static CALresult createLBTrigBuffers(MWCALInfo* ci,
SeparationCALMem* cm,
const AstronomyParameters* ap,
const IntegralArea* ia)
{
CALresult err = CAL_RESULT_OK;
LTrigPair* lTrig;
real* bTrig;
getSplitLBTrig(ap, ia, &lTrig, &bTrig);
err = createConstantBuffer2D(&cm->lTrig, ci, (CALdouble*) lTrig, formatReal2, CAL_FALSE, ia->nu_steps, ia->mu_steps);
if (err != CAL_RESULT_OK)
{
cal_warn("Failed to create l trig buffer", err);
goto fail;
}
err = createConstantBuffer2D(&cm->bTrig, ci, (CALdouble*) bTrig, formatReal1, CAL_FALSE, ia->nu_steps, ia->mu_steps);
if (err != CAL_RESULT_OK)
cal_warn("Failed to create b trig buffer", err);
fail:
mwFreeA(lTrig);
mwFreeA(bTrig);
return err;
}
static cl_int createLBTrigBuffer(CLInfo* ci,
SeparationCLMem* cm,
const AstronomyParameters* ap,
const IntegralArea* ia,
const SeparationSizes* sizes,
const cl_mem_flags constBufFlags)
{
cl_int err = CL_SUCCESS;
LTrigPair* lTrig = NULL;
real* bSin = NULL;
getSplitLBTrig(ap, ia, &lTrig, &bSin);
cm->lTrig = clCreateBuffer(ci->clctx, constBufFlags, sizes->lTrig, lTrig, &err);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Error creating lTrig buffer of size "ZU, sizes->lTrig);
return err;
}
cm->bSin = clCreateBuffer(ci->clctx, constBufFlags, sizes->bSin, bSin, &err);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Error creating bSin buffer of size "ZU, sizes->bSin);
return err;
}
mwFreeA(lTrig);
mwFreeA(bSin);
return CL_SUCCESS;
}
void freeStreamGauss(StreamGauss sg)
{
mwFreeA(sg.dx);
mwFreeA(sg.qgaus_W);
#ifdef ANDROID
mwFreeA(sg.dx_intfp);
#endif
}
static int worker(const SeparationFlags* sf)
{
AstronomyParameters ap;
BackgroundParameters bgp = EMPTY_BACKGROUND_PARAMETERS;
Streams streams = EMPTY_STREAMS;
IntegralArea* ias = NULL;
StreamConstants* sc = NULL;
SeparationResults* results = NULL;
int rc;
CLRequest clr;
memset(&ap, 0, sizeof(ap));
memset(&clr, 0, sizeof(clr));
setCLReqFlags(&clr, sf);
ias = prepareParameters(sf, &ap, &bgp, &streams);
if (!ias)
return 1;
rc = setAstronomyParameters(&ap, &bgp);
if (rc)
{
mwFreeA(ias);
freeStreams(&streams);
return 1;
}
setExpStreamWeights(&ap, &streams);
sc = getStreamConstants(&ap, &streams);
if (!sc)
{
mw_printf("Failed to get stream constants\n");
mwFreeA(ias);
freeStreams(&streams);
return 1;
}
results = newSeparationResults(ap.number_streams);
rc = evaluate(results, &ap, ias, &streams, sc, sf->star_points_file,
&clr, sf->do_separation, sf->ignoreCheckpoint, sf->separation_outfile);
if (rc)
mw_printf("Failed to calculate likelihood\n");
printSeparationResults(results, ap.number_streams);
mwFreeA(ias);
mwFreeA(sc);
freeStreams(&streams);
freeSeparationResults(results);
return rc;
}
void freeEvaluationState(EvaluationState* es)
{
unsigned int i;
for (i = 0; i < es->numberCuts; ++i)
freeCut(&es->cuts[i]);
mwFreeA(es->cuts);
mwFreeA(es->streamSums);
mwFreeA(es->streamTmps);
#ifdef ANDROID
mwFreeA(es->streamTmpsIntFp);
#endif
mwFreeA(es);
}
static void getSplitLBTrig(const AstronomyParameters* ap,
const IntegralArea* ia,
LTrigPair** lTrigOut,
real** bSinOut)
{
cl_uint i, j;
LTrigPair* lTrig;
real* bSin;
LBTrig* lbts;
size_t idx;
lbts = precalculateLBTrig(ap, ia, FALSE);
lTrig = (LTrigPair*) mwMallocA(ia->mu_steps * ia->nu_steps * sizeof(LTrigPair));
bSin = (real*) mwMallocA(ia->mu_steps * ia->nu_steps * sizeof(real));
for (i = 0; i < ia->nu_steps; ++i)
{
for (j = 0; j < ia->mu_steps; ++j)
{
idx = i * ia->mu_steps + j;
lTrig[idx].lCosBCos = lbts[idx].lCosBCos;
lTrig[idx].lSinBCos = lbts[idx].lSinBCos;
bSin[idx] = lbts[idx].bSin;
}
}
mwFreeA(lbts);
*lTrigOut = lTrig;
*bSinOut = bSin;
}
StreamGauss getStreamGauss(const unsigned int convolve)
{
unsigned int i;
StreamGauss sg;
real* qgaus_X;
qgaus_X = (real*) mwMallocA(sizeof(real) * convolve);
sg.qgaus_W = (real*) mwMallocA(sizeof(real) * convolve);
gaussLegendre(-1.0, 1.0, qgaus_X, sg.qgaus_W, convolve);
sg.dx = (real*) mwMallocA(sizeof(real) * convolve);
for (i = 0; i < convolve; ++i)
sg.dx[i] = 3.0 * stdev * qgaus_X[i];
mwFreeA(qgaus_X);
#ifdef ANDROID
sg.dx_intfp = (IntFp*) mwMallocA(sizeof(IntFp) * convolve);
for (int i=0; i < convolve; i++)
fp_to_intfp(sg.dx[i],&sg.dx_intfp[i]);
#endif
return sg;
}
static IntegralArea* prepareParameters(const SeparationFlags* sf,
AstronomyParameters* ap,
BackgroundParameters* bgp,
Streams* streams)
{
IntegralArea* ias;
ias = setupSeparation(ap, bgp, streams, sf);
/* Try the new file first. If that doesn't work, try the old one. */
if (!ias)
{
mw_printf("Error reading astronomy parameters from file '%s'\n"
" Trying old parameters file\n", sf->ap_file);
ias = readParameters(sf->ap_file, ap, bgp, streams);
}
if (!ias)
{
mw_printf("Failed to read parameters file\n");
return NULL;
}
if (sf->numArgs && setParameters(ap, bgp, streams, sf->numArgs, sf->nForwardedArgs))
{
mwFreeA(ias);
freeStreams(streams);
return NULL;
}
return ias;
}
StreamConstants* getStreamConstants(const AstronomyParameters* ap, const Streams* streams)
{
int i;
StreamConstants* sc;
real stream_sigma;
real sigma_sq2;
sc = (StreamConstants*) mwMallocA(streams->number_streams * sizeof(StreamConstants));
for (i = 0; i < streams->number_streams; ++i)
{
stream_sigma = streams->parameters[i].sigma;
if (stream_sigma == 0.0)
{
mw_printf("stream sigma 0.0 is invalid\n");
mwFreeA(sc);
return NULL;
}
sc[i].large_sigma = (stream_sigma > SIGMA_LIMIT || stream_sigma < -SIGMA_LIMIT);
sigma_sq2 = 2.0 * sqr(stream_sigma);
sc[i].sigma_sq2_inv = 1.0 / sigma_sq2;
sc[i].a = streamA(&streams->parameters[i]);
sc[i].c = streamC(ap,
ap->wedge,
streams->parameters[i].mu,
streams->parameters[i].r);
}
return sc;
}
/* Might be more convenient to split l and b stuff for CAL */
static void getSplitLBTrig(const AstronomyParameters* ap,
const IntegralArea* ia,
LTrigPair** lTrigBCosOut,
real** bTrigOut)
{
CALuint i, j;
LTrigPair* lTrigBCos;
real* bTrig;
LBTrig* lbts;
size_t idx;
CALboolean transpose = CAL_TRUE;
lTrigBCos = (LTrigPair*) mwMallocA(ia->mu_steps * ia->nu_steps * sizeof(LTrigPair));
bTrig = (real*) mwMallocA(ia->mu_steps * ia->nu_steps * sizeof(real));
lbts = precalculateLBTrig(ap, ia, transpose);
for (i = 0; i < ia->nu_steps; ++i)
{
for (j = 0; j < ia->mu_steps; ++j)
{
idx = transpose ? j * ia->nu_steps + i : i * ia->mu_steps + j;
lTrigBCos[idx].lCosBCos = lbts[idx].lCosBCos;
lTrigBCos[idx].lSinBCos = lbts[idx].lSinBCos;
bTrig[idx] = lbts[idx].bSin;
}
}
mwFreeA(lbts);
*lTrigBCosOut = lTrigBCos;
*bTrigOut = bTrig;
}
StreamGauss getStreamGauss(int convolve)
{
int i;
StreamGauss sg;
real* qgaus_X;
qgaus_X = (real*) mwMallocA(sizeof(real) * convolve);
sg.qgaus_W = (real*) mwMallocA(sizeof(real) * convolve);
gaussLegendre(-1.0, 1.0, qgaus_X, sg.qgaus_W, convolve);
sg.dx = (real*) mwMallocA(sizeof(real) * convolve);
/*Using old (single-sided gaussian stdev = 0.6) to spread points. This is a small simplification when using
modfit, but does not cause any problems since it is parameter independent. The weights will be calculated
later based on the two-sided gaussian.*/
for (i = 0; i < convolve; ++i)
{
sg.dx[i] = 3.0 * stdev * qgaus_X[i];
}
mwFreeA(qgaus_X);
return sg;
}
/* It will be easier to make this less chaotic in the next release
* when we can dump the old parameters file */
IntegralArea* setupSeparation(AstronomyParameters* ap,
BackgroundParameters* bg,
Streams* streams,
const SeparationFlags* sf)
{
int rc = 0;
lua_State* luaSt;
luaSt = separationOpenLuaStateWithScript(sf);
if (!luaSt)
return NULL;
_ap = ap;
_bg = bg;
_streams = streams;
rc |= evaluateGlobalName(luaSt, evaluateConstants, CONSTANTS_NAME);
rc |= evaluateGlobalName(luaSt, evaluateIntegralAreas, AREAS_NAME);
rc |= evaluateGlobalName(luaSt, evaluateBackground, BACKGROUND_NAME);
rc |= evaluateGlobalName(luaSt, evaluateStreams, STREAMS_NAME);
lua_close(luaSt);
if (rc)
{
free(_streams->parameters);
mwFreeA(_ias);
return NULL;
}
setAPConstants(ap);
return _ias;
}
static cl_int createRBuffers(CLInfo* ci,
SeparationCLMem* cm,
const AstronomyParameters* ap,
const IntegralArea* ia,
const StreamGauss sg,
const SeparationSizes* sizes,
cl_mem_flags constBufFlags)
{
cl_int err;
RPoints* r_pts;
RConsts* rc;
r_pts = precalculateRPts(ap, ia, sg, &rc, FALSE);
cm->rPts = clCreateBuffer(ci->clctx, constBufFlags, sizes->rPts, r_pts, &err);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Error creating stream r points buffer of size "ZU, sizes->rPts);
return err;
}
cm->rc = clCreateBuffer(ci->clctx, constBufFlags, sizes->rc, rc, &err);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Error creating stream r consts buffer of size "ZU, sizes->rc);
return err;
}
cm->sg_dx = clCreateBuffer(ci->clctx, constBufFlags, sizes->sg_dx, sg.dx, &err);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Error creating stream sg_dx buffer of size "ZU, sizes->sg_dx);
return err;
}
mwFreeA(r_pts);
mwFreeA(rc);
return CL_SUCCESS;
}
static void freeFreeCells(NBodyNode* freeCell)
{
NBodyNode* p;
NBodyNode* tmp;
p = freeCell;
while (p)
{
tmp = Next(p);
mwFreeA(p);
p = tmp;
}
}
static CALresult createRBuffers(MWCALInfo* ci,
SeparationCALMem* cm,
const AstronomyParameters* ap,
const IntegralArea* ia,
const StreamGauss sg)
{
RPoints* r_pts;
RConsts* rc;
CALresult err = CAL_RESULT_OK;
r_pts = precalculateRPts(ap, ia, sg, &rc, TRUE);
err = createConstantBuffer2D(&cm->rPts, ci, (CALdouble*) r_pts, formatReal2, CAL_FALSE, ia->r_steps, ap->convolve);
if (err != CAL_RESULT_OK)
{
cal_warn("Failed to create r_pts buffer", err);
goto fail;
}
err = createConstantBuffer1D(&cm->rc, ci, (CALdouble*) rc, formatReal2, ia->r_steps);
if (err != CAL_RESULT_OK)
{
cal_warn("Failed to create rc buffer", err);
goto fail;
}
err = createConstantBuffer1D(&cm->sg_dx, ci, sg.dx, constantFormatReal1, ap->convolve);
if (err != CAL_RESULT_OK)
cal_warn("Failed to create sg_dx buffer", err);
fail:
mwFreeA(r_pts);
mwFreeA(rc);
return err;
}
static cl_int createSCBuffer(CLInfo* ci,
SeparationCLMem* cm,
const StreamConstants* sc,
const SeparationSizes* sizes,
const cl_mem_flags constBufFlags)
{
cl_int err;
real* buf;
cl_int i;
buf = mwCallocA(sizes->nStream * 8, sizeof(real));
/* Pack into format used by kernel */
for (i = 0; i < sizes->nStream; ++i)
{
buf[8 * i + 0] = X(sc[i].a);
buf[8 * i + 1] = X(sc[i].c);
buf[8 * i + 2] = Y(sc[i].a);
buf[8 * i + 3] = Y(sc[i].c);
buf[8 * i + 4] = Z(sc[i].a);
buf[8 * i + 5] = Z(sc[i].c);
buf[8 * i + 6] = sc[i].sigma_sq2_inv;
buf[8 * i + 7] = 0.0;
}
cm->sc = clCreateBuffer(ci->clctx, constBufFlags, sizes->sc, (void*) buf, &err);
mwFreeA(buf);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Error creating stream constants buffer of size "ZU, sizes->sc);
return err;
}
return CL_SUCCESS;
}
static void freeNBodyTree(NBodyTree* t)
{
NBodyNode* p;
NBodyNode* tmp;
p = (NBodyNode*) t->root;
while (p != NULL)
{
if (isCell(p))
{
tmp = More(p);
mwFreeA(p);
p = tmp;
}
else /* skip over bodies */
{
p = Next(p);
}
}
t->root = NULL;
t->cellUsed = 0;
t->maxDepth = 0;
}
int destroyNBodyState(NBodyState* st)
{
int failed = FALSE;
int nThread = nbGetMaxThreads();
int i;
freeNBodyTree(&st->tree);
freeFreeCells(st->freeCell);
mwFreeA(st->bodytab);
mwFreeA(st->acctab);
mwFreeA(st->orbitTrace);
free(st->checkpointResolved);
if (st->potEvalStates)
{
for (i = 0; i < nThread; ++i)
{
lua_close(st->potEvalStates[i]);
}
free(st->potEvalClosures);
free(st->potEvalStates);
}
#if NBODY_OPENCL
if (st->ci)
{
mwDestroyCLInfo(st->ci);
free(st->ci);
st->ci = NULL;
}
if (st->kernels)
{
cl_int err;
err = nbReleaseKernels(st);
free(st->kernels);
failed |= (err != CL_SUCCESS);
}
if (st->workSizes)
{
free(st->workSizes);
}
if (st->nbb)
{
cl_int err;
err = nbReleaseBuffers(st);
free(st->nbb);
st->nbb = NULL;
failed |= (err != CL_SUCCESS);
}
#endif /* NBODY_OPENCL */
failed |= nbDetachSharedScene(st);
return failed;
}
int evaluate(SeparationResults* results,
const AstronomyParameters* ap,
const IntegralArea* ias,
const Streams* streams,
const StreamConstants* sc,
const char* star_points_file,
const CLRequest* clr,
int do_separation,
int ignoreCheckpoint,
const char* separation_outfile)
{
int rc = 0;
EvaluationState* es;
StreamGauss sg;
GPUInfo ci;
StarPoints sp = EMPTY_STAR_POINTS;
int useImages = FALSE; /* Only applies to CL version */
#ifdef ANDROID
StreamConstantsIntFp* sci;
int armExt = mwDetectARMExt();
#endif
memset(&ci, 0, sizeof(ci));
probabilityFunctionDispatch(ap, clr);
es = newEvaluationState(ap);
sg = getStreamGauss(ap->convolve);
#if SEPARATION_GRAPHICS
if (separationInitSharedEvaluationState(es))
warn("Failed to initialize shared evaluation state\n");
#endif /* SEPARATION_GRAPHICS */
if (!ignoreCheckpoint)
{
if (resolveCheckpoint())
fail("Failed to resolve checkpoint file '%s'\n", CHECKPOINT_FILE);
if (maybeResume(es))
fail("Failed to resume checkpoint\n");
}
#if SEPARATION_OPENCL
if (setupSeparationCL(&ci, ap, ias, clr, &useImages) != CL_SUCCESS)
fail("Failed to setup CL\n");
#elif SEPARATION_CAL
if (separationCALInit(&ci, clr) != CAL_RESULT_OK)
fail("Failed to setup CAL\n");
#endif
#ifdef ANDROID
if (armExt==ARM_CPU_NOVFP && ap->fast_h_prob)
{
int i=0;
unsigned int nstreams = ap->number_streams;
unsigned int convolve = ap->convolve;
warn("Use IntFp Engine\n");
sci = (StreamConstantsIntFp*) mwMallocA(sizeof(StreamConstantsIntFp) * ap->number_streams);
for (i=0; i < nstreams; i++)
{
fp_to_intfp(sc[i].a.x,&(sci[i].a[0]));
fp_to_intfp(sc[i].a.y,&sci[i].a[1]);
fp_to_intfp(sc[i].a.z,&sci[i].a[2]);
fp_to_intfp(sc[i].a.w,&sci[i].a[3]);
fp_to_intfp(-sc[i].c.x,&sci[i].c[0]);
fp_to_intfp(-sc[i].c.y,&sci[i].c[1]);
fp_to_intfp(-sc[i].c.z,&sci[i].c[2]);
fp_to_intfp(-sc[i].c.w,&sci[i].c[3]);
fp_to_intfp(sc[i].sigma_sq2_inv,&sci[i].sigma_sq2_inv);
}
}
#endif
#ifdef ANDROID
if (armExt == ARM_CPU_NOVFP && ap->fast_h_prob)
calculateIntegralsIntFp(ap, ias, sci, sg, es, clr, &ci, useImages);
else
calculateIntegrals(ap, ias, sc, sg, es, clr, &ci, useImages);
#else
calculateIntegrals(ap, ias, sc, sg, es, clr, &ci, useImages);
#endif
if (!ignoreCheckpoint)
{
finalCheckpoint(es);
}
getFinalIntegrals(results, es, ap->number_streams, ap->number_integrals);
freeEvaluationState(es);
if (readStarPoints(&sp, star_points_file))
{
rc = 1;
warn("Failed to read star points file\n");
}
else
{
/* TODO: likelihood on GPU with OpenCL. Make this less of a
* mess. The different versions should appear to be the
* same. */
#if SEPARATION_CAL
if (do_separation)
{
/* No separation on GPU */
rc = likelihood(results, ap, &sp, sc, streams, sg, do_separation, separation_outfile);
}
else
{
//rc = likelihoodCAL(results, ap, &sp, sc, streams, sg, clr, &ci);
rc = likelihood(results, ap, &sp, sc, streams, sg, do_separation, separation_outfile);
}
#else
#ifdef ANDROID
if (armExt == ARM_CPU_NOVFP && ap->fast_h_prob)
rc = likelihood_intfp(results, ap, &sp, sci, streams, sg, do_separation, separation_outfile);
else
rc = likelihood(results, ap, &sp, sc, streams, sg, do_separation, separation_outfile);
#else
rc = likelihood(results, ap, &sp, sc, streams, sg, do_separation, separation_outfile);
#endif
#endif /* SEPARATION_CAL */
rc |= checkSeparationResults(results, ap->number_streams);
}
freeStarPoints(&sp);
freeStreamGauss(sg);
#if SEPARATION_OPENCL
mwDestroyCLInfo(&ci);
#elif SEPARATION_CAL
mwCALShutdown(&ci);
#endif
#ifdef ANDROID
if (armExt == ARM_CPU_NOVFP && ap->fast_h_prob)
mwFreeA(sci);
#endif
return rc;
}
/* glut main loop never quits, so actually freeing these is a bad idea */
void nbodyGLCleanup()
{
mwFreeA(color);
color = NULL;
}
//Needs to loop to account for number of WUs being crunched
static int worker(const SeparationFlags* sf)
{
AstronomyParameters ap;
BackgroundParameters bgp = EMPTY_BACKGROUND_PARAMETERS;
Streams streams = EMPTY_STREAMS;
IntegralArea* ias = NULL;
StreamConstants* sc = NULL;
SeparationResults* results = NULL;
int rc;
CLRequest clr;
memset(&ap, 0, sizeof(ap));
memset(&clr, 0, sizeof(clr));
ap.modfit = sf->modfit;
if(sf->background)
{
ap.background_profile = BROKEN_POWER_LAW;
}
else
{
ap.background_profile = FAST_HERNQUIST;
}
setCLReqFlags(&clr, sf);
/*Assume we are crunching at least 1 work unit (These numbers will be properly set in prepareParameters when the parameter file is read)*/
ias = prepareParameters(sf, &ap, &bgp, &streams);
if (!ias)
return 1;
if(sf->nForwardedArgs)
{
ap.totalWUs = sf->nForwardedArgs/ap.params_per_workunit;
}
else
{
ap.totalWUs = 1;
}
mw_printf("<number_WUs> %d </number_WUs>\n", ap.totalWUs);
mw_printf("<number_params_per_WU> %d </number_params_per_WU>\n", ap.params_per_workunit);
int ignoreCheckpoint = sf->ignoreCheckpoint;
for(ap.currentWU = 0; ap.currentWU < ap.totalWUs; ap.currentWU++)
{
if (sf->numArgs && setParameters(&ap, &bgp, &streams, &(sf->numArgs[ap.params_per_workunit * ap.currentWU]), ap.params_per_workunit))
{
mwFreeA(ias);
freeStreams(&streams);
return 1;
}
rc = setAstronomyParameters(&ap, &bgp);
if (rc)
{
mwFreeA(ias);
freeStreams(&streams);
return 1;
}
setExpStreamWeights(&ap, &streams);
sc = getStreamConstants(&ap, &streams);
if (!sc)
{
mw_printf("Failed to get stream constants\n");
mwFreeA(ias);
freeStreams(&streams);
return 1;
}
results = newSeparationResults(ap.number_streams);
int currentWU = ap.currentWU;
rc = evaluate(results, &ap, ias, &streams, sc, sf->LikelihoodToText, sf->star_points_file,
&clr, sf->do_separation, &ignoreCheckpoint, sf->separation_outfile);
if (rc)
mw_printf("Failed to calculate likelihood\n");
}
mwFreeA(ias);
mwFreeA(sc);
freeStreams(&streams);
if(results) freeSeparationResults(results);
return rc;
}
static void freeCut(Cut* i)
{
mwFreeA(i->streamIntegrals);
}
void freeStreamGauss(StreamGauss sg)
{
mwFreeA(sg.dx);
mwFreeA(sg.qgaus_W);
}
void freeRunSizes(RunSizes* sizes)
{
mwFreeA(sizes->chunkBorders);
}