Skip to content
Closed
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
e5fe852
experimental primitives: AVX2 arms for bit_transpose_64bytes
kevaundray Nov 17, 2026
8f7705d
experimental primitives: AVX2 row transpose in BitFold::fold_block
kevaundray Nov 17, 2026
e7d3240
experimental keccak_gkr: table rows by byte maps on AVX2
kevaundray Nov 17, 2026
6e4fea8
keccak256: search grinding nonces with first_nonce
kevaundray Nov 17, 2026
9e6527e
experimental lean_vm_core: bus GKR vertical kernels on AVX2+VPCLMULQDQ
kevaundray Nov 17, 2026
918df52
experimental pcs: register-tiled NTT sweeps on AVX2 without VPCLMULQDQ
kevaundray Nov 17, 2026
c1fc8ef
experimental keccak_gkr: cofactor skip pass on x86 without GFNI
kevaundray Nov 17, 2026
b68a08f
experimental keccak_gkr: GFNI product sums on values at six points
kevaundray Nov 17, 2026
6425dae
experimental keccak_gkr: whole-position parts on AVX2
kevaundray Nov 17, 2026
ae7b9f5
experimental keccak_gkr: PCLMULQDQ row products with ones by lookup
kevaundray Nov 17, 2026
ff71b8a
experimental keccak_gkr: pack AVX2 row products in registers
kevaundray Nov 17, 2026
658fd20
experimental keccak_gkr: fold digit bucket sums by nibbles as they ar…
kevaundray Nov 17, 2026
8a3fe51
experimental pcs: fold level 0's lanes two coefficients a register on…
kevaundray Nov 17, 2026
836ec09
experimental pcs: run the two-lane level-0 kernels on AVX2 without VP…
kevaundray Nov 17, 2026
491743d
experimental keccak_gkr: reduce the skip pass's cofactors modulo the …
kevaundray Nov 17, 2026
c528130
experimental lean_vm_core: bus GKR and leaf lanes on AVX2 without VPC…
kevaundray Nov 17, 2026
9d7e41f
experimental pcs: pad the base encode's rows to four lanes on x86
kevaundray Nov 18, 2026
08209e9
experimental pcs: hash the L0 leaves from their data lanes
kevaundray Nov 18, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
319 changes: 310 additions & 9 deletions crates/experimental_prover/keccak_gkr/src/kernels.rs

Large diffs are not rendered by default.

22 changes: 13 additions & 9 deletions crates/experimental_prover/keccak_gkr/src/lib.rs
Original file line number Diff line number Diff line change
Expand Up @@ -86,9 +86,10 @@
//! of eight positions' weights applied to bit-transposed products (`products_gfni`); elsewhere,
//! the positions grouped by each byte of their weight (`prover`'s digit buckets).
//!
//! The four rounds' sums are binary polynomials of degree at most 173. With AVX2 and GFNI they
//! are taken on residues modulo 24 irreducibles of degree 8, where a product is a byte product
//! (`skip_gfni`).
//! The four rounds' sums are binary polynomials of degree at most 173. With GFNI (x86-64) they are
//! taken on residues modulo 24 irreducibles of degree 8, where a product is a byte product
//! (`skip_gfni`). Without it, x86-64 multiplies out the weights on the packs first, so that a
//! sum takes fewer carry-less products (`skip_clmul`).
//!
//! The rounds that remain fold a table of one field element per half block and position. A pass over it
//! binds two rounds' variables: over each quad of the rows it writes it sums the next round and, as a
Expand All @@ -114,15 +115,18 @@ mod products_gfni;
mod protocol;
mod prover;
pub mod rmfe;
#[cfg(all(target_arch = "x86_64", target_feature = "avx2", target_feature = "gfni"))]
mod skip_gfni;
#[cfg(all(
target_arch = "x86_64",
target_feature = "avx512f",
target_feature = "avx512bw",
target_feature = "gfni"
target_feature = "avx2",
target_feature = "pclmulqdq",
not(target_feature = "gfni"),
not(all(target_feature = "avx512f", target_feature = "vpclmulqdq"))
))]
mod table_gfni;
mod skip_clmul;
#[cfg(all(target_arch = "x86_64", target_feature = "avx2", target_feature = "gfni"))]
mod skip_gfni;
#[cfg(all(target_arch = "x86_64", target_feature = "avx2"))]
mod table_maps;
pub mod witness;

#[cfg(test)]
Expand Down
4 changes: 2 additions & 2 deletions crates/experimental_prover/keccak_gkr/src/power.rs
Original file line number Diff line number Diff line change
Expand Up @@ -44,9 +44,9 @@ macro_rules! row_kernel {
pub struct PowerField {
pub basis: PowerBasis,
/// `s^128`, the modulus without its leading term.
modulus: Element,
pub(crate) modulus: Element,
/// `floor(x^256 / modulus)` without its leading term, for Barrett's reduction.
barrett: Element,
pub(crate) barrett: Element,
eval: PolyEval,
}

Expand Down
Loading
Loading