Make jellyfish wasm compatibility

[alxiong]

This issue aims to achieve wasm compatibility:

• [ ] finish #111
• [x] document slowdown (if any) and limitations in the wasm version.
• [x] add wasm compilation check in CI
• [ ] pay extra attention to entropy sampling (as /dev/random is no longer available, the usage of PRNG, seed length etc has to be more conservative chosen)

[alxiong]

While doing #318, on my local M1,

native benchmark:

$ cargo flamegraph --root --bench=plonk-benches --features test-srs

proving time for Bls12_381, PlonkType::TurboPlonk: 52120 ns/gate
proving time for Bls12_377, PlonkType::TurboPlonk: 53111 ns/gate
proving time for Bn254, PlonkType::TurboPlonk: 44123 ns/gate
proving time for BW6_761, PlonkType::TurboPlonk: 215765 ns/gate
proving time for Bls12_381, PlonkType::UltraPlonk: 79688 ns/gate
proving time for Bls12_377, PlonkType::UltraPlonk: 78137 ns/gate
proving time for Bn254, PlonkType::UltraPlonk: 69546 ns/gate
proving time for BW6_761, PlonkType::UltraPlonk: 341935 ns/gate
verifying time for Bls12_381, PlonkType::TurboPlonk: 2504441 ns
verifying time for Bls12_377, PlonkType::TurboPlonk: 2597979 ns
verifying time for Bn254, PlonkType::TurboPlonk: 3948429 ns
verifying time for BW6_761, PlonkType::TurboPlonk: 8686679 ns
verifying time for Bls12_381, PlonkType::UltraPlonk: 2658983 ns
verifying time for Bls12_377, PlonkType::UltraPlonk: 2570262 ns
verifying time for Bn254, PlonkType::UltraPlonk: 1644525 ns
verifying time for BW6_761, PlonkType::UltraPlonk: 9560779 ns
batch verifying time for Bls12_381, PlonkType::TurboPlonk, 1000 proofs: 15505 ns/proof
batch verifying time for Bls12_377, PlonkType::TurboPlonk, 1000 proofs: 16257 ns/proof
batch verifying time for Bn254, PlonkType::TurboPlonk, 1000 proofs: 16841 ns/proof
batch verifying time for BW6_761, PlonkType::TurboPlonk, 1000 proofs: 31831 ns/proof
batch verifying time for Bls12_381, PlonkType::UltraPlonk, 1000 proofs: 17502 ns/proof
batch verifying time for Bls12_377, PlonkType::UltraPlonk, 1000 proofs: 18538 ns/proof
batch verifying time for Bn254, PlonkType::UltraPlonk, 1000 proofs: 16461 ns/proof
batch verifying time for BW6_761, PlonkType::UltraPlonk, 1000 proofs: 37605 ns/proof

targeting wasm:

$ cargo flamegraph --root -o wasm-flamegraph.svg --bench=plonk-benches --no-default-features --features test-srs -- --target=wasm32-unknown-unknown

proving time for Bls12_381, PlonkType::TurboPlonk: 228241 ns/gate
proving time for Bls12_377, PlonkType::TurboPlonk: 226454 ns/gate
proving time for Bn254, PlonkType::TurboPlonk: 173796 ns/gate
proving time for BW6_761, PlonkType::TurboPlonk: 925769 ns/gate
proving time for Bls12_381, PlonkType::UltraPlonk: 349392 ns/gate
proving time for Bls12_377, PlonkType::UltraPlonk: 338768 ns/gate
proving time for Bn254, PlonkType::UltraPlonk: 275681 ns/gate
proving time for BW6_761, PlonkType::UltraPlonk: 1245619 ns/gate
verifying time for Bls12_381, PlonkType::TurboPlonk: 2516504 ns
verifying time for Bls12_377, PlonkType::TurboPlonk: 2743795 ns
verifying time for Bn254, PlonkType::TurboPlonk: 1684900 ns
verifying time for BW6_761, PlonkType::TurboPlonk: 11598887 ns
verifying time for Bls12_381, PlonkType::UltraPlonk: 2768262 ns
verifying time for Bls12_377, PlonkType::UltraPlonk: 3081320 ns
verifying time for Bn254, PlonkType::UltraPlonk: 1837012 ns
verifying time for BW6_761, PlonkType::UltraPlonk: 12966108 ns
batch verifying time for Bls12_381, PlonkType::TurboPlonk, 1000 proofs: 52437 ns/proof
batch verifying time for Bls12_377, PlonkType::TurboPlonk, 1000 proofs: 52356 ns/proof
batch verifying time for Bn254, PlonkType::TurboPlonk, 1000 proofs: 47650 ns/proof
batch verifying time for BW6_761, PlonkType::TurboPlonk, 1000 proofs: 91287 ns/proof
batch verifying time for Bls12_381, PlonkType::UltraPlonk, 1000 proofs: 61210 ns/proof
batch verifying time for Bls12_377, PlonkType::UltraPlonk, 1000 proofs: 60756 ns/proof
batch verifying time for Bn254, PlonkType::UltraPlonk, 1000 proofs: 55146 ns/proof
batch verifying time for BW6_761, PlonkType::UltraPlonk, 1000 proofs: 106205 ns/proof

The caveats is that I'm not using any wasm runtime like wasmtime, so I'm not sure how accurate is the wasm target binary benchmark, will investigate more later. But on the surface, it seems that inside wasm is 2~2.5x slower.

[mike1729]

You sure that you actually run wasm code? If you don't provide any runtime how the wasm code is executed? My guess is that --target wasm32-unknow-unknow is ignored and you run native bench but by specifying --no-default-features you disabled parallel feature and this is the cause of worse performance. From my experiments, it seems that wasm performance is ~10x worse comparing to native without rayon.

[alxiong]

yeah, I suspect you are right. so we really need to run inside wasmtime to get an accurate benchmark? @mike1729

[mike1729]

I believe so. What I did to get perf data that you can then feed to flamegraph was:

1. cargo build --release --target wasm32-unknown-unknow
2. perf record -k mono wasmtime --profile=jitdump prover.wasm
3. perf inject --jit --input perf.data --output perf.jit.data

Then you can either inspect perf.jit.data with perf perf report --input perf.jit.data or use some frontend.

Note however, that the above method works without parallel feature as it is hard to get rayon working in wasm.