Searched refs:loadu_128 (Results 1 – 2 of 2) sorted by relevance
13 INLINE uint32x4_t loadu_128(const uint8_t src[16]) { in loadu_128() function215 out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()216 out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()217 out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()218 out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()219 out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()220 out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()221 out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()222 out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()223 out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]); in transpose_msg_vecs4()[all …]
9 INLINE __m128i loadu_128(const uint8_t src[16]) { in loadu_128() function123 rows[0] = loadu_128((uint8_t *)&cv[0]); in compress_pre()124 rows[1] = loadu_128((uint8_t *)&cv[4]); in compress_pre()129 __m128i m0 = loadu_128(&block[sizeof(__m128i) * 0]); in compress_pre()130 __m128i m1 = loadu_128(&block[sizeof(__m128i) * 1]); in compress_pre()131 __m128i m2 = loadu_128(&block[sizeof(__m128i) * 2]); in compress_pre()132 __m128i m3 = loadu_128(&block[sizeof(__m128i) * 3]); in compress_pre()299 storeu_128(xor_128(rows[2], loadu_128((uint8_t *)&cv[0])), &out[32]); in blake3_compress_xof_avx512()300 storeu_128(xor_128(rows[3], loadu_128((uint8_t *)&cv[4])), &out[48]); in blake3_compress_xof_avx512()458 out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(__m128i)]); in transpose_msg_vecs4()[all …]