#include <assert.h>
#include <limits.h>
#include <getopt.h>
+#ifdef __AVX__
+#include <immintrin.h>
+#endif
#include <vector>
#include <algorithm>
// SPSA options
#define NUM_FILTER_COEFF 32
+#define NUM_SPSA_VALS (NUM_FILTER_COEFF + 2)
#define NUM_ITER 5000
#define A NUM_ITER/10 // approx
#define INITIAL_A 0.005 // A bit of trial and error...
#define GAMMA 0.166
#define ALPHA 1.0
-static float hysteresis_upper_limit = 3000.0 / 32768.0;
-static float hysteresis_lower_limit = -3000.0 / 32768.0;
+static float hysteresis_upper_limit = 0.1;
+static float hysteresis_lower_limit = -0.1;
static bool do_calibrate = true;
static bool output_cycles_plot = false;
static bool do_crop = false;
static float min_level = 0.05f;
// search for the value <limit> between [x,x+1]
+template<bool fast>
double find_crossing(const std::vector<float> &pcm, int x, float limit)
{
- double upper = x;
- double lower = x + 1;
- while (lower - upper > 1e-3) {
- double mid = 0.5f * (upper + lower);
- if (lanczos_interpolate(pcm, mid) > limit) {
- upper = mid;
- } else {
- lower = mid;
+ if (fast) {
+ // Do simple linear interpolation.
+ return x + (limit - pcm[x]) / (pcm[x + 1] - pcm[x]);
+ } else {
+ // Binary search for the zero crossing as given by Lanczos interpolation.
+ double upper = x;
+ double lower = x + 1;
+ while (lower - upper > 1e-3) {
+ double mid = 0.5f * (upper + lower);
+ if (lanczos_interpolate(pcm, mid) > limit) {
+ upper = mid;
+ } else {
+ lower = mid;
+ }
}
- }
- return 0.5f * (upper + lower);
+ return 0.5f * (upper + lower);
+ }
}
struct pulse {
{"rc-filter", required_argument, 0, 'r' },
{"output-filtered", 0, 0, 'F' },
{"crop", required_argument, 0, 'c' },
+ {"train", required_argument, 0, 't' },
{"quiet", 0, 0, 'q' },
{"help", 0, 0, 'h' },
{0, 0, 0, 0 }
fprintf(stderr, " -a, --auto-level automatically adjust amplitude levels throughout the file\n");
fprintf(stderr, " -b, --auto-level-freq minimum frequency in Hertz of corrected level changes (default 200 Hz)\n");
fprintf(stderr, " -A, --output-leveled output leveled waveform to leveled.raw\n");
- fprintf(stderr, " -m, --min-level minimum estimated sound level (0..32768) for --auto-level\n");
+ fprintf(stderr, " -m, --min-level minimum estimated sound level (0..1) for --auto-level\n");
fprintf(stderr, " -s, --no-calibrate do not try to calibrate on sync pulse length\n");
fprintf(stderr, " -p, --plot-cycles output debugging info to cycles.plot\n");
- fprintf(stderr, " -l, --hysteresis-limit VAL change amplitude threshold for ignoring pulses (0..32768)\n");
+ fprintf(stderr, " -l, --hysteresis-limit U[:L] change amplitude threshold for ignoring pulses (-1..1)\n");
fprintf(stderr, " -f, --filter C1:C2:C3:... specify FIR filter (up to %d coefficients)\n", NUM_FILTER_COEFF);
fprintf(stderr, " -r, --rc-filter FREQ send signal through a highpass RC filter with given frequency (in Hertz)\n");
fprintf(stderr, " -F, --output-filtered output filtered waveform to filtered.raw\n");
break;
case 'm':
- min_level = atof(optarg) / 32768.0;
+ min_level = atof(optarg);
break;
case 's':
case 'l': {
const char *hyststr = strtok(optarg, ": ");
- hysteresis_upper_limit = atof(hyststr) / 32768.0;
+ hysteresis_upper_limit = atof(hyststr);
hyststr = strtok(NULL, ": ");
if (hyststr == NULL) {
hysteresis_lower_limit = -hysteresis_upper_limit;
} else {
- hysteresis_lower_limit = atof(hyststr) / 32768.0;
+ hysteresis_lower_limit = atof(hyststr);
}
break;
}
return std::vector<float>(pcm.begin() + start_sample, pcm.begin() + end_sample);
}
-// TODO: Support AVX here.
std::vector<float> do_fir_filter(const std::vector<float>& pcm, const float* filter)
{
std::vector<float> filtered_pcm;
- filtered_pcm.reserve(pcm.size());
- for (unsigned i = NUM_FILTER_COEFF; i < pcm.size(); ++i) {
+ filtered_pcm.resize(pcm.size());
+ unsigned i = NUM_FILTER_COEFF;
+#ifdef __AVX__
+ unsigned avx_end = i + ((pcm.size() - i) & ~7);
+ for ( ; i < avx_end; i += 8) {
+ __m256 s = _mm256_setzero_ps();
+ for (int j = 0; j < NUM_FILTER_COEFF; ++j) {
+ __m256 f = _mm256_set1_ps(filter[j]);
+ s = _mm256_fmadd_ps(f, _mm256_load_ps(&pcm[i - j]), s);
+ }
+ _mm256_storeu_ps(&filtered_pcm[i], s);
+ }
+#endif
+ // Do what we couldn't do with AVX (which is everything for non-AVX machines)
+ // as scalar code.
+ for (; i < pcm.size(); ++i) {
float s = 0.0f;
for (int j = 0; j < NUM_FILTER_COEFF; ++j) {
s += filter[j] * pcm[i - j];
}
- filtered_pcm.push_back(s);
+ filtered_pcm[i] = s;
}
if (output_filtered) {
return filtered_pcm;
}
-std::vector<pulse> detect_pulses(const std::vector<float> &pcm, int sample_rate)
+template<bool fast>
+std::vector<pulse> detect_pulses(const std::vector<float> &pcm, float hysteresis_upper_limit, float hysteresis_lower_limit, int sample_rate)
{
std::vector<pulse> pulses;
} else if (pcm[i] < hysteresis_lower_limit) {
if (state == ABOVE) {
// down-flank!
- double t = find_crossing(pcm, i - 1, hysteresis_lower_limit) * (1.0 / sample_rate) + crop_start;
+ double t = find_crossing<fast>(pcm, i - 1, hysteresis_lower_limit) * (1.0 / sample_rate) + crop_start;
if (last_downflank > 0) {
pulse p;
p.time = t;
void spsa_train(const std::vector<float> &pcm, int sample_rate)
{
- float filter[NUM_FILTER_COEFF] = { 1.0f }; // The rest is filled with 0.
+ float vals[NUM_SPSA_VALS] = { hysteresis_upper_limit, hysteresis_lower_limit, 1.0f }; // The rest is filled with 0.
float start_c = INITIAL_C;
double best_badness = HUGE_VAL;
float c = start_c * pow(n, -GAMMA);
// find a random perturbation
- float p[NUM_FILTER_COEFF];
- float filter1[NUM_FILTER_COEFF], filter2[NUM_FILTER_COEFF];
- for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
+ float p[NUM_SPSA_VALS];
+ float vals1[NUM_SPSA_VALS], vals2[NUM_SPSA_VALS];
+ for (int i = 0; i < NUM_SPSA_VALS; ++i) {
p[i] = (rand() % 2) ? 1.0 : -1.0;
- filter1[i] = std::max(std::min(filter[i] - c * p[i], 1.0f), -1.0f);
- filter2[i] = std::max(std::min(filter[i] + c * p[i], 1.0f), -1.0f);
+ vals1[i] = std::max(std::min(vals[i] - c * p[i], 1.0f), -1.0f);
+ vals2[i] = std::max(std::min(vals[i] + c * p[i], 1.0f), -1.0f);
}
- std::vector<pulse> pulses1 = detect_pulses(do_fir_filter(pcm, filter1), sample_rate);
- std::vector<pulse> pulses2 = detect_pulses(do_fir_filter(pcm, filter2), sample_rate);
+ std::vector<pulse> pulses1 = detect_pulses<true>(do_fir_filter(pcm, vals1 + 2), vals1[0], vals1[1], sample_rate);
+ std::vector<pulse> pulses2 = detect_pulses<true>(do_fir_filter(pcm, vals2 + 2), vals2[0], vals2[1], sample_rate);
float badness1 = eval_badness(pulses1, 1.0);
float badness2 = eval_badness(pulses2, 1.0);
// Find the gradient estimator
- float g[NUM_FILTER_COEFF];
- for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
+ float g[NUM_SPSA_VALS];
+ for (int i = 0; i < NUM_SPSA_VALS; ++i) {
g[i] = (badness2 - badness1) / (2.0 * c * p[i]);
- filter[i] -= a * g[i];
- filter[i] = std::max(std::min(filter[i], 1.0f), -1.0f);
+ vals[i] -= a * g[i];
+ vals[i] = std::max(std::min(vals[i], 1.0f), -1.0f);
}
if (badness2 < badness1) {
std::swap(badness1, badness2);
- std::swap(filter1, filter2);
+ std::swap(vals1, vals2);
std::swap(pulses1, pulses2);
}
if (badness1 < best_badness) {
- printf("\nNew best filter (badness=%f):", badness1);
+ fprintf(stderr, "\nNew best filter (badness=%f):", badness1);
for (int i = 0; i < NUM_FILTER_COEFF; ++i) {
- printf(" %.5f", filter1[i]);
+ fprintf(stderr, " %.5f", vals1[i + 2]);
}
+ fprintf(stderr, ", hysteresis limits = %f %f\n", vals1[0], vals1[1]);
best_badness = badness1;
- printf("\n");
find_kmeans(pulses1, 1.0, train_snap_points);
output_cycle_plot(pulses1, 1.0);
}
}
- printf("%d ", n);
- fflush(stdout);
+ fprintf(stderr, "%d ", n);
+ fflush(stderr);
}
}
exit(0);
}
- std::vector<pulse> pulses = detect_pulses(pcm, sample_rate);
+ std::vector<pulse> pulses = detect_pulses<false>(pcm, hysteresis_upper_limit, hysteresis_lower_limit, sample_rate);
double calibration_factor = 1.0;
if (do_calibrate) {
projects / c64tapwav / blobdiff