diff options
| author | Gabriel Schneider <[email protected]> | 2026-10-28 21:07:13 -0300 |
|---|---|---|
| committer | Gabriel Schneider <[email protected]> | 2026-11-12 13:22:02 -0300 |
| commit | 590ba7b9b579b97e34b51ab7e3460ebc85713343 (patch) | |
| tree | 6a73a7570ff4f79f41deff078455877b9bc6571b /src/scope | |
| parent | bc8e6eb1bfef0c5e9b3c40fc1e4fd5a1659e2b03 (diff) | |
| download | sketchbook-590ba7b9b579b97e34b51ab7e3460ebc85713343.tar.gz sketchbook-590ba7b9b579b97e34b51ab7e3460ebc85713343.zip | |
sketchbook: Zig 0.17, a Store on disk, requests, a scope (waveform, spectrum, spectrogram), audio codecs and fonts, web host
Diffstat (limited to 'src/scope')
| -rw-r--r-- | src/scope/layouts.zig | 120 | ||||
| -rw-r--r-- | src/scope/root.zig | 394 | ||||
| -rw-r--r-- | src/scope/shaders/spectrogram.comp | 208 | ||||
| -rw-r--r-- | src/scope/shaders/spectrogram.frag | 45 | ||||
| -rw-r--r-- | src/scope/shaders/spectrum.comp | 128 | ||||
| -rw-r--r-- | src/scope/shaders/spectrum.frag | 51 | ||||
| -rw-r--r-- | src/scope/shaders/view.vert | 7 | ||||
| -rw-r--r-- | src/scope/shaders/waveform.comp | 72 | ||||
| -rw-r--r-- | src/scope/shaders/waveform.frag | 56 |
9 files changed, 1081 insertions, 0 deletions
diff --git a/src/scope/layouts.zig b/src/scope/layouts.zig new file mode 100644 index 0000000..d8fc484 --- /dev/null +++ b/src/scope/layouts.zig @@ -0,0 +1,120 @@ +//! What the shaders read, as Zig structs: the only definition. The shaders +//! include GLSL generated from these (see sketchbook's `glsl.zig`). No +//! imports, so the shader build can compile this alone. + +/// FFT length: the analysis window, zero-padded to twice its size so the +/// spectrum is sampled twice as finely as the window resolves. +pub const fft_size = 8192; +pub const window = 4096; +/// The most columns a view can have; the column buffers hold this many. +pub const max_columns = 2048; +/// The tallest spectrogram; its buffer holds this many rows of +/// `max_columns`. +pub const max_rows = 1024; +// All three analyses read samples `offset + i` of their sample buffer for +// `0 <= i < samples`, so one buffer can hold several sounds. +/// The spectrogram's window, zero-padded to `spectrogram_fft` points. +pub const spectrogram_window = 2048; +pub const spectrogram_fft = 4096; + +/// spectrum.comp: the window ending at `end`, and the columns to fill. +pub const SpectrumParams = extern struct { + offset: u32, + /// Sample index one past the window's last sample. + end: u32, + /// Samples on the tape. + samples: u32, + columns: u32, + /// Nonzero: forget the smoothed levels and peaks. + reset: u32, + rate: f32, + min_freq: f32, + max_freq: f32, + /// Seconds since the previous analysis, for smoothing. + dt: f32, + pad0: u32 = 0, + pad1: u32 = 0, + pad2: u32 = 0, + + pub const glsl = .{ .uniform = .{ .set = 2, .instance = "u" } }; +}; + +/// waveform.comp: column `c` covers samples `first + c * spp` up to the next +/// column's start, and is written to slot `(slot + c) % ring`: the live +/// view writes its newest columns into a ring, the others from slot 0. +pub const WaveParams = extern struct { + offset: u32, + first: u32, + samples: u32, + columns: u32, + spp: f32, + slot: u32, + ring: u32, + pad: u32 = 0, + + pub const glsl = .{ .uniform = .{ .set = 2, .instance = "u" } }; +}; + +pub const Tint = enum(u32) { take, recording, live }; + +/// waveform.frag. Pixel column `x` shows slot `(x + scroll) % width`; only +/// the last `filled` columns have anything in them. +pub const WaveView = extern struct { + size: [2]f32, + /// Pixels; negative hides it. + playhead: f32, + gain: f32, + /// A `Tint`. + tint: u32, + scroll: u32, + filled: u32, + pad: u32 = 0, + + pub const glsl = .{ .uniform = .{ .set = 3, .instance = "u" } }; +}; + +/// spectrum.frag. +pub const SpectrumView = extern struct { + size: [2]f32, + floor_db: f32, + ceil_db: f32, + min_freq: f32, + max_freq: f32, + pad: [2]f32 = .{ 0, 0 }, + + pub const glsl = .{ .uniform = .{ .set = 3, .instance = "u" } }; +}; + +/// spectrogram.comp: the same columns as `WaveParams`, each analysed into +/// `rows` log-spaced frequency rows, row 0 the lowest. +/// Columns land in slots as `WaveParams` says. +pub const SpectrogramParams = extern struct { + offset: u32, + first: u32, + samples: u32, + columns: u32, + rows: u32, + spp: f32, + rate: f32, + min_freq: f32, + max_freq: f32, + slot: u32, + ring: u32, + pad: u32 = 0, + + pub const glsl = .{ .uniform = .{ .set = 2, .instance = "u" } }; +}; + +/// spectrogram.frag; `scroll` and `filled` as in `WaveView`. +pub const SpectrogramView = extern struct { + size: [2]f32, + floor_db: f32, + ceil_db: f32, + /// Pixels; negative hides it. + playhead: f32, + scroll: u32, + filled: u32, + pad: f32 = 0, + + pub const glsl = .{ .uniform = .{ .set = 3, .instance = "u" } }; +}; diff --git a/src/scope/root.zig b/src/scope/root.zig new file mode 100644 index 0000000..a71fa53 --- /dev/null +++ b/src/scope/root.zig @@ -0,0 +1,394 @@ +//! Live views of a stream of sound, analysed and drawn on the GPU: the last +//! few seconds as a scrolling waveform over its spectrogram, the smoothed +//! spectrum of the newest samples, and an oscilloscope. A sketch feeds a +//! `Scope` (kept in its Session) with `push` and draws it once a frame with +//! `draw`, giving the rectangles of the views it wants. +//! +//! The analyses (see the shaders for the details): +//! - spectrogram: a 2048-point Blackman-Harris window per column, and below +//! 500 Hz a second analysis decimated 4× (6 Hz resolution) averaged over +//! three windows; log-frequency rows; +//! - spectrum: a 4096-point window zero-padded to 8192, log-frequency +//! columns, fast attack and slow release, falling peaks; +//! - waveform: min, max and RMS per column, joined into a line when zoomed +//! past one sample per column. +//! +//! Columns of the scrolling views are analysed once, when everything their +//! windows reach has arrived, into rings the views read from the oldest; so +//! the spectrogram runs `reach` samples (about 130 ms) behind the newest. +const std = @import("std"); +const sk = @import("sketch"); +const gpu = sk.abi.gpu; +pub const layouts = @import("scope_layouts"); +const shaders = @import("scope_shaders"); + +pub const min_freq = 20.0; +pub const floor_db = -96.0; +/// The samples kept: enough for a frame's new columns and all they read. +pub const history = 1 << 16; +/// How far the spectrogram reads past a column's centre (its long +/// analysis: 2048 points decimated 4×, three windows a quarter apart). +pub const reach = 6160; + +pub const Tint = layouts.Tint; + +/// Where and how to draw; views left null are not analysed. +pub const Views = struct { + /// Seconds the waveform and spectrogram span, ending now. + seconds: f32 = 6, + /// Both scroll together, column for column: give them the same width. + wave: ?sk.Rect = null, + spectrogram: ?sk.Rect = null, + spectrum: ?sk.Rect = null, + oscilloscope: ?Oscilloscope = null, + gain: f32 = 1, + tint: Tint = .live, + + pub const Oscilloscope = struct { + rect: sk.Rect, + /// The samples across it. + samples: u32 = 1024, + /// How many samples before the newest it ends: a sketch steadies a + /// periodic wave by ending it at the same phase each frame. + end: u32 = 0, + }; +}; + +pub const Scope = struct { + samples: [history]f32 = @splat(0), + /// Samples pushed since the start. + heard: u64 = 0, + /// Pushed since the last `draw`, so the GPU copy is stale. + fresh: bool = false, + /// The next column to analyse, counted from the first sample; how many + /// have been since the view was (re)made; their count and width. + next: u64 = 0, + filled: u32 = 0, + columns: u32 = 0, + spp: u32 = 0, + gpu: Gpu = .{}, + + pub fn push(s: *Scope, heard: []const f32) void { + if (heard.len == 0) return; + const h = &s.samples; + if (heard.len >= h.len) { + @memcpy(h, heard[heard.len - h.len ..]); + } else { + std.mem.copyForwards(f32, h[0 .. h.len - heard.len], h[heard.len..]); + @memcpy(h[h.len - heard.len ..], heard); + } + s.heard += heard.len; + s.fresh = true; + } + + /// Analyses what is new and draws the views; needs `f.host.gpu()`. + pub fn draw(s: *Scope, f: *sk.Frame, v: Views) void { + const g = &s.gpu; + const rate: f64 = @floatFromInt(f.in.sample_rate); + g.setup(f); + if (f.in.reloaded) g.forget(); + if (s.fresh) f.upload(g.tape, 0, std.mem.sliceAsBytes(&s.samples)); + defer s.fresh = false; + + if (v.wave orelse v.spectrogram) |r| s.scroll(f, v, r, rate); + if (v.spectrum) |r| s.spectrum(f, r, rate); + if (v.oscilloscope) |o| s.oscilloscope(f, v, o); + } + + /// The waveform and spectrogram, scrolling. + fn scroll(s: *Scope, f: *sk.Frame, v: Views, r: sk.Rect, rate: f64) void { + const g = &s.gpu; + const columns: u32 = @intFromFloat(std.math.clamp(r.w, 1, layouts.max_columns)); + const spp: u32 = @max(1, @as(u32, @intFromFloat(@round(v.seconds * rate / @as(f64, @floatFromInt(columns)))))); + if (columns != s.columns or spp != s.spp or f.in.reloaded) { + s.columns = columns; + s.spp = spp; + s.filled = 0; + s.next = 0; + } + const rows: u32 = if (v.spectrogram) |sr| @intFromFloat(std.math.clamp(sr.h, 1, layouts.max_rows)) else 1; + g.levelsFor(f, rows); + + // History index of sample `i`: `i - base`. A column is ready once + // all it reads has arrived, and analysable while all it reads is kept. + const base: i64 = @as(i64, @intCast(s.heard)) - history; + const ready = (s.heard -| reach) / spp; + const kept: u64 = @intCast(@max(0, @divFloor(base + reach + spp - 1, spp))); + s.next = @max(s.next, kept, ready -| columns); + if (ready > s.next) { + const count: u32 = @intCast(ready - s.next); + const first: u32 = @intCast(@as(i64, @intCast(s.next * spp)) - base); + const slot: u32 = @intCast(s.next % columns); + if (v.wave != null) f.compute(.{ + .label = "scope waveform.comp", + .kernel = waveform_kernel, + .readonly = &.{g.tape}, + .readwrite = &.{g.wave.columns}, + .uniforms = gpu.uniform(f.scratch, layouts.WaveParams{ .offset = 0, .first = first, .samples = history, .columns = count, .spp = @floatFromInt(spp), .slot = slot, .ring = columns }), + .groups = .{ count, 1, 1 }, + }); + if (v.spectrogram != null) f.compute(.{ + .label = "scope spectrogram.comp", + .kernel = spectrogram_kernel, + .readonly = &.{g.tape}, + .readwrite = &.{g.levels}, + .uniforms = gpu.uniform(f.scratch, layouts.SpectrogramParams{ + .offset = 0, + .first = first, + .samples = history, + .columns = count, + .rows = rows, + .spp = @floatFromInt(spp), + .rate = @floatCast(rate), + .min_freq = min_freq, + .max_freq = maxFreq(rate), + .slot = slot, + .ring = columns, + }), + .groups = .{ count, 1, 1 }, + }); + s.filled = @intCast(@min(columns, s.filled + count)); + s.next = ready; + } + + const scroll_at: u32 = @intCast(s.next % columns); + if (v.wave) |wr| { + const size = target(f, &g.wave, wr, layouts.max_rows); + const view: layouts.WaveView = .{ + .size = .{ @floatFromInt(size[0]), @floatFromInt(size[1]) }, + .playhead = -1, + .gain = v.gain, + .tint = @intFromEnum(v.tint), + .scroll = scroll_at, + .filled = s.filled, + }; + if (g.wave.redraw(std.hash.Wyhash.hash(0, std.mem.asBytes(&view)))) wavePass(f, &g.wave, size, view); + f.image(g.wave.image, wr, .{}); + } + if (v.spectrogram) |sr| { + const size = target(f, &g.spectrogram, sr, rows); + const view: layouts.SpectrogramView = .{ + .size = .{ @floatFromInt(size[0]), @floatFromInt(size[1]) }, + .floor_db = floor_db, + .ceil_db = -6, + .playhead = -1, + .scroll = scroll_at, + .filled = s.filled, + }; + if (g.spectrogram.redraw(std.hash.Wyhash.hash(0, std.mem.asBytes(&view)))) f.gpu(.{ + .label = "scope spectrogram.frag", + .image = g.spectrogram.image, + .width = size[0], + .height = size[1], + .clear = null, + .pipelines = &spectrogram_pipelines, + .draws = &.{.{ .pipeline = 0, .vertex_count = 3, .fragment_uniforms = gpu.uniform(f.scratch, view), .fragment_storage = &.{g.levels} }}, + }); + f.image(g.spectrogram.image, sr, .{}); + } + } + + /// The spectrum of the newest samples, smoothed frame to frame. + fn spectrum(s: *Scope, f: *sk.Frame, r: sk.Rect, rate: f64) void { + const g = &s.gpu; + const size = target(f, &g.spectrum, r, layouts.max_rows); + const dt: f64 = @min(@as(f64, @floatFromInt(f.in.dt)) / std.time.ns_per_s, 0.1); + f.compute(.{ + .label = "scope spectrum.comp", + .kernel = spectrum_kernel, + .readonly = &.{g.tape}, + .readwrite = &.{ g.scratch, g.spectrum.columns }, + .uniforms = gpu.uniform(f.scratch, layouts.SpectrumParams{ + .offset = 0, + .end = history, + .samples = history, + .columns = size[0], + .reset = @intFromBool(!g.spectrum_started), + .rate = @floatCast(rate), + .min_freq = min_freq, + .max_freq = maxFreq(rate), + .dt = @floatCast(dt), + }), + }); + g.spectrum_started = true; + f.gpu(.{ + .label = "scope spectrum.frag", + .image = g.spectrum.image, + .width = size[0], + .height = size[1], + .clear = null, + .pipelines = &spectrum_pipelines, + .draws = &.{.{ + .pipeline = 0, + .vertex_count = 3, + .fragment_uniforms = gpu.uniform(f.scratch, layouts.SpectrumView{ + .size = .{ @floatFromInt(size[0]), @floatFromInt(size[1]) }, + .floor_db = floor_db, + .ceil_db = 0, + .min_freq = min_freq, + .max_freq = maxFreq(rate), + }), + .fragment_storage = &.{g.spectrum.columns}, + }}, + }); + f.image(g.spectrum.image, r, .{}); + } + + /// The newest `samples` (less `end`), as a line. + fn oscilloscope(s: *Scope, f: *sk.Frame, v: Views, o: Views.Oscilloscope) void { + const g = &s.gpu; + const size = target(f, &g.oscilloscope, o.rect, layouts.max_rows); + const n: u32 = @min(o.samples, history / 2); + const end: u32 = @min(o.end, history / 2); + const first: u32 = history - n - end; + f.compute(.{ + .label = "scope oscilloscope", + .kernel = waveform_kernel, + .readonly = &.{g.tape}, + .readwrite = &.{g.oscilloscope.columns}, + .uniforms = gpu.uniform(f.scratch, layouts.WaveParams{ + .offset = 0, + .first = first, + .samples = history, + .columns = size[0], + .spp = @as(f32, @floatFromInt(n)) / @as(f32, @floatFromInt(size[0])), + .slot = 0, + .ring = layouts.max_columns, + }), + .groups = .{ size[0], 1, 1 }, + }); + wavePass(f, &g.oscilloscope, size, .{ + .size = .{ @floatFromInt(size[0]), @floatFromInt(size[1]) }, + .playhead = -1, + .gain = v.gain, + .tint = @intFromEnum(v.tint), + .scroll = 0, + .filled = size[0], + }); + f.image(g.oscilloscope.image, o.rect, .{}); + } +}; + +/// The highest frequency the views show. +pub fn maxFreq(rate: f64) f32 { + return @floatCast(@min(20_000, rate / 2)); +} + +/// Where `hz` falls across a spectrum view (or up a spectrogram, from its +/// bottom): 0 to 1, log frequency. +pub fn place(hz: f64, rate: f64) f32 { + return @floatCast(@log(hz / min_freq) / @log(maxFreq(rate) / min_freq)); +} + +// --------------------------------------------------------------------------- +// GPU objects. + +/// One view's image and columns, and the inputs it was last drawn from. +const View = struct { + image: sk.Image = .none, + size: [2]u32 = .{ 0, 0 }, + columns: sk.Buffer = .none, + drawn: u64 = 0, + + fn redraw(v: *View, key: u64) bool { + if (key == v.drawn) return false; + v.drawn = key; + return true; + } +}; + +const Gpu = struct { + ready: bool = false, + tape: sk.Buffer = .none, + scratch: sk.Buffer = .none, + levels: sk.Buffer = .none, + levels_rows: u32 = 0, + wave: View = .{}, + spectrogram: View = .{}, + spectrum: View = .{}, + oscilloscope: View = .{}, + spectrum_started: bool = false, + + fn setup(g: *Gpu, f: *sk.Frame) void { + if (g.ready) return; + g.ready = true; + g.tape = f.host.newBuffer(history * @sizeOf(f32)); + g.scratch = f.host.newBuffer(layouts.fft_size * 2 * @sizeOf([2]f32)); + g.wave.columns = f.host.newBuffer(layouts.max_columns * @sizeOf([4]f32)); + g.oscilloscope.columns = f.host.newBuffer(layouts.max_columns * @sizeOf([4]f32)); + g.spectrum.columns = f.host.newBuffer(layouts.max_columns * @sizeOf([2]f32)); + } + + /// The spectrogram's levels, `rows` of `max_columns`. + fn levelsFor(g: *Gpu, f: *sk.Frame, rows: u32) void { + if (rows <= g.levels_rows) return; + if (g.levels != .none) f.host.freeBuffer(g.levels); + g.levels = f.host.newBuffer(layouts.max_columns * rows * @sizeOf(f32)); + g.levels_rows = rows; + g.forget(); + } + + /// New shaders, or new buffers: draw everything again. + fn forget(g: *Gpu) void { + g.wave.drawn = 0; + g.spectrogram.drawn = 0; + g.spectrum_started = false; + } +}; + +/// The view's image at `r`'s size, (re)made when that changes. +fn target(f: *sk.Frame, v: *View, r: sk.Rect, max_h: u32) [2]u32 { + const want: [2]u32 = .{ + @intFromFloat(std.math.clamp(r.w, 1, layouts.max_columns)), + @intFromFloat(std.math.clamp(r.h, 1, @as(f32, @floatFromInt(max_h)))), + }; + if (v.image == .none or v.size[0] != want[0] or v.size[1] != want[1]) { + if (v.image != .none) f.host.freeImage(v.image); + v.image = f.host.newImage(want[0], want[1]); + v.size = want; + v.drawn = 0; + } + return want; +} + +fn wavePass(f: *sk.Frame, v: *View, size: [2]u32, view: layouts.WaveView) void { + f.gpu(.{ + .label = "scope waveform.frag", + .image = v.image, + .width = size[0], + .height = size[1], + .clear = null, + .pipelines = &wave_pipelines, + .draws = &.{.{ .pipeline = 0, .vertex_count = 3, .fragment_uniforms = gpu.uniform(f.scratch, view), .fragment_storage = &.{v.columns} }}, + }); +} + +fn viewPipeline(fragment: []const u8) gpu.Pipeline { + return .{ + .vertex = .{ .code = shaders.view_vert }, + .fragment = .{ .code = fragment, .uniform_buffers = 1, .storage_buffers = 1 }, + .stride = 0, + .attributes = &.{}, + .blend = .none, + }; +} +const wave_pipelines = [_]gpu.Pipeline{viewPipeline(shaders.waveform_frag)}; +const spectrum_pipelines = [_]gpu.Pipeline{viewPipeline(shaders.spectrum_frag)}; +const spectrogram_pipelines = [_]gpu.Pipeline{viewPipeline(shaders.spectrogram_frag)}; + +const spectrum_kernel: gpu.Kernel = .{ .code = shaders.spectrum_comp, .readonly_buffers = 1, .readwrite_buffers = 2, .uniform_buffers = 1, .threads = .{ 256, 1, 1 } }; +const spectrogram_kernel: gpu.Kernel = .{ .code = shaders.spectrogram_comp, .readonly_buffers = 1, .readwrite_buffers = 1, .uniform_buffers = 1, .threads = .{ 256, 1, 1 } }; +const waveform_kernel: gpu.Kernel = .{ .code = shaders.waveform_comp, .readonly_buffers = 1, .readwrite_buffers = 1, .uniform_buffers = 1, .threads = .{ 64, 1, 1 } }; + +test "the history keeps the newest samples in order" { + var s: Scope = .{}; + s.push(&.{ 1, 2, 3 }); + s.push(&.{4}); + try std.testing.expectEqualSlices(f32, &.{ 1, 2, 3, 4 }, s.samples[history - 4 ..]); + try std.testing.expectEqual(@as(u64, 4), s.heard); +} + +test "frequencies place on a log axis" { + try std.testing.expectEqual(@as(f32, 0), place(min_freq, 48_000)); + try std.testing.expectApproxEqAbs(@as(f32, 1), place(20_000, 48_000), 1e-6); +} diff --git a/src/scope/shaders/spectrogram.comp b/src/scope/shaders/spectrogram.comp new file mode 100644 index 0000000..b32f4c5 --- /dev/null +++ b/src/scope/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/src/scope/shaders/spectrogram.frag b/src/scope/shaders/spectrogram.frag new file mode 100644 index 0000000..9cb5e3d --- /dev/null +++ b/src/scope/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/src/scope/shaders/spectrum.comp b/src/scope/shaders/spectrum.comp new file mode 100644 index 0000000..e576e30 --- /dev/null +++ b/src/scope/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/src/scope/shaders/spectrum.frag b/src/scope/shaders/spectrum.frag new file mode 100644 index 0000000..e88b091 --- /dev/null +++ b/src/scope/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/src/scope/shaders/view.vert b/src/scope/shaders/view.vert new file mode 100644 index 0000000..d9461e9 --- /dev/null +++ b/src/scope/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/src/scope/shaders/waveform.comp b/src/scope/shaders/waveform.comp new file mode 100644 index 0000000..d1d64e5 --- /dev/null +++ b/src/scope/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/src/scope/shaders/waveform.frag b/src/scope/shaders/waveform.frag new file mode 100644 index 0000000..37be7ce --- /dev/null +++ b/src/scope/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); +} |
