On 04/08/2011 09:37 PM, Marko Ristola wrote: > > Here is some statistics without likely and with likely functions. > > It seems that using likely() gives better performance with Phenom I too. I'm not sure whether that's the right conclusion. See below. > % Plain knl % No likely % With likely % > dvb_ringbuffer_write 5,9 62,8 8,7 81,3 5,7 79,2 > dvb_dmx_swfilter_packet 1,2 12,8 0,7 6,5 0,8 11,1 > dvb_dmx_swfilter_204 2,3 24,5 1,3 12,1 0,7 9,7 > > > Here "Plain knl %" is "perf top -d 30" percentage. > 24,5 12,1 and 9,7 are percentages without a patch, with basic patch > and last is with "likely" functions using patch. The varying ratio between dvb_ringbuffer_write and dvb_dmx_swfilter_packet, which haven't been modified by the patch, seems to indicate that the numbers were influenced by other activities on the system or by the data used during the tests. 8,7% for dvb_ringbuffer_write in test #2 could also indicate a higher data rate than in #1 and #3. However, adding likely() certainly won't do any harm. Feel free to send an incremental patch. Regards, Andreas -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
