Thesis Defense: Decoupling Algorithms from the Organization of Computation for High Performance Imaging
, MIT CSAIL
Date: Thursday, December 19, 2013
Time: 2:00 PM to 3:00 PM Note: all times are in the Eastern Time Zone
Refreshments: 1:45 PM
Location: 32-D463 (Star)
Host: Frédo Durand, MIT CSAIL
Contact: Jonathan Ragan-Kelley, firstname.lastname@example.org
Relevant URL: http://people.csail.mit.edu/jrk/
Speaker URL: None
email@example.com, firstname.lastname@example.org, email@example.com, firstname.lastname@example.org
TALK: Thesis Defense: Decoupling Algorithms from the Organization of Computation for High Performance Imaging
This thesis argues that the performance of image processing pipelines is limited by fundamental tradeoffs between parallelism, locality, and the amount of computation performed. These tradeoffs are determined not just by the algorithms used, but by the way their computations and data are organized. We can explicitly model the space of possible organizations for image processing pipelines at once generally enough to describe and compose many state-of-the-art optimizations, and precisely enough to automatically generate code competitive with hand-tuned expert implementations. Programming image processing pipelines this way is better than with existing languages: it enables dramatically simpler code to run faster than hand-tuned implementations, portably across a wide range of different architectures, while maintaining modularity and composability far greater than comparably optimized C or CUDA. I will demonstrate the application of these concepts to the design and implementation of the Halide language and compiler.
Created by Jonathan M Ragan-Kelley at Saturday, December 07, 2013 at 7:21 PM.