413 INLINE void transpose_msg_vecs(const uint8_t *const *inputs,  in transpose_msg_vecs()  argument
415   out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m128i)]);  in transpose_msg_vecs()
416   out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m128i)]);  in transpose_msg_vecs()
417   out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m128i)]);  in transpose_msg_vecs()
418   out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m128i)]);  in transpose_msg_vecs()
419   out[4] = loadu(&inputs[0][block_offset + 1 * sizeof(__m128i)]);  in transpose_msg_vecs()
420   out[5] = loadu(&inputs[1][block_offset + 1 * sizeof(__m128i)]);  in transpose_msg_vecs()
421   out[6] = loadu(&inputs[2][block_offset + 1 * sizeof(__m128i)]);  in transpose_msg_vecs()
422   out[7] = loadu(&inputs[3][block_offset + 1 * sizeof(__m128i)]);  in transpose_msg_vecs()
423   out[8] = loadu(&inputs[0][block_offset + 2 * sizeof(__m128i)]);  in transpose_msg_vecs()
424   out[9] = loadu(&inputs[1][block_offset + 2 * sizeof(__m128i)]);  in transpose_msg_vecs()
425   out[10] = loadu(&inputs[2][block_offset + 2 * sizeof(__m128i)]);  in transpose_msg_vecs()
426   out[11] = loadu(&inputs[3][block_offset + 2 * sizeof(__m128i)]);  in transpose_msg_vecs()
427   out[12] = loadu(&inputs[0][block_offset + 3 * sizeof(__m128i)]);  in transpose_msg_vecs()
428   out[13] = loadu(&inputs[1][block_offset + 3 * sizeof(__m128i)]);  in transpose_msg_vecs()
429   out[14] = loadu(&inputs[2][block_offset + 3 * sizeof(__m128i)]);  in transpose_msg_vecs()
430   out[15] = loadu(&inputs[3][block_offset + 3 * sizeof(__m128i)]);  in transpose_msg_vecs()
432     _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);  in transpose_msg_vecs()
454 void blake3_hash4_sse41(const uint8_t *const *inputs, size_t blocks,  in blake3_hash4_sse41()  argument
474     transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);  in blake3_hash4_sse41()
535 void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs,  in blake3_hash_many_sse41()  argument
541     blake3_hash4_sse41(inputs, blocks, key, counter, increment_counter, flags,  in blake3_hash_many_sse41()
546     inputs += DEGREE;  in blake3_hash_many_sse41()
551     hash_one_sse41(inputs[0], blocks, key, counter, flags, flags_start,  in blake3_hash_many_sse41()
556     inputs += 1;  in blake3_hash_many_sse41()