Port to Slug algorithm by behdad · Pull Request #62 · behdad/glyphy

behdad · 2026-03-21T04:05:34Z

See: https://terathon.com/blog/decade-slug.html

On my setup, runs >2.5x faster than GLyphy algorithm and seems to consume slightly more shader memory. All rendering artifacts are gone!

This was done in two hours with Claude, from SLUG PDFs & shader code, to working demo. Insane.

Strip all arc, SDF, blob encoding, and shader API from glyphy.h. Add glyphy_curve_t (quadratic Bezier with p1/p2/p3) and glyphy_curve_accumulator_t that directly stores quadratic curves from glyph outlines without arc approximation. This is the first step toward Slug-style winding number rendering. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Encodes quadratic Bezier curves into a RGBA16UI blob with: - Horizontal and vertical bands for spatial indexing - Curves sorted by descending max coordinate for early exit - Em-space quantized int16 coordinates (configurable scale) - Horizontal/vertical line filtering (can't intersect parallel rays) Blob layout: [band headers] [curve index lists] [curve data] All offsets are 1D from blob start; shader converts to 2D via atlas width. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Port SlugPixelShader.hlsl to GLSL 1.30 for our single isampler2D atlas: - CalcRootCode: equivalence class via sign-bit extraction (floatBitsToUint) - SolveHorizPoly/SolveVertPoly: quadratic root finding with discriminant clamping - CalcCoverage: weighted combination of horizontal and vertical ray coverage - glyphy_slug_render: main entry point, iterates H-bands and V-bands Adapted for single-texture blob layout with 1D offsets via CalcBandLoc. Curve data decoded from int16 quantized em-space coordinates. Also: cap bands at 16, use signed int16 texels, update meson.build to remove old arc/sdf/blob sources and add new files. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

New vertex shader passes em-space texcoords and flat per-glyph data (band transform, atlas location, band counts) to fragment shader. Fragment shader calls glyphy_slug_render() for coverage. Updated glyph_info_t: replaced nominal_w/h with num_hbands/num_vbands. Updated glyph_vertex_t: new layout with position, texcoord, band transform, and glyph data fields for Slug rendering. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Atlas: RGBA16I texture with isampler2D, texelFetch upload - Font: use curve accumulator + blob encoder instead of arc pipeline - Shader: assemble slug GLSL + demo fragment shader, new vertex layout with position, texcoord, bandTransform (flat), glyphData (flat) - Buffer: set up 4 vertex attributes including glVertexAttribIPointer for integer glyph data - GLState: simplified, removed old SDF uniforms - HarfBuzz/FreeType headers: updated for curve accumulator, use hb_font_draw_glyph, stub cubic_to for now - Removed demo-atlas.glsl (no longer needed) Builds and links successfully. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Use 16384 texels (vs 4096) to handle complex glyphs with many curves. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Bump to #version 330 (needed for floatBitsToUint) - Remove #version from demo-fshader.glsl (prepended by demo-shader.cc) - Remove demo-atlas.glsl from Makefile.am (no longer used) - Update src/Makefile.am for new source files and shader - Slug shader by Eric Lengyel, ported to GLSL by Behdad Esfahbod Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Delete all arc geometry, bezier-to-arc approximation, SDF calculation, grid-based blob encoding, outline manipulation, and glyphy-validate. Fix glyphy-extents.cc to not depend on deleted glyphy-common.hh. Update both autotools and meson build files. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Add close_path callback to HarfBuzz draw funcs. Without it, the last contour's closing segment was never emitted, causing winding number errors (horizontal streak artifacts). - Use floor() instead of (int) cast for band assignment to avoid off-by-one at band boundaries due to floating-point truncation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Lines encoded as degenerate quadratics with p2=midpoint(p1,p3) had a=p1-2*p2+p3=0 mathematically, but float32 rounding of the intermediate p12.zw*2.0 in the shader made a nonzero (~1e-7). This triggered the quadratic path with 1/a producing Inf, corrupting winding numbers across entire scanlines. Fix: encode lines as (p1,p1,p3) instead. This gives a=p3-p1 (full line span, well-conditioned) and b=0. The standard quadratic solver handles this cleanly with no near-zero divisions and no shader changes needed. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

These features were part of the old SDF pipeline and are not implemented in the Slug renderer. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace the 2D RGBA16I texture atlas with a 1D buffer texture (GL_TEXTURE_BUFFER). This eliminates: - CalcBandLoc wrapping logic in the shader - Atlas width constant (GLYPHY_ATLAS_WIDTH, GLYPHY_LOG_ATLAS_WIDTH) - 2D coordinate math for atlas lookups - item_w / item_h_quantum parameters The shader now uses isamplerBuffer with plain integer offsets: texelFetch(u_atlas, glyphLoc + offset) instead of texelFetch(u_atlas, glyphy_calc_blob_loc(glyphLoc, offset), 0). glyphLoc is a single int (1D offset) instead of ivec2. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Store extents and band counts in a 2-texel header at the start of each glyph's blob. The shader reads the header and computes the band transform itself. Client no longer needs to pass bandTransform (vec4) or band counts. Per-glyph vertex data is now just: position, texcoord, and one flat int (atlas_offset). The encoder API drops num_hbands/num_vbands output parameters. ~1% performance cost for much simpler client integration. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Per-glyph data is now just one uint (atlas offset) instead of an ivec4. Vertex size drops from 48 to 20 bytes. The shader API takes uint for the glyph location. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Meson is the only build system now. Remove all autotools files (configure.ac, Makefile.am, autogen.sh, git.mk) and stale generated headers that caused shader compilation issues when both build systems were used. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The encoder outputs extents in font design units. Previously the demo divided by upem then the shader multiplied back. Now everything stays in font units. Screen positions use font_size/upem as the scale factor. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Adjacent curves in a contour share p3/p1. Instead of storing 2 texels per curve (2N total), store N+1 texels per contour: (p1,p2) (p3/p1,p2_next) ... (p3,0) The shader reads curveLoc and curveLoc+1, unchanged. Band indices point to each curve's first texel. Nearly halves curve data storage. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Ubuntu 24.04 has harfbuzz >= 4.0.0, no need to build from source. Remove autotools steps, autotools deps, and harfbuzz cache. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

For each band, store curves sorted in both directions: descending max (for rightward/upward ray) and ascending min (for leftward/ downward ray). A split value per band determines which direction to use based on pixel position. Pixels on the left side of the glyph use leftward ray with ascending sort; pixels on the right use rightward ray with descending sort. Each side exits early after processing only the nearby curves. Coverage formula adjusted for leftward ray: saturate(0.5 - r) instead of saturate(r + 0.5). ~15% speedup (1725 -> 1980 fps). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Instead of using the glyph midpoint for all bands, compute the optimal split per band by finding the x/y value that minimizes max(left_count, right_count). This balances the work between leftward and rightward rays more evenly, especially for asymmetric glyphs. ~2% speedup (1980 -> 2020 fps). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Dilate glyph quads by exactly half a pixel on screen, regardless of zoom level. The vertex shader computes the dilation in clip space and corrects the texcoord using the inverse 2x2 MVP Jacobian. Per-vertex data adds normal direction and inverse Jacobian for the object-to-em-space mapping. Viewport size passed as uniform. Based on SlugDilate from the Slug reference vertex shader, simplified for z=0 geometry. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Add glyphy-vertex.glsl with glyphy_dilate() that computes exact half-pixel dilation using the MVP matrix and viewport size. Client vertex shader just calls one function. Rename shader files and API: glyphy-slug.glsl -> glyphy-fragment.glsl glyphy_slug_shader_source -> glyphy_fragment_shader_source glyphy_slug_render -> glyphy_render New: glyphy_vertex_shader_source, glyphy_dilate Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Drop "(MIT license)" from Slug attribution -- the algorithm is referenced, not the code. Remove glyphy_*_shader_source_path() functions and PKGDATADIR/SHADER_PATH since shaders are embedded. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace GLUT/FreeGLUT windowing with GLFW3: - Explicit main loop instead of glutMainLoop - GLFW native vsync via glfwSwapInterval (removes platform-specific CGLSetParameter/wglSwapIntervalEXT/glXSwapIntervalSGI code) - Proper core profile context hints (no GLUT_3_2_CORE_PROFILE ifdef) - Proper scroll callback instead of GLUT_WHEEL_UP/DOWN hack - Split keyboard handling into key callback (special keys) and char callback (printable characters) - Dirty-flag based redraw with glfwWaitEvents when idle Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

glewInit() calls glGetString(GL_EXTENSIONS) which generates GL_INVALID_ENUM in a core profile context. Ignore the return value and clear the GL error; the GL_VERSION_3_3 check that follows is the real gate. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Set initial viewport from framebuffer size (not window size) - Use window size for mouse coordinate math since GLFW cursor positions are in screen coordinates, not framebuffer pixels - Poll events before first render Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Render an initial frame, poll events to let the Wayland compositor configure the surface at the correct content scale, then render a second frame at the right resolution. Also set viewport from framebuffer size every frame to avoid stale viewport state on HiDPI displays. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Track cursor position continuously from the cursor callback instead of querying glfwGetCursorPos in the mouse button handler. On Wayland, the two can disagree, causing a position jump on the first drag. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Integer millisecond timing caused jitter at high frame rates (vsync off) because frame deltas alternated between 0 and 1ms. Switch to double-precision seconds throughout the animation and FPS reporting code. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

behdad · 2026-03-21T18:51:33Z

As for encoding speed, the new benchmark shows that this encodes NotoSansCJK in 750ms on my machine. Not bad... For reference, this is 8x slower than HarfBuzz extracting the outlines. Shade memory is bad though: over 400MB for this font.

glyphs: 65535 x 1 repeats = 65535 processed (65524 non-empty)
total blob size: 435617984 bytes (425408.19kb)
avg curves per glyph: 83.72
avg blob size per glyph: 6.49kb
outline:   91.773ms total, 1.400us/glyph, 714100 glyphs/s
encode:   764.360ms total, 11.663us/glyph, 85738 glyphs/s, 543.51 MiB/s
wall:     858.766ms total (outline + encode + loop overhead)

Roboto-Regular:

glyphs: 1294 x 1 repeats = 1294 processed (1275 non-empty)
total blob size: 3191680 bytes (3116.88kb)
avg curves per glyph: 21.36
avg blob size per glyph: 2.41kb
outline:    0.694ms total, 0.536us/glyph, 1864704 glyphs/s
encode:     4.303ms total, 3.325us/glyph, 300755 glyphs/s, 707.45 MiB/s
wall:       5.051ms total (outline + encode + loop overhead)

behdad · 2026-03-21T20:14:56Z

This PR also modernizes the codebase, ports to GLFW, and many other improvements. I suggest we merge it after the license is sorted out.

behdad · 2026-03-21T21:01:42Z

Okay, I received written (email) permission from Eric to use the shaders under HarfBuzz's Old MIT license:

Hi --

You have permission to use the Slug reference shaders (https://github.com/EricLengyel/Slug) under the Old MIT license.

-- Eric Lengyel

I'll go ahead and cherry-pick and update the license change from master, then this can be merged.

wipfli · 2026-03-21T21:39:32Z

Great! Don't forget to update the repo description:

GLyphy is a signed-distance-field (SDF) text renderer using OpenGL ES2 shading language. 
``|

behdad · 2026-03-21T21:43:58Z

Thanks. Done.

behdad force-pushed the slug branch 2 times, most recently from 23b6098 to 69250ee Compare March 21, 2026 05:04

behdad and others added 13 commits March 20, 2026 23:22

Fix demo buffer size for complex glyphs

1a7695c

Use 16384 texels (vs 4096) to handle complex glyphs with many curves. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

[meson] Adjust glut finding

11e61e2

Increase quantization precision to 4 units per em unit

684c80e

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Remove dead demo code for boldness, contrast, gamma, outline, debug

859bf0e

These features were part of the old SDF pipeline and are not implemented in the Slug renderer. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

behdad force-pushed the slug branch from de71eca to 859bf0e Compare March 21, 2026 05:22

behdad and others added 14 commits March 20, 2026 23:46

Shrink per-vertex data: ivec4 -> single uint glyphLoc

68b1110

Per-glyph data is now just one uint (atlas offset) instead of an ivec4. Vertex size drops from 48 to 20 bytes. The shader API takes uint for the glyph location. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Update CI: meson only, Ubuntu 24.04, drop harfbuzz build

1dcbbc8

Ubuntu 24.04 has harfbuzz >= 4.0.0, no need to build from source. Remove autotools steps, autotools deps, and harfbuzz cache. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Cleanup: remove win32 build, unused MIN_FONT_SIZE, fix CI apt

8eb14a3

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Update README for Slug-based renderer, remove empty NEWS

9b136b3

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

behdad and others added 17 commits March 21, 2026 11:14

demo: bind a VAO for core profile

8a25f87

demo: detect sRGB framebuffer in core profile

a736fd7

demo: use GLFW sRGB capability hint

b5437ed

demo: probe sRGB by toggling framebuffer state

74b934a

Rename demo shaders to vertex and fragment

d3983b2

Abort on unsupported cubic outlines

1e13881

Report atlas usage in demo stats

29e6f4d

Add bench-encode font benchmark

306534d

Set default Meson buildtype to debugoptimized

97d35d6

Cache curve bounds during blob encoding

47239a8

Reduce band-list allocation churn in blob encoder

09ad9f2

behdad added 5 commits March 21, 2026 12:53

Make band split selection linear-time

726f39a

Fold extents computation into encoder preprocessing

18fceba

Cache band ranges during encoder preprocessing

0c7ada9

Flatten band index storage in blob encoder

1ed8329

Add reusable glyph encoder object

ea69633

behdad added 2 commits March 21, 2026 15:07

Merge branch 'master' into slug

102cd52

Add Eric to copyright holders for the Slug shaders

fb65234

behdad merged commit 926b96d into master Mar 21, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Port to Slug algorithm#62

Port to Slug algorithm#62
behdad merged 69 commits intomasterfrom
slug

behdad commented Mar 21, 2026 •

edited

Loading

Uh oh!

behdad commented Mar 21, 2026 •

edited

Loading

Uh oh!

behdad commented Mar 21, 2026

Uh oh!

behdad commented Mar 21, 2026 •

edited

Loading

Uh oh!

wipfli commented Mar 21, 2026

Uh oh!

behdad commented Mar 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

behdad commented Mar 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

behdad commented Mar 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

behdad commented Mar 21, 2026

Uh oh!

behdad commented Mar 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wipfli commented Mar 21, 2026

Uh oh!

behdad commented Mar 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

behdad commented Mar 21, 2026 •

edited

Loading

behdad commented Mar 21, 2026 •

edited

Loading

behdad commented Mar 21, 2026 •

edited

Loading