void setauxvar(bodyptr btab, int nbody)
{
bodyptr bp;
vector jvec;
real jtot, etot, r0, r1, r;
if (streq(getparam("auxvar"), "mass"))
for (bp = btab; bp < NthBody(btab, nbody); bp = NextBody(bp))
Aux(bp) = mass_gsp(ggsp, absv(Pos(bp)));
else if (streq(getparam("auxvar"), "rperi"))
for (bp = btab; bp < NthBody(btab, nbody); bp = NextBody(bp)) {
CROSSVP(jvec, Pos(bp), Vel(bp));
jtot = absv(jvec);
etot = 0.5 * dotvp(Vel(bp), Vel(bp)) + Phi(bp);
r0 = 0.0;
r1 = absv(Pos(bp));
r = 0.5 * (r0 + r1);
while ((r1 - r0) > TOL * r) {
if (rsqrt(2 * (etot - phi_gsp(ggsp, r))) > jtot/r)
r1 = r;
else
r0 = r;
r = 0.5 * (r0 + r1);
}
Aux(bp) = r;
}
else
error("%s: unknown auxvar option %s\n",
getargv0(), getparam("auxvar"));
}
// -------- begin of function Music::play ---------//
// <int> songId
// <int> playType 0 = non-looped, 1 = looped
int Music::play(int songId, int playType)
{
if( !init_flag )
return 0;
stop();
if( audio.wav_init_flag )
{
String waveFileStr(DIR_MUSIC);
waveFileStr += music_file[songId-1];
if( !DIR_MUSIC[0] || !m.is_file_exist(waveFileStr) || !audio.wav_init_flag )
return 0;
if( playType & MUSIC_PLAY_LOOPED )
{
AbsVolume absv(config.wav_music_volume,0);
music_channel = audio.play_loop_wav(waveFileStr, absv );
}
else
{
AbsVolume absv(config.wav_music_volume,0);
music_channel = audio.play_long_wav(waveFileStr, absv );
}
play_type = playType;
song_id = songId;
return music_channel >= 0;
}
return 0;
}
int radrank(const void *a, const void *b)
{
real Ra, Rb;
Ra = absv(Pos((bodyptr) a));
Rb = absv(Pos((bodyptr) b));
return (Ra < Rb ? -1 : Ra > Rb ? 1 : 0);
}
int len4(a, b, c, d)
{
absv(a); absv(b); /* get the absolute values */
absv(c); absv(d); /* (component magnitudes) */
inorder(a, b); inorder(c, d); /* everyone has a chance to play */
inorder(a, c); inorder(b, d); /* (a,d) are big (winner, loser) */
inorder(b, c); /* playoff for 2nd and 3rd slots */
a += (25*b + 19*c + 16*d)/60; /* compute 4D approximate length */
/* a += (5*b + 4*c + 3*d)/12; .. only .1% worse; easy to eval */
a++; /* Roundoff -> underestimation */
return(a); /* omit the above one bit jitter */
}
// -------- begin of function Music::play ---------//
// <int> songId
// <int> playType 0 = non-looped, 1 = looped
int Music::play(int songId, int playType)
{
if( !init_flag )
return 0;
stop();
#ifdef BUNDLE
// disable CD music
playType &= ~MUSIC_CD_THEN_WAV & ~MUSIC_PLAY_CD;
#endif
if( playType & MUSIC_CD_THEN_WAV )
{
return play(songId, playType & ~MUSIC_CD_THEN_WAV | MUSIC_PLAY_CD)
|| play(songId, playType & ~MUSIC_CD_THEN_WAV & ~MUSIC_PLAY_CD);
}
else if( playType & MUSIC_PLAY_CD )
{
if( audio.cd_init_flag && audio.play_cd(songId +1, config.cd_music_volume) ) // skip the first data track
{
play_type = playType;
song_id = songId;
music_channel = 1;
return 1;
}
return 0;
}
else
{
if( audio.wav_init_flag )
{
String waveFileStr(DIR_MUSIC);
waveFileStr += music_file[songId-1];
if( !DIR_MUSIC[0] || !misc.is_file_exist(waveFileStr) || !audio.wav_init_flag )
return 0;
if( playType & MUSIC_PLAY_LOOPED )
{
AbsVolume absv(config.wav_music_volume,0);
music_channel = audio.play_loop_wav(waveFileStr, 0, absv );
}
else
{
AbsVolume absv(config.wav_music_volume,0);
music_channel = audio.play_long_wav(waveFileStr, absv );
}
play_type = playType;
song_id = songId;
return music_channel != 0;
}
return 0;
}
}
int
multimut (int oribase)
{
int mutbase;
double _uni;
_uni = ran1 (idum);
if (oribase == 3)
_uni *= mutmatrix[3][2];
else
_uni *= mutmatrix[oribase][3];
for (mutbase = 0; mutbase < 4; mutbase++)
{
//printf ("(%f,%f)\t",_uni,mutmatrix[oribase][mutbase]);
if (_uni <= mutmatrix[oribase][mutbase])
{
if ((absv (mutbase - oribase)) == 2)
numutstrans++;
else
numutstranv++;
return (mutbase);
}
}
printf ("error in multimut\n");
myassert (0);
exit (-1);
}
void polymodel(void)
{
gsl_interp_accel *pmsplacc = gsl_interp_accel_alloc();
bodyptr p;
real rad, phi, vel, psi, vr, vp, a, E, J;
vector rhat, vtmp, vper;
for (p = btab; p < NthBody(btab, nbody); p = NextBody(p)) {
rad = rad_m(xrandom(0.0, mtot));
phi = gsl_spline_eval(pmspline, (double) rad, pmsplacc);
vel = pick_v(phi);
psi = pick_psi();
vr = vel * rcos(psi);
vp = vel * rsin(psi);
Mass(p) = mtot / nbody;
pickshell(rhat, NDIM, 1.0);
MULVS(Pos(p), rhat, rad);
pickshell(vtmp, NDIM, 1.0);
a = dotvp(vtmp, rhat);
MULVS(vper, rhat, - a);
ADDV(vper, vper, vtmp);
a = absv(vper);
MULVS(vper, vper, vp / a);
MULVS(Vel(p), rhat, vr);
ADDV(Vel(p), Vel(p), vper);
Phi(p) = phi;
E = phi + 0.5 * rsqr(vel);
J = rad * ABS(vp);
Aux(p) = Kprime * rpow(phi1 - E, npol - 1.5) * rpow(J, 2 * mpol);
}
gsl_interp_accel_free(pmsplacc);
}
int main(int argc, string argv[])
{
stream fstr, istr, ostr;
gsprof *gsp;
bodyptr btab = NULL, p;
int nbody;
real tnow, r;
string intags[MaxBodyFields];
initparam(argv, defv);
layout_body(bodyfields, Precision, NDIM);
fstr = stropen(getparam("gsp"), "r");
get_history(fstr);
gsp = get_gsprof(fstr);
istr = stropen(getparam("in"), "r");
get_history(istr);
if (! get_snap(istr, &btab, &nbody, &tnow, intags, TRUE))
error("%s: snapshot input failed\n", getargv0());
if (! set_member(intags, PosTag))
error("%s: position data missing\n", getargv0());
if (streq(getparam("option"), "rho"))
for (p = btab; p < NthBody(btab, nbody); p = NextBody(p))
Aux(p) = rho_gsp(gsp, absv(Pos(p)));
else if (streq(getparam("option"), "drho"))
for (p = btab; p < NthBody(btab, nbody); p = NextBody(p))
Aux(p) = drho_gsp(gsp, absv(Pos(p)));
else if (streq(getparam("option"), "mass"))
for (p = btab; p < NthBody(btab, nbody); p = NextBody(p))
Aux(p) = mass_gsp(gsp, absv(Pos(p)));
else if (streq(getparam("option"), "phi"))
for (p = btab; p < NthBody(btab, nbody); p = NextBody(p))
Aux(p) = phi_gsp(gsp, absv(Pos(p)));
else
error("%s: unknown option %s\n", getargv0(), getparam("option"));
if (! strnull(getparam("out"))) {
ostr = stropen(getparam("out"), "w");
put_history(ostr);
put_snap(ostr, &btab, &nbody, &tnow, set_union(bodyfields, intags));
strclose(ostr);
}
return (0);
}
// -------- begin of function Music::change_volume --------//
void Music::change_volume(int vol)
{
if( !init_flag )
return;
if( is_playing() )
{
AbsVolume absv(vol,0);
audio.volume_long_wav(music_channel, DsVolume(absv));
}
}
void picktriad(vector x, vector y, vector z)
{
real a;
pickshell(x, NDIM, 1.0);
pickshell(z, NDIM, 1.0);
CROSSVP(y, x, z);
a = absv(y);
DIVVS(y, y, a);
CROSSVP(z, x, y);
}
local void gspforces(bodyptr btab, int nbody, gsprof *gravgsp)
{
bodyptr bp;
real r, mr3i;
for (bp = btab; bp < NthBody(btab, nbody); bp = NextBody(bp)) {
r = absv(Pos(bp));
Phi(bp) = phi_gsp(gravgsp, r);
mr3i = mass_gsp(gravgsp, r) / rqbe(r);
MULVS(Acc(bp), Pos(bp), -mr3i);
}
}
local void hqmforces(bodyptr btab, int nbody, real M, real a, real b,
real tol)
{
bodyptr bp;
double r, mr3i, params[4], phi0, aR0, az0, abserr[3];
static gsl_integration_workspace *wksp = NULL;
gsl_function FPhi, F_aR, F_az;
static double maxerr = 0.0;
int stat[3];
if (a == b) { // spherical case is easy!
for (bp = btab; bp < NthBody(btab, nbody); bp = NextBody(bp)) {
r = absv(Pos(bp));
Phi(bp) = - M / (a + r);
mr3i = M * rsqr(r / (a + r)) / rqbe(r);
MULVS(Acc(bp), Pos(bp), - mr3i);
}
} else { // flattened case is harder
if (wksp == NULL) { // on first call, initialze
wksp = gsl_integration_workspace_alloc(1000);
gsl_set_error_handler_off(); // handle errors below
}
FPhi.function = &intPhi;
F_aR.function = &int_aR;
F_az.function = &int_az;
FPhi.params = F_aR.params = F_az.params = params;
a2(params) = rsqr(a);
b2(params) = rsqr(b);
for (bp = btab; bp < NthBody(btab, nbody); bp = NextBody(bp)) {
R(params) = rsqrt(rsqr(Pos(bp)[0]) + rsqr(Pos(bp)[1]));
z(params) = Pos(bp)[2];
stat[0] = gsl_integration_qagiu(&FPhi, 0.0, tol, 0.0,
1000, wksp, &phi0, &abserr[0]);
stat[1] = gsl_integration_qagiu(&F_aR, 0.0, tol, 0.0,
1000, wksp, &aR0, &abserr[1]);
stat[2] = gsl_integration_qagiu(&F_az, 0.0, tol, 0.0,
1000, wksp, &az0, &abserr[2]);
if (stat[0] || stat[1] || stat[2]) // any errors reported?
for (int i = 0; i < 3; i++)
if (stat[i] != 0 && abserr[i] > maxerr) {
eprintf("[%s.hqmforces: warning: %s abserr[%d] = %g]\n",
getprog(), gsl_strerror(stat[i]), i+1, abserr[i]);
maxerr = abserr[i]; // adjust reporting threshold
}
Phi(bp) = - M * phi0;
Acc(bp)[0] = - M * (Pos(bp)[0] / R(params)) * aR0;
Acc(bp)[1] = - M * (Pos(bp)[1] / R(params)) * aR0;
Acc(bp)[2] = - M * az0;
}
}
}
static int crossn (double *a, double *b, double *c)
{
/* Local variables */
register double *d1;
register double x;
/* Function Body */
cross (a, b, c);
x = absv (c);
if (x >= 1e-30)
for (d1 = c; d1 < c + 3; )
*d1++ /= x;
return (0);
}
// -------- begin of function Music::change_volume --------//
void Music::change_volume(int vol)
{
if( !init_flag )
return;
if( is_playing() )
{
if( play_type & MUSIC_PLAY_CD )
{
audio.set_cd_volume(vol);
}
else
{
AbsVolume absv(vol,0);
audio.volume_long_wav(music_channel, DsVolume(absv));
}
}
}
void setprofile(real *prof1, real *prof2, int nprof, real rrange[],
bodyptr btab, int nbody)
{
real logrmin, logrdif, logr;
bodyptr bp;
int j;
logrmin = rlog10(rrange[0]);
logrdif = rlog10(rrange[1] / rrange[0]);
for (bp = btab; bp < NthBody(btab, nbody); bp = NextBody(bp)) {
logr = rlog10(absv(Pos(bp)));
j = 1 + floor(nprof * (logr - logrmin) / logrdif);
j = MAX(j, 0);
j = MIN(j, nprof + 1);
prof1[j] += Mass(bp);
prof2[j] += rsqr(Mass(bp));
}
}
// treeforce: supervise force calculation.
// _______________________________________
local void treeforce(void) {
bodyptr p1, p2, p;
real r, mr3i;
p1 = bodytab + MAX(nstatic, 0); // set dynamic body range
p2 = bodytab + nbody + MIN(nstatic, 0);
for (p = bodytab; p < bodytab+nbody; p++) // loop over all bodies
Update(p) = (testcalc ? p1 <= p && p < p2 : TRUE);
// flag bodies to update
maketree(bodytab, nbody); // construct tree structure
gravcalc(); // compute current forces
forcereport(); // print force statistics
#if defined(EXTGRAV)
for (p = bodytab; p < bodytab+nbody; p++) // loop over all bodies
if (Update(p) && gravgsp != NULL) { // update in extern field?
r = absv(Pos(p)); // get distance from origin
mr3i = - mass_gsp(gravgsp, r) / (r*r*r);
ADDMULVS(Acc(p), Pos(p), mr3i); // add extern acc and phi
Phi(p) += phi_gsp(gravgsp, r);
}
#endif
}
gsprof *snapgsp(bodyptr btab, int nbody, int npoint, real alpha, real beta)
{
gsprof *gsp;
real *rtab, *dtab, *mtab, *coef, mtot;
int nsamp, i, j;
if (nbody % npoint != 0)
error("%s: npoint must divide nbody\n", getargv0());
gsp = (gsprof *) allocate(sizeof(gsprof));
rtab = (real *) allocate(npoint * sizeof(real));
dtab = (real *) allocate(npoint * sizeof(real));
mtab = (real *) allocate(npoint * sizeof(real));
coef = (real *) allocate(3 * npoint * sizeof(real));
qsort(btab, nbody, SizeofBody, radrank);
nsamp = nbody / npoint;
mtot = 0.0;
j = 0;
for (i = 0; i < nbody; i++) {
mtot += Mass(NthBody(btab, i));
if (i % nsamp == nsamp - 1) {
rtab[j] = absv(Pos(NthBody(btab, i)));
mtab[j] = mtot;
j++;
}
}
spline(coef, rtab, mtab, npoint);
for (j = 0; j < npoint; j++)
dtab[j] =
spldif(rtab[j], rtab, mtab, coef, npoint) / (FOUR_PI * rsqr(rtab[j]));
gsp->npoint = npoint;
gsp->radius = rtab;
gsp->density = dtab;
gsp->alpha = alpha;
gsp->beta = beta;
gsp->mass = mtab;
gsp->mtot = mtot;
return (gsp);
}
void sphrprof(void)
{
bodyptr *rsort;
int skip, j, i;
real r;
if (nspheroid > 1) {
rsort = (bodyptr *) allocate(nspheroid * sizeof(bodyptr));
for (i = 0; i < nspheroid; i++)
rsort[i] = NthBody(spheroid, i);
qsort(rsort, nspheroid, sizeof(bodyptr), rankrad);
skip = (int) rceil(MAX(((real) nspheroid) / (NTAB - 1), 1.0));
msph[0] = rsph[0] = 0.0;
for (j = 1; j < NTAB; j++) {
i = skip * j - 1;
rsph[j] = (i < nspheroid ? absv(Pos(rsort[i])) : 1.0 + rsph[j-1]);
msph[j] = 0.0;
}
for (i = 0; i < nspheroid; i++) {
j = i / skip + 1;
if (j > NTAB-1)
error("%s.sphrprof: table overflow: i = %d skip = %d\n",
getargv0(), i, skip);
msph[j] = msph[j] + Mass(rsort[i]);
}
for (j = 1; j < NTAB; j++)
msph[j] += msph[j - 1];
} else {
for (j = 0; j < NTAB; j++) {
rsph[j] = 1.0 * j;
msph[j] = Mass(NthBody(spheroid, 0));
}
}
eprintf("[%s.sphrprof: rsph = %f %f %f ... %f %f]\n", getargv0(),
rsph[0], rsph[1], rsph[2], rsph[NTAB-2], rsph[NTAB-1]);
spline(mcof, rsph, msph, NTAB);
}
double simplify_angle_180(double angle)
{
double mult;
if (angle > PI) {
mult = angle/(2.0*PI);
mult = floor(mult);
angle -= mult*2.0*PI;
if (angle > PI)
angle -= 2.0 * PI;
} else if (angle < PI) {
mult = angle/(-2.0*PI);
mult = floor(mult);
angle += mult*2.0*PI;
if (angle < -PI)
angle += 2.0 * PI;
}
if (absv(angle+PI) < ANGLE_TOLERANCE)
angle = PI;
return angle;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing cheevd
*/
int main( int argc, char** argv)
{
TESTING_CUDA_INIT();
cuFloatComplex *h_A, *h_R, *h_work;
float *rwork, *w1, *w2;
magma_int_t *iwork;
float gpu_time, cpu_time;
magma_timestr_t start, end;
/* Matrix size */
magma_int_t N=0, n2;
magma_int_t size[8] = {1024,2048,3072,4032,5184,6016,7040,8064};
magma_int_t i, info;
magma_int_t ione = 1, izero = 0;
magma_int_t ISEED[4] = {0,0,0,1};
const char *uplo = MagmaLowerStr;
const char *jobz = MagmaVectorsStr;
magma_int_t checkres;
float result[3], eps = lapackf77_slamch( "E" );
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0) {
N = atoi(argv[++i]);
}
else if ( strcmp("-JV", argv[i]) == 0 ) {
jobz = MagmaVectorsStr;
}
else if ( strcmp("-JN", argv[i]) == 0 ) {
jobz = MagmaNoVectorsStr;
}
}
if (N>0)
printf(" testing_cheevd -N %d [-JV] [-JN]\n\n", (int) N);
else {
printf("\nUsage: \n");
printf(" testing_cheevd -N %d [-JV] [-JN]\n\n", (int) N);
exit(1);
}
}
else {
printf("\nUsage: \n");
printf(" testing_cheevd -N %d [-JV] [-JN]\n\n", 1024);
N = size[7];
}
checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL;
if ( checkres && jobz[0] == MagmaNoVectors ) {
printf( "Cannot check results when vectors are not computed (jobz='N')\n" );
checkres = false;
}
/* Query for workspace sizes */
cuFloatComplex aux_work[1];
float aux_rwork[1];
magma_int_t aux_iwork[1];
magma_cheevd( jobz[0], uplo[0],
N, h_R, N, w1,
aux_work, -1,
aux_rwork, -1,
aux_iwork, -1,
&info );
magma_int_t lwork, lrwork, liwork;
lwork = (magma_int_t) MAGMA_C_REAL( aux_work[0] );
lrwork = (magma_int_t) aux_rwork[0];
liwork = aux_iwork[0];
/* Allocate host memory for the matrix */
TESTING_MALLOC( h_A, cuFloatComplex, N*N );
TESTING_MALLOC( w1, float , N );
TESTING_MALLOC( w2, float , N );
TESTING_HOSTALLOC( h_R, cuFloatComplex, N*N );
TESTING_HOSTALLOC( h_work, cuFloatComplex, lwork );
TESTING_MALLOC( rwork, float, lrwork );
TESTING_MALLOC( iwork, magma_int_t, liwork );
printf(" N CPU Time(s) GPU Time(s) \n");
printf("===================================\n");
for(i=0; i<8; i++){
if (argc==1){
N = size[i];
}
n2 = N*N;
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
for( int i=0; i<N; i++) {
MAGMA_C_SET2REAL( h_A[i*N+i], MAGMA_C_REAL(h_A[i*N+i]) );
}
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
/* warm up run */
magma_cheevd(jobz[0], uplo[0],
N, h_R, N, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
/* query for optimal workspace sizes */
magma_cheevd(jobz[0], uplo[0],
N, h_R, N, w1,
h_work, -1,
rwork, -1,
iwork, -1,
&info);
int lwork_save = lwork;
int lrwork_save = lrwork;
int liwork_save = liwork;
lwork = min( lwork, (magma_int_t) MAGMA_C_REAL( h_work[0] ));
lrwork = min( lrwork, (magma_int_t) rwork[0] );
liwork = min( liwork, iwork[0] );
//printf( "lwork %d, query %d, used %d; liwork %d, query %d, used %d\n",
// lwork_save, (magma_int_t) h_work[0], lwork,
// liwork_save, iwork[0], liwork );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
start = get_current_time();
magma_cheevd(jobz[0], uplo[0],
N, h_R, N, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
end = get_current_time();
gpu_time = GetTimerValue(start,end)/1000.;
lwork = lwork_save;
lrwork = lrwork_save;
liwork = liwork_save;
if ( checkres ) {
/* =====================================================================
Check the results following the LAPACK's [zcds]drvst routine.
A is factored as A = U S U' and the following 3 tests computed:
(1) | A - U S U' | / ( |A| N )
(2) | I - U'U | / ( N )
(3) | S(with U) - S(w/o U) | / | S |
=================================================================== */
float temp1, temp2;
cuFloatComplex *tau;
lapackf77_chet21(&ione, uplo, &N, &izero,
h_A, &N,
w1, w1,
h_R, &N,
h_R, &N,
tau, h_work, rwork, &result[0]);
lapackf77_clacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
magma_cheevd('N', uplo[0],
N, h_R, N, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
temp1 = temp2 = 0;
for(int j=0; j<N; j++){
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[2] = temp2 / temp1;
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
start = get_current_time();
lapackf77_cheevd(jobz, uplo,
&N, h_A, &N, w2,
h_work, &lwork,
rwork, &lrwork,
iwork, &liwork,
&info);
end = get_current_time();
if (info < 0)
printf("Argument %d of cheevd had an illegal value.\n", (int) -info);
cpu_time = GetTimerValue(start,end)/1000.;
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %6.2f %6.2f\n",
(int) N, cpu_time, gpu_time);
if ( checkres ){
printf("Testing the factorization A = U S U' for correctness:\n");
printf("(1) | A - U S U' | / (|A| N) = %e\n", result[0]*eps);
printf("(2) | I - U'U | / N = %e\n", result[1]*eps);
printf("(3) | S(w/ U)-S(w/o U)|/ |S| = %e\n\n", result[2]);
}
if (argc != 1)
break;
}
/* Memory clean up */
TESTING_FREE( h_A);
TESTING_FREE( w1);
TESTING_FREE( w2);
TESTING_FREE( rwork);
TESTING_FREE( iwork);
TESTING_HOSTFREE(h_work);
TESTING_HOSTFREE( h_R);
/* Shutdown */
TESTING_CUDA_FINALIZE();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssygvd
*/
int main( int argc, char** argv)
{
TESTING_INIT_MGPU();
float *h_A, *h_Ainit, *h_B, *h_Binit, *h_work;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
#endif
float *w1, *w2, result;
magma_int_t *iwork;
float mgpu_time, gpu_time, cpu_time;
/* Matrix size */
magma_int_t N=0, n2;
magma_int_t info;
magma_int_t ione = 1;
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t ISEED[4] = {0,0,0,1};
magma_timestr_t start, end;
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
char jobz = opts.jobz;
int checkres = opts.check;
char uplo = opts.uplo;
magma_int_t itype = opts.itype;
if ( checkres && jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
jobz = MagmaVec;
}
printf("using: nrgpu = %d, itype = %d, jobz = %c, uplo = %c, checkres = %d\n",
(int) opts.ngpu, (int) itype, jobz, uplo, (int) checkres);
printf(" N M nr GPU MGPU Time(s) \n");
printf("====================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
n2 = N*N;
#if defined(PRECISION_z) || defined(PRECISION_c)
magma_int_t lwork = 2*N + N*N;
magma_int_t lrwork = 1 + 5*N +2*N*N;
// MKL's ssygvd has a bug for small N - it looks like what is returned by a
// query (consistent with LAPACK's number above) is different from a the memory
// requirement ckeck (that returns info -11). The lwork increase below is needed
// to pass this check.
if (N<32)
lwork = 34*32;
#else
magma_int_t lwork = 1 + 6*N + 2*N*N;
#endif
magma_int_t liwork = 3 + 5*N;
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( w2, float, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
printf(" N CPU Time(s) GPU Time(s) MGPU Time(s) \n");
printf("==================================================\n");
/* 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)||( checkres )){
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 );
}
if(opts.warmup){
// ==================================================================
// Warmup using MAGMA.
// ==================================================================
magma_ssygvd_m( opts.ngpu, itype, jobz, uplo,
N, h_A, N, h_B, N, 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
// ===================================================================
start = get_current_time();
magma_ssygvd_m( opts.ngpu, itype, jobz, uplo,
N, h_A, N, h_B, N, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
end = get_current_time();
if(info != 0)
printf("magma_ssygvd_m returned error %d: %s.\n",
(int) info, magma_strerror( info ));
mgpu_time = GetTimerValue(start,end)/1000.;
if ( checkres ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvd 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",&uplo, &N, h_Ainit, &N, rwork);
result /= lapackf77_slange("1",&N , &N, h_A, &N, rwork);
if (itype == 1){
blasf77_ssymm("L", &uplo, &N, &N, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", &uplo, &N, &N, &c_neg_one, h_Binit, &N, h_A, &N, &c_one, h_work, &N);
result *= lapackf77_slange("1", &N, &N, h_work, &N, rwork)/N;
}
else if (itype == 2){
blasf77_ssymm("L", &uplo, &N, &N, &c_one, h_Binit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", &uplo, &N, &N, &c_one, h_Ainit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_slange("1", &N, &N, h_A, &N, rwork)/N;
}
else if (itype == 3){
blasf77_ssymm("L", &uplo, &N, &N, &c_one, h_Ainit, &N, h_A, &N, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_sscal(&N, &w1[i], &h_A[i*N], &ione);
blasf77_ssymm("L", &uplo, &N, &N, &c_one, h_Binit, &N, h_work, &N, &c_neg_one, h_A, &N);
result *= lapackf77_slange("1", &N, &N, h_A, &N, rwork)/N;
}
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
=================================================================== */
start = get_current_time();
magma_ssygvd(itype, jobz, uplo,
N, h_A, N, h_B, N, w2,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info);
end = get_current_time();
if(info != 0)
printf("magma_ssygvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
gpu_time = GetTimerValue(start,end)/1000.;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
start = get_current_time();
lapackf77_ssygvd(&itype, &jobz, &uplo,
&N, h_Ainit, &N, h_Binit, &N, w2,
h_work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
&info);
end = get_current_time();
if (info != 0)
printf("lapackf77_ssygvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
cpu_time = GetTimerValue(start,end)/1000.;
float temp1 = 0;
float temp2 = 0;
for(int j=0; j<N; j++){
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
float result2 = temp2 / (((float)N)*temp1);
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %6.2f %6.2f %6.2f\n",
(int) N, cpu_time, gpu_time, mgpu_time);
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if(itype==1)
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %8.2e%s\n", result, (result < tol ? "" : " failed") );
else if(itype==2)
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e%s\n", result, (result < tol ? "" : " failed") );
else if(itype==3)
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e%s\n", result, (result < tol ? "" : " failed") );
printf( "(3) | D(MGPU)-D(LAPACK) |/ |D| = %8.2e%s\n\n", result2, (result2 < tolulp ? "" : " failed") );
}
else {
printf("%5d ------ ------ %6.2f\n",
(int) N, mgpu_time);
}
/* Memory clean up */
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( w2 );
TESTING_FREE_CPU( iwork );
if((opts.warmup)||( checkres )){
TESTING_FREE_CPU( h_Ainit );
TESTING_FREE_CPU( h_Binit );
}
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
/* Shutdown */
TESTING_FINALIZE_MGPU();
}
int main(int argc, char **argv)
{
int nxtarg = 0;
int i;
char c;
char tmpstring[TMPSTRINGLEN+1];
char filename[TMPSTRINGLEN+1];
char outfilename[TMPSTRINGLEN+1];
FILE *cpudsptimefile = NULL;
FILE *cpupostimefile = NULL;
FILE *possynctimefile = NULL;
FILE *outfile = NULL;
u32 cpudsp1[3], cpudsp2[3]; // this assumes that each timecheck contains three numbers, the cpu time, the dsp time, and the subsequent cpu time
u32 cpupos;
u32 lastcpupos;
u32 dspframetime;
u32 possynctime = 0;
int ngaps = 0; // the number of > 40 ms gaps
int ncomputed = 0; // the number of times we used the computed timestamp instead of the sync edge time
u32 tmpdiff, lastdiff, maxframetimediff = 0;
while (++nxtarg < argc) {
if (strcmp(argv[nxtarg], "-cd") == 0) {
strcpy(filename, argv[++nxtarg]);
/* open the file */
if ((cpudsptimefile = fopen(filename, "r"))
== NULL) {
fprintf(stderr, "Error opening %s for reading\n", filename);
exit(-1);
}
}
else if (strcmp(argv[nxtarg], "-cp") == 0) {
/* open cpu pos time file */
strcpy(filename, argv[++nxtarg]);
if ((cpupostimefile = fopen(filename, "r")) == NULL) {
fprintf(stderr, "Error opening %s for reading\n", filename);
exit(-1);
}
}
else if (strcmp(argv[nxtarg], "-ps") == 0) {
/* open pos sync times file */
strcpy(filename, argv[++nxtarg]);
if ((possynctimefile = fopen(filename, "r")) == NULL) {
fprintf(stderr, "Error opening %s for reading\n", filename);
exit(-1);
}
}
else if (strcmp(argv[nxtarg], "-o") == 0) {
/* open the output file */
strcpy(outfilename, argv[++nxtarg]);
if ((outfile = fopen(outfilename, "w")) == NULL) {
fprintf(stderr, "Error opening %s for writing\n", outfilename);
exit(-1);
}
}
else {
fprintf(stderr, "Usage: nspike_postimestamp -cd cpudsptimefile -cp cpupostimefile -ps possynctimesfile -o outputfile\n");
exit(1);
}
}
/* check to make sure that all necessary arguments have been specified */
if ((cpudsptimefile == NULL) || (cpupostimefile == NULL) ||
(possynctimefile == NULL) || (outfile == NULL)) {
fprintf(stderr, "Usage: nspike_postimestamp -cd cpudsptimefile -cp cpupostimefile -ps possynctimesfile -o outputfile\n");
exit(1);
}
/* write out an uncompressed header to the output file */
fprintf(outfile, "%%%%BEGINHEADER\n");
fprintf(outfile, "%% File type:\tBinary\n");
fprintf(outfile, "%% Extraction type:\tdsp position frame time stamps \n");
fprintf(outfile, "%% Fields:\t timestamp (unsigned int)\n");
fprintf(outfile, "%%%%ENDHEADER\n");
/* read past the headers of the input files */
do {
fgets(tmpstring, TMPSTRINGLEN, cpudsptimefile);
} while ((strncmp(tmpstring, "%%ENDHEADER", 10) != 0) &&
(strncmp(tmpstring, "%%ENDCONFIG", 10) != 0));
do {
fgets(tmpstring, TMPSTRINGLEN, cpupostimefile);
} while ((strncmp(tmpstring, "%%ENDHEADER", 10) != 0) &&
(strncmp(tmpstring, "%%ENDCONFIG", 10) != 0));
do {
fgets(tmpstring, TMPSTRINGLEN, possynctimefile);
} while ((strncmp(tmpstring, "%%ENDHEADER", 10) != 0) &&
(strncmp(tmpstring, "%%ENDCONFIG", 10) != 0));
if (nextvalidcpudsptime(cpudsptimefile, cpudsp1) == 0) {
exit(-1);
}
cpudsp2[0] = 0;
/* Now we go through the cpupostimefile and find the possync time that
* corresponds to each frame */
lastcpupos = 0;
while (!feof(cpupostimefile)) {
/* read in the next cpupostimefile entry */
if (fread(&cpupos, sizeof(u32), 1, cpupostimefile) != 1) {
fprintf(stderr, "Error reading cpu - position frame time from cpupostimefile at offset %ld, eof = %d\n", ftell(cpupostimefile), feof(cpupostimefile));
break;
}
if (lastcpupos && (cpupos - lastcpupos > 400)) {
ngaps++;
}
lastcpupos = cpupos;
/* make sure that the next cpudsp time is after the current cpu frame
* time */
while ((cpudsp2[0] < cpupos) && nextvalidcpudsptime(cpudsptimefile,
cpudsp2)) {
/* move cpudsp2 to cpudsp1 and move on */
memcpy(cpudsp1, cpudsp2, 3 * sizeof(u32));
}
/* use the closest cpudsp time to convert the current frame time to a
* dsp clock time */
if (absv(cpudsp1[0] - cpupos) < absv(cpudsp2[0] - cpupos)) {
dspframetime = cpupos + (cpudsp1[1] - cpudsp1[0]);
}
else {
dspframetime = cpupos + (cpudsp2[1] - cpudsp2[0]);
}
/* initialize the first possync time so that we read in the correct
* first value */
if (possynctime == 0) {
possynctime = dspframetime;
}
/* find the pos sync time closest to the dspframe time and write it out
* to the file */
tmpdiff = UINT_MAX;
/* check to see if the possynctime is much greater than the current
* dspframetime as can occur when a time offset has been added */
if (possynctime < dspframetime + 10000) {
do {
lastdiff = tmpdiff;
/* get the timestamp */
if (fread(&possynctime, sizeof(u32), 1, possynctimefile) != 1) {
fprintf(stderr, "Error reading position frame sync time from possynctime file\n");
exit(-1);
}
/* get the transition */
if (fread(&c, sizeof(char), 1, possynctimefile) != 1) {
fprintf(stderr, "Error reading position frame sync type from possynctime file\n");
exit(-1);
}
} while (((tmpdiff = abs(possynctime - dspframetime)) > 50) &&
(tmpdiff < lastdiff));
}
if (tmpdiff >= 1000) {
/* we should use the computed time instead of the read in time, as
* we may have missed the edge of a frame */
fwrite(&dspframetime, sizeof(u32), 1, outfile);
ncomputed++;
}
else {
fwrite(&possynctime, sizeof(u32), 1, outfile);
if (tmpdiff > maxframetimediff) {
maxframetimediff = tmpdiff;
}
}
}
fprintf(stderr, "Timestamp file created, %d gaps > 40 ms\nLargest discrepancy between computed and actual dsp frame times = %2.1f ms\nNumber of frames for which the computed timestamp was used = %d\n", ngaps,
(float) maxframetimediff / 10, ncomputed);
fclose(outfile);
fclose(possynctimefile);
fclose(cpudsptimefile);
fclose(cpupostimefile);
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsyevd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
double *h_A, *h_R, *d_R, *h_work;
double *w1, *w2;
magma_int_t *iwork;
magma_int_t N, n2, info, lwork, liwork, lda, ldda, aux_iwork[1];
magma_int_t izero = 0;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double result[3], eps, aux_work[1];
eps = lapackf77_dlamch( "E" );
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf(" N CPU Time (sec) GPU Time (sec)\n");
printf("=======================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
n2 = N*N;
lda = N;
ldda = ((N + 31)/32)*32;
// query for workspace sizes
magma_dsyevd_gpu( opts.jobz, opts.uplo,
N, NULL, ldda, NULL,
NULL, lda,
aux_work, -1,
aux_iwork, -1,
&info );
lwork = (magma_int_t) aux_work[0];
liwork = aux_iwork[0];
/* Allocate host memory for the matrix */
TESTING_MALLOC_CPU( h_A, double, N*lda );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, double, N*lda );
TESTING_MALLOC_PIN( h_work, double, lwork );
TESTING_MALLOC_DEV( d_R, double, N*ldda );
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
magma_dsetmatrix( N, N, h_A, lda, d_R, ldda );
/* warm up run */
if ( opts.warmup ) {
magma_dsyevd_gpu( opts.jobz, opts.uplo,
N, d_R, ldda, w1,
h_R, lda,
h_work, lwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_dsyevd_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_dsetmatrix( N, N, h_A, lda, d_R, ldda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_dsyevd_gpu( opts.jobz, opts.uplo,
N, d_R, ldda, w1,
h_R, lda,
h_work, lwork,
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_dsyevd_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zcds]drvst routine.
A is factored as A = U S U' and the following 3 tests computed:
(1) | A - U S U' | / ( |A| N )
(2) | I - U'U | / ( N )
(3) | S(with U) - S(w/o U) | / | S |
=================================================================== */
double temp1, temp2;
// tau=NULL is unused since itype=1
magma_dgetmatrix( N, N, d_R, ldda, h_R, lda );
lapackf77_dsyt21( &ione, &opts.uplo, &N, &izero,
h_A, &lda,
w1, h_work,
h_R, &lda,
h_R, &lda,
NULL, h_work, &result[0] );
magma_dsetmatrix( N, N, h_A, lda, d_R, ldda );
magma_dsyevd_gpu( MagmaNoVec, opts.uplo,
N, d_R, ldda, w2,
h_R, lda,
h_work, lwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_dsyevd_gpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for( int j=0; j<N; j++ ) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[2] = temp2 / (((double)N)*temp1);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dsyevd( &opts.jobz, &opts.uplo,
&N, h_A, &lda, w2,
h_work, &lwork,
iwork, &liwork,
&info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_dsyevd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf("%5d %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
/* =====================================================================
Print execution time
=================================================================== */
if ( opts.check ) {
printf("Testing the factorization A = U S U' for correctness:\n");
printf("(1) | A - U S U' | / (|A| N) = %8.2e%s\n", result[0]*eps, (result[0]*eps < tol ? "" : " failed") );
printf("(2) | I - U'U | / N = %8.2e%s\n", result[1]*eps, (result[1]*eps < tol ? "" : " failed") );
printf("(3) | S(w/ U) - S(w/o U) | / |S| = %8.2e%s\n\n", result[2] , (result[2] < tolulp ? "" : " failed") );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_work );
TESTING_FREE_DEV( d_R );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvd
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time, cpu_time;
magmaDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work;
double *rwork, *w1, *w2;
double result[4] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, nb, lwork, liwork, lda, lrwork;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
double d_one = 1.;
double d_neg_one = -1.;
//double d_ten = 10.;
//magma_int_t izero = 0;
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 );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %s, uplo = %s\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo));
printf(" N CPU Time (sec) GPU Time(sec)\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;
n2 = N*lda;
nb = magma_get_zhetrd_nb(N);
lwork = 2*N*nb + N*N;
lrwork = 1 + 5*N +2*N*N;
liwork = 3 + 5*N;
TESTING_MALLOC_CPU( h_A, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( h_B, magmaDoubleComplex, n2 );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( rwork, double, lrwork );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_S, magmaDoubleComplex, n2 );
TESTING_MALLOC_PIN( h_work, magmaDoubleComplex, lwork );
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
//lapackf77_zlatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten,
// &d_one, &N, &N, lapack_uplo_const(opts.uplo), h_B, &lda, h_work, &info);
//lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_B, &lda);
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
magma_zmake_hpd( N, h_B, lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
/* warmup */
if ( opts.warmup ) {
magma_zhegvd( opts.itype, opts.jobz, opts.uplo,
N, h_R, lda, h_S, lda, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
gpu_time = magma_wtime();
magma_zhegvd( opts.itype, opts.jobz, opts.uplo,
N, h_R, lda, h_S, lda, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.check ) {
/* =====================================================================
Check the results following the LAPACK's [zc]hegvd 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)
(2) | I - V V' B | / ( N ) (itype = 1,2)
| B - V V' | / ( |B| N ) (itype = 3)
(3) | S(with V) - S(w/o V) | / | S |
=================================================================== */
double temp1, temp2;
//magmaDoubleComplex *tau;
if ( opts.itype == 1 || opts.itype == 2 ) {
lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_S, &lda);
blasf77_zgemm("N", "C", &N, &N, &N, &c_one, h_R, &lda, h_R, &lda, &c_zero, h_work, &N);
blasf77_zhemm("R", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_work, &N, &c_one, h_S, &lda);
result[1] = lapackf77_zlange("1", &N, &N, h_S, &lda, rwork) / N;
}
else if ( opts.itype == 3 ) {
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda);
blasf77_zherk(lapack_uplo_const(opts.uplo), "N", &N, &N, &d_neg_one, h_R, &lda, &d_one, h_S, &lda);
result[1] = lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_S, &lda, rwork) / N
/ lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_B, &lda, rwork);
}
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1", lapack_uplo_const(opts.uplo), &N, h_A, &lda, rwork);
result[0] /= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork);
if ( opts.itype == 1 ) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_neg_one, h_B, &lda, h_R, &lda, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &N, h_work, &lda, rwork)/N;
}
else if ( opts.itype == 2 ) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_R, &lda, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork)/N;
}
else if ( opts.itype == 3 ) {
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_A, &lda, h_R, &lda, &c_zero, h_work, &N);
for(int i=0; i<N; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", lapack_uplo_const(opts.uplo), &N, &N, &c_one, h_B, &lda, h_work, &N, &c_neg_one, h_R, &lda);
result[0] *= lapackf77_zlange("1", &N, &N, h_R, &lda, rwork)/N;
}
/*
lapackf77_zhet21( &ione, lapack_uplo_const(opts.uplo), &N, &izero,
h_A, &lda,
w1, w1,
h_R, &lda,
h_R, &lda,
tau, h_work, rwork, &result[0] );
*/
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &lda, h_S, &lda );
magma_zhegvd( opts.itype, MagmaNoVec, opts.uplo,
N, h_R, lda, h_S, lda, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info );
if (info != 0)
printf("magma_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j<N; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[2] = temp2 / (((double)N)*temp1);
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_zhegvd( &opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
&N, h_A, &lda, h_B, &lda, w2,
h_work, &lwork,
rwork, &lrwork,
iwork, &liwork,
&info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf("lapackf77_zhegvd returned error %d: %s.\n",
(int) info, magma_strerror( info ));
printf("%5d %7.2f %7.2f\n",
(int) N, cpu_time, gpu_time);
}
else {
printf("%5d --- %7.2f\n",
(int) N, gpu_time);
}
/* =====================================================================
Print execution 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[0], (result[0] < tol ? "ok" : "failed") );
}
else if ( opts.itype==2 ) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
else if ( opts.itype==3 ) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed") );
}
if ( opts.itype==1 || opts.itype==2 ) {
printf("(2) | I - Z Z' B | / N = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
}
else {
printf("(2) | B - Z Z' | / (|B| N) = %8.2e %s\n", result[1], (result[1] < tol ? "ok" : "failed") );
}
printf( "(3) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[2], (result[2] < tolulp ? "ok" : "failed") );
status += ! (result[0] < tol && result[1] < tol && result[2] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( rwork );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing zhegvdx
*/
int main( int argc, char** argv)
{
//#define USE_MGPU
#ifdef USE_MGPU
TESTING_CUDA_INIT_MGPU();
#else
TESTING_CUDA_INIT();
#endif
magma_int_t nrgpu =1;
cuDoubleComplex *h_A, *h_R, *h_B, *h_S, *h_work;
double *rwork, *w1, *w2;
magma_int_t *iwork;
double gpu_time, cpu_time;
magma_timestr_t start, end;
/* Matrix size */
magma_int_t N=0, n2;
magma_int_t size[4] = {1024,2048,4100,6001};
magma_int_t i, itype, info;
magma_int_t ione = 1, izero = 0;
magma_int_t five = 5;
cuDoubleComplex c_zero = MAGMA_Z_ZERO;
cuDoubleComplex c_one = MAGMA_Z_ONE;
cuDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
double d_one = 1.;
double d_neg_one = -1.;
double d_ten = 10.;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t il,iu,m1,m2;
double vl,vu;
double fraction_ev = 0;
//const char *uplo = MagmaLowerStr;
char *uplo = (char*)MagmaLowerStr;
//char *uplo = (char*)MagmaUpperStr;
char *jobz = (char*)MagmaVectorsStr;
char range = 'A';
itype = 1;
magma_int_t checkres;
double result[2];
int flagN = 0;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0){
N = atoi(argv[++i]);
if (N>0){
printf(" testing_zhegvdx -N %d\n\n", (int) N);
flagN=1;
}
else {
printf("\nUsage: \n");
printf(" testing_zhegvdx -N %d\n\n", (int) N);
exit(1);
}
}
if (strcmp("-ngpu", argv[i])==0){
nrgpu = atoi(argv[++i]);
if (nrgpu>0){
printf(" testing_zhegvdx -ngpu %d\n\n", (int) nrgpu);
}
else {
printf("\nUsage: \n");
printf(" testing_zhegvdx -ngpu %d\n\n", (int) nrgpu);
exit(1);
}
}
if (strcmp("-itype", argv[i])==0){
itype = atoi(argv[++i]);
if (itype>0 && itype <= 3){
printf(" testing_zhegvdx -itype %d\n\n", (int) itype);
}
else {
printf("\nUsage: \n");
printf(" testing_zhegvdx -itype %d\n\n", (int) itype);
exit(1);
}
}
if (strcmp("-FE", argv[i])==0){
fraction_ev = atof(argv[++i]);
if (fraction_ev > 0 && fraction_ev <= 1){
printf(" testing_zhegvdx -FE %f\n\n", fraction_ev);
}
else {
fraction_ev = 0;
}
}
if (strcmp("-L", argv[i])==0){
uplo = (char*)MagmaLowerStr;
printf(" testing_zhegvdx -L");
}
if (strcmp("-U", argv[i])==0){
uplo = (char*)MagmaUpperStr;
printf(" testing_zhegvdx -U");
}
}
} else {
printf("\nUsage: \n");
printf(" testing_zhegvdx -L/U -N %d -itype %d\n\n", 1024, 1);
}
if(!flagN)
N = size[3];
checkres = getenv("MAGMA_TESTINGS_CHECK") != NULL;
n2 = N * N;
/* Allocate host memory for the matrix */
TESTING_MALLOC( h_A, cuDoubleComplex, n2);
TESTING_MALLOC( h_B, cuDoubleComplex, n2);
TESTING_MALLOC( w1, double , N);
TESTING_MALLOC( w2, double , N);
TESTING_HOSTALLOC(h_R, cuDoubleComplex, n2);
TESTING_HOSTALLOC(h_S, cuDoubleComplex, n2);
magma_int_t nb = magma_get_zhetrd_nb(N);
magma_int_t lwork = magma_zbulge_get_lq2(N) + 2*N + N*N;
magma_int_t lrwork = 1 + 5*N +2*N*N;
magma_int_t liwork = 3 + 5*N;
TESTING_HOSTALLOC(h_work, cuDoubleComplex, lwork);
TESTING_HOSTALLOC( rwork, double, lrwork);
TESTING_MALLOC( iwork, magma_int_t, liwork);
printf(" N M GPU Time(s) \n");
printf("==========================\n");
for(i=0; i<4; i++){
if (!flagN){
N = size[i];
n2 = N*N;
}
if (fraction_ev == 0){
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int)(fraction_ev*N);
if (iu < 1) iu = 1;
}
/* Initialize the matrix */
lapackf77_zlarnv( &ione, ISEED, &n2, h_A );
//lapackf77_zlatms( &N, &N, "U", ISEED, "P", w1, &five, &d_ten,
// &d_one, &N, &N, uplo, h_B, &N, h_work, &info);
//lapackf77_zlaset( "A", &N, &N, &c_zero, &c_one, h_B, &N);
lapackf77_zlarnv( &ione, ISEED, &n2, h_B );
/* increase the diagonal */
{
magma_int_t i, j;
for(i=0; i<N; i++) {
MAGMA_Z_SET2REAL( h_B[i*N+i], ( MAGMA_Z_REAL(h_B[i*N+i]) + 1.*N ) );
}
}
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
#ifdef USE_MGPU
magma_zhegvdx_2stage_m(nrgpu, itype, jobz[0], range, uplo[0],
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
#else
magma_zhegvdx_2stage(itype, jobz[0], range, uplo[0],
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
#endif
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
start = get_current_time();
#ifdef USE_MGPU
magma_zhegvdx_2stage_m(nrgpu, itype, jobz[0], range, uplo[0],
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
#else
magma_zhegvdx_2stage(itype, jobz[0], range, uplo[0],
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
#endif
end = get_current_time();
gpu_time = GetTimerValue(start,end)/1000.;
if ( checkres ) {
/* =====================================================================
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)
(2) | S(with V) - S(w/o V) | / | S |
=================================================================== */
double temp1, temp2;
cuDoubleComplex *tau;
result[0] = 1.;
result[0] /= lapackf77_zlanhe("1",uplo, &N, h_A, &N, rwork);
result[0] /= lapackf77_zlange("1",&N , &m1, h_R, &N, rwork);
if (itype == 1){
blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", uplo, &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (itype == 2){
blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (itype == 3){
blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i<m1; ++i)
blasf77_zdscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_zhemm("L", uplo, &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_zlange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_zlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_zhegvdx(itype, 'N', range, uplo[0],
N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2,
h_work, lwork,
rwork, lrwork,
iwork, liwork,
&info);
temp1 = temp2 = 0;
for(int j=0; j<m2; j++){
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[1] = temp2 / temp1;
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %6.2f\n",
(int) N, (int) m1, gpu_time);
if ( checkres ){
printf("Testing the eigenvalues and eigenvectors for correctness:\n");
if(itype==1)
printf("(1) | A Z - B Z D | / (|A| |Z| N) = %e\n", result[0]);
else if(itype==2)
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %e\n", result[0]);
else if(itype==3)
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %e\n", result[0]);
printf("(2) | D(w/ Z)-D(w/o Z)|/ |D| = %e\n\n", result[1]);
}
if (flagN)
break;
}
cudaSetDevice(0);
/* Memory clean up */
TESTING_FREE( h_A);
TESTING_FREE( h_B);
TESTING_FREE( w1);
TESTING_FREE( w2);
TESTING_HOSTFREE( rwork);
TESTING_FREE( iwork);
TESTING_HOSTFREE(h_work);
TESTING_HOSTFREE( h_R);
TESTING_HOSTFREE( h_S);
/* Shutdown */
#ifdef USE_MGPU
TESTING_CUDA_FINALIZE_MGPU();
#else
TESTING_CUDA_FINALIZE();
#endif
}
double line_segment_distance(double seg1_a[3], double seg1_b[3],
double seg2_a[3], double seg2_b[3])
{
double sc, sd, sn, tc, td, tn;
double uu, uv, vv, uw, vw, det;
double u[3], v[3], w[3];
vec_subt(u, seg1_b, seg1_a);
vec_subt(v, seg2_b, seg2_a);
vec_subt(w, seg1_a, seg2_a);
uu = vec_dot(u, u);
uv = vec_dot(u, v);
vv = vec_dot(v, v);
uw = vec_dot(u, w);
vw = vec_dot(v, w);
det = uu*vv - sqr(uv);
sd = det;
td = det;
if (det < ALMOST_ZERO) {
sn = 0.0;
sd = 1.0;
tn = vw;
td = vv;
} else {
sn = uv*vw - vv*uw;
tn = uu*vw - uv*uw;
if (sn < 0.0) {
sn = 0.0;
tn = vw;
td = vv;
} else if (sn > sd) {
sn = sd;
tn = vw + uv;
td = vv;
}
}
if (tn < 0.0) {
tn = 0.0;
if (-uw < 0.0) {
sn = 0.0;
} else if (-uw > uu) {
sn = sd;
} else {
sn = -uw;
sd = uu;
}
} else if (tn > td) {
tn = td;
if (uv-uw < 0.0) {
sn = 0.0;
} else if (uv-uw > uu) {
sn = sd;
} else {
sn = uv-uw;
sd = uu;
}
}
sc = (absv(sn) < ALMOST_ZERO ? 0.0 : sn / sd);
tc = (absv(tn) < ALMOST_ZERO ? 0.0 : tn / td);
vs_mult(u, sc);
vs_mult(v, tc);
vec_plus(w, w, u);
vec_subt(w, w, v);
return vec_dot(w, w);
}
void igrfmodel(VEC *bv, double *geo, double *Nmax_loc)
{
/* Local variables */
register double *d1, *d2;
int imax, nmax, N;
double f, h[NCOEF]; /* KV */
int i, k, m;
double s, x, y, z;
int ihmax, ih, il;
double xi[3], rq;
int ihm, ilm;
double srq;
srq=absv(geo); //square root of position vector magnitude
rq = srq*srq; //magnitude of vector
N = (int) *Nmax_loc; // store the max order value in a local variable
if (rq < .8) { // 0.8 * earth radius is the limit to determine the altitude to be above or below earth surface
//printf ("igrf call below surface !!!\n");
}
// number of harmonics depends on the distance from the earth
rq = 1. / rq;
//TODO static INLINE arm_status arm_sqrt_f32
srq = sqrt(rq);
if (rq < 0.25)
nmax = (N - 3) * 4.0 * rq + 3;
else
nmax = N;
//what is xi?
for (d1 = xi, d2 = geo; d1 < xi + 3; )
*d1++ = *d2++ * rq;//position unit vector
ihmax = nmax * nmax; //total number of COEFFS including g and h coeffs
imax = nmax + nmax - 2; //max index for m?
il = ihmax + nmax + nmax;//max index for ?
d1 = h + ihmax; //max index address for h[NCOEF]
d2 = Gh + ihmax; //max index address for Gh[NCOEF]
for ( ; d1 <= h + il; )
*d1++ = *d2++;//last 26 entries of h and Gh are made equal
for (k = 0; k < 3; k += 2) {//runs only 2 times ... for 0 and 2
i = imax;
ih = ihmax;
while (i >= k) {
il = ih - i - 1;
f = 2. / (double) (i - k + 2);
x = xi[0] * f;
y = xi[1] * f;
z = xi[2] * (f + f);
i += -2;
if (i >= 2) {
for (m = 3; m <= i + 1; m += 2) {
ihm = ih + m;
ilm = il + m;
h[ilm+1] = Gh[ilm+1]+z*h[ihm+1]+x*(h[ihm+3]-h[ihm-1])-y*(h[ihm+2]+h[ihm-2]);
h[ilm] = Gh[ilm]+z*h[ihm]+x*(h[ihm+2]-h[ihm-2])+y*(h[ihm+3]+h[ihm-1]);
}
h[il+2] = Gh[il+2]+z*h[ih+2]+x*h[ih+4]-y*(h[ih+3]+h[ih]);
h[il+1] = Gh[il+1]+z*h[ih+1]+y*h[ih+4]+x*(h[ih+3]-h[ih]);
} else if (i == 0) {
h[il + 2] = Gh[il+2]+z*h[ih+2]+x*h[ih+4]-y*(h[ih+3]+h[ih]);
h[il+1] = Gh[il+1]+z*h[ih+1]+y*h[ih+4]+x*(h[ih+3]-h[ih]);
}
h[il] = Gh[il]+z*h[ih]+(x*h[ih+1]+y*h[ih+2])*2.;
ih = il;
}
}
s = h[0]*.5+(h[1]*xi[2]+h[2]*xi[0]+h[3]*xi[1])*2.;
f = (rq+rq)*srq;
x = f*(h[2]-s*(*(geo+0)))*1e-9;
y = f*(h[3]-s*(*(geo+1)))*1e-9;
z = f*(h[1]-s*(*(geo+2)))*1e-9;
bv->ve[0] = x;
bv->ve[1] = y;
bv->ve[2] = z;
/* *(bv+3) = sqrt(x*x+y*y+z*z); */
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsygvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time /*cpu_time*/;
double *h_A, *h_R, *h_B, *h_S, *h_work;
double *w1, *w2, vl=0, vu=0;
double result[2] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, il, iu, m1, m2, nb, lwork, liwork;
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
#if defined(PRECISION_z) || defined(PRECISION_c)
double *rwork;
magma_int_t lrwork;
#endif
//double d_one = 1.;
//double d_ten = 10.;
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 );
double tol = opts.tolerance * lapackf77_dlamch("E");
double tolulp = opts.tolerance * lapackf77_dlamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf("using: itype = %d, jobz = %s, uplo = %s, check = %d, fraction = %6.4f\n",
(int) opts.itype, lapack_vec_const(opts.jobz), lapack_uplo_const(opts.uplo),
(int) opts.check, opts.fraction);
printf(" N M GPU Time (sec)\n");
printf("============================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
n2 = N*N;
nb = magma_get_dsytrd_nb(N);
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = 2*N*nb + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = 1 + 6*N*nb + 2* N*N;
#endif
liwork = 3 + 5*N;
if ( opts.fraction == 0 ) {
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int) (opts.fraction*N);
if (iu < 1) iu = 1;
}
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_CPU( h_B, double, n2 );
TESTING_MALLOC_CPU( w1, double, N );
TESTING_MALLOC_CPU( w2, double, N );
TESTING_MALLOC_CPU( iwork, magma_int_t, liwork );
TESTING_MALLOC_PIN( h_R, double, n2 );
TESTING_MALLOC_PIN( h_S, double, n2 );
TESTING_MALLOC_PIN( h_work, double, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_MALLOC_PIN( rwork, double, lrwork);
#endif
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlarnv( &ione, ISEED, &n2, h_B );
magma_dmake_hpd( N, h_B, N );
magma_dmake_symmetric( N, h_A, N );
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if(opts.warmup){
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_dsygvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_dsygvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_dsygvdx( opts.itype, opts.jobz, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_dsygvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
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)
(2) | S(with V) - S(w/o V) | / | S |
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
double *rwork = h_work + N*N;
#endif
double temp1, temp2;
result[0] = 1.;
result[0] /= lapackf77_dlansy("1", lapack_uplo_const(opts.uplo), &N, h_A, &N, rwork);
result[0] /= lapackf77_dlange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_dscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_dlange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_dscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_dlange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_dscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_dsymm("L", lapack_uplo_const(opts.uplo), &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_dlange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_dsygvdx( opts.itype, MagmaNoVec, MagmaRangeI, opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_dsygvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j < m2; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[1] = temp2 / (((double)m2)*temp1);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %7.2f\n",
(int) N, (int) m1, gpu_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[0], (result[0] < tol ? "ok" : "failed"));
}
else if (opts.itype == 2) {
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
else if (opts.itype == 3) {
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e %s\n", result[0], (result[0] < tol ? "ok" : "failed"));
}
printf( "(2) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e %s\n\n", result[1], (result[1] < tolulp ? "ok" : "failed"));
status += ! (result[0] < tol && result[1] < tolulp);
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( w1 );
TESTING_FREE_CPU( w2 );
TESTING_FREE_CPU( iwork );
TESTING_FREE_PIN( h_R );
TESTING_FREE_PIN( h_S );
TESTING_FREE_PIN( h_work );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_FREE_PIN( rwork );
#endif
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing ssygvdx
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gpu_time /*cpu_time*/;
float *h_A, *h_R, *h_B, *h_S, *h_work;
float *w1, *w2, vl=0, vu=0;
float result[2] = {0};
magma_int_t *iwork;
magma_int_t N, n2, info, il, iu, m1, m2, nb, lwork, liwork;
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork;
#endif
//float d_one = 1.;
//float d_ten = 10.;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
float tolulp = opts.tolerance * lapackf77_slamch("P");
if ( opts.check && opts.jobz == MagmaNoVec ) {
fprintf( stderr, "checking results requires vectors; setting jobz=V (option -JV)\n" );
opts.jobz = MagmaVec;
}
printf(" N M GPU Time (sec)\n");
printf("============================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[i];
n2 = N*N;
nb = magma_get_ssytrd_nb(N);
#if defined(PRECISION_z) || defined(PRECISION_c)
lwork = 2*N*nb + N*N;
lrwork = 1 + 5*N +2*N*N;
#else
lwork = 1 + 6*N*nb + 2* N*N;
#endif
liwork = 3 + 5*N;
if ( opts.fraction == 0 ) {
il = N / 10;
iu = N / 5+il;
}
else {
il = 1;
iu = (int) (opts.fraction*N);
if (iu < 1) iu = 1;
}
TESTING_MALLOC( h_A, float, n2 );
TESTING_MALLOC( h_B, float, n2 );
TESTING_MALLOC( w1, float, N );
TESTING_MALLOC( w2, float, N );
TESTING_MALLOC( iwork, magma_int_t, liwork );
TESTING_HOSTALLOC( h_R, float, n2 );
TESTING_HOSTALLOC( h_S, float, n2 );
TESTING_HOSTALLOC( h_work, float, lwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
TESTING_HOSTALLOC( rwork, float, lrwork);
#endif
/* Initialize the matrix */
lapackf77_slarnv( &ione, ISEED, &n2, h_A );
lapackf77_slarnv( &ione, ISEED, &n2, h_B );
/* increase the diagonal */
for(int i=0; i<N; i++) {
MAGMA_S_SET2REAL( h_B[i*N+i], ( MAGMA_S_REAL(h_B[i*N+i]) + 1.*N ) );
MAGMA_S_SET2REAL( h_A[i*N+i], MAGMA_S_REAL(h_A[i*N+i]) );
}
// ==================================================================
// Warmup using MAGMA
// ==================================================================
if(opts.warmup){
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_ssygvdx( opts.itype, opts.jobz, 'I', opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_ssygvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
gpu_time = magma_wtime();
magma_ssygvdx( opts.itype, opts.jobz, 'I', opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m1, w1,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf("magma_ssygvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
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)
(2) | S(with V) - S(w/o V) | / | S |
=================================================================== */
#if defined(PRECISION_d) || defined(PRECISION_s)
float *rwork = h_work + N*N;
#endif
float temp1, temp2;
result[0] = 1.;
result[0] /= lapackf77_slansy("1", &opts.uplo, &N, h_A, &N, rwork);
result[0] /= lapackf77_slange("1", &N, &m1, h_R, &N, rwork);
if (opts.itype == 1) {
blasf77_ssymm("L", &opts.uplo, &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_sscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_ssymm("L", &opts.uplo, &N, &m1, &c_neg_one, h_B, &N, h_R, &N, &c_one, h_work, &N);
result[0] *= lapackf77_slange("1", &N, &m1, h_work, &N, rwork)/N;
}
else if (opts.itype == 2) {
blasf77_ssymm("L", &opts.uplo, &N, &m1, &c_one, h_B, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_sscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_ssymm("L", &opts.uplo, &N, &m1, &c_one, h_A, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_slange("1", &N, &m1, h_R, &N, rwork)/N;
}
else if (opts.itype == 3) {
blasf77_ssymm("L", &opts.uplo, &N, &m1, &c_one, h_A, &N, h_R, &N, &c_zero, h_work, &N);
for(int i=0; i < m1; ++i)
blasf77_sscal(&N, &w1[i], &h_R[i*N], &ione);
blasf77_ssymm("L", &opts.uplo, &N, &m1, &c_one, h_B, &N, h_work, &N, &c_neg_one, h_R, &N);
result[0] *= lapackf77_slange("1", &N, &m1, h_R, &N, rwork)/N;
}
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_A, &N, h_R, &N );
lapackf77_slacpy( MagmaUpperLowerStr, &N, &N, h_B, &N, h_S, &N );
magma_ssygvdx( opts.itype, 'N', 'I', opts.uplo,
N, h_R, N, h_S, N, vl, vu, il, iu, &m2, w2,
h_work, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork,
&info );
if (info != 0)
printf("magma_ssygvdx returned error %d: %s.\n",
(int) info, magma_strerror( info ));
temp1 = temp2 = 0;
for(int j=0; j < m2; j++) {
temp1 = max(temp1, absv(w1[j]));
temp1 = max(temp1, absv(w2[j]));
temp2 = max(temp2, absv(w1[j]-w2[j]));
}
result[1] = temp2 / (((float)m2)*temp1);
}
/* =====================================================================
Print execution time
=================================================================== */
printf("%5d %5d %7.2f\n",
(int) N, (int) m1, gpu_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[0], (result[0] < tol ? "" : " failed"));
else if (opts.itype==2)
printf("(1) | A B Z - Z D | / (|A| |Z| N) = %8.2e%s\n", result[0], (result[0] < tol ? "" : " failed"));
else if (opts.itype==3)
printf("(1) | B A Z - Z D | / (|A| |Z| N) = %8.2e%s\n", result[0], (result[0] < tol ? "" : " failed"));
printf( "(2) | D(w/ Z) - D(w/o Z) | / |D| = %8.2e%s\n\n", result[1], (result[1] < tolulp ? "" : " failed"));
}
TESTING_FREE( h_A );
TESTING_FREE( h_B );
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 );
TESTING_HOSTFREE( h_S );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return 0;
}
int berank(const void *p1, const void *p2)
{
real e1 = 0.5 * rsqr(absv(Vel((bodyptr) p1))) + Phi((bodyptr) p1);
real e2 = 0.5 * rsqr(absv(Vel((bodyptr) p2))) + Phi((bodyptr) p2);
return (e1 < e2 ? -1 : e1 > e2 ? 1 : 0);
}
