1 // Copyright Steinar H. Gunderson <sgunderson@bigfoot.com>
2 // Licensed under the GPL, v2. (See the file COPYING.)
11 #include <immintrin.h>
16 #include "audioreader.h"
17 #include "interpolate.h"
23 #define C64_FREQUENCY 985248
24 #define SYNC_PULSE_START 1000
25 #define SYNC_PULSE_END 20000
26 #define SYNC_PULSE_LENGTH 378.0
27 #define SYNC_TEST_TOLERANCE 1.10
30 #define NUM_FILTER_COEFF 32
31 #define NUM_SPSA_VALS (NUM_FILTER_COEFF + 2)
33 #define A NUM_ITER/10 // approx
34 #define INITIAL_A 0.005 // A bit of trial and error...
35 #define INITIAL_C 0.02 // This too.
39 static float hysteresis_upper_limit = 0.1;
40 static float hysteresis_lower_limit = -0.1;
41 static bool do_calibrate = true;
42 static bool output_cycles_plot = false;
43 static bool do_crop = false;
44 static float crop_start = 0.0f, crop_end = HUGE_VAL;
46 static bool use_fir_filter = false;
47 static float filter_coeff[NUM_FILTER_COEFF] = { 1.0f }; // The rest is filled with 0.
48 static bool use_rc_filter = false;
49 static float rc_filter_freq;
50 static bool output_filtered = false;
52 static bool quiet = false;
53 static bool do_auto_level = false;
54 static bool output_leveled = false;
55 static std::vector<float> train_snap_points;
56 static bool do_train = false;
58 // The frequency to filter on (for do_auto_level), in Hertz.
59 // Larger values makes the compressor react faster, but if it is too large,
60 // you'll ruin the waveforms themselves.
61 static float auto_level_freq = 200.0;
63 // The minimum estimated sound level (for do_auto_level) at any given point.
64 // If you decrease this, you'll be able to amplify really silent signals
65 // by more, but you'll also increase the level of silent (ie. noise-only) segments,
66 // possibly caused misdetected pulses in these segments.
67 static float min_level = 0.05f;
69 // search for the value <limit> between [x,x+1]
71 double find_crossing(const std::vector<float> &pcm, int x, float limit)
74 // Do simple linear interpolation.
75 return x + (limit - pcm[x]) / (pcm[x + 1] - pcm[x]);
77 // Binary search for the zero crossing as given by Lanczos interpolation.
80 while (lower - upper > 1e-3) {
81 double mid = 0.5f * (upper + lower);
82 if (lanczos_interpolate(pcm, mid) > limit) {
89 return 0.5f * (upper + lower);
94 double time; // in seconds from start
95 double len; // in seconds
98 // Calibrate on the first ~25k pulses (skip a few, just to be sure).
99 double calibrate(const std::vector<pulse> &pulses) {
100 if (pulses.size() < SYNC_PULSE_END) {
101 fprintf(stderr, "Too few pulses, not calibrating!\n");
105 int sync_pulse_end = -1;
106 double sync_pulse_stddev = -1.0;
108 // Compute the standard deviation (to check for uneven speeds).
109 // If it suddenly skyrockets, we assume that sync ended earlier
110 // than we thought (it should be 25000 cycles), and that we should
111 // calibrate on fewer cycles.
112 for (int try_end : { 2000, 4000, 5000, 7500, 10000, 15000, SYNC_PULSE_END }) {
114 for (int i = SYNC_PULSE_START; i < try_end; ++i) {
115 double cycles = pulses[i].len * C64_FREQUENCY;
116 sum2 += (cycles - SYNC_PULSE_LENGTH) * (cycles - SYNC_PULSE_LENGTH);
118 double stddev = sqrt(sum2 / (try_end - SYNC_PULSE_START - 1));
119 if (sync_pulse_end != -1 && stddev > 5.0 && stddev / sync_pulse_stddev > 1.3) {
120 fprintf(stderr, "Stopping at %d sync pulses because standard deviation would be too big (%.2f cycles); shorter-than-usual trailer?\n",
121 sync_pulse_end, stddev);
124 sync_pulse_end = try_end;
125 sync_pulse_stddev = stddev;
128 fprintf(stderr, "Sync pulse length standard deviation: %.2f cycles\n",
133 for (int i = SYNC_PULSE_START; i < sync_pulse_end; ++i) {
134 sum += pulses[i].len;
136 double mean_length = C64_FREQUENCY * sum / (sync_pulse_end - SYNC_PULSE_START);
137 double calibration_factor = SYNC_PULSE_LENGTH / mean_length;
139 fprintf(stderr, "Calibrated sync pulse length: %.2f -> %.2f (change %+.2f%%)\n",
140 mean_length, SYNC_PULSE_LENGTH, 100.0 * (calibration_factor - 1.0));
143 // Check for pulses outside +/- 10% (sign of misdetection).
144 for (int i = SYNC_PULSE_START; i < sync_pulse_end; ++i) {
145 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
146 if (cycles < SYNC_PULSE_LENGTH / SYNC_TEST_TOLERANCE || cycles > SYNC_PULSE_LENGTH * SYNC_TEST_TOLERANCE) {
147 fprintf(stderr, "Sync cycle with downflank at %.6f was detected at %.0f cycles; misdetect?\n",
148 pulses[i].time, cycles);
152 return calibration_factor;
155 void output_tap(const std::vector<pulse>& pulses, double calibration_factor)
157 std::vector<char> tap_data;
158 for (unsigned i = 0; i < pulses.size(); ++i) {
159 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
160 int len = lrintf(cycles / TAP_RESOLUTION);
161 if (i > SYNC_PULSE_END && (cycles < 100 || cycles > 800)) {
162 fprintf(stderr, "Cycle with downflank at %.6f was detected at %.0f cycles; misdetect?\n",
163 pulses[i].time, cycles);
166 tap_data.push_back(len);
168 int overflow_len = lrintf(cycles);
169 tap_data.push_back(0);
170 tap_data.push_back(overflow_len & 0xff);
171 tap_data.push_back((overflow_len >> 8) & 0xff);
172 tap_data.push_back(overflow_len >> 16);
177 memcpy(hdr.identifier, "C64-TAPE-RAW", 12);
179 hdr.reserved[0] = hdr.reserved[1] = hdr.reserved[2] = 0;
180 hdr.data_len = tap_data.size();
182 fwrite(&hdr, sizeof(hdr), 1, stdout);
183 fwrite(tap_data.data(), tap_data.size(), 1, stdout);
186 static struct option long_options[] = {
187 {"auto-level", 0, 0, 'a' },
188 {"auto-level-freq", required_argument, 0, 'b' },
189 {"output-leveled", 0, 0, 'A' },
190 {"min-level", required_argument, 0, 'm' },
191 {"no-calibrate", 0, 0, 's' },
192 {"plot-cycles", 0, 0, 'p' },
193 {"hysteresis-limit", required_argument, 0, 'l' },
194 {"filter", required_argument, 0, 'f' },
195 {"rc-filter", required_argument, 0, 'r' },
196 {"output-filtered", 0, 0, 'F' },
197 {"crop", required_argument, 0, 'c' },
198 {"train", required_argument, 0, 't' },
199 {"quiet", 0, 0, 'q' },
200 {"help", 0, 0, 'h' },
206 fprintf(stderr, "decode [OPTIONS] AUDIO-FILE > TAP-FILE\n");
207 fprintf(stderr, "\n");
208 fprintf(stderr, " -a, --auto-level automatically adjust amplitude levels throughout the file\n");
209 fprintf(stderr, " -b, --auto-level-freq minimum frequency in Hertz of corrected level changes (default 200 Hz)\n");
210 fprintf(stderr, " -A, --output-leveled output leveled waveform to leveled.raw\n");
211 fprintf(stderr, " -m, --min-level minimum estimated sound level (0..1) for --auto-level\n");
212 fprintf(stderr, " -s, --no-calibrate do not try to calibrate on sync pulse length\n");
213 fprintf(stderr, " -p, --plot-cycles output debugging info to cycles.plot\n");
214 fprintf(stderr, " -l, --hysteresis-limit U[:L] change amplitude threshold for ignoring pulses (-1..1)\n");
215 fprintf(stderr, " -f, --filter C1:C2:C3:... specify FIR filter (up to %d coefficients)\n", NUM_FILTER_COEFF);
216 fprintf(stderr, " -r, --rc-filter FREQ send signal through a highpass RC filter with given frequency (in Hertz)\n");
217 fprintf(stderr, " -F, --output-filtered output filtered waveform to filtered.raw\n");
218 fprintf(stderr, " -c, --crop START[:END] use only the given part of the file\n");
219 fprintf(stderr, " -t, --train LEN1:LEN2:... train a filter for detecting any of the given number of cycles\n");
220 fprintf(stderr, " (implies --no-calibrate and --quiet unless overridden)\n");
221 fprintf(stderr, " -q, --quiet suppress some informational messages\n");
222 fprintf(stderr, " -h, --help display this help, then exit\n");
226 void parse_options(int argc, char **argv)
229 int option_index = 0;
230 int c = getopt_long(argc, argv, "ab:Am:spl:f:r:Fc:t:qh", long_options, &option_index);
236 do_auto_level = true;
240 auto_level_freq = atof(optarg);
244 output_leveled = true;
248 min_level = atof(optarg);
252 do_calibrate = false;
256 output_cycles_plot = true;
260 const char *hyststr = strtok(optarg, ": ");
261 hysteresis_upper_limit = atof(hyststr);
262 hyststr = strtok(NULL, ": ");
263 if (hyststr == NULL) {
264 hysteresis_lower_limit = -hysteresis_upper_limit;
266 hysteresis_lower_limit = atof(hyststr);
272 const char *coeffstr = strtok(optarg, ": ");
274 while (coeff_index < NUM_FILTER_COEFF && coeffstr != NULL) {
275 filter_coeff[coeff_index++] = atof(coeffstr);
276 coeffstr = strtok(NULL, ": ");
278 use_fir_filter = true;
283 use_rc_filter = true;
284 rc_filter_freq = atof(optarg);
288 output_filtered = true;
292 const char *cropstr = strtok(optarg, ":");
293 crop_start = atof(cropstr);
294 cropstr = strtok(NULL, ":");
295 if (cropstr == NULL) {
298 crop_end = atof(cropstr);
305 const char *cyclestr = strtok(optarg, ":");
306 while (cyclestr != NULL) {
307 train_snap_points.push_back(atof(cyclestr));
308 cyclestr = strtok(NULL, ":");
312 // Set reasonable defaults (can be overridden later on the command line).
313 do_calibrate = false;
330 std::vector<float> crop(const std::vector<float>& pcm, float crop_start, float crop_end, int sample_rate)
332 size_t start_sample, end_sample;
333 if (crop_start >= 0.0f) {
334 start_sample = std::min<size_t>(lrintf(crop_start * sample_rate), pcm.size());
336 if (crop_end >= 0.0f) {
337 end_sample = std::min<size_t>(lrintf(crop_end * sample_rate), pcm.size());
339 return std::vector<float>(pcm.begin() + start_sample, pcm.begin() + end_sample);
342 std::vector<float> do_fir_filter(const std::vector<float>& pcm, const float* filter)
344 std::vector<float> filtered_pcm;
345 filtered_pcm.resize(pcm.size());
346 unsigned i = NUM_FILTER_COEFF;
348 unsigned avx_end = i + ((pcm.size() - i) & ~7);
349 for ( ; i < avx_end; i += 8) {
350 __m256 s = _mm256_setzero_ps();
351 for (int j = 0; j < NUM_FILTER_COEFF; ++j) {
352 __m256 f = _mm256_set1_ps(filter[j]);
353 s = _mm256_fmadd_ps(f, _mm256_load_ps(&pcm[i - j]), s);
355 _mm256_storeu_ps(&filtered_pcm[i], s);
358 // Do what we couldn't do with AVX (which is everything for non-AVX machines)
360 for (; i < pcm.size(); ++i) {
362 for (int j = 0; j < NUM_FILTER_COEFF; ++j) {
363 s += filter[j] * pcm[i - j];
368 if (output_filtered) {
369 FILE *fp = fopen("filtered.raw", "wb");
370 fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
377 std::vector<float> do_rc_filter(const std::vector<float>& pcm, float freq, int sample_rate)
379 // This is only a 6 dB/oct filter, which seemingly works better
380 // than the Filter class, which is a standard biquad (12 dB/oct).
381 // The b/c calculations come from libnyquist (atone.c);
382 // I haven't checked, but I suppose they fall out of the bilinear
383 // transform of the transfer function H(s) = s/(s + w).
384 std::vector<float> filtered_pcm;
385 filtered_pcm.resize(pcm.size());
386 const float b = 2.0f - cos(2.0 * M_PI * freq / sample_rate);
387 const float c = b - sqrt(b * b - 1.0f);
388 float prev_in = 0.0f;
389 float prev_out = 0.0f;
390 for (unsigned i = 0; i < pcm.size(); ++i) {
392 float out = c * (prev_out + in - prev_in);
393 filtered_pcm[i] = out;
398 if (output_filtered) {
399 FILE *fp = fopen("filtered.raw", "wb");
400 fwrite(filtered_pcm.data(), filtered_pcm.size() * sizeof(filtered_pcm[0]), 1, fp);
408 std::vector<pulse> detect_pulses(const std::vector<float> &pcm, float hysteresis_upper_limit, float hysteresis_lower_limit, int sample_rate)
410 std::vector<pulse> pulses;
413 enum State { START, ABOVE, BELOW } state = START;
414 double last_downflank = -1;
415 for (unsigned i = 0; i < pcm.size(); ++i) {
416 if (pcm[i] > hysteresis_upper_limit) {
418 } else if (pcm[i] < hysteresis_lower_limit) {
419 if (state == ABOVE) {
421 double t = find_crossing<fast>(pcm, i - 1, hysteresis_lower_limit) * (1.0 / sample_rate) + crop_start;
422 if (last_downflank > 0) {
425 p.len = t - last_downflank;
436 void output_cycle_plot(const std::vector<pulse> &pulses, double calibration_factor)
438 FILE *fp = fopen("cycles.plot", "w");
439 for (unsigned i = 0; i < pulses.size(); ++i) {
440 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
441 fprintf(fp, "%f %f\n", pulses[i].time, cycles);
446 std::pair<int, double> find_closest_point(double x, const std::vector<float> &points)
449 double best_dist = (x - points[0]) * (x - points[0]);
450 for (unsigned j = 1; j < train_snap_points.size(); ++j) {
451 double dist = (x - points[j]) * (x - points[j]);
452 if (dist < best_dist) {
457 return std::make_pair(best_point, best_dist);
460 float eval_badness(const std::vector<pulse>& pulses, double calibration_factor)
462 double sum_badness = 0.0;
463 for (unsigned i = 0; i < pulses.size(); ++i) {
464 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
465 if (cycles > 2000.0) cycles = 2000.0; // Don't make pauses arbitrarily bad.
466 std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, train_snap_points);
467 sum_badness += selected_point_and_sq_dist.second;
469 return sqrt(sum_badness / (pulses.size() - 1));
472 void find_kmeans(const std::vector<pulse> &pulses, double calibration_factor, const std::vector<float> &initial_centers)
474 std::vector<float> last_centers = initial_centers;
475 std::vector<float> sums;
476 std::vector<float> num;
477 sums.resize(initial_centers.size());
478 num.resize(initial_centers.size());
480 for (unsigned i = 0; i < initial_centers.size(); ++i) {
484 for (unsigned i = 0; i < pulses.size(); ++i) {
485 double cycles = pulses[i].len * calibration_factor * C64_FREQUENCY;
486 // Ignore heavy outliers, which are almost always long pauses.
487 if (cycles > 2000.0) {
490 std::pair<int, double> selected_point_and_sq_dist = find_closest_point(cycles, last_centers);
491 int p = selected_point_and_sq_dist.first;
495 bool any_moved = false;
496 for (unsigned i = 0; i < initial_centers.size(); ++i) {
498 fprintf(stderr, "K-means broke down, can't output new reference training points\n");
501 float new_center = sums[i] / num[i];
502 if (fabs(new_center - last_centers[i]) > 1e-3) {
505 last_centers[i] = new_center;
511 fprintf(stderr, "New reference training points:");
512 for (unsigned i = 0; i < last_centers.size(); ++i) {
513 fprintf(stderr, " %.3f", last_centers[i]);
515 fprintf(stderr, "\n");
518 void spsa_train(const std::vector<float> &pcm, int sample_rate)
520 float vals[NUM_SPSA_VALS] = { hysteresis_upper_limit, hysteresis_lower_limit, 1.0f }; // The rest is filled with 0.
522 float start_c = INITIAL_C;
523 double best_badness = HUGE_VAL;
525 for (int n = 1; n < NUM_ITER; ++n) {
526 float a = INITIAL_A * pow(n + A, -ALPHA);
527 float c = start_c * pow(n, -GAMMA);
529 // find a random perturbation
530 float p[NUM_SPSA_VALS];
531 float vals1[NUM_SPSA_VALS], vals2[NUM_SPSA_VALS];
532 for (int i = 0; i < NUM_SPSA_VALS; ++i) {
533 p[i] = (rand() % 2) ? 1.0 : -1.0;
534 vals1[i] = std::max(std::min(vals[i] - c * p[i], 1.0f), -1.0f);
535 vals2[i] = std::max(std::min(vals[i] + c * p[i], 1.0f), -1.0f);
538 std::vector<pulse> pulses1 = detect_pulses<true>(do_fir_filter(pcm, vals1 + 2), vals1[0], vals1[1], sample_rate);
539 std::vector<pulse> pulses2 = detect_pulses<true>(do_fir_filter(pcm, vals2 + 2), vals2[0], vals2[1], sample_rate);
540 float badness1 = eval_badness(pulses1, 1.0);
541 float badness2 = eval_badness(pulses2, 1.0);
543 // Find the gradient estimator
544 float g[NUM_SPSA_VALS];
545 for (int i = 0; i < NUM_SPSA_VALS; ++i) {
546 g[i] = (badness2 - badness1) / (2.0 * c * p[i]);
548 vals[i] = std::max(std::min(vals[i], 1.0f), -1.0f);
550 if (badness2 < badness1) {
551 std::swap(badness1, badness2);
552 std::swap(vals1, vals2);
553 std::swap(pulses1, pulses2);
555 if (badness1 < best_badness) {
556 fprintf(stderr, "\nNew best filter (badness=%f):", badness1);
557 for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
558 fprintf(stderr, " %.5f", vals1[i + 2]);
560 fprintf(stderr, ", hysteresis limits = %f %f\n", vals1[0], vals1[1]);
561 best_badness = badness1;
563 find_kmeans(pulses1, 1.0, train_snap_points);
565 if (output_cycles_plot) {
566 output_cycle_plot(pulses1, 1.0);
569 fprintf(stderr, "%d ", n);
574 int main(int argc, char **argv)
576 parse_options(argc, argv);
578 make_lanczos_weight_table();
579 std::vector<float> pcm;
581 if (!read_audio_file(argv[optind], &pcm, &sample_rate)) {
586 pcm = crop(pcm, crop_start, crop_end, sample_rate);
589 if (use_fir_filter) {
590 pcm = do_fir_filter(pcm, filter_coeff);
594 pcm = do_rc_filter(pcm, rc_filter_freq, sample_rate);
598 pcm = level_samples(pcm, min_level, auto_level_freq, sample_rate);
599 if (output_leveled) {
600 FILE *fp = fopen("leveled.raw", "wb");
601 fwrite(pcm.data(), pcm.size() * sizeof(pcm[0]), 1, fp);
607 for (int i = 0; i < LEN; ++i) {
608 in[i] += rand() % 10000;
613 for (int i = 0; i < LEN; ++i) {
614 printf("%d\n", in[i]);
619 spsa_train(pcm, sample_rate);
623 std::vector<pulse> pulses = detect_pulses<false>(pcm, hysteresis_upper_limit, hysteresis_lower_limit, sample_rate);
625 double calibration_factor = 1.0;
627 calibration_factor = calibrate(pulses);
630 if (output_cycles_plot) {
631 output_cycle_plot(pulses, calibration_factor);
634 output_tap(pulses, calibration_factor);