cudaError_t cudaLaunch (const char *p1)
{
cudaError_t res;
#if defined(DEBUG)
fprintf (stderr, PACKAGE_NAME": THREAD %d cudaLaunch is at %p\n", THREADID, real_cudaLaunch);
fprintf (stderr, PACKAGE_NAME": THREAD %d cudaLaunch params %p\n", THREADID, p1);
#endif
if (real_cudaLaunch != NULL && mpitrace_on && Extrae_get_trace_CUDA())
{
Extrae_cudaLaunch_Enter (p1);
res = real_cudaLaunch (p1);
Extrae_cudaLaunch_Exit ();
}
else if (real_cudaLaunch != NULL && !(mpitrace_on && Extrae_get_trace_CUDA()))
{
res = real_cudaLaunch (p1);
}
else
{
fprintf (stderr, "Unable to find cudaLaunch in DSOs! Dying...\n");
exit (0);
}
return res;
}
cudaError_t cudaLaunch (const void *entry) {
struct ld_kernel_s *ldKernel;
struct kernel_lookup_s *kernel;
cudaError_t err;
int i;
for (i = 0; cuda_lookup_table[i].address != NULL; i++) {
if (cuda_lookup_table[i].address == entry) {
break;
}
}
if (!cuda_lookup_table[i].address) {
warning("Couldn't find kernel @%p\n", entry);
goto error;
}
kernel = &cuda_lookup_table[i];
ldKernel = find_kernel_entry(kernel);
assert(ldKernel);
dbg_notify_event();
for (i = 0; i < ldKernel->nb_params; i++) {
struct ld_kern_param_s *ldParam = &ldKernel->params[i];
const void **current_arg = (const void **) temp_arguments[i].arg;
if (ldParam->is_pointer) {
size_t offset = 0;
struct ld_mem_s *ldBuffer = find_off_mem_entry(*current_arg, &offset);
/* better handling of invalid buffers ?*/
if (!ldBuffer) {
error("in kernel %s, arg #%d, invalid bufer %s@%p", ldKernel->name, i, ldParam->name, *current_arg);
}
assert(ldBuffer);
kernel_set_buffer_arg_event (ldKernel, ldParam, i, ldBuffer, offset);
} else {
kernel_set_scalar_arg_event (ldKernel, ldParam, i, current_arg);
}
}
kernel_executed_event(ldKernel, configure_get_worksizes(NULL, NULL), 3);
err = real_cudaLaunch(entry);
kernel_finished_event(ldKernel, configure_get_worksizes(NULL, NULL), 3);
return err;
error:
return real_cudaLaunch(entry);
}