7466b18be33e0af3cc205b19cd0540ecb8d87f9d
[libav.git] / libpostproc / postprocess.c
1 /*
2 * Copyright (C) 2001-2003 Michael Niedermayer (michaelni@gmx.at)
3 *
4 * AltiVec optimizations (C) 2004 Romain Dolbeau <romain@dolbeau.org>
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /**
24 * @file
25 * postprocessing.
26 */
27
28 /*
29 C MMX MMX2 3DNow AltiVec
30 isVertDC Ec Ec Ec
31 isVertMinMaxOk Ec Ec Ec
32 doVertLowPass E e e Ec
33 doVertDefFilter Ec Ec e e Ec
34 isHorizDC Ec Ec Ec
35 isHorizMinMaxOk a E Ec
36 doHorizLowPass E e e Ec
37 doHorizDefFilter Ec Ec e e Ec
38 do_a_deblock Ec E Ec E
39 deRing E e e* Ecp
40 Vertical RKAlgo1 E a a
41 Horizontal RKAlgo1 a a
42 Vertical X1# a E E
43 Horizontal X1# a E E
44 LinIpolDeinterlace e E E*
45 CubicIpolDeinterlace a e e*
46 LinBlendDeinterlace e E E*
47 MedianDeinterlace# E Ec Ec
48 TempDeNoiser# E e e Ec
49
50 * I do not have a 3DNow! CPU -> it is untested, but no one said it does not work so it seems to work
51 # more or less selfinvented filters so the exactness is not too meaningful
52 E = Exact implementation
53 e = almost exact implementation (slightly different rounding,...)
54 a = alternative / approximate impl
55 c = checked against the other implementations (-vo md5)
56 p = partially optimized, still some work to do
57 */
58
59 /*
60 TODO:
61 reduce the time wasted on the mem transfer
62 unroll stuff if instructions depend too much on the prior one
63 move YScale thing to the end instead of fixing QP
64 write a faster and higher quality deblocking filter :)
65 make the mainloop more flexible (variable number of blocks at once
66 (the if/else stuff per block is slowing things down)
67 compare the quality & speed of all filters
68 split this huge file
69 optimize c versions
70 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
71 ...
72 */
73
74 //Changelog: use the Subversion log
75
76 #include "config.h"
77 #include "libavutil/avutil.h"
78 #include <inttypes.h>
79 #include <stdio.h>
80 #include <stdlib.h>
81 #include <string.h>
82 //#undef HAVE_MMX2
83 //#define HAVE_AMD3DNOW
84 //#undef HAVE_MMX
85 //#undef ARCH_X86
86 //#define DEBUG_BRIGHTNESS
87 #include "postprocess.h"
88 #include "postprocess_internal.h"
89
90 unsigned postproc_version(void)
91 {
92 return LIBPOSTPROC_VERSION_INT;
93 }
94
95 const char *postproc_configuration(void)
96 {
97 return LIBAV_CONFIGURATION;
98 }
99
100 const char *postproc_license(void)
101 {
102 #define LICENSE_PREFIX "libpostproc license: "
103 return LICENSE_PREFIX FFMPEG_LICENSE + sizeof(LICENSE_PREFIX) - 1;
104 }
105
106 #if HAVE_ALTIVEC_H
107 #include <altivec.h>
108 #endif
109
110 #define GET_MODE_BUFFER_SIZE 500
111 #define OPTIONS_ARRAY_SIZE 10
112 #define BLOCK_SIZE 8
113 #define TEMP_STRIDE 8
114 //#define NUM_BLOCKS_AT_ONCE 16 //not used yet
115
116 #if ARCH_X86
117 DECLARE_ASM_CONST(8, uint64_t, w05)= 0x0005000500050005LL;
118 DECLARE_ASM_CONST(8, uint64_t, w04)= 0x0004000400040004LL;
119 DECLARE_ASM_CONST(8, uint64_t, w20)= 0x0020002000200020LL;
120 DECLARE_ASM_CONST(8, uint64_t, b00)= 0x0000000000000000LL;
121 DECLARE_ASM_CONST(8, uint64_t, b01)= 0x0101010101010101LL;
122 DECLARE_ASM_CONST(8, uint64_t, b02)= 0x0202020202020202LL;
123 DECLARE_ASM_CONST(8, uint64_t, b08)= 0x0808080808080808LL;
124 DECLARE_ASM_CONST(8, uint64_t, b80)= 0x8080808080808080LL;
125 #endif
126
127 DECLARE_ASM_CONST(8, int, deringThreshold)= 20;
128
129
130 static struct PPFilter filters[]=
131 {
132 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK},
133 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK},
134 /* {"hr", "rkhdeblock", 1, 1, 3, H_RK1_FILTER},
135 {"vr", "rkvdeblock", 1, 2, 4, V_RK1_FILTER},*/
136 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER},
137 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER},
138 {"ha", "ahdeblock", 1, 1, 3, H_A_DEBLOCK},
139 {"va", "avdeblock", 1, 2, 4, V_A_DEBLOCK},
140 {"dr", "dering", 1, 5, 6, DERING},
141 {"al", "autolevels", 0, 1, 2, LEVEL_FIX},
142 {"lb", "linblenddeint", 1, 1, 4, LINEAR_BLEND_DEINT_FILTER},
143 {"li", "linipoldeint", 1, 1, 4, LINEAR_IPOL_DEINT_FILTER},
144 {"ci", "cubicipoldeint", 1, 1, 4, CUBIC_IPOL_DEINT_FILTER},
145 {"md", "mediandeint", 1, 1, 4, MEDIAN_DEINT_FILTER},
146 {"fd", "ffmpegdeint", 1, 1, 4, FFMPEG_DEINT_FILTER},
147 {"l5", "lowpass5", 1, 1, 4, LOWPASS5_DEINT_FILTER},
148 {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER},
149 {"fq", "forcequant", 1, 0, 0, FORCE_QUANT},
150 {NULL, NULL,0,0,0,0} //End Marker
151 };
152
153 static const char *replaceTable[]=
154 {
155 "default", "hb:a,vb:a,dr:a",
156 "de", "hb:a,vb:a,dr:a",
157 "fast", "h1:a,v1:a,dr:a",
158 "fa", "h1:a,v1:a,dr:a",
159 "ac", "ha:a:128:7,va:a,dr:a",
160 NULL //End Marker
161 };
162
163
164 #if ARCH_X86
165 static inline void prefetchnta(void *p)
166 {
167 __asm__ volatile( "prefetchnta (%0)\n\t"
168 : : "r" (p)
169 );
170 }
171
172 static inline void prefetcht0(void *p)
173 {
174 __asm__ volatile( "prefetcht0 (%0)\n\t"
175 : : "r" (p)
176 );
177 }
178
179 static inline void prefetcht1(void *p)
180 {
181 __asm__ volatile( "prefetcht1 (%0)\n\t"
182 : : "r" (p)
183 );
184 }
185
186 static inline void prefetcht2(void *p)
187 {
188 __asm__ volatile( "prefetcht2 (%0)\n\t"
189 : : "r" (p)
190 );
191 }
192 #endif
193
194 /* The horizontal functions exist only in C because the MMX
195 * code is faster with vertical filters and transposing. */
196
197 /**
198 * Check if the given 8x8 Block is mostly "flat"
199 */
200 static inline int isHorizDC_C(uint8_t src[], int stride, PPContext *c)
201 {
202 int numEq= 0;
203 int y;
204 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
205 const int dcThreshold= dcOffset*2 + 1;
206
207 for(y=0; y<BLOCK_SIZE; y++){
208 if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
209 if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++;
210 if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++;
211 if(((unsigned)(src[3] - src[4] + dcOffset)) < dcThreshold) numEq++;
212 if(((unsigned)(src[4] - src[5] + dcOffset)) < dcThreshold) numEq++;
213 if(((unsigned)(src[5] - src[6] + dcOffset)) < dcThreshold) numEq++;
214 if(((unsigned)(src[6] - src[7] + dcOffset)) < dcThreshold) numEq++;
215 src+= stride;
216 }
217 return numEq > c->ppMode.flatnessThreshold;
218 }
219
220 /**
221 * Check if the middle 8x8 Block in the given 8x16 block is flat
222 */
223 static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c)
224 {
225 int numEq= 0;
226 int y;
227 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
228 const int dcThreshold= dcOffset*2 + 1;
229
230 src+= stride*4; // src points to begin of the 8x8 Block
231 for(y=0; y<BLOCK_SIZE-1; y++){
232 if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++;
233 if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++;
234 if(((unsigned)(src[2] - src[2+stride] + dcOffset)) < dcThreshold) numEq++;
235 if(((unsigned)(src[3] - src[3+stride] + dcOffset)) < dcThreshold) numEq++;
236 if(((unsigned)(src[4] - src[4+stride] + dcOffset)) < dcThreshold) numEq++;
237 if(((unsigned)(src[5] - src[5+stride] + dcOffset)) < dcThreshold) numEq++;
238 if(((unsigned)(src[6] - src[6+stride] + dcOffset)) < dcThreshold) numEq++;
239 if(((unsigned)(src[7] - src[7+stride] + dcOffset)) < dcThreshold) numEq++;
240 src+= stride;
241 }
242 return numEq > c->ppMode.flatnessThreshold;
243 }
244
245 static inline int isHorizMinMaxOk_C(uint8_t src[], int stride, int QP)
246 {
247 int i;
248 #if 1
249 for(i=0; i<2; i++){
250 if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0;
251 src += stride;
252 if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0;
253 src += stride;
254 if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0;
255 src += stride;
256 if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0;
257 src += stride;
258 }
259 #else
260 for(i=0; i<8; i++){
261 if((unsigned)(src[0] - src[7] + 2*QP) > 4*QP) return 0;
262 src += stride;
263 }
264 #endif
265 return 1;
266 }
267
268 static inline int isVertMinMaxOk_C(uint8_t src[], int stride, int QP)
269 {
270 #if 1
271 #if 1
272 int x;
273 src+= stride*4;
274 for(x=0; x<BLOCK_SIZE; x+=4){
275 if((unsigned)(src[ x + 0*stride] - src[ x + 5*stride] + 2*QP) > 4*QP) return 0;
276 if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0;
277 if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0;
278 if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0;
279 }
280 #else
281 int x;
282 src+= stride*3;
283 for(x=0; x<BLOCK_SIZE; x++){
284 if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
285 }
286 #endif
287 return 1;
288 #else
289 int x;
290 src+= stride*4;
291 for(x=0; x<BLOCK_SIZE; x++){
292 int min=255;
293 int max=0;
294 int y;
295 for(y=0; y<8; y++){
296 int v= src[x + y*stride];
297 if(v>max) max=v;
298 if(v<min) min=v;
299 }
300 if(max-min > 2*QP) return 0;
301 }
302 return 1;
303 #endif
304 }
305
306 static inline int horizClassify_C(uint8_t src[], int stride, PPContext *c)
307 {
308 if( isHorizDC_C(src, stride, c) ){
309 if( isHorizMinMaxOk_C(src, stride, c->QP) )
310 return 1;
311 else
312 return 0;
313 }else{
314 return 2;
315 }
316 }
317
318 static inline int vertClassify_C(uint8_t src[], int stride, PPContext *c)
319 {
320 if( isVertDC_C(src, stride, c) ){
321 if( isVertMinMaxOk_C(src, stride, c->QP) )
322 return 1;
323 else
324 return 0;
325 }else{
326 return 2;
327 }
328 }
329
330 static inline void doHorizDefFilter_C(uint8_t dst[], int stride, PPContext *c)
331 {
332 int y;
333 for(y=0; y<BLOCK_SIZE; y++){
334 const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]);
335
336 if(FFABS(middleEnergy) < 8*c->QP){
337 const int q=(dst[3] - dst[4])/2;
338 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
339 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
340
341 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
342 d= FFMAX(d, 0);
343
344 d= (5*d + 32) >> 6;
345 d*= FFSIGN(-middleEnergy);
346
347 if(q>0)
348 {
349 d= d<0 ? 0 : d;
350 d= d>q ? q : d;
351 }
352 else
353 {
354 d= d>0 ? 0 : d;
355 d= d<q ? q : d;
356 }
357
358 dst[3]-= d;
359 dst[4]+= d;
360 }
361 dst+= stride;
362 }
363 }
364
365 /**
366 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
367 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
368 */
369 static inline void doHorizLowPass_C(uint8_t dst[], int stride, PPContext *c)
370 {
371 int y;
372 for(y=0; y<BLOCK_SIZE; y++){
373 const int first= FFABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0];
374 const int last= FFABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7];
375
376 int sums[10];
377 sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4;
378 sums[1] = sums[0] - first + dst[3];
379 sums[2] = sums[1] - first + dst[4];
380 sums[3] = sums[2] - first + dst[5];
381 sums[4] = sums[3] - first + dst[6];
382 sums[5] = sums[4] - dst[0] + dst[7];
383 sums[6] = sums[5] - dst[1] + last;
384 sums[7] = sums[6] - dst[2] + last;
385 sums[8] = sums[7] - dst[3] + last;
386 sums[9] = sums[8] - dst[4] + last;
387
388 dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4;
389 dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4;
390 dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4;
391 dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4;
392 dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4;
393 dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4;
394 dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4;
395 dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4;
396
397 dst+= stride;
398 }
399 }
400
401 /**
402 * Experimental Filter 1 (Horizontal)
403 * will not damage linear gradients
404 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
405 * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
406 * MMX2 version does correct clipping C version does not
407 * not identical with the vertical one
408 */
409 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
410 {
411 int y;
412 static uint64_t *lut= NULL;
413 if(lut==NULL)
414 {
415 int i;
416 lut = av_malloc(256*8);
417 for(i=0; i<256; i++)
418 {
419 int v= i < 128 ? 2*i : 2*(i-256);
420 /*
421 //Simulate 112242211 9-Tap filter
422 uint64_t a= (v/16) & 0xFF;
423 uint64_t b= (v/8) & 0xFF;
424 uint64_t c= (v/4) & 0xFF;
425 uint64_t d= (3*v/8) & 0xFF;
426 */
427 //Simulate piecewise linear interpolation
428 uint64_t a= (v/16) & 0xFF;
429 uint64_t b= (v*3/16) & 0xFF;
430 uint64_t c= (v*5/16) & 0xFF;
431 uint64_t d= (7*v/16) & 0xFF;
432 uint64_t A= (0x100 - a)&0xFF;
433 uint64_t B= (0x100 - b)&0xFF;
434 uint64_t C= (0x100 - c)&0xFF;
435 uint64_t D= (0x100 - c)&0xFF;
436
437 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
438 (D<<24) | (C<<16) | (B<<8) | (A);
439 //lut[i] = (v<<32) | (v<<24);
440 }
441 }
442
443 for(y=0; y<BLOCK_SIZE; y++){
444 int a= src[1] - src[2];
445 int b= src[3] - src[4];
446 int c= src[5] - src[6];
447
448 int d= FFMAX(FFABS(b) - (FFABS(a) + FFABS(c))/2, 0);
449
450 if(d < QP){
451 int v = d * FFSIGN(-b);
452
453 src[1] +=v/8;
454 src[2] +=v/4;
455 src[3] +=3*v/8;
456 src[4] -=3*v/8;
457 src[5] -=v/4;
458 src[6] -=v/8;
459 }
460 src+=stride;
461 }
462 }
463
464 /**
465 * accurate deblock filter
466 */
467 static av_always_inline void do_a_deblock_C(uint8_t *src, int step, int stride, PPContext *c){
468 int y;
469 const int QP= c->QP;
470 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
471 const int dcThreshold= dcOffset*2 + 1;
472 //START_TIMER
473 src+= step*4; // src points to begin of the 8x8 Block
474 for(y=0; y<8; y++){
475 int numEq= 0;
476
477 if(((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold) numEq++;
478 if(((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold) numEq++;
479 if(((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold) numEq++;
480 if(((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold) numEq++;
481 if(((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold) numEq++;
482 if(((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold) numEq++;
483 if(((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold) numEq++;
484 if(((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold) numEq++;
485 if(((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold) numEq++;
486 if(numEq > c->ppMode.flatnessThreshold){
487 int min, max, x;
488
489 if(src[0] > src[step]){
490 max= src[0];
491 min= src[step];
492 }else{
493 max= src[step];
494 min= src[0];
495 }
496 for(x=2; x<8; x+=2){
497 if(src[x*step] > src[(x+1)*step]){
498 if(src[x *step] > max) max= src[ x *step];
499 if(src[(x+1)*step] < min) min= src[(x+1)*step];
500 }else{
501 if(src[(x+1)*step] > max) max= src[(x+1)*step];
502 if(src[ x *step] < min) min= src[ x *step];
503 }
504 }
505 if(max-min < 2*QP){
506 const int first= FFABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0];
507 const int last= FFABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step];
508
509 int sums[10];
510 sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4;
511 sums[1] = sums[0] - first + src[3*step];
512 sums[2] = sums[1] - first + src[4*step];
513 sums[3] = sums[2] - first + src[5*step];
514 sums[4] = sums[3] - first + src[6*step];
515 sums[5] = sums[4] - src[0*step] + src[7*step];
516 sums[6] = sums[5] - src[1*step] + last;
517 sums[7] = sums[6] - src[2*step] + last;
518 sums[8] = sums[7] - src[3*step] + last;
519 sums[9] = sums[8] - src[4*step] + last;
520
521 src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4;
522 src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4;
523 src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4;
524 src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4;
525 src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4;
526 src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4;
527 src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4;
528 src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4;
529 }
530 }else{
531 const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]);
532
533 if(FFABS(middleEnergy) < 8*QP){
534 const int q=(src[3*step] - src[4*step])/2;
535 const int leftEnergy= 5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]);
536 const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]);
537
538 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
539 d= FFMAX(d, 0);
540
541 d= (5*d + 32) >> 6;
542 d*= FFSIGN(-middleEnergy);
543
544 if(q>0){
545 d= d<0 ? 0 : d;
546 d= d>q ? q : d;
547 }else{
548 d= d>0 ? 0 : d;
549 d= d<q ? q : d;
550 }
551
552 src[3*step]-= d;
553 src[4*step]+= d;
554 }
555 }
556
557 src += stride;
558 }
559 /*if(step==16){
560 STOP_TIMER("step16")
561 }else{
562 STOP_TIMER("stepX")
563 }*/
564 }
565
566 //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
567 //Plain C versions
568 #if !(HAVE_MMX || HAVE_ALTIVEC) || CONFIG_RUNTIME_CPUDETECT
569 #define COMPILE_C
570 #endif
571
572 #if HAVE_ALTIVEC
573 #define COMPILE_ALTIVEC
574 #endif //HAVE_ALTIVEC
575
576 #if ARCH_X86
577
578 #if (HAVE_MMX && !HAVE_AMD3DNOW && !HAVE_MMX2) || CONFIG_RUNTIME_CPUDETECT
579 #define COMPILE_MMX
580 #endif
581
582 #if HAVE_MMX2 || CONFIG_RUNTIME_CPUDETECT
583 #define COMPILE_MMX2
584 #endif
585
586 #if (HAVE_AMD3DNOW && !HAVE_MMX2) || CONFIG_RUNTIME_CPUDETECT
587 #define COMPILE_3DNOW
588 #endif
589 #endif /* ARCH_X86 */
590
591 #undef HAVE_MMX
592 #define HAVE_MMX 0
593 #undef HAVE_MMX2
594 #define HAVE_MMX2 0
595 #undef HAVE_AMD3DNOW
596 #define HAVE_AMD3DNOW 0
597 #undef HAVE_ALTIVEC
598 #define HAVE_ALTIVEC 0
599
600 #ifdef COMPILE_C
601 #define RENAME(a) a ## _C
602 #include "postprocess_template.c"
603 #endif
604
605 #ifdef COMPILE_ALTIVEC
606 #undef RENAME
607 #undef HAVE_ALTIVEC
608 #define HAVE_ALTIVEC 1
609 #define RENAME(a) a ## _altivec
610 #include "postprocess_altivec_template.c"
611 #include "postprocess_template.c"
612 #endif
613
614 //MMX versions
615 #ifdef COMPILE_MMX
616 #undef RENAME
617 #undef HAVE_MMX
618 #define HAVE_MMX 1
619 #define RENAME(a) a ## _MMX
620 #include "postprocess_template.c"
621 #endif
622
623 //MMX2 versions
624 #ifdef COMPILE_MMX2
625 #undef RENAME
626 #undef HAVE_MMX
627 #undef HAVE_MMX2
628 #define HAVE_MMX 1
629 #define HAVE_MMX2 1
630 #define RENAME(a) a ## _MMX2
631 #include "postprocess_template.c"
632 #endif
633
634 //3DNOW versions
635 #ifdef COMPILE_3DNOW
636 #undef RENAME
637 #undef HAVE_MMX
638 #undef HAVE_MMX2
639 #undef HAVE_AMD3DNOW
640 #define HAVE_MMX 1
641 #define HAVE_MMX2 0
642 #define HAVE_AMD3DNOW 1
643 #define RENAME(a) a ## _3DNow
644 #include "postprocess_template.c"
645 #endif
646
647 // minor note: the HAVE_xyz is messed up after that line so do not use it.
648
649 static inline void postProcess(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
650 const QP_STORE_T QPs[], int QPStride, int isColor, pp_mode *vm, pp_context *vc)
651 {
652 PPContext *c= (PPContext *)vc;
653 PPMode *ppMode= (PPMode *)vm;
654 c->ppMode= *ppMode; //FIXME
655
656 // Using ifs here as they are faster than function pointers although the
657 // difference would not be measurable here but it is much better because
658 // someone might exchange the CPU whithout restarting MPlayer ;)
659 #if CONFIG_RUNTIME_CPUDETECT
660 #if ARCH_X86
661 // ordered per speed fastest first
662 if(c->cpuCaps & PP_CPU_CAPS_MMX2)
663 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
664 else if(c->cpuCaps & PP_CPU_CAPS_3DNOW)
665 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
666 else if(c->cpuCaps & PP_CPU_CAPS_MMX)
667 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
668 else
669 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
670 #else
671 #if HAVE_ALTIVEC
672 if(c->cpuCaps & PP_CPU_CAPS_ALTIVEC)
673 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
674 else
675 #endif
676 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
677 #endif
678 #else //CONFIG_RUNTIME_CPUDETECT
679 #if HAVE_MMX2
680 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
681 #elif HAVE_AMD3DNOW
682 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
683 #elif HAVE_MMX
684 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
685 #elif HAVE_ALTIVEC
686 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
687 #else
688 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
689 #endif
690 #endif //!CONFIG_RUNTIME_CPUDETECT
691 }
692
693 //static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
694 // QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
695
696 /* -pp Command line Help
697 */
698 #if LIBPOSTPROC_VERSION_INT < (52<<16)
699 const char *const pp_help=
700 #else
701 const char pp_help[] =
702 #endif
703 "Available postprocessing filters:\n"
704 "Filters Options\n"
705 "short long name short long option Description\n"
706 "* * a autoq CPU power dependent enabler\n"
707 " c chrom chrominance filtering enabled\n"
708 " y nochrom chrominance filtering disabled\n"
709 " n noluma luma filtering disabled\n"
710 "hb hdeblock (2 threshold) horizontal deblocking filter\n"
711 " 1. difference factor: default=32, higher -> more deblocking\n"
712 " 2. flatness threshold: default=39, lower -> more deblocking\n"
713 " the h & v deblocking filters share these\n"
714 " so you can't set different thresholds for h / v\n"
715 "vb vdeblock (2 threshold) vertical deblocking filter\n"
716 "ha hadeblock (2 threshold) horizontal deblocking filter\n"
717 "va vadeblock (2 threshold) vertical deblocking filter\n"
718 "h1 x1hdeblock experimental h deblock filter 1\n"
719 "v1 x1vdeblock experimental v deblock filter 1\n"
720 "dr dering deringing filter\n"
721 "al autolevels automatic brightness / contrast\n"
722 " f fullyrange stretch luminance to (0..255)\n"
723 "lb linblenddeint linear blend deinterlacer\n"
724 "li linipoldeint linear interpolating deinterlace\n"
725 "ci cubicipoldeint cubic interpolating deinterlacer\n"
726 "md mediandeint median deinterlacer\n"
727 "fd ffmpegdeint ffmpeg deinterlacer\n"
728 "l5 lowpass5 FIR lowpass deinterlacer\n"
729 "de default hb:a,vb:a,dr:a\n"
730 "fa fast h1:a,v1:a,dr:a\n"
731 "ac ha:a:128:7,va:a,dr:a\n"
732 "tn tmpnoise (3 threshold) temporal noise reducer\n"
733 " 1. <= 2. <= 3. larger -> stronger filtering\n"
734 "fq forceQuant <quantizer> force quantizer\n"
735 "Usage:\n"
736 "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n"
737 "long form example:\n"
738 "vdeblock:autoq/hdeblock:autoq/linblenddeint default,-vdeblock\n"
739 "short form example:\n"
740 "vb:a/hb:a/lb de,-vb\n"
741 "more examples:\n"
742 "tn:64:128:256\n"
743 "\n"
744 ;
745
746 pp_mode *pp_get_mode_by_name_and_quality(const char *name, int quality)
747 {
748 char temp[GET_MODE_BUFFER_SIZE];
749 char *p= temp;
750 static const char filterDelimiters[] = ",/";
751 static const char optionDelimiters[] = ":";
752 struct PPMode *ppMode;
753 char *filterToken;
754
755 ppMode= av_malloc(sizeof(PPMode));
756
757 ppMode->lumMode= 0;
758 ppMode->chromMode= 0;
759 ppMode->maxTmpNoise[0]= 700;
760 ppMode->maxTmpNoise[1]= 1500;
761 ppMode->maxTmpNoise[2]= 3000;
762 ppMode->maxAllowedY= 234;
763 ppMode->minAllowedY= 16;
764 ppMode->baseDcDiff= 256/8;
765 ppMode->flatnessThreshold= 56-16-1;
766 ppMode->maxClippedThreshold= 0.01;
767 ppMode->error=0;
768
769 strncpy(temp, name, GET_MODE_BUFFER_SIZE);
770
771 av_log(NULL, AV_LOG_DEBUG, "pp: %s\n", name);
772
773 for(;;){
774 char *filterName;
775 int q= 1000000; //PP_QUALITY_MAX;
776 int chrom=-1;
777 int luma=-1;
778 char *option;
779 char *options[OPTIONS_ARRAY_SIZE];
780 int i;
781 int filterNameOk=0;
782 int numOfUnknownOptions=0;
783 int enable=1; //does the user want us to enabled or disabled the filter
784
785 filterToken= strtok(p, filterDelimiters);
786 if(filterToken == NULL) break;
787 p+= strlen(filterToken) + 1; // p points to next filterToken
788 filterName= strtok(filterToken, optionDelimiters);
789 av_log(NULL, AV_LOG_DEBUG, "pp: %s::%s\n", filterToken, filterName);
790
791 if(*filterName == '-'){
792 enable=0;
793 filterName++;
794 }
795
796 for(;;){ //for all options
797 option= strtok(NULL, optionDelimiters);
798 if(option == NULL) break;
799
800 av_log(NULL, AV_LOG_DEBUG, "pp: option: %s\n", option);
801 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
802 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
803 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
804 else if(!strcmp("noluma", option) || !strcmp("n", option)) luma=0;
805 else{
806 options[numOfUnknownOptions] = option;
807 numOfUnknownOptions++;
808 }
809 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
810 }
811 options[numOfUnknownOptions] = NULL;
812
813 /* replace stuff from the replace Table */
814 for(i=0; replaceTable[2*i]!=NULL; i++){
815 if(!strcmp(replaceTable[2*i], filterName)){
816 int newlen= strlen(replaceTable[2*i + 1]);
817 int plen;
818 int spaceLeft;
819
820 if(p==NULL) p= temp, *p=0; //last filter
821 else p--, *p=','; //not last filter
822
823 plen= strlen(p);
824 spaceLeft= p - temp + plen;
825 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE){
826 ppMode->error++;
827 break;
828 }
829 memmove(p + newlen, p, plen+1);
830 memcpy(p, replaceTable[2*i + 1], newlen);
831 filterNameOk=1;
832 }
833 }
834
835 for(i=0; filters[i].shortName!=NULL; i++){
836 if( !strcmp(filters[i].longName, filterName)
837 || !strcmp(filters[i].shortName, filterName)){
838 ppMode->lumMode &= ~filters[i].mask;
839 ppMode->chromMode &= ~filters[i].mask;
840
841 filterNameOk=1;
842 if(!enable) break; // user wants to disable it
843
844 if(q >= filters[i].minLumQuality && luma)
845 ppMode->lumMode|= filters[i].mask;
846 if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
847 if(q >= filters[i].minChromQuality)
848 ppMode->chromMode|= filters[i].mask;
849
850 if(filters[i].mask == LEVEL_FIX){
851 int o;
852 ppMode->minAllowedY= 16;
853 ppMode->maxAllowedY= 234;
854 for(o=0; options[o]!=NULL; o++){
855 if( !strcmp(options[o],"fullyrange")
856 ||!strcmp(options[o],"f")){
857 ppMode->minAllowedY= 0;
858 ppMode->maxAllowedY= 255;
859 numOfUnknownOptions--;
860 }
861 }
862 }
863 else if(filters[i].mask == TEMP_NOISE_FILTER)
864 {
865 int o;
866 int numOfNoises=0;
867
868 for(o=0; options[o]!=NULL; o++){
869 char *tail;
870 ppMode->maxTmpNoise[numOfNoises]=
871 strtol(options[o], &tail, 0);
872 if(tail!=options[o]){
873 numOfNoises++;
874 numOfUnknownOptions--;
875 if(numOfNoises >= 3) break;
876 }
877 }
878 }
879 else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK
880 || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK){
881 int o;
882
883 for(o=0; options[o]!=NULL && o<2; o++){
884 char *tail;
885 int val= strtol(options[o], &tail, 0);
886 if(tail==options[o]) break;
887
888 numOfUnknownOptions--;
889 if(o==0) ppMode->baseDcDiff= val;
890 else ppMode->flatnessThreshold= val;
891 }
892 }
893 else if(filters[i].mask == FORCE_QUANT){
894 int o;
895 ppMode->forcedQuant= 15;
896
897 for(o=0; options[o]!=NULL && o<1; o++){
898 char *tail;
899 int val= strtol(options[o], &tail, 0);
900 if(tail==options[o]) break;
901
902 numOfUnknownOptions--;
903 ppMode->forcedQuant= val;
904 }
905 }
906 }
907 }
908 if(!filterNameOk) ppMode->error++;
909 ppMode->error += numOfUnknownOptions;
910 }
911
912 av_log(NULL, AV_LOG_DEBUG, "pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode);
913 if(ppMode->error){
914 av_log(NULL, AV_LOG_ERROR, "%d errors in postprocess string \"%s\"\n", ppMode->error, name);
915 av_free(ppMode);
916 return NULL;
917 }
918 return ppMode;
919 }
920
921 void pp_free_mode(pp_mode *mode){
922 av_free(mode);
923 }
924
925 static void reallocAlign(void **p, int alignment, int size){
926 av_free(*p);
927 *p= av_mallocz(size);
928 }
929
930 static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){
931 int mbWidth = (width+15)>>4;
932 int mbHeight= (height+15)>>4;
933 int i;
934
935 c->stride= stride;
936 c->qpStride= qpStride;
937
938 reallocAlign((void **)&c->tempDst, 8, stride*24);
939 reallocAlign((void **)&c->tempSrc, 8, stride*24);
940 reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
941 reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));
942 for(i=0; i<256; i++)
943 c->yHistogram[i]= width*height/64*15/256;
944
945 for(i=0; i<3; i++){
946 //Note: The +17*1024 is just there so i do not have to worry about r/w over the end.
947 reallocAlign((void **)&c->tempBlurred[i], 8, stride*mbHeight*16 + 17*1024);
948 reallocAlign((void **)&c->tempBlurredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
949 }
950
951 reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
952 reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
953 reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
954 reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
955 }
956
957 static const char * context_to_name(void * ptr) {
958 return "postproc";
959 }
960
961 static const AVClass av_codec_context_class = { "Postproc", context_to_name, NULL };
962
963 pp_context *pp_get_context(int width, int height, int cpuCaps){
964 PPContext *c= av_malloc(sizeof(PPContext));
965 int stride= FFALIGN(width, 16); //assumed / will realloc if needed
966 int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed
967
968 memset(c, 0, sizeof(PPContext));
969 c->av_class = &av_codec_context_class;
970 c->cpuCaps= cpuCaps;
971 if(cpuCaps&PP_FORMAT){
972 c->hChromaSubSample= cpuCaps&0x3;
973 c->vChromaSubSample= (cpuCaps>>4)&0x3;
974 }else{
975 c->hChromaSubSample= 1;
976 c->vChromaSubSample= 1;
977 }
978
979 reallocBuffers(c, width, height, stride, qpStride);
980
981 c->frameNum=-1;
982
983 return c;
984 }
985
986 void pp_free_context(void *vc){
987 PPContext *c = (PPContext*)vc;
988 int i;
989
990 for(i=0; i<3; i++) av_free(c->tempBlurred[i]);
991 for(i=0; i<3; i++) av_free(c->tempBlurredPast[i]);
992
993 av_free(c->tempBlocks);
994 av_free(c->yHistogram);
995 av_free(c->tempDst);
996 av_free(c->tempSrc);
997 av_free(c->deintTemp);
998 av_free(c->stdQPTable);
999 av_free(c->nonBQPTable);
1000 av_free(c->forcedQPTable);
1001
1002 memset(c, 0, sizeof(PPContext));
1003
1004 av_free(c);
1005 }
1006
1007 void pp_postprocess(const uint8_t * src[3], const int srcStride[3],
1008 uint8_t * dst[3], const int dstStride[3],
1009 int width, int height,
1010 const QP_STORE_T *QP_store, int QPStride,
1011 pp_mode *vm, void *vc, int pict_type)
1012 {
1013 int mbWidth = (width+15)>>4;
1014 int mbHeight= (height+15)>>4;
1015 PPMode *mode = (PPMode*)vm;
1016 PPContext *c = (PPContext*)vc;
1017 int minStride= FFMAX(FFABS(srcStride[0]), FFABS(dstStride[0]));
1018 int absQPStride = FFABS(QPStride);
1019
1020 // c->stride and c->QPStride are always positive
1021 if(c->stride < minStride || c->qpStride < absQPStride)
1022 reallocBuffers(c, width, height,
1023 FFMAX(minStride, c->stride),
1024 FFMAX(c->qpStride, absQPStride));
1025
1026 if(QP_store==NULL || (mode->lumMode & FORCE_QUANT)){
1027 int i;
1028 QP_store= c->forcedQPTable;
1029 absQPStride = QPStride = 0;
1030 if(mode->lumMode & FORCE_QUANT)
1031 for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= mode->forcedQuant;
1032 else
1033 for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= 1;
1034 }
1035
1036 if(pict_type & PP_PICT_TYPE_QP2){
1037 int i;
1038 const int count= mbHeight * absQPStride;
1039 for(i=0; i<(count>>2); i++){
1040 ((uint32_t*)c->stdQPTable)[i] = (((const uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F;
1041 }
1042 for(i<<=2; i<count; i++){
1043 c->stdQPTable[i] = QP_store[i]>>1;
1044 }
1045 QP_store= c->stdQPTable;
1046 QPStride= absQPStride;
1047 }
1048
1049 if(0){
1050 int x,y;
1051 for(y=0; y<mbHeight; y++){
1052 for(x=0; x<mbWidth; x++){
1053 av_log(c, AV_LOG_INFO, "%2d ", QP_store[x + y*QPStride]);
1054 }
1055 av_log(c, AV_LOG_INFO, "\n");
1056 }
1057 av_log(c, AV_LOG_INFO, "\n");
1058 }
1059
1060 if((pict_type&7)!=3){
1061 if (QPStride >= 0){
1062 int i;
1063 const int count= mbHeight * QPStride;
1064 for(i=0; i<(count>>2); i++){
1065 ((uint32_t*)c->nonBQPTable)[i] = ((const uint32_t*)QP_store)[i] & 0x3F3F3F3F;
1066 }
1067 for(i<<=2; i<count; i++){
1068 c->nonBQPTable[i] = QP_store[i] & 0x3F;
1069 }
1070 } else {
1071 int i,j;
1072 for(i=0; i<mbHeight; i++) {
1073 for(j=0; j<absQPStride; j++) {
1074 c->nonBQPTable[i*absQPStride+j] = QP_store[i*QPStride+j] & 0x3F;
1075 }
1076 }
1077 }
1078 }
1079
1080 av_log(c, AV_LOG_DEBUG, "using npp filters 0x%X/0x%X\n",
1081 mode->lumMode, mode->chromMode);
1082
1083 postProcess(src[0], srcStride[0], dst[0], dstStride[0],
1084 width, height, QP_store, QPStride, 0, mode, c);
1085
1086 width = (width )>>c->hChromaSubSample;
1087 height = (height)>>c->vChromaSubSample;
1088
1089 if(mode->chromMode){
1090 postProcess(src[1], srcStride[1], dst[1], dstStride[1],
1091 width, height, QP_store, QPStride, 1, mode, c);
1092 postProcess(src[2], srcStride[2], dst[2], dstStride[2],
1093 width, height, QP_store, QPStride, 2, mode, c);
1094 }
1095 else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2]){
1096 linecpy(dst[1], src[1], height, srcStride[1]);
1097 linecpy(dst[2], src[2], height, srcStride[2]);
1098 }else{
1099 int y;
1100 for(y=0; y<height; y++){
1101 memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
1102 memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
1103 }
1104 }
1105 }
1106