search

EmbarkStudios / rust-gpu

🐉 Making Rust a first-class language and ecosystem for GPU shaders 🚧
https://shader.rs
Apache License 2.0
7.35k stars 244 forks source link

`rustc_codegen_spirv` taking a long time processing my (large) shader #851

Open hatoo opened 2 years ago

hatoo commented 2 years ago

This issue is from Discord conversation https://discordapp.com/channels/750717012564770887/750717499737243679/943879544006381638

Expected Behaviour

I want to compile my shader faster.

Example & Steps To Reproduce

  1. clone my repo https://github.com/hatoo/rene
  2. cargo build --release -Z timings
  3. See compile duration

I tested on af3cf2b74e8ef4f4bd98a3cd2cda64b9379d20d7.

Completed rene v0.1.0 build script (run) in 263.2s

In console in VSCode and rust-analyzer plugin. It takes more time.

Completed rene v0.1.0 build script (run) in 407.3s

System Info

hatoo commented 2 years ago

After I added codes to my shader, it takes very longer time to compile. In this commit, it takes about 1000s.

Completed rene v0.1.0 build script (run) in 1009.6s

And I added one commit to this in this branch. I've waited for few hours and the build didn't end.

hatoo commented 2 years ago

I enabled -Zself-profile and build it and I found rustc is spending almost all time to link (link_block_ordering_pass_and_mem2reg).

Procedure

  1. Enable -Zself-profile on spirv-builder. It can be done by like https://github.com/hatoo/rust-gpu/commit/e34bda56b12cc86533e92ae57a754f6ae896e0b1
  2. Change rene's spirv-builder dependency to use it.
  3. cargo install --git https://github.com/rust-lang/measureme crox flamegraph summarize
  4. cargo build --release on rene
  5. crox .\rene_shader-xxxx.mm_profdata and chrome_profiler.json will be produced.
  6. Open Chrome Developper tool and click the Performance tab. And upload chrome_profiler.json by clicking ⬆️ button.

image

eddyb commented 1 year ago

I've been working on a mem2reg replacement recently, relying on SPIR-T qptrs (based on https://github.com/EmbarkStudios/spirt/pull/29 / https://github.com/EmbarkStudios/spirt/pull/41 - no PR yet for the mem2reg replacement though, but I really hope I can get it into Rust-GPU 0.9, behind the qptr opt-in).

$ rg mem2reg 2023-07-08-spirt-disagg-baseline 
512:time:   0.011; rss:  187MB ->  187MB (   +0MB)      link_block_ordering_pass_and_mem2reg-before-inlining
515:time: 126.704; rss:  189MB ->  158MB (  -31MB)      link_block_ordering_pass_and_mem2reg-after-inlining
$ rg qptr::partition_and_propagate 2023-07-08-spirt-disagg-qptr-pnp 
538:time:   0.570; rss:  220MB ->  221MB (   +1MB)      qptr::partition_and_propagate

so that's a 222x speed-up (for the same version of rene-shader I used to demo the initial impact from SPIR-T, over a year ago)

(RUSTGPU_CODEGEN_ARGS="--spirt-passes=qptr --no-infer-storage-classes --no-legacy-mem2reg" is the opt-in, at least on my local branch, I'm hoping we can at least simplify it to RUSTGPU_CODEGEN_ARGS=--qptr for 0.9)

EDIT: to be perfectly clear: those flags are not useful without the necessary combination of SPIR-T + Rust-GPU changes and I don't even have an up-to-date Rust-GPU branch anywhere comparable to that one.