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.
151 Commits
32 Branches
411MB
Tree: 3057fa6691
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '3057fa6691'
${ noResults }
FFmpeg/postproc/postprocess.c

1552 lines
42KB
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

  21. isVertDC		Ec	Ec

  22. isVertMinMaxOk		Ec	Ec

  23. doVertLowPass		E		e

  24. doVertDefFilter		Ec	Ec	Ec

  25. isHorizDC		Ec	Ec

  26. isHorizMinMaxOk		a

  27. doHorizLowPass		E		a

  28. doHorizDefFilter	E		a

  29. deRing

  30. 

  31. E = Exact implementation

  32. e = allmost exact implementation

  33. a = alternative / approximate impl

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

  35. */

  36. 

  37. /*

  38. TODO:

  39. verify that everything workes as it should

  40. reduce the time wasted on the mem transfer

  41. implement dering

  42. implement everything in C at least

  43. figure range of QP out (assuming <256 for now)

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

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

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

  47. ...

  48. 

  49. Notes:

  50. 

  51. */

  52. 

  53. 

  54. #include <inttypes.h>

  55. #include <stdio.h>

  56. #include "../config.h"

  57. #include "postprocess.h"

  58. //#undef HAVE_MMX2

  59. //#undef HAVE_MMX

  60. 

  61. 

  62. 

  63. static uint64_t packedYOffset=	0x0000000000000000LL;

  64. static uint64_t packedYScale=	0x0100010001000100LL;

  65. static uint64_t w05=		0x0005000500050005LL;

  66. static uint64_t w20=		0x0020002000200020LL;

  67. static uint64_t w1400=		0x1400140014001400LL;

  68. static uint64_t bm00000001=	0x00000000000000FFLL;

  69. static uint64_t bm00010000=	0x000000FF00000000LL;

  70. static uint64_t bm00001000=	0x00000000FF000000LL;

  71. static uint64_t bm10000000=	0xFF00000000000000LL;

  72. static uint64_t bm10000001=	0xFF000000000000FFLL;

  73. static uint64_t bm11000011=	0xFFFF00000000FFFFLL;

  74. static uint64_t bm00011000=	0x000000FFFF000000LL;

  75. static uint64_t bm00110011=	0x0000FFFF0000FFFFLL;

  76. static uint64_t bm11001100=	0xFFFF0000FFFF0000LL;

  77. static uint64_t b00= 		0x0000000000000000LL;

  78. static uint64_t b02= 		0x0202020202020202LL;

  79. static uint64_t b0F= 		0x0F0F0F0F0F0F0F0FLL;

  80. static uint64_t bFF= 		0xFFFFFFFFFFFFFFFFLL;

  81. static uint64_t b7E= 		0x7E7E7E7E7E7E7E7ELL;

  82. static uint64_t b7C= 		0x7C7C7C7C7C7C7C7CLL;

  83. static uint64_t b3F= 		0x3F3F3F3F3F3F3F3FLL;

  84. static uint64_t temp0=0;

  85. static uint64_t temp1=0;

  86. static uint64_t temp2=0;

  87. static uint64_t temp3=0;

  88. static uint64_t temp4=0;

  89. static uint64_t temp5=0;

  90. static uint64_t pQPb=0;

  91. static uint8_t tempBlock[16*16];

  92. 

  93. int hFlatnessThreshold= 56 - 16;

  94. int vFlatnessThreshold= 56 - 16;

  95. 

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

  97. double maxClippedThreshold= 0.01;

  98. 

  99. int maxAllowedY=255;

  100. //FIXME can never make a movie´s black brighter (anyone needs that?)

  101. int minAllowedY=0;

  102. 

  103. 

  104. static inline long long rdtsc()

  105. {

  106. 	long long l;

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

  108. 		: "=A" (l)

  109. 	);

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

  111. 	return l;

  112. }

  113. 

  114. static inline void prefetchnta(void *p)

  115. {

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

  117. 		: : "r" (p)

  118. 	);

  119. }

  120. 

  121. static inline void prefetcht0(void *p)

  122. {

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

  124. 		: : "r" (p)

  125. 	);

  126. }

  127. 

  128. static inline void prefetcht1(void *p)

  129. {

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

  131. 		: : "r" (p)

  132. 	);

  133. }

  134. 

  135. static inline void prefetcht2(void *p)

  136. {

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

  138. 		: : "r" (p)

  139. 	);

  140. }

  141. 

  142. //FIXME? |255-0| = 1 (shouldnt be a problem ...)

  143. /**

  144.  * Check if the middle 8x8 Block in the given 8x10 block is flat

  145.  */

  146. static inline bool isVertDC(uint8_t src[], int stride){

  147. //	return true;

  148. 	int numEq= 0;

  149. 	src+= stride; // src points to begin of the 8x8 Block

  150. #ifdef HAVE_MMX

  151. 	asm volatile(

  152. //		"int $3 \n\t"

  153. 		"pushl %1\n\t"

  154. 		"movq b7E, %%mm7					\n\t" // mm7 = 0x7F

  155. 		"movq b7C, %%mm6					\n\t" // mm6 = 0x7D

  156. 		"movq (%1), %%mm0				\n\t"

  157. 		"addl %2, %1					\n\t"

  158. 		"movq (%1), %%mm1				\n\t"

  159. 		"psubb %%mm1, %%mm0				\n\t" // mm0 = differnece

  160. 		"paddb %%mm7, %%mm0				\n\t"

  161. 		"pcmpgtb %%mm6, %%mm0				\n\t"

  162. 

  163. 		"addl %2, %1					\n\t"

  164. 		"movq (%1), %%mm2				\n\t"

  165. 		"psubb %%mm2, %%mm1				\n\t"

  166. 		"paddb %%mm7, %%mm1				\n\t"

  167. 		"pcmpgtb %%mm6, %%mm1				\n\t"

  168. 		"paddb %%mm1, %%mm0				\n\t"

  169. 

  170. 		"addl %2, %1					\n\t"

  171. 		"movq (%1), %%mm1				\n\t"

  172. 		"psubb %%mm1, %%mm2				\n\t"

  173. 		"paddb %%mm7, %%mm2				\n\t"

  174. 		"pcmpgtb %%mm6, %%mm2				\n\t"

  175. 		"paddb %%mm2, %%mm0				\n\t"

  176. 

  177. 		"addl %2, %1					\n\t"

  178. 		"movq (%1), %%mm2				\n\t"

  179. 		"psubb %%mm2, %%mm1				\n\t"

  180. 		"paddb %%mm7, %%mm1				\n\t"

  181. 		"pcmpgtb %%mm6, %%mm1				\n\t"

  182. 		"paddb %%mm1, %%mm0				\n\t"

  183. 

  184. 		"addl %2, %1					\n\t"

  185. 		"movq (%1), %%mm1				\n\t"

  186. 		"psubb %%mm1, %%mm2				\n\t"

  187. 		"paddb %%mm7, %%mm2				\n\t"

  188. 		"pcmpgtb %%mm6, %%mm2				\n\t"

  189. 		"paddb %%mm2, %%mm0				\n\t"

  190. 

  191. 		"addl %2, %1					\n\t"

  192. 		"movq (%1), %%mm2				\n\t"

  193. 		"psubb %%mm2, %%mm1				\n\t"

  194. 		"paddb %%mm7, %%mm1				\n\t"

  195. 		"pcmpgtb %%mm6, %%mm1				\n\t"

  196. 		"paddb %%mm1, %%mm0				\n\t"

  197. 

  198. 		"addl %2, %1					\n\t"

  199. 		"movq (%1), %%mm1				\n\t"

  200. 		"psubb %%mm1, %%mm2				\n\t"

  201. 		"paddb %%mm7, %%mm2				\n\t"

  202. 		"pcmpgtb %%mm6, %%mm2				\n\t"

  203. 		"paddb %%mm2, %%mm0				\n\t"

  204. 

  205. 		"						\n\t"

  206. 		"movq %%mm0, %%mm1				\n\t"

  207. 		"psrlw $8, %%mm0				\n\t"

  208. 		"paddb %%mm1, %%mm0				\n\t"

  209. 		"movq %%mm0, %%mm1				\n\t"

  210. 		"psrlq $16, %%mm0				\n\t"

  211. 		"paddb %%mm1, %%mm0				\n\t"

  212. 		"movq %%mm0, %%mm1				\n\t"

  213. 		"psrlq $32, %%mm0				\n\t"

  214. 		"paddb %%mm1, %%mm0				\n\t"

  215. 		"popl %1\n\t"

  216. 		"movd %%mm0, %0					\n\t"

  217. 		: "=r" (numEq)

  218. 		: "r" (src), "r" (stride)

  219. 		);

  220. //	printf("%d\n", numEq);

  221. 	numEq= (256 - (numEq & 0xFF)) &0xFF;

  222. 

  223. //	int asmEq= numEq;

  224. //	numEq=0;

  225. //	uint8_t *temp= src;

  226. 

  227. #else

  228. 	for(int y=0; y<BLOCK_SIZE-1; y++)

  229. 	{

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

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

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

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

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

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

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

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

  238. 		src+= stride;

  239. 	}

  240. #endif

  241. /*	if(abs(numEq - asmEq) > 0)

  242. 	{

  243. 		printf("\nasm:%d  c:%d\n", asmEq, numEq);

  244. 		for(int y=0; y<8; y++)

  245. 		{

  246. 			for(int x=0; x<8; x++)

  247. 			{

  248. 				printf("%d ", temp[x + y*stride]);

  249. 			}

  250. 			printf("\n");

  251. 		}

  252. 	}

  253. */

  254. 	return numEq > vFlatnessThreshold;

  255. }

  256. 

  257. static inline bool isVertMinMaxOk(uint8_t src[], int stride, int QP)

  258. {

  259. #ifdef HAVE_MMX

  260. 	int isOk;

  261. 	asm volatile(

  262. //		"int $3 \n\t"

  263. 		"movq (%1, %2), %%mm0				\n\t"

  264. 		"movq (%1, %2, 8), %%mm1			\n\t"

  265. 		"movq %%mm0, %%mm2				\n\t"

  266. 		"psubusb %%mm1, %%mm0				\n\t"

  267. 		"psubusb %%mm2, %%mm1				\n\t"

  268. 		"por %%mm1, %%mm0				\n\t" // ABS Diff

  269. 

  270. 		"movq pQPb, %%mm7				\n\t" // QP,..., QP

  271. 		"paddusb %%mm7, %%mm7				\n\t" // 2QP ... 2QP

  272. 		"psubusb %%mm7, %%mm0				\n\t" // Diff <= 2QP -> 0

  273. 		"pcmpeqd b00, %%mm0				\n\t"

  274. 		"psrlq $16, %%mm0				\n\t"

  275. 		"pcmpeqd bFF, %%mm0				\n\t"

  276. //		"movd %%mm0, (%1, %2, 4)\n\t"

  277. 		"movd %%mm0, %0					\n\t"

  278. 		: "=r" (isOk)

  279. 		: "r" (src), "r" (stride)

  280. 		);

  281. 	return isOk;

  282. #else

  283. 

  284. 	int isOk2= true;

  285. 	for(int x=0; x<BLOCK_SIZE; x++)

  286. 	{

  287. 		if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=false;

  288. 	}

  289. /*	if(isOk && !isOk2 || !isOk && isOk2)

  290. 	{

  291. 		printf("\nasm:%d  c:%d QP:%d\n", isOk, isOk2, QP);

  292. 		for(int y=0; y<9; y++)

  293. 		{

  294. 			for(int x=0; x<8; x++)

  295. 			{

  296. 				printf("%d ", src[x + y*stride]);

  297. 			}

  298. 			printf("\n");

  299. 		}

  300. 	} */

  301. 

  302. 	return isOk2;

  303. #endif

  304. 

  305. }

  306. 

  307. /**

  308.  * Do a vertical low pass filter on the 8x10 block (only write to the 8x8 block in the middle)

  309.  * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16

  310.  */

  311. static inline void doVertLowPass(uint8_t *src, int stride, int QP)

  312. {

  313. //	QP= 64;

  314. 

  315. #ifdef HAVE_MMX2

  316. 	asm volatile(	//"movv %0 %1 %2\n\t"

  317. 		"pushl %0 \n\t"

  318. 		"movq pQPb, %%mm0				\n\t"  // QP,..., QP

  319. //		"movq bFF  , %%mm0				\n\t"  // QP,..., QP

  320. 

  321. 		"movq (%0), %%mm6				\n\t"

  322. 		"movq (%0, %1), %%mm5				\n\t"

  323. 		"movq %%mm5, %%mm1				\n\t"

  324. 		"movq %%mm6, %%mm2				\n\t"

  325. 		"psubusb %%mm6, %%mm5				\n\t"

  326. 		"psubusb %%mm1, %%mm2				\n\t"

  327. 		"por %%mm5, %%mm2				\n\t" // ABS Diff of lines

  328. 		"psubusb %%mm0, %%mm2				\n\t" // diff <= QP -> 0

  329. 		"pcmpeqb b00, %%mm2				\n\t" // diff <= QP -> FF

  330. 

  331. 		"pand %%mm2, %%mm6				\n\t"

  332. 		"pandn %%mm1, %%mm2				\n\t"

  333. 		"por %%mm2, %%mm6				\n\t"// First Line to Filter

  334. 

  335. 		"movq (%0, %1, 8), %%mm5			\n\t"

  336. 		"leal (%0, %1, 4), %%eax			\n\t"

  337. 		"leal (%0, %1, 8), %%ebx			\n\t"

  338. 		"subl %1, %%ebx					\n\t"

  339. 		"addl %1, %0					\n\t" // %0 points to line 1 not 0

  340. 		"movq (%0, %1, 8), %%mm7			\n\t"

  341. 		"movq %%mm5, %%mm1				\n\t"

  342. 		"movq %%mm7, %%mm2				\n\t"

  343. 		"psubusb %%mm7, %%mm5				\n\t"

  344. 		"psubusb %%mm1, %%mm2				\n\t"

  345. 		"por %%mm5, %%mm2				\n\t" // ABS Diff of lines

  346. 		"psubusb %%mm0, %%mm2				\n\t" // diff <= QP -> 0

  347. 		"pcmpeqb b00, %%mm2				\n\t" // diff <= QP -> FF

  348. 

  349. 		"pand %%mm2, %%mm7				\n\t"

  350. 		"pandn %%mm1, %%mm2				\n\t"

  351. 		"por %%mm2, %%mm7				\n\t" // First Line to Filter

  352. 

  353. 

  354. 		// 	1	2	3	4	5	6	7	8

  355. 		//	%0	%0+%1	%0+2%1	eax	%0+4%1	eax+2%1	ebx	eax+4%1

  356. 		// 6 4 2 2 1 1

  357. 		// 6 4 4 2

  358. 		// 6 8 2

  359. /*

  360. 		"movq %%mm6, %%mm2				\n\t" //1

  361. 		"movq %%mm6, %%mm3				\n\t" //1

  362. 		"paddusb b02, %%mm3 				\n\t"

  363. 		"psrlw $2, %%mm3				\n\t" //1	/4

  364. 		"pand b3F, %%mm3				\n\t"

  365. 		"psubb %%mm3, %%mm2				\n\t"

  366. 		"movq (%0, %1), %%mm0				\n\t" //  1

  367. 		"movq %%mm0, %%mm1				\n\t" //  1

  368. 		"paddusb b02, %%mm0				\n\t"

  369. 		"psrlw $2, %%mm0				\n\t" //  1	/4

  370. 		"pand b3F, %%mm0				\n\t"

  371. 		"paddusb %%mm2, %%mm0				\n\t" //3 1	/4

  372. */

  373. 		"movq (%0, %1), %%mm0				\n\t" //  1

  374. 		"movq %%mm0, %%mm1				\n\t" //  1

  375. 		"pavgb %%mm6, %%mm0				\n\t" //1 1	/2

  376. 		"pavgb %%mm6, %%mm0				\n\t" //3 1	/4

  377. 

  378. 		"movq (%0, %1, 4), %%mm2			\n\t" //     1

  379. 		"movq %%mm2, %%mm5				\n\t" //     1

  380. 		"pavgb (%%eax), %%mm2				\n\t" //    11	/2

  381. 		"pavgb (%0, %1, 2), %%mm2			\n\t" //   211	/4

  382. 		"movq %%mm2, %%mm3				\n\t" //   211	/4

  383. 		"movq (%0), %%mm4				\n\t" // 1

  384. 		"pavgb %%mm4, %%mm3				\n\t" // 4 211	/8

  385. 		"pavgb %%mm0, %%mm3				\n\t" //642211	/16

  386. 		"movq %%mm3, (%0)				\n\t" // X

  387. 		// mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9

  388. 		"movq %%mm1, %%mm0				\n\t" //  1

  389. 		"pavgb %%mm6, %%mm0				\n\t" //1 1	/2

  390. 		"movq %%mm4, %%mm3				\n\t" // 1

  391. 		"pavgb (%0,%1,2), %%mm3				\n\t" // 1 1	/2

  392. 		"pavgb (%%eax,%1,2), %%mm5			\n\t" //     11	/2

  393. 		"pavgb (%%eax), %%mm5				\n\t" //    211 /4

  394. 		"pavgb %%mm5, %%mm3				\n\t" // 2 2211 /8

  395. 		"pavgb %%mm0, %%mm3				\n\t" //4242211 /16

  396. 		"movq %%mm3, (%0,%1)				\n\t" //  X

  397. 		// mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9

  398. 		"pavgb %%mm4, %%mm6				\n\t" //11	/2

  399. 		"movq (%%ebx), %%mm0				\n\t" //       1

  400. 		"pavgb (%%eax, %1, 2), %%mm0			\n\t" //      11/2

  401. 		"movq %%mm0, %%mm3				\n\t" //      11/2

  402. 		"pavgb %%mm1, %%mm0				\n\t" //  2   11/4

  403. 		"pavgb %%mm6, %%mm0				\n\t" //222   11/8

  404. 		"pavgb %%mm2, %%mm0				\n\t" //22242211/16

  405. 		"movq (%0, %1, 2), %%mm2			\n\t" //   1

  406. 		"movq %%mm0, (%0, %1, 2)			\n\t" //   X

  407. 		// mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9

  408. 		"movq (%%eax, %1, 4), %%mm0			\n\t" //        1

  409. 		"pavgb (%%ebx), %%mm0				\n\t" //       11	/2

  410. 		"pavgb %%mm0, %%mm6				\n\t" //11     11	/4

  411. 		"pavgb %%mm1, %%mm4				\n\t" // 11		/2

  412. 		"pavgb %%mm2, %%mm1				\n\t" //  11		/2

  413. 		"pavgb %%mm1, %%mm6				\n\t" //1122   11	/8

  414. 		"pavgb %%mm5, %%mm6				\n\t" //112242211	/16

  415. 		"movq (%%eax), %%mm5				\n\t" //    1

  416. 		"movq %%mm6, (%%eax)				\n\t" //    X

  417. 		// mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9

  418. 		"movq (%%eax, %1, 4), %%mm6			\n\t" //        1

  419. 		"pavgb %%mm7, %%mm6				\n\t" //        11	/2

  420. 		"pavgb %%mm4, %%mm6				\n\t" // 11     11	/4

  421. 		"pavgb %%mm3, %%mm6				\n\t" // 11   2211	/8

  422. 		"pavgb %%mm5, %%mm2				\n\t" //   11		/2

  423. 		"movq (%0, %1, 4), %%mm4			\n\t" //     1

  424. 		"pavgb %%mm4, %%mm2				\n\t" //   112		/4

  425. 		"pavgb %%mm2, %%mm6				\n\t" // 112242211	/16

  426. 		"movq %%mm6, (%0, %1, 4)			\n\t" //     X

  427. 		// mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9

  428. 		"pavgb %%mm7, %%mm1				\n\t" //  11     2	/4

  429. 		"pavgb %%mm4, %%mm5				\n\t" //    11		/2

  430. 		"pavgb %%mm5, %%mm0				\n\t" //    11 11	/4

  431. 		"movq (%%eax, %1, 2), %%mm6			\n\t" //      1

  432. 		"pavgb %%mm6, %%mm1				\n\t" //  11  4  2	/8

  433. 		"pavgb %%mm0, %%mm1				\n\t" //  11224222	/16

  434. //		"pxor %%mm1, %%mm1 \n\t"

  435. 		"movq %%mm1, (%%eax, %1, 2)			\n\t" //      X

  436. 		// mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9

  437. 		"pavgb (%%ebx), %%mm2				\n\t" //   112 4	/8

  438. 		"movq (%%eax, %1, 4), %%mm0			\n\t" //        1

  439. 		"pavgb %%mm0, %%mm6				\n\t" //      1 1	/2

  440. 		"pavgb %%mm7, %%mm6				\n\t" //      1 12	/4

  441. 		"pavgb %%mm2, %%mm6				\n\t" //   1122424	/4

  442. //		"pxor %%mm6, %%mm6 \n\t"

  443. 		"movq %%mm6, (%%ebx)				\n\t" //       X

  444. 		// mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9

  445. 		"pavgb %%mm7, %%mm5				\n\t" //    11   2	/4

  446. 		"pavgb %%mm7, %%mm5				\n\t" //    11   6	/8

  447. 

  448. 		"pavgb %%mm3, %%mm0				\n\t" //      112	/4

  449. 		"pavgb %%mm0, %%mm5				\n\t" //    112246	/16

  450. //		"pxor %%mm5, %%mm5 \n\t"

  451. //		"movq pQPb, %%mm5 \n\t"

  452. 		"movq %%mm5, (%%eax, %1, 4)			\n\t" //        X

  453. 		"popl %0\n\t"

  454. 

  455. 		:

  456. 		: "r" (src), "r" (stride)

  457. 		: "%eax", "%ebx"

  458. 	);

  459. 

  460. #else

  461. 	const int l1= stride;

  462. 	const int l2= stride + l1;

  463. 	const int l3= stride + l2;

  464. 	const int l4= stride + l3;

  465. 	const int l5= stride + l4;

  466. 	const int l6= stride + l5;

  467. 	const int l7= stride + l6;

  468. 	const int l8= stride + l7;

  469. 	const int l9= stride + l8;

  470. 

  471. 	for(int x=0; x<BLOCK_SIZE; x++)

  472. 	{

  473. 		const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1];

  474. 		const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8];

  475. 

  476. 		int sums[9];

  477. 		sums[0] = first + src[l1];

  478. 		sums[1] = src[l1] + src[l2];

  479. 		sums[2] = src[l2] + src[l3];

  480. 		sums[3] = src[l3] + src[l4];

  481. 		sums[4] = src[l4] + src[l5];

  482. 		sums[5] = src[l5] + src[l6];

  483. 		sums[6] = src[l6] + src[l7];

  484. 		sums[7] = src[l7] + src[l8];

  485. 		sums[8] = src[l8] + last;

  486. 

  487. 		src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;

  488. 		src[l2]= ((src[l2]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4;

  489. 		src[l3]= ((src[l3]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4;

  490. 		src[l4]= ((src[l4]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4;

  491. 		src[l5]= ((src[l5]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4;

  492. 		src[l6]= ((src[l6]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4;

  493. 		src[l7]= ((last + src[l7]<<2) + (src[l8] + sums[5]<<1) + sums[3] + 8)>>4;

  494. 		src[l8]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4;

  495. 

  496. 		src++;

  497. 	}

  498. 

  499. #endif

  500. }

  501. 

  502. static inline void doVertDefFilter(uint8_t src[], int stride, int QP)

  503. {

  504. #ifdef HAVE_MMX

  505. 	src+= stride;

  506. 	//FIXME try pmul for *5 stuff

  507. //	src[0]=0;

  508. 	asm volatile(

  509. 		"pxor %%mm7, %%mm7				\n\t"

  510. 		"leal (%0, %1), %%eax				\n\t"

  511. 		"leal (%%eax, %1, 4), %%ebx			\n\t"

  512. //	0	1	2	3	4	5	6	7

  513. //	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	ebx+%1	ebx+2%1

  514. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1

  515. 

  516. 		"movq (%0), %%mm0				\n\t"

  517. 		"movq %%mm0, %%mm1				\n\t"

  518. 		"punpcklbw %%mm7, %%mm0				\n\t" // low part of line 0

  519. 		"punpckhbw %%mm7, %%mm1				\n\t" // high part of line 0

  520. 

  521. 		"movq (%%eax), %%mm2				\n\t"

  522. 		"movq %%mm2, %%mm3				\n\t"

  523. 		"punpcklbw %%mm7, %%mm2				\n\t" // low part of line 1

  524. 		"punpckhbw %%mm7, %%mm3				\n\t" // high part of line 1

  525. 

  526. 		"movq (%%eax, %1), %%mm4			\n\t"

  527. 		"movq %%mm4, %%mm5				\n\t"

  528. 		"punpcklbw %%mm7, %%mm4				\n\t" // low part of line 2

  529. 		"punpckhbw %%mm7, %%mm5				\n\t" // high part of line 2

  530. 

  531. 		"paddw %%mm0, %%mm0				\n\t" // 2L0

  532. 		"paddw %%mm1, %%mm1				\n\t" // 2H0

  533. 		"psubw %%mm4, %%mm2				\n\t" // L1 - L2

  534. 		"psubw %%mm5, %%mm3				\n\t" // H1 - H2

  535. 		"psubw %%mm2, %%mm0				\n\t" // 2L0 - L1 + L2

  536. 		"psubw %%mm3, %%mm1				\n\t" // 2H0 - H1 + H2

  537. 

  538. 		"psllw $2, %%mm2				\n\t" // 4L1 - 4L2

  539. 		"psllw $2, %%mm3				\n\t" // 4H1 - 4H2

  540. 		"psubw %%mm2, %%mm0				\n\t" // 2L0 - 5L1 + 5L2

  541. 		"psubw %%mm3, %%mm1				\n\t" // 2H0 - 5H1 + 5H2

  542. 

  543. 		"movq (%%eax, %1, 2), %%mm2			\n\t"

  544. 		"movq %%mm2, %%mm3				\n\t"

  545. 		"punpcklbw %%mm7, %%mm2				\n\t" // L3

  546. 		"punpckhbw %%mm7, %%mm3				\n\t" // H3

  547. 

  548. 		"psubw %%mm2, %%mm0				\n\t" // 2L0 - 5L1 + 5L2 - L3

  549. 		"psubw %%mm3, %%mm1				\n\t" // 2H0 - 5H1 + 5H2 - H3

  550. 		"psubw %%mm2, %%mm0				\n\t" // 2L0 - 5L1 + 5L2 - 2L3

  551. 		"psubw %%mm3, %%mm1				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

  552. 		"movq %%mm0, temp0				\n\t" // 2L0 - 5L1 + 5L2 - 2L3

  553. 		"movq %%mm1, temp1				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

  554. 

  555. 		"movq (%0, %1, 4), %%mm0			\n\t"

  556. 		"movq %%mm0, %%mm1				\n\t"

  557. 		"punpcklbw %%mm7, %%mm0				\n\t" // L4

  558. 		"punpckhbw %%mm7, %%mm1				\n\t" // H4

  559. 

  560. 		"psubw %%mm0, %%mm2				\n\t" // L3 - L4

  561. 		"psubw %%mm1, %%mm3				\n\t" // H3 - H4

  562. 		"movq %%mm2, temp2				\n\t" // L3 - L4

  563. 		"movq %%mm3, temp3				\n\t" // H3 - H4

  564. 		"paddw %%mm4, %%mm4				\n\t" // 2L2

  565. 		"paddw %%mm5, %%mm5				\n\t" // 2H2

  566. 		"psubw %%mm2, %%mm4				\n\t" // 2L2 - L3 + L4

  567. 		"psubw %%mm3, %%mm5				\n\t" // 2H2 - H3 + H4

  568. 

  569. 		"psllw $2, %%mm2				\n\t" // 4L3 - 4L4

  570. 		"psllw $2, %%mm3				\n\t" // 4H3 - 4H4

  571. 		"psubw %%mm2, %%mm4				\n\t" // 2L2 - 5L3 + 5L4

  572. 		"psubw %%mm3, %%mm5				\n\t" // 2H2 - 5H3 + 5H4

  573. //50 opcodes so far

  574. 		"movq (%%ebx), %%mm2				\n\t"

  575. 		"movq %%mm2, %%mm3				\n\t"

  576. 		"punpcklbw %%mm7, %%mm2				\n\t" // L5

  577. 		"punpckhbw %%mm7, %%mm3				\n\t" // H5

  578. 		"psubw %%mm2, %%mm4				\n\t" // 2L2 - 5L3 + 5L4 - L5

  579. 		"psubw %%mm3, %%mm5				\n\t" // 2H2 - 5H3 + 5H4 - H5

  580. 		"psubw %%mm2, %%mm4				\n\t" // 2L2 - 5L3 + 5L4 - 2L5

  581. 		"psubw %%mm3, %%mm5				\n\t" // 2H2 - 5H3 + 5H4 - 2H5

  582. 

  583. 		"movq (%%ebx, %1), %%mm6			\n\t"

  584. 		"punpcklbw %%mm7, %%mm6				\n\t" // L6

  585. 		"psubw %%mm6, %%mm2				\n\t" // L5 - L6

  586. 		"movq (%%ebx, %1), %%mm6			\n\t"

  587. 		"punpckhbw %%mm7, %%mm6				\n\t" // H6

  588. 		"psubw %%mm6, %%mm3				\n\t" // H5 - H6

  589. 

  590. 		"paddw %%mm0, %%mm0				\n\t" // 2L4

  591. 		"paddw %%mm1, %%mm1				\n\t" // 2H4

  592. 		"psubw %%mm2, %%mm0				\n\t" // 2L4 - L5 + L6

  593. 		"psubw %%mm3, %%mm1				\n\t" // 2H4 - H5 + H6

  594. 

  595. 		"psllw $2, %%mm2				\n\t" // 4L5 - 4L6

  596. 		"psllw $2, %%mm3				\n\t" // 4H5 - 4H6

  597. 		"psubw %%mm2, %%mm0				\n\t" // 2L4 - 5L5 + 5L6

  598. 		"psubw %%mm3, %%mm1				\n\t" // 2H4 - 5H5 + 5H6

  599. 

  600. 		"movq (%%ebx, %1, 2), %%mm2			\n\t"

  601. 		"movq %%mm2, %%mm3				\n\t"

  602. 		"punpcklbw %%mm7, %%mm2				\n\t" // L7

  603. 		"punpckhbw %%mm7, %%mm3				\n\t" // H7

  604. 

  605. 		"paddw %%mm2, %%mm2				\n\t" // 2L7

  606. 		"paddw %%mm3, %%mm3				\n\t" // 2H7

  607. 		"psubw %%mm2, %%mm0				\n\t" // 2L4 - 5L5 + 5L6 - 2L7

  608. 		"psubw %%mm3, %%mm1				\n\t" // 2H4 - 5H5 + 5H6 - 2H7

  609. 

  610. 		"movq temp0, %%mm2				\n\t" // 2L0 - 5L1 + 5L2 - 2L3

  611. 		"movq temp1, %%mm3				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

  612. //FIXME pxor, psubw, pmax for abs

  613. 		"movq %%mm7, %%mm6				\n\t" // 0

  614. 		"pcmpgtw %%mm0, %%mm6				\n\t"

  615. 		"pxor %%mm6, %%mm0				\n\t"

  616. 		"psubw %%mm6, %%mm0				\n\t" // |2L4 - 5L5 + 5L6 - 2L7|

  617. 		"movq %%mm7, %%mm6				\n\t" // 0

  618. 		"pcmpgtw %%mm1, %%mm6				\n\t"

  619. 		"pxor %%mm6, %%mm1				\n\t"

  620. 		"psubw %%mm6, %%mm1				\n\t" // |2H4 - 5H5 + 5H6 - 2H7|

  621. 

  622. 		"movq %%mm7, %%mm6				\n\t" // 0

  623. 		"pcmpgtw %%mm2, %%mm6				\n\t"

  624. 		"pxor %%mm6, %%mm2				\n\t"

  625. 		"psubw %%mm6, %%mm2				\n\t" // |2L0 - 5L1 + 5L2 - 2L3|

  626. 		"movq %%mm7, %%mm6				\n\t" // 0

  627. 		"pcmpgtw %%mm3, %%mm6				\n\t"

  628. 		"pxor %%mm6, %%mm3				\n\t"

  629. 		"psubw %%mm6, %%mm3				\n\t" // |2H0 - 5H1 + 5H2 - 2H3|

  630. 

  631. #ifdef HAVE_MMX2

  632. 		"pminsw %%mm2, %%mm0				\n\t"

  633. 		"pminsw %%mm3, %%mm1				\n\t"

  634. #else

  635. 		"movq %%mm0, %%mm6				\n\t"

  636. 		"psubusw %%mm2, %%mm6				\n\t"

  637. 		"psubw %%mm6, %%mm0				\n\t"

  638. 		"movq %%mm1, %%mm6				\n\t"

  639. 		"psubusw %%mm3, %%mm6				\n\t"

  640. 		"psubw %%mm6, %%mm1				\n\t"

  641. #endif

  642. 

  643. 		"movq %%mm7, %%mm6				\n\t" // 0

  644. 		"pcmpgtw %%mm4, %%mm6				\n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)

  645. 		"pxor %%mm6, %%mm4				\n\t"

  646. 		"psubw %%mm6, %%mm4				\n\t" // |2L2 - 5L3 + 5L4 - 2L5|

  647. 		"pcmpgtw %%mm5, %%mm7				\n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)

  648. 		"pxor %%mm7, %%mm5				\n\t"

  649. 		"psubw %%mm7, %%mm5				\n\t" // |2H2 - 5H3 + 5H4 - 2H5|

  650. // 100 opcodes

  651. 		"movd %2, %%mm2					\n\t" // QP

  652. //"pcmpeqb %%mm2, %%mm2\n\t"

  653. 		"punpcklwd %%mm2, %%mm2				\n\t"

  654. 		"punpcklwd %%mm2, %%mm2				\n\t"

  655. 		"psllw $3, %%mm2				\n\t" // 8QP

  656. 		"movq %%mm2, %%mm3				\n\t" // 8QP

  657. 		"pcmpgtw %%mm4, %%mm2				\n\t"

  658. 		"pcmpgtw %%mm5, %%mm3				\n\t"

  659. 		"pand %%mm2, %%mm4				\n\t"

  660. 		"pand %%mm3, %%mm5				\n\t"

  661. 

  662. 

  663. 		"psubusw %%mm0, %%mm4				\n\t" // hd

  664. 		"psubusw %%mm1, %%mm5				\n\t" // ld

  665. 

  666. 

  667. 		"movq w05, %%mm2				\n\t" // 5

  668. 		"pmullw %%mm2, %%mm4				\n\t"

  669. 		"pmullw %%mm2, %%mm5				\n\t"

  670. 		"movq w20, %%mm2				\n\t" // 32

  671. 		"paddw %%mm2, %%mm4				\n\t"

  672. 		"paddw %%mm2, %%mm5				\n\t"

  673. 		"psrlw $6, %%mm4				\n\t"

  674. 		"psrlw $6, %%mm5				\n\t"

  675. 

  676. /*

  677. 		"movq w06, %%mm2				\n\t" // 6

  678. 		"paddw %%mm2, %%mm4				\n\t"

  679. 		"paddw %%mm2, %%mm5				\n\t"

  680. 		"movq w1400, %%mm2				\n\t" // 1400h = 5120 = 5/64*2^16

  681. //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120

  682. 		"pmulhw %%mm2, %%mm4				\n\t" // hd/13

  683. 		"pmulhw %%mm2, %%mm5				\n\t" // ld/13

  684. */

  685. 

  686. 		"movq temp2, %%mm0				\n\t" // L3 - L4

  687. 		"movq temp3, %%mm1				\n\t" // H3 - H4

  688. 

  689. 		"pxor %%mm2, %%mm2				\n\t"

  690. 		"pxor %%mm3, %%mm3				\n\t"

  691. 

  692. 		// FIXME rounding error

  693. 		"psraw $1, %%mm0				\n\t" // (L3 - L4)/2

  694. 		"psraw $1, %%mm1				\n\t" // (H3 - H4)/2

  695. 		"pcmpgtw %%mm0, %%mm2				\n\t" // sign (L3-L4)

  696. 		"pcmpgtw %%mm1, %%mm3				\n\t" // sign (H3-H4)

  697. 		"pxor %%mm2, %%mm0				\n\t"

  698. 		"pxor %%mm3, %%mm1				\n\t"

  699. 		"psubw %%mm2, %%mm0				\n\t" // |L3-L4|

  700. 		"psubw %%mm3, %%mm1				\n\t" // |H3-H4|

  701. //		"psrlw $1, %%mm0				\n\t" // |L3 - L4|/2

  702. //		"psrlw $1, %%mm1				\n\t" // |H3 - H4|/2

  703. 

  704. 		"pxor %%mm6, %%mm2				\n\t"

  705. 		"pxor %%mm7, %%mm3				\n\t"

  706. 		"pand %%mm2, %%mm4				\n\t"

  707. 		"pand %%mm3, %%mm5				\n\t"

  708. 

  709. #ifdef HAVE_MMX2

  710. 		"pminsw %%mm0, %%mm4				\n\t"

  711. 		"pminsw %%mm1, %%mm5				\n\t"

  712. #else

  713. 		"movq %%mm4, %%mm2				\n\t"

  714. 		"psubusw %%mm0, %%mm2				\n\t"

  715. 		"psubw %%mm2, %%mm4				\n\t"

  716. 		"movq %%mm5, %%mm2				\n\t"

  717. 		"psubusw %%mm1, %%mm2				\n\t"

  718. 		"psubw %%mm2, %%mm5				\n\t"

  719. #endif

  720. 		"pxor %%mm6, %%mm4				\n\t"

  721. 		"pxor %%mm7, %%mm5				\n\t"

  722. 		"psubw %%mm6, %%mm4				\n\t"

  723. 		"psubw %%mm7, %%mm5				\n\t"

  724. 		"packsswb %%mm5, %%mm4				\n\t"

  725. 		"movq (%%eax, %1, 2), %%mm0			\n\t"

  726. 		"paddb   %%mm4, %%mm0				\n\t"

  727. 		"movq %%mm0, (%%eax, %1, 2) 			\n\t"

  728. 		"movq (%0, %1, 4), %%mm0			\n\t"

  729. 		"psubb %%mm4, %%mm0				\n\t"

  730. //		"pxor %%mm0, %%mm0 \n\t"

  731. 		"movq %%mm0, (%0, %1, 4) 			\n\t"

  732. 

  733. 		:

  734. 		: "r" (src), "r" (stride), "r" (QP)

  735. 		: "%eax", "%ebx"

  736. 	);

  737. #else

  738. 	const int l1= stride;

  739. 	const int l2= stride + l1;

  740. 	const int l3= stride + l2;

  741. 	const int l4= stride + l3;

  742. 	const int l5= stride + l4;

  743. 	const int l6= stride + l5;

  744. 	const int l7= stride + l6;

  745. 	const int l8= stride + l7;

  746. //	const int l9= stride + l8;

  747. 

  748. 	for(int x=0; x<BLOCK_SIZE; x++)

  749. 	{

  750. 		const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);

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

  752. 		{

  753. 			const int q=(src[l4] - src[l5])/2;

  754. 			const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);

  755. 			const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);

  756. 

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

  758. 			d= MAX(d, 0);

  759. 

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

  761. 			d*= SIGN(-middleEnergy);

  762. 

  763. 			if(q>0)

  764. 			{

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

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

  767. 			}

  768. 			else

  769. 			{

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

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

  772. 			}

  773. 

  774.         		src[l4]-= d;

  775. 	        	src[l5]+= d;

  776. 		}

  777. 		src++;

  778. 	}

  779. #endif

  780. }

  781. 

  782. //FIXME?  |255-0| = 1

  783. /**

  784.  * Check if the given 8x8 Block is mostly "flat" and copy the unaliged data into tempBlock.

  785.  */

  786. static inline bool isHorizDCAndCopy2Temp(uint8_t src[], int stride)

  787. {

  788. //	src++;

  789. 	int numEq= 0;

  790. #ifdef HAVE_MMX

  791. asm volatile (

  792. //		"int $3 \n\t"

  793. 		"pushl %1\n\t"

  794. 		"movq b7E, %%mm7				\n\t" // mm7 = 0x7F

  795. 		"movq b7C, %%mm6				\n\t" // mm6 = 0x7D

  796. 		"leal tempBlock, %%eax				\n\t"

  797. 		"pxor %%mm0, %%mm0				\n\t"

  798. 

  799. #define HDC_CHECK_AND_CPY(i) \

  800. 		"movq -4(%1), %%mm2				\n\t"\

  801. 		"psrlq $32, %%mm2				\n\t"\

  802. 		"punpckldq 4(%1), %%mm2				\n\t" /* (%1) */\

  803. 		"movq %%mm2, %%mm1				\n\t"\

  804. 		"psrlq $8, %%mm2				\n\t"\

  805. 		"psubb %%mm1, %%mm2				\n\t"\

  806. 		"paddb %%mm7, %%mm2				\n\t"\

  807. 		"pcmpgtb %%mm6, %%mm2				\n\t"\

  808. 		"paddb %%mm2, %%mm0				\n\t"\

  809. 		"movq %%mm1," #i "(%%eax)			\n\t"

  810. 

  811. 		HDC_CHECK_AND_CPY(0)

  812. 		"addl %2, %1					\n\t"

  813. 		HDC_CHECK_AND_CPY(8)

  814. 		"addl %2, %1					\n\t"

  815. 		HDC_CHECK_AND_CPY(16)

  816. 		"addl %2, %1					\n\t"

  817. 		HDC_CHECK_AND_CPY(24)

  818. 		"addl %2, %1					\n\t"

  819. 		HDC_CHECK_AND_CPY(32)

  820. 		"addl %2, %1					\n\t"

  821. 		HDC_CHECK_AND_CPY(40)

  822. 		"addl %2, %1					\n\t"

  823. 		HDC_CHECK_AND_CPY(48)

  824. 		"addl %2, %1					\n\t"

  825. 		HDC_CHECK_AND_CPY(56)

  826. 

  827. 		"psllq $8, %%mm0				\n\t" // remove dummy value

  828. 		"movq %%mm0, %%mm1				\n\t"

  829. 		"psrlw $8, %%mm0				\n\t"

  830. 		"paddb %%mm1, %%mm0				\n\t"

  831. 		"movq %%mm0, %%mm1				\n\t"

  832. 		"psrlq $16, %%mm0				\n\t"

  833. 		"paddb %%mm1, %%mm0				\n\t"

  834. 		"movq %%mm0, %%mm1				\n\t"

  835. 		"psrlq $32, %%mm0				\n\t"

  836. 		"paddb %%mm1, %%mm0				\n\t"

  837. 		"popl %1\n\t"

  838. 		"movd %%mm0, %0					\n\t"

  839. 		: "=r" (numEq)

  840. 		: "r" (src), "r" (stride)

  841. 		: "%eax"

  842. 		);

  843. //	printf("%d\n", numEq);

  844. 	numEq= (256 - (numEq & 0xFF)) &0xFF;

  845. #else

  846. 	for(int y=0; y<BLOCK_SIZE; y++)

  847. 	{

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

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

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

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

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

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

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

  855. 		tempBlock[0 + y*TEMP_STRIDE] = src[0];

  856. 		tempBlock[1 + y*TEMP_STRIDE] = src[1];

  857. 		tempBlock[2 + y*TEMP_STRIDE] = src[2];

  858. 		tempBlock[3 + y*TEMP_STRIDE] = src[3];

  859. 		tempBlock[4 + y*TEMP_STRIDE] = src[4];

  860. 		tempBlock[5 + y*TEMP_STRIDE] = src[5];

  861. 		tempBlock[6 + y*TEMP_STRIDE] = src[6];

  862. 		tempBlock[7 + y*TEMP_STRIDE] = src[7];

  863. 		src+= stride;

  864. 	}

  865. #endif

  866. /*	if(abs(numEq - asmEq) > 0)

  867. 	{

  868. //		printf("\nasm:%d  c:%d\n", asmEq, numEq);

  869. 		for(int y=0; y<8; y++)

  870. 		{

  871. 			for(int x=0; x<8; x++)

  872. 			{

  873. 				printf("%d ", src[x + y*stride]);

  874. 			}

  875. 			printf("\n");

  876. 		}

  877. 	}

  878. */

  879. //	printf("%d\n", numEq);

  880. 	return numEq > hFlatnessThreshold;

  881. }

  882. 

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

  884. {

  885. #ifdef MMX_FIXME

  886. FIXME

  887. 	int isOk;

  888. 	asm volatile(

  889. //		"int $3 \n\t"

  890. 		"movq (%1, %2), %%mm0				\n\t"

  891. 		"movq (%1, %2, 8), %%mm1			\n\t"

  892. 		"movq %%mm0, %%mm2				\n\t"

  893. 		"psubusb %%mm1, %%mm0				\n\t"

  894. 		"psubusb %%mm2, %%mm1				\n\t"

  895. 		"por %%mm1, %%mm0				\n\t" // ABS Diff

  896. 

  897. 		"movq pQPb, %%mm7				\n\t" // QP,..., QP

  898. 		"paddusb %%mm7, %%mm7				\n\t" // 2QP ... 2QP

  899. 		"psubusb %%mm7, %%mm0				\n\t" // Diff <= 2QP -> 0

  900. 		"pcmpeqd b00, %%mm0				\n\t"

  901. 		"psrlq $16, %%mm0				\n\t"

  902. 		"pcmpeqd bFF, %%mm0				\n\t"

  903. //		"movd %%mm0, (%1, %2, 4)\n\t"

  904. 		"movd %%mm0, %0					\n\t"

  905. 		: "=r" (isOk)

  906. 		: "r" (src), "r" (stride)

  907. 		);

  908. 	return isOk;

  909. #else

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

  911. 

  912. 	return true;

  913. #endif

  914. }

  915. 

  916. static inline void doHorizDefFilterAndCopyBack(uint8_t dst[], int stride, int QP)

  917. {

  918. #ifdef HAVE_MMX2

  919. 	asm volatile(

  920. 		"pushl %0					\n\t"

  921. 		"pxor %%mm7, %%mm7				\n\t"

  922. 		"movq bm00001000, %%mm6				\n\t"

  923. 		"movd %2, %%mm5					\n\t" // QP

  924. 		"movq %%mm5, %%mm4				\n\t"

  925. 		"paddusb %%mm5, %%mm5				\n\t" // 2QP

  926. 		"paddusb %%mm5, %%mm4				\n\t" // 3QP

  927. 		"psllq $24, %%mm4				\n\t"

  928. 		"pxor %%mm5, %%mm5				\n\t" // 0

  929. 		"psubb %%mm4, %%mm5				\n\t" // -QP

  930. 		"leal tempBlock, %%eax				\n\t"

  931. 

  932. //FIXME? "unroll by 2" and mix

  933. #define HDF(i)	"movq " #i "(%%eax), %%mm0			\n\t"\

  934. 		"movq %%mm0, %%mm1				\n\t"\

  935. 		"movq %%mm0, %%mm2				\n\t"\

  936. 		"psrlq $8, %%mm1				\n\t"\

  937. 		"psubusb %%mm1, %%mm2				\n\t"\

  938. 		"psubusb %%mm0, %%mm1				\n\t"\

  939. 		"por %%mm2, %%mm1				\n\t" /* |px - p(x+1)| */\

  940. 		"pcmpeqb %%mm7, %%mm2				\n\t" /* sgn[px - p(x+1)] */\

  941. 		"pshufw $0xAA, %%mm1, %%mm3			\n\t"\

  942. 		"pminub %%mm1, %%mm3				\n\t"\

  943. 		"psrlq $16, %%mm3				\n\t"\

  944. 		"psubusb %%mm3, %%mm1			\n\t" /* |p3-p4|-min(|p1-p2|,|p5,ü6|) */\

  945. 		"paddb %%mm5, %%mm1				\n\t"\

  946. 		"psubusb %%mm5, %%mm1				\n\t"\

  947. 		"psrlw $2, %%mm1				\n\t"\

  948. 		"pxor %%mm2, %%mm1				\n\t"\

  949. 		"psubb %%mm2, %%mm1				\n\t"\

  950. 		"pand %%mm6, %%mm1				\n\t"\

  951. 		"psubb %%mm1, %%mm0				\n\t"\

  952. 		"psllq $8, %%mm1				\n\t"\

  953. 		"paddb %%mm1, %%mm0				\n\t"\

  954. 		"movd %%mm0, (%0)				\n\t"\

  955. 		"psrlq $32, %%mm0				\n\t"\

  956. 		"movd %%mm0, 4(%0)				\n\t"

  957. 

  958. 		HDF(0)

  959. 		"addl %1, %0					\n\t"

  960. 		HDF(8)

  961. 		"addl %1, %0					\n\t"

  962. 		HDF(16)

  963. 		"addl %1, %0					\n\t"

  964. 		HDF(24)

  965. 		"addl %1, %0					\n\t"

  966. 		HDF(32)

  967. 		"addl %1, %0					\n\t"

  968. 		HDF(40)

  969. 		"addl %1, %0					\n\t"

  970. 		HDF(48)

  971. 		"addl %1, %0					\n\t"

  972. 		HDF(56)

  973. 		"popl %0					\n\t"

  974. 		:

  975. 		: "r" (dst), "r" (stride), "r" (QP)

  976. 		: "%eax"

  977. 	);

  978. #else

  979. 	uint8_t *src= tempBlock;

  980. 

  981. 	for(int y=0; y<BLOCK_SIZE; y++)

  982. 	{

  983. 		dst[0] = src[0];

  984. 		dst[1] = src[1];

  985. 		dst[2] = src[2];

  986. 		dst[3] = src[3];

  987. 		dst[4] = src[4];

  988. 		dst[5] = src[5];

  989. 		dst[6] = src[6];

  990. 		dst[7] = src[7];

  991. 

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

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

  994. 		{

  995. 			const int q=(src[3] - src[4])/2;

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

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

  998. 

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

  1000. 			d= MAX(d, 0);

  1001. 

  1002. 			d= (5*d + 32) >> 6;

  1003. 			d*= SIGN(-middleEnergy);

  1004. 

  1005. 			if(q>0)

  1006. 			{

  1007. 				d= d<0 ? 0 : d;

  1008. 				d= d>q ? q : d;

  1009. 			}

  1010. 			else

  1011. 			{

  1012. 				d= d>0 ? 0 : d;

  1013. 				d= d<q ? q : d;

  1014. 			}

  1015. 

  1016.         		dst[3]-= d;

  1017. 	        	dst[4]+= d;

  1018. 		}

  1019. 		dst+= stride;

  1020. 		src+= TEMP_STRIDE;

  1021. 	}

  1022. #endif

  1023. }

  1024. 

  1025. /**

  1026.  * Do a horizontal low pass filter on the 8x8 block

  1027.  * useing the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16 (C version)

  1028.  * useing approximately the 7-Tap Filter (1,2,3,4,3,2,1)/16 (MMX2 version)

  1029.  */

  1030. static inline void doHorizLowPassAndCopyBack(uint8_t dst[], int stride, int QP)

  1031. {

  1032. //return;

  1033. #ifdef HAVE_MMX2

  1034. 	asm volatile(	//"movv %0 %1 %2\n\t"

  1035. 		"pushl %0\n\t"

  1036. 		"pxor %%mm7, %%mm7					\n\t"

  1037. 		"leal tempBlock, %%eax					\n\t"

  1038. 

  1039. #define HLP1	"movq (%0), %%mm0					\n\t"\

  1040. 		"movq %%mm0, %%mm1					\n\t"\

  1041. 		"psllq $8, %%mm0					\n\t"\

  1042. 		"pavgb %%mm1, %%mm0					\n\t"\

  1043. 		"psrlw $8, %%mm0					\n\t"\

  1044. 		"pxor %%mm1, %%mm1					\n\t"\

  1045. 		"packuswb %%mm1, %%mm0					\n\t"\

  1046. 		"movq %%mm0, %%mm1					\n\t"\

  1047. 		"movq %%mm0, %%mm2					\n\t"\

  1048. 		"psllq $32, %%mm0					\n\t"\

  1049. 		"paddb %%mm0, %%mm1					\n\t"\

  1050. 		"psllq $16, %%mm2					\n\t"\

  1051. 		"pavgb %%mm2, %%mm0					\n\t"\

  1052. 		"movq %%mm0, %%mm3					\n\t"\

  1053. 		"pand bm11001100, %%mm0					\n\t"\

  1054. 		"paddusb %%mm0, %%mm3					\n\t"\

  1055. 		"psrlq $8, %%mm3					\n\t"\

  1056. 		"pavgb %%mm1, %%mm4					\n\t"\

  1057. 		"pavgb %%mm3, %%mm2					\n\t"\

  1058. 		"psrlq $16, %%mm2					\n\t"\

  1059. 		"punpcklbw %%mm2, %%mm2					\n\t"\

  1060. 		"movq %%mm2, (%0)					\n\t"\

  1061. 

  1062. #define HLP2	"movq (%0), %%mm0					\n\t"\

  1063. 		"movq %%mm0, %%mm1					\n\t"\

  1064. 		"psllq $8, %%mm0					\n\t"\

  1065. 		"pavgb %%mm1, %%mm0					\n\t"\

  1066. 		"psrlw $8, %%mm0					\n\t"\

  1067. 		"pxor %%mm1, %%mm1					\n\t"\

  1068. 		"packuswb %%mm1, %%mm0					\n\t"\

  1069. 		"movq %%mm0, %%mm2					\n\t"\

  1070. 		"psllq $32, %%mm0					\n\t"\

  1071. 		"psllq $16, %%mm2					\n\t"\

  1072. 		"pavgb %%mm2, %%mm0					\n\t"\

  1073. 		"movq %%mm0, %%mm3					\n\t"\

  1074. 		"pand bm11001100, %%mm0					\n\t"\

  1075. 		"paddusb %%mm0, %%mm3					\n\t"\

  1076. 		"psrlq $8, %%mm3					\n\t"\

  1077. 		"pavgb %%mm3, %%mm2					\n\t"\

  1078. 		"psrlq $16, %%mm2					\n\t"\

  1079. 		"punpcklbw %%mm2, %%mm2					\n\t"\

  1080. 		"movq %%mm2, (%0)					\n\t"\

  1081. 

  1082. // approximately a 7-Tap Filter with Vector (1,2,3,4,3,2,1)/16

  1083. /*

  1084.  31

  1085.  121

  1086.   121

  1087.    121

  1088.     121

  1089.      121

  1090.       121

  1091.        13

  1092. Implemented	Exact 7-Tap

  1093.  9421		A321

  1094.  36421		64321

  1095.  334321		=

  1096.  1234321	=

  1097.   1234321	=

  1098.    123433	=

  1099.     12463	  12346

  1100.      1249	   123A

  1101. 

  1102. */

  1103. #define HLP3(i)	"movq " #i "(%%eax), %%mm0				\n\t"\

  1104. 		"movq %%mm0, %%mm1					\n\t"\

  1105. 		"movq %%mm0, %%mm2					\n\t"\

  1106. 		"movq %%mm0, %%mm3					\n\t"\

  1107. 		"movq %%mm0, %%mm4					\n\t"\

  1108. 		"psllq $8, %%mm1					\n\t"\

  1109. 		"psrlq $8, %%mm2					\n\t"\

  1110. 		"pand bm00000001, %%mm3					\n\t"\

  1111. 		"pand bm10000000, %%mm4					\n\t"\

  1112. 		"por %%mm3, %%mm1					\n\t"\

  1113. 		"por %%mm4, %%mm2					\n\t"\

  1114. 		"pavgb %%mm2, %%mm1					\n\t"\

  1115. 		"pavgb %%mm1, %%mm0					\n\t"\

  1116. \

  1117. 		"pshufw $0xF9, %%mm0, %%mm3				\n\t"\

  1118. 		"pshufw $0x90, %%mm0, %%mm4				\n\t"\

  1119. 		"pavgb %%mm3, %%mm4					\n\t"\

  1120. 		"pavgb %%mm4, %%mm0					\n\t"\

  1121. 		"movd %%mm0, (%0)					\n\t"\

  1122. 		"psrlq $32, %%mm0					\n\t"\

  1123. 		"movd %%mm0, 4(%0)					\n\t"\

  1124. 

  1125. #define HLP(i) HLP3(i)

  1126. 

  1127. 		HLP(0)

  1128. 		"addl %1, %0						\n\t"

  1129. 		HLP(8)

  1130. 		"addl %1, %0						\n\t"

  1131. 		HLP(16)

  1132. 		"addl %1, %0						\n\t"

  1133. 		HLP(24)

  1134. 		"addl %1, %0						\n\t"

  1135. 		HLP(32)

  1136. 		"addl %1, %0						\n\t"

  1137. 		HLP(40)

  1138. 		"addl %1, %0						\n\t"

  1139. 		HLP(48)

  1140. 		"addl %1, %0						\n\t"

  1141. 		HLP(56)

  1142. 

  1143. 		"popl %0\n\t"

  1144. 		:

  1145. 		: "r" (dst), "r" (stride)

  1146. 		: "%eax", "%ebx"

  1147. 	);

  1148. 

  1149. #else

  1150. 	uint8_t *temp= tempBlock;

  1151. 	for(int y=0; y<BLOCK_SIZE; y++)

  1152. 	{

  1153. 		const int first= ABS(dst[-1] - dst[0]) < QP ? dst[-1] : dst[0];

  1154. 		const int last= ABS(dst[8] - dst[7]) < QP ? dst[8] : dst[7];

  1155. 

  1156. 		int sums[9];

  1157. 		sums[0] = first + temp[0];

  1158. 		sums[1] = temp[0] + temp[1];

  1159. 		sums[2] = temp[1] + temp[2];

  1160. 		sums[3] = temp[2] + temp[3];

  1161. 		sums[4] = temp[3] + temp[4];

  1162. 		sums[5] = temp[4] + temp[5];

  1163. 		sums[6] = temp[5] + temp[6];

  1164. 		sums[7] = temp[6] + temp[7];

  1165. 		sums[8] = temp[7] + last;

  1166. 

  1167. 		dst[0]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;

  1168. 		dst[1]= ((dst[1]<<2) + (first + sums[0] + sums[3]<<1) + sums[5] + 8)>>4;

  1169. 		dst[2]= ((dst[2]<<2) + (first + sums[1] + sums[4]<<1) + sums[6] + 8)>>4;

  1170. 		dst[3]= ((dst[3]<<2) + (sums[2] + sums[5]<<1) + sums[0] + sums[7] + 8)>>4;

  1171. 		dst[4]= ((dst[4]<<2) + (sums[3] + sums[6]<<1) + sums[1] + sums[8] + 8)>>4;

  1172. 		dst[5]= ((dst[5]<<2) + (last + sums[7] + sums[4]<<1) + sums[2] + 8)>>4;

  1173. 		dst[6]= ((last + dst[6]<<2) + (dst[7] + sums[5]<<1) + sums[3] + 8)>>4;

  1174. 		dst[7]= ((sums[8]<<2) + (last + sums[6]<<1) + sums[4] + 8)>>4;

  1175. 

  1176. 		dst+= stride;

  1177. 		temp+= TEMP_STRIDE;

  1178. 	}

  1179. #endif

  1180. }

  1181. 

  1182. 

  1183. static inline void dering(uint8_t src[], int stride, int QP)

  1184. {

  1185. //FIXME

  1186. 

  1187. #ifdef HAVE_MMX2X

  1188. 	asm volatile(

  1189. 		"leal (%0, %1), %%eax				\n\t"

  1190. 		"leal (%%eax, %1, 4), %%ebx			\n\t"

  1191. //	0	1	2	3	4	5	6	7	8	9

  1192. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ebx	ebx+%1	ebx+2%1	%0+8%1	ebx+4%1

  1193. 

  1194. 		"pcmpeq %%mm6, %%mm6				\n\t"

  1195. 		"pxor %%mm7, %%mm7				\n\t"

  1196. 

  1197. #define FIND_MIN_MAX(addr)\

  1198. 		"movq (" #addr "), %%mm0,			\n\t"\

  1199. 		"pminub %%mm0, %%mm6				\n\t"\

  1200. 		"pmaxub %%mm0, %%mm7				\n\t"

  1201. 

  1202. FIND_MIN_MAX(%0)

  1203. FIND_MIN_MAX(%%eax)

  1204. FIND_MIN_MAX(%%eax, %1)

  1205. FIND_MIN_MAX(%%eax, %1, 2)

  1206. FIND_MIN_MAX(%0, %1, 4)

  1207. FIND_MIN_MAX(%%ebx)

  1208. FIND_MIN_MAX(%%ebx, %1)

  1209. FIND_MIN_MAX(%%ebx, %1, 2)

  1210. FIND_MIN_MAX(%0, %1, 8)

  1211. FIND_MIN_MAX(%%ebx, %1, 2)

  1212. 

  1213. 		"movq %%mm6, %%mm4				\n\t"

  1214. 		"psrlq $32, %%mm6				\n\t"

  1215. 		"pminub %%mm4, %%mm6				\n\t"

  1216. 		"movq %%mm6, %%mm4				\n\t"

  1217. 		"psrlq $16, %%mm6				\n\t"

  1218. 		"pminub %%mm4, %%mm6				\n\t"

  1219. 		"movq %%mm6, %%mm4				\n\t"

  1220. 		"psrlq $8, %%mm6				\n\t"

  1221. 		"pminub %%mm4, %%mm6				\n\t" // min of pixels

  1222. 

  1223. 		"movq %%mm7, %%mm4				\n\t"

  1224. 		"psrlq $32, %%mm7				\n\t"

  1225. 		"pmaxub %%mm4, %%mm7				\n\t"

  1226. 		"movq %%mm7, %%mm4				\n\t"

  1227. 		"psrlq $16, %%mm7				\n\t"

  1228. 		"pmaxub %%mm4, %%mm7				\n\t"

  1229. 		"movq %%mm7, %%mm4				\n\t"

  1230. 		"psrlq $8, %%mm7				\n\t"

  1231. 		"pmaxub %%mm4, %%mm7				\n\t" // max of pixels

  1232. 		"pavgb %%mm6, %%mm7				\n\t" // (max + min)/2

  1233. 

  1234. 

  1235. 		: : "r" (src), "r" (stride), "r" (QP)

  1236. 		: "%eax", "%ebx"

  1237. 	);

  1238. #else

  1239. 

  1240. //FIXME

  1241. #endif

  1242. }

  1243. 

  1244. /**

  1245.  * ...

  1246.  */

  1247. extern "C"{

  1248. void  postprocess(unsigned char * src[], int src_stride,

  1249.                  unsigned char * dst[], int dst_stride,

  1250.                  int horizontal_size,   int vertical_size,

  1251.                  QP_STORE_T *QP_store,  int QP_stride,

  1252. 					  int mode)

  1253. {

  1254. /*

  1255. 	long long T= rdtsc();

  1256. 	for(int y=vertical_size-1; y>=0 ; y--)

  1257. 		memcpy(dst[0] + y*src_stride, src[0] + y*src_stride,src_stride);

  1258. //	memcpy(dst[0], src[0],src_stride*vertical_size);

  1259. 	printf("%4dk\r", (rdtsc()-T)/1000);

  1260. 

  1261. 	return;

  1262. */

  1263. /*

  1264. 	long long T= rdtsc();

  1265. 	while( (rdtsc() - T)/1000 < 4000);

  1266. 

  1267. 	return;

  1268. */

  1269. 	postProcess(src[0], src_stride,

  1270. 		dst[0], dst_stride, horizontal_size, vertical_size, QP_store, QP_stride, false);

  1271. 

  1272. 	horizontal_size >>= 1;

  1273. 	vertical_size   >>= 1;

  1274. 	src_stride      >>= 1;

  1275. 	dst_stride      >>= 1;

  1276. 

  1277. 	if(1)

  1278. 	{

  1279. 		postProcess(src[1], src_stride,

  1280. 			dst[1], dst_stride, horizontal_size, vertical_size, QP_store, QP_stride, true);

  1281. 		postProcess(src[2], src_stride,

  1282. 			dst[2], dst_stride, horizontal_size, vertical_size, QP_store, QP_stride, true);

  1283. 	}

  1284. 	else

  1285. 	{

  1286. 		memcpy(dst[1], src[1], src_stride*horizontal_size);

  1287. 		memcpy(dst[2], src[2], src_stride*horizontal_size);

  1288. 	}

  1289. }

  1290. }

  1291. 

  1292. /**

  1293.  * Copies a block from src to dst and fixes the blacklevel

  1294.  */

  1295. static inline void blockCopy(uint8_t dst[], int dstStride, uint8_t src[], int srcStride)

  1296. {

  1297. #ifdef HAVE_MMX

  1298. 					asm volatile(

  1299. 						"pushl %0 \n\t"

  1300. 						"pushl %1 \n\t"

  1301. 						"leal (%2,%2), %%eax	\n\t"

  1302. 						"leal (%3,%3), %%ebx	\n\t"

  1303. 						"movq packedYOffset, %%mm2	\n\t"

  1304. 						"movq packedYScale, %%mm3	\n\t"

  1305. 

  1306. #define SIMPLE_CPY					\

  1307. 						"movq (%0), %%mm0	\n\t"\

  1308. 						"movq (%0,%2), %%mm1	\n\t"\

  1309. 						"psubusb %%mm2, %%mm0	\n\t"\

  1310. 						"psubusb %%mm2, %%mm1	\n\t"\

  1311. 						"movq %%mm0, (%1)	\n\t"\

  1312. 						"movq %%mm1, (%1, %3)	\n\t"\

  1313. 

  1314. #define SCALED_CPY					\

  1315. 						"movq (%0), %%mm0	\n\t"\

  1316. 						"movq (%0,%2), %%mm1	\n\t"\

  1317. 						"psubusb %%mm2, %%mm0	\n\t"\

  1318. 						"psubusb %%mm2, %%mm1	\n\t"\

  1319. 						"pxor %%mm4, %%mm4	\n\t"\

  1320. 						"pxor %%mm5, %%mm5	\n\t"\

  1321. 						"punpcklbw %%mm0, %%mm4 \n\t"\

  1322. 						"punpckhbw %%mm0, %%mm5 \n\t"\

  1323. 						"pmulhuw %%mm3, %%mm4	\n\t"\

  1324. 						"pmulhuw %%mm3, %%mm5	\n\t"\

  1325. 						"packuswb %%mm5, %%mm4	\n\t"\

  1326. 						"movq %%mm4, (%1)	\n\t"\

  1327. 						"pxor %%mm4, %%mm4	\n\t"\

  1328. 						"pxor %%mm5, %%mm5	\n\t"\

  1329. 						"punpcklbw %%mm1, %%mm4 \n\t"\

  1330. 						"punpckhbw %%mm1, %%mm5 \n\t"\

  1331. 						"pmulhuw %%mm3, %%mm4	\n\t"\

  1332. 						"pmulhuw %%mm3, %%mm5	\n\t"\

  1333. 						"packuswb %%mm5, %%mm4	\n\t"\

  1334. 						"movq %%mm4, (%1, %3)	\n\t"\

  1335. 

  1336. 

  1337. #define CPY SCALED_CPY

  1338. //#define CPY SIMPLE_CPY

  1339. //				"prefetchnta 8(%0)\n\t"

  1340. CPY

  1341. 						"addl %%eax, %0		\n\t"

  1342. 						"addl %%ebx, %1		\n\t"

  1343. CPY

  1344. 						"addl %%eax, %0		\n\t"

  1345. 						"addl %%ebx, %1		\n\t"

  1346. CPY

  1347. 						"addl %%eax, %0		\n\t"

  1348. 						"addl %%ebx, %1		\n\t"

  1349. CPY

  1350. 						"popl %1 \n\t"

  1351. 						"popl %0 \n\t"

  1352. 						: : "r" (src),

  1353. 						"r" (dst),

  1354. 						"r" (srcStride),

  1355. 						"r" (dstStride)

  1356. 						: "%eax", "%ebx"

  1357. 					);

  1358. #else

  1359. 				for(int i=0; i<BLOCK_SIZE; i++) // last 10x8 Block is copied allready so +2

  1360. 					memcpy(	&(dst[dstStride*i]),

  1361. 						&(src[srcStride*i]), BLOCK_SIZE);

  1362. #endif

  1363. }

  1364. 

  1365. 

  1366. /**

  1367.  * Filters array of bytes (Y or U or V values)

  1368.  */

  1369. void postProcess(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,

  1370. 	QP_STORE_T QPs[], int QPStride, bool isColor)

  1371. {

  1372. 

  1373. #ifdef TIMEING

  1374. 	long long T0, T1, memcpyTime=0, vertTime=0, horizTime=0, sumTime, diffTime=0;

  1375. 	sumTime= rdtsc();

  1376. #endif

  1377. 

  1378. 	/* we need 64bit here otherwise we´ll going to have a problem

  1379. 	   after watching a black picture for 5 hours*/

  1380. 	static uint64_t *yHistogram= NULL;

  1381. 	if(!yHistogram)

  1382. 	{

  1383. 		yHistogram= new uint64_t[256];

  1384. 		for(int i=0; i<256; i++) yHistogram[i]= width*height/64/256;

  1385. 	}

  1386. 

  1387. 	int black=0, white=255; // blackest black and whitest white in the picture

  1388. 	if(!isColor)

  1389. 	{

  1390. 		uint64_t sum= 0;

  1391. 		for(int i=0; i<256; i++)

  1392. 			sum+= yHistogram[i];

  1393. 

  1394. 		uint64_t maxClipped= (uint64_t)(sum * maxClippedThreshold);

  1395. 

  1396. 		uint64_t clipped= sum;

  1397. 		for(black=255; black>0; black--)

  1398. 		{

  1399. 			if(clipped < maxClipped) break;

  1400. 			clipped-= yHistogram[black];

  1401. 		}

  1402. 

  1403. 		clipped= sum;

  1404. 		for(white=0; white<256; white++)

  1405. 		{

  1406. 			if(clipped < maxClipped) break;

  1407. 			clipped-= yHistogram[white];

  1408. 		}

  1409. 

  1410. 		// we cant handle negative correctures

  1411. 		packedYOffset= MAX(black - minAllowedY, 0);

  1412. 		packedYOffset|= packedYOffset<<32;

  1413. 		packedYOffset|= packedYOffset<<16;

  1414. 		packedYOffset|= packedYOffset<<8;

  1415. 

  1416. //		uint64_t scale= (int)(256.0*256.0/(white-black) + 0.5);

  1417. 		double scale= (double)(maxAllowedY - minAllowedY) / (double)(white-black);

  1418. 

  1419. 		packedYScale= uint16_t(scale*256.0 + 0.5);

  1420. 		packedYScale|= packedYScale<<32;

  1421. 		packedYScale|= packedYScale<<16;

  1422. 	}

  1423. 	else

  1424. 	{

  1425. 		packedYScale= 0x0100010001000100LL;

  1426. 		packedYOffset= 0;

  1427. 	}

  1428. 

  1429. 	for(int x=0; x<width; x+=BLOCK_SIZE)

  1430. 		blockCopy(dst + x, dstStride, src + x, srcStride);

  1431. 

  1432. 	for(int y=0; y<height; y+=BLOCK_SIZE)

  1433. 	{

  1434. 		//1% speedup if these are here instead of the inner loop

  1435. 		uint8_t *srcBlock= &(src[y*srcStride]);

  1436. 		uint8_t *dstBlock= &(dst[y*dstStride]);

  1437. 		uint8_t *vertSrcBlock= &(srcBlock[srcStride*3]); // Blocks are 10x8 -> *3 to start

  1438. 		uint8_t *vertBlock= &(dstBlock[dstStride*3]);

  1439. 

  1440. 		// finish 1 block before the next otherwise we´ll might have a problem

  1441. 		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing

  1442. 		for(int x=0; x<width; x+=BLOCK_SIZE)

  1443. 		{

  1444. 			int QP= isColor ?

  1445. 				QPs[(y>>3)*QPStride + (x>>3)]:

  1446. 				(QPs[(y>>4)*QPStride + (x>>4)] * (packedYScale &0xFFFF))>>8;

  1447. #ifdef HAVE_MMX

  1448. 		asm volatile(

  1449. 			"movd %0, %%mm7					\n\t"

  1450. 			"packuswb %%mm7, %%mm7				\n\t" // 0, 0, 0, QP, 0, 0, 0, QP

  1451. 			"packuswb %%mm7, %%mm7				\n\t" // 0,QP, 0, QP, 0,QP, 0, QP

  1452. 			"packuswb %%mm7, %%mm7				\n\t" // QP,..., QP

  1453. 			"movq %%mm7, pQPb				\n\t"

  1454. 			: : "r" (QP)

  1455. 		);

  1456. #endif

  1457. 

  1458. 

  1459. 			const int stride= dstStride;

  1460. 			if(y + 12 < height)

  1461. 			{

  1462. #ifdef MORE_TIMEING

  1463. 				T0= rdtsc();

  1464. #endif

  1465. #ifdef HAVE_MMX2

  1466. 

  1467. 				prefetchnta(vertSrcBlock + (((x>>3)&3) + 2)*srcStride + 32);

  1468. 				prefetchnta(vertSrcBlock + (((x>>3)&3) + 6)*srcStride + 32);

  1469. 				prefetcht0(vertBlock + (((x>>3)&3) + 2)*dstStride + 32);

  1470. 				prefetcht0(vertBlock + (((x>>3)&3) + 6)*dstStride + 32);

  1471. #endif

  1472. 				if(!isColor) yHistogram[ srcBlock[0] ]++;

  1473. 

  1474. 				blockCopy(vertBlock + dstStride*2, dstStride,

  1475. 					vertSrcBlock + srcStride*2, srcStride);

  1476. 

  1477. 

  1478. #ifdef MORE_TIMEING

  1479. 				T1= rdtsc();

  1480. 				memcpyTime+= T1-T0;

  1481. 				T0=T1;

  1482. #endif

  1483. 

  1484. 				if( isVertDC(vertBlock, stride))

  1485. 				{

  1486. 					if(isVertMinMaxOk(vertBlock, stride, QP))

  1487. 						doVertLowPass(vertBlock, stride, QP);

  1488. 				}

  1489. 				else if(x<width)

  1490. 					doVertDefFilter(vertBlock, stride, QP);

  1491. 

  1492. #ifdef MORE_TIMEING

  1493. 				T1= rdtsc();

  1494. 				vertTime+= T1-T0;

  1495. 				T0=T1;

  1496. #endif

  1497. 			}

  1498. 			else

  1499. 			{

  1500. 				for(int i=2; i<BLOCK_SIZE/2+1; i++) // last 10x8 Block is copied allready so +2

  1501. 					memcpy(	&(vertBlock[dstStride*i]),

  1502. 						&(vertSrcBlock[srcStride*i]), BLOCK_SIZE);

  1503. 

  1504. 			}

  1505. 

  1506. 			if(x - 8 >= 0 && x<width)

  1507. 			{

  1508. #ifdef MORE_TIMEING

  1509. 				T0= rdtsc();

  1510. #endif

  1511. 

  1512. 				if( isHorizDCAndCopy2Temp(dstBlock-4, stride))

  1513. 				{

  1514. 					if(isHorizMinMaxOk(tempBlock, TEMP_STRIDE, QP))

  1515. 						doHorizLowPassAndCopyBack(dstBlock-4, stride, QP);

  1516. 				}

  1517. 				else

  1518. 					doHorizDefFilterAndCopyBack(dstBlock-4, stride, QP);

  1519. 

  1520. #ifdef MORE_TIMEING

  1521. 				T1= rdtsc();

  1522. 				horizTime+= T1-T0;

  1523. 				T0=T1;

  1524. #endif

  1525. 				dering(dstBlock - 9 - stride, stride, QP);

  1526. 			}

  1527. 			else if(y!=0)

  1528. 				dering(dstBlock - stride*9 + width-9, stride, QP);

  1529. 			//FIXME dering filter will not be applied to last block (bottom right)

  1530. 

  1531. 

  1532. 			dstBlock+=8;

  1533. 			srcBlock+=8;

  1534. 			vertBlock+=8;

  1535. 			vertSrcBlock+=8;

  1536. 		}

  1537. 	}

  1538. #ifdef HAVE_MMX

  1539. 	asm volatile("emms");

  1540. #endif

  1541. 

  1542. #ifdef TIMEING

  1543. 	// FIXME diff is mostly the time spent for rdtsc (should subtract that but ...)

  1544. 	sumTime= rdtsc() - sumTime;

  1545. 	if(!isColor)

  1546. 		printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d    \r",

  1547. 			int(memcpyTime/1000), int(vertTime/1000), int(horizTime/1000),

  1548. 			int(sumTime/1000), int((sumTime-memcpyTime-vertTime-horizTime)/1000)

  1549. 			, black, white);

  1550. #endif

  1551. }