Team, Visitors, External Collaborators
Overall Objectives
Research Program
Application Domains
New Software and Platforms
New Results
Bilateral Contracts and Grants with Industry
Partnerships and Cooperations
XML PDF e-pub
PDF e-Pub

Section: New Results

Compiler Optimizations and Analysis

Participants : Fabrice Rastello, Manuel Selva, Fabian Grüber, Diogo Sampaio [CORSE, Inria] , Christophe Guillon [STMicroelectronics] , P. Sadayappan [OSU, USA] , Louis-Noël Pouchet [CSU, USA] , Atanas Rountev [OSU, USA] , Richard Veras [LSU, USA] , Rui Li [UoU, USA] , Aravind Sukumaran-Rajam [OSU, USA] , Tse Meng Low [CMU, USA] .

Our current efforts with regard to code optimization follows two directions. 1. The first consists in improving compiler optimization techniques by considering pattern specific applications such as those related to machine learning. Our first result presented at SC 2019 [10] focuses on tensor contractions. 2. The second consists in developing dynamic analysis based performance debugging tools. Our first results published at PPoPP 2019 [9] and TACO 2020 [7] shows a scalable approach that compresses an execution trace obtained from binary instrumentation and analyses it using a polyhedral compiler.

Analytical Cache Modeling and Tilesize Optimization for Tensor Contractions

Data movement between processor and memory hierarchy is a fundamental bottleneck that limits the performance of many applications on modern computer architectures. Tiling and loop permutation are key techniques for improving data locality. However, selecting effective tile-sizes and loop permutations is particularly challenging for tensor contractions due to the large number of loops. Even state-of-the-art compilers usually produce sub-optimal tile-sizes and loop permutations, as they rely on naïve cost models. In this work we provide an analytical model based approach to multilevel tile size optimization and permutation selection for tensor contractions. Our experimental results show that this approach achieves comparable or better performance than state-of-the-art frameworks and libraries for tensor contractions.

This work is the fruit of the collaboration with OSU. It has been presented at ACM/IEEE International Conference for High Performance Computing, Networking, Storage, and Analysis, SC 2019 [10].

Profiling-based Polyhedral Optimization Feedback

This work addresses the problem of reconstructing a compact (static) representation of a binary execution, automatically detecting hot regions and enabling precise feedback about optimization opportunities potentially missed by the compiler. Our framework handles codes with irregular accesses, pointers with indirections, inter-procedural or recursive loop regions. By enabling binary execution analysis we are able to discover run-time properties (i.e., the ability to form a compact representation) as well as inter-procedural optimization opportunities that cannot be uncovered by standard static analyses. Our design choices were driven towards achieving portability, both in terms of targeted architecture, but also in terms of programming environment (e.g., being robust to arbitrary programming language, compiler, use of third-party binaries, etc.).

A compact and yet precise inter-procedural dynamic dependence graph (DDG) is first computed via: 1. a new instrumentation framework based on QEMU; 2. the use of a new concept of inter-procedural loop-nesting tree; 3. followed by new techniques we introduce for folding, clamping, and widening of the DDG to agglomerate dynamic dependence instances into polyhedra of integer points whenever possible. State-of-the-art polyhedral analysis and transformation systems we specifically modified to provide useful feedback to the user is then used. We extensively evaluate our tool on numerous benchmarks, demonstrating the pratical usefulness of our tool-chain.

This work is the fruit of the collaboration with OSU and and the past collaboration Nano2017 with STMicroelectronics. The main contributions has been presented at the ACM conference on Principles and Practice of Parallel Programming, PPoPP 2019 [9]. The new techniques that allow to build the polyhedral representation from the instrumented execution in a scalable way lead to a separate publication in the ACM Transactions on Architecture and Code Optimization, TACO 2020 [7].