inline static void
magma_clarfxsym_v2(
magma_int_t n,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *V, magmaFloatComplex *TAU,
magmaFloatComplex *work)
{
/*
WORK (workspace) float complex array, dimension N
*/
magma_int_t ione = 1;
magmaFloatComplex dtmp;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_neg_one= MAGMA_C_NEG_ONE;
magmaFloatComplex c_half = MAGMA_C_HALF;
/* X = AVtau */
blasf77_chemv("L",&n, TAU, A, &lda, V, &ione, &c_zero, work, &ione);
/* compute dtmp= X'*V */
dtmp = magma_cblas_cdotc(n, work, ione, V, ione);
/* compute 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * c_half * (*TAU);
/* compute W=X-1/2VX'Vt = X - dtmp*V */
blasf77_caxpy(&n, &dtmp, V, &ione, work, &ione);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_cher2("L", &n, &c_neg_one, work, &ione, V, &ione, A, &lda);
}
extern "C" void
magma_clarfxsym(
magma_int_t N,
magmaFloatComplex *A, magma_int_t LDA,
magmaFloatComplex *V, magmaFloatComplex *TAU)
{
magma_int_t IONE=1;
magmaFloatComplex dtmp;
magmaFloatComplex Z_ZERO = MAGMA_C_ZERO;
//magmaFloatComplex Z_ONE = MAGMA_C_ONE;
magmaFloatComplex Z_MONE = MAGMA_C_NEG_ONE;
magmaFloatComplex Z_HALF = MAGMA_C_HALF;
//magmaFloatComplex WORK[N];
magmaFloatComplex *WORK;
magma_cmalloc_cpu( &WORK, N );
/* apply left and right on A(st:ed,st:ed)*/
//magma_clarfxsym(len,A(st,st),LDX,V(st),TAU(st));
/* X = AVtau */
blasf77_chemv("L",&N, TAU, A, &LDA, V, &IONE, &Z_ZERO, WORK, &IONE);
/* je calcul dtmp= X'*V */
dtmp = magma_cblas_cdotc(N, WORK, IONE, V, IONE);
/* je calcul 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * Z_HALF * (*TAU);
/* je calcul W=X-1/2VX'Vt = X - dtmp*V */
/*
for (j = 0; j < N; j++)
WORK[j] = WORK[j] + (dtmp*V[j]); */
blasf77_caxpy(&N, &dtmp, V, &IONE, WORK, &IONE);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_cher2("L",&N,&Z_MONE,WORK,&IONE,V,&IONE,A,&LDA);
magma_free_cpu(WORK);
}
inline static void
magma_clarfxsym_v2(magma_int_t n,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *V, magmaFloatComplex *TAU,
magmaFloatComplex *work)
{
/*
WORK (workspace) float complex array, dimension N
*/
magma_int_t ione = 1;
magmaFloatComplex dtmp;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_neg_one= MAGMA_C_NEG_ONE;
magmaFloatComplex c_half = MAGMA_C_HALF;
/* X = AVtau */
blasf77_chemv("L",&n, TAU, A, &lda, V, &ione, &c_zero, work, &ione);
/* compute dtmp= X'*V */
#if defined(PRECISION_z) || defined(PRECISION_c)
dtmp = c_zero;
for (magma_int_t j = 0; j < n; j++)
dtmp = dtmp + MAGMA_C_CNJG(work[j]) * V[j];
//cblas_cdotc_sub(n, work, ione, V, ione, &dtmp);
#else
dtmp = cblas_cdotc(n, work, ione, V, ione);
#endif
/* compute 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * c_half * (*TAU);
/* compute W=X-1/2VX'Vt = X - dtmp*V */
blasf77_caxpy(&n, &dtmp, V, &ione, work, &ione);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_cher2("L", &n, &c_neg_one, work, &ione, V, &ione, A, &lda);
}
int main(int argc, char **argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, cublas_perf, cublas_time, cpu_perf, cpu_time;
float 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, sizeA, sizeX, sizeY, blocks, ldwork;
magma_int_t incx = 1;
magma_int_t incy = 1;
magma_int_t nb = 64;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex alpha = MAGMA_C_MAKE( 1.5, -2.3 );
magmaFloatComplex beta = MAGMA_C_MAKE( -0.6, 0.8 );
magmaFloatComplex *A, *X, *Y, *Ycublas, *Ymagma;
magmaFloatComplex *dA, *dX, *dY, *dC_work;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf(" N 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 ) {
N = opts.nsize[i];
lda = ((N + 31)/32)*32;
sizeA = N*lda;
sizeX = N*incx;
sizeY = N*incy;
gflops = FLOPS_CHEMV( N ) / 1e9;
TESTING_MALLOC( A, magmaFloatComplex, sizeA );
TESTING_MALLOC( X, magmaFloatComplex, sizeX );
TESTING_MALLOC( Y, magmaFloatComplex, sizeY );
TESTING_MALLOC( Ycublas, magmaFloatComplex, sizeY );
TESTING_MALLOC( Ymagma, magmaFloatComplex, sizeY );
TESTING_DEVALLOC( dA, magmaFloatComplex, sizeA );
TESTING_DEVALLOC( dX, magmaFloatComplex, sizeX );
TESTING_DEVALLOC( dY, magmaFloatComplex, sizeY );
blocks = (N + nb - 1) / nb;
ldwork = lda * (blocks + 1);
TESTING_DEVALLOC( dC_work, magmaFloatComplex, ldwork );
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &sizeA, A );
magma_cmake_hermitian( N, A, lda );
lapackf77_clarnv( &ione, ISEED, &sizeX, X );
lapackf77_clarnv( &ione, ISEED, &sizeY, Y );
/* =====================================================================
Performs operation using CUBLAS
=================================================================== */
magma_csetmatrix( N, N, A, lda, dA, lda );
magma_csetvector( N, X, incx, dX, incx );
magma_csetvector( N, Y, incy, dY, incy );
cublas_time = magma_sync_wtime( 0 );
cublasChemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy );
cublas_time = magma_sync_wtime( 0 ) - cublas_time;
cublas_perf = gflops / cublas_time;
magma_cgetvector( N, dY, incy, Ycublas, incy );
/* =====================================================================
Performs operation using MAGMA BLAS
=================================================================== */
magma_csetvector( N, Y, incy, dY, incy );
magma_time = magma_sync_wtime( 0 );
#if (GPUSHMEM >= 200)
magmablas_chemv2( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy, dC_work, ldwork );
#else
magmablas_chemv( opts.uplo, N, alpha, dA, lda, dX, incx, beta, dY, incy );
#endif
magma_time = magma_sync_wtime( 0 ) - magma_time;
magma_perf = gflops / magma_time;
magma_cgetvector( N, dY, incy, Ymagma, incy );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
cpu_time = magma_wtime();
blasf77_chemv( &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_caxpy( &N, &c_neg_one, Y, &incy, Ymagma, &incy );
magma_error = lapackf77_clange( "M", &N, &ione, Ymagma, &N, work ) / N;
blasf77_caxpy( &N, &c_neg_one, Y, &incy, Ycublas, &incy );
cublas_error = lapackf77_clange( "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\n",
(int) N,
magma_perf, 1000.*magma_time,
cublas_perf, 1000.*cublas_time,
cpu_perf, 1000.*cpu_time,
magma_error, cublas_error );
TESTING_FREE( A );
TESTING_FREE( X );
TESTING_FREE( Y );
TESTING_FREE( Ycublas );
TESTING_FREE( Ymagma );
TESTING_DEVFREE( dA );
TESTING_DEVFREE( dX );
TESTING_DEVFREE( dY );
TESTING_DEVFREE( dC_work );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
