yading@11
|
1 /*
|
yading@11
|
2 * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
|
yading@11
|
3 *
|
yading@11
|
4 * This file is part of FFmpeg
|
yading@11
|
5 *
|
yading@11
|
6 * FFmpeg is free software; you can redistribute it and/or
|
yading@11
|
7 * modify it under the terms of the GNU Lesser General Public
|
yading@11
|
8 * License as published by the Free Software Foundation; either
|
yading@11
|
9 * version 2.1 of the License, or (at your option) any later version.
|
yading@11
|
10 *
|
yading@11
|
11 * FFmpeg is distributed in the hope that it will be useful,
|
yading@11
|
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
yading@11
|
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
yading@11
|
14 * Lesser General Public License for more details.
|
yading@11
|
15 *
|
yading@11
|
16 * You should have received a copy of the GNU Lesser General Public
|
yading@11
|
17 * License along with FFmpeg; if not, write to the Free Software
|
yading@11
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
yading@11
|
19 */
|
yading@11
|
20
|
yading@11
|
21 #include "config.h"
|
yading@11
|
22 #include "libavutil/arm/asm.S"
|
yading@11
|
23
|
yading@11
|
24 function ff_conv_flt_to_s16_neon, export=1
|
yading@11
|
25 subs r2, r2, #8
|
yading@11
|
26 vld1.32 {q0}, [r1,:128]!
|
yading@11
|
27 vcvt.s32.f32 q8, q0, #31
|
yading@11
|
28 vld1.32 {q1}, [r1,:128]!
|
yading@11
|
29 vcvt.s32.f32 q9, q1, #31
|
yading@11
|
30 beq 3f
|
yading@11
|
31 bics r12, r2, #15
|
yading@11
|
32 beq 2f
|
yading@11
|
33 1: subs r12, r12, #16
|
yading@11
|
34 vqrshrn.s32 d4, q8, #16
|
yading@11
|
35 vld1.32 {q0}, [r1,:128]!
|
yading@11
|
36 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
37 vqrshrn.s32 d5, q9, #16
|
yading@11
|
38 vld1.32 {q1}, [r1,:128]!
|
yading@11
|
39 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
40 vqrshrn.s32 d6, q0, #16
|
yading@11
|
41 vst1.16 {q2}, [r0,:128]!
|
yading@11
|
42 vqrshrn.s32 d7, q1, #16
|
yading@11
|
43 vld1.32 {q8}, [r1,:128]!
|
yading@11
|
44 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
45 vld1.32 {q9}, [r1,:128]!
|
yading@11
|
46 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
47 vst1.16 {q3}, [r0,:128]!
|
yading@11
|
48 bne 1b
|
yading@11
|
49 ands r2, r2, #15
|
yading@11
|
50 beq 3f
|
yading@11
|
51 2: vld1.32 {q0}, [r1,:128]!
|
yading@11
|
52 vqrshrn.s32 d4, q8, #16
|
yading@11
|
53 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
54 vld1.32 {q1}, [r1,:128]!
|
yading@11
|
55 vqrshrn.s32 d5, q9, #16
|
yading@11
|
56 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
57 vqrshrn.s32 d6, q0, #16
|
yading@11
|
58 vst1.16 {q2}, [r0,:128]!
|
yading@11
|
59 vqrshrn.s32 d7, q1, #16
|
yading@11
|
60 vst1.16 {q3}, [r0,:128]!
|
yading@11
|
61 bx lr
|
yading@11
|
62 3: vqrshrn.s32 d4, q8, #16
|
yading@11
|
63 vqrshrn.s32 d5, q9, #16
|
yading@11
|
64 vst1.16 {q2}, [r0,:128]!
|
yading@11
|
65 bx lr
|
yading@11
|
66 endfunc
|
yading@11
|
67
|
yading@11
|
68 function ff_conv_fltp_to_s16_2ch_neon, export=1
|
yading@11
|
69 ldm r1, {r1, r3}
|
yading@11
|
70 subs r2, r2, #8
|
yading@11
|
71 vld1.32 {q0}, [r1,:128]!
|
yading@11
|
72 vcvt.s32.f32 q8, q0, #31
|
yading@11
|
73 vld1.32 {q1}, [r1,:128]!
|
yading@11
|
74 vcvt.s32.f32 q9, q1, #31
|
yading@11
|
75 vld1.32 {q10}, [r3,:128]!
|
yading@11
|
76 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
77 vld1.32 {q11}, [r3,:128]!
|
yading@11
|
78 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
79 beq 3f
|
yading@11
|
80 bics r12, r2, #15
|
yading@11
|
81 beq 2f
|
yading@11
|
82 1: subs r12, r12, #16
|
yading@11
|
83 vld1.32 {q0}, [r1,:128]!
|
yading@11
|
84 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
85 vsri.32 q10, q8, #16
|
yading@11
|
86 vld1.32 {q1}, [r1,:128]!
|
yading@11
|
87 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
88 vld1.32 {q12}, [r3,:128]!
|
yading@11
|
89 vcvt.s32.f32 q12, q12, #31
|
yading@11
|
90 vld1.32 {q13}, [r3,:128]!
|
yading@11
|
91 vsri.32 q11, q9, #16
|
yading@11
|
92 vst1.16 {q10}, [r0,:128]!
|
yading@11
|
93 vcvt.s32.f32 q13, q13, #31
|
yading@11
|
94 vst1.16 {q11}, [r0,:128]!
|
yading@11
|
95 vsri.32 q12, q0, #16
|
yading@11
|
96 vld1.32 {q8}, [r1,:128]!
|
yading@11
|
97 vsri.32 q13, q1, #16
|
yading@11
|
98 vst1.16 {q12}, [r0,:128]!
|
yading@11
|
99 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
100 vld1.32 {q9}, [r1,:128]!
|
yading@11
|
101 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
102 vld1.32 {q10}, [r3,:128]!
|
yading@11
|
103 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
104 vld1.32 {q11}, [r3,:128]!
|
yading@11
|
105 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
106 vst1.16 {q13}, [r0,:128]!
|
yading@11
|
107 bne 1b
|
yading@11
|
108 ands r2, r2, #15
|
yading@11
|
109 beq 3f
|
yading@11
|
110 2: vsri.32 q10, q8, #16
|
yading@11
|
111 vld1.32 {q0}, [r1,:128]!
|
yading@11
|
112 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
113 vld1.32 {q1}, [r1,:128]!
|
yading@11
|
114 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
115 vld1.32 {q12}, [r3,:128]!
|
yading@11
|
116 vcvt.s32.f32 q12, q12, #31
|
yading@11
|
117 vsri.32 q11, q9, #16
|
yading@11
|
118 vld1.32 {q13}, [r3,:128]!
|
yading@11
|
119 vcvt.s32.f32 q13, q13, #31
|
yading@11
|
120 vst1.16 {q10}, [r0,:128]!
|
yading@11
|
121 vsri.32 q12, q0, #16
|
yading@11
|
122 vst1.16 {q11}, [r0,:128]!
|
yading@11
|
123 vsri.32 q13, q1, #16
|
yading@11
|
124 vst1.16 {q12-q13},[r0,:128]!
|
yading@11
|
125 bx lr
|
yading@11
|
126 3: vsri.32 q10, q8, #16
|
yading@11
|
127 vsri.32 q11, q9, #16
|
yading@11
|
128 vst1.16 {q10-q11},[r0,:128]!
|
yading@11
|
129 bx lr
|
yading@11
|
130 endfunc
|
yading@11
|
131
|
yading@11
|
132 function ff_conv_fltp_to_s16_neon, export=1
|
yading@11
|
133 cmp r3, #2
|
yading@11
|
134 itt lt
|
yading@11
|
135 ldrlt r1, [r1]
|
yading@11
|
136 blt ff_conv_flt_to_s16_neon
|
yading@11
|
137 beq ff_conv_fltp_to_s16_2ch_neon
|
yading@11
|
138
|
yading@11
|
139 push {r4-r8, lr}
|
yading@11
|
140 cmp r3, #4
|
yading@11
|
141 lsl r12, r3, #1
|
yading@11
|
142 blt 4f
|
yading@11
|
143
|
yading@11
|
144 @ 4 channels
|
yading@11
|
145 5: ldm r1!, {r4-r7}
|
yading@11
|
146 mov lr, r2
|
yading@11
|
147 mov r8, r0
|
yading@11
|
148 vld1.32 {q8}, [r4,:128]!
|
yading@11
|
149 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
150 vld1.32 {q9}, [r5,:128]!
|
yading@11
|
151 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
152 vld1.32 {q10}, [r6,:128]!
|
yading@11
|
153 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
154 vld1.32 {q11}, [r7,:128]!
|
yading@11
|
155 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
156 6: subs lr, lr, #8
|
yading@11
|
157 vld1.32 {q0}, [r4,:128]!
|
yading@11
|
158 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
159 vsri.32 q9, q8, #16
|
yading@11
|
160 vld1.32 {q1}, [r5,:128]!
|
yading@11
|
161 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
162 vsri.32 q11, q10, #16
|
yading@11
|
163 vld1.32 {q2}, [r6,:128]!
|
yading@11
|
164 vcvt.s32.f32 q2, q2, #31
|
yading@11
|
165 vzip.32 d18, d22
|
yading@11
|
166 vld1.32 {q3}, [r7,:128]!
|
yading@11
|
167 vcvt.s32.f32 q3, q3, #31
|
yading@11
|
168 vzip.32 d19, d23
|
yading@11
|
169 vst1.16 {d18}, [r8], r12
|
yading@11
|
170 vsri.32 q1, q0, #16
|
yading@11
|
171 vst1.16 {d22}, [r8], r12
|
yading@11
|
172 vsri.32 q3, q2, #16
|
yading@11
|
173 vst1.16 {d19}, [r8], r12
|
yading@11
|
174 vzip.32 d2, d6
|
yading@11
|
175 vst1.16 {d23}, [r8], r12
|
yading@11
|
176 vzip.32 d3, d7
|
yading@11
|
177 beq 7f
|
yading@11
|
178 vld1.32 {q8}, [r4,:128]!
|
yading@11
|
179 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
180 vst1.16 {d2}, [r8], r12
|
yading@11
|
181 vld1.32 {q9}, [r5,:128]!
|
yading@11
|
182 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
183 vst1.16 {d6}, [r8], r12
|
yading@11
|
184 vld1.32 {q10}, [r6,:128]!
|
yading@11
|
185 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
186 vst1.16 {d3}, [r8], r12
|
yading@11
|
187 vld1.32 {q11}, [r7,:128]!
|
yading@11
|
188 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
189 vst1.16 {d7}, [r8], r12
|
yading@11
|
190 b 6b
|
yading@11
|
191 7: vst1.16 {d2}, [r8], r12
|
yading@11
|
192 vst1.16 {d6}, [r8], r12
|
yading@11
|
193 vst1.16 {d3}, [r8], r12
|
yading@11
|
194 vst1.16 {d7}, [r8], r12
|
yading@11
|
195 subs r3, r3, #4
|
yading@11
|
196 it eq
|
yading@11
|
197 popeq {r4-r8, pc}
|
yading@11
|
198 cmp r3, #4
|
yading@11
|
199 add r0, r0, #8
|
yading@11
|
200 bge 5b
|
yading@11
|
201
|
yading@11
|
202 @ 2 channels
|
yading@11
|
203 4: cmp r3, #2
|
yading@11
|
204 blt 4f
|
yading@11
|
205 ldm r1!, {r4-r5}
|
yading@11
|
206 mov lr, r2
|
yading@11
|
207 mov r8, r0
|
yading@11
|
208 tst lr, #8
|
yading@11
|
209 vld1.32 {q8}, [r4,:128]!
|
yading@11
|
210 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
211 vld1.32 {q9}, [r5,:128]!
|
yading@11
|
212 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
213 vld1.32 {q10}, [r4,:128]!
|
yading@11
|
214 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
215 vld1.32 {q11}, [r5,:128]!
|
yading@11
|
216 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
217 beq 6f
|
yading@11
|
218 subs lr, lr, #8
|
yading@11
|
219 beq 7f
|
yading@11
|
220 vsri.32 d18, d16, #16
|
yading@11
|
221 vsri.32 d19, d17, #16
|
yading@11
|
222 vld1.32 {q8}, [r4,:128]!
|
yading@11
|
223 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
224 vst1.32 {d18[0]}, [r8], r12
|
yading@11
|
225 vsri.32 d22, d20, #16
|
yading@11
|
226 vst1.32 {d18[1]}, [r8], r12
|
yading@11
|
227 vsri.32 d23, d21, #16
|
yading@11
|
228 vst1.32 {d19[0]}, [r8], r12
|
yading@11
|
229 vst1.32 {d19[1]}, [r8], r12
|
yading@11
|
230 vld1.32 {q9}, [r5,:128]!
|
yading@11
|
231 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
232 vst1.32 {d22[0]}, [r8], r12
|
yading@11
|
233 vst1.32 {d22[1]}, [r8], r12
|
yading@11
|
234 vld1.32 {q10}, [r4,:128]!
|
yading@11
|
235 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
236 vst1.32 {d23[0]}, [r8], r12
|
yading@11
|
237 vst1.32 {d23[1]}, [r8], r12
|
yading@11
|
238 vld1.32 {q11}, [r5,:128]!
|
yading@11
|
239 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
240 6: subs lr, lr, #16
|
yading@11
|
241 vld1.32 {q0}, [r4,:128]!
|
yading@11
|
242 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
243 vsri.32 d18, d16, #16
|
yading@11
|
244 vld1.32 {q1}, [r5,:128]!
|
yading@11
|
245 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
246 vsri.32 d19, d17, #16
|
yading@11
|
247 vld1.32 {q2}, [r4,:128]!
|
yading@11
|
248 vcvt.s32.f32 q2, q2, #31
|
yading@11
|
249 vld1.32 {q3}, [r5,:128]!
|
yading@11
|
250 vcvt.s32.f32 q3, q3, #31
|
yading@11
|
251 vst1.32 {d18[0]}, [r8], r12
|
yading@11
|
252 vsri.32 d22, d20, #16
|
yading@11
|
253 vst1.32 {d18[1]}, [r8], r12
|
yading@11
|
254 vsri.32 d23, d21, #16
|
yading@11
|
255 vst1.32 {d19[0]}, [r8], r12
|
yading@11
|
256 vsri.32 d2, d0, #16
|
yading@11
|
257 vst1.32 {d19[1]}, [r8], r12
|
yading@11
|
258 vsri.32 d3, d1, #16
|
yading@11
|
259 vst1.32 {d22[0]}, [r8], r12
|
yading@11
|
260 vsri.32 d6, d4, #16
|
yading@11
|
261 vst1.32 {d22[1]}, [r8], r12
|
yading@11
|
262 vsri.32 d7, d5, #16
|
yading@11
|
263 vst1.32 {d23[0]}, [r8], r12
|
yading@11
|
264 vst1.32 {d23[1]}, [r8], r12
|
yading@11
|
265 beq 6f
|
yading@11
|
266 vld1.32 {q8}, [r4,:128]!
|
yading@11
|
267 vcvt.s32.f32 q8, q8, #31
|
yading@11
|
268 vst1.32 {d2[0]}, [r8], r12
|
yading@11
|
269 vst1.32 {d2[1]}, [r8], r12
|
yading@11
|
270 vld1.32 {q9}, [r5,:128]!
|
yading@11
|
271 vcvt.s32.f32 q9, q9, #31
|
yading@11
|
272 vst1.32 {d3[0]}, [r8], r12
|
yading@11
|
273 vst1.32 {d3[1]}, [r8], r12
|
yading@11
|
274 vld1.32 {q10}, [r4,:128]!
|
yading@11
|
275 vcvt.s32.f32 q10, q10, #31
|
yading@11
|
276 vst1.32 {d6[0]}, [r8], r12
|
yading@11
|
277 vst1.32 {d6[1]}, [r8], r12
|
yading@11
|
278 vld1.32 {q11}, [r5,:128]!
|
yading@11
|
279 vcvt.s32.f32 q11, q11, #31
|
yading@11
|
280 vst1.32 {d7[0]}, [r8], r12
|
yading@11
|
281 vst1.32 {d7[1]}, [r8], r12
|
yading@11
|
282 bgt 6b
|
yading@11
|
283 6: vst1.32 {d2[0]}, [r8], r12
|
yading@11
|
284 vst1.32 {d2[1]}, [r8], r12
|
yading@11
|
285 vst1.32 {d3[0]}, [r8], r12
|
yading@11
|
286 vst1.32 {d3[1]}, [r8], r12
|
yading@11
|
287 vst1.32 {d6[0]}, [r8], r12
|
yading@11
|
288 vst1.32 {d6[1]}, [r8], r12
|
yading@11
|
289 vst1.32 {d7[0]}, [r8], r12
|
yading@11
|
290 vst1.32 {d7[1]}, [r8], r12
|
yading@11
|
291 b 8f
|
yading@11
|
292 7: vsri.32 d18, d16, #16
|
yading@11
|
293 vsri.32 d19, d17, #16
|
yading@11
|
294 vst1.32 {d18[0]}, [r8], r12
|
yading@11
|
295 vsri.32 d22, d20, #16
|
yading@11
|
296 vst1.32 {d18[1]}, [r8], r12
|
yading@11
|
297 vsri.32 d23, d21, #16
|
yading@11
|
298 vst1.32 {d19[0]}, [r8], r12
|
yading@11
|
299 vst1.32 {d19[1]}, [r8], r12
|
yading@11
|
300 vst1.32 {d22[0]}, [r8], r12
|
yading@11
|
301 vst1.32 {d22[1]}, [r8], r12
|
yading@11
|
302 vst1.32 {d23[0]}, [r8], r12
|
yading@11
|
303 vst1.32 {d23[1]}, [r8], r12
|
yading@11
|
304 8: subs r3, r3, #2
|
yading@11
|
305 add r0, r0, #4
|
yading@11
|
306 it eq
|
yading@11
|
307 popeq {r4-r8, pc}
|
yading@11
|
308
|
yading@11
|
309 @ 1 channel
|
yading@11
|
310 4: ldr r4, [r1]
|
yading@11
|
311 tst r2, #8
|
yading@11
|
312 mov lr, r2
|
yading@11
|
313 mov r5, r0
|
yading@11
|
314 vld1.32 {q0}, [r4,:128]!
|
yading@11
|
315 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
316 vld1.32 {q1}, [r4,:128]!
|
yading@11
|
317 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
318 bne 8f
|
yading@11
|
319 6: subs lr, lr, #16
|
yading@11
|
320 vld1.32 {q2}, [r4,:128]!
|
yading@11
|
321 vcvt.s32.f32 q2, q2, #31
|
yading@11
|
322 vld1.32 {q3}, [r4,:128]!
|
yading@11
|
323 vcvt.s32.f32 q3, q3, #31
|
yading@11
|
324 vst1.16 {d0[1]}, [r5,:16], r12
|
yading@11
|
325 vst1.16 {d0[3]}, [r5,:16], r12
|
yading@11
|
326 vst1.16 {d1[1]}, [r5,:16], r12
|
yading@11
|
327 vst1.16 {d1[3]}, [r5,:16], r12
|
yading@11
|
328 vst1.16 {d2[1]}, [r5,:16], r12
|
yading@11
|
329 vst1.16 {d2[3]}, [r5,:16], r12
|
yading@11
|
330 vst1.16 {d3[1]}, [r5,:16], r12
|
yading@11
|
331 vst1.16 {d3[3]}, [r5,:16], r12
|
yading@11
|
332 beq 7f
|
yading@11
|
333 vld1.32 {q0}, [r4,:128]!
|
yading@11
|
334 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
335 vld1.32 {q1}, [r4,:128]!
|
yading@11
|
336 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
337 7: vst1.16 {d4[1]}, [r5,:16], r12
|
yading@11
|
338 vst1.16 {d4[3]}, [r5,:16], r12
|
yading@11
|
339 vst1.16 {d5[1]}, [r5,:16], r12
|
yading@11
|
340 vst1.16 {d5[3]}, [r5,:16], r12
|
yading@11
|
341 vst1.16 {d6[1]}, [r5,:16], r12
|
yading@11
|
342 vst1.16 {d6[3]}, [r5,:16], r12
|
yading@11
|
343 vst1.16 {d7[1]}, [r5,:16], r12
|
yading@11
|
344 vst1.16 {d7[3]}, [r5,:16], r12
|
yading@11
|
345 bgt 6b
|
yading@11
|
346 pop {r4-r8, pc}
|
yading@11
|
347 8: subs lr, lr, #8
|
yading@11
|
348 vst1.16 {d0[1]}, [r5,:16], r12
|
yading@11
|
349 vst1.16 {d0[3]}, [r5,:16], r12
|
yading@11
|
350 vst1.16 {d1[1]}, [r5,:16], r12
|
yading@11
|
351 vst1.16 {d1[3]}, [r5,:16], r12
|
yading@11
|
352 vst1.16 {d2[1]}, [r5,:16], r12
|
yading@11
|
353 vst1.16 {d2[3]}, [r5,:16], r12
|
yading@11
|
354 vst1.16 {d3[1]}, [r5,:16], r12
|
yading@11
|
355 vst1.16 {d3[3]}, [r5,:16], r12
|
yading@11
|
356 it eq
|
yading@11
|
357 popeq {r4-r8, pc}
|
yading@11
|
358 vld1.32 {q0}, [r4,:128]!
|
yading@11
|
359 vcvt.s32.f32 q0, q0, #31
|
yading@11
|
360 vld1.32 {q1}, [r4,:128]!
|
yading@11
|
361 vcvt.s32.f32 q1, q1, #31
|
yading@11
|
362 b 6b
|
yading@11
|
363 endfunc
|