(1) H(2) . . . H(k) (Forward)
= 'B': H = H(k) . . . H(2) H(1) (Backward)
STOREV (input) CHARACTER
Indicates how the vectors which define the elementary
reflectors are stored:
= 'C': Columnwise
= 'R': Rowwise
M (input) INTEGER
The number of rows of the matrix C.
N (input) INTEGER
The number of columns of the matrix C.
K (input) INTEGER
The order of the matrix T (= the number of elementary
reflectors whose product defines the block reflector).
DV (input) COMPLEX_16 array, dimension (LDV,K)
The matrix V. See further details.
LDV (input) INTEGER
The leading dimension of the array V. LDV >= max(1,M);
DT (input) COMPLEX_16 array, dimension (LDT,K)
The triangular k by k matrix T in the representation of the
block reflector.
LDT (input) INTEGER
The leading dimension of the array T. LDT >= K.
DC (input/output) COMPLEX_16 array, dimension (LDC,N)
On entry, the m by n matrix C.
On exit, C is overwritten by H*C.
LDC (input) INTEGER
The leading dimension of the array C. LDA >= max(1,M).
WORK (workspace) COMPLEX_16 array, dimension (LDWORK,K)
LDWORK (input) INTEGER
The leading dimension of the array WORK. LDWORK >= max(1,N);
=================================================================== */
cuDoubleComplex c_zero = MAGMA_Z_ZERO;
cuDoubleComplex c_one = MAGMA_Z_ONE;
cuDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
/* Function Body */
if (m <= 0 || n <= 0) {
return MAGMA_SUCCESS;
}
char transt;
if (trans == 'N' || trans == 'n')
transt = MagmaConjTrans;
else
transt = MagmaNoTrans;
if ( ( side == 'r' || side == 'R') ) {
fprintf(stderr, "The case (side == right) is not implemented\n");
magma_xerbla( __func__, 1 );
return MAGMA_ERR_ILLEGAL_VALUE;
}
if ( storev == 'c' || storev == 'C') {
/*
if (n==1 && m%32==0){
// This is used when we have to apply H on only one vector
magmablas_zgemvt(m, k, 1., dv_ref(0,0), ldv, dc_ref(0, 0), dwork);
printf("m= %d, n = %d, ldwork = %d\n", m, k, ldwork);
}
else
*/
cublasZgemm( MagmaConjTrans, MagmaNoTrans,
n, k, m,
c_one, dC, ldc,
dV, ldv,
c_zero, dwork, ldwork);
if (direct == 'F' || direct =='f')
cublasZtrmm( MagmaRight, MagmaUpper, transt, MagmaNonUnit,
n, k,
c_one, dT, ldt,
dwork, ldwork);
else
cublasZtrmm( MagmaRight, MagmaLower, transt, MagmaNonUnit,
n, k,
c_one, dT, ldt,
dwork, ldwork);
cublasZgemm( MagmaNoTrans, MagmaConjTrans,
m, n, k,
c_neg_one, dV, ldv,
dwork, ldwork,
c_one, dC, ldc);
}
else {
cublasZgemm( MagmaNoTrans, MagmaConjTrans,
m, k, n,
c_one, dC, ldc,
dV, ldv,
c_zero, dwork, ldwork);
cublasZtrmm( MagmaRight, MagmaUpper, transt, MagmaNonUnit,
m, k,
c_one, dT, ldt,
dwork, ldwork);
cublasZgemm( MagmaNoTrans, MagmaNoTrans,
m, n, k,
c_neg_one, dwork, ldwork,
dV, ldv,
c_one, dC, ldc);
}
return MAGMA_SUCCESS;
} /* magma_zlarfb */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgemm
*/
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, Cnorm, work[1];
magma_int_t M, N, K;
magma_int_t Am, An, Bm, 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_Cmagma, *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 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
printf("If running lapack (option --lapack), MAGMA and CUBLAS error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.\n\n"
"transA = %c, transB = %c\n", opts.transA, opts.transB );
printf(" M N K MAGMA Gflop/s (ms) CUBLAS Gflop/s (ms) CPU Gflop/s (ms) MAGMA error CUBLAS error\n");
printf("=========================================================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
K = opts.ksize[i];
gflops = FLOPS_ZGEMM( M, N, K ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if ( opts.transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*An;
sizeB = ldb*Bn;
sizeC = ldc*N;
TESTING_MALLOC( h_A, magmaDoubleComplex, lda*An );
TESTING_MALLOC( h_B, magmaDoubleComplex, ldb*Bn );
TESTING_MALLOC( h_C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC( h_Cmagma, magmaDoubleComplex, ldc*N );
TESTING_MALLOC( h_Ccublas, magmaDoubleComplex, ldc*N );
TESTING_DEVALLOC( d_A, magmaDoubleComplex, ldda*An );
TESTING_DEVALLOC( d_B, magmaDoubleComplex, lddb*Bn );
TESTING_DEVALLOC( 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 MAGMA-BLAS
=================================================================== */
magma_zsetmatrix( Am, An, h_A, lda, d_A, ldda );
magma_zsetmatrix( Bm, Bn, h_B, ldb, d_B, lddb );
magma_zsetmatrix( M, N, h_C, ldc, d_C, lddc );
magma_time = magma_sync_wtime( NULL );
magmablas_zgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, ldda,
d_B, lddb,
beta, d_C, lddc );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetmatrix( M, N, d_C, lddc, h_Cmagma, ldc );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_zsetmatrix( M, N, h_C, ldc, d_C, lddc );
cublas_time = magma_sync_wtime( NULL );
cublasZgemm( opts.transA, opts.transB, M, 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( M, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_zgemm( &opts.transA, &opts.transB, &M, &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", &M, &N, h_C, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cmagma, &ione );
magma_error = lapackf77_zlange( "M", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_zlange( "M", &M, &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error );
}
else {
// compute relative error for magma, relative to cublas
Cnorm = lapackf77_zlange( "M", &M, &N, h_Ccublas, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_Ccublas, &ione, h_Cmagma, &ione );
magma_error = lapackf77_zlange( "M", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e ---\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
magma_error );
}
TESTING_FREE( h_A );
TESTING_FREE( h_B );
TESTING_FREE( h_C );
TESTING_FREE( h_Cmagma );
TESTING_FREE( h_Ccublas );
TESTING_DEVFREE( d_A );
TESTING_DEVFREE( d_B );
TESTING_DEVFREE( d_C );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
int TEMPLATE2 (CHOLMOD (gpu_lower_potrf))
(
Int nscol2, /* S is nscol2-by-nscol2 */
Int nsrow, /* leading dimension of S */
Int psx, /* S is located at Lx + L_ENTRY*psx */
double *Lx, /* contains S; overwritten with Cholesky factor */
Int *info, /* BLAS info return value */
cholmod_common *Common,
cholmod_gpu_pointers *gpu_p
)
{
double *devPtrA, *devPtrB, *A ;
double alpha, beta ;
cudaError_t cudaStat ;
cublasStatus_t cublasStatus ;
Int j, nsrow2, nb, n, gpu_lda, lda, gpu_ldb ;
int ilda, ijb, iinfo ;
#ifndef NTIMER
double tstart ;
#endif
if (nscol2 * L_ENTRY < CHOLMOD_POTRF_LIMIT)
{
/* too small for the CUDA BLAS; use the CPU instead */
return (0) ;
}
#ifndef NTIMER
tstart = SuiteSparse_time ( ) ;
Common->CHOLMOD_GPU_POTRF_CALLS++ ;
#endif
nsrow2 = nsrow - nscol2 ;
/* ---------------------------------------------------------------------- */
/* heuristic to get the block size depending of the problem size */
/* ---------------------------------------------------------------------- */
nb = 128 ;
if (nscol2 > 4096) nb = 256 ;
if (nscol2 > 8192) nb = 384 ;
n = nscol2 ;
gpu_lda = ((nscol2+31)/32)*32 ;
lda = nsrow ;
A = gpu_p->h_Lx[(Common->ibuffer+CHOLMOD_HOST_SUPERNODE_BUFFERS-1)%
CHOLMOD_HOST_SUPERNODE_BUFFERS];
/* ---------------------------------------------------------------------- */
/* determine the GPU leading dimension of B */
/* ---------------------------------------------------------------------- */
gpu_ldb = 0 ;
if (nsrow2 > 0)
{
gpu_ldb = ((nsrow2+31)/32)*32 ;
}
/* ---------------------------------------------------------------------- */
/* remember where device memory is, to be used by triangular solve later */
/* ---------------------------------------------------------------------- */
devPtrA = gpu_p->d_Lx[0];
devPtrB = gpu_p->d_Lx[1];
/* ---------------------------------------------------------------------- */
/* copy A from device to device */
/* ---------------------------------------------------------------------- */
cudaStat = cudaMemcpy2DAsync ( devPtrA,
gpu_lda * L_ENTRY * sizeof (devPtrA[0]),
gpu_p->d_A[1],
nsrow * L_ENTRY * sizeof (Lx[0]),
nscol2 * L_ENTRY * sizeof (devPtrA[0]),
nscol2,
cudaMemcpyDeviceToDevice,
Common->gpuStream[0] );
if ( cudaStat ) {
ERROR ( CHOLMOD_GPU_PROBLEM, "GPU memcopy device to device");
}
/* ---------------------------------------------------------------------- */
/* copy B in advance, for gpu_triangular_solve */
/* ---------------------------------------------------------------------- */
if (nsrow2 > 0)
{
cudaStat = cudaMemcpy2DAsync (devPtrB,
gpu_ldb * L_ENTRY * sizeof (devPtrB [0]),
gpu_p->d_A[1] + L_ENTRY*nscol2,
nsrow * L_ENTRY * sizeof (Lx [0]),
nsrow2 * L_ENTRY * sizeof (devPtrB [0]),
nscol2,
cudaMemcpyDeviceToDevice,
Common->gpuStream[0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
}
/* ------------------------------------------------------------------ */
/* define the dpotrf stream */
/* ------------------------------------------------------------------ */
cublasStatus = cublasSetStream (Common->cublasHandle,
Common->gpuStream [0]) ;
if (cublasStatus != CUBLAS_STATUS_SUCCESS) {
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS stream") ;
}
/* ---------------------------------------------------------------------- */
/* block Cholesky factorization of S */
/* ---------------------------------------------------------------------- */
for (j = 0 ; j < n ; j += nb)
{
Int jb = nb < (n-j) ? nb : (n-j) ;
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dsyrk */
/* ------------------------------------------------------------------ */
alpha = -1.0 ;
beta = 1.0 ;
#ifdef REAL
cublasStatus = cublasDsyrk (Common->cublasHandle,
CUBLAS_FILL_MODE_LOWER, CUBLAS_OP_N, jb, j,
&alpha, devPtrA + j, gpu_lda,
&beta, devPtrA + j + j*gpu_lda, gpu_lda) ;
#else
cublasStatus = cublasZherk (Common->cublasHandle,
CUBLAS_FILL_MODE_LOWER, CUBLAS_OP_N, jb, j,
&alpha, (cuDoubleComplex*)devPtrA + j,
gpu_lda,
&beta,
(cuDoubleComplex*)devPtrA + j + j*gpu_lda,
gpu_lda) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
/* ------------------------------------------------------------------ */
cudaStat = cudaEventRecord (Common->cublasEventPotrf [0],
Common->gpuStream [0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
cudaStat = cudaStreamWaitEvent (Common->gpuStream [1],
Common->cublasEventPotrf [0], 0) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
/* ------------------------------------------------------------------ */
/* copy back the jb columns on two different streams */
/* ------------------------------------------------------------------ */
cudaStat = cudaMemcpy2DAsync (A + L_ENTRY*(j + j*lda),
lda * L_ENTRY * sizeof (double),
devPtrA + L_ENTRY*(j + j*gpu_lda),
gpu_lda * L_ENTRY * sizeof (double),
L_ENTRY * sizeof (double)*jb,
jb,
cudaMemcpyDeviceToHost,
Common->gpuStream [1]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy from device") ;
}
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dgemm */
/* ------------------------------------------------------------------ */
if ((j+jb) < n)
{
#ifdef REAL
alpha = -1.0 ;
beta = 1.0 ;
cublasStatus = cublasDgemm (Common->cublasHandle,
CUBLAS_OP_N, CUBLAS_OP_T,
(n-j-jb), jb, j,
&alpha,
devPtrA + (j+jb), gpu_lda,
devPtrA + (j) , gpu_lda,
&beta,
devPtrA + (j+jb + j*gpu_lda), gpu_lda) ;
#else
cuDoubleComplex calpha = {-1.0,0.0} ;
cuDoubleComplex cbeta = { 1.0,0.0} ;
cublasStatus = cublasZgemm (Common->cublasHandle,
CUBLAS_OP_N, CUBLAS_OP_C,
(n-j-jb), jb, j,
&calpha,
(cuDoubleComplex*)devPtrA + (j+jb),
gpu_lda,
(cuDoubleComplex*)devPtrA + (j),
gpu_lda,
&cbeta,
(cuDoubleComplex*)devPtrA +
(j+jb + j*gpu_lda),
gpu_lda ) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
}
cudaStat = cudaStreamSynchronize (Common->gpuStream [1]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
/* ------------------------------------------------------------------ */
/* compute the Cholesky factorization of the jbxjb block on the CPU */
/* ------------------------------------------------------------------ */
ilda = (int) lda ;
ijb = jb ;
#ifdef REAL
LAPACK_DPOTRF ("L", &ijb, A + L_ENTRY * (j + j*lda), &ilda, &iinfo) ;
#else
LAPACK_ZPOTRF ("L", &ijb, A + L_ENTRY * (j + j*lda), &ilda, &iinfo) ;
#endif
*info = iinfo ;
if (*info != 0)
{
*info = *info + j ;
break ;
}
/* ------------------------------------------------------------------ */
/* copy the result back to the GPU */
/* ------------------------------------------------------------------ */
cudaStat = cudaMemcpy2DAsync (devPtrA + L_ENTRY*(j + j*gpu_lda),
gpu_lda * L_ENTRY * sizeof (double),
A + L_ENTRY * (j + j*lda),
lda * L_ENTRY * sizeof (double),
L_ENTRY * sizeof (double) * jb,
jb,
cudaMemcpyHostToDevice,
Common->gpuStream [0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dtrsm */
/* ------------------------------------------------------------------ */
if ((j+jb) < n)
{
#ifdef REAL
alpha = 1.0 ;
cublasStatus = cublasDtrsm (Common->cublasHandle,
CUBLAS_SIDE_RIGHT,
CUBLAS_FILL_MODE_LOWER,
CUBLAS_OP_T, CUBLAS_DIAG_NON_UNIT,
(n-j-jb), jb,
&alpha,
devPtrA + (j + j*gpu_lda), gpu_lda,
devPtrA + (j+jb + j*gpu_lda), gpu_lda) ;
#else
cuDoubleComplex calpha = {1.0,0.0};
cublasStatus = cublasZtrsm (Common->cublasHandle,
CUBLAS_SIDE_RIGHT,
CUBLAS_FILL_MODE_LOWER,
CUBLAS_OP_C, CUBLAS_DIAG_NON_UNIT,
(n-j-jb), jb,
&calpha,
(cuDoubleComplex *)devPtrA +
(j + j*gpu_lda),
gpu_lda,
(cuDoubleComplex *)devPtrA +
(j+jb + j*gpu_lda),
gpu_lda) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
/* -------------------------------------------------------------- */
/* Copy factored column back to host. */
/* -------------------------------------------------------------- */
cudaStat = cudaEventRecord (Common->cublasEventPotrf[2],
Common->gpuStream[0]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
cudaStat = cudaStreamWaitEvent (Common->gpuStream[1],
Common->cublasEventPotrf[2], 0) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "CUDA event failure") ;
}
cudaStat = cudaMemcpy2DAsync (A + L_ENTRY*(j + jb + j * lda),
lda * L_ENTRY * sizeof (double),
devPtrA + L_ENTRY*
(j + jb + j * gpu_lda),
gpu_lda * L_ENTRY * sizeof (double),
L_ENTRY * sizeof (double)*
(n - j - jb), jb,
cudaMemcpyDeviceToHost,
Common->gpuStream[1]) ;
if (cudaStat)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU memcopy to device") ;
}
}
}
#ifndef NTIMER
Common->CHOLMOD_GPU_POTRF_TIME += SuiteSparse_time ( ) - tstart ;
#endif
return (1) ;
}
int TEMPLATE2 (CHOLMOD (gpu_updateC))
(
Int ndrow1, /* C is ndrow2-by-ndrow2 */
Int ndrow2,
Int ndrow, /* leading dimension of Lx */
Int ndcol, /* L1 is ndrow1-by-ndcol */
Int nsrow,
Int pdx1, /* L1 starts at Lx + L_ENTRY*pdx1 */
/* L2 starts at Lx + L_ENTRY*(pdx1 + ndrow1) */
Int pdi1,
double *Lx,
double *C,
cholmod_common *Common,
cholmod_gpu_pointers *gpu_p
)
{
double *devPtrLx, *devPtrC ;
double alpha, beta ;
cublasStatus_t cublasStatus ;
cudaError_t cudaStat [2] ;
Int ndrow3 ;
int icol, irow;
int iHostBuff, iDevBuff ;
#ifndef NTIMER
double tstart = 0;
#endif
if ((ndrow2*L_ENTRY < CHOLMOD_ND_ROW_LIMIT) ||
(ndcol*L_ENTRY < CHOLMOD_ND_COL_LIMIT))
{
/* too small for the CUDA BLAS; use the CPU instead */
return (0) ;
}
ndrow3 = ndrow2 - ndrow1 ;
#ifndef NTIMER
Common->syrkStart = SuiteSparse_time ( ) ;
Common->CHOLMOD_GPU_SYRK_CALLS++ ;
#endif
/* ---------------------------------------------------------------------- */
/* allocate workspace on the GPU */
/* ---------------------------------------------------------------------- */
iHostBuff = (Common->ibuffer)%CHOLMOD_HOST_SUPERNODE_BUFFERS;
iDevBuff = (Common->ibuffer)%CHOLMOD_DEVICE_STREAMS;
/* cycle the device Lx buffer, d_Lx, through CHOLMOD_DEVICE_STREAMS,
usually 2, so we can overlap the copy of this descendent supernode
with the compute of the previous descendant supernode */
devPtrLx = (double *)(gpu_p->d_Lx[iDevBuff]);
/* very little overlap between kernels for difference descendant supernodes
(since we enforce the supernodes must be large enough to fill the
device) so we only need one C buffer */
devPtrC = (double *)(gpu_p->d_C);
/* ---------------------------------------------------------------------- */
/* copy Lx to the GPU */
/* ---------------------------------------------------------------------- */
/* copy host data to pinned buffer first for better H2D bandwidth */
#pragma omp parallel for num_threads(CHOLMOD_OMP_NUM_THREADS) if (ndcol > 32)
for ( icol=0; icol<ndcol; icol++ ) {
for ( irow=0; irow<ndrow2*L_ENTRY; irow++ ) {
gpu_p->h_Lx[iHostBuff][icol*ndrow2*L_ENTRY+irow] =
Lx[pdx1*L_ENTRY+icol*ndrow*L_ENTRY + irow];
}
}
cudaStat[0] = cudaMemcpyAsync ( devPtrLx,
gpu_p->h_Lx[iHostBuff],
ndrow2*ndcol*L_ENTRY*sizeof(devPtrLx[0]),
cudaMemcpyHostToDevice,
Common->gpuStream[iDevBuff] );
if ( cudaStat[0] ) {
CHOLMOD_GPU_PRINTF ((" ERROR cudaMemcpyAsync = %d \n", cudaStat[0]));
return (0);
}
/* make the current stream wait for kernels in previous streams */
cudaStreamWaitEvent ( Common->gpuStream[iDevBuff],
Common->updateCKernelsComplete, 0 ) ;
/* ---------------------------------------------------------------------- */
/* create the relative map for this descendant supernode */
/* ---------------------------------------------------------------------- */
createRelativeMapOnDevice ( (Int *)(gpu_p->d_Map),
(Int *)(gpu_p->d_Ls),
(Int *)(gpu_p->d_RelativeMap),
pdi1, ndrow2,
&(Common->gpuStream[iDevBuff]) );
/* ---------------------------------------------------------------------- */
/* do the CUDA SYRK */
/* ---------------------------------------------------------------------- */
cublasStatus = cublasSetStream (Common->cublasHandle,
Common->gpuStream[iDevBuff]) ;
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS stream") ;
}
alpha = 1.0 ;
beta = 0.0 ;
#ifdef REAL
cublasStatus = cublasDsyrk (Common->cublasHandle,
CUBLAS_FILL_MODE_LOWER,
CUBLAS_OP_N,
(int) ndrow1,
(int) ndcol, /* N, K: L1 is ndrow1-by-ndcol */
&alpha, /* ALPHA: 1 */
devPtrLx,
ndrow2, /* A, LDA: L1, ndrow2 */
&beta, /* BETA: 0 */
devPtrC,
ndrow2) ; /* C, LDC: C1 */
#else
cublasStatus = cublasZherk (Common->cublasHandle,
CUBLAS_FILL_MODE_LOWER,
CUBLAS_OP_N,
(int) ndrow1,
(int) ndcol, /* N, K: L1 is ndrow1-by-ndcol*/
&alpha, /* ALPHA: 1 */
(const cuDoubleComplex *) devPtrLx,
ndrow2, /* A, LDA: L1, ndrow2 */
&beta, /* BETA: 0 */
(cuDoubleComplex *) devPtrC,
ndrow2) ; /* C, LDC: C1 */
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
#ifndef NTIMER
Common->CHOLMOD_GPU_SYRK_TIME += SuiteSparse_time() - Common->syrkStart;
#endif
/* ---------------------------------------------------------------------- */
/* compute remaining (ndrow2-ndrow1)-by-ndrow1 block of C, C2 = L2*L1' */
/* ---------------------------------------------------------------------- */
#ifndef NTIMER
Common->CHOLMOD_GPU_GEMM_CALLS++ ;
tstart = SuiteSparse_time();
#endif
if (ndrow3 > 0)
{
#ifndef REAL
cuDoubleComplex calpha = {1.0,0.0} ;
cuDoubleComplex cbeta = {0.0,0.0} ;
#endif
/* ------------------------------------------------------------------ */
/* do the CUDA BLAS dgemm */
/* ------------------------------------------------------------------ */
#ifdef REAL
alpha = 1.0 ;
beta = 0.0 ;
cublasStatus = cublasDgemm (Common->cublasHandle,
CUBLAS_OP_N, CUBLAS_OP_T,
ndrow3, ndrow1, ndcol, /* M, N, K */
&alpha, /* ALPHA: 1 */
devPtrLx + L_ENTRY*(ndrow1), /* A, LDA: L2*/
ndrow2, /* ndrow */
devPtrLx, /* B, LDB: L1 */
ndrow2, /* ndrow */
&beta, /* BETA: 0 */
devPtrC + L_ENTRY*ndrow1, /* C, LDC: C2 */
ndrow2) ;
#else
cublasStatus = cublasZgemm (Common->cublasHandle,
CUBLAS_OP_N, CUBLAS_OP_C,
ndrow3, ndrow1, ndcol, /* M, N, K */
&calpha, /* ALPHA: 1 */
(const cuDoubleComplex*) devPtrLx + ndrow1,
ndrow2, /* ndrow */
(const cuDoubleComplex *) devPtrLx,
ndrow2, /* ndrow */
&cbeta, /* BETA: 0 */
(cuDoubleComplex *)devPtrC + ndrow1,
ndrow2) ;
#endif
if (cublasStatus != CUBLAS_STATUS_SUCCESS)
{
ERROR (CHOLMOD_GPU_PROBLEM, "GPU CUBLAS routine failure") ;
}
}
#ifndef NTIMER
Common->CHOLMOD_GPU_GEMM_TIME += SuiteSparse_time() - tstart;
#endif
/* ------------------------------------------------------------------ */
/* Assemble the update C on the device using the d_RelativeMap */
/* ------------------------------------------------------------------ */
#ifdef REAL
addUpdateOnDevice ( gpu_p->d_A[0], devPtrC,
gpu_p->d_RelativeMap, ndrow1, ndrow2, nsrow,
&(Common->gpuStream[iDevBuff]) );
#else
addComplexUpdateOnDevice ( gpu_p->d_A[0], devPtrC,
gpu_p->d_RelativeMap, ndrow1, ndrow2, nsrow,
&(Common->gpuStream[iDevBuff]) );
#endif
/* Record an event indicating that kernels for
this descendant are complete */
cudaEventRecord ( Common->updateCKernelsComplete,
Common->gpuStream[iDevBuff]);
cudaEventRecord ( Common->updateCBuffersFree[iHostBuff],
Common->gpuStream[iDevBuff]);
return (1) ;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zgemm
*/
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, Cnorm, work[1];
magma_int_t M, N, K;
magma_int_t Am, An, Bm, 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};
magma_int_t status = 0;
magmaDoubleComplex *h_A, *h_B, *h_C, *h_Cmagma, *h_Cdev;
magmaDoubleComplex_ptr d_A, d_B, d_C;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
#ifdef HAVE_CUBLAS
// for CUDA, we can check MAGMA vs. CUBLAS, without running LAPACK
printf("If running lapack (option --lapack), MAGMA and %s error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to %s result.\n\n",
g_platform_str, g_platform_str );
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
// for others, we need LAPACK for check
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K %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];
K = opts.ksize[itest];
gflops = FLOPS_ZGEMM( M, N, K ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if ( opts.transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*An;
sizeB = ldb*Bn;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*An );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*Bn );
TESTING_MALLOC_CPU( h_C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Cmagma, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Cdev, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*An );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*Bn );
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 );
magma_zsetmatrix( Am, An, h_A, lda, d_A, ldda );
magma_zsetmatrix( Bm, Bn, h_B, ldb, d_B, lddb );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetmatrix( M, N, h_C, ldc, d_C, lddc );
magma_time = magma_sync_wtime( NULL );
magmablas_zgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, ldda,
d_B, lddb,
beta, d_C, lddc );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetmatrix( M, N, d_C, lddc, h_Cmagma, ldc );
#endif
/* =====================================================================
Performs operation using CUBLAS / clBLAS / Xeon Phi MKL
=================================================================== */
magma_zsetmatrix( M, N, h_C, ldc, d_C, lddc );
#ifdef HAVE_CUBLAS
dev_time = magma_sync_wtime( NULL );
cublasZgemm( opts.handle, cublas_trans_const(opts.transA), cublas_trans_const(opts.transB), M, N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
dev_time = magma_sync_wtime( NULL ) - dev_time;
#else
dev_time = magma_sync_wtime( opts.queue );
magma_zgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, opts.queue );
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
#endif
dev_perf = gflops / dev_time;
magma_zgetmatrix( M, N, d_C, lddc, h_Cdev, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_zgemm( lapack_trans_const(opts.transA), lapack_trans_const(opts.transB), &M, &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 & dev, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_zlange( "F", &M, &N, h_C, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cdev, &ione );
dev_error = lapackf77_zlange( "F", &M, &N, h_Cdev, &ldc, work ) / Cnorm;
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cmagma, &ione );
magma_error = lapackf77_zlange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(magma_error < tol && dev_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && dev_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(dev_error < tol ? "ok" : "failed"));
status += ! (dev_error < tol);
#endif
}
else {
#ifdef HAVE_CUBLAS
// compute relative error for magma, relative to dev (currently only with CUDA)
Cnorm = lapackf77_zlange( "F", &M, &N, h_Cdev, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_Cdev, &ione, h_Cmagma, &ione );
magma_error = lapackf77_zlange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e --- %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
magma_error,
(magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) --- ( --- ) ---\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time );
#endif
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Cmagma );
TESTING_FREE_CPU( h_Cdev );
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;
}