The Tor Browser / file comparison
comparison: media/libtheora/lib/x86_vc/mmxidct.c
media/libtheora/lib/x86_vc/mmxidct.c
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
equal
deleted
inserted
replaced
| -1:000000000000 | 0:01acd90e95ec |
|---|---|
| 1 /******************************************************************** | |
| 2 * * | |
| 3 * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE. * | |
| 4 * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS * | |
| 5 * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE * | |
| 6 * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. * | |
| 7 * * | |
| 8 * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009 * | |
| 9 * by the Xiph.Org Foundation and contributors http://www.xiph.org/ * | |
| 10 * * | |
| 11 ******************************************************************** | |
| 12 | |
| 13 function: | |
| 14 last mod: $Id: mmxidct.c 17446 2010-09-23 20:06:20Z tterribe $ | |
| 15 | |
| 16 ********************************************************************/ | |
| 17 | |
| 18 /*MMX acceleration of Theora's iDCT. | |
| 19 Originally written by Rudolf Marek, based on code from On2's VP3.*/ | |
| 20 #include "x86int.h" | |
| 21 #include "../dct.h" | |
| 22 | |
| 23 #if defined(OC_X86_ASM) | |
| 24 | |
| 25 /*These are offsets into the table of constants below.*/ | |
| 26 /*7 rows of cosines, in order: pi/16 * (1 ... 7).*/ | |
| 27 #define OC_COSINE_OFFSET (8) | |
| 28 /*A row of 8's.*/ | |
| 29 #define OC_EIGHT_OFFSET (0) | |
| 30 | |
| 31 | |
| 32 | |
| 33 /*A table of constants used by the MMX routines.*/ | |
| 34 static const OC_ALIGN16(ogg_uint16_t) OC_IDCT_CONSTS[(1+7)*4]={ | |
| 35 8, 8, 8, 8, | |
| 36 (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7, | |
| 37 (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7, | |
| 38 (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6, | |
| 39 (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6, | |
| 40 (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5, | |
| 41 (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5, | |
| 42 (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4, | |
| 43 (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4, | |
| 44 (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3, | |
| 45 (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3, | |
| 46 (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2, | |
| 47 (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2, | |
| 48 (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1, | |
| 49 (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1 | |
| 50 }; | |
| 51 | |
| 52 /*38 cycles*/ | |
| 53 #define OC_IDCT_BEGIN(_y,_x) __asm{ \ | |
| 54 __asm movq mm2,OC_I(3,_x) \ | |
| 55 __asm movq mm6,OC_C(3) \ | |
| 56 __asm movq mm4,mm2 \ | |
| 57 __asm movq mm7,OC_J(5,_x) \ | |
| 58 __asm pmulhw mm4,mm6 \ | |
| 59 __asm movq mm1,OC_C(5) \ | |
| 60 __asm pmulhw mm6,mm7 \ | |
| 61 __asm movq mm5,mm1 \ | |
| 62 __asm pmulhw mm1,mm2 \ | |
| 63 __asm movq mm3,OC_I(1,_x) \ | |
| 64 __asm pmulhw mm5,mm7 \ | |
| 65 __asm movq mm0,OC_C(1) \ | |
| 66 __asm paddw mm4,mm2 \ | |
| 67 __asm paddw mm6,mm7 \ | |
| 68 __asm paddw mm2,mm1 \ | |
| 69 __asm movq mm1,OC_J(7,_x) \ | |
| 70 __asm paddw mm7,mm5 \ | |
| 71 __asm movq mm5,mm0 \ | |
| 72 __asm pmulhw mm0,mm3 \ | |
| 73 __asm paddw mm4,mm7 \ | |
| 74 __asm pmulhw mm5,mm1 \ | |
| 75 __asm movq mm7,OC_C(7) \ | |
| 76 __asm psubw mm6,mm2 \ | |
| 77 __asm paddw mm0,mm3 \ | |
| 78 __asm pmulhw mm3,mm7 \ | |
| 79 __asm movq mm2,OC_I(2,_x) \ | |
| 80 __asm pmulhw mm7,mm1 \ | |
| 81 __asm paddw mm5,mm1 \ | |
| 82 __asm movq mm1,mm2 \ | |
| 83 __asm pmulhw mm2,OC_C(2) \ | |
| 84 __asm psubw mm3,mm5 \ | |
| 85 __asm movq mm5,OC_J(6,_x) \ | |
| 86 __asm paddw mm0,mm7 \ | |
| 87 __asm movq mm7,mm5 \ | |
| 88 __asm psubw mm0,mm4 \ | |
| 89 __asm pmulhw mm5,OC_C(2) \ | |
| 90 __asm paddw mm2,mm1 \ | |
| 91 __asm pmulhw mm1,OC_C(6) \ | |
| 92 __asm paddw mm4,mm4 \ | |
| 93 __asm paddw mm4,mm0 \ | |
| 94 __asm psubw mm3,mm6 \ | |
| 95 __asm paddw mm5,mm7 \ | |
| 96 __asm paddw mm6,mm6 \ | |
| 97 __asm pmulhw mm7,OC_C(6) \ | |
| 98 __asm paddw mm6,mm3 \ | |
| 99 __asm movq OC_I(1,_y),mm4 \ | |
| 100 __asm psubw mm1,mm5 \ | |
| 101 __asm movq mm4,OC_C(4) \ | |
| 102 __asm movq mm5,mm3 \ | |
| 103 __asm pmulhw mm3,mm4 \ | |
| 104 __asm paddw mm7,mm2 \ | |
| 105 __asm movq OC_I(2,_y),mm6 \ | |
| 106 __asm movq mm2,mm0 \ | |
| 107 __asm movq mm6,OC_I(0,_x) \ | |
| 108 __asm pmulhw mm0,mm4 \ | |
| 109 __asm paddw mm5,mm3 \ | |
| 110 __asm movq mm3,OC_J(4,_x) \ | |
| 111 __asm psubw mm5,mm1 \ | |
| 112 __asm paddw mm2,mm0 \ | |
| 113 __asm psubw mm6,mm3 \ | |
| 114 __asm movq mm0,mm6 \ | |
| 115 __asm pmulhw mm6,mm4 \ | |
| 116 __asm paddw mm3,mm3 \ | |
| 117 __asm paddw mm1,mm1 \ | |
| 118 __asm paddw mm3,mm0 \ | |
| 119 __asm paddw mm1,mm5 \ | |
| 120 __asm pmulhw mm4,mm3 \ | |
| 121 __asm paddw mm6,mm0 \ | |
| 122 __asm psubw mm6,mm2 \ | |
| 123 __asm paddw mm2,mm2 \ | |
| 124 __asm movq mm0,OC_I(1,_y) \ | |
| 125 __asm paddw mm2,mm6 \ | |
| 126 __asm paddw mm4,mm3 \ | |
| 127 __asm psubw mm2,mm1 \ | |
| 128 } | |
| 129 | |
| 130 /*38+8=46 cycles.*/ | |
| 131 #define OC_ROW_IDCT(_y,_x) __asm{ \ | |
| 132 OC_IDCT_BEGIN(_y,_x) \ | |
| 133 /*r3=D'*/ \ | |
| 134 __asm movq mm3,OC_I(2,_y) \ | |
| 135 /*r4=E'=E-G*/ \ | |
| 136 __asm psubw mm4,mm7 \ | |
| 137 /*r1=H'+H'*/ \ | |
| 138 __asm paddw mm1,mm1 \ | |
| 139 /*r7=G+G*/ \ | |
| 140 __asm paddw mm7,mm7 \ | |
| 141 /*r1=R1=A''+H'*/ \ | |
| 142 __asm paddw mm1,mm2 \ | |
| 143 /*r7=G'=E+G*/ \ | |
| 144 __asm paddw mm7,mm4 \ | |
| 145 /*r4=R4=E'-D'*/ \ | |
| 146 __asm psubw mm4,mm3 \ | |
| 147 __asm paddw mm3,mm3 \ | |
| 148 /*r6=R6=F'-B''*/ \ | |
| 149 __asm psubw mm6,mm5 \ | |
| 150 __asm paddw mm5,mm5 \ | |
| 151 /*r3=R3=E'+D'*/ \ | |
| 152 __asm paddw mm3,mm4 \ | |
| 153 /*r5=R5=F'+B''*/ \ | |
| 154 __asm paddw mm5,mm6 \ | |
| 155 /*r7=R7=G'-C'*/ \ | |
| 156 __asm psubw mm7,mm0 \ | |
| 157 __asm paddw mm0,mm0 \ | |
| 158 /*Save R1.*/ \ | |
| 159 __asm movq OC_I(1,_y),mm1 \ | |
| 160 /*r0=R0=G.+C.*/ \ | |
| 161 __asm paddw mm0,mm7 \ | |
| 162 } | |
| 163 | |
| 164 /*The following macro does two 4x4 transposes in place. | |
| 165 At entry, we assume: | |
| 166 r0 = a3 a2 a1 a0 | |
| 167 I(1) = b3 b2 b1 b0 | |
| 168 r2 = c3 c2 c1 c0 | |
| 169 r3 = d3 d2 d1 d0 | |
| 170 | |
| 171 r4 = e3 e2 e1 e0 | |
| 172 r5 = f3 f2 f1 f0 | |
| 173 r6 = g3 g2 g1 g0 | |
| 174 r7 = h3 h2 h1 h0 | |
| 175 | |
| 176 At exit, we have: | |
| 177 I(0) = d0 c0 b0 a0 | |
| 178 I(1) = d1 c1 b1 a1 | |
| 179 I(2) = d2 c2 b2 a2 | |
| 180 I(3) = d3 c3 b3 a3 | |
| 181 | |
| 182 J(4) = h0 g0 f0 e0 | |
| 183 J(5) = h1 g1 f1 e1 | |
| 184 J(6) = h2 g2 f2 e2 | |
| 185 J(7) = h3 g3 f3 e3 | |
| 186 | |
| 187 I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3. | |
| 188 J(4) J(5) J(6) J(7) is the transpose of r4 r5 r6 r7. | |
| 189 | |
| 190 Since r1 is free at entry, we calculate the Js first.*/ | |
| 191 /*19 cycles.*/ | |
| 192 #define OC_TRANSPOSE(_y) __asm{ \ | |
| 193 __asm movq mm1,mm4 \ | |
| 194 __asm punpcklwd mm4,mm5 \ | |
| 195 __asm movq OC_I(0,_y),mm0 \ | |
| 196 __asm punpckhwd mm1,mm5 \ | |
| 197 __asm movq mm0,mm6 \ | |
| 198 __asm punpcklwd mm6,mm7 \ | |
| 199 __asm movq mm5,mm4 \ | |
| 200 __asm punpckldq mm4,mm6 \ | |
| 201 __asm punpckhdq mm5,mm6 \ | |
| 202 __asm movq mm6,mm1 \ | |
| 203 __asm movq OC_J(4,_y),mm4 \ | |
| 204 __asm punpckhwd mm0,mm7 \ | |
| 205 __asm movq OC_J(5,_y),mm5 \ | |
| 206 __asm punpckhdq mm6,mm0 \ | |
| 207 __asm movq mm4,OC_I(0,_y) \ | |
| 208 __asm punpckldq mm1,mm0 \ | |
| 209 __asm movq mm5,OC_I(1,_y) \ | |
| 210 __asm movq mm0,mm4 \ | |
| 211 __asm movq OC_J(7,_y),mm6 \ | |
| 212 __asm punpcklwd mm0,mm5 \ | |
| 213 __asm movq OC_J(6,_y),mm1 \ | |
| 214 __asm punpckhwd mm4,mm5 \ | |
| 215 __asm movq mm5,mm2 \ | |
| 216 __asm punpcklwd mm2,mm3 \ | |
| 217 __asm movq mm1,mm0 \ | |
| 218 __asm punpckldq mm0,mm2 \ | |
| 219 __asm punpckhdq mm1,mm2 \ | |
| 220 __asm movq mm2,mm4 \ | |
| 221 __asm movq OC_I(0,_y),mm0 \ | |
| 222 __asm punpckhwd mm5,mm3 \ | |
| 223 __asm movq OC_I(1,_y),mm1 \ | |
| 224 __asm punpckhdq mm4,mm5 \ | |
| 225 __asm punpckldq mm2,mm5 \ | |
| 226 __asm movq OC_I(3,_y),mm4 \ | |
| 227 __asm movq OC_I(2,_y),mm2 \ | |
| 228 } | |
| 229 | |
| 230 /*38+19=57 cycles.*/ | |
| 231 #define OC_COLUMN_IDCT(_y) __asm{ \ | |
| 232 OC_IDCT_BEGIN(_y,_y) \ | |
| 233 __asm paddw mm2,OC_8 \ | |
| 234 /*r1=H'+H'*/ \ | |
| 235 __asm paddw mm1,mm1 \ | |
| 236 /*r1=R1=A''+H'*/ \ | |
| 237 __asm paddw mm1,mm2 \ | |
| 238 /*r2=NR2*/ \ | |
| 239 __asm psraw mm2,4 \ | |
| 240 /*r4=E'=E-G*/ \ | |
| 241 __asm psubw mm4,mm7 \ | |
| 242 /*r1=NR1*/ \ | |
| 243 __asm psraw mm1,4 \ | |
| 244 /*r3=D'*/ \ | |
| 245 __asm movq mm3,OC_I(2,_y) \ | |
| 246 /*r7=G+G*/ \ | |
| 247 __asm paddw mm7,mm7 \ | |
| 248 /*Store NR2 at I(2).*/ \ | |
| 249 __asm movq OC_I(2,_y),mm2 \ | |
| 250 /*r7=G'=E+G*/ \ | |
| 251 __asm paddw mm7,mm4 \ | |
| 252 /*Store NR1 at I(1).*/ \ | |
| 253 __asm movq OC_I(1,_y),mm1 \ | |
| 254 /*r4=R4=E'-D'*/ \ | |
| 255 __asm psubw mm4,mm3 \ | |
| 256 __asm paddw mm4,OC_8 \ | |
| 257 /*r3=D'+D'*/ \ | |
| 258 __asm paddw mm3,mm3 \ | |
| 259 /*r3=R3=E'+D'*/ \ | |
| 260 __asm paddw mm3,mm4 \ | |
| 261 /*r4=NR4*/ \ | |
| 262 __asm psraw mm4,4 \ | |
| 263 /*r6=R6=F'-B''*/ \ | |
| 264 __asm psubw mm6,mm5 \ | |
| 265 /*r3=NR3*/ \ | |
| 266 __asm psraw mm3,4 \ | |
| 267 __asm paddw mm6,OC_8 \ | |
| 268 /*r5=B''+B''*/ \ | |
| 269 __asm paddw mm5,mm5 \ | |
| 270 /*r5=R5=F'+B''*/ \ | |
| 271 __asm paddw mm5,mm6 \ | |
| 272 /*r6=NR6*/ \ | |
| 273 __asm psraw mm6,4 \ | |
| 274 /*Store NR4 at J(4).*/ \ | |
| 275 __asm movq OC_J(4,_y),mm4 \ | |
| 276 /*r5=NR5*/ \ | |
| 277 __asm psraw mm5,4 \ | |
| 278 /*Store NR3 at I(3).*/ \ | |
| 279 __asm movq OC_I(3,_y),mm3 \ | |
| 280 /*r7=R7=G'-C'*/ \ | |
| 281 __asm psubw mm7,mm0 \ | |
| 282 __asm paddw mm7,OC_8 \ | |
| 283 /*r0=C'+C'*/ \ | |
| 284 __asm paddw mm0,mm0 \ | |
| 285 /*r0=R0=G'+C'*/ \ | |
| 286 __asm paddw mm0,mm7 \ | |
| 287 /*r7=NR7*/ \ | |
| 288 __asm psraw mm7,4 \ | |
| 289 /*Store NR6 at J(6).*/ \ | |
| 290 __asm movq OC_J(6,_y),mm6 \ | |
| 291 /*r0=NR0*/ \ | |
| 292 __asm psraw mm0,4 \ | |
| 293 /*Store NR5 at J(5).*/ \ | |
| 294 __asm movq OC_J(5,_y),mm5 \ | |
| 295 /*Store NR7 at J(7).*/ \ | |
| 296 __asm movq OC_J(7,_y),mm7 \ | |
| 297 /*Store NR0 at I(0).*/ \ | |
| 298 __asm movq OC_I(0,_y),mm0 \ | |
| 299 } | |
| 300 | |
| 301 #define OC_MID(_m,_i) [CONSTS+_m+(_i)*8] | |
| 302 #define OC_C(_i) OC_MID(OC_COSINE_OFFSET,_i-1) | |
| 303 #define OC_8 OC_MID(OC_EIGHT_OFFSET,0) | |
| 304 | |
| 305 static void oc_idct8x8_slow(ogg_int16_t _y[64],ogg_int16_t _x[64]){ | |
| 306 int i; | |
| 307 /*This routine accepts an 8x8 matrix, but in partially transposed form. | |
| 308 Every 4x4 block is transposed.*/ | |
| 309 __asm{ | |
| 310 #define CONSTS eax | |
| 311 #define Y edx | |
| 312 #define X ecx | |
| 313 mov CONSTS,offset OC_IDCT_CONSTS | |
| 314 mov Y,_y | |
| 315 mov X,_x | |
| 316 #define OC_I(_k,_y) [(_y)+(_k)*16] | |
| 317 #define OC_J(_k,_y) [(_y)+((_k)-4)*16+8] | |
| 318 OC_ROW_IDCT(Y,X) | |
| 319 OC_TRANSPOSE(Y) | |
| 320 #undef OC_I | |
| 321 #undef OC_J | |
| 322 #define OC_I(_k,_y) [(_y)+(_k)*16+64] | |
| 323 #define OC_J(_k,_y) [(_y)+((_k)-4)*16+72] | |
| 324 OC_ROW_IDCT(Y,X) | |
| 325 OC_TRANSPOSE(Y) | |
| 326 #undef OC_I | |
| 327 #undef OC_J | |
| 328 #define OC_I(_k,_y) [(_y)+(_k)*16] | |
| 329 #define OC_J(_k,_y) OC_I(_k,_y) | |
| 330 OC_COLUMN_IDCT(Y) | |
| 331 #undef OC_I | |
| 332 #undef OC_J | |
| 333 #define OC_I(_k,_y) [(_y)+(_k)*16+8] | |
| 334 #define OC_J(_k,_y) OC_I(_k,_y) | |
| 335 OC_COLUMN_IDCT(Y) | |
| 336 #undef OC_I | |
| 337 #undef OC_J | |
| 338 #undef CONSTS | |
| 339 #undef Y | |
| 340 #undef X | |
| 341 } | |
| 342 if(_x!=_y){ | |
| 343 int i; | |
| 344 __asm pxor mm0,mm0; | |
| 345 for(i=0;i<4;i++){ | |
| 346 ogg_int16_t *x; | |
| 347 x=_x+16*i; | |
| 348 #define X ecx | |
| 349 __asm{ | |
| 350 mov X,x | |
| 351 movq [X+0x00],mm0 | |
| 352 movq [X+0x08],mm0 | |
| 353 movq [X+0x10],mm0 | |
| 354 movq [X+0x18],mm0 | |
| 355 } | |
| 356 #undef X | |
| 357 } | |
| 358 } | |
| 359 } | |
| 360 | |
| 361 /*25 cycles.*/ | |
| 362 #define OC_IDCT_BEGIN_10(_y,_x) __asm{ \ | |
| 363 __asm movq mm2,OC_I(3,_x) \ | |
| 364 __asm nop \ | |
| 365 __asm movq mm6,OC_C(3) \ | |
| 366 __asm movq mm4,mm2 \ | |
| 367 __asm movq mm1,OC_C(5) \ | |
| 368 __asm pmulhw mm4,mm6 \ | |
| 369 __asm movq mm3,OC_I(1,_x) \ | |
| 370 __asm pmulhw mm1,mm2 \ | |
| 371 __asm movq mm0,OC_C(1) \ | |
| 372 __asm paddw mm4,mm2 \ | |
| 373 __asm pxor mm6,mm6 \ | |
| 374 __asm paddw mm2,mm1 \ | |
| 375 __asm movq mm5,OC_I(2,_x) \ | |
| 376 __asm pmulhw mm0,mm3 \ | |
| 377 __asm movq mm1,mm5 \ | |
| 378 __asm paddw mm0,mm3 \ | |
| 379 __asm pmulhw mm3,OC_C(7) \ | |
| 380 __asm psubw mm6,mm2 \ | |
| 381 __asm pmulhw mm5,OC_C(2) \ | |
| 382 __asm psubw mm0,mm4 \ | |
| 383 __asm movq mm7,OC_I(2,_x) \ | |
| 384 __asm paddw mm4,mm4 \ | |
| 385 __asm paddw mm7,mm5 \ | |
| 386 __asm paddw mm4,mm0 \ | |
| 387 __asm pmulhw mm1,OC_C(6) \ | |
| 388 __asm psubw mm3,mm6 \ | |
| 389 __asm movq OC_I(1,_y),mm4 \ | |
| 390 __asm paddw mm6,mm6 \ | |
| 391 __asm movq mm4,OC_C(4) \ | |
| 392 __asm paddw mm6,mm3 \ | |
| 393 __asm movq mm5,mm3 \ | |
| 394 __asm pmulhw mm3,mm4 \ | |
| 395 __asm movq OC_I(2,_y),mm6 \ | |
| 396 __asm movq mm2,mm0 \ | |
| 397 __asm movq mm6,OC_I(0,_x) \ | |
| 398 __asm pmulhw mm0,mm4 \ | |
| 399 __asm paddw mm5,mm3 \ | |
| 400 __asm paddw mm2,mm0 \ | |
| 401 __asm psubw mm5,mm1 \ | |
| 402 __asm pmulhw mm6,mm4 \ | |
| 403 __asm paddw mm6,OC_I(0,_x) \ | |
| 404 __asm paddw mm1,mm1 \ | |
| 405 __asm movq mm4,mm6 \ | |
| 406 __asm paddw mm1,mm5 \ | |
| 407 __asm psubw mm6,mm2 \ | |
| 408 __asm paddw mm2,mm2 \ | |
| 409 __asm movq mm0,OC_I(1,_y) \ | |
| 410 __asm paddw mm2,mm6 \ | |
| 411 __asm psubw mm2,mm1 \ | |
| 412 __asm nop \ | |
| 413 } | |
| 414 | |
| 415 /*25+8=33 cycles.*/ | |
| 416 #define OC_ROW_IDCT_10(_y,_x) __asm{ \ | |
| 417 OC_IDCT_BEGIN_10(_y,_x) \ | |
| 418 /*r3=D'*/ \ | |
| 419 __asm movq mm3,OC_I(2,_y) \ | |
| 420 /*r4=E'=E-G*/ \ | |
| 421 __asm psubw mm4,mm7 \ | |
| 422 /*r1=H'+H'*/ \ | |
| 423 __asm paddw mm1,mm1 \ | |
| 424 /*r7=G+G*/ \ | |
| 425 __asm paddw mm7,mm7 \ | |
| 426 /*r1=R1=A''+H'*/ \ | |
| 427 __asm paddw mm1,mm2 \ | |
| 428 /*r7=G'=E+G*/ \ | |
| 429 __asm paddw mm7,mm4 \ | |
| 430 /*r4=R4=E'-D'*/ \ | |
| 431 __asm psubw mm4,mm3 \ | |
| 432 __asm paddw mm3,mm3 \ | |
| 433 /*r6=R6=F'-B''*/ \ | |
| 434 __asm psubw mm6,mm5 \ | |
| 435 __asm paddw mm5,mm5 \ | |
| 436 /*r3=R3=E'+D'*/ \ | |
| 437 __asm paddw mm3,mm4 \ | |
| 438 /*r5=R5=F'+B''*/ \ | |
| 439 __asm paddw mm5,mm6 \ | |
| 440 /*r7=R7=G'-C'*/ \ | |
| 441 __asm psubw mm7,mm0 \ | |
| 442 __asm paddw mm0,mm0 \ | |
| 443 /*Save R1.*/ \ | |
| 444 __asm movq OC_I(1,_y),mm1 \ | |
| 445 /*r0=R0=G'+C'*/ \ | |
| 446 __asm paddw mm0,mm7 \ | |
| 447 } | |
| 448 | |
| 449 /*25+19=44 cycles'*/ | |
| 450 #define OC_COLUMN_IDCT_10(_y) __asm{ \ | |
| 451 OC_IDCT_BEGIN_10(_y,_y) \ | |
| 452 __asm paddw mm2,OC_8 \ | |
| 453 /*r1=H'+H'*/ \ | |
| 454 __asm paddw mm1,mm1 \ | |
| 455 /*r1=R1=A''+H'*/ \ | |
| 456 __asm paddw mm1,mm2 \ | |
| 457 /*r2=NR2*/ \ | |
| 458 __asm psraw mm2,4 \ | |
| 459 /*r4=E'=E-G*/ \ | |
| 460 __asm psubw mm4,mm7 \ | |
| 461 /*r1=NR1*/ \ | |
| 462 __asm psraw mm1,4 \ | |
| 463 /*r3=D'*/ \ | |
| 464 __asm movq mm3,OC_I(2,_y) \ | |
| 465 /*r7=G+G*/ \ | |
| 466 __asm paddw mm7,mm7 \ | |
| 467 /*Store NR2 at I(2).*/ \ | |
| 468 __asm movq OC_I(2,_y),mm2 \ | |
| 469 /*r7=G'=E+G*/ \ | |
| 470 __asm paddw mm7,mm4 \ | |
| 471 /*Store NR1 at I(1).*/ \ | |
| 472 __asm movq OC_I(1,_y),mm1 \ | |
| 473 /*r4=R4=E'-D'*/ \ | |
| 474 __asm psubw mm4,mm3 \ | |
| 475 __asm paddw mm4,OC_8 \ | |
| 476 /*r3=D'+D'*/ \ | |
| 477 __asm paddw mm3,mm3 \ | |
| 478 /*r3=R3=E'+D'*/ \ | |
| 479 __asm paddw mm3,mm4 \ | |
| 480 /*r4=NR4*/ \ | |
| 481 __asm psraw mm4,4 \ | |
| 482 /*r6=R6=F'-B''*/ \ | |
| 483 __asm psubw mm6,mm5 \ | |
| 484 /*r3=NR3*/ \ | |
| 485 __asm psraw mm3,4 \ | |
| 486 __asm paddw mm6,OC_8 \ | |
| 487 /*r5=B''+B''*/ \ | |
| 488 __asm paddw mm5,mm5 \ | |
| 489 /*r5=R5=F'+B''*/ \ | |
| 490 __asm paddw mm5,mm6 \ | |
| 491 /*r6=NR6*/ \ | |
| 492 __asm psraw mm6,4 \ | |
| 493 /*Store NR4 at J(4).*/ \ | |
| 494 __asm movq OC_J(4,_y),mm4 \ | |
| 495 /*r5=NR5*/ \ | |
| 496 __asm psraw mm5,4 \ | |
| 497 /*Store NR3 at I(3).*/ \ | |
| 498 __asm movq OC_I(3,_y),mm3 \ | |
| 499 /*r7=R7=G'-C'*/ \ | |
| 500 __asm psubw mm7,mm0 \ | |
| 501 __asm paddw mm7,OC_8 \ | |
| 502 /*r0=C'+C'*/ \ | |
| 503 __asm paddw mm0,mm0 \ | |
| 504 /*r0=R0=G'+C'*/ \ | |
| 505 __asm paddw mm0,mm7 \ | |
| 506 /*r7=NR7*/ \ | |
| 507 __asm psraw mm7,4 \ | |
| 508 /*Store NR6 at J(6).*/ \ | |
| 509 __asm movq OC_J(6,_y),mm6 \ | |
| 510 /*r0=NR0*/ \ | |
| 511 __asm psraw mm0,4 \ | |
| 512 /*Store NR5 at J(5).*/ \ | |
| 513 __asm movq OC_J(5,_y),mm5 \ | |
| 514 /*Store NR7 at J(7).*/ \ | |
| 515 __asm movq OC_J(7,_y),mm7 \ | |
| 516 /*Store NR0 at I(0).*/ \ | |
| 517 __asm movq OC_I(0,_y),mm0 \ | |
| 518 } | |
| 519 | |
| 520 static void oc_idct8x8_10(ogg_int16_t _y[64],ogg_int16_t _x[64]){ | |
| 521 __asm{ | |
| 522 #define CONSTS eax | |
| 523 #define Y edx | |
| 524 #define X ecx | |
| 525 mov CONSTS,offset OC_IDCT_CONSTS | |
| 526 mov Y,_y | |
| 527 mov X,_x | |
| 528 #define OC_I(_k,_y) [(_y)+(_k)*16] | |
| 529 #define OC_J(_k,_y) [(_y)+((_k)-4)*16+8] | |
| 530 /*Done with dequant, descramble, and partial transpose. | |
| 531 Now do the iDCT itself.*/ | |
| 532 OC_ROW_IDCT_10(Y,X) | |
| 533 OC_TRANSPOSE(Y) | |
| 534 #undef OC_I | |
| 535 #undef OC_J | |
| 536 #define OC_I(_k,_y) [(_y)+(_k)*16] | |
| 537 #define OC_J(_k,_y) OC_I(_k,_y) | |
| 538 OC_COLUMN_IDCT_10(Y) | |
| 539 #undef OC_I | |
| 540 #undef OC_J | |
| 541 #define OC_I(_k,_y) [(_y)+(_k)*16+8] | |
| 542 #define OC_J(_k,_y) OC_I(_k,_y) | |
| 543 OC_COLUMN_IDCT_10(Y) | |
| 544 #undef OC_I | |
| 545 #undef OC_J | |
| 546 #undef CONSTS | |
| 547 #undef Y | |
| 548 #undef X | |
| 549 } | |
| 550 if(_x!=_y){ | |
| 551 #define X ecx | |
| 552 __asm{ | |
| 553 pxor mm0,mm0; | |
| 554 mov X,_x | |
| 555 movq [X+0x00],mm0 | |
| 556 movq [X+0x10],mm0 | |
| 557 movq [X+0x20],mm0 | |
| 558 movq [X+0x30],mm0 | |
| 559 } | |
| 560 #undef X | |
| 561 } | |
| 562 } | |
| 563 | |
| 564 /*Performs an inverse 8x8 Type-II DCT transform. | |
| 565 The input is assumed to be scaled by a factor of 4 relative to orthonormal | |
| 566 version of the transform.*/ | |
| 567 void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){ | |
| 568 /*_last_zzi is subtly different from an actual count of the number of | |
| 569 coefficients we decoded for this block. | |
| 570 It contains the value of zzi BEFORE the final token in the block was | |
| 571 decoded. | |
| 572 In most cases this is an EOB token (the continuation of an EOB run from a | |
| 573 previous block counts), and so this is the same as the coefficient count. | |
| 574 However, in the case that the last token was NOT an EOB token, but filled | |
| 575 the block up with exactly 64 coefficients, _last_zzi will be less than 64. | |
| 576 Provided the last token was not a pure zero run, the minimum value it can | |
| 577 be is 46, and so that doesn't affect any of the cases in this routine. | |
| 578 However, if the last token WAS a pure zero run of length 63, then _last_zzi | |
| 579 will be 1 while the number of coefficients decoded is 64. | |
| 580 Thus, we will trigger the following special case, where the real | |
| 581 coefficient count would not. | |
| 582 Note also that a zero run of length 64 will give _last_zzi a value of 0, | |
| 583 but we still process the DC coefficient, which might have a non-zero value | |
| 584 due to DC prediction. | |
| 585 Although convoluted, this is arguably the correct behavior: it allows us to | |
| 586 use a smaller transform when the block ends with a long zero run instead | |
| 587 of a normal EOB token. | |
| 588 It could be smarter... multiple separate zero runs at the end of a block | |
| 589 will fool it, but an encoder that generates these really deserves what it | |
| 590 gets. | |
| 591 Needless to say we inherited this approach from VP3.*/ | |
| 592 /*Perform the iDCT.*/ | |
| 593 if(_last_zzi<=10)oc_idct8x8_10(_y,_x); | |
| 594 else oc_idct8x8_slow(_y,_x); | |
| 595 } | |
| 596 | |
| 597 #endif |
