Implement shader caching #49050

reduz · 2021-05-25T00:33:02Z

Shader compilation is now cached. Subsequent loads take less than a millisecond.
Improved game, editor and project manager startup time.
Editor uses .godot/shader_cache to store shaders.
Game uses user://shader_cache
Project manager uses $config_dir/shader_cache
Options to tweak shader caching in project settings.
Editor path configuration moved from EditorSettings to new class, EditorPaths, so it can be available early on (before shaders are compiled).
Reworked ShaderCompilerRD to ensure deterministic shader code creation (else shader may change and cache will be invalidated).
Added shader compression with SMOLV: https://github.com/aras-p/smol-v

Please test.

BastiaanOlij · 2021-05-25T00:55:35Z

modules/glslang/register_types.cpp

+static String _get_cache_key_function_glsl(const RenderingDevice::Capabilities *p_capabilities) {
+	String version;
+	version = "SpirVGen=" + itos(glslang::GetSpirvGeneratorVersion()) + ", major=" + itos(p_capabilities->version_major) + ", minor=" + itos(p_capabilities->version_minor) + " , subgroup_size=" + itos(p_capabilities->subgroup_operations) + " , subgroup_ops=" + itos(p_capabilities->subgroup_operations) + " , subgroup_in_shaders=" + itos(p_capabilities->subgroup_in_shaders);
+	return version;


multiview settings need to be added here too but as there are a few more coming I'll do that as part of the multiview stereo render PR

BastiaanOlij

Super cool! looks like a sound approach to me. Couldn't find anything obvious that stuck out just doing a quick code review.

Calinou · 2021-05-25T01:07:38Z

I benchmarked it. Note that the times reported also include shutdown time (which takes about 3.5 seconds on average on the master branch).

I didn't notice any visual regressions so far in my testing.

System information

  PROCESSOR:          Intel Core i7-6700K @ 4.40GHz
    Core Count:       4                                        
    Thread Count:     8                                        
    Extensions:       SSE 4.2 + AVX2 + AVX + RDRAND + FSGSBASE 
    Cache Size:       8 MB                                     
    Microcode:        0xe2                                     
    Core Family:      Skylake                                  
    Scaling Driver:   intel_pstate powersave                   

  GRAPHICS:           Gigabyte NVIDIA GeForce GTX 1080 8GB
    Frequency:        1873/5005MHz     
    OpenGL:           4.6.0            
    Vulkan:           1.2.168          
    Display Driver:   NVIDIA 465.27    
    Monitor:          3 x Q32G1WG4     
    Screen:           7680x1440        

  MOTHERBOARD:        MSI Z170A GAMING PRO CARBON
    BIOS Version:     1.30                          
    Chipset:          Intel Xeon E3-1200 v5/E3-1500 
    Audio:            Intel 100 /C230               
    Network:          Intel I219-V                  

  MEMORY:             32GB

  DISK:               1000GB Samsung SSD 860 + 500GB Samsung SSD 850 + 1000GB Samsung SSD 850
    File-System:      ext4             
    Mount Options:    relatime rw      
    Disk Scheduler:   BFQ              

  OPERATING SYSTEM:   Fedora 33
    Kernel:           5.11.17-200.fc33.x86_64 (x86_64)                                                                     
    Desktop:          KDE Plasma 5.20.5                                                                                    
    Display Server:   X Server 1.20.11                                                                                     
    Compiler:         Clang 11.0.0 + LLVM 11.0.0                                                                           
    Security:         itlb_multihit: KVM: Mitigation of VMX disabled                                                       
                      + l1tf: Mitigation of PTE Inversion; VMX: vulnerable                                                 
                      + mds: Vulnerable; SMT vulnerable                                                                    
                      + meltdown: Vulnerable                                                                               
                      + spec_store_bypass: Vulnerable                                                                      
                      + spectre_v1: Vulnerable: __user pointer sanitization and usercopy barriers only; no swapgs barriers 
                      + spectre_v2: Vulnerable IBPB: disabled STIBP: disabled                                              
                      + srbds: Vulnerable                                                                                  
                      + tsx_async_abort: Vulnerable

Godot versions

All binaries were compiled with the same Clang version and SCons options.

bin/godot.linuxbsd.tools.64.llvm: 4.0.dev.custom_build.f11219175 (this branch)
bin/godot.linuxbsd.tools.64.llvm.vanilla: 4.0.dev.custom_build.af03e9c83 (master branch)
bin/godot.x11.tools.64.llvm: 3.4.beta.custom_build.a38b44741

Project manager

❯ hyperfine -i "bin/godot.linuxbsd.tools.64.llvm --quit" "bin/godot.linuxbsd.tools.64.llvm.vanilla --quit" "bin/godot.x11.tools.64.llvm --quit"

Benchmark #1: bin/godot.linuxbsd.tools.64.llvm --quit
  Time (mean ± σ):      3.959 s ±  0.354 s    [User: 3.090 s, System: 0.159 s]
  Range (min … max):    3.282 s …  4.260 s    10 runs
 
  Warning: Ignoring non-zero exit code.
 
Benchmark #2: bin/godot.linuxbsd.tools.64.llvm.vanilla --quit
  Time (mean ± σ):      5.573 s ±  0.094 s    [User: 8.848 s, System: 0.195 s]
  Range (min … max):    5.417 s …  5.693 s    10 runs
 
  Warning: Ignoring non-zero exit code.
 
Benchmark #3: bin/godot.x11.tools.64.llvm --quit
  Time (mean ± σ):      2.577 s ±  0.145 s    [User: 1.665 s, System: 0.112 s]
  Range (min … max):    2.459 s …  2.868 s    10 runs
 
Summary
  'bin/godot.x11.tools.64.llvm --quit' ran
    1.54 ± 0.16 times faster than 'bin/godot.linuxbsd.tools.64.llvm --quit'
    2.16 ± 0.13 times faster than 'bin/godot.linuxbsd.tools.64.llvm.vanilla --quit'

Empty project

❯ hyperfine -i "bin/godot.linuxbsd.tools.64.llvm /tmp/c/project.godot --quit" "bin/godot.linuxbsd.tools.64.llvm.vanilla /tmp/c/project.godot --quit" "bin/godot.x11.tools.64.llvm /tmp/d/project.godot --quit"

Benchmark #1: bin/godot.linuxbsd.tools.64.llvm /tmp/c/project.godot --quit
  Time (mean ± σ):     12.142 s ±  0.154 s    [User: 10.030 s, System: 0.329 s]
  Range (min … max):   11.909 s … 12.268 s    10 runs
 
  Warning: Ignoring non-zero exit code.
 
Benchmark #2: bin/godot.linuxbsd.tools.64.llvm.vanilla /tmp/c/project.godot --quit
  Time (mean ± σ):     15.639 s ±  1.005 s    [User: 31.154 s, System: 0.389 s]
  Range (min … max):   14.500 s … 18.368 s    10 runs
 
  Warning: Ignoring non-zero exit code.
  Warning: The first benchmarking run for this command was significantly slower than the rest (18.368 s). This could be caused by (filesystem) caches that were not filled until after the first run. You should consider using the '--warmup' option to fill those caches before the actual benchmark. Alternatively, use the '--prepare' option to clear the caches before each timing run.
 
Benchmark #3: bin/godot.x11.tools.64.llvm /tmp/d/project.godot --quit
  Time (mean ± σ):      6.516 s ±  0.274 s    [User: 3.201 s, System: 0.198 s]
  Range (min … max):    5.749 s …  6.653 s    10 runs
 
Summary
  'bin/godot.x11.tools.64.llvm /tmp/d/project.godot --quit' ran
    1.86 ± 0.08 times faster than 'bin/godot.linuxbsd.tools.64.llvm /tmp/c/project.godot --quit'
    2.40 ± 0.18 times faster than 'bin/godot.linuxbsd.tools.64.llvm.vanilla /tmp/c/project.godot --quit'

Conclusion

The new shader caching helps noticeably, but there's still more work to do to get back to 3.x speeds.

TokisanGames · 2021-05-25T01:34:50Z

Conclusion

The new shader caching helps noticeably, but there's still more work to do to get back to 3.x speeds.

While this doesn't sound good, I'm not sure of the value of benchmarking an empty project. You need a complex reference scene with lighting and many materials.

Godot 3 has severe game lag (as in no one would play it), where it will seize the engine to recompile the shaders. No logging, no debugging, monitors don't update. No information unless you build a custom engine with printk logging directly in the renderer. It even lags in the editor, 20-300ms each time.

The work done so far to cache shaders in #46330 hasn't yet been more effective than using hacky work arounds.

My hope is that vulkan and the shader caching here finally eliminates this performance lag. However if it's not tested against a complex reference scene that currently lags in Godot 3, how will we know if the root problem this and other PRs attempt to address is resolved?

editor/editor_node.cpp

nathanfranke · 2021-05-25T11:37:33Z

Edit: Made wording clearer

Note that this PR is a great improvement, but as of now still doesn't fully resolve the underlying startup performance problems. As Calinou noted they are still not optimal and also first time startups can still be improved. Otherwise thank you very much for the hard work that went into this.

akien-mga · 2021-05-25T12:05:08Z

@nathanfranke This is not advertised as a "fix" for startup time issues. It improves startup times significantly, but there are many other fixes that can be made to further reduce startup times.

That doesn't mean that shader caching shouldn't be merged as is.

The real potential gain from this is not so much startup time, but avoid shader compilation stutters in-game (or at most having them happen only the first time the game is played, if we can't do shader pre-compilation at startup).

akien-mga · 2021-05-25T12:43:43Z

Project manager uses $config_dir/shader_cache

Not super important but this should likely use the cache dir instead of config dir? (~/.cache/godot/shader_cache)

servers/rendering_server.cpp

thirdparty/README.md

servers/rendering/renderer_rd/shader_rd.cpp

servers/rendering/renderer_rd/renderer_compositor_rd.cpp

thirdparty/README.md

servers/rendering/renderer_rd/shader_rd.cpp

* Shader compilation is now cached. Subsequent loads take less than a millisecond. * Improved game, editor and project manager startup time. * Editor uses .godot/shader_cache to store shaders. * Game uses user://shader_cache * Project manager uses $config_dir/shader_cache * Options to tweak shader caching in project settings. * Editor path configuration moved from EditorSettings to new class, EditorPaths, so it can be available early on (before shaders are compiled). * Reworked ShaderCompilerRD to ensure deterministic shader code creation (else shader may change and cache will be invalidated). * Added shader compression with SMOLV: https://github.com/aras-p/smol-v

akien-mga · 2021-05-31T09:02:22Z

Thanks!

Zireael07 · 2021-05-31T09:29:29Z

I just noticed that Calinou also reported results for 3.4 - does that mean this is not Vulkan-only, as I originally guessed?

aaronfranke · 2021-05-31T09:42:52Z

@Zireael07 No, Calinou was comparing with the current 3.x branch, which doesn't have this change, for the sake of comparison (just like he compared with the then-current master without this PR for the sake of comparison).

YuriSizov · 2021-05-31T11:17:33Z

Also because we have a huge performance decrease between 3.x and the master and this PR addresses some of that. So a comparison was made to see how close are we getting to the old performance.

With Godot 4 the .godot folder is introduced for import and shader_cache files. See i.e. godotengine/godot#49050 When Godot 4 is the de-facto standard all other entries can most likely be removed.

reduz requested review from a team as code owners May 25, 2021 00:33

Calinou added enhancement topic:rendering labels May 25, 2021

Calinou added this to the 4.0 milestone May 25, 2021

Calinou added the topic:shaders label May 25, 2021

BastiaanOlij reviewed May 25, 2021

View reviewed changes

BastiaanOlij approved these changes May 25, 2021

View reviewed changes

Calinou mentioned this pull request May 25, 2021

Very slow project/editor opening in Godot 4 #43351

Open

clayjohn approved these changes May 25, 2021

View reviewed changes

neikeq reviewed May 25, 2021

View reviewed changes

editor/editor_node.cpp Outdated Show resolved Hide resolved

Calinou reviewed May 25, 2021

View reviewed changes

Chaosus self-requested a review May 25, 2021 17:22

reduz force-pushed the implement-spirv-cache branch from f112191 to dbbf888 Compare May 25, 2021 17:39

reduz requested a review from a team as a code owner May 25, 2021 17:39

Chaosus removed their request for review May 25, 2021 17:45

Chaosus reviewed May 26, 2021

View reviewed changes

thirdparty/README.md Outdated Show resolved Hide resolved

DarkKilauea reviewed May 27, 2021

View reviewed changes

servers/rendering/renderer_rd/shader_rd.cpp Show resolved Hide resolved

akien-mga force-pushed the implement-spirv-cache branch from dbbf888 to 0d2e029 Compare May 31, 2021 08:13

akien-mga approved these changes May 31, 2021

View reviewed changes

akien-mga merged commit 596eb78 into godotengine:master May 31, 2021

This was referenced Jun 3, 2021

Creating editor feature profiles doesn't work #49257

Closed

Replace get_settings_dir with get_config_dir when fetching configuration paths #49282

Merged

madmiraal mentioned this pull request Jun 4, 2021

Add shader_cache folder to .gitignore #49317

Closed

akien-mga mentioned this pull request Jun 7, 2021

Running --doctool wrongly creates shader_cache directory in working directory #49392

Closed

KoBeWi mentioned this pull request Jun 20, 2021

Compile shaders when a scene is loaded #13954

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement shader caching #49050

Implement shader caching #49050

reduz commented May 25, 2021

BastiaanOlij May 25, 2021

BastiaanOlij left a comment

Calinou commented May 25, 2021 •

edited

Loading

TokisanGames commented May 25, 2021

Conclusion

nathanfranke commented May 25, 2021 •

edited

Loading

akien-mga commented May 25, 2021 •

edited

Loading

akien-mga commented May 25, 2021

akien-mga commented May 31, 2021

Zireael07 commented May 31, 2021

aaronfranke commented May 31, 2021 •

edited

Loading

YuriSizov commented May 31, 2021

Implement shader caching #49050

Implement shader caching #49050

Conversation

reduz commented May 25, 2021

BastiaanOlij May 25, 2021

Choose a reason for hiding this comment

BastiaanOlij left a comment

Choose a reason for hiding this comment

Calinou commented May 25, 2021 • edited Loading

System information

Godot versions

Project manager

Empty project

Conclusion

TokisanGames commented May 25, 2021

Conclusion

nathanfranke commented May 25, 2021 • edited Loading

akien-mga commented May 25, 2021 • edited Loading

akien-mga commented May 25, 2021

akien-mga commented May 31, 2021

Zireael07 commented May 31, 2021

aaronfranke commented May 31, 2021 • edited Loading

YuriSizov commented May 31, 2021

Calinou commented May 25, 2021 •

edited

Loading

nathanfranke commented May 25, 2021 •

edited

Loading

akien-mga commented May 25, 2021 •

edited

Loading

aaronfranke commented May 31, 2021 •

edited

Loading