The basic idea is for each thread to save own result in SLM then read back all workgroup results. I've correctted the initial mistake with SLM offset as follows:
Correctness problem 1: /* write result data to SLM with offset using offset (threadID + threadNum*groupID) */ threadNum = sel.selReg(ocl::threadn, ir::TYPE_U32); sel.MUL(addr, groupID, threadNum); sel.ADD(addr, addr, sel.selReg(ocl::threadid, ir::TYPE_U32)); sel.MUL(addr, addr, GenRegister::immud(4)); sel.ADD(addr, addr, GenRegister::immud(slmAddr)); Should this be enough to ensure SLM conflicts don't occur ? Correctness problem 2: It looks that the initial input data pertains only to the initial local workgroup. Hence if I have (1 2 3 ... 128) as input src and 4 threads (x16 elem each) x 2 workgroups only the elements (1 2 3.. 64) are read for each workgroup. Efficiency problem 1: In a configuration of 512 elements (local size 64) only 4 threads launch at a time (threadID register). The execution looks like bellow. Launching 4096 or more with 64 local size issues again only 4 threads. I would expect that this is a waste of hardware...Is this normal ? Or maybe this is for each EU ? 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 _______________________________________________ Beignet mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/beignet
