Various fixes.

Author: chriselrod

benchmark/loadsharedlibs.jl

@@ -36,15 +36,15 @@ end
 eigenfile = joinpath(LOOPVECBENCHDIR, "looptestseigen.cpp")
 if !isfile(LIBEIGENTEST) || mtime(eigenfile) > mtime(LIBEIGENTEST)
     # Clang seems to have trouble finding includes
-    if LoopVectorization.VectorizationBase.has_feature("x86_64_avx512f")
+    if Bool(LoopVectorization.VectorizationBase.has_feature(Val(:x86_64_avx512f)))
         run(`g++ -O3 -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
     else
         run(`g++ -O3 -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
     end
 end
 if !isfile(LIBIEIGENTEST) || mtime(eigenfile) > mtime(LIBIEIGENTEST)
     # run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/c++/9 -I/usr/include/c++/9/x86_64-generic-linux -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
-    if LoopVectorization.VectorizationBase.has_feature("x86_64_avx512f")
+    if Bool(LoopVectorization.VectorizationBase.has_feature(Val(:x86_64_avx512f)))
         run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBIEIGENTEST`)
     else
         run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBIEIGENTEST`)

src/filter.jl

@@ -6,7 +6,7 @@ function vfilter!(f::F, x::Vector{T}, y::AbstractArray{T}) where {F,T <: NativeT
     Nrem = N & (W - 1)
     j = 0
     st = VectorizationBase.static_sizeof(T)
-    zero_index = MM{W}(Static(0), st)
+    zero_index = MM(W, Static(0), st)
     GC.@preserve x y begin
         ptr_x = pointer(x)
         ptr_y = pointer(y)

src/map.jl

@@ -196,7 +196,8 @@ function vmap_multithread!(
     # nt = min(Threads.nthreads(), VectorizationBase.SYS_CPU_THREADS, N >> (Wshift + 3))
     nt = min(Threads.nthreads(), VectorizationBase.num_cores(), N >> (Wshift + 5))
 
-    if !((nt > 1) && iszero(ccall(:jl_in_threaded_region, Cint, ())))
+    # if !((nt > 1) && iszero(ccall(:jl_in_threaded_region, Cint, ())))
+    if nt < 2
         vmap_singlethread!(f, ptry, Zero(), N, Val{NonTemporal}(), ptrargs)
         return
     end

src/user_api_conveniences.jl

@@ -18,6 +18,7 @@ function matmul_params(rs::Int, rc::Int, cls::Int)
     order[5], last(order)
 end
 @generated function matmul_params(::StaticInt{RS}, ::StaticInt{RC}, ::StaticInt{CLS}) where {RS,RC,CLS}
+    mᵣ, nᵣ = matmul_params(RS, RC, CLS)
     Expr(:tuple, Expr(:call, Expr(:curly, :StaticInt, mᵣ)), Expr(:call, Expr(:curly, :StaticInt, nᵣ)))
 end
 matmul_params() = matmul_params(register_size(), register_count(), cache_linesize())

test/dot.jl

@@ -298,7 +298,7 @@ using Test
             @test πest == pi_avx_u4(a, b)
         end
 
-        if !(!LoopVectorization.VectorizationBase.has_feature("x86_64_avx2") && T === Int32)
+        if !(!Bool(LoopVectorization.VectorizationBase.has_feature(Val(:x86_64_avx2))) && T === Int32)
             @test dotloopinductvarpow(a) ≈ dotloopinductvarpowavx(a)
         end
         @test dot_from_n_to_100(a, b, 33) == @views mydotavx(a[33:100], b[33:100])

test/gemv.jl

@@ -206,6 +206,27 @@ using Test
         return out
     end
 
+
+    function multiple_muls!(Y, dY, A, dA, b, db)
+        mul!(dY, dA, b)
+        # much of the cost is in memory bandwidth for traversing `A`, so we group the two together
+        mul!(Y, A, b)
+        mul!(dY, A, db, true, true)
+        nothing
+    end
+    function multiple_muls_avx!(Y, dY, A, dA, b, db)
+        @avx for m ∈ axes(A,1)
+            dy = 0.0
+            y = 0.0
+            for n ∈ axes(A,2)
+                dy += dA[m,n] * b[n] + A[m,n] * db[n]
+                y += A[m,n] * b[n]
+            end
+            dY[m] = dy
+            Y[m] = y
+        end
+    end
+
     M, K, N = 51, 49, 61
     for T ∈ (Float32, Float64, Int32, Int64)
         @show T, @__LINE__
@@ -286,5 +307,16 @@ using Test
         out1 = similar(A, 11); out2 = similar(out1);
         @test reinterpret(T,tuplemul!(out1, A, b)) ≈ reinterpret(T,tuplemulavx!(out2, A, b))
 
+
+        A = rand(R, N, N); dA = rand(R, N, N);
+        b = rand(R, N); db = rand(R, N);
+        Y0 = Vector{TC}(undef, N); Y1 = similar(Y0);
+        dY0 = similar(Y0); dY1 = similar(Y0);
+
+        multiple_muls!(Y0, dY0, A, dA, b, db)
+        multiple_muls_avx!(Y1, dY1, A, dA, b, db)
+        @test Y0 ≈ Y1
+        @test dY0 ≈ dY1
+
     end
 end