Disk Shader Caching #4923

jroweboy · 2019-09-07T22:32:40Z

Based on the excellent work from Rodrigo, this brings over much of the framework that he wrote but with obvious changes to support 3DS shaders instead.

Design:
There are two caches with different reasons to wipe the cache. The transferrable cache is meant to be device agnostic, and store everything from the PICA state needed to generate the shaders. For FS and VS, we store the entire PICA registers, as its simpler to handle storing the same configuration for both types. Additionally, for VS, we store the shader program and shader swizzle data from memory as well. As this cache should be essentially unchanged from version to version, it is only invalidated when either 1) a manual version bump occurs or 2) the cache is corrupted and a shader fails to build

The precompiled cache stores both generated GLSL and the resulting binary cache. This cache is only built during the game launch when the transferrable cache is loaded. If this cache is present, we will read all of the elements from the transferrable cache, and attempt to match up and "inject" the precompiled shaders into the in-memory ShaderCaches. If this cache is present, it should speed up loading as we don't need to run a full shader compile for all of the shaders. This cache is invalidated in the following circumstances: 1) Precompiled shader binary is rejected by the driver 2) any of the long list of files related to shader generation is changed (causing the generated version hash in cmake to not match) 3) the cache is missing both the decompiled glsl and the precompiled binary.

Currently there is NOT a loading screen as I want people to test it out first and focus on the actual implementation. I would rather avoid a loading screen entirely and instead do async compilation on a separate context, but if we can't work that out in time before this gets merged, I can tack on the loading screen I made for yuzu.

UI Screenshot

This change is

flamesage · 2019-09-07T22:48:22Z

Are there any changes to the UI?

Is there an option to enable / disable disk shader caching in settings?
Do we have an option in the game list to open the Disk Shader cache folder like in yuzu?

wwylele · 2019-09-08T11:53:56Z

Posting this message from discord here in case we forget: the introduction of vfs is unnecessary. Can implement a simple vector appender (or not at all) for this use case.

wwylele · 2019-09-08T12:09:51Z

src/video_core/renderer_opengl/gl_shader_manager.cpp

+
+    auto shader = OGLProgram();
+    shader.handle = glCreateProgram();
+    glProgramParameteri(shader.handle, GL_PROGRAM_SEPARABLE, GL_TRUE);


~~This is an extension. Did you make this optional? (Sorry if I missed some code)~~ nvm

The disk cache only loads with separable shader objects enabled. I didn't wanna play around with caching full shader programs at this point in time ^^; Added a log in case they don't support it so theres some feedback. FYI I'm intentionally still saving the shaders to the transferable cache as i don't see a reason not too.

wwylele · 2019-09-08T12:15:49Z

src/video_core/renderer_opengl/gl_shader_disk_cache.h

+    /// Loads current game's precompiled cache. Invalidates on failure.
+    std::pair<ShaderDecompiledMap, ShaderDumpsMap> LoadPrecompiled();
+
+    /// Removes the transferable (and precompiled) cache file.


This comment indicates the function name is bad. Maybe rename it to InvalidateAll?

Hmm, I'm not sure I agree, but I'll change it to save on argument. I think it logically follows that if the Transferable cache is invalidated, then you have to invalidate the Precompiled as well. Renamed it anyway 👍

jroweboy · 2019-09-10T01:54:43Z

@chris062689 Added ui screenshot (Its a single checkbox in graphics)

@wwylele Addressed all concerns and fixed any known issues. Putting it on canary for user testing 👍

FearlessTobi · 2019-09-10T10:01:00Z

Just a nit, but can't we at least keep the checkbox labeling the same as it is in yuzu?

weihuoya · 2019-09-12T14:10:28Z

how about unseparated shader cache?
separated shader doesn't work well on android device.

BreadFish64 · 2019-11-05T00:36:12Z

src/common/CMakeLists.txt

+      -P "${CMAKE_SOURCE_DIR}/CMakeModules/GenerateSCMRev.cmake"
+    DEPENDS
+      # WARNING! It was too much work to try and make a common location for this list,
+      # so if you need to change it, please update CMakeModules/GenerateSCMRev.cmake as well


Should this warning be in CMakeModules/GenerateSCMRev.cmake as well?

Eh, a new developer won't be changing that file before changing src/common so i don't see a reason too.

BreadFish64 · 2019-11-05T00:36:12Z

src/common/file_util.cpp

@@ -720,16 +720,15 @@ const std::string& GetUserPath(UserPath path) {
        SetUserPath();
    return g_paths[path];
 }
-


Accidental?

BreadFish64 · 2019-11-05T00:36:12Z

src/common/file_util.cpp

@@ -546,7 +546,7 @@ std::optional<std::string> GetCurrentDir() {
 #endif
    free(dir);
    return strDir;
-}
+} // namespace FileUtil


Clang format always gets this wrong.

sorry but it really wants it https://travis-ci.org/citra-emu/citra/jobs/637758338#L576

BreadFish64 · 2019-11-05T00:36:12Z

src/common/file_util.cpp

+               path.end());
+    return std::string(RemoveTrailingSlash(path));
+}
+


Are we able to use std::filesystem for these?

we support mac sadly, so i don't want to deal with that. last i checked, mac had it under a different header, and it wasn't complete? but i could be wrong.

BreadFish64 · 2019-11-05T00:36:12Z

src/common/zstd_compression.cpp

+    return decompressed;
+}
+
+} // namespace Common::Compression


BreadFish64 · 2019-11-05T00:36:12Z

src/video_core/renderer_base.h

@@ -6,8 +6,9 @@

 #include <memory>
 #include "common/common_types.h"
-#include "core/core.h"
+#include "core/frontend/emu_window.h"


You have EmuWindow forward declared in this header already.

BreadFish64 · 2019-11-05T00:36:12Z

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -15,6 +15,7 @@
 #include "common/microprofile.h"
 #include "common/scope_exit.h"
 #include "common/vector_math.h"
+#include "core/core.h"


What is this included for?

BreadFish64 · 2019-11-05T00:36:12Z

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

+                LOG_ERROR(Render_OpenGL, "Failed to load transferable raw entry - skipping");
+                return {};
+            }
+            transferable.insert({entry.GetUniqueIdentifier(), {}});


BreadFish64 · 2019-11-05T00:36:12Z

src/video_core/renderer_opengl/gl_shader_manager.cpp

+    // Save VS to the disk cache if its a new shader
+    if (result) {
+        auto& disk_cache = impl->disk_cache;
+        ProgramCode program_code{setup.program_code.begin(), setup.program_code.end()};


Isn't this equivalent to just ProgramCode program_code{setup.program_code};

BreadFish64 · 2019-11-05T00:36:12Z

src/video_core/renderer_opengl/gl_shader_manager.h

 namespace OpenGL {

+class ShaderDiskCacheOpenGL;
+


Neither of these appear to used in this header.

jroweboy · 2019-11-28T18:48:13Z

@weihuoya I'm aware, but any android specific changes aren't a part of this PR and will be published with the rest of the android changes.

John-Gee · 2019-12-05T11:05:51Z

Hello,

I have a weird issue on Linux that I think is related to this PR, at least my bisection seems to say so as it does not happen on Canary with building at the previous commit. But disabling disk caching in the options changes nothing.

When I switch virtual desktops, Citra's performance drop to probably about the same as if I disable the hardware renderer, obviously I cannot see the FPS counter but the audio is super slow/broken apart/jittery. Now interestingly just alt-tabbing out of Citra does not cause that issue, only switching virtual desktops.

I've quickly tried with melonDS and The Dragon's Trap and I don't see that issue there, so I don't think it's my system's fault, but not completely sure either.

Specs wise:
I have a 4790k with a Radeon RX 580, running on Arch Linux 64b, using Mesa for the driver, and my DE is KDE's Plasma 5.17.4.

log: https://gist.github.com/John-Gee/9d38443f53575e1f19f9940d414b2d0d
The game I've tried in this log is Fire Emblem Echoes, but I've just tried with the much simpler (I think) Cave Story and it was the same.

I'll be happy to add any log/information needed or jump on chat if that's more helpful.

Thank you!

John-Gee · 2019-12-06T09:21:58Z

Based on:
#4940 (comment)

If I understand it correctly, I'm guessing the issue is from the 4940 PR, not this one, but since they are merged, it's hard for me to try either way.

jroweboy · 2019-12-15T23:25:02Z

@John-Gee I'm going to be merging #4940 since replicating the issue you mentioned is hard for me since i develop on windows and that PR is holding up a lot of upcoming changes. I'd still like to take a look at fixing that at some point, so please just open a new issue with your findings so we can track it down.

jroweboy · 2019-12-16T07:31:04Z

Known issues:

The cache doesn't take into account accurate mul setting, which can cause some weird situations with games that rely on it, but have a cache that wasn't compiled with it on. Currently investigating what the best route for fixing that is
A small number of games have a large number of shaders so load times are exasperated on those. Tomorrow I plan on looking at a fix for that, and if it doesn't look likely, then I can port my load screen over from yuzu.

B3n30

Did a first round of reviewing,

but I probably need to do a second one to look closer at some logic

B3n30 · 2020-01-12T16:56:55Z

src/common/zstd_compression.cpp

+namespace Common::Compression {
+
+std::vector<u8> CompressDataZSTD(const u8* source, std::size_t source_size, s32 compression_level) {
+    compression_level = std::clamp(compression_level, 1, ZSTD_maxCLevel());


magical number 1, is there no ZSTD_minCLevel()`? is there no level 0?

Funny enough there isn't in the version we used, but the recently added it upstream so i'm updating and using it here now.

B3n30 · 2020-01-12T16:56:55Z

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

+
+namespace OpenGL {
+
+using ShaderCacheVersionHash = std::array<u8, 64>;


magic number

B3n30 · 2020-01-12T16:56:55Z

src/video_core/renderer_opengl/gl_shader_disk_cache.h

+    explicit ShaderDiskCacheRaw(u64 unique_identifier, ProgramType program_type,
+                                RawShaderConfig config, ProgramCode program_code);
+    ShaderDiskCacheRaw();
+    ~ShaderDiskCacheRaw();


= default should be in header not in cpp. Same for other ctors/dtors in this file

B3n30 · 2020-01-12T16:56:55Z

src/video_core/renderer_opengl/gl_shader_disk_cache.h

+    std::string GetBaseDir() const;
+
+    /// Get current game's title id as u64
+    u64 GetProgramID();


Its not actually const. it has to load the program id if it exists from AppLoader.

B3n30 · 2020-01-12T16:56:55Z

src/video_core/renderer_opengl/gl_shader_disk_cache.h

+    std::string GetTitleID();
+
+    template <typename T>
+    bool SaveArrayToPrecompiled(const T* data, std::size_t length) {


I think this doesn't need a template and could just be const u8* data
Same for the function below

B3n30 · 2020-01-12T16:56:55Z

src/video_core/renderer_opengl/gl_shader_manager.cpp

+    std::vector<std::size_t> load_raws_index;
+    // Loads both decompiled and precompiled shaders from the cache. If either one is missing for
+    const auto LoadPrecompiledWorker =
+        [&](std::size_t begin, std::size_t end, const std::vector<ShaderDiskCacheRaw>& raws,


begin is unused

this can probably be made nicer (and maybe faster) by using iterators, same for the other lambdas below

The reason I made them lambdas is for threaded shader loading (which was a feature that I removed). This is also why begin is unused. I don't mind changing it back, but i think its more readable this way personally so you can see which blocks are what. Just my opinion though.

B3n30 · 2020-01-12T16:56:55Z

src/video_core/renderer_opengl/gl_shader_manager.cpp

+
+    compilation_failed = false;
+
+    const auto LoadTransferable = [&](std::size_t begin, std::size_t end,


unused begin

jroweboy · 2020-01-16T04:11:24Z

I've addressed all feedback and fixed the sanitize mul issue. I plan to merge this tomorrow so that we can start testing some other things. Thanks for the feedback!

wwylele reviewed Sep 8, 2019

View changes

jroweboy force-pushed the diskcachelul branch from 60c6186 to 9acadd5 Compare Sep 10, 2019

jroweboy added canary-merge and removed canary-merge labels Sep 10, 2019

WaterFace added a commit to WaterFace/citra that referenced this issue Sep 19, 2019

Merged citra-emu#4923 Disk Shader Cache

5ec28b8

BreadFish64 reviewed Nov 5, 2019

View changes

Jack-Rickwood mentioned this pull request Nov 24, 2019

[Build failing] Ignore format reinterpretation hack #4089 weihuoya/citra#2

Closed

jroweboy force-pushed the diskcachelul branch from e687aef to f642162 Compare Nov 28, 2019

jroweboy added the canary-merge label Nov 28, 2019

jroweboy removed the canary-merge label Nov 28, 2019

jroweboy force-pushed the diskcachelul branch from f642162 to 441f92f Compare Dec 3, 2019

jroweboy added the canary-merge label Dec 3, 2019

jroweboy mentioned this pull request Dec 3, 2019

Split Presentation thread from Render thread #4940

Merged

11 tasks

jroweboy force-pushed the diskcachelul branch from 441f92f to 7827219 Compare Dec 16, 2019

John-Gee mentioned this pull request Dec 16, 2019

PR #4940 Slows down the emulation a lot when switching Virtual Desktop #5026

Closed

B3n30 reviewed Jan 12, 2020

View changes

jroweboy added 3 commits Jan 16, 2020

Add ZSTD compression for precompiled cache

8f67d6a

Add settings for disk shader cache

b5bc057

Add shader cache version generation

b5b6ce5

jroweboy added 13 commits Jan 16, 2020

Add frontend settings

ce3f8bf

Add shader disk caching

4e9ec4e

Add missing changes from yuzu file_util

be52d3a

Add GetRenderer to core

d03d201

citra-qt: Load Disk Resources

d418f04

citra-sdl: Add Load Disk Resources

6f27560

Change VFS vector to regular vector

a20c81d

Properly bind the shader sampler and uniform bindings

061a334

Only add shaders to precompiled cache if they are new

7092ba8

Rename InvalidateTransferable to InvalidateAll to match what it does

2d86bc6

Add a error log for unsupported configurations for disk cache

45bc5b4

Log the number of entries in each cache

936094d

Address review and update zstd

6945b65

jroweboy force-pushed the diskcachelul branch from 7827219 to 6945b65 Compare Jan 16, 2020

Only load precompiled shaders if their sanitize_mul setting matches

cf4125a

jroweboy added 2 commits Jan 16, 2020

Fix title bar

43c9695

Only check for sanitize_mul if theres a shader in the cache

e95bc52

jroweboy merged commit e74a402 into citra-emu:master Jan 18, 2020
3 checks passed

jroweboy mentioned this pull request Jan 22, 2020

[Request] Asynchronous shader compilation and save the shaders cache for low spec systems #4357

Open

tywald mentioned this pull request Apr 8, 2020

Band-aid solution for 'Disk Shader Cache' #5188

Merged


		namespace OpenGL {

		using ShaderCacheVersionHash = std::array<u8, 64>;


		compilation_failed = false;

		const auto LoadTransferable = [&](std::size_t begin, std::size_t end,

Disk Shader Caching #4923

Disk Shader Caching #4923

Conversation

jroweboy commented Sep 7, 2019 • edited

flamesage commented Sep 7, 2019 • edited

wwylele commented Sep 8, 2019

wwylele Sep 8, 2019 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jroweboy commented Sep 10, 2019

FearlessTobi commented Sep 10, 2019

weihuoya commented Sep 12, 2019

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jroweboy commented Nov 28, 2019

John-Gee commented Dec 5, 2019 • edited

John-Gee commented Dec 6, 2019

jroweboy commented Dec 15, 2019

jroweboy commented Dec 16, 2019

B3n30 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jroweboy commented Jan 16, 2020

jroweboy commented Sep 7, 2019 •

edited

flamesage commented Sep 7, 2019 •

edited

wwylele Sep 8, 2019 •

edited

John-Gee commented Dec 5, 2019 •

edited