Update SIMDPirates version and correctly calculate loopdependencies o…

…f variables marked as possible reductions that are not.
JuliaSIMD · Apr 19, 2020 · 5795539 · 5795539 · chriselrod · Apr 19, 2020
1 parent 80dad59
commit 5795539
Show file tree

Hide file tree

Showing 22 changed files with 40 additions and 20 deletions.
diff --git a/Project.toml b/Project.toml
@@ -15,7 +15,7 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 [compat]
 DocStringExtensions = "0.8"
 OffsetArrays = "1"
-SIMDPirates = "0.7.13"
+SIMDPirates = "0.7.14"
 SLEEFPirates = "0.4.4"
 UnPack = "0"
 VectorizationBase = "0.10.4"

diff --git a/benchmark/driver.jl b/benchmark/driver.jl
@@ -45,7 +45,7 @@ logdettriangle_bench = benchmark_logdettriangle(sizes)
 
 v = 1
 filetype = "svg"
-const PICTURES = joinpath(pkgdir("LoopVectorization"), "docs", "src", "assets")
+const PICTURES = joinpath(pkgdir(LoopVectorization), "docs", "src", "assets")
 save(joinpath(PICTURES, "bench_filter2d_dynamic_v$v.$filetype"), plot(filter2d_dynamic_bench));
 save(joinpath(PICTURES, "bench_filter2d_3x3_v$v.$filetype"), plot(filter2d_3x3_bench));
 save(joinpath(PICTURES, "bench_filter2d_unrolled_v$v.$filetype"), plot(filter2d_unrolled_bench));

diff --git a/docs/src/assets/bench_AmulB_v1.svg b/docs/src/assets/bench_AmulB_v1.svg
diff --git a/docs/src/assets/bench_AmulBt_v1.svg b/docs/src/assets/bench_AmulBt_v1.svg
diff --git a/docs/src/assets/bench_Amulvb_v1.svg b/docs/src/assets/bench_Amulvb_v1.svg
diff --git a/docs/src/assets/bench_AplusAt_v1.svg b/docs/src/assets/bench_AplusAt_v1.svg
diff --git a/docs/src/assets/bench_AtmulB_v1.svg b/docs/src/assets/bench_AtmulB_v1.svg
diff --git a/docs/src/assets/bench_AtmulBt_v1.svg b/docs/src/assets/bench_AtmulBt_v1.svg
diff --git a/docs/src/assets/bench_Atmulvb_v1.svg b/docs/src/assets/bench_Atmulvb_v1.svg
diff --git a/docs/src/assets/bench_aplusBc_v1.svg b/docs/src/assets/bench_aplusBc_v1.svg
diff --git a/docs/src/assets/bench_dot3_v1.svg b/docs/src/assets/bench_dot3_v1.svg
diff --git a/docs/src/assets/bench_dot_v1.svg b/docs/src/assets/bench_dot_v1.svg
diff --git a/docs/src/assets/bench_exp_v1.svg b/docs/src/assets/bench_exp_v1.svg
diff --git a/docs/src/assets/bench_filter2d_3x3_v1.svg b/docs/src/assets/bench_filter2d_3x3_v1.svg
diff --git a/docs/src/assets/bench_filter2d_dynamic_v1.svg b/docs/src/assets/bench_filter2d_dynamic_v1.svg
diff --git a/docs/src/assets/bench_filter2d_unrolled_v1.svg b/docs/src/assets/bench_filter2d_unrolled_v1.svg
diff --git a/docs/src/assets/bench_logdettriangle_v1.svg b/docs/src/assets/bench_logdettriangle_v1.svg
diff --git a/docs/src/assets/bench_random_access_v1.svg b/docs/src/assets/bench_random_access_v1.svg
diff --git a/docs/src/assets/bench_selfdot_v1.svg b/docs/src/assets/bench_selfdot_v1.svg
diff --git a/docs/src/assets/bench_sse_v1.svg b/docs/src/assets/bench_sse_v1.svg
diff --git a/src/add_compute.jl b/src/add_compute.jl
@@ -240,6 +240,7 @@ function add_compute!(
         # parent = getop(ls, var, elementbytes)
         if length(reduceddeps) == 0
             insert!(vparents, reduction_ind, parent)
+            mergesetv!(deps, loopdependencies(parent))
             op = Operation(length(operations(ls)), var, elementbytes, instruction(ls,instr), compute, deps, reduceddeps, vparents)
             pushop!(ls, op, var)
         else

diff --git a/test/miscellaneous.jl b/test/miscellaneous.jl
@@ -153,6 +153,20 @@ using Test
         end
     end
 
+    function setcolumstovectorplus100!(Z::AbstractArray{T}, A) where {T}
+        for i = axes(A,1), j = axes(Z,2)
+            acc = zero(T)
+            acc = acc + A[i] + 100
+            Z[i, j] = acc
+        end
+    end
+    function setcolumstovectorplus100avx!(Z::AbstractArray{T}, A) where {T} 
+        @avx for i = axes(A,1), j = axes(Z,2)
+            acc = zero(T)
+            acc = acc + A[i] + 100
+            Z[i, j] = acc
+        end
+    end
 
     function mvp(P, basis, coeffs::Vector{T}) where {T}
         len_c = length(coeffs)
@@ -557,6 +571,11 @@ using Test
         @test dot3v2avx(x, A, y) ≈ d3
         @test dot3_avx(x, A, y) ≈ d3
 
+        A2 = similar(A);
+        setcolumstovectorplus100!(A, x)
+        setcolumstovectorplus100avx!(A2, x)
+        @test A == A2
+
         maxdeg = 20; nbasis = 1_000; dim = 15;
         r = T == Float32 ? (Int32(1):Int32(maxdeg+1)) : (1:maxdeg+1)
         basis = rand(r, (dim, nbasis));