jidctfst.c

jpeg编解码器

        "pmullw          8*12(%%edi),%%mm4 \n\t"        //tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
        "paddw           %%mm7,%%mm2 \n\t"              //tmp12

        "pmulhw          _fix_141,%%mm6 \n\t"           //tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
        "psubw           %%mm3,%%mm2 \n\t"      //tmp6 = tmp12 - tmp7

        "movq            %%mm1,%%mm5 \n\t"                      //copy tmp1
        "paddw           %%mm4,%%mm1 \n\t"                      //tmp13= tmp1 + tmp3;    /* phases 5-3 */

        "psubw           %%mm4,%%mm5 \n\t"                      //tmp1-tmp3
        "psubw           %%mm2,%%mm6 \n\t"      //tmp5 = tmp11 - tmp6;

        "movq            %%mm1,8*0(%%esi) \n\t"         //save tmp13 in workspace
        "psllw           $2,%%mm5    \n\t"                      //shift tmp1-tmp3

        "movq            8*0(%%ebx),%%mm7 \n\t"         //load inptr[DCTSIZE*0]

        "pmulhw          _fix_141,%%mm5 \n\t"           //MULTIPLY(tmp1 - tmp3, FIX_1_414213562)
        "paddw           %%mm6,%%mm0 \n\t"      //tmp4 = tmp10 + tmp5;

        "pmullw          8*0(%%edi),%%mm7 \n\t"         //tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);

        "movq            8*8(%%ebx),%%mm4 \n\t"         //load inptr[DCTSIZE*4]

        "pmullw          8*8(%%edi),%%mm4 \n\t"         //tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
        "psubw           %%mm1,%%mm5 \n\t"                      //tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */

        "movq            %%mm0,8*4(%%esi) \n\t" //save tmp4 in workspace
        "movq            %%mm7,%%mm1 \n\t"              //copy tmp0      /* phase 3 */

        "movq            %%mm5,8*2(%%esi) \n\t" //save tmp12 in workspace
        "psubw           %%mm4,%%mm1 \n\t"              //tmp11 = tmp0 - tmp2; 

        "paddw           %%mm4,%%mm7 \n\t"              //tmp10 = tmp0 + tmp2;
    "movq                %%mm1,%%mm5 \n\t"      //copy tmp11

        "paddw           8*2(%%esi),%%mm1 \n\t" //tmp1 = tmp11 + tmp12;
        "movq            %%mm7,%%mm4 \n\t"      //copy tmp10             /* phase 2 */

        "paddw           8*0(%%esi),%%mm7 \n\t" //tmp0 = tmp10 + tmp13;  

        "psubw           8*0(%%esi),%%mm4 \n\t" //tmp3 = tmp10 - tmp13;
        "movq            %%mm7,%%mm0 \n\t"      //copy tmp0

        "psubw           8*2(%%esi),%%mm5 \n\t" //tmp2 = tmp11 - tmp12;
        "paddw           %%mm3,%%mm7 \n\t"      //wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);

        "psubw           %%mm3,%%mm0 \n\t"              //wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);

        "movq            %%mm7,8*0(%%esi) \n\t" //wsptr[DCTSIZE*0]
        "movq            %%mm1,%%mm3 \n\t"              //copy tmp1

        "movq            %%mm0,8*14(%%esi) \n\t" //wsptr[DCTSIZE*7]
        "paddw           %%mm2,%%mm1 \n\t"              //wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);

        "psubw           %%mm2,%%mm3 \n\t"              //wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);

        "movq            %%mm1,8*2(%%esi) \n\t" //wsptr[DCTSIZE*1]
        "movq            %%mm4,%%mm1 \n\t"              //copy tmp3

        "movq            %%mm3,8*12(%%esi) \n\t" //wsptr[DCTSIZE*6]

        "paddw           8*4(%%esi),%%mm4 \n\t" //wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);

        "psubw           8*4(%%esi),%%mm1 \n\t" //wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);

        "movq            %%mm4,8*8(%%esi) \n\t"
        "movq            %%mm5,%%mm7 \n\t"              //copy tmp2

        "paddw           %%mm6,%%mm5 \n\t"              //wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5)

        "movq            %%mm1,8*6(%%esi) \n\t" //
        "psubw           %%mm6,%%mm7 \n\t"              //wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);

        "movq            %%mm5,8*4(%%esi) \n\t"

        "movq            %%mm7,8*10(%%esi) \n\t"



/*****************************************************************/
        "addl            $8,%%edi    \n\t"
        "addl            $8,%%ebx    \n\t"
        "addl            $8,%%esi    \n\t"

/*****************************************************************/




        "movq            8*10(%%ebx),%%mm1 \n\t"        //load inptr[DCTSIZE*5]

        "pmullw          8*10(%%edi),%%mm1 \n\t"        //tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);

        "movq            8*6(%%ebx),%%mm0 \n\t"         //load inptr[DCTSIZE*3]

        "pmullw          8*6(%%edi),%%mm0 \n\t"         //tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);

        "movq            8*2(%%ebx),%%mm3 \n\t"         //load inptr[DCTSIZE*1]
        "movq    %%mm1,%%mm2         \n\t"                      //copy tmp6      /* phase 6 */

        "pmullw          8*2(%%edi),%%mm3 \n\t"         //tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);

        "movq            8*14(%%ebx),%%mm4 \n\t"        //load inptr[DCTSIZE*1]
        "paddw   %%mm0,%%mm1         \n\t"                      //z13 = tmp6 + tmp5;

        "pmullw          8*14(%%edi),%%mm4 \n\t"    //tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
        "psubw   %%mm0,%%mm2         \n\t"                      //z10 = tmp6 - tmp5   

        "psllw           $2,%%mm2    \n\t"              //shift z10
        "movq            %%mm2,%%mm0 \n\t"              //copy z10

        "pmulhw          _fix_184n261,%%mm2 \n\t" //MULTIPLY( z12, FIX_1_847759065); /* 2*c2 */
        "movq            %%mm3,%%mm5 \n\t"                      //copy tmp4

        "pmulhw          _fix_n184,%%mm0 \n\t"  //MULTIPLY(z10, -FIX_1_847759065); /* 2*c2 */
        "paddw           %%mm4,%%mm3 \n\t"                      //z11 = tmp4 + tmp7;

        "movq            %%mm3,%%mm6 \n\t"                      //copy z11                       /* phase 5 */
        "psubw           %%mm4,%%mm5 \n\t"                      //z12 = tmp4 - tmp7;

        "psubw           %%mm1,%%mm6 \n\t"                      //z11-z13
        "psllw           $2,%%mm5    \n\t"              //shift z12

        "movq            8*12(%%ebx),%%mm4 \n\t"        //load inptr[DCTSIZE*6], even part
        "movq            %%mm5,%%mm7 \n\t"              //copy z12

        "pmulhw          _fix_108n184,%%mm5 \n\t" //MULT(z12, (FIX_1_08-FIX_1_84)) //- z5; /* 2*(c2-c6) */ even part
        "paddw           %%mm1,%%mm3 \n\t"                      //tmp7 = z11 + z13;      


    /* Even part */
        "pmulhw          _fix_184,%%mm7 \n\t"   //MULTIPLY(z10,(FIX_1_847759065 - FIX_2_613125930)) //+ z5; /* -2*(c2+c6) */
        "psllw           $2,%%mm6    \n\t"

        "movq            8*4(%%ebx),%%mm1 \n\t"         //load inptr[DCTSIZE*2]

        "pmullw          8*4(%%edi),%%mm1 \n\t"         //tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
        "paddw           %%mm5,%%mm0 \n\t"              //tmp10

        "pmullw          8*12(%%edi),%%mm4 \n\t"        //tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
        "paddw           %%mm7,%%mm2 \n\t"              //tmp12

        "pmulhw          _fix_141,%%mm6 \n\t"           //tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
        "psubw           %%mm3,%%mm2 \n\t"      //tmp6 = tmp12 - tmp7

        "movq            %%mm1,%%mm5 \n\t"                      //copy tmp1
        "paddw           %%mm4,%%mm1 \n\t"                      //tmp13= tmp1 + tmp3;    /* phases 5-3 */

        "psubw           %%mm4,%%mm5 \n\t"                      //tmp1-tmp3
        "psubw           %%mm2,%%mm6 \n\t"      //tmp5 = tmp11 - tmp6;

        "movq            %%mm1,8*0(%%esi) \n\t"         //save tmp13 in workspace
        "psllw           $2,%%mm5    \n\t"                      //shift tmp1-tmp3

        "movq            8*0(%%ebx),%%mm7 \n\t"         //load inptr[DCTSIZE*0]
        "paddw           %%mm6,%%mm0 \n\t"      //tmp4 = tmp10 + tmp5;

        "pmulhw          _fix_141,%%mm5 \n\t"           //MULTIPLY(tmp1 - tmp3, FIX_1_414213562)

        "pmullw          8*0(%%edi),%%mm7 \n\t"         //tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);

        "movq            8*8(%%ebx),%%mm4 \n\t"         //load inptr[DCTSIZE*4]

        "pmullw          8*8(%%edi),%%mm4 \n\t"         //tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
        "psubw           %%mm1,%%mm5 \n\t"                      //tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */

        "movq            %%mm0,8*4(%%esi) \n\t" //save tmp4 in workspace
        "movq            %%mm7,%%mm1 \n\t"              //copy tmp0      /* phase 3 */

        "movq            %%mm5,8*2(%%esi) \n\t" //save tmp12 in workspace
        "psubw           %%mm4,%%mm1 \n\t"              //tmp11 = tmp0 - tmp2; 

        "paddw           %%mm4,%%mm7 \n\t"              //tmp10 = tmp0 + tmp2;
    "movq                %%mm1,%%mm5 \n\t"      //copy tmp11

        "paddw           8*2(%%esi),%%mm1 \n\t" //tmp1 = tmp11 + tmp12;
        "movq            %%mm7,%%mm4 \n\t"      //copy tmp10             /* phase 2 */

        "paddw           8*0(%%esi),%%mm7 \n\t" //tmp0 = tmp10 + tmp13;  

        "psubw           8*0(%%esi),%%mm4 \n\t" //tmp3 = tmp10 - tmp13;
        "movq            %%mm7,%%mm0 \n\t"      //copy tmp0

        "psubw           8*2(%%esi),%%mm5 \n\t" //tmp2 = tmp11 - tmp12;
        "paddw           %%mm3,%%mm7 \n\t"      //wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);

        "psubw           %%mm3,%%mm0 \n\t"              //wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);

        "movq            %%mm7,8*0(%%esi) \n\t" //wsptr[DCTSIZE*0]
        "movq            %%mm1,%%mm3 \n\t"              //copy tmp1

        "movq            %%mm0,8*14(%%esi) \n\t" //wsptr[DCTSIZE*7]
        "paddw           %%mm2,%%mm1 \n\t"              //wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);

        "psubw           %%mm2,%%mm3 \n\t"              //wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);

        "movq            %%mm1,8*2(%%esi) \n\t" //wsptr[DCTSIZE*1]
        "movq            %%mm4,%%mm1 \n\t"              //copy tmp3

        "movq            %%mm3,8*12(%%esi) \n\t" //wsptr[DCTSIZE*6]

        "paddw           8*4(%%esi),%%mm4 \n\t" //wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);

        "psubw           8*4(%%esi),%%mm1 \n\t" //wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);

        "movq            %%mm4,8*8(%%esi) \n\t"
        "movq            %%mm5,%%mm7 \n\t"              //copy tmp2

        "paddw           %%mm6,%%mm5 \n\t"              //wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5)

        "movq            %%mm1,8*6(%%esi) \n\t" //
        "psubw           %%mm6,%%mm7 \n\t"              //wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);

        "movq            %%mm5,8*4(%%esi) \n\t"

        "movq            %%mm7,8*10(%%esi) \n\t"




/*****************************************************************/

  /* Pass 2: process rows from work array, store into output array. */
  /* Note that we must descale the results by a factor of 8 == 2**3, */
  /* and also undo the PASS1_BITS scaling. */

/*****************************************************************/
    /* Even part */

        "movl                    %%eax,%%esi \n\t"
        "movl                    %3, %%eax \n\t"

//    tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
//    tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
//    tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
//    tmp14 = ((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6]);
        "movq            8*0(%%esi),%%mm0 \n\t" //wsptr[0,0],[0,1],[0,2],[0,3]

        "movq            8*1(%%esi),%%mm1 \n\t" //wsptr[0,4],[0,5],[0,6],[0,7]
        "movq            %%mm0,%%mm2 \n\t"

        "movq            8*2(%%esi),%%mm3 \n\t" //wsptr[1,0],[1,1],[1,2],[1,3]
        "paddw           %%mm1,%%mm0 \n\t"              //wsptr[0,tmp10],[xxx],[0,tmp13],[xxx]

        "movq            8*3(%%esi),%%mm4 \n\t" //wsptr[1,4],[1,5],[1,6],[1,7]
        "psubw           %%mm1,%%mm2 \n\t"              //wsptr[0,tmp11],[xxx],[0,tmp14],[xxx]

        "movq            %%mm0,%%mm6 \n\t"
        "movq            %%mm3,%%mm5 \n\t"

        "paddw           %%mm4,%%mm3 \n\t"              //wsptr[1,tmp10],[xxx],[1,tmp13],[xxx]
        "movq            %%mm2,%%mm1 \n\t"

        "psubw           %%mm4,%%mm5 \n\t"              //wsptr[1,tmp11],[xxx],[1,tmp14],[xxx]
        "punpcklwd       %%mm3,%%mm0 \n\t"              //wsptr[0,tmp10],[1,tmp10],[xxx],[xxx]

        "movq            8*7(%%esi),%%mm7 \n\t" //wsptr[3,4],[3,5],[3,6],[3,7]
        "punpckhwd       %%mm3,%%mm6 \n\t"              //wsptr[0,tmp13],[1,tmp13],[xxx],[xxx]

        "movq            8*4(%%esi),%%mm3 \n\t" //wsptr[2,0],[2,1],[2,2],[2,3]
        "punpckldq       %%mm6,%%mm0 \n\t" //wsptr[0,tmp10],[1,tmp10],[0,tmp13],[1,tmp13]

        "punpcklwd       %%mm5,%%mm1 \n\t"              //wsptr[0,tmp11],[1,tmp11],[xxx],[xxx]
        "movq            %%mm3,%%mm4 \n\t"

        "movq            8*6(%%esi),%%mm6 \n\t" //wsptr[3,0],[3,1],[3,2],[3,3]
        "punpckhwd       %%mm5,%%mm2 \n\t"              //wsptr[0,tmp14],[1,tmp14],[xxx],[xxx]

        "movq            8*5(%%esi),%%mm5 \n\t" //wsptr[2,4],[2,5],[2,6],[2,7]
        "punpckldq       %%mm2,%%mm1 \n\t" //wsptr[0,tmp11],[1,tmp11],[0,tmp14],[1,tmp14]


        "paddw           %%mm5,%%mm3 \n\t"              //wsptr[2,tmp10],[xxx],[2,tmp13],[xxx]
        "movq            %%mm6,%%mm2 \n\t"

        "psubw           %%mm5,%%mm4 \n\t"              //wsptr[2,tmp11],[xxx],[2,tmp14],[xxx]
        "paddw           %%mm7,%%mm6 \n\t"              //wsptr[3,tmp10],[xxx],[3,tmp13],[xxx]

        "movq            %%mm3,%%mm5 \n\t"
        "punpcklwd       %%mm6,%%mm3 \n\t"              //wsptr[2,tmp10],[3,tmp10],[xxx],[xxx]

        "psubw           %%mm7,%%mm2 \n\t"              //wsptr[3,tmp11],[xxx],[3,tmp14],[xxx]
        "punpckhwd       %%mm6,%%mm5 \n\t"              //wsptr[2,tmp13],[3,tmp13],[xxx],[xxx]

        "movq            %%mm4,%%mm7 \n\t"
        "punpckldq       %%mm5,%%mm3 \n\t" //wsptr[2,tmp10],[3,tmp10],[2,tmp13],[3,tmp13]

        "punpcklwd       %%mm2,%%mm4 \n\t"              //wsptr[2,tmp11],[3,tmp11],[xxx],[xxx]

        "punpckhwd       %%mm2,%%mm7 \n\t"              //wsptr[2,tmp14],[3,tmp14],[xxx],[xxx]

        "punpckldq       %%mm7,%%mm4 \n\t" //wsptr[2,tmp11],[3,tmp11],[2,tmp14],[3,tmp14]
        "movq            %%mm1,%%mm6 \n\t"

//      mm0 =   ;wsptr[0,tmp10],[1,tmp10],[0,tmp13],[1,tmp13]
//      mm1 =   ;wsptr[0,tmp11],[1,tmp11],[0,tmp14],[1,tmp14]


        "movq            %%mm0,%%mm2 \n\t"
        "punpckhdq       %%mm4,%%mm6 \n\t" //wsptr[0,tmp14],[1,tmp14],[2,tmp14],[3,tmp14]

        "punpckldq       %%mm4,%%mm1 \n\t" //wsptr[0,tmp11],[1,tmp11],[2,tmp11],[3,tmp11]
        "psllw           $2,%%mm6    \n\t"

        "pmulhw          _fix_141,%%mm6 \n\t"
        "punpckldq       %%mm3,%%mm0 \n\t" //wsptr[0,tmp10],[1,tmp10],[2,tmp10],[3,tmp10]

        "punpckhdq       %%mm3,%%mm2 \n\t" //wsptr[0,tmp13],[1,tmp13],[2,tmp13],[3,tmp13]
        "movq            %%mm0,%%mm7 \n\t"

//    tmp0 = tmp10 + tmp13;
//    tmp3 = tmp10 - tmp13;
        "paddw           %%mm2,%%mm0 \n\t" //[0,tmp0],[1,tmp0],[2,tmp0],[3,tmp0]
        "psubw           %%mm2,%%mm7 \n\t" //[0,tmp3],[1,tmp3],[2,tmp3],[3,tmp3]

//    tmp12 = MULTIPLY(tmp14, FIX_1_414213562) - tmp13;
        "psubw           %%mm2,%%mm6 \n\t" //wsptr[0,tmp12],[1,tmp12],[2,tmp12],[3,tmp12]
//    tmp1 = tmp11 + tmp12;
//    tmp2 = tmp11 - tmp12;
        "movq            %%mm1,%%mm5 \n\t"



    /* Odd part */

//    z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
//    z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
//    z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
//    z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
        "movq            8*0(%%esi),%%mm3 \n\t" //wsptr[0,0],[0,1],[0,2],[0,3]
        "paddw           %%mm6,%%mm1 \n\t" //[0,tmp1],[1,tmp1],[2,tmp1],[3,tmp1]

        "movq            8*1(%%esi),%%mm4 \n\t" //wsptr[0,4],[0,5],[0,6],[0,7]
        "psubw           %%mm6,%%mm5 \n\t" //[0,tmp2],[1,tmp2],[2,tmp2],[3,tmp2]

        "movq            %%mm3,%%mm6 \n\t"
        "punpckldq       %%mm4,%%mm3 \n\t"              //wsptr[0,0],[0,1],[0,4],[0,5]

        "punpckhdq       %%mm6,%%mm4 \n\t"              //wsptr[0,6],[0,7],[0,2],[0,3]
        "movq            %%mm3,%%mm2 \n\t"

//Save tmp0 and tmp1 in wsptr
        "movq            %%mm0,8*0(%%esi) \n\t" //save tmp0
        "paddw           %%mm4,%%mm2 \n\t"              //wsptr[xxx],[0,z11],[xxx],[0,z13]


//Continue with z10 --- z13
        "movq            8*2(%%esi),%%mm6 \n\t" //wsptr[1,0],[1,1],[1,2],[1,3]
        "psubw           %%mm4,%%mm3 \n\t"              //wsptr[xxx],[0,z12],[xxx],[0,z10]

        "movq            8*3(%%esi),%%mm0 \n\t" //wsptr[1,4],[1,5],[1,6],[1,7]
        "movq            %%mm6,%%mm4 \n\t"

        "movq            %%mm1,8*1(%%esi) \n\t" //save tmp1
        "punpckldq       %%mm0,%%mm6 \n\t"              //wsptr[1,0],[1,1],[1,4],[1,5]

        "punpckhdq       %%mm4,%%mm0 \n\t"              //wsptr[1,6],[1,7],[1,2],[1,3]
        "movq            %%mm6,%%mm1 \n\t"

//Save tmp2 and tmp3 in wsptr
        "paddw           %%mm0,%%mm6 \n\t"      //wsptr[xxx],[1,z11],[xxx],[1,z13]
        "movq            %%mm2,%%mm4 \n\t"

//Continue with z10 --- z13
        "movq            %%mm5,8*2(%%esi) \n\t" //save tmp2
        "punpcklwd       %%mm6,%%mm2 \n\t"      //wsptr[xxx],[xxx],[0,z11],[1,z11]

        "psubw           %%mm0,%%mm1 \n\t"      //wsptr[xxx],[1,z12],[xxx],[1,z10]
        "punpckhwd       %%mm6,%%mm4 \n\t"      //wsptr[xxx],[xxx],[0,z13],[1,z13]

        "movq            %%mm3,%%mm0 \n\t"
        "punpcklwd       %%mm1,%%mm3 \n\t"      //wsptr[xxx],[xxx],[0,z12],[1,z12]

        "movq            %%mm7,8*3(%%esi) \n\t" //save tmp3
        "punpckhwd       %%mm1,%%mm0 \n\t"