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.
324 Commits
32 Branches
283MB
Tree: e89952aabe
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'e89952aabe'
${ noResults }
FFmpeg/postproc/postprocess.c

887 lines
26KB
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. int hFlatnessThreshold= 56 - 16;

  147. int vFlatnessThreshold= 56 - 16;

  148. int deringThreshold= 20;

  149. 

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

  151. double maxClippedThreshold= 0.01;

  152. 

  153. int maxAllowedY=234;

  154. int minAllowedY=16;

  155. 

  156. static struct PPFilter filters[]=

  157. {

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

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

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

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

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

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

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

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

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

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

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

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

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

  171. };

  172. 

  173. static char *replaceTable[]=

  174. {

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

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

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

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

  179. 	NULL //End Marker

  180. };

  181. 

  182. #ifdef CAN_COMPILE_X86_ASM

  183. static inline void unusedVariableWarningFixer()

  184. {

  185. if(

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

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

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

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

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

  191. }

  192. #endif

  193. 

  194. #ifdef TIMING

  195. static inline long long rdtsc()

  196. {

  197. 	long long l;

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

  199. 		: "=A" (l)

  200. 	);

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

  202. 	return l;

  203. }

  204. #endif

  205. 

  206. #ifdef CAN_COMPILE_X86_ASM

  207. static inline void prefetchnta(void *p)

  208. {

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

  210. 		: : "r" (p)

  211. 	);

  212. }

  213. 

  214. static inline void prefetcht0(void *p)

  215. {

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

  217. 		: : "r" (p)

  218. 	);

  219. }

  220. 

  221. static inline void prefetcht1(void *p)

  222. {

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

  224. 		: : "r" (p)

  225. 	);

  226. }

  227. 

  228. static inline void prefetcht2(void *p)

  229. {

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

  231. 		: : "r" (p)

  232. 	);

  233. }

  234. #endif

  235. 

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

  237. 

  238. /**

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

  240.  */

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

  242. {

  243. 	int numEq= 0;

  244. 	int y;

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

  246. 	{

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

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

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

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

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

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

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

  254. 		src+= stride;

  255. 	}

  256. 	return numEq > hFlatnessThreshold;

  257. }

  258. 

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

  260. {

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

  262. 

  263. 	return 1;

  264. }

  265. 

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

  267. {

  268. 	int y;

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

  270. 	{

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

  272. 

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

  274. 		{

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

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

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

  278. 

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

  280. 			d= MAX(d, 0);

  281. 

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

  283. 			d*= SIGN(-middleEnergy);

  284. 

  285. 			if(q>0)

  286. 			{

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

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

  289. 			}

  290. 			else

  291. 			{

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

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

  294. 			}

  295. 

  296.         		dst[3]-= d;

  297. 	        	dst[4]+= d;

  298. 		}

  299. 		dst+= stride;

  300. 	}

  301. }

  302. 

  303. /**

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

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

  306.  */

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

  308. {

  309. 

  310. 	int y;

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

  312. 	{

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

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

  315. 

  316. 		int sums[9];

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

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

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

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

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

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

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

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

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

  326. 

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

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

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

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

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

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

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

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

  335. 

  336. 		dst+= stride;

  337. 	}

  338. }

  339. 

  340. /**

  341.  * Experimental Filter 1 (Horizontal)

  342.  * will not damage linear gradients

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

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

  345.  * MMX2 version does correct clipping C version doesnt

  346.  * not identical with the vertical one

  347.  */

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

  349. {

  350. 	int y;

  351. 	static uint64_t *lut= NULL;

  352. 	if(lut==NULL)

  353. 	{

  354. 		int i;

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

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

  357. 		{

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

  359. /*

  360. //Simulate 112242211 9-Tap filter

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

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

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

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

  365. */

  366. //Simulate piecewise linear interpolation

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

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

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

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

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

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

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

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

  375. 

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

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

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

  379. 		}

  380. 	}

  381. 

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

  383. 	{

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

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

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

  387. 

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

  389. 

  390. 		if(d < QP)

  391. 		{

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

  393. 

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

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

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

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

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

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

  400. 

  401. 		}

  402. 		src+=stride;

  403. 	}

  404. }

  405. 

  406. 

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

  408. //Plain C versions

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

  410. #define COMPILE_C

  411. #endif

  412. 

  413. #ifdef CAN_COMPILE_X86_ASM

  414. 

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

  416. #define COMPILE_MMX

  417. #endif

  418. 

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

  420. #define COMPILE_MMX2

  421. #endif

  422. 

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

  424. #define COMPILE_3DNOW

  425. #endif

  426. #endif //CAN_COMPILE_X86_ASM

  427. 

  428. #undef HAVE_MMX

  429. #undef HAVE_MMX2

  430. #undef HAVE_3DNOW

  431. #undef ARCH_X86

  432. 

  433. #ifdef COMPILE_C

  434. #undef HAVE_MMX

  435. #undef HAVE_MMX2

  436. #undef HAVE_3DNOW

  437. #undef ARCH_X86

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

  439. #include "postprocess_template.c"

  440. #endif

  441. 

  442. //MMX versions

  443. #ifdef COMPILE_MMX

  444. #undef RENAME

  445. #define HAVE_MMX

  446. #undef HAVE_MMX2

  447. #undef HAVE_3DNOW

  448. #define ARCH_X86

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

  450. #include "postprocess_template.c"

  451. #endif

  452. 

  453. //MMX2 versions

  454. #ifdef COMPILE_MMX2

  455. #undef RENAME

  456. #define HAVE_MMX

  457. #define HAVE_MMX2

  458. #undef HAVE_3DNOW

  459. #define ARCH_X86

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

  461. #include "postprocess_template.c"

  462. #endif

  463. 

  464. //3DNOW versions

  465. #ifdef COMPILE_3DNOW

  466. #undef RENAME

  467. #define HAVE_MMX

  468. #undef HAVE_MMX2

  469. #define HAVE_3DNOW

  470. #define ARCH_X86

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

  472. #include "postprocess_template.c"

  473. #endif

  474. 

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

  476. 

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

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

  479. {

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

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

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

  483. #ifdef RUNTIME_CPUDETECT

  484. #ifdef CAN_COMPILE_X86_ASM

  485. 	// ordered per speed fasterst first

  486. 	if(gCpuCaps.hasMMX2)

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

  488. 	else if(gCpuCaps.has3DNow)

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

  490. 	else if(gCpuCaps.hasMMX)

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

  492. 	else

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

  494. #else

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

  496. #endif

  497. #else //RUNTIME_CPUDETECT

  498. #ifdef HAVE_MMX2

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

  500. #elif defined (HAVE_3DNOW)

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

  502. #elif defined (HAVE_MMX)

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

  504. #else

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

  506. #endif

  507. #endif //!RUNTIME_CPUDETECT

  508. }

  509. 

  510. #ifdef HAVE_ODIVX_POSTPROCESS

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

  512. int use_old_pp=0;

  513. #endif

  514. 

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

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

  517. 

  518. /* -pp Command line Help

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

  520. 

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

  522. 

  523. long form example:

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

  525. short form example:

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

  527. more examples:

  528. -pp tn:64:128:256

  529. 

  530. Filters			Options

  531. short	long name	short	long option	Description

  532. *	*		a	autoq		cpu power dependant enabler

  533. 			c	chrom		chrominance filtring enabled

  534. 			y	nochrom		chrominance filtring disabled

  535. hb	hdeblock				horizontal deblocking filter

  536. vb	vdeblock				vertical deblocking filter

  537. vr	rkvdeblock

  538. h1	x1hdeblock				Experimental horizontal deblock filter 1

  539. v1	x1vdeblock				Experimental vertical deblock filter 1

  540. dr	dering					not implemented yet

  541. al	autolevels				automatic brightness / contrast fixer

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

  543. lb	linblenddeint				linear blend deinterlacer

  544. li	linipoldeint				linear interpolating deinterlacer

  545. ci	cubicipoldeint				cubic interpolating deinterlacer

  546. md	mediandeint				median deinterlacer

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

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

  549. tn	tmpnoise	(3 Thresholds)		Temporal Noise Reducer

  550. */

  551. 

  552. /**

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

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

  555.  * quality is a number from 0 to GET_PP_QUALITY_MAX

  556.  */

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

  558. {

  559. 	char temp[GET_MODE_BUFFER_SIZE];

  560. 	char *p= temp;

  561. 	char *filterDelimiters= ",";

  562. 	char *optionDelimiters= ":";

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

  564. 	char *filterToken;

  565. 

  566. 	strncpy(temp, name, GET_MODE_BUFFER_SIZE);

  567. 

  568. 	printf("%s\n", name);

  569. 

  570. 	for(;;){

  571. 		char *filterName;

  572. 		int q= 1000000; //GET_PP_QUALITY_MAX;

  573. 		int chrom=-1;

  574. 		char *option;

  575. 		char *options[OPTIONS_ARRAY_SIZE];

  576. 		int i;

  577. 		int filterNameOk=0;

  578. 		int numOfUnknownOptions=0;

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

  580. 

  581. 		filterToken= strtok(p, filterDelimiters);

  582. 		if(filterToken == NULL) break;

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

  584. 		filterName= strtok(filterToken, optionDelimiters);

  585. 		printf("%s::%s\n", filterToken, filterName);

  586. 

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

  588. 		{

  589. 			enable=0;

  590. 			filterName++;

  591. 		}

  592. 

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

  594. 			option= strtok(NULL, optionDelimiters);

  595. 			if(option == NULL) break;

  596. 

  597. 			printf("%s\n", option);

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

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

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

  601. 			else

  602. 			{

  603. 				options[numOfUnknownOptions] = option;

  604. 				numOfUnknownOptions++;

  605. 			}

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

  607. 		}

  608. 		options[numOfUnknownOptions] = NULL;

  609. 

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

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

  612. 		{

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

  614. 			{

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

  616. 				int plen;

  617. 				int spaceLeft;

  618. 

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

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

  621. 

  622. 				plen= strlen(p);

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

  624. 				if(spaceLeft + newlen  >= GET_MODE_BUFFER_SIZE)

  625. 				{

  626. 					ppMode.error++;

  627. 					break;

  628. 				}

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

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

  631. 				filterNameOk=1;

  632. 			}

  633. 		}

  634. 

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

  636. 		{

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

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

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

  640. 			{

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

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

  643. 

  644. 				filterNameOk=1;

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

  646. 

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

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

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

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

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

  652. 

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

  654. 				{

  655. 					int o;

  656. 					ppMode.minAllowedY= 16;

  657. 					ppMode.maxAllowedY= 234;

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

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

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

  661. 						{

  662. 							ppMode.minAllowedY= 0;

  663. 							ppMode.maxAllowedY= 255;

  664. 							numOfUnknownOptions--;

  665. 						}

  666. 				}

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

  668. 				{

  669. 					int o;

  670. 					int numOfNoises=0;

  671. 					ppMode.maxTmpNoise[0]= 150;

  672. 					ppMode.maxTmpNoise[1]= 200;

  673. 					ppMode.maxTmpNoise[2]= 400;

  674. 

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

  676. 					{

  677. 						char *tail;

  678. 						ppMode.maxTmpNoise[numOfNoises]=

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

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

  681. 						{

  682. 							numOfNoises++;

  683. 							numOfUnknownOptions--;

  684. 							if(numOfNoises >= 3) break;

  685. 						}

  686. 					}

  687. 				}

  688. 			}

  689. 		}

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

  691. 		ppMode.error += numOfUnknownOptions;

  692. 	}

  693. 

  694. #ifdef HAVE_ODIVX_POSTPROCESS

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

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

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

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

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

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

  701. #endif

  702. 

  703. 	return ppMode;

  704. }

  705. 

  706. /**

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

  708.  */

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

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

  711.                  int horizontal_size,   int vertical_size,

  712.                  QP_STORE_T *QP_store,  int QP_stride,

  713. 					  int mode)

  714. {

  715. 	struct PPMode ppMode;

  716. 	static QP_STORE_T zeroArray[2048/8];

  717. /*

  718. 	static int qual=0;

  719. 

  720. 	ppMode= getPPModeByNameAndQuality("fast,default,-hdeblock,-vdeblock,tmpnoise:150:200:300", qual);

  721. 	printf("OK\n");

  722. 	qual++;

  723. 	qual%=7;

  724. 	printf("\n%X %X %X %X :%d: %d %d %d\n", ppMode.lumMode, ppMode.chromMode, ppMode.oldMode, ppMode.error,

  725. 		qual, ppMode.maxTmpNoise[0], ppMode.maxTmpNoise[1], ppMode.maxTmpNoise[2]);

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

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

  728. 

  729. 	return;

  730. */

  731. 	if(QP_store==NULL)

  732. 	{

  733. 		QP_store= zeroArray;

  734. 		QP_stride= 0;

  735. 	}

  736. 

  737. 	ppMode.lumMode= mode;

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

  739. 	ppMode.chromMode= mode;

  740. 	ppMode.maxTmpNoise[0]= 700;

  741. 	ppMode.maxTmpNoise[1]= 1500;

  742. 	ppMode.maxTmpNoise[2]= 3000;

  743. 

  744. #ifdef HAVE_ODIVX_POSTPROCESS

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

  746. // more function call...

  747. 	if(use_old_pp){

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

  749. 	    return;

  750. 	}

  751. #endif

  752. 

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

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

  755. 

  756. 	horizontal_size >>= 1;

  757. 	vertical_size   >>= 1;

  758. 	src_stride      >>= 1;

  759. 	dst_stride      >>= 1;

  760. 

  761. 	if(ppMode.chromMode)

  762. 	{

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

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

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

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

  767. 	}

  768. 	else if(src_stride == dst_stride)

  769. 	{

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

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

  772. 	}

  773. 	else

  774. 	{

  775. 		int y;

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

  777. 		{

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

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

  780. 		}

  781. 	}

  782. 

  783. #if 0

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

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

  786. #endif

  787. }

  788. 

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

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

  791.                  int horizontal_size,   int vertical_size,

  792.                  QP_STORE_T *QP_store,  int QP_stride,

  793. 		 struct PPMode *mode)

  794. {

  795. 

  796. 	static QP_STORE_T zeroArray[2048/8];

  797. 	if(QP_store==NULL)

  798. 	{

  799. 		QP_store= zeroArray;

  800. 		QP_stride= 0;

  801. 	}

  802. 

  803. #ifdef HAVE_ODIVX_POSTPROCESS

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

  805. // more function call...

  806. 	if(use_old_pp){

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

  808. 	    mode->oldMode);

  809. 	    return;

  810. 	}

  811. #endif

  812. 

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

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

  815. 

  816. 	horizontal_size >>= 1;

  817. 	vertical_size   >>= 1;

  818. 	src_stride      >>= 1;

  819. 	dst_stride      >>= 1;

  820. 

  821. 	if(mode->chromMode)

  822. 	{

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

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

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

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

  827. 	}

  828. 	else if(src_stride == dst_stride)

  829. 	{

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

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

  832. 	}

  833. 	else

  834. 	{

  835. 		int y;

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

  837. 		{

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

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

  840. 		}

  841. 	}

  842. }

  843. 

  844. 

  845. /**

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

  847.  * 0 <= quality <= 6

  848.  */

  849. int getPpModeForQuality(int quality){

  850. 	int modes[1+GET_PP_QUALITY_MAX]= {

  851. 		0,

  852. #if 1

  853. 		// horizontal filters first

  854. 		LUM_H_DEBLOCK,

  855. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK,

  856. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK,

  857. 		LUM_H_DEBLOCK | LUM_V_DEBLOCK | CHROM_H_DEBLOCK | CHROM_V_DEBLOCK,

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

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

  860. #else

  861. 		// vertical filters first

  862. 		LUM_V_DEBLOCK,

  863. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK,

  864. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK,

  865. 		LUM_V_DEBLOCK | LUM_H_DEBLOCK | CHROM_V_DEBLOCK | CHROM_H_DEBLOCK,

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

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

  868. #endif

  869. 	};

  870. 

  871. #ifdef HAVE_ODIVX_POSTPROCESS

  872. 	int odivx_modes[1+GET_PP_QUALITY_MAX]= {

  873. 		0,

  874. 		PP_DEBLOCK_Y_H,

  875. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V,

  876. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H,

  877. 		PP_DEBLOCK_Y_H|PP_DEBLOCK_Y_V|PP_DEBLOCK_C_H|PP_DEBLOCK_C_V,

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

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

  880. 	};

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

  882. #endif

  883. 	return modes[quality];

  884. }

  885.