From 04ee85dbb7a3060477ca5cc5cdda64d1ad5b58e5 Mon Sep 17 00:00:00 2001
From: Kelly Tang <ygtang@berkeley.edu>
Date: Tue, 9 Jul 2024 14:16:46 -0400
Subject: [PATCH] swaped thread dimension

---
 ext/JACCAMDGPU/JACCAMDGPU.jl | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ext/JACCAMDGPU/JACCAMDGPU.jl b/ext/JACCAMDGPU/JACCAMDGPU.jl
index e65116b..e344702 100644
--- a/ext/JACCAMDGPU/JACCAMDGPU.jl
+++ b/ext/JACCAMDGPU/JACCAMDGPU.jl
@@ -34,8 +34,8 @@ function JACC.parallel_for(
         I <: Integer, F <: Function}
     numThreads = 32
     Lthreads = min(L, numThreads)
-    Mthreads = min(M, numThreads)
-    Nthreads = 1
+    Mthreads = 1
+    Nthreads = min(M, numThreads)
     Lblocks = ceil(Int, L / Lthreads)
     Mblocks = ceil(Int, M / Mthreads)
     Nblocks = ceil(Int, N / Nthreads)