PPU LLVM JIT Repository Early Testing
Started by Ekaseo




358 posts in this topic
ssshadow
Moderator
*****


4
2,494 posts 63 threads Joined: Aug 2017
10-19-2014, 12:48 AM -
(10-18-2014, 09:56 PM)Dante38490 Wrote: Yep Smile is Here http://rghost.net/58597622

Is it just me, or is for instance the Disgaea 3 intro running at ~5-10 fps and about 3-4% cpu usage?

Edit: The main branch does this as well. Someone really messed something up it seems.
jacky400
Unregistered


 
10-19-2014, 04:48 PM -
Disgaea 3 runs pretty quick now than before using latest PPU recompiler build.

Title : 37FPS


In-game : 5-6 FPS

jacky400
Unregistered


 
10-19-2014, 10:20 PM -
The GCM sample , cube_mrt.elf runs very fast in 720x480 now

Previous : 13-16 FPS
Current : 38-45 FPS


leon112211
Unregistered


 
10-20-2014, 08:47 AM -
(10-19-2014, 12:48 AM)ssshadow Wrote:
(10-18-2014, 09:56 PM)Dante38490 Wrote: Yep Smile is Here http://rghost.net/58597622

Is it just me, or is for instance the Disgaea 3 intro running at ~5-10 fps and about 3-4% cpu usage?

Edit: The main branch does this as well. Someone really messed something up it seems.

If i open chrome with the emulator it works fine if i close it works 5-10 fps
and another report with guided fate paradox you need more than 4 cores
with i5 intro works 5-10 with i7 50-60
this game uses all 4 cores with i5
HyPeR
Unregistered


 
10-20-2014, 10:05 AM -
You think it depends on the cores?

My core i5 2500k @ 4,3ghz also runs paradox only witz 7-13fps. But my i7 q3610m in my laptop has 8 threads and its not running at 100% like my 2500k, only with 80-85% on all threads but even there i am still getting only 10-15fps. (well its a laptop maybe its different on desktop, can be that the clock is limiting here but who knows)

I think threads will not help much, its time for intel to release some mainstream 8cores, hopefully within the next two gens..

Need a reason to upgrade my 2500k xD
leon112211
Unregistered


 
10-20-2014, 10:21 AM -
(10-20-2014, 10:05 AM)HyPeR Wrote: You think it depends on the cores?

My core i5 2500k @ 4,3ghz also runs paradox only witz 7-13fps. But my i7 q3610m in my laptop has 8 threads and its not running at 100% like my 2500k, only with 80-85% on all threads but even there i am still getting only 10-15fps. (well its a laptop maybe its different on desktop, can be that the clock is limiting here but who knows)

I think threads will not help much, its time for intel to release some mainstream 8cores, hopefully within the next two gens..

Need a reason to upgrade my 2500k xD

they already released 8 cores but only For i7 extreme
and did you use the newest rpcs3 for test?
you see with disgaea 3 i get 55-70 fps
guided fate paradox only gives me like i mentioned it weird
HyPeR
Unregistered


 
10-20-2014, 11:01 AM -
yes of course i used the newest rpcs3 version^^

well i7 extreme is nowhere near mainstream xD
ssshadow
Moderator
*****


4
2,494 posts 63 threads Joined: Aug 2017
10-20-2014, 09:49 PM -
(10-20-2014, 08:47 AM)leon112211 Wrote:
(10-19-2014, 12:48 AM)ssshadow Wrote:
(10-18-2014, 09:56 PM)Dante38490 Wrote: Yep Smile is Here http://rghost.net/58597622

Is it just me, or is for instance the Disgaea 3 intro running at ~5-10 fps and about 3-4% cpu usage?

Edit: The main branch does this as well. Someone really messed something up it seems.

If i open chrome with the emulator it works fine if i close it works 5-10 fps
and another report with guided fate paradox you need more than 4 cores
with i5 intro works 5-10 with i7 50-60
this game uses all 4 cores with i5

?

This is some strange bug with the emulator. I can get really good speed, like 70 fps on the opening video for some very short or very long period of time, but eventually frame rate just drops to nothing, like 6 fps right now. This happens at the same time as rpcs3.exe frees like 1 gb (!) of memory too for whatever reason.

But whatever, this is a WIP.

EDIT: Performance can go both ways, TGFP went from 25 -> 70 fps while cpu usage went from 60 % -> 40 %. A wild and uneducated guess would be cache misses, branch prediction doing the wrong predictions, or similar. CPU clock was consistently 2,8 ghz.
gopalsr83
Unregistered


 
10-21-2014, 06:12 AM -
(10-20-2014, 09:49 PM)ssshadow Wrote: ?

This is some strange bug with the emulator. I can get really good speed, like 70 fps on the opening video for some very short or very long period of time, but eventually frame rate just drops to nothing, like 6 fps right now. This happens at the same time as rpcs3.exe frees like 1 gb (!) of memory too for whatever reason.

But whatever, this is a WIP.

EDIT: Performance can go both ways, TGFP went from 25 -> 70 fps while cpu usage went from 60 % -> 40 %. A wild and uneducated guess would be cache misses, branch prediction doing the wrong predictions, or similar. CPU clock was consistently 2,8 ghz.

The algorithm the JIT recompiler uses to detect which basic blocks of code need to be compiled is very simplistic and inefficient and results in the same basic block being compiled more than once. Ideally each block must be recompiled only once. This has the following negative effects:

1. The recompiler ends up using a lot of memory
2. It takes more time than what is ideally required to recompile a section of the code.
3. It produces a lot of code which can cause caches to overflow.

I am working on a more intelligent algorithm which should hopefully fix the issues that you are seeing.
Darkriot
Member


0
498 posts 29 threads Joined: Aug 2017
10-21-2014, 08:34 AM -
(10-21-2014, 06:12 AM)gopalsr83 Wrote:
(10-20-2014, 09:49 PM)ssshadow Wrote: ?

This is some strange bug with the emulator. I can get really good speed, like 70 fps on the opening video for some very short or very long period of time, but eventually frame rate just drops to nothing, like 6 fps right now. This happens at the same time as rpcs3.exe frees like 1 gb (!) of memory too for whatever reason.

But whatever, this is a WIP.

EDIT: Performance can go both ways, TGFP went from 25 -> 70 fps while cpu usage went from 60 % -> 40 %. A wild and uneducated guess would be cache misses, branch prediction doing the wrong predictions, or similar. CPU clock was consistently 2,8 ghz.

The algorithm the JIT recompiler uses to detect which basic blocks of code need to be compiled is very simplistic and inefficient and results in the same basic block being compiled more than once. Ideally each block must be recompiled only once. This has the following negative effects:

1. The recompiler ends up using a lot of memory
2. It takes more time than what is ideally required to recompile a section of the code.
3. It produces a lot of code which can cause caches to overflow.

I am working on a more intelligent algorithm which should hopefully fix the issues that you are seeing.

Ooh, thanks. We will be wait for this fix Sleepy


Forum Jump:


Users browsing this thread: 4 Guest(s)