Lines Matching full:rows
89 INLINE void compress_pre(__m128i rows[4], const uint32_t cv[8], in compress_pre()
92 rows[0] = loadu((uint8_t *)&cv[0]); in compress_pre()
93 rows[1] = loadu((uint8_t *)&cv[4]); in compress_pre()
94 rows[2] = set4(IV[0], IV[1], IV[2], IV[3]); in compress_pre()
95 rows[3] = set4(counter_low(counter), counter_high(counter), in compress_pre()
108 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
110 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
111 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
114 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
117 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
118 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
128 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
132 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
133 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
137 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
141 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
142 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
151 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
155 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
156 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
160 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
164 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
165 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
174 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
178 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
179 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
183 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
187 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
188 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
197 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
201 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
202 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
206 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
210 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
211 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
220 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
224 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
225 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
229 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
233 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
234 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
243 g1(&rows[0], &rows[1], &rows[2], &rows[3], t0); in compress_pre()
247 g2(&rows[0], &rows[1], &rows[2], &rows[3], t1); in compress_pre()
248 diagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
252 g1(&rows[0], &rows[1], &rows[2], &rows[3], t2); in compress_pre()
256 g2(&rows[0], &rows[1], &rows[2], &rows[3], t3); in compress_pre()
257 undiagonalize(&rows[0], &rows[2], &rows[3]); in compress_pre()
264 __m128i rows[4]; in blake3_compress_in_place_sse2() local
265 compress_pre(rows, cv, block, block_len, counter, flags); in blake3_compress_in_place_sse2()
266 storeu(xorv(rows[0], rows[2]), (uint8_t *)&cv[0]); in blake3_compress_in_place_sse2()
267 storeu(xorv(rows[1], rows[3]), (uint8_t *)&cv[4]); in blake3_compress_in_place_sse2()
274 __m128i rows[4]; in blake3_compress_xof_sse2() local
275 compress_pre(rows, cv, block, block_len, counter, flags); in blake3_compress_xof_sse2()
276 storeu(xorv(rows[0], rows[2]), &out[0]); in blake3_compress_xof_sse2()
277 storeu(xorv(rows[1], rows[3]), &out[16]); in blake3_compress_xof_sse2()
278 storeu(xorv(rows[2], loadu((uint8_t *)&cv[0])), &out[32]); in blake3_compress_xof_sse2()
279 storeu(xorv(rows[3], loadu((uint8_t *)&cv[4])), &out[48]); in blake3_compress_xof_sse2()