static void
printBuildError(
cl_int error,
cl_device_id device,
SolverKgen kgen,
const SubproblemDim *dims,
const PGranularity *pgran,
const CLBLASKernExtra *kextra,
const char *source,
const char *buildLog)
{
char name[128];
char dimStr[1024];
char pgranStr[1024];
char *p;
MemoryPattern *mempat = NULL;
unsigned int i, j;
const char *s;
name[0] = '\0';
clGetDeviceInfo(device, CL_DEVICE_NAME, sizeof(name), name, NULL);
// lookup memory pattern
s = NULL;
for (i = 0; i < BLAS_FUNCTIONS_NUMBER; i++) {
for (j = 0; j < clblasSolvers[i].nrPatterns; j++) {
mempat = &clblasSolvers[i].memPatterns[j];
if (kgen == mempat->sops->genKernel) {
s = kernelTypeString(kextra->kernType);
break;
}
}
if (s != NULL) {
break;
}
}
// sprintf Subproblem dimensions
p = dimStr;
for (i = 0; i < mempat->nrLevels; i++) {
p = sprintfGranulation(p, dims, i);
strcat(p, "; ");
p += strlen(p);
}
// sprintf data parallelism granularity
sprintf(pgranStr, "pgran->wgDim = %d, pgran->wgSize[0] = %u, "
"pgran->wgSize[1] = %u, pgran->wfSize = %u",
pgran->wgDim, pgran->wgSize[0], pgran->wgSize[1],
pgran->wfSize);
fprintf(stderr, "\n========================================================\n\n");
fprintf(stderr, "AN INTERNAL KERNEL BUILD ERROR OCCURRED!\n");
fprintf(stderr, "device name = %s\n", name);
fprintf(stderr, "error = %d\n", error);
fprintf(stderr, "memory pattern = %s, %s kernel generator\n", mempat->name, s);
fprintf(stderr, "Subproblem dimensions: %s\n", dimStr);
fprintf(stderr, "Parallelism granularity: %s\n", pgranStr);
fprintf(stderr, "Kernel extra flags: %u\n", kextra->flags);
fprintf(stderr, "Source:\n\n%s\n\n", source);
fprintf(stderr, "--------------------------------------------------------\n\n");
if (buildLog) {
fprintf(stderr, "Build log:\n\n%s\n", buildLog);
}
else {
fprintf(stderr, "Build log is unavailable\n");
}
fprintf(stderr, "========================================================\n\n");
}
void
dumpKernel(
const SolutionStep *step,
CLBlasKernelType ktype)
{
FILE *file;
char tmp[1024];
MemoryPattern *pattern;
const char *s;
const CLBlasKargs *kargs = (const CLBlasKargs*)&step->args;
char *srcBuf;
unsigned int i;
fileNameFromSolution(tmp, step->funcID, step);
file = fopen((const char*)tmp, "a+");
pattern = &clblasSolvers[step->funcID].memPatterns[step->patternID];
// now, dump the info
sprintf(tmp, "offset M = %lu, offset N = %lu, offset A = %lu,"
"offset BX = %lu, offset CY = %lu\n",
kargs->offsetM, kargs->offsetN, kargs->offA, kargs->offBX,
kargs->offCY);
fputs(tmp, file);
sprintf(tmp, "Memory pattern = %s\n", pattern->name);
fputs(tmp, file);
s = kernelTypeString(ktype);
sprintf(tmp, "Kernel type = %s\n", s);
fputs(tmp, file);
// data parallelism granularity
if (step->pgran.wgDim == 1) {
sprintf(tmp, "work group size = %u\n", step->pgran.wgSize[0]);
}
else {
sprintf(tmp, "work group size = %u x %u\n", step->pgran.wgSize[0],
step->pgran.wgSize[1]);
}
fputs(tmp, file);
fputs("Problem granulation\n", file);
for (i = 0; i < pattern->nrLevels; i++) {
sprintf(tmp, "[%u]: ", i);
fputs(tmp, file);
sprintfGranulation(tmp, step->subdims, i);
fputs(tmp, file);
fputs("\n", file);
}
srcBuf = malloc(SRC_BUFSIZE);
if (srcBuf != NULL) {
clGetProgramInfo(step->kernels[ktype]->program,
CL_PROGRAM_SOURCE, SRC_BUFSIZE, srcBuf, NULL);
fputs("Kernel source:\n\n", file);
fputs(srcBuf, file);
}
else {
fputs("Kernel source: not available\n", file);
}
free(srcBuf);
fputs("--------------------------------------------------------------"
"------------------------------------------------------------\n",
file);
fclose(file);
}