Texture upscale shader rework (drastic perf improvement) #15109

hrydgard · 2021-11-07T14:57:55Z

This drastically improves the performance of (Vulkan-only) compute-shader based texture upscaling, making it perform stutter-free on much more lower end GPUs.

Texture upscale shaders now each have a fixed scale factor, and instead of computing each 2x2 or 4x4 for every pixel and sampling from that, each individual thread of the compute shader writes all four or sixteen pixels that it computes anyway.

This does mean that we now only have these two variants of compute shader upscaling, but more can be added:

MMPX: 2x for pixel art
4xBRZ: 4x upscaler, more generic

It also fixes #13679 and #15094, and removes a bunch of unused code.

Verified on NV, Intel, Adreno and Mali GPUs, plus Apple through MoltenVK.

…the next change.

…doing redundant work

…ten it for benefit of mobile GPUs

unknownbrackets

Seems like we still have an unused computeCopy bool.

-[Unknown]

unknownbrackets · 2021-11-07T15:04:11Z

GPU/Vulkan/TextureCacheVulkan.cpp


 	// Rachet down scale factor in low-memory mode.
-	if (lowMemoryMode_) {
+	// TODO: I think really we should just turn it off?


Maybe. Nowadays it's most common on mobile:
https://report.ppsspp.org/logs/kind/724

-[Unknown]

Leaving it as is in this PR, but yeah.

GPU/Vulkan/TextureCacheVulkan.cpp

assets/shaders/tex_mmpx.csh

hrydgard · 2021-11-07T15:49:34Z

The 4xBRZ shader works on ARM Mali GPUs now by the way, so I removed the blacklisting.

ghost · 2021-11-07T21:21:59Z

This is also fix my issue #14530
TY ☺️

ghost · 2021-12-11T13:45:04Z

The latest ppsspp build 4xBRZ is slow on my phone compare to this build I don't know which commit cause the lag issue.
I tried to use 4xBRZ in Radiant Mythology in causing
lag unlike before.

recording.ppdmp.zip

hrydgard · 2021-12-12T08:05:20Z

Hm. Do you have Texture Filtering set to "Auto Max Quality"? Just wondering if that might be related, since it'll add some extra work at texture creation time.

ghost · 2021-12-12T10:21:47Z

Hm. Do you have Texture Filtering set to "Auto Max Quality"? Just wondering if that might be related, since it'll add some extra work at texture creation time.

Yes I enable it together with 4xBRZ but after I disable auto max quality the game is still slow/lag if 4xBRZ is enable.

here's some logs idk if this is useful.

<<< log_count = 71 >>>
[12-11 20:19:30.833 23773:23773 F/DEBUG]
pid: 23632, tid: 23748, name: RenderMan  >>> org.ppsspp.ppsspp <<<

[12-11 20:19:30.858 23773:23773 F/DEBUG]
      #04 pc 00000000006ddca0  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (HandleAssert(char const*, char const*, int, char const*, char const*, ...)+248) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:19:30.858 23773:23773 F/DEBUG]
      #05 pc 00000000006bf71c  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (VulkanRenderManager::EndSubmitFrame(int)+300) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:19:30.859 23773:23773 F/DEBUG]
      #06 pc 00000000006bba84  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (VulkanRenderManager::ThreadFunc()+108) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:19:30.859 23773:23773 F/DEBUG]
      #07 pc 00000000006bfd00  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (void* std::__ndk1::__thread_proxy<std::__ndk1::tuple<std::__ndk1::unique_ptr<std::__ndk1::__thread_struct, std::__ndk1::default_delete<std::__ndk1::__thread_struct> >, void (VulkanRenderManager::*)(), VulkanRenderManager*> >(void*)+64) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:20:37.972 23982:23982 F/DEBUG]
pid: 23870, tid: 23953, name: RenderMan  >>> org.ppsspp.ppsspp <<<

[12-11 20:20:37.982 23982:23982 F/DEBUG]
      #04 pc 00000000006ddca0  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (HandleAssert(char const*, char const*, int, char const*, char const*, ...)+248) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:20:37.982 23982:23982 F/DEBUG]
      #05 pc 00000000006bf71c  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (VulkanRenderManager::EndSubmitFrame(int)+300) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:20:37.982 23982:23982 F/DEBUG]
      #06 pc 00000000006bba84  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (VulkanRenderManager::ThreadFunc()+108) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

[12-11 20:20:37.982 23982:23982 F/DEBUG]
      #07 pc 00000000006bfd00  /data/app/~~r-RclirpY_XORGIAwFmxMg==/org.ppsspp.ppsspp-7CjNZogj7YSXOBuBNumSHw==/lib/arm64/libppsspp_jni.so (void* std::__ndk1::__thread_proxy<std::__ndk1::tuple<std::__ndk1::unique_ptr<std::__ndk1::__thread_struct, std::__ndk1::default_delete<std::__ndk1::__thread_struct> >, void (VulkanRenderManager::*)(), VulkanRenderManager*> >(void*)+64) (BuildId: 2f41c1f7a28a1b96f3abdc46a5afe6742381a376)

hrydgard · 2021-12-12T11:15:54Z

Nah, looks like those logs are from some previous crash?

I'm going to look into this properly, starting with #15229 (which also disables shader upscaling of video)

ghost · 2021-12-12T13:45:02Z

GE Dump using ppsspp v1.12.3-453
4xBRZ.ppdmp.zip

hrydgard · 2021-12-12T17:59:30Z

@Gamemulatorer do you think you could try to find about in which commit it slowed down? Or maybe it's just some changed setting?

Anyway, #15230 should make this a little bit faster at least (though Adreno seems to benefit more).

ghost · 2021-12-13T01:18:08Z

Anyway, #15230 should make this a little bit faster at least (though Adreno seems to benefit more).

Why there is no android build in ppsspp buildbot for cd8493f

unknownbrackets · 2021-12-13T03:42:46Z

It's missing due to Gradle / Android SDK stability/quality issues. Requeued again...

-[Unknown]

ghost · 2021-12-13T04:04:50Z

Still slow at 2xBRZ using ppsspp latest build

hrydgard · 2021-12-13T07:56:44Z

OK, thanks.

I might try to enable the GPU profiler automatically when stats are displayed, but I think I know what's going on here.

The game is rendering at 60hz, but it's doing some animation of a large background texture every couple of frames. Unfortunately, running the upscaling shader plus mipmap generation on a 512x512 texture takes ~30ms on hardware similar to yours, while due to the 60hz framerate, we have a 16ms deadline to get it done (and in practice less). So I don't think this could have ever worked smoothly on your hardware with texture upscaling enabled, unless we've changed something that previously caused it to skip upscaling.

One thing I will do is disable mipmap generation for 512x512 textures since those are hardly ever used for 3D things, but that would only happen if you had Auto Max Quality enabled anyway.. I still don't think we can hit the deadline here unless we find some really good optimizations of the upscaling algorithm, or switch to a different one.

ghost · 2021-12-13T09:11:31Z

I just notice in the latest build that only in menu of Tales Radiant Mythology is lag but in game is smooth with 4xBRZ or 2xBRZ maybe this game is behaving the same as #12463
But in other games 4xBRZ is smooth like Tekken 6.

hrydgard · 2021-12-13T09:21:11Z

Yeah, it works fine in many games, this game is just a bit "extreme". All I'm saying is that I don't think texture scaling actually slowed down from previous builds, it's just this particular game that's using textures in a tricky way.

hrydgard · 2021-12-16T01:01:41Z

I've been messing around with it a bit more and it seems like older Mali drivers will aggressively downclock to save power if a heavy compute shader happens only every few frames. If I hack out part of the algorithm so it makes it under a frame, GPU clockspeed seems to stay high, and the scaler runs many, many times faster. So weird.

I think I'm going to set a limit at Mali-G72 or something, and below that we'll use similar rules as the software scaler - not scaling large textures that change a lot.

unknownbrackets · 2021-12-16T01:25:05Z

Maybe it's an anti bitcoin miner malware feature...

-[Unknown]

ghost · 2021-12-16T04:00:49Z

The newest ps2 emu on android have this settings this could possibly be implemented on ppsspp?

Just an idiot question/request 😂

hrydgard · 2021-12-16T08:16:05Z

That might actually be interesting to try, I just yesterday saw it mentioned in Android documentation (search for affinity in https://developer.android.com/games/optimize ).

However since in this case it's the GPU that's clocking down, it's probably not gonna help much...

Applies the same rules as for software upscaling in this case. Should fix the stutters seen in #15109

hrydgard added 4 commits November 7, 2021 12:55

Remove the copy-compute-shader path

c111d6c

Switch texture scaling shaders to a fixed scale model, preparing for …

6349704

…the next change.

Speed up the 4xBRZ upscaler with 16x and the MMPX one with 4x by not …

281ddc2

…doing redundant work

Fix edge conditions in 4xXBZR

5654d3c

hrydgard added Vulkan GE emulation Backend-independent GPU issues labels Nov 7, 2021

hrydgard added this to the v1.13.0 milestone Nov 7, 2021

Remove support for 16-bit textures from the upscaling shader, to shor…

c60feef

…ten it for benefit of mobile GPUs

unknownbrackets reviewed Nov 7, 2021

View reviewed changes

hrydgard added 3 commits November 7, 2021 16:30

Comment fix, remove assert

43dc54b

Remove unused variable

f58e75e

Remove ARM Mali blacklisting for 4xBRZ, works now

b152dc1

hrydgard marked this pull request as ready for review November 7, 2021 15:49

hrydgard merged commit 3f4562c into master Nov 7, 2021

hrydgard deleted the texture-upscale-rework branch November 7, 2021 16:33

hrydgard mentioned this pull request Nov 7, 2021

[Android] [Vulkan] [Texture Scaling] Upscale level auto, 2x, 3x + Texture shader 4xBRZ can cause wacky gfx issue. #14530

Closed

hrydgard mentioned this pull request Dec 12, 2021

Vulkan: 4xBRZ texture upscaling shader optimizations #15230

Merged

hrydgard added a commit that referenced this pull request Dec 16, 2021

Vulkan: Be more restrictive about texture upscaling on "slow" devices.

cbe196a

Applies the same rules as for software upscaling in this case. Should fix the stutters seen in #15109

hrydgard added a commit that referenced this pull request Dec 16, 2021

Vulkan: Be more restrictive about texture upscaling on "slow" devices.

80ae4b0

Applies the same rules as for software upscaling in this case. Should fix the stutters seen in #15109

hrydgard mentioned this pull request Dec 16, 2021

Vulkan: Be more restrictive about hardware texture upscaling on "slow" GPUs #15238

Merged

unknownbrackets mentioned this pull request Feb 7, 2022

libretro core graphical glitches with OpenGL driver #15056

Closed

5 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Texture upscale shader rework (drastic perf improvement) #15109

Texture upscale shader rework (drastic perf improvement) #15109

hrydgard commented Nov 7, 2021 •

edited

Loading

unknownbrackets left a comment

unknownbrackets Nov 7, 2021

hrydgard Nov 7, 2021

hrydgard commented Nov 7, 2021 •

edited

Loading

ghost commented Nov 7, 2021

ghost commented Dec 11, 2021 •

edited by ghost

Loading

hrydgard commented Dec 12, 2021

ghost commented Dec 12, 2021 •

edited by ghost

Loading

hrydgard commented Dec 12, 2021

ghost commented Dec 12, 2021

hrydgard commented Dec 12, 2021 •

edited

Loading

ghost commented Dec 13, 2021

unknownbrackets commented Dec 13, 2021

ghost commented Dec 13, 2021 •

edited by ghost

Loading

hrydgard commented Dec 13, 2021

ghost commented Dec 13, 2021

hrydgard commented Dec 13, 2021

hrydgard commented Dec 16, 2021

unknownbrackets commented Dec 16, 2021

ghost commented Dec 16, 2021 •

edited by ghost

Loading

hrydgard commented Dec 16, 2021

Texture upscale shader rework (drastic perf improvement) #15109

Texture upscale shader rework (drastic perf improvement) #15109

Conversation

hrydgard commented Nov 7, 2021 • edited Loading

unknownbrackets left a comment

Choose a reason for hiding this comment

unknownbrackets Nov 7, 2021

Choose a reason for hiding this comment

hrydgard Nov 7, 2021

Choose a reason for hiding this comment

hrydgard commented Nov 7, 2021 • edited Loading

ghost commented Nov 7, 2021

ghost commented Dec 11, 2021 • edited by ghost Loading

hrydgard commented Dec 12, 2021

ghost commented Dec 12, 2021 • edited by ghost Loading

hrydgard commented Dec 12, 2021

ghost commented Dec 12, 2021

hrydgard commented Dec 12, 2021 • edited Loading

ghost commented Dec 13, 2021

unknownbrackets commented Dec 13, 2021

ghost commented Dec 13, 2021 • edited by ghost Loading

hrydgard commented Dec 13, 2021

ghost commented Dec 13, 2021

hrydgard commented Dec 13, 2021

hrydgard commented Dec 16, 2021

unknownbrackets commented Dec 16, 2021

ghost commented Dec 16, 2021 • edited by ghost Loading

hrydgard commented Dec 16, 2021

hrydgard commented Nov 7, 2021 •

edited

Loading

hrydgard commented Nov 7, 2021 •

edited

Loading

ghost commented Dec 11, 2021 •

edited by ghost

Loading

ghost commented Dec 12, 2021 •

edited by ghost

Loading

hrydgard commented Dec 12, 2021 •

edited

Loading

ghost commented Dec 13, 2021 •

edited by ghost

Loading

ghost commented Dec 16, 2021 •

edited by ghost

Loading