Fix default license header.
[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 postprocess.c
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 dont have a 3dnow CPU -> its untested, but noone said it doesnt work so it seems to work
51 # more or less selfinvented filters so the exactness isnt too meaningfull
52 E = Exact implementation
53 e = allmost 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 "avutil.h"
78 #include <inttypes.h>
79 #include <stdio.h>
80 #include <stdlib.h>
81 #include <string.h>
82 #ifdef HAVE_MALLOC_H
83 #include <malloc.h>
84 #endif
85 //#undef HAVE_MMX2
86 //#define HAVE_3DNOW
87 //#undef HAVE_MMX
88 //#undef ARCH_X86
89 //#define DEBUG_BRIGHTNESS
90 #ifdef USE_FASTMEMCPY
91 #include "libvo/fastmemcpy.h"
92 #endif
93 #include "postprocess.h"
94 #include "postprocess_internal.h"
95
96 #include "mangle.h" //FIXME should be supressed
97
98 #ifdef HAVE_ALTIVEC_H
99 #include <altivec.h>
100 #endif
101
102 #define MIN(a,b) ((a) > (b) ? (b) : (a))
103 #define MAX(a,b) ((a) < (b) ? (b) : (a))
104 #define ABS(a) ((a) > 0 ? (a) : (-(a)))
105 #define SIGN(a) ((a) > 0 ? 1 : -1)
106
107 #define GET_MODE_BUFFER_SIZE 500
108 #define OPTIONS_ARRAY_SIZE 10
109 #define BLOCK_SIZE 8
110 #define TEMP_STRIDE 8
111 //#define NUM_BLOCKS_AT_ONCE 16 //not used yet
112
113 #if defined(__GNUC__) && (__GNUC__ > 3 || __GNUC__ == 3 && __GNUC_MINOR__ > 0)
114 # define attribute_used __attribute__((used))
115 # define always_inline __attribute__((always_inline)) inline
116 #else
117 # define attribute_used
118 # define always_inline inline
119 #endif
120
121 #if defined(ARCH_X86) || defined(ARCH_X86_64)
122 static uint64_t __attribute__((aligned(8))) attribute_used w05= 0x0005000500050005LL;
123 static uint64_t __attribute__((aligned(8))) attribute_used w04= 0x0004000400040004LL;
124 static uint64_t __attribute__((aligned(8))) attribute_used w20= 0x0020002000200020LL;
125 static uint64_t __attribute__((aligned(8))) attribute_used b00= 0x0000000000000000LL;
126 static uint64_t __attribute__((aligned(8))) attribute_used b01= 0x0101010101010101LL;
127 static uint64_t __attribute__((aligned(8))) attribute_used b02= 0x0202020202020202LL;
128 static uint64_t __attribute__((aligned(8))) attribute_used b08= 0x0808080808080808LL;
129 static uint64_t __attribute__((aligned(8))) attribute_used b80= 0x8080808080808080LL;
130 #endif
131
132 static uint8_t clip_table[3*256];
133 static uint8_t * const clip_tab= clip_table + 256;
134
135 static const int verbose= 0;
136
137 static const int attribute_used deringThreshold= 20;
138
139
140 static struct PPFilter filters[]=
141 {
142 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK},
143 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK},
144 /* {"hr", "rkhdeblock", 1, 1, 3, H_RK1_FILTER},
145 {"vr", "rkvdeblock", 1, 2, 4, V_RK1_FILTER},*/
146 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER},
147 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER},
148 {"ha", "ahdeblock", 1, 1, 3, H_A_DEBLOCK},
149 {"va", "avdeblock", 1, 2, 4, V_A_DEBLOCK},
150 {"dr", "dering", 1, 5, 6, DERING},
151 {"al", "autolevels", 0, 1, 2, LEVEL_FIX},
152 {"lb", "linblenddeint", 1, 1, 4, LINEAR_BLEND_DEINT_FILTER},
153 {"li", "linipoldeint", 1, 1, 4, LINEAR_IPOL_DEINT_FILTER},
154 {"ci", "cubicipoldeint", 1, 1, 4, CUBIC_IPOL_DEINT_FILTER},
155 {"md", "mediandeint", 1, 1, 4, MEDIAN_DEINT_FILTER},
156 {"fd", "ffmpegdeint", 1, 1, 4, FFMPEG_DEINT_FILTER},
157 {"l5", "lowpass5", 1, 1, 4, LOWPASS5_DEINT_FILTER},
158 {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER},
159 {"fq", "forcequant", 1, 0, 0, FORCE_QUANT},
160 {NULL, NULL,0,0,0,0} //End Marker
161 };
162
163 static const char *replaceTable[]=
164 {
165 "default", "hdeblock:a,vdeblock:a,dering:a",
166 "de", "hdeblock:a,vdeblock:a,dering:a",
167 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a",
168 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a",
169 "ac", "ha:a:128:7,va:a,dering:a",
170 NULL //End Marker
171 };
172
173
174 #if defined(ARCH_X86) || defined(ARCH_X86_64)
175 static inline void prefetchnta(void *p)
176 {
177 asm volatile( "prefetchnta (%0)\n\t"
178 : : "r" (p)
179 );
180 }
181
182 static inline void prefetcht0(void *p)
183 {
184 asm volatile( "prefetcht0 (%0)\n\t"
185 : : "r" (p)
186 );
187 }
188
189 static inline void prefetcht1(void *p)
190 {
191 asm volatile( "prefetcht1 (%0)\n\t"
192 : : "r" (p)
193 );
194 }
195
196 static inline void prefetcht2(void *p)
197 {
198 asm volatile( "prefetcht2 (%0)\n\t"
199 : : "r" (p)
200 );
201 }
202 #endif
203
204 // The horizontal Functions exist only in C cuz the MMX code is faster with vertical filters and transposing
205
206 /**
207 * Check if the given 8x8 Block is mostly "flat"
208 */
209 static inline int isHorizDC_C(uint8_t src[], int stride, PPContext *c)
210 {
211 int numEq= 0;
212 int y;
213 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
214 const int dcThreshold= dcOffset*2 + 1;
215
216 for(y=0; y<BLOCK_SIZE; y++)
217 {
218 if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
219 if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++;
220 if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++;
221 if(((unsigned)(src[3] - src[4] + dcOffset)) < dcThreshold) numEq++;
222 if(((unsigned)(src[4] - src[5] + dcOffset)) < dcThreshold) numEq++;
223 if(((unsigned)(src[5] - src[6] + dcOffset)) < dcThreshold) numEq++;
224 if(((unsigned)(src[6] - src[7] + dcOffset)) < dcThreshold) numEq++;
225 src+= stride;
226 }
227 return numEq > c->ppMode.flatnessThreshold;
228 }
229
230 /**
231 * Check if the middle 8x8 Block in the given 8x16 block is flat
232 */
233 static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c){
234 int numEq= 0;
235 int y;
236 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
237 const int dcThreshold= dcOffset*2 + 1;
238
239 src+= stride*4; // src points to begin of the 8x8 Block
240 for(y=0; y<BLOCK_SIZE-1; y++)
241 {
242 if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++;
243 if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++;
244 if(((unsigned)(src[2] - src[2+stride] + dcOffset)) < dcThreshold) numEq++;
245 if(((unsigned)(src[3] - src[3+stride] + dcOffset)) < dcThreshold) numEq++;
246 if(((unsigned)(src[4] - src[4+stride] + dcOffset)) < dcThreshold) numEq++;
247 if(((unsigned)(src[5] - src[5+stride] + dcOffset)) < dcThreshold) numEq++;
248 if(((unsigned)(src[6] - src[6+stride] + dcOffset)) < dcThreshold) numEq++;
249 if(((unsigned)(src[7] - src[7+stride] + dcOffset)) < dcThreshold) numEq++;
250 src+= stride;
251 }
252 return numEq > c->ppMode.flatnessThreshold;
253 }
254
255 static inline int isHorizMinMaxOk_C(uint8_t src[], int stride, int QP)
256 {
257 int i;
258 #if 1
259 for(i=0; i<2; i++){
260 if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0;
261 src += stride;
262 if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0;
263 src += stride;
264 if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0;
265 src += stride;
266 if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0;
267 src += stride;
268 }
269 #else
270 for(i=0; i<8; i++){
271 if((unsigned)(src[0] - src[7] + 2*QP) > 4*QP) return 0;
272 src += stride;
273 }
274 #endif
275 return 1;
276 }
277
278 static inline int isVertMinMaxOk_C(uint8_t src[], int stride, int QP)
279 {
280 #if 1
281 #if 1
282 int x;
283 src+= stride*4;
284 for(x=0; x<BLOCK_SIZE; x+=4)
285 {
286 if((unsigned)(src[ x + 0*stride] - src[ x + 5*stride] + 2*QP) > 4*QP) return 0;
287 if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0;
288 if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0;
289 if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0;
290 }
291 #else
292 int x;
293 src+= stride*3;
294 for(x=0; x<BLOCK_SIZE; x++)
295 {
296 if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
297 }
298 #endif
299 return 1;
300 #else
301 int x;
302 src+= stride*4;
303 for(x=0; x<BLOCK_SIZE; x++)
304 {
305 int min=255;
306 int max=0;
307 int y;
308 for(y=0; y<8; y++){
309 int v= src[x + y*stride];
310 if(v>max) max=v;
311 if(v<min) min=v;
312 }
313 if(max-min > 2*QP) return 0;
314 }
315 return 1;
316 #endif
317 }
318
319 static inline int horizClassify_C(uint8_t src[], int stride, PPContext *c){
320 if( isHorizDC_C(src, stride, c) ){
321 if( isHorizMinMaxOk_C(src, stride, c->QP) )
322 return 1;
323 else
324 return 0;
325 }else{
326 return 2;
327 }
328 }
329
330 static inline int vertClassify_C(uint8_t src[], int stride, PPContext *c){
331 if( isVertDC_C(src, stride, c) ){
332 if( isVertMinMaxOk_C(src, stride, c->QP) )
333 return 1;
334 else
335 return 0;
336 }else{
337 return 2;
338 }
339 }
340
341 static inline void doHorizDefFilter_C(uint8_t dst[], int stride, PPContext *c)
342 {
343 int y;
344 for(y=0; y<BLOCK_SIZE; y++)
345 {
346 const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]);
347
348 if(ABS(middleEnergy) < 8*c->QP)
349 {
350 const int q=(dst[3] - dst[4])/2;
351 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
352 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
353
354 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
355 d= MAX(d, 0);
356
357 d= (5*d + 32) >> 6;
358 d*= SIGN(-middleEnergy);
359
360 if(q>0)
361 {
362 d= d<0 ? 0 : d;
363 d= d>q ? q : d;
364 }
365 else
366 {
367 d= d>0 ? 0 : d;
368 d= d<q ? q : d;
369 }
370
371 dst[3]-= d;
372 dst[4]+= d;
373 }
374 dst+= stride;
375 }
376 }
377
378 /**
379 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
380 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
381 */
382 static inline void doHorizLowPass_C(uint8_t dst[], int stride, PPContext *c)
383 {
384 int y;
385 for(y=0; y<BLOCK_SIZE; y++)
386 {
387 const int first= ABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0];
388 const int last= ABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7];
389
390 int sums[10];
391 sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4;
392 sums[1] = sums[0] - first + dst[3];
393 sums[2] = sums[1] - first + dst[4];
394 sums[3] = sums[2] - first + dst[5];
395 sums[4] = sums[3] - first + dst[6];
396 sums[5] = sums[4] - dst[0] + dst[7];
397 sums[6] = sums[5] - dst[1] + last;
398 sums[7] = sums[6] - dst[2] + last;
399 sums[8] = sums[7] - dst[3] + last;
400 sums[9] = sums[8] - dst[4] + last;
401
402 dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4;
403 dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4;
404 dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4;
405 dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4;
406 dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4;
407 dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4;
408 dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4;
409 dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4;
410
411 dst+= stride;
412 }
413 }
414
415 /**
416 * Experimental Filter 1 (Horizontal)
417 * will not damage linear gradients
418 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
419 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
420 * MMX2 version does correct clipping C version doesnt
421 * not identical with the vertical one
422 */
423 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
424 {
425 int y;
426 static uint64_t *lut= NULL;
427 if(lut==NULL)
428 {
429 int i;
430 lut = av_malloc(256*8);
431 for(i=0; i<256; i++)
432 {
433 int v= i < 128 ? 2*i : 2*(i-256);
434 /*
435 //Simulate 112242211 9-Tap filter
436 uint64_t a= (v/16) & 0xFF;
437 uint64_t b= (v/8) & 0xFF;
438 uint64_t c= (v/4) & 0xFF;
439 uint64_t d= (3*v/8) & 0xFF;
440 */
441 //Simulate piecewise linear interpolation
442 uint64_t a= (v/16) & 0xFF;
443 uint64_t b= (v*3/16) & 0xFF;
444 uint64_t c= (v*5/16) & 0xFF;
445 uint64_t d= (7*v/16) & 0xFF;
446 uint64_t A= (0x100 - a)&0xFF;
447 uint64_t B= (0x100 - b)&0xFF;
448 uint64_t C= (0x100 - c)&0xFF;
449 uint64_t D= (0x100 - c)&0xFF;
450
451 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
452 (D<<24) | (C<<16) | (B<<8) | (A);
453 //lut[i] = (v<<32) | (v<<24);
454 }
455 }
456
457 for(y=0; y<BLOCK_SIZE; y++)
458 {
459 int a= src[1] - src[2];
460 int b= src[3] - src[4];
461 int c= src[5] - src[6];
462
463 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
464
465 if(d < QP)
466 {
467 int v = d * SIGN(-b);
468
469 src[1] +=v/8;
470 src[2] +=v/4;
471 src[3] +=3*v/8;
472 src[4] -=3*v/8;
473 src[5] -=v/4;
474 src[6] -=v/8;
475
476 }
477 src+=stride;
478 }
479 }
480
481 /**
482 * accurate deblock filter
483 */
484 static always_inline void do_a_deblock_C(uint8_t *src, int step, int stride, PPContext *c){
485 int y;
486 const int QP= c->QP;
487 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
488 const int dcThreshold= dcOffset*2 + 1;
489 //START_TIMER
490 src+= step*4; // src points to begin of the 8x8 Block
491 for(y=0; y<8; y++){
492 int numEq= 0;
493
494 if(((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold) numEq++;
495 if(((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold) numEq++;
496 if(((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold) numEq++;
497 if(((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold) numEq++;
498 if(((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold) numEq++;
499 if(((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold) numEq++;
500 if(((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold) numEq++;
501 if(((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold) numEq++;
502 if(((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold) numEq++;
503 if(numEq > c->ppMode.flatnessThreshold){
504 int min, max, x;
505
506 if(src[0] > src[step]){
507 max= src[0];
508 min= src[step];
509 }else{
510 max= src[step];
511 min= src[0];
512 }
513 for(x=2; x<8; x+=2){
514 if(src[x*step] > src[(x+1)*step]){
515 if(src[x *step] > max) max= src[ x *step];
516 if(src[(x+1)*step] < min) min= src[(x+1)*step];
517 }else{
518 if(src[(x+1)*step] > max) max= src[(x+1)*step];
519 if(src[ x *step] < min) min= src[ x *step];
520 }
521 }
522 if(max-min < 2*QP){
523 const int first= ABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0];
524 const int last= ABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step];
525
526 int sums[10];
527 sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4;
528 sums[1] = sums[0] - first + src[3*step];
529 sums[2] = sums[1] - first + src[4*step];
530 sums[3] = sums[2] - first + src[5*step];
531 sums[4] = sums[3] - first + src[6*step];
532 sums[5] = sums[4] - src[0*step] + src[7*step];
533 sums[6] = sums[5] - src[1*step] + last;
534 sums[7] = sums[6] - src[2*step] + last;
535 sums[8] = sums[7] - src[3*step] + last;
536 sums[9] = sums[8] - src[4*step] + last;
537
538 src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4;
539 src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4;
540 src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4;
541 src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4;
542 src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4;
543 src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4;
544 src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4;
545 src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4;
546 }
547 }else{
548 const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]);
549
550 if(ABS(middleEnergy) < 8*QP)
551 {
552 const int q=(src[3*step] - src[4*step])/2;
553 const int leftEnergy= 5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]);
554 const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]);
555
556 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
557 d= MAX(d, 0);
558
559 d= (5*d + 32) >> 6;
560 d*= SIGN(-middleEnergy);
561
562 if(q>0)
563 {
564 d= d<0 ? 0 : d;
565 d= d>q ? q : d;
566 }
567 else
568 {
569 d= d>0 ? 0 : d;
570 d= d<q ? q : d;
571 }
572
573 src[3*step]-= d;
574 src[4*step]+= d;
575 }
576 }
577
578 src += stride;
579 }
580 /*if(step==16){
581 STOP_TIMER("step16")
582 }else{
583 STOP_TIMER("stepX")
584 }*/
585 }
586
587 //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
588 //Plain C versions
589 #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
590 #define COMPILE_C
591 #endif
592
593 #ifdef ARCH_POWERPC
594 #ifdef HAVE_ALTIVEC
595 #define COMPILE_ALTIVEC
596 #endif //HAVE_ALTIVEC
597 #endif //ARCH_POWERPC
598
599 #if defined(ARCH_X86) || defined(ARCH_X86_64)
600
601 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
602 #define COMPILE_MMX
603 #endif
604
605 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
606 #define COMPILE_MMX2
607 #endif
608
609 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
610 #define COMPILE_3DNOW
611 #endif
612 #endif //ARCH_X86
613
614 #undef HAVE_MMX
615 #undef HAVE_MMX2
616 #undef HAVE_3DNOW
617 #undef HAVE_ALTIVEC
618
619 #ifdef COMPILE_C
620 #undef HAVE_MMX
621 #undef HAVE_MMX2
622 #undef HAVE_3DNOW
623 #define RENAME(a) a ## _C
624 #include "postprocess_template.c"
625 #endif
626
627 #ifdef ARCH_POWERPC
628 #ifdef COMPILE_ALTIVEC
629 #undef RENAME
630 #define HAVE_ALTIVEC
631 #define RENAME(a) a ## _altivec
632 #include "postprocess_altivec_template.c"
633 #include "postprocess_template.c"
634 #endif
635 #endif //ARCH_POWERPC
636
637 //MMX versions
638 #ifdef COMPILE_MMX
639 #undef RENAME
640 #define HAVE_MMX
641 #undef HAVE_MMX2
642 #undef HAVE_3DNOW
643 #define RENAME(a) a ## _MMX
644 #include "postprocess_template.c"
645 #endif
646
647 //MMX2 versions
648 #ifdef COMPILE_MMX2
649 #undef RENAME
650 #define HAVE_MMX
651 #define HAVE_MMX2
652 #undef HAVE_3DNOW
653 #define RENAME(a) a ## _MMX2
654 #include "postprocess_template.c"
655 #endif
656
657 //3DNOW versions
658 #ifdef COMPILE_3DNOW
659 #undef RENAME
660 #define HAVE_MMX
661 #undef HAVE_MMX2
662 #define HAVE_3DNOW
663 #define RENAME(a) a ## _3DNow
664 #include "postprocess_template.c"
665 #endif
666
667 // minor note: the HAVE_xyz is messed up after that line so dont use it
668
669 static inline void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
670 QP_STORE_T QPs[], int QPStride, int isColor, pp_mode_t *vm, pp_context_t *vc)
671 {
672 PPContext *c= (PPContext *)vc;
673 PPMode *ppMode= (PPMode *)vm;
674 c->ppMode= *ppMode; //FIXME
675
676 // useing ifs here as they are faster than function pointers allthough the
677 // difference wouldnt be messureable here but its much better because
678 // someone might exchange the cpu whithout restarting mplayer ;)
679 #ifdef RUNTIME_CPUDETECT
680 #if defined(ARCH_X86) || defined(ARCH_X86_64)
681 // ordered per speed fasterst first
682 if(c->cpuCaps & PP_CPU_CAPS_MMX2)
683 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
684 else if(c->cpuCaps & PP_CPU_CAPS_3DNOW)
685 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
686 else if(c->cpuCaps & PP_CPU_CAPS_MMX)
687 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
688 else
689 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
690 #else
691 #ifdef ARCH_POWERPC
692 #ifdef HAVE_ALTIVEC
693 if(c->cpuCaps & PP_CPU_CAPS_ALTIVEC)
694 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
695 else
696 #endif
697 #endif
698 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
699 #endif
700 #else //RUNTIME_CPUDETECT
701 #ifdef HAVE_MMX2
702 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
703 #elif defined (HAVE_3DNOW)
704 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
705 #elif defined (HAVE_MMX)
706 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
707 #elif defined (HAVE_ALTIVEC)
708 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
709 #else
710 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
711 #endif
712 #endif //!RUNTIME_CPUDETECT
713 }
714
715 //static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
716 // QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
717
718 /* -pp Command line Help
719 */
720 char *pp_help=
721 "Available postprocessing filters:\n"
722 "Filters Options\n"
723 "short long name short long option Description\n"
724 "* * a autoq CPU power dependent enabler\n"
725 " c chrom chrominance filtering enabled\n"
726 " y nochrom chrominance filtering disabled\n"
727 " n noluma luma 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 "ac ha:a:128:7,va:a,dr:a\n"
750 "tn tmpnoise (3 threshold) temporal noise reducer\n"
751 " 1. <= 2. <= 3. larger -> stronger filtering\n"
752 "fq forceQuant <quantizer> force quantizer\n"
753 "Usage:\n"
754 "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n"
755 "long form example:\n"
756 "vdeblock:autoq/hdeblock:autoq/linblenddeint default,-vdeblock\n"
757 "short form example:\n"
758 "vb:a/hb:a/lb de,-vb\n"
759 "more examples:\n"
760 "tn:64:128:256\n"
761 "\n"
762 ;
763
764 pp_mode_t *pp_get_mode_by_name_and_quality(char *name, int quality)
765 {
766 char temp[GET_MODE_BUFFER_SIZE];
767 char *p= temp;
768 const char *filterDelimiters= ",/";
769 const char *optionDelimiters= ":";
770 struct PPMode *ppMode;
771 char *filterToken;
772
773 ppMode= av_malloc(sizeof(PPMode));
774
775 ppMode->lumMode= 0;
776 ppMode->chromMode= 0;
777 ppMode->maxTmpNoise[0]= 700;
778 ppMode->maxTmpNoise[1]= 1500;
779 ppMode->maxTmpNoise[2]= 3000;
780 ppMode->maxAllowedY= 234;
781 ppMode->minAllowedY= 16;
782 ppMode->baseDcDiff= 256/8;
783 ppMode->flatnessThreshold= 56-16-1;
784 ppMode->maxClippedThreshold= 0.01;
785 ppMode->error=0;
786
787 strncpy(temp, name, GET_MODE_BUFFER_SIZE);
788
789 if(verbose>1) printf("pp: %s\n", name);
790
791 for(;;){
792 char *filterName;
793 int q= 1000000; //PP_QUALITY_MAX;
794 int chrom=-1;
795 int luma=-1;
796 char *option;
797 char *options[OPTIONS_ARRAY_SIZE];
798 int i;
799 int filterNameOk=0;
800 int numOfUnknownOptions=0;
801 int enable=1; //does the user want us to enabled or disabled the filter
802
803 filterToken= strtok(p, filterDelimiters);
804 if(filterToken == NULL) break;
805 p+= strlen(filterToken) + 1; // p points to next filterToken
806 filterName= strtok(filterToken, optionDelimiters);
807 if(verbose>1) printf("pp: %s::%s\n", filterToken, filterName);
808
809 if(*filterName == '-')
810 {
811 enable=0;
812 filterName++;
813 }
814
815 for(;;){ //for all options
816 option= strtok(NULL, optionDelimiters);
817 if(option == NULL) break;
818
819 if(verbose>1) printf("pp: option: %s\n", option);
820 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
821 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
822 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
823 else if(!strcmp("noluma", option) || !strcmp("n", option)) luma=0;
824 else
825 {
826 options[numOfUnknownOptions] = option;
827 numOfUnknownOptions++;
828 }
829 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
830 }
831 options[numOfUnknownOptions] = NULL;
832
833 /* replace stuff from the replace Table */
834 for(i=0; replaceTable[2*i]!=NULL; i++)
835 {
836 if(!strcmp(replaceTable[2*i], filterName))
837 {
838 int newlen= strlen(replaceTable[2*i + 1]);
839 int plen;
840 int spaceLeft;
841
842 if(p==NULL) p= temp, *p=0; //last filter
843 else p--, *p=','; //not last filter
844
845 plen= strlen(p);
846 spaceLeft= p - temp + plen;
847 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE)
848 {
849 ppMode->error++;
850 break;
851 }
852 memmove(p + newlen, p, plen+1);
853 memcpy(p, replaceTable[2*i + 1], newlen);
854 filterNameOk=1;
855 }
856 }
857
858 for(i=0; filters[i].shortName!=NULL; i++)
859 {
860 // printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName);
861 if( !strcmp(filters[i].longName, filterName)
862 || !strcmp(filters[i].shortName, filterName))
863 {
864 ppMode->lumMode &= ~filters[i].mask;
865 ppMode->chromMode &= ~filters[i].mask;
866
867 filterNameOk=1;
868 if(!enable) break; // user wants to disable it
869
870 if(q >= filters[i].minLumQuality && luma)
871 ppMode->lumMode|= filters[i].mask;
872 if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
873 if(q >= filters[i].minChromQuality)
874 ppMode->chromMode|= filters[i].mask;
875
876 if(filters[i].mask == LEVEL_FIX)
877 {
878 int o;
879 ppMode->minAllowedY= 16;
880 ppMode->maxAllowedY= 234;
881 for(o=0; options[o]!=NULL; o++)
882 {
883 if( !strcmp(options[o],"fullyrange")
884 ||!strcmp(options[o],"f"))
885 {
886 ppMode->minAllowedY= 0;
887 ppMode->maxAllowedY= 255;
888 numOfUnknownOptions--;
889 }
890 }
891 }
892 else if(filters[i].mask == TEMP_NOISE_FILTER)
893 {
894 int o;
895 int numOfNoises=0;
896
897 for(o=0; options[o]!=NULL; o++)
898 {
899 char *tail;
900 ppMode->maxTmpNoise[numOfNoises]=
901 strtol(options[o], &tail, 0);
902 if(tail!=options[o])
903 {
904 numOfNoises++;
905 numOfUnknownOptions--;
906 if(numOfNoises >= 3) break;
907 }
908 }
909 }
910 else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK
911 || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK)
912 {
913 int o;
914
915 for(o=0; options[o]!=NULL && o<2; o++)
916 {
917 char *tail;
918 int val= strtol(options[o], &tail, 0);
919 if(tail==options[o]) break;
920
921 numOfUnknownOptions--;
922 if(o==0) ppMode->baseDcDiff= val;
923 else ppMode->flatnessThreshold= val;
924 }
925 }
926 else if(filters[i].mask == FORCE_QUANT)
927 {
928 int o;
929 ppMode->forcedQuant= 15;
930
931 for(o=0; options[o]!=NULL && o<1; o++)
932 {
933 char *tail;
934 int val= strtol(options[o], &tail, 0);
935 if(tail==options[o]) break;
936
937 numOfUnknownOptions--;
938 ppMode->forcedQuant= val;
939 }
940 }
941 }
942 }
943 if(!filterNameOk) ppMode->error++;
944 ppMode->error += numOfUnknownOptions;
945 }
946
947 if(verbose>1) printf("pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode);
948 if(ppMode->error)
949 {
950 fprintf(stderr, "%d errors in postprocess string \"%s\"\n", ppMode->error, name);
951 av_free(ppMode);
952 return NULL;
953 }
954 return ppMode;
955 }
956
957 void pp_free_mode(pp_mode_t *mode){
958 av_free(mode);
959 }
960
961 static void reallocAlign(void **p, int alignment, int size){
962 av_free(*p);
963 *p= av_mallocz(size);
964 }
965
966 static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){
967 int mbWidth = (width+15)>>4;
968 int mbHeight= (height+15)>>4;
969 int i;
970
971 c->stride= stride;
972 c->qpStride= qpStride;
973
974 reallocAlign((void **)&c->tempDst, 8, stride*24);
975 reallocAlign((void **)&c->tempSrc, 8, stride*24);
976 reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
977 reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));
978 for(i=0; i<256; i++)
979 c->yHistogram[i]= width*height/64*15/256;
980
981 for(i=0; i<3; i++)
982 {
983 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end
984 reallocAlign((void **)&c->tempBlured[i], 8, stride*mbHeight*16 + 17*1024);
985 reallocAlign((void **)&c->tempBluredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
986 }
987
988 reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
989 reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
990 reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
991 reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
992 }
993
994 static void global_init(void){
995 int i;
996 memset(clip_table, 0, 256);
997 for(i=256; i<512; i++)
998 clip_table[i]= i;
999 memset(clip_table+512, 0, 256);
1000 }
1001
1002 pp_context_t *pp_get_context(int width, int height, int cpuCaps){
1003 PPContext *c= av_malloc(sizeof(PPContext));
1004 int stride= (width+15)&(~15); //assumed / will realloc if needed
1005 int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed
1006
1007 global_init();
1008
1009 memset(c, 0, sizeof(PPContext));
1010 c->cpuCaps= cpuCaps;
1011 if(cpuCaps&PP_FORMAT){
1012 c->hChromaSubSample= cpuCaps&0x3;
1013 c->vChromaSubSample= (cpuCaps>>4)&0x3;
1014 }else{
1015 c->hChromaSubSample= 1;
1016 c->vChromaSubSample= 1;
1017 }
1018
1019 reallocBuffers(c, width, height, stride, qpStride);
1020
1021 c->frameNum=-1;
1022
1023 return c;
1024 }
1025
1026 void pp_free_context(void *vc){
1027 PPContext *c = (PPContext*)vc;
1028 int i;
1029
1030 for(i=0; i<3; i++) av_free(c->tempBlured[i]);
1031 for(i=0; i<3; i++) av_free(c->tempBluredPast[i]);
1032
1033 av_free(c->tempBlocks);
1034 av_free(c->yHistogram);
1035 av_free(c->tempDst);
1036 av_free(c->tempSrc);
1037 av_free(c->deintTemp);
1038 av_free(c->stdQPTable);
1039 av_free(c->nonBQPTable);
1040 av_free(c->forcedQPTable);
1041
1042 memset(c, 0, sizeof(PPContext));
1043
1044 av_free(c);
1045 }
1046
1047 void pp_postprocess(uint8_t * src[3], int srcStride[3],
1048 uint8_t * dst[3], int dstStride[3],
1049 int width, int height,
1050 QP_STORE_T *QP_store, int QPStride,
1051 pp_mode_t *vm, void *vc, int pict_type)
1052 {
1053 int mbWidth = (width+15)>>4;
1054 int mbHeight= (height+15)>>4;
1055 PPMode *mode = (PPMode*)vm;
1056 PPContext *c = (PPContext*)vc;
1057 int minStride= MAX(ABS(srcStride[0]), ABS(dstStride[0]));
1058 int absQPStride = ABS(QPStride);
1059
1060 // c->stride and c->QPStride are always positive
1061 if(c->stride < minStride || c->qpStride < absQPStride)
1062 reallocBuffers(c, width, height,
1063 MAX(minStride, c->stride),
1064 MAX(c->qpStride, absQPStride));
1065
1066 if(QP_store==NULL || (mode->lumMode & FORCE_QUANT))
1067 {
1068 int i;
1069 QP_store= c->forcedQPTable;
1070 absQPStride = QPStride = 0;
1071 if(mode->lumMode & FORCE_QUANT)
1072 for(i=0; i<mbWidth; i++) QP_store[i]= mode->forcedQuant;
1073 else
1074 for(i=0; i<mbWidth; i++) QP_store[i]= 1;
1075 }
1076 //printf("pict_type:%d\n", pict_type);
1077
1078 if(pict_type & PP_PICT_TYPE_QP2){
1079 int i;
1080 const int count= mbHeight * absQPStride;
1081 for(i=0; i<(count>>2); i++){
1082 ((uint32_t*)c->stdQPTable)[i] = (((uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F;
1083 }
1084 for(i<<=2; i<count; i++){
1085 c->stdQPTable[i] = QP_store[i]>>1;
1086 }
1087 QP_store= c->stdQPTable;
1088 QPStride= absQPStride;
1089 }
1090
1091 if(0){
1092 int x,y;
1093 for(y=0; y<mbHeight; y++){
1094 for(x=0; x<mbWidth; x++){
1095 printf("%2d ", QP_store[x + y*QPStride]);
1096 }
1097 printf("\n");
1098 }
1099 printf("\n");
1100 }
1101
1102 if((pict_type&7)!=3)
1103 {
1104 if (QPStride >= 0) {
1105 int i;
1106 const int count= mbHeight * QPStride;
1107 for(i=0; i<(count>>2); i++){
1108 ((uint32_t*)c->nonBQPTable)[i] = ((uint32_t*)QP_store)[i] & 0x3F3F3F3F;
1109 }
1110 for(i<<=2; i<count; i++){
1111 c->nonBQPTable[i] = QP_store[i] & 0x3F;
1112 }
1113 } else {
1114 int i,j;
1115 for(i=0; i<mbHeight; i++) {
1116 for(j=0; j<absQPStride; j++) {
1117 c->nonBQPTable[i*absQPStride+j] = QP_store[i*QPStride+j] & 0x3F;
1118 }
1119 }
1120 }
1121 }
1122
1123 if(verbose>2)
1124 {
1125 printf("using npp filters 0x%X/0x%X\n", mode->lumMode, mode->chromMode);
1126 }
1127
1128 postProcess(src[0], srcStride[0], dst[0], dstStride[0],
1129 width, height, QP_store, QPStride, 0, mode, c);
1130
1131 width = (width )>>c->hChromaSubSample;
1132 height = (height)>>c->vChromaSubSample;
1133
1134 if(mode->chromMode)
1135 {
1136 postProcess(src[1], srcStride[1], dst[1], dstStride[1],
1137 width, height, QP_store, QPStride, 1, mode, c);
1138 postProcess(src[2], srcStride[2], dst[2], dstStride[2],
1139 width, height, QP_store, QPStride, 2, mode, c);
1140 }
1141 else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2])
1142 {
1143 linecpy(dst[1], src[1], height, srcStride[1]);
1144 linecpy(dst[2], src[2], height, srcStride[2]);
1145 }
1146 else
1147 {
1148 int y;
1149 for(y=0; y<height; y++)
1150 {
1151 memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
1152 memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
1153 }
1154 }
1155 }
1156