192 INLINE void transpose_msg_vecs(const uint8_t *const *inputs,  in transpose_msg_vecs()  argument
194   out[0] = loadu(&inputs[0][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
195   out[1] = loadu(&inputs[1][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
196   out[2] = loadu(&inputs[2][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
197   out[3] = loadu(&inputs[3][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
198   out[4] = loadu(&inputs[4][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
199   out[5] = loadu(&inputs[5][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
200   out[6] = loadu(&inputs[6][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
201   out[7] = loadu(&inputs[7][block_offset + 0 * sizeof(__m256i)]);  in transpose_msg_vecs()
202   out[8] = loadu(&inputs[0][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
203   out[9] = loadu(&inputs[1][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
204   out[10] = loadu(&inputs[2][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
205   out[11] = loadu(&inputs[3][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
206   out[12] = loadu(&inputs[4][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
207   out[13] = loadu(&inputs[5][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
208   out[14] = loadu(&inputs[6][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
209   out[15] = loadu(&inputs[7][block_offset + 1 * sizeof(__m256i)]);  in transpose_msg_vecs()
211     _mm_prefetch((const void *)&inputs[i][block_offset + 256], _MM_HINT_T0);  in transpose_msg_vecs()
231 void blake3_hash8_avx2(const uint8_t *const *inputs, size_t blocks,  in blake3_hash8_avx2()  argument
251     transpose_msg_vecs(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);  in blake3_hash8_avx2()
290 void blake3_hash_many_sse41(const uint8_t *const *inputs, size_t num_inputs,
296 void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs,
303 void blake3_hash_many_avx2(const uint8_t *const *inputs, size_t num_inputs,  in blake3_hash_many_avx2()  argument
309     blake3_hash8_avx2(inputs, blocks, key, counter, increment_counter, flags,  in blake3_hash_many_avx2()
314     inputs += DEGREE;  in blake3_hash_many_avx2()
319   blake3_hash_many_sse41(inputs, num_inputs, blocks, key, counter,  in blake3_hash_many_avx2()
322   blake3_hash_many_portable(inputs, num_inputs, blocks, key, counter,  in blake3_hash_many_avx2()