Hi Kyrill,
On 07/26/2018 11:34 AM, Kyrill Tkachov wrote:
Hi Luis,
On 25/07/18 19:10, Luis Machado wrote:
The adjusted vector costs give Falkor a reasonable boost in
performance for FP
benchmarks (both CPU2017 and CPU2006) and doesn't change INT
benchmarks that
much. About 0.7% for CPU2017 FP and 1.54% for CPU2006 FP.
OK for trunk?
The patch looks ok and safe to me (though you'll need approval from the
maintainers).
I'd be interested to see what workloads in CPU2017 were affected by this.
Any chance you could post the breakdown in numbers from CPU2017?
Sure. Here it is (speed):
605.mcf_s: -1.8%
620.omnetpp_s: -2% (tends to be noisy)
623.xalancbmk_s: 2%
654.roms_s: 7%
INT mean: -0.09%
FP mean: 0.70%
It is worth mentioning i noticed bigger improvements in CPU2017 rate,
but i did not record those numbers for the final run. The speed
benchmarks seem to have a slightly different performance profile.
Here's a breakdown of the biggest changes from CPU2006 in case you're
interested:
410.bwaves: 5.4%
434.zeusmp: 9.7%
436.cactusADM: -12.3%
437.leslie3d: 5.2%
459.GemsFDTD: 16.9%
cactusADM seems to have a pretty big loop that is a win if vectorized,
but experimentation showed me it is tricky to get GCC to vectorize that
specific loop without also vectorizing particular loops from the other
benchmarks.
It would be nice to get cactusADM back up though.
