Artifact of the paper: Abstraction at What Cost? Evaluating Python DSLs for Transformer GPU Kernel Primitives on Nvidia Blackwell

The artifact comprises BF16 and FP32 benchmark implementations, Nsight Systems profiling scripts, and SQL-based verification tools designed to validate Transformer GPU kernel performance on NVIDIA Blackwell hardware. The README.md outlines the steps to replicate the benchmark results from the associated paper, which evaluates Python DSLs for Transformer GPU kernels on NVIDIA Blackwell. The process requires an NVIDIA RTX 5070 Ti GPU, CUDA 13.1 or newer, and a Debian-based environment, with an optional Docker container setup provided. To reproduce the data, users execute the main script run.sh, which runs the BF16 and FP32 benchmarks, profiles them using Nsight Systems, and exports the results to SQLite databases. Users then run five separate verify_table*.sh scripts to query the SQLite files. These verification scripts print kernel latencies that can be compared directly against the tables in the paper to confirm performance consistency.