workaround for #353 (disable threading in case of CartesianIndex-ing)

Author: chriselrod

src/reconstruct_loopset.jl

@@ -718,7 +718,7 @@ Execute an `@turbo` block. The block's code is represented via the arguments:
   pushfirst!(ls.preamble.args, :(var"#lv#tuple#args#" = reassemble_tuple(Tuple{var"#LB#",var"#V#"}, var"#flattened#var#arguments#")))
   post = hoist_constant_memory_accesses!(ls)
   # return @show avx_body(ls, var"#UNROLL#")
-  q = if last(var"#UNROLL#") > 1
+  q = if (last(var"#UNROLL#") > 1) && length(var"#LPSYM#") == length(ls.loops)
     inline, u₁, u₂, v, isbroadcast, W, rs, rc, cls, l1, l2, l3, nt = var"#UNROLL#"
     # wrap in `var"#OPS#", var"#ARF#", var"#AM#", var"#LPSYM#"` in `Expr` to homogenize types
     avx_threads_expr(

test/threading.jl

@@ -102,6 +102,14 @@ function convlayer_direct!(
     out
 end
 
+function thread_cartesian_indices!(dst, src)
+  @turbo thread=true for i in eachindex(dst, src)
+    dst[i] = src[i]
+  end
+end
+
+
+
 @testset "Threading" begin
   @show @__LINE__
   dcd = DenseConvDims{2,(5,5),3,6}()
@@ -139,6 +147,10 @@ end
     out2 = similar(out1);
     @test conv!(out1, A, kern) ≈ conv_baseline!(out2, A, kern)
   end
+
+  src_big = rand(4, 10); dst_big = similar(src_big); src = view(src_big, 2:3, :); dst = view(dst_big, 2:3, :);
+  thread_cartesian_indices!(dst, src); dst ≈ src
+
 end