A signed saturating add through NAK →
From CUDA's five-instruction sequence to PLOP3 .SIGN, mixed-source encodings, and CTS.
Debugging notes, experiments, and patch write-ups.
From CUDA's five-instruction sequence to PLOP3 .SIGN, mixed-source encodings, and CTS.
Removing impossible GRF state from a compiler dataflow pass, then measuring the tradeoff.
An Arm assembler bug caused by reading ISA state at a late .type directive.
The initial Orange Pi 5 Pro port, from U-Boot handoff to NSH.
Running SPIR-V through RADV and ACO using drm-shim, without physical AMD hardware.
Small builds, verified driver loading, source search, sanitizers, and bisecting.
The boundaries between a Vulkan application, Mesa, DRM, and GPU machine code.