1 // Copyright Steinar H. Gunderson <sgunderson@bigfoot.com>
2 // Licensed under the GPL, v2. (See the file COPYING.)
13 #include "audioreader.h"
14 #include "interpolate.h"
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
27 #define NUM_FILTER_COEFF 32
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.
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;
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;
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;
54 // The frequency to filter on (for do_auto_level), in Hertz.
55 // Larger values makes the compressor react faster, but if it is too large,
56 // you'll ruin the waveforms themselves.
57 static float auto_level_freq = 200.0;
59 // The minimum estimated sound level (for do_auto_level) at any given point.
60 // If you decrease this, you'll be able to amplify really silent signals
61 // by more, but you'll also increase the level of silent (ie. noise-only) segments,
62 // possibly caused misdetected pulses in these segments.
63 static float min_level = 0.05f;
65 // search for the value <limit> between [x,x+1]
66 double find_crossing(const std::vector<float> &pcm, int x, float limit)
70 while (lower - upper > 1e-3) {
71 double mid = 0.5f * (upper + lower);
72 if (lanczos_interpolate(pcm, mid) > limit) {
79 return 0.5f * (upper + lower);
83 double time; // in seconds from start
84 double len; // in seconds
87 // Calibrate on the first ~25k pulses (skip a few, just to be sure).
88 double calibrate(const std::vector<pulse> &pulses) {
89 if (pulses.size() < SYNC_PULSE_END) {
90 fprintf(stderr, "Too few pulses, not calibrating!\n");
94 int sync_pulse_end = -1;
95 double sync_pulse_stddev = -1.0;
97 // Compute the standard deviation (to check for uneven speeds).
98 // If it suddenly skyrockets, we assume that sync ended earlier
99 // than we thought (it should be 25000 cycles), and that we should
100 // calibrate on fewer cycles.
101 for (int try_end : { 2000, 4000, 5000, 7500, 10000, 15000, SYNC_PULSE_END }) {
103 for (int i = SYNC_PULSE_START; i < try_end; ++i) {
104 double cycles = pulses[i].len * C64_FREQUENCY;
105 sum2 += (cycles - SYNC_PULSE_LENGTH) * (cycles - SYNC_PULSE_LENGTH);
107 double stddev = sqrt(sum2 / (try_end - SYNC_PULSE_START - 1));
108 if (sync_pulse_end != -1 && stddev > 5.0 && stddev / sync_pulse_stddev > 1.3) {
109 fprintf(stderr, "Stopping at %d sync pulses because standard deviation would be too big (%.2f cycles); shorter-than-usual trailer?\n",
110 sync_pulse_end, stddev);
113 sync_pulse_end = try_end;
114 sync_pulse_stddev = stddev;
117 fprintf(stderr, "Sync pulse length standard deviation: %.2f cycles\n",
122 for (int i = SYNC_PULSE_START; i < sync_pulse_end; ++i) {
123 sum += pulses[i].len;
125 double mean_length = C64_FREQUENCY * sum / (sync_pulse_end - SYNC_PULSE_START);
126 double calibration_factor = SYNC_PULSE_LENGTH / mean_length;
128 fprintf(stderr, "Calibrated sync pulse length: %.2f -> %.2f (change %+.2f%%)\n",
129 mean_length, SYNC_PULSE_LENGTH, 100.0 * (calibration_factor - 1.0));
132 // Check for pulses outside +/- 10% (sign of misdetection).
133 for (int i = SYNC_PULSE_START; i < sync_pulse_end; ++i) {
134 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
135 if (cycles < SYNC_PULSE_LENGTH / SYNC_TEST_TOLERANCE || cycles > SYNC_PULSE_LENGTH * SYNC_TEST_TOLERANCE) {
136 fprintf(stderr, "Sync cycle with downflank at %.6f was detected at %.0f cycles; misdetect?\n",
137 pulses[i].time, cycles);
141 return calibration_factor;
144 void output_tap(const std::vector<pulse>& pulses, double calibration_factor)
146 std::vector<char> tap_data;
147 for (unsigned i = 0; i < pulses.size(); ++i) {
148 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
149 int len = lrintf(cycles / TAP_RESOLUTION);
150 if (i > SYNC_PULSE_END && (cycles < 100 || cycles > 800)) {
151 fprintf(stderr, "Cycle with downflank at %.6f was detected at %.0f cycles; misdetect?\n",
152 pulses[i].time, cycles);
155 tap_data.push_back(len);
157 int overflow_len = lrintf(cycles);
158 tap_data.push_back(0);
159 tap_data.push_back(overflow_len & 0xff);
160 tap_data.push_back((overflow_len >> 8) & 0xff);
161 tap_data.push_back(overflow_len >> 16);
166 memcpy(hdr.identifier, "C64-TAPE-RAW", 12);
168 hdr.reserved[0] = hdr.reserved[1] = hdr.reserved[2] = 0;
169 hdr.data_len = tap_data.size();
171 fwrite(&hdr, sizeof(hdr), 1, stdout);
172 fwrite(tap_data.data(), tap_data.size(), 1, stdout);
175 static struct option long_options[] = {
176 {"auto-level", 0, 0, 'a' },
177 {"auto-level-freq", required_argument, 0, 'b' },
178 {"output-leveled", 0, 0, 'A' },
179 {"min-level", required_argument, 0, 'm' },
180 {"no-calibrate", 0, 0, 's' },
181 {"plot-cycles", 0, 0, 'p' },
182 {"hysteresis-limit", required_argument, 0, 'l' },
183 {"filter", required_argument, 0, 'f' },
184 {"rc-filter", required_argument, 0, 'r' },
185 {"output-filtered", 0, 0, 'F' },
186 {"crop", required_argument, 0, 'c' },
187 {"quiet", 0, 0, 'q' },
188 {"help", 0, 0, 'h' },
194 fprintf(stderr, "decode [OPTIONS] AUDIO-FILE > TAP-FILE\n");
195 fprintf(stderr, "\n");
196 fprintf(stderr, " -a, --auto-level automatically adjust amplitude levels throughout the file\n");
197 fprintf(stderr, " -b, --auto-level-freq minimum frequency in Hertz of corrected level changes (default 200 Hz)\n");
198 fprintf(stderr, " -A, --output-leveled output leveled waveform to leveled.raw\n");
199 fprintf(stderr, " -m, --min-level minimum estimated sound level (0..32768) for --auto-level\n");
200 fprintf(stderr, " -s, --no-calibrate do not try to calibrate on sync pulse length\n");
201 fprintf(stderr, " -p, --plot-cycles output debugging info to cycles.plot\n");
202 fprintf(stderr, " -l, --hysteresis-limit VAL change amplitude threshold for ignoring pulses (0..32768)\n");
203 fprintf(stderr, " -f, --filter C1:C2:C3:... specify FIR filter (up to %d coefficients)\n", NUM_FILTER_COEFF);
204 fprintf(stderr, " -r, --rc-filter FREQ send signal through a highpass RC filter with given frequency (in Hertz)\n");
205 fprintf(stderr, " -F, --output-filtered output filtered waveform to filtered.raw\n");
206 fprintf(stderr, " -c, --crop START[:END] use only the given part of the file\n");
207 fprintf(stderr, " -t, --train LEN1:LEN2:... train a filter for detecting any of the given number of cycles\n");
208 fprintf(stderr, " (implies --no-calibrate and --quiet unless overridden)\n");
209 fprintf(stderr, " -q, --quiet suppress some informational messages\n");
210 fprintf(stderr, " -h, --help display this help, then exit\n");
214 void parse_options(int argc, char **argv)
217 int option_index = 0;
218 int c = getopt_long(argc, argv, "ab:Am:spl:f:r:Fc:t:qh", long_options, &option_index);
224 do_auto_level = true;
228 auto_level_freq = atof(optarg);
232 output_leveled = true;
236 min_level = atof(optarg) / 32768.0;
240 do_calibrate = false;
244 output_cycles_plot = true;
248 const char *hyststr = strtok(optarg, ": ");
249 hysteresis_upper_limit = atof(hyststr) / 32768.0;
250 hyststr = strtok(NULL, ": ");
251 if (hyststr == NULL) {
252 hysteresis_lower_limit = -hysteresis_upper_limit;
254 hysteresis_lower_limit = atof(hyststr) / 32768.0;
260 const char *coeffstr = strtok(optarg, ": ");
262 while (coeff_index < NUM_FILTER_COEFF && coeffstr != NULL) {
263 filter_coeff[coeff_index++] = atof(coeffstr);
264 coeffstr = strtok(NULL, ": ");
266 use_fir_filter = true;
271 use_rc_filter = true;
272 rc_filter_freq = atof(optarg);
276 output_filtered = true;
280 const char *cropstr = strtok(optarg, ":");
281 crop_start = atof(cropstr);
282 cropstr = strtok(NULL, ":");
283 if (cropstr == NULL) {
286 crop_end = atof(cropstr);
293 const char *cyclestr = strtok(optarg, ":");
294 while (cyclestr != NULL) {
295 train_snap_points.push_back(atof(cyclestr));
296 cyclestr = strtok(NULL, ":");
300 // Set reasonable defaults (can be overridden later on the command line).
301 do_calibrate = false;
318 std::vector<float> crop(const std::vector<float>& pcm, float crop_start, float crop_end, int sample_rate)
320 size_t start_sample, end_sample;
321 if (crop_start >= 0.0f) {
322 start_sample = std::min<size_t>(lrintf(crop_start * sample_rate), pcm.size());
324 if (crop_end >= 0.0f) {
325 end_sample = std::min<size_t>(lrintf(crop_end * sample_rate), pcm.size());
327 return std::vector<float>(pcm.begin() + start_sample, pcm.begin() + end_sample);
330 // TODO: Support AVX here.
331 std::vector<float> do_fir_filter(const std::vector<float>& pcm, const float* filter)
333 std::vector<float> filtered_pcm;
334 filtered_pcm.reserve(pcm.size());
335 for (unsigned i = NUM_FILTER_COEFF; i < pcm.size(); ++i) {
337 for (int j = 0; j < NUM_FILTER_COEFF; ++j) {
338 s += filter[j] * pcm[i - j];
340 filtered_pcm.push_back(s);
343 if (output_filtered) {
344 FILE *fp = fopen("filtered.raw", "wb");
345 fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
352 std::vector<float> do_rc_filter(const std::vector<float>& pcm, float freq, int sample_rate)
354 // This is only a 6 dB/oct filter, which seemingly works better
355 // than the Filter class, which is a standard biquad (12 dB/oct).
356 // The b/c calculations come from libnyquist (atone.c);
357 // I haven't checked, but I suppose they fall out of the bilinear
358 // transform of the transfer function H(s) = s/(s + w).
359 std::vector<float> filtered_pcm;
360 filtered_pcm.resize(pcm.size());
361 const float b = 2.0f - cos(2.0 * M_PI * freq / sample_rate);
362 const float c = b - sqrt(b * b - 1.0f);
363 float prev_in = 0.0f;
364 float prev_out = 0.0f;
365 for (unsigned i = 0; i < pcm.size(); ++i) {
367 float out = c * (prev_out + in - prev_in);
368 filtered_pcm[i] = out;
373 if (output_filtered) {
374 FILE *fp = fopen("filtered.raw", "wb");
375 fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
382 std::vector<pulse> detect_pulses(const std::vector<float> &pcm, int sample_rate)
384 std::vector<pulse> pulses;
387 enum State { START, ABOVE, BELOW } state = START;
388 double last_downflank = -1;
389 for (unsigned i = 0; i < pcm.size(); ++i) {
390 if (pcm[i] > hysteresis_upper_limit) {
392 } else if (pcm[i] < hysteresis_lower_limit) {
393 if (state == ABOVE) {
395 double t = find_crossing(pcm, i - 1, hysteresis_lower_limit) * (1.0 / sample_rate) + crop_start;
396 if (last_downflank > 0) {
399 p.len = t - last_downflank;
410 void output_cycle_plot(const std::vector<pulse> &pulses, double calibration_factor)
412 FILE *fp = fopen("cycles.plot", "w");
413 for (unsigned i = 0; i < pulses.size(); ++i) {
414 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
415 fprintf(fp, "%f %f\n", pulses[i].time, cycles);
420 std::pair<int, double> find_closest_point(double x, const std::vector<float> &points)
423 double best_dist = (x - points[0]) * (x - points[0]);
424 for (unsigned j = 1; j < train_snap_points.size(); ++j) {
425 double dist = (x - points[j]) * (x - points[j]);
426 if (dist < best_dist) {
431 return std::make_pair(best_point, best_dist);
434 float eval_badness(const std::vector<pulse>& pulses, double calibration_factor)
436 double sum_badness = 0.0;
437 for (unsigned i = 0; i < pulses.size(); ++i) {
438 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
439 if (cycles > 2000.0) cycles = 2000.0; // Don't make pauses arbitrarily bad.
440 std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, train_snap_points);
441 sum_badness += selected_point_and_sq_dist.second;
443 return sqrt(sum_badness / (pulses.size() - 1));
446 void find_kmeans(const std::vector<pulse> &pulses, double calibration_factor, const std::vector<float> &initial_centers)
448 std::vector<float> last_centers = initial_centers;
449 std::vector<float> sums;
450 std::vector<float> num;
451 sums.resize(initial_centers.size());
452 num.resize(initial_centers.size());
454 for (unsigned i = 0; i < initial_centers.size(); ++i) {
458 for (unsigned i = 0; i < pulses.size(); ++i) {
459 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
460 // Ignore heavy outliers, which are almost always long pauses.
461 if (cycles > 2000.0) {
464 std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, last_centers);
465 int p = selected_point_and_sq_dist.first;
469 bool any_moved = false;
470 for (unsigned i = 0; i < initial_centers.size(); ++i) {
472 fprintf(stderr, "K-means broke down, can't output new reference training points\n");
475 float new_center = sums[i] / num[i];
476 if (fabs(new_center - last_centers[i]) > 1e-3) {
479 last_centers[i] = new_center;
485 fprintf(stderr, "New reference training points:");
486 for (unsigned i = 0; i < last_centers.size(); ++i) {
487 fprintf(stderr, " %.3f", last_centers[i]);
489 fprintf(stderr, "\n");
492 void spsa_train(const std::vector<float> &pcm, int sample_rate)
494 float filter[NUM_FILTER_COEFF] = { 1.0f }; // The rest is filled with 0.
496 float start_c = INITIAL_C;
497 double best_badness = HUGE_VAL;
499 for (int n = 1; n < NUM_ITER; ++n) {
500 float a = INITIAL_A * pow(n + A, -ALPHA);
501 float c = start_c * pow(n, -GAMMA);
503 // find a random perturbation
504 float p[NUM_FILTER_COEFF];
505 float filter1[NUM_FILTER_COEFF], filter2[NUM_FILTER_COEFF];
506 for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
507 p[i] = (rand() % 2) ? 1.0 : -1.0;
508 filter1[i] = std::max(std::min(filter[i] - c * p[i], 1.0f), -1.0f);
509 filter2[i] = std::max(std::min(filter[i] + c * p[i], 1.0f), -1.0f);
512 std::vector<pulse> pulses1 = detect_pulses(do_fir_filter(pcm, filter1), sample_rate);
513 std::vector<pulse> pulses2 = detect_pulses(do_fir_filter(pcm, filter2), sample_rate);
514 float badness1 = eval_badness(pulses1, 1.0);
515 float badness2 = eval_badness(pulses2, 1.0);
517 // Find the gradient estimator
518 float g[NUM_FILTER_COEFF];
519 for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
520 g[i] = (badness2 - badness1) / (2.0 * c * p[i]);
521 filter[i] -= a * g[i];
522 filter[i] = std::max(std::min(filter[i], 1.0f), -1.0f);
524 if (badness2 < badness1) {
525 std::swap(badness1, badness2);
526 std::swap(filter1, filter2);
527 std::swap(pulses1, pulses2);
529 if (badness1 < best_badness) {
530 printf("\nNew best filter (badness=%f):", badness1);
531 for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
532 printf(" %.5f", filter1[i]);
534 best_badness = badness1;
537 find_kmeans(pulses1, 1.0, train_snap_points);
539 if (output_cycles_plot) {
540 output_cycle_plot(pulses1, 1.0);
548 int main(int argc, char **argv)
550 parse_options(argc, argv);
552 make_lanczos_weight_table();
553 std::vector<float> pcm;
555 if (!read_audio_file(argv[optind], &pcm, &sample_rate)) {
560 pcm = crop(pcm, crop_start, crop_end, sample_rate);
563 if (use_fir_filter) {
564 pcm = do_fir_filter(pcm, filter_coeff);
568 pcm = do_rc_filter(pcm, rc_filter_freq, sample_rate);
572 pcm = level_samples(pcm, min_level, auto_level_freq, sample_rate);
573 if (output_leveled) {
574 FILE *fp = fopen("leveled.raw", "wb");
575 fwrite(pcm.data(), pcm.size() * sizeof(pcm[0]), 1, fp);
581 for (int i = 0; i < LEN; ++i) {
582 in[i] += rand() % 10000;
587 for (int i = 0; i < LEN; ++i) {
588 printf("%d\n", in[i]);
593 spsa_train(pcm, sample_rate);
597 std::vector<pulse> pulses = detect_pulses(pcm, sample_rate);
599 double calibration_factor = 1.0;
601 calibration_factor = calibrate(pulses);
604 if (output_cycles_plot) {
605 output_cycle_plot(pulses, calibration_factor);
608 output_tap(pulses, calibration_factor);