When you are "testing" is that on a desktop PC? It seems likely that there are numpy optimizations for your desktop processor, but not for the E310.
I've been able to write VOLK wrapped python code to easily do some operations 10x faster than numpy code. Since there's a bunch of NEON stuff in VOLK, that would probably translate well. Or you could just write your code in C++ if it's really a concern and get the python runtime out of the way of signal processing (but that's not much fun) nw On Fri, Jun 16, 2017 at 3:18 PM, Zach Morris <[email protected]> wrote: > I would like to know what interferer signals occur at different > frequencies in a spectrum. Since I want to get an idea of the average > behaviour of the spectrum, I'd like to look at the spectrum over a long > period of time (hours?). > > This would be too much data to store, so I wrote a custom Python block. It > updates power statistics (specifically the Rayleigh parameter) while the > magnitude squared of an FFT is above a threshold, then writes the power > statistics and the time to a file whenever an element drops below the > threshold. This cuts down the data from tens of MBs to kBs per second. > > I got an overflow (the 0 character) about once every five seconds on the > Ettus 310, running at 4MHz and 1024 vector length. To speed up the Python > code, I vectorized the threshold and update code with numpy, and switched > from write_line to numpy's save_txt to save the power statistics. In > testing the new code chewed through 10,000 x 1024 vectors about thirty > times faster than the old code. Feeling very clever, I ran the new code on > the E310, only for it to overflow twice a second. > > In short, my "optimized" Python code seems to run faster in testing, but > overflow more in practice. Could you help me understand why this happens? I > copied some of the relevant code below, and attached a picture of my > flowgraph. I'm new to Python and GNU Radio, so I apologize if my code is > particularly un-Pythonic, or if I've missed something basic. :) > > ### Old code ### > def write_line(self, bin, theta, n): > now = time.time() > > line = str(now) + ',' + str(bin) + ',' + str(theta) + ',' + str(n) > + '\n' > > self.f.write(line) > > def general_work(self, input_items, output_items): > in0 = input_items[0] > > for i in range(len(in0[0])): > if in0[0, i] > self.threshold: > self.last_state[0, i] += 1 > self.theta[0, i] += in0[0, i] ** 2 > elif (self.last_state[0, i] > 0).any(): > self.write_line(i, self.theta[0, i], self.last_state[0, i]) > self.last_state[0, i] = 0 > self.theta[0, i] = 0 > > will_consume = len(input_items) > > self.consume(0, will_consume) > return len(input_items) > > ### Vectorized code ### > def write_line(self, bin, theta, count): > np.savetxt(self.f, self.fft_num, fmt='%i', newline='\n') > np.savetxt(self.f, np.transpose([bin, theta, count]), fmt='%i > %0.7f %i', newline='\n') > > # Update fft_num > self.fft_num += 1 > > def general_work(self, input_items, output_items): > curr_input = input_items[0][0] > > # Find entries that just dropped below the threshold > just_dropped = (self.last_state > 0) & (curr_input <= > self.threshold) > > # Update theta and count > self.last_state[0, curr_input > self.threshold] += 1 > self.theta[0, curr_input > self.threshold] += > curr_input[curr_input > self.threshold]**2 > > # Write bins, thetas and counts for this vector > self.write_line(np.nonzero(just_dropped[0])[0], > self.theta[just_dropped], self.last_state[just_dropped]) > > # Reset last_state and theta if below threshold > self.last_state[0, curr_input <= self.threshold] = 0 > self.theta[0, curr_input <= self.threshold] = 0 > > will_consume = len(input_items) > > self.consume(0, will_consume) > return len(input_items) > > _______________________________________________ > Discuss-gnuradio mailing list > [email protected] > https://lists.gnu.org/mailman/listinfo/discuss-gnuradio > >
_______________________________________________ Discuss-gnuradio mailing list [email protected] https://lists.gnu.org/mailman/listinfo/discuss-gnuradio
