Merge branch 'master' of https://github.com/chriselrod/LoopVectorization.jl

Author: chriselrod

src/determinestrategy.jl

@@ -548,8 +548,11 @@ function stride_penalty(ls::LoopSet, order::Vector{Symbol})
             push!(v, stride_penalty(ls, op, order, loopfreqs))
         end
     end
-    # 1 / 1024 = 0.0009765625
-    10.0sum(maximum, values(stridepenaltydict)) * Base.power_by_squaring(0.0009765625, length(order))
+    if iszero(length(values(stridepenaltydict)))
+        0.0
+    else # 1 / 1024 = 0.0009765625
+        10.0sum(maximum, values(stridepenaltydict)) * Base.power_by_squaring(0.0009765625, length(order))
+    end
 end
 function isoptranslation(ls::LoopSet, op::Operation, unrollsyms::UnrollSymbols)
     @unpack u₁loopsym, u₂loopsym, vectorized = unrollsyms

src/loopstartstopmanager.jl

@@ -157,10 +157,11 @@ function add_loop_start_stop_manager!(ls::LoopSet)
             end
         end
         loopstarts[i] = loopstartᵢ
-        terminators[i] = if (loopsym ∈ loopinductvars) || (any(r -> any(isequal(-i), r), use_livs))
+        terminators[i] = if (loopsym ∈ loopinductvars) || (any(r -> any(isequal(-i), r), use_livs)) || iszero(length(loopstartᵢ))
             0
         else
-            @assert !iszero(length(loopstartᵢ))
+            # @show i, loopsym loopdependencies.(operations(ls)) operations(ls)
+            # @assert !iszero(length(loopstartᵢ))
             last(ric[argmin(first.(ric))]) # index corresponds to array ref's position in loopstart
         end
     end

src/lower_compute.jl

@@ -251,7 +251,7 @@ function lower_compute!(
             if last(instrcall.args) == varsym
                 pushfirst!(instrcall.args, lv(:vifelse))
                 insert!(instrcall.args, 3, mask)
-            else
+            elseif all(in(loopdependencies(op)), reduceddeps) || any(opp -> mangledvar(opp) === mangledvar(op), parents_op)
                 push!(q.args, Expr(:(=), varsym, Expr(:call, lv(:vifelse), mask, instrcall, varsym)))
                 continue
             end

src/reconstruct_loopset.jl

@@ -494,7 +494,7 @@ Execute an `@avx` block. The block's code is represented via the arguments:
 @generated function _avx_!(::Val{UNROLL}, ::Type{OPS}, ::Type{ARF}, ::Type{AM}, ::Type{LPSYM}, lb::LB, vargs...) where {UNROLL, OPS, ARF, AM, LPSYM, LB}
     # 1 + 1 # Irrelevant line you can comment out/in to force recompilation...
     ls = _avx_loopset(OPS.parameters, ARF.parameters, AM.parameters, LPSYM.parameters, LB.parameters, vargs)
-    # @show avx_body(ls, UNROLL)
+    # return @show avx_body(ls, UNROLL)
     # @show UNROLL, OPS, ARF, AM, LPSYM, LB
     avx_body(ls, UNROLL)
 end

test/miscellaneous.jl

@@ -660,6 +660,38 @@ function maxavx!(R::AbstractArray{T}, Q, keep=nothing) where T
     end
     R
 end
+function splitintonoloop(U = randn(2,2), E1 = randn(2))
+    t = 1
+    a = 1.0
+    _s = 0.0
+    n, k = size(U)
+    @avx for j = 1:k
+        for i = 1:n
+            u = tanh(a * U[i,j])
+            v = a * (1 - t * t)
+            U[i,j] = u
+            _s += v
+        end
+        E1[j] = _s / n
+    end
+    U, E1
+end
+function splitintonoloop_reference(U = randn(2,2), E1 = randn(2))
+    t = 1
+    a = 1.0
+    _s = 0.0
+    n, k = size(U)
+    for j = 1:k
+        for i = 1:n
+            u = tanh(a * U[i,j])
+            v = a * (1 - t * t)
+            U[i,j] = u
+            _s += v
+        end
+        E1[j] = _s / n
+    end
+    U, E1
+end
 
 
 
@@ -865,7 +897,12 @@ end
         @test maxavx!(R, Q) == vec(maximum(Q, dims=(2,3)))
         R .+= randn.(T); Rc = copy(R);
         @test maxavx!(R, Q, true) == max.(vec(maximum(Q, dims=(2,3))), Rc)
-   
+
+        U1 = randn(5,7); E1 = randn(7);
+        U2, E2 = splitintonoloop_reference(copy(U1), copy(E1));
+        splitintonoloop(U1, E1);
+        @test U1 ≈ U2
+        @test E1 ≈ E2
     end
     for T ∈ [Int16, Int32, Int64]
         n = 8sizeof(T) - 1

test/offsetarrays.jl

@@ -1,6 +1,6 @@
 using LoopVectorization, OffsetArrays, Test
 using LoopVectorization.VectorizationBase: StaticUnitRange
-T = Float64
+# T = Float64
 # T = Float32
 
 @testset "OffsetArrays" begin
@@ -201,55 +201,61 @@ T = Float64
     for T ∈ (Float32, Float64)
         @show T, @__LINE__
         A = rand(T, 100, 100); At = copy(A');
-        kern = OffsetArray(rand(T, 3, 3), -1:1, -1:1);
-        out1 = OffsetArray(view(similar(A, size(A) .+ 32), (1:98) .+ 32, (1:98) .+ 32), 1, 1);   # stay away from the edges of A
-        # out1 = OffsetArray(similar(A, size(A).-2), 1, 1);   # stay away from the edges of A
-        out2 = similar(out1); out3 = similar(out1); out4 = similar(out1);
-        skern = SizedOffsetMatrix{T,-1,1,-1,1}(parent(kern));
+        for r ∈ (-1:1, -2:2)
+            @show r
+            fr = first(r); lr = last(r); 
+            kern = OffsetArray(rand(T, length(r), length(r)), r, r);
+            out1 = OffsetArray(view(similar(A, size(A) .+ 32), (1+lr:100-lr) .+ 32, (1+lr:100-lr) .+ 32), lr, lr);   # stay away from the edges of A
+            # out1 = OffsetArray(similar(A, size(A).-2), 1, 1);   # stay away from the edges of A
+            out2 = similar(out1); out3 = similar(out1); out4 = similar(out1);
+            skern = SizedOffsetMatrix{T,fr,lr,fr,lr}(parent(kern));
 
-        old2d!(out1, A, kern);
-        avx2d!(out2, A, kern);
-        @test out1 ≈ out2
+            old2d!(out1, A, kern);
+            avx2d!(out2, A, kern);
+            @test out1 ≈ out2
 
-        avx2douter!(out3, A, kern);
-        @test out1 ≈ out3
+            avx2douter!(out3, A, kern);
+            @test out1 ≈ out3
 
-        fill!(out2, NaN); avx2d!(out2, A, skern);
-        @test out1 ≈ out2
+            fill!(out2, NaN); avx2d!(out2, A, skern);
+            @test out1 ≈ out2
 
-        fill!(out2, NaN); avx2douter!(out2, At', kern);
-        @test out1 ≈ out2
+            fill!(out2, NaN); avx2douter!(out2, At', kern);
+            @test out1 ≈ out2
 
-        fill!(out2, NaN); avx2douter!(out2', A, kern);
-        @test out1 ≈ out2'
+            fill!(out2, NaN); avx2douter!(out2', A, kern);
+            @test out1 ≈ out2'
 
-        fill!(out2, NaN); avx2douter!(out2', At', kern);
-        @test out1 ≈ out2'
+            fill!(out2, NaN); avx2douter!(out2', At', kern);
+            @test out1 ≈ out2'
 
-        fill!(out3, NaN); avx2douter!(out3, A, skern);
-        @test out1 ≈ out3
+            fill!(out3, NaN); avx2douter!(out3, A, skern);
+            @test out1 ≈ out3
 
-        fill!(out3, NaN); avx2dunrolled!(out3, A, skern);
-        @test out1 ≈ out3
+            if r == -1:1
+                fill!(out3, NaN); avx2dunrolled!(out3, A, skern);
+                @test out1 ≈ out3
 
-        fill!(out3, NaN); avx2dunrolled2x2!(out3, A, skern);
-        @test out1 ≈ out3
+                fill!(out3, NaN); avx2dunrolled2x2!(out3, A, skern);
+                @test out1 ≈ out3
 
-        fill!(out3, NaN); avx2dunrolled3x3!(out3, A, skern);
-        @test out1 ≈ out3
-
-        @test avxgeneric!(out4, A, kern) ≈ out1
-        fill!(out4, NaN);
-        @test avxgeneric!(out4, A, skern) ≈ out1
+                fill!(out3, NaN); avx2dunrolled3x3!(out3, A, skern);
+                @test out1 ≈ out3
+            end
+            
+            @test avxgeneric!(out4, A, kern) ≈ out1
+            fill!(out4, NaN);
+            @test avxgeneric!(out4, A, skern) ≈ out1
 
-        fill!(out4, NaN); @test avxgeneric2!(out4, A, kern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4, A, skern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4, At', kern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4, At', skern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4', A, kern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4', A, skern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4', At', kern) ≈ out1
-        fill!(out4, NaN); @test avxgeneric2!(out4', At', skern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4, A, kern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4, A, skern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4, At', kern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4, At', skern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4', A, kern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4', A, skern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4', At', kern) ≈ out1
+            fill!(out4, NaN); @test avxgeneric2!(out4', At', skern) ≈ out1
+        end
     end