Update compat and Manifest for SIMDPirates version adding ::Mask + ::Vec{W,<:Integer} addition; fixes #61. Also now using given type of reduction vars for temporary accumulators.

Author: chriselrod

Manifest.toml

@@ -49,9 +49,9 @@ uuid = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 
 [[SIMDPirates]]
 deps = ["VectorizationBase"]
-git-tree-sha1 = "839625f8699855a7d5ca96be25bc24d71c5c00ff"
+git-tree-sha1 = "0be25063d6e4306eb656778bb613d32c2ed9268b"
 uuid = "21efa798-c60a-11e8-04d3-e1a92915a26a"
-version = "0.6.0"
+version = "0.6.1"
 
 [[SLEEFPirates]]
 deps = ["Libdl", "SIMDPirates", "VectorizationBase"]

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.6.16"
+version = "0.6.17"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
@@ -12,7 +12,7 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
 Parameters = "0"
-SIMDPirates = "~0.6"
+SIMDPirates = "~0.6.1"
 SLEEFPirates = "~0.4"
 VectorizationBase = "~0.6.1"
 julia = "1.1"

src/condense_loopset.jl

@@ -335,6 +335,7 @@ function setup_call_noinline(ls::LoopSet, U = zero(Int8), T = zero(Int8))
             mvar = mangledvar(op)
             out = Symbol(mvar, 0)
             push!(outer_reducts.args, out)
+            # push!(call.args, Symbol("##TYPEOF##", var))
         end
         push!(q.args, outer_reducts)
         retv = loopset_return_value(ls, Val(false))
@@ -356,15 +357,18 @@ end
 function setup_call_inline(ls::LoopSet, U = zero(Int8), T = zero(Int8))
     call = generate_call(ls, (true,U,T))
     hasouterreductions = length(ls.outer_reductions) > 0
-    if hasouterreductions
-        retv = loopset_return_value(ls, Val(false))
-        call = Expr(:(=), retv, call)
+    if !hasouterreductions
+        q = Expr(:block,gc_preserve(ls, call))
+        append!(ls.preamble.args, q.args)
+        return ls.preamble
     end
-    q = Expr(:block,gc_preserve(ls, call))
+    retv = loopset_return_value(ls, Val(false))
     outer_reducts = Expr(:local)
+    q = Expr(:block,gc_preserve(ls, Expr(:(=), retv, call)))
     for or ∈ ls.outer_reductions
         op = ls.operations[or]
         var = name(op)
+        # push!(call.args, Symbol("##TYPEOF##", var))
         mvar = mangledvar(op)
         instr = instruction(op)
         out = Symbol(mvar, 0)

src/lowering.jl

@@ -154,14 +154,16 @@ function initialize_outer_reductions!(
 )
     reduct_zero = reduction_zero(op.instruction)
     isvectorized = vectorized ∈ reduceddependencies(op)
+    # typeTr = Symbol("##TYPEOF##", name(op))
+    typeTr = Expr(:call, :typeof, mangledvar(op))
     z = if isvectorized
         if reduct_zero === :zero
-            Expr(:call, lv(:vzero), W, typeT)
+            Expr(:call, lv(:vzero), W, typeTr)
         else
-            Expr(:call, lv(:vbroadcast), W, Expr(:call, reduct_zero, typeT))
+            Expr(:call, lv(:vbroadcast), W, Expr(:call, reduct_zero, typeTr))
         end
     else
-        Expr(:call, reduct_zero, typeT)
+        Expr(:call, reduct_zero, typeTr)
     end
     mvar = variable_name(op, suffix)
     for u ∈ Umin:Umax-1

src/reconstruct_loopset.jl

@@ -216,7 +216,6 @@ function sizeofeltypes(v, num_arrays)::Int
     sizeof(T)
 end
 
-
 function avx_loopset(instr, ops, arf, AM, LB, vargs)
     ls = LoopSet(:LoopVectorization)
     num_arrays = length(arf)
@@ -241,6 +240,7 @@ function avx_body(ls, UT)
 end
 
 function _avx_loopset_debug(::Type{OPS}, ::Type{ARF}, ::Type{AM}, ::Type{LB}, vargs...) where {UT, OPS, ARF, AM, LB}
+    @show OPS ARF AM LB vargs
     _avx_loopset(OPS.parameters, ARF.parameters, AM.parameters, LB.parameters, typeof.(vargs))
 end
 function _avx_loopset(OPSsv, ARFsv, AMsv, LBsv, vargs) where {UT, OPS, ARF, AM, LB}

test/dot.jl

@@ -130,6 +130,44 @@
             c_im[i] = b_re[i] * a_im[i + 1] + b_im[i] * a_re[i + 1]
         end
     end
+
+
+    function pi(x, y)
+        acc = 0
+        @inbounds @simd for i ∈ eachindex(x)
+            acc += (x[i]*x[i] + y[i]*y[i]) < 1.0
+        end
+        4acc/length(x)
+    end
+    function piavx(x, y)
+        acc = 0
+        @avx for i ∈ eachindex(x)
+            acc += (x[i]*x[i] + y[i]*y[i]) < 1.0
+        end
+        4acc/length(x)
+    end
+    function piavx_u4(x, y)
+        acc = 0
+        @avx unroll=4 for i ∈ eachindex(x)
+            acc += (x[i]*x[i] + y[i]*y[i]) < 1.0
+        end
+        4acc/length(x)
+    end
+    function pi_avx(x, y)
+        acc = 0
+        @_avx for i ∈ eachindex(x)
+            acc += (x[i]*x[i] + y[i]*y[i]) < 1.0
+        end
+        4acc/length(x)
+    end
+    function pi_avx_u4(x, y)
+        acc = 0
+        @_avx unroll=4 for i ∈ eachindex(x)
+            acc += (x[i]*x[i] + y[i]*y[i]) < 1.0
+        end
+        4acc/length(x)
+    end
+    
     # @macroexpand @_avx for i = 1:length(a_re) - 1
     #     c_re[i] = b_re[i] * a_re[i + 1] - b_im[i] * a_im[i + 1]
     #     c_im[i] = b_re[i] * a_im[i + 1] + b_im[i] * a_re[i + 1]
@@ -155,6 +193,14 @@
         @test myselfdot_avx(a) ≈ s
         @test myselfdotavx(a) ≈ s
 
+        if T <: Union{Float32,Float64}
+            πest = pi(a, b)
+            @test πest == piavx(a, b)
+            @test πest == piavx_u4(a, b)
+            @test πest == pi_avx(a, b)
+            @test πest == pi_avx_u4(a, b)
+        end
+
         a_re = rand(R, N); a_im = rand(R, N);
         b_re = rand(R, N); b_im = rand(R, N);
         ac = Complex.(a_re, a_im);