Updates to improve performance for loops where LLVM compiler needs aliasing information to eliminate loads.

Author: chriselrod

Manifest.toml

@@ -76,6 +76,6 @@ uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [[VectorizationBase]]
 deps = ["CpuId", "LinearAlgebra"]
-git-tree-sha1 = "1e8a90888ec61405ea345c1ac2bdc7d86b99bd69"
+git-tree-sha1 = "b68b3234127d7839280f39bd668fd0025633aa01"
 uuid = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
-version = "0.8.2"
+version = "0.8.5"

Project.toml

@@ -16,7 +16,7 @@ OffsetArrays = "1"
 Parameters = "0"
 SIMDPirates = "0.7"
 SLEEFPirates = "0.4"
-VectorizationBase = "0.8"
+VectorizationBase = "0.8.5"
 julia = "1.1"
 
 [extras]

benchmark/driver.jl

@@ -23,9 +23,9 @@ end
 # sizes = 23:23
 sizes = 256:-1:2
 
-filter2d_dynamic_bench = benchmark_filter2ddynamic(512:-1:2)
-filter2d_3x3_bench = benchmark_filter2d3x3(512:-1:2)
-filter2d_unrolled_bench = benchmark_filter2dunrolled(512:-1:2)
+filter2d_dynamic_bench = benchmark_filter2ddynamic(sizes)#512:-1:2)
+filter2d_3x3_bench = benchmark_filter2d3x3(sizes)#512:-1:2)
+filter2d_unrolled_bench = benchmark_filter2dunrolled(sizes)#512:-1:2)
 
 AmulB_bench = benchmark_AmulB(sizes)
 AmulBt_bench = benchmark_AmulBt(sizes)

docs/src/assets/bench_AmulB_v1.svg

@@ -210,7 +210,7 @@ end
         LHS = ind === nothing ? gensym() : vptrs[ind]
         assigned_names[i] = LHS
         d = (D[i])::Union{Nothing,Int}
-        if d === nothing # stridedpointer
+        if d === nothing # stridedpointer instead of noaliasstridedpointer, because alias info will be lost across function boundary...
             num_arrays += 1
             RHS = Expr(:call, lv(:stridedpointer), Expr(:ref, :vargs, ari), Expr(:ref, :arraydescript, ari))
         else #subsetview
@@ -284,7 +284,7 @@ function setup_call_noinline(ls::LoopSet, U = zero(Int8), T = zero(Int8))
         if ex isa Expr && ex.head === :(=) && length(ex.args) == 2
             if ex.args[2] isa Expr && ex.args[2].head === :call
                 gr = first(ex.args[2].args)
-                if gr == lv(:stridedpointer)
+                if gr == lv(:noaliasstridedpointer)
                     array = ex.args[2].args[2]
                     arrayid = findfirst(a -> a === array, ls.includedactualarrays)
                     if arrayid isa Int

docs/src/assets/bench_AmulBt_v1.svg

@@ -15,14 +15,16 @@ end
 
 add_vptr!(ls::LoopSet, op::Operation) = add_vptr!(ls, op.ref)
 add_vptr!(ls::LoopSet, mref::ArrayReferenceMeta) = add_vptr!(ls, mref.ref.array, vptr(mref))
+using VectorizationBase: noaliasstridedpointer
 function add_vptr!(ls::LoopSet, array::Symbol, vptrarray::Symbol = vptr(array), actualarray::Bool = true, broadcast::Bool = false)
     if !includesarray(ls, array)
         push!(ls.includedarrays, array)
         actualarray && push!(ls.includedactualarrays, array)
         if broadcast
             pushpreamble!(ls, Expr(:(=), vptrarray, Expr(:call, lv(:stridedpointer_for_broadcast), array)))
         else
-            pushpreamble!(ls, Expr(:(=), vptrarray, Expr(:call, lv(:stridedpointer), array)))
+            # pushpreamble!(ls, Expr(:(=), vptrarray, Expr(:call, lv(:stridedpointer), array)))
+            pushpreamble!(ls, Expr(:(=), vptrarray, Expr(:call, lv(:noaliasstridedpointer), array)))
         end
     end
     nothing

docs/src/assets/bench_Amulvb_v1.svg

@@ -23,6 +23,7 @@ end
 # Tuple of length > 1, subtract offsets.
 @inline VectorizationBase.offset(ptr::OffsetStridedPointer{<:Any,N}, ind::Tuple) where {N} = ntuple(n -> ind[n] - ptr.offsets[n], Val{N}())
 @inline Base.similar(p::OffsetStridedPointer, ptr::Ptr) = OffsetStridedPointer(similar(p.ptr, ptr), p.offsets)
+@inline Base.pointer(p::OffsetStridedPointer) = pointer(p.ptr)
 
 # If an OffsetArray is getting indexed by a (loop-)constant value, then this particular vptr object cannot also be eachindexed, so we can safely return a stridedpointer
 @inline function VectorizationBase.subsetview(ptr::OffsetStridedPointer{<:Any,N}, ::Val{I}, i) where {I,N}

docs/src/assets/bench_AplusAt_v1.svg

@@ -33,7 +33,7 @@
                 B[j,i] = A[j,i] - x[j]
                 end)
     lssubcol = LoopVectorization.LoopSet(subcolq);
-    @test LoopVectorization.choose_order(lssubcol) == (Symbol[:j,:i], :i, :j, :j, Unum, Tnum)
+    @test LoopVectorization.choose_order(lssubcol) == (Symbol[:j,:i], :i, :j, :j, 4, 4)
     ## @avx is SLOWER!!!!
     ## need to fix!
     function mysubcol!(B, A, x)
@@ -58,7 +58,7 @@
                 x[j] += A[j,i] - 0.25
                 end)
     lscolsum = LoopVectorization.LoopSet(colsumq);
-    @test LoopVectorization.choose_order(lscolsum) == (Symbol[:j,:i], :i, :j, :j, Unum, Tnum)
+    @test LoopVectorization.choose_order(lscolsum) == (Symbol[:j,:i], :i, :j, :j, 4, 4)
 
     # my colsum is wrong (by 0.25), but slightly more interesting
     function mycolsum!(x, A)
@@ -95,7 +95,7 @@
     lsvar = LoopVectorization.LoopSet(varq);
     # LoopVectorization.choose_order(lsvar)
     # @test LoopVectorization.choose_order(lscolsum) == (Symbol[:j,:i], :j, Symbol("##undefined##"), :j, 4, -1)
-    @test LoopVectorization.choose_order(lsvar) == (Symbol[:j,:i], :i, :j, :j, Unum, Tnum)
+    @test LoopVectorization.choose_order(lsvar) == (Symbol[:j,:i], :i, :j, :j, 4, 4)
 
     function myvar!(s², A, x̄)
         @. s² = 0