i am trying to use opencl (api 1.2) on a bit older hardware and software - socket am3 cpu/a 5770 and a 6550 gpu, debian/squeeze w. 3.2 backports kernel and amd-driver version 12.4 <-- i use this outdated driver because of the old xorg version in debian/squeeze.
i use the enqueueNDRangeKernel method from the queue object to start a simple matrix multiplication kernel with different numbers of threads to get a feeling for the necessary number of threads for this device and i am experiencing the following problem:
if i use the following call it works like a charm, but only utilizing one compute module on the gpu/one core on the CPU:
queue->enqueueNDRangeKernel(*kernel, cl::NullRange, numThreads, cl::nullRange, NULL, &event);
numThreads ... total number of threads.
if i try to speficy the global and local dimenstion:
queue->enqueueNDRangeKernel(*kernel, cl::NullRange, numThreads, threadsPerWorkgroup, NULL, &event);
numThreads ... still total number of threads
threadsPerWorkgroup ... at the moment fixated to 16
i get an error "Invalid work group size".
i thought the opencl library will select a proper layout for the device automatically
can anyone give me some hints to resolve this problem?
global work size must be divisible by local work group size } in each and every dimension.
Are you making sure of this?
Also, 16 is not a nice number on GPU. Use multiples of 64 in order to use the GPU hardware effectively lest you should wither away hardware cycles for non-existing workitems
yes, thank you! you are right ... it was the problem of the number of global work-item dimensions has to be divisable by the local dimensions!