fastfft.c

MPEG-4编解码的实现(包括MPEG4视音频编解码)

     st8 = st8 + st5;
     st8 = st8 * K707106781[0];
     st5 = st4 - st1;
     st5 = st5 * K707106781[0];
     st4 = st4 + st1;
     st4 = st4 * K707106781[0];
     st1 = st2 - st6;
     tmp704 = st8;
     st8 = st1 * K195090322[0];
     st2 = st2 + st6;
     st1 = st1 * K980785280[0];
     st6 = st3 - st4;
     tmp705 = st1;
     st1 = st6 * K555570233[0];
     st3 = st3 + st4;
     st6 = st6 * K831469612[0];
     st4 = st7 - st5;
     tmp706 = st8;
     st8 = st4 * K980785280[0];
     st7 = st7 + st5;
     st4 = st4 * K195090322[0];
     st5 = tmp703 - tmp704;
     tmp707 = st4;
     st4 = st5 * K831469612[0];
     tmp708 = st8;
     st8 = tmp703 + tmp704;
     st5 = st5 * K555570233[0];
     st1 = st1 - st4;
     st4 = st2 * K831469612[0];
     tmp709 = st4;
     st4 = st1 - tmp702;
     tmp710 = st4;
     st4 = st7 * K555570233[0];
     st1 = st1 + tmp702;
     st2 = st2 * K555570233[0];
     st5 = st5 + st6;
     st7 = st7 * K831469612[0];
     st6 = st5 + tmp701;
     tmp711 = st2;
     st2 = st3 * K980785280[0];
     st5 = tmp701 - st5;
     tmp712 = st2;
     st2 = st8 * K195090322[0];
     tmp713 = st2;
     st2 = tmp673 - st6;
     st8 = st8 * K980785280[0];
     st3 = st3 * K195090322[0];
     st6 = tmp673 + st6;
     c_re(input[21]) = st2;
     st2 = tmp672 - st5;
     st5 = tmp672 + st5;
     c_re(input[5]) = st6;
     st6 = tmp659 - tmp710;
     c_im(input[29]) = st2;
     st2 = tmp659 + tmp710;
     c_im(input[13]) = st5;
     st5 = tmp658 - st1;
     st1 = tmp658 + st1;
     c_re(input[29]) = st6;
     st6 = tmp706 - tmp708;
     c_re(input[13]) = st2;
     st2 = st6 - tmp695;
     st6 = st6 + tmp695;
     c_im(input[21]) = st5;
     st5 = tmp705 + tmp707;
     c_im(input[5]) = st1;
     st1 = st5 + tmp700;
     st5 = tmp700 - st5;
     tmp714 = st8;
     st8 = tmp661 - st1;
     st1 = tmp661 + st1;
     c_re(input[23]) = st8;
     st8 = tmp671 - st5;
     st5 = tmp671 + st5;
     c_re(input[7]) = st1;
     st1 = tmp660 - st2;
     st2 = tmp660 + st2;
     c_im(input[31]) = st8;
     st8 = tmp662 - st6;
     st6 = tmp662 + st6;
     st4 = tmp709 - st4;
     c_im(input[15]) = st5;
     st5 = st4 - tmp699;
     st4 = st4 + tmp699;
     st7 = tmp711 + st7;
     c_re(input[31]) = st1;
     st1 = st7 + tmp698;
     st7 = tmp698 - st7;
     c_re(input[15]) = st2;
     st2 = tmp670 - st1;
     st1 = tmp670 + st1;
     c_im(input[23]) = st8;
     st8 = tmp669 - st7;
     st7 = tmp669 + st7;
     c_im(input[7]) = st6;
     st6 = tmp663 - st5;
     st5 = tmp663 + st5;
     c_re(input[19]) = st2;
     st2 = tmp664 - st4;
     st4 = tmp664 + st4;
     c_re(input[3]) = st1;
     st1 = tmp712 - tmp713;
     c_im(input[27]) = st8;
     st8 = st1 - tmp696;
     st1 = st1 + tmp696;
     st3 = tmp714 + st3;
     c_im(input[11]) = st7;
     st7 = st3 + tmp697;
     st3 = tmp697 - st3;
     c_re(input[27]) = st6;
     st6 = tmp668 - st7;
     st7 = tmp668 + st7;
     c_re(input[11]) = st5;
     st5 = tmp667 - st3;
     st3 = tmp667 + st3;
     c_im(input[19]) = st2;
     st2 = tmp665 - st8;
     st8 = tmp665 + st8;
     c_im(input[3]) = st4;
     st4 = tmp666 - st1;
     st1 = tmp666 + st1;
     c_re(input[17]) = st6;
     c_re(input[1]) = st7;
     c_im(input[25]) = st5;
     c_im(input[9]) = st3;
     c_re(input[25]) = st2;
     c_re(input[9]) = st8;
     c_im(input[17]) = st4;
     c_im(input[1]) = st1;
}

void  PFFTW(64)(fftw_complex *input)
{
     PFFTW(twiddle_4)(input, PFFTW(W_64), 16);
     PFFTW(16)(input );
     PFFTW(16)(input + 16);
     PFFTW(16)(input + 32);
     PFFTW(16)(input + 48);
}

void PFFTW(128)(fftw_complex *input)
{
     PFFTW(twiddle_4)(input, PFFTW(W_128), 32);
     PFFTW(32)(input );
     PFFTW(32)(input + 32);
     PFFTW(32)(input + 64);
     PFFTW(32)(input + 96);
}

void PFFTW(512)(fftw_complex *input)
{
     PFFTW(twiddle_4)(input, PFFTW(W_512), 128);
     PFFTW(128)(input );
     PFFTW(128)(input + 128);
     PFFTW(128)(input + 256);
     PFFTW(128)(input + 384);
}
///////////////////////////////////////////////////////////////
void PFFTWI(16) (fftw_complex * input) {
     fftw_real tmp333;
     fftw_real tmp332;
     fftw_real tmp331;
     fftw_real tmp330;
     fftw_real tmp329;
     fftw_real tmp328;
     fftw_real tmp327;
     fftw_real tmp326;
     fftw_real tmp325;
     fftw_real tmp324;
     fftw_real tmp323;
     fftw_real tmp322;
     fftw_real tmp321;
     fftw_real tmp320;
     fftw_real tmp319;
     fftw_real tmp318;
     fftw_real tmp317;
     fftw_real tmp316;
     fftw_real tmp315;
     fftw_real tmp314;
     fftw_real tmp313;
     fftw_real tmp312;
     fftw_real tmp311;
     fftw_real tmp310;
     fftw_real tmp309;
     fftw_real tmp308;
     fftw_real tmp307;
     fftw_real tmp306;
     fftw_real tmp305;
     fftw_real tmp304;
     fftw_real tmp303;
     fftw_real tmp302;
     fftw_real tmp301;
     fftw_real st1;
     fftw_real st2;
     fftw_real st3;
     fftw_real st4;
     fftw_real st5;
     fftw_real st6;
     fftw_real st7;
     fftw_real st8;
     st8 = c_re(input[4]);
     st8 = st8 - c_re(input[12]);
     st7 = c_im(input[0]);
     st7 = st7 - c_im(input[8]);
     st6 = c_re(input[0]);
     st6 = st6 - c_re(input[8]);
     st5 = st8 + st7;
     st7 = st7 - st8;
     st4 = c_im(input[4]);
     st4 = st4 - c_im(input[12]);
     st3 = c_re(input[0]);
     st3 = st3 + c_re(input[8]);
     st2 = st6 - st4;
     st6 = st6 + st4;
     st1 = c_re(input[4]);
     st1 = st1 + c_re(input[12]);
     st8 = c_im(input[0]);
     st8 = st8 + c_im(input[8]);
     st4 = st3 + st1;
     st3 = st3 - st1;
     st1 = c_im(input[4]);
     st1 = st1 + c_im(input[12]);
     tmp301 = st6;
     st6 = c_re(input[2]);
     st6 = st6 + c_re(input[10]);
     tmp302 = st7;
     st7 = st8 + st1;
     st8 = st8 - st1;
     st1 = c_re(input[6]);
     st1 = st1 + c_re(input[14]);
     tmp303 = st2;
     st2 = c_im(input[2]);
     st2 = st2 + c_im(input[10]);
     tmp304 = st5;
     st5 = st6 + st1;
     st6 = st6 - st1;
     st1 = c_im(input[6]);
     st1 = st1 + c_im(input[14]);
     tmp305 = st3;
     st3 = st4 + st5;
     st4 = st4 - st5;
     st5 = st2 + st1;
     st2 = st2 - st1;
     st1 = st6 + st8;
     tmp306 = st1;
     st1 = st7 - st5;
     st7 = st7 + st5;
     st5 = tmp305 - st2;
     st2 = tmp305 + st2;
     st8 = st8 - st6;
     st6 = c_re(input[6]);
     st6 = st6 - c_re(input[14]);
     tmp307 = st8;
     st8 = c_im(input[2]);
     st8 = st8 - c_im(input[10]);
     tmp308 = st2;
     st2 = c_re(input[2]);
     st2 = st2 - c_re(input[10]);
     tmp309 = st4;
     st4 = st6 + st8;
     st8 = st8 - st6;
     st6 = c_im(input[6]);
     st6 = st6 - c_im(input[14]);
     tmp310 = st5;
     st5 = c_re(input[1]);
     st5 = st5 - c_re(input[9]);
     tmp311 = st7;
     st7 = st2 - st6;
     st2 = st2 + st6;
     st6 = c_im(input[5]);
     tmp312 = st1;
     st1 = st4 + st7;
     st7 = st7 - st4;
     st4 = st2 - st8;
     st1 = st1 * K707106781[0];
     st8 = st8 + st2;
     st7 = st7 * K707106781[0];
     st6 = st6 - c_im(input[13]);
     st4 = st4 * K707106781[0];
     st2 = tmp304 - st1;
     st8 = st8 * K707106781[0];
     tmp313 = st3;
     st3 = st5 - st6;
     st5 = st5 + st6;
     st6 = tmp303 + st7;
     tmp314 = st6;
     st6 = st3 * K923879532[0];
     st7 = tmp303 - st7;
     st3 = st3 * K382683432[0];
     st1 = tmp304 + st1;
     tmp315 = st1;
     st1 = st5 * K382683432[0];
     tmp316 = st7;
     st7 = tmp302 - st4;
     st5 = st5 * K923879532[0];
     st4 = tmp302 + st4;
     tmp317 = st4;
     st4 = tmp301 - st8;
     st8 = tmp301 + st8;
     tmp318 = st8;
     st8 = c_re(input[5]);
     st8 = st8 - c_re(input[13]);
     tmp319 = st4;
     st4 = c_im(input[1]);
     st4 = st4 - c_im(input[9]);
     tmp320 = st7;
     st7 = c_re(input[3]);
     st7 = st7 - c_re(input[11]);
     tmp321 = st2;
     st2 = st8 + st4;
     st4 = st4 - st8;
     st8 = c_im(input[7]);
     tmp322 = st5;
     st5 = st2 * K382683432[0];
     st8 = st8 - c_im(input[15]);
     st2 = st2 * K923879532[0];
     st6 = st6 - st5;
     st5 = st4 * K923879532[0];
     st2 = st2 + st3;
     st4 = st4 * K382683432[0];
     st1 = st1 - st5;
     st3 = st7 - st8;
     st4 = st4 + tmp322;
     st7 = st7 + st8;
     st8 = st3 * K382683432[0];
     st5 = c_re(input[7]);
     st3 = st3 * K923879532[0];
     st5 = st5 - c_re(input[15]);
     tmp323 = st4;
     st4 = st7 * K923879532[0];
     tmp324 = st1;
     st1 = c_im(input[3]);
     st7 = st7 * K382683432[0];
     st1 = st1 - c_im(input[11]);
     tmp325 = st2;
     st2 = c_re(input[1]);
     st2 = st2 + c_re(input[9]);
     tmp326 = st6;
     st6 = st5 + st1;
     st1 = st1 - st5;
     st5 = c_re(input[5]);
     tmp327 = st7;
     st7 = st6 * K923879532[0];
     st5 = st5 + c_re(input[13]);
     st6 = st6 * K382683432[0];
     st8 = st8 - st7;
     st7 = st1 * K382683432[0];
     st6 = st6 + st3;
     st1 = st1 * K923879532[0];
     st7 = st7 - st4;
     st3 = st2 + st5;
     st1 = st1 + tmp327;
     st2 = st2 - st5;
     st4 = tmp326 - st8;
     st8 = tmp326 + st8;
     st5 = tmp325 - st6;
     tmp328 = st2;
     st2 = tmp321 - st4;
     st4 = tmp321 + st4;
     tmp329 = st3;
     st3 = tmp314 - st8;
     st8 = tmp314 + st8;
     tmp330 = st2;
     st2 = tmp316 - st5;
     st5 = tmp316 + st5;
     c_re(input[9]) = st3;
     c_re(input[1]) = st8;
     c_re(input[5]) = st2;
     c_re(input[13]) = st5;
     st6 = tmp325 + st6;
     st3 = tmp324 - st7;
     st8 = tmp323 - st1;
     st2 = tmp315 - st6;
     st6 = tmp315 + st6;
     st5 = tmp320 - st3;
     st3 = tmp320 + st3;
     tmp331 = st5;
     st5 = tmp317 - st8;
     st8 = tmp317 + st8;
     st7 = tmp324 + st7;
     st1 = tmp323 + st1;
     tmp332 = st3;
     st3 = c_im(input[1]);
     c_im(input[1]) = st6;
     st3 = st3 + c_im(input[9]);
     c_im(input[9]) = st2;
     st2 = tmp319 - st7;
     st7 = tmp319 + st7;
     st6 = tmp318 - st1;
     st1 = tmp318 + st1;
     tmp333 = st5;
     st5 = c_im(input[5]);
     c_im(input[5]) = st4;
     st5 = st5 + c_im(input[13]);
     c_im(input[13]) = tmp330;
     st4 = st3 - st5;
     st3 = st3 + st5;
     st5 = c_re(input[3]);
     c_re(input[3]) = st7;
     st5 = st5 + c_re(input[11]);
     c_re(input[11]) = st2;
     st2 = c_re(input[7]);
     c_re(input[7]) = st6;
     st2 = st2 + c_re(input[15]);
     c_re(input[15]) = st1;
     st7 = st5 + st2;
     st5 = st5 - st2;
     st6 = c_im(input[3]);
     c_im(input[3]) = st8;
     st6 = st6 + c_im(input[11]);
     c_im(input[11]) = tmp333;
     st8 = tmp329 + st7;
     st7 = tmp329 - st7;
     st1 = st5 + st4;
     st4 = st4 - st5;
     st2 = tmp313 - st8;
     st8 = tmp313 + st8;
     st5 = st7 + tmp312;
     st7 = tmp312 - st7;
     c_re(input[8]) = st2;
     c_re(input[0]) = st8;
     c_im(input[4]) = st5;
     c_im(input[12]) = st7;
     st2 = c_im(input[7]);
     c_im(input[7]) = tmp332;
     st2 = st2 + c_im(input[15]);
     c_im(input[15]) = tmp331;
     st8 = st6 - st2;
     st6 = st6 + st2;
     st5 = tmp328 - st8;
     st8 = tmp328 + st8;
     st7 = st3 - st6;
     st2 = st5 - st1;
     st1 = st1 + st5;
     st3 = st3 + st6;
     st2 = st2 * K707106781[0];
     st6 = st4 + st8;
     st1 = st1 * K707106781[0];
     st8 = st8 - st4;
     st6 = st6 * K707106781[0];
     st4 = tmp311 + st3;
     st8 = st8 * K707106781[0];
     st3 = tmp311 - st3;
     st5 = tmp310 - st2;
     c_im(input[0]) = st4;
     c_im(input[8]) = st3;
     c_re(input[10]) = st5;
     st2 = tmp310 + st2;
     st4 = tmp309 + st7;
     st7 = tmp309 - st7;
     st3 = tmp308 - st6;
     c_re(input[2]) = st2;
     c_re(input[12]) = st4;