Lines Matching full:smem
606 "__global float* block_hists, __local float* smem)\n"
618 "__local float* hists = smem + lp * cnbins * (CELLS_PER_BLOCK_X *\n"
685 "__local float* smem = squares + boffset;\n"
686 "float sum = smem[hid];\n"
688 "smem[hid] = sum = sum + smem[hid + 18];\n"
691 "smem[hid] = sum = sum + smem[hid + 9];\n"
694 "smem[hid] = sum + smem[hid + 4];\n"
696 "sum = smem[0] + smem[1] + smem[2] + smem[3] + smem[8];\n"
702 "sum = smem[hid];\n"
704 "smem[hid] = sum = sum + smem[hid + 18];\n"
707 "smem[hid] = sum = sum + smem[hid + 9];\n"
710 "smem[hid] = sum + smem[hid + 4];\n"
712 "sum = smem[0] + smem[1] + smem[2] + smem[3] + smem[8];\n"
715 "inline float reduce_smem(volatile __local float* smem, int size)\n"
718 "float sum = smem[tid];\n"
719 "if (size >= 512) { if (tid < 256) smem[tid] = sum = sum + smem[tid + 256];\n"
721 "if (size >= 256) { if (tid < 128) smem[tid] = sum = sum + smem[tid + 128];\n"
723 "if (size >= 128) { if (tid < 64) smem[tid] = sum = sum + smem[tid + 64];\n"
726 "if (size >= 64) { if (tid < 32) smem[tid] = sum = sum + smem[tid + 32];\n"
728 "if (size >= 32) { if (tid < 16) smem[tid] = sum = sum + smem[tid + 16];\n"
730 "if (size >= 16) { if (tid < 8) smem[tid] = sum = sum + smem[tid + 8];\n"
732 "if (size >= 8) { if (tid < 4) smem[tid] = sum = sum + smem[tid + 4];\n"
734 "if (size >= 4) { if (tid < 2) smem[tid] = sum = sum + smem[tid + 2];\n"
736 "if (size >= 2) { if (tid < 1) smem[tid] = sum = sum + smem[tid + 1];\n"
741 "if (size >= 64) smem[tid] = sum = sum + smem[tid + 32];\n"
746 "if (size >= 32) smem[tid] = sum = sum + smem[tid + 16];\n"
747 "if (size >= 16) smem[tid] = sum = sum + smem[tid + 8];\n"
748 "if (size >= 8) smem[tid] = sum = sum + smem[tid + 4];\n"
749 "if (size >= 4) smem[tid] = sum = sum + smem[tid + 2];\n"
750 "if (size >= 2) smem[tid] = sum = sum + smem[tid + 1];\n"
805 "volatile __local float* smem = products;\n"
807 "if (tid < 13) smem[tid] = product = product + smem[tid + 32];\n"
809 "if (tid < 16) smem[tid] = product = product + smem[tid + 16];\n"
811 "if(tid<8) smem[tid] = product = product + smem[tid + 8];\n"
813 "if(tid<4) smem[tid] = product = product + smem[tid + 4];\n"
815 "if(tid<2) smem[tid] = product = product + smem[tid + 2];\n"
820 "smem[tid] = product = product + smem[tid + 32];\n"
827 "smem[tid] = product = product + smem[tid + 16];\n"
828 "smem[tid] = product = product + smem[tid + 8];\n"
829 "smem[tid] = product = product + smem[tid + 4];\n"
830 "smem[tid] = product = product + smem[tid + 2];\n"
834 "product = product + smem[tid + 1];\n"
864 "volatile __local float* smem = products;\n"
866 "if(tid<32) smem[tid] = product = product + smem[tid + 32];\n"
868 "if(tid<16) smem[tid] = product = product + smem[tid + 16];\n"
870 "if(tid<8) smem[tid] = product = product + smem[tid + 8];\n"
872 "if(tid<4) smem[tid] = product = product + smem[tid + 4];\n"
874 "if(tid<2) smem[tid] = product = product + smem[tid + 2];\n"
879 "smem[tid] = product = product + smem[tid + 32];\n"
884 "smem[tid] = product = product + smem[tid + 16];\n"
885 "smem[tid] = product = product + smem[tid + 8];\n"
886 "smem[tid] = product = product + smem[tid + 4];\n"
887 "smem[tid] = product = product + smem[tid + 2];\n"
891 "product = product + smem[tid + 1];\n"
922 "volatile __local float* smem = products;\n"
924 "if(tid<32) smem[tid] = product = product + smem[tid + 32];\n"
926 "if(tid<16) smem[tid] = product = product + smem[tid + 16];\n"
928 "if(tid<8) smem[tid] = product = product + smem[tid + 8];\n"
930 "if(tid<4) smem[tid] = product = product + smem[tid + 4];\n"
932 "if(tid<2) smem[tid] = product = product + smem[tid + 2];\n"
937 "smem[tid] = product = product + smem[tid + 32];\n"
942 "smem[tid] = product = product + smem[tid + 16];\n"
943 "smem[tid] = product = product + smem[tid + 8];\n"
944 "smem[tid] = product = product + smem[tid + 4];\n"
945 "smem[tid] = product = product + smem[tid + 2];\n"
949 "smem[tid] = product = product + smem[tid + 1];\n"
