it's ca. 3 months now and after trying out all the drivers i could get somehow, upto 12.4 preview, the crashing issue that happens only on HD4xxx GPUs with max. workgroup size of 128 is still not resolved. Hopefully with one of the next drivers the fix you were taking about in january gets in the driver.
Patiently (a)waiting ...
Are you querying the 'CL_KERNEL_WORK_GROUP_SIZE' attribute of the device you are executing on? In some cases we must limit the size on a per kernel basis.
int getMaxKernelWorkGroupSize(cl_fft_plan *plan, unsigned int *max_wg_size, unsigned int num_devices, cl_device_id *devices) the
if(wg_size < kInfo->num_workitems_per_workgroup)" is replaced with "if(wg_size < plan->max_work_item_per_workgroup)", where "plan->max_work_item_per_workgroup" is constantly set to 32, for
as Raistmer has sent you the sources already, I've put some example together that should make reproducing the crashes possible (Linux Testcase for LHD4K crashes) :
edit: link removed
Please inform me if you got it, so i can free the server space again.
Nachricht wurde geändert durch: Urs Echternacht to remove a link.