void magma_sgemm(
magma_trans_t transA, magma_trans_t transB,
magma_int_t m, magma_int_t n, magma_int_t k,
float alpha, float const* dA, magma_int_t lda,
float const* dB, magma_int_t ldb,
float beta, float* dC, magma_int_t ldc )
{
cublasSgemm(
cublas_trans_const( transA ),
cublas_trans_const( transB ),
m, n, k,
alpha, dA, lda,
dB, ldb,
beta, dC, ldc );
}
void magma_dgemm(
magma_trans_t transA, magma_trans_t transB,
magma_int_t m, magma_int_t n, magma_int_t k,
double alpha, double const* dA, magma_int_t lda,
double const* dB, magma_int_t ldb,
double beta, double* dC, magma_int_t ldc )
{
cublasDgemm(
cublas_trans_const( transA ),
cublas_trans_const( transB ),
m, n, k,
alpha, dA, lda,
dB, ldb,
beta, dC, ldc );
}
void magma_cgemm(
magma_trans_t transA, magma_trans_t transB,
magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloatComplex alpha,
const magmaFloatComplex *dA, magma_int_t ldda,
const magmaFloatComplex *dB, magma_int_t lddb,
magmaFloatComplex beta,
magmaFloatComplex *dC, magma_int_t lddc )
{
cublasCgemm(
cublas_trans_const( transA ),
cublas_trans_const( transB ),
m, n, k,
alpha, dA, ldda,
dB, lddb,
beta, dC, lddc );
}
/** 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 );
}
void magma_zsyrk(
magma_uplo_t uplo, magma_trans_t trans,
magma_int_t n, magma_int_t k,
cuDoubleComplex alpha, cuDoubleComplex const* dA, magma_int_t lda,
cuDoubleComplex beta, cuDoubleComplex* dC, magma_int_t ldc )
{
cublasZsyrk(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, lda,
beta, dC, ldc );
}
void magma_dsyrk(
magma_uplo_t uplo, magma_trans_t trans,
magma_int_t n, magma_int_t k,
double alpha, double const* dA, magma_int_t lda,
double beta, double* dC, magma_int_t ldc )
{
cublasDsyrk(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, lda,
beta, dC, ldc );
}
void magma_ssyrk(
magma_uplo_t uplo, magma_trans_t trans,
magma_int_t n, magma_int_t k,
float alpha, float const* dA, magma_int_t lda,
float beta, float* dC, magma_int_t ldc )
{
cublasSsyrk(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, lda,
beta, dC, ldc );
}
void magma_strsv(
magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t n,
float const *dA, magma_int_t lda,
float *dx, magma_int_t incx )
{
cublasStrsv(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
n,
dA, lda,
dx, incx );
}
void magma_dgemv(
magma_trans_t transA,
magma_int_t m, magma_int_t n,
double alpha, double const* dA, magma_int_t lda,
double const* dx, magma_int_t incx,
double beta, double* dy, magma_int_t incy )
{
cublasDgemv(
cublas_trans_const( transA ),
m, n,
alpha, dA, lda,
dx, incx,
beta, dy, incy );
}
void magma_dtrsv(
magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t n,
const double *dA, magma_int_t ldda,
double *dx, magma_int_t incx )
{
cublasDtrsv(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
n,
dA, ldda,
dx, incx );
}
void magma_ctrsv(
magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t n,
const magmaFloatComplex *dA, magma_int_t ldda,
magmaFloatComplex *dx, magma_int_t incx )
{
cublasCtrsv(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
n,
dA, ldda,
dx, incx );
}
void magma_sgemv(
magma_trans_t transA,
magma_int_t m, magma_int_t n,
float alpha, float const* dA, magma_int_t lda,
float const* dx, magma_int_t incx,
float beta, float* dy, magma_int_t incy )
{
cublasSgemv(
cublas_trans_const( transA ),
m, n,
alpha, dA, lda,
dx, incx,
beta, dy, incy );
}
/** Solve triangular matrix-vector system (one right-hand side).
\f$ A x = b \f$ (trans == MagmaNoTrans), or \n
\f$ A^T x = b \f$ (trans == MagmaTrans), or \n
\f$ A^H x = b \f$ (trans == MagmaConjTrans).
@param[in]
uplo Whether the upper or lower triangle of A is referenced.
@param[in]
trans Operation to perform on A.
@param[in]
diag Whether the diagonal of A is assumed to be unit or non-unit.
@param[in]
n Number of rows and columns of A. n >= 0.
@param[in]
dA COMPLEX_16 array of dimension (ldda,n), ldda >= max(1,n).
The n-by-n matrix A, on GPU device.
@param[in]
ldda Leading dimension of dA.
@param[in,out]
dx COMPLEX_16 array on GPU device.
On entry, the n element RHS vector b of dimension (1 + (n-1)*incx).
On exit, overwritten with the solution vector x.
@param[in]
incx Stride between consecutive elements of dx. incx != 0.
@ingroup magma_zblas2
*/
extern "C" void
magma_ztrsv(
magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t n,
magmaDoubleComplex_const_ptr dA, magma_int_t ldda,
magmaDoubleComplex_ptr dx, magma_int_t incx )
{
cublasZtrsv(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
n,
dA, ldda,
dx, incx );
}
void magma_cherk(
magma_uplo_t uplo, magma_trans_t trans,
magma_int_t n, magma_int_t k,
float alpha,
const magmaFloatComplex *dA, magma_int_t ldda,
float beta,
magmaFloatComplex *dC, magma_int_t lddc )
{
cublasCherk(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, ldda,
beta, dC, lddc );
}
void magma_strsm(
magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t m, magma_int_t n,
float alpha, float const* dA, magma_int_t lda,
float* dB, magma_int_t ldb )
{
cublasStrsm(
cublas_side_const( side ),
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
m, n,
alpha, dA, lda,
dB, ldb );
}
void magma_dtrmm(
magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t m, magma_int_t n,
double alpha, double const *dA, magma_int_t lda,
double *dB, magma_int_t ldb )
{
cublasDtrmm(
cublas_side_const( side ),
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
m, n,
alpha, dA, lda,
dB, ldb );
}
void magma_ctrsm(
magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t m, magma_int_t n,
magmaFloatComplex alpha,
const magmaFloatComplex *dA, magma_int_t ldda,
magmaFloatComplex *dB, magma_int_t lddb )
{
cublasCtrsm(
cublas_side_const( side ),
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
m, n,
alpha, dA, ldda,
dB, lddb );
}
/** Perform Hermitian rank-k update.
\f$ C = \alpha A A^T + \beta C \f$ (trans == MagmaNoTrans), or \n
\f$ C = \alpha A^T A + \beta C \f$ (trans == MagmaTrans), \n
where \f$ C \f$ is Hermitian.
@param[in]
uplo Whether the upper or lower triangle of C is referenced.
@param[in]
trans Operation to perform on A.
@param[in]
n Number of rows and columns of C. n >= 0.
@param[in]
k Number of columns of A (for MagmaNoTrans) or rows of A (for MagmaTrans). k >= 0.
@param[in]
alpha Scalar \f$ \alpha \f$
@param[in]
dA COMPLEX_16 array on GPU device.
If trans == MagmaNoTrans, the n-by-k matrix A of dimension (ldda,k), ldda >= max(1,n); \n
otherwise, the k-by-n matrix A of dimension (ldda,n), ldda >= max(1,k).
@param[in]
ldda Leading dimension of dA.
@param[in]
beta Scalar \f$ \beta \f$
@param[in,out]
dC COMPLEX_16 array on GPU device.
The n-by-n Hermitian matrix C of dimension (lddc,n), lddc >= max(1,n).
@param[in]
lddc Leading dimension of dC.
@ingroup magma_zblas3
*/
extern "C" void
magma_zherk(
magma_uplo_t uplo, magma_trans_t trans,
magma_int_t n, magma_int_t k,
double alpha,
magmaDoubleComplex_const_ptr dA, magma_int_t ldda,
double beta,
magmaDoubleComplex_ptr dC, magma_int_t lddc )
{
cublasZherk(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, ldda,
beta, dC, lddc );
}
void magma_cgemv(
magma_trans_t transA,
magma_int_t m, magma_int_t n,
magmaFloatComplex alpha,
const magmaFloatComplex *dA, magma_int_t ldda,
const magmaFloatComplex *dx, magma_int_t incx,
magmaFloatComplex beta,
magmaFloatComplex *dy, magma_int_t incy )
{
cublasCgemv(
cublas_trans_const( transA ),
m, n,
alpha, dA, ldda,
dx, incx,
beta, dy, incy );
}
void magma_csyr2k(
magma_uplo_t uplo, magma_trans_t trans,
magma_int_t n, magma_int_t k,
magmaFloatComplex alpha,
const magmaFloatComplex *dA, magma_int_t ldda,
const magmaFloatComplex *dB, magma_int_t lddb,
magmaFloatComplex beta,
magmaFloatComplex *dC, magma_int_t lddc )
{
cublasCsyr2k(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, ldda,
dB, lddb,
beta, dC, lddc );
}
/** Solve triangular matrix-matrix system (multiple right-hand sides).
\f$ op(A) X = \alpha B \f$ (side == MagmaLeft), or \n
\f$ X op(A) = \alpha B \f$ (side == MagmaRight), \n
where \f$ A \f$ is triangular.
@param[in]
side Whether A is on the left or right.
@param[in]
uplo Whether A is upper or lower triangular.
@param[in]
trans Operation to perform on A.
@param[in]
diag Whether the diagonal of A is assumed to be unit or non-unit.
@param[in]
m Number of rows of B. m >= 0.
@param[in]
n Number of columns of B. n >= 0.
@param[in]
alpha Scalar \f$ \alpha \f$
@param[in]
dA COMPLEX_16 array on GPU device.
If side == MagmaLeft, the m-by-m triangular matrix A of dimension (ldda,m), ldda >= max(1,m); \n
otherwise, the n-by-n triangular matrix A of dimension (ldda,n), ldda >= max(1,n).
@param[in]
ldda Leading dimension of dA.
@param[in,out]
dB COMPLEX_16 array on GPU device.
On entry, m-by-n matrix B of dimension (lddb,n), lddb >= max(1,m).
On exit, overwritten with the solution matrix X.
@param[in]
lddb Leading dimension of dB.
@ingroup magma_zblas3
*/
extern "C" void
magma_ztrsm(
magma_side_t side, magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag,
magma_int_t m, magma_int_t n,
magmaDoubleComplex alpha,
magmaDoubleComplex_const_ptr dA, magma_int_t ldda,
magmaDoubleComplex_ptr dB, magma_int_t lddb )
{
cublasZtrsm(
cublas_side_const( side ),
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
cublas_diag_const( diag ),
m, n,
alpha, dA, ldda,
dB, lddb );
}
/** Perform matrix-vector product.
\f$ y = \alpha A x + \beta y \f$ (transA == MagmaNoTrans), or \n
\f$ y = \alpha A^T x + \beta y \f$ (transA == MagmaTrans), or \n
\f$ y = \alpha A^H x + \beta y \f$ (transA == MagmaConjTrans).
@param[in]
transA Operation to perform on A.
@param[in]
m Number of rows of A. m >= 0.
@param[in]
n Number of columns of A. n >= 0.
@param[in]
alpha Scalar \f$ \alpha \f$
@param[in]
dA COMPLEX_16 array of dimension (ldda,n), ldda >= max(1,m).
The m-by-n matrix A, on GPU device.
@param[in]
ldda Leading dimension of dA.
@param[in]
dx COMPLEX_16 array on GPU device.
If transA == MagmaNoTrans, the n element vector x of dimension (1 + (n-1)*incx); \n
otherwise, the m element vector x of dimension (1 + (m-1)*incx).
@param[in]
incx Stride between consecutive elements of dx. incx != 0.
@param[in]
beta Scalar \f$ \beta \f$
@param[in,out]
dy COMPLEX_16 array on GPU device.
If transA == MagmaNoTrans, the m element vector y of dimension (1 + (m-1)*incy); \n
otherwise, the n element vector y of dimension (1 + (n-1)*incy).
@param[in]
incy Stride between consecutive elements of dy. incy != 0.
@ingroup magma_zblas2
*/
extern "C" void
magma_zgemv(
magma_trans_t transA,
magma_int_t m, magma_int_t n,
magmaDoubleComplex alpha,
magmaDoubleComplex_const_ptr dA, magma_int_t ldda,
magmaDoubleComplex_const_ptr dx, magma_int_t incx,
magmaDoubleComplex beta,
magmaDoubleComplex_ptr dy, magma_int_t incy )
{
cublasZgemv(
cublas_trans_const( transA ),
m, n,
alpha, dA, ldda,
dx, incx,
beta, dy, incy );
}
/** Perform symmetric rank-2k update.
\f$ C = \alpha A B^T + \alpha B A^T \beta C \f$ (trans == MagmaNoTrans), or \n
\f$ C = \alpha A^T B + \alpha B^T A \beta C \f$ (trans == MagmaTrans), \n
where \f$ C \f$ is symmetric.
@param[in]
uplo Whether the upper or lower triangle of C is referenced.
@param[in]
trans Operation to perform on A and B.
@param[in]
n Number of rows and columns of C. n >= 0.
@param[in]
k Number of columns of A and B (for MagmaNoTrans) or rows of A and B (for MagmaTrans). k >= 0.
@param[in]
alpha Scalar \f$ \alpha \f$
@param[in]
dA COMPLEX_16 array on GPU device.
If trans == MagmaNoTrans, the n-by-k matrix A of dimension (ldda,k), ldda >= max(1,n); \n
otherwise, the k-by-n matrix A of dimension (ldda,n), ldda >= max(1,k).
@param[in]
ldda Leading dimension of dA.
@param[in]
dB COMPLEX_16 array on GPU device.
If trans == MagmaNoTrans, the n-by-k matrix B of dimension (lddb,k), lddb >= max(1,n); \n
otherwise, the k-by-n matrix B of dimension (lddb,n), lddb >= max(1,k).
@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 n-by-n symmetric matrix C of dimension (lddc,n), lddc >= max(1,n).
@param[in]
lddc Leading dimension of dC.
@ingroup magma_zblas3
*/
extern "C" void
magma_zsyr2k(
magma_uplo_t uplo, magma_trans_t trans,
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 )
{
cublasZsyr2k(
cublas_uplo_const( uplo ),
cublas_trans_const( trans ),
n, k,
alpha, dA, ldda,
dB, lddb,
beta, dC, lddc );
}
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_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;
double c_neg_one = MAGMA_D_NEG_ONE;
double alpha = MAGMA_D_MAKE( 1.5, -2.3 );
double beta = MAGMA_D_MAKE( -0.6, 0.8 );
double *A, *X, *Y, *Ycublas, *Ymagma;
double *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_DGEMV( 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, double, sizeA );
TESTING_MALLOC_CPU( X, double, sizeX );
TESTING_MALLOC_CPU( Y, double, sizeY );
TESTING_MALLOC_CPU( Ycublas, double, sizeY );
TESTING_MALLOC_CPU( Ymagma, double, sizeY );
TESTING_MALLOC_DEV( dA, double, sizeA );
TESTING_MALLOC_DEV( dX, double, sizeX );
TESTING_MALLOC_DEV( dY, double, sizeY );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &sizeA, A );
lapackf77_dlarnv( &ione, ISEED, &sizeX, X );
lapackf77_dlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_dsetmatrix( M, N, A, lda, dA, lda );
magma_dsetvector( Xm, X, incx, dX, incx );
magma_dsetvector( Ym, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasDgemv( 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_dgetvector( Ym, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_dsetvector( Ym, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
magmablas_dgemv( 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_dgetvector( Ym, dY, incx, Ymagma, incx );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_dgemv( 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_daxpy( &Ym, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_dlange( "M", &Ym, &ione, Ymagma, &Ym, work ) / Ym;
blasf77_daxpy( &Ym, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_dlange( "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 sgemm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
float 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;
float *h_A, *h_B, *h_C, *h_Cmagma, *h_Cdev;
magmaFloat_ptr d_A, d_B, d_C;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 0.29, -0.86 );
float beta = MAGMA_S_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("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_SGEMM( 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, float, lda*An );
TESTING_MALLOC_CPU( h_B, float, ldb*Bn );
TESTING_MALLOC_CPU( h_C, float, ldc*N );
TESTING_MALLOC_CPU( h_Cmagma, float, ldc*N );
TESTING_MALLOC_CPU( h_Cdev, float, ldc*N );
TESTING_MALLOC_DEV( d_A, float, ldda*An );
TESTING_MALLOC_DEV( d_B, float, lddb*Bn );
TESTING_MALLOC_DEV( d_C, float, lddc*N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_slarnv( &ione, ISEED, &sizeC, h_C );
magma_ssetmatrix( Am, An, h_A, lda, d_A, 0, ldda, opts.queue );
magma_ssetmatrix( Bm, Bn, h_B, ldb, d_B, 0, lddb, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_ssetmatrix( M, N, h_C, ldc, d_C, lddc );
magma_time = magma_sync_wtime( NULL );
magmablas_sgemm( 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_sgetmatrix( M, N, d_C, lddc, h_Cmagma, ldc );
#endif
/* =====================================================================
Performs operation using CUBLAS / clBLAS / Xeon Phi MKL
=================================================================== */
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, opts.queue );
#ifdef HAVE_CUBLAS
dev_time = magma_sync_wtime( NULL );
cublasSgemm( 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_sgemm( 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_sgetmatrix( M, N, d_C, 0, lddc, h_Cdev, ldc, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_sgemm( 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_slange( "F", &M, &N, h_C, &ldc, work );
blasf77_saxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cdev, &ione );
dev_error = lapackf77_slange( "F", &M, &N, h_Cdev, &ldc, work ) / Cnorm;
#ifdef HAVE_CUBLAS
blasf77_saxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cmagma, &ione );
magma_error = lapackf77_slange( "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_slange( "F", &M, &N, h_Cdev, &ldc, work );
blasf77_saxpy( &sizeC, &c_neg_one, h_Cdev, &ione, h_Cmagma, &ione );
magma_error = lapackf77_slange( "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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zher2k
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
double cublas_error, Cnorm, work[1];
magma_int_t N, K;
magma_int_t Ak, An, Bk, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex *h_A, *h_B, *h_C, *h_Ccublas;
magmaDoubleComplex *d_A, *d_B, *d_C;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.29, -0.86 );
double beta = MAGMA_D_MAKE( -0.48, 0.38 );
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("If running lapack (option --lapack), CUBLAS error is computed\n"
"relative to CPU BLAS result.\n\n");
printf("uplo = %s, transA = %s\n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA) );
printf(" N K CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("==================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.msize[itest];
K = opts.ksize[itest];
gflops = FLOPS_ZHER2K(K, N) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = An = N;
Ak = K;
ldb = Bn = N;
Bk = K;
} else {
lda = An = K;
Ak = N;
ldb = Bn = K;
Bk = N;
}
ldc = N;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*Ak;
sizeB = ldb*Ak;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, lda*Ak );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, ldb*Bk );
TESTING_MALLOC_CPU( h_C, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_CPU( h_Ccublas, magmaDoubleComplex, ldc*N );
TESTING_MALLOC_DEV( d_A, magmaDoubleComplex, ldda*Ak );
TESTING_MALLOC_DEV( d_B, magmaDoubleComplex, lddb*Bk );
TESTING_MALLOC_DEV( d_C, magmaDoubleComplex, lddc*N );
/* Initialize the matrices */
lapackf77_zlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_zlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_zlarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_zsetmatrix( Bn, Bk, h_B, ldb, d_B, lddb );
magma_zsetmatrix( N, N, h_C, ldc, d_C, lddc );
cublas_time = magma_sync_wtime( NULL );
cublasZher2k( handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA), N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
cublas_time = magma_sync_wtime( NULL ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_zher2k( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_zlange( "M", &N, &N, h_C, &ldc, work );
blasf77_zaxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_zlange( "M", &N, &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Ccublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsyr2k
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
double cublas_error, Cnorm, work[1];
magma_int_t N, K;
magma_int_t Ak, An, Bk, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double *h_A, *h_B, *h_C, *h_Ccublas;
magmaDouble_ptr d_A, d_B, d_C;
double c_neg_one = MAGMA_D_NEG_ONE;
double alpha = MAGMA_D_MAKE( 0.29, -0.86 );
double beta = MAGMA_D_MAKE( -0.48, 0.38 );
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
#ifdef COMPLEX
if (opts.transA == MagmaTrans) {
opts.transA = MagmaConjTrans;
printf("%% WARNING: transA = MagmaTrans changed to MagmaConjTrans\n");
}
#endif
printf("%% If running lapack (option --lapack), CUBLAS error is computed\n"
"%% relative to CPU BLAS result.\n\n");
printf("%% uplo = %s, transA = %s\n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA) );
printf("%% N K CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("%%=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.msize[itest];
K = opts.ksize[itest];
gflops = FLOPS_DSYR2K(K, N) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = An = N;
Ak = K;
ldb = Bn = N;
Bk = K;
} else {
lda = An = K;
Ak = N;
ldb = Bn = K;
Bk = N;
}
ldc = N;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
lddb = magma_roundup( ldb, opts.align ); // multiple of 32 by default
lddc = magma_roundup( ldc, opts.align ); // multiple of 32 by default
sizeA = lda*Ak;
sizeB = ldb*Ak;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, double, lda*Ak );
TESTING_MALLOC_CPU( h_B, double, ldb*Bk );
TESTING_MALLOC_CPU( h_C, double, ldc*N );
TESTING_MALLOC_CPU( h_Ccublas, double, ldc*N );
TESTING_MALLOC_DEV( d_A, double, ldda*Ak );
TESTING_MALLOC_DEV( d_B, double, lddb*Bk );
TESTING_MALLOC_DEV( d_C, double, lddc*N );
/* Initialize the matrices */
lapackf77_dlarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_dlarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_dlarnv( &ione, ISEED, &sizeC, h_C );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_dsetmatrix( An, Ak, h_A, lda, d_A, ldda );
magma_dsetmatrix( Bn, Bk, h_B, ldb, d_B, lddb );
magma_dsetmatrix( N, N, h_C, ldc, d_C, lddc );
magmablasSetKernelStream( opts.queue ); // opts.handle also uses opts.queue
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasDsyr2k( opts.handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA), N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
#else
magma_dsyr2k( opts.uplo, opts.transA, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_dgetmatrix( N, N, d_C, lddc, h_Ccublas, ldc );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_dsyr2k( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & cublas, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_dlange( "M", &N, &N, h_C, &ldc, work );
blasf77_daxpy( &sizeC, &c_neg_one, h_C, &ione, h_Ccublas, &ione );
cublas_error = lapackf77_dlange( "M", &N, &N, h_Ccublas, &ldc, work ) / Cnorm;
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N, (int) K,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Ccublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
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;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing strmv
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, cublas_perf, cublas_time, cpu_perf, cpu_time;
float cublas_error, Cnorm, work[1];
magma_int_t N;
magma_int_t Ak;
magma_int_t sizeA;
magma_int_t lda, ldda;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float *h_A, *h_x, *h_xcublas;
magmaFloat_ptr d_A, d_x;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
float tol = opts.tolerance * lapackf77_slamch("E");
printf("%% If running lapack (option --lapack), CUBLAS error is computed\n"
"%% relative to CPU BLAS result.\n\n");
printf("%% uplo = %s, transA = %s, diag = %s \n",
lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA),
lapack_diag_const(opts.diag) );
printf("%% N CUBLAS Gflop/s (ms) CPU Gflop/s (ms) CUBLAS error\n");
printf("%%=================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
gflops = FLOPS_STRMM(opts.side, N, 1) / 1e9;
lda = N;
Ak = N;
ldda = magma_roundup( lda, opts.align ); // multiple of 32 by default
sizeA = lda*Ak;
TESTING_MALLOC_CPU( h_A, float, lda*Ak );
TESTING_MALLOC_CPU( h_x, float, N );
TESTING_MALLOC_CPU( h_xcublas, float, N );
TESTING_MALLOC_DEV( d_A, float, ldda*Ak );
TESTING_MALLOC_DEV( d_x, float, N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &N, h_x );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_ssetmatrix( Ak, Ak, h_A, lda, d_A, ldda, opts.queue );
magma_ssetvector( N, h_x, 1, d_x, 1, opts.queue );
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasStrmv( opts.handle, cublas_uplo_const(opts.uplo), cublas_trans_const(opts.transA),
cublas_diag_const(opts.diag),
N,
d_A, ldda,
d_x, 1 );
#else
magma_strmv( opts.uplo, opts.transA, opts.diag,
N,
d_A, 0, ldda,
d_x, 0, 1, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_sgetvector( N, d_x, 1, h_xcublas, 1, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_strmv( lapack_uplo_const(opts.uplo), lapack_trans_const(opts.transA), lapack_diag_const(opts.diag),
&N,
h_A, &lda,
h_x, &ione );
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_slange( "M", &N, &ione, h_x, &N, work );
blasf77_saxpy( &N, &c_neg_one, h_x, &ione, h_xcublas, &ione );
cublas_error = lapackf77_slange( "M", &N, &ione, h_xcublas, &N, work ) / Cnorm;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
cublas_error, (cublas_error < tol ? "ok" : "failed"));
status += ! (cublas_error < tol);
}
else {
printf("%5d %7.2f (%7.2f) --- ( --- ) --- ---\n",
(int) N,
cublas_perf, 1000.*cublas_time);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_x );
TESTING_FREE_CPU( h_xcublas );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_x );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
