358656b76b264bbe120679a758975128f0608e83
[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
105 #define GET_MODE_BUFFER_SIZE 500
106 #define OPTIONS_ARRAY_SIZE 10
107 #define BLOCK_SIZE 8
108 #define TEMP_STRIDE 8
109 //#define NUM_BLOCKS_AT_ONCE 16 //not used yet
110
111 #if defined(ARCH_X86) || defined(ARCH_X86_64)
112 static uint64_t __attribute__((aligned(8))) attribute_used w05= 0x0005000500050005LL;
113 static uint64_t __attribute__((aligned(8))) attribute_used w04= 0x0004000400040004LL;
114 static uint64_t __attribute__((aligned(8))) attribute_used w20= 0x0020002000200020LL;
115 static uint64_t __attribute__((aligned(8))) attribute_used b00= 0x0000000000000000LL;
116 static uint64_t __attribute__((aligned(8))) attribute_used b01= 0x0101010101010101LL;
117 static uint64_t __attribute__((aligned(8))) attribute_used b02= 0x0202020202020202LL;
118 static uint64_t __attribute__((aligned(8))) attribute_used b08= 0x0808080808080808LL;
119 static uint64_t __attribute__((aligned(8))) attribute_used b80= 0x8080808080808080LL;
120 #endif
121
122 static uint8_t clip_table[3*256];
123 static uint8_t * const clip_tab= clip_table + 256;
124
125 static const int verbose= 0;
126
127 static const int attribute_used 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", "hdeblock:a,vdeblock:a,dering:a",
156 "de", "hdeblock:a,vdeblock:a,dering:a",
157 "fast", "x1hdeblock:a,x1vdeblock:a,dering:a",
158 "fa", "x1hdeblock:a,x1vdeblock:a,dering:a",
159 "ac", "ha:a:128:7,va:a,dering:a",
160 NULL //End Marker
161 };
162
163
164 #if defined(ARCH_X86) || defined(ARCH_X86_64)
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 cuz the MMX code is faster with vertical filters and transposing
195
196 /**
197 * Check if the given 8x8 Block is mostly "flat"
198 */
199 static inline int isHorizDC_C(uint8_t src[], int stride, PPContext *c)
200 {
201 int numEq= 0;
202 int y;
203 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
204 const int dcThreshold= dcOffset*2 + 1;
205
206 for(y=0; y<BLOCK_SIZE; y++)
207 {
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 int numEq= 0;
225 int y;
226 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
227 const int dcThreshold= dcOffset*2 + 1;
228
229 src+= stride*4; // src points to begin of the 8x8 Block
230 for(y=0; y<BLOCK_SIZE-1; y++)
231 {
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 {
276 if((unsigned)(src[ x + 0*stride] - src[ x + 5*stride] + 2*QP) > 4*QP) return 0;
277 if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0;
278 if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0;
279 if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0;
280 }
281 #else
282 int x;
283 src+= stride*3;
284 for(x=0; x<BLOCK_SIZE; x++)
285 {
286 if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
287 }
288 #endif
289 return 1;
290 #else
291 int x;
292 src+= stride*4;
293 for(x=0; x<BLOCK_SIZE; x++)
294 {
295 int min=255;
296 int max=0;
297 int y;
298 for(y=0; y<8; y++){
299 int v= src[x + y*stride];
300 if(v>max) max=v;
301 if(v<min) min=v;
302 }
303 if(max-min > 2*QP) return 0;
304 }
305 return 1;
306 #endif
307 }
308
309 static inline int horizClassify_C(uint8_t src[], int stride, PPContext *c){
310 if( isHorizDC_C(src, stride, c) ){
311 if( isHorizMinMaxOk_C(src, stride, c->QP) )
312 return 1;
313 else
314 return 0;
315 }else{
316 return 2;
317 }
318 }
319
320 static inline int vertClassify_C(uint8_t src[], int stride, PPContext *c){
321 if( isVertDC_C(src, stride, c) ){
322 if( isVertMinMaxOk_C(src, stride, c->QP) )
323 return 1;
324 else
325 return 0;
326 }else{
327 return 2;
328 }
329 }
330
331 static inline void doHorizDefFilter_C(uint8_t dst[], int stride, PPContext *c)
332 {
333 int y;
334 for(y=0; y<BLOCK_SIZE; y++)
335 {
336 const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]);
337
338 if(ABS(middleEnergy) < 8*c->QP)
339 {
340 const int q=(dst[3] - dst[4])/2;
341 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
342 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
343
344 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
345 d= MAX(d, 0);
346
347 d= (5*d + 32) >> 6;
348 d*= SIGN(-middleEnergy);
349
350 if(q>0)
351 {
352 d= d<0 ? 0 : d;
353 d= d>q ? q : d;
354 }
355 else
356 {
357 d= d>0 ? 0 : d;
358 d= d<q ? q : d;
359 }
360
361 dst[3]-= d;
362 dst[4]+= d;
363 }
364 dst+= stride;
365 }
366 }
367
368 /**
369 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
370 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
371 */
372 static inline void doHorizLowPass_C(uint8_t dst[], int stride, PPContext *c)
373 {
374 int y;
375 for(y=0; y<BLOCK_SIZE; y++)
376 {
377 const int first= ABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0];
378 const int last= ABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7];
379
380 int sums[10];
381 sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4;
382 sums[1] = sums[0] - first + dst[3];
383 sums[2] = sums[1] - first + dst[4];
384 sums[3] = sums[2] - first + dst[5];
385 sums[4] = sums[3] - first + dst[6];
386 sums[5] = sums[4] - dst[0] + dst[7];
387 sums[6] = sums[5] - dst[1] + last;
388 sums[7] = sums[6] - dst[2] + last;
389 sums[8] = sums[7] - dst[3] + last;
390 sums[9] = sums[8] - dst[4] + last;
391
392 dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4;
393 dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4;
394 dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4;
395 dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4;
396 dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4;
397 dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4;
398 dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4;
399 dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4;
400
401 dst+= stride;
402 }
403 }
404
405 /**
406 * Experimental Filter 1 (Horizontal)
407 * will not damage linear gradients
408 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
409 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
410 * MMX2 version does correct clipping C version doesnt
411 * not identical with the vertical one
412 */
413 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
414 {
415 int y;
416 static uint64_t *lut= NULL;
417 if(lut==NULL)
418 {
419 int i;
420 lut = av_malloc(256*8);
421 for(i=0; i<256; i++)
422 {
423 int v= i < 128 ? 2*i : 2*(i-256);
424 /*
425 //Simulate 112242211 9-Tap filter
426 uint64_t a= (v/16) & 0xFF;
427 uint64_t b= (v/8) & 0xFF;
428 uint64_t c= (v/4) & 0xFF;
429 uint64_t d= (3*v/8) & 0xFF;
430 */
431 //Simulate piecewise linear interpolation
432 uint64_t a= (v/16) & 0xFF;
433 uint64_t b= (v*3/16) & 0xFF;
434 uint64_t c= (v*5/16) & 0xFF;
435 uint64_t d= (7*v/16) & 0xFF;
436 uint64_t A= (0x100 - a)&0xFF;
437 uint64_t B= (0x100 - b)&0xFF;
438 uint64_t C= (0x100 - c)&0xFF;
439 uint64_t D= (0x100 - c)&0xFF;
440
441 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
442 (D<<24) | (C<<16) | (B<<8) | (A);
443 //lut[i] = (v<<32) | (v<<24);
444 }
445 }
446
447 for(y=0; y<BLOCK_SIZE; y++)
448 {
449 int a= src[1] - src[2];
450 int b= src[3] - src[4];
451 int c= src[5] - src[6];
452
453 int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);
454
455 if(d < QP)
456 {
457 int v = d * SIGN(-b);
458
459 src[1] +=v/8;
460 src[2] +=v/4;
461 src[3] +=3*v/8;
462 src[4] -=3*v/8;
463 src[5] -=v/4;
464 src[6] -=v/8;
465
466 }
467 src+=stride;
468 }
469 }
470
471 /**
472 * accurate deblock filter
473 */
474 static always_inline void do_a_deblock_C(uint8_t *src, int step, int stride, PPContext *c){
475 int y;
476 const int QP= c->QP;
477 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
478 const int dcThreshold= dcOffset*2 + 1;
479 //START_TIMER
480 src+= step*4; // src points to begin of the 8x8 Block
481 for(y=0; y<8; y++){
482 int numEq= 0;
483
484 if(((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold) numEq++;
485 if(((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold) numEq++;
486 if(((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold) numEq++;
487 if(((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold) numEq++;
488 if(((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold) numEq++;
489 if(((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold) numEq++;
490 if(((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold) numEq++;
491 if(((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold) numEq++;
492 if(((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold) numEq++;
493 if(numEq > c->ppMode.flatnessThreshold){
494 int min, max, x;
495
496 if(src[0] > src[step]){
497 max= src[0];
498 min= src[step];
499 }else{
500 max= src[step];
501 min= src[0];
502 }
503 for(x=2; x<8; x+=2){
504 if(src[x*step] > src[(x+1)*step]){
505 if(src[x *step] > max) max= src[ x *step];
506 if(src[(x+1)*step] < min) min= src[(x+1)*step];
507 }else{
508 if(src[(x+1)*step] > max) max= src[(x+1)*step];
509 if(src[ x *step] < min) min= src[ x *step];
510 }
511 }
512 if(max-min < 2*QP){
513 const int first= ABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0];
514 const int last= ABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step];
515
516 int sums[10];
517 sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4;
518 sums[1] = sums[0] - first + src[3*step];
519 sums[2] = sums[1] - first + src[4*step];
520 sums[3] = sums[2] - first + src[5*step];
521 sums[4] = sums[3] - first + src[6*step];
522 sums[5] = sums[4] - src[0*step] + src[7*step];
523 sums[6] = sums[5] - src[1*step] + last;
524 sums[7] = sums[6] - src[2*step] + last;
525 sums[8] = sums[7] - src[3*step] + last;
526 sums[9] = sums[8] - src[4*step] + last;
527
528 src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4;
529 src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4;
530 src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4;
531 src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4;
532 src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4;
533 src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4;
534 src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4;
535 src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4;
536 }
537 }else{
538 const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]);
539
540 if(ABS(middleEnergy) < 8*QP)
541 {
542 const int q=(src[3*step] - src[4*step])/2;
543 const int leftEnergy= 5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]);
544 const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]);
545
546 int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
547 d= MAX(d, 0);
548
549 d= (5*d + 32) >> 6;
550 d*= SIGN(-middleEnergy);
551
552 if(q>0)
553 {
554 d= d<0 ? 0 : d;
555 d= d>q ? q : d;
556 }
557 else
558 {
559 d= d>0 ? 0 : d;
560 d= d<q ? q : d;
561 }
562
563 src[3*step]-= d;
564 src[4*step]+= d;
565 }
566 }
567
568 src += stride;
569 }
570 /*if(step==16){
571 STOP_TIMER("step16")
572 }else{
573 STOP_TIMER("stepX")
574 }*/
575 }
576
577 //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
578 //Plain C versions
579 #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
580 #define COMPILE_C
581 #endif
582
583 #ifdef ARCH_POWERPC
584 #ifdef HAVE_ALTIVEC
585 #define COMPILE_ALTIVEC
586 #endif //HAVE_ALTIVEC
587 #endif //ARCH_POWERPC
588
589 #if defined(ARCH_X86) || defined(ARCH_X86_64)
590
591 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
592 #define COMPILE_MMX
593 #endif
594
595 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
596 #define COMPILE_MMX2
597 #endif
598
599 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
600 #define COMPILE_3DNOW
601 #endif
602 #endif //ARCH_X86
603
604 #undef HAVE_MMX
605 #undef HAVE_MMX2
606 #undef HAVE_3DNOW
607 #undef HAVE_ALTIVEC
608
609 #ifdef COMPILE_C
610 #undef HAVE_MMX
611 #undef HAVE_MMX2
612 #undef HAVE_3DNOW
613 #define RENAME(a) a ## _C
614 #include "postprocess_template.c"
615 #endif
616
617 #ifdef ARCH_POWERPC
618 #ifdef COMPILE_ALTIVEC
619 #undef RENAME
620 #define HAVE_ALTIVEC
621 #define RENAME(a) a ## _altivec
622 #include "postprocess_altivec_template.c"
623 #include "postprocess_template.c"
624 #endif
625 #endif //ARCH_POWERPC
626
627 //MMX versions
628 #ifdef COMPILE_MMX
629 #undef RENAME
630 #define HAVE_MMX
631 #undef HAVE_MMX2
632 #undef HAVE_3DNOW
633 #define RENAME(a) a ## _MMX
634 #include "postprocess_template.c"
635 #endif
636
637 //MMX2 versions
638 #ifdef COMPILE_MMX2
639 #undef RENAME
640 #define HAVE_MMX
641 #define HAVE_MMX2
642 #undef HAVE_3DNOW
643 #define RENAME(a) a ## _MMX2
644 #include "postprocess_template.c"
645 #endif
646
647 //3DNOW versions
648 #ifdef COMPILE_3DNOW
649 #undef RENAME
650 #define HAVE_MMX
651 #undef HAVE_MMX2
652 #define HAVE_3DNOW
653 #define RENAME(a) a ## _3DNow
654 #include "postprocess_template.c"
655 #endif
656
657 // minor note: the HAVE_xyz is messed up after that line so dont use it
658
659 static inline void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
660 QP_STORE_T QPs[], int QPStride, int isColor, pp_mode_t *vm, pp_context_t *vc)
661 {
662 PPContext *c= (PPContext *)vc;
663 PPMode *ppMode= (PPMode *)vm;
664 c->ppMode= *ppMode; //FIXME
665
666 // useing ifs here as they are faster than function pointers allthough the
667 // difference wouldnt be messureable here but its much better because
668 // someone might exchange the cpu whithout restarting mplayer ;)
669 #ifdef RUNTIME_CPUDETECT
670 #if defined(ARCH_X86) || defined(ARCH_X86_64)
671 // ordered per speed fasterst first
672 if(c->cpuCaps & PP_CPU_CAPS_MMX2)
673 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
674 else if(c->cpuCaps & PP_CPU_CAPS_3DNOW)
675 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
676 else if(c->cpuCaps & PP_CPU_CAPS_MMX)
677 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
678 else
679 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
680 #else
681 #ifdef ARCH_POWERPC
682 #ifdef HAVE_ALTIVEC
683 if(c->cpuCaps & PP_CPU_CAPS_ALTIVEC)
684 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
685 else
686 #endif
687 #endif
688 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
689 #endif
690 #else //RUNTIME_CPUDETECT
691 #ifdef HAVE_MMX2
692 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
693 #elif defined (HAVE_3DNOW)
694 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
695 #elif defined (HAVE_MMX)
696 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
697 #elif defined (HAVE_ALTIVEC)
698 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
699 #else
700 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
701 #endif
702 #endif //!RUNTIME_CPUDETECT
703 }
704
705 //static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
706 // QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
707
708 /* -pp Command line Help
709 */
710 char *pp_help=
711 "Available postprocessing filters:\n"
712 "Filters Options\n"
713 "short long name short long option Description\n"
714 "* * a autoq CPU power dependent enabler\n"
715 " c chrom chrominance filtering enabled\n"
716 " y nochrom chrominance filtering disabled\n"
717 " n noluma luma filtering disabled\n"
718 "hb hdeblock (2 threshold) horizontal deblocking filter\n"
719 " 1. difference factor: default=32, higher -> more deblocking\n"
720 " 2. flatness threshold: default=39, lower -> more deblocking\n"
721 " the h & v deblocking filters share these\n"
722 " so you can't set different thresholds for h / v\n"
723 "vb vdeblock (2 threshold) vertical deblocking filter\n"
724 "ha hadeblock (2 threshold) horizontal deblocking filter\n"
725 "va vadeblock (2 threshold) vertical deblocking filter\n"
726 "h1 x1hdeblock experimental h deblock filter 1\n"
727 "v1 x1vdeblock experimental v deblock filter 1\n"
728 "dr dering deringing filter\n"
729 "al autolevels automatic brightness / contrast\n"
730 " f fullyrange stretch luminance to (0..255)\n"
731 "lb linblenddeint linear blend deinterlacer\n"
732 "li linipoldeint linear interpolating deinterlace\n"
733 "ci cubicipoldeint cubic interpolating deinterlacer\n"
734 "md mediandeint median deinterlacer\n"
735 "fd ffmpegdeint ffmpeg deinterlacer\n"
736 "l5 lowpass5 FIR lowpass deinterlacer\n"
737 "de default hb:a,vb:a,dr:a\n"
738 "fa fast h1:a,v1:a,dr:a\n"
739 "ac ha:a:128:7,va:a,dr:a\n"
740 "tn tmpnoise (3 threshold) temporal noise reducer\n"
741 " 1. <= 2. <= 3. larger -> stronger filtering\n"
742 "fq forceQuant <quantizer> force quantizer\n"
743 "Usage:\n"
744 "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n"
745 "long form example:\n"
746 "vdeblock:autoq/hdeblock:autoq/linblenddeint default,-vdeblock\n"
747 "short form example:\n"
748 "vb:a/hb:a/lb de,-vb\n"
749 "more examples:\n"
750 "tn:64:128:256\n"
751 "\n"
752 ;
753
754 pp_mode_t *pp_get_mode_by_name_and_quality(char *name, int quality)
755 {
756 char temp[GET_MODE_BUFFER_SIZE];
757 char *p= temp;
758 const char *filterDelimiters= ",/";
759 const char *optionDelimiters= ":";
760 struct PPMode *ppMode;
761 char *filterToken;
762
763 ppMode= av_malloc(sizeof(PPMode));
764
765 ppMode->lumMode= 0;
766 ppMode->chromMode= 0;
767 ppMode->maxTmpNoise[0]= 700;
768 ppMode->maxTmpNoise[1]= 1500;
769 ppMode->maxTmpNoise[2]= 3000;
770 ppMode->maxAllowedY= 234;
771 ppMode->minAllowedY= 16;
772 ppMode->baseDcDiff= 256/8;
773 ppMode->flatnessThreshold= 56-16-1;
774 ppMode->maxClippedThreshold= 0.01;
775 ppMode->error=0;
776
777 strncpy(temp, name, GET_MODE_BUFFER_SIZE);
778
779 if(verbose>1) printf("pp: %s\n", name);
780
781 for(;;){
782 char *filterName;
783 int q= 1000000; //PP_QUALITY_MAX;
784 int chrom=-1;
785 int luma=-1;
786 char *option;
787 char *options[OPTIONS_ARRAY_SIZE];
788 int i;
789 int filterNameOk=0;
790 int numOfUnknownOptions=0;
791 int enable=1; //does the user want us to enabled or disabled the filter
792
793 filterToken= strtok(p, filterDelimiters);
794 if(filterToken == NULL) break;
795 p+= strlen(filterToken) + 1; // p points to next filterToken
796 filterName= strtok(filterToken, optionDelimiters);
797 if(verbose>1) printf("pp: %s::%s\n", filterToken, filterName);
798
799 if(*filterName == '-')
800 {
801 enable=0;
802 filterName++;
803 }
804
805 for(;;){ //for all options
806 option= strtok(NULL, optionDelimiters);
807 if(option == NULL) break;
808
809 if(verbose>1) printf("pp: option: %s\n", option);
810 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
811 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
812 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
813 else if(!strcmp("noluma", option) || !strcmp("n", option)) luma=0;
814 else
815 {
816 options[numOfUnknownOptions] = option;
817 numOfUnknownOptions++;
818 }
819 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
820 }
821 options[numOfUnknownOptions] = NULL;
822
823 /* replace stuff from the replace Table */
824 for(i=0; replaceTable[2*i]!=NULL; i++)
825 {
826 if(!strcmp(replaceTable[2*i], filterName))
827 {
828 int newlen= strlen(replaceTable[2*i + 1]);
829 int plen;
830 int spaceLeft;
831
832 if(p==NULL) p= temp, *p=0; //last filter
833 else p--, *p=','; //not last filter
834
835 plen= strlen(p);
836 spaceLeft= p - temp + plen;
837 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE)
838 {
839 ppMode->error++;
840 break;
841 }
842 memmove(p + newlen, p, plen+1);
843 memcpy(p, replaceTable[2*i + 1], newlen);
844 filterNameOk=1;
845 }
846 }
847
848 for(i=0; filters[i].shortName!=NULL; i++)
849 {
850 // printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName);
851 if( !strcmp(filters[i].longName, filterName)
852 || !strcmp(filters[i].shortName, filterName))
853 {
854 ppMode->lumMode &= ~filters[i].mask;
855 ppMode->chromMode &= ~filters[i].mask;
856
857 filterNameOk=1;
858 if(!enable) break; // user wants to disable it
859
860 if(q >= filters[i].minLumQuality && luma)
861 ppMode->lumMode|= filters[i].mask;
862 if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
863 if(q >= filters[i].minChromQuality)
864 ppMode->chromMode|= filters[i].mask;
865
866 if(filters[i].mask == LEVEL_FIX)
867 {
868 int o;
869 ppMode->minAllowedY= 16;
870 ppMode->maxAllowedY= 234;
871 for(o=0; options[o]!=NULL; o++)
872 {
873 if( !strcmp(options[o],"fullyrange")
874 ||!strcmp(options[o],"f"))
875 {
876 ppMode->minAllowedY= 0;
877 ppMode->maxAllowedY= 255;
878 numOfUnknownOptions--;
879 }
880 }
881 }
882 else if(filters[i].mask == TEMP_NOISE_FILTER)
883 {
884 int o;
885 int numOfNoises=0;
886
887 for(o=0; options[o]!=NULL; o++)
888 {
889 char *tail;
890 ppMode->maxTmpNoise[numOfNoises]=
891 strtol(options[o], &tail, 0);
892 if(tail!=options[o])
893 {
894 numOfNoises++;
895 numOfUnknownOptions--;
896 if(numOfNoises >= 3) break;
897 }
898 }
899 }
900 else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK
901 || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK)
902 {
903 int o;
904
905 for(o=0; options[o]!=NULL && o<2; o++)
906 {
907 char *tail;
908 int val= strtol(options[o], &tail, 0);
909 if(tail==options[o]) break;
910
911 numOfUnknownOptions--;
912 if(o==0) ppMode->baseDcDiff= val;
913 else ppMode->flatnessThreshold= val;
914 }
915 }
916 else if(filters[i].mask == FORCE_QUANT)
917 {
918 int o;
919 ppMode->forcedQuant= 15;
920
921 for(o=0; options[o]!=NULL && o<1; o++)
922 {
923 char *tail;
924 int val= strtol(options[o], &tail, 0);
925 if(tail==options[o]) break;
926
927 numOfUnknownOptions--;
928 ppMode->forcedQuant= val;
929 }
930 }
931 }
932 }
933 if(!filterNameOk) ppMode->error++;
934 ppMode->error += numOfUnknownOptions;
935 }
936
937 if(verbose>1) printf("pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode);
938 if(ppMode->error)
939 {
940 fprintf(stderr, "%d errors in postprocess string \"%s\"\n", ppMode->error, name);
941 av_free(ppMode);
942 return NULL;
943 }
944 return ppMode;
945 }
946
947 void pp_free_mode(pp_mode_t *mode){
948 av_free(mode);
949 }
950
951 static void reallocAlign(void **p, int alignment, int size){
952 av_free(*p);
953 *p= av_mallocz(size);
954 }
955
956 static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){
957 int mbWidth = (width+15)>>4;
958 int mbHeight= (height+15)>>4;
959 int i;
960
961 c->stride= stride;
962 c->qpStride= qpStride;
963
964 reallocAlign((void **)&c->tempDst, 8, stride*24);
965 reallocAlign((void **)&c->tempSrc, 8, stride*24);
966 reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
967 reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));
968 for(i=0; i<256; i++)
969 c->yHistogram[i]= width*height/64*15/256;
970
971 for(i=0; i<3; i++)
972 {
973 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end
974 reallocAlign((void **)&c->tempBlured[i], 8, stride*mbHeight*16 + 17*1024);
975 reallocAlign((void **)&c->tempBluredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
976 }
977
978 reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
979 reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
980 reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
981 reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
982 }
983
984 static void global_init(void){
985 int i;
986 memset(clip_table, 0, 256);
987 for(i=256; i<512; i++)
988 clip_table[i]= i;
989 memset(clip_table+512, 0, 256);
990 }
991
992 pp_context_t *pp_get_context(int width, int height, int cpuCaps){
993 PPContext *c= av_malloc(sizeof(PPContext));
994 int stride= (width+15)&(~15); //assumed / will realloc if needed
995 int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed
996
997 global_init();
998
999 memset(c, 0, sizeof(PPContext));
1000 c->cpuCaps= cpuCaps;
1001 if(cpuCaps&PP_FORMAT){
1002 c->hChromaSubSample= cpuCaps&0x3;
1003 c->vChromaSubSample= (cpuCaps>>4)&0x3;
1004 }else{
1005 c->hChromaSubSample= 1;
1006 c->vChromaSubSample= 1;
1007 }
1008
1009 reallocBuffers(c, width, height, stride, qpStride);
1010
1011 c->frameNum=-1;
1012
1013 return c;
1014 }
1015
1016 void pp_free_context(void *vc){
1017 PPContext *c = (PPContext*)vc;
1018 int i;
1019
1020 for(i=0; i<3; i++) av_free(c->tempBlured[i]);
1021 for(i=0; i<3; i++) av_free(c->tempBluredPast[i]);
1022
1023 av_free(c->tempBlocks);
1024 av_free(c->yHistogram);
1025 av_free(c->tempDst);
1026 av_free(c->tempSrc);
1027 av_free(c->deintTemp);
1028 av_free(c->stdQPTable);
1029 av_free(c->nonBQPTable);
1030 av_free(c->forcedQPTable);
1031
1032 memset(c, 0, sizeof(PPContext));
1033
1034 av_free(c);
1035 }
1036
1037 void pp_postprocess(uint8_t * src[3], int srcStride[3],
1038 uint8_t * dst[3], int dstStride[3],
1039 int width, int height,
1040 QP_STORE_T *QP_store, int QPStride,
1041 pp_mode_t *vm, void *vc, int pict_type)
1042 {
1043 int mbWidth = (width+15)>>4;
1044 int mbHeight= (height+15)>>4;
1045 PPMode *mode = (PPMode*)vm;
1046 PPContext *c = (PPContext*)vc;
1047 int minStride= MAX(ABS(srcStride[0]), ABS(dstStride[0]));
1048 int absQPStride = ABS(QPStride);
1049
1050 // c->stride and c->QPStride are always positive
1051 if(c->stride < minStride || c->qpStride < absQPStride)
1052 reallocBuffers(c, width, height,
1053 MAX(minStride, c->stride),
1054 MAX(c->qpStride, absQPStride));
1055
1056 if(QP_store==NULL || (mode->lumMode & FORCE_QUANT))
1057 {
1058 int i;
1059 QP_store= c->forcedQPTable;
1060 absQPStride = QPStride = 0;
1061 if(mode->lumMode & FORCE_QUANT)
1062 for(i=0; i<mbWidth; i++) QP_store[i]= mode->forcedQuant;
1063 else
1064 for(i=0; i<mbWidth; i++) QP_store[i]= 1;
1065 }
1066 //printf("pict_type:%d\n", pict_type);
1067
1068 if(pict_type & PP_PICT_TYPE_QP2){
1069 int i;
1070 const int count= mbHeight * absQPStride;
1071 for(i=0; i<(count>>2); i++){
1072 ((uint32_t*)c->stdQPTable)[i] = (((uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F;
1073 }
1074 for(i<<=2; i<count; i++){
1075 c->stdQPTable[i] = QP_store[i]>>1;
1076 }
1077 QP_store= c->stdQPTable;
1078 QPStride= absQPStride;
1079 }
1080
1081 if(0){
1082 int x,y;
1083 for(y=0; y<mbHeight; y++){
1084 for(x=0; x<mbWidth; x++){
1085 printf("%2d ", QP_store[x + y*QPStride]);
1086 }
1087 printf("\n");
1088 }
1089 printf("\n");
1090 }
1091
1092 if((pict_type&7)!=3)
1093 {
1094 if (QPStride >= 0) {
1095 int i;
1096 const int count= mbHeight * QPStride;
1097 for(i=0; i<(count>>2); i++){
1098 ((uint32_t*)c->nonBQPTable)[i] = ((uint32_t*)QP_store)[i] & 0x3F3F3F3F;
1099 }
1100 for(i<<=2; i<count; i++){
1101 c->nonBQPTable[i] = QP_store[i] & 0x3F;
1102 }
1103 } else {
1104 int i,j;
1105 for(i=0; i<mbHeight; i++) {
1106 for(j=0; j<absQPStride; j++) {
1107 c->nonBQPTable[i*absQPStride+j] = QP_store[i*QPStride+j] & 0x3F;
1108 }
1109 }
1110 }
1111 }
1112
1113 if(verbose>2)
1114 {
1115 printf("using npp filters 0x%X/0x%X\n", mode->lumMode, mode->chromMode);
1116 }
1117
1118 postProcess(src[0], srcStride[0], dst[0], dstStride[0],
1119 width, height, QP_store, QPStride, 0, mode, c);
1120
1121 width = (width )>>c->hChromaSubSample;
1122 height = (height)>>c->vChromaSubSample;
1123
1124 if(mode->chromMode)
1125 {
1126 postProcess(src[1], srcStride[1], dst[1], dstStride[1],
1127 width, height, QP_store, QPStride, 1, mode, c);
1128 postProcess(src[2], srcStride[2], dst[2], dstStride[2],
1129 width, height, QP_store, QPStride, 2, mode, c);
1130 }
1131 else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2])
1132 {
1133 linecpy(dst[1], src[1], height, srcStride[1]);
1134 linecpy(dst[2], src[2], height, srcStride[2]);
1135 }
1136 else
1137 {
1138 int y;
1139 for(y=0; y<height; y++)
1140 {
1141 memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
1142 memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
1143 }
1144 }
1145 }
1146