Fix bug where the LHS ref was used instead of the RHS on one-line load-stores, and insert an extra symbol with a new name so that they hve separate entries in syms/refs_aliasing_refs/syms. Fixes #35.

Author: chriselrod

.travis.yml

@@ -4,6 +4,7 @@ os:
   - linux
   - osx
 julia:
+  - 1.1
   - 1.3
   - nightly
 matrix:

Project.toml

@@ -17,7 +17,7 @@ Parameters = "0"
 SIMDPirates = "0.3"
 SLEEFPirates = "0.3"
 VectorizationBase = "0.2"
-julia = "1.3"
+julia = "1.1"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

benchmark/driver.jl

@@ -47,7 +47,7 @@ AtmulB_bench = fetch(AtmulB_future)
 AmulBt_bench = fetch(AmulBt_future)
 Atmulvb_bench = fetch(Atmulvb_future)
 
-v = 3
+v = 4
 const PICTURES = "/home/chriselrod/Pictures"
 save(joinpath(PICTURES, "bench_gemm_v$v.png"), plot(gemm_bench));
 save(joinpath(PICTURES, "bench_AtmulB_v$v.png"), plot(AtmulB_bench));

src/add_stores.jl

@@ -13,33 +13,43 @@ function add_store!(ls::LoopSet, op::Operation)
     id = op.identifier
     id == nops ? add_unique_store!(ls, op) : cse_store!(ls, op)
 end
+function add_copystore!(
+    ls::LoopSet, parent::Operation, mpref::ArrayReferenceMetaPosition, elementbytes::Int
+)
+    op = add_compute!(ls, gensym(), :identity, [parent], elementbytes)
+    add_store!(ls, name(op), mpref, elementbytes, op)
+end
+
 function add_store!(
-    ls::LoopSet, var::Symbol, mpref::ArrayReferenceMetaPosition, elementbytes::Int = 8
+    ls::LoopSet, var::Symbol, mpref::ArrayReferenceMetaPosition, elementbytes::Int = 8, parent = getop(ls, var, mpref.loopdependencies, elementbytes)
 )
+    isload(parent) && return add_copystore!(ls, parent, mpref, elementbytes)
     parents = mpref.parents
     ldref = mpref.loopdependencies
     reduceddeps = mpref.reduceddeps
-    parent = getop(ls, var, ldref, elementbytes)
-    pvar = parent.variable
+    pvar = name(parent)
     id = length(ls.operations)
     if pvar ∉ ls.syms_aliasing_refs
         push!(ls.syms_aliasing_refs, pvar)
         push!(ls.refs_aliasing_syms, mpref.mref)
     else
         # try to cse store, by replacing the previous one
-         ref = mpref.mref.ref
+        ref = mpref.mref.ref
         for opp ∈ operations(ls)
             isstore(opp) || continue
             if ref == opp.ref.ref
                 id = opp.identifier
                 break
+            else
+                @show ref opp.ref.ref
             end
         end
     end
     pushparent!(parents, ldref, reduceddeps, parent)
     op = Operation( id, name(mpref), elementbytes, :setindex!, memstore, mpref )#loopdependencies, reduceddeps, parents, mpref.mref )
     add_store!(ls, op)
 end
+
 function add_store!(
     ls::LoopSet, var::Symbol, array::Symbol, rawindices, elementbytes::Int = 8
 )

src/graphs.jl

@@ -192,12 +192,6 @@ function save_tilecost!(ls::LoopSet)
 end
 
 
-# function op_to_ref(ls::LoopSet, op::Operation)
-    # s = op.variable
-    # id = findfirst(ls.syms_aliasing_regs)
-    # @assert id !== nothing
-    # ls.refs_aliasing_syms[id]
-# end
 function pushpreamble!(ls::LoopSet, op::Operation, v::Symbol)
     if v !== mangledvar(op)
         push!(ls.preamble_symsym, (identifier(op),v))
@@ -450,7 +444,10 @@ function add_operation!(
     ls::LoopSet, LHS_sym::Symbol, RHS::Expr, LHS_ref::ArrayReferenceMetaPosition, elementbytes::Int = 8
 )
     if RHS.head === :ref# || (RHS.head === :call && first(RHS.args) === :getindex)
-        add_load!(ls, LHS_sym, LHS_ref, elementbytes)
+        array, rawindices = ref_from_expr(RHS)
+        RHS_ref = array_reference_meta!(ls, array, rawindices, elementbytes)
+        op = add_load!(ls, gensym(LHS_sym), RHS_ref, elementbytes)
+        add_compute!(ls, LHS_sym, :identity, [op], elementbytes)
     elseif RHS.head === :call
         f = first(RHS.args)
         if f === :getindex

test/runtests.jl

@@ -1285,6 +1285,29 @@ end
         lse, qq, xx, tmpmax, maxk, nk = softmax3_setup!(q, lse, tmpmax, x, maxk)
         softmax3_core_avx2!(lse, qq, xx, tmpmax, maxk, nk)
     end
+
+    function copyavx1!(x, y)
+        @avx for i ∈ eachindex(x)
+            x[i] = y[i]
+        end
+    end
+    function copy_avx1!(x, y)
+        @_avx for i ∈ eachindex(x)
+            x[i] = y[i]
+        end
+    end
+    function copyavx2!(x, y)
+        @avx for i ∈ eachindex(x)
+            yᵢ = y[i]
+            x[i] = yᵢ
+        end
+    end
+    function copy_avx2!(x, y)
+        @_avx for i ∈ eachindex(x)
+            yᵢ = y[i]
+            x[i] = yᵢ
+        end
+    end
 
     for T ∈ (Float32, Float64)
         @show T, @__LINE__
@@ -1366,7 +1389,15 @@ end
         fill!(q2, 0); fill!(lse, 0);  softmax3_avx2!(q2, lse, tmpmax, x);
         @test q1 ≈ q2
         @test sum(q2; dims=3) ≈ ones(T,ni,nj)
-        
+
+        fill!(q2, NaN); copyavx1!(q2, x)
+        @test x == q2
+        fill!(q2, NaN); copy_avx1!(q2, x)
+        @test x == q2
+        fill!(q2, NaN); copyavx2!(q2, x)
+        @test x == q2
+        fill!(q2, NaN); copy_avx2!(q2, x)
+        @test x == q2
     end
 end