feat(bb): consider polynomial end_index when constructing partially evaluated multivariates by fcarreiro · Pull Request #12530 · AztecProtocol/aztec-packages

fcarreiro · 2025-03-06T15:03:08Z

This fixes the sumcheck memory peak.

Takes AVM memory usage for a trivial transaction from 33GB (projected 100GB on full VM) to 5GB.

Hopefully also useful for other flavors, when the trace is not full.

This also improves sumcheck performance (thanks @jeanmon and @lucasxia01 for the suggestions on this)

AVM (16 cores, short trace) - Before: 28s
AVM (16 cores, short trace) - After: 16s

fcarreiro · 2025-03-06T15:03:25Z

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

fcarreiro · 2025-03-06T15:06:01Z

barretenberg/cpp/src/barretenberg/eccvm/eccvm_flavor.hpp

        }
    };

+    /**


Just moved because this class refers to ProverPolynomials.

barretenberg/cpp/src/barretenberg/polynomials/polynomial.hpp

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

barretenberg/cpp/src/barretenberg/stdlib_circuit_builders/ultra_flavor.hpp

lucasxia01 · 2025-03-06T17:21:30Z

barretenberg/cpp/src/barretenberg/sumcheck/partial_evaluation.test.cpp

    FF expected_lo = v00 * (FF(1) - round_challenge_0) + v10 * round_challenge_0;
    FF expected_hi = v01 * (FF(1) - round_challenge_0) + v11 * round_challenge_0;

+    sumcheck.partially_evaluated_polynomials = typename Flavor::PartiallyEvaluatedMultivariates(multivariate_n);


why reassign? should probably just change the constructor in sumcheck itself

oh I see, that constructor was just removed and moved to the prove() function

Yeah this test itself is not very nice. It tests "partially_evaluate" on itself, but that's not how the usual flow goes.

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

lucasxia01

Thanks for taking this on; the memory savings are fantastic. There's still a little room for optimization I believe. You actually allocate up to double the memory necessary. I also realize that a lot of these changes were made with the assumption that the polynomials are nonzero starting from 0, rather than some arbitrary start_index. That's fine though, we can keep it like that for now since that's the main use case for this optimization.

lucasxia01 · 2025-03-06T18:08:43Z

barretenberg/cpp/src/barretenberg/eccvm/eccvm_flavor.hpp

            for (auto [poly, full_poly] : zip_view(get_all(), full_polynomials.get_all())) {
-                size_t desired_size = std::min(full_poly.end_index(), halved_circuit_size);
-                poly = Polynomial(desired_size, halved_circuit_size);
+                poly = Polynomial(full_poly.end_index() / 2, circuit_size / 2);


looking at this again, I would be wary of off-by-issues. I think if the end_index was odd like it was 7, then, we would actually want the poly size to be 4, not 3.

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

lucasxia01 · 2025-03-06T18:35:59Z

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

            multivariate_challenge.emplace_back(round_challenge);
            // Prepare sumcheck book-keeping table for the next round
-            partially_evaluate(partially_evaluated_polynomials, round.round_size, round_challenge);
+            partially_evaluate(partially_evaluated_polynomials, round_idx, round_challenge);


there's confusingly 2 prove functions (and many other functions) - one for normal, one with zk. So both need to be updated similarly

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

lucasxia01 · 2025-03-06T18:48:43Z

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

-                pep_view[j].at(i >> 1) =
-                    polynomials[j][i] + round_challenge * (polynomials[j][i + 1] - polynomials[j][i]);
+            const auto& poly = polynomials[j];
+            size_t actual_end_index = (poly.end_index() + 1) / (1 << (round_index - 1));


I think this is the wrong partially_evaluate to update? No idea what PolynomialT is and where this is used...

is this only used in the partially evaluation test? if so, maybe we can just get rid of this function...

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

lucasxia01 · 2025-03-06T23:05:17Z

barretenberg/cpp/src/barretenberg/numeric/bitop/division.hpp

@@ -0,0 +1,10 @@
+#pragma once


nit: not really a bitop but its fine

he yes I know, but didn't want to create a whole new directory. Give me a name for it and I'll create it :)

If we create a function for this, I would like it to be "overflow/underflow" safe.
If we restrict such a function to work on "unsigned types" then one elegant solution is:

q = x/y + (x % y != 0); (Credit to https://stackoverflow.com/a/14878734)

Or for any signed type integer:

int div = a / b; if (((a ^ b) >= 0) && (a % b != 0)) div++;

(Credit to https://stackoverflow.com/a/924160)

jeanmon

LGTM and great improvement!

lucasxia01

Looks good! Agree with Jean that the optimizations we were suggesting should be possible, but also agree that we can just get this in first.

lucasxia01 · 2025-03-10T15:57:21Z

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp

-            // Prepare sumcheck book-keeping table for the next round
-            partially_evaluate(partially_evaluated_polynomials, round.round_size, round_challenge);
+            // Prepare sumcheck book-keeping table for the next round.
+            partially_evaluate(partially_evaluated_polynomials, multivariate_n >> round_idx, round_challenge);


why this change? to make it more clear what this is?

Ah, it is kind of a leftover of many changes that I had and rolled back. However now that you mention it, it is kind of nice to use multivariate_n in round 0 and multivariate_n >> round_idx here. I'll keep it just because CI is green but feel free to change!

lucasxia01 · 2025-03-10T16:02:28Z

Partially handles AztecProtocol/barretenberg#1120, but maybe a followup will fully close it

🤖 I have created a new Aztec Packages release --- ## [0.79.0](v0.78.1...v0.79.0) (2025-03-11) ### ⚠ BREAKING CHANGES * aggregate data for batch calls ([#12562](#12562)) ### Features * add extra attributes to target_info ([#12583](#12583)) ([c296422](c296422)) * add optional oracle resolver url in `acvm_cli` (noir-lang/noir#7630) ([cc6cdbb](cc6cdbb)) * allow to pay via sponsored fpc from cli ([#12598](#12598)) ([877de5c](877de5c)) * array concat method (noir-lang/noir#7199) ([cc6cdbb](cc6cdbb)) * **avm:** ToRadix gadget ([#12528](#12528)) ([02a7171](02a7171)) * aztec-up -v flag ([#12590](#12590)) ([6a41565](6a41565)) * **bb:** consider polynomial end_index when constructing partially evaluated multivariates ([#12530](#12530)) ([abd22cd](abd22cd)) * **config:** add fallbacks ([#12593](#12593)) ([f2f9ef3](f2f9ef3)) * **p2p:** add trusted peers mechanics ([#12447](#12447)) ([d67f7e8](d67f7e8)) * **p2p:** peer manager peer count metrics ([#12575](#12575)) ([b4891c1](b4891c1)) * provision alerts ([#12561](#12561)) ([2ea1767](2ea1767)) * Resolve callstacks in protocol circuit errors on wasm ([#12573](#12573)) ([657299b](657299b)) ### Bug Fixes * aggregate data for batch calls ([#12562](#12562)) ([bd0b3b6](bd0b3b6)) * broken kind transfer test ([#12611](#12611)) ([6e91934](6e91934)) * Cl/release fixes 2 ([#12595](#12595)) ([fc597f4](fc597f4)) * Cl/release noir refs ([#12597](#12597)) ([fdcfcaf](fdcfcaf)) * demote log ([#12626](#12626)) ([bec8953](bec8953)) * deploy method test ([#12609](#12609)) ([f2c06c2](f2c06c2)) * Do not report epoch as complete until blocks have synced ([#12638](#12638)) ([2ddfa76](2ddfa76)), closes [#12625](#12625) * Error on infinitely recursive types (noir-lang/noir#7579) ([cc6cdbb](cc6cdbb)) * get L1 tx utils config from env ([#12620](#12620)) ([d930c01](d930c01)) * Log overflow handling in reset ([#12579](#12579)) ([283b624](283b624)) * metrics update ([#12571](#12571)) ([80a5df2](80a5df2)) * **sandbox:** query release please manifest for version if in a docker container ([#12591](#12591)) ([db8ebc6](db8ebc6)) * **spartan:** setup needs kubectl ([#12580](#12580)) ([753cb33](753cb33)) * update dead partial notes link ([#12629](#12629)) ([5a1dc4c](5a1dc4c)) * update error message to display 128 bits as valid bit size (noir-lang/noir#7626) ([cc6cdbb](cc6cdbb)) * update fallback transport ([#12470](#12470)) ([88f0711](88f0711)) ### Miscellaneous * bump external pinned commits (noir-lang/noir#7640) ([cc6cdbb](cc6cdbb)) * **ci3:** add helper for uncached test introspection ([#12618](#12618)) ([9ac518b](9ac518b)) * **ci3:** better memsuspend_limit comment ([#12622](#12622)) ([de84187](de84187)) * clean up upgrade test and other small things ([#12558](#12558)) ([c28abe1](c28abe1)) * cleanup eth artifacts + misc aztec.js reorg ([#12563](#12563)) ([6623244](6623244)) * **docs:** Updated accounts page ([#12019](#12019)) ([d45dac9](d45dac9)) * Fix mac build ([#12610](#12610)) ([adceed6](adceed6)) * gemini soundness regression test ([#12570](#12570)) ([c654106](c654106)) * more sane e2e_prover/full timeout ([#12619](#12619)) ([add9d35](add9d35)) * reactivate acir_test for `regression_5045` ([#12548](#12548)) ([c89f89c](c89f89c)) * remove unnecessary trait bounds (noir-lang/noir#7635) ([cc6cdbb](cc6cdbb)) * Rename `StructDefinition` to `TypeDefinition` (noir-lang/noir#7614) ([cc6cdbb](cc6cdbb)) * replace relative paths to noir-protocol-circuits ([4f7f5c3](4f7f5c3)) * replace relative paths to noir-protocol-circuits ([0f68d11](0f68d11)) * replace relative paths to noir-protocol-circuits ([8f593ce](8f593ce)) * replace relative paths to noir-protocol-circuits ([251ae38](251ae38)) * rollup library cleanup ([#12621](#12621)) ([361fc59](361fc59)) * **sandbox:** drop cheat-codes log level ([#12586](#12586)) ([24f04c7](24f04c7)) * **sandbox:** expose anvil port ([#12599](#12599)) ([955f1b0](955f1b0)) * **testnet:** updating script for ignition, change naming ([#12566](#12566)) ([2d7b69d](2d7b69d)) * turn on masking in eccvm ([#12467](#12467)) ([aacb91a](aacb91a)) * Update Bb line counting script ([#12350](#12350)) ([7a41843](7a41843)) * update docs to reflect u128 type (noir-lang/noir#7623) ([cc6cdbb](cc6cdbb)) * Validate blobs posted to sink belong to our L2 ([#12587](#12587)) ([9578f1e](9578f1e)), closes [#12497](#12497) ### Documentation * update cli-wallet commands in profiler doc ([#12568](#12568)) ([239a4fb](239a4fb)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please).

fcarreiro requested review from IlyasRidhuan, Maddiaa0 and jeanmon as code owners March 6, 2025 15:03

fcarreiro force-pushed the fc/bb-sumcheck-memory branch from 164245b to 77cd95d Compare March 6, 2025 15:03

fcarreiro removed request for IlyasRidhuan, Maddiaa0 and jeanmon March 6, 2025 15:04

fcarreiro force-pushed the fc/bb-sumcheck-memory branch from 77cd95d to caf6579 Compare March 6, 2025 15:05

fcarreiro commented Mar 6, 2025

View reviewed changes

fcarreiro force-pushed the fc/bb-sumcheck-memory branch from caf6579 to 90bed9f Compare March 6, 2025 15:49

jeanmon reviewed Mar 6, 2025

View reviewed changes

barretenberg/cpp/src/barretenberg/polynomials/polynomial.hpp Outdated Show resolved Hide resolved

fcarreiro force-pushed the fc/bb-sumcheck-memory branch from 90bed9f to 4c353b0 Compare March 6, 2025 16:35

fcarreiro requested review from ledwards2225 and lucasxia01 March 6, 2025 16:35

jeanmon reviewed Mar 6, 2025

View reviewed changes

barretenberg/cpp/src/barretenberg/sumcheck/sumcheck.hpp Outdated Show resolved Hide resolved