511     __m128i msg_vecs[16];  in blake3_hash4_avx512()  local
512     transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);  in blake3_hash4_avx512()
520     round_fn4(v, msg_vecs, 0);  in blake3_hash4_avx512()
521     round_fn4(v, msg_vecs, 1);  in blake3_hash4_avx512()
522     round_fn4(v, msg_vecs, 2);  in blake3_hash4_avx512()
523     round_fn4(v, msg_vecs, 3);  in blake3_hash4_avx512()
524     round_fn4(v, msg_vecs, 4);  in blake3_hash4_avx512()
525     round_fn4(v, msg_vecs, 5);  in blake3_hash4_avx512()
526     round_fn4(v, msg_vecs, 6);  in blake3_hash4_avx512()
766     __m256i msg_vecs[16];  in blake3_hash8_avx512()  local
767     transpose_msg_vecs8(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);  in blake3_hash8_avx512()
775     round_fn8(v, msg_vecs, 0);  in blake3_hash8_avx512()
776     round_fn8(v, msg_vecs, 1);  in blake3_hash8_avx512()
777     round_fn8(v, msg_vecs, 2);  in blake3_hash8_avx512()
778     round_fn8(v, msg_vecs, 3);  in blake3_hash8_avx512()
779     round_fn8(v, msg_vecs, 4);  in blake3_hash8_avx512()
780     round_fn8(v, msg_vecs, 5);  in blake3_hash8_avx512()
781     round_fn8(v, msg_vecs, 6);  in blake3_hash8_avx512()
1080     __m512i msg_vecs[16];  in blake3_hash16_avx512()  local
1081     transpose_msg_vecs16(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);  in blake3_hash16_avx512()
1089     round_fn16(v, msg_vecs, 0);  in blake3_hash16_avx512()
1090     round_fn16(v, msg_vecs, 1);  in blake3_hash16_avx512()
1091     round_fn16(v, msg_vecs, 2);  in blake3_hash16_avx512()
1092     round_fn16(v, msg_vecs, 3);  in blake3_hash16_avx512()
1093     round_fn16(v, msg_vecs, 4);  in blake3_hash16_avx512()
1094     round_fn16(v, msg_vecs, 5);  in blake3_hash16_avx512()
1095     round_fn16(v, msg_vecs, 6);  in blake3_hash16_avx512()