Lines Matching full:s
28 #include "libavutil/aarch64/asm.S"
39 ld1 {v0.2s,v1.2s,v2.2s,v3.2s}, [x0]
41 fadd v4.2s, v0.2s, v1.2s // r0+r1,i0+i1
42 fsub v6.2s, v0.2s, v1.2s // r0-r1,i0-i1
47 fadd v5.2s, v2.2s, v3.2s // i2+i3,r2+r3
48 fsub v7.2s, v16.2s, v17.2s // r3-r2,i2-i3
50 fadd v0.2s, v4.2s, v5.2s
51 fsub v2.2s, v4.2s, v5.2s
52 fadd v1.2s, v6.2s, v7.2s
53 fsub v3.2s, v6.2s, v7.2s
55 st1 {v0.2s,v1.2s,v2.2s,v3.2s}, [x0]
62 ld1 {v0.2s, v1.2s, v2.2s, v3.2s}, [x0], #32
63 ld1 {v16.2s,v17.2s,v18.2s,v19.2s}, [x0]
66 fadd v4.2s, v16.2s, v17.2s // r4+r5,i4+i5
67 fadd v5.2s, v18.2s, v19.2s // r6+r7,i6+i7
68 fsub v17.2s, v16.2s, v17.2s // r4-r5,i4-i5
69 fsub v19.2s, v18.2s, v19.2s // r6-r7,i6-i7
70 rev64 v27.2s, v28.2s // ???
71 fadd v20.2s, v0.2s, v1.2s // r0+r1,i0+i1
72 fadd v21.2s, v2.2s, v3.2s // r2+r3,i2+i3
73 fmul v26.2s, v17.2s, v28.2s // -a2r*w,a2i*w
76 fmul v27.2s, v19.2s, v27.2s // a3r*w,-a3i*w
77 fsub v23.2s, v22.2s, v23.2s // i2-i3,r3-r2
78 fsub v22.2s, v0.2s, v1.2s // r0-r1,i0-i1
79 fmul v24.2s, v17.2s, v28.s[1] // a2r*w,a2i*w
80 fmul v25.2s, v19.2s, v28.s[1] // a3r*w,a3i*w
81 fadd v0.2s, v20.2s, v21.2s
82 fsub v2.2s, v20.2s, v21.2s
83 fadd v1.2s, v22.2s, v23.2s
84 rev64 v26.2s, v26.2s
85 rev64 v27.2s, v27.2s
86 fsub v3.2s, v22.2s, v23.2s
87 fsub v6.2s, v6.2s, v7.2s
88 fadd v24.2s, v24.2s, v26.2s // a2r+a2i,a2i-a2r t1,t2
89 fadd v25.2s, v25.2s, v27.2s // a3r-a3i,a3i+a3r t5,t6
90 fadd v7.2s, v4.2s, v5.2s
91 fsub v18.2s, v2.2s, v6.2s
94 fadd v2.2s, v2.2s, v6.2s
95 fsub v16.2s, v0.2s, v7.2s
96 fadd v5.2s, v25.2s, v24.2s
97 fsub v4.2s, v26.2s, v27.2s
98 fadd v0.2s, v0.2s, v7.2s
99 fsub v17.2s, v1.2s, v5.2s
100 fsub v19.2s, v3.2s, v4.2s
101 fadd v3.2s, v3.2s, v4.2s
102 fadd v1.2s, v1.2s, v5.2s
104 st1 {v16.2s,v17.2s,v18.2s,v19.2s}, [x0]
105 st1 {v0.2s, v1.2s, v2.2s, v3.2s}, [x1]
112 ld1 {v0.2s, v1.2s, v2.2s, v3.2s}, [x0], #32
113 ld1 {v16.2s,v17.2s,v18.2s,v19.2s}, [x0], #32
116 fadd v4.2s, v16.2s, v17.2s // r4+r5,i4+i5
117 fadd v5.2s, v18.2s, v19.2s // r6+r7,i6+i7
118 fsub v17.2s, v16.2s, v17.2s // r4-r5,i4-i5
119 fsub v19.2s, v18.2s, v19.2s // r6-r7,i6-i7
120 rev64 v27.2s, v28.2s // ???
121 fadd v20.2s, v0.2s, v1.2s // r0+r1,i0+i1
122 fadd v21.2s, v2.2s, v3.2s // r2+r3,i2+i3
123 fmul v26.2s, v17.2s, v28.2s // -a2r*w,a2i*w
126 fmul v27.2s, v19.2s, v27.2s // a3r*w,-a3i*w
127 fsub v23.2s, v22.2s, v23.2s // i2-i3,r3-r2
128 fsub v22.2s, v0.2s, v1.2s // r0-r1,i0-i1
129 fmul v24.2s, v17.2s, v28.s[1] // a2r*w,a2i*w
130 fmul v25.2s, v19.2s, v28.s[1] // a3r*w,a3i*w
131 fadd v0.2s, v20.2s, v21.2s
132 fsub v2.2s, v20.2s, v21.2s
133 fadd v1.2s, v22.2s, v23.2s
134 rev64 v26.2s, v26.2s
135 rev64 v27.2s, v27.2s
136 fsub v3.2s, v22.2s, v23.2s
137 fsub v6.2s, v6.2s, v7.2s
138 fadd v24.2s, v24.2s, v26.2s // a2r+a2i,a2i-a2r t1,t2
139 fadd v25.2s, v25.2s, v27.2s // a3r-a3i,a3i+a3r t5,t6
140 fadd v7.2s, v4.2s, v5.2s
141 fsub v18.2s, v2.2s, v6.2s
142 ld1 {v20.4s,v21.4s}, [x0], #32
143 ld1 {v22.4s,v23.4s}, [x0], #32
146 fadd v2.2s, v2.2s, v6.2s
147 fsub v16.2s, v0.2s, v7.2s
148 fadd v5.2s, v25.2s, v24.2s
149 fsub v4.2s, v26.2s, v27.2s
152 fadd v0.2s, v0.2s, v7.2s
153 fsub v17.2s, v1.2s, v5.2s
154 fsub v19.2s, v3.2s, v4.2s
155 fadd v3.2s, v3.2s, v4.2s
156 fadd v1.2s, v1.2s, v5.2s
168 fadd v4.4s, v24.4s, v25.4s
169 fadd v5.4s, v26.4s, v27.4s
170 fsub v6.4s, v24.4s, v25.4s
171 fsub v7.4s, v22.4s, v23.4s
173 ld1 {v23.4s}, [x14]
175 fadd v24.4s, v4.4s, v5.4s // {z[o2+0],z[o2+1]}
176 fsub v26.4s, v4.4s, v5.4s // {z[o2+2],z[o2+3]}
177 fadd v25.4s, v6.4s, v7.4s // {z[o3+0],z[o3+1]}
178 fsub v27.4s, v6.4s, v7.4s // {z[o3+2],z[o3+3]}
181 rev64 v7.4s, v25.4s
182 fmul v25.4s, v25.4s, v23.s[1]
183 fmul v7.4s, v7.4s, v29.4s
184 fmla v25.4s, v7.4s, v23.s[3] // {t1a,t2a,t5a,t6a}
186 zip1 v20.4s, v24.4s, v25.4s
187 zip2 v21.4s, v24.4s, v25.4s
188 fneg v22.4s, v20.4s
189 fadd v4.4s, v21.4s, v20.4s
190 fsub v6.4s, v20.4s, v21.4s // just the second half
191 fadd v5.4s, v21.4s, v22.4s // just the first half
196 fsub v20.4s, v0.4s, v4.4s // {z[o2],z[o2+1]}
197 fadd v16.4s, v0.4s, v4.4s // {z[0], z[1]}
198 fsub v22.4s, v2.4s, v5.4s // {z[o3],z[o3+1]}
199 fadd v18.4s, v2.4s, v5.4s // {z[o1],z[o1+1]}
202 rev64 v6.4s, v26.4s
203 fmul v26.4s, v26.4s, v23.s[2]
204 rev64 v7.4s, v27.4s
205 fmul v27.4s, v27.4s, v23.s[3]
206 fmul v6.4s, v6.4s, v29.4s
207 fmul v7.4s, v7.4s, v29.4s
208 fmla v26.4s, v6.4s, v23.s[2] // {t1,t2,t5,t6}
209 fmla v27.4s, v7.4s, v23.s[1] // {t1a,t2a,t5a,t6a}
211 zip1 v24.4s, v26.4s, v27.4s
212 zip2 v25.4s, v26.4s, v27.4s
213 fneg v26.4s, v24.4s
214 fadd v4.4s, v25.4s, v24.4s
215 fsub v6.4s, v24.4s, v25.4s // just the second half
216 fadd v5.4s, v25.4s, v26.4s // just the first half
221 fadd v17.4s, v1.4s, v4.4s // {z[2], z[3]}
222 fsub v21.4s, v1.4s, v4.4s // {z[o2+2],z[o2+3]}
223 fadd v19.4s, v3.4s, v5.4s // {z[o1+2],z[o1+3]}
224 fsub v23.4s, v3.4s, v5.4s // {z[o3+2],z[o3+3]}
226 st1 {v16.4s,v17.4s}, [x1], #32
227 st1 {v18.4s,v19.4s}, [x1], #32
228 st1 {v20.4s,v21.4s}, [x1], #32
229 st1 {v22.4s,v23.4s}, [x1], #32
258 ld1 {v20.4s},[x2] // {z[o2],z[o2+1]}
259 ld1 {v22.4s},[x3] // {z[o3],z[o3+1]}
260 ld1 {v4.2s}, [x4], #8 // {wre[0],wre[1]}
264 ld1 {v5.s}[0], [x5], x7 // d5[0] = wim[-1]
265 rev64 v7.4s, v25.4s
266 fmul v25.4s, v25.4s, v4.s[1]
267 ld1 {v16.4s}, [x0] // {z[0],z[1]}
268 fmul v7.4s, v7.4s, v29.4s
269 ld1 {v17.4s}, [x1] // {z[o1],z[o1+1]}
272 fmla v25.4s, v7.4s, v5.s[0] // {t1a,t2a,t5a,t6a}
276 zip1 v20.4s, v24.4s, v25.4s
277 zip2 v21.4s, v24.4s, v25.4s
278 fneg v22.4s, v20.4s
279 fadd v4.4s, v21.4s, v20.4s
280 fsub v6.4s, v20.4s, v21.4s // just the second half
281 fadd v5.4s, v21.4s, v22.4s // just the first half
286 fadd v20.4s, v16.4s, v4.4s
287 fsub v22.4s, v16.4s, v4.4s
288 fadd v21.4s, v17.4s, v5.4s
289 st1 {v20.4s}, [x0], #16 // {z[0], z[1]}
290 fsub v23.4s, v17.4s, v5.4s
292 st1 {v21.4s}, [x1], #16 // {z[o1],z[o1+1]}
293 st1 {v22.4s}, [x2], #16 // {z[o2],z[o2+1]}
294 st1 {v23.4s}, [x3], #16 // {z[o3],z[o3+1]}
296 ld1 {v20.4s},[x2] // {z[o2],z[o2+1]}
297 ld1 {v22.4s},[x3] // {z[o3],z[o3+1]}
298 ld1 {v4.2s}, [x4], #8 // {wre[0],wre[1]}
300 ld1 {v5.2s}, [x5], x7 // {wim[-1],wim[0]}
301 rev64 v6.4s, v26.4s
302 fmul v26.4s, v26.4s, v4.s[0]
303 rev64 v7.4s, v27.4s
304 fmul v27.4s, v27.4s, v4.s[1]
305 fmul v6.4s, v6.4s, v29.4s
306 fmul v7.4s, v7.4s, v29.4s
307 ld1 {v16.4s},[x0] // {z[0],z[1]}
308 fmla v26.4s, v6.4s, v5.s[1] // {t1,t2,t5,t6}
309 fmla v27.4s, v7.4s, v5.s[0] // {t1a,t2a,t5a,t6a}
310 ld1 {v17.4s},[x1] // {z[o1],z[o1+1]}
314 zip1 v20.4s, v26.4s, v27.4s
315 zip2 v21.4s, v26.4s, v27.4s
316 fneg v22.4s, v20.4s
317 fadd v4.4s, v21.4s, v20.4s
318 fsub v6.4s, v20.4s, v21.4s // just the second half
319 fadd v5.4s, v21.4s, v22.4s // just the first half
324 fadd v20.4s, v16.4s, v4.4s
325 fsub v22.4s, v16.4s, v4.4s
326 fadd v21.4s, v17.4s, v5.4s
327 st1 {v20.4s}, [x0], #16 // {z[0], z[1]}
328 fsub v23.4s, v17.4s, v5.4s
330 st1 {v21.4s}, [x1], #16 // {z[o1],z[o1+1]}
331 st1 {v22.4s}, [x2], #16 // {z[o2],z[o2+1]}
332 st1 {v23.4s}, [x3], #16 // {z[o3],z[o3+1]}
384 ld1 {v29.4s}, [x12] // pmmp
385 ld1 {v28.4s}, [x13]
397 ld1 {v0.2s,v1.2s}, [x1], #16
403 st1 {v0.2s}, [x5]
404 st1 {v1.2s}, [x4]
410 ld1 {v0.4s,v1.4s}, [x3], #32
411 st1 {v0.4s,v1.4s}, [x1], #32