English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Tools annual

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

While parallelism remains the main source of performance, architecturalimplementations and programming models change with each new hardwaregeneration, often leading to costly application re-engineering. Most tools forperformance portability require manual and costly application porting to yetanother programming model.  We propose an alternative approach that automatically translates programswritten in one programming model (CUDA), into another (CPU threads) based onPolygeist/MLIR. Our approach includes a representation of parallel constructsthat allows conventional compiler transformations to apply transparently andwithout modification and enables parallelism-specific optimizations. Weevaluate our framework by transpiling and optimizing the CUDA Rodinia benchmarksuite for a multi-core CPU and achieve a 76% geomean speedup over handwrittenOpenMP code. Further, we show how CUDA kernels from PyTorch can efficiently runand scale on the CPU-only Supercomputer Fugaku without user intervention. OurPyTorch compatibility layer making use of transpiled CUDA PyTorch kernelsoutperforms the PyTorch CPU native backend by 2.7$\times$.

High-Performance GPU-to-CPU Transpilation and Optimization via High-Level Parallel Constructs