#131  
Old 10th December 2014, 05:30 PM
V1del V1del is offline
Project Supporter
Senior Member
 
Join Date: Feb 2012
Posts: 442
Default

It up to date with what you have in master right now, I'd have to go and compare myself to check if there is indeed such a big regression, it could also be that the build environment of the daily builds is rather old and therefore fails some optimisations or something might be worth looking into, a bit busy right now so I might not get to it until holidays start (which is 2 weeks from now)

EDIT: Went ahead and played some rogue squadron SSSE3 does indeed seem a bit slower (although not to the degrees mentioned by Ambient on my machine) running both on linux compiled with gcc 4.9.2

Last edited by V1del; 10th December 2014 at 07:16 PM.
Reply With Quote
  #132  
Old 10th December 2014, 07:45 PM
HatCat's Avatar
HatCat HatCat is offline
Alpha Tester
Project Supporter
Senior Member
 
Join Date: Feb 2007
Location: In my hat.
Posts: 16,255
Default

Hm. Well I never touched any of the SSSE3 code since the public release I did; the only thing that's written in SSSE3 is the shuffling. So I think it's not SSSE3-related.

Then again, I have no GCC 4.9.2 because MinGW slacks behind. I do however have Linux and got my plugin to compile on it for Mupen64 0.5 (fixed a compiler error with the Linux port too, wonder if they synced that cause I know I haven't pushed that to master yet). I don't remember any of the games seeming any slower than how I got it to build with MinGW on Windows; maybe it's something with their makefile or something experimental about the updated compiler they use.
Reply With Quote
  #133  
Old 10th December 2014, 08:28 PM
RPGMaster's Avatar
RPGMaster RPGMaster is offline
Alpha Tester
Project Supporter
Super Moderator
 
Join Date: Dec 2013
Posts: 2,001
Default

Lol, I was about to report that it was slower on mine too. But I went and enabled Whole Program Optimization, for intel. So either that fixed it for me, or the binary I had in my PJ64 2.1 folder was a testing build that I forgot about ;/ . Either way, it's faster on my end for Rogue Squadron. So it must be something they did wrong.
Reply With Quote
  #134  
Old 10th December 2014, 10:40 PM
HatCat's Avatar
HatCat HatCat is offline
Alpha Tester
Project Supporter
Senior Member
 
Join Date: Feb 2007
Location: In my hat.
Posts: 16,255
Default

Who said anything about MSVC haha. This is news to me?! Your Microsoft Visual Studio build of my RSP plugin is faster than my MinGW public release 6? I wasn't expecting that, but I guess it's apt to expect, considering what drastic bottlenecks the vector multiply and packs/unpacks without temporary intrinsic solutions were.

I'm only talking about GCC here which is what they're using on Linux. It's certainly faster for me (I'm sure I've lost credibility to say that though after the whole SSSE3 interpreter > recompiler generalization I made ), so I have no idea what happened to make it slower for them. If somebody could isolate as of what commit this happened I'd greatly appreciate it.
Reply With Quote
  #135  
Old 11th December 2014, 12:55 AM
Ambient_Malice Ambient_Malice is offline
Senior Member
 
Join Date: Dec 2014
Posts: 179
Default

Quote:
Originally Posted by V1del View Post
It up to date with what you have in master right now, I'd have to go and compare myself to check if there is indeed such a big regression, it could also be that the build environment of the daily builds is rather old and therefore fails some optimisations or something might be worth looking into, a bit busy right now so I might not get to it until holidays start (which is 2 weeks from now)

EDIT: Went ahead and played some rogue squadron SSSE3 does indeed seem a bit slower (although not to the degrees mentioned by Ambient on my machine) running both on linux compiled with gcc 4.9.2
Naboo/Infernal Machine seem worse affected, especially when it comes to rendering particle effects, which are RSP-based.

During the intro for Infernal Machine, the new SSSE3 drops to 24-30Hz on my Core Quad @ 3.2Ghz. The new SSE2 runs at a steady 49-50Hz. (Rom is PAL.) The old SSSE3 from m64py dips to 40Hz at spots during this intro.
Reply With Quote
  #136  
Old 11th December 2014, 12:55 AM
RPGMaster's Avatar
RPGMaster RPGMaster is offline
Alpha Tester
Project Supporter
Super Moderator
 
Join Date: Dec 2013
Posts: 2,001
Default

Actually Intel is integrated into MSVC. Same with Clang. If GCC was integrated into MSVC, I'd probably use that more too .

I just tried MSVC, it couldn't even compile SSSE3 ;/ .

Error 66 error C2284: '_mm_shuffle_epi8': illegal argument to intrinsic function, parameter 1

it was refering to
vd = _mm_shuffle_epi8(vd, key);

MSVC is too slow ;/ . I'm actually surprised how bad it was. You'd probably have to do more than just altering the vector instructions, to even catch up with build #6, for MSVC.

I'm almost 100% certain it's some compiler issue. Wouldn't make sense why my Intel build would be faster, when HatCat optimized his code for GCC. + HatCat said it was faster for him too .
Reply With Quote
  #137  
Old 11th December 2014, 12:56 AM
V1del V1del is offline
Project Supporter
Senior Member
 
Join Date: Feb 2012
Posts: 442
Default

It didn't really test thouroughly could also just've been some placebo or something *shrug*

EDIT just read all those posts, may test indiana then

Last edited by V1del; 11th December 2014 at 01:05 AM.
Reply With Quote
  #138  
Old 11th December 2014, 12:58 AM
RPGMaster's Avatar
RPGMaster RPGMaster is offline
Alpha Tester
Project Supporter
Super Moderator
 
Join Date: Dec 2013
Posts: 2,001
Default

I just realized, it could just be the compiler version they used. For example, VCH was terrible with Intel.

Also MSVC was just really bad ;/ .
Reply With Quote
  #139  
Old 11th December 2014, 01:31 AM
HatCat's Avatar
HatCat HatCat is offline
Alpha Tester
Project Supporter
Senior Member
 
Join Date: Feb 2007
Location: In my hat.
Posts: 16,255
Default

Quote:
Originally Posted by Ambient_Malice View Post
During the intro for Infernal Machine, the new SSSE3 drops to 24-30Hz on my Core Quad @ 3.2Ghz. The new SSE2 runs at a steady 49-50Hz.

the new SSSE3 drops to 24-30Hz on my Core Quad @ 3.2Ghz. The new SSE2 runs at a steady 49-50Hz.

the new SSSE3 drops to 24-30Hz on my Core Quad @ 3.2Ghz. The new SSE2 runs at a steady 49-50Hz.

the new SSSE3 drops to 24-30Hz.... The new SSE2 runs at a steady 49-50Hz.
WAT?!

Their new m64+ SSSE3 build of my plugin is slower than their new SSE2 build of it?

lol they must have a fucked up makefile now then, when ecsv was still on the team I'd helped him fix the makefile, maybe someone changed it again

Last edited by HatCat; 11th December 2014 at 01:36 AM.
Reply With Quote
  #140  
Old 11th December 2014, 03:03 AM
Ambient_Malice Ambient_Malice is offline
Senior Member
 
Join Date: Dec 2014
Posts: 179
Default

Quote:
Originally Posted by V1del View Post
It didn't really test thouroughly could also just've been some placebo or something *shrug*

EDIT just read all those posts, may test indiana then
I have no clue why, but the 64 bit version of m64p runs Naboo/Infernal much better than the 32 bit version, particularly in dynarec mode. Both games suffer extreme rendering errors resulting in geometry flickering and portals rendering incorrectly, but the 64 bit version is completely unaffected. I've mentioned this fact on the mupen64plus google group, along with my RSP findings. Hopefully this results in some improvements.

I think the m64p team need to set up a proper automated build system, because the bitbucket thing is just... there. Chugging along like some ghostly compiler attempting to fulfill the dying wish of its creators. There's no way to download old binaries for testing, which is particularly annoying for obsessives like me.
Reply With Quote
Reply
Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Forum Jump


All times are GMT. The time now is 03:16 AM.


Powered by vBulletin® Version 3.7.3
Copyright ©2000 - 2018, Jelsoft Enterprises Ltd.