refactor(avm): kernel trace and finalization

[fcarreiro]

The original objective of this PR was to just remove kernel finalize() from trace.cpp. Other things were done on top but I do not claim to have solved or constrained everything.

Something that I realized while working on this: right now gas is broken for SLOAD/SSTORE/UNENCRYPTEDLOG because they still either use SET or take multiple lines (or should). However, if they take multiple lines, we have to move them to their own gadget, and if we do so, then their kernel accesses (1 per item) will make the kernel trace not be 1-1 anymore. So, we might need to rethink this whole trace.

---

PIL

• Made kernel trace officially virtual
• Moved most constraints, lookups, etc from main to kernel
• Changed a few constraints; now we will constrain all rows of the execution trace (this bleeds to "last + 1" using the shifts, on purpose)
• Changed some names and added some comments to make it more understandable, but still needs love

CPP

• Encapsulated kernel finalization in its own finalize
• Changed sel_last to actually be the last row of the execution trace, as PIL says. (in any case, it's not used now)
• Other misc changes:
  • error reporting in check circuit now only reports first failure for each subrelation (otherwise rows were spammed, but it didn't mean that the actual row didn't satisfy it, just that the accumulator was already "dirty")
  • Moved execution hints to their own hpp

[fcarreiro]

• refactor(avm): kernel trace and finalization #8049 👈
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.

Join @fcarreiro and the rest of your teammates on Graphite

[AztecBot]

Benchmark results

Metrics with a significant change:

• avm_simulation_time_ms (Token:transfer_public): 28.4 (+38%)

Detailed results

All benchmarks are run on txs on the Benchmarking contract on the repository. Each tx consists of a batch call to create_note and increment_balance, which guarantees that each tx has a private call, a nested private call, a public call, and a nested public call, as well as an emitted private note, an unencrypted log, and public storage read and write.

This benchmark source data is available in JSON format on S3 here.

Proof generation

Each column represents the number of threads used in proof generation.

Metric	1 threads	4 threads	16 threads	32 threads	64 threads
proof_construction_time_sha256_ms	5,789	1,562	712 (-3%)	751 (-7%)	769 (-4%)
proof_construction_time_sha256_30_ms	11,935	3,204	1,428 (-2%)	1,446 (-2%)	1,480 (-6%)
proof_construction_time_sha256_100_ms	45,865	12,036	5,495 (-3%)	5,442 (-3%)	5,408 (-4%)
proof_construction_time_poseidon_hash_ms	79.0	34.0 (-3%)	34.0 (-3%)	59.0 (-3%)	88.0 (-2%)
proof_construction_time_poseidon_hash_30_ms	1,539	424	204 (-1%)	226 (-3%)	271 (-3%)
proof_construction_time_poseidon_hash_100_ms	5,666	1,520	681 (-2%)	727 (-4%)	749 (-3%)

L2 block published to L1

Each column represents the number of txs on an L2 block published to L1.

Metric	4 txs	8 txs	16 txs
l1_rollup_calldata_size_in_bytes	4,324	7,844	14,852
l1_rollup_calldata_gas	49,696	92,474	177,596
l1_rollup_execution_gas	1,383,190	2,130,936	3,958,097
l2_block_processing_time_in_ms	256 (-1%)	452 (+2%)	816 (+3%)
l2_block_building_time_in_ms	9,341 (+1%)	18,223 (+1%)	36,039
l2_block_rollup_simulation_time_in_ms	9,341 (+1%)	18,222 (+1%)	36,039
l2_block_public_tx_process_time_in_ms	7,841	16,658 (+1%)	34,461

L2 chain processing

Each column represents the number of blocks on the L2 chain where each block has 8 txs.

Metric	3 blocks	5 blocks
node_history_sync_time_in_ms	3,012 (+5%)	3,953 (+3%)
node_database_size_in_bytes	12,628,048	16,740,432
pxe_database_size_in_bytes	16,254	26,813

Circuits stats

Stats on running time and I/O sizes collected for every kernel circuit run across all benchmarks.

Circuit	simulation_time_in_ms	witness_generation_time_in_ms	input_size_in_bytes	output_size_in_bytes	proving_time_in_ms
private-kernel-init	96.7 (+2%)	432 (+7%)	21,673	44,858	N/A
private-kernel-inner	163	717 (-5%)	72,309	45,005	N/A
private-kernel-reset-tiny	501	897 (-1%)	65,974	44,844	N/A
private-kernel-tail	207	166 (-1%)	50,728	52,256	N/A
base-parity	5.56 (-1%)	N/A	160	96.0	N/A
root-parity	33.5	N/A	69,084	96.0	N/A
base-rollup	2,858 (+1%)	N/A	187,817	664	N/A
root-rollup	38.9 (-2%)	N/A	54,525	716	N/A
public-kernel-setup	98.4 (-1%)	N/A	103,760	71,222	N/A
public-kernel-app-logic	104 (+1%)	N/A	103,599	71,222	N/A
public-kernel-tail	576	N/A	409,190	16,414	N/A
private-kernel-reset-small	495	N/A	66,085	45,629	N/A
private-kernel-tail-to-public	1,062	662 (-1%)	492,960	1,697	N/A
public-kernel-teardown	92.7	N/A	104,005	71,222	N/A
merge-rollup	19.4 (+2%)	N/A	35,742	664	N/A
undefined	N/A	N/A	N/A	N/A	66,845 (+4%)

Stats on running time collected for app circuits

Function	input_size_in_bytes	output_size_in_bytes	witness_generation_time_in_ms
ContractClassRegisterer:register	1,344	11,731	345 (+1%)
ContractInstanceDeployer:deploy	1,408	11,731	18.6 (+2%)
MultiCallEntrypoint:entrypoint	1,920	11,731	427 (+1%)
FeeJuice:deploy	1,376	11,731	392
SchnorrAccount:constructor	1,312	11,731	108 (+3%)
SchnorrAccount:entrypoint	2,304	11,731	435 (+1%)
Token:privately_mint_private_note	1,280	11,731	139 (-3%)
FPC:fee_entrypoint_public	1,344	11,731	29.2 (+10%)
Token:transfer	1,312	11,731	276 (+1%)
Benchmarking:create_note	1,344	11,731	105 (+5%)
SchnorrAccount:verify_private_authwit	1,280	11,731	28.4 (-28%)
Token:unshield	1,376	11,731	581 (+3%)
FPC:fee_entrypoint_private	1,376	11,731	772 (+1%)

AVM Simulation

Time to simulate various public functions in the AVM.

Function	time_ms	bytecode_size_in_bytes
FeeJuice:_increase_public_balance	53.5 (-1%)	8,139
FeeJuice:set_portal	11.7 (+5%)	2,362
Token:constructor	81.9 (-2%)	31,107
FPC:constructor	54.0 (+1%)	22,380
FeeJuice:mint_public	38.3 (-1%)	6,150
Token:mint_public	44.0	11,720
Token:assert_minter_and_mint	68.7 (-3%)	8,028
AuthRegistry:set_authorized	46.1 (+22%)	4,537
FPC:prepare_fee	236 (-2%)	8,812
Token:transfer_public	⚠️ 28.4 (+38%)	47,358
FPC:pay_refund	58.9 (-15%)	12,082
Benchmarking:increment_balance	973 (+1%)	7,450
Token:_increase_public_balance	39.4	8,960
FPC:pay_refund_with_shielded_rebate	67.1 (-1%)	12,631

Public DB Access

Time to access various public DBs.

Function	time_ms
get-nullifier-index	0.158

Tree insertion stats

The duration to insert a fixed batch of leaves into each tree type.

Metric	1 leaves	16 leaves	64 leaves	128 leaves	256 leaves	512 leaves	1024 leaves
batch_insert_into_append_only_tree_16_depth_ms	2.23 (+1%)	4.03 (+3%)	N/A	N/A	N/A	N/A	N/A
batch_insert_into_append_only_tree_16_depth_hash_count	16.8	31.7	N/A	N/A	N/A	N/A	N/A
batch_insert_into_append_only_tree_16_depth_hash_ms	0.116 (+1%)	0.114 (+3%)	N/A	N/A	N/A	N/A	N/A
batch_insert_into_append_only_tree_32_depth_ms	N/A	N/A	11.6 (+1%)	18.9 (+5%)	34.9 (+10%)	60.8 (+3%)	115 (+3%)
batch_insert_into_append_only_tree_32_depth_hash_count	N/A	N/A	95.9	159	287	543	1,055
batch_insert_into_append_only_tree_32_depth_hash_ms	N/A	N/A	0.112 (+1%)	0.110 (+5%)	0.112 (+10%)	0.105 (+4%)	0.102 (+2%)
batch_insert_into_indexed_tree_20_depth_ms	N/A	N/A	14.9 (+2%)	26.5 (+3%)	44.6 (+1%)	85.3 (+3%)	164 (+4%)
batch_insert_into_indexed_tree_20_depth_hash_count	N/A	N/A	109	207	355	691	1,363
batch_insert_into_indexed_tree_20_depth_hash_ms	N/A	N/A	0.112 (+2%)	0.107 (+4%)	0.108 (+5%)	0.106 (+4%)	0.104 (+4%)
batch_insert_into_indexed_tree_40_depth_ms	N/A	N/A	16.8 (+3%)	N/A	N/A	N/A	N/A
batch_insert_into_indexed_tree_40_depth_hash_count	N/A	N/A	132	N/A	N/A	N/A	N/A
batch_insert_into_indexed_tree_40_depth_hash_ms	N/A	N/A	0.109 (+3%)	N/A	N/A	N/A	N/A

Miscellaneous

Transaction sizes based on how many contract classes are registered in the tx.

Metric	0 registered classes	1 registered classes
tx_size_in_bytes	64,779	668,997

Transaction size based on fee payment method

| Metric | |
| - | |

[AztecBot]

⚠️ Performance Alert ⚠️

Possible performance regression was detected for benchmark 'C++ Benchmark'.
Benchmark result of this commit is worse than the previous benchmark result exceeding threshold 1.05.

Benchmark suite	Current: ebaf8de	Previous: b7b76ff	Ratio
nativeClientIVCBench/Full/6	14290.349651000013 ms/iter	13266.748604000015 ms/iter	1.08
nativeconstruct_proof_ultrahonk_power_of_2/20	5637.750612000005 ms/iter	4913.0781789999955 ms/iter	1.15
wasmconstruct_proof_ultrahonk_power_of_2/20	16483.796306999997 ms/iter	14482.776602 ms/iter	1.14

This comment was automatically generated by workflow using github-action-benchmark.

CC: @ludamad @codygunton

[jeanmon]

Nice refactoring and I like the way we process finalization. Please have a look at my feedback before merging.

[jeanmon]

My understanding is that this algorithm is assuming that clk fields are sorted for both traces. It might be worth to mention this above. Typically, memory trace will not be sorted by clk in its final form.

[fcarreiro]

That's right! Added a comment.

[jeanmon]

I am reading 'fall' like fallthrough whil we meant "for all rows in exec trace". Could we find another wording?

[fcarreiro]

Ah good point. It was meant to be func_all; I renamed.

[jeanmon]

Destination field is named main_trace, in this case, should not we remove the template type D and use the type of main_trace? If this is not always used for the main trace, then let us rename the argument by dest.

[jeanmon]

Is the D template used to avoid import of AvmFullRow?

[fcarreiro]

I changed it to std::vector<AvmFullrow<FF>>. It was both to avoid the full row and the vector (anything that can be iterated and has size() would've worked). In any case, we can live with this and will generalize when the need comes.

[jeanmon]

Very good!

[fcarreiro]

Merge activity

• Aug 19, 7:10 AM EDT: @fcarreiro merged this pull request with Graphite.