How to use the GPU backend

How to use the GPU backend

The GPU backend runs the rasterizer as a kernel with one thread per pixel. It needs MAX and a supported accelerator. Nothing else in the project needs either.

Install MAX

uv pip install "max==26.6.0"

On macOS, also install Apple's Metal toolchain. Xcode does not install it by default:

xcodebuild -downloadComponent MetalToolchain

Without it every kernel fails with Metal Compiler failed to compile metallib. That message looks like a code error and is not.

The same message also comes when Apple's Metal linker, air-lld, crashes. Mojo shows no other text. To find the cause, look for a new air-lld-*.ips report in ~/Library/Logs/DiagnosticReports/. Then compile one kernel at a time with DeviceContext.compile_function to find the kernel that crashes it. Two known causes are a whole struct loaded through a cast device pointer, and a branch in level_filter. The code now avoids both.

Check that a GPU is visible

make gpu-status

The status command runs tools/gpu_status.mojo. It checks physical devices through MAX. A compiler target alone does not establish device availability.

Run the GPU checks

make check-gpu

Without a physical device, the command reports the device tests as skipped and runs the host layout tests. With a device, it also compiles and runs the GPU parity suite. A detection error fails the check.

The suite runs under a time budget of 300 seconds. A hang fails the run instead of looking slow. Raise the budget on a machine with a slow first kernel compile:

make test-gpu GPU_BUDGET=600

Render on the GPU

Prepare the triangles with the CPU renderer, then draw them on the device:

from render.gpu import render_triangles

var corners = renderer.prepare(scene, assets, camera)
var lighting = Lighting(scene)
var image = render_triangles(
    corners, width, height, background, SHADE_TEXTURE, assets.textures, lighting
)

For many frames, hold a GpuRenderer so the device buffers survive between frames. See GPU backend.

Compare the backends

make bench

This times the CPU and GPU rasterizers across image sizes. The GPU numbers include allocation and the copy back.

Windows

The backend runs under CUDA on WSL 2. Read The CUDA teardown hang if the second renderer in a process hangs.