theora-old/lib/x86_32/dct_decode_mmx.c

   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) 2002-2003                *
   9  * by the Xiph.Org Foundation http://www.xiph.org/                  *
  10  *                                                                  *
  11  ********************************************************************
  12
  13   function:
  14   last mod: $Id: dsp_mmx.c 12440 2007-02-06 16:36:26Z j $
  15
  16  ********************************************************************/
  17
  18 #include <stdlib.h>
  19
  20 #include "codec_internal.h"
  21
  22 #if defined(USE_ASM)
  23
  24 static const __attribute__((aligned(8),used)) ogg_int64_t V3= 0x0003000300030003LL;
  25 static const __attribute__((aligned(8),used)) ogg_int64_t V804= 0x0804080408040804LL;
  26
  27 #if defined(__APPLE__)
  28 #define MANGLE(x) "_"#x
  29 #else
  30 #define MANGLE(x) #x
  31 #endif
  32
  33 static void FilterHoriz__mmx(unsigned char * PixelPtr,
  34                         ogg_int32_t LineLength,
  35                         ogg_int16_t *BoundingValuePtr){
  36
  37 #define OC_LOOP_H_4x4                                                   \
  38     __asm__ __volatile__(                                               \
  39     "lea (%1,%1,2),%%esi\n"     /* esi = ystride*3 */                   \
  40     "movd (%0), %%mm0\n"        /* 0 0 0 0 3 2 1 0 */                   \
  41     "movd (%0,%1),%%mm1\n"      /* 0 0 0 0 7 6 5 4 */                   \
  42     "movd (%0,%1,2),%%mm2\n"    /* 0 0 0 0 b a 9 8 */                   \
  43     "movd (%0,%%esi),%%mm3\n"   /* 0 0 0 0 f e d c */                   \
  44     "punpcklbw %%mm1,%%mm0\n"   /* mm0 = 7 3 6 2 5 1 4 0 */             \
  45     "punpcklbw %%mm3,%%mm2\n"   /* mm2 = f b e a d 9 c 8 */             \
  46     "movq %%mm0,%%mm1\n"        /* mm1 = 7 3 6 2 5 1 4 0 */             \
  47     "punpcklwd %%mm2,%%mm1\n"   /* mm1 = d 9 5 1 c 8 4 0 */             \
  48     "punpckhwd %%mm2,%%mm0\n"   /* mm0 = f b 7 3 e a 6 2 */             \
  49     "pxor %%mm7,%%mm7\n"                                                \
  50     "movq %%mm1,%%mm5\n"        /* mm5 = d 9 5 1 c 8 4 0 */             \
  51     "punpckhbw %%mm7,%%mm5\n"   /* mm5 = 0 d 0 9 0 5 0 1 = pix[1]*/     \
  52     "punpcklbw %%mm7,%%mm1\n"   /* mm1 = 0 c 0 8 0 4 0 0 = pix[0]*/     \
  53     "movq %%mm0,%%mm3\n"        /* mm3 = f b 7 3 e a 6 2 */             \
  54     "punpckhbw %%mm7,%%mm3\n"   /* mm3 = 0 f 0 b 0 7 0 3 = pix[3]*/     \
  55     "punpcklbw %%mm7,%%mm0\n"       /* mm0 = 0 e 0 a 0 6 0 2 = pix[2]*/ \
  56                                                                         \
  57     "psubw %%mm3,%%mm1\n"       /* mm1 = pix[0]-pix[3] mm1 - mm3 */     \
  58     "movq %%mm0,%%mm7\n"        /* mm7 = pix[2]*/                       \
  59     "psubw %%mm5,%%mm0\n"       /* mm0 = pix[2]-pix[1] mm0 - mm5*/      \
  60     "PMULLW "MANGLE(V3)",%%mm0\n" /* *3 */                              \
  61     "paddw %%mm0,%%mm1\n"         /* mm1 has f[0] ... f[4]*/            \
  62     "paddw "MANGLE(V804)",%%mm1\n"/* add 4 */ /* add 256 after shift */ \
  63     "psraw $3,%%mm1\n"          /* >>3 */                               \
  64     " pextrw $0,%%mm1,%%esi\n"  /* In MM1 we have 4 f coefs (16bits) */ \
  65     " pextrw $1,%%mm1,%%edi\n"  /* now perform MM4 = *(_bv+ f) */       \
  66     " pinsrw $0,(%2,%%esi,2),%%mm4\n"                                   \
  67     " pextrw $2,%%mm1,%%esi\n"                                          \
  68     " pinsrw $1,(%2,%%edi,2),%%mm4\n"                                   \
  69     " pextrw $3,%%mm1,%%edi\n"                                          \
  70     " pinsrw $2,(%2,%%esi,2),%%mm4\n"                                   \
  71     " pinsrw $3,(%2,%%edi,2),%%mm4\n" /* new f vals loaded */           \
  72     "pxor %%mm0,%%mm0\n"                                                \
  73     " paddw %%mm4,%%mm5\n"      /*(pix[1]+f);*/                         \
  74     " psubw %%mm4,%%mm7\n"      /* (pix[2]-f); */                       \
  75     " packuswb %%mm0,%%mm5\n"   /* mm5 = x x x x newpix1 */             \
  76     " packuswb %%mm0,%%mm7\n"   /* mm7 = x x x x newpix2 */             \
  77     " punpcklbw %%mm7,%%mm5\n"  /* 2 1 2 1 2 1 2 1 */                   \
  78     " movd %%mm5,%%eax\n"       /* eax = newpix21 */                    \
  79     " movw %%ax,1(%0)\n"                                                \
  80     " psrlq $32,%%mm5\n"        /* why is so big stall here ? */        \
  81     " shrl $16,%%eax\n"                                                 \
  82     " lea 1(%0,%1,2),%%edi\n"                                           \
  83     " movw %%ax,1(%0,%1,1)\n"                                           \
  84     " movd %%mm5,%%eax\n"       /* eax = newpix21 high part */          \
  85     " lea (%1,%1,2),%%esi\n"                                            \
  86     " movw %%ax,(%%edi)\n"                                              \
  87     " shrl $16,%%eax\n"                                                 \
  88     " movw %%ax,1(%0,%%esi)\n"                                          \
  89     :                                                                   \
  90     : "r" (PixelPtr), "r" (LineLength), "r" (BoundingValuePtr-256)      \
  91     : "esi", "edi" , "memory", "eax"                                    \
  92     );
  93
  94     OC_LOOP_H_4x4
  95     PixelPtr += LineLength*4;
  96     OC_LOOP_H_4x4
  97     __asm__ __volatile__("emms\n");
  98 }
  99
 100 static void FilterVert__mmx(unsigned char * PixelPtr,
 101                 ogg_int32_t LineLength,
 102                 ogg_int16_t *BoundingValuePtr){
 103     __asm__ __volatile__(
 104     "pxor %%mm0,%%mm0\n"        /* mm0 = 0 */
 105     "movq (%0),%%mm7\n"         /* mm7 = pix[0..7] */
 106     "lea (%1,%1,2),%%esi\n"     /* esi = ystride*3 */
 107     "movq (%0,%%esi),%%mm4\n"   /* mm4 = pix[0..7+ystride*3] */
 108     "movq %%mm7,%%mm6\n"        /* mm6 = pix[0..7] */
 109     "punpcklbw %%mm0,%%mm6\n"   /* expand unsigned pix[0..3] to 16 bits */
 110     "movq %%mm4,%%mm5\n"
 111     "punpckhbw %%mm0,%%mm7\n"   /* expand unsigned pix[4..7] to 16 bits */
 112     "punpcklbw %%mm0,%%mm4\n"   /* expand other arrays too */
 113     "punpckhbw %%mm0,%%mm5\n"
 114     "psubw %%mm4,%%mm6\n"       /* mm6 = mm6 - mm4 */
 115     "psubw %%mm5,%%mm7\n"       /* mm7 = mm7 - mm5 */
 116                 /* mm7:mm6 = _p[0]-_p[ystride*3] */
 117     "movq (%0,%1),%%mm4\n"      /* mm4 = pix[0..7+ystride] */
 118     "movq %%mm4,%%mm5\n"
 119     "movq (%0,%1,2),%%mm2\n"    /* mm2 = pix[0..7+ystride*2] */
 120     "movq %%mm2,%%mm3\n"
 121     "movq %%mm2,%%mm1\n"        //ystride*2
 122     "punpckhbw %%mm0,%%mm5\n"
 123     "punpcklbw %%mm0,%%mm4\n"
 124     "punpckhbw %%mm0,%%mm3\n"
 125     "punpcklbw %%mm0,%%mm2\n"
 126     "psubw %%mm5,%%mm3\n"
 127     "psubw %%mm4,%%mm2\n"
 128                 /* mm3:mm2 = (pix[ystride*2]-pix[ystride]); */
 129     "PMULLW "MANGLE(V3)",%%mm3\n"    /* *3 */
 130     "PMULLW "MANGLE(V3)",%%mm2\n"    /* *3 */
 131     "paddw %%mm7,%%mm3\n"            /* highpart */
 132     "paddw %%mm6,%%mm2\n"            /* lowpart of pix[0]-pix[ystride*3]+3*(pix[ystride*2]-pix[ystride]);  */
 133     "paddw "MANGLE(V804)",%%mm3\n"   /* add 4 */ /* add 256 after shift */
 134     "paddw "MANGLE(V804)",%%mm2\n"   /* add 4 */ /* add 256 after shift */
 135     "psraw $3,%%mm3\n"               /* >>3 f coefs high */
 136     "psraw $3,%%mm2\n"               /* >>3 f coefs low */
 137
 138     " pextrw $0,%%mm2,%%esi\n"  /* In MM3:MM2 we have f coefs (16bits) */
 139     " pextrw $1,%%mm2,%%edi\n"  /* now perform MM7:MM6 = *(_bv+ f) */
 140     " pinsrw $0,(%2,%%esi,2),%%mm6\n"
 141     " pinsrw $1,(%2,%%edi,2),%%mm6\n"
 142
 143     " pextrw $2,%%mm2,%%esi\n"
 144     " pextrw $3,%%mm2,%%edi\n"
 145     " pinsrw $2,(%2,%%esi,2),%%mm6\n"
 146     " pinsrw $3,(%2,%%edi,2),%%mm6\n"
 147
 148     " pextrw $0,%%mm3,%%esi\n"
 149     " pextrw $1,%%mm3,%%edi\n"
 150     " pinsrw $0,(%2,%%esi,2),%%mm7\n"
 151     " pinsrw $1,(%2,%%edi,2),%%mm7\n"
 152
 153     " pextrw $2,%%mm3,%%esi\n"
 154     " pextrw $3,%%mm3,%%edi\n"
 155     " pinsrw $2,(%2,%%esi,2),%%mm7\n"
 156     " pinsrw $3,(%2,%%edi,2),%%mm7\n"   //MM7 MM6   f=*(_bv+(f+4>>3));
 157
 158     "paddw %%mm6,%%mm4\n"       /* (pix[ystride]+f); */
 159     "paddw %%mm7,%%mm5\n"       /* (pix[ystride]+f); */
 160     "movq %%mm1,%%mm2\n"
 161     "punpcklbw %%mm0,%%mm1\n"
 162     "punpckhbw %%mm0,%%mm2\n"   //[ystride*2]
 163     "psubw %%mm6,%%mm1\n"       /* (pix[ystride*2]-f); */
 164     "psubw %%mm7,%%mm2\n"       /* (pix[ystride*2]-f); */
 165     "packuswb %%mm2,%%mm1\n"
 166     "packuswb %%mm5,%%mm4\n"
 167     "movq %%mm1,(%0,%1,2)\n"    /* pix[ystride*2]= */
 168     "movq %%mm4,(%0,%1)\n"      /* pix[ystride]= */
 169     "emms\n"
 170     :
 171     : "r" (PixelPtr-2*LineLength), "r" (LineLength), "r" (BoundingValuePtr-256)
 172     : "esi", "edi" , "memory"
 173     );
 174 }
 175
 176 /* install our implementation in the function table */
 177 void dsp_mmx_dct_decode_init(DspFunctions *funcs)
 178 {
 179   TH_DEBUG("enabling accelerated x86_32 mmx dct decode functions.\n");
 180   funcs->FilterVert = FilterVert__mmx;
 181   funcs->FilterHoriz = FilterHoriz__mmx;
 182 }
 183
 184 #endif /* USE_ASM */