~/talk

Cocktailer: Analyzing and Optimizing Dynamic Control Flow in Deep Learning

on YouTube 14m
Talkat OSDI 2023 · 2023-07-10..07-12· [machine-learning, compilers-and-language-design]

speakers

name:Chen Zhang
links:none on file
name:Lingxiao Ma
links:none on file
name:Jilong Xue
links:none on file
name:Yining Shi
links:none on file
name:Ziming Miao
links:none on file
name:Fan Yang
linkedin:fyang26
name:Jidong Zhai
links:none on file
name:Zhi Yang
links:none on file
name:Mao Yang
links:none on file

resources

video:https://youtu.be/aX_yReWfE4o
slides:none on file

summary

Cocktailer is a DNN compiler that targets models with dynamic control flow (loops, branches, recursion) and aims to run both control and data flow on accelerators like GPUs instead of handling control on the CPU. It introduces a uTask abstraction to represent control and data uniformly, exposes a global scheduling space to map control flow onto accelerator parallelism, and uses a heuristic scheduler to move control into accelerator kernels for cross-boundary optimizations, claiming up to 8.2× speedups over current frameworks.

AI-written summary · about

related talks

~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.