Automated Operator Fusion for GPU-Accelerated SQL: Generation and Evolutionary Refinement

GPU-accelerated SQL engines can achieve significant throughput and cost improvements over CPU-based ones, by lowering operations, e.g., relational operators, to GPU kernels, which can, in turn, be massively parallelized. However, for each GPU kernel invocation, a launch cost is incurred, and intermediate data must perform a round-trip through global memory. Kernel fusion reclaims these costs, but two questions remain central for data-processing workloads: at what granularity should fusion be applied, and how can fused kernels be produced automatically rather than hand-written? This materializes in a design space spanning the specificity of the fused kernels, the automation/runtime costs for their generation, and their performance. We study this design space in the context of the cuDF backend of Velox, targeting the middle ground between hand-written kernels for individual operators and bespoke implementations synthesized for an entire workload or execution engine. Our approach first lowers query plans into structural fusion dispatch rules and fused-operator scaffolds, then uses an agentic workflow to evaluate four code-generation models spanning frontier, efficient closed-model, and open-weight settings for producing functionally correct fused Triton kernels — seeds. These seeds are subsequently optimized with evolutionary search through SkyDiscover. Across TPC-H-derived pipeline windows, we find that strong code models can often generate buildable and equivalent fused Triton kernels, and that these kernels can improve end-to-end runtime when deployed opportunistically with fallback. We also show, on selected representative seeds, that evolutionary refinement can further improve performance. These results suggest that pipeline-granularity fusion is a promising intermediate point between per-operator GPU lowering and full workload-specific synthesis.