commit 55e4c2ce7d6c144d9fd54466896f39e702c29dfe
author Scott Todd <scotttodd@google.com> Mon May 16 09:48:52 2022 -0700
committer GitHub <noreply@github.com> Mon May 16 09:48:52 2022 -0700
tree 9a8d78dc57df924f2292f9efb33f0bc0343b0d36
parent 31fd07157fef2c3cf5dee727685e69fdfde3f36c

Add initial benchmarking features to WebAssembly experiments. (#9100)

### Background

I'm working towards identifying and quantifying key metrics for our web platform support. Metrics of interest thus far include:
* runtime binary size
* program binary size
* runtime startup time
* program load time
* total function call time (as seen from a JavaScript application)
* no-overhead function call time (`iree_runtime_call_invoke()`)

Once we have baselines for those, we can start deeper analysis and optimization work. Some optimization work may be purely in the IREE compiler (e.g. codegen, Flow/Stream/HAL dialects, etc.), while other work may be in the web port itself (e.g. Emscripten flags, use of web APIs, runtime JS bindings).

### Current Status

This is still all experimental, but I'd like to checkpoint what I've built so far and let other people try it out. Expect rough edges (e.g. some Windows/Linux paths specific to my setup that can be overwritten).

Summary of changes:

* `sample_dynamic/index.html` now supports interactive benchmarking
  * a "benchmark iterations" form input drives a loop around `iree_runtime_call_invoke()` down in C/Wasm
  * timing information is rendered onto the page itself, instead of being logged to stdout / the console
  * sample screenshot: https://user-images.githubusercontent.com/4010439/167958736-228a1541-8ed6-4b2c-9af9-55ef0b10bf74.png
* `generate_web_metrics.sh` imports and compiles programs from our existing benchmark suite, preserving all sorts of artifacts for further manual inspection or automated use
* `run_native_benchmarks.sh` executes the compiled native programs from `generate_web_metrics.sh`
* `sample_dynamic/benchmarks.html` loads and runs each compiled Wasm program from `generate_web_metrics.sh`
  * Sample output: https://gist.github.com/ScottTodd/f2bb1f274c5895c8f979400abd6d2b67

Also of note: I haven't really looked at the browser profiling tools yet. I expect those will help with detailed analysis while the general scaffolding and `performance.now()` measurements will help with comparisons between frameworks and across browsers/devices.