nroberts changed the topic of #videocore to: Raspberry Pi Mesa drivers discussion - Logs http://freenode.irclog.whitequark.org/videocore
apinheiro has quit [Ping timeout: 246 seconds]
jcea has quit [Ping timeout: 250 seconds]
chema has joined #videocore
txenoo has quit [Ping timeout: 252 seconds]
gpoo has quit [Ping timeout: 240 seconds]
itoral has joined #videocore
<itoral> apinheiro[m]: FYI, I am experimenting with adding compile strategies for thread counts other than 4
<itoral> right now, if we can't compile with 4 threads, we basically give up on some of the optimizations
<itoral> but some times we can get to compile with some of these optimizations at 2 threads
<itoral> I am seeing this happening for a bunch of UE4 shaders for example
<itoral> so it is probably worth a try
<itoral> since lowering thread count already makes a large impact in reducing register pressure
itoral has quit [Read error: Connection reset by peer]
itoral has joined #videocore
<apinheiro[m]> itoral ok, thanks for the ping
<apinheiro[m]> as we were mentioning the other day
<apinheiro[m]> the issue now is how to balance having so many strategies (with different combination of optimizations) and getting a reasonable build time
<apinheiro[m]> specially on the register allocation
<apinheiro[m]> one of this days it would be good to measure how much time is used on each part
<itoral> I also implemented a mechanism to skip strategies that won't make progress
<itoral> so that should alleviate things
<apinheiro[m]> I have the feeling that a lot goes to the register allocation part, but I could be wrong
<itoral> nope
<itoral> compiles are typically fast enough
<apinheiro[m]> <itoral "I also implemented a mechanism t"> ok, nice
<itoral> except when we spill
<itoral> but we only allow spilling as a last resort in the last 2 strategies
<itoral> so impact should be limited
<itoral> spilling can take very long though
<itoral> because it restarts register allocation for every spill
<itoral> they had the same problem in intel
<itoral> and jason managed to fix that in their compiler
<itoral> by avoiding a full restart of the register allocation for each spill
<itoral> and manually fixing its graph for the spilled regiter
<itoral> we should do that one that and it should have a big impact for us too
apinheiro has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Read error: Connection reset by peer]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
gpoo has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
<itoral> nice, the new compile strategies work does lead to visible perf improvements in spoza and UE4
<itoral> somehwere between 1% and 3%
itoral has quit [Remote host closed the connection]
itoral has joined #videocore
jcea has joined #videocore
jcea has quit [Ping timeout: 250 seconds]
itoral has quit [Quit: Leaving]
jcea has joined #videocore
apinheiro has quit [Ping timeout: 260 seconds]
apinheiro has joined #videocore
jcea1 has joined #videocore
jcea has quit [Quit: jcea]
jcea1 is now known as jcea
jcea has quit [Quit: jcea]
jcea has joined #videocore
apinheiro has quit [Ping timeout: 268 seconds]