From d5cbe10c0e2298b0e40161607a3da158249bdb82 Mon Sep 17 00:00:00 2001 From: "Matthias P. Braendli" Date: Tue, 4 Dec 2018 10:18:33 +0100 Subject: Move python stuff to folder --- python/gui/dpd/Align.py | 166 +++++++++++++++++++++++++++++ python/gui/dpd/Capture.py | 253 +++++++++++++++++++++++++++++++++++++++++++++ python/gui/dpd/__init__.py | 93 +++++++++++++++++ 3 files changed, 512 insertions(+) create mode 100644 python/gui/dpd/Align.py create mode 100644 python/gui/dpd/Capture.py create mode 100644 python/gui/dpd/__init__.py (limited to 'python/gui/dpd') diff --git a/python/gui/dpd/Align.py b/python/gui/dpd/Align.py new file mode 100644 index 0000000..1634ec8 --- /dev/null +++ b/python/gui/dpd/Align.py @@ -0,0 +1,166 @@ +# -*- coding: utf-8 -*- +# +# DPD Computation Engine, utility to do subsample alignment. +# +# Copyright (c) 2017 Andreas Steger +# Copyright (c) 2018 Matthias P. Braendli +# +# http://www.opendigitalradio.org +# +# This file is part of ODR-DabMod. +# +# ODR-DabMod is free software: you can redistribute it and/or modify +# it under the terms of the GNU General Public License as +# published by the Free Software Foundation, either version 3 of the +# License, or (at your option) any later version. +# +# ODR-DabMod is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU General Public License for more details. +# +# You should have received a copy of the GNU General Public License +# along with ODR-DabMod. If not, see . +import datetime +import os +import numpy as np +from scipy import optimize +import matplotlib.pyplot as plt + +def gen_omega(length): + if (length % 2) == 1: + raise ValueError("Needs an even length array.") + + halflength = int(length / 2) + factor = 2.0 * np.pi / length + + omega = np.zeros(length, dtype=np.float) + for i in range(halflength): + omega[i] = factor * i + + for i in range(halflength, length): + omega[i] = factor * (i - length) + + return omega + + +def subsample_align(sig, ref_sig, plot_location=None): + """Do subsample alignment for sig relative to the reference signal + ref_sig. The delay between the two must be less than sample + + Returns the aligned signal""" + + n = len(sig) + if (n % 2) == 1: + raise ValueError("Needs an even length signal.") + halflen = int(n / 2) + + fft_sig = np.fft.fft(sig) + + omega = gen_omega(n) + + def correlate_for_delay(tau): + # A subsample offset between two signals corresponds, in the frequency + # domain, to a linearly increasing phase shift, whose slope + # corresponds to the delay. + # + # Here, we build this phase shift in rotate_vec, and multiply it with + # our signal. + + rotate_vec = np.exp(1j * tau * omega) + # zero-frequency is rotate_vec[0], so rotate_vec[N/2] is the + # bin corresponding to the [-1, 1, -1, 1, ...] time signal, which + # is both the maximum positive and negative frequency. + # I don't remember why we handle it differently. + rotate_vec[halflen] = np.cos(np.pi * tau) + + corr_sig = np.fft.ifft(rotate_vec * fft_sig) + + return -np.abs(np.sum(np.conj(corr_sig) * ref_sig)) + + optim_result = optimize.minimize_scalar(correlate_for_delay, bounds=(-1, 1), method='bounded', + options={'disp': True}) + + if optim_result.success: + best_tau = optim_result.x + + if plot_location is not None: + corr = np.vectorize(correlate_for_delay) + ixs = np.linspace(-1, 1, 100) + taus = corr(ixs) + + dt = datetime.datetime.now().isoformat() + tau_path = (plot_location + "/" + dt + "_tau.png") + plt.plot(ixs, taus) + plt.title("Subsample correlation, minimum is best: {}".format(best_tau)) + plt.savefig(tau_path) + plt.close() + + # Prepare rotate_vec = fft_sig with rotated phase + rotate_vec = np.exp(1j * best_tau * omega) + rotate_vec[halflen] = np.cos(np.pi * best_tau) + return np.fft.ifft(rotate_vec * fft_sig).astype(np.complex64) + else: + # print("Could not optimize: " + optim_result.message) + return np.zeros(0, dtype=np.complex64) + +def phase_align(sig, ref_sig, plot_location=None): + """Do phase alignment for sig relative to the reference signal + ref_sig. + + Returns the aligned signal""" + + angle_diff = (np.angle(sig) - np.angle(ref_sig)) % (2. * np.pi) + + real_diffs = np.cos(angle_diff) + imag_diffs = np.sin(angle_diff) + + if plot_location is not None: + dt = datetime.datetime.now().isoformat() + fig_path = plot_location + "/" + dt + "_phase_align.png" + + plt.subplot(511) + plt.hist(angle_diff, bins=60, label="Angle Diff") + plt.xlabel("Angle") + plt.ylabel("Count") + plt.legend(loc=4) + + plt.subplot(512) + plt.hist(real_diffs, bins=60, label="Real Diff") + plt.xlabel("Real Part") + plt.ylabel("Count") + plt.legend(loc=4) + + plt.subplot(513) + plt.hist(imag_diffs, bins=60, label="Imaginary Diff") + plt.xlabel("Imaginary Part") + plt.ylabel("Count") + plt.legend(loc=4) + + plt.subplot(514) + plt.plot(np.angle(ref_sig[:128]), label="ref_sig") + plt.plot(np.angle(sig[:128]), label="sig") + plt.xlabel("Angle") + plt.ylabel("Sample") + plt.legend(loc=4) + + real_diff = np.median(real_diffs) + imag_diff = np.median(imag_diffs) + + angle = np.angle(real_diff + 1j * imag_diff) + + #logging.debug( "Compensating phase by {} rad, {} degree. real median {}, imag median {}".format( angle, angle*180./np.pi, real_diff, imag_diff)) + sig = sig * np.exp(1j * -angle) + + if plot_location is not None: + plt.subplot(515) + plt.plot(np.angle(ref_sig[:128]), label="ref_sig") + plt.plot(np.angle(sig[:128]), label="sig") + plt.xlabel("Angle") + plt.ylabel("Sample") + plt.legend(loc=4) + plt.tight_layout() + plt.savefig(fig_path) + plt.close() + + return sig diff --git a/python/gui/dpd/Capture.py b/python/gui/dpd/Capture.py new file mode 100644 index 0000000..7d95f90 --- /dev/null +++ b/python/gui/dpd/Capture.py @@ -0,0 +1,253 @@ +# -*- coding: utf-8 -*- +# +# DPD Computation Engine, Capture TX signal and RX feedback using ODR-DabMod's +# DPD Server. +# +# Copyright (c) 2017 Andreas Steger +# Copyright (c) 2018 Matthias P. Braendli +# +# http://www.opendigitalradio.org +# +# This file is part of ODR-DabMod. +# +# ODR-DabMod is free software: you can redistribute it and/or modify +# it under the terms of the GNU General Public License as +# published by the Free Software Foundation, either version 3 of the +# License, or (at your option) any later version. +# +# ODR-DabMod is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU General Public License for more details. +# +# You should have received a copy of the GNU General Public License +# along with ODR-DabMod. If not, see . + +import socket +import struct +import os.path +import logging +import numpy as np +from scipy import signal +import matplotlib +matplotlib.use('Agg') +import matplotlib.pyplot as plt +import io + +from . import Align as sa + +def correlation_coefficient(sig_tx, sig_rx): + return np.corrcoef(sig_tx, sig_rx)[0, 1] + +def align_samples(sig_tx, sig_rx): + """ + Returns an aligned version of sig_tx and sig_rx by cropping, subsample alignment and + correct phase offset + """ + + # Coarse sample-level alignment + c = np.abs(signal.correlate(sig_rx, sig_tx)) + off_meas = np.argmax(c) - sig_tx.shape[0] + 1 + off = int(abs(off_meas)) + + if off_meas > 0: + sig_tx = sig_tx[:-off] + sig_rx = sig_rx[off:] + elif off_meas < 0: + sig_tx = sig_tx[off:] + sig_rx = sig_rx[:-off] + + if off % 2 == 1: + sig_tx = sig_tx[:-1] + sig_rx = sig_rx[:-1] + + # Fine subsample alignment and phase offset + sig_rx = sa.subsample_align(sig_rx, sig_tx) + sig_rx = sa.phase_align(sig_rx, sig_tx) + return sig_tx, sig_rx, abs(off_meas) + +class Capture: + """Capture samples from ODR-DabMod""" + def __init__(self, samplerate, port, num_samples_to_request, plot_dir): + self.samplerate = samplerate + self.sizeof_sample = 8 # complex floats + self.port = port + self.num_samples_to_request = num_samples_to_request + self.plot_dir = plot_dir + + # Before we run the samples through the model, we want to accumulate + # them into bins depending on their amplitude, and keep only n_per_bin + # samples to avoid that the polynomial gets overfitted in the low-amplitude + # part, which is less interesting than the high-amplitude part, where + # non-linearities become apparent. + self.binning_n_bins = 64 # Number of bins between binning_start and binning_end + self.binning_n_per_bin = 128 # Number of measurements pre bin + + self.rx_normalisation = 1.0 + + self.clear_accumulated() + + def clear_accumulated(self): + self.binning_start = 0.0 + self.binning_end = 1.0 + + # axis 0: bins + # axis 1: 0=tx, 1=rx + self.accumulated_bins = [np.zeros((0, 2), dtype=np.complex64) for i in range(self.binning_n_bins)] + + def _recv_exact(self, sock, num_bytes): + """Receive an exact number of bytes from a socket. This is + a wrapper around sock.recv() that can return less than the number + of requested bytes. + + Args: + sock (socket): Socket to receive data from. + num_bytes (int): Number of bytes that will be returned. + """ + bufs = [] + while num_bytes > 0: + b = sock.recv(num_bytes) + if len(b) == 0: + break + num_bytes -= len(b) + bufs.append(b) + return b''.join(bufs) + + def receive_tcp(self): + s = socket.socket(socket.AF_INET, socket.SOCK_STREAM) + s.settimeout(4) + s.connect(('localhost', self.port)) + + logging.debug("Send version") + s.sendall(b"\x01") + + logging.debug("Send request for {} samples".format(self.num_samples_to_request)) + s.sendall(struct.pack("=I", self.num_samples_to_request)) + + logging.debug("Wait for TX metadata") + num_samps, tx_second, tx_pps = struct.unpack("=III", self._recv_exact(s, 12)) + tx_ts = tx_second + tx_pps / 16384000.0 + + if num_samps > 0: + logging.debug("Receiving {} TX samples".format(num_samps)) + txframe_bytes = self._recv_exact(s, num_samps * self.sizeof_sample) + txframe = np.fromstring(txframe_bytes, dtype=np.complex64) + else: + txframe = np.array([], dtype=np.complex64) + + logging.debug("Wait for RX metadata") + rx_second, rx_pps = struct.unpack("=II", self._recv_exact(s, 8)) + rx_ts = rx_second + rx_pps / 16384000.0 + + if num_samps > 0: + logging.debug("Receiving {} RX samples".format(num_samps)) + rxframe_bytes = self._recv_exact(s, num_samps * self.sizeof_sample) + rxframe = np.fromstring(rxframe_bytes, dtype=np.complex64) + else: + rxframe = np.array([], dtype=np.complex64) + + if logging.getLogger().getEffectiveLevel() == logging.DEBUG: + logging.debug('txframe: min {}, max {}, median {}'.format( + np.min(np.abs(txframe)), + np.max(np.abs(txframe)), + np.median(np.abs(txframe)))) + + logging.debug('rxframe: min {}, max {}, median {}'.format( + np.min(np.abs(rxframe)), + np.max(np.abs(rxframe)), + np.median(np.abs(rxframe)))) + + logging.debug("Disconnecting") + s.close() + + return txframe, tx_ts, rxframe, rx_ts + + def _plot_spectrum(self, signal, filename, title): + fig = plt.figure() + ax = plt.subplot(1, 1, 1) + + fft = np.fft.fftshift(np.fft.fft(signal)) + fft_db = 20 * np.log10(np.abs(fft)) + + ax.plot(fft_db) + ax.set_title(title) + fig.tight_layout() + fig.savefig(os.path.join(self.plot_dir, filename)) + plt.close(fig) + + def calibrate(self): + txframe, tx_ts, rxframe, rx_ts = self.receive_tcp() + + # Normalize received signal with sent signal + tx_median = np.median(np.abs(txframe)) + rx_median = np.median(np.abs(rxframe)) + self.rx_normalisation = tx_median / rx_median + + rxframe = rxframe * self.rx_normalisation + txframe_aligned, rxframe_aligned, coarse_offset = align_samples(txframe, rxframe) + + self._plot_spectrum(rxframe[:8192], "rxframe.png", "RX Frame") + self._plot_spectrum(txframe[:8192], "txframe.png", "RX Frame") + + return tx_ts, tx_median, rx_ts, rx_median, np.abs(coarse_offset), correlation_coefficient(txframe_aligned, rxframe_aligned) + + def get_samples(self): + """Connect to ODR-DabMod, retrieve TX and RX samples, load + into numpy arrays, and return a tuple + (txframe_aligned, tx_ts, tx_median, rxframe_aligned, rx_ts, rx_median) + """ + + txframe, tx_ts, rxframe, rx_ts = self.receive_tcp() + + # Normalize received signal with calibrated normalisation + rxframe = rxframe * self.rx_normalisation + txframe_aligned, rxframe_aligned, coarse_offset = align_samples(txframe, rxframe) + self._bin_and_accumulate(txframe_aligned, rxframe_aligned) + return txframe_aligned, tx_ts, tx_median, rxframe_aligned, rx_ts, rx_median + + def bin_histogram(self): + return [b.shape[0] for b in self.accumulated_bins] + + def pointcloud_png(self): + fig = plt.figure() + ax = plt.subplot(1, 1, 1) + for b in self.accumulated_bins: + if b: + ax.scatter( + np.abs(b[0]), + np.abs(b[1]), + s=0.1, + color="black") + ax.set_title("Captured and Binned Samples") + ax.set_xlabel("TX Amplitude") + ax.set_ylabel("RX Amplitude") + ax.set_ylim(0, 0.8) + ax.set_xlim(0, 1.1) + ax.legend(loc=4) + fig.tight_layout() + fig.savefig(os.path.join(self.plot_dir, "pointcloud.png")) + plt.close(fig) + + def _bin_and_accumulate(self, txframe, rxframe): + """Bin the samples and extend the accumulated samples""" + + bin_edges = np.linspace(self.binning_start, self.binning_end, self.binning_n_bins) + + minsize = self.num_samples_to_request + + for i, (tx_start, tx_end) in enumerate(zip(bin_edges, bin_edges[1:])): + txframe_abs = np.abs(txframe) + indices = np.bitwise_and(tx_start < txframe_abs, txframe_abs <= tx_end) + txsamples = np.asmatrix(txframe[indices]) + rxsamples = np.asmatrix(rxframe[indices]) + binned_sample_pairs = np.concatenate((txsamples, rxsamples)).T + + missing_in_bin = self.binning_n_per_bin - self.accumulated_bins[i].shape[0] + num_to_append = min(missing_in_bin, binned_sample_pairs.shape[0]) + print("Handling bin {} {}-{}, {} available, {} missing".format(i, tx_start, tx_end, binned_sample_pairs.shape[0], missing_in_bin)) + if num_to_append: + print("Appending {} to bin {} with shape {}".format(num_to_append, i, self.accumulated_bins[i].shape)) + + self.accumulated_bins[i] = np.concatenate((self.accumulated_bins[i], binned_sample_pairs[:num_to_append,...])) + print("{} now has shape {}".format(i, self.accumulated_bins[i].shape)) + diff --git a/python/gui/dpd/__init__.py b/python/gui/dpd/__init__.py new file mode 100644 index 0000000..9009436 --- /dev/null +++ b/python/gui/dpd/__init__.py @@ -0,0 +1,93 @@ +# -*- coding: utf-8 -*- +# +# DPD Computation Engine module +# +# Copyright (c) 2017 Andreas Steger +# Copyright (c) 2018 Matthias P. Braendli +# +# http://www.opendigitalradio.org +# +# This file is part of ODR-DabMod. +# +# ODR-DabMod is free software: you can redistribute it and/or modify +# it under the terms of the GNU General Public License as +# published by the Free Software Foundation, either version 3 of the +# License, or (at your option) any later version. +# +# ODR-DabMod is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU General Public License for more details. +# +# You should have received a copy of the GNU General Public License +# along with ODR-DabMod. If not, see . + +from . import Capture +import numpy as np + +class DPD: + def __init__(self, plot_dir, samplerate=8192000): + self.samplerate = samplerate + + oversample = int(self.samplerate / 2048000) + self.T_F = oversample * 196608 # Transmission frame duration + self.T_NULL = oversample * 2656 # Null symbol duration + self.T_S = oversample * 2552 # Duration of OFDM symbols of indices l = 1, 2, 3,... L; + self.T_U = oversample * 2048 # Inverse of carrier spacing + self.T_C = oversample * 504 # Duration of cyclic prefix + + self.last_capture_info = {} + + port = 50055 + samples_to_capture = 81920 + self.capture = Capture.Capture(self.samplerate, port, samples_to_capture, plot_dir) + + def status(self): + r = {} + r['histogram'] = self.capture.bin_histogram() + r['capture'] = self.last_capture_info + return r + + def pointcloud_png(self): + return self.capture.pointcloud_png() + + def clear_accumulated(self): + return self.capture.clear_accumulated() + + def capture_calibration(self): + tx_ts, tx_median, rx_ts, rx_median, coarse_offset, correlation_coefficient = self.capture.calibrate() + result = {'status': "ok"} + result['tx_median'] = "{:.2f}dB".format(20*np.log10(tx_median)) + result['rx_median'] = "{:.2f}dB".format(20*np.log10(rx_median)) + result['tx_ts'] = tx_ts + result['rx_ts'] = rx_ts + result['coarse_offset'] = int(coarse_offset) + result['correlation'] = float(correlation_coefficient) + return result + + def capture_samples(self): + """Captures samples and store them in the accumulated samples, + returns a dict with some info""" + result = {} + try: + txframe_aligned, tx_ts, tx_median, rxframe_aligned, rx_ts, rx_median = self.capture.get_samples() + result['status'] = "ok" + result['length'] = len(txframe_aligned) + result['tx_median'] = float(tx_median) + result['rx_median'] = float(rx_median) + result['tx_ts'] = tx_ts + result['rx_ts'] = rx_ts + except ValueError as e: + result['status'] = "Capture failed: {}".format(e) + + self.last_capture_info = result + + # tx, rx, phase_diff, n_per_bin = extStat.extract(txframe_aligned, rxframe_aligned) + # off = SA.calc_offset(txframe_aligned) + # print("off {}".format(off)) + # tx_mer = MER.calc_mer(txframe_aligned[off:off + c.T_U], debug_name='TX') + # print("tx_mer {}".format(tx_mer)) + # rx_mer = MER.calc_mer(rxframe_aligned[off:off + c.T_U], debug_name='RX') + # print("rx_mer {}".format(rx_mer)) + # mse = np.mean(np.abs((txframe_aligned - rxframe_aligned) ** 2)) + # print("mse {}".format(mse)) -- cgit v1.2.3