PLSE Seminar - Learned x86 Cost Models: Steps Towards a Learned Compiler Backend

Speaker: Alex Renda , MIT CSAIL

Date: Thursday, March 25, 2021

Time: 11:00 AM to 12:00 PM Note: all times are in the Eastern Time Zone

Public: Yes

Location: https://mit.zoom.us/j/93308858252

Event Type:

Room Description:

Host:

Contact: Eric H. Atkinson, eatkinson@csail.mit.edu

Relevant URL: http://people.csail.mit.edu/eatkinson/plse-seminar/seminars.html

Speaker URL: None

Speaker Photo:
None

Reminders to: plse-seminar@csail.mit.edu, pl@csail.mit.edu, seminars@csail.mit.edu

Reminder Subject: TALK: PLSE Seminar - Learned x86 Cost Models: Steps Towards a Learned Compiler Backend

Abtract:
I will talk about some our efforts to automate construction of x86 basic block cost models used by performance engineers and compilers. First we present Ithemal, a neural-network that can learn to predict the execution time of basic blocks, halving the error of state-of-the-art hand-written x86 basic block cost models. However, the predictions from Ithemal are not interpretable, meaning that performance engineers don’t get signal about what to change in their code to make it run faster. To this end we present DiffTune, a methodology for learning parameters of existing hand-written basic block cost models (e.g., the execution time of each instruction), allowing for learning accurate and interpretable cost models. These components and techniques are steps towards the end goal of learning the entire compiler backend.

Bio:
Alex Renda is a PhD candidate in Michael Carbin's group at MIT. His research interests are at the intersection of programming methodologies and deep learning.

Research Areas:
Programming Languages & Software

Impact Areas:

This event is not part of a series.

Created by Eric H. Atkinson at Friday, March 19, 2021 at 10:55 PM.