IBM open-sources KTIR and Dataflow Scheduler MLIR backend for Spyre accelerator
Overview
IBM's PyTorch Blog describes KTIR, the Dataflow Scheduler and SpyreCode, an MLIR backend that compiles kernels for the IBM Spyre accelerator.
The open-source Dataflow Scheduler turns KTIR kernels into double-buffered load, compute and store pipelines routed through the memory hierarchy. The Dataflow Scheduler builds against upstream LLVM/MLIR, and the post says KTIR and the scheduler are open source.
The authors say their Version 1 handles elementwise ops. They report that the first op ran about 4x faster than on the previous stack, and later ops about 8x faster. These speedups are the authors' own figures and come from a first-party post, not an independent benchmark.
Written by AI from the articles below · updated Oct 20, 12:56 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- PyTorch BlogOfficialPickIBM explains how KTIR and the Dataflow Scheduler compile kernels for Spyre
AIIBM's PyTorch Blog post describes the MLIR backend that compiles Spyre kernels, covering KTIR, the Dataflow Scheduler, DataflowIR and SpyreCode. The open-source scheduler turns KTIR kernels into double-buffered load, compute and store pipelines routed through the memory hierarchy. The authors say their Version 1 handles elementwise ops and that the first op came up about 4x faster than on the previous stack, with later ops about 8x faster.
Heat Trend
Not enough continuous observations to show a trend yet.