p3-goldilocks 0.5.3

//! Poseidon1 permutation for Goldilocks.
//!
//! # Overview
//!
//! This module provides the Poseidon1 hash permutation instantiated for the
//! Goldilocks field (p = 2^64 - 2^32 + 1). The public API is a single type
//! alias that transparently dispatches to the best available implementation.
//!
//! # Platform Dispatch
//!
//! On **aarch64**, the type alias resolves to a dual-dispatch wrapper:
//! scalar permutations use NEON-accelerated MDS for full rounds with
//! LLVM-optimized sparse partial rounds, while packed NEON permutations
//! use the fused dual-lane ASM path (w8) or per-lane scalar path (w12).
//!
//! On **all other platforms**, it resolves to the generic Poseidon1
//! implementation with Karatsuba MDS convolution.
//!
//! No `#[cfg]` is needed in calling code.
//!
//! # MDS Matrix
//!
//! The MDS matrix is a **circulant** matrix sourced from the MDS crate.
//! At runtime, it is applied via fast Karatsuba convolution (sub-O(t^2)).
//! During initialization only, it is expanded to dense form for the
//! sparse matrix decomposition of partial rounds.
//!
//! # Round Constants
//!
//! Generated by the Grain LFSR (Poseidon1 paper, Appendix E) with SBOX=0 (x^alpha encoding).

use p3_poseidon1::{
    Poseidon1, Poseidon1Constants, Poseidon1ExternalLayerGeneric, Poseidon1InternalLayerGeneric,
};

use crate::mds::{MATRIX_CIRC_MDS_8_COL, MATRIX_CIRC_MDS_12_COL};
use crate::{Goldilocks, MdsMatrixGoldilocks};

/// S-box degree for Goldilocks Poseidon1.
///
/// The S-box raises each element to this power. The Goldilocks prime
/// factors as `p - 1 = 2^32 * 3 * 5 * 17 * 257 * 65537`. Neither 3 nor 5
/// are coprime to `p - 1`, so the smallest valid exponent is 7.
pub const GOLDILOCKS_S_BOX_DEGREE: u64 = 7;

/// Number of full rounds per half for Goldilocks Poseidon (`RF / 2`).
///
/// The total number of full rounds is `RF = 8` (4 beginning + 4 ending).
/// Follows the Poseidon paper's security analysis (Section 5.4) with a +2 RF margin.
pub const GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS: usize = 4;

/// Number of partial rounds for Goldilocks Poseidon (width 8).
///
/// Derived from the interpolation bound in the Poseidon paper (Eq. 3):
///
///   R_interp ≥ ⌈min{κ,n}/log_2(α)⌉ + ⌈log_α(t)⌉ − 5
///            = ⌈64/log_2(7)⌉ + ⌈log_7(8)⌉ − 5 = 23 + 2 − 5 = 20
///
/// With the +7.5% security margin (Section 5.4): ⌈1.075 × 20⌉ = 22.
pub const GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8: usize = 22;

/// Number of partial rounds for Goldilocks Poseidon (width 12).
///
/// Same interpolation bound as width 8:
///
///   R_interp ≥ ⌈64/log_2(7)⌉ + ⌈log_7(12)⌉ − 5 = 23 + 2 − 5 = 20
///
/// With the +7.5% security margin: ⌈1.075 × 20⌉ = 22.
pub const GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_12: usize = 22;

/// Generic (non-fused) Poseidon1 permutation for Goldilocks.
///
/// Uses the platform-independent Poseidon1 implementation with Karatsuba
/// MDS convolution. Used directly for widths not supported by the fused
/// type (e.g. 16, 24) and as the non-aarch64 fallback for widths 8 and 12.
pub type Poseidon1GoldilocksGeneric<const WIDTH: usize> = Poseidon1<
    Goldilocks,
    Poseidon1ExternalLayerGeneric<Goldilocks, MdsMatrixGoldilocks, WIDTH>,
    Poseidon1InternalLayerGeneric<Goldilocks, WIDTH>,
    WIDTH,
    GOLDILOCKS_S_BOX_DEGREE,
>;

/// Unified Poseidon1 permutation for Goldilocks.
///
/// On aarch64, resolves to a dual-dispatch wrapper: scalar permutations
/// use NEON MDS for full rounds with sparse partial rounds, packed NEON
/// permutations use fused dual-lane ASM (w8) or per-lane scalar (w12).
///
/// On all other platforms, resolves to the generic implementation with
/// Karatsuba MDS convolution.
///
/// Supports both scalar and packed state representations transparently.
#[cfg(target_arch = "aarch64")]
pub type Poseidon1Goldilocks<const WIDTH: usize> = crate::Poseidon1GoldilocksDispatch<WIDTH>;

/// Unified Poseidon1 permutation for Goldilocks.
///
/// On aarch64, resolves to the fused ASM-optimized implementation that
/// uses inline assembly and dual-lane NEON processing.
///
/// On all other platforms, resolves to the generic implementation with
/// Karatsuba MDS convolution.
///
/// Supports both scalar and packed state representations transparently.
#[cfg(not(target_arch = "aarch64"))]
pub type Poseidon1Goldilocks<const WIDTH: usize> = Poseidon1GoldilocksGeneric<WIDTH>;

/// Round constants for width-8 Poseidon1 on Goldilocks.
///
/// Generated by the Grain LFSR with parameters:
///     field_type=1, alpha=7 (exp_flag=0), n=64, t=8, R_F=8, R_P=22
///
/// Generated by `poseidon/generate_constants.py --field goldilocks --width 8`.
///
/// Layout: [initial_full (4 rounds), partial (22 rounds), terminal_full (4 rounds)].
pub const GOLDILOCKS_POSEIDON1_RC_8: [[Goldilocks; 8]; 30] = Goldilocks::new_2d_array([
    // Initial full rounds (4)
    [
        0xdd5743e7f2a5a5d9,
        0xcb3a864e58ada44b,
        0xffa2449ed32f8cdc,
        0x42025f65d6bd13ee,
        0x7889175e25506323,
        0x34b98bb03d24b737,
        0xbdcc535ecc4faa2a,
        0x5b20ad869fc0d033,
    ],
    [
        0xf1dda5b9259dfcb4,
        0x27515210be112d59,
        0x4227d1718c766c3f,
        0x26d333161a5bd794,
        0x49b938957bf4b026,
        0x4a56b5938b213669,
        0x1120426b48c8353d,
        0x6b323c3f10a56cad,
    ],
    [
        0xce57d6245ddca6b2,
        0xb1fc8d402bba1eb1,
        0xb5c5096ca959bd04,
        0x6db55cd306d31f7f,
        0xc49d293a81cb9641,
        0x1ce55a4fe979719f,
        0xa92e60a9d178a4d1,
        0x002cc64973bcfd8c,
    ],
    [
        0xcea721cce82fb11b,
        0xe5b55eb8098ece81,
        0x4e30525c6f1ddd66,
        0x43c6702827070987,
        0xaca68430a7b5762a,
        0x3674238634df9c93,
        0x88cee1c825e33433,
        0xde99ae8d74b57176,
    ],
    // Partial rounds (22)
    [
        0x488897d85ff51f56,
        0x1140737ccb162218,
        0xa7eeb9215866ed35,
        0x9bd2976fee49fcc9,
        0xc0c8f0de580a3fcc,
        0x4fb2dae6ee8fc793,
        0x343a89f35f37395b,
        0x223b525a77ca72c8,
    ],
    [
        0x56ccb62574aaa918,
        0xc4d507d8027af9ed,
        0xa080673cf0b7e95c,
        0xf0184884eb70dcf8,
        0x044f10b0cb3d5c69,
        0xe9e3f7993938f186,
        0x1b761c80e772f459,
        0x606cec607a1b5fac,
    ],
    [
        0x14a0c2e1d45f03cd,
        0x4eace8855398574f,
        0xf905ca7103eff3e6,
        0xf8c8f8d20862c059,
        0xb524fe8bdd678e5a,
        0xfbb7865901a1ec41,
        0x014ef1197d341346,
        0x9725e20825d07394,
    ],
    [
        0xfdb25aef2c5bae3b,
        0xbe5402dc598c971e,
        0x93a5711f04cdca3d,
        0xc45a9a5b2f8fb97b,
        0xfe8946a924933545,
        0x2af997a27369091c,
        0xaa62c88e0b294011,
        0x058eb9d810ce9f74,
    ],
    [
        0xb3cb23eced349ae4,
        0xa3648177a77b4a84,
        0x43153d905992d95d,
        0xf4e2a97cda44aa4b,
        0x5baa2702b908682f,
        0x082923bdf4f750d1,
        0x98ae09a325893803,
        0xf8a6475077968838,
    ],
    [
        0xceb0735bf00b2c5f,
        0x0a1a5d953888e072,
        0x2fcb190489f94475,
        0xb5be06270dec69fc,
        0x739cb934b09acf8b,
        0x537750b75ec7f25b,
        0xe9dd318bae1f3961,
        0xf7462137299efe1a,
    ],
    [
        0xb1f6b8eee9adb940,
        0xbdebcc8a809dfe6b,
        0x40fc1f791b178113,
        0x3ac1c3362d014864,
        0x9a016184bdb8aeba,
        0x95f2394459fbc25e,
        0xe3f34a07a76a66c2,
        0x8df25f9ad98b1b96,
    ],
    [
        0x85ffc27171439d9d,
        0xddcb9a2dcfd26910,
        0x26b5ba4bf3afb94e,
        0xffff9cc7c7651e2f,
        0x8c88364698280b55,
        0xebc114167b910501,
        0x2d77b4d89ecfb516,
        0x332e0828eba151f2,
    ],
    [
        0x46fa6a6450dd4735,
        0xd00db7dd92384a33,
        0x5fd4fb751f3a5fc5,
        0x496fb90c0bb65ea2,
        0xf3baec0bb87cc5c7,
        0x862a3c0a7d4c7713,
        0xbf5f38336a3f47d8,
        0x41ad9dbc1394a20c,
    ],
    [
        0xcc535945b7dbf0f7,
        0x82af2bc93685bcec,
        0x8e4c8d0c8cebfccd,
        0x17cb39417e84597e,
        0xd4a965a8c749b232,
        0xa2cab040f33f3ee5,
        0xa98811a1fed4e3a6,
        0x1cc48b54f377e2a1,
    ],
    [
        0xe40cd4f6c5609a27,
        0x11de79ebca97a4a4,
        0x9177c73d8b7e929d,
        0x2a6fe8085797e792,
        0x3de6e93329f8d5ae,
        0x3f7af9125da962ff,
        0xd710682cfc77d3ac,
        0x48faf05f3b053cf4,
    ],
    [
        0x287db8630da89c8b,
        0x4d0de32053cb30e9,
        0x8b37a4f20c5ada7b,
        0xe7cc6ebe78c84ecf,
        0x240bdc0a66a2610d,
        0x8299e7f02caa1650,
        0x380a53fefb6e754e,
        0x684a1d8cf8eb6810,
    ],
    [
        0xe839452eb4b8a5e1,
        0xb03fa62e90626af4,
        0x11a688602fbc5efc,
        0x30dda75c355a2d62,
        0x0f712adcb73810de,
        0xffdc1102187f1ae1,
        0x40c34f398254b99c,
        0xede021b9dc289a4a,
    ],
    [
        0x8b7b05225c4e7dad,
        0x3bc794346f9d9ff9,
        0xfccb5a57f2ca86ff,
        0xbb1502015a7da9d4,
        0xd7e0a35d4352a015,
        0x27af7a44f8160931,
        0xc37442f6782f4615,
        0xbdf392a9bd095dcb,
    ],
    [
        0xc17f55037cf00de9,
        0xbcffedd34c71a874,
        0x5eb45d2a8133d1f2,
        0xbabe251e1612ebdf,
        0x3efeb9fbe438c536,
        0x2d7cef97b4afe1cf,
        0xe5de1b4660016c0b,
        0xcdcc26c332f5657c,
    ],
    [
        0xe01dd653daf15809,
        0xb0a6bdd4b41094b5,
        0x27eac858b0b03a05,
        0x51d43b5e93adbdc0,
        0x8b89a23b0fea5fc9,
        0xdc8ac3b14f7f2fc1,
        0xe793f82f1efec039,
        0x9f6f2cf8969e7b80,
    ],
    [
        0x49d45382e0f21d4a,
        0x5f4ad1797cd72786,
        0x4dc3dbebfd45f795,
        0x03a3ef84dba6e1bc,
        0x204bc9b3d3fc4c01,
        0x9ad706081e89b9ba,
        0x638bfb4d840e9f89,
        0x5ef2938cd095ae35,
    ],
    [
        0x42cca18ebeb265c8,
        0xb7b2ec5c29aecbf8,
        0x0d84f9535dc78f0f,
        0x04e64ad942e77b8c,
        0xb4880dffffc9da0b,
        0x16db16d9c29adeb1,
        0x09bbaf2a0590cd1e,
        0x76460e74961fcf8d,
    ],
    [
        0xed12a2276dfa1553,
        0x0b5acec5de0436fd,
        0x3c6cfea033a1f0a8,
        0x2b5ecefe546cac15,
        0x6e2d82884cd3bf6f,
        0xc134878d1add7b83,
        0x997963422eb7a280,
        0x5e834537ac648cf6,
    ],
    [
        0x89e779214737c0b7,
        0x1a8c05e8581ad95b,
        0x8d18b72796437cf7,
        0xe7252c949e04b106,
        0x53267c4fd174585a,
        0xa16ef5d9c81dad47,
        0xda65191937270a46,
        0xcb2a5b55f2df664c,
    ],
    [
        0x854aee2dc1924137,
        0xf37013c9d479ece6,
        0x0e163bc0630c4696,
        0x384ee64955048f76,
        0xf65d814e28ee4ec5,
        0xe57bc564fd82f1b1,
        0x4b338937b6876614,
        0x66ee0b04ed43cd8d,
    ],
    [
        0x49884bf25f4ef15d,
        0xeb51fe28de1c6f54,
        0x2cd64e84fce8dfcc,
        0x29164a96a541a013,
        0x173ce7558f4cacb8,
        0xeb5b1ce5877c89e9,
        0x5faff4b0f5217bf6,
        0xac42d0b1c20f205e,
    ],
    // Terminal full rounds (4)
    [
        0xfb1d6bf0ca43221b,
        0x97b0a1b01d6a2955,
        0x08c60bd622952b30,
        0x43f2be0f9e24147c,
        0xfa7268b7d3730f5d,
        0x43a6c419a23983bb,
        0xcd77c1f7b29b113c,
        0xcfa43c9db8eec29f,
    ],
    [
        0xcaaa95a6c7365dec,
        0x0a91193f798f3be0,
        0x1104497652735dc6,
        0x35aecb93663b515e,
        0x8dbc9916065aa858,
        0xada8f7a0266579ed,
        0x524dee7bec1ea789,
        0xa93aee9dd5af9521,
    ],
    [
        0x9d1f1b54750d707e,
        0x7c9feab87096d5dc,
        0xa2e1fb19f9d4261b,
        0xb714deb448de6346,
        0x225d1f0d011c5403,
        0x1549b7f1d28cedc0,
        0xaef3e46f97d43942,
        0x6dfc7ffe0b38bf08,
    ],
    [
        0x7de853fdc542b663,
        0xa68ecc96610657b2,
        0xe88bb5428af289b1,
        0xd7cfa1504c5569f5,
        0x78a9aad0d642d30a,
        0xd68315f2353dce52,
        0x46e56300f86fcfd5,
        0x323d95332b145fd6,
    ],
]);

/// Round constants for width-12 Poseidon1 on Goldilocks.
///
/// Generated by the Grain LFSR with parameters:
///     field_type=1, alpha=7 (exp_flag=0), n=64, t=12, R_F=8, R_P=22
///
/// Generated by `poseidon/generate_constants.py --field goldilocks --width 12`.
///
/// Layout: [initial_full (4 rounds), partial (22 rounds), terminal_full (4 rounds)].
pub const GOLDILOCKS_POSEIDON1_RC_12: [[Goldilocks; 12]; 30] = Goldilocks::new_2d_array([
    // Initial full rounds (4)
    [
        0x13dcf33aba214f46,
        0x30b3b654a1da6d83,
        0x1fc634ada6159b56,
        0x937459964dc03466,
        0xedd2ef2ca7949924,
        0xede9affde0e22f68,
        0x8515b9d6bac9282d,
        0x6b5c07b4e9e900d8,
        0x1ec66368838c8a08,
        0x9042367d80d1fbab,
        0x400283564a3c3799,
        0x4a00be0466bca75e,
    ],
    [
        0x7913beee58e3817f,
        0xf545e88532237d90,
        0x22f8cb8736042005,
        0x6f04990e247a2623,
        0xfe22e87ba37c38cd,
        0xd20e32c85ffe2815,
        0x117227674048fe73,
        0x4e9fb7ea98a6b145,
        0xe0866c232b8af08b,
        0x00bbc77916884964,
        0x7031c0fb990d7116,
        0x240a9e87cf35108f,
    ],
    [
        0x2e6363a5a12244b3,
        0x5e1c3787d1b5011c,
        0x4132660e2a196e8b,
        0x3a013b648d3d4327,
        0xf79839f49888ea43,
        0xfe85658ebafe1439,
        0xb6889825a14240bd,
        0x578453605541382b,
        0x4508cda8f6b63ce9,
        0x9c3ef35848684c91,
        0x0812bde23c87178c,
        0xfe49638f7f722c14,
    ],
    [
        0x8e3f688ce885cbf5,
        0xb8e110acf746a87d,
        0xb4b2e8973a6dabef,
        0x9e714c5da3d462ec,
        0x6438f9033d3d0c15,
        0x24312f7cf1a27199,
        0x23f843bb47acbf71,
        0x9183f11a34be9f01,
        0x839062fbb9d45dbf,
        0x24b56e7e6c2e43fa,
        0xe1683da61c962a72,
        0xa95c63971a19bfa7,
    ],
    // Partial rounds (22)
    [
        0x4adf842aa75d4316,
        0xf8fbb871aa4ab4eb,
        0x68e85b6eb2dd6aeb,
        0x07a0b06b2d270380,
        0xd94e0228bd282de4,
        0x8bdd91d3250c5278,
        0x209c68b88bba778f,
        0xb5e18cdab77f3877,
        0xb296a3e808da93fa,
        0x8370ecbda11a327e,
        0x3f9075283775dad8,
        0xb78095bb23c6aa84,
    ],
    [
        0x3f36b9fe72ad4e5f,
        0x69bc96780b10b553,
        0x3f1d341f2eb7b881,
        0x4e939e9815838818,
        0xda366b3ae2a31604,
        0xbc89db1e7287d509,
        0x6102f411f9ef5659,
        0x58725c5e7ac1f0ab,
        0x0df5856c798883e7,
        0xf7bb62a8da4c961b,
        0xc68be7c94882a24d,
        0xaf996d5d5cdaedd9,
    ],
    [
        0x9717f025e7daf6a5,
        0x6436679e6e7216f4,
        0x8a223d99047af267,
        0xbb512e35a133ba9a,
        0xfbbf44097671aa03,
        0xf04058ebf6811e61,
        0x5cca84703fac7ffb,
        0x9b55c7945de6469f,
        0x8e05bf09808e934f,
        0x2ea900de876307d7,
        0x7748fff2b38dfb89,
        0x6b99a676dd3b5d81,
    ],
    [
        0xac4bb7c627cf7c13,
        0xadb6ebe5e9e2f5ba,
        0x2d33378cafa24ae3,
        0x1e5b73807543f8c2,
        0x09208814bfebb10f,
        0x782e64b6bb5b93dd,
        0xadd5a48eac90b50f,
        0xadd4c54c736ea4b1,
        0xd58dbb86ed817fd8,
        0x6d5ed1a533f34ddd,
        0x28686aa3e36b7cb9,
        0x591abd3476689f36,
    ],
    [
        0x047d766678f13875,
        0xa2a11112625f5b49,
        0x21fd10a3f8304958,
        0xf9b40711443b0280,
        0xd2697eb8b2bde88e,
        0x3493790b51731b3f,
        0x11caf9dd73764023,
        0x7acfb8f72878164e,
        0x744ec4db23cefc26,
        0x1e00e58f422c6340,
        0x21dd28d906a62dda,
        0xf32a46ab5f465b5f,
    ],
    [
        0xbfce13201f3f7e6b,
        0xf30d2e7adb5304e2,
        0xecdf4ee4abad48e9,
        0xf94e82182d395019,
        0x4ee52e3744d887c5,
        0xa1341c7cac0083b2,
        0x2302fb26c30c834a,
        0xaea3c587273bf7d3,
        0xf798e24961823ec7,
        0x962deba3e9a2cd94,
        0xb36ee79485ca4707,
        0xd380199eddd2de52,
    ],
    [
        0x70971fc4e6f85305,
        0x8e722f6e5dc32699,
        0xa0883df133052b92,
        0x8f86c6a3eb7d01a4,
        0x763649c8b670bdc5,
        0x830d5c82b808759b,
        0xaa1da8bb91da02e7,
        0x9bc9bf629e211c4d,
        0x0f0a899b10a4dea8,
        0xb883bdcee7c6b356,
        0x78c7101e7496ae1e,
        0x2fd6c5a8bf1e5ca6,
    ],
    [
        0xe2a6e06e61fcec9c,
        0xebfce7d5c5b3dbd5,
        0xca2eeca4bb485d85,
        0xc2b875537c42eb69,
        0x6faf849976873328,
        0xfc3fcb6e81ad4cc3,
        0x180dd95503955a28,
        0xd40f19a3c9fe1520,
        0x49d178ddbf7fd96d,
        0x3950bee2e10e0297,
        0x437b90cf295be062,
        0xa5cd126edffad23b,
    ],
    [
        0xdf58134c134491c2,
        0x0677eca229d9f7bd,
        0x492200a1f7d83a3c,
        0xafb58c9810a43645,
        0x7659077c5a9c208e,
        0x30b4bc83706995cd,
        0xc98fa77bbbef3a3b,
        0x84a82905750b3109,
        0x72f2a02326aeb69b,
        0x8d27a2a2d73a848a,
        0xaa9e30a80bde4b68,
        0x63abb1415e050474,
    ],
    [
        0x1c4bd1e816050a7e,
        0x15d1502e4f469dfd,
        0x53989d594b0c4cd8,
        0x7a1a4c83cb7e377e,
        0x1b52f8a9944e480e,
        0xeb7b03f76a91a79e,
        0x0073a4fc9328c69e,
        0x2c7b16f8620d9de4,
        0x950d052963e46bc4,
        0x8d201ba1a9c89fac,
        0xd3502941bdf35503,
        0x7c6dfcd5af8676fb,
    ],
    [
        0xf8a6cd02e92cdb0b,
        0x6e7500f3a5464b22,
        0x07637eabba4bdd20,
        0x88b82717beee0e14,
        0xbaa2b1cd3dd4c79a,
        0xdfecc3aebec4cfa6,
        0x7561087b0cff0166,
        0x538fcac317a703a6,
        0xd7d6c6eeeeeeea19,
        0xd647b1ee441658a0,
        0xdf4442110236c546,
        0x559ef2c6dd73ec15,
    ],
    [
        0x4c0f5fc6c0dda3d1,
        0x685010cc3100cea7,
        0x2fb6ba8aa0344440,
        0xb515f0a3ca75f1fb,
        0x886887eaecb87c10,
        0xf03ec3fd710abb04,
        0xd3b4763e17f543ef,
        0x50d9e5716e78083a,
        0x0bce2385cf8d74ff,
        0xaf23032cd5f0e04b,
        0xd366aa112b6159d9,
        0x810a3ad3ac7979db,
    ],
    [
        0x0a4a11d794be40a2,
        0xeebf0cf23b668a3f,
        0x600873fb011d761b,
        0x0bfb5591a02ff618,
        0xa16e2a528910af52,
        0xf6553653e2878421,
        0xccbe7c7a601a30c0,
        0xb18b214fe489f5b3,
        0xe21017ab9e153425,
        0x586099ede17af9a6,
        0x385078b514f50647,
        0xc02b3a9afb89883d,
    ],
    [
        0x6d3fbd3b4a9f1de6,
        0x4b4d40a41b0f473c,
        0x838f1887b8f31711,
        0x9396895be5c58a41,
        0x6247a479d66fc2e3,
        0x13fe228a98f2d0a2,
        0x5ba5fde765f9481e,
        0xafb89fa62267e117,
        0xfa4dc1bebcaa6333,
        0xdbab590882b87289,
        0xc3b6c08e23ba9301,
        0xd84b5de94a324fb7,
    ],
    [
        0x0d0c371c5b35b850,
        0x7964f570e7188038,
        0x5daf18bbd996604c,
        0x6743bc47b9595258,
        0x5528b9362c59bb71,
        0xac45e25b7127b68c,
        0xa2077d7dfbb606b6,
        0xf3faac6faee378af,
        0x0c6388b51545e884,
        0xd27dbb6944917b61,
        0x89bcac584344c104,
        0x856bab802ce7402d,
    ],
    [
        0x2cff3000be1fcd0a,
        0x765f2977fa72a917,
        0x1443711329f5f9d5,
        0xd35cd0261af2f951,
        0x2a1bb986084ec281,
        0x2334a54b758f23f2,
        0xa9b8cb612caf706b,
        0xb6ba11c4ab1a1017,
        0xde96b0824b4b46e2,
        0xc59d4272c6d92e2c,
        0x389bb5107611754d,
        0x23647fbc77657372,
    ],
    [
        0xd5ef60d6f76a42fa,
        0xebb406bb79ac9819,
        0x55faccc709a2f423,
        0xd9d6ea97490091cd,
        0xef3ce5069647a7e4,
        0xdf31625d3fa78464,
        0x242e60fd68f10f66,
        0x39c966cc815f084d,
        0x20e2e22e02bae3f7,
        0xb38919d3f1173d7c,
        0xf17769f6c77084d9,
        0xcc051d8094cac41f,
    ],
    [
        0x942069f5d6eece7e,
        0x8d61d3e6f141c572,
        0xc5cef9d85dd605f4,
        0x938f2ac2bf885997,
        0x23bddbace7c48f6c,
        0xc90a6c5ba98537e4,
        0x0be6ee2cca90f6ae,
        0xa026175394ae0e90,
        0x29fca3e314c77628,
        0x2aa2aa8738ab7b77,
        0xe11bbd31fbb8cac6,
        0xb5bbbef1b78a23af,
    ],
    [
        0x8b62a5551e9a9797,
        0x3f91073d4d491c80,
        0x4cfa44976396424a,
        0xf8dcb2dfb3aa1b44,
        0x3849409eba1a95f5,
        0x070845799f234380,
        0x184c0093667da1ba,
        0xbd66aafccd51601e,
        0xee6d14e92155b490,
        0x626f2ec1865bc544,
        0x1bd2854bf6485986,
        0x368b8497472f12ef,
    ],
    [
        0x4f88cdcdfb791921,
        0xe2c0acfeda9ae781,
        0x9739bc21773469b3,
        0x00ce3ad64dc4bb8f,
        0xaab85a321ee7a4c8,
        0xd5de825be97004f4,
        0x48d676d3a043b1c6,
        0x9c6180b1ff643097,
        0x34882a89dd590b09,
        0xae7e6b0d249c3b1d,
        0x8c016908a04885a1,
        0x83ebaaebc9ae0721,
    ],
    [
        0xab21b42e0f642307,
        0xdb46631f62bb29c1,
        0xef29f0399e09b5d9,
        0x5b52fbb3613b8ba1,
        0x57e129fcc96922e6,
        0xcdeb14c9d9204b3a,
        0x1341ef0da8536e34,
        0xd7e3400f2bacde63,
        0x6911eeb42f70d7e5,
        0xc3a2a910a4679767,
        0x1773cbe4a0f6bb28,
        0xe17b0d53e843eab5,
    ],
    [
        0x587fa39990b62800,
        0x0d5d32788135879d,
        0x277f7b31fd3a4cdb,
        0xa435290ee56d7efa,
        0xea6f40be35159925,
        0xcb73377a506171cb,
        0xe43c367ce731d82a,
        0x6eb305031ca10c43,
        0xc019a8c622cc84cb,
        0xd5614f5658c612e6,
        0x7b1ecbe957c3ff98,
        0x60db6ee9651a8478,
    ],
    // Terminal full rounds (4)
    [
        0x9271d450fc9b4117,
        0xcffeea06b6e3aac1,
        0xfa4a44c748d1cd8e,
        0xe64db01ba569b469,
        0xd31005160e4045fe,
        0x39e0fa013e025f79,
        0xe243be574196a956,
        0x205b2a681e3d2642,
        0x79cae5ad93486bab,
        0xfdf567844e32c295,
        0x331679589bfb7189,
        0xaf06ee32297b89c2,
    ],
    [
        0xa6bcae311e498491,
        0x9d16f52c96ac8b3e,
        0x48a674b59393fa35,
        0x0f9e65da3fde3796,
        0x1e098310fc84578c,
        0x559ae5fab1ae8dad,
        0x56bd4d624078881d,
        0xfd8bbbf8fbe817b5,
        0x82d30695c44df534,
        0x3ec0a97bc41127c5,
        0x1eb8b64adaa22078,
        0x82c45e418d60c983,
    ],
    [
        0xb092280f484d55bf,
        0xcd317c9537697939,
        0xd3be2e352feb79f3,
        0xca6d866539a390e5,
        0xb5efb1a494e55ee6,
        0xfa9013ac89756e9e,
        0xaeb88efd1e981242,
        0x13ee477cdab6e0dc,
        0xce7df902c40da2d3,
        0xf3fbaf0d4e6f5f34,
        0xf96354ada6785f38,
        0x13b5692812406886,
    ],
    [
        0xf03cae030a0f4418,
        0x7d3172887aa98e1a,
        0x8a2c2644f2faf7b9,
        0x80d721abee696d00,
        0x27c8b903a4d68267,
        0xaf0b7b12f90291b8,
        0x00acd08cfdff3817,
        0x4659ee496c634328,
        0xf5b25c10730dbff1,
        0xdde3a153297329c2,
        0x50c0b70d6910a44b,
        0x23c7426af725a6a0,
    ],
]);

/// Create the default width-8 Poseidon1 permutation for Goldilocks.
///
/// Returns the platform-optimal implementation: dual-dispatch on aarch64
/// (generic for scalar, fused ASM for packed), generic Karatsuba on all
/// other platforms.
#[cfg(target_arch = "aarch64")]
pub fn default_goldilocks_poseidon1_8() -> Poseidon1Goldilocks<8> {
    let constants = Poseidon1Constants {
        rounds_f: 2 * GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
        rounds_p: GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
        mds_circ_col: MATRIX_CIRC_MDS_8_COL,
        round_constants: GOLDILOCKS_POSEIDON1_RC_8.to_vec(),
    };
    let (full, partial) = constants.to_optimized();
    let fused = crate::Poseidon1GoldilocksFused::new(&full, &partial);
    crate::Poseidon1GoldilocksDispatch::new(fused, full, partial)
}

/// Create the default width-8 Poseidon1 permutation for Goldilocks.
///
/// Returns the platform-optimal implementation: fused ASM on aarch64,
/// generic Karatsuba on all other platforms.
#[cfg(not(target_arch = "aarch64"))]
pub fn default_goldilocks_poseidon1_8() -> Poseidon1Goldilocks<8> {
    Poseidon1::new(&Poseidon1Constants {
        rounds_f: 2 * GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
        rounds_p: GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
        mds_circ_col: MATRIX_CIRC_MDS_8_COL,
        round_constants: GOLDILOCKS_POSEIDON1_RC_8.to_vec(),
    })
}

/// Create the default width-12 Poseidon1 permutation for Goldilocks.
///
/// Returns the platform-optimal implementation: dual-dispatch on aarch64
/// (generic for scalar, fused ASM for packed), generic Karatsuba on all
/// other platforms.
#[cfg(target_arch = "aarch64")]
pub fn default_goldilocks_poseidon1_12() -> Poseidon1Goldilocks<12> {
    let constants = Poseidon1Constants {
        rounds_f: 2 * GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
        rounds_p: GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_12,
        mds_circ_col: MATRIX_CIRC_MDS_12_COL,
        round_constants: GOLDILOCKS_POSEIDON1_RC_12.to_vec(),
    };
    let (full, partial) = constants.to_optimized();
    let fused = crate::Poseidon1GoldilocksFused::new(&full, &partial);
    crate::Poseidon1GoldilocksDispatch::new(fused, full, partial)
}

/// Create the default width-12 Poseidon1 permutation for Goldilocks.
///
/// Returns the platform-optimal implementation: fused ASM on aarch64,
/// generic Karatsuba on all other platforms.
#[cfg(not(target_arch = "aarch64"))]
pub fn default_goldilocks_poseidon1_12() -> Poseidon1Goldilocks<12> {
    Poseidon1::new(&Poseidon1Constants {
        rounds_f: 2 * GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
        rounds_p: GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_12,
        mds_circ_col: MATRIX_CIRC_MDS_12_COL,
        round_constants: GOLDILOCKS_POSEIDON1_RC_12.to_vec(),
    })
}

#[cfg(test)]
mod tests {
    use p3_symmetric::Permutation;
    use rand::SeedableRng;
    use rand::rngs::SmallRng;

    use super::*;

    type F = Goldilocks;

    /// Known-answer test for width 8 (sequential 0..7 input).
    #[test]
    fn test_poseidon_goldilocks_width_8() {
        let perm = default_goldilocks_poseidon1_8();

        let mut input: [F; 8] = F::new_array([0, 1, 2, 3, 4, 5, 6, 7]);
        perm.permute_mut(&mut input);

        let expected: [F; 8] = F::new_array([
            2431226948502761687,
            9427563026145807618,
            6827549936272051660,
            16907684411084503785,
            10131745626715172913,
            17448305483431576765,
            9066501914269485014,
            12095238468458521303,
        ]);
        assert_eq!(input, expected);
    }

    /// Known-answer test for width 12 (sequential 0..11 input).
    #[test]
    fn test_poseidon_goldilocks_width_12() {
        let perm = default_goldilocks_poseidon1_12();

        let mut input: [F; 12] = F::new_array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]);
        perm.permute_mut(&mut input);

        let expected: [F; 12] = F::new_array([
            15595088881848875364,
            9564850329150784619,
            13607005230761744521,
            12117102595842533385,
            2814257411756993122,
            11640647689983397089,
            14363867760831937423,
            13323891071259596526,
            11219803511311150468,
            9221595262780869902,
            5898229059046891887,
            18181291031484020550,
        ]);
        assert_eq!(input, expected);
    }

    /// Smoke test for width 16 with random constants.
    /// Uses the generic type directly since the fused type only supports 8 and 12.
    #[test]
    fn test_poseidon_goldilocks_width_16() {
        let mut rng = SmallRng::seed_from_u64(1);
        let poseidon = Poseidon1GoldilocksGeneric::<16>::new_from_rng(
            GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
            GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
            &MdsMatrixGoldilocks,
            &mut rng,
        );
        let input: [F; 16] = rand::RngExt::random(&mut rng);
        let output = poseidon.permute(input);
        assert_ne!(output, input);
    }

    /// Smoke test for width 24 with random constants.
    #[test]
    fn test_poseidon_goldilocks_width_24() {
        let mut rng = SmallRng::seed_from_u64(1);
        let poseidon = Poseidon1GoldilocksGeneric::<24>::new_from_rng(
            GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
            GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
            &MdsMatrixGoldilocks,
            &mut rng,
        );
        let input: [F; 24] = rand::RngExt::random(&mut rng);
        let output = poseidon.permute(input);
        assert_ne!(output, input);
    }

    #[cfg(all(target_arch = "x86_64", target_feature = "avx512f"))]
    mod avx512 {
        use super::*;
        use crate::PackedGoldilocksAVX512;

        #[test]
        fn test_avx512_poseidon_width_16() {
            let mut rng = SmallRng::seed_from_u64(1);
            let poseidon = Poseidon1GoldilocksGeneric::<16>::new_from_rng(
                GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
                GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
                &MdsMatrixGoldilocks,
                &mut rng,
            );
            let input: [F; 16] = rand::RngExt::random(&mut rng);

            let mut expected = input;
            poseidon.permute_mut(&mut expected);

            let mut avx512_input = input.map(Into::<PackedGoldilocksAVX512>::into);
            poseidon.permute_mut(&mut avx512_input);

            let avx512_output = avx512_input.map(|x| x.0[0]);
            assert_eq!(avx512_output, expected);
        }

        #[test]
        fn test_avx512_poseidon_width_24() {
            let mut rng = SmallRng::seed_from_u64(1);
            let poseidon = Poseidon1GoldilocksGeneric::<24>::new_from_rng(
                GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
                GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
                &MdsMatrixGoldilocks,
                &mut rng,
            );
            let input: [F; 24] = rand::RngExt::random(&mut rng);

            let mut expected = input;
            poseidon.permute_mut(&mut expected);

            let mut avx512_input = input.map(Into::<PackedGoldilocksAVX512>::into);
            poseidon.permute_mut(&mut avx512_input);

            let avx512_output = avx512_input.map(|x| x.0[0]);
            assert_eq!(avx512_output, expected);
        }
    }

    #[cfg(all(
        target_arch = "x86_64",
        target_feature = "avx2",
        not(target_feature = "avx512f")
    ))]
    mod avx2 {
        use super::*;
        use crate::PackedGoldilocksAVX2;

        #[test]
        fn test_avx2_poseidon_width_16() {
            let mut rng = SmallRng::seed_from_u64(1);
            let poseidon = Poseidon1GoldilocksGeneric::<16>::new_from_rng(
                GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
                GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
                &MdsMatrixGoldilocks,
                &mut rng,
            );
            let input: [F; 16] = rand::RngExt::random(&mut rng);

            let mut expected = input;
            poseidon.permute_mut(&mut expected);

            let mut avx2_input = input.map(Into::<PackedGoldilocksAVX2>::into);
            poseidon.permute_mut(&mut avx2_input);

            let avx2_output = avx2_input.map(|x| x.0[0]);
            assert_eq!(avx2_output, expected);
        }

        #[test]
        fn test_avx2_poseidon_width_24() {
            let mut rng = SmallRng::seed_from_u64(1);
            let poseidon = Poseidon1GoldilocksGeneric::<24>::new_from_rng(
                GOLDILOCKS_POSEIDON_HALF_FULL_ROUNDS,
                GOLDILOCKS_POSEIDON_PARTIAL_ROUNDS_8,
                &MdsMatrixGoldilocks,
                &mut rng,
            );
            let input: [F; 24] = rand::RngExt::random(&mut rng);

            let mut expected = input;
            poseidon.permute_mut(&mut expected);

            let mut avx2_input = input.map(Into::<PackedGoldilocksAVX2>::into);
            poseidon.permute_mut(&mut avx2_input);

            let avx2_output = avx2_input.map(|x| x.0[0]);
            assert_eq!(avx2_output, expected);
        }
    }

    #[cfg(target_arch = "aarch64")]
    mod neon {
        use super::*;
        use crate::PackedGoldilocksNeon;

        #[test]
        fn test_neon_poseidon_width_8() {
            let perm = default_goldilocks_poseidon1_8();
            let input: [F; 8] = F::new_array([0, 1, 2, 3, 4, 5, 6, 7]);

            let mut expected = input;
            perm.permute_mut(&mut expected);

            let mut neon_input = input.map(Into::<PackedGoldilocksNeon>::into);
            perm.permute_mut(&mut neon_input);

            let neon_output = neon_input.map(|x| x.0[0]);
            assert_eq!(neon_output, expected);
        }

        #[test]
        fn test_neon_poseidon_width_12() {
            let perm = default_goldilocks_poseidon1_12();
            let input: [F; 12] = F::new_array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11]);

            let mut expected = input;
            perm.permute_mut(&mut expected);

            let mut neon_input = input.map(Into::<PackedGoldilocksNeon>::into);
            perm.permute_mut(&mut neon_input);

            let neon_output = neon_input.map(|x| x.0[0]);
            assert_eq!(neon_output, expected);
        }
    }
}