Indexing changes and more precompiles (#254)

* CSE indices outside of _avx_!

* rangesym is an empty symbol for static loops

* Fix vcmpend bounds

* No longer mark stores as noalias.

* Update precompile statements, reduce some specialization, split deeply nested reductions to stop confusing LLVM with too many phi nodes, fix some reduction naming code.

* Inline if there are constant compute ops to workaround #245

* esc fallback macro, remove Tullio workaround as it didn't work anyway, mangle parameters in `_avx_!` to make sure they don't clash with user's symbols from loops

* More phi-splits for deep reductions, add energy(spin_conf) test (#244).

* Add broutine2x2 to tests (fixes #239)

* Fix typo.

* Don't pass unnecessary arguments

* latency

* Print sizes in staticsizes bench to find out which are failing; cannot reproduce locally.

* Require VectorizationBase 0.19.34

* Quantum tests on >=1.6 only (uses reinterpret(reshape,...))

Author: chriselrod

Project.toml

@@ -13,12 +13,13 @@ OffsetArrays = "6fe1bfb0-de20-5000-8ca7-80f57d26f881"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
 SLEEFPirates = "476501e8-09a2-5ece-8869-fb82de89a1fa"
 Static = "aedffcd0-7271-4cad-89d0-dc628f76c6d3"
+StrideArraysCore = "7792a7ef-975c-4747-a70f-980b88e8d1da"
 ThreadingUtilities = "8290d209-cae3-49c0-8002-c8c24d57dab5"
 UnPack = "3a884ed6-31ef-47d7-9d2a-63182c4928ed"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
-ArrayInterface = "3.1.7"
+ArrayInterface = "3.1.9"
 CheapThreads = "0.1.3,0.2"
 DocStringExtensions = "0.8"
 IfElse = "0.1"
@@ -28,7 +29,7 @@ SLEEFPirates = "0.6.14"
 Static = "0.2"
 ThreadingUtilities = "0.4.1"
 UnPack = "1"
-VectorizationBase = "0.19.30"
+VectorizationBase = "0.19.34"
 julia = "1.5"
 
 [extras]

docs/src/examples/filtering.md

@@ -4,6 +4,8 @@ Here, we convolve a small matrix `kern` with a larger matrix `A`, storing the re
 ```julia
 using LoopVectorization, OffsetArrays, Images
 kern = Images.Kernel.gaussian((1, 1), (3, 3))
+A = rand(130,130);
+out = OffsetArray(similar(A, size(A) .- size(kernel) .+ 1), -1 .- kernel.offsets);
 function filter2davx!(out::AbstractMatrix, A::AbstractMatrix, kern)
     @avx for J in CartesianIndices(out)
         tmp = zero(eltype(out))

src/LoopVectorization.jl

@@ -9,7 +9,7 @@ using VectorizationBase: register_size, register_count, cache_linesize, cache_si
     maybestaticfirst, maybestaticlast, scalar_less, scalar_greaterequal, gep, gesp, NativeTypes, #llvmptr,
     vfmadd, vfmsub, vfnmadd, vfnmsub, vfmadd_fast, vfmsub_fast, vfnmadd_fast, vfnmsub_fast, vfmadd231, vfmsub231, vfnmadd231, vfnmsub231,
     vfma_fast, vmuladd_fast, vdiv_fast, vadd_fast, vsub_fast, vmul_fast,
-    relu, stridedpointer, StridedPointer, StridedBitPointer, AbstractStridedPointer, _vload, _vstore!,
+    relu, stridedpointer, stridedpointer_preserve, StridedPointer, StridedBitPointer, AbstractStridedPointer, _vload, _vstore!,
     reduced_add, reduced_prod, reduce_to_add, reduce_to_prod, reduced_max, reduced_min, reduce_to_max, reduce_to_min,
     reduced_all, reduced_any, reduce_to_all, reduce_to_any,
     vsum, vprod, vmaximum, vminimum, vany, vall, unwrap, Unroll, VecUnroll,
@@ -40,13 +40,14 @@ using SLEEFPirates: log_fast, log2_fast, log10_fast
 
 using ArrayInterface
 using ArrayInterface: OptionallyStaticUnitRange, OptionallyStaticRange, Zero, One, StaticBool, True, False, reduce_tup, indices, UpTri, LoTri
-@static if VERSION ≥ v"1.6.0-rc1" #TODO: delete `else` when dropping 1.5 support
-    using ArrayInterface: static_step
-else # Julia 1.5 did not define `step` on CartesianIndices
-    @inline static_step(x) = ArrayInterface.static_step
-    static_step(x::CartesianIndices) = CartesianIndex(map(step, x.indices))
-end
-const Static = ArrayInterface.StaticInt
+using StrideArraysCore: CloseOpen
+# @static if VERSION ≥ v"1.6.0-rc1" #TODO: delete `else` when dropping 1.5 support
+    # using ArrayInterface: static_step
+# else # Julia 1.5 did not define `step` on CartesianIndices
+@inline static_step(x) = ArrayInterface.static_step(x)
+@inline static_step(x::CartesianIndices) = VectorizationBase.CartesianVIndex(map(static_step, x.indices))
+# end
+const Static = StaticInt
 
 using Requires
 
@@ -62,7 +63,6 @@ const VECTORWIDTHSYMBOL, ELTYPESYMBOL, MASKSYMBOL = Symbol("##Wvecwidth##"), Sym
 
 include("vectorizationbase_compat/contract_pass.jl")
 include("vectorizationbase_compat/subsetview.jl")
-include("closeopen.jl")
 include("getconstindexes.jl")
 include("predicates.jl")
 include("simdfunctionals/map.jl")

src/broadcast.jl

@@ -273,7 +273,7 @@ function add_broadcast!(
     end
     fulldims = Symbol[loopsyms[n] for n ∈ 1:N if ((Dlen < n) || D[n]::Bool)]
     ref = ArrayReference(bcname, fulldims)
-    loadop = add_simple_load!(ls, destname, ref, elementbytes, true, true )::Operation
+    loadop = add_simple_load!(ls, destname, ref, elementbytes, true )::Operation
     doaddref!(ls, loadop)
 end
 function add_broadcast_adjoint_array!(
@@ -283,7 +283,7 @@ function add_broadcast_adjoint_array!(
     # pushprepreamble!(ls, Expr(:(=), parent, Expr(:call, :parent, bcname)))
     # isone(length(loopsyms)) && return extract_all_1_array!(ls, bcname, N, elementbytes)
     ref = ArrayReference(bcname, Symbol[loopsyms[N + 1 - n] for n ∈ 1:N])
-    loadop = add_simple_load!( ls, destname, ref, elementbytes, true, true )::Operation
+    loadop = add_simple_load!( ls, destname, ref, elementbytes, true )::Operation
     doaddref!(ls, loadop)
 end
 function add_broadcast_adjoint_array!(
@@ -294,7 +294,7 @@ function add_broadcast_adjoint_array!(
     # pushprepreamble!(ls, Expr(:(=), parent, Expr(:call, :parent, bcname)))
 
     ref = ArrayReference(bcname, Symbol[loopsyms[2]])
-    loadop = add_simple_load!( ls, destname, ref, elementbytes, true, true )::Operation
+    loadop = add_simple_load!( ls, destname, ref, elementbytes, true )::Operation
     doaddref!(ls, loadop)
 end
 function add_broadcast!(
@@ -313,7 +313,7 @@ function add_broadcast!(
     ls::LoopSet, destname::Symbol, bcname::Symbol, loopsyms::Vector{Symbol},
     ::Type{<:AbstractArray{T,N}}, elementbytes::Int
 ) where {T,N}
-    loadop = add_simple_load!(ls, destname, ArrayReference(bcname, @view(loopsyms[1:N])), elementbytes, true, true)
+    loadop = add_simple_load!(ls, destname, ArrayReference(bcname, @view(loopsyms[1:N])), elementbytes, true)
     doaddref!(ls, loadop)
 end
 function add_broadcast!(
@@ -335,7 +335,7 @@ function add_broadcast!(
     inds = Vector{Symbol}(undef, N+1)
     inds[1] = DISCONTIGUOUS
     inds[2:end] .= @view(loopsyms[1:N])
-    loadop = add_simple_load!(ls, destname, ArrayReference(bcname, inds), elementbytes, true, true)
+    loadop = add_simple_load!(ls, destname, ArrayReference(bcname, inds), elementbytes, true)
     doaddref!(ls, loadop)
 end
 const BroadcastedArray{S<:Broadcast.AbstractArrayStyle,F,A} = Broadcasted{S,Nothing,F,A}
@@ -408,8 +408,10 @@ end
     storeop = add_simple_store!(ls, :dest, ArrayReference(:dest, loopsyms), elementbytes)
     doaddref!(ls, storeop)
     resize!(ls.loop_order, num_loops(ls)) # num_loops may be greater than N, eg Product
-    # return ls
-    Expr(:block, Expr(:meta,:inline), setup_call(ls, :(Base.Broadcast.materialize!(dest, bc)), LineNumberNode(0), inline, false, u₁, u₂, threads%Int), :dest)
+  # return ls
+  sc = setup_call(ls, :(Base.Broadcast.materialize!(dest, bc)), LineNumberNode(0), inline, false, u₁, u₂, threads%Int)
+  # return sc
+    Expr(:block, Expr(:meta,:inline), sc, :dest)
 end
 @generated function vmaterialize!(
     dest′::Union{Adjoint{T,A},Transpose{T,A}}, bc::BC, ::Val{Mod}, ::Val{UNROLL}