kernekmod.cl

                uint wei_stg_off = wei_stg_base_off +
                                   o_c * MLO_N_IN_TILES_PERSTACK * MLO_FILTER_SZ +
                                   k_act * MLO_FILTER_SIZE0;
                if(wei_stg_off + MLO_FILTER_SIZE0 > MLO_WEIGHTS_SZ)
                    printf("%d\n",wei_stg_off);
                for(uint i = 0; i < MLO_FILTER_SIZE0; ++i)
                {
                    //Comment offset out so it don't crash on NV/Intel OpenCL
                    //This might not be needed for a AMD system
                    pvt_wei_stage[i] = lcl_wei[/*wei_stg_off + */i];
                }