/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssygvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
real_Double_t mgpu_time;
float *h_A, *h_Ainit, *h_B, *h_Binit, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
#endif
float *w1, result=0;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
magma_range_t range = MagmaRangeAll;
if (opts.fraction != 1)
range = MagmaRangeI;
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: ngpu = %d, itype = %d, jobz = %s, range = %s, uplo = %s, opts.check = %d, fraction = %6.4f\n",
(int) opts.ngpu, (int) opts.itype,
lapack_vec_const(opts.jobz), lapack_range_const(range), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M ngpu MGPU Time (sec)\n");
printf("====================================\n");
magma_int_t threads = magma_get_parallel_numthreads();
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = magma_sbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_sbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
//magma_int_t NB = 96;//magma_bulge_get_nb(N);
//magma_int_t sizvblg = magma_sbulge_get_lq2(N, threads);
//magma_int_t siz = max(sizvblg,n2)+2*(N*NB+N)+24*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC_PIN( h_A, float, n2 );
TESTING_MALLOC_PIN( h_B, float, n2 );
TESTING_MALLOC_PIN( h_work, float, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, float, lrwork);
#endif
TESTING_MALLOC_CPU( w1, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork);
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slarnv( &ione, ISEED, &n2, h_B );
magma_smake_hpd( N, h_B, N );
magma_smake_symmetric( N, h_A, N );
if ( opts.warmup || opts.check ) {
TESTING_MALLOC_CPU( h_Ainit, float, n2 );
TESTING_MALLOC_CPU( h_Binit, float, n2 );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_Ainit, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_Binit, &N );
}
magma_int_t m1 = 0;
float vl = 0;
float vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == MagmaRangeI) {
il = 1;
iu = (int) (opts.fraction*N);
}
if ( opts.warmup ) {
// ==================================================================
// Warmup using MAGMA. I prefer to use smalltest to warmup A-
// ==================================================================
magma_ssygvdx_2stage_m(opts.ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_Ainit, &N, h_A, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_Binit, &N, h_B, &N );
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
mgpu_time = magma_wtime();
magma_ssygvdx_2stage_m(opts.ngpu, opts.itype, opts.jobz, range, opts.uplo,
N, h_A, N, h_B, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
mgpu_time = magma_wtime() - mgpu_time;
if ( opts.check ) {
// ===================================================================
// Check the results following the LAPACK's [zc]hegvdx routine.
// A x = lambda B x is solved
// and the following 3 tests computed:
// (1) | A Z - B Z D | / ( |A||Z| N ) (itype = 1)
// | A B Z - Z D | / ( |A||Z| N ) (itype = 2)
// | B A Z - Z D | / ( |A||Z| N ) (itype = 3)
// ===================================================================
#if defined(PRECISION_d) || defined(PRECISION_s)
float *rwork = h_work + N*N;
#endif
result = 1.;
result /= lapackf77_slansy("1", lapack_uplo_const(opts.uplo), &N, h_Ainit, &N, rwork);
result /= lapackf77_slange("1", &N , &m1, h_A, &N, rwork);
if (opts.itype == 1) {
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N);
result *= lapackf77_slange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_slange("1", &N, &m1, h_A, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_slange("1", &N, &m1, h_A, &N, rwork)/N;
}
}
// ===================================================================
// Print execution time
// ===================================================================
printf("%5d %5d %4d %7.2f\n",
(int) N, (int) m1, (int) opts.ngpu, mgpu_time);
if ( opts.check ) {
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if (opts.itype==1) {
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==2) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
else if (opts.itype==3) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result, (result < tol ? "ok" : "failed") );
}
printf("\n");
status += ! (result < tol);
}
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_B );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( iwork );
if ( opts.warmup || opts.check ) {
TESTING_FREE_CPU( h_Ainit );
TESTING_FREE_CPU( h_Binit );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE_MGPU();
return status;
}
extern "C" magma_int_t
magma_ssygvdx_2stage(magma_int_t itype, char jobz, char range, char uplo, magma_int_t n,
float *a, magma_int_t lda, float *b, magma_int_t ldb,
float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t *m, float *w, float *work, magma_int_t lwork,
magma_int_t *iwork, magma_int_t liwork, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
SSYGVDX_2STAGE computes all the eigenvalues, and optionally, the eigenvectors
of a complex generalized Hermitian-definite eigenproblem, of the form
A*x=(lambda)*B*x, A*Bx=(lambda)*x, or B*A*x=(lambda)*x. Here A and
B are assumed to be Hermitian and B is also positive definite.
It uses a two-stage algorithm for the tridiagonalization.
If eigenvectors are desired, it uses a divide and conquer algorithm.
The divide and conquer algorithm makes very mild assumptions about
floating point arithmetic. It will work on machines with a guard
digit in add/subtract, or on those binary machines without guard
digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or
Cray-2. It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
=========
ITYPE (input) INTEGER
Specifies the problem type to be solved:
= 1: A*x = (lambda)*B*x
= 2: A*B*x = (lambda)*x
= 3: B*A*x = (lambda)*x
RANGE (input) CHARACTER*1
= 'A': all eigenvalues will be found.
= 'V': all eigenvalues in the half-open interval (VL,VU]
will be found.
= 'I': the IL-th through IU-th eigenvalues will be found.
JOBZ (input) CHARACTER*1
= 'N': Compute eigenvalues only;
= 'V': Compute eigenvalues and eigenvectors.
UPLO (input) CHARACTER*1
= 'U': Upper triangles of A and B are stored;
= 'L': Lower triangles of A and B are stored.
N (input) INTEGER
The order of the matrices A and B. N >= 0.
A (input/output) DOUBLE PRECISION array, dimension (LDA, N)
On entry, the Hermitian matrix A. If UPLO = 'U', the
leading N-by-N upper triangular part of A contains the
upper triangular part of the matrix A. If UPLO = 'L',
the leading N-by-N lower triangular part of A contains
the lower triangular part of the matrix A.
On exit, if JOBZ = 'V', then if INFO = 0, A contains the
matrix Z of eigenvectors. The eigenvectors are normalized
as follows:
if ITYPE = 1 or 2, Z**H*B*Z = I;
if ITYPE = 3, Z**H*inv(B)*Z = I.
If JOBZ = 'N', then on exit the upper triangle (if UPLO='U')
or the lower triangle (if UPLO='L') of A, including the
diagonal, is destroyed.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,N).
B (input/output) DOUBLE PRECISION array, dimension (LDB, N)
On entry, the Hermitian matrix B. If UPLO = 'U', the
leading N-by-N upper triangular part of B contains the
upper triangular part of the matrix B. If UPLO = 'L',
the leading N-by-N lower triangular part of B contains
the lower triangular part of the matrix B.
On exit, if INFO <= N, the part of B containing the matrix is
overwritten by the triangular factor U or L from the Cholesky
factorization B = U**H*U or B = L*L**H.
LDB (input) INTEGER
The leading dimension of the array B. LDB >= max(1,N).
VL (input) DOUBLE PRECISION
VU (input) DOUBLE PRECISION
If RANGE='V', the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = 'A' or 'I'.
IL (input) INTEGER
IU (input) INTEGER
If RANGE='I', the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = 'A' or 'V'.
M (output) INTEGER
The total number of eigenvalues found. 0 <= M <= N.
If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
W (output) DOUBLE PRECISION array, dimension (N)
If INFO = 0, the eigenvalues in ascending order.
WORK (workspace/output) DOUBLE PRECISION array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The length of the array WORK.
If N <= 1, LWORK >= 1.
If JOBZ = 'N' and N > 1, LWORK >= LQ2 + N * (NB + 2).
If JOBZ = 'V' and N > 1, LWORK >= LQ2 + 1 + 6*N + 2*N**2.
where LQ2 is the size needed to store
the Q2 matrix and is returned by
MAGMA_BULGE_GET_LQ2.
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal sizes of the WORK, RWORK and
IWORK arrays, returns these values as the first entries of
the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK))
On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
LIWORK (input) INTEGER
The dimension of the array IWORK.
If N <= 1, LIWORK >= 1.
If JOBZ = 'N' and N > 1, LIWORK >= 1.
If JOBZ = 'V' and N > 1, LIWORK >= 3 + 5*N.
If LIWORK = -1, then a workspace query is assumed; the
routine only calculates the optimal sizes of the WORK, RWORK
and IWORK arrays, returns these values as the first entries
of the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
> 0: ZPOTRF or ZHEEVD returned an error code:
<= N: if INFO = i and JOBZ = 'N', then the algorithm
failed to converge; i off-diagonal elements of an
intermediate tridiagonal form did not converge to
zero;
if INFO = i and JOBZ = 'V', then the algorithm
failed to compute an eigenvalue while working on
the submatrix lying in rows and columns INFO/(N+1)
through mod(INFO,N+1);
> N: if INFO = N + i, for 1 <= i <= N, then the leading
minor of order i of B is not positive definite.
The factorization of B could not be completed and
no eigenvalues or eigenvectors were computed.
Further Details
===============
Based on contributions by
Mark Fahey, Department of Mathematics, Univ. of Kentucky, USA
Modified so that no backsubstitution is performed if ZHEEVD fails to
converge (NEIG in old code could be greater than N causing out of
bounds reference to A - reported by Ralf Meyer). Also corrected the
description of INFO and the test on ITYPE. Sven, 16 Feb 05.
===================================================================== */
char uplo_[2] = {uplo, 0};
char jobz_[2] = {jobz, 0};
char range_[2] = {range, 0};
float d_one = MAGMA_S_ONE;
float *da;
float *db;
magma_int_t ldda = n;
magma_int_t lddb = n;
magma_int_t lower;
char trans[1];
magma_int_t wantz;
magma_int_t lquery;
magma_int_t alleig, valeig, indeig;
magma_int_t lwmin;
magma_int_t liwmin;
magma_queue_t stream;
magma_queue_create( &stream );
/* determine the number of threads */
magma_int_t threads = magma_get_numthreads();
magma_setlapack_numthreads(threads);
wantz = lapackf77_lsame(jobz_, MagmaVecStr);
lower = lapackf77_lsame(uplo_, MagmaLowerStr);
alleig = lapackf77_lsame(range_, "A");
valeig = lapackf77_lsame(range_, "V");
indeig = lapackf77_lsame(range_, "I");
lquery = lwork == -1 || liwork == -1;
*info = 0;
if (itype < 1 || itype > 3) {
*info = -1;
} else if (! (alleig || valeig || indeig)) {
*info = -2;
} else if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) {
*info = -3;
} else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) {
*info = -4;
} else if (n < 0) {
*info = -5;
} else if (lda < max(1,n)) {
*info = -7;
} else if (ldb < max(1,n)) {
*info = -9;
} else {
if (valeig) {
if (n > 0 && vu <= vl) {
*info = -11;
}
} else if (indeig) {
if (il < 1 || il > max(1,n)) {
*info = -12;
} else if (iu < min(n,il) || iu > n) {
*info = -13;
}
}
}
magma_int_t nb = magma_get_sbulge_nb(n, threads);
magma_int_t lq2 = magma_sbulge_get_lq2(n, threads);
if (wantz) {
lwmin = lq2 + 1 + 6*n + 2*n*n;
liwmin = 3 + 5*n;
} else {
lwmin = n * (nb + 2);
liwmin = 1;
}
work[0] = lwmin * (1. + lapackf77_slamch("Epsilon"));
iwork[0] = liwmin;
if (lwork < lwmin && ! lquery) {
*info = -17;
} else if (liwork < liwmin && ! lquery) {
*info = -19;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info));
return *info;
} else if (lquery) {
return *info;
}
/* Quick return if possible */
if (n == 0) {
return *info;
}
/* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */
if (n <= 128){
#ifdef ENABLE_DEBUG
printf("--------------------------------------------------------------\n");
printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb);
printf("--------------------------------------------------------------\n");
#endif
lapackf77_ssygvd(&itype, jobz_, uplo_,
&n, a, &lda, b, &ldb,
w, work, &lwork,
iwork, &liwork, info);
*m = n;
return *info;
}
if (MAGMA_SUCCESS != magma_smalloc( &da, n*ldda ) ||
MAGMA_SUCCESS != magma_smalloc( &db, n*lddb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* Form a Cholesky factorization of B. */
magma_ssetmatrix( n, n, b, ldb, db, lddb );
magma_ssetmatrix_async( n, n,
a, lda,
da, ldda, stream );
#ifdef ENABLE_TIMER
magma_timestr_t start, end;
start = get_current_time();
#endif
magma_spotrf_gpu(uplo_[0], n, db, lddb, info);
if (*info != 0) {
*info = n + *info;
return *info;
}
#ifdef ENABLE_TIMER
end = get_current_time();
printf("time spotrf_gpu = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
magma_queue_sync( stream );
magma_sgetmatrix_async( n, n,
db, lddb,
b, ldb, stream );
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
/* Transform problem to standard eigenvalue problem and solve. */
magma_ssygst_gpu(itype, uplo, n, da, ldda, db, lddb, info);
#ifdef ENABLE_TIMER
end = get_current_time();
printf("time ssygst_gpu = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
magma_sgetmatrix( n, n, da, ldda, a, lda );
magma_queue_sync( stream );
magma_free( da );
magma_free( db );
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
magma_ssyevdx_2stage(jobz, range, uplo, n, a, lda, vl, vu, il, iu, m, w, work, lwork, iwork, liwork, info);
#ifdef ENABLE_TIMER
end = get_current_time();
printf("time ssyevdx_2stage = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
if (wantz && *info == 0) {
if (MAGMA_SUCCESS != magma_smalloc( &da, n*ldda ) ||
MAGMA_SUCCESS != magma_smalloc( &db, n*lddb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
magma_ssetmatrix( n, *m, a, lda, da, ldda );
magma_ssetmatrix( n, n, b, ldb, db, lddb );
/* Backtransform eigenvectors to the original problem. */
if (itype == 1 || itype == 2) {
/* For A*x=(lambda)*B*x and A*B*x=(lambda)*x;
backtransform eigenvectors: x = inv(L)'*y or inv(U)*y */
if (lower) {
*(unsigned char *)trans = MagmaConjTrans;
} else {
*(unsigned char *)trans = MagmaNoTrans;
}
magma_strsm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, d_one, db, lddb, da, ldda);
}
else if (itype == 3) {
/* For B*A*x=(lambda)*x;
backtransform eigenvectors: x = L*y or U'*y */
if (lower) {
*(unsigned char *)trans = MagmaNoTrans;
} else {
*(unsigned char *)trans = MagmaConjTrans;
}
magma_strmm(MagmaLeft, uplo, *trans, MagmaNonUnit, n, *m, d_one, db, lddb, da, ldda);
}
magma_sgetmatrix( n, *m, da, ldda, a, lda );
#ifdef ENABLE_TIMER
end = get_current_time();
printf("time strsm/mm + getmatrix = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
magma_free( da );
magma_free( db );
}
magma_queue_destroy( stream );
work[0] = lwmin * (1. + lapackf77_slamch("Epsilon"));
iwork[0] = liwmin;
return *info;
} /* ssygvdx_2stage */
extern "C" magma_int_t
magma_ssyevdx_2stage(char jobz, char range, char uplo,
magma_int_t n,
float *a, magma_int_t lda,
float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t *m, float *w,
float *work, magma_int_t lwork,
magma_int_t *iwork, magma_int_t liwork,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
ZHEEVD_2STAGE computes all eigenvalues and, optionally, eigenvectors of a
complex Hermitian matrix A. It uses a two-stage algorithm for the tridiagonalization.
If eigenvectors are desired, it uses a divide and conquer algorithm.
The divide and conquer algorithm makes very mild assumptions about
floating point arithmetic. It will work on machines with a guard
digit in add/subtract, or on those binary machines without guard
digits which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or
Cray-2. It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
=========
JOBZ (input) CHARACTER*1
= 'N': Compute eigenvalues only;
= 'V': Compute eigenvalues and eigenvectors.
RANGE (input) CHARACTER*1
= 'A': all eigenvalues will be found.
= 'V': all eigenvalues in the half-open interval (VL,VU]
will be found.
= 'I': the IL-th through IU-th eigenvalues will be found.
UPLO (input) CHARACTER*1
= 'U': Upper triangle of A is stored;
= 'L': Lower triangle of A is stored.
N (input) INTEGER
The order of the matrix A. N >= 0.
A (input/output) COMPLEX_16 array, dimension (LDA, N)
On entry, the Hermitian matrix A. If UPLO = 'U', the
leading N-by-N upper triangular part of A contains the
upper triangular part of the matrix A. If UPLO = 'L',
the leading N-by-N lower triangular part of A contains
the lower triangular part of the matrix A.
On exit, if JOBZ = 'V', then if INFO = 0, the first m columns
of A contains the required
orthonormal eigenvectors of the matrix A.
If JOBZ = 'N', then on exit the lower triangle (if UPLO='L')
or the upper triangle (if UPLO='U') of A, including the
diagonal, is destroyed.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,N).
VL (input) REAL
VU (input) REAL
If RANGE='V', the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = 'A' or 'I'.
IL (input) INTEGER
IU (input) INTEGER
If RANGE='I', the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = 'A' or 'V'.
M (output) INTEGER
The total number of eigenvalues found. 0 <= M <= N.
If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
W (output) REAL array, dimension (N)
If INFO = 0, the required m eigenvalues in ascending order.
WORK (workspace/output) COMPLEX_16 array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The length of the array WORK.
If N <= 1, LWORK >= 1.
If JOBZ = 'N' and N > 1, LWORK >= LQ2 + N * (NB + 2).
If JOBZ = 'V' and N > 1, LWORK >= LQ2 + 1 + 6*N + 2*N**2.
where LQ2 is the size needed to store
the Q2 matrix and is returned by
MAGMA_BULGE_GET_LQ2.
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal sizes of the WORK, RWORK and
IWORK arrays, returns these values as the first entries of
the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK))
On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
LIWORK (input) INTEGER
The dimension of the array IWORK.
If N <= 1, LIWORK >= 1.
If JOBZ = 'N' and N > 1, LIWORK >= 1.
If JOBZ = 'V' and N > 1, LIWORK >= 3 + 5*N.
If LIWORK = -1, then a workspace query is assumed; the
routine only calculates the optimal sizes of the WORK, RWORK
and IWORK arrays, returns these values as the first entries
of the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
> 0: if INFO = i and JOBZ = 'N', then the algorithm failed
to converge; i off-diagonal elements of an intermediate
tridiagonal form did not converge to zero;
if INFO = i and JOBZ = 'V', then the algorithm failed
to compute an eigenvalue while working on the submatrix
lying in rows and columns INFO/(N+1) through
mod(INFO,N+1).
Further Details
===============
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified description of INFO. Sven, 16 Feb 05.
===================================================================== */
char uplo_[2] = {uplo, 0};
char jobz_[2] = {jobz, 0};
char range_[2] = {range, 0};
float d_one = 1.;
magma_int_t ione = 1;
magma_int_t izero = 0;
float d__1;
float eps;
float anrm;
magma_int_t imax;
float rmin, rmax;
float sigma;
magma_int_t lwmin, liwmin;
magma_int_t lower;
magma_int_t wantz;
magma_int_t iscale;
float safmin;
float bignum;
float smlnum;
magma_int_t lquery;
magma_int_t alleig, valeig, indeig;
float* dwork;
/* determine the number of threads */
magma_int_t threads = magma_get_numthreads();
magma_setlapack_numthreads(threads);
wantz = lapackf77_lsame(jobz_, MagmaVecStr);
lower = lapackf77_lsame(uplo_, MagmaLowerStr);
alleig = lapackf77_lsame( range_, "A" );
valeig = lapackf77_lsame( range_, "V" );
indeig = lapackf77_lsame( range_, "I" );
lquery = lwork == -1 || liwork == -1;
*info = 0;
if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) {
*info = -1;
} else if (! (alleig || valeig || indeig)) {
*info = -2;
} else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) {
*info = -3;
} else if (n < 0) {
*info = -4;
} else if (lda < max(1,n)) {
*info = -6;
} else {
if (valeig) {
if (n > 0 && vu <= vl) {
*info = -8;
}
} else if (indeig) {
if (il < 1 || il > max(1,n)) {
*info = -9;
} else if (iu < min(n,il) || iu > n) {
*info = -10;
}
}
}
magma_int_t nb = magma_get_sbulge_nb(n, threads);
magma_int_t Vblksiz = magma_sbulge_get_Vblksiz(n, nb, threads);
magma_int_t ldt = Vblksiz;
magma_int_t ldv = nb + Vblksiz;
magma_int_t blkcnt = magma_bulge_get_blkcnt(n, nb, Vblksiz);
magma_int_t lq2 = magma_sbulge_get_lq2(n, threads);
if (wantz) {
lwmin = lq2 + 1 + 6 * n + 2 * n * n;
liwmin = 5 * n + 3;
} else {
lwmin = lq2 + n * (nb + 1);
liwmin = 1;
}
work[0] = lwmin * (1. + lapackf77_slamch("Epsilon"));
iwork[0] = liwmin;
if ((lwork < lwmin) && !lquery) {
*info = -14;
} else if ((liwork < liwmin) && ! lquery) {
*info = -16;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (n == 0) {
return *info;
}
if (n == 1) {
w[0] = a[0];
if (wantz) {
a[0] = MAGMA_S_ONE;
}
return *info;
}
#ifdef ENABLE_TIMER
printf("using %d threads\n", threads);
#endif
/* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */
magma_int_t ntiles = n/nb;
if( ( ntiles < 2 ) || ( n <= 128 ) ){
#ifdef ENABLE_DEBUG
printf("--------------------------------------------------------------\n");
printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int ) nb);
printf("--------------------------------------------------------------\n");
#endif
lapackf77_ssyevd(jobz_, uplo_, &n,
a, &lda, w,
work, &lwork,
iwork, &liwork,
info);
*m = n;
return *info;
}
/* Get machine constants. */
safmin = lapackf77_slamch("Safe minimum");
eps = lapackf77_slamch("Precision");
smlnum = safmin / eps;
bignum = 1. / smlnum;
rmin = magma_ssqrt(smlnum);
rmax = magma_ssqrt(bignum);
/* Scale matrix to allowable range, if necessary. */
anrm = lapackf77_slansy("M", uplo_, &n, a, &lda, work);
iscale = 0;
if (anrm > 0. && anrm < rmin) {
iscale = 1;
sigma = rmin / anrm;
} else if (anrm > rmax) {
iscale = 1;
sigma = rmax / anrm;
}
if (iscale == 1) {
lapackf77_slascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, a,
&lda, info);
}
magma_int_t inde = 0;
magma_int_t indT2 = inde + n;
magma_int_t indTAU2 = indT2 + blkcnt*ldt*Vblksiz;
magma_int_t indV2 = indTAU2+ blkcnt*Vblksiz;
magma_int_t indtau1 = indV2 + blkcnt*ldv*Vblksiz;
magma_int_t indwrk = indtau1+ n;
magma_int_t indwk2 = indwrk + n * n;
magma_int_t llwork = lwork - indwrk;
magma_int_t llwrk2 = lwork - indwk2;
#ifdef ENABLE_TIMER
magma_timestr_t start, st1, st2, end;
start = get_current_time();
#endif
float *dT1;
if (MAGMA_SUCCESS != magma_smalloc( &dT1, n*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_ssytrd_sy2sb(uplo, n, nb, a, lda, &work[indtau1], &work[indwrk], llwork, dT1, threads, info);
#ifdef ENABLE_TIMER
st1 = get_current_time();
printf(" time ssytrd_sy2sb = %6.2f\n" , GetTimerValue(start,st1)/1000.);
#endif
/* copy the input matrix into WORK(INDWRK) with band storage */
/* PAY ATTENTION THAT work[indwrk] should be able to be of size lda2*n which it should be checked in any future modification of lwork.*/
magma_int_t lda2 = 2*nb; //nb+1+(nb-1);
float* A2 = &work[indwrk];
memset(A2 , 0, n*lda2*sizeof(float));
for (magma_int_t j = 0; j < n-nb; j++)
{
cblas_scopy(nb+1, &a[j*(lda+1)], 1, &A2[j*lda2], 1);
memset(&a[j*(lda+1)], 0, (nb+1)*sizeof(float));
a[nb + j*(lda+1)] = d_one;
}
for (magma_int_t j = 0; j < nb; j++)
{
cblas_scopy(nb-j, &a[(j+n-nb)*(lda+1)], 1, &A2[(j+n-nb)*lda2], 1);
memset(&a[(j+n-nb)*(lda+1)], 0, (nb-j)*sizeof(float));
}
#ifdef ENABLE_TIMER
st2 = get_current_time();
printf(" time ssytrd_convert = %6.2f\n" , GetTimerValue(st1,st2)/1000.);
#endif
magma_ssytrd_sb2st(threads, uplo, n, nb, Vblksiz, A2, lda2, w, &work[inde], &work[indV2], ldv, &work[indTAU2], wantz, &work[indT2], ldt);
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time ssytrd_sy2st = %6.2f\n" , GetTimerValue(st2,end)/1000.);
printf(" time ssytrd = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
/* For eigenvalues only, call SSTERF. For eigenvectors, first call
ZSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the
tridiagonal matrix, then call ZUNMTR to multiply it to the Householder
transformations represented as Householder vectors in A. */
if (! wantz) {
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
lapackf77_ssterf(&n, w, &work[inde], info);
magma_smove_eig(range, n, w, &il, &iu, vl, vu, m);
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time sstedc = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
} else {
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
if (MAGMA_SUCCESS != magma_smalloc( &dwork, 3*n*(n/2 + 1) )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_sstedx(range, n, vl, vu, il, iu, w, &work[inde],
&work[indwrk], n, &work[indwk2],
llwrk2, iwork, liwork, dwork, info);
magma_free( dwork );
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time sstedx = %6.2f\n", GetTimerValue(start,end)/1000.);
start = get_current_time();
#endif
float *dZ;
magma_int_t lddz = n;
float *da;
magma_int_t ldda = n;
magma_smove_eig(range, n, w, &il, &iu, vl, vu, m);
if (MAGMA_SUCCESS != magma_smalloc( &dZ, *m*lddz)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_smalloc( &da, n*ldda )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_sbulge_back(threads, uplo, n, nb, *m, Vblksiz, &work[indwrk + n * (il-1)], n, dZ, lddz,
&work[indV2], ldv, &work[indTAU2], &work[indT2], ldt, info);
#ifdef ENABLE_TIMER
st1 = get_current_time();
printf(" time sbulge_back = %6.2f\n" , GetTimerValue(start,st1)/1000.);
#endif
magma_ssetmatrix( n, n, a, lda, da, ldda );
magma_sormqr_gpu_2stages(MagmaLeft, MagmaNoTrans, n-nb, *m, n-nb, da+nb, ldda,
dZ+nb, n, dT1, nb, info);
magma_sgetmatrix( n, *m, dZ, lddz, a, lda );
magma_free(dT1);
magma_free(dZ);
magma_free(da);
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time sormqr + copy = %6.2f\n", GetTimerValue(st1,end)/1000.);
printf(" time eigenvectors backtransf. = %6.2f\n" , GetTimerValue(start,end)/1000.);
#endif
}
/* If matrix was scaled, then rescale eigenvalues appropriately. */
if (iscale == 1) {
if (*info == 0) {
imax = n;
} else {
imax = *info - 1;
}
d__1 = 1. / sigma;
blasf77_sscal(&imax, &d__1, w, &ione);
}
work[0] = lwmin * (1. + lapackf77_slamch("Epsilon"));
iwork[0] = liwmin;
return *info;
} /* magma_zheevdx_2stage */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssygvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time;
float *h_A, *h_R, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
#endif
/* Matrix size */
float *w1, *w2;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};;
magma_int_t info_ortho = 0;
magma_int_t info_solution = 0;
magma_int_t info_reduction = 0;
magma_timestr_t start, end;
magma_opts opts;
parse_opts( argc, argv, &opts );
magma_int_t ngpu = opts.ngpu;
char jobz = opts.jobz;
magma_int_t checkres = opts.check;
char range = 'A';
char uplo = opts.uplo;
magma_int_t itype = opts.itype;
float f = opts.fraction;
if (f != 1)
range='I';
if ( checkres && jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %c, range = %c, uplo = %c, checkres = %d, fraction = %6.4f\n",
(int) itype, jobz, range, uplo, (int) checkres, f);
printf(" N M GPU Time(s) \n");
printf("==========================\n");
magma_int_t threads = magma_get_numthreads();
for( magma_int_t i = 0; i < opts.ntest; ++i ) {
for( magma_int_t iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = magma_sbulge_get_lq2(N, threads) + 2*N + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = magma_sbulge_get_lq2(N, threads) + 1 + 6*N + 2*N*N;
#endif
liwork = 3 + 5*N;
/* Allocate host memory for the matrix */
TESTING_MALLOC( h_A, float, n2);
TESTING_MALLOC( w1, float , N);
TESTING_MALLOC( w2, float , N);
TESTING_HOSTALLOC(h_R, float, n2);
TESTING_HOSTALLOC(h_work, float, lwork);
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_HOSTALLOC( rwork, float, lrwork);
#endif
TESTING_MALLOC( iwork, magma_int_t, liwork);
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
/* Make diagonal real */
for(int i=0; i<N; i++) {
MAGMA_S_SET2REAL( h_A[i*N+i], MAGMA_S_REAL(h_A[i*N+i]) );
}
magma_int_t m1 = 0;
float vl = 0;
float vu = 0;
magma_int_t il = 0;
magma_int_t iu = 0;
if (range == 'I'){
il = 1;
iu = (int) (f*N);
}
if(opts.warmup){
// ==================================================================
// Warmup using MAGMA
// ==================================================================
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
if(ngpu==1){
printf("calling ssyevdx_2stage 1 GPU\n");
magma_ssyevdx_2stage(jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}else{
printf("calling ssyevdx_2stage_m %d GPU\n", (int) ngpu);
magma_ssyevdx_2stage_m(ngpu, jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
}
// ===================================================================
// Performs operation using MAGMA
// ===================================================================
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
start = get_current_time();
if(ngpu==1){
printf("calling ssyevdx_2stage 1 GPU\n");
magma_ssyevdx_2stage(jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}else{
printf("calling ssyevdx_2stage_m %d GPU\n", (int) ngpu);
magma_ssyevdx_2stage_m(ngpu, jobz, range, uplo, N,
h_R, N,
vl, vu, il, iu,
&m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
}
end = get_current_time();
gpu_time = GetTimerValue(start,end)/1000.;
if ( checkres ) {
float eps = lapackf77_slamch("E");
printf("\n");
printf("------ TESTS FOR MAGMA SSYEVD ROUTINE ------- \n");
printf(" Size of the Matrix %d by %d\n", (int) N, (int) N);
printf("\n");
printf(" The matrix A is randomly generated for each test.\n");
printf("============\n");
printf(" The relative machine precision (eps) is %8.2e\n",eps);
printf(" Computational tests pass if scaled residuals are less than 60.\n");
/* Check the orthogonality, reduction and the eigen solutions */
if (jobz == MagmaVec) {
info_ortho = check_orthogonality(N, N, h_R, N, eps);
info_reduction = check_reduction(uplo, N, 1, h_A, w1, N, h_R, eps);
}
printf("------ CALLING LAPACK SSYEVD TO COMPUTE only eigenvalue and verify elementswise ------- \n");
lapackf77_ssyevd("N", "L", &N,
h_A, &N, w2,
h_work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
info_solution = check_solution(N, w2, w1, eps);
if ( (info_solution == 0) & (info_ortho == 0) & (info_reduction == 0) ) {
printf("***************************************************\n");
printf(" ---- TESTING SSYEVD ...................... PASSED !\n");
printf("***************************************************\n");
}
else {
printf("************************************************\n");
printf(" - TESTING SSYEVD ... FAILED !\n");
printf("************************************************\n");
}
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %6.2f\n",
(int) N, (int) m1, gpu_time);
TESTING_FREE( h_A);
TESTING_FREE( w1);
TESTING_FREE( w2);
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_HOSTFREE( rwork);
#endif
TESTING_FREE( iwork);
TESTING_HOSTFREE(h_work);
TESTING_HOSTFREE( h_R);
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE();
return 0;
}
