Fix fill! memory

Author: Github action runner

src/array.jl

@@ -505,8 +505,17 @@ fill(v, dims...) = fill!(oneArray{typeof(v)}(undef, dims...), v)
 fill(v, dims::Dims) = fill!(oneArray{typeof(v)}(undef, dims...), v)
 
 function Base.fill!(A::oneDenseArray{T}, val) where T
-  B = [convert(T, val)]
-  unsafe_fill!(context(A), device(), pointer(A), pointer(B), length(A))
+  length(A) == 0 && return A
+  val = convert(T, val)
+  sizeof(T) == 0 && return A
+
+  # execute! is async, so we need to allocate the pattern in USM memory
+  # and keep it alive until the operation completes.
+  buf = oneL0.host_alloc(context(A), sizeof(T), Base.datatype_alignment(T))
+  unsafe_store!(convert(Ptr{T}, buf), val)
+  unsafe_fill!(context(A), device(), pointer(A), convert(ZePtr{T}, buf), length(A))
+  synchronize(global_queue(context(A), device()))
+  oneL0.free(buf)
   A
 end
 

test/level-zero.jl

@@ -271,13 +271,23 @@ let src = rand(Int, 1024)
     synchronize(queue)
     @test chk == src
 
+    # FIX: Allocate pattern in USM Host Memory
+    # Standard Host memory (stack/heap) is not accessible by discrete GPUs for fill patterns.
+    # We must use USM Host Memory.
+    pattern_val = 42
+    pattern_buf = oneL0.host_alloc(ctx, sizeof(Int), Base.datatype_alignment(Int))
+    unsafe_store!(convert(Ptr{Int}, pattern_buf), pattern_val)
+
     execute!(queue) do list
-        pattern = [42]
-        append_fill!(list, pointer(dst), pointer(pattern), sizeof(pattern), sizeof(src))
+        # Use the USM pointer (converted to ZePtr)
+        append_fill!(list, pointer(dst), convert(ZePtr{Int}, pattern_buf), sizeof(Int), sizeof(src))
         append_barrier!(list)
         append_copy!(list, pointer(chk), pointer(dst), sizeof(src))
     end
     synchronize(queue)
+
+    oneL0.free(pattern_buf)
+
     @test all(isequal(42), chk)
 
     free(dst)