Removed precompile body, and reduced read code generation in presence of non-vectorized static loops.

Author: chriselrod

Manifest.toml

@@ -36,18 +36,6 @@ git-tree-sha1 = "6a35d9446b40ae5004cd7bd0f1ae3505528c7fd6"
 uuid = "6fe1bfb0-de20-5000-8ca7-80f57d26f881"
 version = "1.0.3"
 
-[[OrderedCollections]]
-deps = ["Random", "Serialization", "Test"]
-git-tree-sha1 = "c4c13474d23c60d20a67b217f1d7f22a40edf8f1"
-uuid = "bac558e1-5e72-5ebc-8fee-abe8a469f55d"
-version = "1.1.0"
-
-[[Parameters]]
-deps = ["OrderedCollections"]
-git-tree-sha1 = "b62b2558efb1eef1fa44e4be5ff58a515c287e38"
-uuid = "d96e819e-fc66-5662-9728-84c9c7592b0a"
-version = "0.12.0"
-
 [[Random]]
 deps = ["Serialization"]
 uuid = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
@@ -74,6 +62,11 @@ uuid = "6462fe0b-24de-5631-8697-dd941f90decc"
 deps = ["Distributed", "InteractiveUtils", "Logging", "Random"]
 uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
+[[UnPack]]
+git-tree-sha1 = "e0cb9715adda456f7657e45377fd3063bf87179a"
+uuid = "3a884ed6-31ef-47d7-9d2a-63182c4928ed"
+version = "0.1.0"
+
 [[VectorizationBase]]
 deps = ["CpuId", "LinearAlgebra"]
 git-tree-sha1 = "0ca41b27f0a918c8ee51bd495818fb7a12b9a19f"

Project.toml

@@ -6,16 +6,16 @@ version = "0.6.22"
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 OffsetArrays = "6fe1bfb0-de20-5000-8ca7-80f57d26f881"
-Parameters = "d96e819e-fc66-5662-9728-84c9c7592b0a"
 SIMDPirates = "21efa798-c60a-11e8-04d3-e1a92915a26a"
 SLEEFPirates = "476501e8-09a2-5ece-8869-fb82de89a1fa"
+UnPack = "3a884ed6-31ef-47d7-9d2a-63182c4928ed"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
 OffsetArrays = "1"
-Parameters = "0"
 SIMDPirates = "0.7.1"
 SLEEFPirates = "0.4"
+UnPack = "0"
 VectorizationBase = "0.9"
 julia = "1.1"
 

benchmark/driver.jl

@@ -10,7 +10,7 @@ include(joinpath(LOOPVECBENCHDIR, "benchmarkflops.jl"))
 include(joinpath(LOOPVECBENCHDIR, "plotbenchmarks.jl"))
 
 
-addprocs((Sys.CPU_THREADS >> 1)-1); nprocs()
+addprocs((Sys.CPU_THREADS >> 1)-1); nworkers()
 
 @everywhere begin
     pkgdir(pkg::String) = abspath(joinpath(dirname(Base.find_package(pkg)), ".."))

src/LoopVectorization.jl

@@ -1,6 +1,6 @@
 module LoopVectorization
 
-using VectorizationBase, SIMDPirates, SLEEFPirates, Parameters
+using VectorizationBase, SIMDPirates, SLEEFPirates, UnPack
 using VectorizationBase: REGISTER_SIZE, REGISTER_COUNT, extract_data, num_vector_load_expr,
     mask, masktable, pick_vector_width_val, valmul, valrem, valmuladd, valadd, valsub, _MM,
     maybestaticlength, maybestaticsize, staticm1, subsetview, vzero, stridedpointer_for_broadcast,

src/lowering.jl

@@ -156,6 +156,7 @@ function lower_unrolled_dynamic(ls::LoopSet, us::UnrollSpecification, n::Int, in
     vectorized = order[vectorizedloopnum]
     nisunrolled = isunrolled(us, n)
     nisvectorized = isvectorized(us, n)
+    loopisstatic = isstaticloop(loop) & (!nisvectorized)
 
     remmask = inclmask | nisvectorized
     Ureduct = (n == num_loops(ls)) ? calc_Ureduct(ls, us) : -1
@@ -178,8 +179,12 @@ function lower_unrolled_dynamic(ls::LoopSet, us::UnrollSpecification, n::Int, in
     end
     remblock = init_remblock(loop, loopsym)
     push!(q.args, remblock)
-    UFt = 1
-    while true
+    UFt = if loopisstatic
+        length(loop) % UF
+    else
+        1
+    end
+    while !iszero(UFt)
         comparison = if nisvectorized
             itercount = if loop.stopexact
                 Expr(:call, :-, loop.stophint, Expr(:call, lv(:valmul), ls.W, UFt))
@@ -196,7 +201,7 @@ function lower_unrolled_dynamic(ls::LoopSet, us::UnrollSpecification, n::Int, in
         remblocknew = Expr(:elseif, comparison, lower_block(ls, ust, n, remmask, UFt))
         push!(remblock.args, remblocknew)
         remblock = remblocknew
-        if !(UFt < UF - 1 + nisvectorized) || UFt == Ureduct
+        if !(UFt < UF - 1 + nisvectorized) || UFt == Ureduct || loopisstatic
             break
         else
             UFt += 1