Scimark and the sparse matrix mul test
- From: Daniel Kolesa <quaker66@xxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Tue, 18 Mar 2014 16:48:53 +0000
Hello,
I was recently messing around with LuaJIT benchmarking and used the Scimark
test as one of the benchmarks, only to find out that the performance for
most tests is fine (within 20% slower compared to gcc -O2, which is very
nice) with one test being considerably slower, the sparse matmul test.
So I decided to create a very minimal implementation of the test and ended
up with http://codepad.org/hEJOhpAJ - the resulting performance is the same
as the "regular" LuaJIT scimark rewrite by Mike. I deided to replace the
inner loop with a very equivalent C function called through ffi - simply
put this http://codepad.org/gmjl15by into a file called loopc.c and
compiled with gcc loopc.c -o loopc.so -O2 -fPIC -shared. As you can see the
code is direct rewrite of the Lua inner loop. The performance is however
vastly different - getting 1800 MFLOPS rather than 400 something (just like
native SciMark).
Can anyone explain this behavior? It's the same on LuaJIT 2.0.3 and LuaJIT
2.1-alpha.
Thanks,
Daniel
Other related posts:
