Updates for VectorizationBase 0.18

Author: chriselrod

Project.toml

@@ -24,8 +24,8 @@ Requires = "1"
 SLEEFPirates = "0.6.7"
 ThreadingUtilities = "0.2.3"
 UnPack = "1"
-VectorizationBase = "0.18"
-julia = "1.6"
+VectorizationBase = "0.18.1"
+julia = "1.5"
 
 [extras]
 Aqua = "4c88cf16-eb10-579e-8560-4a9242c79595"

src/LoopVectorization.jl

@@ -14,7 +14,7 @@ using VectorizationBase: register_size, register_count, cache_linesize, has_opma
     vfma_fast, vmuladd_fast, vdiv_fast, vadd_fast, vsub_fast, vmul_fast,
     relu, stridedpointer, StridedPointer, StridedBitPointer, AbstractStridedPointer,
     reduced_add, reduced_prod, reduce_to_add, reduce_to_prod, reduced_max, reduced_min, reduce_to_max, reduce_to_min,
-    vsum, vprod, vmaximum, vminimum, vstorent!
+    vsum, vprod, vmaximum, vminimum, unwrap
 
 using IfElse: ifelse
 
@@ -41,10 +41,6 @@ export LowDimArray, stridedpointer,
     vmap, vmap!, vmapt, vmapt!, vmapnt, vmapnt!, vmapntt, vmapntt!, tanh_fast, sigmoid_fast,
     vfilter, vfilter!, vmapreduce, vreduce
 
-@inline unwrap(::Val{N}) where {N} = N
-@inline unwrap(::Static{N}) where {N} = N
-@inline unwrap(x) = x
-
 const VECTORWIDTHSYMBOL, ELTYPESYMBOL = Symbol("##Wvecwidth##"), Symbol("##Tloopeltype##")
 
 include("vectorizationbase_compat/contract_pass.jl")

src/constructors.jl

@@ -31,7 +31,7 @@ function add_ci_call!(q::Expr, @nospecialize(f), args, syms, i, mod = nothing)
     end
     if mod !== nothing # indicates it's `vmaterialize(!)`
         reg_size = Expr(:call, lv(:register_size))
-        reg_count = Expr(:call, lv(:register_size))
+        reg_count = Expr(:call, lv(:register_count))
         cache_lnsze = Expr(:call, lv(:cache_linesize))
         push!(call.args, Expr(:call, Expr(:curly, :Val, QuoteNode(mod))), reg_size, reg_count, cache_lnsze)
     end

src/map.jl

@@ -56,7 +56,7 @@ function vmap_singlethread!(
     UNROLL = 4
     LOG2UNROLL = 2
     while i < N - ((W << LOG2UNROLL) - 1)
-        index = VectorizationBase.Unroll{1,1,UNROLL,1,W,0x0000000000000000}((i,))
+        index = VectorizationBase.Unroll{1,W,UNROLL,1,W,0x0000000000000000}((i,))
         v = f(vload.(ptrargs, index)...)
         if NonTemporal
             vstore!(ptry, v, index, True(), True(), True(), register_size())

src/mapreduce.jl

@@ -48,10 +48,10 @@ end
 @inline function _vmapreduce(f::F, op::OP, ::StaticInt{W}, N, ::Type{T}, args::Vararg{AbstractArray{<:NativeTypes},A}) where {F,OP,A,W,T}
     ptrargs = VectorizationBase.zero_offsets.(stridedpointer.(args))
     if N ≥ 4W
-        index = VectorizationBase.Unroll{1,1,4,1,W,0x0000000000000000}((Zero(),)); i = 4W
+        index = VectorizationBase.Unroll{1,W,4,1,W,0x0000000000000000}((Zero(),)); i = 4W
         au = f(vload.(ptrargs, index)...)
         while i < N - ((W << 2) - 1)
-            index = VectorizationBase.Unroll{1,1,4,1,W,0x0000000000000000}((i,)); i += 4W
+            index = VectorizationBase.Unroll{1,W,4,1,W,0x0000000000000000}((i,)); i += 4W
             au = op(au, f(vload.(ptrargs, index)...))
         end
         a_0 = VectorizationBase.reduce_to_onevec(op, au)

src/vmap_grad.jl

@@ -102,7 +102,7 @@ function ∂vmap_singlethread!(
     st = VectorizationBase.static_sizeof(T)
     zero_index = MM{W}(StaticInt(0), st)
     while i < N - ((W << 2) - 1)
-        index = VectorizationBase.Unroll{1,1,4,1,W,0x0000000000000000}((i,))
+        index = VectorizationBase.Unroll{1,W,4,1,W,0x0000000000000000}((i,))
         v = f(init_dual(vload.(ptrargs, index))...)
         dual_store!(ptr∂y, ptry, v, index)
         i = vadd_fast(i, 4W)