feat: use assign_regions api to speed up the witness assignment

[alannotnerd]

See #23

Bechmarking

For each phase (in AMD EPYC 7702 64-Core Processor x 128 threads):

before

ASSIGNMENT_TYPE: default Elapsed: 22.257506277s

after:

[2023-06-16T06:58:10Z INFO  poseidon_circuit::hash] hash table parallel version took 445.565348ms
[2023-06-16T07:03:29Z INFO  poseidon_circuit::hash] final state took 110.563839ms

around 556ms

[kunxian-xia]

Overall LGTM!

[kunxian-xia]

@noel2004 Is it true that each ((inp, control), check) occupies exactly one row in the fill_hash_tbl_body region?

poseidon-circuit/src/hash.rs

Line 507 in 89e9c04

for (i, ((inp, control), check)) in inputs_i.zip(controls_i).zip(checks_i).enumerate() {

[noel2004]

@noel2004 Is it true that each ((inp, control), check) occupies exactly one row in the fill_hash_tbl_body region?

poseidon-circuit/src/hash.rs

Line 507 in 89e9c04

for (i, ((inp, control), check)) in inputs_i.zip(controls_i).zip(checks_i).enumerate() {

Yes. Each entry occupies one row.

[kunxian-xia]

There remains two problems to be solved:

• parallelize all these permute to get final state (10s out of 60s),
• reduce 1st phase of fill_hash_tbl_body_partial to 0.

[kunxian-xia]

@noel2004 Is it true that each ((inp, control), check) occupies exactly one row in the fill_hash_tbl_body region?

poseidon-circuit/src/hash.rs

Line 507 in 89e9c04

for (i, ((inp, control), check)) in inputs_i.zip(controls_i).zip(checks_i).enumerate() {

Yes. Each entry occupies one row.

@alannotnerd The region height for each fill_hash_tble_body_partial is equal to data.len().

[alannotnerd]

Latest benchmark:

[2023-06-13T11:42:40Z INFO  halo2_proofs::circuit::floor_planner::single_pass] 25 sub_regions of hash table 2nd pass synthesis took 966.003307ms
[2023-06-13T11:42:40Z INFO  poseidon_circuit::hash] hash table parallel version took 972.602083ms
[2023-06-13T11:42:40Z INFO  halo2_proofs::circuit::floor_planner::single_pass] 24 sub_regions of permute state 2nd pass synthesis took 41.37005ms
[2023-06-13T11:42:40Z INFO  poseidon_circuit::hash] final state took 834.965283ms

[kunxian-xia]

On a machine with 128 threads, we get the following logs:

[2023-06-13T09:57:52Z INFO  halo2_proofs::circuit::floor_planner::single_pass] 128 sub_regions of hash table 2nd pass synthesis took 403.969017ms
[2023-06-13T09:57:52Z INFO  poseidon_circuit::hash] hash table parallel version took 456.107107ms

The running time of fill_hash_tbl_body part is reduced to about 0.5s.

[kunxian-xia]

We got the following stats when testing in scroll-zkevm (tested on a machine with 128 threads):

[2023-06-16T06:58:10Z INFO  poseidon_circuit::hash] hash table parallel version took 445.565348ms
[2023-06-16T07:00:47Z INFO  poseidon_circuit::hash] hash table parallel version took 452.829501ms
[2023-06-16T07:03:29Z INFO  poseidon_circuit::hash] hash table parallel version took 490.576667ms
[2023-06-16T06:58:10Z INFO  poseidon_circuit::hash] final state took 105.097895ms
[2023-06-16T07:00:47Z INFO  poseidon_circuit::hash] final state took 103.493175ms
[2023-06-16T07:03:29Z INFO  poseidon_circuit::hash] final state took 110.563839ms

[noel2004]

Has been merged into main