The SDL3 browser that reads BML v2.1 bytecode directly: it walks the stream, lays out the page and plays video, with no HTML parsing.
Stunt Grok Jockey Tony (@tswain555 on X)
BlitzBrowse (BlitzBrowse7) is the BML client. It opens .bml files or bml:// URLs, renders them in an SDL3 window, and decodes video in its own threads.
Builds: Linux (fully static), Linux GUI (static except glibc, so SDL can open an X11 window and use PulseAudio), and Windows 64-bit (UCRT64).
BlitzBrowse is a local test build. The executables are not on this web server and there is no download here. The pages below are HTML renders of the same .bml sources BlitzBrowse opens.
BlitzBrowse covers 117 of 128 tags; the remaining 11 are deprecated and deliberately skipped.
The v5 update added 8 tags, which took coverage from the earlier count to 117: data hgroup math menu search selectedcontent slot svg.
The 11 skipped are applet and the deprecated tags:
acronym applet big center dir font frame frameset marquee strike tt
The v6 fixes below add rendering for slot / shadow templates, math and SVG. The tag test suite passes 111/111.
Windows (UCRT64 static exe), from the demo folder:
cd /d I:\dev\channels7\demos\blitzbrowse7
run_demo.bat
run_demo.bat pages\newtags.bml
Or from I:\dev\channels7 with the run script: ./run blitzbrowse [page.bml], e.g. ./run blitzbrowse pages/newtags.bml.
With no argument it opens index.bml. Linux:
./run_demo.sh pages/newtags.bml
SDL_VIDEO_DRIVER=dummy ./run_demo.sh page.bml --shot out.png # headless
Online renders of the same pages: newtags.html (newtags.bml), one video, four videos.
Measured headless on the build box (dummy video, software renderer, dummy audio, 20 s, three runs each, averaged). The shared box is noisy by about ±0.3%.
| Page | Before (phase 3) | Now | Target |
|---|---|---|---|
| idle | 1.1% | 0.0% (no draw, no present) | - |
| 1 video | 4.9% | 1.9% | <2% |
| 4 videos | 14.7% | 6.4% | <6% |
No frames were dropped. One video meets its target. Four videos are just over: about 3.5% is h264 decode and scaling across four threads, about 2.5% is main-thread blits. The gains came from damage-rect redraw (only changed video boxes repaint, one present per tick), vsync only on GPU renderers, no busy loops, and uploading textures only when a frame changes.
<template shadowrootmode> renders in flow; slot fallback and slotted content are shown. This is an approximation: there is no real slot assignment, so fallback and light DOM both show. A plain template stays inert.Tests (ASan + UBSan): tags 111/111, elements 144/0, input 143/0, tables 90/0, layout 1419/0, new tags 24/0. Four-video CPU did not change in this phase (still 6.4%).