aleleba/lightcraft
1
0
Fork 0
mirror of https://github.com/storytold/lightcraft.git synced 2026-10-24 10:09:01 -06:00
Code Issues Packages Projects Releases Wiki Activity

Merge PR #1663

Browse Source 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Brandon Thomas
2026-10-15 03:24:51 -04:00
co-authored by Claude Opus 5.5
parent 211255770f f625c74c7a
commit 2179473536
30 changed files with 2717 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/faces.md
+37
View File
@@ -143,6 +143,43 @@ SFace's builtin speed metadata now uses the measured 6.9× ratio. User `catalog.
Older published tables for MobileFaceNet/R50/R100 came from a different runtime and are not measurements of this runner; architecture
names alone do not guarantee the same speed, and those additional model files were not verified in this task.
### On the GPU
On a machine with a usable GPU, the whole detector and recogniser run on it through wgpu (`crates/gpu/src/faces.rs`).
The interpreter lowers the graph to a flat program for the input's shape (`crates/faces/src/net/program.rs`): each
convolution (`wgsl/faces_conv.wgsl`: an implicit matrix product that gathers the input straight into workgroup memory,
cuts K into slices when a map is small, and a direct kernel for depthwise layers) and every other operator (activations,
batch norm, PRelu, broadcasting add/sub/mul, max pooling, nearest upscaling, transposes, global pooling and the fully
connected layer, `wgsl/faces_ops.wgsl`) is a kernel, every value stays on the device between layers in reused
buffers, and a run is one upload, one submission and one read-back. Shape-only operators make no step. A model that
needs something the program cannot express runs its convolutions one by one on the GPU instead. Several faces of one
photo go together (`Embedder::embed_batch`, up to 16 per submission).
The CPU interpreter remains the reference. After a model loads, a background check (a row in the activity stack, "Checking
the GPU for faces") runs it once on the CPU and twice through the GPU on a fixed test input: the GPU is used only when
every output is within 1e-3 of the CPU's (relative to the output's largest value) **and** it was faster; until then,
and whenever the GPU is switched off (`LIGHTCRAFT_GPU=0`, `LIGHTCRAFT_GPU_BACKEND=off`, the GPU preference, or **Settings ▸
Faces ▸ Run face models on the CPU only**), missing, wrong or slower, the model runs on the CPU. The first GPU error,
or a result of the wrong size or with values that are not finite, drops it for that model. Settings ▸ Faces shows
which device each model runs on (`faces.gpu`). The weights stay on a device of the scan's own (separate from the
editor's) while the model is loaded, at most 768 MB per model.
The background scan's pace also limits the GPU: paused or light, one model run at a time and the GPU busy at most a
quarter of the time (the editor renders on the same GPU); normal, two at once and three quarters; full, as fast as it
goes.
Measured on Linux with a Radeon (Vulkan) and a 32-thread CPU, one call at a time, machine under very heavy load from
other jobs (load average 100–250, so the absolute numbers are noisy):
| Model | CPU | GPU, one layer at a time | GPU, whole network | Batch of 16, per face |
| --- | ---: | ---: | ---: | ---: |
| YuNet (640², one photo) | 58–60 ms | 85–89 ms | 21 ms | – |
| SFace 2021dec | 40–42 ms | 47–51 ms | 9.6 ms | 8.1 ms |
| AuraFace v1 | 295–308 ms | 165–172 ms | 70–73 ms | 71 ms |
`LC_FACE_MODELS=<folder> cargo test --release -p lightcraft-gpu faces::tests -- --include-ignored --nocapture` checks
agreement on the real models and every operator, and prints these timings.
### Agreement and regression coverage
The public-crop embeddings match the reference runtime's outputs: