24 Feb
2012
24 Feb
'12
4:10 p.m.
Just did an MX6 build with skia and agl built with -ftree-parallelize-loops=4
The results are not very encouraging - visible tearing got worse, and 0xbench 3D results improve by merely 1.5%. (This was done with the 4.7 toolchain, given -ftree-parallelize-loops should be slightly more effective there).
Either gcc's support for it just isn't where we want it yet, or skia/agl code is simply not using any bits that could be parallelized meaningfully.
Let's see how it goes with other parts of the tree...
ttyl bero