Hi this can be seen as crazy but some research of year 96 can be useful in thinking what Intel thought
were heavy useful algos that could offer improved perf using SSE,MMX,AVX!
For AVX there is an AVX site containing a lot of posts:
some new are from January offering general CRC perf spee using pcmuldq on Westemere!
also some AVX report numbers using Sandy Bridge silicon!
For SSE see:
http://www.datasheetarchive.com/datasheet-pdf/1070.html
especially intel reports 802-833 here you can see
"Increasing the Accuracy of the Results from the Reciprocal and Reciprocal Square Root"
Instructions using the Newton-Raphson Method..
which in fact is redeferenced in gpu gems3 nbody
MMX manuals here:
http://www.tommesani.com/IntelAppNotes.html
http://software.intel.com/en-us/articles/mmxt-technology-manuals-and-application-notes/
Thursday, 25 February 2010
Ideas for porting algos to GPU:AVX SSE and MMX ports!
Posted on 07:09 by Unknown
Subscribe to:
Post Comments (Atom)
0 comments:
Post a Comment