Skip to content

Windows: export xml's paths open in Resolve (#195); DX12 with DXC renders 3D again (#196) - #202

Merged
shinyobjectz merged 4 commits into
mainfrom
win-export-xml
Nov 19, 2026
Merged

shinyobjectz merged 4 commits into
mainfrom
win-export-xml

Conversation

@shinyobjectz

Copy link
Copy Markdown
Owner

Fixes #195. Fixes #196.

#195. raea export xml percent-encoded a Windows path whole into <pathurl> (file://localhostD%3A%5CWork%5C...), which Resolve refused. A drive's path is now file:///D:/Work/..., the form the reporter confirmed Resolve imports; a share's is file://nas/clips/...; Windows' \\?\ prefix comes off; Mac and Linux paths are unchanged. There's a test for each kind of path.

#196. wgpu passes a shader module's label to DXC as its source's file name. On Windows, LLVM reads hanabi:shader:utils as a root and a file, so DXC failed with "error reading 'hanabi:shader:utils'" and every Bevy app failed with it. That happens on any PC where dxcompiler.dll is on the PATH (wgpu loads DXC when it can, else FXC) and raea runs on DX12 (an Intel or AMD GPU, or WGPU_BACKEND=dx12). It reproduced on the RTX 4070 box with rc8, DXC 1.9.2609 on the PATH and DX12: a one-box 3D render exited 101, and the same render passed under FXC. Bevy gives its modules no label; hanabi's utils module and its contract tests' were the only labels with a colon, and they now use underscores. Also, an export refuses an alpha codec this machine can't write when it starts, not after the reference render (Codec::chosen).

Checked: cargo test -p raea-app-layers -p raea-app-capcut -p raea-io-alpha, and every CI gate on main with this branch merged. v0.1.8-rc10 is that merge, so it can be checked on Windows under DX12 with both DXC and FXC.

🤖 Generated with Claude Code

0.1.8-rc8 percent-encoded a Windows path whole into <pathurl>: file://localhostD%3A%5CWork%5C... for D:\Work\...,
which DaVinci Resolve refused. The same XML imported at once with file:///D:/Work/... (the reporter's test). A drive's
path is now that, a share's file://nas/clips/..., separators as slashes and the \\?\ prefix off; a Mac's and
Linux's are as before. cargo test -p raea-app-layers passes, with a test of each kind of path.
…ort refuses a codec before rendering (#196)

wgpu hands DXC a shader module's label as its source's file name, and on Windows LLVM reads "hanabi:shader:utils"
as a root and a file, so DXC said "error reading 'hanabi:shader:utils'" and every Bevy app failed: on a PC where
dxcompiler.dll is on the PATH (wgpu loads it when it can, else FXC), rc8 rendered no 3D at all on DX12 (exit 101 on
an RTX 4070 with WGPU_BACKEND=dx12 and DXC 1.9.2609; the same render passed with FXC). raea export xml's passes were
where the reporter saw it, on a PC whose Intel iGPU puts raea on DX12. Bevy names its modules nothing; hanabi's
utils module and its contract tests' were the only labels with a colon, now underscores.

And an export refuses an alpha codec this machine can't write (ProRes or HEVC off a Mac) when it starts, not after the
reference render: Codec::chosen, which layers, capcut and render_movie share.
… pack failed on it)

GitHub's Windows runner checks text files out with CRLF, as the Models step already allows for (tr -d '\r' <
models.tsv), so release.sh looked for "vcruntime140.dll\r" in target/dist, beside the vcruntime140.dll win-bundle.sh
had just put there, and stopped. release.sh and release-check.sh, the list's two readers, now strip it.
…, where it reads it

windows-verify checks the zip on a runner with only release-check.sh and lavapipe.sh checked out, and runs
release-check.sh from RUNNER_TEMP; since the Windows bundle it reads release/win-bundle.files beside itself, which
wasn't there, so the step would have stopped on it once rc9's pack step stopped failing. It's checked out and copied
there now.
@shinyobjectz
shinyobjectz merged commit 69fb3d8 into main Nov 19, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

1 participant