(03-24-2016, 10:24 PM)Tino Wrote: [ -> ]you can help find the problem, please compare dx12 performance, if performance is the same then the problem is gpu texture decoder, only diference with dx11, if it's slower then it's a different issue in videocommon
Can't test / compare to D3D12 at the moment. Not with an OS that's still buggy / doesn't work properly on certain hardware.
If Dolphin had a Vulkan backend, I could test that
But Ishiiruka's
OpenGL performance is the same as master. This means some D3D-specific changes are slowing down the D3D11 backend.
D3D9 can't be tested and compared to master, since that backend is deprecated/removed.
updated the latest folder with a new version: added options to Enable/Disable Compute shader encoding/decoding to allow better performance profiling.
@Kirbypuff: your turn
@Tino
I will try it with the latest version, but I haven't made a manual save in a while (the save state won't work with the new version and I stupidly forgot to make manual saves) so I'll have to do the whole dungeon again. Not a big deal since its a fairly short dungeon, but it may be a little bit until I can follow up due to work this weekend. I will let you know what happens with it!
Thanks for making this unofficial build Ishiiruka Team. I couldn't play The Legend of Zelda: The Wind Waker on the official Dolphin build with custom textures at 6x Internal Resolution. Now with Ishiiruka I can play it without stuttering with that quality. The option to Cache Custom Texture on GPU may have to do something with it. Custom Textures are 4Gb and my GTX 780 is 6Gb so am enjoying it right now.
(03-25-2016, 01:42 AM)Tino Wrote: [ -> ]updated the latest folder with a new version: added options to Enable/Disable the Compute shader texture codec.
I tested the latest version with both ComputeShader options disabled and guess what - the performance is exactly the same (not a single FPS difference between the GPU and CPU path).
BUT switching to the CPU-based codec *finally* restored the missing zigzag dither pattern effect on certain textures, so now the accuracy is the same as in master/dev builds. There's a bug in Ishiiruka's Direct3D11 ComputShader decoder that fails to render those patterns.
The mystery of why Ishiiruka renders certain textures differently is finally solved. Some of you actually preferred the inaccurate "clean" look of those textures without dithering (but with lots of banding/posterization).
So that leaves one thing that may cause the peformance issue.
It's a certain D3D "optimization patch" by Galop1n that Tino added to Ishiiruka a long time ago. It improved the performance of NSMB (w1-1, tested by Tino), but *slowed down everything else*. Now, with the latest changes from master integrated into Ishiiruka, that optimization slows down NSMB as well.
@Kirbypuff can you do me a favor, please set the option LogRenderTimeToFile and generate log files for master and ishiiruka at the points you see the performance difference?
@Tino: Here's a quick benchmark [2 tests, 4 logs] (see attachment)
Ishiiruka-v602 and master-4.0-9146 @6xIR, GPU texture processing OFF, Fast EFB Access turned off.
Test1: master is ~7% faster
Test2: master is ~13% faster
@kirbypuff: thanks for the test, can you be a little more specific in what you tested, game, location?
Using Build 602 in DBZBT3:
Disabling "Compute Texture Decoding" removes some effects, and I don't get the flicker(when using Store EFB copies to Texture Only) that I keep bugging you about.
Thanks for the report, now will be ablot easyer to isolate the issue