Mention the l5 pp filter in 'mplayer -pphelp' output.
[ffmpeg.git] / libavcodec / 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 program is free software; you can redistribute it and/or modify
7     it under the terms of the GNU General Public License as published by
8     the Free Software Foundation; either version 2 of the License, or
9     (at your option) any later version.
10
11     This program 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
14     GNU General Public License for more details.
15
16     You should have received a copy of the GNU General Public License
17     along with this program; if not, write to the Free Software
18     Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
19 */
20
21 /**
22  * @file postprocess.c
23  * postprocessing.
24  */
25  
26 /*
27                         C       MMX     MMX2    3DNow   AltiVec
28 isVertDC                Ec      Ec                      Ec
29 isVertMinMaxOk          Ec      Ec                      Ec
30 doVertLowPass           E               e       e       Ec
31 doVertDefFilter         Ec      Ec      e       e       Ec
32 isHorizDC               Ec      Ec                      Ec
33 isHorizMinMaxOk         a       E                       Ec
34 doHorizLowPass          E               e       e       Ec
35 doHorizDefFilter        Ec      Ec      e       e       Ec
36 do_a_deblock            Ec      E       Ec      E
37 deRing                  E               e       e*      Ecp
38 Vertical RKAlgo1        E               a       a
39 Horizontal RKAlgo1                      a       a
40 Vertical X1#            a               E       E
41 Horizontal X1#          a               E       E
42 LinIpolDeinterlace      e               E       E*
43 CubicIpolDeinterlace    a               e       e*
44 LinBlendDeinterlace     e               E       E*
45 MedianDeinterlace#      E       Ec      Ec
46 TempDeNoiser#           E               e       e       Ec
47
48 * i dont have a 3dnow CPU -> its untested, but noone said it doesnt work so it seems to work
49 # more or less selfinvented filters so the exactness isnt too meaningfull
50 E = Exact implementation
51 e = allmost exact implementation (slightly different rounding,...)
52 a = alternative / approximate impl
53 c = checked against the other implementations (-vo md5)
54 p = partially optimized, still some work to do
55 */
56
57 /*
58 TODO:
59 reduce the time wasted on the mem transfer
60 unroll stuff if instructions depend too much on the prior one
61 move YScale thing to the end instead of fixing QP
62 write a faster and higher quality deblocking filter :)
63 make the mainloop more flexible (variable number of blocks at once
64         (the if/else stuff per block is slowing things down)
65 compare the quality & speed of all filters
66 split this huge file
67 optimize c versions
68 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
69 ...
70 */
71
72 //Changelog: use the CVS log
73
74 #include "config.h"
75 #include <inttypes.h>
76 #include <stdio.h>
77 #include <stdlib.h>
78 #include <string.h>
79 #ifdef HAVE_MALLOC_H
80 #include <malloc.h>
81 #endif
82 //#undef HAVE_MMX2
83 //#define HAVE_3DNOW
84 //#undef HAVE_MMX
85 //#undef ARCH_X86
86 //#define DEBUG_BRIGHTNESS
87 #ifdef USE_FASTMEMCPY
88 #include "fastmemcpy.h"
89 #endif
90 #include "postprocess.h"
91 #include "postprocess_internal.h"
92
93 #include "mangle.h" //FIXME should be supressed
94
95 #ifdef HAVE_ALTIVEC_H
96 #include <altivec.h>
97 #endif
98
99 #ifndef HAVE_MEMALIGN
100 #define memalign(a,b) malloc(b)
101 #endif
102
103 #define MIN(a,b) ((a) > (b) ? (b) : (a))
104 #define MAX(a,b) ((a) < (b) ? (b) : (a))
105 #define ABS(a) ((a) > 0 ? (a) : (-(a)))
106 #define SIGN(a) ((a) > 0 ? 1 : -1)
107
108 #define GET_MODE_BUFFER_SIZE 500
109 #define OPTIONS_ARRAY_SIZE 10
110 #define BLOCK_SIZE 8
111 #define TEMP_STRIDE 8
112 //#define NUM_BLOCKS_AT_ONCE 16 //not used yet
113
114 #if defined(__GNUC__) && (__GNUC__ > 3 || __GNUC__ == 3 && __GNUC_MINOR__ > 0)
115 #    define attribute_used __attribute__((used))
116 #    define always_inline __attribute__((always_inline)) inline
117 #else
118 #    define attribute_used
119 #    define always_inline inline
120 #endif
121
122 #if defined(ARCH_X86) || defined(ARCH_X86_64)
123 static uint64_t __attribute__((aligned(8))) attribute_used w05=         0x0005000500050005LL;
124 static uint64_t __attribute__((aligned(8))) attribute_used w04=         0x0004000400040004LL;
125 static uint64_t __attribute__((aligned(8))) attribute_used w20=         0x0020002000200020LL;
126 static uint64_t __attribute__((aligned(8))) attribute_used b00=                 0x0000000000000000LL;
127 static uint64_t __attribute__((aligned(8))) attribute_used b01=                 0x0101010101010101LL;
128 static uint64_t __attribute__((aligned(8))) attribute_used b02=                 0x0202020202020202LL;
129 static uint64_t __attribute__((aligned(8))) attribute_used b08=                 0x0808080808080808LL;
130 static uint64_t __attribute__((aligned(8))) attribute_used b80=                 0x8080808080808080LL;
131 #endif
132
133 static uint8_t clip_table[3*256];
134 static uint8_t * const clip_tab= clip_table + 256;
135
136 static const int verbose= 0;
137
138 static const int attribute_used deringThreshold= 20;
139
140
141 static struct PPFilter filters[]=
142 {
143         {"hb", "hdeblock",              1, 1, 3, H_DEBLOCK},
144         {"vb", "vdeblock",              1, 2, 4, V_DEBLOCK},
145 /*      {"hr", "rkhdeblock",            1, 1, 3, H_RK1_FILTER},
146         {"vr", "rkvdeblock",            1, 2, 4, V_RK1_FILTER},*/
147         {"h1", "x1hdeblock",            1, 1, 3, H_X1_FILTER},
148         {"v1", "x1vdeblock",            1, 2, 4, V_X1_FILTER},
149         {"ha", "ahdeblock",             1, 1, 3, H_A_DEBLOCK},
150         {"va", "avdeblock",             1, 2, 4, V_A_DEBLOCK},
151         {"dr", "dering",                1, 5, 6, DERING},
152         {"al", "autolevels",            0, 1, 2, LEVEL_FIX},
153         {"lb", "linblenddeint",         1, 1, 4, LINEAR_BLEND_DEINT_FILTER},
154         {"li", "linipoldeint",          1, 1, 4, LINEAR_IPOL_DEINT_FILTER},
155         {"ci", "cubicipoldeint",        1, 1, 4, CUBIC_IPOL_DEINT_FILTER},
156         {"md", "mediandeint",           1, 1, 4, MEDIAN_DEINT_FILTER},
157         {"fd", "ffmpegdeint",           1, 1, 4, FFMPEG_DEINT_FILTER},
158         {"l5", "lowpass5",              1, 1, 4, LOWPASS5_DEINT_FILTER},
159         {"tn", "tmpnoise",              1, 7, 8, TEMP_NOISE_FILTER},
160         {"fq", "forcequant",            1, 0, 0, FORCE_QUANT},
161         {NULL, NULL,0,0,0,0} //End Marker
162 };
163
164 static char *replaceTable[]=
165 {
166         "default",      "hdeblock:a,vdeblock:a,dering:a",
167         "de",           "hdeblock:a,vdeblock:a,dering:a",
168         "fast",         "x1hdeblock:a,x1vdeblock:a,dering:a",
169         "fa",           "x1hdeblock:a,x1vdeblock:a,dering:a",
170         "ac",           "ha:a:128:7,va:a,dering:a",
171         NULL //End Marker
172 };
173
174
175 #if defined(ARCH_X86) || defined(ARCH_X86_64)
176 static inline void prefetchnta(void *p)
177 {
178         asm volatile(   "prefetchnta (%0)\n\t"
179                 : : "r" (p)
180         );
181 }
182
183 static inline void prefetcht0(void *p)
184 {
185         asm volatile(   "prefetcht0 (%0)\n\t"
186                 : : "r" (p)
187         );
188 }
189
190 static inline void prefetcht1(void *p)
191 {
192         asm volatile(   "prefetcht1 (%0)\n\t"
193                 : : "r" (p)
194         );
195 }
196
197 static inline void prefetcht2(void *p)
198 {
199         asm volatile(   "prefetcht2 (%0)\n\t"
200                 : : "r" (p)
201         );
202 }
203 #endif
204
205 // The horizontal Functions exist only in C cuz the MMX code is faster with vertical filters and transposing
206
207 /**
208  * Check if the given 8x8 Block is mostly "flat"
209  */
210 static inline int isHorizDC_C(uint8_t src[], int stride, PPContext *c)
211 {
212         int numEq= 0;
213         int y;
214         const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
215         const int dcThreshold= dcOffset*2 + 1;
216
217         for(y=0; y<BLOCK_SIZE; y++)
218         {
219                 if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
220                 if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++;
221                 if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++;
222                 if(((unsigned)(src[3] - src[4] + dcOffset)) < dcThreshold) numEq++;
223                 if(((unsigned)(src[4] - src[5] + dcOffset)) < dcThreshold) numEq++;
224                 if(((unsigned)(src[5] - src[6] + dcOffset)) < dcThreshold) numEq++;
225                 if(((unsigned)(src[6] - src[7] + dcOffset)) < dcThreshold) numEq++;
226                 src+= stride;
227         }
228         return numEq > c->ppMode.flatnessThreshold;
229 }
230
231 /**
232  * Check if the middle 8x8 Block in the given 8x16 block is flat
233  */
234 static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c){
235         int numEq= 0;
236         int y;
237         const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
238         const int dcThreshold= dcOffset*2 + 1;
239
240         src+= stride*4; // src points to begin of the 8x8 Block
241         for(y=0; y<BLOCK_SIZE-1; y++)
242         {
243                 if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++;
244                 if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++;
245                 if(((unsigned)(src[2] - src[2+stride] + dcOffset)) < dcThreshold) numEq++;
246                 if(((unsigned)(src[3] - src[3+stride] + dcOffset)) < dcThreshold) numEq++;
247                 if(((unsigned)(src[4] - src[4+stride] + dcOffset)) < dcThreshold) numEq++;
248                 if(((unsigned)(src[5] - src[5+stride] + dcOffset)) < dcThreshold) numEq++;
249                 if(((unsigned)(src[6] - src[6+stride] + dcOffset)) < dcThreshold) numEq++;
250                 if(((unsigned)(src[7] - src[7+stride] + dcOffset)) < dcThreshold) numEq++;
251                 src+= stride;
252         }
253         return numEq > c->ppMode.flatnessThreshold;
254 }
255
256 static inline int isHorizMinMaxOk_C(uint8_t src[], int stride, int QP)
257 {
258         int i;
259 #if 1
260         for(i=0; i<2; i++){
261                 if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0;
262                 src += stride;
263                 if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0;
264                 src += stride;
265                 if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0;
266                 src += stride;
267                 if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0;
268                 src += stride;
269         }
270 #else        
271         for(i=0; i<8; i++){
272                 if((unsigned)(src[0] - src[7] + 2*QP) > 4*QP) return 0;
273                 src += stride;
274         }
275 #endif
276         return 1;
277 }
278
279 static inline int isVertMinMaxOk_C(uint8_t src[], int stride, int QP)
280 {
281 #if 1
282 #if 1
283         int x;
284         src+= stride*4;
285         for(x=0; x<BLOCK_SIZE; x+=4)
286         {
287                 if((unsigned)(src[  x + 0*stride] - src[  x + 5*stride] + 2*QP) > 4*QP) return 0;
288                 if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0;
289                 if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0;
290                 if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0;
291         }
292 #else
293         int x;
294         src+= stride*3;
295         for(x=0; x<BLOCK_SIZE; x++)
296         {
297                 if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
298         }
299 #endif
300         return 1;
301 #else
302         int x;
303         src+= stride*4;
304         for(x=0; x<BLOCK_SIZE; x++)
305         {
306                 int min=255;
307                 int max=0;
308                 int y;
309                 for(y=0; y<8; y++){
310                         int v= src[x + y*stride];
311                         if(v>max) max=v;
312                         if(v<min) min=v;
313                 }
314                 if(max-min > 2*QP) return 0;
315         }
316         return 1;
317 #endif
318 }
319
320 static inline int horizClassify_C(uint8_t src[], int stride, PPContext *c){
321         if( isHorizDC_C(src, stride, c) ){
322                 if( isHorizMinMaxOk_C(src, stride, c->QP) )
323                         return 1;
324                 else
325                         return 0;
326         }else{
327                 return 2;
328         }
329 }
330
331 static inline int vertClassify_C(uint8_t src[], int stride, PPContext *c){
332         if( isVertDC_C(src, stride, c) ){
333                 if( isVertMinMaxOk_C(src, stride, c->QP) )
334                         return 1;
335                 else
336                         return 0;
337         }else{
338                 return 2;
339         }
340 }
341
342 static inline void doHorizDefFilter_C(uint8_t dst[], int stride, PPContext *c)
343 {
344         int y;
345         for(y=0; y<BLOCK_SIZE; y++)
346         {
347                 const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]);
348
349                 if(ABS(middleEnergy) < 8*c->QP)
350                 {
351                         const int q=(dst[3] - dst[4])/2;
352                         const int leftEnergy=  5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
353                         const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
354
355                         int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
356                         d= MAX(d, 0);
357
358                         d= (5*d + 32) >> 6;
359                         d*= SIGN(-middleEnergy);
360
361                         if(q>0)
362                         {
363                                 d= d<0 ? 0 : d;
364                                 d= d>q ? q : d;
365                         }
366                         else
367                         {
368                                 d= d>0 ? 0 : d;
369                                 d= d<q ? q : d;
370                         }
371
372                         dst[3]-= d;
373                         dst[4]+= d;
374                 }
375                 dst+= stride;
376         }
377 }
378
379 /**
380  * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
381  * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
382  */
383 static inline void doHorizLowPass_C(uint8_t dst[], int stride, PPContext *c)
384 {
385         int y;
386         for(y=0; y<BLOCK_SIZE; y++)
387         {
388                 const int first= ABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0];
389                 const int last= ABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7];
390
391                 int sums[10];
392                 sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4;
393                 sums[1] = sums[0] - first  + dst[3];
394                 sums[2] = sums[1] - first  + dst[4];
395                 sums[3] = sums[2] - first  + dst[5];
396                 sums[4] = sums[3] - first  + dst[6];
397                 sums[5] = sums[4] - dst[0] + dst[7];
398                 sums[6] = sums[5] - dst[1] + last;
399                 sums[7] = sums[6] - dst[2] + last;
400                 sums[8] = sums[7] - dst[3] + last;
401                 sums[9] = sums[8] - dst[4] + last;
402
403                 dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4;
404                 dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4;
405                 dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4;
406                 dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4;
407                 dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4;
408                 dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4;
409                 dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4;
410                 dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4;
411
412                 dst+= stride;
413         }
414 }
415
416 /**
417  * Experimental Filter 1 (Horizontal)
418  * will not damage linear gradients
419  * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
420  * can only smooth blocks at the expected locations (it cant smooth them if they did move)
421  * MMX2 version does correct clipping C version doesnt
422  * not identical with the vertical one
423  */
424 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
425 {
426         int y;
427         static uint64_t *lut= NULL;
428         if(lut==NULL)
429         {
430                 int i;
431                 lut= (uint64_t*)memalign(8, 256*8);
432                 for(i=0; i<256; i++)
433                 {
434                         int v= i < 128 ? 2*i : 2*(i-256);
435 /*
436 //Simulate 112242211 9-Tap filter
437                         uint64_t a= (v/16) & 0xFF;
438                         uint64_t b= (v/8) & 0xFF;
439                         uint64_t c= (v/4) & 0xFF;
440                         uint64_t d= (3*v/8) & 0xFF;
441 */
442 //Simulate piecewise linear interpolation
443                         uint64_t a= (v/16) & 0xFF;
444                         uint64_t b= (v*3/16) & 0xFF;
445                         uint64_t c= (v*5/16) & 0xFF;
446                         uint64_t d= (7*v/16) & 0xFF;
447                         uint64_t A= (0x100 - a)&0xFF;
448                         uint64_t B= (0x100 - b)&0xFF;
449                         uint64_t C= (0x100 - c)&0xFF;
450                         uint64_t D= (0x100 - c)&0xFF;
451
452                         lut[i]   = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
453                                 (D<<24) | (C<<16) | (B<<8) | (A);
454                         //lut[i] = (v<<32) | (v<<24);
455                 }
456         }
457
458         for(y=0; y<BLOCK_SIZE; y++)
459         {
460                 int a= src[1] - src[2];
461                 int b= src[3] - src[4];
462                 int c= src[5] - src[6];
463
464                 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
465
466                 if(d < QP)
467                 {
468                         int v = d * SIGN(-b);
469
470                         src[1] +=v/8;
471                         src[2] +=v/4;
472                         src[3] +=3*v/8;
473                         src[4] -=3*v/8;
474                         src[5] -=v/4;
475                         src[6] -=v/8;
476
477                 }
478                 src+=stride;
479         }
480 }
481
482 /**
483  * accurate deblock filter
484  */
485 static always_inline void do_a_deblock_C(uint8_t *src, int step, int stride, PPContext *c){
486         int y;
487         const int QP= c->QP;
488         const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
489         const int dcThreshold= dcOffset*2 + 1;
490 //START_TIMER
491         src+= step*4; // src points to begin of the 8x8 Block
492         for(y=0; y<8; y++){
493                 int numEq= 0;
494
495                 if(((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold) numEq++;
496                 if(((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold) numEq++;
497                 if(((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold) numEq++;
498                 if(((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold) numEq++;
499                 if(((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold) numEq++;
500                 if(((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold) numEq++;
501                 if(((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold) numEq++;
502                 if(((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold) numEq++;
503                 if(((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold) numEq++;
504                 if(numEq > c->ppMode.flatnessThreshold){
505                         int min, max, x;
506                         
507                         if(src[0] > src[step]){
508                             max= src[0];
509                             min= src[step];
510                         }else{
511                             max= src[step];
512                             min= src[0];
513                         }
514                         for(x=2; x<8; x+=2){
515                                 if(src[x*step] > src[(x+1)*step]){
516                                         if(src[x    *step] > max) max= src[ x   *step];
517                                         if(src[(x+1)*step] < min) min= src[(x+1)*step];
518                                 }else{
519                                         if(src[(x+1)*step] > max) max= src[(x+1)*step];
520                                         if(src[ x   *step] < min) min= src[ x   *step];
521                                 }
522                         }
523                         if(max-min < 2*QP){
524                                 const int first= ABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0];
525                                 const int last= ABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step];
526                                 
527                                 int sums[10];
528                                 sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4;
529                                 sums[1] = sums[0] - first       + src[3*step];
530                                 sums[2] = sums[1] - first       + src[4*step];
531                                 sums[3] = sums[2] - first       + src[5*step];
532                                 sums[4] = sums[3] - first       + src[6*step];
533                                 sums[5] = sums[4] - src[0*step] + src[7*step];
534                                 sums[6] = sums[5] - src[1*step] + last;
535                                 sums[7] = sums[6] - src[2*step] + last;
536                                 sums[8] = sums[7] - src[3*step] + last;
537                                 sums[9] = sums[8] - src[4*step] + last;
538
539                                 src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4;
540                                 src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4;
541                                 src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4;
542                                 src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4;
543                                 src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4;
544                                 src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4;
545                                 src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4;
546                                 src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4;
547                         }
548                 }else{
549                         const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]);
550
551                         if(ABS(middleEnergy) < 8*QP)
552                         {
553                                 const int q=(src[3*step] - src[4*step])/2;
554                                 const int leftEnergy=  5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]);
555                                 const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]);
556
557                                 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
558                                 d= MAX(d, 0);
559         
560                                 d= (5*d + 32) >> 6;
561                                 d*= SIGN(-middleEnergy);
562         
563                                 if(q>0)
564                                 {
565                                         d= d<0 ? 0 : d;
566                                         d= d>q ? q : d;
567                                 }
568                                 else
569                                 {
570                                         d= d>0 ? 0 : d;
571                                         d= d<q ? q : d;
572                                 }
573         
574                                 src[3*step]-= d;
575                                 src[4*step]+= d;
576                         }
577                 }
578
579                 src += stride;
580         }
581 /*if(step==16){
582     STOP_TIMER("step16")
583 }else{
584     STOP_TIMER("stepX")
585 }*/
586 }
587
588 //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
589 //Plain C versions
590 #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
591 #define COMPILE_C
592 #endif
593
594 #ifdef ARCH_POWERPC
595 #ifdef HAVE_ALTIVEC
596 #define COMPILE_ALTIVEC
597 #endif //HAVE_ALTIVEC
598 #endif //ARCH_POWERPC
599
600 #if defined(ARCH_X86) || defined(ARCH_X86_64)
601
602 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
603 #define COMPILE_MMX
604 #endif
605
606 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
607 #define COMPILE_MMX2
608 #endif
609
610 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
611 #define COMPILE_3DNOW
612 #endif
613 #endif //ARCH_X86
614
615 #undef HAVE_MMX
616 #undef HAVE_MMX2
617 #undef HAVE_3DNOW
618 #undef HAVE_ALTIVEC
619
620 #ifdef COMPILE_C
621 #undef HAVE_MMX
622 #undef HAVE_MMX2
623 #undef HAVE_3DNOW
624 #define RENAME(a) a ## _C
625 #include "postprocess_template.c"
626 #endif
627
628 #ifdef ARCH_POWERPC
629 #ifdef COMPILE_ALTIVEC
630 #undef RENAME
631 #define HAVE_ALTIVEC
632 #define RENAME(a) a ## _altivec
633 #include "postprocess_altivec_template.c"
634 #include "postprocess_template.c"
635 #endif
636 #endif //ARCH_POWERPC
637
638 //MMX versions
639 #ifdef COMPILE_MMX
640 #undef RENAME
641 #define HAVE_MMX
642 #undef HAVE_MMX2
643 #undef HAVE_3DNOW
644 #define RENAME(a) a ## _MMX
645 #include "postprocess_template.c"
646 #endif
647
648 //MMX2 versions
649 #ifdef COMPILE_MMX2
650 #undef RENAME
651 #define HAVE_MMX
652 #define HAVE_MMX2
653 #undef HAVE_3DNOW
654 #define RENAME(a) a ## _MMX2
655 #include "postprocess_template.c"
656 #endif
657
658 //3DNOW versions
659 #ifdef COMPILE_3DNOW
660 #undef RENAME
661 #define HAVE_MMX
662 #undef HAVE_MMX2
663 #define HAVE_3DNOW
664 #define RENAME(a) a ## _3DNow
665 #include "postprocess_template.c"
666 #endif
667
668 // minor note: the HAVE_xyz is messed up after that line so dont use it
669
670 static inline void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
671         QP_STORE_T QPs[], int QPStride, int isColor, pp_mode_t *vm, pp_context_t *vc)
672 {
673         PPContext *c= (PPContext *)vc;
674         PPMode *ppMode= (PPMode *)vm;
675         c->ppMode= *ppMode; //FIXME
676
677         // useing ifs here as they are faster than function pointers allthough the
678         // difference wouldnt be messureable here but its much better because
679         // someone might exchange the cpu whithout restarting mplayer ;)
680 #ifdef RUNTIME_CPUDETECT
681 #if defined(ARCH_X86) || defined(ARCH_X86_64)
682         // ordered per speed fasterst first
683         if(c->cpuCaps & PP_CPU_CAPS_MMX2)
684                 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
685         else if(c->cpuCaps & PP_CPU_CAPS_3DNOW)
686                 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
687         else if(c->cpuCaps & PP_CPU_CAPS_MMX)
688                 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
689         else
690                 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
691 #else
692 #ifdef ARCH_POWERPC
693 #ifdef HAVE_ALTIVEC
694         if(c->cpuCaps & PP_CPU_CAPS_ALTIVEC)
695                 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
696         else
697 #endif
698 #endif
699                 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
700 #endif
701 #else //RUNTIME_CPUDETECT
702 #ifdef HAVE_MMX2
703                 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
704 #elif defined (HAVE_3DNOW)
705                 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
706 #elif defined (HAVE_MMX)
707                 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
708 #elif defined (HAVE_ALTIVEC)
709                 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
710 #else
711                 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
712 #endif
713 #endif //!RUNTIME_CPUDETECT
714 }
715
716 //static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
717 //      QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
718
719 /* -pp Command line Help
720 */
721 char *pp_help=
722 "Available postprocessing filters:\n"
723 "Filters                        Options\n"
724 "short  long name       short   long option     Description\n"
725 "*      *               a       autoq           CPU power dependent enabler\n"
726 "                       c       chrom           chrominance filtering enabled\n"
727 "                       y       nochrom         chrominance filtering disabled\n"
728 "hb     hdeblock        (2 threshold)           horizontal deblocking filter\n"
729 "       1. difference factor: default=32, higher -> more deblocking\n"
730 "       2. flatness threshold: default=39, lower -> more deblocking\n"
731 "                       the h & v deblocking filters share these\n"
732 "                       so you can't set different thresholds for h / v\n"
733 "vb     vdeblock        (2 threshold)           vertical deblocking filter\n"
734 "ha     hadeblock       (2 threshold)           horizontal deblocking filter\n"
735 "va     vadeblock       (2 threshold)           vertical deblocking filter\n"
736 "h1     x1hdeblock                              experimental h deblock filter 1\n"
737 "v1     x1vdeblock                              experimental v deblock filter 1\n"
738 "dr     dering                                  deringing filter\n"
739 "al     autolevels                              automatic brightness / contrast\n"
740 "                       f       fullyrange      stretch luminance to (0..255)\n"
741 "lb     linblenddeint                           linear blend deinterlacer\n"
742 "li     linipoldeint                            linear interpolating deinterlace\n"
743 "ci     cubicipoldeint                          cubic interpolating deinterlacer\n"
744 "md     mediandeint                             median deinterlacer\n"
745 "fd     ffmpegdeint                             ffmpeg deinterlacer\n"
746 "l5     lowpass5                                FIR lowpass deinterlacer\n"
747 "de     default                                 hb:a,vb:a,dr:a\n"
748 "fa     fast                                    h1:a,v1:a,dr:a\n"
749 "tn     tmpnoise        (3 threshold)           temporal noise reducer\n"
750 "                       1. <= 2. <= 3.          larger -> stronger filtering\n"
751 "fq     forceQuant      <quantizer>             force quantizer\n"
752 "Usage:\n"
753 "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n"
754 "long form example:\n"
755 "vdeblock:autoq/hdeblock:autoq/linblenddeint    default,-vdeblock\n"
756 "short form example:\n"
757 "vb:a/hb:a/lb                                   de,-vb\n"
758 "more examples:\n"
759 "tn:64:128:256\n"
760 ;
761
762 pp_mode_t *pp_get_mode_by_name_and_quality(char *name, int quality)
763 {
764         char temp[GET_MODE_BUFFER_SIZE];
765         char *p= temp;
766         char *filterDelimiters= ",/";
767         char *optionDelimiters= ":";
768         struct PPMode *ppMode;
769         char *filterToken;
770
771         ppMode= memalign(8, sizeof(PPMode));
772         
773         ppMode->lumMode= 0;
774         ppMode->chromMode= 0;
775         ppMode->maxTmpNoise[0]= 700;
776         ppMode->maxTmpNoise[1]= 1500;
777         ppMode->maxTmpNoise[2]= 3000;
778         ppMode->maxAllowedY= 234;
779         ppMode->minAllowedY= 16;
780         ppMode->baseDcDiff= 256/8;
781         ppMode->flatnessThreshold= 56-16-1;
782         ppMode->maxClippedThreshold= 0.01;
783         ppMode->error=0;
784
785         strncpy(temp, name, GET_MODE_BUFFER_SIZE);
786
787         if(verbose>1) printf("pp: %s\n", name);
788
789         for(;;){
790                 char *filterName;
791                 int q= 1000000; //PP_QUALITY_MAX;
792                 int chrom=-1;
793                 char *option;
794                 char *options[OPTIONS_ARRAY_SIZE];
795                 int i;
796                 int filterNameOk=0;
797                 int numOfUnknownOptions=0;
798                 int enable=1; //does the user want us to enabled or disabled the filter
799
800                 filterToken= strtok(p, filterDelimiters);
801                 if(filterToken == NULL) break;
802                 p+= strlen(filterToken) + 1; // p points to next filterToken
803                 filterName= strtok(filterToken, optionDelimiters);
804                 if(verbose>1) printf("pp: %s::%s\n", filterToken, filterName);
805
806                 if(*filterName == '-')
807                 {
808                         enable=0;
809                         filterName++;
810                 }
811
812                 for(;;){ //for all options
813                         option= strtok(NULL, optionDelimiters);
814                         if(option == NULL) break;
815
816                         if(verbose>1) printf("pp: option: %s\n", option);
817                         if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
818                         else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
819                         else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
820                         else
821                         {
822                                 options[numOfUnknownOptions] = option;
823                                 numOfUnknownOptions++;
824                         }
825                         if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
826                 }
827                 options[numOfUnknownOptions] = NULL;
828
829                 /* replace stuff from the replace Table */
830                 for(i=0; replaceTable[2*i]!=NULL; i++)
831                 {
832                         if(!strcmp(replaceTable[2*i], filterName))
833                         {
834                                 int newlen= strlen(replaceTable[2*i + 1]);
835                                 int plen;
836                                 int spaceLeft;
837
838                                 if(p==NULL) p= temp, *p=0;      //last filter
839                                 else p--, *p=',';               //not last filter
840
841                                 plen= strlen(p);
842                                 spaceLeft= p - temp + plen;
843                                 if(spaceLeft + newlen  >= GET_MODE_BUFFER_SIZE)
844                                 {
845                                         ppMode->error++;
846                                         break;
847                                 }
848                                 memmove(p + newlen, p, plen+1);
849                                 memcpy(p, replaceTable[2*i + 1], newlen);
850                                 filterNameOk=1;
851                         }
852                 }
853
854                 for(i=0; filters[i].shortName!=NULL; i++)
855                 {
856 //                      printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName);
857                         if(   !strcmp(filters[i].longName, filterName)
858                            || !strcmp(filters[i].shortName, filterName))
859                         {
860                                 ppMode->lumMode &= ~filters[i].mask;
861                                 ppMode->chromMode &= ~filters[i].mask;
862
863                                 filterNameOk=1;
864                                 if(!enable) break; // user wants to disable it
865
866                                 if(q >= filters[i].minLumQuality)
867                                         ppMode->lumMode|= filters[i].mask;
868                                 if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
869                                         if(q >= filters[i].minChromQuality)
870                                                 ppMode->chromMode|= filters[i].mask;
871
872                                 if(filters[i].mask == LEVEL_FIX)
873                                 {
874                                         int o;
875                                         ppMode->minAllowedY= 16;
876                                         ppMode->maxAllowedY= 234;
877                                         for(o=0; options[o]!=NULL; o++)
878                                         {
879                                                 if(  !strcmp(options[o],"fullyrange")
880                                                    ||!strcmp(options[o],"f"))
881                                                 {
882                                                         ppMode->minAllowedY= 0;
883                                                         ppMode->maxAllowedY= 255;
884                                                         numOfUnknownOptions--;
885                                                 }
886                                         }
887                                 }
888                                 else if(filters[i].mask == TEMP_NOISE_FILTER)
889                                 {
890                                         int o;
891                                         int numOfNoises=0;
892
893                                         for(o=0; options[o]!=NULL; o++)
894                                         {
895                                                 char *tail;
896                                                 ppMode->maxTmpNoise[numOfNoises]=
897                                                         strtol(options[o], &tail, 0);
898                                                 if(tail!=options[o])
899                                                 {
900                                                         numOfNoises++;
901                                                         numOfUnknownOptions--;
902                                                         if(numOfNoises >= 3) break;
903                                                 }
904                                         }
905                                 }
906                                 else if(filters[i].mask == V_DEBLOCK   || filters[i].mask == H_DEBLOCK 
907                                      || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK)
908                                 {
909                                         int o;
910
911                                         for(o=0; options[o]!=NULL && o<2; o++)
912                                         {
913                                                 char *tail;
914                                                 int val= strtol(options[o], &tail, 0);
915                                                 if(tail==options[o]) break;
916
917                                                 numOfUnknownOptions--;
918                                                 if(o==0) ppMode->baseDcDiff= val;
919                                                 else ppMode->flatnessThreshold= val;
920                                         }
921                                 }
922                                 else if(filters[i].mask == FORCE_QUANT)
923                                 {
924                                         int o;
925                                         ppMode->forcedQuant= 15;
926
927                                         for(o=0; options[o]!=NULL && o<1; o++)
928                                         {
929                                                 char *tail;
930                                                 int val= strtol(options[o], &tail, 0);
931                                                 if(tail==options[o]) break;
932
933                                                 numOfUnknownOptions--;
934                                                 ppMode->forcedQuant= val;
935                                         }
936                                 }
937                         }
938                 }
939                 if(!filterNameOk) ppMode->error++;
940                 ppMode->error += numOfUnknownOptions;
941         }
942
943         if(verbose>1) printf("pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode);
944         if(ppMode->error)
945         {
946                 fprintf(stderr, "%d errors in postprocess string \"%s\"\n", ppMode->error, name);
947                 free(ppMode);
948                 return NULL;
949         }
950         return ppMode;
951 }
952
953 void pp_free_mode(pp_mode_t *mode){
954     if(mode) free(mode);
955 }
956
957 static void reallocAlign(void **p, int alignment, int size){
958         if(*p) free(*p);
959         *p= memalign(alignment, size);
960         memset(*p, 0, size);
961 }
962
963 static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){
964         int mbWidth = (width+15)>>4;
965         int mbHeight= (height+15)>>4;
966         int i;
967
968         c->stride= stride;
969         c->qpStride= qpStride;
970
971         reallocAlign((void **)&c->tempDst, 8, stride*24);
972         reallocAlign((void **)&c->tempSrc, 8, stride*24);
973         reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
974         reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));
975         for(i=0; i<256; i++)
976                 c->yHistogram[i]= width*height/64*15/256;
977
978         for(i=0; i<3; i++)
979         {
980                 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end
981                 reallocAlign((void **)&c->tempBlured[i], 8, stride*mbHeight*16 + 17*1024);
982                 reallocAlign((void **)&c->tempBluredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
983         }
984
985         reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
986         reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
987         reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
988         reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
989 }
990
991 static void global_init(void){
992         int i;
993         memset(clip_table, 0, 256);
994         for(i=256; i<512; i++)
995                 clip_table[i]= i;
996         memset(clip_table+512, 0, 256);
997 }
998
999 pp_context_t *pp_get_context(int width, int height, int cpuCaps){
1000         PPContext *c= memalign(32, sizeof(PPContext));
1001         int stride= (width+15)&(~15); //assumed / will realloc if needed
1002         int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed
1003         
1004         global_init();
1005
1006         memset(c, 0, sizeof(PPContext));
1007         c->cpuCaps= cpuCaps;
1008         if(cpuCaps&PP_FORMAT){
1009                 c->hChromaSubSample= cpuCaps&0x3;
1010                 c->vChromaSubSample= (cpuCaps>>4)&0x3;
1011         }else{
1012                 c->hChromaSubSample= 1;
1013                 c->vChromaSubSample= 1;
1014         }
1015
1016         reallocBuffers(c, width, height, stride, qpStride);
1017         
1018         c->frameNum=-1;
1019
1020         return c;
1021 }
1022
1023 void pp_free_context(void *vc){
1024         PPContext *c = (PPContext*)vc;
1025         int i;
1026         
1027         for(i=0; i<3; i++) free(c->tempBlured[i]);
1028         for(i=0; i<3; i++) free(c->tempBluredPast[i]);
1029         
1030         free(c->tempBlocks);
1031         free(c->yHistogram);
1032         free(c->tempDst);
1033         free(c->tempSrc);
1034         free(c->deintTemp);
1035         free(c->stdQPTable);
1036         free(c->nonBQPTable);
1037         free(c->forcedQPTable);
1038         
1039         memset(c, 0, sizeof(PPContext));
1040
1041         free(c);
1042 }
1043
1044 void  pp_postprocess(uint8_t * src[3], int srcStride[3],
1045                  uint8_t * dst[3], int dstStride[3],
1046                  int width, int height,
1047                  QP_STORE_T *QP_store,  int QPStride,
1048                  pp_mode_t *vm,  void *vc, int pict_type)
1049 {
1050         int mbWidth = (width+15)>>4;
1051         int mbHeight= (height+15)>>4;
1052         PPMode *mode = (PPMode*)vm;
1053         PPContext *c = (PPContext*)vc;
1054         int minStride= MAX(srcStride[0], dstStride[0]);
1055
1056         if(c->stride < minStride || c->qpStride < QPStride)
1057                 reallocBuffers(c, width, height, 
1058                                 MAX(minStride, c->stride), 
1059                                 MAX(c->qpStride, QPStride));
1060
1061         if(QP_store==NULL || (mode->lumMode & FORCE_QUANT)) 
1062         {
1063                 int i;
1064                 QP_store= c->forcedQPTable;
1065                 QPStride= 0;
1066                 if(mode->lumMode & FORCE_QUANT)
1067                         for(i=0; i<mbWidth; i++) QP_store[i]= mode->forcedQuant;
1068                 else
1069                         for(i=0; i<mbWidth; i++) QP_store[i]= 1;
1070         }
1071 //printf("pict_type:%d\n", pict_type);
1072
1073         if(pict_type & PP_PICT_TYPE_QP2){
1074                 int i;
1075                 const int count= mbHeight * QPStride;
1076                 for(i=0; i<(count>>2); i++){
1077                         ((uint32_t*)c->stdQPTable)[i] = (((uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F;
1078                 }
1079                 for(i<<=2; i<count; i++){
1080                         c->stdQPTable[i] = QP_store[i]>>1;
1081                 }
1082                 QP_store= c->stdQPTable;
1083         }
1084
1085 if(0){
1086 int x,y;
1087 for(y=0; y<mbHeight; y++){
1088         for(x=0; x<mbWidth; x++){
1089                 printf("%2d ", QP_store[x + y*QPStride]);
1090         }
1091         printf("\n");
1092 }
1093         printf("\n");
1094 }
1095
1096         if((pict_type&7)!=3)
1097         {
1098                 int i;
1099                 const int count= mbHeight * QPStride;
1100                 for(i=0; i<(count>>2); i++){
1101                         ((uint32_t*)c->nonBQPTable)[i] = ((uint32_t*)QP_store)[i] & 0x3F3F3F3F;
1102                 }
1103                 for(i<<=2; i<count; i++){
1104                         c->nonBQPTable[i] = QP_store[i] & 0x3F;
1105                 }
1106         }
1107
1108         if(verbose>2)
1109         {
1110                 printf("using npp filters 0x%X/0x%X\n", mode->lumMode, mode->chromMode);
1111         }
1112
1113         postProcess(src[0], srcStride[0], dst[0], dstStride[0],
1114                 width, height, QP_store, QPStride, 0, mode, c);
1115
1116         width  = (width )>>c->hChromaSubSample;
1117         height = (height)>>c->vChromaSubSample;
1118
1119         if(mode->chromMode)
1120         {
1121                 postProcess(src[1], srcStride[1], dst[1], dstStride[1],
1122                         width, height, QP_store, QPStride, 1, mode, c);
1123                 postProcess(src[2], srcStride[2], dst[2], dstStride[2],
1124                         width, height, QP_store, QPStride, 2, mode, c);
1125         }
1126         else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2])
1127         {
1128                 memcpy(dst[1], src[1], srcStride[1]*height);
1129                 memcpy(dst[2], src[2], srcStride[2]*height);
1130         }
1131         else
1132         {
1133                 int y;
1134                 for(y=0; y<height; y++)
1135                 {
1136                         memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
1137                         memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
1138                 }
1139         }
1140 }
1141