bench: instrument hash counts for verifier

[jonathanpwang]

I was able to find a way to do this without forking Plonky3. The two traits that need to be implemented for PCS hashes are PseudoCompressionFunction and CryptographicHasher. I make a wrapper struct Instrumented that does the normal operation and also updates the counter. For logging / debugging purposes, I store the lengths of inputs by the input type using type_name. To translate into actual counts, some specific logic about the absorb rate, etc of the specific hash (currently poseidon2) are used.

Notes:

• CryptographicHasher is for flat hashing long arrays of inputs. I think this is mostly used on the merkle leaf.
• PseudoCompressionFunction is what is used on the other layers of merkle tree, since it's designed so 2-to-1 hashes have same input and output shape.

What is currently benchmarked: given a matrix with height rows and field_width + 1 columns, we make two AIRs:

1. Just one big trace matrix, and with a receive interaction (this adds 2 virtual columns and gates for logup partial sum) where the receive is on the vector of field_width with count the column 0 value.
2. Treat the height-by-field_width matrix as a cached trace, so it is committed to separately. The count matrix that is height-by-1 is a separate matrix that is committed.

We run both of these and log the hashes used, where we count based on the number of Poseidon2 permutations done.

• Add output to a csv so we can just on a bunch and post-process in spreadsheet
• Modify bench to time stages of proving. This should not be instrumented because the counter is pretty slow.
• Instrument challenger -- this is where grinding is done

[jonathanpwang]

Some preliminary stats:

log_blowup_factor: 3, num_queries: 33, proof_of_work_bits: 16

Field width: 1, height: 65536 = 1 << 16
Without cached trace:
Hash: (p3_baby_bear::baby_bear::BabyBear, [p3_baby_bear::baby_bear::BabyBear; 8]), Count: 627
Compress: [p3_baby_bear::baby_bear::BabyBear; 8], Count: 7425
Total Count: 8052
With cached trace:
Hash: (p3_baby_bear::baby_bear::BabyBear, [p3_baby_bear::baby_bear::BabyBear; 8]), Count: 660
Compress: [p3_baby_bear::baby_bear::BabyBear; 8], Count: 8052
Total Count: 8712
(Difference in total count: 660)


Field width: 1, height: 1048576 = 1 << 20
Without cached trace:
Hash: (p3_baby_bear::baby_bear::BabyBear, [p3_baby_bear::baby_bear::BabyBear; 8]), Count: 759
Compress: [p3_baby_bear::baby_bear::BabyBear; 8], Count: 10527
Total Count: 11286
With cached trace:
test cached_lookup::instrumented::bench_cached_trace has been running for over 60 seconds
Hash: (p3_baby_bear::baby_bear::BabyBear, [p3_baby_bear::baby_bear::BabyBear; 8]), Count: 792
Compress: [p3_baby_bear::baby_bear::BabyBear; 8], Count: 11286
Total Count: 12078
(Difference in total count: 792)

[yi-sun]

To confirm, in case 2, is the logUp still done?

[jonathanpwang]

To confirm, in case 2, is the logUp still done?

In both cases, the logUp is exactly the same: same 2 virtual columns, same constraints. The logUp treats the main trace as one big matrix, and doesn't care about the partition.

[yi-sun]

Great, I think that's the correct relevant comparison

[jonathanpwang]

Uploaded instrumented verifier csv to https://docs.google.com/spreadsheets/d/1AEwbUvsge07qq9ASi0BL4Or0bbuwq8Iy9cy5cNxiLj4/edit?usp=sharing

[jonathanpwang]

@OsamaAlkhodairy FYI I'm going to merge this one.

I added some stuff to test-utils/src/config to start some tooling. I added a StarkEngine trait with the idea we can make a different engine for each field x hash combo. I moved your run_simple_test function to config::baby_bear_poseidon2 because the test depends on the choice of field (babybear) and hash (poseidon2), so I added it both as a function in StarkEngine trait but also it's still nice to have the do-it-all command you had before, so I simply moved its location.