GPU — the device lane
A function marked @ kernel is a compute kernel. The device lane links the compiler to real Metal and CUDA execution.
The shader text below is the lowering surface. Real device execution — faber run --backend metal|cuda — is the narrower product proof, recorded in the device kernel support summary.
Targets#
A target showing fewer scenarios than the others is not broken. It means the emitter declines that shape, which the pages state directly rather than hiding.
Measured support#
Device-kernel emitters are not scored against the general corpus — they lower a kernel surface and nothing else. Their measured support is the device kernel support summary.
---