?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Print just [no] or [yes] in the help output instead of [default=no/yes].
[FFMpeg-mirror/ffmpeg-vdpau.git] / libpostproc / postprocess.c
blob56de4f1be31d4dfe0323c9e3025e7280f64b1a98
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 do not have a 3DNow! CPU -> it is untested, but no one said it does not work so it seems to work
51 # more or less selfinvented filters so the exactness is not too meaningful
52 E = Exact implementation
53 e = almost exact implementation (slightly different rounding,...)
54 a = alternative / approximate impl
55 c = checked against the other implementations (-vo md5)
56 p = partially optimized, still some work to do
57 */
58
59 /*
60 TODO:
61 reduce the time wasted on the mem transfer
62 unroll stuff if instructions depend too much on the prior one
63 move YScale thing to the end instead of fixing QP
64 write a faster and higher quality deblocking filter :)
65 make the mainloop more flexible (variable number of blocks at once
66 (the if/else stuff per block is slowing things down)
67 compare the quality & speed of all filters
68 split this huge file
69 optimize c versions
70 try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks
71 ...
72 */
73
74 //Changelog: use the Subversion log
75
76 #include "config.h"
77 #include "libavutil/avutil.h"
78 #include <inttypes.h>
79 #include <stdio.h>
80 #include <stdlib.h>
81 #include <string.h>
82 #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 #include "postprocess.h"
91 #include "postprocess_internal.h"
92
93 unsigned postproc_version(void)
94 {
95 return LIBPOSTPROC_VERSION_INT;
96 }
97
98 #ifdef HAVE_ALTIVEC_H
99 #include <altivec.h>
100 #endif
101
102 #define GET_MODE_BUFFER_SIZE 500
103 #define OPTIONS_ARRAY_SIZE 10
104 #define BLOCK_SIZE 8
105 #define TEMP_STRIDE 8
106 //#define NUM_BLOCKS_AT_ONCE 16 //not used yet
107
108 #if defined(ARCH_X86)
109 DECLARE_ASM_CONST(8, uint64_t, w05)= 0x0005000500050005LL;
110 DECLARE_ASM_CONST(8, uint64_t, w04)= 0x0004000400040004LL;
111 DECLARE_ASM_CONST(8, uint64_t, w20)= 0x0020002000200020LL;
112 DECLARE_ASM_CONST(8, uint64_t, b00)= 0x0000000000000000LL;
113 DECLARE_ASM_CONST(8, uint64_t, b01)= 0x0101010101010101LL;
114 DECLARE_ASM_CONST(8, uint64_t, b02)= 0x0202020202020202LL;
115 DECLARE_ASM_CONST(8, uint64_t, b08)= 0x0808080808080808LL;
116 DECLARE_ASM_CONST(8, uint64_t, b80)= 0x8080808080808080LL;
117 #endif
118
119 DECLARE_ASM_CONST(8, int, deringThreshold)= 20;
120
121
122 static struct PPFilter filters[]=
123 {
124 {"hb", "hdeblock", 1, 1, 3, H_DEBLOCK},
125 {"vb", "vdeblock", 1, 2, 4, V_DEBLOCK},
126 /* {"hr", "rkhdeblock", 1, 1, 3, H_RK1_FILTER},
127 {"vr", "rkvdeblock", 1, 2, 4, V_RK1_FILTER},*/
128 {"h1", "x1hdeblock", 1, 1, 3, H_X1_FILTER},
129 {"v1", "x1vdeblock", 1, 2, 4, V_X1_FILTER},
130 {"ha", "ahdeblock", 1, 1, 3, H_A_DEBLOCK},
131 {"va", "avdeblock", 1, 2, 4, V_A_DEBLOCK},
132 {"dr", "dering", 1, 5, 6, DERING},
133 {"al", "autolevels", 0, 1, 2, LEVEL_FIX},
134 {"lb", "linblenddeint", 1, 1, 4, LINEAR_BLEND_DEINT_FILTER},
135 {"li", "linipoldeint", 1, 1, 4, LINEAR_IPOL_DEINT_FILTER},
136 {"ci", "cubicipoldeint", 1, 1, 4, CUBIC_IPOL_DEINT_FILTER},
137 {"md", "mediandeint", 1, 1, 4, MEDIAN_DEINT_FILTER},
138 {"fd", "ffmpegdeint", 1, 1, 4, FFMPEG_DEINT_FILTER},
139 {"l5", "lowpass5", 1, 1, 4, LOWPASS5_DEINT_FILTER},
140 {"tn", "tmpnoise", 1, 7, 8, TEMP_NOISE_FILTER},
141 {"fq", "forcequant", 1, 0, 0, FORCE_QUANT},
142 {NULL, NULL,0,0,0,0} //End Marker
143 };
144
145 static const char *replaceTable[]=
146 {
147 "default", "hb:a,vb:a,dr:a",
148 "de", "hb:a,vb:a,dr:a",
149 "fast", "h1:a,v1:a,dr:a",
150 "fa", "h1:a,v1:a,dr:a",
151 "ac", "ha:a:128:7,va:a,dr:a",
152 NULL //End Marker
153 };
154
155
156 #if defined(ARCH_X86)
157 static inline void prefetchnta(void *p)
158 {
159 __asm__ volatile( "prefetchnta (%0)\n\t"
160 : : "r" (p)
161 );
162 }
163
164 static inline void prefetcht0(void *p)
165 {
166 __asm__ volatile( "prefetcht0 (%0)\n\t"
167 : : "r" (p)
168 );
169 }
170
171 static inline void prefetcht1(void *p)
172 {
173 __asm__ volatile( "prefetcht1 (%0)\n\t"
174 : : "r" (p)
175 );
176 }
177
178 static inline void prefetcht2(void *p)
179 {
180 __asm__ volatile( "prefetcht2 (%0)\n\t"
181 : : "r" (p)
182 );
183 }
184 #endif
185
186 /* The horizontal functions exist only in C because the MMX
187 * code is faster with vertical filters and transposing. */
188
189 /**
190 * Check if the given 8x8 Block is mostly "flat"
191 */
192 static inline int isHorizDC_C(uint8_t src[], int stride, PPContext *c)
193 {
194 int numEq= 0;
195 int y;
196 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
197 const int dcThreshold= dcOffset*2 + 1;
198
199 for(y=0; y<BLOCK_SIZE; y++){
200 if(((unsigned)(src[0] - src[1] + dcOffset)) < dcThreshold) numEq++;
201 if(((unsigned)(src[1] - src[2] + dcOffset)) < dcThreshold) numEq++;
202 if(((unsigned)(src[2] - src[3] + dcOffset)) < dcThreshold) numEq++;
203 if(((unsigned)(src[3] - src[4] + dcOffset)) < dcThreshold) numEq++;
204 if(((unsigned)(src[4] - src[5] + dcOffset)) < dcThreshold) numEq++;
205 if(((unsigned)(src[5] - src[6] + dcOffset)) < dcThreshold) numEq++;
206 if(((unsigned)(src[6] - src[7] + dcOffset)) < dcThreshold) numEq++;
207 src+= stride;
208 }
209 return numEq > c->ppMode.flatnessThreshold;
210 }
211
212 /**
213 * Check if the middle 8x8 Block in the given 8x16 block is flat
214 */
215 static inline int isVertDC_C(uint8_t src[], int stride, PPContext *c)
216 {
217 int numEq= 0;
218 int y;
219 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
220 const int dcThreshold= dcOffset*2 + 1;
221
222 src+= stride*4; // src points to begin of the 8x8 Block
223 for(y=0; y<BLOCK_SIZE-1; y++){
224 if(((unsigned)(src[0] - src[0+stride] + dcOffset)) < dcThreshold) numEq++;
225 if(((unsigned)(src[1] - src[1+stride] + dcOffset)) < dcThreshold) numEq++;
226 if(((unsigned)(src[2] - src[2+stride] + dcOffset)) < dcThreshold) numEq++;
227 if(((unsigned)(src[3] - src[3+stride] + dcOffset)) < dcThreshold) numEq++;
228 if(((unsigned)(src[4] - src[4+stride] + dcOffset)) < dcThreshold) numEq++;
229 if(((unsigned)(src[5] - src[5+stride] + dcOffset)) < dcThreshold) numEq++;
230 if(((unsigned)(src[6] - src[6+stride] + dcOffset)) < dcThreshold) numEq++;
231 if(((unsigned)(src[7] - src[7+stride] + dcOffset)) < dcThreshold) numEq++;
232 src+= stride;
233 }
234 return numEq > c->ppMode.flatnessThreshold;
235 }
236
237 static inline int isHorizMinMaxOk_C(uint8_t src[], int stride, int QP)
238 {
239 int i;
240 #if 1
241 for(i=0; i<2; i++){
242 if((unsigned)(src[0] - src[5] + 2*QP) > 4*QP) return 0;
243 src += stride;
244 if((unsigned)(src[2] - src[7] + 2*QP) > 4*QP) return 0;
245 src += stride;
246 if((unsigned)(src[4] - src[1] + 2*QP) > 4*QP) return 0;
247 src += stride;
248 if((unsigned)(src[6] - src[3] + 2*QP) > 4*QP) return 0;
249 src += stride;
250 }
251 #else
252 for(i=0; i<8; i++){
253 if((unsigned)(src[0] - src[7] + 2*QP) > 4*QP) return 0;
254 src += stride;
255 }
256 #endif
257 return 1;
258 }
259
260 static inline int isVertMinMaxOk_C(uint8_t src[], int stride, int QP)
261 {
262 #if 1
263 #if 1
264 int x;
265 src+= stride*4;
266 for(x=0; x<BLOCK_SIZE; x+=4){
267 if((unsigned)(src[ x + 0*stride] - src[ x + 5*stride] + 2*QP) > 4*QP) return 0;
268 if((unsigned)(src[1+x + 2*stride] - src[1+x + 7*stride] + 2*QP) > 4*QP) return 0;
269 if((unsigned)(src[2+x + 4*stride] - src[2+x + 1*stride] + 2*QP) > 4*QP) return 0;
270 if((unsigned)(src[3+x + 6*stride] - src[3+x + 3*stride] + 2*QP) > 4*QP) return 0;
271 }
272 #else
273 int x;
274 src+= stride*3;
275 for(x=0; x<BLOCK_SIZE; x++){
276 if((unsigned)(src[x + stride] - src[x + (stride<<3)] + 2*QP) > 4*QP) return 0;
277 }
278 #endif
279 return 1;
280 #else
281 int x;
282 src+= stride*4;
283 for(x=0; x<BLOCK_SIZE; x++){
284 int min=255;
285 int max=0;
286 int y;
287 for(y=0; y<8; y++){
288 int v= src[x + y*stride];
289 if(v>max) max=v;
290 if(v<min) min=v;
291 }
292 if(max-min > 2*QP) return 0;
293 }
294 return 1;
295 #endif
296 }
297
298 static inline int horizClassify_C(uint8_t src[], int stride, PPContext *c)
299 {
300 if( isHorizDC_C(src, stride, c) ){
301 if( isHorizMinMaxOk_C(src, stride, c->QP) )
302 return 1;
303 else
304 return 0;
305 }else{
306 return 2;
307 }
308 }
309
310 static inline int vertClassify_C(uint8_t src[], int stride, PPContext *c)
311 {
312 if( isVertDC_C(src, stride, c) ){
313 if( isVertMinMaxOk_C(src, stride, c->QP) )
314 return 1;
315 else
316 return 0;
317 }else{
318 return 2;
319 }
320 }
321
322 static inline void doHorizDefFilter_C(uint8_t dst[], int stride, PPContext *c)
323 {
324 int y;
325 for(y=0; y<BLOCK_SIZE; y++){
326 const int middleEnergy= 5*(dst[4] - dst[3]) + 2*(dst[2] - dst[5]);
327
328 if(FFABS(middleEnergy) < 8*c->QP){
329 const int q=(dst[3] - dst[4])/2;
330 const int leftEnergy= 5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);
331 const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);
332
333 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
334 d= FFMAX(d, 0);
335
336 d= (5*d + 32) >> 6;
337 d*= FFSIGN(-middleEnergy);
338
339 if(q>0)
340 {
341 d= d<0 ? 0 : d;
342 d= d>q ? q : d;
343 }
344 else
345 {
346 d= d>0 ? 0 : d;
347 d= d<q ? q : d;
348 }
349
350 dst[3]-= d;
351 dst[4]+= d;
352 }
353 dst+= stride;
354 }
355 }
356
357 /**
358 * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)
359 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)
360 */
361 static inline void doHorizLowPass_C(uint8_t dst[], int stride, PPContext *c)
362 {
363 int y;
364 for(y=0; y<BLOCK_SIZE; y++){
365 const int first= FFABS(dst[-1] - dst[0]) < c->QP ? dst[-1] : dst[0];
366 const int last= FFABS(dst[8] - dst[7]) < c->QP ? dst[8] : dst[7];
367
368 int sums[10];
369 sums[0] = 4*first + dst[0] + dst[1] + dst[2] + 4;
370 sums[1] = sums[0] - first + dst[3];
371 sums[2] = sums[1] - first + dst[4];
372 sums[3] = sums[2] - first + dst[5];
373 sums[4] = sums[3] - first + dst[6];
374 sums[5] = sums[4] - dst[0] + dst[7];
375 sums[6] = sums[5] - dst[1] + last;
376 sums[7] = sums[6] - dst[2] + last;
377 sums[8] = sums[7] - dst[3] + last;
378 sums[9] = sums[8] - dst[4] + last;
379
380 dst[0]= (sums[0] + sums[2] + 2*dst[0])>>4;
381 dst[1]= (sums[1] + sums[3] + 2*dst[1])>>4;
382 dst[2]= (sums[2] + sums[4] + 2*dst[2])>>4;
383 dst[3]= (sums[3] + sums[5] + 2*dst[3])>>4;
384 dst[4]= (sums[4] + sums[6] + 2*dst[4])>>4;
385 dst[5]= (sums[5] + sums[7] + 2*dst[5])>>4;
386 dst[6]= (sums[6] + sums[8] + 2*dst[6])>>4;
387 dst[7]= (sums[7] + sums[9] + 2*dst[7])>>4;
388
389 dst+= stride;
390 }
391 }
392
393 /**
394 * Experimental Filter 1 (Horizontal)
395 * will not damage linear gradients
396 * Flat blocks should look like they were passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
397 * can only smooth blocks at the expected locations (it cannot smooth them if they did move)
398 * MMX2 version does correct clipping C version does not
399 * not identical with the vertical one
400 */
401 static inline void horizX1Filter(uint8_t *src, int stride, int QP)
402 {
403 int y;
404 static uint64_t *lut= NULL;
405 if(lut==NULL)
406 {
407 int i;
408 lut = av_malloc(256*8);
409 for(i=0; i<256; i++)
410 {
411 int v= i < 128 ? 2*i : 2*(i-256);
412 /*
413 //Simulate 112242211 9-Tap filter
414 uint64_t a= (v/16) & 0xFF;
415 uint64_t b= (v/8) & 0xFF;
416 uint64_t c= (v/4) & 0xFF;
417 uint64_t d= (3*v/8) & 0xFF;
418 */
419 //Simulate piecewise linear interpolation
420 uint64_t a= (v/16) & 0xFF;
421 uint64_t b= (v*3/16) & 0xFF;
422 uint64_t c= (v*5/16) & 0xFF;
423 uint64_t d= (7*v/16) & 0xFF;
424 uint64_t A= (0x100 - a)&0xFF;
425 uint64_t B= (0x100 - b)&0xFF;
426 uint64_t C= (0x100 - c)&0xFF;
427 uint64_t D= (0x100 - c)&0xFF;
428
429 lut[i] = (a<<56) | (b<<48) | (c<<40) | (d<<32) |
430 (D<<24) | (C<<16) | (B<<8) | (A);
431 //lut[i] = (v<<32) | (v<<24);
432 }
433 }
434
435 for(y=0; y<BLOCK_SIZE; y++){
436 int a= src[1] - src[2];
437 int b= src[3] - src[4];
438 int c= src[5] - src[6];
439
440 int d= FFMAX(FFABS(b) - (FFABS(a) + FFABS(c))/2, 0);
441
442 if(d < QP){
443 int v = d * FFSIGN(-b);
444
445 src[1] +=v/8;
446 src[2] +=v/4;
447 src[3] +=3*v/8;
448 src[4] -=3*v/8;
449 src[5] -=v/4;
450 src[6] -=v/8;
451 }
452 src+=stride;
453 }
454 }
455
456 /**
457 * accurate deblock filter
458 */
459 static av_always_inline void do_a_deblock_C(uint8_t *src, int step, int stride, PPContext *c){
460 int y;
461 const int QP= c->QP;
462 const int dcOffset= ((c->nonBQP*c->ppMode.baseDcDiff)>>8) + 1;
463 const int dcThreshold= dcOffset*2 + 1;
464 //START_TIMER
465 src+= step*4; // src points to begin of the 8x8 Block
466 for(y=0; y<8; y++){
467 int numEq= 0;
468
469 if(((unsigned)(src[-1*step] - src[0*step] + dcOffset)) < dcThreshold) numEq++;
470 if(((unsigned)(src[ 0*step] - src[1*step] + dcOffset)) < dcThreshold) numEq++;
471 if(((unsigned)(src[ 1*step] - src[2*step] + dcOffset)) < dcThreshold) numEq++;
472 if(((unsigned)(src[ 2*step] - src[3*step] + dcOffset)) < dcThreshold) numEq++;
473 if(((unsigned)(src[ 3*step] - src[4*step] + dcOffset)) < dcThreshold) numEq++;
474 if(((unsigned)(src[ 4*step] - src[5*step] + dcOffset)) < dcThreshold) numEq++;
475 if(((unsigned)(src[ 5*step] - src[6*step] + dcOffset)) < dcThreshold) numEq++;
476 if(((unsigned)(src[ 6*step] - src[7*step] + dcOffset)) < dcThreshold) numEq++;
477 if(((unsigned)(src[ 7*step] - src[8*step] + dcOffset)) < dcThreshold) numEq++;
478 if(numEq > c->ppMode.flatnessThreshold){
479 int min, max, x;
480
481 if(src[0] > src[step]){
482 max= src[0];
483 min= src[step];
484 }else{
485 max= src[step];
486 min= src[0];
487 }
488 for(x=2; x<8; x+=2){
489 if(src[x*step] > src[(x+1)*step]){
490 if(src[x *step] > max) max= src[ x *step];
491 if(src[(x+1)*step] < min) min= src[(x+1)*step];
492 }else{
493 if(src[(x+1)*step] > max) max= src[(x+1)*step];
494 if(src[ x *step] < min) min= src[ x *step];
495 }
496 }
497 if(max-min < 2*QP){
498 const int first= FFABS(src[-1*step] - src[0]) < QP ? src[-1*step] : src[0];
499 const int last= FFABS(src[8*step] - src[7*step]) < QP ? src[8*step] : src[7*step];
500
501 int sums[10];
502 sums[0] = 4*first + src[0*step] + src[1*step] + src[2*step] + 4;
503 sums[1] = sums[0] - first + src[3*step];
504 sums[2] = sums[1] - first + src[4*step];
505 sums[3] = sums[2] - first + src[5*step];
506 sums[4] = sums[3] - first + src[6*step];
507 sums[5] = sums[4] - src[0*step] + src[7*step];
508 sums[6] = sums[5] - src[1*step] + last;
509 sums[7] = sums[6] - src[2*step] + last;
510 sums[8] = sums[7] - src[3*step] + last;
511 sums[9] = sums[8] - src[4*step] + last;
512
513 src[0*step]= (sums[0] + sums[2] + 2*src[0*step])>>4;
514 src[1*step]= (sums[1] + sums[3] + 2*src[1*step])>>4;
515 src[2*step]= (sums[2] + sums[4] + 2*src[2*step])>>4;
516 src[3*step]= (sums[3] + sums[5] + 2*src[3*step])>>4;
517 src[4*step]= (sums[4] + sums[6] + 2*src[4*step])>>4;
518 src[5*step]= (sums[5] + sums[7] + 2*src[5*step])>>4;
519 src[6*step]= (sums[6] + sums[8] + 2*src[6*step])>>4;
520 src[7*step]= (sums[7] + sums[9] + 2*src[7*step])>>4;
521 }
522 }else{
523 const int middleEnergy= 5*(src[4*step] - src[3*step]) + 2*(src[2*step] - src[5*step]);
524
525 if(FFABS(middleEnergy) < 8*QP){
526 const int q=(src[3*step] - src[4*step])/2;
527 const int leftEnergy= 5*(src[2*step] - src[1*step]) + 2*(src[0*step] - src[3*step]);
528 const int rightEnergy= 5*(src[6*step] - src[5*step]) + 2*(src[4*step] - src[7*step]);
529
530 int d= FFABS(middleEnergy) - FFMIN( FFABS(leftEnergy), FFABS(rightEnergy) );
531 d= FFMAX(d, 0);
532
533 d= (5*d + 32) >> 6;
534 d*= FFSIGN(-middleEnergy);
535
536 if(q>0){
537 d= d<0 ? 0 : d;
538 d= d>q ? q : d;
539 }else{
540 d= d>0 ? 0 : d;
541 d= d<q ? q : d;
542 }
543
544 src[3*step]-= d;
545 src[4*step]+= d;
546 }
547 }
548
549 src += stride;
550 }
551 /*if(step==16){
552 STOP_TIMER("step16")
553 }else{
554 STOP_TIMER("stepX")
555 }*/
556 }
557
558 //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one
559 //Plain C versions
560 #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
561 #define COMPILE_C
562 #endif
563
564 #ifdef HAVE_ALTIVEC
565 #define COMPILE_ALTIVEC
566 #endif //HAVE_ALTIVEC
567
568 #if defined(ARCH_X86)
569
570 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
571 #define COMPILE_MMX
572 #endif
573
574 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
575 #define COMPILE_MMX2
576 #endif
577
578 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
579 #define COMPILE_3DNOW
580 #endif
581 #endif /* defined(ARCH_X86) */
582
583 #undef HAVE_MMX
584 #undef HAVE_MMX2
585 #undef HAVE_3DNOW
586 #undef HAVE_ALTIVEC
587
588 #ifdef COMPILE_C
589 #undef HAVE_MMX
590 #undef HAVE_MMX2
591 #undef HAVE_3DNOW
592 #define RENAME(a) a ## _C
593 #include "postprocess_template.c"
594 #endif
595
596 #ifdef COMPILE_ALTIVEC
597 #undef RENAME
598 #define HAVE_ALTIVEC
599 #define RENAME(a) a ## _altivec
600 #include "postprocess_altivec_template.c"
601 #include "postprocess_template.c"
602 #endif
603
604 //MMX versions
605 #ifdef COMPILE_MMX
606 #undef RENAME
607 #define HAVE_MMX
608 #undef HAVE_MMX2
609 #undef HAVE_3DNOW
610 #define RENAME(a) a ## _MMX
611 #include "postprocess_template.c"
612 #endif
613
614 //MMX2 versions
615 #ifdef COMPILE_MMX2
616 #undef RENAME
617 #define HAVE_MMX
618 #define HAVE_MMX2
619 #undef HAVE_3DNOW
620 #define RENAME(a) a ## _MMX2
621 #include "postprocess_template.c"
622 #endif
623
624 //3DNOW versions
625 #ifdef COMPILE_3DNOW
626 #undef RENAME
627 #define HAVE_MMX
628 #undef HAVE_MMX2
629 #define HAVE_3DNOW
630 #define RENAME(a) a ## _3DNow
631 #include "postprocess_template.c"
632 #endif
633
634 // minor note: the HAVE_xyz is messed up after that line so do not use it.
635
636 static inline void postProcess(const uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
637 const QP_STORE_T QPs[], int QPStride, int isColor, pp_mode_t *vm, pp_context_t *vc)
638 {
639 PPContext *c= (PPContext *)vc;
640 PPMode *ppMode= (PPMode *)vm;
641 c->ppMode= *ppMode; //FIXME
642
643 // Using ifs here as they are faster than function pointers although the
644 // difference would not be measurable here but it is much better because
645 // someone might exchange the CPU whithout restarting MPlayer ;)
646 #ifdef RUNTIME_CPUDETECT
647 #if defined(ARCH_X86)
648 // ordered per speed fastest first
649 if(c->cpuCaps & PP_CPU_CAPS_MMX2)
650 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
651 else if(c->cpuCaps & PP_CPU_CAPS_3DNOW)
652 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
653 else if(c->cpuCaps & PP_CPU_CAPS_MMX)
654 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
655 else
656 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
657 #else
658 #ifdef HAVE_ALTIVEC
659 if(c->cpuCaps & PP_CPU_CAPS_ALTIVEC)
660 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
661 else
662 #endif
663 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
664 #endif
665 #else //RUNTIME_CPUDETECT
666 #ifdef HAVE_MMX2
667 postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
668 #elif defined (HAVE_3DNOW)
669 postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
670 #elif defined (HAVE_MMX)
671 postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
672 #elif defined (HAVE_ALTIVEC)
673 postProcess_altivec(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
674 #else
675 postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, c);
676 #endif
677 #endif //!RUNTIME_CPUDETECT
678 }
679
680 //static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
681 // QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
682
683 /* -pp Command line Help
684 */
685 #if LIBPOSTPROC_VERSION_INT < (52<<16)
686 const char *const pp_help=
687 #else
688 const char pp_help[] =
689 #endif
690 "Available postprocessing filters:\n"
691 "Filters Options\n"
692 "short long name short long option Description\n"
693 "* * a autoq CPU power dependent enabler\n"
694 " c chrom chrominance filtering enabled\n"
695 " y nochrom chrominance filtering disabled\n"
696 " n noluma luma filtering disabled\n"
697 "hb hdeblock (2 threshold) horizontal deblocking filter\n"
698 " 1. difference factor: default=32, higher -> more deblocking\n"
699 " 2. flatness threshold: default=39, lower -> more deblocking\n"
700 " the h & v deblocking filters share these\n"
701 " so you can't set different thresholds for h / v\n"
702 "vb vdeblock (2 threshold) vertical deblocking filter\n"
703 "ha hadeblock (2 threshold) horizontal deblocking filter\n"
704 "va vadeblock (2 threshold) vertical deblocking filter\n"
705 "h1 x1hdeblock experimental h deblock filter 1\n"
706 "v1 x1vdeblock experimental v deblock filter 1\n"
707 "dr dering deringing filter\n"
708 "al autolevels automatic brightness / contrast\n"
709 " f fullyrange stretch luminance to (0..255)\n"
710 "lb linblenddeint linear blend deinterlacer\n"
711 "li linipoldeint linear interpolating deinterlace\n"
712 "ci cubicipoldeint cubic interpolating deinterlacer\n"
713 "md mediandeint median deinterlacer\n"
714 "fd ffmpegdeint ffmpeg deinterlacer\n"
715 "l5 lowpass5 FIR lowpass deinterlacer\n"
716 "de default hb:a,vb:a,dr:a\n"
717 "fa fast h1:a,v1:a,dr:a\n"
718 "ac ha:a:128:7,va:a,dr:a\n"
719 "tn tmpnoise (3 threshold) temporal noise reducer\n"
720 " 1. <= 2. <= 3. larger -> stronger filtering\n"
721 "fq forceQuant <quantizer> force quantizer\n"
722 "Usage:\n"
723 "<filterName>[:<option>[:<option>...]][[,|/][-]<filterName>[:<option>...]]...\n"
724 "long form example:\n"
725 "vdeblock:autoq/hdeblock:autoq/linblenddeint default,-vdeblock\n"
726 "short form example:\n"
727 "vb:a/hb:a/lb de,-vb\n"
728 "more examples:\n"
729 "tn:64:128:256\n"
730 "\n"
731 ;
732
733 pp_mode_t *pp_get_mode_by_name_and_quality(const char *name, int quality)
734 {
735 char temp[GET_MODE_BUFFER_SIZE];
736 char *p= temp;
737 static const char filterDelimiters[] = ",/";
738 static const char optionDelimiters[] = ":";
739 struct PPMode *ppMode;
740 char *filterToken;
741
742 ppMode= av_malloc(sizeof(PPMode));
743
744 ppMode->lumMode= 0;
745 ppMode->chromMode= 0;
746 ppMode->maxTmpNoise[0]= 700;
747 ppMode->maxTmpNoise[1]= 1500;
748 ppMode->maxTmpNoise[2]= 3000;
749 ppMode->maxAllowedY= 234;
750 ppMode->minAllowedY= 16;
751 ppMode->baseDcDiff= 256/8;
752 ppMode->flatnessThreshold= 56-16-1;
753 ppMode->maxClippedThreshold= 0.01;
754 ppMode->error=0;
755
756 strncpy(temp, name, GET_MODE_BUFFER_SIZE);
757
758 av_log(NULL, AV_LOG_DEBUG, "pp: %s\n", name);
759
760 for(;;){
761 char *filterName;
762 int q= 1000000; //PP_QUALITY_MAX;
763 int chrom=-1;
764 int luma=-1;
765 char *option;
766 char *options[OPTIONS_ARRAY_SIZE];
767 int i;
768 int filterNameOk=0;
769 int numOfUnknownOptions=0;
770 int enable=1; //does the user want us to enabled or disabled the filter
771
772 filterToken= strtok(p, filterDelimiters);
773 if(filterToken == NULL) break;
774 p+= strlen(filterToken) + 1; // p points to next filterToken
775 filterName= strtok(filterToken, optionDelimiters);
776 av_log(NULL, AV_LOG_DEBUG, "pp: %s::%s\n", filterToken, filterName);
777
778 if(*filterName == '-'){
779 enable=0;
780 filterName++;
781 }
782
783 for(;;){ //for all options
784 option= strtok(NULL, optionDelimiters);
785 if(option == NULL) break;
786
787 av_log(NULL, AV_LOG_DEBUG, "pp: option: %s\n", option);
788 if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;
789 else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;
790 else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;
791 else if(!strcmp("noluma", option) || !strcmp("n", option)) luma=0;
792 else{
793 options[numOfUnknownOptions] = option;
794 numOfUnknownOptions++;
795 }
796 if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;
797 }
798 options[numOfUnknownOptions] = NULL;
799
800 /* replace stuff from the replace Table */
801 for(i=0; replaceTable[2*i]!=NULL; i++){
802 if(!strcmp(replaceTable[2*i], filterName)){
803 int newlen= strlen(replaceTable[2*i + 1]);
804 int plen;
805 int spaceLeft;
806
807 if(p==NULL) p= temp, *p=0; //last filter
808 else p--, *p=','; //not last filter
809
810 plen= strlen(p);
811 spaceLeft= p - temp + plen;
812 if(spaceLeft + newlen >= GET_MODE_BUFFER_SIZE){
813 ppMode->error++;
814 break;
815 }
816 memmove(p + newlen, p, plen+1);
817 memcpy(p, replaceTable[2*i + 1], newlen);
818 filterNameOk=1;
819 }
820 }
821
822 for(i=0; filters[i].shortName!=NULL; i++){
823 if( !strcmp(filters[i].longName, filterName)
824 || !strcmp(filters[i].shortName, filterName)){
825 ppMode->lumMode &= ~filters[i].mask;
826 ppMode->chromMode &= ~filters[i].mask;
827
828 filterNameOk=1;
829 if(!enable) break; // user wants to disable it
830
831 if(q >= filters[i].minLumQuality && luma)
832 ppMode->lumMode|= filters[i].mask;
833 if(chrom==1 || (chrom==-1 && filters[i].chromDefault))
834 if(q >= filters[i].minChromQuality)
835 ppMode->chromMode|= filters[i].mask;
836
837 if(filters[i].mask == LEVEL_FIX){
838 int o;
839 ppMode->minAllowedY= 16;
840 ppMode->maxAllowedY= 234;
841 for(o=0; options[o]!=NULL; o++){
842 if( !strcmp(options[o],"fullyrange")
843 ||!strcmp(options[o],"f")){
844 ppMode->minAllowedY= 0;
845 ppMode->maxAllowedY= 255;
846 numOfUnknownOptions--;
847 }
848 }
849 }
850 else if(filters[i].mask == TEMP_NOISE_FILTER)
851 {
852 int o;
853 int numOfNoises=0;
854
855 for(o=0; options[o]!=NULL; o++){
856 char *tail;
857 ppMode->maxTmpNoise[numOfNoises]=
858 strtol(options[o], &tail, 0);
859 if(tail!=options[o]){
860 numOfNoises++;
861 numOfUnknownOptions--;
862 if(numOfNoises >= 3) break;
863 }
864 }
865 }
866 else if(filters[i].mask == V_DEBLOCK || filters[i].mask == H_DEBLOCK
867 || filters[i].mask == V_A_DEBLOCK || filters[i].mask == H_A_DEBLOCK){
868 int o;
869
870 for(o=0; options[o]!=NULL && o<2; o++){
871 char *tail;
872 int val= strtol(options[o], &tail, 0);
873 if(tail==options[o]) break;
874
875 numOfUnknownOptions--;
876 if(o==0) ppMode->baseDcDiff= val;
877 else ppMode->flatnessThreshold= val;
878 }
879 }
880 else if(filters[i].mask == FORCE_QUANT){
881 int o;
882 ppMode->forcedQuant= 15;
883
884 for(o=0; options[o]!=NULL && o<1; o++){
885 char *tail;
886 int val= strtol(options[o], &tail, 0);
887 if(tail==options[o]) break;
888
889 numOfUnknownOptions--;
890 ppMode->forcedQuant= val;
891 }
892 }
893 }
894 }
895 if(!filterNameOk) ppMode->error++;
896 ppMode->error += numOfUnknownOptions;
897 }
898
899 av_log(NULL, AV_LOG_DEBUG, "pp: lumMode=%X, chromMode=%X\n", ppMode->lumMode, ppMode->chromMode);
900 if(ppMode->error){
901 av_log(NULL, AV_LOG_ERROR, "%d errors in postprocess string \"%s\"\n", ppMode->error, name);
902 av_free(ppMode);
903 return NULL;
904 }
905 return ppMode;
906 }
907
908 void pp_free_mode(pp_mode_t *mode){
909 av_free(mode);
910 }
911
912 static void reallocAlign(void **p, int alignment, int size){
913 av_free(*p);
914 *p= av_mallocz(size);
915 }
916
917 static void reallocBuffers(PPContext *c, int width, int height, int stride, int qpStride){
918 int mbWidth = (width+15)>>4;
919 int mbHeight= (height+15)>>4;
920 int i;
921
922 c->stride= stride;
923 c->qpStride= qpStride;
924
925 reallocAlign((void **)&c->tempDst, 8, stride*24);
926 reallocAlign((void **)&c->tempSrc, 8, stride*24);
927 reallocAlign((void **)&c->tempBlocks, 8, 2*16*8);
928 reallocAlign((void **)&c->yHistogram, 8, 256*sizeof(uint64_t));
929 for(i=0; i<256; i++)
930 c->yHistogram[i]= width*height/64*15/256;
931
932 for(i=0; i<3; i++){
933 //Note: The +17*1024 is just there so i do not have to worry about r/w over the end.
934 reallocAlign((void **)&c->tempBlurred[i], 8, stride*mbHeight*16 + 17*1024);
935 reallocAlign((void **)&c->tempBlurredPast[i], 8, 256*((height+7)&(~7))/2 + 17*1024);//FIXME size
936 }
937
938 reallocAlign((void **)&c->deintTemp, 8, 2*width+32);
939 reallocAlign((void **)&c->nonBQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
940 reallocAlign((void **)&c->stdQPTable, 8, qpStride*mbHeight*sizeof(QP_STORE_T));
941 reallocAlign((void **)&c->forcedQPTable, 8, mbWidth*sizeof(QP_STORE_T));
942 }
943
944 static const char * context_to_name(void * ptr) {
945 return "postproc";
946 }
947
948 static const AVClass av_codec_context_class = { "Postproc", context_to_name, NULL };
949
950 pp_context_t *pp_get_context(int width, int height, int cpuCaps){
951 PPContext *c= av_malloc(sizeof(PPContext));
952 int stride= (width+15)&(~15); //assumed / will realloc if needed
953 int qpStride= (width+15)/16 + 2; //assumed / will realloc if needed
954
955 memset(c, 0, sizeof(PPContext));
956 c->av_class = &av_codec_context_class;
957 c->cpuCaps= cpuCaps;
958 if(cpuCaps&PP_FORMAT){
959 c->hChromaSubSample= cpuCaps&0x3;
960 c->vChromaSubSample= (cpuCaps>>4)&0x3;
961 }else{
962 c->hChromaSubSample= 1;
963 c->vChromaSubSample= 1;
964 }
965
966 reallocBuffers(c, width, height, stride, qpStride);
967
968 c->frameNum=-1;
969
970 return c;
971 }
972
973 void pp_free_context(void *vc){
974 PPContext *c = (PPContext*)vc;
975 int i;
976
977 for(i=0; i<3; i++) av_free(c->tempBlurred[i]);
978 for(i=0; i<3; i++) av_free(c->tempBlurredPast[i]);
979
980 av_free(c->tempBlocks);
981 av_free(c->yHistogram);
982 av_free(c->tempDst);
983 av_free(c->tempSrc);
984 av_free(c->deintTemp);
985 av_free(c->stdQPTable);
986 av_free(c->nonBQPTable);
987 av_free(c->forcedQPTable);
988
989 memset(c, 0, sizeof(PPContext));
990
991 av_free(c);
992 }
993
994 void pp_postprocess(const uint8_t * src[3], const int srcStride[3],
995 uint8_t * dst[3], const int dstStride[3],
996 int width, int height,
997 const QP_STORE_T *QP_store, int QPStride,
998 pp_mode_t *vm, void *vc, int pict_type)
999 {
1000 int mbWidth = (width+15)>>4;
1001 int mbHeight= (height+15)>>4;
1002 PPMode *mode = (PPMode*)vm;
1003 PPContext *c = (PPContext*)vc;
1004 int minStride= FFMAX(FFABS(srcStride[0]), FFABS(dstStride[0]));
1005 int absQPStride = FFABS(QPStride);
1006
1007 // c->stride and c->QPStride are always positive
1008 if(c->stride < minStride || c->qpStride < absQPStride)
1009 reallocBuffers(c, width, height,
1010 FFMAX(minStride, c->stride),
1011 FFMAX(c->qpStride, absQPStride));
1012
1013 if(QP_store==NULL || (mode->lumMode & FORCE_QUANT)){
1014 int i;
1015 QP_store= c->forcedQPTable;
1016 absQPStride = QPStride = 0;
1017 if(mode->lumMode & FORCE_QUANT)
1018 for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= mode->forcedQuant;
1019 else
1020 for(i=0; i<mbWidth; i++) c->forcedQPTable[i]= 1;
1021 }
1022
1023 if(pict_type & PP_PICT_TYPE_QP2){
1024 int i;
1025 const int count= mbHeight * absQPStride;
1026 for(i=0; i<(count>>2); i++){
1027 ((uint32_t*)c->stdQPTable)[i] = (((const uint32_t*)QP_store)[i]>>1) & 0x7F7F7F7F;
1028 }
1029 for(i<<=2; i<count; i++){
1030 c->stdQPTable[i] = QP_store[i]>>1;
1031 }
1032 QP_store= c->stdQPTable;
1033 QPStride= absQPStride;
1034 }
1035
1036 if(0){
1037 int x,y;
1038 for(y=0; y<mbHeight; y++){
1039 for(x=0; x<mbWidth; x++){
1040 av_log(c, AV_LOG_INFO, "%2d ", QP_store[x + y*QPStride]);
1041 }
1042 av_log(c, AV_LOG_INFO, "\n");
1043 }
1044 av_log(c, AV_LOG_INFO, "\n");
1045 }
1046
1047 if((pict_type&7)!=3){
1048 if (QPStride >= 0){
1049 int i;
1050 const int count= mbHeight * QPStride;
1051 for(i=0; i<(count>>2); i++){
1052 ((uint32_t*)c->nonBQPTable)[i] = ((const uint32_t*)QP_store)[i] & 0x3F3F3F3F;
1053 }
1054 for(i<<=2; i<count; i++){
1055 c->nonBQPTable[i] = QP_store[i] & 0x3F;
1056 }
1057 } else {
1058 int i,j;
1059 for(i=0; i<mbHeight; i++) {
1060 for(j=0; j<absQPStride; j++) {
1061 c->nonBQPTable[i*absQPStride+j] = QP_store[i*QPStride+j] & 0x3F;
1062 }
1063 }
1064 }
1065 }
1066
1067 av_log(c, AV_LOG_DEBUG, "using npp filters 0x%X/0x%X\n",
1068 mode->lumMode, mode->chromMode);
1069
1070 postProcess(src[0], srcStride[0], dst[0], dstStride[0],
1071 width, height, QP_store, QPStride, 0, mode, c);
1072
1073 width = (width )>>c->hChromaSubSample;
1074 height = (height)>>c->vChromaSubSample;
1075
1076 if(mode->chromMode){
1077 postProcess(src[1], srcStride[1], dst[1], dstStride[1],
1078 width, height, QP_store, QPStride, 1, mode, c);
1079 postProcess(src[2], srcStride[2], dst[2], dstStride[2],
1080 width, height, QP_store, QPStride, 2, mode, c);
1081 }
1082 else if(srcStride[1] == dstStride[1] && srcStride[2] == dstStride[2]){
1083 linecpy(dst[1], src[1], height, srcStride[1]);
1084 linecpy(dst[2], src[2], height, srcStride[2]);
1085 }else{
1086 int y;
1087 for(y=0; y<height; y++){
1088 memcpy(&(dst[1][y*dstStride[1]]), &(src[1][y*srcStride[1]]), width);
1089 memcpy(&(dst[2][y*dstStride[2]]), &(src[2][y*srcStride[2]]), width);
1090 }
1091 }
1092 }
1093
VDPAU for FFmpeg