diff options
| author | Gabriel Schneider <[email protected]> | 2026-11-17 00:15:15 -0300 |
|---|---|---|
| committer | Gabriel Schneider <[email protected]> | 2026-11-17 12:15:01 -0300 |
| commit | 9bd77e54943d954291b5faef287ccca5dcebd7a4 (patch) | |
| tree | 3f2d628781441568620b8108fab920713f26f649 /examples/micscope/shaders | |
| parent | e997fc3e6b312b30f5af90012fc22365bab701d4 (diff) | |
| download | sketchbook-9bd77e54943d954291b5faef287ccca5dcebd7a4.tar.gz sketchbook-9bd77e54943d954291b5faef287ccca5dcebd7a4.zip | |
examples: micscope and spheres come home, as they were, building
Both lived outside any repository, beside sketchbook, with manifests that
pointed at ../01-drat-meout/lib/sketchbook, which no longer exists. They
are copied into examples/micscope and examples/spheres (their sources,
shaders, tests, test material and web icons; not their caches, builds,
packages, state files or .gitignore, which the root's covers), their
manifests depend on "../.." as the other examples do, and their
READMEs link sketchbook there. Nothing else changed: both build against
today's library and pass their own zig build test (micscope 62/62 steps,
11/11 tests, its transport, record, render and tuner scripts; spheres
75/75 steps, 3/3 tests, its play script). The originals are untouched.
micscope's test/reference.flac (3.7 MiB, CC BY-SA 3.0, credited in its
test/README.md) is the rendering reference its render script opens. The
README lists both examples.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_016Y9MNBECMWcJD5pdS4SR6T
Diffstat (limited to 'examples/micscope/shaders')
| -rw-r--r-- | examples/micscope/shaders/spectrogram.comp | 208 | ||||
| -rw-r--r-- | examples/micscope/shaders/spectrogram.frag | 45 | ||||
| -rw-r--r-- | examples/micscope/shaders/spectrum.comp | 128 | ||||
| -rw-r--r-- | examples/micscope/shaders/spectrum.frag | 51 | ||||
| -rw-r--r-- | examples/micscope/shaders/tuner.comp | 134 | ||||
| -rw-r--r-- | examples/micscope/shaders/tuner_difference.comp | 49 | ||||
| -rw-r--r-- | examples/micscope/shaders/view.vert | 7 | ||||
| -rw-r--r-- | examples/micscope/shaders/waveform.comp | 72 | ||||
| -rw-r--r-- | examples/micscope/shaders/waveform.frag | 56 |
9 files changed, 750 insertions, 0 deletions
diff --git a/examples/micscope/shaders/spectrogram.comp b/examples/micscope/shaders/spectrogram.comp new file mode 100644 index 0000000..b32f4c5 --- /dev/null +++ b/examples/micscope/shaders/spectrogram.comp @@ -0,0 +1,208 @@ +#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; + } +} diff --git a/examples/micscope/shaders/spectrogram.frag b/examples/micscope/shaders/spectrogram.frag new file mode 100644 index 0000000..9cb5e3d --- /dev/null +++ b/examples/micscope/shaders/spectrogram.frag @@ -0,0 +1,45 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// The spectrogram (spectrogram.comp): time across, as the waveform above +// it, log frequency up, each level through a heat gradient (inferno: +// black, violet, red, orange, pale yellow), with the playhead. +layout(location = 0) out vec4 out_color; + +#include "SpectrogramView.glsl" + +layout(set = 2, binding = 0) readonly buffer Levels { float levels[]; }; + +// Match layouts.zig and spectrogram.comp. +const uint MAX_COLUMNS = 2048; +const float NO_TAPE = 1000.0; + +// Matplotlib's inferno, as a polynomial fit (Matt Zucker's, CC0). +vec3 inferno(float t) { + const vec3 c0 = vec3(0.0002189403691192265, 0.001651004631001012, -0.01948089843709184); + const vec3 c1 = vec3(0.1065134194856116, 0.5639564367884091, 3.932712388889277); + const vec3 c2 = vec3(11.60249308247187, -3.972853965665698, -15.9423941062914); + const vec3 c3 = vec3(-41.70399613139459, 17.43639888205313, 44.35414519872813); + const vec3 c4 = vec3(77.162935699427, -33.40235894210092, -81.80730925738993); + const vec3 c5 = vec3(-71.31942824499214, 32.62606426397723, 73.20951985803202); + const vec3 c6 = vec3(25.13112622477341, -12.24266895238567, -23.07032500287172); + return clamp(c0 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * (c5 + t * c6))))), 0.0, 1.0); +} + +void main() { + vec2 p = gl_FragCoord.xy; + uint row = uint(u.size.y - p.y); + uint x = uint(p.x); + float db = levels[row * MAX_COLUMNS + (x + u.scroll) % uint(u.size.x)]; + if (x + u.filled < uint(u.size.x)) db = NO_TAPE; + vec3 color = vec3(0.035, 0.03, 0.05); + if (db < NO_TAPE) { + float t = clamp((db - u.floor_db) / (u.ceil_db - u.floor_db), 0.0, 1.0); + // A touch of contrast: quiet stays dark, loud blooms. + color = inferno(pow(t, 1.3)); + } + if (u.playhead >= 0.0) { + float d = abs(p.x - u.playhead); + color = mix(color, vec3(1.0, 0.85, 0.3), clamp(1.5 - d, 0.0, 1.0) * 0.8); + } + out_color = vec4(color, 1.0); +} diff --git a/examples/micscope/shaders/spectrum.comp b/examples/micscope/shaders/spectrum.comp new file mode 100644 index 0000000..e576e30 --- /dev/null +++ b/examples/micscope/shaders/spectrum.comp @@ -0,0 +1,128 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// The spectrum at the playhead, entirely on the GPU, in one workgroup: +// +// 1. a Blackman-Harris window of the tape ending at `u.end`, zero-padded to +// N. Its sidelobes sit below -92 dB, so the finely sampled spectrum shows +// the signal rather than the window's ripples; +// 2. a radix-2 Stockham FFT, ping-ponging between the scratch halves; +// 3. levels in dBFS per bin; +// 4. one value per log-spaced column. Where columns are narrower than a bin +// (the low end), a Catmull-Rom spline through the bins' levels; where a +// column spans several bins (the high end), their loudest. The zero +// padding already interpolates between the window's own bins, so the +// curve is continuous at every zoom; +// 5. smoothing in time (fast attack, slow release) and falling peaks, kept +// in the column buffer between frames. +layout(local_size_x = 256) in; + +#include "SpectrumParams.glsl" + +layout(set = 0, binding = 0) readonly buffer Tape { float tape[]; }; +// Two halves of N complex values, which the workgroup's threads read from +// each other: coherent, for Vulkan; WGSL's storageBarrier covers it. +#ifdef SKETCHBOOK_WEB +layout(set = 1, binding = 0) buffer Scratch { vec2 scratch[]; }; +#else +layout(set = 1, binding = 0) coherent buffer Scratch { vec2 scratch[]; }; +#endif +// x: smoothed level, y: peak, both dBFS. +layout(set = 1, binding = 1) buffer Columns { vec2 columns[]; }; + +// Match layouts.zig. +const uint N = 8192; +const uint LOG2_N = 13; +const uint W = 4096; +const uint THREADS = 256; +const float PI = 3.14159265358979; +const float SILENCE = -160.0; +// Blackman-Harris, 4 terms. A0 is also its coherent gain. +const float A0 = 0.35875; +const float A1 = 0.48829; +const float A2 = 0.14128; +const float A3 = 0.01168; + +void sync() { + memoryBarrierBuffer(); + barrier(); +} + +void main() { + uint t = gl_LocalInvocationID.x; + + // 1. Window. + int first = int(u.end) - int(W); + for (uint i = t; i < N; i += THREADS) { + float x = 0.0; + int at = first + int(i); + if (i < W && at >= 0 && at < int(u.samples)) { + float a = 2.0 * PI * float(i) / float(W - 1); + x = tape[u.offset + uint(at)] * (A0 - A1 * cos(a) + A2 * cos(2.0 * a) - A3 * cos(3.0 * a)); + } + scratch[i] = vec2(x, 0.0); + } + sync(); + + // 2. FFT: after stage s, sub-transforms of 2p = 2^(s+1) points, in order. + for (uint s = 0; s < LOG2_N; s++) { + uint p = 1u << s; + uint from = (s & 1u) * N; + uint to = N - from; + for (uint i = t; i < N / 2; i += THREADS) { + uint k = i & (p - 1u); + vec2 a = scratch[from + i]; + vec2 b = scratch[from + i + N / 2]; + float angle = -PI * float(k) / float(p); + vec2 w = vec2(cos(angle), sin(angle)); + b = vec2(b.x * w.x - b.y * w.y, b.x * w.y + b.y * w.x); + uint j = (i << 1u) - k; + scratch[to + j] = a + b; + scratch[to + j + p] = a - b; + } + sync(); + } + uint result = (LOG2_N & 1u) * N; + + // 3. Levels into the other half. A full-scale sine peaks at W * A0 / 2: + // the window sums to W * A0, and a real sine splits between ±f. + uint levels = N - result; + for (uint i = t; i <= N / 2; i += THREADS) { + float amplitude = length(scratch[result + i]) * 2.0 / (float(W) * A0); + scratch[levels + i].x = amplitude > 0.0 ? max(20.0 * log(amplitude) / log(10.0), SILENCE) : SILENCE; + } + sync(); + + // 4 and 5. Columns. + float bin_hz = u.rate / float(N); + float ratio = log(u.max_freq / u.min_freq); + for (uint c = t; c < u.columns; c += THREADS) { + float lo = u.min_freq * exp(ratio * float(c) / float(u.columns)) / bin_hz; + float hi = u.min_freq * exp(ratio * float(c + 1u) / float(u.columns)) / bin_hz; + float mid = sqrt(lo * hi); + float v; + if (hi - lo < 1.0) { + uint i = uint(floor(mid)); + float f = mid - floor(mid); + float p0 = scratch[levels + min(max(i, 1u) - 1u, N / 2)].x; + float p1 = scratch[levels + min(i, N / 2)].x; + float p2 = scratch[levels + min(i + 1u, N / 2)].x; + float p3 = scratch[levels + min(i + 2u, N / 2)].x; + 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))); + v = clamp(v, SILENCE, max(max(p1, p2), max(p0, p3))); + } else { + v = SILENCE; + uint end = min(uint(ceil(hi)), N / 2); + for (uint i = uint(floor(lo)); i <= end; i++) v = max(v, scratch[levels + i].x); + } + vec2 old = columns[c]; + vec2 now; + if (u.reset != 0u) { + now = vec2(v, v); + } else { + float rate = v > old.x ? 40.0 : 6.0; + now.x = mix(old.x, v, 1.0 - exp(-rate * u.dt)); + now.y = max(now.x, old.y - 24.0 * u.dt); + } + columns[c] = now; + } +} diff --git a/examples/micscope/shaders/spectrum.frag b/examples/micscope/shaders/spectrum.frag new file mode 100644 index 0000000..e88b091 --- /dev/null +++ b/examples/micscope/shaders/spectrum.frag @@ -0,0 +1,51 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// The spectrum from its columns (spectrum.comp): one column per pixel on a +// log frequency axis. The curve is drawn by its distance along the normal, +// from the neighbours' slope, so steep parts stay as thick as flat ones. +layout(location = 0) out vec4 out_color; + +#include "SpectrumView.glsl" + +layout(set = 2, binding = 0) readonly buffer Columns { vec2 columns[]; }; + +float y_of(float db) { + return u.size.y * (1.0 - clamp((db - u.floor_db) / (u.ceil_db - u.floor_db), 0.0, 1.0)); +} + +float x_of(float hz) { + return u.size.x * log(hz / u.min_freq) / log(u.max_freq / u.min_freq); +} + +// How close `p` is to a grid line, 0..1. +float grid(vec2 p) { + const float lines[10] = float[](20.0, 50.0, 100.0, 200.0, 500.0, 1000.0, 2000.0, 5000.0, 10000.0, 20000.0); + float near = 1e9; + for (int i = 0; i < 10; i++) near = min(near, abs(p.x - x_of(lines[i]))); + float step = u.size.y * 12.0 / (u.ceil_db - u.floor_db); + float db = mod(p.y, step); + near = min(near, min(db, step - db)); + return clamp(1.0 - near, 0.0, 1.0); +} + +void main() { + vec2 p = gl_FragCoord.xy; + int n = int(u.size.x); + int x = int(p.x); + float y = y_of(columns[x].x); + float slope = (y_of(columns[min(x + 1, n - 1)].x) - y_of(columns[max(x - 1, 0)].x)) * 0.5; + float distance = abs(p.y - y) / sqrt(1.0 + slope * slope); + + vec3 color = vec3(0.055, 0.06, 0.07); + color = mix(color, vec3(0.16, 0.17, 0.2), grid(p)); + + // Fill under the curve, by frequency, fading downwards. + float along = p.x / u.size.x; + vec3 hue = mix(vec3(0.2, 0.75, 0.9), vec3(0.85, 0.35, 0.9), along); + if (p.y > y) color = mix(color, hue, 0.35 * (1.0 - (p.y - y) / max(u.size.y - y, 1.0)) + 0.08); + // Peaks, then the curve. + float peak = abs(p.y - y_of(columns[x].y)); + color = mix(color, vec3(0.9, 0.9, 0.95), clamp(1.0 - peak, 0.0, 1.0) * 0.45); + color = mix(color, mix(hue, vec3(1.0), 0.5), clamp(1.6 - distance, 0.0, 1.0)); + out_color = vec4(color, 1.0); +} diff --git a/examples/micscope/shaders/tuner.comp b/examples/micscope/shaders/tuner.comp new file mode 100644 index 0000000..1907f7f --- /dev/null +++ b/examples/micscope/shaders/tuner.comp @@ -0,0 +1,134 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// The pitch of the newest (or the playhead's) sound, by YIN (de Cheveigné +// and Kawahara, 2002), in one workgroup. Picking the strongest FFT peak +// finds the loudest harmonic, which is often not the note: low strings, +// voices and brass put more energy in the 2nd or 3rd harmonic, and a +// missing fundamental has no peak at all. YIN looks for the period instead: +// +// 1. the difference function d(tau) = sum over the window of +// (x[j] - x[j + tau])^2, which dips to 0 at every multiple of the +// period, whatever the harmonics' balance; +// 2. normalised by its running mean, d'(tau) = d(tau) * tau / sum d(1..tau), +// which starts at 1 and removes the dip-at-zero bias, so the dips are +// comparable; +// 3. the first dip below an absolute threshold, then the bottom of it: the +// shortest period that explains the signal, which keeps it from +// dropping an octave onto a multiple of the period; +// 4. a parabola through that dip of d and its neighbours, for a period +// between samples: the semitone is 6 %, the cent 0.06 %, and whole samples +// alone would be 30 cents apart at 1 kHz. +layout(local_size_x = 256) in; + +#include "TunerParams.glsl" + +layout(set = 0, binding = 0) readonly buffer Tape { float tape[]; }; +// d(tau), from tuner_difference.comp, which does the bulk of the work. +layout(set = 0, binding = 1) readonly buffer Difference { float difference[]; }; +// freq, clarity, level, period: layouts.Pitch. +layout(set = 1, binding = 0) buffer Result { vec4 result; }; + +// Match layouts.zig. +const uint W = 2048; +const uint LAGS = 1700; +const uint THREADS = 256; + +// Each thread owns CHUNK consecutive lags. +const uint CHUNK = (LAGS + THREADS - 1u) / THREADS; +shared float d[THREADS * CHUNK]; +// d before normalising: the parabola goes through these, since the +// normalisation's factor tau tilts a dip by more than a cent at high pitch. +shared float raw[THREADS * CHUNK]; +shared float squares[THREADS]; +shared float sums[THREADS]; +shared uint first_dip; +shared float deepest[THREADS]; +shared uint deepest_at[THREADS]; + +void main() { + uint t = gl_LocalInvocationID.x; + uint tau_max = min(LAGS - 2u, uint(ceil(u.rate / u.min_freq))); + uint tau_min = max(2u, uint(floor(u.rate / u.max_freq))); + + // 1. The difference function (tuner_difference.comp), and the window's + // energy. + int first = int(u.end) - int(W); + float sq = 0.0; + for (uint j = t; j < W; j += THREADS) { + int at = first + int(j); + float v = at >= 0 && at < int(u.samples) ? tape[u.offset + uint(at)] : 0.0; + sq += v * v; + } + squares[t] = sq; + float own = 0.0; + for (uint i = 0u; i < CHUNK; i++) { + uint tau = t * CHUNK + i; + float v = tau >= 1u && tau <= tau_max + 1u ? difference[tau] : 0.0; + raw[tau] = v; + own += v; + } + sums[t] = own; + if (t == 0u) first_dip = 0xffffffffu; + barrier(); + + // 2. Normalised by the running sum: a scan of the threads' chunk sums + // (Hillis-Steele), then each chunk in order. + for (uint stride = 1u; stride < THREADS; stride <<= 1u) { + float add = t >= stride ? sums[t - stride] : 0.0; + barrier(); + sums[t] += add; + barrier(); + } + float running = t > 0u ? sums[t - 1u] : 0.0; + float low = 2.0; + uint low_at = 0u; + for (uint i = 0u; i < CHUNK; i++) { + uint tau = t * CHUNK + i; + running += raw[tau]; + float v = tau == 0u ? 1.0 : (running > 0.0 ? raw[tau] * float(tau) / running : 1.0); + d[tau] = v; + if (tau >= tau_min && tau <= tau_max) { + // 3. The first dip under the threshold, and the deepest. + if (v < u.threshold) atomicMin(first_dip, tau); + if (v < low) { + low = v; + low_at = tau; + } + } + } + deepest[t] = low; + deepest_at[t] = low_at; + barrier(); + for (uint stride = THREADS / 2u; stride > 0u; stride >>= 1u) { + if (t < stride && deepest[t + stride] < deepest[t]) { + deepest[t] = deepest[t + stride]; + deepest_at[t] = deepest_at[t + stride]; + } + barrier(); + } + if (t != 0u) return; + + float energy = 0.0; + for (uint i = 0u; i < THREADS; i++) energy += squares[i]; + float level = energy > 0.0 ? 10.0 * log(energy / float(W)) / log(10.0) : -160.0; + + // The first dip followed to its bottom; failing one, the deepest, which + // the CPU may still reject as unclear. + uint best = first_dip; + if (best == 0xffffffffu) { + best = max(deepest_at[0], tau_min); + } else { + while (best + 1u <= tau_max && d[best + 1u] < d[best]) best++; + } + + // 4. The parabola through the dip. + float a = raw[best - 1u]; + float b = raw[best]; + float c = raw[best + 1u]; + float bend = a - 2.0 * b + c; + float shift = bend > 0.0 ? clamp(0.5 * (a - c) / bend, -0.5, 0.5) : 0.0; + float period = float(best) + shift; + float depth = d[best]; + bool silent = energy <= 0.0; + result = vec4(silent ? 0.0 : u.rate / period, silent ? 0.0 : clamp(1.0 - depth, 0.0, 1.0), level, period); +} diff --git a/examples/micscope/shaders/tuner_difference.comp b/examples/micscope/shaders/tuner_difference.comp new file mode 100644 index 0000000..c7d1b6e --- /dev/null +++ b/examples/micscope/shaders/tuner_difference.comp @@ -0,0 +1,49 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// YIN's difference function for tuner.comp, spread over many workgroups: +// d(tau) = sum over the newest W samples of (x[j] - x[j - tau])^2. Each +// workgroup takes LAGS_PER_GROUP lags, with THREADS / LAGS_PER_GROUP +// threads summing every so-many term of each, then adding up. +layout(local_size_x = 256) in; + +#include "TunerParams.glsl" + +layout(set = 0, binding = 0) readonly buffer Tape { float tape[]; }; +layout(set = 1, binding = 0) buffer Difference { float d[]; }; + +// Match layouts.zig and tuner.comp. +const uint W = 2048; +const uint LAGS = 1700; +const uint THREADS = 256; +const uint LAGS_PER_GROUP = 16; +const uint PER_LAG = THREADS / LAGS_PER_GROUP; + +shared float x[W + LAGS]; +shared float partial[THREADS]; + +void main() { + uint t = gl_LocalInvocationID.x; + int first = int(u.end) - int(W + LAGS); + for (uint i = t; i < W + LAGS; i += THREADS) { + int at = first + int(i); + x[i] = at >= 0 && at < int(u.samples) ? tape[u.offset + uint(at)] : 0.0; + } + barrier(); + + uint tau = gl_WorkGroupID.x * LAGS_PER_GROUP + t / PER_LAG; + uint lane = t % PER_LAG; + float sum = 0.0; + if (tau < LAGS) { + for (uint j = lane; j < W; j += PER_LAG) { + float diff = x[LAGS + j] - x[LAGS + j - tau]; + sum += diff * diff; + } + } + partial[t] = sum; + barrier(); + for (uint stride = PER_LAG / 2u; stride > 0u; stride >>= 1u) { + if (lane < stride) partial[t] += partial[t + stride]; + barrier(); + } + if (lane == 0u && tau < LAGS) d[tau] = partial[t]; +} diff --git a/examples/micscope/shaders/view.vert b/examples/micscope/shaders/view.vert new file mode 100644 index 0000000..d9461e9 --- /dev/null +++ b/examples/micscope/shaders/view.vert @@ -0,0 +1,7 @@ +#version 450 +// One triangle over the whole target, from the vertex index alone: the +// fragment shaders work in pixels (gl_FragCoord). +void main() { + vec2 p = vec2((gl_VertexIndex << 1) & 2, gl_VertexIndex & 2); + gl_Position = vec4(p * 2.0 - 1.0, 0.0, 1.0); +} diff --git a/examples/micscope/shaders/waveform.comp b/examples/micscope/shaders/waveform.comp new file mode 100644 index 0000000..d1d64e5 --- /dev/null +++ b/examples/micscope/shaders/waveform.comp @@ -0,0 +1,72 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// One workgroup per waveform column: the lowest and highest the signal goes +// in it, and its RMS. Each column also includes the linearly interpolated +// signal at its own edges, which it shares with its neighbours, so that +// zoomed in past one sample per column the columns still join into one +// continuous line. +layout(local_size_x = 64) in; + +#include "WaveParams.glsl" + +layout(set = 0, binding = 0) readonly buffer Tape { float tape[]; }; +// x: min, y: max, z: RMS, w: 1 where the column has tape under it. +layout(set = 1, binding = 0) buffer Columns { vec4 columns[]; }; + +const uint THREADS = 64; +shared float lows[THREADS]; +shared float highs[THREADS]; +shared float squares[THREADS]; + +float sample_at(int i) { + return i >= 0 && i < int(u.samples) ? tape[u.offset + uint(i)] : 0.0; +} + +// The signal at `t` samples after `u.first`, between samples linearly. +float signal(float t) { + float whole = floor(t); + int i = int(u.first) + int(whole); + return mix(sample_at(i), sample_at(i + 1), t - whole); +} + +void main() { + uint c = gl_WorkGroupID.x; + uint t = gl_LocalInvocationID.x; + float start = float(c) * u.spp; + float end = start + u.spp; + int first = int(ceil(start)); + int last = min(int(floor(end)), int(u.samples) - int(u.first) - 1); + + float lo = 1e9; + float hi = -1e9; + float sq = 0.0; + for (int i = first + int(t); i <= last; i += int(THREADS)) { + float v = tape[u.offset + u.first + uint(i)]; + lo = min(lo, v); + hi = max(hi, v); + sq += v * v; + } + if (t == 0u) { + float a = signal(start); + float b = signal(min(end, float(u.samples - u.first - 1u))); + lo = min(lo, min(a, b)); + hi = max(hi, max(a, b)); + } + lows[t] = lo; + highs[t] = hi; + squares[t] = sq; + barrier(); + for (uint stride = THREADS / 2u; stride > 0u; stride >>= 1u) { + if (t < stride) { + lows[t] = min(lows[t], lows[t + stride]); + highs[t] = max(highs[t], highs[t + stride]); + squares[t] += squares[t + stride]; + } + barrier(); + } + if (t == 0u) { + float count = float(max(last - first + 1, 1)); + float inside = float(u.first) + start < float(u.samples) ? 1.0 : 0.0; + columns[(u.slot + c) % u.ring] = vec4(lows[0], highs[0], sqrt(squares[0] / count), inside); + } +} diff --git a/examples/micscope/shaders/waveform.frag b/examples/micscope/shaders/waveform.frag new file mode 100644 index 0000000..37be7ce --- /dev/null +++ b/examples/micscope/shaders/waveform.frag @@ -0,0 +1,56 @@ +#version 450 +#extension GL_GOOGLE_include_directive : require +// The waveform from its columns (waveform.comp): one column per pixel, the +// band between its low and high filled, anti-aliased, at least a pixel +// thick; the RMS inside it brighter; played and unplayed in two colors, +// red while recording, green for the live microphone. +layout(location = 0) out vec4 out_color; + +#include "WaveView.glsl" + +layout(set = 2, binding = 0) readonly buffer Columns { vec4 columns[]; }; + +// Match layouts.Tint. +const uint TINT_RECORDING = 1u; +const uint TINT_LIVE = 2u; + +void main() { + vec2 p = gl_FragCoord.xy; + uint x = uint(p.x); + vec4 col = columns[(x + u.scroll) % uint(u.size.x)]; + if (x + u.filled < uint(u.size.x)) col.w = 0.0; + float mid = u.size.y * 0.5; + float scale = u.size.y * 0.46 * u.gain; + + vec3 color = vec3(0.055, 0.06, 0.07); + if (col.w > 0.0) color = vec3(0.075, 0.08, 0.095); + // Center line. + color = mix(color, vec3(0.2, 0.22, 0.26), clamp(1.0 - abs(p.y - mid), 0.0, 1.0) * 0.6); + + if (col.w > 0.0) { + float top = mid - col.y * scale - 0.5; + float bottom = mid - col.x * scale + 0.5; + float outside = max(top - p.y, p.y - bottom); + float cover = clamp(0.5 - outside, 0.0, 1.0); + float rms = col.z * scale; + float inner = clamp(rms - abs(p.y - mid) + 0.5, 0.0, 1.0); + bool played = u.playhead >= 0.0 && p.x < u.playhead; + vec3 band = played ? vec3(0.35, 0.8, 0.95) : vec3(0.45, 0.5, 0.58); + vec3 core = played ? vec3(0.7, 0.95, 1.0) : vec3(0.65, 0.7, 0.78); + if (u.tint == TINT_RECORDING) { + band = vec3(0.95, 0.35, 0.35); + core = vec3(1.0, 0.6, 0.55); + } else if (u.tint == TINT_LIVE) { + band = vec3(0.35, 0.75, 0.5); + core = vec3(0.65, 1.0, 0.75); + } + color = mix(color, mix(band, core, inner), cover); + } + + // Playhead. + if (u.playhead >= 0.0) { + float d = abs(p.x - u.playhead); + color = mix(color, vec3(1.0, 0.85, 0.3), clamp(1.5 - d, 0.0, 1.0)); + } + out_color = vec4(color, 1.0); +} |
