// --------------------
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, cpu_time=0, cpu_perf=0, gpu_time, gpu_perf, mgpu_time, mgpu_perf, cuda_time, cuda_perf;
double Ynorm, error=0, error2=0, work[1];
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t n_local[MagmaMaxGPUs];
magma_int_t N, Noffset, lda, ldda, blocks, lhwork, ldwork, matsize, vecsize;
magma_int_t incx = 1;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ylapack, *Ycublas, *Ymagma, *Ymagma1, *hwork;
magmaDoubleComplex_ptr dA, dX, dY;
magmaDoubleComplex_ptr d_lA[MagmaMaxGPUs], dwork[MagmaMaxGPUs];
magma_device_t dev;
magma_queue_t queues[MagmaMaxGPUs];
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
double tol = opts.tolerance * lapackf77_dlamch("E");
magma_int_t nb = 64; // required by magmablas_zhemv_mgpu implementation
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_create( dev, &queues[dev] );
}
// currently, tests all offsets in the offsets array;
// comment out loop below to test a specific offset.
magma_int_t offset = opts.offset;
magma_int_t offsets[] = { 0, 1, 31, 32, 33, 63, 64, 65, 100, 200 };
magma_int_t noffsets = sizeof(offsets) / sizeof(*offsets);
printf("%% uplo = %s, ngpu %d, block size = %d, offset %d\n",
lapack_uplo_const(opts.uplo), (int) opts.ngpu, (int) nb, (int) offset );
printf( "%% BLAS CUBLAS MAGMA 1 GPU MAGMA MGPU Error rel Error rel\n"
"%% N offset Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) Gflop/s (msec) to CUBLAS to LAPACK\n"
"%%==================================================================================================================\n" );
for( int itest = 0; itest < opts.ntest; ++itest ) {
// comment out these two lines & end of loop to test a specific offset
for( int ioffset=0; ioffset < noffsets; ioffset += 1 ) {
offset = offsets[ioffset];
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
Noffset = N + offset;
lda = Noffset;
ldda = magma_roundup( Noffset, opts.align ); // multiple of 32 by default
matsize = Noffset*ldda;
vecsize = (Noffset-1)*incx + 1;
gflops = FLOPS_ZHEMV( N ) / 1e9;
blocks = magma_ceildiv( N + (offset % nb), nb );
lhwork = N*opts.ngpu;
ldwork = ldda*(blocks + 1);
TESTING_MALLOC_CPU( A, magmaDoubleComplex, matsize );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ymagma1, magmaDoubleComplex, vecsize );
TESTING_MALLOC_CPU( Ylapack, magmaDoubleComplex, vecsize );
TESTING_MALLOC_PIN( X, magmaDoubleComplex, vecsize );
TESTING_MALLOC_PIN( hwork, magmaDoubleComplex, lhwork );
magma_setdevice( opts.device );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, matsize );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, vecsize );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, vecsize );
// TODO make magma_zmalloc_bcyclic helper function?
for( dev=0; dev < opts.ngpu; dev++ ) {
n_local[dev] = ((Noffset/nb)/opts.ngpu)*nb;
if (dev < (Noffset/nb) % opts.ngpu)
n_local[dev] += nb;
else if (dev == (Noffset/nb) % opts.ngpu)
n_local[dev] += Noffset % nb;
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], magmaDoubleComplex, ldda*n_local[dev] );
TESTING_MALLOC_DEV( dwork[dev], magmaDoubleComplex, ldwork );
}
//////////////////////////////////////////////////////////////////////////
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &matsize, A );
magma_zmake_hermitian( Noffset, A, lda );
lapackf77_zlarnv( &ione, ISEED, &vecsize, X );
lapackf77_zlarnv( &ione, ISEED, &vecsize, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_setdevice( opts.device );
magma_zsetmatrix( Noffset, Noffset, A, lda, dA, ldda, opts.queue );
magma_zsetvector( Noffset, X, incx, dX, incx, opts.queue );
magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue );
cuda_time = magma_sync_wtime(0);
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo), N,
&alpha, dA + offset + offset*ldda, ldda,
dX + offset, incx,
&beta, dY + offset, incx );
cuda_time = magma_sync_wtime(0) - cuda_time;
cuda_perf = gflops / cuda_time;
magma_zgetvector( Noffset, dY, incx, Ycublas, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (1 GPU)
=================================================================== */
magma_setdevice( opts.device );
magma_zsetvector( Noffset, Y, incx, dY, incx, opts.queue );
gpu_time = magma_sync_wtime( opts.queue );
magmablas_zhemv_work( opts.uplo, N,
alpha, dA + offset + offset*ldda, ldda,
dX + offset, incx,
beta, dY + offset, incx, dwork[ opts.device ], ldwork,
opts.queue );
gpu_time = magma_sync_wtime( opts.queue ) - gpu_time;
gpu_perf = gflops / gpu_time;
magma_zgetvector( Noffset, dY, incx, Ymagma1, incx, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (multi-GPU)
=================================================================== */
magma_zsetmatrix_1D_col_bcyclic( Noffset, Noffset, A, lda, d_lA, ldda, opts.ngpu, nb, queues );
blasf77_zcopy( &Noffset, Y, &incx, Ymagma, &incx );
// workspaces do NOT need to be zero -- set to NAN to prove
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_setdevice( dev );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork[dev], ldwork, opts.queue );
}
lapackf77_zlaset( "Full", &lhwork, &ione, &MAGMA_Z_NAN, &MAGMA_Z_NAN, hwork, &lhwork );
mgpu_time = magma_sync_wtime(0);
magma_int_t info;
info = magmablas_zhemv_mgpu(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_zhemv_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
info = magmablas_zhemv_mgpu_sync(
opts.uplo, N,
alpha,
d_lA, ldda, offset,
X + offset, incx,
beta,
Ymagma + offset, incx,
hwork, lhwork,
dwork, ldwork,
opts.ngpu, nb, queues );
if (info != 0) {
printf("magmablas_zhemv_sync returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
mgpu_time = magma_sync_wtime(0) - mgpu_time;
mgpu_perf = gflops / mgpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
blasf77_zcopy( &Noffset, Y, &incx, Ylapack, &incx );
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(opts.uplo), &N,
&alpha, A + offset + offset*lda, &lda,
X + offset, &incx,
&beta, Ylapack + offset, &incx );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Compute the Difference LAPACK vs. Magma
=================================================================== */
Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work );
blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ylapack, &incx );
error2 = lapackf77_zlange( "F", &Noffset, &ione, Ylapack, &Noffset, work ) / Ynorm;
}
/* =====================================================================
Compute the Difference Cublas vs. Magma
=================================================================== */
Ynorm = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work );
blasf77_zaxpy( &Noffset, &c_neg_one, Ymagma, &incx, Ycublas, &incx );
error = lapackf77_zlange( "F", &Noffset, &ione, Ycublas, &Noffset, work ) / Ynorm;
bool okay = (error < tol && error2 < tol);
status += ! okay;
if ( opts.lapack ) {
printf( "%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) N, (int) offset,
cpu_perf, cpu_time*1000.,
cuda_perf, cuda_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, error2, (okay ? "ok" : "failed") );
}
else {
printf( "%5d %5d --- ( --- ) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e --- %s\n",
(int) N, (int) offset,
cuda_perf, cuda_time*1000.,
gpu_perf, gpu_time*1000.,
mgpu_perf, mgpu_time*1000.,
error, (okay ? "ok" : "failed") );
}
/* Free Memory */
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_CPU( Ymagma1 );
TESTING_FREE_CPU( Ylapack );
TESTING_FREE_PIN( X );
TESTING_FREE_PIN( hwork );
magma_setdevice( opts.device );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
for( dev=0; dev < opts.ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
TESTING_FREE_DEV( dwork[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
// comment out these two lines line & top of loop test a specific offset
} // end for ioffset
printf( "\n" );
}
for( dev=0; dev < opts.ngpu; ++dev ) {
magma_queue_destroy( queues[dev] );
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
ZLATRD2 reduces NB rows and columns of a complex Hermitian matrix A to
Hermitian tridiagonal form by an orthogonal similarity
transformation Q' * A * Q, and returns the matrices V and W which are
needed to apply the transformation to the unreduced part of A.
If UPLO = MagmaUpper, ZLATRD reduces the last NB rows and columns of a
matrix, of which the upper triangle is supplied;
if UPLO = MagmaLower, ZLATRD reduces the first NB rows and columns of a
matrix, of which the lower triangle is supplied.
This is an auxiliary routine called by ZHETRD2_GPU. It uses an
accelerated HEMV that needs extra memory.
Arguments
---------
@param[in]
uplo magma_uplo_t
Specifies whether the upper or lower triangular part of the
Hermitian matrix A is stored:
- = MagmaUpper: Upper triangular
- = MagmaLower: Lower triangular
@param[in]
n INTEGER
The order of the matrix A.
@param[in]
nb INTEGER
The number of rows and columns to be reduced.
@param[in,out]
A COMPLEX_16 array, dimension (LDA,N)
On entry, the Hermitian matrix A. If UPLO = MagmaUpper, the leading
n-by-n upper triangular part of A contains the upper
triangular part of the matrix A, and the strictly lower
triangular part of A is not referenced. If UPLO = MagmaLower, the
leading n-by-n lower triangular part of A contains the lower
triangular part of the matrix A, and the strictly upper
triangular part of A is not referenced.
On exit:
- if UPLO = MagmaUpper, the last NB columns have been reduced to
tridiagonal form, with the diagonal elements overwriting
the diagonal elements of A; the elements above the diagonal
with the array TAU, represent the orthogonal matrix Q as a
product of elementary reflectors;
- if UPLO = MagmaLower, the first NB columns have been reduced to
tridiagonal form, with the diagonal elements overwriting
the diagonal elements of A; the elements below the diagonal
with the array TAU, represent the orthogonal matrix Q as a
product of elementary reflectors.
See Further Details.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= (1,N).
@param[out]
e COMPLEX_16 array, dimension (N-1)
If UPLO = MagmaUpper, E(n-nb:n-1) contains the superdiagonal
elements of the last NB columns of the reduced matrix;
if UPLO = MagmaLower, E(1:nb) contains the subdiagonal elements of
the first NB columns of the reduced matrix.
@param[out]
tau COMPLEX_16 array, dimension (N-1)
The scalar factors of the elementary reflectors, stored in
TAU(n-nb:n-1) if UPLO = MagmaUpper, and in TAU(1:nb) if UPLO = MagmaLower.
See Further Details.
@param[out]
W COMPLEX_16 array, dimension (LDW,NB)
The n-by-nb matrix W required to update the unreduced part
of A.
@param[in]
ldw INTEGER
The leading dimension of the array W. LDW >= max(1,N).
Further Details
---------------
If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary
reflectors
Q = H(n) H(n-1) . . . H(n-nb+1).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(i:n) = 0 and v(i-1) = 1; v(1:i-1) is stored on exit in A(1:i-1,i),
and tau in TAU(i-1).
If UPLO = MagmaLower, the matrix Q is represented as a product of elementary
reflectors
Q = H(1) H(2) . . . H(nb).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(1:i) = 0 and v(i+1) = 1; v(i+1:n) is stored on exit in A(i+1:n,i),
and tau in TAU(i).
The elements of the vectors v together form the n-by-nb matrix V
which is needed, with W, to apply the transformation to the unreduced
part of the matrix, using a Hermitian rank-2k update of the form:
A := A - V*W' - W*V'.
The contents of A on exit are illustrated by the following examples
with n = 5 and nb = 2:
if UPLO = MagmaUpper: if UPLO = MagmaLower:
( a a a v4 v5 ) ( d )
( a a v4 v5 ) ( 1 d )
( a 1 v5 ) ( v1 1 a )
( d 1 ) ( v1 v2 a a )
( d ) ( v1 v2 a a a )
where d denotes a diagonal element of the reduced matrix, a denotes
an element of the original matrix that is unchanged, and vi denotes
an element of the vector defining H(i).
@ingroup magma_zheev_aux
********************************************************************/
extern "C" magma_int_t
magma_zlatrd2(magma_uplo_t uplo, magma_int_t n, magma_int_t nb,
magmaDoubleComplex *A, magma_int_t lda,
double *e, magmaDoubleComplex *tau,
magmaDoubleComplex *W, magma_int_t ldw,
magmaDoubleComplex *dA, magma_int_t ldda,
magmaDoubleComplex *dW, magma_int_t lddw,
magmaDoubleComplex *dwork, magma_int_t ldwork)
{
#define A(i, j) (A + (j)*lda + (i))
#define W(i, j) (W + (j)*ldw + (i))
#define dA(i, j) (dA + (j)*ldda + (i))
#define dW(i, j) (dW + (j)*lddw + (i))
magma_int_t i;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex value = MAGMA_Z_ZERO;
magma_int_t ione = 1;
magma_int_t i_n, i_1, iw;
magmaDoubleComplex alpha;
magmaDoubleComplex *f;
if (n <= 0) {
return 0;
}
magma_queue_t stream;
magma_queue_create( &stream );
magma_zmalloc_cpu( &f, n );
assert( f != NULL ); // TODO return error, or allocate outside zlatrd
if (uplo == MagmaUpper) {
/* Reduce last NB columns of upper triangle */
for (i = n-1; i >= n - nb; --i) {
i_1 = i + 1;
i_n = n - i - 1;
iw = i - n + nb;
if (i < n-1) {
/* Update A(1:i,i) */
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_zlacgv(&i_n, W(i, iw+1), &ldw);
#endif
blasf77_zgemv("No transpose", &i_1, &i_n, &c_neg_one, A(0, i+1), &lda,
W(i, iw+1), &ldw, &c_one, A(0, i), &ione);
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_zlacgv(&i_n, W(i, iw+1), &ldw);
lapackf77_zlacgv(&i_n, A(i, i+1), &ldw);
#endif
blasf77_zgemv("No transpose", &i_1, &i_n, &c_neg_one, W(0, iw+1), &ldw,
A(i, i+1), &lda, &c_one, A(0, i), &ione);
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_zlacgv(&i_n, A(i, i+1), &ldw);
#endif
}
if (i > 0) {
/* Generate elementary reflector H(i) to annihilate A(1:i-2,i) */
alpha = *A(i-1, i);
lapackf77_zlarfg(&i, &alpha, A(0, i), &ione, &tau[i - 1]);
e[i-1] = MAGMA_Z_REAL( alpha );
*A(i-1,i) = MAGMA_Z_MAKE( 1, 0 );
/* Compute W(1:i-1,i) */
// 1. Send the block reflector A(0:n-i-1,i) to the GPU
magma_zsetvector( i, A(0, i), 1, dA(0, i), 1 );
//#if (GPUSHMEM < 200)
//magma_zhemv(MagmaUpper, i, c_one, dA(0, 0), ldda,
// dA(0, i), ione, c_zero, dW(0, iw), ione);
//#else
magmablas_zhemv_work(MagmaUpper, i, c_one, dA(0, 0), ldda,
dA(0, i), ione, c_zero, dW(0, iw), ione,
dwork, ldwork);
//#endif
// 2. Start putting the result back (asynchronously)
magma_zgetmatrix_async( i, 1,
dW(0, iw), lddw,
W(0, iw) /*test*/, ldw, stream );
if (i < n-1) {
blasf77_zgemv(MagmaConjTransStr, &i, &i_n, &c_one, W(0, iw+1), &ldw,
A(0, i), &ione, &c_zero, W(i+1, iw), &ione);
}
// 3. Here is where we need it // TODO find the right place
magma_queue_sync( stream );
if (i < n-1) {
blasf77_zgemv("No transpose", &i, &i_n, &c_neg_one, A(0, i+1), &lda,
W(i+1, iw), &ione, &c_one, W(0, iw), &ione);
blasf77_zgemv(MagmaConjTransStr, &i, &i_n, &c_one, A(0, i+1), &lda,
A(0, i), &ione, &c_zero, W(i+1, iw), &ione);
blasf77_zgemv("No transpose", &i, &i_n, &c_neg_one, W(0, iw+1), &ldw,
W(i+1, iw), &ione, &c_one, W(0, iw), &ione);
}
blasf77_zscal(&i, &tau[i - 1], W(0, iw), &ione);
#if defined(PRECISION_z) || defined(PRECISION_c)
cblas_zdotc_sub( i, W(0,iw), ione, A(0,i), ione, &value );
#else
value = cblas_zdotc( i, W(0,iw), ione, A(0,i), ione );
#endif
alpha = tau[i - 1] * -0.5f * value;
blasf77_zaxpy(&i, &alpha, A(0, i), &ione,
W(0, iw), &ione);
}
}
}
else {
/* Reduce first NB columns of lower triangle */
for (i = 0; i < nb; ++i) {
/* Update A(i:n,i) */
i_n = n - i;
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_zlacgv(&i, W(i, 0), &ldw);
#endif
blasf77_zgemv("No transpose", &i_n, &i, &c_neg_one, A(i, 0), &lda,
W(i, 0), &ldw, &c_one, A(i, i), &ione);
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_zlacgv(&i, W(i, 0), &ldw);
lapackf77_zlacgv(&i, A(i, 0), &lda);
#endif
blasf77_zgemv("No transpose", &i_n, &i, &c_neg_one, W(i, 0), &ldw,
A(i, 0), &lda, &c_one, A(i, i), &ione);
#if defined(PRECISION_z) || defined(PRECISION_c)
lapackf77_zlacgv(&i, A(i, 0), &lda);
#endif
if (i < n-1) {
/* Generate elementary reflector H(i) to annihilate A(i+2:n,i) */
i_n = n - i - 1;
alpha = *A(i+1, i);
lapackf77_zlarfg(&i_n, &alpha, A(min(i+2,n-1), i), &ione, &tau[i]);
e[i] = MAGMA_Z_REAL( alpha );
*A(i+1,i) = MAGMA_Z_MAKE( 1, 0 );
/* Compute W(i+1:n,i) */
// 1. Send the block reflector A(i+1:n,i) to the GPU
magma_zsetvector( i_n, A(i+1, i), 1, dA(i+1, i), 1 );
//#if (GPUSHMEM < 200)
//magma_zhemv(MagmaLower, i_n, c_one, dA(i+1, i+1), ldda, dA(i+1, i), ione, c_zero,
// dW(i+1, i), ione);
//#else
magmablas_zhemv_work(MagmaLower, i_n, c_one, dA(i+1, i+1), ldda, dA(i+1, i), ione, c_zero,
dW(i+1, i), ione,
dwork, ldwork);
//#endif
// 2. Start putting the result back (asynchronously)
magma_zgetmatrix_async( i_n, 1,
dW(i+1, i), lddw,
W(i+1, i), ldw, stream );
blasf77_zgemv(MagmaConjTransStr, &i_n, &i, &c_one, W(i+1, 0), &ldw,
A(i+1, i), &ione, &c_zero, W(0, i), &ione);
blasf77_zgemv("No transpose", &i_n, &i, &c_neg_one, A(i+1, 0), &lda,
W(0, i), &ione, &c_zero, f, &ione);
blasf77_zgemv(MagmaConjTransStr, &i_n, &i, &c_one, A(i+1, 0), &lda,
A(i+1, i), &ione, &c_zero, W(0, i), &ione);
// 3. Here is where we need it
magma_queue_sync( stream );
if (i != 0)
blasf77_zaxpy(&i_n, &c_one, f, &ione, W(i+1, i), &ione);
blasf77_zgemv("No transpose", &i_n, &i, &c_neg_one, W(i+1, 0), &ldw,
W(0, i), &ione, &c_one, W(i+1, i), &ione);
blasf77_zscal(&i_n, &tau[i], W(i+1,i), &ione);
#if defined(PRECISION_z) || defined(PRECISION_c)
cblas_zdotc_sub( i_n, W(i+1,i), ione, A(i+1,i), ione, &value );
#else
value = cblas_zdotc( i_n, W(i+1,i), ione, A(i+1,i), ione );
#endif
alpha = tau[i] * -0.5f * value;
blasf77_zaxpy(&i_n, &alpha, A(i+1, i), &ione, W(i+1,i), &ione);
}
}
}
magma_free_cpu(f);
magma_queue_destroy( stream );
return 0;
} /* magma_zlatrd */
int main(int argc, char **argv)
{
TESTING_INIT();
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magma_int_t ione = 1;
real_Double_t atomics_perf=0, atomics_time=0;
real_Double_t gflops, magma_perf=0, magma_time=0, cublas_perf, cublas_time, cpu_perf, cpu_time;
double magma_error=0, atomics_error=0, cublas_error, work[1];
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Yatomics, *Ycublas, *Ymagma;
magmaDoubleComplex_ptr dA, dX, dY, dwork;
magma_int_t status = 0;
magma_opts opts;
opts.parse_opts( argc, argv );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf("%% N MAGMA Gflop/s (ms) Atomics Gflop/s CUBLAS Gflop/s CPU Gflop/s MAGMA error Atomics CUBLAS\n");
printf("%%=====================================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = magma_roundup( N, opts.align ); // multiple of 32 by default
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZHEMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Yatomics, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ycublas, 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 );
blocks = magma_ceildiv( N, nb );
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork );
magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
// should not use data from the opposite triangle -- fill with NAN to check
magma_int_t N1 = N-1;
if ( opts.uplo == MagmaUpper ) {
lapackf77_zlaset( "Lower", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[1], &lda );
}
else {
lapackf77_zlaset( "Upper", &N1, &N1, &MAGMA_Z_NAN, &MAGMA_Z_NAN, &A[lda], &lda );
}
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, ldda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
magmablasSetKernelStream( opts.queue ); // opts.handle also uses opts.queue
cublas_time = magma_sync_wtime( opts.queue );
#ifdef HAVE_CUBLAS
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
#else
magma_zhemv( opts.uplo, N, alpha, dA, 0, ldda, dX, 0, incx, beta, dY, 0, incy, opts.queue );
#endif
cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using CUBLAS - using atomics
=================================================================== */
#ifdef HAVE_CUBLAS
cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_ALLOWED );
magma_zsetvector( N, Y, incy, dY, incy );
// sync on queue doesn't work -- need device sync or use NULL stream -- bug in CUBLAS?
atomics_time = magma_sync_wtime( NULL /*opts.queue*/ );
cublasZhemv( opts.handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
atomics_time = magma_sync_wtime( NULL /*opts.queue*/ ) - atomics_time;
atomics_perf = gflops / atomics_time;
magma_zgetvector( N, dY, incy, Yatomics, incy );
cublasSetAtomicsMode( opts.handle, CUBLAS_ATOMICS_NOT_ALLOWED );
#endif
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
#ifdef HAVE_CUBLAS
magma_zsetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( opts.queue );
if ( opts.version == 1 ) {
magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork, opts.queue );
}
else {
// non-work interface (has added overhead)
magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
}
magma_time = magma_sync_wtime( opts.queue ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
#endif
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N;
#ifdef HAVE_CUBLAS
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Yatomics, &incy );
atomics_error = lapackf77_zlange( "M", &N, &ione, Yatomics, &N, work ) / N;
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
#endif
bool okay = (magma_error < tol && cublas_error < tol && atomics_error < tol);
status += ! okay;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %8.2e %s\n",
(int) N,
magma_perf, 1000.*magma_time,
atomics_perf, 1000.*atomics_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error, atomics_error,
(okay ? "ok" : "failed"));
TESTING_FREE_CPU( A );
TESTING_FREE_CPU( X );
TESTING_FREE_CPU( Y );
TESTING_FREE_CPU( Ycublas );
TESTING_FREE_CPU( Yatomics );
TESTING_FREE_CPU( Ymagma );
TESTING_FREE_DEV( dA );
TESTING_FREE_DEV( dX );
TESTING_FREE_DEV( dY );
TESTING_FREE_DEV( dwork );
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 N, lda, ldda, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 1.5, -2.3 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( -0.6, 0.8 );
magmaDoubleComplex *A, *X, *Y, *Ycublas, *Ymagma;
magmaDoubleComplex *dA, *dX, *dY, *dwork;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
if ( opts.uplo == MagmaUpper ) {
printf("** for uplo=MagmaUpper, magmablas_zhemv simply calls cublas_zhemv.\n");
}
printf(" 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 ) {
N = opts.nsize[itest];
lda = N;
ldda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_ZHEMV( N ) / 1e9;
TESTING_MALLOC_CPU( A, magmaDoubleComplex, sizeA );
TESTING_MALLOC_CPU( X, magmaDoubleComplex, sizeX );
TESTING_MALLOC_CPU( Y, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ycublas, magmaDoubleComplex, sizeY );
TESTING_MALLOC_CPU( Ymagma, magmaDoubleComplex, sizeY );
TESTING_MALLOC_DEV( dA, magmaDoubleComplex, ldda*N );
TESTING_MALLOC_DEV( dX, magmaDoubleComplex, sizeX );
TESTING_MALLOC_DEV( dY, magmaDoubleComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = ldda*blocks;
TESTING_MALLOC_DEV( dwork, magmaDoubleComplex, ldwork );
magmablas_zlaset( MagmaFull, ldwork, 1, MAGMA_Z_NAN, MAGMA_Z_NAN, dwork, ldwork );
magmablas_zlaset( MagmaFull, ldda, N, MAGMA_Z_NAN, MAGMA_Z_NAN, dA, ldda );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &sizeA, A );
magma_zmake_hermitian( N, A, lda );
lapackf77_zlarnv( &ione, ISEED, &sizeX, X );
lapackf77_zlarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_zsetmatrix( N, N, A, lda, dA, ldda );
magma_zsetvector( N, X, incx, dX, incx );
magma_zsetvector( N, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasZhemv( handle, cublas_uplo_const(opts.uplo),
N, &alpha, dA, ldda, dX, incx, &beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_zgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMABLAS
=================================================================== */
magma_zsetvector( N, Y, incy, dY, incy );
//magma_zprint_gpu( ldda, blocks, dwork, ldda );
magma_time = magma_sync_wtime( 0 );
magmablas_zhemv_work( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy, dwork, ldwork );
// TODO provide option to test non-work interface
//magmablas_zhemv( opts.uplo, N, alpha, dA, ldda, dX, incx, beta, dY, incy );
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_zgetvector( N, dY, incy, Ymagma, incy );
//magma_zprint_gpu( ldda, blocks, dwork, ldda );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_zhemv( lapack_uplo_const(opts.uplo), &N, &alpha, A, &lda, X, &incx, &beta, Y, &incy );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_zlange( "M", &N, &ione, Ymagma, &N, work ) / N;
blasf77_zaxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_zlange( "M", &N, &ione, Ycublas, &N, work ) / N;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(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 );
TESTING_FREE_DEV( dwork );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}