HPVM 1.0 Released As LLVM-Based Compiler For CPUs / GPUs / FPGAs / Accelerators

The latest open-source compiler infrastructure effort seeking to target a wide spectrum of devices from CPUs through GPUs, FPGAs, and accelerators is HPVM. The HPVM project today celebrated its 1.0 milestone.

Like most compiler projects these days, HPVM is based on the LLVM compiler stack. HPVM was also born at the University of Illinois where LLVM itself was first started. We covered the initial work on HPVM more than a year ago in University of Illinois Releases HPVM As Heterogeneous Parallel Systems Compiler.

HPVM uses a target-independent IR based on LLVM (v9.0) compiler IR and with the 1.0 milestone now supports linear algebra tensor operations, PyTorch and Keras front-ends, and other improvements over their prior sub-1.0 release. HPVM 1.0 is currently working with CPUs and NVIDIA GPUs using a combination of cuDNN and OpenCL.

"The intended audience for HPVM includes researchers and developers interested in heterogeneous parallel computing, including those working in the areas of compilers, programming languages, approximate computing, software optimization, static and dynamic program analysis, autotuning, and systems for machine learning," notes the 1.0 announcement.

Those interested in learning more about the Heterogeneous Parallel Virtual Machine can do so via the project site at hpvm.cs.illinois.edu.