试图让 KissFFT 定点实现与 DSPIC 保持一致。问题是 Kiss 中的定点实现是真正的定点，但 dspic 在 40 位寄存器中进行乘法和求和，然后在舍入后向下移动到 16 位。KissFFT 可以是 16/32 位定点或浮点。到目前为止，浮点数是最接近的匹配，但我需要它们准确。我不确定在 ASM 代码中每个频率 bin 更新了多少次，但正如您所见，每次更新 bin 时，累加器结果都会移位和舍入。我没有足够的 FFT 知识来解决这个问题。如果有人能指出我正确的方向，将不胜感激。
这是 ASM 代码：
.global _FFT
_FFT:
push.d w8
push.d w10
push.d w12
push w14
push CORCON
mov #0x00f1, w7
mov w7, CORCON
push PSVPAG
push w1 ; save return value
mov #0xff00, w7 ; check if w3==COEFFS_IN_DATA
cp w7, w3
bra z, $+6
bset CORCON, #2
mov w3, PSVPAG
mov #1, w3
sl w3, w0, w3 ; w3 = N (1<<log2N)
mov #0x8000, w14
dec2 w0, w12 ; w12 is # of non-trivial stages
mov #4, w0 ; w0 = 4 * (1, 2, 4, ..., (N/2))
mov #0x0018, w9 ; w9->w12
mov #0x8000, w6
lsr w3, w3 ; start of outer loop, w3 = N/2, N/4, N/8, ..., 1
sl w3, #2, w1 ; w1 = 4 * (N/2, N/4, N/8, ..., 1)
mov [w15-2], w10 ; w10->start of butterfly
lsr w0, #2, w4 ; w4 = groups per stage
dec w4, w4
do w4, $+88 ; first butterfly in group has trivial multiplications
add w0, w2, w8
mov w10, w13
add w1, w10, w11
mov [w10++], w4
mpy.n w4*w6, a, [w11]+=2, w5 ; a = Ar/2
msc w5*w6, a, [w10]+=2, w7 ; a = (Ar+Br)/2
mpy.n w6*w7, b, [w11]+=2, w4 ; b = Ai/2
msc w4*w6, b, [w13]+=2 ; b = (Ai+Bi)/2
mac w5*w6, a, [w13]+=2
sub w11, #4, w13
mac w5*w6, a, [w11]+=2, w5 ; a = (Ar-Br)/2
mac w4*w6, b
mac w4*w6, b, [w8]+=2, w6, [w11]+=2, w7, [w13]+=2 ; b = (Ai-Bi)/2
sub w3, #3, w4
do w4, $+28 ; b = previous Bi, w5 = Br, w6 = Wr, w7 = Bi, w8-> Wi, w10-> Ar, w11-> next Br, w13-> previous Bi
lac [w10], a ; a=Ar
add w5, a ; a=Ar+Br
subr w5, [w10], w4 ; w4=Ar-Br
sac.r a, #1, [w10] ; *w10++=__real__(A+B)
lac [++w10], a ; a=Ai
add w7, a ; a=Ai+Bi
subr w7, [w10], w5 ; w5=Ai-Bi
sac.r a, #1, [w10++] ; *w10++=__imag__(A+B)
mpy w4*w6, a, [w8]-=2, w7 ; a=(Ar-Br)*Wr, w7=Wi
msc w5*w7, a, [w13]+=2 ; a=(Ar-Br)*Wr-(Ai-Bi)*Wi, *w13++ = previous Bi
add w0, w8, w8 ; w8->next Wr
mpy w5*w6, b, [w11]+=2, w5 ; b=(Ai-Bi)*Wr, w5=next Br, w11->next Bi
mac w4*w7, b, [w8]+=2, w6, [w11]+=2, w7, [w13]+=2 ; b=(Ai-Bi)*Wr+(Ar-Br)*Wi, w6=next Wr=*w8++, w7=next Bi=*w11++, *w13++=__real__(A-B)*W
lac [w10], a ; epilog
add w5, a
subr w5, [w10], w4
sac.r a, #1, [w10]
lac [++w10], a
add w7, a
subr w7, [w10], w5
sac.r a, #1, [w10++]
mpy w4*w6, a, [w8]+=2, w7
msc w5*w7, a, [w13]+=2
mpy w5*w6, b, [w9]+=4, w6
mac w4*w7, b, [w9]-=4, w6, [w13]+=2
clr a, [w13]+=2
mov w11, w10 ; last instruction in group
sl w0, w0 ; next stage, double twiddle factor offset
dec w12, w12
bra gt, $-104 ; if w12 > 0, do next stage
mov [w15-2], w10 ; last two stages are done simultaneously
mov [w15-2], w13
add w10, #8, w11
lsr w0, #2, w3
dec w3, w3
clr w8
mov #0x4000, w12
clr a, [w9]+=4, w6, [w10]+=2, w4 ; initialize Ar, w6=0x4000
mov [w10++], w5 ; initialize Ai, w10->Br
do w3, $+58
mov #12, w0 ; adjust DOSTART to run prolog only once
add DOSTARTL
bra NC, $+4
inc DOSTARTH
sub w4, [w11], w0 ; w0 = Ar-Cr
bra $+10 ; w4 = Ar, w5 = Ai, w6 = 0x4000, w8->w0, w9->w14, w10->Br, w11->Cr, w12= 0x4000, w13->last Di, w14= 0x8000
add #12, w11 ; start of 22-cycle do loop
msc w5*w7, b, [w10]+=2, w4, [w13]+=2 ; b = new Di
sub w4, [w11], w0 ; w0 = Ar-Cr
clr a, [w9]+=4, w6, [w10]+=2, w5, [w13]+=2
add w4, [w11], w4 ; w4 = Ar+Cr
sub w5, [++w11], w1 ; w1 = Ai-Ci
add w5, [w11++], w5 ; w5 = Ai+Ci, w11->Dr
mpy w4*w6, a, [w10]+=2, w4 ; a = Ar+Cr, w4 = Br, *w13++ = Di
mpy w5*w6, b, [w9]-=4, w7, [w10]+=6, w5 ; b = Ai+Ci, w5 = Bi
sub w4, [w10], w3 ; w3 = Br-Dr
add w4, [w10], w4 ; w4 = Br+Dr
sub w5, [++w10], w2 ; w2 = Bi-Di
add w5, [w10++], w5 ; w5 = Bi+Di, w10->next Ar
mac w4*w6, a ; a = new Ar
mac w5*w6, b, [w13]+=2 ; b = new Ai, *w13++ = Ar
mac w4*w7, a, [w8]+=2, w4, [w13]+=2 ; a = new Br, w4=Ar-Cr, *w13++ = Ai
mac w5*w7, b, [w8]+=2, w5, [w13]+=2 ; b = new Bi, w5=Ai-Ci, *w13++ = Br
mpy w4*w6, a, [w8]+=2, w4 ; a = Ar-Cr, w4 = Bi-Di
mac w4*w6, a, [w13]+=2 ; a = new Cr, *w13++ = Bi
mpy w5*w6, b, [w8]-=6, w5 ; b = Ai-Ci, w5 = Br-Dr
msc w5*w6, b, [w13]+=2 ; b = new Ci, w6 = *w10++, *w13++ = Cr
mac w4*w7, a, [w13]+=2 ; a = new Dr (last instruction of do loop)
msc w5*w7, b, [w13]+=2 ; epilog
sac.r b, [w13]
pop w0 ; cleanup
pop PSVPAG
pop CORCON
pop w14
pop.d w12
pop.d w10
pop.d w8
return
Kiss Code
http://sourceforge.net/projects/kissfft/
This is where I think I need to modify kiss to line up with dspic
# define S_MUL(a,b) ( (a)*(b) )
#define C_MUL(m,a,b) \
do{ (m).r = (a).r*(b).r - (a).i*(b).i;\
(m).i = (a).r*(b).i + (a).i*(b).r; }while(0)
# define C_FIXDIV(c,div) /* NOOP */
# define C_MULBYSCALAR( c, s ) \
do{ (c).r *= (s);\
(c).i *= (s); }while(0)
#define C_ADD( res, a,b)\
do { \
CHECK_OVERFLOW_OP((a).r,+,(b).r)\
CHECK_OVERFLOW_OP((a).i,+,(b).i)\
(res).r=(a).r+(b).r; (res).i=(a).i+(b).i; \
}while(0)
#define C_SUB( res, a,b)\
do { \
CHECK_OVERFLOW_OP((a).r,-,(b).r)\
CHECK_OVERFLOW_OP((a).i,-,(b).i)\
(res).r=(a).r-(b).r; (res).i=(a).i-(b).i; \
}while(0)
#define C_ADDTO( res , a)\
do { \
CHECK_OVERFLOW_OP((res).r,+,(a).r)\
CHECK_OVERFLOW_OP((res).i,+,(a).i)\
(res).r += (a).r; (res).i += (a).i;\
}while(0)
#define C_SUBFROM( res , a)\
do {\
CHECK_OVERFLOW_OP((res).r,-,(a).r)\
CHECK_OVERFLOW_OP((res).i,-,(a).i)\
(res).r -= (a).r; (res).i -= (a).i; \
}while(0)