diff options
Diffstat (limited to 'tools')
| -rw-r--r-- | tools/bench.zig | 96 | ||||
| -rw-r--r-- | tools/svgdiff.py | 116 |
2 files changed, 212 insertions, 0 deletions
diff --git a/tools/bench.zig b/tools/bench.zig new file mode 100644 index 0000000..faa769e --- /dev/null +++ b/tools/bench.zig @@ -0,0 +1,96 @@ +//! How long a frame of vector graphics takes, per stage, in release mode: +//! parsing the SVG, turning it into polygons and coverage (CPU backend), or +//! into a GPU pass description (GPU backend; the GPU's own time is not +//! measured here). +const std = @import("std"); +const vector = @import("vector"); +const scene = @import("scene").scene; +const typst = @import("typst"); + +const width = 1280; +const height = 720; +const runs = 300; +/// Drawings up to this many shapes count as simple: they must take < 1ms. +const simple_shapes = 30; + +/// Stand-in for `sketch.Host`: the renderers only need images (and to free +/// the GPU backend's paint buffer, which this bench never makes). +const Host = struct { + pub fn newImage(_: Host, _: u32, _: u32) @import("abi").Image { + return @fromBackingInt(@intCast(1)); + } + pub fn freeImage(_: Host, _: @import("abi").Image) void {} + pub fn freeBuffer(_: Host, _: @import("abi").Buffer) void {} +}; + +pub fn main(init: std.process.Init) !void { + const gpa = std.heap.smp_allocator; + const io = init.io; + var failed = false; + std.debug.print("{s:<24} {s:>7} {s:>10} {s:>10} {s:>10} {s:>10}\n", .{ "object", "shapes", "parse", "cpu draw", "gpu record", "cpu total" }); + // Every Typst object of the scene being played (a live scene adds them + // all whatever the time). + var objects_arena: std.heap.ArenaAllocator = .init(gpa); + defer objects_arena.deinit(); + const shot = try scene.shot(objects_arena.allocator(), .{ .t = 0 }); + for (shot.objects, 0..) |object, index| if (object.source == .typst) { + const svg = typst.find(object.source.typst).?; + var arena: std.heap.ArenaAllocator = .init(gpa); + defer arena.deinit(); + var cpu: vector.Renderer(.cpu) = try .init(gpa, Host{}, width, height); + defer cpu.deinit(Host{}); + var gpu: vector.Renderer(.gpu) = try .init(gpa, Host{}, width, height); + defer gpu.deinit(Host{}); + const Frame = struct { scratch: std.mem.Allocator }; + + var parse_ns: [runs]u64 = undefined; + var cpu_ns: [runs]u64 = undefined; + var gpu_ns: [runs]u64 = undefined; + var shapes: usize = 0; + for (0..runs) |i| { + _ = arena.reset(.retain_capacity); + const t0 = now(io); + const d = try vector.svg.parse(arena.allocator(), svg); + const t1 = now(io); + // Fit to the canvas, as large as it goes: the worst case. + const m = vector.fit(d, .{ .w = width, .h = height }); + cpu.begin(Frame{ .scratch = arena.allocator() }, .{ 0, 0, 0, 1 }); + try cpu.draw(d, m, 1); + const t2 = now(io); + gpu.begin(Frame{ .scratch = arena.allocator() }, .{ 0, 0, 0, 1 }); + try gpu.draw(d, m, 1); + std.mem.doNotOptimizeAway(gpu.target.pass()); + const t3 = now(io); + parse_ns[i] = t1 - t0; + cpu_ns[i] = t2 - t1; + gpu_ns[i] = t3 - t2; + shapes = d.shapes.len; + } + const p = median(&parse_ns); + const c = median(&cpu_ns); + const g = median(&gpu_ns); + std.debug.print("{d:>2} {s:<21} {d:>7} {d:>8.3}ms {d:>8.3}ms {d:>8.3}ms {d:>8.3}ms\n", .{ index, label(object.source.typst), shapes, ms(p), ms(c), ms(g), ms(p + c) }); + if (shapes <= simple_shapes and (p + c > std.time.ns_per_ms or p + g > std.time.ns_per_ms)) failed = true; + }; + std.debug.print("\nup to {d} shapes must parse and draw in under 1ms at {d}x{d}, filling the canvas.\n", .{ simple_shapes, width, height }); + if (failed) return error.TooSlow; +} + +/// The start of a Typst source's first line. +fn label(source: []const u8) []const u8 { + const line = source[0 .. std.mem.indexOfScalar(u8, source, '\n') orelse source.len]; + return line[0..@min(line.len, 21)]; +} + +fn now(io: std.Io) u64 { + return @intCast(std.Io.Timestamp.now(io, .awake).nanoseconds); +} + +fn median(samples: []u64) u64 { + std.mem.sort(u64, samples, {}, std.sort.asc(u64)); + return samples[samples.len / 2]; +} + +fn ms(ns: u64) f64 { + return @as(f64, @floatFromInt(ns)) / std.time.ns_per_ms; +} diff --git a/tools/svgdiff.py b/tools/svgdiff.py new file mode 100644 index 0000000..80e15a9 --- /dev/null +++ b/tools/svgdiff.py @@ -0,0 +1,116 @@ +#!/usr/bin/env python3 +"""Differential test of the vector renderers against resvg. + +Every SVG (and every Typst document, compiled to SVG first) in the corpus is +rendered by resvg as the reference, then by sketchbook's CPU and GPU +backends through the real host, headless. Each result is compared with the +reference, and GPU with CPU; side-by-side images of every case land in the +output directory. + +usage: svgdiff.py SVGVIEW_EXE CORPUS_DIR OUT_DIR +""" +import re +import subprocess +import sys +from pathlib import Path + +import numpy as np +from PIL import Image + +WIDTH = 480 +BACKGROUND = "#808080" +# resvg samples coverage on four sub-scanlines, so thin edges come out in +# quarter steps. Rendering it 4x larger and box-filtering down gives the +# reference exact-enough coverage to hold both renderers to. +SUPERSAMPLE = 4 +# Per case: mean absolute channel error, and the share of pixels whose worst +# channel is off by more than `EDGE` (anti-aliasing differences sit below it). +MEAN_LIMIT = 1.0 +EDGE = 48 +BAD_LIMIT = 0.005 + + +def load(path): + return np.asarray(Image.open(path).convert("RGB"), dtype=np.int16) + + +def compare(a, b): + if a.shape != b.shape: + return float("inf"), 1.0 + diff = np.abs(a - b) + return float(diff.mean()), float((diff.max(axis=2) > EDGE).mean()) + + +def strip(img, ref, ours): + """Reference | ours | difference, amplified.""" + diff = np.clip(np.abs(ref - ours) * 4, 0, 255) + return Image.fromarray(np.concatenate([ref, ours, diff], axis=1).astype(np.uint8)).save(img) + + +def main(): + exe, corpus, out = Path(sys.argv[1]).resolve(), Path(sys.argv[2]), Path(sys.argv[3]) + out.mkdir(parents=True, exist_ok=True) + cases = [] + for source in sorted(corpus.iterdir()): + if source.suffix == ".typ": + svg = out / (source.stem + ".svg") + subprocess.run(["typst", "compile", "--format", "svg", source, svg], check=True) + # Sizes in pt make resvg round to whole pixels before scaling, + # which shifts its picture by a fraction; give it plain units. + text = svg.read_text() + box = re.search(r'viewBox="0 0 ([\d.]+) ([\d.]+)"', text) + text = re.sub(r'width="[\d.]+pt" height="[\d.]+pt"', f'width="{box[1]}" height="{box[2]}"', text, count=1) + svg.write_text(text) + elif source.suffix == ".svg": + svg = source + else: + continue + ref = out / f"{source.stem}-resvg.png" + big = out / f"{source.stem}-resvg-x{SUPERSAMPLE}.png" + subprocess.run(["resvg", "-w", str(WIDTH * SUPERSAMPLE), "--background", BACKGROUND, svg, big], check=True) + hi = Image.open(big).convert("RGB") + # Whole blocks only: a height that is no multiple is padded, not + # squeezed, which would shift every edge below by a fraction. + w, h = WIDTH, -(-hi.size[1] // SUPERSAMPLE) + whole = Image.new("RGB", (w * SUPERSAMPLE, h * SUPERSAMPLE), BACKGROUND) + whole.paste(hi, (0, 0)) + whole.resize((w, h), Image.BOX).save(ref) + cases.append((source.stem, svg.resolve(), ref, w, h)) + + listing = out / "list.txt" + listing.write_text("".join(f"{svg} {w} {h}\n" for _, svg, _, w, h in cases)) + script = out / "render.script" + lines = [] + for i, (stem, *_rest) in enumerate(cases): + for backend in ("gpu", "cpu"): + lines += [f"console set index {i}", f"console set backend {backend}", "frames 2", + f"screenshot {out.resolve()}/{stem}-{backend}.png"] + script.write_text("\n".join(lines) + "\n") + subprocess.run([exe, "--headless", "--fresh", "--state", out / "svgview.state", "--script", script], + check=True, env={"SVGVIEW_LIST": str(listing.resolve()), **__import__("os").environ}) + + failed = [] + print(f"{'case':<18} {'cpu mean':>9} {'cpu bad':>8} {'gpu mean':>9} {'gpu bad':>8} {'gpu~cpu':>8}") + for stem, _svg, ref_path, _w, _h in cases: + ref = load(ref_path) + row = [stem] + results = {} + for backend in ("cpu", "gpu"): + ours = load(out / f"{stem}-{backend}.png") + mean, bad = compare(ref, ours) + results[backend] = (mean, bad) + strip(out / f"{stem}-{backend}-diff.png", ref, ours) + row += [f"{mean:9.3f}", f"{bad * 100:7.2f}%"] + if mean > MEAN_LIMIT or bad > BAD_LIMIT: + failed.append(f"{stem} ({backend})") + mean, _ = compare(load(out / f"{stem}-cpu.png"), load(out / f"{stem}-gpu.png")) + row.append(f"{mean:8.3f}") + print(f"{row[0]:<18} " + " ".join(row[1:])) + print(f"\nlimits: mean < {MEAN_LIMIT}, pixels off by > {EDGE}: < {BAD_LIMIT * 100:.0f}%. Images in {out}") + if failed: + print("FAILED: " + ", ".join(failed)) + sys.exit(1) + + +if __name__ == "__main__": + main() |
