Ideal versus Reality: Optimal Parallelism and Offloading Support in LLVM
Explicit parallelism and offloading support is an important and growing part of LLVM’s eco-system for CPUs, GPUs, FPGAs and accelerators. LLVM's optimizer has not traditionally been involved in explicit parallelism and offloading support, and specifically, the outlining logic and lowering translation into runtime-library calls resides in Clang. While there are several reasons why the optimizer must be involved in parallelization in order to suitably handle a wide set of applications, the design an appropriate parallel IR for LLVM remains unsettled. Several groups (ANL, Intel, MIT, UIUC) have been experimenting with implementation techniques that push this transformation process into LLVM's IR-level optimization passes [1, 2, 3, 4, 5]. These efforts all aim to allow the optimizer to leverage language properties and optimize parallel constructs before they're transformed into runtime calls and outlined functions. Over the past couple of years, these groups have implemented out-of-tree extensions to LLVM IR to represent and optimize parallelism, and these designs have been influenced by community RFCs [6] and discussions on this topic. In this BoF, we will discuss the use cases we'd like to address and several of the open design questions, including:
* Is a canonical loop form necessary for parallelization and vectorization? * What is the SSA update requirements for extensive loop parallelization and transformation? * What are the required changes to, and impact on, existing LLVM optimization passes and analyses? E.g. inlining and aliasing-information propagation * How to represent and leverage language properties of parallel constructs in LLVM IR? * Where is the proper place in the pipeline to lower these constructs?
The purpose of this BoF is to bring together all parties interested in optimizing parallelism and offloading support in LLVM, as well as the experts in parts of the compiler that will need to be modified. Our goal is to discuss the gap between ideal and reality and identify the pieces that are still missing. In the best case we expect interested parties to agree on the next steps towards better parallelism support Clang and LLVM.
