[kd-11]

Relates to #17277

Changes

• Remove all per-draw UBO objects and replace with SSBO array-of-structs model. This removes a lot of descriptor bind operations completely.
• Avoid using more than a few bytes of data in push constants. Instead, we write a per-draw config and push a pointer offset to the next one. Restricting push constants memory to 4 bytes per draw has decent gains.
• Replace std::vector with rsx::simple_array for trivial storage. The latter has small-vector optimizations that drastically reduce the number of memory allocations done on the fly and gives a decent perf bump to both Vulkan and OpenGL.
• Other minor optimizations related to resource bindings.

Overall, a 10-20% boost is observed depending on the game. MESA drivers benefit the most here since they doesn't keep a state cache like the properietary ones so state changes are more expensive on those drivers.

[AniLeo]

Reduces performance gap between radeonsi and radv as described on #15926

radv: 319 FPS -> 349 FPS (+30 FPS)

radeonsi: 385 FPS -> 390 FPS (+5 FPS)

System information

• Kernel: 6.15.7-1-MANJARO
• CPU: AMD Ryzen 7 9800X3D
• GPU: AMD Radeon RX 6800 XT (RADV NAVI21)
• Driver: mesa 25.1.6

[digant73]

quick test on KZ3 but perf seems worse (more gpu usage and lower fps).
Attached log files and screenshots on both this PR and current build.

CPU: 9950X3D
GPU: RTX 5090 (latest driver 580.88)
RAM: 64 GB

Current build:

RPCS3.log_current.gz

PR:

RPCS3.log.gz

[AniLeo]

No system specifications, no cpu/gpu/driver info, no performance overlay with averages, can't do anything just from that

[digant73]

No system specifications, no cpu/gpu/driver info, no performance overlay with averages, can't do anything just from that

Added the requested info. As I said, it was a very quick test. I see also R3 has more GPU usage. I will test the PR more in deep tomorrow.

[LucioXerus]

I tested 10 games. (Call of Duty Black Ops, CN Punchtime Explosion, Goldeneye Reloaded, Playstation All Stars Battle Royale, Street Fighter X Tekken, Super Street Fighter 2 Turbo, Need for Speed Carbon, Tekken 6, Virtua Fighter 5 and WWE All Stars). They all went ingame without crashes or graphical glitches. 4 player splitscreen with Call of Duty i saw the fps go from approximately 15fps average to 18 fps average. So all good on my end so far.
Master:

PR:

Intel i9 11900K
16GB RAM
AMD Radeon RX 7600
Fedora 42 Kde Plasma

[digant73]

Got a DEVICE_LOST in Resistance 3 after few minutes of gameplay.
Attached also the log file.

NOTE: I confirm also the worse perf on KZ3 (previous post edited adding onscreen stats)

RPCS3.zip

[kd-11]

I think the perf loss is just because I forgot to restore optimizations in the shader compiler.

[digant73]

Infamous 2 demo crashing in the same place with the following error just when shooting the first enemy:

RPCS3.zip

[digant73]

same crash as in Infamous 2 demo also in UC1 just in a place where current build has an avg of 40 fps (on top of the u-boat):

[kd-11]

The crashes are fixed so I pushed the commits though I have drafted the PR due to serious issues with vulkan shader interpreter.

[digant73]

I think the perf loss is just because I forgot to restore optimizations in the shader compiler.

Hmm no difference on perf on KZ3 with latest commit.
Also the crashes on shaders are still there

[kd-11]

I think the perf loss is just because I forgot to restore optimizations in the shader compiler.

Hmm no difference on perf on KZ3 with latest commit. Also the crashes on shaders are still there

Compilation crashes in the log you provided previously should be fixed already. If there are new crashes, upload the log files.

[digant73]

I think the perf loss is just because I forgot to restore optimizations in the shader compiler.

Hmm no difference on perf on KZ3 with latest commit. Also the crashes on shaders are still there

Compilation crashes in the log you provided previously should be fixed already. If there are new crashes, upload the log files.

Here the log for Infamous 2 demo. It crashes in the same place

RPCS3.log

[Megamouse]

Suggested change

	index =::size32(properties.constant_offsets);
	index = ::size32(properties.constant_offsets);

[kd-11]

I think the perf loss is just because I forgot to restore optimizations in the shader compiler.

Hmm no difference on perf on KZ3 with latest commit. Also the crashes on shaders are still there

Compilation crashes in the log you provided previously should be fixed already. If there are new crashes, upload the log files.

Here the log for Infamous 2 demo. It crashes in the same place

RPCS3.log

Looks like a glslang bug? Same shader compiles fine for me. I'll circle back to this once I finish the interpreter work.

[digant73]

with latest two commits the crashes are still present even in other places were it was not previously present such as in the following, in UC1 (u-boat chapter):

RPCS3.zip

[EmulationChannell]

[RPCS3.log](https://github.com/user-attachments/files/21780949/RPCS3.log)

[bigol83]

GTA IV crashes with this log just before reaching ingame

RPCS3.log

F 0:00:45.700683 {RSX [0x024179c]} SIG: Thread terminated due to fatal error: Failed to compile fragment shader
(in file C:/a/rpcs3/rpcs3/rpcs3/Emu/RSX/VK/VKProgramPipeline.cpp:109[:5], in function 'VkShaderModule vk::glsl::shader::compile()')

[kd-11]

Compiler crash is definitely a glslang bug. I'll figure out a workaround.

[kd-11]

@digant73 @bigol83 Time to retest once the artifacts are ready

[digant73]

the crashes seem fixed but perf are overall the same or worse on my rtx 5090, especially on KZ2 and KZ3. In all games, I see more GPU usage (in RivaTuner SS) in the PR than in current build. E.g. in Infamous 1 and 2 it is 33% against 30%. In KZ1 and 2 the difference is more evident with high GPU usage (99% against 82%) with lower FPS (53 against solid 60) as reported in my previous post for KZ3

[bigol83]

GTA IV also fixed for me, performance is about the same, maybe slightly lower in PR compared to master, like 1-2 fps lower in PR compared to master. I am on rtx 3060 gpu

[kd-11]

Not sure what is bugging with the NV driver, definitely something from their end. CPU usage should be much lower, and I don't have any plans on moving stuff back to the CPU. In fact, the roadmap offloads even more work to the GPU. I'll benchmark a few things, I probably just need to remove the UAVs from the FS stage. I'll see what nsight has to say about it.

[Valtekken]

Not sure what is bugging with the NV driver, definitely something from their end. CPU usage should be much lower, and I don't have any plans on moving stuff back to the CPU. In fact, the roadmap offloads even more work to the GPU. I'll benchmark a few things, I probably just need to remove the UAVs from the FS stage. I'll see what nsight has to say about it.

Offloading work to the GPU sounds good, would this result in lower requirements on the CPU side by any chance?

[digant73]

Not sure what is bugging with the NV driver, definitely something from their end. CPU usage should be much lower, and I don't have any plans on moving stuff back to the CPU. In fact, the roadmap offloads even more work to the GPU. I'll benchmark a few things, I probably just need to remove the UAVs from the FS stage. I'll see what nsight has to say about it.

ok. CPU usage remained basically the same in my 9950X3D (otherwise I also reported it). Only GPU usage is worse

[kd-11]

This is ready for testing again. A workaround is in place for the crashing issue.

[Jonathan44062]

Completed the whole level of Manchester in Resistance without any crashes
playing a few races in GT5 and GT6 seems fine
GOW 3 seems fine too

[digant73]

I confirm no more crash on FOM after some long runs.
@woj1993 Hopefully even the glitches on your games could be fixed. Please retest

[kd-11]

Let's track any visual bugs in a separate ticket if any are found. Otherwise we'll never merge this monster of a PR.

[woj1993]

Well after last commit Asura's Wrath has crashed RPCS3. It seems his wrath was too powerful. Log:
log.zip

[kd-11]

Well after last commit Asura's Wrath has crashed RPCS3. It seems his wrath was too powerful. Log: log.zip

I don't see a RSX crash in the log. What happened when you ran it? It could be some random issue.
If you can reproduce it, just open a ticket with the log attached and tag me, I'll take it over.