Building an open source RDNA3 compiler backend for tinygrad. Performs instruction selection, register allocation, control flow lowering, instruction field encoding, and architecture specific analysis passes.

Currently achieves ~2x performance ratio to LLVM in less then 700loc on MNIST/llama 1B

Cool features include: global pose estimation with kalman filter over multi-ocular vision estimates and drive train odometry; iterative shoot on the move turret tracking; autonomous motion planning with custom control systems and state machine management; complete offline hardware signal logging and replayability; 2d physics simulation and mechanism visualization