Enabling Composition of Kernels and Compilers
summary
Shows concrete MLIR work inside IREE and OpenXLA and backs claims with numbers: a found library kernel is 4× faster locally yet integration causes 29% end-to-end slowdown because it breaks fusion and tiling, forces extra copies and syncs, and mismatches execution models. Argues for smaller microkernels as intrinsics and for lowering custom kernels via MLIR so compiler passes still run, and demonstrates a Mosaic/TPU-style low‑level DSL that fits existing pipelines rather than replacing them.
AI-written summary · about
chapters (8)
- 0:00ML introduction and goals
- 1:04ML customization vision evolution
- 2:48Libraries versus compilers
- 3:30Kernel integration challenges
- 5:58Microkernels solve problems
- 7:42Microkernels in compiler pipelines
- 9:08TPU and XLA custom kernels
- 10:54Compilers and kernels complement
related talks
48mCross-Domain Compilation: Exploiting Synergies Across the CS Community
41mML applications in Linux kernel
34mOptimizing AI Inference with ML Compilers & Hardware
38mFireside chat
32mToward Seamless Model Compilation Integration
44mIntel Math Kernel Library Perspectives & Latest Advances
26mAn MLIR Dialect for WebAssembly
34mBridging the Gap: Compiling and Optimizing Triton Kernels Onto RISC-V Targets Based on MLIR