cl_double mwDeviceEstimateGFLOPs(const DevInfo* di, cl_bool useDouble)
{
cl_double gflops = 0.0;
if (di->devType == CL_DEVICE_TYPE_GPU)
{
if (mwIsNvidiaGPUDevice(di))
{
gflops = mwCUDAEstimateGFLOPs(di, useDouble);
}
else if (mwIsAMDGPUDevice(di))
{
gflops = mwAMDEstimateGFLOPs(di, useDouble);
}
else
{
mw_printf("Unhandled GPU vendor '%s' (0x%x)\n", di->vendor, di->vendorID);
gflops = 100.0;
}
}
else
{
mw_printf("Missing flops estimate for device type %s\n", showCLDeviceType(di->devType));
return 1.0;
}
return gflops;
}
/* Query one device specified by type, create a context and command
* queue, as well as retrieve device information
*/
cl_int mwSetupCL(CLInfo* ci, const CLRequest* clr)
{
cl_int err;
err = mwGetCLInfo(ci, clr);
if (err != CL_SUCCESS)
{
mw_printf("Failed to get information about device\n");
return err;
}
err = mwGetDevInfo(&ci->di, ci->dev);
if (err != CL_SUCCESS)
{
mw_printf("Failed to get device info\n");
return err;
}
if (mwIsFirstRun())
{
if (clr->verbose)
{
mwPrintDevInfo(&ci->di);
}
else
{
mwPrintDevInfoShort(&ci->di);
}
}
return mwCreateCtxQueue(ci, CL_FALSE, clr->enableProfiling);
}
void nbReportTreeIncest(const NBodyCtx* ctx, NBodyState* st)
{
if (!st->treeIncest) /* don't repeat warning */
{
st->treeIncest = TRUE;
if (!ctx->quietErrors) /* Avoid massive printing of tests causing incest */
{
if (ctx->allowIncest)
{
mw_printf("[tree-incest detected at step %u / %u (%f%%)]\n",
st->step,
ctx->nStep,
100.0 * (real) st->step / (real) ctx->nStep
);
}
else
{
mw_printf("tree-incest detected (fatal) at step %u / %u (%f%%)\n",
st->step,
ctx->nStep,
100.0 * (real) st->step / (real) ctx->nStep
);
}
}
}
}
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;
}
int equalBody(const Body* a, const Body* b)
{
if (Mass(a) != Mass(b))
{
mw_printf("mass differ\n");
return FALSE;
}
if (Type(a) != Type(b))
{
mw_printf("type ndiffer\n");
return FALSE;
}
if (!equalVector(&Pos(a), &Pos(b)))
{
mw_printf("pos differ\n");
return FALSE;
}
if (!equalVector(&Vel(a), &Vel(b)))
{
mw_printf("VElocity differ\n");
return FALSE;
}
return TRUE;
}
static int nbglGetExclusiveSceneAccess(scene_t* scene)
{
int pid = (int) getpid();
int oldPID = OPA_cas_int(&scene->attachedLock, 0, pid);
if (oldPID != 0)
{
if (mwProcessIsAlive(oldPID))
{
mw_printf("Could not get exclusive access to simulation shared segment "
"(Owned by process %d)\n",
oldPID);
return 1;
}
else
{
mw_printf("Simulation shared segment owned by dead process %d, stealing it\n",
oldPID);
/* Process is dead, steal the lock */
nbglReleaseSceneLocks(scene);
return 0;
}
}
else
{
OPA_store_int(&scene->attachedPID, 0);
return 0;
}
}
static void nbPrintVersion(int boincTag, int verbose)
{
char versionStr[2048];
snprintf(versionStr, sizeof(versionStr),
"%s %u.%02u %s %s %s %s %s, %s",
NBODY_PROJECT_NAME,
NBODY_VERSION_MAJOR, NBODY_VERSION_MINOR,
MILKYWAY_SYSTEM_NAME,
ARCH_STRING,
PRECSTRING,
DENORMAL_STRING,
NBODY_EXTRAVER,
NBODY_EXTRALIB);
if (boincTag)
{
mw_printf("<search_application> %s </search_application>\n", versionStr);
}
else
{
mw_printf("%s %s\n",
versionStr,
BOINC_APPLICATION ? "BOINC" : "");
}
if (verbose)
{
mw_printf("Commit %s\n", MILKYWAY_GIT_COMMIT_ID);
}
}
cl_double mwAMDEstimateGFLOPs(const DevInfo* di, cl_bool useDouble)
{
cl_ulong flops, flopsFloat, flopsDouble;
cl_double gflops;
flopsFloat = 2 * (di->maxCompUnits * di->aluPerCU) * (cl_ulong) di->clockFreq * 1000000;
flopsDouble = flopsFloat / di->doubleFrac;
mw_printf("Estimated AMD GPU GFLOP/s: %.0f SP GFLOP/s, %.0f DP FLOP/s\n",
1.0e-9 * (cl_double) flopsFloat,
1.0e-9 * (cl_double) flopsDouble);
flops = useDouble ? flopsDouble : flopsFloat;
gflops = floor(1.0e-9 * (cl_double) flops);
/* At different times the AMD drivers have reported 0 as the clock
* speed, so try to catch that. We could test the GPU and figure
* out what the FLOPs should be to get a better estimate.
*/
if (gflops <= 100.0)
{
mw_printf("Warning: Bizarrely low flops (%.0f). Defaulting to %.0f\n", gflops, 100.0);
gflops = 100.0;
}
return gflops;
}
static int nbSetNumThreads(int numThreads)
{
#ifdef _OPENMP
int nProc = omp_get_num_procs();
int nBoinc = mwGetBoincNumCPU();
if (nProc <= 0) /* It's happened before... */
{
mw_printf("Number of processors %d is crazy\n", nProc);
return 1;
}
/* If command line argument not given, and BOINC gives us a value use that */
if (numThreads <= 0 && nBoinc > 0)
{
numThreads = nBoinc;
}
if (numThreads != 0)
{
omp_set_num_threads(numThreads);
mw_printf("Using OpenMP %d max threads on a system with %d processors\n",
omp_get_max_threads(),
nProc);
}
#endif
return 0;
}
/* If possible, resume from a checkpoint. Otherwise do the necessary
* initialization for a new run */
static NBodyStatus nbResumeOrNewRun(NBodyCtx* ctx, NBodyState* st, const NBodyFlags* nbf)
{
if (nbResolveCheckpoint(st, nbf->checkpointFileName))
{
mw_printf("Failed to resolve checkpoint\n");
return NBODY_ERROR;
}
/* If the checkpoint exists (and we want to use it), try to use it */
if (nbf->ignoreCheckpoint || !nbResolvedCheckpointExists(st))
{
if (!nbf->inputFile)
{
mw_printf("No input file and no checkpoint\n");
return NBODY_USER_ERROR;
}
if (nbSetup(ctx, st, nbf))
{
mw_printf("Failed to read input parameters file\n");
return NBODY_PARAM_FILE_ERROR;
}
}
else /* Resume from checkpoint */
{
if (nbf->inputFile && !BOINC_APPLICATION)
{
mw_printf("Warning: input file '%s' unused\n", nbf->inputFile);
}
if (nbReadCheckpoint(ctx, st))
{
mw_report("Failed to read checkpoint\n");
return NBODY_CHECKPOINT_ERROR;
}
else
{
mw_report("Resumed from checkpoint '%s'\n", nbf->checkpointFileName);
}
}
if (ctx->potentialType == EXTERNAL_POTENTIAL_CUSTOM_LUA)
{
/* We're using a custom potential, so we'll reevaluate the
* script. We must do this once per thread.
*/
if (nbOpenPotentialEvalStatePerThread(st, nbf))
{
return NBODY_PARAM_FILE_ERROR;
}
}
return NBODY_SUCCESS;
}
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;
}
static cl_int runIntegral(CLInfo* ci,
SeparationCLMem* cm,
RunSizes* runSizes,
EvaluationState* es,
const CLRequest* clr,
const AstronomyParameters* ap,
const IntegralArea* ia)
{
cl_int err = CL_SUCCESS;
double t1, t2, dt;
double tAcc = 0.0;
for (; es->nu_step < ia->nu_steps; es->nu_step++)
{
if (clr->enableCheckpointing && timeToCheckpointGPU(es, ia))
{
err = checkpointCL(ci, cm, ia, es);
if (err != CL_SUCCESS)
break;
}
t1 = mwGetTimeMilli();
err = runNuStep(ci, ia, runSizes, es->nu_step);
if (err != CL_SUCCESS)
{
mwPerrorCL(err, "Failed to run nu step");
return err;
}
t2 = mwGetTimeMilli();
dt = t2 - t1;
tAcc += dt;
reportProgress(ap, ia, es, es->nu_step + 1, dt);
}
es->nu_step = 0;
mw_printf("Integration time: %f s. Average time per iteration = %f ms\n",
tAcc / 1000.0, tAcc / (double) ia->nu_steps);
if (err == CL_SUCCESS)
{
err = readKernelResults(ci, cm, es, ia);
if (err != CL_SUCCESS)
mw_printf("Failed to read final kernel results\n");
/* Add final episode to running totals */
addTmpCheckpointSums(es);
}
return err;
}
/* Set any flags based on project preferences that weren't specified
* on the command line.
*
* This is a bit convoluted since we need to boinc_init before we read
* the preferences, but that needs to be delayed until after argument
* reading in case we want to disable output redirection, and then we
* still want the command line to supersede the project prefs / the
* device specified by app_init_data
*/
static void setFlagsFromPreferences(SeparationFlags* flags, const SeparationPrefs* prefs, const char* progName)
{
if (flags->useDevNumber < 0)
{
/* Try to use BOINC's suggestion from app_init_data stuff;
We might not get it so just use the first device. */
flags->useDevNumber = mwGetBoincOpenCLDeviceIndex();
if (flags->useDevNumber < 0)
{
flags->useDevNumber = 0;
}
}
if (!flags->preferredPlatformVendor)
{
const char* vendor = mwGetBoincOpenCLPlatformVendor();
if (vendor)
{
mw_printf("BOINC GPU type suggests using OpenCL vendor '%s'\n", vendor);
}
else
{
/* If BOINC doesn't tell us, guess based on the binary name */
vendor = mwGuessPreferredPlatform(progName);
if (vendor)
{
mw_printf("Guessing preferred OpenCL vendor '%s'\n", vendor);
}
}
flags->preferredPlatformVendor = vendor ? strdup(vendor) : NULL;
}
if (flags->targetFrequency <= 0.0)
{
flags->targetFrequency = prefs->gpuTargetFrequency;
}
if (flags->nonResponsive < 0)
{
flags->nonResponsive = prefs->gpuNonResponsive;
}
if (flags->processPriority == MW_PRIORITY_INVALID)
{
/* For GPU versions, default to using a higher process priority if not set */
if (SEPARATION_OPENCL && !flags->forceNoOpenCL)
{
flags->processPriority = prefs->gpuProcessPriority;
}
}
}
static int readIntegralArea(lua_State* luaSt, IntegralArea* iaOut, int table)
{
uint64_t r, mu, nu;
static IntegralArea ia;
static real nuStepsf, muStepsf, rStepsf;
static const MWNamedArg iaArgTable[] =
{
{ "nu_min", LUA_TNUMBER, NULL, TRUE, &ia.nu_min },
{ "nu_max", LUA_TNUMBER, NULL, TRUE, &ia.nu_max },
{ "nu_steps", LUA_TNUMBER, NULL, TRUE, &nuStepsf },
{ "mu_min", LUA_TNUMBER, NULL, TRUE, &ia.mu_min },
{ "mu_max", LUA_TNUMBER, NULL, TRUE, &ia.mu_max },
{ "mu_steps", LUA_TNUMBER, NULL, TRUE, &muStepsf },
{ "r_min", LUA_TNUMBER, NULL, TRUE, &ia.r_min },
{ "r_max", LUA_TNUMBER, NULL, TRUE, &ia.r_max },
{ "r_steps", LUA_TNUMBER, NULL, TRUE, &rStepsf },
END_MW_NAMED_ARG
};
handleNamedArgumentTable(luaSt, iaArgTable, table);
ia.nu_steps = (unsigned int) nuStepsf;
ia.mu_steps = (unsigned int) muStepsf;
ia.r_steps = (unsigned int) rStepsf;
r = (uint64_t) ia.r_steps;
mu = (uint64_t) ia.mu_steps;
nu = (uint64_t) ia.nu_steps;
if (nu == 0 || mu == 0 || r == 0)
{
mw_printf("Integral size { %u, %u, %u } cannot be 0\n", nu, mu, r);
return 1;
}
if ((r > UINT64_MAX / mu) || ((r * mu) > UINT64_MAX / nu))
{
mw_printf("Integral size { %u, %u, %u } will overflow progress calculation\n",
ia.nu_steps, ia.mu_steps, ia.r_steps);
return 1;
}
calcIntegralStepSizes(&ia);
*iaOut = ia;
return 0;
}
static int emdIterateSolution(EMDState* state)
{
int result;
float min_delta;
float eps = EMD_EPS * state->max_cost;
/* if ssize = 1 or dsize = 1 then we are done, else ... */
if (state->ssize > 1 && state->dsize > 1)
{
int itr;
for (itr = 1; itr < MAX_ITERATIONS; itr++)
{
/* find basic variables */
result = emdFindBasicVariables(state->cost, state->is_x,
state->u, state->v, state->ssize, state->dsize);
if (result < 0)
{
break;
}
/* check for optimality */
min_delta = emdIsOptimal(state->cost, state->is_x,
state->u, state->v,
state->ssize, state->dsize, state->enter_x);
if (min_delta == EMD_INF)
{
mw_printf("Iteration didn't converge");
return 1;
}
/* if no negative deltamin, we found the optimal solution */
if (min_delta >= -eps)
{
break;
}
/* improve solution */
if (!emdNewSolution(state))
{
mw_printf("Iteration didn't converge");
return 1;
}
}
}
return 0;
}
int pushType(lua_State* luaSt, const char* typeName, size_t typeSize, void* p)
{
void* lp;
lp = lua_newuserdata(luaSt, typeSize);
if (!lp)
{
mw_printf("Creating userdata '%s' failed\n", typeName);
return 0;
}
assert((uintptr_t) lp % MW_LUA_ALIGN == 0); /* This must be true for dSFMT intrinsics stuff to work */
luaL_getmetatable(luaSt, typeName);
lua_setmetatable(luaSt, -2);
#if 0
/* Give this object a new function environment; for installing
* arbitrary lua functions into a type */
lua_newtable(luaSt);
lua_setfenv(luaSt, -2);
#endif
memcpy(lp, p, typeSize);
return 1;
}
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;
}
/*Functions for reading parameters from files */
real* fread_double_array(FILE* file, const char* array_name, unsigned int* sizeOut)
{
unsigned int i, size;
int rc;
real* arr;
fscanf(file, array_name);
fscanf(file, "[%u]: ", &size);
arr = (real*) mwMalloc(sizeof(real) * size);
for (i = 0; i < size; i++)
{
rc = fscanf(file, READ_DOUBLE_ARRAY_READ_STR, &arr[i]);
if (rc != 1)
{
mw_printf("Error reading into %s\n", array_name);
free(arr);
return NULL;
}
if (i < size - 1)
fscanf(file, ", ");
}
fscanf(file, "\n");
if (sizeOut)
*sizeOut = size;
return arr;
}
/* If using BOINC try reading a few of the settings from the project
* preferences. If command line arguments are used, those will
* override the preferences. The command line arguments will also
* still work without BOINC */
static void separationReadPreferences(SeparationPrefs* prefsOut)
{
static SeparationPrefs prefs;
static MWProjectPrefs sepPrefs[] =
{
{ "gpu_target_frequency", MW_PREF_DOUBLE, FALSE, &prefs.gpuTargetFrequency },
{ "gpu_wait_factor", MW_PREF_DOUBLE, FALSE, &prefs.gpuWaitFactor },
{ "gpu_non_responsive", MW_PREF_BOOL, FALSE, &prefs.gpuNonResponsive },
{ "gpu_process_priority", MW_PREF_INT, FALSE, &prefs.gpuProcessPriority },
{ "no_gpu_checkpoint", MW_PREF_BOOL, FALSE, &prefs.gpuDisableCheckpoint },
END_MW_PROJECT_PREFS
};
prefs.gpuTargetFrequency = DEFAULT_TARGET_FREQUENCY;
prefs.gpuWaitFactor = DEFAULT_WAIT_FACTOR;
prefs.gpuNonResponsive = DEFAULT_NON_RESPONSIVE;
prefs.gpuProcessPriority = DEFAULT_GPU_PRIORITY;
prefs.gpuDisableCheckpoint = DEFAULT_DISABLE_GPU_CHECKPOINTING;
if (BOINC_APPLICATION)
{
if (mwGetAppInitData())
{
mw_printf("Error reading app init data. Project preferences will not be used\n");
}
else
{
mwReadProjectPrefs(sepPrefs, mwGetProjectPrefs());
}
}
*prefsOut = prefs;
}
static cl_int runNuStep(CLInfo* ci, const IntegralArea* ia, const RunSizes* runSizes, cl_uint nu_step)
{
cl_uint i;
cl_int err = CL_SUCCESS;
size_t offset[1];
err = setNuKernelArgs(ia, nu_step);
if (err != CL_SUCCESS)
{
mw_printf("Failed to set nu kernel argument\n");
return err;
}
mw_begin_critical_section();
offset[0] = 0;
for (i = 0; i < runSizes->nChunk && err == CL_SUCCESS; ++i)
{
err = runIntegralKernel(ci, runSizes, offset);
checkQuitRequest(); /* Kernel has finished by now */
offset[0] += runSizes->global[0];
}
mw_end_critical_section();
return err;
}
void mwPrintDevInfoShort(const DevInfo* di)
{
mw_printf("Device '%s' (%s:0x%x) (%s)\n"
"Driver version: %s\n"
"Version: %s\n"
"Compute capability: %u.%u\n"
"Max compute units: %u\n"
"Clock frequency: %u Mhz\n"
"Global mem size: "LLU"\n"
"Local mem size: "LLU"\n"
"Max const buf size: "LLU"\n"
"Double extension: %s\n",
di->devName,
di->vendor, di->vendorID,
showCLDeviceType(di->devType),
di->driver,
di->version,
di->computeCapabilityMajor, di->computeCapabilityMinor,
di->maxCompUnits,
di->clockFreq,
di->memSize,
di->localMemSize,
di->maxConstBufSize,
showMWDoubleExts(di->doubleExts)
);
}
int* fread_int_array(FILE *file, const char *array_name, unsigned int* sizeOut)
{
unsigned int i, size;
int* arr;
fscanf(file, array_name);
fscanf(file, "[%u]: ", &size);
arr = (int*) mwMalloc(sizeof(int) * size);
for (i = 0; i < size; i++)
{
if (fscanf(file, "%d", &arr[i]) != 1)
{
mw_printf("Error reading into %s\n", array_name);
free(arr);
return NULL;
}
if (i < size - 1)
fscanf(file, ", ");
}
fscanf(file, "\n");
if (sizeOut)
*sizeOut = size;
return arr;
}
/* The transactional stuff is only available on Vista and later */
static void initW32TransactionalFunctions()
{
HMODULE ktm32Lib;
HMODULE kernel32Lib;
transactionFuncsInit = TRUE;
kernel32Lib = LoadLibrary("Kernel32.dll");
if (!kernel32Lib)
{
mwPerrorW32("Could not load Kernel32.dll");
return;
}
ktm32Lib = LoadLibrary("KtmW32.dll");
if (!ktm32Lib)
{
mwPerrorW32("Could not load Ktm32.dll");
return;
}
__CreateTransaction = GetProcAddress(ktm32Lib, "CreateTransaction");
__CommitTransaction = GetProcAddress(ktm32Lib, "CommitTransaction");
__MoveFileTransacted = GetProcAddress(kernel32Lib, "MoveFileTransactedA");
transactionFuncsOK = (__CreateTransaction && __MoveFileTransacted && __CommitTransaction);
if (!transactionFuncsOK)
{
mw_printf("Failed to get transaction functions\n");
}
}
NBodyStatus nbRunSystemPlain(const NBodyCtx* ctx, NBodyState* st)
{
NBodyStatus rc = NBODY_SUCCESS;
rc |= nbGravMap(ctx, st); /* Calculate accelerations for 1st step this episode */
if (nbStatusIsFatal(rc))
return rc;
while (st->step < ctx->nStep)
{
nbAddTracePoint(ctx, st);
nbUpdateDisplayedBodies(ctx, st);
rc |= nbStepSystemPlain(ctx, st);
if (nbStatusIsFatal(rc)) /* advance N-body system */
return rc;
rc |= nbCheckpoint(ctx, st);
if (nbStatusIsFatal(rc))
return rc;
nbReportProgress(ctx, st);
}
if (BOINC_APPLICATION || ctx->checkpointT >= 0)
{
mw_report("Making final checkpoint\n");
if (nbWriteCheckpoint(ctx, st))
{
mw_printf("Failed to write final checkpoint\n");
return NBODY_CHECKPOINT_ERROR;
}
}
return rc;
}
cl_int mwSelectDevice(CLInfo* ci, const cl_device_id* devs, const CLRequest* clr, const cl_uint nDev)
{
cl_int err = CL_SUCCESS;
if (clr->devNum >= nDev)
{
mw_printf("Requested device is out of range of number found devices\n");
return MW_CL_ERROR;
}
ci->dev = devs[clr->devNum];
err = mwGetDeviceType(ci->dev, &ci->devType);
if (err != CL_SUCCESS)
mw_printf("Failed to find type of device %u\n", clr->devNum);
return err;
}
static void CL_CALLBACK contextCallback(const char* errInfo,
const void* privateInfo,
size_t cb,
void* userData)
{
(void) privateInfo, (void) cb, (void) userData;
mw_printf("CL context error: %s\n", errInfo);
}
NBodyStatus nbInitNBodyStateCL(NBodyState* st, const NBodyCtx* ctx, const CLRequest* clr)
{
cl_int err;
/* Bodies must be set before trying to use this */
if (!st->bodytab)
{
return NBODY_CONSISTENCY_ERROR;
}
if (ctx->potentialType == EXTERNAL_POTENTIAL_CUSTOM_LUA)
{
mw_printf("Cannot use Lua potential with OpenCL\n");
return NBODY_UNSUPPORTED;
}
st->usesQuad = ctx->useQuad;
st->usesExact = (ctx->criterion == Exact);
st->usesCL = TRUE;
st->ci = mwCalloc(1, sizeof(CLInfo));
st->nbb = mwCalloc(1, sizeof(NBodyBuffers));
st->workSizes = mwCalloc(1, sizeof(NBodyWorkSizes));
st->kernels = mwCalloc(1, sizeof(NBodyKernels));
err = mwSetupCL(st->ci, clr);
if (err != CL_SUCCESS)
return NBODY_CL_ERROR;
if (!nbCheckDevCapabilities(&st->ci->di, ctx, st->nbody))
return NBODY_CAPABILITY_ERROR;
if (nbSetThreadCounts(st->workSizes, &st->ci->di, ctx) || nbSetWorkSizes(st->workSizes, &st->ci->di))
return NBODY_ERROR;
st->effNBody = nbFindEffectiveNBody(st->workSizes, st->usesExact, st->nbody);
if (nbLoadKernels(ctx, st))
return NBODY_CL_ERROR;
err = nbCreateBuffers(ctx, st);
if (err != CL_SUCCESS)
return NBODY_CL_ERROR;
err = nbSetInitialTreeStatus(st);
if (err != CL_SUCCESS)
return NBODY_CL_ERROR;
err = nbSetAllKernelArguments(st);
if (err != CL_SUCCESS)
return NBODY_CL_ERROR;
err = nbMarshalBodies(st, CL_TRUE);
if (err != CL_SUCCESS)
return NBODY_CL_ERROR;
return NBODY_SUCCESS;
}
int mwDisableDenormalsSSE(void)
{
int oldMXCSR = _mm_getcsr();
int newMXCSR = oldMXCSR | 0x8040;
_mm_setcsr(newMXCSR);
mw_printf("Disabled denormals\n");
return oldMXCSR;
}
int resolveCheckpoint(void)
{
int rc;
rc = mw_resolve_filename(CHECKPOINT_FILE, resolvedCheckpointPath, sizeof(resolvedCheckpointPath));
if (rc)
mw_printf("Error resolving checkpoint file '%s': %d\n", CHECKPOINT_FILE, rc);
return rc;
}
static void emdPrintFlowMatrix(const float* flow, int size1, int size2)
{
int i, j;
const int flowStep = 1;
if (!flow)
{
mw_printf("Empty flow\n");
return;
}
for (i = 0; i < size1; ++i)
{
for (j = 0; j < size2; ++j)
{
mw_printf("Flow[ % d][ % d] = % f\n", i, j, flow[flowStep * j + i]);
}
}
}