michael@0: /********************************************************************
michael@0:  *                                                                  *
michael@0:  * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE.   *
michael@0:  *                                                                  *
michael@0:  * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
michael@0:  * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
michael@0:  * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
michael@0:  *                                                                  *
michael@0:  * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002    *
michael@0:  * BY THE Xiph.Org FOUNDATION http://www.xiph.org/                  *
michael@0:  *                                                                  *
michael@0:  ********************************************************************
michael@0: 
michael@0:  function: normalized modified discrete cosine transform
michael@0:            power of two length transform only [64 <= n ]
michael@0:  last mod: $Id: mdct.c,v 1.9 2002/10/16 09:17:39 xiphmont Exp $
michael@0: 
michael@0:  Original algorithm adapted long ago from _The use of multirate filter
michael@0:  banks for coding of high quality digital audio_, by T. Sporer,
michael@0:  K. Brandenburg and B. Edler, collection of the European Signal
michael@0:  Processing Conference (EUSIPCO), Amsterdam, June 1992, Vol.1, pp
michael@0:  211-214
michael@0: 
michael@0:  The below code implements an algorithm that no longer looks much like
michael@0:  that presented in the paper, but the basic structure remains if you
michael@0:  dig deep enough to see it.
michael@0: 
michael@0:  This module DOES NOT INCLUDE code to generate/apply the window
michael@0:  function.  Everybody has their own weird favorite including me... I
michael@0:  happen to like the properties of y=sin(.5PI*sin^2(x)), but others may
michael@0:  vehemently disagree.
michael@0: 
michael@0:  ********************************************************************/
michael@0: 
michael@0: #include "ivorbiscodec.h"
michael@0: #include "codebook.h"
michael@0: #include "misc.h"
michael@0: #include "mdct.h"
michael@0: #include "mdct_lookup.h"
michael@0: 
michael@0: 
michael@0: /* 8 point butterfly (in place) */
michael@0: STIN void mdct_butterfly_8(DATA_TYPE *x){
michael@0: 
michael@0:   REG_TYPE r0   = x[4] + x[0];
michael@0:   REG_TYPE r1   = x[4] - x[0];
michael@0:   REG_TYPE r2   = x[5] + x[1];
michael@0:   REG_TYPE r3   = x[5] - x[1];
michael@0:   REG_TYPE r4   = x[6] + x[2];
michael@0:   REG_TYPE r5   = x[6] - x[2];
michael@0:   REG_TYPE r6   = x[7] + x[3];
michael@0:   REG_TYPE r7   = x[7] - x[3];
michael@0: 
michael@0: 	   x[0] = r5   + r3;
michael@0: 	   x[1] = r7   - r1;
michael@0: 	   x[2] = r5   - r3;
michael@0: 	   x[3] = r7   + r1;
michael@0:            x[4] = r4   - r0;
michael@0: 	   x[5] = r6   - r2;
michael@0:            x[6] = r4   + r0;
michael@0: 	   x[7] = r6   + r2;
michael@0: 	   MB();
michael@0: }
michael@0: 
michael@0: /* 16 point butterfly (in place, 4 register) */
michael@0: STIN void mdct_butterfly_16(DATA_TYPE *x){
michael@0: 
michael@0:   REG_TYPE r0, r1;
michael@0: 
michael@0: 	   r0 = x[ 0] - x[ 8]; x[ 8] += x[ 0];
michael@0: 	   r1 = x[ 1] - x[ 9]; x[ 9] += x[ 1];
michael@0: 	   x[ 0] = MULT31((r0 + r1) , cPI2_8);
michael@0: 	   x[ 1] = MULT31((r1 - r0) , cPI2_8);
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[10] - x[ 2]; x[10] += x[ 2];
michael@0: 	   r1 = x[ 3] - x[11]; x[11] += x[ 3];
michael@0: 	   x[ 2] = r1; x[ 3] = r0;
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[12] - x[ 4]; x[12] += x[ 4];
michael@0: 	   r1 = x[13] - x[ 5]; x[13] += x[ 5];
michael@0: 	   x[ 4] = MULT31((r0 - r1) , cPI2_8);
michael@0: 	   x[ 5] = MULT31((r0 + r1) , cPI2_8);
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[14] - x[ 6]; x[14] += x[ 6];
michael@0: 	   r1 = x[15] - x[ 7]; x[15] += x[ 7];
michael@0: 	   x[ 6] = r0; x[ 7] = r1;
michael@0: 	   MB();
michael@0: 
michael@0: 	   mdct_butterfly_8(x);
michael@0: 	   mdct_butterfly_8(x+8);
michael@0: }
michael@0: 
michael@0: /* 32 point butterfly (in place, 4 register) */
michael@0: STIN void mdct_butterfly_32(DATA_TYPE *x){
michael@0: 
michael@0:   REG_TYPE r0, r1;
michael@0: 
michael@0: 	   r0 = x[30] - x[14]; x[30] += x[14];           
michael@0: 	   r1 = x[31] - x[15]; x[31] += x[15];
michael@0: 	   x[14] = r0; x[15] = r1;
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[28] - x[12]; x[28] += x[12];           
michael@0: 	   r1 = x[29] - x[13]; x[29] += x[13];
michael@0: 	   XNPROD31( r0, r1, cPI1_8, cPI3_8, &x[12], &x[13] );
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[26] - x[10]; x[26] += x[10];
michael@0: 	   r1 = x[27] - x[11]; x[27] += x[11];
michael@0: 	   x[10] = MULT31((r0 - r1) , cPI2_8);
michael@0: 	   x[11] = MULT31((r0 + r1) , cPI2_8);
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[24] - x[ 8]; x[24] += x[ 8];
michael@0: 	   r1 = x[25] - x[ 9]; x[25] += x[ 9];
michael@0: 	   XNPROD31( r0, r1, cPI3_8, cPI1_8, &x[ 8], &x[ 9] );
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[22] - x[ 6]; x[22] += x[ 6];
michael@0: 	   r1 = x[ 7] - x[23]; x[23] += x[ 7];
michael@0: 	   x[ 6] = r1; x[ 7] = r0;
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[ 4] - x[20]; x[20] += x[ 4];
michael@0: 	   r1 = x[ 5] - x[21]; x[21] += x[ 5];
michael@0: 	   XPROD31 ( r0, r1, cPI3_8, cPI1_8, &x[ 4], &x[ 5] );
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[ 2] - x[18]; x[18] += x[ 2];
michael@0: 	   r1 = x[ 3] - x[19]; x[19] += x[ 3];
michael@0: 	   x[ 2] = MULT31((r1 + r0) , cPI2_8);
michael@0: 	   x[ 3] = MULT31((r1 - r0) , cPI2_8);
michael@0: 	   MB();
michael@0: 
michael@0: 	   r0 = x[ 0] - x[16]; x[16] += x[ 0];
michael@0: 	   r1 = x[ 1] - x[17]; x[17] += x[ 1];
michael@0: 	   XPROD31 ( r0, r1, cPI1_8, cPI3_8, &x[ 0], &x[ 1] );
michael@0: 	   MB();
michael@0: 
michael@0: 	   mdct_butterfly_16(x);
michael@0: 	   mdct_butterfly_16(x+16);
michael@0: }
michael@0: 
michael@0: /* N/stage point generic N stage butterfly (in place, 2 register) */
michael@0: STIN void mdct_butterfly_generic(DATA_TYPE *x,int points,int step){
michael@0: 
michael@0:   LOOKUP_T *T   = sincos_lookup0;
michael@0:   DATA_TYPE *x1        = x + points      - 8;
michael@0:   DATA_TYPE *x2        = x + (points>>1) - 8;
michael@0:   REG_TYPE   r0;
michael@0:   REG_TYPE   r1;
michael@0: 
michael@0:   do{
michael@0:     r0 = x1[6] - x2[6]; x1[6] += x2[6];
michael@0:     r1 = x2[7] - x1[7]; x1[7] += x2[7];
michael@0:     XPROD31( r1, r0, T[0], T[1], &x2[6], &x2[7] ); T+=step;
michael@0: 
michael@0:     r0 = x1[4] - x2[4]; x1[4] += x2[4];
michael@0:     r1 = x2[5] - x1[5]; x1[5] += x2[5];
michael@0:     XPROD31( r1, r0, T[0], T[1], &x2[4], &x2[5] ); T+=step;
michael@0: 
michael@0:     r0 = x1[2] - x2[2]; x1[2] += x2[2];
michael@0:     r1 = x2[3] - x1[3]; x1[3] += x2[3];
michael@0:     XPROD31( r1, r0, T[0], T[1], &x2[2], &x2[3] ); T+=step;
michael@0: 
michael@0:     r0 = x1[0] - x2[0]; x1[0] += x2[0];
michael@0:     r1 = x2[1] - x1[1]; x1[1] += x2[1];
michael@0:     XPROD31( r1, r0, T[0], T[1], &x2[0], &x2[1] ); T+=step;
michael@0: 
michael@0:     x1-=8; x2-=8;
michael@0:   }while(T<sincos_lookup0+1024);
michael@0:   do{
michael@0:     r0 = x1[6] - x2[6]; x1[6] += x2[6];
michael@0:     r1 = x1[7] - x2[7]; x1[7] += x2[7];
michael@0:     XNPROD31( r0, r1, T[0], T[1], &x2[6], &x2[7] ); T-=step;
michael@0: 
michael@0:     r0 = x1[4] - x2[4]; x1[4] += x2[4];
michael@0:     r1 = x1[5] - x2[5]; x1[5] += x2[5];
michael@0:     XNPROD31( r0, r1, T[0], T[1], &x2[4], &x2[5] ); T-=step;
michael@0: 
michael@0:     r0 = x1[2] - x2[2]; x1[2] += x2[2];
michael@0:     r1 = x1[3] - x2[3]; x1[3] += x2[3];
michael@0:     XNPROD31( r0, r1, T[0], T[1], &x2[2], &x2[3] ); T-=step;
michael@0: 
michael@0:     r0 = x1[0] - x2[0]; x1[0] += x2[0];
michael@0:     r1 = x1[1] - x2[1]; x1[1] += x2[1];
michael@0:     XNPROD31( r0, r1, T[0], T[1], &x2[0], &x2[1] ); T-=step;
michael@0: 
michael@0:     x1-=8; x2-=8;
michael@0:   }while(T>sincos_lookup0);
michael@0:   do{
michael@0:     r0 = x2[6] - x1[6]; x1[6] += x2[6];
michael@0:     r1 = x2[7] - x1[7]; x1[7] += x2[7];
michael@0:     XPROD31( r0, r1, T[0], T[1], &x2[6], &x2[7] ); T+=step;
michael@0: 
michael@0:     r0 = x2[4] - x1[4]; x1[4] += x2[4];
michael@0:     r1 = x2[5] - x1[5]; x1[5] += x2[5];
michael@0:     XPROD31( r0, r1, T[0], T[1], &x2[4], &x2[5] ); T+=step;
michael@0: 
michael@0:     r0 = x2[2] - x1[2]; x1[2] += x2[2];
michael@0:     r1 = x2[3] - x1[3]; x1[3] += x2[3];
michael@0:     XPROD31( r0, r1, T[0], T[1], &x2[2], &x2[3] ); T+=step;
michael@0: 
michael@0:     r0 = x2[0] - x1[0]; x1[0] += x2[0];
michael@0:     r1 = x2[1] - x1[1]; x1[1] += x2[1];
michael@0:     XPROD31( r0, r1, T[0], T[1], &x2[0], &x2[1] ); T+=step;
michael@0: 
michael@0:     x1-=8; x2-=8;
michael@0:   }while(T<sincos_lookup0+1024);
michael@0:   do{
michael@0:     r0 = x1[6] - x2[6]; x1[6] += x2[6];
michael@0:     r1 = x2[7] - x1[7]; x1[7] += x2[7];
michael@0:     XNPROD31( r1, r0, T[0], T[1], &x2[6], &x2[7] ); T-=step;
michael@0: 
michael@0:     r0 = x1[4] - x2[4]; x1[4] += x2[4];
michael@0:     r1 = x2[5] - x1[5]; x1[5] += x2[5];
michael@0:     XNPROD31( r1, r0, T[0], T[1], &x2[4], &x2[5] ); T-=step;
michael@0: 
michael@0:     r0 = x1[2] - x2[2]; x1[2] += x2[2];
michael@0:     r1 = x2[3] - x1[3]; x1[3] += x2[3];
michael@0:     XNPROD31( r1, r0, T[0], T[1], &x2[2], &x2[3] ); T-=step;
michael@0: 
michael@0:     r0 = x1[0] - x2[0]; x1[0] += x2[0];
michael@0:     r1 = x2[1] - x1[1]; x1[1] += x2[1];
michael@0:     XNPROD31( r1, r0, T[0], T[1], &x2[0], &x2[1] ); T-=step;
michael@0: 
michael@0:     x1-=8; x2-=8;
michael@0:   }while(T>sincos_lookup0);
michael@0: }
michael@0: 
michael@0: STIN void mdct_butterflies(DATA_TYPE *x,int points,int shift){
michael@0: 
michael@0:   int stages=8-shift;
michael@0:   int i,j;
michael@0:   
michael@0:   for(i=0;--stages>0;i++){
michael@0:     for(j=0;j<(1<<i);j++)
michael@0:       mdct_butterfly_generic(x+(points>>i)*j,points>>i,4<<(i+shift));
michael@0:   }
michael@0: 
michael@0:   for(j=0;j<points;j+=32)
michael@0:     mdct_butterfly_32(x+j);
michael@0: 
michael@0: }
michael@0: 
michael@0: static unsigned char bitrev[16]={0,8,4,12,2,10,6,14,1,9,5,13,3,11,7,15};
michael@0: 
michael@0: STIN int bitrev12(int x){
michael@0:   return bitrev[x>>8]|(bitrev[(x&0x0f0)>>4]<<4)|(((int)bitrev[x&0x00f])<<8);
michael@0: }
michael@0: 
michael@0: STIN void mdct_bitreverse(DATA_TYPE *x,int n,int step,int shift){
michael@0: 
michael@0:   int          bit   = 0;
michael@0:   DATA_TYPE   *w0    = x;
michael@0:   DATA_TYPE   *w1    = x = w0+(n>>1);
michael@0:   LOOKUP_T    *T = (step>=4)?(sincos_lookup0+(step>>1)):sincos_lookup1;
michael@0:   LOOKUP_T    *Ttop  = T+1024;
michael@0:   DATA_TYPE    r2;
michael@0: 
michael@0:   do{
michael@0:     DATA_TYPE r3     = bitrev12(bit++);
michael@0:     DATA_TYPE *x0    = x + ((r3 ^ 0xfff)>>shift) -1;
michael@0:     DATA_TYPE *x1    = x + (r3>>shift);
michael@0: 
michael@0:     REG_TYPE  r0     = x0[0]  + x1[0];
michael@0:     REG_TYPE  r1     = x1[1]  - x0[1];
michael@0: 
michael@0: 	      XPROD32( r0, r1, T[1], T[0], &r2, &r3 ); T+=step;
michael@0: 
michael@0: 	      w1    -= 4;
michael@0: 
michael@0: 	      r0     = (x0[1] + x1[1])>>1;
michael@0:               r1     = (x0[0] - x1[0])>>1;
michael@0: 	      w0[0]  = r0     + r2;
michael@0: 	      w0[1]  = r1     + r3;
michael@0: 	      w1[2]  = r0     - r2;
michael@0: 	      w1[3]  = r3     - r1;
michael@0: 
michael@0: 	      r3     = bitrev12(bit++);
michael@0:               x0     = x + ((r3 ^ 0xfff)>>shift) -1;
michael@0:               x1     = x + (r3>>shift);
michael@0: 
michael@0:               r0     = x0[0]  + x1[0];
michael@0:               r1     = x1[1]  - x0[1];
michael@0: 
michael@0: 	      XPROD32( r0, r1, T[1], T[0], &r2, &r3 ); T+=step;
michael@0: 
michael@0:               r0     = (x0[1] + x1[1])>>1;
michael@0:               r1     = (x0[0] - x1[0])>>1;
michael@0: 	      w0[2]  = r0     + r2;
michael@0: 	      w0[3]  = r1     + r3;
michael@0: 	      w1[0]  = r0     - r2;
michael@0: 	      w1[1]  = r3     - r1;
michael@0: 
michael@0: 	      w0    += 4;
michael@0:   }while(T<Ttop);
michael@0:   do{
michael@0:     DATA_TYPE r3     = bitrev12(bit++);
michael@0:     DATA_TYPE *x0    = x + ((r3 ^ 0xfff)>>shift) -1;
michael@0:     DATA_TYPE *x1    = x + (r3>>shift);
michael@0: 
michael@0:     REG_TYPE  r0     = x0[0]  + x1[0];
michael@0:     REG_TYPE  r1     = x1[1]  - x0[1];
michael@0: 
michael@0: 	      T-=step; XPROD32( r0, r1, T[0], T[1], &r2, &r3 );
michael@0: 
michael@0: 	      w1    -= 4;
michael@0: 
michael@0: 	      r0     = (x0[1] + x1[1])>>1;
michael@0:               r1     = (x0[0] - x1[0])>>1;
michael@0: 	      w0[0]  = r0     + r2;
michael@0: 	      w0[1]  = r1     + r3;
michael@0: 	      w1[2]  = r0     - r2;
michael@0: 	      w1[3]  = r3     - r1;
michael@0: 
michael@0: 	      r3     = bitrev12(bit++);
michael@0:               x0     = x + ((r3 ^ 0xfff)>>shift) -1;
michael@0:               x1     = x + (r3>>shift);
michael@0: 
michael@0:               r0     = x0[0]  + x1[0];
michael@0:               r1     = x1[1]  - x0[1];
michael@0: 
michael@0: 	      T-=step; XPROD32( r0, r1, T[0], T[1], &r2, &r3 );
michael@0: 
michael@0:               r0     = (x0[1] + x1[1])>>1;
michael@0:               r1     = (x0[0] - x1[0])>>1;
michael@0: 	      w0[2]  = r0     + r2;
michael@0: 	      w0[3]  = r1     + r3;
michael@0: 	      w1[0]  = r0     - r2;
michael@0: 	      w1[1]  = r3     - r1;
michael@0: 
michael@0: 	      w0    += 4;
michael@0:   }while(w0<w1);
michael@0: }
michael@0: 
michael@0: void mdct_backward(int n, DATA_TYPE *in, DATA_TYPE *out){
michael@0:   int n2=n>>1;
michael@0:   int n4=n>>2;
michael@0:   DATA_TYPE *iX;
michael@0:   DATA_TYPE *oX;
michael@0:   LOOKUP_T *T;
michael@0:   LOOKUP_T *V;
michael@0:   int shift;
michael@0:   int step;
michael@0: 
michael@0:   for (shift=6;!(n&(1<<shift));shift++);
michael@0:   shift=13-shift;
michael@0:   step=2<<shift;
michael@0:    
michael@0:   /* rotate */
michael@0: 
michael@0:   iX            = in+n2-7;
michael@0:   oX            = out+n2+n4;
michael@0:   T             = sincos_lookup0;
michael@0: 
michael@0:   do{
michael@0:     oX-=4;
michael@0:     XPROD31( iX[4], iX[6], T[0], T[1], &oX[2], &oX[3] ); T+=step;
michael@0:     XPROD31( iX[0], iX[2], T[0], T[1], &oX[0], &oX[1] ); T+=step;
michael@0:     iX-=8;
michael@0:   }while(iX>=in+n4);
michael@0:   do{
michael@0:     oX-=4;
michael@0:     XPROD31( iX[4], iX[6], T[1], T[0], &oX[2], &oX[3] ); T-=step;
michael@0:     XPROD31( iX[0], iX[2], T[1], T[0], &oX[0], &oX[1] ); T-=step;
michael@0:     iX-=8;
michael@0:   }while(iX>=in);
michael@0: 
michael@0:   iX            = in+n2-8;
michael@0:   oX            = out+n2+n4;
michael@0:   T             = sincos_lookup0;
michael@0: 
michael@0:   do{
michael@0:     T+=step; XNPROD31( iX[6], iX[4], T[0], T[1], &oX[0], &oX[1] );
michael@0:     T+=step; XNPROD31( iX[2], iX[0], T[0], T[1], &oX[2], &oX[3] );
michael@0:     iX-=8;
michael@0:     oX+=4;
michael@0:   }while(iX>=in+n4);
michael@0:   do{
michael@0:     T-=step; XNPROD31( iX[6], iX[4], T[1], T[0], &oX[0], &oX[1] );
michael@0:     T-=step; XNPROD31( iX[2], iX[0], T[1], T[0], &oX[2], &oX[3] );
michael@0:     iX-=8;
michael@0:     oX+=4;
michael@0:   }while(iX>=in);
michael@0: 
michael@0:   mdct_butterflies(out+n2,n2,shift);
michael@0:   mdct_bitreverse(out,n,step,shift);
michael@0: 
michael@0:   /* rotate + window */
michael@0: 
michael@0:   step>>=2;
michael@0:   {
michael@0:     DATA_TYPE *oX1=out+n2+n4;
michael@0:     DATA_TYPE *oX2=out+n2+n4;
michael@0:     DATA_TYPE *iX =out;
michael@0: 
michael@0:     switch(step) {
michael@0:       default: {
michael@0:         T=(step>=4)?(sincos_lookup0+(step>>1)):sincos_lookup1;
michael@0:         do{
michael@0:           oX1-=4;
michael@0: 	  XPROD31( iX[0], -iX[1], T[0], T[1], &oX1[3], &oX2[0] ); T+=step;
michael@0: 	  XPROD31( iX[2], -iX[3], T[0], T[1], &oX1[2], &oX2[1] ); T+=step;
michael@0: 	  XPROD31( iX[4], -iX[5], T[0], T[1], &oX1[1], &oX2[2] ); T+=step;
michael@0: 	  XPROD31( iX[6], -iX[7], T[0], T[1], &oX1[0], &oX2[3] ); T+=step;
michael@0: 	  oX2+=4;
michael@0: 	  iX+=8;
michael@0: 	}while(iX<oX1);
michael@0: 	break;
michael@0:       }
michael@0: 
michael@0:       case 1: {
michael@0:         /* linear interpolation between table values: offset=0.5, step=1 */
michael@0: 	REG_TYPE  t0,t1,v0,v1;
michael@0:         T         = sincos_lookup0;
michael@0:         V         = sincos_lookup1;
michael@0: 	t0        = (*T++)>>1;
michael@0: 	t1        = (*T++)>>1;
michael@0:         do{
michael@0:           oX1-=4;
michael@0: 
michael@0: 	  t0 += (v0 = (*V++)>>1);
michael@0: 	  t1 += (v1 = (*V++)>>1);
michael@0: 	  XPROD31( iX[0], -iX[1], t0, t1, &oX1[3], &oX2[0] );
michael@0: 	  v0 += (t0 = (*T++)>>1);
michael@0: 	  v1 += (t1 = (*T++)>>1);
michael@0: 	  XPROD31( iX[2], -iX[3], v0, v1, &oX1[2], &oX2[1] );
michael@0: 	  t0 += (v0 = (*V++)>>1);
michael@0: 	  t1 += (v1 = (*V++)>>1);
michael@0: 	  XPROD31( iX[4], -iX[5], t0, t1, &oX1[1], &oX2[2] );
michael@0: 	  v0 += (t0 = (*T++)>>1);
michael@0: 	  v1 += (t1 = (*T++)>>1);
michael@0: 	  XPROD31( iX[6], -iX[7], v0, v1, &oX1[0], &oX2[3] );
michael@0: 
michael@0: 	  oX2+=4;
michael@0: 	  iX+=8;
michael@0: 	}while(iX<oX1);
michael@0: 	break;
michael@0:       }
michael@0: 
michael@0:       case 0: {
michael@0:         /* linear interpolation between table values: offset=0.25, step=0.5 */
michael@0: 	REG_TYPE  t0,t1,v0,v1,q0,q1;
michael@0:         T         = sincos_lookup0;
michael@0:         V         = sincos_lookup1;
michael@0: 	t0        = *T++;
michael@0: 	t1        = *T++;
michael@0:         do{
michael@0:           oX1-=4;
michael@0: 
michael@0: 	  v0  = *V++;
michael@0: 	  v1  = *V++;
michael@0: 	  t0 +=  (q0 = (v0-t0)>>2);
michael@0: 	  t1 +=  (q1 = (v1-t1)>>2);
michael@0: 	  XPROD31( iX[0], -iX[1], t0, t1, &oX1[3], &oX2[0] );
michael@0: 	  t0  = v0-q0;
michael@0: 	  t1  = v1-q1;
michael@0: 	  XPROD31( iX[2], -iX[3], t0, t1, &oX1[2], &oX2[1] );
michael@0: 
michael@0: 	  t0  = *T++;
michael@0: 	  t1  = *T++;
michael@0: 	  v0 += (q0 = (t0-v0)>>2);
michael@0: 	  v1 += (q1 = (t1-v1)>>2);
michael@0: 	  XPROD31( iX[4], -iX[5], v0, v1, &oX1[1], &oX2[2] );
michael@0: 	  v0  = t0-q0;
michael@0: 	  v1  = t1-q1;
michael@0: 	  XPROD31( iX[6], -iX[7], v0, v1, &oX1[0], &oX2[3] );
michael@0: 
michael@0: 	  oX2+=4;
michael@0: 	  iX+=8;
michael@0: 	}while(iX<oX1);
michael@0: 	break;
michael@0:       }
michael@0:     }
michael@0: 
michael@0:     iX=out+n2+n4;
michael@0:     oX1=out+n4;
michael@0:     oX2=oX1;
michael@0: 
michael@0:     do{
michael@0:       oX1-=4;
michael@0:       iX-=4;
michael@0: 
michael@0:       oX2[0] = -(oX1[3] = iX[3]);
michael@0:       oX2[1] = -(oX1[2] = iX[2]);
michael@0:       oX2[2] = -(oX1[1] = iX[1]);
michael@0:       oX2[3] = -(oX1[0] = iX[0]);
michael@0: 
michael@0:       oX2+=4;
michael@0:     }while(oX2<iX);
michael@0: 
michael@0:     iX=out+n2+n4;
michael@0:     oX1=out+n2+n4;
michael@0:     oX2=out+n2;
michael@0: 
michael@0:     do{
michael@0:       oX1-=4;
michael@0:       oX1[0]= iX[3];
michael@0:       oX1[1]= iX[2];
michael@0:       oX1[2]= iX[1];
michael@0:       oX1[3]= iX[0];
michael@0:       iX+=4;
michael@0:     }while(oX1>oX2);
michael@0:   }
michael@0: }
michael@0: