#version 450 #extension GL_GOOGLE_include_directive : require // The spectrogram, one workgroup per column, the same columns as the // waveform (waveform.comp), so the two line up in time. // // Two resolutions per column, because one window cannot serve both ends of // a log axis: 2048 samples resolve 23 Hz, which is sharp in time and fine // in the kilohertz, but leaves 20 to 100 Hz (a quarter of the height) to // four bins. So the low end is analysed again from the signal low-passed // and decimated to a quarter of the rate: the same 2048 points then span // 170 ms and resolve 6 Hz. Rows take the long analysis below 250 Hz, the // short one above 500 Hz, and blend in between. // // A single window's spectrum of noise scatters as much as its level (each // bin is chi-squared with 2 degrees of freedom), which in dB shows as // speckle full of black holes, and finer bins only make more of them. So // the long analysis averages three windows a quarter window apart in time // (Welch): steady lines stay as sharp, noise settles. Multitaper averaging // was tried and measured against a reference (see test/README.md): sine // tapers and DPSS both widen narrow lines into slabs and leak, which a // 96 dB scale shows. // // Each analysis is a Blackman-Harris window zero-padded to 4096 points, // transformed as 2048 complex points (even and odd samples as real and // imaginary parts) by an in-place radix-2 FFT in shared memory, then split // into the real signal's bins. Zoomed out, a column averages the power of // up to four windows spread across its samples, so it neither skips sound // nor flickers as the view moves; zoomed in, neighbouring windows overlap. // Rows are log-spaced: a Catmull-Rom spline through the bins' levels where // rows are finer than bins, the loudest bin where they are coarser. layout(local_size_x = 256) in; #include "SpectrogramParams.glsl" layout(set = 0, binding = 0) readonly buffer Tape { float tape[]; }; // MAX_ROWS rows of MAX_COLUMNS levels, dBFS; row 0 is the lowest frequency. layout(set = 1, binding = 0) buffer Levels { float levels[]; }; // Match layouts.zig. const uint MAX_COLUMNS = 2048; const uint W = 2048; const uint N = 4096; const uint HALF = N / 2; const uint LOG2_HALF = 11; const uint THREADS = 256; // The long analysis: every DECIMATE-th sample of the low-passed signal. const uint DECIMATE = 4; // Its bins needed below the blend: 600 Hz at 12 kHz / 4096 is bin 205. const uint LONG_BINS = 256; // Windows the long analysis averages per column, a quarter window apart. const uint WELCH = 3; const float BLEND_LO = 250.0; const float BLEND_HI = 500.0; const float PI = 3.14159265358979; const float SILENCE = -160.0; // Columns past the end of the tape. const float NO_TAPE = 1000.0; const float A0 = 0.35875; const float A1 = 0.48829; const float A2 = 0.14128; const float A3 = 0.01168; shared vec2 z[HALF]; // Power per bin, summed over windows; then dBFS. shared float short_bins[HALF + 1]; shared float long_bins[LONG_BINS]; float sample_at(int i) { return i >= 0 && i < int(u.samples) ? tape[u.offset + uint(i)] : 0.0; } // The signal low-passed for decimation: a triangle over 7 samples (two // boxes of 4), whose zeros sit on 12 and 24 kHz, exactly where anything // that would fold onto the low bins comes from. float low_passed(int i) { return (sample_at(i - 3) + 2.0 * sample_at(i - 2) + 3.0 * sample_at(i - 1) + 4.0 * sample_at(i) + 3.0 * sample_at(i + 1) + 2.0 * sample_at(i + 2) + sample_at(i + 3)) / 16.0; } float blackman_harris(uint i) { float a = 2.0 * PI * float(i) / float(W - 1); return A0 - A1 * cos(a) + A2 * cos(2.0 * a) - A3 * cos(3.0 * a); } vec2 mul(vec2 a, vec2 b) { return vec2(a.x * b.x - a.y * b.y, a.x * b.y + a.y * b.x); } float windowed(int centre, uint i, bool long_window) { if (long_window) return low_passed(centre + (int(i) - int(W / 2)) * int(DECIMATE)) * blackman_harris(i); return sample_at(centre + int(i) - int(W / 2)) * blackman_harris(i); } // Adds the power spectrum of the window centred on sample `centre`. void analyse(int centre, bool long_window) { uint t = gl_LocalInvocationID.x; // Even and odd samples packed as one complex sequence, in bit-reversed // order for the in-place transform. Past W it is all padding. for (uint i = t; i < HALF; i += THREADS) { vec2 v = vec2(0.0); if (2u * i < W) v = vec2(windowed(centre, 2u * i, long_window), windowed(centre, 2u * i + 1u, long_window)); z[bitfieldReverse(i) >> (32u - LOG2_HALF)] = v; } barrier(); for (uint s = 1u; s <= LOG2_HALF; s++) { uint half_size = 1u << (s - 1u); for (uint j = t; j < HALF / 2u; j += THREADS) { uint k = j & (half_size - 1u); uint i0 = ((j >> (s - 1u)) << s) + k; uint i1 = i0 + half_size; float angle = -PI * float(k) / float(half_size); vec2 b = mul(vec2(cos(angle), sin(angle)), z[i1]); vec2 a = z[i0]; z[i0] = a + b; z[i1] = a - b; } barrier(); } // The real signal's bins from the packed transform: // X[k] = E[k] + e^(-2 pi i k / N) O[k], with E and O the even and odd // samples' spectra, recovered from Z[k] and conj(Z[HALF - k]). uint count = long_window ? LONG_BINS : HALF + 1u; for (uint k = t; k < count; k += THREADS) { vec2 zk = z[k % HALF]; vec2 zc = z[(HALF - k) % HALF]; zc.y = -zc.y; vec2 even = 0.5 * (zk + zc); vec2 d = 0.5 * (zk - zc); vec2 odd = vec2(d.y, -d.x); float angle = -2.0 * PI * float(k) / float(N); vec2 x = even + mul(vec2(cos(angle), sin(angle)), odd); if (long_window) long_bins[k] += dot(x, x); else short_bins[k] += dot(x, x); } barrier(); } float bin_level(uint k, bool long_window) { return long_window ? long_bins[min(k, LONG_BINS - 1u)] : short_bins[min(k, HALF)]; } // The level of the row spanning bins `lo` to `hi` of one analysis. float row_level(float lo, float hi, bool long_window) { if (hi - lo < 1.0) { float mid = sqrt(lo * hi); uint i = uint(floor(mid)); float f = mid - floor(mid); float p0 = bin_level(max(i, 1u) - 1u, long_window); float p1 = bin_level(i, long_window); float p2 = bin_level(i + 1u, long_window); float p3 = bin_level(i + 2u, long_window); float v = p1 + 0.5 * f * (p2 - p0 + f * (2.0 * p0 - 5.0 * p1 + 4.0 * p2 - p3 + f * (3.0 * (p1 - p2) + p3 - p0))); return clamp(v, SILENCE, max(max(p1, p2), max(p0, p3))); } float v = SILENCE; for (uint i = uint(floor(lo)); i <= uint(ceil(hi)); i++) v = max(v, bin_level(i, long_window)); return v; } // Average power over `count` windows to dBFS: a full-scale sine at a bin's // centre gives (W * A0 / 2) squared. float to_db(float power, uint count) { float ratio = power / float(count) / (0.25 * float(W) * A0 * float(W) * A0); return ratio > 0.0 ? max(10.0 * log(ratio) / log(10.0), SILENCE) : SILENCE; } void main() { uint c = gl_WorkGroupID.x; uint t = gl_LocalInvocationID.x; for (uint k = t; k <= HALF; k += THREADS) short_bins[k] = 0.0; for (uint k = t; k < LONG_BINS; k += THREADS) long_bins[k] = 0.0; barrier(); float start = float(c) * u.spp; uint short_windows = uint(clamp(ceil(u.spp / float(W / 2u)), 1.0, 4.0)); for (uint w = 0u; w < short_windows; w++) { analyse(int(float(u.first) + start + u.spp * (float(w) + 0.5) / float(short_windows)), false); } // Zoomed out, the windows spread across the column average already; // zoomed in, WELCH windows a quarter window apart around its centre. uint long_windows = uint(clamp(ceil(u.spp / float(W * DECIMATE / 2u)), 1.0, 4.0)); if (long_windows == 1u) { int centre = int(float(u.first) + start + 0.5 * u.spp); for (uint w = 0u; w < WELCH; w++) analyse(centre + (int(w) - int(WELCH / 2u)) * int(W * DECIMATE / 4u), true); long_windows = WELCH; } else { for (uint w = 0u; w < long_windows; w++) { analyse(int(float(u.first) + start + u.spp * (float(w) + 0.5) / float(long_windows)), true); } } for (uint k = t; k <= HALF; k += THREADS) short_bins[k] = to_db(short_bins[k], short_windows); for (uint k = t; k < LONG_BINS; k += THREADS) long_bins[k] = to_db(long_bins[k], long_windows); barrier(); bool inside = float(u.first) + start < float(u.samples); float short_hz = u.rate / float(N); float long_hz = short_hz / float(DECIMATE); float ratio = log(u.max_freq / u.min_freq); for (uint r = t; r < u.rows; r += THREADS) { float lo = u.min_freq * exp(ratio * float(r) / float(u.rows)); float hi = u.min_freq * exp(ratio * float(r + 1u) / float(u.rows)); float blend = smoothstep(BLEND_LO, BLEND_HI, sqrt(lo * hi)); float v = 0.0; if (blend < 1.0) v += (1.0 - blend) * row_level(lo / long_hz, hi / long_hz, true); if (blend > 0.0) v += blend * row_level(lo / short_hz, hi / short_hz, false); levels[r * MAX_COLUMNS + (u.slot + c) % u.ring] = inside ? v : NO_TAPE; } }