kernel-designer

To design an algorithm sketch for kernel generation, you provide the operator name, task description, backend, and selected DSL. The system then generates a structured UnifiedSketch with symbols, tensors, data flows, and optimization hints.

Supported DSLs for generating operator algorithm sketches include triton_cuda, triton_ascend, cpp, cuda_c, tilelang_cuda, and pypto. You select one based on your target backend hardware constraints.

Yes, you can use hint mode to configure parameter spaces for kernel optimization. This mode delivers explicit optimization hints via the @llm_hint mechanism while respecting hardware and DSL constraints.

The best way to optimize an existing kernel algorithm sketch is using evolutionary optimization mode. This mode adapts to your current sketch inputs and iteratively refines data flows like alloc, load, compute, and store.

Yes, you need to specify the backend and task description to generate a kernel sketch. These inputs ensure the resulting algorithm sketch respects hardware constraints and applies the correct DSL syntax.

Your kernel algorithm sketch needs explicit optimization hints to guide downstream code generation effectively. Providing hints via the @llm_hint mechanism ensures the generated code respects specific hardware and DSL constraints.