Complex "dtype" error happen when certain fractional precision

[VincentChen1017]

Hi, I'm trying to use fixed point to simulate the FFT and generate the golden pattern for hardware design.

I set my input signal to be 'fxp-s32/23', and to avoid the possible overflow situation, so I try to large the temp variable's size.

When I set the temp variable to be 'fxp-s65/46', it occur ERROR message like : AttributeError: 'complex' object has no attribute 'dtype'.

But if I test the complex operation with 'fxp-s65/46 using other simple equation, or change the temp variable in my fft to be 'fxp-s65/23' (or reduce the n_frac part lower than 36), then it will work without error.

I'm new to the fxpmath library, I tried hard to solve it but still not work :( Following is my code, is there any wrong operation with the "Fxp" ? Thanks !

input_size = Fxp(None, dtype='fxp-s32/23')
f = [0,10+7j,20-0.65j,30]
f = Fxp(f, like = input_size)

def FFT(f):
N = len(f)
if N <= 1:
    return f

# division: decompose N point FFT into N/2 point FFT
even= FFT(f[0::2])
odd = FFT(f[1::2])

# store combination of results
temp = np.zeros(N, dtype=complex)
temp = Fxp(temp, dtype='fxp-s65/23')

for u in range(N//2):
    W =  Fxp(exp(-2j*pi*u/N), like=input_size) 
    temp[u] = even[u] + W* odd[u] 
    temp[u+N//2] = even[u] - W*odd[u]  

return temp

# testing the function to see if it matches the manual computation
F_fft = FFT(f)
print(F_fft)

[francof2a]

Hello, I ran your example but I did not get any errors using 'fxp-s65/46' or 'fxp-s32/23'

This message AttributeError: 'complex' object has no attribute 'dtype' is confusing, it is like a python complex variable wanted to be changed instead a Fxp or Numpy array. Could you copy/paste the whole error report?

Might you also tell the OS and fxpmath version are you using?

[VincentChen1017]

Hello,

when I ran code like this, it's OK.

temp = Fxp(temp, dtype='fxp-s65/23')

but I change n_frac to 46 like:

temp = Fxp(temp, dtype='fxp-s65/46')

It's error report in the following:

---

AttributeError Traceback (most recent call last)

in 24 25 # testing the function to see if it matches the manual computation ---> 26 F_fft = FFT(f) 27 print(F_fft) in FFT(f) 9 10 # division: decompose N point FFT into N/2 point FFT ---> 11 even= FFT(f[0::2]) 12 odd = FFT(f[1::2]) 13 in FFT(f) 18 for u in range(N//2): 19 W = Fxp(exp(-2j*pi*u/N), like=input_size) ---> 20 temp[u] = even[u] + W* odd[u] 21 temp[u+N//2] = even[u] - W*odd[u] 22 ~/opt/anaconda3/lib/python3.7/site-packages/fxpmath/objects.py in __mul__(self, x) 1223 _sizing = self.config.op_sizing 1224 -> 1225 return mul(self, x, out=self.config.op_out, out_like=self.config.op_out_like, sizing=_sizing, method=self.config.op_method) 1226 1227 __rmul__ = __mul__ ~/opt/anaconda3/lib/python3.7/site-packages/fxpmath/functions.py in mul(x, y, out, out_like, sizing, method, **kwargs) 362 optimal_size = (signed, n_word, n_int, n_frac) 363 --> 364 return _function_over_two_vars(repr_func=np.multiply, raw_func=_mul_raw, x=x, y=y, out=out, out_like=out_like, sizing=sizing, method=method, optimal_size=optimal_size, **kwargs) 365 366 @implements(np.floor_divide) ~/opt/anaconda3/lib/python3.7/site-packages/fxpmath/functions.py in _function_over_two_vars(repr_func, raw_func, x, y, out, out_like, sizing, method, optimal_size, **kwargs) 176 z = out.set_val(val, raw=raw) 177 else: --> 178 z = Fxp(val, signed=signed, n_int=n_int, n_frac=n_frac, like=out_like, raw=raw, config=config) 179 180 return z ~/opt/anaconda3/lib/python3.7/site-packages/fxpmath/objects.py in __init__(self, val, signed, n_word, n_frac, n_int, like, dtype, **kwargs) 235 236 # store the value --> 237 self.set_val(val, raw=raw) 238 239 # --- ~/opt/anaconda3/lib/python3.7/site-packages/fxpmath/objects.py in set_val(self, val, raw, vdtype, index) 870 871 # update dtype --> 872 self._update_dtype() 873 874 # vdtype ~/opt/anaconda3/lib/python3.7/site-packages/fxpmath/objects.py in _update_dtype(self, notation) 751 nword=self.n_word, 752 nfrac=self.n_frac, --> 753 comp='-complex' if (self.val.dtype == complex or self.vdtype == complex) else '') 754 else: 755 self._dtype = 'fxp-{sign}{nword}/{nfrac}'.format(sign='s' if self.signed else 'u', AttributeError: 'complex' object has no attribute 'dtype' --------------------------------------------------------------------------- My OS version is macOS 10.13.6, fxpmath version is 0.4.8, python version is 3.7.6, numpy version is 1.18.6. If need more information just let me know. Thank you !

[francof2a]

solved in v0.4.9