Repository navigation
v8x16.shuffle optimizations needed #2531
Description
I translated the IDCT SSE code into Wasm. The algorithm uses lots of various punpckxxxx instructions, though WebAssembly has v8x16.shuffle. The v8 lowers into native SSE2 equivalents by matching immediate argument. I cannot find if we do it for any of the cranelift backends.
STR:
- Use test case at https://github.com/yurydelendik/zbar-wasm/raw/0083a9a48c8c06e5555424d85f71ce5a4b560145/zbar_jpeg/test.wasm
- Run
time wasmtime run --enable-simd test.wasm --invoke test500
Observe the time; it is about 15 sec here. Node runs test.wasm (_initialize + test500) in about 11 sec here.
It is expected that wasmtime/cranelift will improve the performance by using specialized SSE2 instructions by 40-50%.
Activity
abrown commentedon Dec 30, 2020
I agree; perhaps this benchmark could be adapted for sightglass?
yurydelendik commentedon Dec 30, 2020
Not familiar with the sightglass. Can you sketch what needs to be done?
abrown commentedon Dec 30, 2020
Here is a high-level document describing the basic idea and here is an example.
alexcrichton commentedon Jun 22, 2023
Can confirm that all these shuffles are now implemented, even on aarch64 too. All
i8x16.shuffleinstructions present in the above module are compiled to single-instruction lowerings on both x86_64 and aarch64. In that case I'm going to close this.