Skip to content

v3.1.4

  • v3.1.4
  • 8986029
  • Verified

    This commit was created on GitHub.com and signed with GitHub’s verified signature.
  • Choose a tag to compare

  • v3.1.4
  • 8986029
  • Choose a tag to compare

  • Verified

    This commit was created on GitHub.com and signed with GitHub’s verified signature.
@eason1981 eason1981 tagged this 16 Nov 10:37
* fix: prove EF guests and the eth-act standards tests on the GPU (#135)

* perf(sdk): zkvm_keccak256 absorbs whole lanes

On the rv64 target, u64::from_le_bytes over a byte slice compiles to eight
byte loads joined by shifts and ors. In the reth stateless guest the state-root
chunk runs about 20k keccaks, and those joins pushed its Bitwise and ShiftLeft
tables past every whitelist row, so the chunk took the catch-all shape that no
card proves.

absorb now reads words: one load per lane when the block is on a lane boundary,
and otherwise two aligned loads joined with one shift each (inline asm on the
zkvm target, so no read goes through the reference past the block). A
misaligned load is not an option, since no whitelist shape has room for the
unaligned-memory chip. The padding block and the digest use an aligned buffer.
The C API is unchanged.

* fix(cuda-mle): stream the RS encode by default

The prover picks 2^23-row chunks on a large-RAM host, and the monolithic RS
encode only fits a 32 GB card up to 2^22: such a chunk asks for one ~17 GiB
codeword buffer and runs out of memory. Encoding in column blocks (8 unless
RS_STREAM_BLOCKS says otherwise) keeps the peak at about one block. The
streamed commit and opening are bit-identical to the monolithic ones.

This is the setting the single-node and ethproofs scripts already export, so
library callers (the SDK, ere) now run what those scripts run, including
cache_rs_code_matrix having no effect while streaming is on.
RS_STREAM_BLOCKS=0 (or empty) still selects the monolithic path; a value that
is not a count keeps the default and logs a warning.

* feat(vm): core ladder rows for EF guests and word-op heavy programs

Two kinds of chunk fit no core ladder row, so they select the catch-all
shape, which a GPU build refuses:

- Row 30 is row 25 with Bitwise at 2^22 (TI 0.79 x 2^29). The EF ethrex
  stateless guest cuts one chunk in each of Sepolia blocks 12054750,
  12054853 and 12054913 with 2.10-2.20 M Bitwise rows (hash-table probing
  in the guest). The eth-act standards blake2f test lands there too.
- Row 31 is row 18 with Addw at 2^21 and Subw at 2^19 (TI 0.55 x 2^29).
  The eth-act standards suite's byte-offset memcpy/memcmp/memmove/memset
  tests and io-read-no-side-effects run word-op heavy loops, up to 2.01 M
  Addw and 0.33 M Subw rows per chunk.

Both rows go last, so they take only chunks no other row holds: no
chunk of the reth or ethrex blocks or of the monitor ISA and standards
programs changes row. Both convert shapes fit convert row 1; no
recursion row moves, the compress key is unchanged.

Whitelist rebuilt (mle_build_vk_map, CHUNK_SIZE=8388608,
SPLIT_THRESHOLD=2097152, 8 workers, no checkpoint, exit 0):
- leaves: 15606 convert + 36 combine + 1 deferred + 1 compress = 15644
- MLE_VK_ROOT="1076840983 318352835 1970043533 1657057240 410155794
  1624808160 1975281382 1077636617"
- (a) 143 chunks over 4 blocks, (b) fib-10: layouts and convert VKs match

The soundness manifest is regenerated (root, commit, one shape count);
contracts/test/dev keep the earlier root until the release.

* feat: BLS12-381 G2 add and double precompiles (#137)

* chore(perf): bench_apps lockfile follows the bls12_381 pin
Assets 2
Loading