1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libtheora/lib/x86_vc/mmxidct.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,597 @@
1.4 +/********************************************************************
1.5 + * *
1.6 + * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE. *
1.7 + * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS *
1.8 + * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
1.9 + * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. *
1.10 + * *
1.11 + * THE Theora SOURCE CODE IS COPYRIGHT (C) 2002-2009 *
1.12 + * by the Xiph.Org Foundation and contributors http://www.xiph.org/ *
1.13 + * *
1.14 + ********************************************************************
1.15 +
1.16 + function:
1.17 + last mod: $Id: mmxidct.c 17446 2010-09-23 20:06:20Z tterribe $
1.18 +
1.19 + ********************************************************************/
1.20 +
1.21 +/*MMX acceleration of Theora's iDCT.
1.22 + Originally written by Rudolf Marek, based on code from On2's VP3.*/
1.23 +#include "x86int.h"
1.24 +#include "../dct.h"
1.25 +
1.26 +#if defined(OC_X86_ASM)
1.27 +
1.28 +/*These are offsets into the table of constants below.*/
1.29 +/*7 rows of cosines, in order: pi/16 * (1 ... 7).*/
1.30 +#define OC_COSINE_OFFSET (8)
1.31 +/*A row of 8's.*/
1.32 +#define OC_EIGHT_OFFSET (0)
1.33 +
1.34 +
1.35 +
1.36 +/*A table of constants used by the MMX routines.*/
1.37 +static const OC_ALIGN16(ogg_uint16_t) OC_IDCT_CONSTS[(1+7)*4]={
1.38 + 8, 8, 8, 8,
1.39 + (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7,
1.40 + (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7,
1.41 + (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6,
1.42 + (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6,
1.43 + (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5,
1.44 + (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5,
1.45 + (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4,
1.46 + (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4,
1.47 + (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3,
1.48 + (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3,
1.49 + (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2,
1.50 + (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2,
1.51 + (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1,
1.52 + (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1
1.53 +};
1.54 +
1.55 +/*38 cycles*/
1.56 +#define OC_IDCT_BEGIN(_y,_x) __asm{ \
1.57 + __asm movq mm2,OC_I(3,_x) \
1.58 + __asm movq mm6,OC_C(3) \
1.59 + __asm movq mm4,mm2 \
1.60 + __asm movq mm7,OC_J(5,_x) \
1.61 + __asm pmulhw mm4,mm6 \
1.62 + __asm movq mm1,OC_C(5) \
1.63 + __asm pmulhw mm6,mm7 \
1.64 + __asm movq mm5,mm1 \
1.65 + __asm pmulhw mm1,mm2 \
1.66 + __asm movq mm3,OC_I(1,_x) \
1.67 + __asm pmulhw mm5,mm7 \
1.68 + __asm movq mm0,OC_C(1) \
1.69 + __asm paddw mm4,mm2 \
1.70 + __asm paddw mm6,mm7 \
1.71 + __asm paddw mm2,mm1 \
1.72 + __asm movq mm1,OC_J(7,_x) \
1.73 + __asm paddw mm7,mm5 \
1.74 + __asm movq mm5,mm0 \
1.75 + __asm pmulhw mm0,mm3 \
1.76 + __asm paddw mm4,mm7 \
1.77 + __asm pmulhw mm5,mm1 \
1.78 + __asm movq mm7,OC_C(7) \
1.79 + __asm psubw mm6,mm2 \
1.80 + __asm paddw mm0,mm3 \
1.81 + __asm pmulhw mm3,mm7 \
1.82 + __asm movq mm2,OC_I(2,_x) \
1.83 + __asm pmulhw mm7,mm1 \
1.84 + __asm paddw mm5,mm1 \
1.85 + __asm movq mm1,mm2 \
1.86 + __asm pmulhw mm2,OC_C(2) \
1.87 + __asm psubw mm3,mm5 \
1.88 + __asm movq mm5,OC_J(6,_x) \
1.89 + __asm paddw mm0,mm7 \
1.90 + __asm movq mm7,mm5 \
1.91 + __asm psubw mm0,mm4 \
1.92 + __asm pmulhw mm5,OC_C(2) \
1.93 + __asm paddw mm2,mm1 \
1.94 + __asm pmulhw mm1,OC_C(6) \
1.95 + __asm paddw mm4,mm4 \
1.96 + __asm paddw mm4,mm0 \
1.97 + __asm psubw mm3,mm6 \
1.98 + __asm paddw mm5,mm7 \
1.99 + __asm paddw mm6,mm6 \
1.100 + __asm pmulhw mm7,OC_C(6) \
1.101 + __asm paddw mm6,mm3 \
1.102 + __asm movq OC_I(1,_y),mm4 \
1.103 + __asm psubw mm1,mm5 \
1.104 + __asm movq mm4,OC_C(4) \
1.105 + __asm movq mm5,mm3 \
1.106 + __asm pmulhw mm3,mm4 \
1.107 + __asm paddw mm7,mm2 \
1.108 + __asm movq OC_I(2,_y),mm6 \
1.109 + __asm movq mm2,mm0 \
1.110 + __asm movq mm6,OC_I(0,_x) \
1.111 + __asm pmulhw mm0,mm4 \
1.112 + __asm paddw mm5,mm3 \
1.113 + __asm movq mm3,OC_J(4,_x) \
1.114 + __asm psubw mm5,mm1 \
1.115 + __asm paddw mm2,mm0 \
1.116 + __asm psubw mm6,mm3 \
1.117 + __asm movq mm0,mm6 \
1.118 + __asm pmulhw mm6,mm4 \
1.119 + __asm paddw mm3,mm3 \
1.120 + __asm paddw mm1,mm1 \
1.121 + __asm paddw mm3,mm0 \
1.122 + __asm paddw mm1,mm5 \
1.123 + __asm pmulhw mm4,mm3 \
1.124 + __asm paddw mm6,mm0 \
1.125 + __asm psubw mm6,mm2 \
1.126 + __asm paddw mm2,mm2 \
1.127 + __asm movq mm0,OC_I(1,_y) \
1.128 + __asm paddw mm2,mm6 \
1.129 + __asm paddw mm4,mm3 \
1.130 + __asm psubw mm2,mm1 \
1.131 +}
1.132 +
1.133 +/*38+8=46 cycles.*/
1.134 +#define OC_ROW_IDCT(_y,_x) __asm{ \
1.135 + OC_IDCT_BEGIN(_y,_x) \
1.136 + /*r3=D'*/ \
1.137 + __asm movq mm3,OC_I(2,_y) \
1.138 + /*r4=E'=E-G*/ \
1.139 + __asm psubw mm4,mm7 \
1.140 + /*r1=H'+H'*/ \
1.141 + __asm paddw mm1,mm1 \
1.142 + /*r7=G+G*/ \
1.143 + __asm paddw mm7,mm7 \
1.144 + /*r1=R1=A''+H'*/ \
1.145 + __asm paddw mm1,mm2 \
1.146 + /*r7=G'=E+G*/ \
1.147 + __asm paddw mm7,mm4 \
1.148 + /*r4=R4=E'-D'*/ \
1.149 + __asm psubw mm4,mm3 \
1.150 + __asm paddw mm3,mm3 \
1.151 + /*r6=R6=F'-B''*/ \
1.152 + __asm psubw mm6,mm5 \
1.153 + __asm paddw mm5,mm5 \
1.154 + /*r3=R3=E'+D'*/ \
1.155 + __asm paddw mm3,mm4 \
1.156 + /*r5=R5=F'+B''*/ \
1.157 + __asm paddw mm5,mm6 \
1.158 + /*r7=R7=G'-C'*/ \
1.159 + __asm psubw mm7,mm0 \
1.160 + __asm paddw mm0,mm0 \
1.161 + /*Save R1.*/ \
1.162 + __asm movq OC_I(1,_y),mm1 \
1.163 + /*r0=R0=G.+C.*/ \
1.164 + __asm paddw mm0,mm7 \
1.165 +}
1.166 +
1.167 +/*The following macro does two 4x4 transposes in place.
1.168 + At entry, we assume:
1.169 + r0 = a3 a2 a1 a0
1.170 + I(1) = b3 b2 b1 b0
1.171 + r2 = c3 c2 c1 c0
1.172 + r3 = d3 d2 d1 d0
1.173 +
1.174 + r4 = e3 e2 e1 e0
1.175 + r5 = f3 f2 f1 f0
1.176 + r6 = g3 g2 g1 g0
1.177 + r7 = h3 h2 h1 h0
1.178 +
1.179 + At exit, we have:
1.180 + I(0) = d0 c0 b0 a0
1.181 + I(1) = d1 c1 b1 a1
1.182 + I(2) = d2 c2 b2 a2
1.183 + I(3) = d3 c3 b3 a3
1.184 +
1.185 + J(4) = h0 g0 f0 e0
1.186 + J(5) = h1 g1 f1 e1
1.187 + J(6) = h2 g2 f2 e2
1.188 + J(7) = h3 g3 f3 e3
1.189 +
1.190 + I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3.
1.191 + J(4) J(5) J(6) J(7) is the transpose of r4 r5 r6 r7.
1.192 +
1.193 + Since r1 is free at entry, we calculate the Js first.*/
1.194 +/*19 cycles.*/
1.195 +#define OC_TRANSPOSE(_y) __asm{ \
1.196 + __asm movq mm1,mm4 \
1.197 + __asm punpcklwd mm4,mm5 \
1.198 + __asm movq OC_I(0,_y),mm0 \
1.199 + __asm punpckhwd mm1,mm5 \
1.200 + __asm movq mm0,mm6 \
1.201 + __asm punpcklwd mm6,mm7 \
1.202 + __asm movq mm5,mm4 \
1.203 + __asm punpckldq mm4,mm6 \
1.204 + __asm punpckhdq mm5,mm6 \
1.205 + __asm movq mm6,mm1 \
1.206 + __asm movq OC_J(4,_y),mm4 \
1.207 + __asm punpckhwd mm0,mm7 \
1.208 + __asm movq OC_J(5,_y),mm5 \
1.209 + __asm punpckhdq mm6,mm0 \
1.210 + __asm movq mm4,OC_I(0,_y) \
1.211 + __asm punpckldq mm1,mm0 \
1.212 + __asm movq mm5,OC_I(1,_y) \
1.213 + __asm movq mm0,mm4 \
1.214 + __asm movq OC_J(7,_y),mm6 \
1.215 + __asm punpcklwd mm0,mm5 \
1.216 + __asm movq OC_J(6,_y),mm1 \
1.217 + __asm punpckhwd mm4,mm5 \
1.218 + __asm movq mm5,mm2 \
1.219 + __asm punpcklwd mm2,mm3 \
1.220 + __asm movq mm1,mm0 \
1.221 + __asm punpckldq mm0,mm2 \
1.222 + __asm punpckhdq mm1,mm2 \
1.223 + __asm movq mm2,mm4 \
1.224 + __asm movq OC_I(0,_y),mm0 \
1.225 + __asm punpckhwd mm5,mm3 \
1.226 + __asm movq OC_I(1,_y),mm1 \
1.227 + __asm punpckhdq mm4,mm5 \
1.228 + __asm punpckldq mm2,mm5 \
1.229 + __asm movq OC_I(3,_y),mm4 \
1.230 + __asm movq OC_I(2,_y),mm2 \
1.231 +}
1.232 +
1.233 +/*38+19=57 cycles.*/
1.234 +#define OC_COLUMN_IDCT(_y) __asm{ \
1.235 + OC_IDCT_BEGIN(_y,_y) \
1.236 + __asm paddw mm2,OC_8 \
1.237 + /*r1=H'+H'*/ \
1.238 + __asm paddw mm1,mm1 \
1.239 + /*r1=R1=A''+H'*/ \
1.240 + __asm paddw mm1,mm2 \
1.241 + /*r2=NR2*/ \
1.242 + __asm psraw mm2,4 \
1.243 + /*r4=E'=E-G*/ \
1.244 + __asm psubw mm4,mm7 \
1.245 + /*r1=NR1*/ \
1.246 + __asm psraw mm1,4 \
1.247 + /*r3=D'*/ \
1.248 + __asm movq mm3,OC_I(2,_y) \
1.249 + /*r7=G+G*/ \
1.250 + __asm paddw mm7,mm7 \
1.251 + /*Store NR2 at I(2).*/ \
1.252 + __asm movq OC_I(2,_y),mm2 \
1.253 + /*r7=G'=E+G*/ \
1.254 + __asm paddw mm7,mm4 \
1.255 + /*Store NR1 at I(1).*/ \
1.256 + __asm movq OC_I(1,_y),mm1 \
1.257 + /*r4=R4=E'-D'*/ \
1.258 + __asm psubw mm4,mm3 \
1.259 + __asm paddw mm4,OC_8 \
1.260 + /*r3=D'+D'*/ \
1.261 + __asm paddw mm3,mm3 \
1.262 + /*r3=R3=E'+D'*/ \
1.263 + __asm paddw mm3,mm4 \
1.264 + /*r4=NR4*/ \
1.265 + __asm psraw mm4,4 \
1.266 + /*r6=R6=F'-B''*/ \
1.267 + __asm psubw mm6,mm5 \
1.268 + /*r3=NR3*/ \
1.269 + __asm psraw mm3,4 \
1.270 + __asm paddw mm6,OC_8 \
1.271 + /*r5=B''+B''*/ \
1.272 + __asm paddw mm5,mm5 \
1.273 + /*r5=R5=F'+B''*/ \
1.274 + __asm paddw mm5,mm6 \
1.275 + /*r6=NR6*/ \
1.276 + __asm psraw mm6,4 \
1.277 + /*Store NR4 at J(4).*/ \
1.278 + __asm movq OC_J(4,_y),mm4 \
1.279 + /*r5=NR5*/ \
1.280 + __asm psraw mm5,4 \
1.281 + /*Store NR3 at I(3).*/ \
1.282 + __asm movq OC_I(3,_y),mm3 \
1.283 + /*r7=R7=G'-C'*/ \
1.284 + __asm psubw mm7,mm0 \
1.285 + __asm paddw mm7,OC_8 \
1.286 + /*r0=C'+C'*/ \
1.287 + __asm paddw mm0,mm0 \
1.288 + /*r0=R0=G'+C'*/ \
1.289 + __asm paddw mm0,mm7 \
1.290 + /*r7=NR7*/ \
1.291 + __asm psraw mm7,4 \
1.292 + /*Store NR6 at J(6).*/ \
1.293 + __asm movq OC_J(6,_y),mm6 \
1.294 + /*r0=NR0*/ \
1.295 + __asm psraw mm0,4 \
1.296 + /*Store NR5 at J(5).*/ \
1.297 + __asm movq OC_J(5,_y),mm5 \
1.298 + /*Store NR7 at J(7).*/ \
1.299 + __asm movq OC_J(7,_y),mm7 \
1.300 + /*Store NR0 at I(0).*/ \
1.301 + __asm movq OC_I(0,_y),mm0 \
1.302 +}
1.303 +
1.304 +#define OC_MID(_m,_i) [CONSTS+_m+(_i)*8]
1.305 +#define OC_C(_i) OC_MID(OC_COSINE_OFFSET,_i-1)
1.306 +#define OC_8 OC_MID(OC_EIGHT_OFFSET,0)
1.307 +
1.308 +static void oc_idct8x8_slow(ogg_int16_t _y[64],ogg_int16_t _x[64]){
1.309 + int i;
1.310 + /*This routine accepts an 8x8 matrix, but in partially transposed form.
1.311 + Every 4x4 block is transposed.*/
1.312 + __asm{
1.313 +#define CONSTS eax
1.314 +#define Y edx
1.315 +#define X ecx
1.316 + mov CONSTS,offset OC_IDCT_CONSTS
1.317 + mov Y,_y
1.318 + mov X,_x
1.319 +#define OC_I(_k,_y) [(_y)+(_k)*16]
1.320 +#define OC_J(_k,_y) [(_y)+((_k)-4)*16+8]
1.321 + OC_ROW_IDCT(Y,X)
1.322 + OC_TRANSPOSE(Y)
1.323 +#undef OC_I
1.324 +#undef OC_J
1.325 +#define OC_I(_k,_y) [(_y)+(_k)*16+64]
1.326 +#define OC_J(_k,_y) [(_y)+((_k)-4)*16+72]
1.327 + OC_ROW_IDCT(Y,X)
1.328 + OC_TRANSPOSE(Y)
1.329 +#undef OC_I
1.330 +#undef OC_J
1.331 +#define OC_I(_k,_y) [(_y)+(_k)*16]
1.332 +#define OC_J(_k,_y) OC_I(_k,_y)
1.333 + OC_COLUMN_IDCT(Y)
1.334 +#undef OC_I
1.335 +#undef OC_J
1.336 +#define OC_I(_k,_y) [(_y)+(_k)*16+8]
1.337 +#define OC_J(_k,_y) OC_I(_k,_y)
1.338 + OC_COLUMN_IDCT(Y)
1.339 +#undef OC_I
1.340 +#undef OC_J
1.341 +#undef CONSTS
1.342 +#undef Y
1.343 +#undef X
1.344 + }
1.345 + if(_x!=_y){
1.346 + int i;
1.347 + __asm pxor mm0,mm0;
1.348 + for(i=0;i<4;i++){
1.349 + ogg_int16_t *x;
1.350 + x=_x+16*i;
1.351 +#define X ecx
1.352 + __asm{
1.353 + mov X,x
1.354 + movq [X+0x00],mm0
1.355 + movq [X+0x08],mm0
1.356 + movq [X+0x10],mm0
1.357 + movq [X+0x18],mm0
1.358 + }
1.359 +#undef X
1.360 + }
1.361 + }
1.362 +}
1.363 +
1.364 +/*25 cycles.*/
1.365 +#define OC_IDCT_BEGIN_10(_y,_x) __asm{ \
1.366 + __asm movq mm2,OC_I(3,_x) \
1.367 + __asm nop \
1.368 + __asm movq mm6,OC_C(3) \
1.369 + __asm movq mm4,mm2 \
1.370 + __asm movq mm1,OC_C(5) \
1.371 + __asm pmulhw mm4,mm6 \
1.372 + __asm movq mm3,OC_I(1,_x) \
1.373 + __asm pmulhw mm1,mm2 \
1.374 + __asm movq mm0,OC_C(1) \
1.375 + __asm paddw mm4,mm2 \
1.376 + __asm pxor mm6,mm6 \
1.377 + __asm paddw mm2,mm1 \
1.378 + __asm movq mm5,OC_I(2,_x) \
1.379 + __asm pmulhw mm0,mm3 \
1.380 + __asm movq mm1,mm5 \
1.381 + __asm paddw mm0,mm3 \
1.382 + __asm pmulhw mm3,OC_C(7) \
1.383 + __asm psubw mm6,mm2 \
1.384 + __asm pmulhw mm5,OC_C(2) \
1.385 + __asm psubw mm0,mm4 \
1.386 + __asm movq mm7,OC_I(2,_x) \
1.387 + __asm paddw mm4,mm4 \
1.388 + __asm paddw mm7,mm5 \
1.389 + __asm paddw mm4,mm0 \
1.390 + __asm pmulhw mm1,OC_C(6) \
1.391 + __asm psubw mm3,mm6 \
1.392 + __asm movq OC_I(1,_y),mm4 \
1.393 + __asm paddw mm6,mm6 \
1.394 + __asm movq mm4,OC_C(4) \
1.395 + __asm paddw mm6,mm3 \
1.396 + __asm movq mm5,mm3 \
1.397 + __asm pmulhw mm3,mm4 \
1.398 + __asm movq OC_I(2,_y),mm6 \
1.399 + __asm movq mm2,mm0 \
1.400 + __asm movq mm6,OC_I(0,_x) \
1.401 + __asm pmulhw mm0,mm4 \
1.402 + __asm paddw mm5,mm3 \
1.403 + __asm paddw mm2,mm0 \
1.404 + __asm psubw mm5,mm1 \
1.405 + __asm pmulhw mm6,mm4 \
1.406 + __asm paddw mm6,OC_I(0,_x) \
1.407 + __asm paddw mm1,mm1 \
1.408 + __asm movq mm4,mm6 \
1.409 + __asm paddw mm1,mm5 \
1.410 + __asm psubw mm6,mm2 \
1.411 + __asm paddw mm2,mm2 \
1.412 + __asm movq mm0,OC_I(1,_y) \
1.413 + __asm paddw mm2,mm6 \
1.414 + __asm psubw mm2,mm1 \
1.415 + __asm nop \
1.416 +}
1.417 +
1.418 +/*25+8=33 cycles.*/
1.419 +#define OC_ROW_IDCT_10(_y,_x) __asm{ \
1.420 + OC_IDCT_BEGIN_10(_y,_x) \
1.421 + /*r3=D'*/ \
1.422 + __asm movq mm3,OC_I(2,_y) \
1.423 + /*r4=E'=E-G*/ \
1.424 + __asm psubw mm4,mm7 \
1.425 + /*r1=H'+H'*/ \
1.426 + __asm paddw mm1,mm1 \
1.427 + /*r7=G+G*/ \
1.428 + __asm paddw mm7,mm7 \
1.429 + /*r1=R1=A''+H'*/ \
1.430 + __asm paddw mm1,mm2 \
1.431 + /*r7=G'=E+G*/ \
1.432 + __asm paddw mm7,mm4 \
1.433 + /*r4=R4=E'-D'*/ \
1.434 + __asm psubw mm4,mm3 \
1.435 + __asm paddw mm3,mm3 \
1.436 + /*r6=R6=F'-B''*/ \
1.437 + __asm psubw mm6,mm5 \
1.438 + __asm paddw mm5,mm5 \
1.439 + /*r3=R3=E'+D'*/ \
1.440 + __asm paddw mm3,mm4 \
1.441 + /*r5=R5=F'+B''*/ \
1.442 + __asm paddw mm5,mm6 \
1.443 + /*r7=R7=G'-C'*/ \
1.444 + __asm psubw mm7,mm0 \
1.445 + __asm paddw mm0,mm0 \
1.446 + /*Save R1.*/ \
1.447 + __asm movq OC_I(1,_y),mm1 \
1.448 + /*r0=R0=G'+C'*/ \
1.449 + __asm paddw mm0,mm7 \
1.450 +}
1.451 +
1.452 +/*25+19=44 cycles'*/
1.453 +#define OC_COLUMN_IDCT_10(_y) __asm{ \
1.454 + OC_IDCT_BEGIN_10(_y,_y) \
1.455 + __asm paddw mm2,OC_8 \
1.456 + /*r1=H'+H'*/ \
1.457 + __asm paddw mm1,mm1 \
1.458 + /*r1=R1=A''+H'*/ \
1.459 + __asm paddw mm1,mm2 \
1.460 + /*r2=NR2*/ \
1.461 + __asm psraw mm2,4 \
1.462 + /*r4=E'=E-G*/ \
1.463 + __asm psubw mm4,mm7 \
1.464 + /*r1=NR1*/ \
1.465 + __asm psraw mm1,4 \
1.466 + /*r3=D'*/ \
1.467 + __asm movq mm3,OC_I(2,_y) \
1.468 + /*r7=G+G*/ \
1.469 + __asm paddw mm7,mm7 \
1.470 + /*Store NR2 at I(2).*/ \
1.471 + __asm movq OC_I(2,_y),mm2 \
1.472 + /*r7=G'=E+G*/ \
1.473 + __asm paddw mm7,mm4 \
1.474 + /*Store NR1 at I(1).*/ \
1.475 + __asm movq OC_I(1,_y),mm1 \
1.476 + /*r4=R4=E'-D'*/ \
1.477 + __asm psubw mm4,mm3 \
1.478 + __asm paddw mm4,OC_8 \
1.479 + /*r3=D'+D'*/ \
1.480 + __asm paddw mm3,mm3 \
1.481 + /*r3=R3=E'+D'*/ \
1.482 + __asm paddw mm3,mm4 \
1.483 + /*r4=NR4*/ \
1.484 + __asm psraw mm4,4 \
1.485 + /*r6=R6=F'-B''*/ \
1.486 + __asm psubw mm6,mm5 \
1.487 + /*r3=NR3*/ \
1.488 + __asm psraw mm3,4 \
1.489 + __asm paddw mm6,OC_8 \
1.490 + /*r5=B''+B''*/ \
1.491 + __asm paddw mm5,mm5 \
1.492 + /*r5=R5=F'+B''*/ \
1.493 + __asm paddw mm5,mm6 \
1.494 + /*r6=NR6*/ \
1.495 + __asm psraw mm6,4 \
1.496 + /*Store NR4 at J(4).*/ \
1.497 + __asm movq OC_J(4,_y),mm4 \
1.498 + /*r5=NR5*/ \
1.499 + __asm psraw mm5,4 \
1.500 + /*Store NR3 at I(3).*/ \
1.501 + __asm movq OC_I(3,_y),mm3 \
1.502 + /*r7=R7=G'-C'*/ \
1.503 + __asm psubw mm7,mm0 \
1.504 + __asm paddw mm7,OC_8 \
1.505 + /*r0=C'+C'*/ \
1.506 + __asm paddw mm0,mm0 \
1.507 + /*r0=R0=G'+C'*/ \
1.508 + __asm paddw mm0,mm7 \
1.509 + /*r7=NR7*/ \
1.510 + __asm psraw mm7,4 \
1.511 + /*Store NR6 at J(6).*/ \
1.512 + __asm movq OC_J(6,_y),mm6 \
1.513 + /*r0=NR0*/ \
1.514 + __asm psraw mm0,4 \
1.515 + /*Store NR5 at J(5).*/ \
1.516 + __asm movq OC_J(5,_y),mm5 \
1.517 + /*Store NR7 at J(7).*/ \
1.518 + __asm movq OC_J(7,_y),mm7 \
1.519 + /*Store NR0 at I(0).*/ \
1.520 + __asm movq OC_I(0,_y),mm0 \
1.521 +}
1.522 +
1.523 +static void oc_idct8x8_10(ogg_int16_t _y[64],ogg_int16_t _x[64]){
1.524 + __asm{
1.525 +#define CONSTS eax
1.526 +#define Y edx
1.527 +#define X ecx
1.528 + mov CONSTS,offset OC_IDCT_CONSTS
1.529 + mov Y,_y
1.530 + mov X,_x
1.531 +#define OC_I(_k,_y) [(_y)+(_k)*16]
1.532 +#define OC_J(_k,_y) [(_y)+((_k)-4)*16+8]
1.533 + /*Done with dequant, descramble, and partial transpose.
1.534 + Now do the iDCT itself.*/
1.535 + OC_ROW_IDCT_10(Y,X)
1.536 + OC_TRANSPOSE(Y)
1.537 +#undef OC_I
1.538 +#undef OC_J
1.539 +#define OC_I(_k,_y) [(_y)+(_k)*16]
1.540 +#define OC_J(_k,_y) OC_I(_k,_y)
1.541 + OC_COLUMN_IDCT_10(Y)
1.542 +#undef OC_I
1.543 +#undef OC_J
1.544 +#define OC_I(_k,_y) [(_y)+(_k)*16+8]
1.545 +#define OC_J(_k,_y) OC_I(_k,_y)
1.546 + OC_COLUMN_IDCT_10(Y)
1.547 +#undef OC_I
1.548 +#undef OC_J
1.549 +#undef CONSTS
1.550 +#undef Y
1.551 +#undef X
1.552 + }
1.553 + if(_x!=_y){
1.554 +#define X ecx
1.555 + __asm{
1.556 + pxor mm0,mm0;
1.557 + mov X,_x
1.558 + movq [X+0x00],mm0
1.559 + movq [X+0x10],mm0
1.560 + movq [X+0x20],mm0
1.561 + movq [X+0x30],mm0
1.562 + }
1.563 +#undef X
1.564 + }
1.565 +}
1.566 +
1.567 +/*Performs an inverse 8x8 Type-II DCT transform.
1.568 + The input is assumed to be scaled by a factor of 4 relative to orthonormal
1.569 + version of the transform.*/
1.570 +void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){
1.571 + /*_last_zzi is subtly different from an actual count of the number of
1.572 + coefficients we decoded for this block.
1.573 + It contains the value of zzi BEFORE the final token in the block was
1.574 + decoded.
1.575 + In most cases this is an EOB token (the continuation of an EOB run from a
1.576 + previous block counts), and so this is the same as the coefficient count.
1.577 + However, in the case that the last token was NOT an EOB token, but filled
1.578 + the block up with exactly 64 coefficients, _last_zzi will be less than 64.
1.579 + Provided the last token was not a pure zero run, the minimum value it can
1.580 + be is 46, and so that doesn't affect any of the cases in this routine.
1.581 + However, if the last token WAS a pure zero run of length 63, then _last_zzi
1.582 + will be 1 while the number of coefficients decoded is 64.
1.583 + Thus, we will trigger the following special case, where the real
1.584 + coefficient count would not.
1.585 + Note also that a zero run of length 64 will give _last_zzi a value of 0,
1.586 + but we still process the DC coefficient, which might have a non-zero value
1.587 + due to DC prediction.
1.588 + Although convoluted, this is arguably the correct behavior: it allows us to
1.589 + use a smaller transform when the block ends with a long zero run instead
1.590 + of a normal EOB token.
1.591 + It could be smarter... multiple separate zero runs at the end of a block
1.592 + will fool it, but an encoder that generates these really deserves what it
1.593 + gets.
1.594 + Needless to say we inherited this approach from VP3.*/
1.595 + /*Perform the iDCT.*/
1.596 + if(_last_zzi<=10)oc_idct8x8_10(_y,_x);
1.597 + else oc_idct8x8_slow(_y,_x);
1.598 +}
1.599 +
1.600 +#endif
