5
|
1 /*
|
|
2 * SpanDSP - a series of DSP components for telephony
|
|
3 *
|
|
4 * v17rx.h - ITU V.17 modem receive part
|
|
5 *
|
|
6 * Written by Steve Underwood <steveu@coppice.org>
|
|
7 *
|
|
8 * Copyright (C) 2003 Steve Underwood
|
|
9 *
|
|
10 * All rights reserved.
|
|
11 *
|
|
12 * This program is free software; you can redistribute it and/or modify
|
|
13 * it under the terms of the GNU General Public License version 2, as
|
|
14 * published by the Free Software Foundation.
|
|
15 *
|
|
16 * This program is distributed in the hope that it will be useful,
|
|
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
19 * GNU General Public License for more details.
|
|
20 *
|
|
21 * You should have received a copy of the GNU General Public License
|
|
22 * along with this program; if not, write to the Free Software
|
|
23 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
|
|
24 *
|
|
25 * $Id: v17rx.h,v 1.35 2006/10/24 13:45:28 steveu Exp $
|
|
26 */
|
|
27
|
|
28 /*! \file */
|
|
29
|
|
30 #if !defined(_V17RX_H_)
|
|
31 #define _V17RX_H_
|
|
32
|
|
33 /*! \page v17rx_page The V.17 receiver
|
|
34 \section v17rx_page_sec_1 What does it do?
|
|
35 The V.17 receiver implements the receive side of a V.17 modem. This can operate
|
|
36 at data rates of 14400, 12000, 9600 and 7200 bits/second. The audio input is a stream
|
|
37 of 16 bit samples, at 8000 samples/second. The transmit and receive side of V.17
|
|
38 modems operate independantly. V.17 is mostly used for FAX transmission over PSTN
|
|
39 lines, where it provides the standard 14400 bits/second rate.
|
|
40
|
|
41 \section v17rx_page_sec_2 How does it work?
|
|
42 V.17 uses QAM modulation, at 2400 baud, and trellis coding. Constellations with
|
|
43 16, 32, 64, and 128 points are defined. After one bit per baud is absorbed by the
|
|
44 trellis coding, this gives usable bit rates of 7200, 9600, 12000, and 14400 per
|
|
45 second.
|
|
46
|
|
47 V.17 specifies a training sequence at the start of transmission, which makes the
|
|
48 design of a V.17 receiver relatively straightforward. The first stage of the
|
|
49 training sequence consists of 256
|
|
50 symbols, alternating between two constellation positions. The receiver monitors
|
|
51 the signal power, to sense the possible presence of a valid carrier. When the
|
|
52 alternating signal begins, the power rising above a minimum threshold (-43dBm0)
|
|
53 causes the main receiver computation to begin. The initial measured power is
|
|
54 used to quickly set the gain of the receiver. After this initial settling, the
|
|
55 front end gain is locked, and the adaptive equalizer tracks any subsequent
|
|
56 signal level variation. The signal is oversampled to 24000 samples/second (i.e.
|
|
57 signal, zero, zero, signal, zero, zero, ...) and fed to a complex root raised
|
|
58 cosine pulse shaping filter. This filter has been modified from the conventional
|
|
59 root raised cosine filter, by shifting it up the band, to be centred at the nominal
|
|
60 carrier frequency. This filter interpolates the samples, pulse shapes, and performs
|
|
61 a fractional sample delay at the same time. 192 sets of filter coefficients are used
|
|
62 to achieve a set of finely spaces fractional sample delays, between zero and
|
|
63 one sample. By choosing every fifth sample, and the appropriate set of filter
|
|
64 coefficients, the properly tuned symbol tracker can select data samples at 4800
|
|
65 samples/second from points within 0.28 degrees of the centre and mid-points of
|
|
66 each symbol. The output of the filter is multiplied by a complex carrier, generated
|
|
67 by a DDS. The result is a baseband signal, requiring no further filtering, apart from
|
|
68 an adaptive equalizer. The baseband signal is fed to a T/2 adaptive equalizer.
|
|
69 A band edge component maximisation algorithm is used to tune the sampling, so the samples
|
|
70 fed to the equalizer are close to the mid point and edges of each symbol. Initially
|
|
71 the algorithm is very lightly damped, to ensure the symbol alignment pulls in
|
|
72 quickly. Because the sampling rate will not be precisely the same as the
|
|
73 transmitter's (the spec. says the symbol timing should be within 0.01%), the
|
|
74 receiver constantly evaluates and corrects this sampling throughout its
|
|
75 operation. During the symbol timing maintainence phase, the algorithm uses
|
|
76 a heavier damping.
|
|
77
|
|
78 The carrier is specified as 1800Hz +- 1Hz at the transmitter, and 1800 +-7Hz at
|
|
79 the receiver. The receive carrier would only be this inaccurate if the link
|
|
80 includes FDM sections. These are being phased out, but the design must still
|
|
81 allow for the worst case. Using an initial 1800Hz signal for demodulation gives
|
|
82 a worst case rotation rate for the constellation of about one degree per symbol.
|
|
83 Once the symbol timing synchronisation algorithm has been given time to lock to the
|
|
84 symbol timing of the initial alternating pattern, the phase of the demodulated signal
|
|
85 is recorded on two successive symbols - once for each of the constellation positions.
|
|
86 The receiver then tracks the symbol alternations, until a large phase jump occurs.
|
|
87 This signifies the start of the next phase of the training sequence. At this
|
|
88 point the total phase shift between the original recorded symbol phase, and the
|
|
89 symbol phase just before the phase jump occurred is used to provide a coarse
|
|
90 estimation of the rotation rate of the constellation, and it current absolute
|
|
91 angle of rotation. These are used to update the current carrier phase and phase
|
|
92 update rate in the carrier DDS. The working data already in the pulse shaping
|
|
93 filter and equalizer buffers is given a similar step rotation to pull it all
|
|
94 into line. From this point on, a heavily damped integrate and dump approach,
|
|
95 based on the angular difference between each received constellation position and
|
|
96 its expected position, is sufficient to track the carrier, and maintain phase
|
|
97 alignment. A fast rough approximator for the arc-tangent function is adequate
|
|
98 for the estimation of the angular error.
|
|
99
|
|
100 The next phase of the training sequence is a scrambled sequence of two
|
|
101 particular symbols. We train the T/2 adaptive equalizer using this sequence. The
|
|
102 scrambling makes the signal sufficiently diverse to ensure the equalizer
|
|
103 converges to the proper generalised solution. At the end of this sequence, the
|
|
104 equalizer should be sufficiently well adapted that is can correctly resolve the
|
|
105 full QAM constellation. However, the equalizer continues to adapt throughout
|
|
106 operation of the modem, fine tuning on the more complex data patterns of the
|
|
107 full QAM constellation.
|
|
108
|
|
109 In the last phase of the training sequence, the modem enters normal data
|
|
110 operation, with a short defined period of all ones as data. As in most high
|
|
111 speed modems, data in a V.17 modem passes through a scrambler, to whiten the
|
|
112 spectrum of the signal. The transmitter should initialise its data scrambler,
|
|
113 and pass the ones through it. At the end of the ones, real data begins to pass
|
|
114 through the scrambler, and the transmit modem is in normal operation. The
|
|
115 receiver tests that ones are really received, in order to verify the modem
|
|
116 trained correctly. If all is well, the data following the ones is fed to the
|
|
117 application, and the receive modem is up and running. Unfortunately, some
|
|
118 transmit side of some real V.17 modems fail to initialise their scrambler before
|
|
119 sending the ones. This means the first 23 received bits (the length of the
|
|
120 scrambler register) cannot be trusted for the test. The receive modem,
|
|
121 therefore, only tests that bits starting at bit 24 are really ones.
|
|
122
|
|
123 The V.17 signal is trellis coded. Two bits of each symbol are convolutionally coded
|
|
124 to form a 3 bit trellis code - the two original bits, plus an extra redundant bit. It
|
|
125 is possible to ignore the trellis coding, and just decode the non-redundant bits.
|
|
126 However, the noise performance of the receiver would suffer. Using a proper
|
|
127 trellis decoder adds several dB to the noise tolerance to the receiving modem. Trellis
|
|
128 coding seems quite complex at first sight, but is fairly straightforward once you
|
|
129 get to grips with it.
|
|
130
|
|
131 Trellis decoding tracks the data in terms of the possible states of the convolutional
|
|
132 coder at the transmitter. There are 8 possible states of the V.17 coder. The first
|
|
133 step in trellis decoding is to find the best candidate constellation point
|
|
134 for each of these 8 states. One of thse will be our final answer. The constellation
|
|
135 has been designed so groups of 8 are spread fairly evenly across it. Locating them
|
|
136 is achieved is a reasonably fast manner, by looking up the answers in a set of space
|
|
137 map tables. The disadvantage is the tables are potentially large enough to affect
|
|
138 cache performance. The trellis decoder works over 16 successive symbols. The result
|
|
139 of decoding is not known until 16 symbols after the data enters the decoder. The
|
|
140 minimum total accumulated mismatch between each received point and the actual
|
|
141 constellation (termed the distance) is assessed for each of the 8 states. A little
|
|
142 analysis of the coder shows that each of the 8 current states could be arrived at
|
|
143 from 4 different previous states, through 4 different constellation bit patterns.
|
|
144 For each new state, the running total distance is arrived at by inspecting a previous
|
|
145 total plus a new distance for the appropriate 4 previous states. The minimum of the 4
|
|
146 values becomes the new distance for the state. Clearly, a mechanism is needed to stop
|
|
147 this distance from growing indefinitely. A sliding window, and several other schemes
|
|
148 are possible. However, a simple single pole IIR is very simple, and provides adequate
|
|
149 results.
|
|
150
|
|
151 For each new state we store the constellation bit pattern, or path, to that state, and
|
|
152 the number of the previous state. We find the minimum distance amongst the 8 new
|
|
153 states for each new symbol. We then trace back through the states, until we reach the
|
|
154 one 16 states ago which leads to the current minimum distance. The bit pattern stored
|
|
155 there is the error corrected bit pattern for that symbol.
|
|
156
|
|
157 So, what does Trellis coding actually achieve? TCM is easier to understand by looking
|
|
158 at the V.23bis modem spec. The V.32bis spec. is very similar to V.17, except that it
|
|
159 is a full duplex modem and has non-TCM options, as well as the TCM ones in V.17.
|
|
160
|
|
161 V32bis defines two options for pumping 9600 bits per second down a phone line - one
|
|
162 with and one without TCM. Both run at 2400 baud. The non-TCM one uses simple 16 point
|
|
163 QAM on the raw data. The other takes two out of every four raw bits, and convolutionally
|
|
164 encodes them to 3. Now we have 5 bits per symbol, and we need 32 point QAM to send the
|
|
165 data.
|
|
166
|
|
167 The raw error rate from simple decoding of the 32 point QAM is horrible compared to
|
|
168 decoding the 16 point QAM. If a point decoded from the 32 point QAM is wrong, the likely
|
|
169 correct choice should be one of the adjacent ones. It is unlikely to have been one that
|
|
170 is far away across the constellation, unless there was a huge noise spike, interference,
|
|
171 or something equally nasty. Now, the 32 point symbols do not exist in isolation. There
|
|
172 was a kind of temporal smearing in the convolutional coding. It created a well defined
|
|
173 dependency between successive symbols. If we knew for sure what the last few symbols
|
|
174 were, they would lead us to a limited group of possible values for the current symbol,
|
|
175 constrained by the behaviour of the convolutional coder. If you look at how the symbols
|
|
176 were mapped to constellation points, you will see the mapping tries to spread those
|
|
177 possible symbols as far apart as possible. This will leave only one that is pretty
|
|
178 close to the received point, which must be the correct choice. However, this assumes
|
|
179 we know the last few symbols for sure. Since we don't, we have a bit more work to do
|
|
180 to achieve reliable decoding.
|
|
181
|
|
182 Instead of decoding to the nearest point on the constellation, we decode to a group of
|
|
183 likely constellation points in the neighbourhood of the received point. We record the
|
|
184 mismatch for each - that is the distance across the constellation between the received
|
|
185 point and the group of nearby points. To avoid square roots, recording x2 + y2 can be
|
|
186 good enough. Symbol by symbol, we record this information. After a few symbols we can
|
|
187 stand back and look at the recorded information.
|
|
188
|
|
189 For each symbol we have a set of possible symbol values and error metric pairs. The
|
|
190 dependency between symbols, created by the convolutional coder, means some paths from
|
|
191 symbol to symbol are possible and some are not. It we trace back through the possible
|
|
192 symbol to symbol paths, and total up the error metric through those paths, we end up
|
|
193 with a set of figures of merit (or more accurately figures of demerit, since
|
|
194 larger == worse) for the likelihood of each path being the correct one. The path with
|
|
195 the lowest total metric is the most likely, and gives us our final choice for what we
|
|
196 think the current symbol really is.
|
|
197
|
|
198 That was hard work. It takes considerable computation to do this selection and traceback,
|
|
199 symbol by symbol. We need to get quite a lot from this. It needs to drive the error rate
|
|
200 down so far that is compensates for the much higher error rate due to the larger
|
|
201 constellation, and then buys us some actual benefit. Well in the example we are looking
|
|
202 at - V.32bis at 9600bps - it works out the error rate from the TCM option is like using
|
|
203 the non-TCM option with several dB more signal to noise ratio. That's nice. The non-TCM
|
|
204 option is pretty reasonable on most phone lines, but a better error rate is always a
|
|
205 good thing. However, V32bis includes a 14,400bps option. That uses 2400 baud, and 6 bit
|
|
206 symbols. Convolutional encoding increases that to 7 bits per symbol, by taking 2 bits and
|
|
207 encoding them to 3. This give a 128 point QAM constellation. Again, the difference between
|
|
208 using this, and using just an uncoded 64 point constellation is equivalent to maybe 5dB of
|
|
209 extra signal to noise ratio. However, in this case it is the difference between the modem
|
|
210 working only on the most optimal lines, and being widely usable across most phone lines.
|
|
211 TCM absolutely transformed the phone line modem business.
|
|
212 */
|
|
213
|
|
214 /* Target length for the equalizer is about 63 taps, to deal with the worst stuff
|
|
215 in V.56bis. */
|
|
216 #define V17_EQUALIZER_PRE_LEN 7 /* this much before the real event */
|
|
217 #define V17_EQUALIZER_POST_LEN 7 /* this much after the real event */
|
|
218 #define V17_EQUALIZER_MASK 63 /* one less than a power of 2 >= (V17_EQUALIZER_PRE_LEN + 1 + V17_EQUALIZER_POST_LEN) */
|
|
219
|
|
220 #define V17_RX_FILTER_STEPS 27
|
|
221
|
|
222 /* We can store more trellis depth that we look back over, so that we can push out a group
|
|
223 of symbols in one go, giving greater processing efficiency, at the expense of a bit more
|
|
224 latency through the modem. */
|
|
225 /* Right now we don't take advantage of this optimisation. */
|
|
226 #define V17_TRELLIS_STORAGE_DEPTH 16
|
|
227 #define V17_TRELLIS_LOOKBACK_DEPTH 16
|
|
228
|
|
229 /*!
|
|
230 V.17 modem receive side descriptor. This defines the working state for a
|
|
231 single instance of a V.17 modem receiver.
|
|
232 */
|
|
233 typedef struct
|
|
234 {
|
|
235 /*! \brief The bit rate of the modem. Valid values are 7200 9600, 12000 and 14400. */
|
|
236 int bit_rate;
|
|
237 /*! \brief The callback function used to put each bit received. */
|
|
238 put_bit_func_t put_bit;
|
|
239 /*! \brief A user specified opaque pointer passed to the put_but routine. */
|
|
240 void *user_data;
|
|
241 /*! \brief A callback function which may be enabled to report every symbol's
|
|
242 constellation position. */
|
|
243 qam_report_handler_t *qam_report;
|
|
244 /*! \brief A user specified opaque pointer passed to the qam_report callback
|
|
245 routine. */
|
|
246 void *qam_user_data;
|
|
247
|
|
248 /*! \brief The route raised cosine (RRC) pulse shaping filter buffer. */
|
|
249 float rrc_filter[2*V17_RX_FILTER_STEPS];
|
|
250 /*! \brief Current offset into the RRC pulse shaping filter buffer. */
|
|
251 int rrc_filter_step;
|
|
252
|
|
253 /*! \brief The state of the differential decoder */
|
|
254 int diff;
|
|
255 /*! \brief The register for the data scrambler. */
|
|
256 unsigned int scramble_reg;
|
|
257 /*! \brief TRUE if the short training sequence is to be used. */
|
|
258 int short_train;
|
|
259 int in_training;
|
|
260 int training_count;
|
|
261 float training_error;
|
|
262 int carrier_present;
|
|
263 int16_t last_sample;
|
|
264
|
|
265 /*! \brief The current phase of the carrier (i.e. the DDS parameter). */
|
|
266 uint32_t carrier_phase;
|
|
267 /*! \brief The update rate for the phase of the carrier (i.e. the DDS increment). */
|
|
268 int32_t carrier_phase_rate;
|
|
269 /*! \brief The carrier update rate saved for reuse when using short training. */
|
|
270 int32_t carrier_phase_rate_save;
|
|
271 float carrier_track_p;
|
|
272 float carrier_track_i;
|
|
273
|
|
274 /*! \brief The received signal power monitor. */
|
|
275 power_meter_t power;
|
|
276 int32_t carrier_on_power;
|
|
277 int32_t carrier_off_power;
|
|
278 float agc_scaling;
|
|
279 float agc_scaling_save;
|
|
280
|
|
281 float eq_delta;
|
|
282 /*! \brief The adaptive equalizer coefficients */
|
|
283 complexf_t eq_coeff[V17_EQUALIZER_PRE_LEN + 1 + V17_EQUALIZER_POST_LEN];
|
|
284 complexf_t eq_coeff_save[V17_EQUALIZER_PRE_LEN + 1 + V17_EQUALIZER_POST_LEN];
|
|
285 complexf_t eq_buf[V17_EQUALIZER_MASK + 1];
|
|
286 /*! \brief Current offset into equalizer buffer. */
|
|
287 int eq_step;
|
|
288 int eq_put_step;
|
|
289
|
|
290 /*! \brief The current half of the baud. */
|
|
291 int baud_half;
|
|
292 /*! \brief Band edge symbol sync. filter state. */
|
|
293 float symbol_sync_low[2];
|
|
294 float symbol_sync_high[2];
|
|
295 float symbol_sync_dc_filter[2];
|
|
296 float baud_phase;
|
|
297 /*! \brief The total symbol timing correction since the carrier came up.
|
|
298 This is only for performance analysis purposes. */
|
|
299 int total_baud_timing_correction;
|
|
300
|
|
301 /*! \brief Starting phase angles for the coarse carrier aquisition step. */
|
|
302 int32_t start_angles[2];
|
|
303 /*! \brief History list of phase angles for the coarse carrier aquisition step. */
|
|
304 int32_t angles[16];
|
|
305 /*! \brief A pointer to the current constellation. */
|
|
306 const complexf_t *constellation;
|
|
307 /*! \brief A pointer to the current space map. There is a space map for
|
|
308 each trellis state. */
|
|
309 int space_map;
|
|
310 /*! \brief The number of bits in each symbol at the current bit rate. */
|
|
311 int bits_per_symbol;
|
|
312
|
|
313 /*! \brief Current pointer to the trellis buffers */
|
|
314 int trellis_ptr;
|
|
315 /*! \brief The trellis. */
|
|
316 int full_path_to_past_state_locations[V17_TRELLIS_STORAGE_DEPTH][8];
|
|
317 /*! \brief The trellis. */
|
|
318 int past_state_locations[V17_TRELLIS_STORAGE_DEPTH][8];
|
|
319 /*! \brief Euclidean distances (actually the sqaures of the distances)
|
|
320 from the last states of the trellis. */
|
|
321 float distances[8];
|
|
322 /*! \brief Error and flow logging control */
|
|
323 logging_state_t logging;
|
|
324 } v17_rx_state_t;
|
|
325
|
|
326 extern const complexf_t v17_14400_constellation[128];
|
|
327 extern const complexf_t v17_12000_constellation[64];
|
|
328 extern const complexf_t v17_9600_constellation[32];
|
|
329 extern const complexf_t v17_7200_constellation[16];
|
|
330
|
|
331 #ifdef __cplusplus
|
|
332 extern "C" {
|
|
333 #endif
|
|
334
|
|
335 /*! Initialise a V.17 modem receive context.
|
|
336 \brief Initialise a V.17 modem receive context.
|
|
337 \param s The modem context.
|
|
338 \param rate The bit rate of the modem. Valid values are 7200, 9600, 12000 and 14400.
|
|
339 \param put_bit The callback routine used to put the received data.
|
|
340 \param user_data An opaque pointer passed to the put_bit routine.
|
|
341 \return A pointer to the modem context, or NULL if there was a problem. */
|
|
342 v17_rx_state_t *v17_rx_init(v17_rx_state_t *s, int rate, put_bit_func_t put_bit, void *user_data);
|
|
343
|
|
344 /*! Reinitialise an existing V.17 modem receive context.
|
|
345 \brief Reinitialise an existing V.17 modem receive context.
|
|
346 \param s The modem context.
|
|
347 \param rate The bit rate of the modem. Valid values are 7200, 9600, 12000 and 14400.
|
|
348 \param short_train TRUE if a short training sequence is expected.
|
|
349 \return 0 for OK, -1 for bad parameter */
|
|
350 int v17_rx_restart(v17_rx_state_t *s, int rate, int short_train);
|
|
351
|
|
352 /*! Release a V.17 modem receive context.
|
|
353 \brief Release a V.17 modem receive context.
|
|
354 \param s The modem context.
|
|
355 \return 0 for OK */
|
|
356 int v17_rx_release(v17_rx_state_t *s);
|
|
357
|
|
358 /*! Change the put_bit function associated with a V.17 modem receive context.
|
|
359 \brief Change the put_bit function associated with a V.17 modem receive context.
|
|
360 \param s The modem context.
|
|
361 \param put_bit The callback routine used to handle received bits.
|
|
362 \param user_data An opaque pointer. */
|
|
363 void v17_rx_set_put_bit(v17_rx_state_t *s, put_bit_func_t put_bit, void *user_data);
|
|
364
|
|
365 /*! Process a block of received V.17 modem audio samples.
|
|
366 \brief Process a block of received V.17 modem audio samples.
|
|
367 \param s The modem context.
|
|
368 \param amp The audio sample buffer.
|
|
369 \param len The number of samples in the buffer.
|
|
370 */
|
|
371 void v17_rx(v17_rx_state_t *s, const int16_t amp[], int len);
|
|
372
|
|
373 /*! Get a snapshot of the current equalizer coefficients.
|
|
374 \brief Get a snapshot of the current equalizer coefficients.
|
|
375 \param s The modem context.
|
|
376 \param coeffs The vector of complex coefficients.
|
|
377 \return The number of coefficients in the vector. */
|
|
378 int v17_rx_equalizer_state(v17_rx_state_t *s, complexf_t **coeffs);
|
|
379
|
|
380 /*! Get the current received carrier frequency.
|
|
381 \param s The modem context.
|
|
382 \return The frequency, in Hertz. */
|
|
383 float v17_rx_carrier_frequency(v17_rx_state_t *s);
|
|
384
|
|
385 /*! Get the current symbol timing correction since startup.
|
|
386 \param s The modem context.
|
|
387 \return The correction. */
|
|
388 float v17_rx_symbol_timing_correction(v17_rx_state_t *s);
|
|
389
|
|
390 /*! Get a current received signal power.
|
|
391 \param s The modem context.
|
|
392 \return The signal power, in dBm0. */
|
|
393 float v17_rx_signal_power(v17_rx_state_t *s);
|
|
394
|
|
395 /*! Set the power level at which the carrier detection will cut in
|
|
396 \param s The modem context.
|
|
397 \param cutoff The signal cutoff power, in dBm0. */
|
|
398 void v17_rx_signal_cutoff(v17_rx_state_t *s, float cutoff);
|
|
399
|
|
400 /*! Set a handler routine to process QAM status reports
|
|
401 \param s The modem context.
|
|
402 \param handler The handler routine.
|
|
403 \param user_data An opaque pointer passed to the handler routine. */
|
|
404 void v17_rx_set_qam_report_handler(v17_rx_state_t *s, qam_report_handler_t *handler, void *user_data);
|
|
405
|
|
406 #ifdef __cplusplus
|
|
407 }
|
|
408 #endif
|
|
409
|
|
410 #endif
|
|
411 /*- End of file ------------------------------------------------------------*/
|