Hackathon news (#18)

PTsolvers · Oct 25, 2024 · 557975f · 557975f
2 parents c9b47cf + c742df7
commit 557975f
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 0 deletions.
diff --git a/_assets/images/AdjointSensitivities.png b/_assets/images/AdjointSensitivities.png
diff --git a/posts/julia-hackathon-v6-2024-news_1.md b/posts/julia-hackathon-v6-2024-news_1.md
@@ -73,3 +73,17 @@ iteration         1    1.43s  100.0%   1.43s   56.2MiB  100.0%  56.2MiB
   forward        10   73.0μs    0.0%  7.30μs     0.00B    0.0%    0.00B
 ───────────────────────────────────────────────────────────────────────
 ```
+
+## Using Enzyme.jl in order to calculate adjoint sensitivies with JustRelax.jl
+*Christian Schuler, Valentin Churavy, Albert de Montserrat, Pascal Aellig*
+
+[ParallelStencil.jl](https://github.com/omlins/ParallelStencil.jl) has been made compatible with the latest [Enzyme.jl](https://github.com/EnzymeAD/Enzyme.jl) version (PR [#169](https://github.com/omlins/ParallelStencil.jl/pull/169) and PR [#170](https://github.com/omlins/ParallelStencil.jl/pull/170)). With the help of [Enzyme.jl](https://github.com/EnzymeAD/Enzyme.jl) and [ParallelStencil.jl](https://github.com/omlins/ParallelStencil.jl) the neccessary vector-Jacobian products (VJP) for the adjoint solve in [JustRelax.jl](https://github.com/PTsolvers/JustRelax.jl) can be calculated. Work has also been done to make the adjoint solve work on multiple GPUs/CPUs. The figure shows a viscoelastic falling block example with adjoint sensitivities w.r.t. to the viscosity and density. 
+
+~~~
+<center>
+    <img src="../../assets/images/AdjointSensitivities.png" title="Falling Block" alt="FallingBlock" width="75%">
+</center>
+~~~
+
+
+