... but we are back in the convention center. Some of us continued their work in their hotel room untill 3 am in order to do some benchmarks for memcpy. We got a hint from a former team about using a static preallocated buffer to make usage of AMD Geodes internal cache. According to their tests they seem to need only 3 ms to copy a whole video frame. Copying seems to be the bottleneck of our video processing code: to copy one image by using a shared segment, the libc memcpy surprisingly requires approx. 40 ms.
After testing we can't confirm the statement of the other team, but next to some benchmarking, testing and lecture about the CPU we aligned variables to 16 byte adresses, page-aligned our regions that needs to be copied and implemented a MMX powered memcpy.
Apropos today: we received a test match invitation from NaoDevils that is scheduled for this evening. Yesterday was a one vs. one of some other teams and we lend our keeper to them. He did very well. The actual matches are starting tomorrow. Our first match will be on monday at 11 am local time (5 am in Germany) against Northern Bites.
No comments:
Post a Comment