mtd6.asm

Tiger sharec系列DSP的短点FFT代码(4,6,8,16点FFT),全汇编实现,将DSP的性能发挥到极致,在超高实时要求应用下有意义.

#include "TS101_def.h"
.section M1DATA;
.align 4;
.var fft6Result[12]; //order is re,im,re,im...
.var win[12]="win12.dat";
.section M2DATA;

.align 4;
.var fft6input[12]="fft6input.dat";//order is re,im,re,im...

.var W[8]=-0.5,-0.5,0.866,-0.866,-0.5,-0.5,-0.866,0.866;
.section M0CODE;
.global _main;
_main: 
		k0=fft6input;;
		k0=k0+6;;
		j0=win;;
		j0=win+6;;
		j1=fft6Result;;
		k1=fft6Result+6;;
	    r20=l[j31+W];;//正的W(3)
		r21=l[k31+W+2];;//负的W(3)
		r22=l[j31+W+4];;
		r23=l[j31+W+6];;
		r4=l[k0+=4];r6=l[j0+=4];;  //yr0=real(0),XR0=img(0) r6=加权系数
		r5=l[k0+=-6];r7=l[j0+=-6];;//yr1=real(2),XR1=img(2)
		r1=l[k0+=4];r8=l[j0+=4];fr4=r4*r6;;//yr2=real(4),xr2=img(4)
		r2=l[k0+=-6];r9=l[j0+=-6];fr5=r5*r7;;//yr3=real(1),yr3=img(1)
		r3=l[k0+=-2];r10=l[j0+=-2];fr1=r1*r8;;//yr4=real(3)
		r0=l[k0+=0];r11=l[j0+=0];fr2=r2*r9;;//yr5=real(5)
		fr4=r4+r5,fr12=r4-r5;fr3=r3*r10;;
		fr1=r1+r2,fr13=r1-r2;fr0=r0*r11;;
		yr24=xr12;fr16=r4*r20;fr6=r3+r4;;
		xr24=yr12;fr9=r0+r1;;
		fr25=r3+r16;yr26=xr13;;
	    fr17=r24*r21;xr26=yr13;;
	    fr18=r24*r23;;
		fr12=r1*r20;fr7=r25+r17;;
	    fr8=r25+r18;yr24=xr7;;
	    fr2=r7*r20;xr24=yr7;;
		fr0=r26*r21;fr1=r0+r12;yr25=xr8;;
	    fr7=r24*r21;xr25=yr8;;
		fr10=r1+r0;fr29=r8*r20;;
		fr8=r25*r21;fr7=r7-r2;;
		fr12=r9+r6,fr15=r9-r6;;
		fr11=r26*r23;fr8=r8+r29;;
		fr13=r10+r7,fr16=r10-r7;l[j1+=2]=r12;;
		fr11=r11+r1;l[k1+=2]=r15;;
		l[j1+=2]=r13;;
		fr14=r11+r8,fr17=r11-r8;;
		l[k1+=2]=r16;;
		l[j1+=0]=r14;;
		l[k1+=0]=r17;;
		nop;;
_main.end: