TITLE: HPVM: Extending LLVM For Compiling to Heterogeneous Parallel Systems
SPEAKER: Vikram Adve, University of Illinois at Urbana-Champaign
LLVM has been extraordinarily successful as a compiler infrastructure for enabling a wide range of compilers and compiler-based tools for scalar and vector processors, and for supporting GPU compilers for OpenCL and CUDA. LLVM has seen only limited use, however, for other classes of target architectures, such as reconfigurable hardware (FPGAs) and domain-specific accelerators such as for machine learning, image processing, signal processing, graph processing, and other emerging domains. More generally, heterogeneous system-on-chip (SoC) architectures are becoming increasingly important, especially in "edge computing," but LLVM has largely been limited to the host CPU and GPU on such SoCs, even though the number of other programmable components on these systems has been steadily increasing.
In this talk, I will describe an extension of LLVM for developing a compiler infrastructure -- Heterogeneous Parallel Virtual Machine, or HPVM -- for heterogeneous parallel systems . I will briefly describe at a high-level the key parallel abstraction of hierarchical dataflow graphs used in HPVM to describe heterogeneous parallelism, where ordinary LLVM code is used to represent the computatational tasks. The main focus of this part of the talk is how HPVM is integrated on top of LLVM. First, HPVM has been implemented as a set of intrinsic functions that extend the LLVM infrastructure. Second, the HPVM code generation framework reuses existing LLVM (and other) back-ends, in order to leverage existing (often well-tuned) code generators for individual programmable hardware elements, such as NVPTX for NVIDIA GPUs, Intel's SPIR-V code generator for Intel SSE and AVX vector hardware, and Altera's AOCL compiler for targeting Altera's FPGAs.
A second part of the talk will briefly describe how we are extending HPVM to enable greater energy efficiency and performance by taking advantage of approximation opportunities in application domains such as machine learning and image processing. In particular, we are currently developing ApproxHPVM, an extension of HPVM that supports a range of algorithmic and hardware-level approximation mechanisms . Moreover, ApproxHPVM only requires application programmers to specify high-level, "end-to-end" design goals such as the maximum allowable accuracy loss in a neural network or loss of image quality (e.g., PSNR) and the system automatically selects, optimizes and maps approximation choices for individual coarse-grain tensor operations in the application. The goal is to make sophisticated and well-tested approximation techniques widely accessible to application developers.
To conclude, I will briefly discuss how HPVM and ApproxHPVM might be added as a dialect in MLIR so that other MLIR dialects and MLIR-based compilers can use HPVM for diverse heterogeneous hardware targets, including GPUs, FPGAs, and custom accelerators.
The intended target audience for this talk falls into broadly two classes. The first includes compiler practitioners and researchers interested in compiling to heterogeneous systems, such as SoCs, FPGAs, and other "edge-compute" hardware. The second includes language implementers interested in implementing or porting domain-specific languages such as TensorFlow, Halide, SPIRAL, and others to heterogeneous parallel systems.
We envision several takeaways for the audience: (1) Understand how to develop an extension of LLVM that makes it easier to target emerging hardware platforms not sufficiently well-supported by the existing LLVM IR and code generation framework. (2) Expose attendees to the opportunities and challenges in supporting and reasoning about approximate computations in a compiler framework. (3) Discuss the opportunities and limitations of using HPVM for supporting heterogeneous parallel systems in the context of MLIR.
Web Site and Software Availability
More information about HPVM is available at http://hpvm.cs.illinois.edu/. The HPVM infrastructure is implemented as an extension to LLVM. To date, the software is being developed using an internal Git repository at Illinois and has been shared with collaborators at IBM and at Harvard University. We will make it available publicly in open-source form on Github before the FOSDEM conference.
 Maria Kotsifakou, Prakalp Srivastava, Matthew D. Sinclair, Rakesh Komuravelli, Vikram S. Adve and Sarita V. Adve, “HPVM: Heterogeneous Parallel Virtual Machine.” Proceedings of Principles and Practice of Parallel Programming (PPoPP), Feb 2018, Vösendorf / Wien, Austria.
 Hashim Sharif, Prakalp Srivastava, Mohammed Huzaifa, Maria Kotsifakou, Yasmin Sarita, Nathan Zhou, Keyur Joshi, Vikram S. Adve, Sasa Misailovic and Sarita V. Adve, “ApproxHPVM: A Portable Compiler IR for Accuracy-aware Optimizations,” OOPSLA 2019, October 2019, Athens, Greece.