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.

425 lines
20KB

  1. /*
  2. * MPEG video MMX templates
  3. *
  4. * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
  5. *
  6. * This file is part of FFmpeg.
  7. *
  8. * FFmpeg is free software; you can redistribute it and/or
  9. * modify it under the terms of the GNU Lesser General Public
  10. * License as published by the Free Software Foundation; either
  11. * version 2.1 of the License, or (at your option) any later version.
  12. *
  13. * FFmpeg is distributed in the hope that it will be useful,
  14. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  16. * Lesser General Public License for more details.
  17. *
  18. * You should have received a copy of the GNU Lesser General Public
  19. * License along with FFmpeg; if not, write to the Free Software
  20. * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21. */
  22. #include <stdint.h>
  23. #include "libavutil/internal.h"
  24. #include "libavutil/mem_internal.h"
  25. #include "libavutil/x86/asm.h"
  26. #include "libavcodec/mpegutils.h"
  27. #include "libavcodec/mpegvideo.h"
  28. #include "fdct.h"
  29. #undef MMREG_WIDTH
  30. #undef MM
  31. #undef MOVQ
  32. #undef SPREADW
  33. #undef PMAXW
  34. #undef PMAX
  35. #undef SAVE_SIGN
  36. #undef RESTORE_SIGN
  37. #if COMPILE_TEMPLATE_SSE2
  38. #define MMREG_WIDTH "16"
  39. #define MM "%%xmm"
  40. #define MOVQ "movdqa"
  41. #define SPREADW(a) \
  42. "pshuflw $0, "a", "a" \n\t"\
  43. "punpcklwd "a", "a" \n\t"
  44. #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
  45. #define PMAX(a,b) \
  46. "movhlps "a", "b" \n\t"\
  47. PMAXW(b, a)\
  48. "pshuflw $0x0E, "a", "b" \n\t"\
  49. PMAXW(b, a)\
  50. "pshuflw $0x01, "a", "b" \n\t"\
  51. PMAXW(b, a)
  52. #else
  53. #define MMREG_WIDTH "8"
  54. #define MM "%%mm"
  55. #define MOVQ "movq"
  56. #if COMPILE_TEMPLATE_MMXEXT
  57. #define SPREADW(a) "pshufw $0, "a", "a" \n\t"
  58. #define PMAXW(a,b) "pmaxsw "a", "b" \n\t"
  59. #define PMAX(a,b) \
  60. "pshufw $0x0E, "a", "b" \n\t"\
  61. PMAXW(b, a)\
  62. "pshufw $0x01, "a", "b" \n\t"\
  63. PMAXW(b, a)
  64. #else
  65. #define SPREADW(a) \
  66. "punpcklwd "a", "a" \n\t"\
  67. "punpcklwd "a", "a" \n\t"
  68. #define PMAXW(a,b) \
  69. "psubusw "a", "b" \n\t"\
  70. "paddw "a", "b" \n\t"
  71. #define PMAX(a,b) \
  72. "movq "a", "b" \n\t"\
  73. "psrlq $32, "a" \n\t"\
  74. PMAXW(b, a)\
  75. "movq "a", "b" \n\t"\
  76. "psrlq $16, "a" \n\t"\
  77. PMAXW(b, a)
  78. #endif
  79. #endif
  80. #if COMPILE_TEMPLATE_SSSE3
  81. #define SAVE_SIGN(a,b) \
  82. "movdqa "b", "a" \n\t"\
  83. "pabsw "b", "b" \n\t"
  84. #define RESTORE_SIGN(a,b) \
  85. "psignw "a", "b" \n\t"
  86. #else
  87. #define SAVE_SIGN(a,b) \
  88. "pxor "a", "a" \n\t"\
  89. "pcmpgtw "b", "a" \n\t" /* block[i] <= 0 ? 0xFF : 0x00 */\
  90. "pxor "a", "b" \n\t"\
  91. "psubw "a", "b" \n\t" /* ABS(block[i]) */
  92. #define RESTORE_SIGN(a,b) \
  93. "pxor "a", "b" \n\t"\
  94. "psubw "a", "b" \n\t" // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
  95. #endif
  96. static int RENAME(dct_quantize)(MpegEncContext *s,
  97. int16_t *block, int n,
  98. int qscale, int *overflow)
  99. {
  100. x86_reg last_non_zero_p1;
  101. int level=0, q; //=0 is because gcc says uninitialized ...
  102. const uint16_t *qmat, *bias;
  103. LOCAL_ALIGNED_16(int16_t, temp_block, [64]);
  104. av_assert2((7&(int)(&temp_block[0])) == 0); //did gcc align it correctly?
  105. //s->fdct (block);
  106. RENAME_FDCT(ff_fdct)(block); // cannot be anything else ...
  107. if(s->dct_error_sum)
  108. s->denoise_dct(s, block);
  109. if (s->mb_intra) {
  110. int dummy;
  111. if (n < 4){
  112. q = s->y_dc_scale;
  113. bias = s->q_intra_matrix16[qscale][1];
  114. qmat = s->q_intra_matrix16[qscale][0];
  115. }else{
  116. q = s->c_dc_scale;
  117. bias = s->q_chroma_intra_matrix16[qscale][1];
  118. qmat = s->q_chroma_intra_matrix16[qscale][0];
  119. }
  120. /* note: block[0] is assumed to be positive */
  121. if (!s->h263_aic) {
  122. __asm__ volatile (
  123. "mul %%ecx \n\t"
  124. : "=d" (level), "=a"(dummy)
  125. : "a" ((block[0]>>2) + q), "c" (ff_inverse[q<<1])
  126. );
  127. } else
  128. /* For AIC we skip quant/dequant of INTRADC */
  129. level = (block[0] + 4)>>3;
  130. block[0]=0; //avoid fake overflow
  131. // temp_block[0] = (block[0] + (q >> 1)) / q;
  132. last_non_zero_p1 = 1;
  133. } else {
  134. last_non_zero_p1 = 0;
  135. bias = s->q_inter_matrix16[qscale][1];
  136. qmat = s->q_inter_matrix16[qscale][0];
  137. }
  138. if((s->out_format == FMT_H263 || s->out_format == FMT_H261) && s->mpeg_quant==0){
  139. __asm__ volatile(
  140. "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
  141. SPREADW(MM"3")
  142. "pxor "MM"7, "MM"7 \n\t" // 0
  143. "pxor "MM"4, "MM"4 \n\t" // 0
  144. MOVQ" (%2), "MM"5 \n\t" // qmat[0]
  145. "pxor "MM"6, "MM"6 \n\t"
  146. "psubw (%3), "MM"6 \n\t" // -bias[0]
  147. "mov $-128, %%"FF_REG_a" \n\t"
  148. ".p2align 4 \n\t"
  149. "1: \n\t"
  150. MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
  151. SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
  152. "psubusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
  153. "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16
  154. "por "MM"0, "MM"4 \n\t"
  155. RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
  156. MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
  157. "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
  158. MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
  159. MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
  160. "pandn "MM"1, "MM"0 \n\t"
  161. PMAXW(MM"0", MM"3")
  162. "add $"MMREG_WIDTH", %%"FF_REG_a" \n\t"
  163. " js 1b \n\t"
  164. PMAX(MM"3", MM"0")
  165. "movd "MM"3, %%"FF_REG_a" \n\t"
  166. "movzbl %%al, %%eax \n\t" // last_non_zero_p1
  167. : "+a" (last_non_zero_p1)
  168. : "r" (block+64), "r" (qmat), "r" (bias),
  169. "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
  170. XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
  171. "%xmm4", "%xmm5", "%xmm6", "%xmm7")
  172. );
  173. }else{ // FMT_H263
  174. __asm__ volatile(
  175. "movd %%"FF_REG_a", "MM"3 \n\t" // last_non_zero_p1
  176. SPREADW(MM"3")
  177. "pxor "MM"7, "MM"7 \n\t" // 0
  178. "pxor "MM"4, "MM"4 \n\t" // 0
  179. "mov $-128, %%"FF_REG_a" \n\t"
  180. ".p2align 4 \n\t"
  181. "1: \n\t"
  182. MOVQ" (%1, %%"FF_REG_a"), "MM"0 \n\t" // block[i]
  183. SAVE_SIGN(MM"1", MM"0") // ABS(block[i])
  184. MOVQ" (%3, %%"FF_REG_a"), "MM"6 \n\t" // bias[0]
  185. "paddusw "MM"6, "MM"0 \n\t" // ABS(block[i]) + bias[0]
  186. MOVQ" (%2, %%"FF_REG_a"), "MM"5 \n\t" // qmat[i]
  187. "pmulhw "MM"5, "MM"0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
  188. "por "MM"0, "MM"4 \n\t"
  189. RESTORE_SIGN(MM"1", MM"0") // out=((ABS(block[i])*qmat[0] - bias[0]*qmat[0])>>16)*sign(block[i])
  190. MOVQ" "MM"0, (%5, %%"FF_REG_a") \n\t"
  191. "pcmpeqw "MM"7, "MM"0 \n\t" // out==0 ? 0xFF : 0x00
  192. MOVQ" (%4, %%"FF_REG_a"), "MM"1 \n\t"
  193. MOVQ" "MM"7, (%1, %%"FF_REG_a") \n\t" // 0
  194. "pandn "MM"1, "MM"0 \n\t"
  195. PMAXW(MM"0", MM"3")
  196. "add $"MMREG_WIDTH", %%"FF_REG_a" \n\t"
  197. " js 1b \n\t"
  198. PMAX(MM"3", MM"0")
  199. "movd "MM"3, %%"FF_REG_a" \n\t"
  200. "movzbl %%al, %%eax \n\t" // last_non_zero_p1
  201. : "+a" (last_non_zero_p1)
  202. : "r" (block+64), "r" (qmat+64), "r" (bias+64),
  203. "r" (inv_zigzag_direct16 + 64), "r" (temp_block + 64)
  204. XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3",
  205. "%xmm4", "%xmm5", "%xmm6", "%xmm7")
  206. );
  207. }
  208. __asm__ volatile(
  209. "movd %1, "MM"1 \n\t" // max_qcoeff
  210. SPREADW(MM"1")
  211. "psubusw "MM"1, "MM"4 \n\t"
  212. "packuswb "MM"4, "MM"4 \n\t"
  213. #if COMPILE_TEMPLATE_SSE2
  214. "packsswb "MM"4, "MM"4 \n\t"
  215. #endif
  216. "movd "MM"4, %0 \n\t" // *overflow
  217. : "=g" (*overflow)
  218. : "g" (s->max_qcoeff)
  219. );
  220. if(s->mb_intra) block[0]= level;
  221. else block[0]= temp_block[0];
  222. if (s->idsp.perm_type == FF_IDCT_PERM_SIMPLE) {
  223. if(last_non_zero_p1 <= 1) goto end;
  224. block[0x08] = temp_block[0x01]; block[0x10] = temp_block[0x08];
  225. block[0x20] = temp_block[0x10];
  226. if(last_non_zero_p1 <= 4) goto end;
  227. block[0x18] = temp_block[0x09]; block[0x04] = temp_block[0x02];
  228. block[0x09] = temp_block[0x03];
  229. if(last_non_zero_p1 <= 7) goto end;
  230. block[0x14] = temp_block[0x0A]; block[0x28] = temp_block[0x11];
  231. block[0x12] = temp_block[0x18]; block[0x02] = temp_block[0x20];
  232. if(last_non_zero_p1 <= 11) goto end;
  233. block[0x1A] = temp_block[0x19]; block[0x24] = temp_block[0x12];
  234. block[0x19] = temp_block[0x0B]; block[0x01] = temp_block[0x04];
  235. block[0x0C] = temp_block[0x05];
  236. if(last_non_zero_p1 <= 16) goto end;
  237. block[0x11] = temp_block[0x0C]; block[0x29] = temp_block[0x13];
  238. block[0x16] = temp_block[0x1A]; block[0x0A] = temp_block[0x21];
  239. block[0x30] = temp_block[0x28]; block[0x22] = temp_block[0x30];
  240. block[0x38] = temp_block[0x29]; block[0x06] = temp_block[0x22];
  241. if(last_non_zero_p1 <= 24) goto end;
  242. block[0x1B] = temp_block[0x1B]; block[0x21] = temp_block[0x14];
  243. block[0x1C] = temp_block[0x0D]; block[0x05] = temp_block[0x06];
  244. block[0x0D] = temp_block[0x07]; block[0x15] = temp_block[0x0E];
  245. block[0x2C] = temp_block[0x15]; block[0x13] = temp_block[0x1C];
  246. if(last_non_zero_p1 <= 32) goto end;
  247. block[0x0B] = temp_block[0x23]; block[0x34] = temp_block[0x2A];
  248. block[0x2A] = temp_block[0x31]; block[0x32] = temp_block[0x38];
  249. block[0x3A] = temp_block[0x39]; block[0x26] = temp_block[0x32];
  250. block[0x39] = temp_block[0x2B]; block[0x03] = temp_block[0x24];
  251. if(last_non_zero_p1 <= 40) goto end;
  252. block[0x1E] = temp_block[0x1D]; block[0x25] = temp_block[0x16];
  253. block[0x1D] = temp_block[0x0F]; block[0x2D] = temp_block[0x17];
  254. block[0x17] = temp_block[0x1E]; block[0x0E] = temp_block[0x25];
  255. block[0x31] = temp_block[0x2C]; block[0x2B] = temp_block[0x33];
  256. if(last_non_zero_p1 <= 48) goto end;
  257. block[0x36] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
  258. block[0x23] = temp_block[0x34]; block[0x3C] = temp_block[0x2D];
  259. block[0x07] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
  260. block[0x0F] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
  261. if(last_non_zero_p1 <= 56) goto end;
  262. block[0x2E] = temp_block[0x35]; block[0x33] = temp_block[0x3C];
  263. block[0x3E] = temp_block[0x3D]; block[0x27] = temp_block[0x36];
  264. block[0x3D] = temp_block[0x2F]; block[0x2F] = temp_block[0x37];
  265. block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
  266. }else if(s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2){
  267. if(last_non_zero_p1 <= 1) goto end;
  268. block[0x04] = temp_block[0x01];
  269. block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
  270. if(last_non_zero_p1 <= 4) goto end;
  271. block[0x0C] = temp_block[0x09]; block[0x01] = temp_block[0x02];
  272. block[0x05] = temp_block[0x03];
  273. if(last_non_zero_p1 <= 7) goto end;
  274. block[0x09] = temp_block[0x0A]; block[0x14] = temp_block[0x11];
  275. block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
  276. if(last_non_zero_p1 <= 11) goto end;
  277. block[0x1C] = temp_block[0x19];
  278. block[0x11] = temp_block[0x12]; block[0x0D] = temp_block[0x0B];
  279. block[0x02] = temp_block[0x04]; block[0x06] = temp_block[0x05];
  280. if(last_non_zero_p1 <= 16) goto end;
  281. block[0x0A] = temp_block[0x0C]; block[0x15] = temp_block[0x13];
  282. block[0x19] = temp_block[0x1A]; block[0x24] = temp_block[0x21];
  283. block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
  284. block[0x2C] = temp_block[0x29]; block[0x21] = temp_block[0x22];
  285. if(last_non_zero_p1 <= 24) goto end;
  286. block[0x1D] = temp_block[0x1B]; block[0x12] = temp_block[0x14];
  287. block[0x0E] = temp_block[0x0D]; block[0x03] = temp_block[0x06];
  288. block[0x07] = temp_block[0x07]; block[0x0B] = temp_block[0x0E];
  289. block[0x16] = temp_block[0x15]; block[0x1A] = temp_block[0x1C];
  290. if(last_non_zero_p1 <= 32) goto end;
  291. block[0x25] = temp_block[0x23]; block[0x29] = temp_block[0x2A];
  292. block[0x34] = temp_block[0x31]; block[0x38] = temp_block[0x38];
  293. block[0x3C] = temp_block[0x39]; block[0x31] = temp_block[0x32];
  294. block[0x2D] = temp_block[0x2B]; block[0x22] = temp_block[0x24];
  295. if(last_non_zero_p1 <= 40) goto end;
  296. block[0x1E] = temp_block[0x1D]; block[0x13] = temp_block[0x16];
  297. block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
  298. block[0x1B] = temp_block[0x1E]; block[0x26] = temp_block[0x25];
  299. block[0x2A] = temp_block[0x2C]; block[0x35] = temp_block[0x33];
  300. if(last_non_zero_p1 <= 48) goto end;
  301. block[0x39] = temp_block[0x3A]; block[0x3D] = temp_block[0x3B];
  302. block[0x32] = temp_block[0x34]; block[0x2E] = temp_block[0x2D];
  303. block[0x23] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
  304. block[0x27] = temp_block[0x27]; block[0x2B] = temp_block[0x2E];
  305. if(last_non_zero_p1 <= 56) goto end;
  306. block[0x36] = temp_block[0x35]; block[0x3A] = temp_block[0x3C];
  307. block[0x3E] = temp_block[0x3D]; block[0x33] = temp_block[0x36];
  308. block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
  309. block[0x3B] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
  310. } else if (s->idsp.perm_type == FF_IDCT_PERM_NONE) {
  311. if(last_non_zero_p1 <= 1) goto end;
  312. block[0x01] = temp_block[0x01];
  313. block[0x08] = temp_block[0x08]; block[0x10] = temp_block[0x10];
  314. if(last_non_zero_p1 <= 4) goto end;
  315. block[0x09] = temp_block[0x09]; block[0x02] = temp_block[0x02];
  316. block[0x03] = temp_block[0x03];
  317. if(last_non_zero_p1 <= 7) goto end;
  318. block[0x0A] = temp_block[0x0A]; block[0x11] = temp_block[0x11];
  319. block[0x18] = temp_block[0x18]; block[0x20] = temp_block[0x20];
  320. if(last_non_zero_p1 <= 11) goto end;
  321. block[0x19] = temp_block[0x19];
  322. block[0x12] = temp_block[0x12]; block[0x0B] = temp_block[0x0B];
  323. block[0x04] = temp_block[0x04]; block[0x05] = temp_block[0x05];
  324. if(last_non_zero_p1 <= 16) goto end;
  325. block[0x0C] = temp_block[0x0C]; block[0x13] = temp_block[0x13];
  326. block[0x1A] = temp_block[0x1A]; block[0x21] = temp_block[0x21];
  327. block[0x28] = temp_block[0x28]; block[0x30] = temp_block[0x30];
  328. block[0x29] = temp_block[0x29]; block[0x22] = temp_block[0x22];
  329. if(last_non_zero_p1 <= 24) goto end;
  330. block[0x1B] = temp_block[0x1B]; block[0x14] = temp_block[0x14];
  331. block[0x0D] = temp_block[0x0D]; block[0x06] = temp_block[0x06];
  332. block[0x07] = temp_block[0x07]; block[0x0E] = temp_block[0x0E];
  333. block[0x15] = temp_block[0x15]; block[0x1C] = temp_block[0x1C];
  334. if(last_non_zero_p1 <= 32) goto end;
  335. block[0x23] = temp_block[0x23]; block[0x2A] = temp_block[0x2A];
  336. block[0x31] = temp_block[0x31]; block[0x38] = temp_block[0x38];
  337. block[0x39] = temp_block[0x39]; block[0x32] = temp_block[0x32];
  338. block[0x2B] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
  339. if(last_non_zero_p1 <= 40) goto end;
  340. block[0x1D] = temp_block[0x1D]; block[0x16] = temp_block[0x16];
  341. block[0x0F] = temp_block[0x0F]; block[0x17] = temp_block[0x17];
  342. block[0x1E] = temp_block[0x1E]; block[0x25] = temp_block[0x25];
  343. block[0x2C] = temp_block[0x2C]; block[0x33] = temp_block[0x33];
  344. if(last_non_zero_p1 <= 48) goto end;
  345. block[0x3A] = temp_block[0x3A]; block[0x3B] = temp_block[0x3B];
  346. block[0x34] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
  347. block[0x26] = temp_block[0x26]; block[0x1F] = temp_block[0x1F];
  348. block[0x27] = temp_block[0x27]; block[0x2E] = temp_block[0x2E];
  349. if(last_non_zero_p1 <= 56) goto end;
  350. block[0x35] = temp_block[0x35]; block[0x3C] = temp_block[0x3C];
  351. block[0x3D] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
  352. block[0x2F] = temp_block[0x2F]; block[0x37] = temp_block[0x37];
  353. block[0x3E] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
  354. } else if (s->idsp.perm_type == FF_IDCT_PERM_TRANSPOSE) {
  355. if(last_non_zero_p1 <= 1) goto end;
  356. block[0x08] = temp_block[0x01];
  357. block[0x01] = temp_block[0x08]; block[0x02] = temp_block[0x10];
  358. if(last_non_zero_p1 <= 4) goto end;
  359. block[0x09] = temp_block[0x09]; block[0x10] = temp_block[0x02];
  360. block[0x18] = temp_block[0x03];
  361. if(last_non_zero_p1 <= 7) goto end;
  362. block[0x11] = temp_block[0x0A]; block[0x0A] = temp_block[0x11];
  363. block[0x03] = temp_block[0x18]; block[0x04] = temp_block[0x20];
  364. if(last_non_zero_p1 <= 11) goto end;
  365. block[0x0B] = temp_block[0x19];
  366. block[0x12] = temp_block[0x12]; block[0x19] = temp_block[0x0B];
  367. block[0x20] = temp_block[0x04]; block[0x28] = temp_block[0x05];
  368. if(last_non_zero_p1 <= 16) goto end;
  369. block[0x21] = temp_block[0x0C]; block[0x1A] = temp_block[0x13];
  370. block[0x13] = temp_block[0x1A]; block[0x0C] = temp_block[0x21];
  371. block[0x05] = temp_block[0x28]; block[0x06] = temp_block[0x30];
  372. block[0x0D] = temp_block[0x29]; block[0x14] = temp_block[0x22];
  373. if(last_non_zero_p1 <= 24) goto end;
  374. block[0x1B] = temp_block[0x1B]; block[0x22] = temp_block[0x14];
  375. block[0x29] = temp_block[0x0D]; block[0x30] = temp_block[0x06];
  376. block[0x38] = temp_block[0x07]; block[0x31] = temp_block[0x0E];
  377. block[0x2A] = temp_block[0x15]; block[0x23] = temp_block[0x1C];
  378. if(last_non_zero_p1 <= 32) goto end;
  379. block[0x1C] = temp_block[0x23]; block[0x15] = temp_block[0x2A];
  380. block[0x0E] = temp_block[0x31]; block[0x07] = temp_block[0x38];
  381. block[0x0F] = temp_block[0x39]; block[0x16] = temp_block[0x32];
  382. block[0x1D] = temp_block[0x2B]; block[0x24] = temp_block[0x24];
  383. if(last_non_zero_p1 <= 40) goto end;
  384. block[0x2B] = temp_block[0x1D]; block[0x32] = temp_block[0x16];
  385. block[0x39] = temp_block[0x0F]; block[0x3A] = temp_block[0x17];
  386. block[0x33] = temp_block[0x1E]; block[0x2C] = temp_block[0x25];
  387. block[0x25] = temp_block[0x2C]; block[0x1E] = temp_block[0x33];
  388. if(last_non_zero_p1 <= 48) goto end;
  389. block[0x17] = temp_block[0x3A]; block[0x1F] = temp_block[0x3B];
  390. block[0x26] = temp_block[0x34]; block[0x2D] = temp_block[0x2D];
  391. block[0x34] = temp_block[0x26]; block[0x3B] = temp_block[0x1F];
  392. block[0x3C] = temp_block[0x27]; block[0x35] = temp_block[0x2E];
  393. if(last_non_zero_p1 <= 56) goto end;
  394. block[0x2E] = temp_block[0x35]; block[0x27] = temp_block[0x3C];
  395. block[0x2F] = temp_block[0x3D]; block[0x36] = temp_block[0x36];
  396. block[0x3D] = temp_block[0x2F]; block[0x3E] = temp_block[0x37];
  397. block[0x37] = temp_block[0x3E]; block[0x3F] = temp_block[0x3F];
  398. } else {
  399. av_log(s, AV_LOG_DEBUG, "s->idsp.perm_type: %d\n",
  400. (int)s->idsp.perm_type);
  401. av_assert0(s->idsp.perm_type == FF_IDCT_PERM_NONE ||
  402. s->idsp.perm_type == FF_IDCT_PERM_LIBMPEG2 ||
  403. s->idsp.perm_type == FF_IDCT_PERM_SIMPLE ||
  404. s->idsp.perm_type == FF_IDCT_PERM_TRANSPOSE);
  405. }
  406. end:
  407. return last_non_zero_p1 - 1;
  408. }