nx_argon2_block.nx

nx_argon2_block.nx -- RFC 9106 sec 3.5/3.6 Argon2 G compression + P permutation. The third and final brick in the Argon2id arc (before the orchestrator that ties indexing + memory matrix together). G is what gives Argon2 its memory hardness: it's intentionally slow on GPU/ASIC because it requires 1024-byte block-at-a-time mixing that can't be unrolled or pipelined deeply. ===== G compression function (RFC 9106 sec 3.5) ================= G(X, Y) -> Z where X, Y, Z are each 1024-byte blocks. 1. R = X XOR Y (1024 bytes) 2. Save R as R_save for the final XOR. 3. View R as 8x8 matrix of u128 registers (each u128 = 2 u64). Equivalently: 8 rows of 16 u64 words (a row = 128 bytes). 4. Row-wise: apply P to each of 8 rows. 5. Column-wise: apply P to each of 8 "columns" (gathered slice of 16 u64 from rows i = 0..7 at columns 2j, 2j+1). 6. Output Z = R XOR R_save. ===== P permutation (RFC 9106 sec 3.6) ========================== P operates on 16 u64 words v_0..v_15 in-place. Same call pattern as BLAKE2b's round but using GB (modified G with 2*L(a)*L(b) instead of message words): GB(v_0, v_4, v_8, v_12) GB(v_1, v_5, v_9, v_13) GB(v_2, v_6, v_10, v_14) GB(v_3, v_7, v_11, v_15) GB(v_0, v_5, v_10, v_15) GB(v_1, v_6, v_11, v_12) GB(v_2, v_7, v_8, v_13) GB(v_3, v_4, v_9, v_14) ===== GB function (RFC 9106 sec 3.6) ============================ GB(a, b, c, d):