forward_progress_litmus_tests

Github Pages Website

This branch is linked to the github pages website. Changes pushed to this branch will automatically be published to the webpage

Dependancies

• rust ($ curl --proto '=https' --tlsv1.2 https://sh.rustup.rs -sSf | sh)
• rust crates (cargo update in this directory)

Running in the Browser

Steps to run:

• Navigate to the root of this repository
• Compile using the command wasm-pack build --target web. This will generate the webGPU bindings and javascript in the pkg/ directory.
• Make sure webGPU is enabled in chrome (chrome://gpu, webGPU should have hardware acceleration enabled)
• Start a local server in the root of this repo with python3 -m http.server
  • if you see something like 'failed to initialize webGPU' it probably means chrome is blocking the gpu because this is an http server. Launch a secure server (https, use the serve.py script) instead.
• view in the browser at localhost:8000/ (use chrome!).
• Select a test and run with the run button.
• If a test case finishes it will display the number of threads that terminated, otherwise it should display test failed after 15 seconds. It may not display this if the test case is seriously messing with chrome or the computer.

Running Natively

Steps to run:

• Navigate to the root of this repository
• to run the default test (2 threads 2 instructions test 5) run cargo run
• to run a specific test run cargo run <path_to_wgsl> <num_workgroups>
  • The wgsls are in src/tests/ and are sorted by progress model
  • Currently there is 1 thread per workgroup

Generating wgsls and the website

compile_wgsl.py takes a .txt file from the alloy_forward_progress directory and generates a wgsl. To use: python3 compile_wgsl.py -g 1 -tf /path/to/test.txt -o /path/to/outfile.wgsl

gen_website.py contains the script for generating all of the wgsls from the alloy_forward_progress repo and generating the HTML files for the website.

To generate everything: (wgsls, rust lib.rs, and HTML): python3 gen_website.py -c 1 --alloyfp_path <path_to_webtest_dir> -r 1 -o <path to src/lib.rs> -i 1

To generate the wgsls: python3 gen_website.py -c 1 --alloyfp_path <path_to_webtest_dir>

To generate the rust/wgpu code: python3 gen_website.py -r 1 -o <path to src/lib.rs>

To generate the HTML files: python3 gen_website.py -i 1

Note that wasm-pack should be re-run whenever changes to the wgsls or rust code are made.

To-Do (in no particular order)

• make the number of global buffer variables passed based on the max mem locations in the test cases (currently hard coded and some aren't used)
• Add support for intra-workgroup tests
• Create an all_runner for all of the tests so people only have to press one button
• Figure out how to actually terminate tests that hang in the browser. Currently if the test is unresponsive for 15 seconds the website will say the test failed, but won't directly terminate the process (the driver probably will but its not guarunteed)
• Reduce size of website
  • only compile a test once and have a csv or something that says which models it passes under, and call the kernel with that (currently a given wgsl can have multiple copies)
  • Put all of the javascript stuff in separate files instead of within the HTML
  • Find a way to dynamically load the wgsls instead of statically loading them at compile time
• Fix number of workgroups called (currently hardcoded as 2 0.o)