Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Bug Fix and use of new GPUCompiler persistent caching (8x improvement to loading time) #760

Closed
wants to merge 2 commits into from

Conversation

collinwarner
Copy link

@collinwarner collinwarner commented Apr 23, 2023

Add functionality of GPUCompiler JuliaGPU/GPUCompiler.jl#425. Sees 8x improvement of cached loading times of the enzyme package.

Additionally latent bug found and fixed related to precompilation of Enzyme functions

@collinwarner
Copy link
Author

collinwarner commented Apr 23, 2023

image

@collinwarner collinwarner changed the title add GPUCompiler persistent caching Bug Fix and use of new GPUCompiler persistent caching May 3, 2023
@collinwarner
Copy link
Author

collinwarner commented May 3, 2023

Bug Found:

The following script will error out with a segfault in the current Enzyme implementation. This occurs because the generated functions for thunk and deferred_codegen do no retrigger GPUCompiler and instead use stale values. Simply dropping the generated tag fixes the segfault but we are no longer able to look up unique method instances as the returned world ages are incorrect. Instead we should switch to using GPUCompiler.methodinstance directly dropping the need to filter on world ages.

module EnzymeTest                                                               
using Enzyme                                                                    
                                                                                
f1(x) = x*x                                                                     
                                                                                
@show autodiff(Reverse, f1, Active(1.0))                                        
                                                                                
                                                                                
function __init__()                                                             
    @show autodiff(Reverse, f1, Active(1.0))                                    
end                                                                             
                                                                                
end # module EnzymeTest 

With the following segfault:


julia> using EnzymeTest
[ Info: Precompiling EnzymeTest [87ed9522-ec2b-4441-9f14-f86404eed9ba]
autodiff(Reverse, f1, Active(1.0)) = ((2.0,),)

[1013783] signal (11.1): Segmentation fault
in expression starting at REPL[1]:1
unknown function (ip: 0x7f86ad477020)
macro expansion at /home/collinw/.julia/dev/Enzyme/src/compiler.jl:8620 [inlined]
enzyme_call at /home/collinw/.julia/dev/Enzyme/src/compiler.jl:8312 [inlined]
CombinedAdjointThunk at /home/collinw/.julia/dev/Enzyme/src/compiler.jl:8275 [inlined]
autodiff at /home/collinw/.julia/dev/Enzyme/src/Enzyme.jl:205 [inlined]
autodiff at /home/collinw/.julia/dev/Enzyme/src/Enzyme.jl:228 [inlined]
autodiff at /home/collinw/.julia/dev/Enzyme/src/Enzyme.jl:214 [inlined]
macro expansion at ./show.jl:1153 [inlined]
__init__ at /home/collinw/julia_master/EnzymeTest/src/EnzymeTest.jl:10
Allocations: 2883 (Pool: 2874; Big: 9); GC: 0
Segmentation fault (core dumped)

@collinwarner collinwarner changed the title Bug Fix and use of new GPUCompiler persistent caching Bug Fix and use of new GPUCompiler persistent caching (8x improvement to loading time) May 4, 2023
@vchuravy
Copy link
Member

vchuravy commented May 4, 2023

Can you split that last change out into a new PR?

@collinwarner
Copy link
Author

Changing to new PR, cleaning

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants