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.
2225 Commits
32 Branches
2.4GB
Tree: 89ebf4e85e
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '89ebf4e85e'
${ noResults }
FFmpeg/libavcodec/libpostproc/postprocess_template.c

3241 lines
96KB
Raw Blame History 

  1. /*

  2.     Copyright (C) 2001-2002 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.  * @file postprocess_template.c

  21.  * mmx/mmx2/3dnow postprocess code.

  22.  */

  23. 

  24. 

  25. #undef PAVGB

  26. #undef PMINUB

  27. #undef PMAXUB

  28. 

  29. #ifdef HAVE_MMX2

  30. #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"

  31. #elif defined (HAVE_3DNOW)

  32. #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"

  33. #endif

  34. 

  35. #ifdef HAVE_MMX2

  36. #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"

  37. #elif defined (HAVE_MMX)

  38. #define PMINUB(b,a,t) \

  39. 	"movq " #a ", " #t " \n\t"\

  40. 	"psubusb " #b ", " #t " \n\t"\

  41. 	"psubb " #t ", " #a " \n\t"

  42. #endif

  43. 

  44. #ifdef HAVE_MMX2

  45. #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"

  46. #elif defined (HAVE_MMX)

  47. #define PMAXUB(a,b) \

  48. 	"psubusb " #a ", " #b " \n\t"\

  49. 	"paddb " #a ", " #b " \n\t"

  50. #endif

  51. 

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

  53. #ifdef HAVE_MMX

  54. /**

  55.  * Check if the middle 8x8 Block in the given 8x16 block is flat

  56.  */

  57. static inline int RENAME(vertClassify)(uint8_t src[], int stride, PPContext *c){

  58. 	int numEq= 0, dcOk;

  59. 	src+= stride*4; // src points to begin of the 8x8 Block

  60. asm volatile(

  61. 		"leal (%2, %3), %%eax				\n\t"

  62. //	0	1	2	3	4	5	6	7	8	9

  63. //	%1	eax	eax+%2	eax+2%2	%1+4%2	ecx	ecx+%2	ecx+2%2	%1+8%2	ecx+4%2

  64. 		"movq %4, %%mm7					\n\t" 

  65. 		"movq %5, %%mm6					\n\t" 

  66. 

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

  68. 		"movq (%%eax), %%mm1				\n\t"

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

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

  71.                 PMAXUB(%%mm1, %%mm4)

  72.                 PMINUB(%%mm1, %%mm3, %%mm5)

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

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

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

  76. 

  77. 		"movq (%%eax,%3), %%mm2				\n\t"

  78.                 PMAXUB(%%mm2, %%mm4)

  79.                 PMINUB(%%mm2, %%mm3, %%mm5)

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

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

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

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

  84. 

  85. 		"movq (%%eax, %3, 2), %%mm1			\n\t"

  86.                 PMAXUB(%%mm1, %%mm4)

  87.                 PMINUB(%%mm1, %%mm3, %%mm5)

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

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

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

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

  92. 		

  93. 		"leal (%%eax, %3, 4), %%eax			\n\t"

  94. 

  95. 		"movq (%2, %3, 4), %%mm2			\n\t"

  96.                 PMAXUB(%%mm2, %%mm4)

  97.                 PMINUB(%%mm2, %%mm3, %%mm5)

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

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

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

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

  102. 

  103. 		"movq (%%eax), %%mm1				\n\t"

  104.                 PMAXUB(%%mm1, %%mm4)

  105.                 PMINUB(%%mm1, %%mm3, %%mm5)

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

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

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

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

  110. 

  111. 		"movq (%%eax, %3), %%mm2			\n\t"

  112.                 PMAXUB(%%mm2, %%mm4)

  113.                 PMINUB(%%mm2, %%mm3, %%mm5)

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

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

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

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

  118. 

  119. 		"movq (%%eax, %3, 2), %%mm1			\n\t"

  120.                 PMAXUB(%%mm1, %%mm4)

  121.                 PMINUB(%%mm1, %%mm3, %%mm5)

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

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

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

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

  126. 		"psubusb %%mm3, %%mm4				\n\t"

  127. 

  128. 		"						\n\t"

  129. #ifdef HAVE_MMX2

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

  131. 		"psadbw %%mm7, %%mm0				\n\t"

  132. #else

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

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

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

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

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

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

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

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

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

  142. #endif

  143.                 "movq %6, %%mm7					\n\t" // QP,..., QP

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

  145. 		"psubusb %%mm7, %%mm4				\n\t" // Diff <= 2QP -> 0

  146. 		"packssdw %%mm4, %%mm4				\n\t"

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

  148. 		"movd %%mm4, %1					\n\t"

  149. 

  150. 		: "=r" (numEq), "=r" (dcOk)

  151. 		: "r" (src), "r" (stride), "m" (c->mmxDcOffset[c->nonBQP]),  "m" (c->mmxDcThreshold[c->nonBQP]), "m" (c->pQPb)

  152. 		: "%eax"

  153. 		);

  154. 

  155. 	numEq= (-numEq) &0xFF;

  156. 	if(numEq > c->ppMode.flatnessThreshold){

  157.             if(dcOk) return 0;

  158.             else     return 1;

  159.         }else{

  160.             return 2;

  161.         }

  162. }

  163. #endif

  164. 

  165. /**

  166.  * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)

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

  168.  */

  169. static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)

  170. {

  171. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  172. 	src+= stride*3;

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

  174. 		"movq %2, %%mm0			\n\t"  // QP,..., QP

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

  176. 

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

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

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

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

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

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

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

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

  185. 		"pcmpeqb %%mm4, %%mm2			\n\t" // diff <= QP -> FF

  186. 

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

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

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

  190. 

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

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

  193. 		"leal (%0, %1, 8), %%ecx			\n\t"

  194. 		"subl %1, %%ecx					\n\t"

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

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

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

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

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

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

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

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

  203. 		"pcmpeqb %%mm4, %%mm2			\n\t" // diff <= QP -> FF

  204. 

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

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

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

  208. 

  209. 

  210. 		// 	1	2	3	4	5	6	7	8

  211. 		//	%0	%0+%1	%0+2%1	eax	%0+4%1	eax+2%1	ecx	eax+4%1

  212. 		// 6 4 2 2 1 1

  213. 		// 6 4 4 2

  214. 		// 6 8 2

  215. 

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

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

  218. 		PAVGB(%%mm6, %%mm0)				      //1 1	/2

  219. 		PAVGB(%%mm6, %%mm0)				      //3 1	/4

  220. 

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

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

  223. 		PAVGB((%%eax), %%mm2)				      //    11	/2

  224. 		PAVGB((%0, %1, 2), %%mm2)			      //   211	/4

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

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

  227. 		PAVGB(%%mm4, %%mm3)				      // 4 211	/8

  228. 		PAVGB(%%mm0, %%mm3)				      //642211	/16

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

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

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

  232. 		PAVGB(%%mm6, %%mm0)				      //1 1	/2

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

  234. 		PAVGB((%0,%1,2), %%mm3)				      // 1 1	/2

  235. 		PAVGB((%%eax,%1,2), %%mm5)			      //     11	/2

  236. 		PAVGB((%%eax), %%mm5)				      //    211 /4

  237. 		PAVGB(%%mm5, %%mm3)				      // 2 2211 /8

  238. 		PAVGB(%%mm0, %%mm3)				      //4242211 /16

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

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

  241. 		PAVGB(%%mm4, %%mm6)				      //11	/2

  242. 		"movq (%%ecx), %%mm0				\n\t" //       1

  243. 		PAVGB((%%eax, %1, 2), %%mm0)			      //      11/2

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

  245. 		PAVGB(%%mm1, %%mm0)				      //  2   11/4

  246. 		PAVGB(%%mm6, %%mm0)				      //222   11/8

  247. 		PAVGB(%%mm2, %%mm0)				      //22242211/16

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

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

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

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

  252. 		PAVGB((%%ecx), %%mm0)				      //       11	/2

  253. 		PAVGB(%%mm0, %%mm6)				      //11     11	/4

  254. 		PAVGB(%%mm1, %%mm4)				      // 11		/2

  255. 		PAVGB(%%mm2, %%mm1)				      //  11		/2

  256. 		PAVGB(%%mm1, %%mm6)				      //1122   11	/8

  257. 		PAVGB(%%mm5, %%mm6)				      //112242211	/16

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

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

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

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

  262. 		PAVGB(%%mm7, %%mm6)				      //        11	/2

  263. 		PAVGB(%%mm4, %%mm6)				      // 11     11	/4

  264. 		PAVGB(%%mm3, %%mm6)				      // 11   2211	/8

  265. 		PAVGB(%%mm5, %%mm2)				      //   11		/2

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

  267. 		PAVGB(%%mm4, %%mm2)				      //   112		/4

  268. 		PAVGB(%%mm2, %%mm6)				      // 112242211	/16

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

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

  271. 		PAVGB(%%mm7, %%mm1)				      //  11     2	/4

  272. 		PAVGB(%%mm4, %%mm5)				      //    11		/2

  273. 		PAVGB(%%mm5, %%mm0)				      //    11 11	/4

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

  275. 		PAVGB(%%mm6, %%mm1)				      //  11  4  2	/8

  276. 		PAVGB(%%mm0, %%mm1)				      //  11224222	/16

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

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

  279. 		PAVGB((%%ecx), %%mm2)				      //   112 4	/8

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

  281. 		PAVGB(%%mm0, %%mm6)				      //      1 1	/2

  282. 		PAVGB(%%mm7, %%mm6)				      //      1 12	/4

  283. 		PAVGB(%%mm2, %%mm6)				      //   1122424	/4

  284. 		"movq %%mm6, (%%ecx)				\n\t" //       X

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

  286. 		PAVGB(%%mm7, %%mm5)				      //    11   2	/4

  287. 		PAVGB(%%mm7, %%mm5)				      //    11   6	/8

  288. 

  289. 		PAVGB(%%mm3, %%mm0)				      //      112	/4

  290. 		PAVGB(%%mm0, %%mm5)				      //    112246	/16

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

  292. 		"subl %1, %0					\n\t"

  293. 

  294. 		:

  295. 		: "r" (src), "r" (stride), "m" (c->pQPb)

  296. 		: "%eax", "%ecx"

  297. 	);

  298. #else

  299. 	const int l1= stride;

  300. 	const int l2= stride + l1;

  301. 	const int l3= stride + l2;

  302. 	const int l4= stride + l3;

  303. 	const int l5= stride + l4;

  304. 	const int l6= stride + l5;

  305. 	const int l7= stride + l6;

  306. 	const int l8= stride + l7;

  307. 	const int l9= stride + l8;

  308. 	int x;

  309. 	src+= stride*3;

  310. 	for(x=0; x<BLOCK_SIZE; x++)

  311. 	{

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

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

  314. 

  315. 		int sums[9];

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

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

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

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

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

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

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

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

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

  325. 

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

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

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

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

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

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

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

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

  334. 

  335. 		src++;

  336. 	}

  337. #endif

  338. }

  339. 

  340. #if 0

  341. /**

  342.  * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar

  343.  * values are correctly clipped (MMX2)

  344.  * values are wraparound (C)

  345.  * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient

  346. 	0 8 16 24

  347. 	x = 8

  348. 	x/2 = 4

  349. 	x/8 = 1

  350. 	1 12 12 23

  351.  */

  352. static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)

  353. {

  354. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  355. 	src+= stride*3;

  356. // FIXME rounding

  357. 	asm volatile(

  358. 		"pxor %%mm7, %%mm7				\n\t" // 0

  359. 		"movq "MANGLE(b80)", %%mm6			\n\t" // MIN_SIGNED_BYTE

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

  361. 		"leal (%%eax, %1, 4), %%ecx			\n\t"

  362. //	0	1	2	3	4	5	6	7	8	9

  363. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1	%0+8%1	ecx+4%1

  364. 		"movq "MANGLE(pQPb)", %%mm0			\n\t" // QP,..., QP

  365. 		"movq %%mm0, %%mm1				\n\t" // QP,..., QP

  366. 		"paddusb "MANGLE(b02)", %%mm0			\n\t"

  367. 		"psrlw $2, %%mm0				\n\t"

  368. 		"pand "MANGLE(b3F)", %%mm0			\n\t" // QP/4,..., QP/4

  369. 		"paddusb %%mm1, %%mm0				\n\t" // QP*1.25 ...

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

  371. 		"movq (%%ecx), %%mm3				\n\t" // line 5

  372. 		"movq %%mm2, %%mm4				\n\t" // line 4

  373. 		"pcmpeqb %%mm5, %%mm5				\n\t" // -1

  374. 		"pxor %%mm2, %%mm5				\n\t" // -line 4 - 1

  375. 		PAVGB(%%mm3, %%mm5)

  376. 		"paddb %%mm6, %%mm5				\n\t" // (l5-l4)/2

  377. 		"psubusb %%mm3, %%mm4				\n\t"

  378. 		"psubusb %%mm2, %%mm3				\n\t"

  379. 		"por %%mm3, %%mm4				\n\t" // |l4 - l5|

  380. 		"psubusb %%mm0, %%mm4				\n\t"

  381. 		"pcmpeqb %%mm7, %%mm4				\n\t"

  382. 		"pand %%mm4, %%mm5				\n\t" // d/2

  383. 

  384. //		"paddb %%mm6, %%mm2				\n\t" // line 4 + 0x80

  385. 		"paddb %%mm5, %%mm2				\n\t"

  386. //		"psubb %%mm6, %%mm2				\n\t"

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

  388. 

  389. 		"movq (%%ecx), %%mm2				\n\t"

  390. //		"paddb %%mm6, %%mm2				\n\t" // line 5 + 0x80

  391. 		"psubb %%mm5, %%mm2				\n\t"

  392. //		"psubb %%mm6, %%mm2				\n\t"

  393. 		"movq %%mm2, (%%ecx)				\n\t"

  394. 

  395. 		"paddb %%mm6, %%mm5				\n\t"

  396. 		"psrlw $2, %%mm5				\n\t"

  397. 		"pand "MANGLE(b3F)", %%mm5			\n\t"

  398. 		"psubb "MANGLE(b20)", %%mm5			\n\t" // (l5-l4)/8

  399. 

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

  401. 		"paddb %%mm6, %%mm2				\n\t" // line 3 + 0x80

  402. 		"paddsb %%mm5, %%mm2				\n\t"

  403. 		"psubb %%mm6, %%mm2				\n\t"

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

  405. 

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

  407. 		"paddb %%mm6, %%mm2				\n\t" // line 6 + 0x80

  408. 		"psubsb %%mm5, %%mm2				\n\t"

  409. 		"psubb %%mm6, %%mm2				\n\t"

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

  411. 

  412. 		:

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

  414. 		: "%eax", "%ecx"

  415. 	);

  416. #else

  417.  	const int l1= stride;

  418. 	const int l2= stride + l1;

  419. 	const int l3= stride + l2;

  420. 	const int l4= stride + l3;

  421. 	const int l5= stride + l4;

  422. 	const int l6= stride + l5;

  423. //	const int l7= stride + l6;

  424. //	const int l8= stride + l7;

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

  426. 	int x;

  427. 	const int QP15= QP + (QP>>2);

  428. 	src+= stride*3;

  429. 	for(x=0; x<BLOCK_SIZE; x++)

  430. 	{

  431. 		const int v = (src[x+l5] - src[x+l4]);

  432. 		if(ABS(v) < QP15)

  433. 		{

  434. 			src[x+l3] +=v>>3;

  435. 			src[x+l4] +=v>>1;

  436. 			src[x+l5] -=v>>1;

  437. 			src[x+l6] -=v>>3;

  438. 

  439. 		}

  440. 	}

  441. 

  442. #endif

  443. }

  444. #endif

  445. 

  446. /**

  447.  * Experimental Filter 1

  448.  * will not damage linear gradients

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

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

  451.  * MMX2 version does correct clipping C version doesnt

  452.  */

  453. static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)

  454. {

  455. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  456. 	src+= stride*3;

  457. 

  458. 	asm volatile(

  459. 		"pxor %%mm7, %%mm7				\n\t" // 0

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

  461. 		"leal (%%eax, %1, 4), %%ecx			\n\t"

  462. //	0	1	2	3	4	5	6	7	8	9

  463. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1	%0+8%1	ecx+4%1

  464. 		"movq (%%eax, %1, 2), %%mm0			\n\t" // line 3

  465. 		"movq (%0, %1, 4), %%mm1			\n\t" // line 4

  466. 		"movq %%mm1, %%mm2				\n\t" // line 4

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

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

  469. 		"por %%mm1, %%mm0				\n\t" // |l2 - l3|

  470. 		"movq (%%ecx), %%mm3				\n\t" // line 5

  471. 		"movq (%%ecx, %1), %%mm4			\n\t" // line 6

  472. 		"movq %%mm3, %%mm5				\n\t" // line 5

  473. 		"psubusb %%mm4, %%mm3				\n\t"

  474. 		"psubusb %%mm5, %%mm4				\n\t"

  475. 		"por %%mm4, %%mm3				\n\t" // |l5 - l6|

  476. 		PAVGB(%%mm3, %%mm0)				      // (|l2 - l3| + |l5 - l6|)/2

  477. 		"movq %%mm2, %%mm1				\n\t" // line 4

  478. 		"psubusb %%mm5, %%mm2				\n\t"

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

  480. 		"pcmpeqb %%mm7, %%mm2				\n\t" // (l4 - l5) <= 0 ? -1 : 0

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

  482. 		"por %%mm5, %%mm4				\n\t" // |l4 - l5|

  483. 		"psubusb %%mm0, %%mm4		\n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)

  484. 		"movq %%mm4, %%mm3				\n\t" // d

  485. 		"movq %2, %%mm0			\n\t"

  486.                 "paddusb %%mm0, %%mm0				\n\t"

  487. 		"psubusb %%mm0, %%mm4				\n\t"

  488. 		"pcmpeqb %%mm7, %%mm4				\n\t" // d <= QP ? -1 : 0

  489. 		"psubusb "MANGLE(b01)", %%mm3			\n\t"

  490. 		"pand %%mm4, %%mm3				\n\t" // d <= QP ? d : 0

  491. 

  492. 		PAVGB(%%mm7, %%mm3)				      // d/2

  493. 		"movq %%mm3, %%mm1				\n\t" // d/2

  494. 		PAVGB(%%mm7, %%mm3)				      // d/4

  495. 		PAVGB(%%mm1, %%mm3)				      // 3*d/8

  496. 

  497. 		"movq (%0, %1, 4), %%mm0			\n\t" // line 4

  498. 		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4

  499. 		"psubusb %%mm3, %%mm0				\n\t"

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

  501. 		"movq %%mm0, (%0, %1, 4)			\n\t" // line 4

  502. 

  503. 		"movq (%%ecx), %%mm0				\n\t" // line 5

  504. 		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l5-1 : l5

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

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

  507. 		"movq %%mm0, (%%ecx)				\n\t" // line 5

  508. 

  509. 		PAVGB(%%mm7, %%mm1)				      // d/4

  510. 

  511. 		"movq (%%eax, %1, 2), %%mm0			\n\t" // line 3

  512. 		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4

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

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

  515. 		"movq %%mm0, (%%eax, %1, 2)			\n\t" // line 3

  516. 

  517. 		"movq (%%ecx, %1), %%mm0			\n\t" // line 6

  518. 		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l5-1 : l5

  519. 		"paddusb %%mm1, %%mm0				\n\t"

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

  521. 		"movq %%mm0, (%%ecx, %1)			\n\t" // line 6

  522. 

  523. 		PAVGB(%%mm7, %%mm1)				      // d/8

  524. 

  525. 		"movq (%%eax, %1), %%mm0			\n\t" // line 2

  526. 		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l2-1 : l2

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

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

  529. 		"movq %%mm0, (%%eax, %1)			\n\t" // line 2

  530. 

  531. 		"movq (%%ecx, %1, 2), %%mm0			\n\t" // line 7

  532. 		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l7-1 : l7

  533. 		"paddusb %%mm1, %%mm0				\n\t"

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

  535. 		"movq %%mm0, (%%ecx, %1, 2)			\n\t" // line 7

  536. 

  537. 		:

  538. 		: "r" (src), "r" (stride), "m" (co->pQPb)

  539. 		: "%eax", "%ecx"

  540. 	);

  541. #else

  542. 

  543.  	const int l1= stride;

  544. 	const int l2= stride + l1;

  545. 	const int l3= stride + l2;

  546. 	const int l4= stride + l3;

  547. 	const int l5= stride + l4;

  548. 	const int l6= stride + l5;

  549. 	const int l7= stride + l6;

  550. //	const int l8= stride + l7;

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

  552. 	int x;

  553. 

  554. 	src+= stride*3;

  555. 	for(x=0; x<BLOCK_SIZE; x++)

  556. 	{

  557. 		int a= src[l3] - src[l4];

  558. 		int b= src[l4] - src[l5];

  559. 		int c= src[l5] - src[l6];

  560. 

  561. 		int d= ABS(b) - ((ABS(a) + ABS(c))>>1);

  562. 		d= MAX(d, 0);

  563. 

  564. 		if(d < co->QP*2)

  565. 		{

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

  567. 

  568. 			src[l2] +=v>>3;

  569. 			src[l3] +=v>>2;

  570. 			src[l4] +=(3*v)>>3;

  571. 			src[l5] -=(3*v)>>3;

  572. 			src[l6] -=v>>2;

  573. 			src[l7] -=v>>3;

  574. 

  575. 		}

  576. 		src++;

  577. 	}

  578. #endif

  579. }

  580. 

  581. static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)

  582. {

  583. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  584. /*

  585. 	uint8_t tmp[16];

  586. 	const int l1= stride;

  587. 	const int l2= stride + l1;

  588. 	const int l3= stride + l2;

  589. 	const int l4= (int)tmp - (int)src - stride*3;

  590. 	const int l5= (int)tmp - (int)src - stride*3 + 8;

  591. 	const int l6= stride*3 + l3;

  592. 	const int l7= stride + l6;

  593. 	const int l8= stride + l7;

  594. 

  595. 	memcpy(tmp, src+stride*7, 8);

  596. 	memcpy(tmp+8, src+stride*8, 8);

  597. */

  598. 	src+= stride*4;

  599. 	asm volatile(

  600. 

  601. #if 0 //sligtly more accurate and slightly slower

  602. 		"pxor %%mm7, %%mm7				\n\t" // 0

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

  604. 		"leal (%%eax, %1, 4), %%ecx			\n\t"

  605. //	0	1	2	3	4	5	6	7

  606. //	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	ecx+%1	ecx+2%1

  607. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1

  608. 

  609. 

  610. 		"movq (%0, %1, 2), %%mm0			\n\t" // l2

  611. 		"movq (%0), %%mm1				\n\t" // l0

  612. 		"movq %%mm0, %%mm2				\n\t" // l2

  613. 		PAVGB(%%mm7, %%mm0)				      // ~l2/2

  614. 		PAVGB(%%mm1, %%mm0)				      // ~(l2 + 2l0)/4

  615. 		PAVGB(%%mm2, %%mm0)				      // ~(5l2 + 2l0)/8

  616. 

  617. 		"movq (%%eax), %%mm1				\n\t" // l1

  618. 		"movq (%%eax, %1, 2), %%mm3			\n\t" // l3

  619. 		"movq %%mm1, %%mm4				\n\t" // l1

  620. 		PAVGB(%%mm7, %%mm1)				      // ~l1/2

  621. 		PAVGB(%%mm3, %%mm1)				      // ~(l1 + 2l3)/4

  622. 		PAVGB(%%mm4, %%mm1)				      // ~(5l1 + 2l3)/8

  623. 

  624. 		"movq %%mm0, %%mm4				\n\t" // ~(5l2 + 2l0)/8

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

  626. 		"psubusb %%mm4, %%mm1				\n\t"

  627. 		"por %%mm0, %%mm1				\n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8

  628. // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0

  629. 

  630. 		"movq (%0, %1, 4), %%mm0			\n\t" // l4

  631. 		"movq %%mm0, %%mm4				\n\t" // l4

  632. 		PAVGB(%%mm7, %%mm0)				      // ~l4/2

  633. 		PAVGB(%%mm2, %%mm0)				      // ~(l4 + 2l2)/4

  634. 		PAVGB(%%mm4, %%mm0)				      // ~(5l4 + 2l2)/8

  635. 

  636. 		"movq (%%ecx), %%mm2				\n\t" // l5

  637. 		"movq %%mm3, %%mm5				\n\t" // l3

  638. 		PAVGB(%%mm7, %%mm3)				      // ~l3/2

  639. 		PAVGB(%%mm2, %%mm3)				      // ~(l3 + 2l5)/4

  640. 		PAVGB(%%mm5, %%mm3)				      // ~(5l3 + 2l5)/8

  641. 

  642. 		"movq %%mm0, %%mm6				\n\t" // ~(5l4 + 2l2)/8

  643. 		"psubusb %%mm3, %%mm0				\n\t"

  644. 		"psubusb %%mm6, %%mm3				\n\t"

  645. 		"por %%mm0, %%mm3				\n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8

  646. 		"pcmpeqb %%mm7, %%mm0				\n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)

  647. // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0

  648. 

  649. 		"movq (%%ecx, %1), %%mm6			\n\t" // l6

  650. 		"movq %%mm6, %%mm5				\n\t" // l6

  651. 		PAVGB(%%mm7, %%mm6)				      // ~l6/2

  652. 		PAVGB(%%mm4, %%mm6)				      // ~(l6 + 2l4)/4

  653. 		PAVGB(%%mm5, %%mm6)				      // ~(5l6 + 2l4)/8

  654. 

  655. 		"movq (%%ecx, %1, 2), %%mm5			\n\t" // l7

  656. 		"movq %%mm2, %%mm4				\n\t" // l5

  657. 		PAVGB(%%mm7, %%mm2)				      // ~l5/2

  658. 		PAVGB(%%mm5, %%mm2)				      // ~(l5 + 2l7)/4

  659. 		PAVGB(%%mm4, %%mm2)				      // ~(5l5 + 2l7)/8

  660. 

  661. 		"movq %%mm6, %%mm4				\n\t" // ~(5l6 + 2l4)/8

  662. 		"psubusb %%mm2, %%mm6				\n\t"

  663. 		"psubusb %%mm4, %%mm2				\n\t"

  664. 		"por %%mm6, %%mm2				\n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8

  665. // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0

  666. 

  667. 

  668. 		PMINUB(%%mm2, %%mm1, %%mm4)			      // MIN(|lenergy|,|renergy|)/8

  669. 		"movq %2, %%mm4					\n\t" // QP //FIXME QP+1 ?

  670. 		"paddusb "MANGLE(b01)", %%mm4			\n\t"

  671. 		"pcmpgtb %%mm3, %%mm4				\n\t" // |menergy|/8 < QP

  672. 		"psubusb %%mm1, %%mm3				\n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8

  673. 		"pand %%mm4, %%mm3				\n\t"

  674. 

  675. 		"movq %%mm3, %%mm1				\n\t"

  676. //		"psubusb "MANGLE(b01)", %%mm3			\n\t"

  677. 		PAVGB(%%mm7, %%mm3)

  678. 		PAVGB(%%mm7, %%mm3)

  679. 		"paddusb %%mm1, %%mm3				\n\t"

  680. //		"paddusb "MANGLE(b01)", %%mm3			\n\t"

  681. 

  682. 		"movq (%%eax, %1, 2), %%mm6			\n\t" //l3

  683. 		"movq (%0, %1, 4), %%mm5			\n\t" //l4

  684. 		"movq (%0, %1, 4), %%mm4			\n\t" //l4

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

  686. 		"psubusb %%mm4, %%mm6				\n\t"

  687. 		"por %%mm6, %%mm5				\n\t" // |l3-l4|

  688. 		"pcmpeqb %%mm7, %%mm6				\n\t" // SIGN(l3-l4)

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

  690. 		"pand %%mm0, %%mm3				\n\t"

  691. 		PMINUB(%%mm5, %%mm3, %%mm0)

  692. 

  693. 		"psubusb "MANGLE(b01)", %%mm3			\n\t"

  694. 		PAVGB(%%mm7, %%mm3)

  695. 

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

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

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

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

  700. 		"psubb %%mm3, %%mm0				\n\t"

  701. 		"paddb %%mm3, %%mm2				\n\t"

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

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

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

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

  706. #endif

  707. 

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

  709. 		"pcmpeqb %%mm6, %%mm6				\n\t" // -1

  710. //	0	1	2	3	4	5	6	7

  711. //	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	ecx+%1	ecx+2%1

  712. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1

  713. 

  714. 

  715. 		"movq (%%eax, %1, 2), %%mm1			\n\t" // l3

  716. 		"movq (%0, %1, 4), %%mm0			\n\t" // l4

  717. 		"pxor %%mm6, %%mm1				\n\t" // -l3-1

  718. 		PAVGB(%%mm1, %%mm0)				      // -q+128 = (l4-l3+256)/2

  719. // mm1=-l3-1, mm0=128-q

  720. 

  721. 		"movq (%%eax, %1, 4), %%mm2			\n\t" // l5

  722. 		"movq (%%eax, %1), %%mm3			\n\t" // l2

  723. 		"pxor %%mm6, %%mm2				\n\t" // -l5-1

  724. 		"movq %%mm2, %%mm5				\n\t" // -l5-1

  725. 		"movq "MANGLE(b80)", %%mm4			\n\t" // 128

  726. 		"leal (%%eax, %1, 4), %%ecx			\n\t"

  727. 		PAVGB(%%mm3, %%mm2)				      // (l2-l5+256)/2

  728. 		PAVGB(%%mm0, %%mm4)				      // ~(l4-l3)/4 + 128

  729. 		PAVGB(%%mm2, %%mm4)				      // ~(l2-l5)/4 +(l4-l3)/8 + 128

  730. 		PAVGB(%%mm0, %%mm4)				      // ~(l2-l5)/8 +5(l4-l3)/16 + 128

  731. // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1

  732. 

  733. 		"movq (%%eax), %%mm2				\n\t" // l1

  734. 		"pxor %%mm6, %%mm2				\n\t" // -l1-1

  735. 		PAVGB(%%mm3, %%mm2)				      // (l2-l1+256)/2

  736. 		PAVGB((%0), %%mm1)				      // (l0-l3+256)/2

  737. 		"movq "MANGLE(b80)", %%mm3			\n\t" // 128

  738. 		PAVGB(%%mm2, %%mm3)				      // ~(l2-l1)/4 + 128

  739. 		PAVGB(%%mm1, %%mm3)				      // ~(l0-l3)/4 +(l2-l1)/8 + 128

  740. 		PAVGB(%%mm2, %%mm3)				      // ~(l0-l3)/8 +5(l2-l1)/16 + 128

  741. // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1

  742. 

  743. 		PAVGB((%%ecx, %1), %%mm5)			      // (l6-l5+256)/2

  744. 		"movq (%%ecx, %1, 2), %%mm1			\n\t" // l7

  745. 		"pxor %%mm6, %%mm1				\n\t" // -l7-1

  746. 		PAVGB((%0, %1, 4), %%mm1)			      // (l4-l7+256)/2

  747. 		"movq "MANGLE(b80)", %%mm2			\n\t" // 128

  748. 		PAVGB(%%mm5, %%mm2)				      // ~(l6-l5)/4 + 128

  749. 		PAVGB(%%mm1, %%mm2)				      // ~(l4-l7)/4 +(l6-l5)/8 + 128

  750. 		PAVGB(%%mm5, %%mm2)				      // ~(l4-l7)/8 +5(l6-l5)/16 + 128

  751. // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128

  752. 

  753. 		"movq "MANGLE(b00)", %%mm1			\n\t" // 0

  754. 		"movq "MANGLE(b00)", %%mm5			\n\t" // 0

  755. 		"psubb %%mm2, %%mm1				\n\t" // 128 - renergy/16

  756. 		"psubb %%mm3, %%mm5				\n\t" // 128 - lenergy/16

  757. 		PMAXUB(%%mm1, %%mm2)				      // 128 + |renergy/16|

  758.  		PMAXUB(%%mm5, %%mm3)				      // 128 + |lenergy/16|

  759. 		PMINUB(%%mm2, %%mm3, %%mm1)			      // 128 + MIN(|lenergy|,|renergy|)/16

  760. 

  761. // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128

  762. 

  763. 		"movq "MANGLE(b00)", %%mm7			\n\t" // 0

  764. 		"movq %2, %%mm2					\n\t" // QP

  765. 		PAVGB(%%mm6, %%mm2)				      // 128 + QP/2

  766. 		"psubb %%mm6, %%mm2				\n\t"

  767. 

  768. 		"movq %%mm4, %%mm1				\n\t"

  769. 		"pcmpgtb %%mm7, %%mm1				\n\t" // SIGN(menergy)

  770. 		"pxor %%mm1, %%mm4				\n\t"

  771. 		"psubb %%mm1, %%mm4				\n\t" // 128 + |menergy|/16

  772. 		"pcmpgtb %%mm4, %%mm2				\n\t" // |menergy|/16 < QP/2

  773. 		"psubusb %%mm3, %%mm4				\n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16

  774. // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16

  775. 

  776. 		"movq %%mm4, %%mm3				\n\t" // d

  777. 		"psubusb "MANGLE(b01)", %%mm4			\n\t"

  778. 		PAVGB(%%mm7, %%mm4)				      // d/32

  779. 		PAVGB(%%mm7, %%mm4)				      // (d + 32)/64

  780. 		"paddb %%mm3, %%mm4				\n\t" // 5d/64

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

  782. 

  783. 		"movq "MANGLE(b80)", %%mm5			\n\t" // 128

  784. 		"psubb %%mm0, %%mm5				\n\t" // q

  785. 		"paddsb %%mm6, %%mm5				\n\t" // fix bad rounding

  786. 		"pcmpgtb %%mm5, %%mm7				\n\t" // SIGN(q)

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

  788. 

  789. 		PMINUB(%%mm5, %%mm4, %%mm3)			      // MIN(|q|, 5d/64)

  790. 		"pxor %%mm1, %%mm7				\n\t" // SIGN(d*q)

  791. 

  792. 		"pand %%mm7, %%mm4				\n\t"

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

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

  795. 		"pxor %%mm1, %%mm0				\n\t"

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

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

  798. 		"psubb %%mm4, %%mm2				\n\t"

  799. 		"pxor %%mm1, %%mm0				\n\t"

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

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

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

  803. 

  804. 		:

  805. 		: "r" (src), "r" (stride), "m" (c->pQPb)

  806. 		: "%eax", "%ecx"

  807. 	);

  808. 

  809. /*

  810. 	{

  811. 	int x;

  812. 	src-= stride;

  813. 	for(x=0; x<BLOCK_SIZE; x++)

  814. 	{

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

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

  817. 		{

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

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

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

  821. 

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

  823. 			d= MAX(d, 0);

  824. 

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

  826. 			d*= SIGN(-middleEnergy);

  827. 

  828. 			if(q>0)

  829. 			{

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

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

  832. 			}

  833. 			else

  834. 			{

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

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

  837. 			}

  838. 

  839.         		src[l4]-= d;

  840. 	        	src[l5]+= d;

  841. 		}

  842. 		src++;

  843. 	}

  844. src-=8;

  845. 	for(x=0; x<8; x++)

  846. 	{

  847. 		int y;

  848. 		for(y=4; y<6; y++)

  849. 		{

  850. 			int d= src[x+y*stride] - tmp[x+(y-4)*8];

  851. 			int ad= ABS(d);

  852. 			static int max=0;

  853. 			static int sum=0;

  854. 			static int num=0;

  855. 			static int bias=0;

  856. 

  857. 			if(max<ad) max=ad;

  858. 			sum+= ad>3 ? 1 : 0;

  859. 			if(ad>3)

  860. 			{

  861. 				src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;

  862. 			}

  863. 			if(y==4) bias+=d;

  864. 			num++;

  865. 			if(num%1000000 == 0)

  866. 			{

  867. 				printf(" %d %d %d %d\n", num, sum, max, bias);

  868. 			}

  869. 		}

  870. 	}

  871. }

  872. */

  873. #elif defined (HAVE_MMX)

  874. 	src+= stride*4;

  875. 	asm volatile(

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

  877. 		"leal -40(%%esp), %%ecx				\n\t" // make space for 4 8-byte vars

  878. 		"andl $0xFFFFFFF8, %%ecx			\n\t" // align

  879. //	0	1	2	3	4	5	6	7

  880. //	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	edx+%1	edx+2%1

  881. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1

  882. 

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

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

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

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

  887. 

  888. 		"movq (%0, %1), %%mm2				\n\t"

  889. 		"leal (%0, %1, 2), %%eax			\n\t"

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

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

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

  893. 

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

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

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

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

  898. 

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

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

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

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

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

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

  905. 

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

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

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

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

  910. 

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

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

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

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

  915. 

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

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

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

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

  920. 		"movq %%mm0, (%%ecx)				\n\t" // 2L0 - 5L1 + 5L2 - 2L3

  921. 		"movq %%mm1, 8(%%ecx)				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

  922. 

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

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

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

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

  927. 

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

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

  930. 		"movq %%mm2, 16(%%ecx)				\n\t" // L3 - L4

  931. 		"movq %%mm3, 24(%%ecx)				\n\t" // H3 - H4

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

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

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

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

  936. 

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

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

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

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

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

  942. //50 opcodes so far

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

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

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

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

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

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

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

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

  951. 

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

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

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

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

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

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

  958. 

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

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

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

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

  963. 

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

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

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

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

  968. 

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

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

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

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

  973. 

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

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

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

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

  978. 

  979. 		"movq (%%ecx), %%mm2				\n\t" // 2L0 - 5L1 + 5L2 - 2L3

  980. 		"movq 8(%%ecx), %%mm3				\n\t" // 2H0 - 5H1 + 5H2 - 2H3

  981. 

  982. #ifdef HAVE_MMX2

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

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

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

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

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

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

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

  990. 		"psubw %%mm2, %%mm6				\n\t"

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

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

  993. 		"psubw %%mm3, %%mm6				\n\t"

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

  995. #else

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

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

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

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

  1000. 		"movq %%mm7, %%mm6				\n\t" // 0

  1001. 		"pcmpgtw %%mm1, %%mm6				\n\t"

  1002. 		"pxor %%mm6, %%mm1				\n\t"

  1003. 		"psubw %%mm6, %%mm1				\n\t" // |2H4 - 5H5 + 5H6 - 2H7|

  1004. 		"movq %%mm7, %%mm6				\n\t" // 0

  1005. 		"pcmpgtw %%mm2, %%mm6				\n\t"

  1006. 		"pxor %%mm6, %%mm2				\n\t"

  1007. 		"psubw %%mm6, %%mm2				\n\t" // |2L0 - 5L1 + 5L2 - 2L3|

  1008. 		"movq %%mm7, %%mm6				\n\t" // 0

  1009. 		"pcmpgtw %%mm3, %%mm6				\n\t"

  1010. 		"pxor %%mm6, %%mm3				\n\t"

  1011. 		"psubw %%mm6, %%mm3				\n\t" // |2H0 - 5H1 + 5H2 - 2H3|

  1012. #endif

  1013. 

  1014. #ifdef HAVE_MMX2

  1015. 		"pminsw %%mm2, %%mm0				\n\t"

  1016. 		"pminsw %%mm3, %%mm1				\n\t"

  1017. #else

  1018. 		"movq %%mm0, %%mm6				\n\t"

  1019. 		"psubusw %%mm2, %%mm6				\n\t"

  1020. 		"psubw %%mm6, %%mm0				\n\t"

  1021. 		"movq %%mm1, %%mm6				\n\t"

  1022. 		"psubusw %%mm3, %%mm6				\n\t"

  1023. 		"psubw %%mm6, %%mm1				\n\t"

  1024. #endif

  1025. 

  1026. 		"movq %%mm7, %%mm6				\n\t" // 0

  1027. 		"pcmpgtw %%mm4, %%mm6				\n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)

  1028. 		"pxor %%mm6, %%mm4				\n\t"

  1029. 		"psubw %%mm6, %%mm4				\n\t" // |2L2 - 5L3 + 5L4 - 2L5|

  1030. 		"pcmpgtw %%mm5, %%mm7				\n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)

  1031. 		"pxor %%mm7, %%mm5				\n\t"

  1032. 		"psubw %%mm7, %%mm5				\n\t" // |2H2 - 5H3 + 5H4 - 2H5|

  1033. // 100 opcodes

  1034. 		"movd %2, %%mm2					\n\t" // QP

  1035. 		"psllw $3, %%mm2				\n\t" // 8QP

  1036. 		"movq %%mm2, %%mm3				\n\t" // 8QP

  1037. 		"pcmpgtw %%mm4, %%mm2				\n\t"

  1038. 		"pcmpgtw %%mm5, %%mm3				\n\t"

  1039. 		"pand %%mm2, %%mm4				\n\t"

  1040. 		"pand %%mm3, %%mm5				\n\t"

  1041. 

  1042. 

  1043. 		"psubusw %%mm0, %%mm4				\n\t" // hd

  1044. 		"psubusw %%mm1, %%mm5				\n\t" // ld

  1045. 

  1046. 

  1047. 		"movq "MANGLE(w05)", %%mm2			\n\t" // 5

  1048. 		"pmullw %%mm2, %%mm4				\n\t"

  1049. 		"pmullw %%mm2, %%mm5				\n\t"

  1050. 		"movq "MANGLE(w20)", %%mm2			\n\t" // 32

  1051. 		"paddw %%mm2, %%mm4				\n\t"

  1052. 		"paddw %%mm2, %%mm5				\n\t"

  1053. 		"psrlw $6, %%mm4				\n\t"

  1054. 		"psrlw $6, %%mm5				\n\t"

  1055. 

  1056. 		"movq 16(%%ecx), %%mm0				\n\t" // L3 - L4

  1057. 		"movq 24(%%ecx), %%mm1				\n\t" // H3 - H4

  1058. 

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

  1060. 		"pxor %%mm3, %%mm3				\n\t"

  1061. 

  1062. 		"pcmpgtw %%mm0, %%mm2				\n\t" // sign (L3-L4)

  1063. 		"pcmpgtw %%mm1, %%mm3				\n\t" // sign (H3-H4)

  1064. 		"pxor %%mm2, %%mm0				\n\t"

  1065. 		"pxor %%mm3, %%mm1				\n\t"

  1066. 		"psubw %%mm2, %%mm0				\n\t" // |L3-L4|

  1067. 		"psubw %%mm3, %%mm1				\n\t" // |H3-H4|

  1068. 		"psrlw $1, %%mm0				\n\t" // |L3 - L4|/2

  1069. 		"psrlw $1, %%mm1				\n\t" // |H3 - H4|/2

  1070. 

  1071. 		"pxor %%mm6, %%mm2				\n\t"

  1072. 		"pxor %%mm7, %%mm3				\n\t"

  1073. 		"pand %%mm2, %%mm4				\n\t"

  1074. 		"pand %%mm3, %%mm5				\n\t"

  1075. 

  1076. #ifdef HAVE_MMX2

  1077. 		"pminsw %%mm0, %%mm4				\n\t"

  1078. 		"pminsw %%mm1, %%mm5				\n\t"

  1079. #else

  1080. 		"movq %%mm4, %%mm2				\n\t"

  1081. 		"psubusw %%mm0, %%mm2				\n\t"

  1082. 		"psubw %%mm2, %%mm4				\n\t"

  1083. 		"movq %%mm5, %%mm2				\n\t"

  1084. 		"psubusw %%mm1, %%mm2				\n\t"

  1085. 		"psubw %%mm2, %%mm5				\n\t"

  1086. #endif

  1087. 		"pxor %%mm6, %%mm4				\n\t"

  1088. 		"pxor %%mm7, %%mm5				\n\t"

  1089. 		"psubw %%mm6, %%mm4				\n\t"

  1090. 		"psubw %%mm7, %%mm5				\n\t"

  1091. 		"packsswb %%mm5, %%mm4				\n\t"

  1092. 		"movq (%0), %%mm0				\n\t"

  1093. 		"paddb   %%mm4, %%mm0				\n\t"

  1094. 		"movq %%mm0, (%0)				\n\t"

  1095. 		"movq (%0, %1), %%mm0				\n\t"

  1096. 		"psubb %%mm4, %%mm0				\n\t"

  1097. 		"movq %%mm0, (%0, %1)				\n\t"

  1098. 

  1099. 		: "+r" (src)

  1100. 		: "r" (stride), "m" (c->pQPb)

  1101. 		: "%eax", "%ecx"

  1102. 	);

  1103. #else

  1104. 	const int l1= stride;

  1105. 	const int l2= stride + l1;

  1106. 	const int l3= stride + l2;

  1107. 	const int l4= stride + l3;

  1108. 	const int l5= stride + l4;

  1109. 	const int l6= stride + l5;

  1110. 	const int l7= stride + l6;

  1111. 	const int l8= stride + l7;

  1112. //	const int l9= stride + l8;

  1113. 	int x;

  1114. 	src+= stride*3;

  1115. 	for(x=0; x<BLOCK_SIZE; x++)

  1116. 	{

  1117. 		const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);

  1118. 		if(ABS(middleEnergy) < 8*c->QP)

  1119. 		{

  1120. 			const int q=(src[l4] - src[l5])/2;

  1121. 			const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);

  1122. 			const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);

  1123. 

  1124. 			int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );

  1125. 			d= MAX(d, 0);

  1126. 

  1127. 			d= (5*d + 32) >> 6;

  1128. 			d*= SIGN(-middleEnergy);

  1129. 

  1130. 			if(q>0)

  1131. 			{

  1132. 				d= d<0 ? 0 : d;

  1133. 				d= d>q ? q : d;

  1134. 			}

  1135. 			else

  1136. 			{

  1137. 				d= d>0 ? 0 : d;

  1138. 				d= d<q ? q : d;

  1139. 			}

  1140. 

  1141.         		src[l4]-= d;

  1142. 	        	src[l5]+= d;

  1143. 		}

  1144. 		src++;

  1145. 	}

  1146. #endif

  1147. }

  1148. 

  1149. static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)

  1150. {

  1151. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  1152. 	asm volatile(

  1153. 		"pxor %%mm6, %%mm6				\n\t"

  1154. 		"pcmpeqb %%mm7, %%mm7				\n\t"

  1155. 		"movq %2, %%mm0					\n\t"

  1156. 		"punpcklbw %%mm6, %%mm0				\n\t"

  1157. 		"psrlw $1, %%mm0				\n\t"

  1158. 		"psubw %%mm7, %%mm0				\n\t"

  1159. 		"packuswb %%mm0, %%mm0				\n\t"

  1160. 		"movq %%mm0, %3					\n\t"

  1161. 

  1162. 		"leal (%0, %1), %%eax				\n\t"

  1163. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1164. 		

  1165. //	0	1	2	3	4	5	6	7	8	9

  1166. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

  1167. 

  1168. #undef FIND_MIN_MAX

  1169. #ifdef HAVE_MMX2

  1170. #define FIND_MIN_MAX(addr)\

  1171. 		"movq " #addr ", %%mm0				\n\t"\

  1172. 		"pminub %%mm0, %%mm7				\n\t"\

  1173. 		"pmaxub %%mm0, %%mm6				\n\t"

  1174. #else

  1175. #define FIND_MIN_MAX(addr)\

  1176. 		"movq " #addr ", %%mm0				\n\t"\

  1177. 		"movq %%mm7, %%mm1				\n\t"\

  1178. 		"psubusb %%mm0, %%mm6				\n\t"\

  1179. 		"paddb %%mm0, %%mm6				\n\t"\

  1180. 		"psubusb %%mm0, %%mm1				\n\t"\

  1181. 		"psubb %%mm1, %%mm7				\n\t"

  1182. #endif

  1183. 

  1184. FIND_MIN_MAX((%%eax))

  1185. FIND_MIN_MAX((%%eax, %1))

  1186. FIND_MIN_MAX((%%eax, %1, 2))

  1187. FIND_MIN_MAX((%0, %1, 4))

  1188. FIND_MIN_MAX((%%edx))

  1189. FIND_MIN_MAX((%%edx, %1))

  1190. FIND_MIN_MAX((%%edx, %1, 2))

  1191. FIND_MIN_MAX((%0, %1, 8))

  1192. 

  1193. 		"movq %%mm7, %%mm4				\n\t"

  1194. 		"psrlq $8, %%mm7				\n\t"

  1195. #ifdef HAVE_MMX2

  1196. 		"pminub %%mm4, %%mm7				\n\t" // min of pixels

  1197. 		"pshufw $0xF9, %%mm7, %%mm4			\n\t"

  1198. 		"pminub %%mm4, %%mm7				\n\t" // min of pixels

  1199. 		"pshufw $0xFE, %%mm7, %%mm4			\n\t"

  1200. 		"pminub %%mm4, %%mm7				\n\t"

  1201. #else

  1202. 		"movq %%mm7, %%mm1				\n\t"

  1203. 		"psubusb %%mm4, %%mm1				\n\t"

  1204. 		"psubb %%mm1, %%mm7				\n\t"

  1205. 		"movq %%mm7, %%mm4				\n\t"

  1206. 		"psrlq $16, %%mm7				\n\t"

  1207. 		"movq %%mm7, %%mm1				\n\t"

  1208. 		"psubusb %%mm4, %%mm1				\n\t"

  1209. 		"psubb %%mm1, %%mm7				\n\t"

  1210. 		"movq %%mm7, %%mm4				\n\t"

  1211. 		"psrlq $32, %%mm7				\n\t"

  1212. 		"movq %%mm7, %%mm1				\n\t"

  1213. 		"psubusb %%mm4, %%mm1				\n\t"

  1214. 		"psubb %%mm1, %%mm7				\n\t"

  1215. #endif

  1216. 

  1217. 

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

  1219. 		"psrlq $8, %%mm6				\n\t"

  1220. #ifdef HAVE_MMX2

  1221. 		"pmaxub %%mm4, %%mm6				\n\t" // max of pixels

  1222. 		"pshufw $0xF9, %%mm6, %%mm4			\n\t"

  1223. 		"pmaxub %%mm4, %%mm6				\n\t"

  1224. 		"pshufw $0xFE, %%mm6, %%mm4			\n\t"

  1225. 		"pmaxub %%mm4, %%mm6				\n\t"

  1226. #else

  1227. 		"psubusb %%mm4, %%mm6				\n\t"

  1228. 		"paddb %%mm4, %%mm6				\n\t"

  1229. 		"movq %%mm6, %%mm4				\n\t"

  1230. 		"psrlq $16, %%mm6				\n\t"

  1231. 		"psubusb %%mm4, %%mm6				\n\t"

  1232. 		"paddb %%mm4, %%mm6				\n\t"

  1233. 		"movq %%mm6, %%mm4				\n\t"

  1234. 		"psrlq $32, %%mm6				\n\t"

  1235. 		"psubusb %%mm4, %%mm6				\n\t"

  1236. 		"paddb %%mm4, %%mm6				\n\t"

  1237. #endif

  1238. 		"movq %%mm6, %%mm0				\n\t" // max

  1239. 		"psubb %%mm7, %%mm6				\n\t" // max - min

  1240. 		"movd %%mm6, %%ecx				\n\t"

  1241. 		"cmpb "MANGLE(deringThreshold)", %%cl		\n\t"

  1242. 		" jb 1f						\n\t"

  1243. 		"leal -24(%%esp), %%ecx				\n\t"

  1244. 		"andl $0xFFFFFFF8, %%ecx			\n\t" 

  1245. 		PAVGB(%%mm0, %%mm7)				      // a=(max + min)/2

  1246. 		"punpcklbw %%mm7, %%mm7				\n\t"

  1247. 		"punpcklbw %%mm7, %%mm7				\n\t"

  1248. 		"punpcklbw %%mm7, %%mm7				\n\t"

  1249. 		"movq %%mm7, (%%ecx)				\n\t"

  1250. 

  1251. 		"movq (%0), %%mm0				\n\t" // L10

  1252. 		"movq %%mm0, %%mm1				\n\t" // L10

  1253. 		"movq %%mm0, %%mm2				\n\t" // L10

  1254. 		"psllq $8, %%mm1				\n\t"

  1255. 		"psrlq $8, %%mm2				\n\t"

  1256. 		"movd -4(%0), %%mm3				\n\t"

  1257. 		"movd 8(%0), %%mm4				\n\t"

  1258. 		"psrlq $24, %%mm3				\n\t"

  1259. 		"psllq $56, %%mm4				\n\t"

  1260. 		"por %%mm3, %%mm1				\n\t" // L00

  1261. 		"por %%mm4, %%mm2				\n\t" // L20

  1262. 		"movq %%mm1, %%mm3				\n\t" // L00

  1263. 		PAVGB(%%mm2, %%mm1)				      // (L20 + L00)/2

  1264. 		PAVGB(%%mm0, %%mm1)				      // (L20 + L00 + 2L10)/4

  1265. 		"psubusb %%mm7, %%mm0				\n\t"

  1266. 		"psubusb %%mm7, %%mm2				\n\t"

  1267. 		"psubusb %%mm7, %%mm3				\n\t"

  1268. 		"pcmpeqb "MANGLE(b00)", %%mm0			\n\t" // L10 > a ? 0 : -1

  1269. 		"pcmpeqb "MANGLE(b00)", %%mm2			\n\t" // L20 > a ? 0 : -1

  1270. 		"pcmpeqb "MANGLE(b00)", %%mm3			\n\t" // L00 > a ? 0 : -1

  1271. 		"paddb %%mm2, %%mm0				\n\t"

  1272. 		"paddb %%mm3, %%mm0				\n\t"

  1273. 

  1274. 		"movq (%%eax), %%mm2				\n\t" // L11

  1275. 		"movq %%mm2, %%mm3				\n\t" // L11

  1276. 		"movq %%mm2, %%mm4				\n\t" // L11

  1277. 		"psllq $8, %%mm3				\n\t"

  1278. 		"psrlq $8, %%mm4				\n\t"

  1279. 		"movd -4(%%eax), %%mm5				\n\t"

  1280. 		"movd 8(%%eax), %%mm6				\n\t"

  1281. 		"psrlq $24, %%mm5				\n\t"

  1282. 		"psllq $56, %%mm6				\n\t"

  1283. 		"por %%mm5, %%mm3				\n\t" // L01

  1284. 		"por %%mm6, %%mm4				\n\t" // L21

  1285. 		"movq %%mm3, %%mm5				\n\t" // L01

  1286. 		PAVGB(%%mm4, %%mm3)				      // (L21 + L01)/2

  1287. 		PAVGB(%%mm2, %%mm3)				      // (L21 + L01 + 2L11)/4

  1288. 		"psubusb %%mm7, %%mm2				\n\t"

  1289. 		"psubusb %%mm7, %%mm4				\n\t"

  1290. 		"psubusb %%mm7, %%mm5				\n\t"

  1291. 		"pcmpeqb "MANGLE(b00)", %%mm2			\n\t" // L11 > a ? 0 : -1

  1292. 		"pcmpeqb "MANGLE(b00)", %%mm4			\n\t" // L21 > a ? 0 : -1

  1293. 		"pcmpeqb "MANGLE(b00)", %%mm5			\n\t" // L01 > a ? 0 : -1

  1294. 		"paddb %%mm4, %%mm2				\n\t"

  1295. 		"paddb %%mm5, %%mm2				\n\t"

  1296. // 0, 2, 3, 1

  1297. #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \

  1298. 		"movq " #src ", " #sx "				\n\t" /* src[0] */\

  1299. 		"movq " #sx ", " #lx "				\n\t" /* src[0] */\

  1300. 		"movq " #sx ", " #t0 "				\n\t" /* src[0] */\

  1301. 		"psllq $8, " #lx "				\n\t"\

  1302. 		"psrlq $8, " #t0 "				\n\t"\

  1303. 		"movd -4" #src ", " #t1 "			\n\t"\

  1304. 		"psrlq $24, " #t1 "				\n\t"\

  1305. 		"por " #t1 ", " #lx "				\n\t" /* src[-1] */\

  1306. 		"movd 8" #src ", " #t1 "			\n\t"\

  1307. 		"psllq $56, " #t1 "				\n\t"\

  1308. 		"por " #t1 ", " #t0 "				\n\t" /* src[+1] */\

  1309. 		"movq " #lx ", " #t1 "				\n\t" /* src[-1] */\

  1310. 		PAVGB(t0, lx)				              /* (src[-1] + src[+1])/2 */\

  1311. 		PAVGB(sx, lx)				      /* (src[-1] + 2src[0] + src[+1])/4 */\

  1312. 		PAVGB(lx, pplx)					     \

  1313. 		"movq " #lx ", 8(%%ecx)				\n\t"\

  1314. 		"movq (%%ecx), " #lx "				\n\t"\

  1315. 		"psubusb " #lx ", " #t1 "			\n\t"\

  1316. 		"psubusb " #lx ", " #t0 "			\n\t"\

  1317. 		"psubusb " #lx ", " #sx "			\n\t"\

  1318. 		"movq "MANGLE(b00)", " #lx "			\n\t"\

  1319. 		"pcmpeqb " #lx ", " #t1 "			\n\t" /* src[-1] > a ? 0 : -1*/\

  1320. 		"pcmpeqb " #lx ", " #t0 "			\n\t" /* src[+1] > a ? 0 : -1*/\

  1321. 		"pcmpeqb " #lx ", " #sx "			\n\t" /* src[0]  > a ? 0 : -1*/\

  1322. 		"paddb " #t1 ", " #t0 "				\n\t"\

  1323. 		"paddb " #t0 ", " #sx "				\n\t"\

  1324. \

  1325. 		PAVGB(plx, pplx)				      /* filtered */\

  1326. 		"movq " #dst ", " #t0 "				\n\t" /* dst */\

  1327. 		"movq " #t0 ", " #t1 "				\n\t" /* dst */\

  1328. 		"psubusb %3, " #t0 "				\n\t"\

  1329. 		"paddusb %3, " #t1 "				\n\t"\

  1330. 		PMAXUB(t0, pplx)\

  1331. 		PMINUB(t1, pplx, t0)\

  1332. 		"paddb " #sx ", " #ppsx "			\n\t"\

  1333. 		"paddb " #psx ", " #ppsx "			\n\t"\

  1334. 		"#paddb "MANGLE(b02)", " #ppsx "		\n\t"\

  1335. 		"pand "MANGLE(b08)", " #ppsx "			\n\t"\

  1336. 		"pcmpeqb " #lx ", " #ppsx "			\n\t"\

  1337. 		"pand " #ppsx ", " #pplx "			\n\t"\

  1338. 		"pandn " #dst ", " #ppsx "			\n\t"\

  1339. 		"por " #pplx ", " #ppsx "			\n\t"\

  1340. 		"movq " #ppsx ", " #dst "			\n\t"\

  1341. 		"movq 8(%%ecx), " #lx "				\n\t"

  1342. 

  1343. /*

  1344. 0000000

  1345. 1111111

  1346. 

  1347. 1111110

  1348. 1111101

  1349. 1111100

  1350. 1111011

  1351. 1111010

  1352. 1111001

  1353. 

  1354. 1111000

  1355. 1110111

  1356. 

  1357. */

  1358. //DERING_CORE(dst,src                  ,ppsx ,psx  ,sx   ,pplx ,plx  ,lx   ,t0   ,t1)

  1359. DERING_CORE((%%eax),(%%eax, %1)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)

  1360. DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)

  1361. DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)

  1362. DERING_CORE((%0, %1, 4),(%%edx)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)

  1363. DERING_CORE((%%edx),(%%edx, %1)        ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)

  1364. DERING_CORE((%%edx, %1), (%%edx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)

  1365. DERING_CORE((%%edx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)

  1366. DERING_CORE((%0, %1, 8),(%%edx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)

  1367. 

  1368. 		"1:			\n\t"

  1369. 		: : "r" (src), "r" (stride), "m" (c->pQPb), "m"(c->pQPb2)

  1370. 		: "%eax", "%edx", "%ecx"

  1371. 	);

  1372. #else

  1373. 	int y;

  1374. 	int min=255;

  1375. 	int max=0;

  1376. 	int avg;

  1377. 	uint8_t *p;

  1378. 	int s[10];

  1379. 	const int QP2= c->QP/2 + 1;

  1380. 

  1381. 	for(y=1; y<9; y++)

  1382. 	{

  1383. 		int x;

  1384. 		p= src + stride*y;

  1385. 		for(x=1; x<9; x++)

  1386. 		{

  1387. 			p++;

  1388. 			if(*p > max) max= *p;

  1389. 			if(*p < min) min= *p;

  1390. 		}

  1391. 	}

  1392. 	avg= (min + max + 1)>>1;

  1393. 

  1394. 	if(max - min <deringThreshold) return;

  1395. 

  1396. 	for(y=0; y<10; y++)

  1397. 	{

  1398. 		int t = 0;

  1399. 

  1400. 		if(src[stride*y + 0] > avg) t+= 1;

  1401. 		if(src[stride*y + 1] > avg) t+= 2;

  1402. 		if(src[stride*y + 2] > avg) t+= 4;

  1403. 		if(src[stride*y + 3] > avg) t+= 8;

  1404. 		if(src[stride*y + 4] > avg) t+= 16;

  1405. 		if(src[stride*y + 5] > avg) t+= 32;

  1406. 		if(src[stride*y + 6] > avg) t+= 64;

  1407. 		if(src[stride*y + 7] > avg) t+= 128;

  1408. 		if(src[stride*y + 8] > avg) t+= 256;

  1409. 		if(src[stride*y + 9] > avg) t+= 512;

  1410. 		

  1411. 		t |= (~t)<<16;

  1412. 		t &= (t<<1) & (t>>1);

  1413. 		s[y] = t;

  1414. 	}

  1415. 	

  1416. 	for(y=1; y<9; y++)

  1417. 	{

  1418. 		int t = s[y-1] & s[y] & s[y+1];

  1419. 		t|= t>>16;

  1420. 		s[y-1]= t;

  1421. 	}

  1422. 

  1423. 	for(y=1; y<9; y++)

  1424. 	{

  1425. 		int x;

  1426. 		int t = s[y-1];

  1427. 

  1428. 		p= src + stride*y;

  1429. 		for(x=1; x<9; x++)

  1430. 		{

  1431. 			p++;

  1432. 			if(t & (1<<x))

  1433. 			{

  1434. 				int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))

  1435. 				      +2*(*(p     -1)) + 4*(*p         ) + 2*(*(p     +1))

  1436. 				      +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));

  1437. 				f= (f + 8)>>4;

  1438. 

  1439. #ifdef DEBUG_DERING_THRESHOLD

  1440. 				asm volatile("emms\n\t":);

  1441. 				{

  1442. 				static long long numPixels=0;

  1443. 				if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;

  1444. //				if((max-min)<20 || (max-min)*QP<200)

  1445. //				if((max-min)*QP < 500)

  1446. //				if(max-min<QP/2)

  1447. 				if(max-min < 20)

  1448. 				{

  1449. 					static int numSkiped=0;

  1450. 					static int errorSum=0;

  1451. 					static int worstQP=0;

  1452. 					static int worstRange=0;

  1453. 					static int worstDiff=0;

  1454. 					int diff= (f - *p);

  1455. 					int absDiff= ABS(diff);

  1456. 					int error= diff*diff;

  1457. 

  1458. 					if(x==1 || x==8 || y==1 || y==8) continue;

  1459. 

  1460. 					numSkiped++;

  1461. 					if(absDiff > worstDiff)

  1462. 					{

  1463. 						worstDiff= absDiff;

  1464. 						worstQP= QP;

  1465. 						worstRange= max-min;

  1466. 					}

  1467. 					errorSum+= error;

  1468. 

  1469. 					if(1024LL*1024LL*1024LL % numSkiped == 0)

  1470. 					{

  1471. 						printf( "sum:%1.3f, skip:%d, wQP:%d, "

  1472. 							"wRange:%d, wDiff:%d, relSkip:%1.3f\n",

  1473. 							(float)errorSum/numSkiped, numSkiped, worstQP, worstRange,

  1474. 							worstDiff, (float)numSkiped/numPixels);

  1475. 					}

  1476. 				}

  1477. 				}

  1478. #endif

  1479. 				if     (*p + QP2 < f) *p= *p + QP2;

  1480. 				else if(*p - QP2 > f) *p= *p - QP2;

  1481. 				else *p=f;

  1482. 			}

  1483. 		}

  1484. 	}

  1485. #ifdef DEBUG_DERING_THRESHOLD

  1486. 	if(max-min < 20)

  1487. 	{

  1488. 		for(y=1; y<9; y++)

  1489. 		{

  1490. 			int x;

  1491. 			int t = 0;

  1492. 			p= src + stride*y;

  1493. 			for(x=1; x<9; x++)

  1494. 			{

  1495. 				p++;

  1496. 				*p = MIN(*p + 20, 255);

  1497. 			}

  1498. 		}

  1499. //		src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;

  1500. 	}

  1501. #endif

  1502. #endif

  1503. }

  1504. 

  1505. /**

  1506.  * Deinterlaces the given block by linearly interpolating every second line.

  1507.  * will be called for every 8x8 block and can read & write from line 4-15

  1508.  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

  1509.  * lines 4-12 will be read into the deblocking filter and should be deinterlaced

  1510.  */

  1511. static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)

  1512. {

  1513. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  1514. 	src+= 4*stride;

  1515. 	asm volatile(

  1516. 		"leal (%0, %1), %%eax				\n\t"

  1517. 		"leal (%%eax, %1, 4), %%ecx			\n\t"

  1518. //	0	1	2	3	4	5	6	7	8	9

  1519. //	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1	%0+8%1	ecx+4%1

  1520. 

  1521. 		"movq (%0), %%mm0				\n\t"

  1522. 		"movq (%%eax, %1), %%mm1			\n\t"

  1523. 		PAVGB(%%mm1, %%mm0)

  1524. 		"movq %%mm0, (%%eax)				\n\t"

  1525. 		"movq (%0, %1, 4), %%mm0			\n\t"

  1526. 		PAVGB(%%mm0, %%mm1)

  1527. 		"movq %%mm1, (%%eax, %1, 2)			\n\t"

  1528. 		"movq (%%ecx, %1), %%mm1			\n\t"

  1529. 		PAVGB(%%mm1, %%mm0)

  1530. 		"movq %%mm0, (%%ecx)				\n\t"

  1531. 		"movq (%0, %1, 8), %%mm0			\n\t"

  1532. 		PAVGB(%%mm0, %%mm1)

  1533. 		"movq %%mm1, (%%ecx, %1, 2)			\n\t"

  1534. 

  1535. 		: : "r" (src), "r" (stride)

  1536. 		: "%eax", "%ecx"

  1537. 	);

  1538. #else

  1539. 	int a, b, x;

  1540. 	src+= 4*stride;

  1541. 

  1542. 	for(x=0; x<2; x++){

  1543. 		a= *(uint32_t*)&src[stride*0];

  1544. 		b= *(uint32_t*)&src[stride*2];

  1545. 		*(uint32_t*)&src[stride*1]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1546. 		a= *(uint32_t*)&src[stride*4];

  1547. 		*(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1548. 		b= *(uint32_t*)&src[stride*6];

  1549. 		*(uint32_t*)&src[stride*5]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1550. 		a= *(uint32_t*)&src[stride*8];

  1551. 		*(uint32_t*)&src[stride*7]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1552. 		src += 4;

  1553. 	}

  1554. #endif

  1555. }

  1556. 

  1557. /**

  1558.  * Deinterlaces the given block by cubic interpolating every second line.

  1559.  * will be called for every 8x8 block and can read & write from line 4-15

  1560.  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

  1561.  * lines 4-12 will be read into the deblocking filter and should be deinterlaced

  1562.  * this filter will read lines 3-15 and write 7-13

  1563.  */

  1564. static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)

  1565. {

  1566. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  1567. 	src+= stride*3;

  1568. 	asm volatile(

  1569. 		"leal (%0, %1), %%eax				\n\t"

  1570. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1571. 		"leal (%%edx, %1, 4), %%ecx			\n\t"

  1572. 		"addl %1, %%ecx					\n\t"

  1573. 		"pxor %%mm7, %%mm7				\n\t"

  1574. //	0	1	2	3	4	5	6	7	8	9	10

  1575. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1 ecx

  1576. 

  1577. #define DEINT_CUBIC(a,b,c,d,e)\

  1578. 		"movq " #a ", %%mm0				\n\t"\

  1579. 		"movq " #b ", %%mm1				\n\t"\

  1580. 		"movq " #d ", %%mm2				\n\t"\

  1581. 		"movq " #e ", %%mm3				\n\t"\

  1582. 		PAVGB(%%mm2, %%mm1)					/* (b+d) /2 */\

  1583. 		PAVGB(%%mm3, %%mm0)					/* a(a+e) /2 */\

  1584. 		"movq %%mm0, %%mm2				\n\t"\

  1585. 		"punpcklbw %%mm7, %%mm0				\n\t"\

  1586. 		"punpckhbw %%mm7, %%mm2				\n\t"\

  1587. 		"movq %%mm1, %%mm3				\n\t"\

  1588. 		"punpcklbw %%mm7, %%mm1				\n\t"\

  1589. 		"punpckhbw %%mm7, %%mm3				\n\t"\

  1590. 		"psubw %%mm1, %%mm0				\n\t"	/* L(a+e - (b+d))/2 */\

  1591. 		"psubw %%mm3, %%mm2				\n\t"	/* H(a+e - (b+d))/2 */\

  1592. 		"psraw $3, %%mm0				\n\t"	/* L(a+e - (b+d))/16 */\

  1593. 		"psraw $3, %%mm2				\n\t"	/* H(a+e - (b+d))/16 */\

  1594. 		"psubw %%mm0, %%mm1				\n\t"	/* L(9b + 9d - a - e)/16 */\

  1595. 		"psubw %%mm2, %%mm3				\n\t"	/* H(9b + 9d - a - e)/16 */\

  1596. 		"packuswb %%mm3, %%mm1				\n\t"\

  1597. 		"movq %%mm1, " #c "				\n\t"

  1598. 

  1599. DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx, %1))

  1600. DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%edx), (%%edx, %1), (%0, %1, 8))

  1601. DEINT_CUBIC((%0, %1, 4), (%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%ecx))

  1602. DEINT_CUBIC((%%edx, %1), (%0, %1, 8), (%%edx, %1, 4), (%%ecx), (%%ecx, %1, 2))

  1603. 

  1604. 		: : "r" (src), "r" (stride)

  1605. 		: "%eax", "%edx", "ecx"

  1606. 	);

  1607. #else

  1608. 	int x;

  1609. 	src+= stride*3;

  1610. 	for(x=0; x<8; x++)

  1611. 	{

  1612. 		src[stride*3] = CLIP((-src[0]        + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4);

  1613. 		src[stride*5] = CLIP((-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4);

  1614. 		src[stride*7] = CLIP((-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4);

  1615. 		src[stride*9] = CLIP((-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4);

  1616. 		src++;

  1617. 	}

  1618. #endif

  1619. }

  1620. 

  1621. /**

  1622.  * Deinterlaces the given block by filtering every second line with a (-1 4 2 4 -1) filter.

  1623.  * will be called for every 8x8 block and can read & write from line 4-15

  1624.  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

  1625.  * lines 4-12 will be read into the deblocking filter and should be deinterlaced

  1626.  * this filter will read lines 4-13 and write 5-11

  1627.  */

  1628. static inline void RENAME(deInterlaceFF)(uint8_t src[], int stride, uint8_t *tmp)

  1629. {

  1630. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  1631. 	src+= stride*4;

  1632. 	asm volatile(

  1633. 		"leal (%0, %1), %%eax				\n\t"

  1634. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1635. 		"pxor %%mm7, %%mm7				\n\t"

  1636. 		"movq (%2), %%mm0				\n\t"

  1637. //	0	1	2	3	4	5	6	7	8	9	10

  1638. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1 ecx

  1639. 

  1640. #define DEINT_FF(a,b,c,d)\

  1641. 		"movq " #a ", %%mm1				\n\t"\

  1642. 		"movq " #b ", %%mm2				\n\t"\

  1643. 		"movq " #c ", %%mm3				\n\t"\

  1644. 		"movq " #d ", %%mm4				\n\t"\

  1645. 		PAVGB(%%mm3, %%mm1)					\

  1646. 		PAVGB(%%mm4, %%mm0)					\

  1647. 		"movq %%mm0, %%mm3				\n\t"\

  1648. 		"punpcklbw %%mm7, %%mm0				\n\t"\

  1649. 		"punpckhbw %%mm7, %%mm3				\n\t"\

  1650. 		"movq %%mm1, %%mm4				\n\t"\

  1651. 		"punpcklbw %%mm7, %%mm1				\n\t"\

  1652. 		"punpckhbw %%mm7, %%mm4				\n\t"\

  1653. 		"psllw $2, %%mm1				\n\t"\

  1654. 		"psllw $2, %%mm4				\n\t"\

  1655. 		"psubw %%mm0, %%mm1				\n\t"\

  1656. 		"psubw %%mm3, %%mm4				\n\t"\

  1657. 		"movq %%mm2, %%mm5				\n\t"\

  1658. 		"movq %%mm2, %%mm0				\n\t"\

  1659. 		"punpcklbw %%mm7, %%mm2				\n\t"\

  1660. 		"punpckhbw %%mm7, %%mm5				\n\t"\

  1661. 		"paddw %%mm2, %%mm1				\n\t"\

  1662. 		"paddw %%mm5, %%mm4				\n\t"\

  1663. 		"psraw $2, %%mm1				\n\t"\

  1664. 		"psraw $2, %%mm4				\n\t"\

  1665. 		"packuswb %%mm4, %%mm1				\n\t"\

  1666. 		"movq %%mm1, " #b "				\n\t"\

  1667. 

  1668. DEINT_FF((%0)       , (%%eax)       , (%%eax, %1), (%%eax, %1, 2))

  1669. DEINT_FF((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%edx)       )

  1670. DEINT_FF((%0, %1, 4), (%%edx)       , (%%edx, %1), (%%edx, %1, 2))

  1671. DEINT_FF((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8), (%%edx, %1, 4))

  1672. 

  1673. 		"movq %%mm0, (%2)				\n\t"

  1674. 		: : "r" (src), "r" (stride), "r"(tmp)

  1675. 		: "%eax", "%edx"

  1676. 	);

  1677. #else

  1678. 	int x;

  1679. 	src+= stride*4;

  1680. 	for(x=0; x<8; x++)

  1681. 	{

  1682. 		int t1= tmp[x];

  1683. 		int t2= src[stride*1];

  1684. 

  1685. 		src[stride*1]= CLIP((-t1 + 4*src[stride*0] + 2*t2 + 4*src[stride*2] - src[stride*3] + 4)>>3);

  1686. 		t1= src[stride*4];

  1687. 		src[stride*3]= CLIP((-t2 + 4*src[stride*2] + 2*t1 + 4*src[stride*4] - src[stride*5] + 4)>>3);

  1688. 		t2= src[stride*6];

  1689. 		src[stride*5]= CLIP((-t1 + 4*src[stride*4] + 2*t2 + 4*src[stride*6] - src[stride*7] + 4)>>3);

  1690. 		t1= src[stride*8];

  1691. 		src[stride*7]= CLIP((-t2 + 4*src[stride*6] + 2*t1 + 4*src[stride*8] - src[stride*9] + 4)>>3);

  1692. 		tmp[x]= t1;

  1693. 

  1694. 		src++;

  1695. 	}

  1696. #endif

  1697. }

  1698. 

  1699. /**

  1700.  * Deinterlaces the given block by filtering every line with a (-1 2 6 2 -1) filter.

  1701.  * will be called for every 8x8 block and can read & write from line 4-15

  1702.  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

  1703.  * lines 4-12 will be read into the deblocking filter and should be deinterlaced

  1704.  * this filter will read lines 4-13 and write 4-11

  1705.  */

  1706. static inline void RENAME(deInterlaceL5)(uint8_t src[], int stride, uint8_t *tmp, uint8_t *tmp2)

  1707. {

  1708. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  1709. 	src+= stride*4;

  1710. 	asm volatile(

  1711. 		"leal (%0, %1), %%eax				\n\t"

  1712. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1713. 		"pxor %%mm7, %%mm7				\n\t"

  1714. 		"movq (%2), %%mm0				\n\t"

  1715. 		"movq (%3), %%mm1				\n\t"

  1716. //	0	1	2	3	4	5	6	7	8	9	10

  1717. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1 ecx

  1718. 

  1719. #define DEINT_L5(t1,t2,a,b,c)\

  1720. 		"movq " #a ", %%mm2				\n\t"\

  1721. 		"movq " #b ", %%mm3				\n\t"\

  1722. 		"movq " #c ", %%mm4				\n\t"\

  1723. 		PAVGB(t2, %%mm3)					\

  1724. 		PAVGB(t1, %%mm4)					\

  1725. 		"movq %%mm2, %%mm5				\n\t"\

  1726. 		"movq %%mm2, " #t1 "				\n\t"\

  1727. 		"punpcklbw %%mm7, %%mm2				\n\t"\

  1728. 		"punpckhbw %%mm7, %%mm5				\n\t"\

  1729. 		"movq %%mm2, %%mm6				\n\t"\

  1730. 		"paddw %%mm2, %%mm2				\n\t"\

  1731. 		"paddw %%mm6, %%mm2				\n\t"\

  1732. 		"movq %%mm5, %%mm6				\n\t"\

  1733. 		"paddw %%mm5, %%mm5				\n\t"\

  1734. 		"paddw %%mm6, %%mm5				\n\t"\

  1735. 		"movq %%mm3, %%mm6				\n\t"\

  1736. 		"punpcklbw %%mm7, %%mm3				\n\t"\

  1737. 		"punpckhbw %%mm7, %%mm6				\n\t"\

  1738. 		"paddw %%mm3, %%mm3				\n\t"\

  1739. 		"paddw %%mm6, %%mm6				\n\t"\

  1740. 		"paddw %%mm3, %%mm2				\n\t"\

  1741. 		"paddw %%mm6, %%mm5				\n\t"\

  1742. 		"movq %%mm4, %%mm6				\n\t"\

  1743. 		"punpcklbw %%mm7, %%mm4				\n\t"\

  1744. 		"punpckhbw %%mm7, %%mm6				\n\t"\

  1745. 		"psubw %%mm4, %%mm2				\n\t"\

  1746. 		"psubw %%mm6, %%mm5				\n\t"\

  1747. 		"psraw $2, %%mm2				\n\t"\

  1748. 		"psraw $2, %%mm5				\n\t"\

  1749. 		"packuswb %%mm5, %%mm2				\n\t"\

  1750. 		"movq %%mm2, " #a "				\n\t"\

  1751. 

  1752. DEINT_L5(%%mm0, %%mm1, (%0)          , (%%eax)       , (%%eax, %1)   )

  1753. DEINT_L5(%%mm1, %%mm0, (%%eax)       , (%%eax, %1)   , (%%eax, %1, 2))

  1754. DEINT_L5(%%mm0, %%mm1, (%%eax, %1)   , (%%eax, %1, 2), (%0, %1, 4)   )

  1755. DEINT_L5(%%mm1, %%mm0, (%%eax, %1, 2), (%0, %1, 4)   , (%%edx)       )

  1756. DEINT_L5(%%mm0, %%mm1, (%0, %1, 4)   , (%%edx)       , (%%edx, %1)   )  

  1757. DEINT_L5(%%mm1, %%mm0, (%%edx)       , (%%edx, %1)   , (%%edx, %1, 2))

  1758. DEINT_L5(%%mm0, %%mm1, (%%edx, %1)   , (%%edx, %1, 2), (%0, %1, 8)   )

  1759. DEINT_L5(%%mm1, %%mm0, (%%edx, %1, 2), (%0, %1, 8)   , (%%edx, %1, 4))

  1760. 

  1761. 		"movq %%mm0, (%2)				\n\t"

  1762. 		"movq %%mm1, (%3)				\n\t"

  1763. 		: : "r" (src), "r" (stride), "r"(tmp), "r"(tmp2)

  1764. 		: "%eax", "%edx"

  1765. 	);

  1766. #else

  1767. 	int x;

  1768. 	src+= stride*4;

  1769. 	for(x=0; x<8; x++)

  1770. 	{

  1771. 		int t1= tmp[x];

  1772. 		int t2= tmp2[x];

  1773. 		int t3= src[0];

  1774. 

  1775. 		src[stride*0]= CLIP((-(t1 + src[stride*2]) + 2*(t2 + src[stride*1]) + 6*t3 + 4)>>3);

  1776. 		t1= src[stride*1];

  1777. 		src[stride*1]= CLIP((-(t2 + src[stride*3]) + 2*(t3 + src[stride*2]) + 6*t1 + 4)>>3);

  1778. 		t2= src[stride*2];

  1779. 		src[stride*2]= CLIP((-(t3 + src[stride*4]) + 2*(t1 + src[stride*3]) + 6*t2 + 4)>>3);

  1780. 		t3= src[stride*3];

  1781. 		src[stride*3]= CLIP((-(t1 + src[stride*5]) + 2*(t2 + src[stride*4]) + 6*t3 + 4)>>3);

  1782. 		t1= src[stride*4];

  1783. 		src[stride*4]= CLIP((-(t2 + src[stride*6]) + 2*(t3 + src[stride*5]) + 6*t1 + 4)>>3);

  1784. 		t2= src[stride*5];

  1785. 		src[stride*5]= CLIP((-(t3 + src[stride*7]) + 2*(t1 + src[stride*6]) + 6*t2 + 4)>>3);

  1786. 		t3= src[stride*6];

  1787. 		src[stride*6]= CLIP((-(t1 + src[stride*8]) + 2*(t2 + src[stride*7]) + 6*t3 + 4)>>3);

  1788. 		t1= src[stride*7];

  1789. 		src[stride*7]= CLIP((-(t2 + src[stride*9]) + 2*(t3 + src[stride*8]) + 6*t1 + 4)>>3);

  1790. 

  1791. 		tmp[x]= t3;

  1792. 		tmp2[x]= t1;

  1793. 

  1794. 		src++;

  1795. 	}

  1796. #endif

  1797. }

  1798. 

  1799. /**

  1800.  * Deinterlaces the given block by filtering all lines with a (1 2 1) filter.

  1801.  * will be called for every 8x8 block and can read & write from line 4-15

  1802.  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

  1803.  * lines 4-12 will be read into the deblocking filter and should be deinterlaced

  1804.  * will shift the image up by 1 line (FIXME if this is a problem)

  1805.  * this filter will read lines 4-13 and write 4-11

  1806.  */

  1807. static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride)

  1808. {

  1809. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  1810. 	src+= 4*stride;

  1811. 	asm volatile(

  1812. 		"leal (%0, %1), %%eax				\n\t"

  1813. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1814. //	0	1	2	3	4	5	6	7	8	9

  1815. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

  1816. 

  1817. 		"movq (%0), %%mm0				\n\t" // L0

  1818. 		"movq (%%eax, %1), %%mm1			\n\t" // L2

  1819. 		PAVGB(%%mm1, %%mm0)				      // L0+L2

  1820. 		"movq (%%eax), %%mm2				\n\t" // L1

  1821. 		PAVGB(%%mm2, %%mm0)

  1822. 		"movq %%mm0, (%0)				\n\t"

  1823. 		"movq (%%eax, %1, 2), %%mm0			\n\t" // L3

  1824. 		PAVGB(%%mm0, %%mm2)				      // L1+L3

  1825. 		PAVGB(%%mm1, %%mm2)				      // 2L2 + L1 + L3

  1826. 		"movq %%mm2, (%%eax)				\n\t"

  1827. 		"movq (%0, %1, 4), %%mm2			\n\t" // L4

  1828. 		PAVGB(%%mm2, %%mm1)				      // L2+L4

  1829. 		PAVGB(%%mm0, %%mm1)				      // 2L3 + L2 + L4

  1830. 		"movq %%mm1, (%%eax, %1)			\n\t"

  1831. 		"movq (%%edx), %%mm1				\n\t" // L5

  1832. 		PAVGB(%%mm1, %%mm0)				      // L3+L5

  1833. 		PAVGB(%%mm2, %%mm0)				      // 2L4 + L3 + L5

  1834. 		"movq %%mm0, (%%eax, %1, 2)			\n\t"

  1835. 		"movq (%%edx, %1), %%mm0			\n\t" // L6

  1836. 		PAVGB(%%mm0, %%mm2)				      // L4+L6

  1837. 		PAVGB(%%mm1, %%mm2)				      // 2L5 + L4 + L6

  1838. 		"movq %%mm2, (%0, %1, 4)			\n\t"

  1839. 		"movq (%%edx, %1, 2), %%mm2			\n\t" // L7

  1840. 		PAVGB(%%mm2, %%mm1)				      // L5+L7

  1841. 		PAVGB(%%mm0, %%mm1)				      // 2L6 + L5 + L7

  1842. 		"movq %%mm1, (%%edx)				\n\t"

  1843. 		"movq (%0, %1, 8), %%mm1			\n\t" // L8

  1844. 		PAVGB(%%mm1, %%mm0)				      // L6+L8

  1845. 		PAVGB(%%mm2, %%mm0)				      // 2L7 + L6 + L8

  1846. 		"movq %%mm0, (%%edx, %1)			\n\t"

  1847. 		"movq (%%edx, %1, 4), %%mm0			\n\t" // L9

  1848. 		PAVGB(%%mm0, %%mm2)				      // L7+L9

  1849. 		PAVGB(%%mm1, %%mm2)				      // 2L8 + L7 + L9

  1850. 		"movq %%mm2, (%%edx, %1, 2)			\n\t"

  1851. 

  1852. 

  1853. 		: : "r" (src), "r" (stride)

  1854. 		: "%eax", "%edx"

  1855. 	);

  1856. #else

  1857. 	int a, b, c, x;

  1858. 	src+= 4*stride;

  1859. 

  1860. 	for(x=0; x<2; x++){

  1861. 		a= *(uint32_t*)&src[stride*0];

  1862. 		b= *(uint32_t*)&src[stride*1];

  1863. 		c= *(uint32_t*)&src[stride*2];

  1864. 		a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);

  1865. 		*(uint32_t*)&src[stride*0]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1866. 

  1867. 		a= *(uint32_t*)&src[stride*3];

  1868. 		b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);

  1869. 		*(uint32_t*)&src[stride*1]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);

  1870. 

  1871. 		b= *(uint32_t*)&src[stride*4];

  1872. 		c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);

  1873. 		*(uint32_t*)&src[stride*2]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);

  1874. 

  1875. 		c= *(uint32_t*)&src[stride*5];

  1876. 		a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);

  1877. 		*(uint32_t*)&src[stride*3]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1878. 

  1879. 		a= *(uint32_t*)&src[stride*6];

  1880. 		b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);

  1881. 		*(uint32_t*)&src[stride*4]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);

  1882. 

  1883. 		b= *(uint32_t*)&src[stride*7];

  1884. 		c= (b&c) + (((b^c)&0xFEFEFEFEUL)>>1);

  1885. 		*(uint32_t*)&src[stride*5]= (c|a) - (((c^a)&0xFEFEFEFEUL)>>1);

  1886. 

  1887. 		c= *(uint32_t*)&src[stride*8];

  1888. 		a= (a&c) + (((a^c)&0xFEFEFEFEUL)>>1);

  1889. 		*(uint32_t*)&src[stride*6]= (a|b) - (((a^b)&0xFEFEFEFEUL)>>1);

  1890. 

  1891. 		a= *(uint32_t*)&src[stride*9];

  1892. 		b= (a&b) + (((a^b)&0xFEFEFEFEUL)>>1);

  1893. 		*(uint32_t*)&src[stride*7]= (c|b) - (((c^b)&0xFEFEFEFEUL)>>1);

  1894. 

  1895. 		src += 4;

  1896. 	}

  1897. #endif

  1898. }

  1899. 

  1900. /**

  1901.  * Deinterlaces the given block by applying a median filter to every second line.

  1902.  * will be called for every 8x8 block and can read & write from line 4-15,

  1903.  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

  1904.  * lines 4-12 will be read into the deblocking filter and should be deinterlaced

  1905.  */

  1906. static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)

  1907. {

  1908. #ifdef HAVE_MMX

  1909. 	src+= 4*stride;

  1910. #ifdef HAVE_MMX2

  1911. 	asm volatile(

  1912. 		"leal (%0, %1), %%eax				\n\t"

  1913. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1914. //	0	1	2	3	4	5	6	7	8	9

  1915. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

  1916. 

  1917. 		"movq (%0), %%mm0				\n\t" //

  1918. 		"movq (%%eax, %1), %%mm2			\n\t" //

  1919. 		"movq (%%eax), %%mm1				\n\t" //

  1920. 		"movq %%mm0, %%mm3				\n\t"

  1921. 		"pmaxub %%mm1, %%mm0				\n\t" //

  1922. 		"pminub %%mm3, %%mm1				\n\t" //

  1923. 		"pmaxub %%mm2, %%mm1				\n\t" //

  1924. 		"pminub %%mm1, %%mm0				\n\t"

  1925. 		"movq %%mm0, (%%eax)				\n\t"

  1926. 

  1927. 		"movq (%0, %1, 4), %%mm0			\n\t" //

  1928. 		"movq (%%eax, %1, 2), %%mm1			\n\t" //

  1929. 		"movq %%mm2, %%mm3				\n\t"

  1930. 		"pmaxub %%mm1, %%mm2				\n\t" //

  1931. 		"pminub %%mm3, %%mm1				\n\t" //

  1932. 		"pmaxub %%mm0, %%mm1				\n\t" //

  1933. 		"pminub %%mm1, %%mm2				\n\t"

  1934. 		"movq %%mm2, (%%eax, %1, 2)			\n\t"

  1935. 

  1936. 		"movq (%%edx), %%mm2				\n\t" //

  1937. 		"movq (%%edx, %1), %%mm1			\n\t" //

  1938. 		"movq %%mm2, %%mm3				\n\t"

  1939. 		"pmaxub %%mm0, %%mm2				\n\t" //

  1940. 		"pminub %%mm3, %%mm0				\n\t" //

  1941. 		"pmaxub %%mm1, %%mm0				\n\t" //

  1942. 		"pminub %%mm0, %%mm2				\n\t"

  1943. 		"movq %%mm2, (%%edx)				\n\t"

  1944. 

  1945. 		"movq (%%edx, %1, 2), %%mm2			\n\t" //

  1946. 		"movq (%0, %1, 8), %%mm0			\n\t" //

  1947. 		"movq %%mm2, %%mm3				\n\t"

  1948. 		"pmaxub %%mm0, %%mm2				\n\t" //

  1949. 		"pminub %%mm3, %%mm0				\n\t" //

  1950. 		"pmaxub %%mm1, %%mm0				\n\t" //

  1951. 		"pminub %%mm0, %%mm2				\n\t"

  1952. 		"movq %%mm2, (%%edx, %1, 2)			\n\t"

  1953. 

  1954. 

  1955. 		: : "r" (src), "r" (stride)

  1956. 		: "%eax", "%edx"

  1957. 	);

  1958. 

  1959. #else // MMX without MMX2

  1960. 	asm volatile(

  1961. 		"leal (%0, %1), %%eax				\n\t"

  1962. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  1963. //	0	1	2	3	4	5	6	7	8	9

  1964. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

  1965. 		"pxor %%mm7, %%mm7				\n\t"

  1966. 

  1967. #define MEDIAN(a,b,c)\

  1968. 		"movq " #a ", %%mm0				\n\t"\

  1969. 		"movq " #b ", %%mm2				\n\t"\

  1970. 		"movq " #c ", %%mm1				\n\t"\

  1971. 		"movq %%mm0, %%mm3				\n\t"\

  1972. 		"movq %%mm1, %%mm4				\n\t"\

  1973. 		"movq %%mm2, %%mm5				\n\t"\

  1974. 		"psubusb %%mm1, %%mm3				\n\t"\

  1975. 		"psubusb %%mm2, %%mm4				\n\t"\

  1976. 		"psubusb %%mm0, %%mm5				\n\t"\

  1977. 		"pcmpeqb %%mm7, %%mm3				\n\t"\

  1978. 		"pcmpeqb %%mm7, %%mm4				\n\t"\

  1979. 		"pcmpeqb %%mm7, %%mm5				\n\t"\

  1980. 		"movq %%mm3, %%mm6				\n\t"\

  1981. 		"pxor %%mm4, %%mm3				\n\t"\

  1982. 		"pxor %%mm5, %%mm4				\n\t"\

  1983. 		"pxor %%mm6, %%mm5				\n\t"\

  1984. 		"por %%mm3, %%mm1				\n\t"\

  1985. 		"por %%mm4, %%mm2				\n\t"\

  1986. 		"por %%mm5, %%mm0				\n\t"\

  1987. 		"pand %%mm2, %%mm0				\n\t"\

  1988. 		"pand %%mm1, %%mm0				\n\t"\

  1989. 		"movq %%mm0, " #b "				\n\t"

  1990. 

  1991. MEDIAN((%0), (%%eax), (%%eax, %1))

  1992. MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))

  1993. MEDIAN((%0, %1, 4), (%%edx), (%%edx, %1))

  1994. MEDIAN((%%edx, %1), (%%edx, %1, 2), (%0, %1, 8))

  1995. 

  1996. 		: : "r" (src), "r" (stride)

  1997. 		: "%eax", "%edx"

  1998. 	);

  1999. #endif // MMX

  2000. #else

  2001. 	int x, y;

  2002. 	src+= 4*stride;

  2003. 	// FIXME - there should be a way to do a few columns in parallel like w/mmx

  2004. 	for(x=0; x<8; x++)

  2005. 	{

  2006. 		uint8_t *colsrc = src;

  2007. 		for (y=0; y<4; y++)

  2008. 		{

  2009. 			int a, b, c, d, e, f;

  2010. 			a = colsrc[0       ];

  2011. 			b = colsrc[stride  ];

  2012. 			c = colsrc[stride*2];

  2013. 			d = (a-b)>>31;

  2014. 			e = (b-c)>>31;

  2015. 			f = (c-a)>>31;

  2016. 			colsrc[stride  ] = (a|(d^f)) & (b|(d^e)) & (c|(e^f));

  2017. 			colsrc += stride*2;

  2018. 		}

  2019. 		src++;

  2020. 	}

  2021. #endif

  2022. }

  2023. 

  2024. #ifdef HAVE_MMX

  2025. /**

  2026.  * transposes and shift the given 8x8 Block into dst1 and dst2

  2027.  */

  2028. static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)

  2029. {

  2030. 	asm(

  2031. 		"leal (%0, %1), %%eax				\n\t"

  2032. //	0	1	2	3	4	5	6	7	8	9

  2033. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

  2034. 		"movq (%0), %%mm0		\n\t" // 12345678

  2035. 		"movq (%%eax), %%mm1		\n\t" // abcdefgh

  2036. 		"movq %%mm0, %%mm2		\n\t" // 12345678

  2037. 		"punpcklbw %%mm1, %%mm0		\n\t" // 1a2b3c4d

  2038. 		"punpckhbw %%mm1, %%mm2		\n\t" // 5e6f7g8h

  2039. 

  2040. 		"movq (%%eax, %1), %%mm1	\n\t"

  2041. 		"movq (%%eax, %1, 2), %%mm3	\n\t"

  2042. 		"movq %%mm1, %%mm4		\n\t"

  2043. 		"punpcklbw %%mm3, %%mm1		\n\t"

  2044. 		"punpckhbw %%mm3, %%mm4		\n\t"

  2045. 

  2046. 		"movq %%mm0, %%mm3		\n\t"

  2047. 		"punpcklwd %%mm1, %%mm0		\n\t"

  2048. 		"punpckhwd %%mm1, %%mm3		\n\t"

  2049. 		"movq %%mm2, %%mm1		\n\t"

  2050. 		"punpcklwd %%mm4, %%mm2		\n\t"

  2051. 		"punpckhwd %%mm4, %%mm1		\n\t"

  2052. 

  2053. 		"movd %%mm0, 128(%2)		\n\t"

  2054. 		"psrlq $32, %%mm0		\n\t"

  2055. 		"movd %%mm0, 144(%2)		\n\t"

  2056. 		"movd %%mm3, 160(%2)		\n\t"

  2057. 		"psrlq $32, %%mm3		\n\t"

  2058. 		"movd %%mm3, 176(%2)		\n\t"

  2059. 		"movd %%mm3, 48(%3)		\n\t"

  2060. 		"movd %%mm2, 192(%2)		\n\t"

  2061. 		"movd %%mm2, 64(%3)		\n\t"

  2062. 		"psrlq $32, %%mm2		\n\t"

  2063. 		"movd %%mm2, 80(%3)		\n\t"

  2064. 		"movd %%mm1, 96(%3)		\n\t"

  2065. 		"psrlq $32, %%mm1		\n\t"

  2066. 		"movd %%mm1, 112(%3)		\n\t"

  2067. 

  2068. 		"leal (%%eax, %1, 4), %%eax	\n\t"

  2069. 		

  2070. 		"movq (%0, %1, 4), %%mm0	\n\t" // 12345678

  2071. 		"movq (%%eax), %%mm1		\n\t" // abcdefgh

  2072. 		"movq %%mm0, %%mm2		\n\t" // 12345678

  2073. 		"punpcklbw %%mm1, %%mm0		\n\t" // 1a2b3c4d

  2074. 		"punpckhbw %%mm1, %%mm2		\n\t" // 5e6f7g8h

  2075. 

  2076. 		"movq (%%eax, %1), %%mm1	\n\t"

  2077. 		"movq (%%eax, %1, 2), %%mm3	\n\t"

  2078. 		"movq %%mm1, %%mm4		\n\t"

  2079. 		"punpcklbw %%mm3, %%mm1		\n\t"

  2080. 		"punpckhbw %%mm3, %%mm4		\n\t"

  2081. 

  2082. 		"movq %%mm0, %%mm3		\n\t"

  2083. 		"punpcklwd %%mm1, %%mm0		\n\t"

  2084. 		"punpckhwd %%mm1, %%mm3		\n\t"

  2085. 		"movq %%mm2, %%mm1		\n\t"

  2086. 		"punpcklwd %%mm4, %%mm2		\n\t"

  2087. 		"punpckhwd %%mm4, %%mm1		\n\t"

  2088. 

  2089. 		"movd %%mm0, 132(%2)		\n\t"

  2090. 		"psrlq $32, %%mm0		\n\t"

  2091. 		"movd %%mm0, 148(%2)		\n\t"

  2092. 		"movd %%mm3, 164(%2)		\n\t"

  2093. 		"psrlq $32, %%mm3		\n\t"

  2094. 		"movd %%mm3, 180(%2)		\n\t"

  2095. 		"movd %%mm3, 52(%3)		\n\t"

  2096. 		"movd %%mm2, 196(%2)		\n\t"

  2097. 		"movd %%mm2, 68(%3)		\n\t"

  2098. 		"psrlq $32, %%mm2		\n\t"

  2099. 		"movd %%mm2, 84(%3)		\n\t"

  2100. 		"movd %%mm1, 100(%3)		\n\t"

  2101. 		"psrlq $32, %%mm1		\n\t"

  2102. 		"movd %%mm1, 116(%3)		\n\t"

  2103. 

  2104. 

  2105. 	:: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)

  2106. 	: "%eax"

  2107. 	);

  2108. }

  2109. 

  2110. /**

  2111.  * transposes the given 8x8 block

  2112.  */

  2113. static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)

  2114. {

  2115. 	asm(

  2116. 		"leal (%0, %1), %%eax				\n\t"

  2117. 		"leal (%%eax, %1, 4), %%edx			\n\t"

  2118. //	0	1	2	3	4	5	6	7	8	9

  2119. //	%0	eax	eax+%1	eax+2%1	%0+4%1	edx	edx+%1	edx+2%1	%0+8%1	edx+4%1

  2120. 		"movq (%2), %%mm0		\n\t" // 12345678

  2121. 		"movq 16(%2), %%mm1		\n\t" // abcdefgh

  2122. 		"movq %%mm0, %%mm2		\n\t" // 12345678

  2123. 		"punpcklbw %%mm1, %%mm0		\n\t" // 1a2b3c4d

  2124. 		"punpckhbw %%mm1, %%mm2		\n\t" // 5e6f7g8h

  2125. 

  2126. 		"movq 32(%2), %%mm1		\n\t"

  2127. 		"movq 48(%2), %%mm3		\n\t"

  2128. 		"movq %%mm1, %%mm4		\n\t"

  2129. 		"punpcklbw %%mm3, %%mm1		\n\t"

  2130. 		"punpckhbw %%mm3, %%mm4		\n\t"

  2131. 

  2132. 		"movq %%mm0, %%mm3		\n\t"

  2133. 		"punpcklwd %%mm1, %%mm0		\n\t"

  2134. 		"punpckhwd %%mm1, %%mm3		\n\t"

  2135. 		"movq %%mm2, %%mm1		\n\t"

  2136. 		"punpcklwd %%mm4, %%mm2		\n\t"

  2137. 		"punpckhwd %%mm4, %%mm1		\n\t"

  2138. 

  2139. 		"movd %%mm0, (%0)		\n\t"

  2140. 		"psrlq $32, %%mm0		\n\t"

  2141. 		"movd %%mm0, (%%eax)		\n\t"

  2142. 		"movd %%mm3, (%%eax, %1)	\n\t"

  2143. 		"psrlq $32, %%mm3		\n\t"

  2144. 		"movd %%mm3, (%%eax, %1, 2)	\n\t"

  2145. 		"movd %%mm2, (%0, %1, 4)	\n\t"

  2146. 		"psrlq $32, %%mm2		\n\t"

  2147. 		"movd %%mm2, (%%edx)		\n\t"

  2148. 		"movd %%mm1, (%%edx, %1)	\n\t"

  2149. 		"psrlq $32, %%mm1		\n\t"

  2150. 		"movd %%mm1, (%%edx, %1, 2)	\n\t"

  2151. 

  2152. 

  2153. 		"movq 64(%2), %%mm0		\n\t" // 12345678

  2154. 		"movq 80(%2), %%mm1		\n\t" // abcdefgh

  2155. 		"movq %%mm0, %%mm2		\n\t" // 12345678

  2156. 		"punpcklbw %%mm1, %%mm0		\n\t" // 1a2b3c4d

  2157. 		"punpckhbw %%mm1, %%mm2		\n\t" // 5e6f7g8h

  2158. 

  2159. 		"movq 96(%2), %%mm1		\n\t"

  2160. 		"movq 112(%2), %%mm3		\n\t"

  2161. 		"movq %%mm1, %%mm4		\n\t"

  2162. 		"punpcklbw %%mm3, %%mm1		\n\t"

  2163. 		"punpckhbw %%mm3, %%mm4		\n\t"

  2164. 

  2165. 		"movq %%mm0, %%mm3		\n\t"

  2166. 		"punpcklwd %%mm1, %%mm0		\n\t"

  2167. 		"punpckhwd %%mm1, %%mm3		\n\t"

  2168. 		"movq %%mm2, %%mm1		\n\t"

  2169. 		"punpcklwd %%mm4, %%mm2		\n\t"

  2170. 		"punpckhwd %%mm4, %%mm1		\n\t"

  2171. 

  2172. 		"movd %%mm0, 4(%0)		\n\t"

  2173. 		"psrlq $32, %%mm0		\n\t"

  2174. 		"movd %%mm0, 4(%%eax)		\n\t"

  2175. 		"movd %%mm3, 4(%%eax, %1)	\n\t"

  2176. 		"psrlq $32, %%mm3		\n\t"

  2177. 		"movd %%mm3, 4(%%eax, %1, 2)	\n\t"

  2178. 		"movd %%mm2, 4(%0, %1, 4)	\n\t"

  2179. 		"psrlq $32, %%mm2		\n\t"

  2180. 		"movd %%mm2, 4(%%edx)		\n\t"

  2181. 		"movd %%mm1, 4(%%edx, %1)	\n\t"

  2182. 		"psrlq $32, %%mm1		\n\t"

  2183. 		"movd %%mm1, 4(%%edx, %1, 2)	\n\t"

  2184. 

  2185. 	:: "r" (dst), "r" (dstStride), "r" (src)

  2186. 	: "%eax", "%edx"

  2187. 	);

  2188. }

  2189. #endif

  2190. //static int test=0;

  2191. 

  2192. static inline void RENAME(tempNoiseReducer)(uint8_t *src, int stride,

  2193. 				    uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)

  2194. {

  2195. 	// to save a register (FIXME do this outside of the loops)

  2196. 	tempBluredPast[127]= maxNoise[0];

  2197. 	tempBluredPast[128]= maxNoise[1];

  2198. 	tempBluredPast[129]= maxNoise[2];

  2199.         

  2200. #define FAST_L2_DIFF

  2201. //#define L1_DIFF //u should change the thresholds too if u try that one

  2202. #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)

  2203. 	asm volatile(

  2204. 		"leal (%2, %2, 2), %%eax			\n\t" // 3*stride

  2205. 		"leal (%2, %2, 4), %%edx			\n\t" // 5*stride

  2206. 		"leal (%%edx, %2, 2), %%ecx			\n\t" // 7*stride

  2207. //	0	1	2	3	4	5	6	7	8	9

  2208. //	%x	%x+%2	%x+2%2	%x+eax	%x+4%2	%x+edx	%x+2eax	%x+ecx	%x+8%2

  2209. //FIXME reorder?

  2210. #ifdef L1_DIFF //needs mmx2

  2211. 		"movq (%0), %%mm0				\n\t" // L0

  2212. 		"psadbw (%1), %%mm0				\n\t" // |L0-R0|

  2213. 		"movq (%0, %2), %%mm1				\n\t" // L1

  2214. 		"psadbw (%1, %2), %%mm1				\n\t" // |L1-R1|

  2215. 		"movq (%0, %2, 2), %%mm2			\n\t" // L2

  2216. 		"psadbw (%1, %2, 2), %%mm2			\n\t" // |L2-R2|

  2217. 		"movq (%0, %%eax), %%mm3			\n\t" // L3

  2218. 		"psadbw (%1, %%eax), %%mm3			\n\t" // |L3-R3|

  2219. 

  2220. 		"movq (%0, %2, 4), %%mm4			\n\t" // L4

  2221. 		"paddw %%mm1, %%mm0				\n\t"

  2222. 		"psadbw (%1, %2, 4), %%mm4			\n\t" // |L4-R4|

  2223. 		"movq (%0, %%edx), %%mm5			\n\t" // L5

  2224. 		"paddw %%mm2, %%mm0				\n\t"

  2225. 		"psadbw (%1, %%edx), %%mm5			\n\t" // |L5-R5|

  2226. 		"movq (%0, %%eax, 2), %%mm6			\n\t" // L6

  2227. 		"paddw %%mm3, %%mm0				\n\t"

  2228. 		"psadbw (%1, %%eax, 2), %%mm6			\n\t" // |L6-R6|

  2229. 		"movq (%0, %%ecx), %%mm7			\n\t" // L7

  2230. 		"paddw %%mm4, %%mm0				\n\t"

  2231. 		"psadbw (%1, %%ecx), %%mm7			\n\t" // |L7-R7|

  2232. 		"paddw %%mm5, %%mm6				\n\t"

  2233. 		"paddw %%mm7, %%mm6				\n\t"

  2234. 		"paddw %%mm6, %%mm0				\n\t"

  2235. #elif defined (FAST_L2_DIFF)

  2236. 		"pcmpeqb %%mm7, %%mm7				\n\t"

  2237. 		"movq "MANGLE(b80)", %%mm6			\n\t"

  2238. 		"pxor %%mm0, %%mm0				\n\t"

  2239. #define L2_DIFF_CORE(a, b)\

  2240. 		"movq " #a ", %%mm5				\n\t"\

  2241. 		"movq " #b ", %%mm2				\n\t"\

  2242. 		"pxor %%mm7, %%mm2				\n\t"\

  2243. 		PAVGB(%%mm2, %%mm5)\

  2244. 		"paddb %%mm6, %%mm5				\n\t"\

  2245. 		"movq %%mm5, %%mm2				\n\t"\

  2246. 		"psllw $8, %%mm5				\n\t"\

  2247. 		"pmaddwd %%mm5, %%mm5				\n\t"\

  2248. 		"pmaddwd %%mm2, %%mm2				\n\t"\

  2249. 		"paddd %%mm2, %%mm5				\n\t"\

  2250. 		"psrld $14, %%mm5				\n\t"\

  2251. 		"paddd %%mm5, %%mm0				\n\t"

  2252. 

  2253. L2_DIFF_CORE((%0), (%1))

  2254. L2_DIFF_CORE((%0, %2), (%1, %2))

  2255. L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))

  2256. L2_DIFF_CORE((%0, %%eax), (%1, %%eax))

  2257. L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))

  2258. L2_DIFF_CORE((%0, %%edx), (%1, %%edx))

  2259. L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))

  2260. L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))

  2261. 

  2262. #else

  2263. 		"pxor %%mm7, %%mm7				\n\t"

  2264. 		"pxor %%mm0, %%mm0				\n\t"

  2265. #define L2_DIFF_CORE(a, b)\

  2266. 		"movq " #a ", %%mm5				\n\t"\

  2267. 		"movq " #b ", %%mm2				\n\t"\

  2268. 		"movq %%mm5, %%mm1				\n\t"\

  2269. 		"movq %%mm2, %%mm3				\n\t"\

  2270. 		"punpcklbw %%mm7, %%mm5				\n\t"\

  2271. 		"punpckhbw %%mm7, %%mm1				\n\t"\

  2272. 		"punpcklbw %%mm7, %%mm2				\n\t"\

  2273. 		"punpckhbw %%mm7, %%mm3				\n\t"\

  2274. 		"psubw %%mm2, %%mm5				\n\t"\

  2275. 		"psubw %%mm3, %%mm1				\n\t"\

  2276. 		"pmaddwd %%mm5, %%mm5				\n\t"\

  2277. 		"pmaddwd %%mm1, %%mm1				\n\t"\

  2278. 		"paddd %%mm1, %%mm5				\n\t"\

  2279. 		"paddd %%mm5, %%mm0				\n\t"

  2280. 

  2281. L2_DIFF_CORE((%0), (%1))

  2282. L2_DIFF_CORE((%0, %2), (%1, %2))

  2283. L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))

  2284. L2_DIFF_CORE((%0, %%eax), (%1, %%eax))

  2285. L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))

  2286. L2_DIFF_CORE((%0, %%edx), (%1, %%edx))

  2287. L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))

  2288. L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))

  2289. 

  2290. #endif

  2291. 

  2292. 		"movq %%mm0, %%mm4				\n\t"

  2293. 		"psrlq $32, %%mm0				\n\t"

  2294. 		"paddd %%mm0, %%mm4				\n\t"

  2295. 		"movd %%mm4, %%ecx				\n\t"

  2296. 		"shll $2, %%ecx					\n\t"

  2297. 		"movl %3, %%edx					\n\t"

  2298. 		"addl -4(%%edx), %%ecx				\n\t"

  2299. 		"addl 4(%%edx), %%ecx				\n\t"

  2300. 		"addl -1024(%%edx), %%ecx			\n\t"

  2301. 		"addl $4, %%ecx					\n\t"

  2302. 		"addl 1024(%%edx), %%ecx			\n\t"

  2303. 		"shrl $3, %%ecx					\n\t"

  2304. 		"movl %%ecx, (%%edx)				\n\t"

  2305. 

  2306. //		"movl %3, %%ecx					\n\t"

  2307. //		"movl %%ecx, test				\n\t"

  2308. //		"jmp 4f \n\t"

  2309. 		"cmpl 512(%%edx), %%ecx				\n\t"

  2310. 		" jb 2f						\n\t"

  2311. 		"cmpl 516(%%edx), %%ecx				\n\t"

  2312. 		" jb 1f						\n\t"

  2313. 

  2314. 		"leal (%%eax, %2, 2), %%edx			\n\t" // 5*stride

  2315. 		"leal (%%edx, %2, 2), %%ecx			\n\t" // 7*stride

  2316. 		"movq (%0), %%mm0				\n\t" // L0

  2317. 		"movq (%0, %2), %%mm1				\n\t" // L1

  2318. 		"movq (%0, %2, 2), %%mm2			\n\t" // L2

  2319. 		"movq (%0, %%eax), %%mm3			\n\t" // L3

  2320. 		"movq (%0, %2, 4), %%mm4			\n\t" // L4

  2321. 		"movq (%0, %%edx), %%mm5			\n\t" // L5

  2322. 		"movq (%0, %%eax, 2), %%mm6			\n\t" // L6

  2323. 		"movq (%0, %%ecx), %%mm7			\n\t" // L7

  2324. 		"movq %%mm0, (%1)				\n\t" // L0

  2325. 		"movq %%mm1, (%1, %2)				\n\t" // L1

  2326. 		"movq %%mm2, (%1, %2, 2)			\n\t" // L2

  2327. 		"movq %%mm3, (%1, %%eax)			\n\t" // L3

  2328. 		"movq %%mm4, (%1, %2, 4)			\n\t" // L4

  2329. 		"movq %%mm5, (%1, %%edx)			\n\t" // L5

  2330. 		"movq %%mm6, (%1, %%eax, 2)			\n\t" // L6

  2331. 		"movq %%mm7, (%1, %%ecx)			\n\t" // L7

  2332. 		"jmp 4f						\n\t"

  2333. 

  2334. 		"1:						\n\t"

  2335. 		"leal (%%eax, %2, 2), %%edx			\n\t" // 5*stride

  2336. 		"leal (%%edx, %2, 2), %%ecx			\n\t" // 7*stride

  2337. 		"movq (%0), %%mm0				\n\t" // L0

  2338. 		PAVGB((%1), %%mm0)				      // L0

  2339. 		"movq (%0, %2), %%mm1				\n\t" // L1

  2340. 		PAVGB((%1, %2), %%mm1)				      // L1

  2341. 		"movq (%0, %2, 2), %%mm2			\n\t" // L2

  2342. 		PAVGB((%1, %2, 2), %%mm2)			      // L2

  2343. 		"movq (%0, %%eax), %%mm3			\n\t" // L3

  2344. 		PAVGB((%1, %%eax), %%mm3)			      // L3

  2345. 		"movq (%0, %2, 4), %%mm4			\n\t" // L4

  2346. 		PAVGB((%1, %2, 4), %%mm4)			      // L4

  2347. 		"movq (%0, %%edx), %%mm5			\n\t" // L5

  2348. 		PAVGB((%1, %%edx), %%mm5)			      // L5

  2349. 		"movq (%0, %%eax, 2), %%mm6			\n\t" // L6

  2350. 		PAVGB((%1, %%eax, 2), %%mm6)			      // L6

  2351. 		"movq (%0, %%ecx), %%mm7			\n\t" // L7

  2352. 		PAVGB((%1, %%ecx), %%mm7)			      // L7

  2353. 		"movq %%mm0, (%1)				\n\t" // R0

  2354. 		"movq %%mm1, (%1, %2)				\n\t" // R1

  2355. 		"movq %%mm2, (%1, %2, 2)			\n\t" // R2

  2356. 		"movq %%mm3, (%1, %%eax)			\n\t" // R3

  2357. 		"movq %%mm4, (%1, %2, 4)			\n\t" // R4

  2358. 		"movq %%mm5, (%1, %%edx)			\n\t" // R5

  2359. 		"movq %%mm6, (%1, %%eax, 2)			\n\t" // R6

  2360. 		"movq %%mm7, (%1, %%ecx)			\n\t" // R7

  2361. 		"movq %%mm0, (%0)				\n\t" // L0

  2362. 		"movq %%mm1, (%0, %2)				\n\t" // L1

  2363. 		"movq %%mm2, (%0, %2, 2)			\n\t" // L2

  2364. 		"movq %%mm3, (%0, %%eax)			\n\t" // L3

  2365. 		"movq %%mm4, (%0, %2, 4)			\n\t" // L4

  2366. 		"movq %%mm5, (%0, %%edx)			\n\t" // L5

  2367. 		"movq %%mm6, (%0, %%eax, 2)			\n\t" // L6

  2368. 		"movq %%mm7, (%0, %%ecx)			\n\t" // L7

  2369. 		"jmp 4f						\n\t"

  2370. 

  2371. 		"2:						\n\t"

  2372. 		"cmpl 508(%%edx), %%ecx				\n\t"

  2373. 		" jb 3f						\n\t"

  2374. 

  2375. 		"leal (%%eax, %2, 2), %%edx			\n\t" // 5*stride

  2376. 		"leal (%%edx, %2, 2), %%ecx			\n\t" // 7*stride

  2377. 		"movq (%0), %%mm0				\n\t" // L0

  2378. 		"movq (%0, %2), %%mm1				\n\t" // L1

  2379. 		"movq (%0, %2, 2), %%mm2			\n\t" // L2

  2380. 		"movq (%0, %%eax), %%mm3			\n\t" // L3

  2381. 		"movq (%1), %%mm4				\n\t" // R0

  2382. 		"movq (%1, %2), %%mm5				\n\t" // R1

  2383. 		"movq (%1, %2, 2), %%mm6			\n\t" // R2

  2384. 		"movq (%1, %%eax), %%mm7			\n\t" // R3

  2385. 		PAVGB(%%mm4, %%mm0)

  2386. 		PAVGB(%%mm5, %%mm1)

  2387. 		PAVGB(%%mm6, %%mm2)

  2388. 		PAVGB(%%mm7, %%mm3)

  2389. 		PAVGB(%%mm4, %%mm0)

  2390. 		PAVGB(%%mm5, %%mm1)

  2391. 		PAVGB(%%mm6, %%mm2)

  2392. 		PAVGB(%%mm7, %%mm3)

  2393. 		"movq %%mm0, (%1)				\n\t" // R0

  2394. 		"movq %%mm1, (%1, %2)				\n\t" // R1

  2395. 		"movq %%mm2, (%1, %2, 2)			\n\t" // R2

  2396. 		"movq %%mm3, (%1, %%eax)			\n\t" // R3

  2397. 		"movq %%mm0, (%0)				\n\t" // L0

  2398. 		"movq %%mm1, (%0, %2)				\n\t" // L1

  2399. 		"movq %%mm2, (%0, %2, 2)			\n\t" // L2

  2400. 		"movq %%mm3, (%0, %%eax)			\n\t" // L3

  2401. 

  2402. 		"movq (%0, %2, 4), %%mm0			\n\t" // L4

  2403. 		"movq (%0, %%edx), %%mm1			\n\t" // L5

  2404. 		"movq (%0, %%eax, 2), %%mm2			\n\t" // L6

  2405. 		"movq (%0, %%ecx), %%mm3			\n\t" // L7

  2406. 		"movq (%1, %2, 4), %%mm4			\n\t" // R4

  2407. 		"movq (%1, %%edx), %%mm5			\n\t" // R5

  2408. 		"movq (%1, %%eax, 2), %%mm6			\n\t" // R6

  2409. 		"movq (%1, %%ecx), %%mm7			\n\t" // R7

  2410. 		PAVGB(%%mm4, %%mm0)

  2411. 		PAVGB(%%mm5, %%mm1)

  2412. 		PAVGB(%%mm6, %%mm2)

  2413. 		PAVGB(%%mm7, %%mm3)

  2414. 		PAVGB(%%mm4, %%mm0)

  2415. 		PAVGB(%%mm5, %%mm1)

  2416. 		PAVGB(%%mm6, %%mm2)

  2417. 		PAVGB(%%mm7, %%mm3)

  2418. 		"movq %%mm0, (%1, %2, 4)			\n\t" // R4

  2419. 		"movq %%mm1, (%1, %%edx)			\n\t" // R5

  2420. 		"movq %%mm2, (%1, %%eax, 2)			\n\t" // R6

  2421. 		"movq %%mm3, (%1, %%ecx)			\n\t" // R7

  2422. 		"movq %%mm0, (%0, %2, 4)			\n\t" // L4

  2423. 		"movq %%mm1, (%0, %%edx)			\n\t" // L5

  2424. 		"movq %%mm2, (%0, %%eax, 2)			\n\t" // L6

  2425. 		"movq %%mm3, (%0, %%ecx)			\n\t" // L7

  2426. 		"jmp 4f						\n\t"

  2427. 

  2428. 		"3:						\n\t"

  2429. 		"leal (%%eax, %2, 2), %%edx			\n\t" // 5*stride

  2430. 		"leal (%%edx, %2, 2), %%ecx			\n\t" // 7*stride

  2431. 		"movq (%0), %%mm0				\n\t" // L0

  2432. 		"movq (%0, %2), %%mm1				\n\t" // L1

  2433. 		"movq (%0, %2, 2), %%mm2			\n\t" // L2

  2434. 		"movq (%0, %%eax), %%mm3			\n\t" // L3

  2435. 		"movq (%1), %%mm4				\n\t" // R0

  2436. 		"movq (%1, %2), %%mm5				\n\t" // R1

  2437. 		"movq (%1, %2, 2), %%mm6			\n\t" // R2

  2438. 		"movq (%1, %%eax), %%mm7			\n\t" // R3

  2439. 		PAVGB(%%mm4, %%mm0)

  2440. 		PAVGB(%%mm5, %%mm1)

  2441. 		PAVGB(%%mm6, %%mm2)

  2442. 		PAVGB(%%mm7, %%mm3)

  2443. 		PAVGB(%%mm4, %%mm0)

  2444. 		PAVGB(%%mm5, %%mm1)

  2445. 		PAVGB(%%mm6, %%mm2)

  2446. 		PAVGB(%%mm7, %%mm3)

  2447. 		PAVGB(%%mm4, %%mm0)

  2448. 		PAVGB(%%mm5, %%mm1)

  2449. 		PAVGB(%%mm6, %%mm2)

  2450. 		PAVGB(%%mm7, %%mm3)

  2451. 		"movq %%mm0, (%1)				\n\t" // R0

  2452. 		"movq %%mm1, (%1, %2)				\n\t" // R1

  2453. 		"movq %%mm2, (%1, %2, 2)			\n\t" // R2

  2454. 		"movq %%mm3, (%1, %%eax)			\n\t" // R3

  2455. 		"movq %%mm0, (%0)				\n\t" // L0

  2456. 		"movq %%mm1, (%0, %2)				\n\t" // L1

  2457. 		"movq %%mm2, (%0, %2, 2)			\n\t" // L2

  2458. 		"movq %%mm3, (%0, %%eax)			\n\t" // L3

  2459. 

  2460. 		"movq (%0, %2, 4), %%mm0			\n\t" // L4

  2461. 		"movq (%0, %%edx), %%mm1			\n\t" // L5

  2462. 		"movq (%0, %%eax, 2), %%mm2			\n\t" // L6

  2463. 		"movq (%0, %%ecx), %%mm3			\n\t" // L7

  2464. 		"movq (%1, %2, 4), %%mm4			\n\t" // R4

  2465. 		"movq (%1, %%edx), %%mm5			\n\t" // R5

  2466. 		"movq (%1, %%eax, 2), %%mm6			\n\t" // R6

  2467. 		"movq (%1, %%ecx), %%mm7			\n\t" // R7

  2468. 		PAVGB(%%mm4, %%mm0)

  2469. 		PAVGB(%%mm5, %%mm1)

  2470. 		PAVGB(%%mm6, %%mm2)

  2471. 		PAVGB(%%mm7, %%mm3)

  2472. 		PAVGB(%%mm4, %%mm0)

  2473. 		PAVGB(%%mm5, %%mm1)

  2474. 		PAVGB(%%mm6, %%mm2)

  2475. 		PAVGB(%%mm7, %%mm3)

  2476. 		PAVGB(%%mm4, %%mm0)

  2477. 		PAVGB(%%mm5, %%mm1)

  2478. 		PAVGB(%%mm6, %%mm2)

  2479. 		PAVGB(%%mm7, %%mm3)

  2480. 		"movq %%mm0, (%1, %2, 4)			\n\t" // R4

  2481. 		"movq %%mm1, (%1, %%edx)			\n\t" // R5

  2482. 		"movq %%mm2, (%1, %%eax, 2)			\n\t" // R6

  2483. 		"movq %%mm3, (%1, %%ecx)			\n\t" // R7

  2484. 		"movq %%mm0, (%0, %2, 4)			\n\t" // L4

  2485. 		"movq %%mm1, (%0, %%edx)			\n\t" // L5

  2486. 		"movq %%mm2, (%0, %%eax, 2)			\n\t" // L6

  2487. 		"movq %%mm3, (%0, %%ecx)			\n\t" // L7

  2488. 

  2489. 		"4:						\n\t"

  2490. 

  2491. 		:: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast)

  2492. 		: "%eax", "%edx", "%ecx", "memory"

  2493. 		);

  2494. //printf("%d\n", test);

  2495. #else

  2496. {

  2497. 	int y;

  2498. 	int d=0;

  2499. 	int sysd=0;

  2500. 	int i;

  2501. 

  2502. 	for(y=0; y<8; y++)

  2503. 	{

  2504. 		int x;

  2505. 		for(x=0; x<8; x++)

  2506. 		{

  2507. 			int ref= tempBlured[ x + y*stride ];

  2508. 			int cur= src[ x + y*stride ];

  2509. 			int d1=ref - cur;

  2510. //			if(x==0 || x==7) d1+= d1>>1;

  2511. //			if(y==0 || y==7) d1+= d1>>1;

  2512. //			d+= ABS(d1);

  2513. 			d+= d1*d1;

  2514. 			sysd+= d1;

  2515. 		}

  2516. 	}

  2517. 	i=d;

  2518. 	d= 	(

  2519. 		4*d

  2520. 		+(*(tempBluredPast-256))

  2521. 		+(*(tempBluredPast-1))+ (*(tempBluredPast+1))

  2522. 		+(*(tempBluredPast+256))

  2523. 		+4)>>3;

  2524. 	*tempBluredPast=i;

  2525. //	((*tempBluredPast)*3 + d + 2)>>2;

  2526. 

  2527. //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]);

  2528. /*

  2529. Switch between

  2530.  1  0  0  0  0  0  0  (0)

  2531. 64 32 16  8  4  2  1  (1)

  2532. 64 48 36 27 20 15 11 (33) (approx)

  2533. 64 56 49 43 37 33 29 (200) (approx)

  2534. */

  2535. 	if(d > maxNoise[1])

  2536. 	{

  2537. 		if(d < maxNoise[2])

  2538. 		{

  2539. 			for(y=0; y<8; y++)

  2540. 			{

  2541. 				int x;

  2542. 				for(x=0; x<8; x++)

  2543. 				{

  2544. 					int ref= tempBlured[ x + y*stride ];

  2545. 					int cur= src[ x + y*stride ];

  2546. 					tempBlured[ x + y*stride ]=

  2547. 					src[ x + y*stride ]=

  2548. 						(ref + cur + 1)>>1;

  2549. 				}

  2550. 			}

  2551. 		}

  2552. 		else

  2553. 		{

  2554. 			for(y=0; y<8; y++)

  2555. 			{

  2556. 				int x;

  2557. 				for(x=0; x<8; x++)

  2558. 				{

  2559. 					tempBlured[ x + y*stride ]= src[ x + y*stride ];

  2560. 				}

  2561. 			}

  2562. 		}

  2563. 	}

  2564. 	else

  2565. 	{

  2566. 		if(d < maxNoise[0])

  2567. 		{

  2568. 			for(y=0; y<8; y++)

  2569. 			{

  2570. 				int x;

  2571. 				for(x=0; x<8; x++)

  2572. 				{

  2573. 					int ref= tempBlured[ x + y*stride ];

  2574. 					int cur= src[ x + y*stride ];

  2575. 					tempBlured[ x + y*stride ]=

  2576. 					src[ x + y*stride ]=

  2577. 						(ref*7 + cur + 4)>>3;

  2578. 				}

  2579. 			}

  2580. 		}

  2581. 		else

  2582. 		{

  2583. 			for(y=0; y<8; y++)

  2584. 			{

  2585. 				int x;

  2586. 				for(x=0; x<8; x++)

  2587. 				{

  2588. 					int ref= tempBlured[ x + y*stride ];

  2589. 					int cur= src[ x + y*stride ];

  2590. 					tempBlured[ x + y*stride ]=

  2591. 					src[ x + y*stride ]=

  2592. 						(ref*3 + cur + 2)>>2;

  2593. 				}

  2594. 			}

  2595. 		}

  2596. 	}

  2597. }

  2598. #endif

  2599. }

  2600. 

  2601. static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,

  2602. 	QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c);

  2603. 

  2604. /**

  2605.  * Copies a block from src to dst and fixes the blacklevel

  2606.  * levelFix == 0 -> dont touch the brighness & contrast

  2607.  */

  2608. #undef SCALED_CPY

  2609. 

  2610. static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,

  2611. 	int levelFix, int64_t *packedOffsetAndScale)

  2612. {

  2613. #ifndef HAVE_MMX

  2614. 	int i;

  2615. #endif

  2616. 	if(levelFix)

  2617. 	{

  2618. #ifdef HAVE_MMX

  2619. 					asm volatile(

  2620. 						"movq (%%eax), %%mm2	\n\t" // packedYOffset

  2621. 						"movq 8(%%eax), %%mm3	\n\t" // packedYScale

  2622. 						"leal (%2,%4), %%eax	\n\t"

  2623. 						"leal (%3,%5), %%edx	\n\t"

  2624. 						"pxor %%mm4, %%mm4	\n\t"

  2625. #ifdef HAVE_MMX2

  2626. #define SCALED_CPY(src1, src2, dst1, dst2)					\

  2627. 						"movq " #src1 ", %%mm0	\n\t"\

  2628. 						"movq " #src1 ", %%mm5	\n\t"\

  2629. 						"movq " #src2 ", %%mm1	\n\t"\

  2630. 						"movq " #src2 ", %%mm6	\n\t"\

  2631. 						"punpcklbw %%mm0, %%mm0 \n\t"\

  2632. 						"punpckhbw %%mm5, %%mm5 \n\t"\

  2633. 						"punpcklbw %%mm1, %%mm1 \n\t"\

  2634. 						"punpckhbw %%mm6, %%mm6 \n\t"\

  2635. 						"pmulhuw %%mm3, %%mm0	\n\t"\

  2636. 						"pmulhuw %%mm3, %%mm5	\n\t"\

  2637. 						"pmulhuw %%mm3, %%mm1	\n\t"\

  2638. 						"pmulhuw %%mm3, %%mm6	\n\t"\

  2639. 						"psubw %%mm2, %%mm0	\n\t"\

  2640. 						"psubw %%mm2, %%mm5	\n\t"\

  2641. 						"psubw %%mm2, %%mm1	\n\t"\

  2642. 						"psubw %%mm2, %%mm6	\n\t"\

  2643. 						"packuswb %%mm5, %%mm0	\n\t"\

  2644. 						"packuswb %%mm6, %%mm1	\n\t"\

  2645. 						"movq %%mm0, " #dst1 "	\n\t"\

  2646. 						"movq %%mm1, " #dst2 "	\n\t"\

  2647. 

  2648. #else //HAVE_MMX2

  2649. #define SCALED_CPY(src1, src2, dst1, dst2)					\

  2650. 						"movq " #src1 ", %%mm0	\n\t"\

  2651. 						"movq " #src1 ", %%mm5	\n\t"\

  2652. 						"punpcklbw %%mm4, %%mm0 \n\t"\

  2653. 						"punpckhbw %%mm4, %%mm5 \n\t"\

  2654. 						"psubw %%mm2, %%mm0	\n\t"\

  2655. 						"psubw %%mm2, %%mm5	\n\t"\

  2656. 						"movq " #src2 ", %%mm1	\n\t"\

  2657. 						"psllw $6, %%mm0	\n\t"\

  2658. 						"psllw $6, %%mm5	\n\t"\

  2659. 						"pmulhw %%mm3, %%mm0	\n\t"\

  2660. 						"movq " #src2 ", %%mm6	\n\t"\

  2661. 						"pmulhw %%mm3, %%mm5	\n\t"\

  2662. 						"punpcklbw %%mm4, %%mm1 \n\t"\

  2663. 						"punpckhbw %%mm4, %%mm6 \n\t"\

  2664. 						"psubw %%mm2, %%mm1	\n\t"\

  2665. 						"psubw %%mm2, %%mm6	\n\t"\

  2666. 						"psllw $6, %%mm1	\n\t"\

  2667. 						"psllw $6, %%mm6	\n\t"\

  2668. 						"pmulhw %%mm3, %%mm1	\n\t"\

  2669. 						"pmulhw %%mm3, %%mm6	\n\t"\

  2670. 						"packuswb %%mm5, %%mm0	\n\t"\

  2671. 						"packuswb %%mm6, %%mm1	\n\t"\

  2672. 						"movq %%mm0, " #dst1 "	\n\t"\

  2673. 						"movq %%mm1, " #dst2 "	\n\t"\

  2674. 

  2675. #endif //!HAVE_MMX2

  2676. 

  2677. SCALED_CPY((%2)       , (%2, %4)      , (%3)       , (%3, %5))

  2678. SCALED_CPY((%2, %4, 2), (%%eax, %4, 2), (%3, %5, 2), (%%edx, %5, 2))

  2679. SCALED_CPY((%2, %4, 4), (%%eax, %4, 4), (%3, %5, 4), (%%edx, %5, 4))

  2680. 						"leal (%%eax,%4,4), %%eax	\n\t"

  2681. 						"leal (%%edx,%5,4), %%edx	\n\t"

  2682. SCALED_CPY((%%eax, %4), (%%eax, %4, 2), (%%edx, %5), (%%edx, %5, 2))

  2683. 

  2684. 

  2685. 						: "=&a" (packedOffsetAndScale)

  2686. 						: "0" (packedOffsetAndScale),

  2687. 						"r"(src),

  2688. 						"r"(dst),

  2689. 						"r" (srcStride),

  2690. 						"r" (dstStride)

  2691. 						: "%edx"

  2692. 					);

  2693. #else

  2694. 				for(i=0; i<8; i++)

  2695. 					memcpy(	&(dst[dstStride*i]),

  2696. 						&(src[srcStride*i]), BLOCK_SIZE);

  2697. #endif

  2698. 	}

  2699. 	else

  2700. 	{

  2701. #ifdef HAVE_MMX

  2702. 					asm volatile(

  2703. 						"leal (%0,%2), %%eax	\n\t"

  2704. 						"leal (%1,%3), %%edx	\n\t"

  2705. 

  2706. #define SIMPLE_CPY(src1, src2, dst1, dst2)				\

  2707. 						"movq " #src1 ", %%mm0	\n\t"\

  2708. 						"movq " #src2 ", %%mm1	\n\t"\

  2709. 						"movq %%mm0, " #dst1 "	\n\t"\

  2710. 						"movq %%mm1, " #dst2 "	\n\t"\

  2711. 

  2712. SIMPLE_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))

  2713. SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%edx, %3, 2))

  2714. SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%edx, %3, 4))

  2715. 						"leal (%%eax,%2,4), %%eax	\n\t"

  2716. 						"leal (%%edx,%3,4), %%edx	\n\t"

  2717. SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%edx, %3), (%%edx, %3, 2))

  2718. 

  2719. 						: : "r" (src),

  2720. 						"r" (dst),

  2721. 						"r" (srcStride),

  2722. 						"r" (dstStride)

  2723. 						: "%eax", "%edx"

  2724. 					);

  2725. #else

  2726. 				for(i=0; i<8; i++)

  2727. 					memcpy(	&(dst[dstStride*i]),

  2728. 						&(src[srcStride*i]), BLOCK_SIZE);

  2729. #endif

  2730. 	}

  2731. }

  2732. 

  2733. /**

  2734.  * Duplicates the given 8 src pixels ? times upward

  2735.  */

  2736. static inline void RENAME(duplicate)(uint8_t src[], int stride)

  2737. {

  2738. #ifdef HAVE_MMX

  2739. 	asm volatile(

  2740. 		"movq (%0), %%mm0		\n\t"

  2741. 		"addl %1, %0			\n\t"

  2742. 		"movq %%mm0, (%0)		\n\t"

  2743. 		"movq %%mm0, (%0, %1)		\n\t"

  2744. 		"movq %%mm0, (%0, %1, 2)	\n\t"

  2745. 		: "+r" (src)

  2746. 		: "r" (-stride)

  2747. 	);

  2748. #else

  2749. 	int i;

  2750. 	uint8_t *p=src;

  2751. 	for(i=0; i<3; i++)

  2752. 	{

  2753. 		p-= stride;

  2754. 		memcpy(p, src, 8);

  2755. 	}

  2756. #endif

  2757. }

  2758. 

  2759. /**

  2760.  * Filters array of bytes (Y or U or V values)

  2761.  */

  2762. static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,

  2763. 	QP_STORE_T QPs[], int QPStride, int isColor, PPContext *c2)

  2764. {

  2765. 	PPContext __attribute__((aligned(8))) c= *c2; //copy to stack for faster access

  2766. 	int x,y;

  2767. #ifdef COMPILE_TIME_MODE

  2768. 	const int mode= COMPILE_TIME_MODE;

  2769. #else

  2770. 	const int mode= isColor ? c.ppMode.chromMode : c.ppMode.lumMode;

  2771. #endif

  2772. 	int black=0, white=255; // blackest black and whitest white in the picture

  2773. 	int QPCorrecture= 256*256;

  2774. 

  2775. 	int copyAhead;

  2776. #ifdef HAVE_MMX

  2777. 	int i;

  2778. #endif

  2779. 

  2780. 	const int qpHShift= isColor ? 4-c.hChromaSubSample : 4;

  2781. 	const int qpVShift= isColor ? 4-c.vChromaSubSample : 4;

  2782. 

  2783. 	//FIXME remove

  2784. 	uint64_t * const yHistogram= c.yHistogram;

  2785. 	uint8_t * const tempSrc= c.tempSrc;

  2786. 	uint8_t * const tempDst= c.tempDst;

  2787. 	const int mbWidth= isColor ? (width+7)>>3 : (width+15)>>4;

  2788. 

  2789. #ifdef HAVE_MMX

  2790. 	for(i=0; i<32; i++){

  2791. 		int offset= ((i*c.ppMode.baseDcDiff)>>8) + 1;

  2792. 		int threshold= offset*2 + 1;

  2793. 		c.mmxDcOffset[i]= 0x7F - offset;

  2794. 		c.mmxDcThreshold[i]= 0x7F - threshold;

  2795. 		c.mmxDcOffset[i]*= 0x0101010101010101LL;

  2796. 		c.mmxDcThreshold[i]*= 0x0101010101010101LL;

  2797. 	}

  2798. #endif

  2799. 

  2800. 	if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;

  2801. 	else if(   (mode & LINEAR_BLEND_DEINT_FILTER)

  2802. 		|| (mode & FFMPEG_DEINT_FILTER)

  2803. 		|| (mode & LOWPASS5_DEINT_FILTER)) copyAhead=14;

  2804. 	else if(   (mode & V_DEBLOCK)

  2805. 		|| (mode & LINEAR_IPOL_DEINT_FILTER)

  2806. 		|| (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;

  2807. 	else if(mode & V_X1_FILTER) copyAhead=11;

  2808. //	else if(mode & V_RK1_FILTER) copyAhead=10;

  2809. 	else if(mode & DERING) copyAhead=9;

  2810. 	else copyAhead=8;

  2811. 

  2812. 	copyAhead-= 8;

  2813. 

  2814. 	if(!isColor)

  2815. 	{

  2816. 		uint64_t sum= 0;

  2817. 		int i;

  2818. 		uint64_t maxClipped;

  2819. 		uint64_t clipped;

  2820. 		double scale;

  2821. 

  2822. 		c.frameNum++;

  2823. 		// first frame is fscked so we ignore it

  2824. 		if(c.frameNum == 1) yHistogram[0]= width*height/64*15/256;

  2825. 

  2826. 		for(i=0; i<256; i++)

  2827. 		{

  2828. 			sum+= yHistogram[i];

  2829. //			printf("%d ", yHistogram[i]);

  2830. 		}

  2831. //		printf("\n\n");

  2832. 

  2833. 		/* we allways get a completly black picture first */

  2834. 		maxClipped= (uint64_t)(sum * c.ppMode.maxClippedThreshold);

  2835. 

  2836. 		clipped= sum;

  2837. 		for(black=255; black>0; black--)

  2838. 		{

  2839. 			if(clipped < maxClipped) break;

  2840. 			clipped-= yHistogram[black];

  2841. 		}

  2842. 

  2843. 		clipped= sum;

  2844. 		for(white=0; white<256; white++)

  2845. 		{

  2846. 			if(clipped < maxClipped) break;

  2847. 			clipped-= yHistogram[white];

  2848. 		}

  2849. 

  2850. 		scale= (double)(c.ppMode.maxAllowedY - c.ppMode.minAllowedY) / (double)(white-black);

  2851. 

  2852. #ifdef HAVE_MMX2

  2853. 		c.packedYScale= (uint16_t)(scale*256.0 + 0.5);

  2854. 		c.packedYOffset= (((black*c.packedYScale)>>8) - c.ppMode.minAllowedY) & 0xFFFF;

  2855. #else

  2856. 		c.packedYScale= (uint16_t)(scale*1024.0 + 0.5);

  2857. 		c.packedYOffset= (black - c.ppMode.minAllowedY) & 0xFFFF;

  2858. #endif

  2859. 

  2860. 		c.packedYOffset|= c.packedYOffset<<32;

  2861. 		c.packedYOffset|= c.packedYOffset<<16;

  2862. 

  2863. 		c.packedYScale|= c.packedYScale<<32;

  2864. 		c.packedYScale|= c.packedYScale<<16;

  2865. 		

  2866. 		if(mode & LEVEL_FIX)	QPCorrecture= (int)(scale*256*256 + 0.5);

  2867. 		else			QPCorrecture= 256*256;

  2868. 	}

  2869. 	else

  2870. 	{

  2871. 		c.packedYScale= 0x0100010001000100LL;

  2872. 		c.packedYOffset= 0;

  2873. 		QPCorrecture= 256*256;

  2874. 	}

  2875. 

  2876. 	/* copy & deinterlace first row of blocks */

  2877. 	y=-BLOCK_SIZE;

  2878. 	{

  2879. 		uint8_t *srcBlock= &(src[y*srcStride]);

  2880. 		uint8_t *dstBlock= tempDst + dstStride;

  2881. 

  2882. 		// From this point on it is guranteed that we can read and write 16 lines downward

  2883. 		// finish 1 block before the next otherwise we´ll might have a problem

  2884. 		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing

  2885. 		for(x=0; x<width; x+=BLOCK_SIZE)

  2886. 		{

  2887. 

  2888. #ifdef HAVE_MMX2

  2889. /*

  2890. 			prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);

  2891. 			prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);

  2892. 			prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);

  2893. 			prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);

  2894. */

  2895. 

  2896. 			asm(

  2897. 				"movl %4, %%eax			\n\t"

  2898. 				"shrl $2, %%eax			\n\t"

  2899. 				"andl $6, %%eax			\n\t"

  2900. 				"addl %5, %%eax			\n\t"

  2901. 				"movl %%eax, %%edx		\n\t"

  2902. 				"imul %1, %%eax			\n\t"

  2903. 				"imul %3, %%edx			\n\t"

  2904. 				"prefetchnta 32(%%eax, %0)	\n\t"

  2905. 				"prefetcht0 32(%%edx, %2)	\n\t"

  2906. 				"addl %1, %%eax			\n\t"

  2907. 				"addl %3, %%edx			\n\t"

  2908. 				"prefetchnta 32(%%eax, %0)	\n\t"

  2909. 				"prefetcht0 32(%%edx, %2)	\n\t"

  2910. 			:: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),

  2911. 			"m" (x), "m" (copyAhead)

  2912. 			: "%eax", "%edx"

  2913. 			);

  2914. 

  2915. #elif defined(HAVE_3DNOW)

  2916. //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...

  2917. /*			prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);

  2918. 			prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);

  2919. 			prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);

  2920. 			prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);

  2921. */

  2922. #endif

  2923. 

  2924. 			RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,

  2925. 				srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX, &c.packedYOffset);

  2926. 

  2927. 			RENAME(duplicate)(dstBlock + dstStride*8, dstStride);

  2928. 

  2929. 			if(mode & LINEAR_IPOL_DEINT_FILTER)

  2930. 				RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);

  2931. 			else if(mode & LINEAR_BLEND_DEINT_FILTER)

  2932. 				RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);

  2933. 			else if(mode & MEDIAN_DEINT_FILTER)

  2934. 				RENAME(deInterlaceMedian)(dstBlock, dstStride);

  2935. 			else if(mode & CUBIC_IPOL_DEINT_FILTER)

  2936. 				RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);

  2937. 			else if(mode & FFMPEG_DEINT_FILTER)

  2938. 				RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);

  2939. 			else if(mode & LOWPASS5_DEINT_FILTER)

  2940. 				RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);

  2941. /*			else if(mode & CUBIC_BLEND_DEINT_FILTER)

  2942. 				RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);

  2943. */

  2944. 			dstBlock+=8;

  2945. 			srcBlock+=8;

  2946. 		}

  2947. 		if(width==dstStride)

  2948. 			memcpy(dst, tempDst + 9*dstStride, copyAhead*dstStride);

  2949. 		else

  2950. 		{

  2951. 			int i;

  2952. 			for(i=0; i<copyAhead; i++)

  2953. 			{

  2954. 				memcpy(dst + i*dstStride, tempDst + (9+i)*dstStride, width);

  2955. 			}

  2956. 		}

  2957. 	}

  2958. 

  2959. //printf("\n");

  2960. 	for(y=0; y<height; y+=BLOCK_SIZE)

  2961. 	{

  2962. 		//1% speedup if these are here instead of the inner loop

  2963. 		uint8_t *srcBlock= &(src[y*srcStride]);

  2964. 		uint8_t *dstBlock= &(dst[y*dstStride]);

  2965. #ifdef HAVE_MMX

  2966. 		uint8_t *tempBlock1= c.tempBlocks;

  2967. 		uint8_t *tempBlock2= c.tempBlocks + 8;

  2968. #endif

  2969. 		int8_t *QPptr= &QPs[(y>>qpVShift)*QPStride];

  2970. 		int8_t *nonBQPptr= &c.nonBQPTable[(y>>qpVShift)*QPStride];

  2971. 		int QP=0;

  2972. 		/* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards

  2973. 		   if not than use a temporary buffer */

  2974. 		if(y+15 >= height)

  2975. 		{

  2976. 			int i;

  2977. 			/* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with

  2978. 			   blockcopy to dst later */

  2979. 			memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,

  2980. 				srcStride*MAX(height-y-copyAhead, 0) );

  2981. 

  2982. 			/* duplicate last line of src to fill the void upto line (copyAhead+7) */

  2983. 			for(i=MAX(height-y, 8); i<copyAhead+8; i++)

  2984. 				memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);

  2985. 

  2986. 			/* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/

  2987. 			memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );

  2988. 

  2989. 			/* duplicate last line of dst to fill the void upto line (copyAhead) */

  2990. 			for(i=height-y+1; i<=copyAhead; i++)

  2991. 				memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);

  2992. 

  2993. 			dstBlock= tempDst + dstStride;

  2994. 			srcBlock= tempSrc;

  2995. 		}

  2996. //printf("\n");

  2997. 

  2998. 		// From this point on it is guranteed that we can read and write 16 lines downward

  2999. 		// finish 1 block before the next otherwise we´ll might have a problem

  3000. 		// with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing

  3001. 		for(x=0; x<width; x+=BLOCK_SIZE)

  3002. 		{

  3003. 			const int stride= dstStride;

  3004. #ifdef HAVE_MMX

  3005. 			uint8_t *tmpXchg;

  3006. #endif

  3007. 			if(isColor)

  3008. 			{

  3009. 				QP= QPptr[x>>qpHShift];

  3010. 				c.nonBQP= nonBQPptr[x>>qpHShift];

  3011. 			}

  3012. 			else

  3013. 			{

  3014. 				QP= QPptr[x>>4];

  3015. 				QP= (QP* QPCorrecture + 256*128)>>16;

  3016. 				c.nonBQP= nonBQPptr[x>>4];

  3017. 				c.nonBQP= (c.nonBQP* QPCorrecture + 256*128)>>16;

  3018. 				yHistogram[ srcBlock[srcStride*12 + 4] ]++;

  3019. 			}

  3020. 			c.QP= QP;

  3021. #ifdef HAVE_MMX

  3022. 			asm volatile(

  3023. 				"movd %1, %%mm7					\n\t"

  3024. 				"packuswb %%mm7, %%mm7				\n\t" // 0, 0, 0, QP, 0, 0, 0, QP

  3025. 				"packuswb %%mm7, %%mm7				\n\t" // 0,QP, 0, QP, 0,QP, 0, QP

  3026. 				"packuswb %%mm7, %%mm7				\n\t" // QP,..., QP

  3027. 				"movq %%mm7, %0			\n\t"

  3028. 				: "=m" (c.pQPb) 

  3029. 				: "r" (QP)

  3030. 			);

  3031. #endif

  3032. 

  3033. 

  3034. #ifdef HAVE_MMX2

  3035. /*

  3036. 			prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);

  3037. 			prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);

  3038. 			prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);

  3039. 			prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);

  3040. */

  3041. 

  3042. 			asm(

  3043. 				"movl %4, %%eax			\n\t"

  3044. 				"shrl $2, %%eax			\n\t"

  3045. 				"andl $6, %%eax			\n\t"

  3046. 				"addl %5, %%eax			\n\t"

  3047. 				"movl %%eax, %%edx		\n\t"

  3048. 				"imul %1, %%eax			\n\t"

  3049. 				"imul %3, %%edx			\n\t"

  3050. 				"prefetchnta 32(%%eax, %0)	\n\t"

  3051. 				"prefetcht0 32(%%edx, %2)	\n\t"

  3052. 				"addl %1, %%eax			\n\t"

  3053. 				"addl %3, %%edx			\n\t"

  3054. 				"prefetchnta 32(%%eax, %0)	\n\t"

  3055. 				"prefetcht0 32(%%edx, %2)	\n\t"

  3056. 			:: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),

  3057. 			"m" (x), "m" (copyAhead)

  3058. 			: "%eax", "%edx"

  3059. 			);

  3060. 

  3061. #elif defined(HAVE_3DNOW)

  3062. //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...

  3063. /*			prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);

  3064. 			prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);

  3065. 			prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);

  3066. 			prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);

  3067. */

  3068. #endif

  3069. 

  3070. 			RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,

  3071. 				srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX, &c.packedYOffset);

  3072. 

  3073. 			if(mode & LINEAR_IPOL_DEINT_FILTER)

  3074. 				RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);

  3075. 			else if(mode & LINEAR_BLEND_DEINT_FILTER)

  3076. 				RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);

  3077. 			else if(mode & MEDIAN_DEINT_FILTER)

  3078. 				RENAME(deInterlaceMedian)(dstBlock, dstStride);

  3079. 			else if(mode & CUBIC_IPOL_DEINT_FILTER)

  3080. 				RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);

  3081. 			else if(mode & FFMPEG_DEINT_FILTER)

  3082. 				RENAME(deInterlaceFF)(dstBlock, dstStride, c.deintTemp + x);

  3083. 			else if(mode & LOWPASS5_DEINT_FILTER)

  3084. 				RENAME(deInterlaceL5)(dstBlock, dstStride, c.deintTemp + x, c.deintTemp + width + x);

  3085. /*			else if(mode & CUBIC_BLEND_DEINT_FILTER)

  3086. 				RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);

  3087. */

  3088. 

  3089. 			/* only deblock if we have 2 blocks */

  3090. 			if(y + 8 < height)

  3091. 			{

  3092. 				if(mode & V_X1_FILTER)

  3093. 					RENAME(vertX1Filter)(dstBlock, stride, &c);

  3094. 				else if(mode & V_DEBLOCK)

  3095. 				{

  3096. 					const int t= RENAME(vertClassify)(dstBlock, stride, &c);

  3097. 

  3098. 					if(t==1)

  3099. 						RENAME(doVertLowPass)(dstBlock, stride, &c);

  3100. 					else if(t==2)

  3101. 						RENAME(doVertDefFilter)(dstBlock, stride, &c);

  3102. 				}

  3103. 			}

  3104. 

  3105. #ifdef HAVE_MMX

  3106. 			RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);

  3107. #endif

  3108. 			/* check if we have a previous block to deblock it with dstBlock */

  3109. 			if(x - 8 >= 0)

  3110. 			{

  3111. #ifdef HAVE_MMX

  3112. 				if(mode & H_X1_FILTER)

  3113. 					RENAME(vertX1Filter)(tempBlock1, 16, &c);

  3114. 				else if(mode & H_DEBLOCK)

  3115. 				{

  3116. //START_TIMER

  3117. 					const int t= RENAME(vertClassify)(tempBlock1, 16, &c);

  3118. //STOP_TIMER("dc & minmax")

  3119.                                         if(t==1)

  3120. 						RENAME(doVertLowPass)(tempBlock1, 16, &c);

  3121. 					else if(t==2)

  3122. 						RENAME(doVertDefFilter)(tempBlock1, 16, &c);

  3123. 				}

  3124. 

  3125. 				RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);

  3126. 

  3127. #else

  3128. 				if(mode & H_X1_FILTER)

  3129. 					horizX1Filter(dstBlock-4, stride, QP);

  3130. 				else if(mode & H_DEBLOCK)

  3131. 				{

  3132. 					if( isHorizDC(dstBlock-4, stride, &c))

  3133. 					{

  3134. 						if(isHorizMinMaxOk(dstBlock-4, stride, QP))

  3135. 							doHorizLowPass(dstBlock-4, stride, QP);

  3136. 					}

  3137. 					else

  3138. 						doHorizDefFilter(dstBlock-4, stride, QP);

  3139. 				}

  3140. #endif

  3141. 				if(mode & DERING)

  3142. 				{

  3143. 				//FIXME filter first line

  3144. 					if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, &c);

  3145. 				}

  3146. 

  3147. 				if(mode & TEMP_NOISE_FILTER)

  3148. 				{

  3149. 					RENAME(tempNoiseReducer)(dstBlock-8, stride,

  3150. 						c.tempBlured[isColor] + y*dstStride + x,

  3151. 						c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),

  3152. 						c.ppMode.maxTmpNoise);

  3153. 				}

  3154. 			}

  3155. 

  3156. 			dstBlock+=8;

  3157. 			srcBlock+=8;

  3158. 

  3159. #ifdef HAVE_MMX

  3160. 			tmpXchg= tempBlock1;

  3161. 			tempBlock1= tempBlock2;

  3162. 			tempBlock2 = tmpXchg;

  3163. #endif

  3164. 		}

  3165. 

  3166. 		if(mode & DERING)

  3167. 		{

  3168. 				if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, &c);

  3169. 		}

  3170. 

  3171. 		if((mode & TEMP_NOISE_FILTER))

  3172. 		{

  3173. 			RENAME(tempNoiseReducer)(dstBlock-8, dstStride,

  3174. 				c.tempBlured[isColor] + y*dstStride + x,

  3175. 				c.tempBluredPast[isColor] + (y>>3)*256 + (x>>3),

  3176. 				c.ppMode.maxTmpNoise);

  3177. 		}

  3178. 

  3179. 		/* did we use a tmp buffer for the last lines*/

  3180. 		if(y+15 >= height)

  3181. 		{

  3182. 			uint8_t *dstBlock= &(dst[y*dstStride]);

  3183. 			if(width==dstStride)

  3184. 				memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y));

  3185. 			else

  3186. 			{

  3187. 				int i;

  3188. 				for(i=0; i<height-y; i++)

  3189. 				{

  3190. 					memcpy(dstBlock + i*dstStride, tempDst + (i+1)*dstStride, width);

  3191. 				}

  3192. 			}

  3193. 		}

  3194. /*

  3195. 		for(x=0; x<width; x+=32)

  3196. 		{

  3197. 			volatile int i;

  3198. 			i+=	+ dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]

  3199. 				+ dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]

  3200. 				+ dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];

  3201. //				+ dstBlock[x +13*dstStride]

  3202. //				+ dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];

  3203. 		}*/

  3204. 	}

  3205. #ifdef HAVE_3DNOW

  3206. 	asm volatile("femms");

  3207. #elif defined (HAVE_MMX)

  3208. 	asm volatile("emms");

  3209. #endif

  3210. 

  3211. #ifdef DEBUG_BRIGHTNESS

  3212. 	if(!isColor)

  3213. 	{

  3214. 		int max=1;

  3215. 		int i;

  3216. 		for(i=0; i<256; i++)

  3217. 			if(yHistogram[i] > max) max=yHistogram[i];

  3218. 

  3219. 		for(i=1; i<256; i++)

  3220. 		{

  3221. 			int x;

  3222. 			int start=yHistogram[i-1]/(max/256+1);

  3223. 			int end=yHistogram[i]/(max/256+1);

  3224. 			int inc= end > start ? 1 : -1;

  3225. 			for(x=start; x!=end+inc; x+=inc)

  3226. 				dst[ i*dstStride + x]+=128;

  3227. 		}

  3228. 

  3229. 		for(i=0; i<100; i+=2)

  3230. 		{

  3231. 			dst[ (white)*dstStride + i]+=128;

  3232. 			dst[ (black)*dstStride + i]+=128;

  3233. 		}

  3234. 

  3235. 	}

  3236. #endif

  3237. 

  3238. 	*c2= c; //copy local context back

  3239. 

  3240. }