Tweak backend selection

src/accumulate/accumulate.jl

@@ -175,7 +175,7 @@ function _accumulate_impl!(
     temp_flags::Union{Nothing, AbstractArray}=nothing,
 )
     if isnothing(dims)
-        return if use_KA_algo(v, prefer_threads)
+        return if use_gpu_algo(backend, prefer_threads)
             accumulate_1d_gpu!(
                 op, v, backend, alg;
                 init, neutral, inclusive,

src/accumulate/accumulate_nd.jl

@@ -35,7 +35,7 @@ function accumulate_nd!(
 
     # Degenerate cases end
 
-    if !use_KA_algo(v, prefer_threads)
+    if !use_gpu_algo(backend, prefer_threads)
         _accumulate_nd_cpu_sections!(op, v; init, dims, inclusive, max_tasks, min_elems)
     else
         # On GPUs we have two parallelisation approaches, based on which dimension has more elements:

src/foreachindex.jl

@@ -130,7 +130,7 @@ function foreachindex(
     # GPU settings
     block_size=256,
 )
-    if use_KA_algo(itr, prefer_threads)
+    if use_gpu_algo(backend, prefer_threads)
         _forindices_gpu(f, eachindex(itr), backend; block_size)
     else
         _forindices_threads(f, eachindex(itr); max_tasks, min_elems)
@@ -232,7 +232,7 @@ function foraxes(
         )
     end
 
-    if use_KA_algo(itr, prefer_threads)
+    if use_gpu_algo(backend, prefer_threads)
         _forindices_gpu(f, axes(itr, dims), backend; block_size)
     else
         _forindices_threads(f, axes(itr, dims); max_tasks, min_elems)

src/predicates.jl

@@ -119,7 +119,7 @@ function _any_impl(
     # GPU settings
     block_size::Int=256,
 )
-    if use_KA_algo(v, prefer_threads)
+    if use_gpu_algo(backend, prefer_threads)
         @argcheck block_size > 0
 
         # Some platforms crash when multiple threads write to the same memory location in a global
@@ -253,7 +253,7 @@ function _all_impl(
     # GPU settings
     block_size::Int=256,
 )
-    if use_KA_algo(v, prefer_threads)
+    if use_gpu_algo(backend, prefer_threads)
         @argcheck block_size > 0
 
         # Some platforms crash when multiple threads write to the same memory location in a global

src/reduce/mapreduce_nd.jl

@@ -114,7 +114,7 @@ function mapreduce_nd(
     end
     dst_size = length(dst)
 
-    if !use_KA_algo(src, prefer_threads)
+    if !use_gpu_algo(backend, prefer_threads)
         _mapreduce_nd_cpu_sections!(
             f, op, dst, src;
             init,

src/reduce/reduce.jl

@@ -183,7 +183,7 @@ function _mapreduce_impl(
     switch_below::Int=0,
 )
     if isnothing(dims)
-        if use_KA_algo(src, prefer_threads)
+        if use_gpu_algo(backend, prefer_threads)
             mapreduce_1d_gpu(
                 f, op, src, backend;
                 init, neutral,

src/sort/sort.jl

@@ -96,7 +96,7 @@ function _sort_impl!(
     # Temporary buffer, same size as `v`
     temp::Union{Nothing, AbstractArray}=nothing,
 )
-    if use_KA_algo(v, prefer_threads)
+    if use_gpu_algo(backend, prefer_threads)
         merge_sort!(
             v, backend;
             lt, by, rev, order,
@@ -207,7 +207,7 @@ function _sortperm_impl!(
     # Temporary buffer, same size as `v`
     temp::Union{Nothing, AbstractArray}=nothing,
 )
-    if use_KA_algo(v, prefer_threads)
+    if use_gpu_algo(backend, prefer_threads)
         merge_sortperm_lowmem!(
             ix, v, backend;
             lt, by, rev, order,

src/utils.jl

@@ -3,8 +3,9 @@ function ispow2(x)
 end
 
 # Helper function to check whether the package cpu implementation of an algorithm should be used
-@inline function use_KA_algo(output_array, prefer_threads)
-    return output_array isa AnyGPUArray || !prefer_threads
+const CPU_BACKEND = get_backend([])
+@inline function use_gpu_algo(backend, prefer_threads)
+    return backend != CPU_BACKEND || !prefer_threads
 end
 
 """