falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
343 Commits
32 Branches
944MB
Tree: 45b4f28550
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '45b4f28550'
${ noResults }
FFmpeg/postproc/postprocess.c

915 lines
27KB
Raw Blame History 

  1. /*

  2.     Copyright (C) 2001 Michael Niedermayer (michaelni@gmx.at)

  3. 

  4.     This program is free software; you can redistribute it and/or modify

  5.     it under the terms of the GNU General Public License as published by

  6.     the Free Software Foundation; either version 2 of the License, or

  7.     (at your option) any later version.

  8. 

  9.     This program is distributed in the hope that it will be useful,

  10.     but WITHOUT ANY WARRANTY; without even the implied warranty of

  11.     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

  12.     GNU General Public License for more details.

  13. 

  14.     You should have received a copy of the GNU General Public License

  15.     along with this program; if not, write to the Free Software

  16.     Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA

  17. */

  18. 

  19. /*

  20. 			C	MMX	MMX2	3DNow

  21. isVertDC		Ec	Ec

  22. isVertMinMaxOk		Ec	Ec

  23. doVertLowPass		E		e	e

  24. doVertDefFilter		Ec	Ec	e	e

  25. isHorizDC		Ec	Ec

  26. isHorizMinMaxOk		a	E

  27. doHorizLowPass		E		e	e

  28. doHorizDefFilter	Ec	Ec	e	e

  29. deRing			E		e	e*

  30. Vertical RKAlgo1	E		a	a

  31. Horizontal RKAlgo1			a	a

  32. Vertical X1#		a		E	E

  33. Horizontal X1#		a		E	E

  34. LinIpolDeinterlace	e		E	E*

  35. CubicIpolDeinterlace	a		e	e*

  36. LinBlendDeinterlace	e		E	E*

  37. MedianDeinterlace#	 	Ec	Ec

  38. TempDeNoiser#		E		e	e

  39. 

  40. * i dont have a 3dnow CPU -> its untested, but noone said it doesnt work so it seems to work

  41. # more or less selfinvented filters so the exactness isnt too meaningfull

  42. E = Exact implementation

  43. e = allmost exact implementation (slightly different rounding,...)

  44. a = alternative / approximate impl

  45. c = checked against the other implementations (-vo md5)

  46. */

  47. 

  48. /*

  49. TODO:

  50. reduce the time wasted on the mem transfer

  51. implement everything in C at least (done at the moment but ...)

  52. unroll stuff if instructions depend too much on the prior one

  53. we use 8x8 blocks for the horizontal filters, opendivx seems to use 8x4?

  54. move YScale thing to the end instead of fixing QP

  55. write a faster and higher quality deblocking filter :)

  56. make the mainloop more flexible (variable number of blocks at once

  57. 	(the if/else stuff per block is slowing things down)

  58. compare the quality & speed of all filters

  59. split this huge file

  60. border remover

  61. optimize c versions

  62. try to unroll inner for(x=0 ... loop to avoid these damn if(x ... checks

  63. smart blur

  64. commandline option for   the deblock / dering thresholds

  65. put fastmemcpy back

  66. dont use #ifdef ARCH_X86 for the asm stuff ... cross compilers? (note cpudetect uses ARCH_X86)

  67. ...

  68. */

  69. 

  70. //Changelog: use the CVS log

  71. 

  72. #include "../config.h"

  73. #include <inttypes.h>

  74. #include <stdio.h>

  75. #include <stdlib.h>

  76. #include <string.h>

  77. #ifdef HAVE_MALLOC_H

  78. #include <malloc.h>

  79. #endif

  80. //#undef HAVE_MMX2

  81. //#define HAVE_3DNOW

  82. //#undef HAVE_MMX

  83. //#undef ARCH_X86

  84. //#define DEBUG_BRIGHTNESS

  85. //#include "../libvo/fastmemcpy.h"

  86. #include "postprocess.h"

  87. #include "../cpudetect.h"

  88. 

  89. #define MIN(a,b) ((a) > (b) ? (b) : (a))

  90. #define MAX(a,b) ((a) < (b) ? (b) : (a))

  91. #define ABS(a) ((a) > 0 ? (a) : (-(a)))

  92. #define SIGN(a) ((a) > 0 ? 1 : -1)

  93. 

  94. #define GET_MODE_BUFFER_SIZE 500

  95. #define OPTIONS_ARRAY_SIZE 10

  96. 

  97. #ifdef ARCH_X86

  98. #define CAN_COMPILE_X86_ASM

  99. #endif

  100. 

  101. #ifdef CAN_COMPILE_X86_ASM

  102. static volatile uint64_t __attribute__((aligned(8))) packedYOffset=	0x0000000000000000LL;

  103. static volatile uint64_t __attribute__((aligned(8))) packedYScale=	0x0100010001000100LL;

  104. static uint64_t __attribute__((aligned(8))) w05=		0x0005000500050005LL;

  105. static uint64_t __attribute__((aligned(8))) w20=		0x0020002000200020LL;

  106. static uint64_t __attribute__((aligned(8))) w1400=		0x1400140014001400LL;

  107. static uint64_t __attribute__((aligned(8))) bm00000001=		0x00000000000000FFLL;

  108. static uint64_t __attribute__((aligned(8))) bm00010000=		0x000000FF00000000LL;

  109. static uint64_t __attribute__((aligned(8))) bm00001000=		0x00000000FF000000LL;

  110. static uint64_t __attribute__((aligned(8))) bm10000000=		0xFF00000000000000LL;

  111. static uint64_t __attribute__((aligned(8))) bm10000001=		0xFF000000000000FFLL;

  112. static uint64_t __attribute__((aligned(8))) bm11000011=		0xFFFF00000000FFFFLL;

  113. static uint64_t __attribute__((aligned(8))) bm00000011=		0x000000000000FFFFLL;

  114. static uint64_t __attribute__((aligned(8))) bm11111110=		0xFFFFFFFFFFFFFF00LL;

  115. static uint64_t __attribute__((aligned(8))) bm11000000=		0xFFFF000000000000LL;

  116. static uint64_t __attribute__((aligned(8))) bm00011000=		0x000000FFFF000000LL;

  117. static uint64_t __attribute__((aligned(8))) bm00110011=		0x0000FFFF0000FFFFLL;

  118. static uint64_t __attribute__((aligned(8))) bm11001100=		0xFFFF0000FFFF0000LL;

  119. static uint64_t __attribute__((aligned(8))) b00= 		0x0000000000000000LL;

  120. static uint64_t __attribute__((aligned(8))) b01= 		0x0101010101010101LL;

  121. static uint64_t __attribute__((aligned(8))) b02= 		0x0202020202020202LL;

  122. static uint64_t __attribute__((aligned(8))) b0F= 		0x0F0F0F0F0F0F0F0FLL;

  123. static uint64_t __attribute__((aligned(8))) b04= 		0x0404040404040404LL;

  124. static uint64_t __attribute__((aligned(8))) b08= 		0x0808080808080808LL;

  125. static uint64_t __attribute__((aligned(8))) bFF= 		0xFFFFFFFFFFFFFFFFLL;

  126. static uint64_t __attribute__((aligned(8))) b20= 		0x2020202020202020LL;

  127. static uint64_t __attribute__((aligned(8))) b80= 		0x8080808080808080LL;

  128. static uint64_t __attribute__((aligned(8))) b7E= 		0x7E7E7E7E7E7E7E7ELL;

  129. static uint64_t __attribute__((aligned(8))) b7C= 		0x7C7C7C7C7C7C7C7CLL;

  130. static uint64_t __attribute__((aligned(8))) b3F= 		0x3F3F3F3F3F3F3F3FLL;

  131. static uint64_t __attribute__((aligned(8))) temp0=0;

  132. static uint64_t __attribute__((aligned(8))) temp1=0;

  133. static uint64_t __attribute__((aligned(8))) temp2=0;

  134. static uint64_t __attribute__((aligned(8))) temp3=0;

  135. static uint64_t __attribute__((aligned(8))) temp4=0;

  136. static uint64_t __attribute__((aligned(8))) temp5=0;

  137. static uint64_t __attribute__((aligned(8))) pQPb=0;

  138. static uint64_t __attribute__((aligned(8))) pQPb2=0;

  139. static uint8_t __attribute__((aligned(8))) tempBlocks[8*16*2]; //used for the horizontal code

  140. static uint32_t __attribute__((aligned(4))) maxTmpNoise[4];

  141. #else

  142. static uint64_t packedYOffset=	0x0000000000000000LL;

  143. static uint64_t packedYScale=	0x0100010001000100LL;

  144. #endif

  145. 

  146. extern int divx_quality;

  147. int newPPFlag=0; //is set if -npp is used

  148. struct PPMode gPPMode[GET_PP_QUALITY_MAX+1];

  149. 

  150. extern int verbose;

  151. 

  152. int hFlatnessThreshold= 56 - 16;

  153. int vFlatnessThreshold= 56 - 16;

  154. int deringThreshold= 20;

  155. 

  156. //amount of "black" u r willing to loose to get a brightness corrected picture

  157. double maxClippedThreshold= 0.01;

  158. 

  159. int maxAllowedY=234;

  160. int minAllowedY=16;

  161. 

  162. static struct PPFilter filters[]=

  163. {

  164. 	{"hb", "hdeblock", 		1, 1, 3, H_DEBLOCK},

  165. 	{"vb", "vdeblock", 		1, 2, 4, V_DEBLOCK},

  166. 	{"vr", "rkvdeblock", 		1, 2, 4, H_RK1_FILTER},

  167. 	{"h1", "x1hdeblock", 		1, 1, 3, H_X1_FILTER},

  168. 	{"v1", "x1vdeblock", 		1, 2, 4, V_X1_FILTER},

  169. 	{"dr", "dering", 		1, 5, 6, DERING},

  170. 	{"al", "autolevels", 		0, 1, 2, LEVEL_FIX},

  171. 	{"lb", "linblenddeint", 	0, 1, 6, LINEAR_BLEND_DEINT_FILTER},

  172. 	{"li", "linipoldeint", 		0, 1, 6, LINEAR_IPOL_DEINT_FILTER},

  173. 	{"ci", "cubicipoldeint",	0, 1, 6, CUBIC_IPOL_DEINT_FILTER},

  174. 	{"md", "mediandeint", 		0, 1, 6, MEDIAN_DEINT_FILTER},

  175. 	{"tn", "tmpnoise", 		1, 7, 8, TEMP_NOISE_FILTER},

  176. 	{NULL, NULL,0,0,0,0} //End Marker

  177. };

  178. 

  179. static char *replaceTable[]=

  180. {

  181. 	"default", 	"hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",

  182. 	"de", 		"hdeblock:a,vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",

  183. 	"fast", 	"x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",

  184. 	"fa", 		"x1hdeblock:a,x1vdeblock:a,dering:a,autolevels,tmpnoise:a:150:200:400",

  185. 	NULL //End Marker

  186. };

  187. 

  188. #ifdef CAN_COMPILE_X86_ASM

  189. static inline void unusedVariableWarningFixer()

  190. {

  191. if(

  192.  packedYOffset + packedYScale + w05 + w20 + w1400 + bm00000001 + bm00010000

  193.  + bm00001000 + bm10000000 + bm10000001 + bm11000011 + bm00000011 + bm11111110

  194.  + bm11000000 + bm00011000 + bm00110011 + bm11001100 + b00 + b01 + b02 + b0F

  195.  + bFF + b20 + b04+ b08 + pQPb2 + b80 + b7E + b7C + b3F + temp0 + temp1 + temp2 + temp3 + temp4

  196.  + temp5 + pQPb== 0) b00=0;

  197. }

  198. #endif

  199. 

  200. #ifdef TIMING

  201. static inline long long rdtsc()

  202. {

  203. 	long long l;

  204. 	asm volatile(	"rdtsc\n\t"

  205. 		: "=A" (l)

  206. 	);

  207. //	printf("%d\n", int(l/1000));

  208. 	return l;

  209. }

  210. #endif

  211. 

  212. #ifdef CAN_COMPILE_X86_ASM

  213. static inline void prefetchnta(void *p)

  214. {

  215. 	asm volatile(	"prefetchnta (%0)\n\t"

  216. 		: : "r" (p)

  217. 	);

  218. }

  219. 

  220. static inline void prefetcht0(void *p)

  221. {

  222. 	asm volatile(	"prefetcht0 (%0)\n\t"

  223. 		: : "r" (p)

  224. 	);

  225. }

  226. 

  227. static inline void prefetcht1(void *p)

  228. {

  229. 	asm volatile(	"prefetcht1 (%0)\n\t"

  230. 		: : "r" (p)

  231. 	);

  232. }

  233. 

  234. static inline void prefetcht2(void *p)

  235. {

  236. 	asm volatile(	"prefetcht2 (%0)\n\t"

  237. 		: : "r" (p)

  238. 	);

  239. }

  240. #endif

  241. 

  242. // The horizontal Functions exist only in C cuz the MMX code is faster with vertical filters and transposing

  243. 

  244. /**

  245.  * Check if the given 8x8 Block is mostly "flat"

  246.  */

  247. static inline int isHorizDC(uint8_t src[], int stride)

  248. {

  249. 	int numEq= 0;

  250. 	int y;

  251. 	for(y=0; y<BLOCK_SIZE; y++)

  252. 	{

  253. 		if(((src[0] - src[1] + 1) & 0xFFFF) < 3) numEq++;

  254. 		if(((src[1] - src[2] + 1) & 0xFFFF) < 3) numEq++;

  255. 		if(((src[2] - src[3] + 1) & 0xFFFF) < 3) numEq++;

  256. 		if(((src[3] - src[4] + 1) & 0xFFFF) < 3) numEq++;

  257. 		if(((src[4] - src[5] + 1) & 0xFFFF) < 3) numEq++;

  258. 		if(((src[5] - src[6] + 1) & 0xFFFF) < 3) numEq++;

  259. 		if(((src[6] - src[7] + 1) & 0xFFFF) < 3) numEq++;

  260. 		src+= stride;

  261. 	}

  262. 	return numEq > hFlatnessThreshold;

  263. }

  264. 

  265. static inline int isHorizMinMaxOk(uint8_t src[], int stride, int QP)

  266. {

  267. 	if(abs(src[0] - src[7]) > 2*QP) return 0;

  268. 

  269. 	return 1;

  270. }

  271. 

  272. static inline void doHorizDefFilter(uint8_t dst[], int stride, int QP)

  273. {

  274. 	int y;

  275. 	for(y=0; y<BLOCK_SIZE; y++)

  276. 	{

  277. 		const int middleEnergy= 5*(dst[4] - dst[5]) + 2*(dst[2] - dst[5]);

  278. 

  279. 		if(ABS(middleEnergy) < 8*QP)

  280. 		{

  281. 			const int q=(dst[3] - dst[4])/2;

  282. 			const int leftEnergy=  5*(dst[2] - dst[1]) + 2*(dst[0] - dst[3]);

  283. 			const int rightEnergy= 5*(dst[6] - dst[5]) + 2*(dst[4] - dst[7]);

  284. 

  285. 			int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );

  286. 			d= MAX(d, 0);

  287. 

  288. 			d= (5*d + 32) >> 6;

  289. 			d*= SIGN(-middleEnergy);

  290. 

  291. 			if(q>0)

  292. 			{

  293. 				d= d<0 ? 0 : d;

  294. 				d= d>q ? q : d;

  295. 			}

  296. 			else

  297. 			{

  298. 				d= d>0 ? 0 : d;

  299. 				d= d<q ? q : d;

  300. 			}

  301. 

  302.         		dst[3]-= d;

  303. 	        	dst[4]+= d;

  304. 		}

  305. 		dst+= stride;

  306. 	}

  307. }

  308. 

  309. /**

  310.  * Do a horizontal low pass filter on the 10x8 block (dst points to middle 8x8 Block)

  311.  * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)

  312.  */

  313. static inline void doHorizLowPass(uint8_t dst[], int stride, int QP)

  314. {

  315. 

  316. 	int y;

  317. 	for(y=0; y<BLOCK_SIZE; y++)

  318. 	{

  319. 		const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];

  320. 		const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];

  321. 

  322. 		int sums[9];

  323. 		sums[0] = first + dst[0];

  324. 		sums[1] = dst[0] + dst[1];

  325. 		sums[2] = dst[1] + dst[2];

  326. 		sums[3] = dst[2] + dst[3];

  327. 		sums[4] = dst[3] + dst[4];

  328. 		sums[5] = dst[4] + dst[5];

  329. 		sums[6] = dst[5] + dst[6];

  330. 		sums[7] = dst[6] + dst[7];

  331. 		sums[8] = dst[7] + last;

  332. 

  333. 		dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;

  334. 		dst[1]= ((dst[1]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;

  335. 		dst[2]= ((dst[2]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;

  336. 		dst[3]= ((dst[3]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;

  337. 		dst[4]= ((dst[4]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;

  338. 		dst[5]= ((dst[5]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;

  339. 		dst[6]= (((last + dst[6])<<2) + ((dst[7] + sums[5])<<1) + sums[3] + 8)>>4;

  340. 		dst[7]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;

  341. 

  342. 		dst+= stride;

  343. 	}

  344. }

  345. 

  346. /**

  347.  * Experimental Filter 1 (Horizontal)

  348.  * will not damage linear gradients

  349.  * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter

  350.  * can only smooth blocks at the expected locations (it cant smooth them if they did move)

  351.  * MMX2 version does correct clipping C version doesnt

  352.  * not identical with the vertical one

  353.  */

  354. static inline void horizX1Filter(uint8_t *src, int stride, int QP)

  355. {

  356. 	int y;

  357. 	static uint64_t *lut= NULL;

  358. 	if(lut==NULL)

  359. 	{

  360. 		int i;

  361. 		lut= (uint64_t*)memalign(8, 256*8);

  362. 		for(i=0; i<256; i++)

  363. 		{

  364. 			int v= i < 128 ? 2*i : 2*(i-256);

  365. /*

  366. //Simulate 112242211 9-Tap filter

  367. 			uint64_t a= (v/16) & 0xFF;

  368. 			uint64_t b= (v/8) & 0xFF;

  369. 			uint64_t c= (v/4) & 0xFF;

  370. 			uint64_t d= (3*v/8) & 0xFF;

  371. */

  372. //Simulate piecewise linear interpolation

  373. 			uint64_t a= (v/16) & 0xFF;

  374. 			uint64_t b= (v*3/16) & 0xFF;

  375. 			uint64_t c= (v*5/16) & 0xFF;

  376. 			uint64_t d= (7*v/16) & 0xFF;

  377. 			uint64_t A= (0x100 - a)&0xFF;

  378. 			uint64_t B= (0x100 - b)&0xFF;

  379. 			uint64_t C= (0x100 - c)&0xFF;

  380. 			uint64_t D= (0x100 - c)&0xFF;

  381. 

  382. 			lut[i]   = (a<<56) | (b<<48) | (c<<40) | (d<<32) |

  383. 				(D<<24) | (C<<16) | (B<<8) | (A);

  384. 			//lut[i] = (v<<32) | (v<<24);

  385. 		}

  386. 	}

  387. 

  388. 	for(y=0; y<BLOCK_SIZE; y++)

  389. 	{

  390. 		int a= src[1] - src[2];

  391. 		int b= src[3] - src[4];

  392. 		int c= src[5] - src[6];

  393. 

  394. 		int d= MAX(ABS(b) - (ABS(a) + ABS(c))/2, 0);

  395. 

  396. 		if(d < QP)

  397. 		{

  398. 			int v = d * SIGN(-b);

  399. 

  400. 			src[1] +=v/8;

  401. 			src[2] +=v/4;

  402. 			src[3] +=3*v/8;

  403. 			src[4] -=3*v/8;

  404. 			src[5] -=v/4;

  405. 			src[6] -=v/8;

  406. 

  407. 		}

  408. 		src+=stride;

  409. 	}

  410. }

  411. 

  412. 

  413. //Note: we have C, MMX, MMX2, 3DNOW version there is no 3DNOW+MMX2 one

  414. //Plain C versions

  415. #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)

  416. #define COMPILE_C

  417. #endif

  418. 

  419. #ifdef CAN_COMPILE_X86_ASM

  420. 

  421. #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)

  422. #define COMPILE_MMX

  423. #endif

  424. 

  425. #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)

  426. #define COMPILE_MMX2

  427. #endif

  428. 

  429. #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)

  430. #define COMPILE_3DNOW

  431. #endif

  432. #endif //CAN_COMPILE_X86_ASM

  433. 

  434. #undef HAVE_MMX

  435. #undef HAVE_MMX2

  436. #undef HAVE_3DNOW

  437. #undef ARCH_X86

  438. 

  439. #ifdef COMPILE_C

  440. #undef HAVE_MMX

  441. #undef HAVE_MMX2

  442. #undef HAVE_3DNOW

  443. #undef ARCH_X86

  444. #define RENAME(a) a ## _C

  445. #include "postprocess_template.c"

  446. #endif

  447. 

  448. //MMX versions

  449. #ifdef COMPILE_MMX

  450. #undef RENAME

  451. #define HAVE_MMX

  452. #undef HAVE_MMX2

  453. #undef HAVE_3DNOW

  454. #define ARCH_X86

  455. #define RENAME(a) a ## _MMX

  456. #include "postprocess_template.c"

  457. #endif

  458. 

  459. //MMX2 versions

  460. #ifdef COMPILE_MMX2

  461. #undef RENAME

  462. #define HAVE_MMX

  463. #define HAVE_MMX2

  464. #undef HAVE_3DNOW

  465. #define ARCH_X86

  466. #define RENAME(a) a ## _MMX2

  467. #include "postprocess_template.c"

  468. #endif

  469. 

  470. //3DNOW versions

  471. #ifdef COMPILE_3DNOW

  472. #undef RENAME

  473. #define HAVE_MMX

  474. #undef HAVE_MMX2

  475. #define HAVE_3DNOW

  476. #define ARCH_X86

  477. #define RENAME(a) a ## _3DNow

  478. #include "postprocess_template.c"

  479. #endif

  480. 

  481. // minor note: the HAVE_xyz is messed up after that line so dont use it

  482. 

  483. static inline void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,

  484. 	QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode)

  485. {

  486. 	// useing ifs here as they are faster than function pointers allthough the

  487. 	// difference wouldnt be messureable here but its much better because

  488. 	// someone might exchange the cpu whithout restarting mplayer ;)

  489. #ifdef RUNTIME_CPUDETECT

  490. #ifdef CAN_COMPILE_X86_ASM

  491. 	// ordered per speed fasterst first

  492. 	if(gCpuCaps.hasMMX2)

  493. 		postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  494. 	else if(gCpuCaps.has3DNow)

  495. 		postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  496. 	else if(gCpuCaps.hasMMX)

  497. 		postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  498. 	else

  499. 		postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  500. #else

  501. 		postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  502. #endif

  503. #else //RUNTIME_CPUDETECT

  504. #ifdef HAVE_MMX2

  505. 		postProcess_MMX2(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  506. #elif defined (HAVE_3DNOW)

  507. 		postProcess_3DNow(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  508. #elif defined (HAVE_MMX)

  509. 		postProcess_MMX(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  510. #else

  511. 		postProcess_C(src, srcStride, dst, dstStride, width, height, QPs, QPStride, isColor, ppMode);

  512. #endif

  513. #endif //!RUNTIME_CPUDETECT

  514. }

  515. 

  516. #ifdef HAVE_ODIVX_POSTPROCESS

  517. #include "../opendivx/postprocess.h"

  518. int use_old_pp=0;

  519. #endif

  520. 

  521. //static void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,

  522. //	QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);

  523. 

  524. /* -pp Command line Help

  525. NOTE/FIXME: put this at an appropriate place (--help, html docs, man mplayer)?

  526. 

  527. -pp <filterName>[:<option>[:<option>...]][,[-]<filterName>[:<option>...]]...

  528. 

  529. long form example:

  530. -pp vdeblock:autoq,hdeblock:autoq,linblenddeint		-pp default,-vdeblock

  531. short form example:

  532. -pp vb:a,hb:a,lb					-pp de,-vb

  533. more examples:

  534. -pp tn:64:128:256

  535. 

  536. Filters			Options

  537. short	long name	short	long option	Description

  538. *	*		a	autoq		cpu power dependant enabler

  539. 			c	chrom		chrominance filtring enabled

  540. 			y	nochrom		chrominance filtring disabled

  541. hb	hdeblock				horizontal deblocking filter

  542. vb	vdeblock				vertical deblocking filter

  543. vr	rkvdeblock

  544. h1	x1hdeblock				Experimental horizontal deblock filter 1

  545. v1	x1vdeblock				Experimental vertical deblock filter 1

  546. dr	dering					not implemented yet

  547. al	autolevels				automatic brightness / contrast fixer

  548. 			f	fullyrange	stretch luminance range to (0..255)

  549. lb	linblenddeint				linear blend deinterlacer

  550. li	linipoldeint				linear interpolating deinterlacer

  551. ci	cubicipoldeint				cubic interpolating deinterlacer

  552. md	mediandeint				median deinterlacer

  553. de	default					hdeblock:a,vdeblock:a,dering:a,autolevels

  554. fa	fast					x1hdeblock:a,x1vdeblock:a,dering:a,autolevels

  555. tn	tmpnoise	(3 Thresholds)		Temporal Noise Reducer

  556. */

  557. 

  558. /**

  559.  * returns a PPMode struct which will have a non 0 error variable if an error occured

  560.  * name is the string after "-pp" on the command line

  561.  * quality is a number from 0 to GET_PP_QUALITY_MAX

  562.  */

  563. struct PPMode getPPModeByNameAndQuality(char *name, int quality)

  564. {

  565. 	char temp[GET_MODE_BUFFER_SIZE];

  566. 	char *p= temp;

  567. 	char *filterDelimiters= ",";

  568. 	char *optionDelimiters= ":";

  569. 	struct PPMode ppMode= {0,0,0,0,0,0,{150,200,400}};

  570. 	char *filterToken;

  571. 

  572. 	strncpy(temp, name, GET_MODE_BUFFER_SIZE);

  573. 

  574. 	if(verbose) printf("%s\n", name);

  575. 

  576. 	for(;;){

  577. 		char *filterName;

  578. 		int q= 1000000; //GET_PP_QUALITY_MAX;

  579. 		int chrom=-1;

  580. 		char *option;

  581. 		char *options[OPTIONS_ARRAY_SIZE];

  582. 		int i;

  583. 		int filterNameOk=0;

  584. 		int numOfUnknownOptions=0;

  585. 		int enable=1; //does the user want us to enabled or disabled the filter

  586. 

  587. 		filterToken= strtok(p, filterDelimiters);

  588. 		if(filterToken == NULL) break;

  589. 		p+= strlen(filterToken) + 1; // p points to next filterToken

  590. 		filterName= strtok(filterToken, optionDelimiters);

  591. 		if(verbose) printf("%s::%s\n", filterToken, filterName);

  592. 

  593. 		if(*filterName == '-')

  594. 		{

  595. 			enable=0;

  596. 			filterName++;

  597. 		}

  598. 

  599. 		for(;;){ //for all options

  600. 			option= strtok(NULL, optionDelimiters);

  601. 			if(option == NULL) break;

  602. 

  603. 			if(verbose) printf(" option: %s\n", option);

  604. 			if(!strcmp("autoq", option) || !strcmp("a", option)) q= quality;

  605. 			else if(!strcmp("nochrom", option) || !strcmp("y", option)) chrom=0;

  606. 			else if(!strcmp("chrom", option) || !strcmp("c", option)) chrom=1;

  607. 			else

  608. 			{

  609. 				options[numOfUnknownOptions] = option;

  610. 				numOfUnknownOptions++;

  611. 			}

  612. 			if(numOfUnknownOptions >= OPTIONS_ARRAY_SIZE-1) break;

  613. 		}

  614. 		options[numOfUnknownOptions] = NULL;

  615. 

  616. 		/* replace stuff from the replace Table */

  617. 		for(i=0; replaceTable[2*i]!=NULL; i++)

  618. 		{

  619. 			if(!strcmp(replaceTable[2*i], filterName))

  620. 			{

  621. 				int newlen= strlen(replaceTable[2*i + 1]);

  622. 				int plen;

  623. 				int spaceLeft;

  624. 

  625. 				if(p==NULL) p= temp, *p=0; 	//last filter

  626. 				else p--, *p=',';		//not last filter

  627. 

  628. 				plen= strlen(p);

  629. 				spaceLeft= (int)p - (int)temp + plen;

  630. 				if(spaceLeft + newlen  >= GET_MODE_BUFFER_SIZE)

  631. 				{

  632. 					ppMode.error++;

  633. 					break;

  634. 				}

  635. 				memmove(p + newlen, p, plen+1);

  636. 				memcpy(p, replaceTable[2*i + 1], newlen);

  637. 				filterNameOk=1;

  638. 			}

  639. 		}

  640. 

  641. 		for(i=0; filters[i].shortName!=NULL; i++)

  642. 		{

  643. //			printf("Compareing %s, %s, %s\n", filters[i].shortName,filters[i].longName, filterName);

  644. 			if(   !strcmp(filters[i].longName, filterName)

  645. 			   || !strcmp(filters[i].shortName, filterName))

  646. 			{

  647. 				ppMode.lumMode &= ~filters[i].mask;

  648. 				ppMode.chromMode &= ~filters[i].mask;

  649. 

  650. 				filterNameOk=1;

  651. 				if(!enable) break; // user wants to disable it

  652. 

  653. 				if(q >= filters[i].minLumQuality)

  654. 					ppMode.lumMode|= filters[i].mask;

  655. 				if(chrom==1 || (chrom==-1 && filters[i].chromDefault))

  656. 					if(q >= filters[i].minChromQuality)

  657. 						ppMode.chromMode|= filters[i].mask;

  658. 

  659. 				if(filters[i].mask == LEVEL_FIX)

  660. 				{

  661. 					int o;

  662. 					ppMode.minAllowedY= 16;

  663. 					ppMode.maxAllowedY= 234;

  664. 					for(o=0; options[o]!=NULL; o++)

  665. 						if(  !strcmp(options[o],"fullyrange")

  666. 						   ||!strcmp(options[o],"f"))

  667. 						{

  668. 							ppMode.minAllowedY= 0;

  669. 							ppMode.maxAllowedY= 255;

  670. 							numOfUnknownOptions--;

  671. 						}

  672. 				}

  673. 				else if(filters[i].mask == TEMP_NOISE_FILTER)

  674. 				{

  675. 					int o;

  676. 					int numOfNoises=0;

  677. 					ppMode.maxTmpNoise[0]= 150;

  678. 					ppMode.maxTmpNoise[1]= 200;

  679. 					ppMode.maxTmpNoise[2]= 400;

  680. 

  681. 					for(o=0; options[o]!=NULL; o++)

  682. 					{

  683. 						char *tail;

  684. 						ppMode.maxTmpNoise[numOfNoises]=

  685. 							strtol(options[o], &tail, 0);

  686. 						if(tail!=options[o])

  687. 						{

  688. 							numOfNoises++;

  689. 							numOfUnknownOptions--;

  690. 							if(numOfNoises >= 3) break;

  691. 						}

  692. 					}

  693. 				}

  694. 			}

  695. 		}

  696. 		if(!filterNameOk) ppMode.error++;

  697. 		ppMode.error += numOfUnknownOptions;

  698. 	}

  699. 

  700. #ifdef HAVE_ODIVX_POSTPROCESS

  701. 	if(ppMode.lumMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_H;

  702. 	if(ppMode.lumMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_Y_V;

  703. 	if(ppMode.chromMode & H_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_H;

  704. 	if(ppMode.chromMode & V_DEBLOCK) ppMode.oldMode |= PP_DEBLOCK_C_V;

  705. 	if(ppMode.lumMode & DERING) ppMode.oldMode |= PP_DERING_Y;

  706. 	if(ppMode.chromMode & DERING) ppMode.oldMode |= PP_DERING_C;

  707. #endif

  708. 

  709. 	if(verbose) printf("lumMode=%X, chromMode=%X\n", ppMode.lumMode, ppMode.chromMode);

  710. 	return ppMode;

  711. }

  712. 

  713. /**

  714.  * Check and load the -npp part of the cmd line

  715.  */

  716. int readPPOpt(void *conf, char *arg)

  717. {

  718. 	int quality;

  719. 	for(quality=0; quality<GET_PP_QUALITY_MAX+1; quality++)

  720. 	{

  721. 		gPPMode[quality]= getPPModeByNameAndQuality(arg, quality);

  722. 

  723. 		if(gPPMode[quality].error) return -1;

  724. 	}

  725. 	newPPFlag=1;

  726. 

  727. // it shouldnt matter what we set divx_quality to as long as its not 0 

  728. // (we dont use it if newPPFlag is set)

  729. 	divx_quality=1; 

  730. 	return 1;

  731. }

  732. 

  733. /**

  734.  * Obsolete, dont use it, use postprocess2() instead

  735.  * this will check newPPFlag automatically and use postprocess2 if it is set

  736.  * mode = quality if newPPFlag

  737.  */

  738. void  postprocess(unsigned char * src[], int src_stride,

  739.                  unsigned char * dst[], int dst_stride,

  740.                  int horizontal_size,   int vertical_size,

  741.                  QP_STORE_T *QP_store,  int QP_stride,

  742. 					  int mode)

  743. {

  744. 	struct PPMode ppMode;

  745. 	static QP_STORE_T zeroArray[2048/8];

  746. 

  747. 	if(newPPFlag)

  748. 	{

  749. 		ppMode= gPPMode[mode];

  750. 		

  751. 		postprocess2(src, src_stride, dst, dst_stride,

  752. 			horizontal_size, vertical_size, QP_store, QP_stride, &ppMode);

  753. 

  754. 		return;

  755. 	}

  756. 	

  757. 	if(QP_store==NULL)

  758. 	{

  759. 		QP_store= zeroArray;

  760. 		QP_stride= 0;

  761. 	}

  762. 

  763. 	ppMode.lumMode= mode;

  764. 	mode= ((mode&0xFF)>>4) | (mode&0xFFFFFF00);

  765. 	ppMode.chromMode= mode;

  766. 	ppMode.maxTmpNoise[0]= 700;

  767. 	ppMode.maxTmpNoise[1]= 1500;

  768. 	ppMode.maxTmpNoise[2]= 3000;

  769. 

  770. #ifdef HAVE_ODIVX_POSTPROCESS

  771. // Note: I could make this shit outside of this file, but it would mean one

  772. // more function call...

  773. 	if(use_old_pp){

  774. 	    odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,mode);

  775. 	    return;

  776. 	}

  777. #endif

  778. 

  779. 	postProcess(src[0], src_stride, dst[0], dst_stride,

  780. 		horizontal_size, vertical_size, QP_store, QP_stride, 0, &ppMode);

  781. 

  782. 	horizontal_size >>= 1;

  783. 	vertical_size   >>= 1;

  784. 	src_stride      >>= 1;

  785. 	dst_stride      >>= 1;

  786. 

  787. 	if(ppMode.chromMode)

  788. 	{

  789. 		postProcess(src[1], src_stride, dst[1], dst_stride,

  790. 			horizontal_size, vertical_size, QP_store, QP_stride, 1, &ppMode);

  791. 		postProcess(src[2], src_stride, dst[2], dst_stride,

  792. 			horizontal_size, vertical_size, QP_store, QP_stride, 2, &ppMode);

  793. 	}

  794. 	else if(src_stride == dst_stride)

  795. 	{

  796. 		memcpy(dst[1], src[1], src_stride*vertical_size);

  797. 		memcpy(dst[2], src[2], src_stride*vertical_size);

  798. 	}

  799. 	else

  800. 	{

  801. 		int y;

  802. 		for(y=0; y<vertical_size; y++)

  803. 		{

  804. 			memcpy(&(dst[1][y*dst_stride]), &(src[1][y*src_stride]), horizontal_size);

  805. 			memcpy(&(dst[2][y*dst_stride]), &(src[2][y*src_stride]), horizontal_size);

  806. 		}

  807. 	}

  808. 

  809. #if 0

  810. 		memset(dst[1], 128, dst_stride*vertical_size);

  811. 		memset(dst[2], 128, dst_stride*vertical_size);

  812. #endif

  813. }

  814. 

  815. void  postprocess2(unsigned char * src[], int src_stride,

  816.                  unsigned char * dst[], int dst_stride,

  817.                  int horizontal_size,   int vertical_size,

  818.                  QP_STORE_T *QP_store,  int QP_stride,

  819. 		 struct PPMode *mode)

  820. {

  821. 

  822. 	static QP_STORE_T zeroArray[2048/8];

  823. 	if(QP_store==NULL)

  824. 	{

  825. 		QP_store= zeroArray;

  826. 		QP_stride= 0;

  827. 	}

  828. 

  829. #ifdef HAVE_ODIVX_POSTPROCESS

  830. // Note: I could make this shit outside of this file, but it would mean one

  831. // more function call...

  832. 	if(use_old_pp){

  833. 	    odivx_postprocess(src,src_stride,dst,dst_stride,horizontal_size,vertical_size,QP_store,QP_stride,

  834. 	    mode->oldMode);

  835. 	    return;

  836. 	}

  837. #endif

  838. 

  839. 	postProcess(src[0], src_stride, dst[0], dst_stride,

  840. 		horizontal_size, vertical_size, QP_store, QP_stride, 0, mode);

  841. 

  842. 	horizontal_size >>= 1;

  843. 	vertical_size   >>= 1;

  844. 	src_stride      >>= 1;

  845. 	dst_stride      >>= 1;

  846. 

  847. 	if(mode->chromMode)

  848. 	{

  849. 		postProcess(src[1], src_stride, dst[1], dst_stride,

  850. 			horizontal_size, vertical_size, QP_store, QP_stride, 1, mode);

  851. 		postProcess(src[2], src_stride, dst[2], dst_stride,

  852. 			horizontal_size, vertical_size, QP_store, QP_stride, 2, mode);

  853. 	}

  854. 	else if(src_stride == dst_stride)

  855. 	{

  856. 		memcpy(dst[1], src[1], src_stride*vertical_size);

  857. 		memcpy(dst[2], src[2], src_stride*vertical_size);

  858. 	}

  859. 	else

  860. 	{

  861. 		int y;

  862. 		for(y=0; y<vertical_size; y++)

  863. 		{

  864. 			memcpy(&(dst[1][y*dst_stride]), &(src[1][y*src_stride]), horizontal_size);

  865. 			memcpy(&(dst[2][y*dst_stride]), &(src[2][y*src_stride]), horizontal_size);

  866. 		}

  867. 	}

  868. }

  869. 

  870. 

  871. /**

  872.  * gets the mode flags for a given quality (larger values mean slower but better postprocessing)

  873.  * with -npp it simply returns quality 

  874.  * 0 <= quality <= 6

  875.  */

  876. int getPpModeForQuality(int quality){

  877. 	int modes[1+GET_PP_QUALITY_MAX]= {

  878. 		0,

  879. #if 1

  880. 		// horizontal filters first

  881. 		LUM_H_DEBLOCK,

  882. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK,

  883. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK,

  884. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK,

  885. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING,

  886. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK | LUM_DERING | CHROM_DERING

  887. #else

  888. 		// vertical filters first

  889. 		LUM_V_DEBLOCK,

  890. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK,

  891. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK,

  892. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK,

  893. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING,

  894. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK | LUM_DERING | CHROM_DERING

  895. #endif

  896. 	};

  897. 

  898. #ifdef HAVE_ODIVX_POSTPROCESS

  899. 	int odivx_modes[1+GET_PP_QUALITY_MAX]= {

  900. 		0,

  901. 		PP_DEBLOCK_Y_H,

  902. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V,

  903. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H,

  904. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V,

  905. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y,

  906. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V|PP_DERING_Y|PP_DERING_C

  907. 	};

  908. 	if(use_old_pp) return odivx_modes[quality];

  909. #endif

  910. 	if(newPPFlag)	return quality;

  911. 	else		return modes[quality];

  912. }

  913.