1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libtheora/lib/x86/mmxfrag.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,368 @@
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: mmxfrag.c 17410 2010-09-21 21:53:48Z tterribe $
1.18 +
1.19 + ********************************************************************/
1.20 +
1.21 +/*MMX acceleration of fragment reconstruction for motion compensation.
1.22 + Originally written by Rudolf Marek.
1.23 + Additional optimization by Nils Pipenbrinck.
1.24 + Note: Loops are unrolled for best performance.
1.25 + The iteration each instruction belongs to is marked in the comments as #i.*/
1.26 +#include <stddef.h>
1.27 +#include "x86int.h"
1.28 +
1.29 +#if defined(OC_X86_ASM)
1.30 +
1.31 +/*Copies an 8x8 block of pixels from _src to _dst, assuming _ystride bytes
1.32 + between rows.*/
1.33 +# define OC_FRAG_COPY_MMX(_dst,_src,_ystride) \
1.34 + do{ \
1.35 + const unsigned char *src; \
1.36 + unsigned char *dst; \
1.37 + ptrdiff_t ystride3; \
1.38 + src=(_src); \
1.39 + dst=(_dst); \
1.40 + __asm__ __volatile__( \
1.41 + /*src+0*ystride*/ \
1.42 + "movq (%[src]),%%mm0\n\t" \
1.43 + /*src+1*ystride*/ \
1.44 + "movq (%[src],%[ystride]),%%mm1\n\t" \
1.45 + /*ystride3=ystride*3*/ \
1.46 + "lea (%[ystride],%[ystride],2),%[ystride3]\n\t" \
1.47 + /*src+2*ystride*/ \
1.48 + "movq (%[src],%[ystride],2),%%mm2\n\t" \
1.49 + /*src+3*ystride*/ \
1.50 + "movq (%[src],%[ystride3]),%%mm3\n\t" \
1.51 + /*dst+0*ystride*/ \
1.52 + "movq %%mm0,(%[dst])\n\t" \
1.53 + /*dst+1*ystride*/ \
1.54 + "movq %%mm1,(%[dst],%[ystride])\n\t" \
1.55 + /*Pointer to next 4.*/ \
1.56 + "lea (%[src],%[ystride],4),%[src]\n\t" \
1.57 + /*dst+2*ystride*/ \
1.58 + "movq %%mm2,(%[dst],%[ystride],2)\n\t" \
1.59 + /*dst+3*ystride*/ \
1.60 + "movq %%mm3,(%[dst],%[ystride3])\n\t" \
1.61 + /*Pointer to next 4.*/ \
1.62 + "lea (%[dst],%[ystride],4),%[dst]\n\t" \
1.63 + /*src+0*ystride*/ \
1.64 + "movq (%[src]),%%mm0\n\t" \
1.65 + /*src+1*ystride*/ \
1.66 + "movq (%[src],%[ystride]),%%mm1\n\t" \
1.67 + /*src+2*ystride*/ \
1.68 + "movq (%[src],%[ystride],2),%%mm2\n\t" \
1.69 + /*src+3*ystride*/ \
1.70 + "movq (%[src],%[ystride3]),%%mm3\n\t" \
1.71 + /*dst+0*ystride*/ \
1.72 + "movq %%mm0,(%[dst])\n\t" \
1.73 + /*dst+1*ystride*/ \
1.74 + "movq %%mm1,(%[dst],%[ystride])\n\t" \
1.75 + /*dst+2*ystride*/ \
1.76 + "movq %%mm2,(%[dst],%[ystride],2)\n\t" \
1.77 + /*dst+3*ystride*/ \
1.78 + "movq %%mm3,(%[dst],%[ystride3])\n\t" \
1.79 + :[dst]"+r"(dst),[src]"+r"(src),[ystride3]"=&r"(ystride3) \
1.80 + :[ystride]"r"((ptrdiff_t)(_ystride)) \
1.81 + :"memory" \
1.82 + ); \
1.83 + } \
1.84 + while(0)
1.85 +
1.86 +/*Copies an 8x8 block of pixels from _src to _dst, assuming _ystride bytes
1.87 + between rows.*/
1.88 +void oc_frag_copy_mmx(unsigned char *_dst,
1.89 + const unsigned char *_src,int _ystride){
1.90 + OC_FRAG_COPY_MMX(_dst,_src,_ystride);
1.91 +}
1.92 +
1.93 +/*Copies the fragments specified by the lists of fragment indices from one
1.94 + frame to another.
1.95 + _dst_frame: The reference frame to copy to.
1.96 + _src_frame: The reference frame to copy from.
1.97 + _ystride: The row stride of the reference frames.
1.98 + _fragis: A pointer to a list of fragment indices.
1.99 + _nfragis: The number of fragment indices to copy.
1.100 + _frag_buf_offs: The offsets of fragments in the reference frames.*/
1.101 +void oc_frag_copy_list_mmx(unsigned char *_dst_frame,
1.102 + const unsigned char *_src_frame,int _ystride,
1.103 + const ptrdiff_t *_fragis,ptrdiff_t _nfragis,const ptrdiff_t *_frag_buf_offs){
1.104 + ptrdiff_t fragii;
1.105 + for(fragii=0;fragii<_nfragis;fragii++){
1.106 + ptrdiff_t frag_buf_off;
1.107 + frag_buf_off=_frag_buf_offs[_fragis[fragii]];
1.108 + OC_FRAG_COPY_MMX(_dst_frame+frag_buf_off,
1.109 + _src_frame+frag_buf_off,_ystride);
1.110 + }
1.111 +}
1.112 +
1.113 +
1.114 +void oc_frag_recon_intra_mmx(unsigned char *_dst,int _ystride,
1.115 + const ogg_int16_t *_residue){
1.116 + __asm__ __volatile__(
1.117 + /*Set mm0 to 0xFFFFFFFFFFFFFFFF.*/
1.118 + "pcmpeqw %%mm0,%%mm0\n\t"
1.119 + /*#0 Load low residue.*/
1.120 + "movq 0*8(%[residue]),%%mm1\n\t"
1.121 + /*#0 Load high residue.*/
1.122 + "movq 1*8(%[residue]),%%mm2\n\t"
1.123 + /*Set mm0 to 0x8000800080008000.*/
1.124 + "psllw $15,%%mm0\n\t"
1.125 + /*#1 Load low residue.*/
1.126 + "movq 2*8(%[residue]),%%mm3\n\t"
1.127 + /*#1 Load high residue.*/
1.128 + "movq 3*8(%[residue]),%%mm4\n\t"
1.129 + /*Set mm0 to 0x0080008000800080.*/
1.130 + "psrlw $8,%%mm0\n\t"
1.131 + /*#2 Load low residue.*/
1.132 + "movq 4*8(%[residue]),%%mm5\n\t"
1.133 + /*#2 Load high residue.*/
1.134 + "movq 5*8(%[residue]),%%mm6\n\t"
1.135 + /*#0 Bias low residue.*/
1.136 + "paddsw %%mm0,%%mm1\n\t"
1.137 + /*#0 Bias high residue.*/
1.138 + "paddsw %%mm0,%%mm2\n\t"
1.139 + /*#0 Pack to byte.*/
1.140 + "packuswb %%mm2,%%mm1\n\t"
1.141 + /*#1 Bias low residue.*/
1.142 + "paddsw %%mm0,%%mm3\n\t"
1.143 + /*#1 Bias high residue.*/
1.144 + "paddsw %%mm0,%%mm4\n\t"
1.145 + /*#1 Pack to byte.*/
1.146 + "packuswb %%mm4,%%mm3\n\t"
1.147 + /*#2 Bias low residue.*/
1.148 + "paddsw %%mm0,%%mm5\n\t"
1.149 + /*#2 Bias high residue.*/
1.150 + "paddsw %%mm0,%%mm6\n\t"
1.151 + /*#2 Pack to byte.*/
1.152 + "packuswb %%mm6,%%mm5\n\t"
1.153 + /*#0 Write row.*/
1.154 + "movq %%mm1,(%[dst])\n\t"
1.155 + /*#1 Write row.*/
1.156 + "movq %%mm3,(%[dst],%[ystride])\n\t"
1.157 + /*#2 Write row.*/
1.158 + "movq %%mm5,(%[dst],%[ystride],2)\n\t"
1.159 + /*#3 Load low residue.*/
1.160 + "movq 6*8(%[residue]),%%mm1\n\t"
1.161 + /*#3 Load high residue.*/
1.162 + "movq 7*8(%[residue]),%%mm2\n\t"
1.163 + /*#4 Load high residue.*/
1.164 + "movq 8*8(%[residue]),%%mm3\n\t"
1.165 + /*#4 Load high residue.*/
1.166 + "movq 9*8(%[residue]),%%mm4\n\t"
1.167 + /*#5 Load high residue.*/
1.168 + "movq 10*8(%[residue]),%%mm5\n\t"
1.169 + /*#5 Load high residue.*/
1.170 + "movq 11*8(%[residue]),%%mm6\n\t"
1.171 + /*#3 Bias low residue.*/
1.172 + "paddsw %%mm0,%%mm1\n\t"
1.173 + /*#3 Bias high residue.*/
1.174 + "paddsw %%mm0,%%mm2\n\t"
1.175 + /*#3 Pack to byte.*/
1.176 + "packuswb %%mm2,%%mm1\n\t"
1.177 + /*#4 Bias low residue.*/
1.178 + "paddsw %%mm0,%%mm3\n\t"
1.179 + /*#4 Bias high residue.*/
1.180 + "paddsw %%mm0,%%mm4\n\t"
1.181 + /*#4 Pack to byte.*/
1.182 + "packuswb %%mm4,%%mm3\n\t"
1.183 + /*#5 Bias low residue.*/
1.184 + "paddsw %%mm0,%%mm5\n\t"
1.185 + /*#5 Bias high residue.*/
1.186 + "paddsw %%mm0,%%mm6\n\t"
1.187 + /*#5 Pack to byte.*/
1.188 + "packuswb %%mm6,%%mm5\n\t"
1.189 + /*#3 Write row.*/
1.190 + "movq %%mm1,(%[dst],%[ystride3])\n\t"
1.191 + /*#4 Write row.*/
1.192 + "movq %%mm3,(%[dst4])\n\t"
1.193 + /*#5 Write row.*/
1.194 + "movq %%mm5,(%[dst4],%[ystride])\n\t"
1.195 + /*#6 Load low residue.*/
1.196 + "movq 12*8(%[residue]),%%mm1\n\t"
1.197 + /*#6 Load high residue.*/
1.198 + "movq 13*8(%[residue]),%%mm2\n\t"
1.199 + /*#7 Load low residue.*/
1.200 + "movq 14*8(%[residue]),%%mm3\n\t"
1.201 + /*#7 Load high residue.*/
1.202 + "movq 15*8(%[residue]),%%mm4\n\t"
1.203 + /*#6 Bias low residue.*/
1.204 + "paddsw %%mm0,%%mm1\n\t"
1.205 + /*#6 Bias high residue.*/
1.206 + "paddsw %%mm0,%%mm2\n\t"
1.207 + /*#6 Pack to byte.*/
1.208 + "packuswb %%mm2,%%mm1\n\t"
1.209 + /*#7 Bias low residue.*/
1.210 + "paddsw %%mm0,%%mm3\n\t"
1.211 + /*#7 Bias high residue.*/
1.212 + "paddsw %%mm0,%%mm4\n\t"
1.213 + /*#7 Pack to byte.*/
1.214 + "packuswb %%mm4,%%mm3\n\t"
1.215 + /*#6 Write row.*/
1.216 + "movq %%mm1,(%[dst4],%[ystride],2)\n\t"
1.217 + /*#7 Write row.*/
1.218 + "movq %%mm3,(%[dst4],%[ystride3])\n\t"
1.219 + :
1.220 + :[residue]"r"(_residue),
1.221 + [dst]"r"(_dst),
1.222 + [dst4]"r"(_dst+(_ystride<<2)),
1.223 + [ystride]"r"((ptrdiff_t)_ystride),
1.224 + [ystride3]"r"((ptrdiff_t)_ystride*3)
1.225 + :"memory"
1.226 + );
1.227 +}
1.228 +
1.229 +void oc_frag_recon_inter_mmx(unsigned char *_dst,const unsigned char *_src,
1.230 + int _ystride,const ogg_int16_t *_residue){
1.231 + int i;
1.232 + /*Zero mm0.*/
1.233 + __asm__ __volatile__("pxor %%mm0,%%mm0\n\t"::);
1.234 + for(i=4;i-->0;){
1.235 + __asm__ __volatile__(
1.236 + /*#0 Load source.*/
1.237 + "movq (%[src]),%%mm3\n\t"
1.238 + /*#1 Load source.*/
1.239 + "movq (%[src],%[ystride]),%%mm7\n\t"
1.240 + /*#0 Get copy of src.*/
1.241 + "movq %%mm3,%%mm4\n\t"
1.242 + /*#0 Expand high source.*/
1.243 + "punpckhbw %%mm0,%%mm4\n\t"
1.244 + /*#0 Expand low source.*/
1.245 + "punpcklbw %%mm0,%%mm3\n\t"
1.246 + /*#0 Add residue high.*/
1.247 + "paddsw 8(%[residue]),%%mm4\n\t"
1.248 + /*#1 Get copy of src.*/
1.249 + "movq %%mm7,%%mm2\n\t"
1.250 + /*#0 Add residue low.*/
1.251 + "paddsw (%[residue]), %%mm3\n\t"
1.252 + /*#1 Expand high source.*/
1.253 + "punpckhbw %%mm0,%%mm2\n\t"
1.254 + /*#0 Pack final row pixels.*/
1.255 + "packuswb %%mm4,%%mm3\n\t"
1.256 + /*#1 Expand low source.*/
1.257 + "punpcklbw %%mm0,%%mm7\n\t"
1.258 + /*#1 Add residue low.*/
1.259 + "paddsw 16(%[residue]),%%mm7\n\t"
1.260 + /*#1 Add residue high.*/
1.261 + "paddsw 24(%[residue]),%%mm2\n\t"
1.262 + /*Advance residue.*/
1.263 + "lea 32(%[residue]),%[residue]\n\t"
1.264 + /*#1 Pack final row pixels.*/
1.265 + "packuswb %%mm2,%%mm7\n\t"
1.266 + /*Advance src.*/
1.267 + "lea (%[src],%[ystride],2),%[src]\n\t"
1.268 + /*#0 Write row.*/
1.269 + "movq %%mm3,(%[dst])\n\t"
1.270 + /*#1 Write row.*/
1.271 + "movq %%mm7,(%[dst],%[ystride])\n\t"
1.272 + /*Advance dst.*/
1.273 + "lea (%[dst],%[ystride],2),%[dst]\n\t"
1.274 + :[residue]"+r"(_residue),[dst]"+r"(_dst),[src]"+r"(_src)
1.275 + :[ystride]"r"((ptrdiff_t)_ystride)
1.276 + :"memory"
1.277 + );
1.278 + }
1.279 +}
1.280 +
1.281 +void oc_frag_recon_inter2_mmx(unsigned char *_dst,const unsigned char *_src1,
1.282 + const unsigned char *_src2,int _ystride,const ogg_int16_t *_residue){
1.283 + int i;
1.284 + /*Zero mm7.*/
1.285 + __asm__ __volatile__("pxor %%mm7,%%mm7\n\t"::);
1.286 + for(i=4;i-->0;){
1.287 + __asm__ __volatile__(
1.288 + /*#0 Load src1.*/
1.289 + "movq (%[src1]),%%mm0\n\t"
1.290 + /*#0 Load src2.*/
1.291 + "movq (%[src2]),%%mm2\n\t"
1.292 + /*#0 Copy src1.*/
1.293 + "movq %%mm0,%%mm1\n\t"
1.294 + /*#0 Copy src2.*/
1.295 + "movq %%mm2,%%mm3\n\t"
1.296 + /*#1 Load src1.*/
1.297 + "movq (%[src1],%[ystride]),%%mm4\n\t"
1.298 + /*#0 Unpack lower src1.*/
1.299 + "punpcklbw %%mm7,%%mm0\n\t"
1.300 + /*#1 Load src2.*/
1.301 + "movq (%[src2],%[ystride]),%%mm5\n\t"
1.302 + /*#0 Unpack higher src1.*/
1.303 + "punpckhbw %%mm7,%%mm1\n\t"
1.304 + /*#0 Unpack lower src2.*/
1.305 + "punpcklbw %%mm7,%%mm2\n\t"
1.306 + /*#0 Unpack higher src2.*/
1.307 + "punpckhbw %%mm7,%%mm3\n\t"
1.308 + /*Advance src1 ptr.*/
1.309 + "lea (%[src1],%[ystride],2),%[src1]\n\t"
1.310 + /*Advance src2 ptr.*/
1.311 + "lea (%[src2],%[ystride],2),%[src2]\n\t"
1.312 + /*#0 Lower src1+src2.*/
1.313 + "paddsw %%mm2,%%mm0\n\t"
1.314 + /*#0 Higher src1+src2.*/
1.315 + "paddsw %%mm3,%%mm1\n\t"
1.316 + /*#1 Copy src1.*/
1.317 + "movq %%mm4,%%mm2\n\t"
1.318 + /*#0 Build lo average.*/
1.319 + "psraw $1,%%mm0\n\t"
1.320 + /*#1 Copy src2.*/
1.321 + "movq %%mm5,%%mm3\n\t"
1.322 + /*#1 Unpack lower src1.*/
1.323 + "punpcklbw %%mm7,%%mm4\n\t"
1.324 + /*#0 Build hi average.*/
1.325 + "psraw $1,%%mm1\n\t"
1.326 + /*#1 Unpack higher src1.*/
1.327 + "punpckhbw %%mm7,%%mm2\n\t"
1.328 + /*#0 low+=residue.*/
1.329 + "paddsw (%[residue]),%%mm0\n\t"
1.330 + /*#1 Unpack lower src2.*/
1.331 + "punpcklbw %%mm7,%%mm5\n\t"
1.332 + /*#0 high+=residue.*/
1.333 + "paddsw 8(%[residue]),%%mm1\n\t"
1.334 + /*#1 Unpack higher src2.*/
1.335 + "punpckhbw %%mm7,%%mm3\n\t"
1.336 + /*#1 Lower src1+src2.*/
1.337 + "paddsw %%mm4,%%mm5\n\t"
1.338 + /*#0 Pack and saturate.*/
1.339 + "packuswb %%mm1,%%mm0\n\t"
1.340 + /*#1 Higher src1+src2.*/
1.341 + "paddsw %%mm2,%%mm3\n\t"
1.342 + /*#0 Write row.*/
1.343 + "movq %%mm0,(%[dst])\n\t"
1.344 + /*#1 Build lo average.*/
1.345 + "psraw $1,%%mm5\n\t"
1.346 + /*#1 Build hi average.*/
1.347 + "psraw $1,%%mm3\n\t"
1.348 + /*#1 low+=residue.*/
1.349 + "paddsw 16(%[residue]),%%mm5\n\t"
1.350 + /*#1 high+=residue.*/
1.351 + "paddsw 24(%[residue]),%%mm3\n\t"
1.352 + /*#1 Pack and saturate.*/
1.353 + "packuswb %%mm3,%%mm5\n\t"
1.354 + /*#1 Write row ptr.*/
1.355 + "movq %%mm5,(%[dst],%[ystride])\n\t"
1.356 + /*Advance residue ptr.*/
1.357 + "add $32,%[residue]\n\t"
1.358 + /*Advance dest ptr.*/
1.359 + "lea (%[dst],%[ystride],2),%[dst]\n\t"
1.360 + :[dst]"+r"(_dst),[residue]"+r"(_residue),
1.361 + [src1]"+%r"(_src1),[src2]"+r"(_src2)
1.362 + :[ystride]"r"((ptrdiff_t)_ystride)
1.363 + :"memory"
1.364 + );
1.365 + }
1.366 +}
1.367 +
1.368 +void oc_restore_fpu_mmx(void){
1.369 + __asm__ __volatile__("emms\n\t");
1.370 +}
1.371 +#endif
