View | Details | Raw Unified
Collapse All | Expand All

(-) libtheora-1.0beta2/lib/enc/x86_32/dct_decode_mmx.c (-8 / +8 lines)
 Lines 57-65   static void FilterHoriz__mmx(unsigned ch Link Here  
    "psubw %%mm3,%%mm1\n"       /* mm1 = pix[0]-pix[3] mm1 - mm3 */     \
    "psubw %%mm3,%%mm1\n"       /* mm1 = pix[0]-pix[3] mm1 - mm3 */     \
    "movq %%mm0,%%mm7\n"        /* mm7 = pix[2]*/                       \
    "movq %%mm0,%%mm7\n"        /* mm7 = pix[2]*/                       \
    "psubw %%mm5,%%mm0\n"       /* mm0 = pix[2]-pix[1] mm0 - mm5*/      \
    "psubw %%mm5,%%mm0\n"       /* mm0 = pix[2]-pix[1] mm0 - mm5*/      \
    "PMULLW "MANGLE(V3)",%%mm0\n" /* *3 */                              \
 
    "PMULLW %3,%%mm0\n" /* *3 */                              \
 
    "paddw %%mm0,%%mm1\n"         /* mm1 has f[0] ... f[4]*/            \
    "paddw %%mm0,%%mm1\n"         /* mm1 has f[0] ... f[4]*/            \
    "paddw "MANGLE(V804)",%%mm1\n"/* add 4 */ /* add 256 after shift */ \
 
    "paddw %4,%%mm1\n"/* add 4 */ /* add 256 after shift */ \
 
    "psraw $3,%%mm1\n"          /* >>3 */                               \
    "psraw $3,%%mm1\n"          /* >>3 */                               \
    " pextrw $0,%%mm1,%%esi\n"  /* In MM1 we have 4 f coefs (16bits) */ \
    " pextrw $0,%%mm1,%%esi\n"  /* In MM1 we have 4 f coefs (16bits) */ \
    " pextrw $1,%%mm1,%%edi\n"  /* now perform MM4 = *(_bv+ f) */       \
    " pextrw $1,%%mm1,%%edi\n"  /* now perform MM4 = *(_bv+ f) */       \
 Lines 87-93   static void FilterHoriz__mmx(unsigned ch Link Here  
    " shrl $16,%%eax\n"                                                 \
    " shrl $16,%%eax\n"                                                 \
    " movw %%ax,1(%0,%%esi)\n"                                          \
    " movw %%ax,1(%0,%%esi)\n"                                          \
    :                                                                   \
    :                                                                   \
    : "r" (PixelPtr), "r" (LineLength), "r" (BoundingValuePtr-256)      \
 
    : "r" (PixelPtr), "r" (LineLength), "r" (BoundingValuePtr-256), "m" (V3), "m" (V804) \
 
    : "esi", "edi" , "memory", "eax"                                    \
    : "esi", "edi" , "memory", "eax"                                    \
    );
    );

  

    

    

       Lines 126-137
      static void FilterVert__mmx(unsigned cha 
    
  Link Here 
    

  

        
    "psubw %%mm5,%%mm3\n"
    "psubw %%mm5,%%mm3\n"

        
    "psubw %%mm4,%%mm2\n"
    "psubw %%mm4,%%mm2\n"

        
                /* mm3:mm2 = (pix[ystride*2]-pix[ystride]); */
                /* mm3:mm2 = (pix[ystride*2]-pix[ystride]); */

            

              
    "PMULLW "MANGLE(V3)",%%mm3\n"    /* *3 */

              
    "PMULLW %3,%%mm3\n"    /* *3 */

            

            

              
    "PMULLW "MANGLE(V3)",%%mm2\n"    /* *3 */

              
    "PMULLW %3,%%mm2\n"    /* *3 */

            

        
    "paddw %%mm7,%%mm3\n"            /* highpart */
    "paddw %%mm7,%%mm3\n"            /* highpart */

        
    "paddw %%mm6,%%mm2\n"            /* lowpart of pix[0]-pix[ystride*3]+3*(pix[ystride*2]-pix[ystride]);  */
    "paddw %%mm6,%%mm2\n"            /* lowpart of pix[0]-pix[ystride*3]+3*(pix[ystride*2]-pix[ystride]);  */

            

              
    "paddw "MANGLE(V804)",%%mm3\n"   /* add 4 */ /* add 256 after shift */

              
    "paddw %4,%%mm3\n"   /* add 4 */ /* add 256 after shift */

            

            

              
    "paddw "MANGLE(V804)",%%mm2\n"   /* add 4 */ /* add 256 after shift */

              
    "paddw %4,%%mm2\n"   /* add 4 */ /* add 256 after shift */

            

        
    "psraw $3,%%mm3\n"               /* >>3 f coefs high */
    "psraw $3,%%mm3\n"               /* >>3 f coefs high */

        
    "psraw $3,%%mm2\n"               /* >>3 f coefs low */
    "psraw $3,%%mm2\n"               /* >>3 f coefs low */

        

  

    

    

       Lines 168-174
      static void FilterVert__mmx(unsigned cha 
    
  Link Here 
    

  

        
    "movq %%mm4,(%0,%1)\n"      /* pix[ystride]= */
    "movq %%mm4,(%0,%1)\n"      /* pix[ystride]= */

        
    "emms\n"
    "emms\n"

        
    :
    :

            

              
    : "r" (PixelPtr-2*LineLength), "r" (LineLength), "r" (BoundingValuePtr-256)

              
    : "r" (PixelPtr-2*LineLength), "r" (LineLength), "r" (BoundingValuePtr-256), "m" (V3), "m" (V804)

            

        
    : "esi", "edi" , "memory"
    : "esi", "edi" , "memory"

        
    );
    );

        
}
}