ffft.s

用avr调试t6369的240128并将采到的数据用于fft分析的程序

	.dc.w	1*4, 129*4, 65*4, 193*4, 33*4, 161*4, 97*4, 225*4, 17*4, 145*4, 81*4, 209*4, 49*4, 177*4, 113*4, 241*4
	.dc.w	9*4, 137*4, 73*4, 201*4, 41*4, 169*4, 105*4, 233*4, 25*4, 153*4, 89*4, 217*4, 57*4, 185*4, 121*4, 249*4
	.dc.w	5*4, 133*4, 69*4, 197*4, 37*4, 165*4, 101*4, 229*4, 21*4, 149*4, 85*4, 213*4, 53*4, 181*4, 117*4, 245*4
	.dc.w	13*4, 141*4, 77*4, 205*4, 45*4, 173*4, 109*4, 237*4, 29*4, 157*4, 93*4, 221*4, 61*4, 189*4, 125*4, 253*4
	.dc.w	3*4, 131*4, 67*4, 195*4, 35*4, 163*4, 99*4, 227*4, 19*4, 147*4, 83*4, 211*4, 51*4, 179*4, 115*4, 243*4
	.dc.w	11*4, 139*4, 75*4, 203*4, 43*4, 171*4, 107*4, 235*4, 27*4, 155*4, 91*4, 219*4, 59*4, 187*4, 123*4, 251*4
	.dc.w	7*4, 135*4, 71*4, 199*4, 39*4, 167*4, 103*4, 231*4, 23*4, 151*4, 87*4, 215*4, 55*4, 183*4, 119*4, 247*4
	.dc.w	15*4, 143*4, 79*4, 207*4, 47*4, 175*4, 111*4, 239*4, 31*4, 159*4, 95*4, 223*4, 63*4, 191*4, 127*4, 255*4
#endif
	.dc.w	0*4, 128*4, 64*4, 192*4, 32*4, 160*4, 96*4, 224*4, 16*4, 144*4, 80*4, 208*4, 48*4, 176*4, 112*4, 240*4
	.dc.w	8*4, 136*4, 72*4, 200*4, 40*4, 168*4, 104*4, 232*4, 24*4, 152*4, 88*4, 216*4, 56*4, 184*4, 120*4, 248*4
	.dc.w	4*4, 132*4, 68*4, 196*4, 36*4, 164*4, 100*4, 228*4, 20*4, 148*4, 84*4, 212*4, 52*4, 180*4, 116*4, 244*4
	.dc.w	12*4, 140*4, 76*4, 204*4, 44*4, 172*4, 108*4, 236*4, 28*4, 156*4, 92*4, 220*4, 60*4, 188*4, 124*4, 252*4
	.dc.w	2*4, 130*4, 66*4, 194*4, 34*4, 162*4, 98*4, 226*4, 18*4, 146*4, 82*4, 210*4, 50*4, 178*4, 114*4, 242*4
	.dc.w	10*4, 138*4, 74*4, 202*4, 42*4, 170*4, 106*4, 234*4, 26*4, 154*4, 90*4, 218*4, 58*4, 186*4, 122*4, 250*4
	.dc.w	6*4, 134*4, 70*4, 198*4, 38*4, 166*4, 102*4, 230*4, 22*4, 150*4, 86*4, 214*4, 54*4, 182*4, 118*4, 246*4
	.dc.w	14*4, 142*4, 78*4, 206*4, 46*4, 174*4, 110*4, 238*4, 30*4, 158*4, 94*4, 222*4, 62*4, 190*4, 126*4, 254*4

#elif FFT_N == 128
#ifdef INPUT_IQ
	.dc.w	1*4, 65*4, 33*4, 97*4, 17*4, 81*4, 49*4, 113*4, 9*4, 73*4, 41*4, 105*4, 25*4, 89*4, 57*4, 121*4
	.dc.w	5*4, 69*4, 37*4, 101*4, 21*4, 85*4, 53*4, 117*4, 13*4, 77*4, 45*4, 109*4, 29*4, 93*4, 61*4, 125*4
	.dc.w	3*4, 67*4, 35*4, 99*4, 19*4, 83*4, 51*4, 115*4, 11*4, 75*4, 43*4, 107*4, 27*4, 91*4, 59*4, 123*4
	.dc.w	7*4, 71*4, 39*4, 103*4, 23*4, 87*4, 55*4, 119*4, 15*4, 79*4, 47*4, 111*4, 31*4, 95*4, 63*4, 127*4
#endif
	.dc.w	0*4, 64*4, 32*4, 96*4, 16*4, 80*4, 48*4, 112*4, 8*4, 72*4, 40*4, 104*4, 24*4, 88*4, 56*4, 120*4
	.dc.w	4*4, 68*4, 36*4, 100*4, 20*4, 84*4, 52*4, 116*4, 12*4, 76*4, 44*4, 108*4, 28*4, 92*4, 60*4, 124*4
	.dc.w	2*4, 66*4, 34*4, 98*4, 18*4, 82*4, 50*4, 114*4, 10*4, 74*4, 42*4, 106*4, 26*4, 90*4, 58*4, 122*4
	.dc.w	6*4, 70*4, 38*4, 102*4, 22*4, 86*4, 54*4, 118*4, 14*4, 78*4, 46*4, 110*4, 30*4, 94*4, 62*4, 126*4

#elif FFT_N == 64
#ifdef INPUT_IQ
	.dc.w	1*4, 33*4, 17*4, 49*4, 9*4, 41*4, 25*4, 57*4, 5*4, 37*4, 21*4, 53*4, 13*4, 45*4, 29*4, 61*4
	.dc.w	3*4, 35*4, 19*4, 51*4, 11*4, 43*4, 27*4, 59*4, 7*4, 39*4, 23*4, 55*4, 15*4, 47*4, 31*4, 63*4
#endif
	.dc.w	0*4, 32*4, 16*4, 48*4, 8*4, 40*4, 24*4, 56*4, 4*4, 36*4, 20*4, 52*4, 12*4, 44*4, 28*4, 60*4
	.dc.w	2*4, 34*4, 18*4, 50*4, 10*4, 42*4, 26*4, 58*4, 6*4, 38*4, 22*4, 54*4, 14*4, 46*4, 30*4, 62*4
#endif



;----------------------------------------------------------------------------;
#ifndef INPUT_NOUSE
.global fft_input
.func fft_input
fft_input:
	pushw	T2H,T2L
	pushw	AH,AL
	pushw	YH,YL

	movw	XL, EL				;X = array_src;
	movw	YL, DL				;Y = array_bfly;
	clr	EH				;Zero
	ldiw	ZH,ZL, tbl_window		;Z = &tbl_window[0];
	ldiw	AH,AL, FFT_N			;A = FFT_N;
1:	lpmw	BH,BL, Z+			;B = *Z++; (window)
	ldw	CH,CL, X+			;C = *X++; (I-axis)
	FMULS16	DH,DL,T2H,T2L, BH,BL, CH,CL	;D = B * C;
	stw	Y+, DH,DL			;*Y++ = D;
#ifdef INPUT_IQ
	ldw	CH,CL, X+			;C = *X++; (Q-axis)
	FMULS16	DH,DL,T2H,T2L, BH,BL, CH,CL	;D = B * C;
#endif
	stw	Y+, DH,DL			;*Y++ = D;
	subiw	AH,AL, 1			;while(--A)
	brne	1b				;/

	popw	YH,YL
	popw	AH,AL
	popw	T2H,T2L
	clr	r1
	ret
.endfunc
#endif	/* INPUT_NOUSE */



;----------------------------------------------------------------------------;
.global fft_execute
.func fft_execute
fft_execute:
	pushw	T2H,T2L
	pushw	T4H,T4L
	pushw	T6H,T6L
	pushw	T8H,T8L
	pushw	T10H,T10L
	pushw	T12H,T12L
	pushw	T14H,T14L
	pushw	AH,AL
	pushw	YH,YL

	movw	ZL, EL				;Z = array_bfly;
	ldiw	EH,EL, 1			;E = 1;
	ldiw	XH,XL, FFT_N/2			;X = FFT_N/2;
1:	ldi	AL, 4				;T12 = E; (angular speed)
	mul	EL, AL				;
	movw	T12L, T0L			;
	mul	EH, AL				;
	add	T12H, T0L			;/
	movw	T14L, EL			;T14 = E;
	pushw	EH,EL
	movw	YL, ZL				;Z = &array_bfly[0];
	mul	XL, AL				;Y = &array_bfly[X];
	addw	YH,YL, T0H,T0L			;
	mul	XH, AL				;
	add	YH, T0L				;/
	pushw	ZH,ZL
2:	clrw	T10H,T10L			;T10 = 0 (angle)
	clr	EH				;Zero reg.
3:	lddw	AH,AL, Z+0			;A = *Z - *Y; *Z++ += *Y;
	asrw	AH,AL				;
	lddw	DH,DL, Y+0			;
	asrw	DH,DL				;
	movw	CL, AL				;
	subw	AH,AL, DH,DL			;
	addw	CH,CL, DH,DL			;
	stw	Z+, CH,CL			;/
	lddw	BH,BL, Z+0			;B = *Z - *Y; *Z++ += *Y;
	asrw	BH,BL				;
	lddw	DH,DL, Y+2			;
	asrw	DH,DL				;
	movw	CL, BL				;
	subw	BH,BL, DH,DL			;
	addw	CH,CL, DH,DL			;
	stw	Z+, CH,CL			;/
	movw	T0L, ZL
	ldiw	ZH,ZL, tbl_cos_sin		;C = cos(T10); D = sin(T10);
	addw	ZH,ZL, T10H,T10L		;
	lpmw	CH,CL, Z+			;
	lpmw	DH,DL, Z+			;/
	movw	ZL, T0L
	FMULS16	T4H,T4L,T2H,T2L, AH,AL, CH,CL	;*Y++ = A * C + B * D;
	FMULS16	T8H,T8L,T6H,T6L, BH,BL, DH,DL	;
	addd	T4H,T4L,T2H,T2L, T8H,T8L,T6H,T6L;
	stw	Y+, T4H,T4L			;/
	FMULS16	T4H,T4L,T2H,T2L, BH,BL, CH,CL 	;*Y++ = B * C - A * D;
	FMULS16	T8H,T8L,T6H,T6L, AH,AL, DH,DL 	;
	subd	T4H,T4L,T2H,T2L, T8H,T8L,T6H,T6L;
	stw	Y+, T4H,T4L			;/
	addw	T10H,T10L, T12H,T12L		;T10 += T12; (next angle)
#if FFT_N >= 128
	sbrs	T10H, FFT_B - 7			;while(T10 < pi)
#else
	sbrs	T10L, FFT_B + 1
#endif
	rjmp	3b				;/
	ldi	AL, 4				;Y += X; Z += X; (skip split segment)
	mul	XL, AL
	addw	YH,YL, T0H,T0L			;
	addw	ZH,ZL, T0H,T0L			;
	mul	XH, AL				;
	add	YH, T0L				;
	add	ZH, T0L				;/
	ldi	EL, 1				;while(--T14)
	subw	T14H,T14L, EH,EL		;
	rjne	2b				;/
	popw	ZH,ZL
	popw	EH,EL
	lslw	EH,EL				;E *= 2;
	lsrw	XH,XL				;while(X /= 2)
	adiw	XL, 0				;
	rjne	1b				;/

	popw	YH,YL
	popw	AH,AL
	popw	T14H,T14L
	popw	T12H,T12L
	popw	T10H,T10L
	popw	T8H,T8L
	popw	T6H,T6L
	popw	T4H,T4L
	popw	T2H,T2L
;	clr	r1
	ret
.endfunc



;----------------------------------------------------------------------------;
.global fft_output
.func fft_output
fft_output:
	pushw	T2H,T2L
	pushw	T4H,T4L
	pushw	T6H,T6L
	pushw	T8H,T8L
	pushw	T10H,T10L
	pushw	AH,AL
	pushw	YH,YL

	movw	T10L, EL			;T10 = array_bfly;
	movw	YL, DL				;Y = array_output;
	ldiw	ZH,ZL, tbl_bitrev		;Z = tbl_bitrev;
	clr	EH				;Zero
#ifdef INPUT_IQ
	ldiw	AH,AL, FFT_N			;A = FFT_N; (pluse/minus)
#else
	ldiw	AH,AL, FFT_N / 2		;A = FFT_N / 2; (pluse only)
#endif
1:	lpmw	XH,XL, Z+			;X = *Z++;
	addw	XH,XL, T10H,T10L		;X += array_bfly;
	ldw	BH,BL, X+			;B = *X++;
	ldw	CH,CL, X+			;C = *X++;
	FMULS16	T4H,T4L,T2H,T2L, BH,BL, BH,BL	;T4:T2 = B * B;
	FMULS16	T8H,T8L,T6H,T6L, CH,CL, CH,CL	;T8:T6 = C * C;
	addd	T4H,T4L,T2H,T2L, T8H,T8L,T6H,T6L;T4:T2 += T8:T6;
	SQRT32					;B = sqrt(T4:T2);
	stw	Y+, BH,BL			;*Y++ = B;
	subiw	AH,AL, 1			;while(--A)
	rjne	1b				;/

	popw	YH,YL
	popw	AH,AL
	popw	T10H,T10L
	popw	T8H,T8L
	popw	T6H,T6L
	popw	T4H,T4L
	popw	T2H,T2L
	clr	r1
	ret
.endfunc



;----------------------------------------------------------------------------;
.global fmuls_f
.func fmuls_f
fmuls_f:
	movw	CL, EL				;C = E;
	clr	EH	;Zero
	FMULS16	ZH,ZL,XH,XL, CH,CL, DH,DL	;Z:X = C * D;
	movw	EL, ZL
	clr	r1
	ret
.endfunc