We are living through a fundamental shift in computing, where performance and efficiency gains increasingly come from specialized accelerators tailored to "hot" domains like AI. Yet as accelerator-centric computing proliferates, it continues to build atop a longstanding disconnect between the way we design these systems and the way we program them. This divide slows hardware innovation, complicates the software stack, and makes accelerators far harder to evolve than the rapidly changing applications they are meant to serve.

In this talk, I will present our recent research on closing this gap through a unified hardware-software co-design stack. At the core is a new abstraction that generalizes single program multiple data beyond largely independent threads to encompass spatial structure and communication. It offers a single, composable model for expressing both how accelerators are constructed and how they are programmed. I will show how this abstraction is realized in Allo, an open-source, Python-based, MLIR-powered framework to enable concise specifications that map efficiently across FPGA, ASIC, NPU, and GPU platforms. I will conclude with a forward-looking perspective on automated compiler construction, differentiable hardware synthesis, and agentic design automation, and discuss how these directions may help move us toward truly evolvable heterogeneous computing.