Lines Matching refs:b
65 umull v1.8h, v8.8b, v24.8b // g0 = y0 * 149
66 umull v17.8h, v9.8b, v24.8b // g1 = y1 * 149
67 umull2 v5.8h, v8.16b, v24.16b // g0_hi = y0_hi * 149
68 umull2 v21.8h, v9.16b, v24.16b // g1_hi = y1_hi * 149
70 umull v8.8h, \regu\().8b, v25.8b // g2 = u * 50 + v * 104
71 umlal v8.8h, \regv\().8b, v26.8b
72 umull2 v9.8h, \regu\().16b, v25.16b // g2_hi = u_hi * 50 + v_hi * 104
73 umlal2 v9.8h, \regv\().16b, v26.16b
75 ushr v19.16b, \regv\().16b, #1
76 uaddw v0.8h, v1.8h, v19.8b // r0 = g0 + (v >> 1)
77 uaddw v16.8h, v17.8h, v19.8b // r1 = g1 + (v >> 1)
79 uaddw2 v4.8h, v5.8h, v19.16b // r0_hi = g0_hi + (v_hi >> 1)
80 uaddw2 v20.8h, v21.8h, v19.16b // r1_hi = g1_hi + (v_hi >> 1)
82 ushll v19.8h, \regu\().8b, #2
83 ushll2 v23.8h, \regu\().16b, #2
90 umull v12.8h, \regv\().8b, v27.8b // r2 = v * 204
91 umull v13.8h, \regu\().8b, v28.8b // b2 = u * 254
93 umull2 v14.8h, \regv\().16b, v27.16b // r2_hi = v_hi * 204
94 umull2 v15.8h, \regu\().16b, v28.16b // b2_hi = u_hi * 254
124 uqrshrn v0.8b, v0.8h, #6
125 uqrshrn v16.8b, v16.8h, #6
126 uqrshrn v1.8b, v1.8h, #7
127 uqrshrn v17.8b, v17.8h, #7
128 uqrshrn v2.8b, v2.8h, #6
129 uqrshrn v18.8b, v18.8h, #6
131 uqrshrn v4.8b, v4.8h, #6
132 uqrshrn v20.8b, v20.8h, #6
133 uqrshrn v5.8b, v5.8h, #7
134 uqrshrn v21.8b, v21.8h, #7
135 uqrshrn v6.8b, v6.8h, #6
136 uqrshrn v22.8b, v22.8h, #6
138 zip1 v0.16b, v0.16b, v16.16b
139 zip1 v1.16b, v1.16b, v17.16b
140 zip1 v2.16b, v2.16b, v18.16b
142 zip1 v4.16b, v4.16b, v20.16b
143 zip1 v5.16b, v5.16b, v21.16b
144 zip1 v6.16b, v6.16b, v22.16b
153 movi v24.16b, #149
154 movi v25.16b, #50
155 movi v26.16b, #104
156 movi v27.16b, #204
157 movi v28.16b, #254
165 movi v3.16b, #0xff
166 movi v7.16b, #0xff
170 b 2f
173 1: ld2 {v8.16b,v9.16b}, [x1], #32
175 ld2 {v10.16b,v11.16b}, [x3], #32
177 ld1 {v10.16b}, [x3], #16
178 ld1 {v11.16b}, [x4], #16
189 st4 {v0.16b - v3.16b}, [x0], #64
190 st4 {v4.16b - v7.16b}, [x0], #64
192 bhs 1b
204 movi v8.8b, #0
205 movi v9.8b, #0
206 movi v10.8b, #0
207 movi v11.8b, #0
210 ld1 {v9.16b}, [x1], #16
212 ld1 {v11.16b}, [x3], #16
238 ld1 {v10.b}[1], [x3], #1
239 ld1 {v11.b}[1], [x4], #1
242 ld1 {v8.b}[1], [x1], #1
246 ld1 {v10.b}[0], [x3], #1
247 ld1 {v11.b}[0], [x4], #1
255 1: mov v12.16b, v8.16b
256 uzp1 v8.16b, v12.16b, v9.16b
257 uzp2 v9.16b, v12.16b, v9.16b
259 mov v12.16b, v10.16b
260 uzp1 v10.16b, v12.16b, v11.16b
261 uzp2 v11.16b, v12.16b, v11.16b
273 zip1 v16.16b, v0.16b, v2.16b
274 zip2 v18.16b, v0.16b, v2.16b
275 zip1 v17.16b, v1.16b, v3.16b
276 zip2 v19.16b, v1.16b, v3.16b
277 zip1 v0.16b, v16.16b, v17.16b
278 zip2 v1.16b, v16.16b, v17.16b
279 zip1 v2.16b, v18.16b, v19.16b
280 zip2 v3.16b, v18.16b, v19.16b
286 st4 {v4.16b - v7.16b}, [x0], #64
288 st1 {v2.16b,v3.16b}, [x0], #32
290 st1 {v1.16b}, [x0], #16