Merge pull request #653 from JuliaGPU/vc/accumulate

Aliasscope miss-compile for 1.11+

Author: vchuravy

src/macros.jl

@@ -87,10 +87,14 @@ function transform_gpu!(def, constargs, force_inbounds, unsafe_indices)
             push!(let_constargs, :($arg = $constify($arg)))
         end
     end
+    has_constargs = !isempty(let_constargs)
     pushfirst!(def[:args], :__ctx__)
     new_stmts = Expr[]
     body = MacroTools.flatten(def[:body])
-    push!(new_stmts, Expr(:aliasscope))
+    # On 1.11 and later having this aliasscope causes issues
+    # even with kernels that don't use `@Const` on arguments
+    # See https://github.com/JuliaGPU/KernelAbstractions.jl/issues/652
+    has_constargs && push!(new_stmts, Expr(:aliasscope))
     if !unsafe_indices
         push!(new_stmts, :(__active_lane__ = $__validindex(__ctx__)))
     end
@@ -105,7 +109,7 @@ function transform_gpu!(def, constargs, force_inbounds, unsafe_indices)
     if force_inbounds
         push!(new_stmts, Expr(:inbounds, :pop))
     end
-    push!(new_stmts, Expr(:popaliasscope))
+    has_constargs && push!(new_stmts, Expr(:popaliasscope))
     push!(new_stmts, :(return nothing))
     def[:body] = Expr(
         :let,
@@ -129,18 +133,22 @@ function transform_cpu!(def, constargs, force_inbounds)
             push!(let_constargs, :($arg = $constify($arg)))
         end
     end
+    has_constargs = !isempty(let_constargs)
     pushfirst!(def[:args], :__ctx__)
     new_stmts = Expr[]
     body = MacroTools.flatten(def[:body])
-    push!(new_stmts, Expr(:aliasscope))
+    # On 1.11 and later having this aliasscope causes issues
+    # even with kernels that don't use `@Const` on arguments
+    # See https://github.com/JuliaGPU/KernelAbstractions.jl/issues/652
+    has_constargs && push!(new_stmts, Expr(:aliasscope))
     if force_inbounds
         push!(new_stmts, Expr(:inbounds, true))
     end
     append!(new_stmts, split(emit_cpu, body.args))
     if force_inbounds
         push!(new_stmts, Expr(:inbounds, :pop))
     end
-    push!(new_stmts, Expr(:popaliasscope))
+    has_constargs && push!(new_stmts, Expr(:popaliasscope))
     push!(new_stmts, :(return nothing))
     def[:body] = Expr(
         :let,

test/test.jl

@@ -371,5 +371,54 @@ function unittest_testsuite(Backend, backend_str, backend_mod, BackendArrayT; sk
         end
     end
 
+    # from https://github.com/JuliaGPU/KernelAbstractions.jl/issues/652
+    @kernel function unaliased_accumulate!(output, input, n)
+        i, j = @index(Global, NTuple)
+
+        for k in j:n
+            output[i, j] += input[i, k]
+        end
+    end
+
+    @kernel function unaliased_accumulate_local!(output, input, n)
+        i, j = @index(Global, NTuple)
+
+        # Use local accumulator
+        sum_val = zero(eltype(output))
+        for k in j:n
+            sum_val += input[i, k]
+        end
+        output[i, j] = sum_val
+    end
+
+    @testset "unaliased accumulate" begin
+        backend = Backend()
+        N = 8
+        M = 5
+
+        input = Float32[i + k for i in 1:M, k in 1:N]
+
+        reference = zeros(Float32, M, N)
+        for i in 1:M
+            for j in 1:N
+                for k in j:N
+                    reference[i, j] += input[i, k]
+                end
+            end
+        end
+
+        # Allocate device arrays
+        input = adapt(backend, input)
+        output = KernelAbstractions.zeros(backend, Float32, M, N)
+
+        # Perform accumulation
+        unaliased_accumulate!(backend)(output, input, N; ndrange = size(output))
+        @test adapt(Array, output) == reference
+
+        fill!(output, 0)
+        unaliased_accumulate_local!(backend)(output, input, N; ndrange = size(output))
+        @test adapt(Array, output) == reference
+    end
+
     return
 end