Apply thread number fix for AMDGPU and oneAPI backends

JuliaORNL · Jan 9, 2025 · f975fe3 · f975fe3
1 parent 2d9e6df
commit f975fe3
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 6 deletions.
diff --git a/ext/JACCAMDGPU/JACCAMDGPU.jl b/ext/JACCAMDGPU/JACCAMDGPU.jl
@@ -68,7 +68,7 @@ end
 function JACC.parallel_reduce(
         ::AMDGPUBackend, N::Integer, op, f::Function, x...; init)
     numThreads = 512
-    threads = min(N, numThreads)
+    threads = numThreads
     blocks = ceil(Int, N / threads)
     ret = fill!(AMDGPU.ROCArray{typeof(init)}(undef, blocks), init)
     rret = AMDGPU.ROCArray([init])
@@ -84,8 +84,8 @@ end
 function JACC.parallel_reduce(
         ::AMDGPUBackend, (M, N)::Tuple{Integer, Integer}, op, f::Function, x...; init)
     numThreads = 16
-    Mthreads = min(M, numThreads)
-    Nthreads = min(N, numThreads)
+    Mthreads = numThreads
+    Nthreads = numThreads
     Mblocks = ceil(Int, M / Mthreads)
     Nblocks = ceil(Int, N / Nthreads)
     ret = fill!(AMDGPU.ROCArray{typeof(init)}(undef, (Mblocks, Nblocks)), init)

diff --git a/ext/JACCONEAPI/JACCONEAPI.jl b/ext/JACCONEAPI/JACCONEAPI.jl
@@ -56,7 +56,7 @@ end
 function JACC.parallel_reduce(
         ::oneAPIBackend, N::Integer, op, f::Function, x...; init)
     numItems = 256
-    items = min(N, numItems)
+    items = numItems
     groups = ceil(Int, N / items)
     ret = oneAPI.zeros(typeof(init), groups)
     rret = oneAPI.zeros(typeof(init), 1)
@@ -70,8 +70,8 @@ end
 function JACC.parallel_reduce(
         ::oneAPIBackend, (M, N)::Tuple{Integer, Integer}, op, f::Function, x...; init)
     numItems = 16
-    Mitems = min(M, numItems)
-    Nitems = min(N, numItems)
+    Mitems = numItems
+    Nitems = numItems
     Mgroups = ceil(Int, M / Mitems)
     Ngroups = ceil(Int, N / Nitems)
     ret = oneAPI.zeros(typeof(init), (Mgroups, Ngroups))