1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libtheora/lib/x86/mmxidct.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,562 @@
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 (0)
1.31 +/*A row of 8's.*/
1.32 +#define OC_EIGHT_OFFSET (56)
1.33 +
1.34 +
1.35 +
1.36 +/*38 cycles*/
1.37 +#define OC_IDCT_BEGIN(_y,_x) \
1.38 + "#OC_IDCT_BEGIN\n\t" \
1.39 + "movq "OC_I(3,_x)",%%mm2\n\t" \
1.40 + "movq "OC_MEM_OFFS(0x30,c)",%%mm6\n\t" \
1.41 + "movq %%mm2,%%mm4\n\t" \
1.42 + "movq "OC_J(5,_x)",%%mm7\n\t" \
1.43 + "pmulhw %%mm6,%%mm4\n\t" \
1.44 + "movq "OC_MEM_OFFS(0x50,c)",%%mm1\n\t" \
1.45 + "pmulhw %%mm7,%%mm6\n\t" \
1.46 + "movq %%mm1,%%mm5\n\t" \
1.47 + "pmulhw %%mm2,%%mm1\n\t" \
1.48 + "movq "OC_I(1,_x)",%%mm3\n\t" \
1.49 + "pmulhw %%mm7,%%mm5\n\t" \
1.50 + "movq "OC_MEM_OFFS(0x10,c)",%%mm0\n\t" \
1.51 + "paddw %%mm2,%%mm4\n\t" \
1.52 + "paddw %%mm7,%%mm6\n\t" \
1.53 + "paddw %%mm1,%%mm2\n\t" \
1.54 + "movq "OC_J(7,_x)",%%mm1\n\t" \
1.55 + "paddw %%mm5,%%mm7\n\t" \
1.56 + "movq %%mm0,%%mm5\n\t" \
1.57 + "pmulhw %%mm3,%%mm0\n\t" \
1.58 + "paddw %%mm7,%%mm4\n\t" \
1.59 + "pmulhw %%mm1,%%mm5\n\t" \
1.60 + "movq "OC_MEM_OFFS(0x70,c)",%%mm7\n\t" \
1.61 + "psubw %%mm2,%%mm6\n\t" \
1.62 + "paddw %%mm3,%%mm0\n\t" \
1.63 + "pmulhw %%mm7,%%mm3\n\t" \
1.64 + "movq "OC_I(2,_x)",%%mm2\n\t" \
1.65 + "pmulhw %%mm1,%%mm7\n\t" \
1.66 + "paddw %%mm1,%%mm5\n\t" \
1.67 + "movq %%mm2,%%mm1\n\t" \
1.68 + "pmulhw "OC_MEM_OFFS(0x20,c)",%%mm2\n\t" \
1.69 + "psubw %%mm5,%%mm3\n\t" \
1.70 + "movq "OC_J(6,_x)",%%mm5\n\t" \
1.71 + "paddw %%mm7,%%mm0\n\t" \
1.72 + "movq %%mm5,%%mm7\n\t" \
1.73 + "psubw %%mm4,%%mm0\n\t" \
1.74 + "pmulhw "OC_MEM_OFFS(0x20,c)",%%mm5\n\t" \
1.75 + "paddw %%mm1,%%mm2\n\t" \
1.76 + "pmulhw "OC_MEM_OFFS(0x60,c)",%%mm1\n\t" \
1.77 + "paddw %%mm4,%%mm4\n\t" \
1.78 + "paddw %%mm0,%%mm4\n\t" \
1.79 + "psubw %%mm6,%%mm3\n\t" \
1.80 + "paddw %%mm7,%%mm5\n\t" \
1.81 + "paddw %%mm6,%%mm6\n\t" \
1.82 + "pmulhw "OC_MEM_OFFS(0x60,c)",%%mm7\n\t" \
1.83 + "paddw %%mm3,%%mm6\n\t" \
1.84 + "movq %%mm4,"OC_I(1,_y)"\n\t" \
1.85 + "psubw %%mm5,%%mm1\n\t" \
1.86 + "movq "OC_MEM_OFFS(0x40,c)",%%mm4\n\t" \
1.87 + "movq %%mm3,%%mm5\n\t" \
1.88 + "pmulhw %%mm4,%%mm3\n\t" \
1.89 + "paddw %%mm2,%%mm7\n\t" \
1.90 + "movq %%mm6,"OC_I(2,_y)"\n\t" \
1.91 + "movq %%mm0,%%mm2\n\t" \
1.92 + "movq "OC_I(0,_x)",%%mm6\n\t" \
1.93 + "pmulhw %%mm4,%%mm0\n\t" \
1.94 + "paddw %%mm3,%%mm5\n\t" \
1.95 + "movq "OC_J(4,_x)",%%mm3\n\t" \
1.96 + "psubw %%mm1,%%mm5\n\t" \
1.97 + "paddw %%mm0,%%mm2\n\t" \
1.98 + "psubw %%mm3,%%mm6\n\t" \
1.99 + "movq %%mm6,%%mm0\n\t" \
1.100 + "pmulhw %%mm4,%%mm6\n\t" \
1.101 + "paddw %%mm3,%%mm3\n\t" \
1.102 + "paddw %%mm1,%%mm1\n\t" \
1.103 + "paddw %%mm0,%%mm3\n\t" \
1.104 + "paddw %%mm5,%%mm1\n\t" \
1.105 + "pmulhw %%mm3,%%mm4\n\t" \
1.106 + "paddw %%mm0,%%mm6\n\t" \
1.107 + "psubw %%mm2,%%mm6\n\t" \
1.108 + "paddw %%mm2,%%mm2\n\t" \
1.109 + "movq "OC_I(1,_y)",%%mm0\n\t" \
1.110 + "paddw %%mm6,%%mm2\n\t" \
1.111 + "paddw %%mm3,%%mm4\n\t" \
1.112 + "psubw %%mm1,%%mm2\n\t" \
1.113 + "#end OC_IDCT_BEGIN\n\t" \
1.114 +
1.115 +/*38+8=46 cycles.*/
1.116 +#define OC_ROW_IDCT(_y,_x) \
1.117 + "#OC_ROW_IDCT\n" \
1.118 + OC_IDCT_BEGIN(_y,_x) \
1.119 + /*r3=D'*/ \
1.120 + "movq "OC_I(2,_y)",%%mm3\n\t" \
1.121 + /*r4=E'=E-G*/ \
1.122 + "psubw %%mm7,%%mm4\n\t" \
1.123 + /*r1=H'+H'*/ \
1.124 + "paddw %%mm1,%%mm1\n\t" \
1.125 + /*r7=G+G*/ \
1.126 + "paddw %%mm7,%%mm7\n\t" \
1.127 + /*r1=R1=A''+H'*/ \
1.128 + "paddw %%mm2,%%mm1\n\t" \
1.129 + /*r7=G'=E+G*/ \
1.130 + "paddw %%mm4,%%mm7\n\t" \
1.131 + /*r4=R4=E'-D'*/ \
1.132 + "psubw %%mm3,%%mm4\n\t" \
1.133 + "paddw %%mm3,%%mm3\n\t" \
1.134 + /*r6=R6=F'-B''*/ \
1.135 + "psubw %%mm5,%%mm6\n\t" \
1.136 + "paddw %%mm5,%%mm5\n\t" \
1.137 + /*r3=R3=E'+D'*/ \
1.138 + "paddw %%mm4,%%mm3\n\t" \
1.139 + /*r5=R5=F'+B''*/ \
1.140 + "paddw %%mm6,%%mm5\n\t" \
1.141 + /*r7=R7=G'-C'*/ \
1.142 + "psubw %%mm0,%%mm7\n\t" \
1.143 + "paddw %%mm0,%%mm0\n\t" \
1.144 + /*Save R1.*/ \
1.145 + "movq %%mm1,"OC_I(1,_y)"\n\t" \
1.146 + /*r0=R0=G.+C.*/ \
1.147 + "paddw %%mm7,%%mm0\n\t" \
1.148 + "#end OC_ROW_IDCT\n\t" \
1.149 +
1.150 +/*The following macro does two 4x4 transposes in place.
1.151 + At entry, we assume:
1.152 + r0 = a3 a2 a1 a0
1.153 + I(1) = b3 b2 b1 b0
1.154 + r2 = c3 c2 c1 c0
1.155 + r3 = d3 d2 d1 d0
1.156 +
1.157 + r4 = e3 e2 e1 e0
1.158 + r5 = f3 f2 f1 f0
1.159 + r6 = g3 g2 g1 g0
1.160 + r7 = h3 h2 h1 h0
1.161 +
1.162 + At exit, we have:
1.163 + I(0) = d0 c0 b0 a0
1.164 + I(1) = d1 c1 b1 a1
1.165 + I(2) = d2 c2 b2 a2
1.166 + I(3) = d3 c3 b3 a3
1.167 +
1.168 + J(4) = h0 g0 f0 e0
1.169 + J(5) = h1 g1 f1 e1
1.170 + J(6) = h2 g2 f2 e2
1.171 + J(7) = h3 g3 f3 e3
1.172 +
1.173 + I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3.
1.174 + J(4) J(5) J(6) J(7) is the transpose of r4 r5 r6 r7.
1.175 +
1.176 + Since r1 is free at entry, we calculate the Js first.*/
1.177 +/*19 cycles.*/
1.178 +#define OC_TRANSPOSE(_y) \
1.179 + "#OC_TRANSPOSE\n\t" \
1.180 + "movq %%mm4,%%mm1\n\t" \
1.181 + "punpcklwd %%mm5,%%mm4\n\t" \
1.182 + "movq %%mm0,"OC_I(0,_y)"\n\t" \
1.183 + "punpckhwd %%mm5,%%mm1\n\t" \
1.184 + "movq %%mm6,%%mm0\n\t" \
1.185 + "punpcklwd %%mm7,%%mm6\n\t" \
1.186 + "movq %%mm4,%%mm5\n\t" \
1.187 + "punpckldq %%mm6,%%mm4\n\t" \
1.188 + "punpckhdq %%mm6,%%mm5\n\t" \
1.189 + "movq %%mm1,%%mm6\n\t" \
1.190 + "movq %%mm4,"OC_J(4,_y)"\n\t" \
1.191 + "punpckhwd %%mm7,%%mm0\n\t" \
1.192 + "movq %%mm5,"OC_J(5,_y)"\n\t" \
1.193 + "punpckhdq %%mm0,%%mm6\n\t" \
1.194 + "movq "OC_I(0,_y)",%%mm4\n\t" \
1.195 + "punpckldq %%mm0,%%mm1\n\t" \
1.196 + "movq "OC_I(1,_y)",%%mm5\n\t" \
1.197 + "movq %%mm4,%%mm0\n\t" \
1.198 + "movq %%mm6,"OC_J(7,_y)"\n\t" \
1.199 + "punpcklwd %%mm5,%%mm0\n\t" \
1.200 + "movq %%mm1,"OC_J(6,_y)"\n\t" \
1.201 + "punpckhwd %%mm5,%%mm4\n\t" \
1.202 + "movq %%mm2,%%mm5\n\t" \
1.203 + "punpcklwd %%mm3,%%mm2\n\t" \
1.204 + "movq %%mm0,%%mm1\n\t" \
1.205 + "punpckldq %%mm2,%%mm0\n\t" \
1.206 + "punpckhdq %%mm2,%%mm1\n\t" \
1.207 + "movq %%mm4,%%mm2\n\t" \
1.208 + "movq %%mm0,"OC_I(0,_y)"\n\t" \
1.209 + "punpckhwd %%mm3,%%mm5\n\t" \
1.210 + "movq %%mm1,"OC_I(1,_y)"\n\t" \
1.211 + "punpckhdq %%mm5,%%mm4\n\t" \
1.212 + "punpckldq %%mm5,%%mm2\n\t" \
1.213 + "movq %%mm4,"OC_I(3,_y)"\n\t" \
1.214 + "movq %%mm2,"OC_I(2,_y)"\n\t" \
1.215 + "#end OC_TRANSPOSE\n\t" \
1.216 +
1.217 +/*38+19=57 cycles.*/
1.218 +#define OC_COLUMN_IDCT(_y) \
1.219 + "#OC_COLUMN_IDCT\n" \
1.220 + OC_IDCT_BEGIN(_y,_y) \
1.221 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm2\n\t" \
1.222 + /*r1=H'+H'*/ \
1.223 + "paddw %%mm1,%%mm1\n\t" \
1.224 + /*r1=R1=A''+H'*/ \
1.225 + "paddw %%mm2,%%mm1\n\t" \
1.226 + /*r2=NR2*/ \
1.227 + "psraw $4,%%mm2\n\t" \
1.228 + /*r4=E'=E-G*/ \
1.229 + "psubw %%mm7,%%mm4\n\t" \
1.230 + /*r1=NR1*/ \
1.231 + "psraw $4,%%mm1\n\t" \
1.232 + /*r3=D'*/ \
1.233 + "movq "OC_I(2,_y)",%%mm3\n\t" \
1.234 + /*r7=G+G*/ \
1.235 + "paddw %%mm7,%%mm7\n\t" \
1.236 + /*Store NR2 at I(2).*/ \
1.237 + "movq %%mm2,"OC_I(2,_y)"\n\t" \
1.238 + /*r7=G'=E+G*/ \
1.239 + "paddw %%mm4,%%mm7\n\t" \
1.240 + /*Store NR1 at I(1).*/ \
1.241 + "movq %%mm1,"OC_I(1,_y)"\n\t" \
1.242 + /*r4=R4=E'-D'*/ \
1.243 + "psubw %%mm3,%%mm4\n\t" \
1.244 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm4\n\t" \
1.245 + /*r3=D'+D'*/ \
1.246 + "paddw %%mm3,%%mm3\n\t" \
1.247 + /*r3=R3=E'+D'*/ \
1.248 + "paddw %%mm4,%%mm3\n\t" \
1.249 + /*r4=NR4*/ \
1.250 + "psraw $4,%%mm4\n\t" \
1.251 + /*r6=R6=F'-B''*/ \
1.252 + "psubw %%mm5,%%mm6\n\t" \
1.253 + /*r3=NR3*/ \
1.254 + "psraw $4,%%mm3\n\t" \
1.255 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm6\n\t" \
1.256 + /*r5=B''+B''*/ \
1.257 + "paddw %%mm5,%%mm5\n\t" \
1.258 + /*r5=R5=F'+B''*/ \
1.259 + "paddw %%mm6,%%mm5\n\t" \
1.260 + /*r6=NR6*/ \
1.261 + "psraw $4,%%mm6\n\t" \
1.262 + /*Store NR4 at J(4).*/ \
1.263 + "movq %%mm4,"OC_J(4,_y)"\n\t" \
1.264 + /*r5=NR5*/ \
1.265 + "psraw $4,%%mm5\n\t" \
1.266 + /*Store NR3 at I(3).*/ \
1.267 + "movq %%mm3,"OC_I(3,_y)"\n\t" \
1.268 + /*r7=R7=G'-C'*/ \
1.269 + "psubw %%mm0,%%mm7\n\t" \
1.270 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm7\n\t" \
1.271 + /*r0=C'+C'*/ \
1.272 + "paddw %%mm0,%%mm0\n\t" \
1.273 + /*r0=R0=G'+C'*/ \
1.274 + "paddw %%mm7,%%mm0\n\t" \
1.275 + /*r7=NR7*/ \
1.276 + "psraw $4,%%mm7\n\t" \
1.277 + /*Store NR6 at J(6).*/ \
1.278 + "movq %%mm6,"OC_J(6,_y)"\n\t" \
1.279 + /*r0=NR0*/ \
1.280 + "psraw $4,%%mm0\n\t" \
1.281 + /*Store NR5 at J(5).*/ \
1.282 + "movq %%mm5,"OC_J(5,_y)"\n\t" \
1.283 + /*Store NR7 at J(7).*/ \
1.284 + "movq %%mm7,"OC_J(7,_y)"\n\t" \
1.285 + /*Store NR0 at I(0).*/ \
1.286 + "movq %%mm0,"OC_I(0,_y)"\n\t" \
1.287 + "#end OC_COLUMN_IDCT\n\t" \
1.288 +
1.289 +static void oc_idct8x8_slow_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64]){
1.290 + /*This routine accepts an 8x8 matrix, but in partially transposed form.
1.291 + Every 4x4 block is transposed.*/
1.292 + __asm__ __volatile__(
1.293 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16,_y)
1.294 +#define OC_J(_k,_y) OC_MEM_OFFS(((_k)-4)*16+8,_y)
1.295 + OC_ROW_IDCT(y,x)
1.296 + OC_TRANSPOSE(y)
1.297 +#undef OC_I
1.298 +#undef OC_J
1.299 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16+64,_y)
1.300 +#define OC_J(_k,_y) OC_MEM_OFFS(((_k)-4)*16+72,_y)
1.301 + OC_ROW_IDCT(y,x)
1.302 + OC_TRANSPOSE(y)
1.303 +#undef OC_I
1.304 +#undef OC_J
1.305 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16,_y)
1.306 +#define OC_J(_k,_y) OC_I(_k,_y)
1.307 + OC_COLUMN_IDCT(y)
1.308 +#undef OC_I
1.309 +#undef OC_J
1.310 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16+8,_y)
1.311 +#define OC_J(_k,_y) OC_I(_k,_y)
1.312 + OC_COLUMN_IDCT(y)
1.313 +#undef OC_I
1.314 +#undef OC_J
1.315 + :[y]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_y,64)
1.316 + :[x]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,_x,64),
1.317 + [c]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,OC_IDCT_CONSTS,128)
1.318 + );
1.319 + if(_x!=_y){
1.320 + int i;
1.321 + __asm__ __volatile__("pxor %%mm0,%%mm0\n\t"::);
1.322 + for(i=0;i<4;i++){
1.323 + __asm__ __volatile__(
1.324 + "movq %%mm0,"OC_MEM_OFFS(0x00,x)"\n\t"
1.325 + "movq %%mm0,"OC_MEM_OFFS(0x08,x)"\n\t"
1.326 + "movq %%mm0,"OC_MEM_OFFS(0x10,x)"\n\t"
1.327 + "movq %%mm0,"OC_MEM_OFFS(0x18,x)"\n\t"
1.328 + :[x]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_x+16*i,16)
1.329 + );
1.330 + }
1.331 + }
1.332 +}
1.333 +
1.334 +/*25 cycles.*/
1.335 +#define OC_IDCT_BEGIN_10(_y,_x) \
1.336 + "#OC_IDCT_BEGIN_10\n\t" \
1.337 + "movq "OC_I(3,_x)",%%mm2\n\t" \
1.338 + "nop\n\t" \
1.339 + "movq "OC_MEM_OFFS(0x30,c)",%%mm6\n\t" \
1.340 + "movq %%mm2,%%mm4\n\t" \
1.341 + "movq "OC_MEM_OFFS(0x50,c)",%%mm1\n\t" \
1.342 + "pmulhw %%mm6,%%mm4\n\t" \
1.343 + "movq "OC_I(1,_x)",%%mm3\n\t" \
1.344 + "pmulhw %%mm2,%%mm1\n\t" \
1.345 + "movq "OC_MEM_OFFS(0x10,c)",%%mm0\n\t" \
1.346 + "paddw %%mm2,%%mm4\n\t" \
1.347 + "pxor %%mm6,%%mm6\n\t" \
1.348 + "paddw %%mm1,%%mm2\n\t" \
1.349 + "movq "OC_I(2,_x)",%%mm5\n\t" \
1.350 + "pmulhw %%mm3,%%mm0\n\t" \
1.351 + "movq %%mm5,%%mm1\n\t" \
1.352 + "paddw %%mm3,%%mm0\n\t" \
1.353 + "pmulhw "OC_MEM_OFFS(0x70,c)",%%mm3\n\t" \
1.354 + "psubw %%mm2,%%mm6\n\t" \
1.355 + "pmulhw "OC_MEM_OFFS(0x20,c)",%%mm5\n\t" \
1.356 + "psubw %%mm4,%%mm0\n\t" \
1.357 + "movq "OC_I(2,_x)",%%mm7\n\t" \
1.358 + "paddw %%mm4,%%mm4\n\t" \
1.359 + "paddw %%mm5,%%mm7\n\t" \
1.360 + "paddw %%mm0,%%mm4\n\t" \
1.361 + "pmulhw "OC_MEM_OFFS(0x60,c)",%%mm1\n\t" \
1.362 + "psubw %%mm6,%%mm3\n\t" \
1.363 + "movq %%mm4,"OC_I(1,_y)"\n\t" \
1.364 + "paddw %%mm6,%%mm6\n\t" \
1.365 + "movq "OC_MEM_OFFS(0x40,c)",%%mm4\n\t" \
1.366 + "paddw %%mm3,%%mm6\n\t" \
1.367 + "movq %%mm3,%%mm5\n\t" \
1.368 + "pmulhw %%mm4,%%mm3\n\t" \
1.369 + "movq %%mm6,"OC_I(2,_y)"\n\t" \
1.370 + "movq %%mm0,%%mm2\n\t" \
1.371 + "movq "OC_I(0,_x)",%%mm6\n\t" \
1.372 + "pmulhw %%mm4,%%mm0\n\t" \
1.373 + "paddw %%mm3,%%mm5\n\t" \
1.374 + "paddw %%mm0,%%mm2\n\t" \
1.375 + "psubw %%mm1,%%mm5\n\t" \
1.376 + "pmulhw %%mm4,%%mm6\n\t" \
1.377 + "paddw "OC_I(0,_x)",%%mm6\n\t" \
1.378 + "paddw %%mm1,%%mm1\n\t" \
1.379 + "movq %%mm6,%%mm4\n\t" \
1.380 + "paddw %%mm5,%%mm1\n\t" \
1.381 + "psubw %%mm2,%%mm6\n\t" \
1.382 + "paddw %%mm2,%%mm2\n\t" \
1.383 + "movq "OC_I(1,_y)",%%mm0\n\t" \
1.384 + "paddw %%mm6,%%mm2\n\t" \
1.385 + "psubw %%mm1,%%mm2\n\t" \
1.386 + "nop\n\t" \
1.387 + "#end OC_IDCT_BEGIN_10\n\t" \
1.388 +
1.389 +/*25+8=33 cycles.*/
1.390 +#define OC_ROW_IDCT_10(_y,_x) \
1.391 + "#OC_ROW_IDCT_10\n\t" \
1.392 + OC_IDCT_BEGIN_10(_y,_x) \
1.393 + /*r3=D'*/ \
1.394 + "movq "OC_I(2,_y)",%%mm3\n\t" \
1.395 + /*r4=E'=E-G*/ \
1.396 + "psubw %%mm7,%%mm4\n\t" \
1.397 + /*r1=H'+H'*/ \
1.398 + "paddw %%mm1,%%mm1\n\t" \
1.399 + /*r7=G+G*/ \
1.400 + "paddw %%mm7,%%mm7\n\t" \
1.401 + /*r1=R1=A''+H'*/ \
1.402 + "paddw %%mm2,%%mm1\n\t" \
1.403 + /*r7=G'=E+G*/ \
1.404 + "paddw %%mm4,%%mm7\n\t" \
1.405 + /*r4=R4=E'-D'*/ \
1.406 + "psubw %%mm3,%%mm4\n\t" \
1.407 + "paddw %%mm3,%%mm3\n\t" \
1.408 + /*r6=R6=F'-B''*/ \
1.409 + "psubw %%mm5,%%mm6\n\t" \
1.410 + "paddw %%mm5,%%mm5\n\t" \
1.411 + /*r3=R3=E'+D'*/ \
1.412 + "paddw %%mm4,%%mm3\n\t" \
1.413 + /*r5=R5=F'+B''*/ \
1.414 + "paddw %%mm6,%%mm5\n\t" \
1.415 + /*r7=R7=G'-C'*/ \
1.416 + "psubw %%mm0,%%mm7\n\t" \
1.417 + "paddw %%mm0,%%mm0\n\t" \
1.418 + /*Save R1.*/ \
1.419 + "movq %%mm1,"OC_I(1,_y)"\n\t" \
1.420 + /*r0=R0=G'+C'*/ \
1.421 + "paddw %%mm7,%%mm0\n\t" \
1.422 + "#end OC_ROW_IDCT_10\n\t" \
1.423 +
1.424 +/*25+19=44 cycles'*/
1.425 +#define OC_COLUMN_IDCT_10(_y) \
1.426 + "#OC_COLUMN_IDCT_10\n\t" \
1.427 + OC_IDCT_BEGIN_10(_y,_y) \
1.428 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm2\n\t" \
1.429 + /*r1=H'+H'*/ \
1.430 + "paddw %%mm1,%%mm1\n\t" \
1.431 + /*r1=R1=A''+H'*/ \
1.432 + "paddw %%mm2,%%mm1\n\t" \
1.433 + /*r2=NR2*/ \
1.434 + "psraw $4,%%mm2\n\t" \
1.435 + /*r4=E'=E-G*/ \
1.436 + "psubw %%mm7,%%mm4\n\t" \
1.437 + /*r1=NR1*/ \
1.438 + "psraw $4,%%mm1\n\t" \
1.439 + /*r3=D'*/ \
1.440 + "movq "OC_I(2,_y)",%%mm3\n\t" \
1.441 + /*r7=G+G*/ \
1.442 + "paddw %%mm7,%%mm7\n\t" \
1.443 + /*Store NR2 at I(2).*/ \
1.444 + "movq %%mm2,"OC_I(2,_y)"\n\t" \
1.445 + /*r7=G'=E+G*/ \
1.446 + "paddw %%mm4,%%mm7\n\t" \
1.447 + /*Store NR1 at I(1).*/ \
1.448 + "movq %%mm1,"OC_I(1,_y)"\n\t" \
1.449 + /*r4=R4=E'-D'*/ \
1.450 + "psubw %%mm3,%%mm4\n\t" \
1.451 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm4\n\t" \
1.452 + /*r3=D'+D'*/ \
1.453 + "paddw %%mm3,%%mm3\n\t" \
1.454 + /*r3=R3=E'+D'*/ \
1.455 + "paddw %%mm4,%%mm3\n\t" \
1.456 + /*r4=NR4*/ \
1.457 + "psraw $4,%%mm4\n\t" \
1.458 + /*r6=R6=F'-B''*/ \
1.459 + "psubw %%mm5,%%mm6\n\t" \
1.460 + /*r3=NR3*/ \
1.461 + "psraw $4,%%mm3\n\t" \
1.462 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm6\n\t" \
1.463 + /*r5=B''+B''*/ \
1.464 + "paddw %%mm5,%%mm5\n\t" \
1.465 + /*r5=R5=F'+B''*/ \
1.466 + "paddw %%mm6,%%mm5\n\t" \
1.467 + /*r6=NR6*/ \
1.468 + "psraw $4,%%mm6\n\t" \
1.469 + /*Store NR4 at J(4).*/ \
1.470 + "movq %%mm4,"OC_J(4,_y)"\n\t" \
1.471 + /*r5=NR5*/ \
1.472 + "psraw $4,%%mm5\n\t" \
1.473 + /*Store NR3 at I(3).*/ \
1.474 + "movq %%mm3,"OC_I(3,_y)"\n\t" \
1.475 + /*r7=R7=G'-C'*/ \
1.476 + "psubw %%mm0,%%mm7\n\t" \
1.477 + "paddw "OC_MEM_OFFS(0x00,c)",%%mm7\n\t" \
1.478 + /*r0=C'+C'*/ \
1.479 + "paddw %%mm0,%%mm0\n\t" \
1.480 + /*r0=R0=G'+C'*/ \
1.481 + "paddw %%mm7,%%mm0\n\t" \
1.482 + /*r7=NR7*/ \
1.483 + "psraw $4,%%mm7\n\t" \
1.484 + /*Store NR6 at J(6).*/ \
1.485 + "movq %%mm6,"OC_J(6,_y)"\n\t" \
1.486 + /*r0=NR0*/ \
1.487 + "psraw $4,%%mm0\n\t" \
1.488 + /*Store NR5 at J(5).*/ \
1.489 + "movq %%mm5,"OC_J(5,_y)"\n\t" \
1.490 + /*Store NR7 at J(7).*/ \
1.491 + "movq %%mm7,"OC_J(7,_y)"\n\t" \
1.492 + /*Store NR0 at I(0).*/ \
1.493 + "movq %%mm0,"OC_I(0,_y)"\n\t" \
1.494 + "#end OC_COLUMN_IDCT_10\n\t" \
1.495 +
1.496 +static void oc_idct8x8_10_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64]){
1.497 + __asm__ __volatile__(
1.498 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16,_y)
1.499 +#define OC_J(_k,_y) OC_MEM_OFFS(((_k)-4)*16+8,_y)
1.500 + /*Done with dequant, descramble, and partial transpose.
1.501 + Now do the iDCT itself.*/
1.502 + OC_ROW_IDCT_10(y,x)
1.503 + OC_TRANSPOSE(y)
1.504 +#undef OC_I
1.505 +#undef OC_J
1.506 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16,_y)
1.507 +#define OC_J(_k,_y) OC_I(_k,_y)
1.508 + OC_COLUMN_IDCT_10(y)
1.509 +#undef OC_I
1.510 +#undef OC_J
1.511 +#define OC_I(_k,_y) OC_MEM_OFFS((_k)*16+8,_y)
1.512 +#define OC_J(_k,_y) OC_I(_k,_y)
1.513 + OC_COLUMN_IDCT_10(y)
1.514 +#undef OC_I
1.515 +#undef OC_J
1.516 + :[y]"=m"OC_ARRAY_OPERAND(ogg_int16_t,_y,64)
1.517 + :[x]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,_x,64),
1.518 + [c]"m"OC_CONST_ARRAY_OPERAND(ogg_int16_t,OC_IDCT_CONSTS,128)
1.519 + );
1.520 + if(_x!=_y){
1.521 + __asm__ __volatile__(
1.522 + "pxor %%mm0,%%mm0\n\t"
1.523 + "movq %%mm0,"OC_MEM_OFFS(0x00,x)"\n\t"
1.524 + "movq %%mm0,"OC_MEM_OFFS(0x10,x)"\n\t"
1.525 + "movq %%mm0,"OC_MEM_OFFS(0x20,x)"\n\t"
1.526 + "movq %%mm0,"OC_MEM_OFFS(0x30,x)"\n\t"
1.527 + :[x]"+m"OC_ARRAY_OPERAND(ogg_int16_t,_x,28)
1.528 + );
1.529 + }
1.530 +}
1.531 +
1.532 +/*Performs an inverse 8x8 Type-II DCT transform.
1.533 + The input is assumed to be scaled by a factor of 4 relative to orthonormal
1.534 + version of the transform.*/
1.535 +void oc_idct8x8_mmx(ogg_int16_t _y[64],ogg_int16_t _x[64],int _last_zzi){
1.536 + /*_last_zzi is subtly different from an actual count of the number of
1.537 + coefficients we decoded for this block.
1.538 + It contains the value of zzi BEFORE the final token in the block was
1.539 + decoded.
1.540 + In most cases this is an EOB token (the continuation of an EOB run from a
1.541 + previous block counts), and so this is the same as the coefficient count.
1.542 + However, in the case that the last token was NOT an EOB token, but filled
1.543 + the block up with exactly 64 coefficients, _last_zzi will be less than 64.
1.544 + Provided the last token was not a pure zero run, the minimum value it can
1.545 + be is 46, and so that doesn't affect any of the cases in this routine.
1.546 + However, if the last token WAS a pure zero run of length 63, then _last_zzi
1.547 + will be 1 while the number of coefficients decoded is 64.
1.548 + Thus, we will trigger the following special case, where the real
1.549 + coefficient count would not.
1.550 + Note also that a zero run of length 64 will give _last_zzi a value of 0,
1.551 + but we still process the DC coefficient, which might have a non-zero value
1.552 + due to DC prediction.
1.553 + Although convoluted, this is arguably the correct behavior: it allows us to
1.554 + use a smaller transform when the block ends with a long zero run instead
1.555 + of a normal EOB token.
1.556 + It could be smarter... multiple separate zero runs at the end of a block
1.557 + will fool it, but an encoder that generates these really deserves what it
1.558 + gets.
1.559 + Needless to say we inherited this approach from VP3.*/
1.560 + /*Then perform the iDCT.*/
1.561 + if(_last_zzi<=10)oc_idct8x8_10_mmx(_y,_x);
1.562 + else oc_idct8x8_slow_mmx(_y,_x);
1.563 +}
1.564 +
1.565 +#endif
