Revert changes back to vload/vstore! instead of load/store!.

Author: chriselrod

Manifest.toml

@@ -49,15 +49,15 @@ uuid = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 
 [[SIMDPirates]]
 deps = ["VectorizationBase"]
-git-tree-sha1 = "ecacd3f808e559d9e363f2620041c6286c8efaca"
+git-tree-sha1 = "4b1e0b1442fb4af5e6b93b9c7fdeacf287d2653b"
 uuid = "21efa798-c60a-11e8-04d3-e1a92915a26a"
-version = "0.4.0"
+version = "0.5.0"
 
 [[SLEEFPirates]]
 deps = ["Libdl", "SIMDPirates", "VectorizationBase"]
-git-tree-sha1 = "bb99e28c0284de9c3233258a93882429752faa55"
+git-tree-sha1 = "769fd039d0835e8e628d61e2f0c80822ba668497"
 uuid = "476501e8-09a2-5ece-8869-fb82de89a1fa"
-version = "0.3.8"
+version = "0.3.9"
 
 [[Serialization]]
 uuid = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
@@ -71,6 +71,6 @@ uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [[VectorizationBase]]
 deps = ["CpuId", "LinearAlgebra"]
-git-tree-sha1 = "b9b5c8fa55e9b859989e759f405624d16b0b0ca2"
+git-tree-sha1 = "9f8caaa5d033f88e188f62a3dba0dab5f429447a"
 uuid = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
-version = "0.4.2"
+version = "0.5.0"

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.6.14"
+version = "0.6.15"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
@@ -12,9 +12,9 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
 Parameters = "0"
-SIMDPirates = "~0.4"
-SLEEFPirates = "~0.3.8"
-VectorizationBase = "~0.4.2"
+SIMDPirates = "~0.5"
+SLEEFPirates = "~0.3.9"
+VectorizationBase = "~0.5"
 julia = "1.1"
 
 [extras]

README.md

@@ -21,6 +21,10 @@ It then tries to vectorize the loop to improve runtime performance.
 
 The macro assumes that loop iterations can be reordered. It also currently supports simple nested loops, where loop bounds of inner loops are constant across iterations of the outer loop, and only a single loop at each level of noop lest. These limitations should be removed in a future version.
 
+## Benchmarks
+
+Please see the documentation for benchmarks versus base Julia, Clang-Polly, icc, ifort, gfortran, and Eigen. If you would believe any code or compiler flags can be improved, would like to submit your own benchmarks, or have Julia code using LoopVectorization that you would like to be tested for performance regressions on a semi-regular basis, please feel file an issue or PR with the code sample.
+
 ## Examples
 ### Dot Product
 <details>

src/add_constants.jl

@@ -39,7 +39,7 @@ function add_constant!(ls::LoopSet, var::Symbol, mpref::ArrayReferenceMetaPositi
     op = Operation(length(operations(ls)), var, elementbytes, LOOPCONSTANT, constant, NODEPENDENCY, Symbol[], NOPARENTS, mpref.mref)
     add_vptr!(ls, op)
     temp = gensym(:intermediateconstref)
-    pushpreamble!(ls, Expr(:(=), temp, Expr(:call, lv(:load), mpref.mref.ptr, mem_offset(op, UnrollArgs(0, Symbol(""), Symbol(""), nothing)))))
+    pushpreamble!(ls, Expr(:(=), temp, Expr(:call, lv(:vload), mpref.mref.ptr, mem_offset(op, UnrollArgs(0, Symbol(""), Symbol(""), nothing)))))
     pushpreamble!(ls, op, temp)
     pushop!(ls, op, temp)
 end

src/add_loads.jl

@@ -73,9 +73,9 @@ end
 
 struct LoopValue end
 @inline VectorizationBase.stridedpointer(::LoopValue) = LoopValue()
-@inline VectorizationBase.load(::LoopValue, i::Tuple{_MM{W}}) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
-@inline VectorizationBase.load(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
-@inline VectorizationBase.load(::LoopValue, i::Integer) = i + one(i)
-@inline VectorizationBase.load(::LoopValue, i::Tuple{I}) where {I<:Integer} = @inbounds(i[1]) + one(I)
+@inline VectorizationBase.vload(::LoopValue, i::Tuple{_MM{W}}) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
+@inline VectorizationBase.vload(::LoopValue, i::Tuple{_MM{W}}, ::Unsigned) where {W} = _MM{W}(@inbounds(i[1].i) + 1)
+@inline VectorizationBase.vload(::LoopValue, i::Integer) = i + one(i)
+@inline VectorizationBase.vload(::LoopValue, i::Tuple{I}) where {I<:Integer} = @inbounds(i[1]) + one(I)
 @inline Base.eltype(::LoopValue) = Int8
 

src/filter.jl

@@ -10,14 +10,14 @@ function vfilter!(f::F, x::Vector{T}, y::AbstractArray{T}) where {F,T <: SUPPORT
         ptr_x = pointer(x)
         ptr_y = pointer(y)
         for _ ∈ 1:Nrep
-            vy = load(Vec{W,T}, ptr_y, i)
+            vy = vload(Vec{W,T}, ptr_y, i)
             mask = f(SVec(vy))
             SIMDPirates.compressstore!(gep(ptr_x, j), vy, mask)
             i += W
             j += count_ones(mask)
         end
         rem_mask = VectorizationBase.mask(T, Nrem)
-        vy = load(Vec{W,T}, gep(ptr_y, i), rem_mask)
+        vy = vload(Vec{W,T}, gep(ptr_y, i), rem_mask)
         mask = rem_mask & f(SVec(vy))
         SIMDPirates.compressstore!(gep(ptr_x, j), vy, mask)
         j += count_ones(mask)

src/lower_load.jl

@@ -2,7 +2,7 @@ function pushvectorload!(q::Expr, op::Operation, var::Symbol, td::UnrollArgs, U:
     @unpack u, unrolled = td
     ptr = refname(op)
     name, mo = name_memoffset(var, op, td, W, vecnotunrolled)
-    instrcall = Expr(:call, lv(:load), ptr, mo)
+    instrcall = Expr(:call, lv(:vload), ptr, mo)
     if mask !== nothing && (vecnotunrolled || u == U - 1)
         push!(instrcall.args, mask)
     end
@@ -21,7 +21,7 @@ function lower_load_scalar!(
     for u ∈ 0:U-1
         varname = varassignname(var, u, isunrolled)
         td = UnrollArgs(u, unrolled, tiled, suffix)
-        push!(q.args, Expr(:(=), varname, Expr(:call, lv(:load), ptr, mem_offset_u(op, td))))
+        push!(q.args, Expr(:(=), varname, Expr(:call, lv(:vload), ptr, mem_offset_u(op, td))))
     end
     nothing
 end

src/lower_store.jl

@@ -66,7 +66,7 @@ function lower_conditionalstore_scalar!(
         varname = varassignname(var, u, parentisunrolled)
         condvarname = varassignname(condvar, u, condunrolled)
         td = UnrollArgs(u, unrolled, tiled, suffix)
-        push!(q.args, Expr(:&&, condvarname, Expr(:call, lv(:store!), ptr, varname, mem_offset_u(op, td))))
+        push!(q.args, Expr(:&&, condvarname, Expr(:call, lv(:vstore!), ptr, varname, mem_offset_u(op, td))))
     end
     nothing
 end
@@ -103,7 +103,7 @@ function lower_conditionalstore_vectorized!(
         td = UnrollArgs(u, unrolled, tiled, suffix)
         name, mo = name_memoffset(var, op, td, W, vecnotunrolled, parentisunrolled)
         condvarname = varassignname(condvar, u, condunrolled)
-        instrcall = Expr(:call, lv(:store!), ptr, name, mo)
+        instrcall = Expr(:call, lv(:vstore!), ptr, name, mo)
         if mask !== nothing && (vecnotunrolled || u == U - 1)
             push!(instrcall.args, Expr(:call, lv(:combinemasks), condvarname, mask))
         else
@@ -123,7 +123,7 @@ function lower_store_scalar!(
     for u ∈ 0:U-1
         varname = varassignname(var, u, parentisunrolled)
         td = UnrollArgs(u, unrolled, tiled, suffix)
-        push!(q.args, Expr(:call, lv(:store!), ptr, varname, mem_offset_u(op, td)))
+        push!(q.args, Expr(:call, lv(:vstore!), ptr, varname, mem_offset_u(op, td)))
     end
     nothing
 end
@@ -147,7 +147,7 @@ function lower_store_vectorized!(
     for u ∈ 0:U-1
         td = UnrollArgs(u, unrolled, tiled, suffix)
         name, mo = name_memoffset(var, op, td, W, vecnotunrolled, parentisunrolled)
-        instrcall = Expr(:call, lv(:store!), ptr, name, mo)
+        instrcall = Expr(:call, lv(:vstore!), ptr, name, mo)
         if mask !== nothing && (vecnotunrolled || u == U - 1)
             push!(instrcall.args, mask)
         end

src/map.jl

@@ -5,14 +5,14 @@ function vmap_quote(N, ::Type{T}) where {T}
     val = Expr(:call, Expr(:curly, :Val, W))
     q = Expr(:block, Expr(:(=), :M, Expr(:call, :length, :dest)), Expr(:(=), :vdest, Expr(:call, :pointer, :dest)), Expr(:(=), :m, 0))
     fcall = Expr(:call, :f)
-    loopbody = Expr(:block, Expr(:call, :store!, :vdest, fcall, :m), Expr(:(+=), :m, W))
+    loopbody = Expr(:block, Expr(:call, :vstore!, :vdest, fcall, :m), Expr(:(+=), :m, W))
     fcallmask = Expr(:call, :f)
-    bodymask = Expr(:block, Expr(:(=), :__mask__, Expr(:call, :mask, val, Expr(:call, :&, :M, W-1))), Expr(:call, :store!, :vdest, fcallmask, :m, :__mask__))
+    bodymask = Expr(:block, Expr(:(=), :__mask__, Expr(:call, :mask, val, Expr(:call, :&, :M, W-1))), Expr(:call, :vstore!, :vdest, fcallmask, :m, :__mask__))
     for n ∈ 1:N
         arg_n = Symbol(:varg_,n)
         push!(q.args, Expr(:(=), arg_n, Expr(:macrocall, Symbol("@inbounds"), LineNumberNode(@__LINE__,Symbol(@__FILE__)), Expr(:call, :pointer, Expr(:ref, :args, n)))))
-        push!(fcall.args, Expr(:call, :load, val, arg_n, :m))
-        push!(fcallmask.args, Expr(:call, :load, val, arg_n, :m, :__mask__))
+        push!(fcall.args, Expr(:call, :vload, val, arg_n, :m))
+        push!(fcallmask.args, Expr(:call, :vload, val, arg_n, :m, :__mask__))
     end
     loop = Expr(:for, Expr(:(=), :_, Expr(:call, :(:), 0, Expr(:call, :-, Expr(:call, :(>>>), :M, Wshift), 1))), loopbody)
     push!(q.args, loop)