Don't consider loops statically known to have a single iteration as valid for threading.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.12.31"
+version = "0.12.32"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/codegen/lower_threads.jl

@@ -644,26 +644,32 @@ function thread_two_loops_expr(
 end
 
 function valid_thread_loops(ls::LoopSet)
-    order, u₁loop, u₂loop, vectorized, u₁, u₂, c, shouldinline = choose_order_cost(ls)
-    # NOTE: `names` are being placed in the opposite order here versus normal lowering!
-    copyto!(names(ls), order); init_loop_map!(ls)
-    u₁loop = getloop(ls, u₁loop)
-    _u₂loop = getloopid_or_nothing(ls, u₂loop)
-    u₂loop = _u₂loop === nothing ? u₁loop : getloop_from_id(ls, _u₂loop)
-    ua = UnrollArgs(u₁loop, u₂loop, getloop(ls, vectorized), u₁, u₂, u₂)
-    valid_thread_loop = fill(true, length(order))
-    for op ∈ operations(ls)
-        if isstore(op) && (length(reduceddependencies(op)) > 0)
-            for reduceddep ∈ reduceddependencies(op)
-                for (i,o) ∈ enumerate(order)
-                    if o === reduceddep
-                        valid_thread_loop[i] = false
-                    end
-                end
-            end
+  order, u₁loop, u₂loop, vectorized, u₁, u₂, c, shouldinline = choose_order_cost(ls)
+  # NOTE: `names` are being placed in the opposite order here versus normal lowering!
+  copyto!(names(ls), order); init_loop_map!(ls)
+  u₁loop = getloop(ls, u₁loop)
+  _u₂loop = getloopid_or_nothing(ls, u₂loop)
+  u₂loop = _u₂loop === nothing ? u₁loop : getloop_from_id(ls, _u₂loop)
+  ua = UnrollArgs(u₁loop, u₂loop, getloop(ls, vectorized), u₁, u₂, u₂)
+  valid_thread_loop = fill(true, length(order))
+  for op ∈ operations(ls)
+    if isstore(op) && (length(reduceddependencies(op)) > 0)
+      for reduceddep ∈ reduceddependencies(op)
+        for (i,o) ∈ enumerate(order)
+          if o === reduceddep
+            valid_thread_loop[i] = false
+          end
         end
+      end
     end
-    valid_thread_loop, ua, c
+  end
+  for (i,o) ∈ enumerate(order)
+    loop = getloop(ls, o)
+    if isstaticloop(loop) & length(loop) ≤ 1
+      valid_thread_loop[i] = false
+    end
+  end
+  valid_thread_loop, ua, c
 end
 function avx_threads_expr(
     ls::LoopSet, UNROLL::Tuple{Bool,Int8,Int8,Bool,Int,Int,Int,Int,Int,Int,Int,UInt},

test/threading.jl

@@ -96,40 +96,42 @@ function convlayer_direct!(
 end
 
 @testset "Threading" begin
-    @show @__LINE__
-    dcd = DenseConvDims{2,(5,5),3,6}()
-    kern4 = rand(Float32, 5, 5, 3, 6);
-    @time for M ∈ 17:50:267
-        img = rand(Float32, M, M, 3, 100);
-        outimage1 = Array{Float32}(undef, size(img,1)+1-size(kern4,1), size(img,2)+1-size(kern4,2), size(kern4,4), size(img,4));
-        outimage2 = similar(outimage1);
+  @show @__LINE__
+  dcd = DenseConvDims{2,(5,5),3,6}()
+  kern4 = rand(Float32, 5, 5, 3, 6);
+  @time for M ∈ 17:50:267
+    img = rand(Float32, M, M, 3, 100);
+    outimage1 = Array{Float32}(undef, size(img,1)+1-size(kern4,1), size(img,2)+1-size(kern4,2), size(kern4,4), size(img,4));
+    outimage2 = similar(outimage1);
 
-        convlayer!(outimage1, img, kern4, dcd);
-        convlayer_direct!(outimage2, img, kern4, dcd);
-        @test outimage1 ≈ outimage2
-    end
-
-    @time for M ∈ 17:399
-        # @show M
-        K = M; N = M;
-        A = rand(M,K); B = rand(K,N);
-        @test dot(A,B) ≈ mydotavx(A,B)
+    convlayer!(outimage1, img, kern4, dcd);
+    convlayer_direct!(outimage2, img, kern4, dcd);
+    @test outimage1 ≈ outimage2
+  end
 
-        C1 = A * B; C0 = similar(C1);
-        @test AmulB!(C0, A, B) ≈ C1
+  @time for M ∈ 17:399
+    # @show M
+    K = M; N = M;
+    A = rand(M,K); B = rand(K,N); b = rand(K);
+    @test dot(A,B) ≈ mydotavx(A,B)
 
-		if VERSION ≥ v"1.6"
-	        x = randn(Complex{Float64}, 3M-1);
-    	    W = randn(Complex{Float64}, 3M-1, 3M+1);
-        	y = randn(Complex{Float64}, 3M+1);
-        	@test dot(x,W,y) ≈ dot3(x,W,y)
-		end
+    C1 = A * B;
+    @test AmulB!(similar(C1), A, B) ≈ C1
+    c1 = A * b;
+    @test AmulB!(similar(c1), A, b) ≈ c1
 
-        kern = OffsetArray(randn(3,3),-2,-2)
-        out1 = OffsetArray(randn(size(A) .- 2), 1, 1)
-        out2 = similar(out1);
-        @test conv!(out1, A, kern) ≈ conv_baseline!(out2, A, kern)
+    if VERSION ≥ v"1.6"
+      x = randn(Complex{Float64}, 3M-1);
+      W = randn(Complex{Float64}, 3M-1, 3M+1);
+      y = randn(Complex{Float64}, 3M+1);
+      @test dot(x,W,y) ≈ dot3(x,W,y)
     end
+
+    kern = OffsetArray(randn(3,3),-2,-2)
+    out1 = OffsetArray(randn(size(A) .- 2), 1, 1)
+    out2 = similar(out1);
+    @test conv!(out1, A, kern) ≈ conv_baseline!(out2, A, kern)
+  end
 end