MIT Libraries logoDSpace@MIT

MIT
View Item 
  • DSpace@MIT Home
  • MIT Libraries
  • MIT Theses
  • Doctoral Theses
  • View Item
  • DSpace@MIT Home
  • MIT Libraries
  • MIT Theses
  • Doctoral Theses
  • View Item
JavaScript is disabled for your browser. Some features of this site may not work without it.

The Taichi High-Performance and Differentiable Programming Language for Sparse and Quantized Visual Computing

Author(s)
Hu, Yuanming
Thumbnail
DownloadThesis PDF (24.78Mb)
Advisor
Durand, Frédo
Freeman, William T.
Terms of use
In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/
Metadata
Show full item record
Abstract
Using traditional programming languages such as C++ and CUDA, writing high-performance visual computing code is often laborious and requires deep expertise in performance engineering. This implies an undesirable trade-off between performance and productivity. Emerging visual computing workloads such as sparse data structure operations, differentiable programming, and quantized computation, lead to further development difficulties with existing programming systems. To address these issues, we propose Taichi, an imperative and parallel programming language, tailored for developing high-performance visual computing systems. Taichi leverages domain-specific features of visual computing tasks, providing first-class abstraction and support for spatially sparse computation, differentiable programming, and quantization. With Taichi’s optimizing compiler that has a high-level understanding of these domain-specific language constructs and automatically optimizes Taichi programs, we achieve performance and productivity simultaneously in various visual computing tasks, especially physical simulation. For example, with Taichi we can easily achieve 4.55× higher performance using 1/10 lines of code on sparse computations, effortlessly develop 10 differentiable physical simulators, and simulate unprecedented 235 million material point method (MPM) particles on a single GPU.
Date issued
2021-06
URI
https://hdl.handle.net/1721.1/139327
Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Publisher
Massachusetts Institute of Technology

Collections
  • Doctoral Theses

Browse

All of DSpaceCommunities & CollectionsBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

My Account

Login

Statistics

OA StatisticsStatistics by CountryStatistics by Department
MIT Libraries
PrivacyPermissionsAccessibilityContact us
MIT
Content created by the MIT Libraries, CC BY-NC unless otherwise noted. Notify us about copyright concerns.