5d7669af232af9febb802494df77afa021cf9b36
[c64tapwav] / decode.cpp
1 // Copyright Steinar H. Gunderson <sgunderson@bigfoot.com>
2 // Licensed under the GPL, v2. (See the file COPYING.)
3
4 #include <stdio.h>
5 #include <string.h>
6 #include <math.h>
7 #include <assert.h>
8 #include <limits.h>
9 #include <getopt.h>
10 #include <vector>
11 #include <algorithm>
12
13 #include "audioreader.h"
14 #include "interpolate.h"
15 #include "level.h"
16 #include "tap.h"
17 #include "filter.h"
18
19 #define BUFSIZE 4096
20 #define C64_FREQUENCY 985248
21 #define SYNC_PULSE_START 1000
22 #define SYNC_PULSE_END 20000
23 #define SYNC_PULSE_LENGTH 378.0
24 #define SYNC_TEST_TOLERANCE 1.10
25
26 // SPSA options
27 #define NUM_FILTER_COEFF 32
28 #define NUM_ITER 5000
29 #define A NUM_ITER/10  // approx
30 #define INITIAL_A 0.005 // A bit of trial and error...
31 #define INITIAL_C 0.02  // This too.
32 #define GAMMA 0.166
33 #define ALPHA 1.0
34
35 static float hysteresis_upper_limit = 3000.0 / 32768.0;
36 static float hysteresis_lower_limit = -3000.0 / 32768.0;
37 static bool do_calibrate = true;
38 static bool output_cycles_plot = false;
39 static bool do_crop = false;
40 static float crop_start = 0.0f, crop_end = HUGE_VAL;
41
42 static bool use_fir_filter = false;
43 static float filter_coeff[NUM_FILTER_COEFF] = { 1.0f };  // The rest is filled with 0.
44 static bool use_rc_filter = false;
45 static float rc_filter_freq;
46 static bool output_filtered = false;
47
48 static bool quiet = false;
49 static bool do_auto_level = false;
50 static bool output_leveled = false;
51 static std::vector<float> train_snap_points;
52 static bool do_train = false;
53
54 // The minimum estimated sound level (for do_auto_level) at any given point.
55 // If you decrease this, you'll be able to amplify really silent signals
56 // by more, but you'll also increase the level of silent (ie. noise-only) segments,
57 // possibly caused misdetected pulses in these segments.
58 static float min_level = 0.05f;
59
60 // search for the value <limit> between [x,x+1]
61 double find_crossing(const std::vector<float> &pcm, int x, float limit)
62 {
63         double upper = x;
64         double lower = x + 1;
65         while (lower - upper > 1e-3) {
66                 double mid = 0.5f * (upper + lower);
67                 if (lanczos_interpolate(pcm, mid) > limit) {
68                         upper = mid;
69                 } else {
70                         lower = mid;
71                 }
72         }
73
74         return 0.5f * (upper + lower);
75 }
76
77 struct pulse {
78         double time;  // in seconds from start
79         double len;   // in seconds
80 };
81
82 // Calibrate on the first ~25k pulses (skip a few, just to be sure).
83 double calibrate(const std::vector<pulse> &pulses) {
84         if (pulses.size() < SYNC_PULSE_END) {
85                 fprintf(stderr, "Too few pulses, not calibrating!\n");
86                 return 1.0;
87         }
88
89         int sync_pulse_end = -1;
90         double sync_pulse_stddev = -1.0;
91
92         // Compute the standard deviation (to check for uneven speeds).
93         // If it suddenly skyrockets, we assume that sync ended earlier
94         // than we thought (it should be 25000 cycles), and that we should
95         // calibrate on fewer cycles.
96         for (int try_end : { 2000, 4000, 5000, 7500, 10000, 15000, SYNC_PULSE_END }) {
97                 double sum2 = 0.0;
98                 for (int i = SYNC_PULSE_START; i < try_end; ++i) {
99                         double cycles = pulses[i].len * C64_FREQUENCY;
100                         sum2 += (cycles - SYNC_PULSE_LENGTH) * (cycles - SYNC_PULSE_LENGTH);
101                 }
102                 double stddev = sqrt(sum2 / (try_end - SYNC_PULSE_START - 1));
103                 if (sync_pulse_end != -1 && stddev > 5.0 && stddev / sync_pulse_stddev > 1.3) {
104                         fprintf(stderr, "Stopping at %d sync pulses because standard deviation would be too big (%.2f cycles); shorter-than-usual trailer?\n",
105                                 sync_pulse_end, stddev);
106                         break;
107                 }
108                 sync_pulse_end = try_end;
109                 sync_pulse_stddev = stddev;
110         }
111         if (!quiet) {
112                 fprintf(stderr, "Sync pulse length standard deviation: %.2f cycles\n",
113                         sync_pulse_stddev);
114         }
115
116         double sum = 0.0;
117         for (int i = SYNC_PULSE_START; i < sync_pulse_end; ++i) {
118                 sum += pulses[i].len;
119         }
120         double mean_length = C64_FREQUENCY * sum / (sync_pulse_end - SYNC_PULSE_START);
121         double calibration_factor = SYNC_PULSE_LENGTH / mean_length;
122         if (!quiet) {
123                 fprintf(stderr, "Calibrated sync pulse length: %.2f -> %.2f (change %+.2f%%)\n",
124                         mean_length, SYNC_PULSE_LENGTH, 100.0 * (calibration_factor - 1.0));
125         }
126
127         // Check for pulses outside +/- 10% (sign of misdetection).
128         for (int i = SYNC_PULSE_START; i < sync_pulse_end; ++i) {
129                 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
130                 if (cycles < SYNC_PULSE_LENGTH / SYNC_TEST_TOLERANCE || cycles > SYNC_PULSE_LENGTH * SYNC_TEST_TOLERANCE) {
131                         fprintf(stderr, "Sync cycle with downflank at %.6f was detected at %.0f cycles; misdetect?\n",
132                                 pulses[i].time, cycles);
133                 }
134         }
135
136         return calibration_factor;
137 }
138
139 void output_tap(const std::vector<pulse>& pulses, double calibration_factor)
140 {
141         std::vector<char> tap_data;
142         for (unsigned i = 0; i < pulses.size(); ++i) {
143                 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
144                 int len = lrintf(cycles / TAP_RESOLUTION);
145                 if (i > SYNC_PULSE_END && (cycles < 100 || cycles > 800)) {
146                         fprintf(stderr, "Cycle with downflank at %.6f was detected at %.0f cycles; misdetect?\n",
147                                         pulses[i].time, cycles);
148                 }
149                 if (len <= 255) {
150                         tap_data.push_back(len);
151                 } else {
152                         int overflow_len = lrintf(cycles);
153                         tap_data.push_back(0);
154                         tap_data.push_back(overflow_len & 0xff);
155                         tap_data.push_back((overflow_len >> 8) & 0xff);
156                         tap_data.push_back(overflow_len >> 16);
157                 }
158         }
159
160         tap_header hdr;
161         memcpy(hdr.identifier, "C64-TAPE-RAW", 12);
162         hdr.version = 1;
163         hdr.reserved[0] = hdr.reserved[1] = hdr.reserved[2] = 0;
164         hdr.data_len = tap_data.size();
165
166         fwrite(&hdr, sizeof(hdr), 1, stdout);
167         fwrite(tap_data.data(), tap_data.size(), 1, stdout);
168 }
169
170 static struct option long_options[] = {
171         {"auto-level",       0,                 0, 'a' },
172         {"output-leveled",   0,                 0, 'A' },
173         {"min-level",        required_argument, 0, 'm' },
174         {"no-calibrate",     0,                 0, 's' },
175         {"plot-cycles",      0,                 0, 'p' },
176         {"hysteresis-limit", required_argument, 0, 'l' },
177         {"filter",           required_argument, 0, 'f' },
178         {"rc-filter",        required_argument, 0, 'r' },
179         {"output-filtered",  0,                 0, 'F' },
180         {"crop",             required_argument, 0, 'c' },
181         {"quiet",            0,                 0, 'q' },
182         {"help",             0,                 0, 'h' },
183         {0,                  0,                 0, 0   }
184 };
185
186 void help()
187 {
188         fprintf(stderr, "decode [OPTIONS] AUDIO-FILE > TAP-FILE\n");
189         fprintf(stderr, "\n");
190         fprintf(stderr, "  -a, --auto-level             automatically adjust amplitude levels throughout the file\n");
191         fprintf(stderr, "  -A, --output-leveled         output leveled waveform to leveled.raw\n");
192         fprintf(stderr, "  -m, --min-level              minimum estimated sound level (0..32768) for --auto-level\n");
193         fprintf(stderr, "  -s, --no-calibrate           do not try to calibrate on sync pulse length\n");
194         fprintf(stderr, "  -p, --plot-cycles            output debugging info to cycles.plot\n");
195         fprintf(stderr, "  -l, --hysteresis-limit VAL   change amplitude threshold for ignoring pulses (0..32768)\n");
196         fprintf(stderr, "  -f, --filter C1:C2:C3:...    specify FIR filter (up to %d coefficients)\n", NUM_FILTER_COEFF);
197         fprintf(stderr, "  -r, --rc-filter FREQ         send signal through a highpass RC filter with given frequency (in Hertz)\n");
198         fprintf(stderr, "  -F, --output-filtered        output filtered waveform to filtered.raw\n");
199         fprintf(stderr, "  -c, --crop START[:END]       use only the given part of the file\n");
200         fprintf(stderr, "  -t, --train LEN1:LEN2:...    train a filter for detecting any of the given number of cycles\n");
201         fprintf(stderr, "                               (implies --no-calibrate and --quiet unless overridden)\n");
202         fprintf(stderr, "  -q, --quiet                  suppress some informational messages\n");
203         fprintf(stderr, "  -h, --help                   display this help, then exit\n");
204         exit(1);
205 }
206
207 void parse_options(int argc, char **argv)
208 {
209         for ( ;; ) {
210                 int option_index = 0;
211                 int c = getopt_long(argc, argv, "aAm:spl:f:r:Fc:t:qh", long_options, &option_index);
212                 if (c == -1)
213                         break;
214
215                 switch (c) {
216                 case 'a':
217                         do_auto_level = true;
218                         break;
219
220                 case 'A':
221                         output_leveled = true;
222                         break;
223
224                 case 'm':
225                         min_level = atof(optarg) / 32768.0;
226                         break;
227
228                 case 's':
229                         do_calibrate = false;
230                         break;
231
232                 case 'p':
233                         output_cycles_plot = true;
234                         break;
235
236                 case 'l': {
237                         const char *hyststr = strtok(optarg, ": ");
238                         hysteresis_upper_limit = atof(hyststr) / 32768.0;
239                         hyststr = strtok(NULL, ": ");
240                         if (hyststr == NULL) {
241                                 hysteresis_lower_limit = -hysteresis_upper_limit;
242                         } else {
243                                 hysteresis_lower_limit = atof(hyststr) / 32768.0;
244                         }
245                         break;
246                 }
247
248                 case 'f': {
249                         const char *coeffstr = strtok(optarg, ": ");
250                         int coeff_index = 0;
251                         while (coeff_index < NUM_FILTER_COEFF && coeffstr != NULL) {
252                                 filter_coeff[coeff_index++] = atof(coeffstr);
253                                 coeffstr = strtok(NULL, ": ");
254                         }
255                         use_fir_filter = true;
256                         break;
257                 }
258
259                 case 'r':
260                         use_rc_filter = true;
261                         rc_filter_freq = atof(optarg);
262                         break;
263
264                 case 'F':
265                         output_filtered = true;
266                         break;
267
268                 case 'c': {
269                         const char *cropstr = strtok(optarg, ":");
270                         crop_start = atof(cropstr);
271                         cropstr = strtok(NULL, ":");
272                         if (cropstr == NULL) {
273                                 crop_end = HUGE_VAL;
274                         } else {
275                                 crop_end = atof(cropstr);
276                         }
277                         do_crop = true;
278                         break;
279                 }
280
281                 case 't': {
282                         const char *cyclestr = strtok(optarg, ":");
283                         while (cyclestr != NULL) {
284                                 train_snap_points.push_back(atof(cyclestr));
285                                 cyclestr = strtok(NULL, ":");
286                         }
287                         do_train = true;
288
289                         // Set reasonable defaults (can be overridden later on the command line).
290                         do_calibrate = false;
291                         quiet = true;
292                         break;
293                 }
294
295                 case 'q':
296                         quiet = true;
297                         break;
298
299                 case 'h':
300                 default:
301                         help();
302                         exit(1);
303                 }
304         }
305 }
306
307 std::vector<float> crop(const std::vector<float>& pcm, float crop_start, float crop_end, int sample_rate)
308 {
309         size_t start_sample, end_sample;
310         if (crop_start >= 0.0f) {
311                 start_sample = std::min<size_t>(lrintf(crop_start * sample_rate), pcm.size());
312         }
313         if (crop_end >= 0.0f) {
314                 end_sample = std::min<size_t>(lrintf(crop_end * sample_rate), pcm.size());
315         }
316         return std::vector<float>(pcm.begin() + start_sample, pcm.begin() + end_sample);
317 }
318
319 // TODO: Support AVX here.
320 std::vector<float> do_fir_filter(const std::vector<float>& pcm, const float* filter)
321 {
322         std::vector<float> filtered_pcm;
323         filtered_pcm.reserve(pcm.size());
324         for (unsigned i = NUM_FILTER_COEFF; i < pcm.size(); ++i) {
325                 float s = 0.0f;
326                 for (int j = 0; j < NUM_FILTER_COEFF; ++j) {
327                         s += filter[j] * pcm[i - j];
328                 }
329                 filtered_pcm.push_back(s);
330         }
331
332         if (output_filtered) {
333                 FILE *fp = fopen("filtered.raw", "wb");
334                 fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
335                 fclose(fp);
336         }
337
338         return filtered_pcm;
339 }
340
341 std::vector<float> do_rc_filter(const std::vector<float>& pcm, float freq, int sample_rate)
342 {
343         // This is only a 6 dB/oct filter, which seemingly works better
344         // than the Filter class, which is a standard biquad (12 dB/oct).
345         // The b/c calculations come from libnyquist (atone.c);
346         // I haven't checked, but I suppose they fall out of the bilinear
347         // transform of the transfer function H(s) = s/(s + w).
348         std::vector<float> filtered_pcm;
349         filtered_pcm.resize(pcm.size());
350         const float b = 2.0f - cos(2.0 * M_PI * freq / sample_rate);
351         const float c = b - sqrt(b * b - 1.0f);
352         float prev_in = 0.0f;
353         float prev_out = 0.0f;
354         for (unsigned i = 0; i < pcm.size(); ++i) {
355                 float in = pcm[i];
356                 float out = c * (prev_out + in - prev_in);
357                 filtered_pcm[i] = out;
358                 prev_in = in;
359                 prev_out = out;
360         }
361
362         if (output_filtered) {
363                 FILE *fp = fopen("filtered.raw", "wb");
364                 fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
365                 fclose(fp);
366         }
367
368         return filtered_pcm;
369 }
370
371 std::vector<pulse> detect_pulses(const std::vector<float> &pcm, int sample_rate)
372 {
373         std::vector<pulse> pulses;
374
375         // Find the flanks.
376         enum State { START, ABOVE, BELOW } state = START;
377         double last_downflank = -1;
378         for (unsigned i = 0; i < pcm.size(); ++i) {
379                 if (pcm[i] > hysteresis_upper_limit) {
380                         state = ABOVE;
381                 } else if (pcm[i] < hysteresis_lower_limit) {
382                         if (state == ABOVE) {
383                                 // down-flank!
384                                 double t = find_crossing(pcm, i - 1, hysteresis_lower_limit) * (1.0 / sample_rate) + crop_start;
385                                 if (last_downflank > 0) {
386                                         pulse p;
387                                         p.time = t;
388                                         p.len = t - last_downflank;
389                                         pulses.push_back(p);
390                                 }
391                                 last_downflank = t;
392                         }
393                         state = BELOW;
394                 }
395         }
396         return pulses;
397 }
398
399 void output_cycle_plot(const std::vector<pulse> &pulses, double calibration_factor)
400 {
401         FILE *fp = fopen("cycles.plot", "w");
402         for (unsigned i = 0; i < pulses.size(); ++i) {
403                 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
404                 fprintf(fp, "%f %f\n", pulses[i].time, cycles);
405         }
406         fclose(fp);
407 }
408
409 std::pair<int, double> find_closest_point(double x, const std::vector<float> &points)
410 {
411         int best_point = 0;
412         double best_dist = (x - points[0]) * (x - points[0]);
413         for (unsigned j = 1; j < train_snap_points.size(); ++j) {
414                 double dist = (x - points[j]) * (x - points[j]);
415                 if (dist < best_dist) {
416                         best_point = j;
417                         best_dist = dist;
418                 }
419         }
420         return std::make_pair(best_point, best_dist);
421 }
422
423 float eval_badness(const std::vector<pulse>& pulses, double calibration_factor)
424 {
425         double sum_badness = 0.0;
426         for (unsigned i = 0; i < pulses.size(); ++i) {
427                 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
428                 if (cycles > 2000.0) cycles = 2000.0;  // Don't make pauses arbitrarily bad.
429                 std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, train_snap_points);
430                 sum_badness += selected_point_and_sq_dist.second;
431         }
432         return sqrt(sum_badness / (pulses.size() - 1));
433 }
434
435 void find_kmeans(const std::vector<pulse> &pulses, double calibration_factor, const std::vector<float> &initial_centers)
436 {
437         std::vector<float> last_centers = initial_centers;
438         std::vector<float> sums;
439         std::vector<float> num;
440         sums.resize(initial_centers.size());
441         num.resize(initial_centers.size());
442         for ( ;; ) {
443                 for (unsigned i = 0; i < initial_centers.size(); ++i) {
444                         sums[i] = 0.0f;
445                         num[i] = 0;
446                 }
447                 for (unsigned i = 0; i < pulses.size(); ++i) {
448                         double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
449                         // Ignore heavy outliers, which are almost always long pauses.
450                         if (cycles > 2000.0) {
451                                 continue;
452                         }
453                         std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, last_centers);
454                         int p = selected_point_and_sq_dist.first;
455                         sums[p] += cycles;
456                         ++num[p];
457                 }
458                 bool any_moved = false;
459                 for (unsigned i = 0; i < initial_centers.size(); ++i) {
460                         if (num[i] == 0) {
461                                 fprintf(stderr, "K-means broke down, can't output new reference training points\n");
462                                 return;
463                         }
464                         float new_center = sums[i] / num[i];
465                         if (fabs(new_center - last_centers[i]) > 1e-3) {
466                                 any_moved = true;
467                         }
468                         last_centers[i] = new_center;
469                 }
470                 if (!any_moved) {
471                         break;
472                 }
473         }
474         fprintf(stderr, "New reference training points:");
475         for (unsigned i = 0; i < last_centers.size(); ++i) {
476                 fprintf(stderr, " %.3f", last_centers[i]);
477         }
478         fprintf(stderr, "\n");
479 }
480
481 void spsa_train(const std::vector<float> &pcm, int sample_rate)
482 {
483         float filter[NUM_FILTER_COEFF] = { 1.0f };  // The rest is filled with 0.
484
485         float start_c = INITIAL_C;
486         double best_badness = HUGE_VAL;
487
488         for (int n = 1; n < NUM_ITER; ++n) {
489                 float a = INITIAL_A * pow(n + A, -ALPHA);
490                 float c = start_c * pow(n, -GAMMA);
491
492                 // find a random perturbation
493                 float p[NUM_FILTER_COEFF];
494                 float filter1[NUM_FILTER_COEFF], filter2[NUM_FILTER_COEFF];
495                 for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
496                         p[i] = (rand() % 2) ? 1.0 : -1.0;
497                         filter1[i] = std::max(std::min(filter[i] - c * p[i], 1.0f), -1.0f);
498                         filter2[i] = std::max(std::min(filter[i] + c * p[i], 1.0f), -1.0f);
499                 }
500
501                 std::vector<pulse> pulses1 = detect_pulses(do_fir_filter(pcm, filter1), sample_rate);
502                 std::vector<pulse> pulses2 = detect_pulses(do_fir_filter(pcm, filter2), sample_rate);
503                 float badness1 = eval_badness(pulses1, 1.0);
504                 float badness2 = eval_badness(pulses2, 1.0);
505
506                 // Find the gradient estimator
507                 float g[NUM_FILTER_COEFF];
508                 for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
509                         g[i] = (badness2 - badness1) / (2.0 * c * p[i]);
510                         filter[i] -= a * g[i];
511                         filter[i] = std::max(std::min(filter[i], 1.0f), -1.0f);
512                 }
513                 if (badness2 < badness1) {
514                         std::swap(badness1, badness2);
515                         std::swap(filter1, filter2);
516                         std::swap(pulses1, pulses2);
517                 }
518                 if (badness1 < best_badness) {
519                         printf("\nNew best filter (badness=%f):", badness1);
520                         for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
521                                 printf(" %.5f", filter1[i]);
522                         }
523                         best_badness = badness1;
524                         printf("\n");
525
526                         find_kmeans(pulses1, 1.0, train_snap_points);
527
528                         if (output_cycles_plot) {
529                                 output_cycle_plot(pulses1, 1.0);
530                         }
531                 }
532                 printf("%d ", n);
533                 fflush(stdout);
534         }
535 }
536
537 int main(int argc, char **argv)
538 {
539         parse_options(argc, argv);
540
541         make_lanczos_weight_table();
542         std::vector<float> pcm;
543         int sample_rate;
544         if (!read_audio_file(argv[optind], &pcm, &sample_rate)) {
545                 exit(1);
546         }
547
548         if (do_crop) {
549                 pcm = crop(pcm, crop_start, crop_end, sample_rate);
550         }
551
552         if (use_fir_filter) {
553                 pcm = do_fir_filter(pcm, filter_coeff);
554         }
555
556         if (use_rc_filter) {
557                 pcm = do_rc_filter(pcm, rc_filter_freq, sample_rate);
558         }
559
560         if (do_auto_level) {
561                 pcm = level_samples(pcm, min_level, sample_rate);
562                 if (output_leveled) {
563                         FILE *fp = fopen("leveled.raw", "wb");
564                         fwrite(pcm.data(), pcm.size() * sizeof(pcm[0]), 1, fp);
565                         fclose(fp);
566                 }
567         }
568
569 #if 0
570         for (int i = 0; i < LEN; ++i) {
571                 in[i] += rand() % 10000;
572         }
573 #endif
574
575 #if 0
576         for (int i = 0; i < LEN; ++i) {
577                 printf("%d\n", in[i]);
578         }
579 #endif
580
581         if (do_train) {
582                 spsa_train(pcm, sample_rate);
583                 exit(0);
584         }
585
586         std::vector<pulse> pulses = detect_pulses(pcm, sample_rate);
587
588         double calibration_factor = 1.0;
589         if (do_calibrate) {
590                 calibration_factor = calibrate(pulses);
591         }
592
593         if (output_cycles_plot) {
594                 output_cycle_plot(pulses, calibration_factor);
595         }
596
597         output_tap(pulses, calibration_factor);
598 }