Mercurial > libavcodec.hg
annotate x86/h264dsp_mmx.c @ 12488:351a81a23343 libavcodec
Set a constant frame size for encoding G.726 audio.
author | jbr |
---|---|
date | Sat, 11 Sep 2010 19:52:09 +0000 |
parents | 14896fa76003 |
children | 58a960d6e34c |
rev | line source |
---|---|
8430 | 1 /* |
2 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt | |
3 * | |
4 * This file is part of FFmpeg. | |
5 * | |
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. | |
10 * | |
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. | |
15 * | |
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 | |
19 */ | |
20 | |
12475
9fef0a8ddd63
Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents:
12456
diff
changeset
|
21 #include "libavutil/cpu.h" |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
22 #include "libavutil/x86_cpu.h" |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
23 #include "libavcodec/h264dsp.h" |
8430 | 24 #include "dsputil_mmx.h" |
25 | |
11369 | 26 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_3_1 ) = 0x0103010301030103ULL; |
27 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_7_3 ) = 0x0307030703070307ULL; | |
8430 | 28 |
29 /***********************************/ | |
30 /* IDCT */ | |
31 | |
32 #define SUMSUB_BADC( a, b, c, d ) \ | |
33 "paddw "#b", "#a" \n\t"\ | |
34 "paddw "#d", "#c" \n\t"\ | |
35 "paddw "#b", "#b" \n\t"\ | |
36 "paddw "#d", "#d" \n\t"\ | |
37 "psubw "#a", "#b" \n\t"\ | |
38 "psubw "#c", "#d" \n\t" | |
39 | |
40 #define SUMSUBD2_AB( a, b, t ) \ | |
41 "movq "#b", "#t" \n\t"\ | |
42 "psraw $1 , "#b" \n\t"\ | |
43 "paddw "#a", "#b" \n\t"\ | |
44 "psraw $1 , "#a" \n\t"\ | |
45 "psubw "#t", "#a" \n\t" | |
46 | |
47 #define IDCT4_1D( s02, s13, d02, d13, t ) \ | |
48 SUMSUB_BA ( s02, d02 )\ | |
49 SUMSUBD2_AB( s13, d13, t )\ | |
50 SUMSUB_BADC( d13, s02, s13, d02 ) | |
51 | |
52 #define STORE_DIFF_4P( p, t, z ) \ | |
53 "psraw $6, "#p" \n\t"\ | |
54 "movd (%0), "#t" \n\t"\ | |
55 "punpcklbw "#z", "#t" \n\t"\ | |
56 "paddsw "#t", "#p" \n\t"\ | |
57 "packuswb "#z", "#p" \n\t"\ | |
58 "movd "#p", (%0) \n\t" | |
59 | |
60 static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride) | |
61 { | |
62 /* Load dct coeffs */ | |
63 __asm__ volatile( | |
64 "movq (%0), %%mm0 \n\t" | |
65 "movq 8(%0), %%mm1 \n\t" | |
66 "movq 16(%0), %%mm2 \n\t" | |
67 "movq 24(%0), %%mm3 \n\t" | |
68 :: "r"(block) ); | |
69 | |
70 __asm__ volatile( | |
71 /* mm1=s02+s13 mm2=s02-s13 mm4=d02+d13 mm0=d02-d13 */ | |
72 IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 ) | |
73 | |
74 "movq %0, %%mm6 \n\t" | |
75 /* in: 1,4,0,2 out: 1,2,3,0 */ | |
76 TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 ) | |
77 | |
78 "paddw %%mm6, %%mm3 \n\t" | |
79 | |
80 /* mm2=s02+s13 mm3=s02-s13 mm4=d02+d13 mm1=d02-d13 */ | |
81 IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 ) | |
82 | |
83 "pxor %%mm7, %%mm7 \n\t" | |
84 :: "m"(ff_pw_32)); | |
85 | |
86 __asm__ volatile( | |
87 STORE_DIFF_4P( %%mm0, %%mm1, %%mm7) | |
88 "add %1, %0 \n\t" | |
89 STORE_DIFF_4P( %%mm2, %%mm1, %%mm7) | |
90 "add %1, %0 \n\t" | |
91 STORE_DIFF_4P( %%mm3, %%mm1, %%mm7) | |
92 "add %1, %0 \n\t" | |
93 STORE_DIFF_4P( %%mm4, %%mm1, %%mm7) | |
94 : "+r"(dst) | |
95 : "r" ((x86_reg)stride) | |
96 ); | |
97 } | |
98 | |
99 static inline void h264_idct8_1d(int16_t *block) | |
100 { | |
101 __asm__ volatile( | |
102 "movq 112(%0), %%mm7 \n\t" | |
103 "movq 80(%0), %%mm0 \n\t" | |
104 "movq 48(%0), %%mm3 \n\t" | |
105 "movq 16(%0), %%mm5 \n\t" | |
106 | |
107 "movq %%mm0, %%mm4 \n\t" | |
108 "movq %%mm5, %%mm1 \n\t" | |
109 "psraw $1, %%mm4 \n\t" | |
110 "psraw $1, %%mm1 \n\t" | |
111 "paddw %%mm0, %%mm4 \n\t" | |
112 "paddw %%mm5, %%mm1 \n\t" | |
113 "paddw %%mm7, %%mm4 \n\t" | |
114 "paddw %%mm0, %%mm1 \n\t" | |
115 "psubw %%mm5, %%mm4 \n\t" | |
116 "paddw %%mm3, %%mm1 \n\t" | |
117 | |
118 "psubw %%mm3, %%mm5 \n\t" | |
119 "psubw %%mm3, %%mm0 \n\t" | |
120 "paddw %%mm7, %%mm5 \n\t" | |
121 "psubw %%mm7, %%mm0 \n\t" | |
122 "psraw $1, %%mm3 \n\t" | |
123 "psraw $1, %%mm7 \n\t" | |
124 "psubw %%mm3, %%mm5 \n\t" | |
125 "psubw %%mm7, %%mm0 \n\t" | |
126 | |
127 "movq %%mm4, %%mm3 \n\t" | |
128 "movq %%mm1, %%mm7 \n\t" | |
129 "psraw $2, %%mm1 \n\t" | |
130 "psraw $2, %%mm3 \n\t" | |
131 "paddw %%mm5, %%mm3 \n\t" | |
132 "psraw $2, %%mm5 \n\t" | |
133 "paddw %%mm0, %%mm1 \n\t" | |
134 "psraw $2, %%mm0 \n\t" | |
135 "psubw %%mm4, %%mm5 \n\t" | |
136 "psubw %%mm0, %%mm7 \n\t" | |
137 | |
138 "movq 32(%0), %%mm2 \n\t" | |
139 "movq 96(%0), %%mm6 \n\t" | |
140 "movq %%mm2, %%mm4 \n\t" | |
141 "movq %%mm6, %%mm0 \n\t" | |
142 "psraw $1, %%mm4 \n\t" | |
143 "psraw $1, %%mm6 \n\t" | |
144 "psubw %%mm0, %%mm4 \n\t" | |
145 "paddw %%mm2, %%mm6 \n\t" | |
146 | |
147 "movq (%0), %%mm2 \n\t" | |
148 "movq 64(%0), %%mm0 \n\t" | |
149 SUMSUB_BA( %%mm0, %%mm2 ) | |
150 SUMSUB_BA( %%mm6, %%mm0 ) | |
151 SUMSUB_BA( %%mm4, %%mm2 ) | |
152 SUMSUB_BA( %%mm7, %%mm6 ) | |
153 SUMSUB_BA( %%mm5, %%mm4 ) | |
154 SUMSUB_BA( %%mm3, %%mm2 ) | |
155 SUMSUB_BA( %%mm1, %%mm0 ) | |
156 :: "r"(block) | |
157 ); | |
158 } | |
159 | |
160 static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride) | |
161 { | |
162 int i; | |
11369 | 163 DECLARE_ALIGNED(8, int16_t, b2)[64]; |
8430 | 164 |
165 block[0] += 32; | |
166 | |
167 for(i=0; i<2; i++){ | |
11369 | 168 DECLARE_ALIGNED(8, uint64_t, tmp); |
8430 | 169 |
170 h264_idct8_1d(block+4*i); | |
171 | |
172 __asm__ volatile( | |
173 "movq %%mm7, %0 \n\t" | |
174 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 ) | |
175 "movq %%mm0, 8(%1) \n\t" | |
176 "movq %%mm6, 24(%1) \n\t" | |
177 "movq %%mm7, 40(%1) \n\t" | |
178 "movq %%mm4, 56(%1) \n\t" | |
179 "movq %0, %%mm7 \n\t" | |
180 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 ) | |
181 "movq %%mm7, (%1) \n\t" | |
182 "movq %%mm1, 16(%1) \n\t" | |
183 "movq %%mm0, 32(%1) \n\t" | |
184 "movq %%mm3, 48(%1) \n\t" | |
185 : "=m"(tmp) | |
186 : "r"(b2+32*i) | |
187 : "memory" | |
188 ); | |
189 } | |
190 | |
191 for(i=0; i<2; i++){ | |
192 h264_idct8_1d(b2+4*i); | |
193 | |
194 __asm__ volatile( | |
195 "psraw $6, %%mm7 \n\t" | |
196 "psraw $6, %%mm6 \n\t" | |
197 "psraw $6, %%mm5 \n\t" | |
198 "psraw $6, %%mm4 \n\t" | |
199 "psraw $6, %%mm3 \n\t" | |
200 "psraw $6, %%mm2 \n\t" | |
201 "psraw $6, %%mm1 \n\t" | |
202 "psraw $6, %%mm0 \n\t" | |
203 | |
204 "movq %%mm7, (%0) \n\t" | |
205 "movq %%mm5, 16(%0) \n\t" | |
206 "movq %%mm3, 32(%0) \n\t" | |
207 "movq %%mm1, 48(%0) \n\t" | |
208 "movq %%mm0, 64(%0) \n\t" | |
209 "movq %%mm2, 80(%0) \n\t" | |
210 "movq %%mm4, 96(%0) \n\t" | |
211 "movq %%mm6, 112(%0) \n\t" | |
212 :: "r"(b2+4*i) | |
213 : "memory" | |
214 ); | |
215 } | |
216 | |
12435
fe78a4548d12
Put ff_ prefix on non-static {put_signed,put,add}_pixels_clamped_mmx()
rbultje
parents:
12414
diff
changeset
|
217 ff_add_pixels_clamped_mmx(b2, dst, stride); |
8430 | 218 } |
219 | |
220 #define STORE_DIFF_8P( p, d, t, z )\ | |
221 "movq "#d", "#t" \n"\ | |
222 "psraw $6, "#p" \n"\ | |
223 "punpcklbw "#z", "#t" \n"\ | |
224 "paddsw "#t", "#p" \n"\ | |
225 "packuswb "#p", "#p" \n"\ | |
226 "movq "#p", "#d" \n" | |
227 | |
228 #define H264_IDCT8_1D_SSE2(a,b,c,d,e,f,g,h)\ | |
229 "movdqa "#c", "#a" \n"\ | |
230 "movdqa "#g", "#e" \n"\ | |
231 "psraw $1, "#c" \n"\ | |
232 "psraw $1, "#g" \n"\ | |
233 "psubw "#e", "#c" \n"\ | |
234 "paddw "#a", "#g" \n"\ | |
235 "movdqa "#b", "#e" \n"\ | |
236 "psraw $1, "#e" \n"\ | |
237 "paddw "#b", "#e" \n"\ | |
238 "paddw "#d", "#e" \n"\ | |
239 "paddw "#f", "#e" \n"\ | |
240 "movdqa "#f", "#a" \n"\ | |
241 "psraw $1, "#a" \n"\ | |
242 "paddw "#f", "#a" \n"\ | |
243 "paddw "#h", "#a" \n"\ | |
244 "psubw "#b", "#a" \n"\ | |
245 "psubw "#d", "#b" \n"\ | |
246 "psubw "#d", "#f" \n"\ | |
247 "paddw "#h", "#b" \n"\ | |
248 "psubw "#h", "#f" \n"\ | |
249 "psraw $1, "#d" \n"\ | |
250 "psraw $1, "#h" \n"\ | |
251 "psubw "#d", "#b" \n"\ | |
252 "psubw "#h", "#f" \n"\ | |
253 "movdqa "#e", "#d" \n"\ | |
254 "movdqa "#a", "#h" \n"\ | |
255 "psraw $2, "#d" \n"\ | |
256 "psraw $2, "#h" \n"\ | |
257 "paddw "#f", "#d" \n"\ | |
258 "paddw "#b", "#h" \n"\ | |
259 "psraw $2, "#f" \n"\ | |
260 "psraw $2, "#b" \n"\ | |
261 "psubw "#f", "#e" \n"\ | |
262 "psubw "#a", "#b" \n"\ | |
263 "movdqa 0x00(%1), "#a" \n"\ | |
264 "movdqa 0x40(%1), "#f" \n"\ | |
265 SUMSUB_BA(f, a)\ | |
266 SUMSUB_BA(g, f)\ | |
267 SUMSUB_BA(c, a)\ | |
268 SUMSUB_BA(e, g)\ | |
269 SUMSUB_BA(b, c)\ | |
270 SUMSUB_BA(h, a)\ | |
271 SUMSUB_BA(d, f) | |
272 | |
273 static void ff_h264_idct8_add_sse2(uint8_t *dst, int16_t *block, int stride) | |
274 { | |
275 __asm__ volatile( | |
276 "movdqa 0x10(%1), %%xmm1 \n" | |
277 "movdqa 0x20(%1), %%xmm2 \n" | |
278 "movdqa 0x30(%1), %%xmm3 \n" | |
279 "movdqa 0x50(%1), %%xmm5 \n" | |
280 "movdqa 0x60(%1), %%xmm6 \n" | |
281 "movdqa 0x70(%1), %%xmm7 \n" | |
282 H264_IDCT8_1D_SSE2(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7) | |
283 TRANSPOSE8(%%xmm4, %%xmm1, %%xmm7, %%xmm3, %%xmm5, %%xmm0, %%xmm2, %%xmm6, (%1)) | |
284 "paddw %4, %%xmm4 \n" | |
285 "movdqa %%xmm4, 0x00(%1) \n" | |
286 "movdqa %%xmm2, 0x40(%1) \n" | |
287 H264_IDCT8_1D_SSE2(%%xmm4, %%xmm0, %%xmm6, %%xmm3, %%xmm2, %%xmm5, %%xmm7, %%xmm1) | |
288 "movdqa %%xmm6, 0x60(%1) \n" | |
289 "movdqa %%xmm7, 0x70(%1) \n" | |
290 "pxor %%xmm7, %%xmm7 \n" | |
291 STORE_DIFF_8P(%%xmm2, (%0), %%xmm6, %%xmm7) | |
292 STORE_DIFF_8P(%%xmm0, (%0,%2), %%xmm6, %%xmm7) | |
293 STORE_DIFF_8P(%%xmm1, (%0,%2,2), %%xmm6, %%xmm7) | |
294 STORE_DIFF_8P(%%xmm3, (%0,%3), %%xmm6, %%xmm7) | |
295 "lea (%0,%2,4), %0 \n" | |
296 STORE_DIFF_8P(%%xmm5, (%0), %%xmm6, %%xmm7) | |
297 STORE_DIFF_8P(%%xmm4, (%0,%2), %%xmm6, %%xmm7) | |
298 "movdqa 0x60(%1), %%xmm0 \n" | |
299 "movdqa 0x70(%1), %%xmm1 \n" | |
300 STORE_DIFF_8P(%%xmm0, (%0,%2,2), %%xmm6, %%xmm7) | |
301 STORE_DIFF_8P(%%xmm1, (%0,%3), %%xmm6, %%xmm7) | |
302 :"+r"(dst) | |
303 :"r"(block), "r"((x86_reg)stride), "r"((x86_reg)3L*stride), "m"(ff_pw_32) | |
304 ); | |
305 } | |
306 | |
307 static void ff_h264_idct_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride) | |
308 { | |
309 int dc = (block[0] + 32) >> 6; | |
310 __asm__ volatile( | |
311 "movd %0, %%mm0 \n\t" | |
312 "pshufw $0, %%mm0, %%mm0 \n\t" | |
313 "pxor %%mm1, %%mm1 \n\t" | |
314 "psubw %%mm0, %%mm1 \n\t" | |
315 "packuswb %%mm0, %%mm0 \n\t" | |
316 "packuswb %%mm1, %%mm1 \n\t" | |
317 ::"r"(dc) | |
318 ); | |
319 __asm__ volatile( | |
320 "movd %0, %%mm2 \n\t" | |
321 "movd %1, %%mm3 \n\t" | |
322 "movd %2, %%mm4 \n\t" | |
323 "movd %3, %%mm5 \n\t" | |
324 "paddusb %%mm0, %%mm2 \n\t" | |
325 "paddusb %%mm0, %%mm3 \n\t" | |
326 "paddusb %%mm0, %%mm4 \n\t" | |
327 "paddusb %%mm0, %%mm5 \n\t" | |
328 "psubusb %%mm1, %%mm2 \n\t" | |
329 "psubusb %%mm1, %%mm3 \n\t" | |
330 "psubusb %%mm1, %%mm4 \n\t" | |
331 "psubusb %%mm1, %%mm5 \n\t" | |
332 "movd %%mm2, %0 \n\t" | |
333 "movd %%mm3, %1 \n\t" | |
334 "movd %%mm4, %2 \n\t" | |
335 "movd %%mm5, %3 \n\t" | |
336 :"+m"(*(uint32_t*)(dst+0*stride)), | |
337 "+m"(*(uint32_t*)(dst+1*stride)), | |
338 "+m"(*(uint32_t*)(dst+2*stride)), | |
339 "+m"(*(uint32_t*)(dst+3*stride)) | |
340 ); | |
341 } | |
342 | |
343 static void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride) | |
344 { | |
345 int dc = (block[0] + 32) >> 6; | |
346 int y; | |
347 __asm__ volatile( | |
348 "movd %0, %%mm0 \n\t" | |
349 "pshufw $0, %%mm0, %%mm0 \n\t" | |
350 "pxor %%mm1, %%mm1 \n\t" | |
351 "psubw %%mm0, %%mm1 \n\t" | |
352 "packuswb %%mm0, %%mm0 \n\t" | |
353 "packuswb %%mm1, %%mm1 \n\t" | |
354 ::"r"(dc) | |
355 ); | |
356 for(y=2; y--; dst += 4*stride){ | |
357 __asm__ volatile( | |
358 "movq %0, %%mm2 \n\t" | |
359 "movq %1, %%mm3 \n\t" | |
360 "movq %2, %%mm4 \n\t" | |
361 "movq %3, %%mm5 \n\t" | |
362 "paddusb %%mm0, %%mm2 \n\t" | |
363 "paddusb %%mm0, %%mm3 \n\t" | |
364 "paddusb %%mm0, %%mm4 \n\t" | |
365 "paddusb %%mm0, %%mm5 \n\t" | |
366 "psubusb %%mm1, %%mm2 \n\t" | |
367 "psubusb %%mm1, %%mm3 \n\t" | |
368 "psubusb %%mm1, %%mm4 \n\t" | |
369 "psubusb %%mm1, %%mm5 \n\t" | |
370 "movq %%mm2, %0 \n\t" | |
371 "movq %%mm3, %1 \n\t" | |
372 "movq %%mm4, %2 \n\t" | |
373 "movq %%mm5, %3 \n\t" | |
374 :"+m"(*(uint64_t*)(dst+0*stride)), | |
375 "+m"(*(uint64_t*)(dst+1*stride)), | |
376 "+m"(*(uint64_t*)(dst+2*stride)), | |
377 "+m"(*(uint64_t*)(dst+3*stride)) | |
378 ); | |
379 } | |
380 } | |
381 | |
382 //FIXME this table is a duplicate from h264data.h, and will be removed once the tables from, h264 have been split | |
383 static const uint8_t scan8[16 + 2*4]={ | |
384 4+1*8, 5+1*8, 4+2*8, 5+2*8, | |
385 6+1*8, 7+1*8, 6+2*8, 7+2*8, | |
386 4+3*8, 5+3*8, 4+4*8, 5+4*8, | |
387 6+3*8, 7+3*8, 6+4*8, 7+4*8, | |
388 1+1*8, 2+1*8, | |
389 1+2*8, 2+2*8, | |
390 1+4*8, 2+4*8, | |
391 1+5*8, 2+5*8, | |
392 }; | |
393 | |
394 static void ff_h264_idct_add16_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
395 int i; | |
396 for(i=0; i<16; i++){ | |
397 if(nnzc[ scan8[i] ]) | |
398 ff_h264_idct_add_mmx(dst + block_offset[i], block + i*16, stride); | |
399 } | |
400 } | |
401 | |
402 static void ff_h264_idct8_add4_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
403 int i; | |
404 for(i=0; i<16; i+=4){ | |
405 if(nnzc[ scan8[i] ]) | |
406 ff_h264_idct8_add_mmx(dst + block_offset[i], block + i*16, stride); | |
407 } | |
408 } | |
409 | |
410 | |
411 static void ff_h264_idct_add16_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
412 int i; | |
413 for(i=0; i<16; i++){ | |
414 int nnz = nnzc[ scan8[i] ]; | |
415 if(nnz){ | |
416 if(nnz==1 && block[i*16]) ff_h264_idct_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); | |
417 else ff_h264_idct_add_mmx (dst + block_offset[i], block + i*16, stride); | |
418 } | |
419 } | |
420 } | |
421 | |
422 static void ff_h264_idct_add16intra_mmx(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
423 int i; | |
424 for(i=0; i<16; i++){ | |
425 if(nnzc[ scan8[i] ] || block[i*16]) | |
426 ff_h264_idct_add_mmx(dst + block_offset[i], block + i*16, stride); | |
427 } | |
428 } | |
429 | |
430 static void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
431 int i; | |
432 for(i=0; i<16; i++){ | |
433 if(nnzc[ scan8[i] ]) ff_h264_idct_add_mmx (dst + block_offset[i], block + i*16, stride); | |
434 else if(block[i*16]) ff_h264_idct_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); | |
435 } | |
436 } | |
437 | |
438 static void ff_h264_idct8_add4_mmx2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
439 int i; | |
440 for(i=0; i<16; i+=4){ | |
441 int nnz = nnzc[ scan8[i] ]; | |
442 if(nnz){ | |
443 if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); | |
444 else ff_h264_idct8_add_mmx (dst + block_offset[i], block + i*16, stride); | |
445 } | |
446 } | |
447 } | |
448 | |
449 static void ff_h264_idct8_add4_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
450 int i; | |
451 for(i=0; i<16; i+=4){ | |
452 int nnz = nnzc[ scan8[i] ]; | |
453 if(nnz){ | |
454 if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_mmx2(dst + block_offset[i], block + i*16, stride); | |
455 else ff_h264_idct8_add_sse2 (dst + block_offset[i], block + i*16, stride); | |
456 } | |
457 } | |
458 } | |
459 | |
460 static void ff_h264_idct_add8_mmx(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
461 int i; | |
462 for(i=16; i<16+8; i++){ | |
463 if(nnzc[ scan8[i] ] || block[i*16]) | |
464 ff_h264_idct_add_mmx (dest[(i&4)>>2] + block_offset[i], block + i*16, stride); | |
465 } | |
466 } | |
467 | |
468 static void ff_h264_idct_add8_mmx2(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ | |
469 int i; | |
470 for(i=16; i<16+8; i++){ | |
471 if(nnzc[ scan8[i] ]) | |
472 ff_h264_idct_add_mmx (dest[(i&4)>>2] + block_offset[i], block + i*16, stride); | |
473 else if(block[i*16]) | |
474 ff_h264_idct_dc_add_mmx2(dest[(i&4)>>2] + block_offset[i], block + i*16, stride); | |
475 } | |
476 } | |
477 | |
12480 | 478 #if HAVE_YASM |
8510
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
479 static void ff_h264_idct_dc_add8_mmx2(uint8_t *dst, int16_t *block, int stride) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
480 { |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
481 __asm__ volatile( |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
482 "movd %0, %%mm0 \n\t" // 0 0 X D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
483 "punpcklwd %1, %%mm0 \n\t" // x X d D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
484 "paddsw %2, %%mm0 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
485 "psraw $6, %%mm0 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
486 "punpcklwd %%mm0, %%mm0 \n\t" // d d D D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
487 "pxor %%mm1, %%mm1 \n\t" // 0 0 0 0 |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
488 "psubw %%mm0, %%mm1 \n\t" // -d-d-D-D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
489 "packuswb %%mm1, %%mm0 \n\t" // -d-d-D-D d d D D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
490 "pshufw $0xFA, %%mm0, %%mm1 \n\t" // -d-d-d-d-D-D-D-D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
491 "punpcklwd %%mm0, %%mm0 \n\t" // d d d d D D D D |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
492 ::"m"(block[ 0]), |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
493 "m"(block[16]), |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
494 "m"(ff_pw_32) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
495 ); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
496 __asm__ volatile( |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
497 "movq %0, %%mm2 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
498 "movq %1, %%mm3 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
499 "movq %2, %%mm4 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
500 "movq %3, %%mm5 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
501 "paddusb %%mm0, %%mm2 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
502 "paddusb %%mm0, %%mm3 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
503 "paddusb %%mm0, %%mm4 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
504 "paddusb %%mm0, %%mm5 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
505 "psubusb %%mm1, %%mm2 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
506 "psubusb %%mm1, %%mm3 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
507 "psubusb %%mm1, %%mm4 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
508 "psubusb %%mm1, %%mm5 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
509 "movq %%mm2, %0 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
510 "movq %%mm3, %1 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
511 "movq %%mm4, %2 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
512 "movq %%mm5, %3 \n\t" |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
513 :"+m"(*(uint64_t*)(dst+0*stride)), |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
514 "+m"(*(uint64_t*)(dst+1*stride)), |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
515 "+m"(*(uint64_t*)(dst+2*stride)), |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
516 "+m"(*(uint64_t*)(dst+3*stride)) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
517 ); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
518 } |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
519 |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
520 extern void ff_x264_add8x4_idct_sse2(uint8_t *dst, int16_t *block, int stride); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
521 |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
522 static void ff_h264_idct_add16_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
523 int i; |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
524 for(i=0; i<16; i+=2) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
525 if(nnzc[ scan8[i+0] ]|nnzc[ scan8[i+1] ]) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
526 ff_x264_add8x4_idct_sse2 (dst + block_offset[i], block + i*16, stride); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
527 } |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
528 |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
529 static void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
530 int i; |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
531 for(i=0; i<16; i+=2){ |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
532 if(nnzc[ scan8[i+0] ]|nnzc[ scan8[i+1] ]) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
533 ff_x264_add8x4_idct_sse2 (dst + block_offset[i], block + i*16, stride); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
534 else if(block[i*16]|block[i*16+16]) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
535 ff_h264_idct_dc_add8_mmx2(dst + block_offset[i], block + i*16, stride); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
536 } |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
537 } |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
538 |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
539 static void ff_h264_idct_add8_sse2(uint8_t **dest, const int *block_offset, DCTELEM *block, int stride, const uint8_t nnzc[6*8]){ |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
540 int i; |
8758
93980b03673e
fix typo in h264dsp_mmx (no effect currently as the function is not used), approved by Dark Shikari on IRC
bcoudurier
parents:
8590
diff
changeset
|
541 for(i=16; i<16+8; i+=2){ |
8510
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
542 if(nnzc[ scan8[i+0] ]|nnzc[ scan8[i+1] ]) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
543 ff_x264_add8x4_idct_sse2 (dest[(i&4)>>2] + block_offset[i], block + i*16, stride); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
544 else if(block[i*16]|block[i*16+16]) |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
545 ff_h264_idct_dc_add8_mmx2(dest[(i&4)>>2] + block_offset[i], block + i*16, stride); |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
546 } |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
547 } |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
548 #endif |
cea216e44ee3
Add x264 SSE2 iDCT functions to H.264 decoder.
darkshikari
parents:
8430
diff
changeset
|
549 |
8430 | 550 /***********************************/ |
551 /* deblocking */ | |
552 | |
553 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2], | |
554 int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) { | |
555 int dir; | |
556 __asm__ volatile( | |
11017 | 557 "movq %0, %%mm7 \n" |
558 "movq %1, %%mm6 \n" | |
559 ::"m"(ff_pb_1), "m"(ff_pb_3) | |
8430 | 560 ); |
561 if(field) | |
562 __asm__ volatile( | |
11017 | 563 "movq %0, %%mm6 \n" |
564 ::"m"(ff_pb_3_1) | |
8430 | 565 ); |
11017 | 566 __asm__ volatile( |
567 "movq %%mm6, %%mm5 \n" | |
568 "paddb %%mm5, %%mm5 \n" | |
569 :); | |
8430 | 570 |
571 // could do a special case for dir==0 && edges==1, but it only reduces the | |
572 // average filter time by 1.2% | |
573 for( dir=1; dir>=0; dir-- ) { | |
10953 | 574 const x86_reg d_idx = dir ? -8 : -1; |
8430 | 575 const int mask_mv = dir ? mask_mv1 : mask_mv0; |
11369 | 576 DECLARE_ALIGNED(8, const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL; |
11017 | 577 int b_idx, edge; |
8430 | 578 for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) { |
579 __asm__ volatile( | |
580 "pand %0, %%mm0 \n\t" | |
581 ::"m"(mask_dir) | |
582 ); | |
583 if(!(mask_mv & edge)) { | |
11017 | 584 if(bidir) { |
8430 | 585 __asm__ volatile( |
11017 | 586 "movd (%1,%0), %%mm2 \n" |
587 "punpckldq 40(%1,%0), %%mm2 \n" // { ref0[bn], ref1[bn] } | |
588 "pshufw $0x44, (%1), %%mm0 \n" // { ref0[b], ref0[b] } | |
589 "pshufw $0x44, 40(%1), %%mm1 \n" // { ref1[b], ref1[b] } | |
590 "pshufw $0x4E, %%mm2, %%mm3 \n" | |
591 "psubb %%mm2, %%mm0 \n" // { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] } | |
592 "psubb %%mm3, %%mm1 \n" // { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] } | |
593 "1: \n" | |
594 "por %%mm1, %%mm0 \n" | |
595 "movq (%2,%0,4), %%mm1 \n" | |
596 "movq 8(%2,%0,4), %%mm2 \n" | |
597 "movq %%mm1, %%mm3 \n" | |
598 "movq %%mm2, %%mm4 \n" | |
599 "psubw (%2), %%mm1 \n" | |
600 "psubw 8(%2), %%mm2 \n" | |
601 "psubw 160(%2), %%mm3 \n" | |
602 "psubw 168(%2), %%mm4 \n" | |
603 "packsswb %%mm2, %%mm1 \n" | |
604 "packsswb %%mm4, %%mm3 \n" | |
605 "paddb %%mm6, %%mm1 \n" | |
606 "paddb %%mm6, %%mm3 \n" | |
607 "psubusb %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit | |
608 "psubusb %%mm5, %%mm3 \n" | |
609 "packsswb %%mm3, %%mm1 \n" | |
610 "add $40, %0 \n" | |
611 "cmp $40, %0 \n" | |
612 "jl 1b \n" | |
613 "sub $80, %0 \n" | |
614 "pshufw $0x4E, %%mm1, %%mm1 \n" | |
615 "por %%mm1, %%mm0 \n" | |
616 "pshufw $0x4E, %%mm0, %%mm1 \n" | |
617 "pminub %%mm1, %%mm0 \n" | |
618 ::"r"(d_idx), | |
619 "r"(ref[0]+b_idx), | |
620 "r"(mv[0]+b_idx) | |
8430 | 621 ); |
11017 | 622 } else { |
10938
2a50b786b888
Fix h264_loop_filter_strength_mmx2() so it works with b frames.
michael
parents:
10937
diff
changeset
|
623 __asm__ volatile( |
11017 | 624 "movd (%1), %%mm0 \n" |
625 "psubb (%1,%0), %%mm0 \n" // ref[b] != ref[bn] | |
626 "movq (%2), %%mm1 \n" | |
627 "movq 8(%2), %%mm2 \n" | |
628 "psubw (%2,%0,4), %%mm1 \n" | |
629 "psubw 8(%2,%0,4), %%mm2 \n" | |
630 "packsswb %%mm2, %%mm1 \n" | |
631 "paddb %%mm6, %%mm1 \n" | |
632 "psubusb %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit | |
633 "packsswb %%mm1, %%mm1 \n" | |
634 "por %%mm1, %%mm0 \n" | |
635 ::"r"(d_idx), | |
636 "r"(ref[0]+b_idx), | |
637 "r"(mv[0]+b_idx) | |
10938
2a50b786b888
Fix h264_loop_filter_strength_mmx2() so it works with b frames.
michael
parents:
10937
diff
changeset
|
638 ); |
2a50b786b888
Fix h264_loop_filter_strength_mmx2() so it works with b frames.
michael
parents:
10937
diff
changeset
|
639 } |
8430 | 640 } |
641 __asm__ volatile( | |
11017 | 642 "movd %0, %%mm1 \n" |
643 "por %1, %%mm1 \n" // nnz[b] || nnz[bn] | |
8430 | 644 ::"m"(nnz[b_idx]), |
645 "m"(nnz[b_idx+d_idx]) | |
646 ); | |
647 __asm__ volatile( | |
11017 | 648 "pminub %%mm7, %%mm1 \n" |
649 "pminub %%mm7, %%mm0 \n" | |
650 "psllw $1, %%mm1 \n" | |
651 "pxor %%mm2, %%mm2 \n" | |
652 "pmaxub %%mm0, %%mm1 \n" | |
653 "punpcklbw %%mm2, %%mm1 \n" | |
654 "movq %%mm1, %0 \n" | |
8430 | 655 :"=m"(*bS[dir][edge]) |
656 ::"memory" | |
657 ); | |
658 } | |
659 edges = 4; | |
660 step = 1; | |
661 } | |
662 __asm__ volatile( | |
663 "movq (%0), %%mm0 \n\t" | |
664 "movq 8(%0), %%mm1 \n\t" | |
665 "movq 16(%0), %%mm2 \n\t" | |
666 "movq 24(%0), %%mm3 \n\t" | |
667 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4) | |
668 "movq %%mm0, (%0) \n\t" | |
669 "movq %%mm3, 8(%0) \n\t" | |
670 "movq %%mm4, 16(%0) \n\t" | |
671 "movq %%mm2, 24(%0) \n\t" | |
672 ::"r"(bS[0]) | |
673 :"memory" | |
674 ); | |
675 } | |
676 | |
12454
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
677 #define LF_FUNC(DIR, TYPE, OPT) \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
678 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
679 int alpha, int beta, int8_t *tc0); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
680 #define LF_IFUNC(DIR, TYPE, OPT) \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
681 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \ |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
682 int alpha, int beta); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
683 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
684 LF_FUNC (h, chroma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
685 LF_IFUNC(h, chroma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
686 LF_FUNC (v, chroma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
687 LF_IFUNC(v, chroma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
688 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
689 LF_FUNC (h, luma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
690 LF_IFUNC(h, luma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
691 #if HAVE_YASM && ARCH_X86_32 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
692 LF_FUNC (v8, luma, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
693 static void ff_x264_deblock_v_luma_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
694 { |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
695 if((tc0[0] & tc0[1]) >= 0) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
696 ff_x264_deblock_v8_luma_mmxext(pix+0, stride, alpha, beta, tc0); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
697 if((tc0[2] & tc0[3]) >= 0) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
698 ff_x264_deblock_v8_luma_mmxext(pix+8, stride, alpha, beta, tc0+2); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
699 } |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
700 LF_IFUNC(v8, luma_intra, mmxext) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
701 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
702 { |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
703 ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
704 ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta); |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
705 } |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
706 #endif |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
707 |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
708 LF_FUNC (h, luma, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
709 LF_IFUNC(h, luma_intra, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
710 LF_FUNC (v, luma, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
711 LF_IFUNC(v, luma_intra, sse2) |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
712 |
8430 | 713 /***********************************/ |
714 /* weighted prediction */ | |
715 | |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
716 #define H264_WEIGHT(W, H, OPT) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
717 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
718 int stride, int log2_denom, int weight, int offset); |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
719 |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
720 #define H264_BIWEIGHT(W, H, OPT) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
721 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
722 uint8_t *src, int stride, int log2_denom, int weightd, \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
723 int weights, int offset); |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
724 |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
725 #define H264_BIWEIGHT_MMX(W,H) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
726 H264_WEIGHT (W, H, mmx2) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
727 H264_BIWEIGHT(W, H, mmx2) |
8430 | 728 |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
729 #define H264_BIWEIGHT_MMX_SSE(W,H) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
730 H264_BIWEIGHT_MMX(W, H) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
731 H264_WEIGHT (W, H, sse2) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
732 H264_BIWEIGHT (W, H, sse2) \ |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
733 H264_BIWEIGHT (W, H, ssse3) |
8430 | 734 |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
735 H264_BIWEIGHT_MMX_SSE(16, 16) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
736 H264_BIWEIGHT_MMX_SSE(16, 8) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
737 H264_BIWEIGHT_MMX_SSE( 8, 16) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
738 H264_BIWEIGHT_MMX_SSE( 8, 8) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
739 H264_BIWEIGHT_MMX_SSE( 8, 4) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
740 H264_BIWEIGHT_MMX ( 4, 8) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
741 H264_BIWEIGHT_MMX ( 4, 4) |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
742 H264_BIWEIGHT_MMX ( 4, 2) |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
743 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
744 void ff_h264dsp_init_x86(H264DSPContext *c) |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
745 { |
12475
9fef0a8ddd63
Move mm_support() from libavcodec to libavutil, make it a public
stefano
parents:
12456
diff
changeset
|
746 int mm_flags = av_get_cpu_flags(); |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
747 |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
748 if (mm_flags & AV_CPU_FLAG_MMX) { |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
749 c->h264_idct_dc_add= |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
750 c->h264_idct_add= ff_h264_idct_add_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
751 c->h264_idct8_dc_add= |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
752 c->h264_idct8_add= ff_h264_idct8_add_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
753 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
754 c->h264_idct_add16 = ff_h264_idct_add16_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
755 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
756 c->h264_idct_add8 = ff_h264_idct_add8_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
757 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
758 |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
759 if (mm_flags & AV_CPU_FLAG_MMX2) { |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
760 c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
761 c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
762 c->h264_idct_add16 = ff_h264_idct_add16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
763 c->h264_idct8_add4 = ff_h264_idct8_add4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
764 c->h264_idct_add8 = ff_h264_idct_add8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
765 c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
766 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
767 c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
768 } |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
769 if(mm_flags & AV_CPU_FLAG_SSE2){ |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
770 c->h264_idct8_add = ff_h264_idct8_add_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
771 c->h264_idct8_add4= ff_h264_idct8_add4_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
772 } |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
773 |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
774 #if HAVE_YASM |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
775 if (mm_flags & AV_CPU_FLAG_MMX2){ |
12454
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
776 c->h264_v_loop_filter_chroma= ff_x264_deblock_v_chroma_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
777 c->h264_h_loop_filter_chroma= ff_x264_deblock_h_chroma_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
778 c->h264_v_loop_filter_chroma_intra= ff_x264_deblock_v_chroma_intra_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
779 c->h264_h_loop_filter_chroma_intra= ff_x264_deblock_h_chroma_intra_mmxext; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
780 #if ARCH_X86_32 |
12454
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
781 c->h264_v_loop_filter_luma= ff_x264_deblock_v_luma_mmxext; |
f4355cd85faa
Port latest x264 deblock asm (before they moved to using NV12 as internal
rbultje
parents:
12451
diff
changeset
|
782 c->h264_h_loop_filter_luma= ff_x264_deblock_h_luma_mmxext; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
783 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
784 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
785 #endif |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
786 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
787 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
788 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
789 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
790 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
791 c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
792 c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
793 c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
794 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
795 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
796 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
797 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
798 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
799 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
800 c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
801 c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
802 c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
803 |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
804 if (mm_flags&AV_CPU_FLAG_SSE2) { |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
805 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
806 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
807 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
808 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
809 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
810 |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
811 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
812 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
813 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2; |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
814 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
815 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
816 |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
817 #if ARCH_X86_64 || !defined(__ICC) || __ICC > 1110 |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
818 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
819 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
820 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
821 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
822 #endif |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
823 c->h264_idct_add16 = ff_h264_idct_add16_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
824 c->h264_idct_add8 = ff_h264_idct_add8_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
825 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2; |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
826 } |
12456
a5ddb39627fd
Rename FF_MM_ symbols related to CPU features flags as AV_CPU_FLAG_
stefano
parents:
12454
diff
changeset
|
827 if (mm_flags&AV_CPU_FLAG_SSSE3) { |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
828 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
829 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3; |
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
830 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3; |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
831 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3; |
12451
4c3e6ff1237e
Rename h264_weight_sse2.asm to h264_weight.asm; add 16x8/8x16/8x4 non-square
rbultje
parents:
12450
diff
changeset
|
832 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3; |
12450
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
833 } |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
834 } |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
835 #endif |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
836 } |
3941687b4fa9
Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
12438
diff
changeset
|
837 } |