Need to wait for non-temporal stores to actually sync, and fig bug where storing reductions was offset by register size, but loading them was hardcoded to 64 (cacheline size)

Author: chriselrod

src/codegen/lower_threads.jl

@@ -202,7 +202,8 @@ end
 
 function outer_reduct_combine_expressions(ls::LoopSet, retv)
     gf = GlobalRef(Core, :getfield)
-    q = Expr(:block, :(var"#load#thread#ret#" = $gf(ThreadingUtilities.load(var"#thread#ptr#", typeof($retv), 64),2,false)))
+    q = Expr(:block, :(var"#load#thread#ret#" = $gf(ThreadingUtilities.load(var"#thread#ptr#", typeof($retv), $(reg_size(ls))),2,false)))
+    # push!(q.args, :(@show var"#load#thread#ret#"))
     for (i,or) ∈ enumerate(ls.outer_reductions)
         op = ls.operations[or]
         var = name(op)
@@ -559,6 +560,7 @@ function thread_two_loops_expr(
         end
         # @show $lastboundexpr
         $_avx_call_
+        # @show $retv
         var"#thread#id#" = 0x00000000
         var"#thread#mask#" = CheapThreads.mask(var"#threads#")
         var"#threads#remain#" = true

test/map.jl

@@ -12,10 +12,12 @@
             c2 = vmapnt(foo, a, b);
             @test c1 ≈ c2
             c2 = vmapntt(foo, a, b);
+            sleep(1e-3) # non-temporal stores won't be automatically synced/coherant, so need to wait!
             @test c1 ≈ c2
             fill!(c2, NaN); @views vmapnt!(foo, c2[2:end], a[2:end], b[2:end]);
             @test @views c1[2:end] ≈ c2[2:end]
             fill!(c2, NaN); @views vmapntt!(foo, c2[2:end], a[2:end], b[2:end]);
+            sleep(1e-3) # non-temporal stores won't be automatically synced/coherant, so need to wait!
             @test @views c1[2:end] ≈ c2[2:end]
         end
 

test/threading.jl

@@ -57,16 +57,17 @@ end
 
 @testset "Threading" begin
     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)
 
         C1 = A * B; C0 = similar(C1);
         @test AmulB!(C0, A, B) ≈ C1
 
-        x = randn(Complex{Float64}, 1783);
-        W = randn(Complex{Float64}, 1783, 1577);
-        y = randn(Complex{Float64}, 1577);
+        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)
 
         kern = OffsetArray(randn(3,3),-2,-2)