Merge pull request #105 from williamfgc/reorder-amdgpu

Reorder AMDGPU gridsize
JuliaORNL · Jul 1, 2024 · ef85431 · ef85431
2 parents 01068b7 + 3852a35
commit ef85431
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/ext/JACCAMDGPU/JACCAMDGPU.jl b/ext/JACCAMDGPU/JACCAMDGPU.jl
@@ -33,9 +33,9 @@ function JACC.parallel_for(
         (L, M, N)::Tuple{I, I, I}, f::F, x...) where {
         I <: Integer, F <: Function}
     numThreads = 32
-    Lthreads = min(L, numThreads)
+    Lthreads = 1
     Mthreads = min(M, numThreads)
-    Nthreads = 1
+    Nthreads = min(N, numThreads)
     Lblocks = ceil(Int, L / Lthreads)
     Mblocks = ceil(Int, M / Mthreads)
     Nblocks = ceil(Int, N / Nthreads)