EvanOLeary/pallasbench-robust-gpu-a100
收藏资源简介:
PallasBench是一个针对JAX Pallas GPU内核的鲁棒基准测试数据集,专注于NVIDIA A100 80GB GPU。该数据集包含45个JAX Pallas内核,覆盖3个难度级别(L1:单操作,L2:融合模式,L3:架构级),这些内核最初为TPU设计,但通过2D分块等修复方法适配GPU兼容性。数据集提供了内核的正确性检查、执行时间、编译性能分析结果,其中39个内核通过测试。它旨在填补Pallas在GPU上基准测试的空白,作为进化内核优化(如ShinkaEvolve)的初始种群,并遵循SakanaAI的评估方法和数据格式。数据集包括原始TPU内核代码、GPU修复后代码、评估框架和完整结果文件。
PallasBench is a robust benchmark dataset for JAX Pallas GPU kernels, focused on NVIDIA A100 80GB GPU. It includes 45 JAX Pallas kernels across 3 difficulty levels (L1: single ops, L2: fused patterns, L3: architecture), originally designed for TPU but adapted for GPU compatibility via fixes like 2D tiling. The dataset provides correctness checking, execution timing, and compilation profiling results, with 39 kernels passing tests. It aims to fill the gap in Pallas benchmarking for GPUs, serving as an initial population for evolutionary kernel optimization (e.g., ShinkaEvolve), and follows SakanaAIs evaluation methodology and data format. The dataset includes original TPU kernel code, GPU-fixed code, evaluation framework, and complete result files.




