Merge branch 'nikopj-fold' of github.com:nikopj/NNlibCUDA.jl into nikopj-fold

HPC merge.

Author: Nikola Janjusevic

ext/NNlibCUDA/.buildkite/pipeline.yml

@@ -1,5 +1,5 @@
 steps:
-  - label: "GPU integration with julia v1"
+  - label: "GPU integration with julia v1.6"
     plugins:
       - JuliaCI/julia#v1:
           version: "1.6" 
@@ -9,10 +9,10 @@ steps:
       cuda: "*"
     timeout_in_minutes: 60
 
-  - label: "GPU integration with julia v1.7-nightly"
+  - label: "GPU integration with julia v1"
     plugins:
       - JuliaCI/julia#v1:
-          version: "1.7-nightly"
+          version: "1.8"
       - JuliaCI/julia-test#v1: ~
     agents:
       queue: "juliagpu"

ext/NNlibCUDA/test/fold.jl

@@ -25,7 +25,7 @@
             x = rand(Float64, fill(8, spatial_rank)..., C_in, batch_size)
             w = rand(Float64, fill(2, spatial_rank)..., C_in, C_out)
             cdims = DenseConvDims(x, w; opts...)
-            y = unfold(x, cdims)
+            y = NNlib.unfold(x, cdims)
 
             # test equivalence of fold/unfold across GPU/CPU
             gputest(x -> NNlib.unfold(x, cdims), x)