2 * Copyright (c) 2008 Mans Rullgard <mans@mansr.com>
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 #include "libavutil/arm/asm.S"
24 /* H.264 loop filter */
26 .macro h264_loop_filter_start
33 and r12, r12, r12, lsl #16
36 ands r12, r12, r12, lsl #8
41 .macro h264_loop_filter_luma
42 vdup.8 q11, r2 @ alpha
44 vabd.u8 q6, q8, q0 @ abs(p0 - q0)
46 vabd.u8 q14, q9, q8 @ abs(p1 - p0)
48 vabd.u8 q15, q1, q0 @ abs(q1 - q0)
50 vclt.u8 q6, q6, q11 @ < alpha
53 vclt.u8 q14, q14, q11 @ < beta
54 vclt.u8 q15, q15, q11 @ < beta
56 vabd.u8 q4, q10, q8 @ abs(p2 - p0)
58 vabd.u8 q5, q2, q0 @ abs(q2 - q0)
59 vclt.u8 q4, q4, q11 @ < beta
61 vclt.u8 q5, q5, q11 @ < beta
68 vhadd.u8 q10, q10, q14
81 vsubw.u8 q10, q10, d17
85 vaddw.u8 q10, q10, d19
89 vrshrn.i16 d5, q10, #3
101 vsubw.s8 q11, q11, d4
102 vsubw.s8 q12, q12, d5
109 function ff_h264_v_loop_filter_luma_neon, export=1
110 h264_loop_filter_start
112 vld1.8 {d0, d1}, [r0,:128], r1
113 vld1.8 {d2, d3}, [r0,:128], r1
114 vld1.8 {d4, d5}, [r0,:128], r1
115 sub r0, r0, r1, lsl #2
116 sub r0, r0, r1, lsl #1
117 vld1.8 {d20,d21}, [r0,:128], r1
118 vld1.8 {d18,d19}, [r0,:128], r1
119 vld1.8 {d16,d17}, [r0,:128], r1
123 h264_loop_filter_luma
125 sub r0, r0, r1, lsl #1
126 vst1.8 {d8, d9}, [r0,:128], r1
127 vst1.8 {d16,d17}, [r0,:128], r1
128 vst1.8 {d0, d1}, [r0,:128], r1
129 vst1.8 {d10,d11}, [r0,:128]
135 function ff_h264_h_loop_filter_luma_neon, export=1
136 h264_loop_filter_start
139 vld1.8 {d6}, [r0], r1
140 vld1.8 {d20}, [r0], r1
141 vld1.8 {d18}, [r0], r1
142 vld1.8 {d16}, [r0], r1
143 vld1.8 {d0}, [r0], r1
144 vld1.8 {d2}, [r0], r1
145 vld1.8 {d4}, [r0], r1
146 vld1.8 {d26}, [r0], r1
147 vld1.8 {d7}, [r0], r1
148 vld1.8 {d21}, [r0], r1
149 vld1.8 {d19}, [r0], r1
150 vld1.8 {d17}, [r0], r1
151 vld1.8 {d1}, [r0], r1
152 vld1.8 {d3}, [r0], r1
153 vld1.8 {d5}, [r0], r1
154 vld1.8 {d27}, [r0], r1
156 transpose_8x8 q3, q10, q9, q8, q0, q1, q2, q13
160 h264_loop_filter_luma
162 transpose_4x4 q4, q8, q0, q5
164 sub r0, r0, r1, lsl #4
166 vst1.32 {d8[0]}, [r0], r1
167 vst1.32 {d16[0]}, [r0], r1
168 vst1.32 {d0[0]}, [r0], r1
169 vst1.32 {d10[0]}, [r0], r1
170 vst1.32 {d8[1]}, [r0], r1
171 vst1.32 {d16[1]}, [r0], r1
172 vst1.32 {d0[1]}, [r0], r1
173 vst1.32 {d10[1]}, [r0], r1
174 vst1.32 {d9[0]}, [r0], r1
175 vst1.32 {d17[0]}, [r0], r1
176 vst1.32 {d1[0]}, [r0], r1
177 vst1.32 {d11[0]}, [r0], r1
178 vst1.32 {d9[1]}, [r0], r1
179 vst1.32 {d17[1]}, [r0], r1
180 vst1.32 {d1[1]}, [r0], r1
181 vst1.32 {d11[1]}, [r0], r1
187 .macro h264_loop_filter_chroma
188 vdup.8 d22, r2 @ alpha
190 vabd.u8 d26, d16, d0 @ abs(p0 - q0)
192 vabd.u8 d28, d18, d16 @ abs(p1 - p0)
196 vabd.u8 d30, d2, d0 @ abs(q1 - q0)
198 vclt.u8 d26, d26, d22 @ < alpha
200 vdup.8 d22, r3 @ beta
201 vrshrn.i16 d4, q2, #3
202 vclt.u8 d28, d28, d22 @ < beta
203 vclt.u8 d30, d30, d22 @ < beta
212 vaddw.s8 q14, q14, d4
213 vsubw.s8 q11, q11, d4
218 function ff_h264_v_loop_filter_chroma_neon, export=1
219 h264_loop_filter_start
221 sub r0, r0, r1, lsl #1
222 vld1.8 {d18}, [r0,:64], r1
223 vld1.8 {d16}, [r0,:64], r1
224 vld1.8 {d0}, [r0,:64], r1
225 vld1.8 {d2}, [r0,:64]
227 h264_loop_filter_chroma
229 sub r0, r0, r1, lsl #1
230 vst1.8 {d16}, [r0,:64], r1
231 vst1.8 {d0}, [r0,:64], r1
236 function ff_h264_h_loop_filter_chroma_neon, export=1
237 h264_loop_filter_start
240 vld1.32 {d18[0]}, [r0], r1
241 vld1.32 {d16[0]}, [r0], r1
242 vld1.32 {d0[0]}, [r0], r1
243 vld1.32 {d2[0]}, [r0], r1
244 vld1.32 {d18[1]}, [r0], r1
245 vld1.32 {d16[1]}, [r0], r1
246 vld1.32 {d0[1]}, [r0], r1
247 vld1.32 {d2[1]}, [r0], r1
254 h264_loop_filter_chroma
261 sub r0, r0, r1, lsl #3
262 vst1.32 {d18[0]}, [r0], r1
263 vst1.32 {d16[0]}, [r0], r1
264 vst1.32 {d0[0]}, [r0], r1
265 vst1.32 {d2[0]}, [r0], r1
266 vst1.32 {d18[1]}, [r0], r1
267 vst1.32 {d16[1]}, [r0], r1
268 vst1.32 {d0[1]}, [r0], r1
269 vst1.32 {d2[1]}, [r0], r1
274 @ Biweighted prediction
276 .macro biweight_16 macs, macd
282 vld1.8 {d20-d21},[r0,:128], r2
286 vld1.8 {d22-d23},[r1,:128], r2
291 vld1.8 {d28-d29},[r0,:128], r2
296 vld1.8 {d30-d31},[r1,:128], r2
304 vshl.s16 q12, q12, q9
305 vshl.s16 q13, q13, q9
309 vst1.8 {d4- d5}, [r6,:128], r2
311 vst1.8 {d24-d25},[r6,:128], r2
316 .macro biweight_8 macs, macd
322 vld1.8 {d4},[r0,:64], r2
325 vld1.8 {d5},[r1,:64], r2
328 vld1.8 {d6},[r0,:64], r2
331 vld1.8 {d7},[r1,:64], r2
336 vshl.s16 q10, q10, q9
339 vst1.8 {d2},[r6,:64], r2
341 vst1.8 {d4},[r6,:64], r2
346 .macro biweight_4 macs, macd
352 vld1.32 {d4[0]},[r0,:32], r2
353 vld1.32 {d4[1]},[r0,:32], r2
356 vld1.32 {d5[0]},[r1,:32], r2
357 vld1.32 {d5[1]},[r1,:32], r2
361 vld1.32 {d6[0]},[r0,:32], r2
362 vld1.32 {d6[1]},[r0,:32], r2
365 vld1.32 {d7[0]},[r1,:32], r2
366 vld1.32 {d7[1]},[r1,:32], r2
371 vshl.s16 q10, q10, q9
374 vst1.32 {d2[0]},[r6,:32], r2
375 vst1.32 {d2[1]},[r6,:32], r2
377 vst1.32 {d4[0]},[r6,:32], r2
378 vst1.32 {d4[1]},[r6,:32], r2
381 2: vshl.s16 q1, q1, q9
383 vst1.32 {d2[0]},[r6,:32], r2
384 vst1.32 {d2[1]},[r6,:32], r2
388 .macro biweight_func w
389 function ff_biweight_h264_pixels_\w\()_neon, export=1
396 eors lr, lr, r5, lsr #30
409 10: biweight_\w vmlal.u8, vmlal.u8
411 biweight_\w vmlal.u8, vmlsl.u8
414 biweight_\w vmlsl.u8, vmlsl.u8
416 biweight_\w vmlsl.u8, vmlal.u8
424 @ Weighted prediction
429 vld1.8 {d20-d21},[r0,:128], r1
433 vld1.8 {d28-d29},[r0,:128], r1
434 vmull.u8 q12, d0, d28
436 vmull.u8 q13, d0, d29
444 vrshl.s16 q12, q12, q9
446 vrshl.s16 q13, q13, q9
449 vst1.8 {d4- d5}, [r4,:128], r1
450 vst1.8 {d24-d25},[r4,:128], r1
458 vld1.8 {d4},[r0,:64], r1
461 vld1.8 {d6},[r0,:64], r1
468 vrshl.s16 q10, q10, q9
470 vst1.8 {d2},[r4,:64], r1
471 vst1.8 {d4},[r4,:64], r1
481 vld1.32 {d4[0]},[r0,:32], r1
482 vld1.32 {d4[1]},[r0,:32], r1
486 vld1.32 {d6[0]},[r0,:32], r1
487 vld1.32 {d6[1]},[r0,:32], r1
494 vrshl.s16 q10, q10, q9
497 vst1.32 {d2[0]},[r4,:32], r1
498 vst1.32 {d2[1]},[r4,:32], r1
500 vst1.32 {d4[0]},[r4,:32], r1
501 vst1.32 {d4[1]},[r4,:32], r1
507 vst1.32 {d2[0]},[r4,:32], r1
508 vst1.32 {d2[1]},[r4,:32], r1
513 function ff_weight_h264_pixels_\w\()_neon, export=1