Changed where masks and eltype are defined for reconstructed loopsets.

Author: chriselrod

src/add_loads.jl

@@ -30,12 +30,19 @@ end
 
 # for use with broadcasting
 function add_simple_load!(
-    ls::LoopSet, var::Symbol, ref::ArrayReference, elementbytes::Int, actualarray::Bool = true, broadcast::Bool = false
+    ls::LoopSet, var::Symbol, ref::ArrayReference, elementbytes::Int,
+    actualarray::Bool = true, broadcast::Bool = false
 )
     loopdeps = Symbol[s for s ∈ ref.indices]
     mref = ArrayReferenceMeta(
         ref, fill(true, length(loopdeps))
     )
+    add_simple_load!(ls, var, mref, loopdeps, elementbytes, actualarray, broadcast)
+end
+function add_simple_load!(
+    ls::LoopSet, var::Symbol, mref::ArrayReferenceMeta, loopdeps::Vector{Symbol},
+    elementbytes::Int, actualarray::Bool = true, broadcast::Bool = false
+)
     op = Operation(
         length(operations(ls)), var, elementbytes,
         :getindex, memload, loopdeps,

src/lowering.jl

@@ -303,20 +303,22 @@ function determine_eltype(ls::LoopSet)
     end
     promote_q
 end
-function determine_width(ls::LoopSet, vectorized::Symbol)
+function determine_width(
+    ls::LoopSet, vectorized::Symbol
+)
     vloop = getloop(ls, vectorized)
     vwidth_q = Expr(:call, lv(:pick_vector_width_val))
     if isstaticloop(vloop)
         push!(vwidth_q.args, Expr(:call, Expr(:curly, :Val, length(vloop))))
     end
     # push!(vwidth_q.args, ls.T)
-    # if length(ls.includedactualarrays) < 2
-    push!(vwidth_q.args, ls.T)
-    # else
-    #     for array ∈ ls.includedactualarrays
-    #         push!(vwidth_q.args, Expr(:call, :eltype, array))
-    #     end
-    # end
+    if length(ls.includedactualarrays) < 2
+        push!(vwidth_q.args, ls.T)
+    else
+        for array ∈ ls.includedactualarrays
+            push!(vwidth_q.args, Expr(:call, :eltype, array))
+        end
+    end
     vwidth_q
 end
 function init_remblock(unrolledloop::Loop, u₁loop::Symbol = unrolledloop.itersymbol)
@@ -358,10 +360,10 @@ function setup_preamble!(ls::LoopSet, us::UnrollSpecification)
     vectorized = order[vectorizedloopnum]
     # println("Setup preamble")
     W = ls.W; typeT = ls.T
-    if length(ls.includedarrays) > 0
+    if length(ls.includedactualarrays) > 0
         push!(ls.preamble.args, Expr(:(=), typeT, determine_eltype(ls)))
+        push!(ls.preamble.args, Expr(:(=), W, determine_width(ls, vectorized)))
     end
-    push!(ls.preamble.args, Expr(:(=), W, determine_width(ls, vectorized)))
     lower_licm_constants!(ls)
     pushpreamble!(ls, definemask(getloop(ls, vectorized), W))#, u₁ > 1 && u₁loopnum == vectorizedloopnum))
     for op ∈ operations(ls)

src/reconstruct_loopset.jl

@@ -387,8 +387,11 @@ function avx_loopset(instr, ops, arf, AM, LPSYM, LB, @nospecialize(vargs))
     nopsv = NOpsType[calcnops(ls, op) for op in ops]
     expandedv = [isexpanded(ls, ops, nopsv, i) for i ∈ eachindex(ops)]
     mrefs = create_mrefs!(ls, arf, arraysymbolinds, opsymbols, nopsv, expandedv, vargs)
-    pushpreamble!(ls, Expr(:(=), ls.T, Expr(:call, :promote_type, [Expr(:call, :eltype, vptr(mref)) for mref ∈ mrefs]...)))
-    # pushpreamble!(ls, Expr(:(=), ls.W, Expr(:call, lv(:pick_vector_width_val), [Expr(:call, :eltype, vptr(mref)) for mref ∈ mrefs]...)))
+    append!(ls.includedactualarrays, (vptr(mref) for mref ∈ mrefs))
+    # eltypes = [Expr(:call, :eltype, vptr(mref)) for mref ∈ mrefs]
+    # pushpreamble!(ls, Expr(:(=), ls.T, Expr(:call, :promote_type, eltypes...)))
+    # pushpreamble!(ls, Expr(:(=), ls.W, determine_width))
+                           # Expr(:call, lv(:pick_vector_width_val), [Expr(:call, :eltype, vptr(mref)) for mref ∈ mrefs]...)))
     num_params = num_arrays + num_parameters(AM)
     add_ops!(ls, instr, ops, mrefs, opsymbols, num_params, nopsv, expandedv, elementbytes)
     process_metadata!(ls, AM, length(arf))

src/split_loops.jl

@@ -11,6 +11,7 @@ function add_operation!(ls_new::LoopSet, included::Vector{Int}, ls::LoopSet, op:
         length(operations(ls_new)), name(op), op.elementbytes, instruction(op), op.node_type,
         loopdependencies(op), reduceddependencies(op), vparents, op.ref, reducedchildren(op)
     )
+    accesses_memory(op) && addsetv!(ls_new.includedactualarrays, vptr(op.ref))
     push!(operations(ls_new), opnew)
     included[identifier(op)] = identifier(opnew)
     opnew

test/gemv.jl

@@ -178,28 +178,28 @@ using Test
         A = rand(R, M, K);
         x = rand(R, K);
         y1 = Vector{TC}(undef, M); y2 = similar(y1);
-        mygemv!(y1, A, x)
-        mygemvavx!(y2, A, x)
+        mygemv!(y1, A, x);
+        mygemvavx!(y2, A, x);
         @test y1 ≈ y2
-        fill!(y2, -999.9); mygemv_avx!(y2, A, x)
+        fill!(y2, -999.9); mygemv_avx!(y2, A, x);
         @test y1 ≈ y2
         fill!(y2, -999.9);
         mygemvavx_range!(y2, A, x)
         @test y1 ≈ y2
 
-        Abit = A .> 0.5
-        fill!(y2, -999.9); mygemv_avx!(y2, Abit, x)
+        Abit = A .> 0.5;
+        fill!(y2, -999.9); mygemv_avx!(y2, Abit, x);
         @test y2 ≈ Abit * x
-        xbit = x .> 0.5
-        fill!(y2, -999.9); mygemv_avx!(y2, A, xbit)
+        xbit = x .> 0.5;
+        fill!(y2, -999.9); mygemv_avx!(y2, A, xbit);
         @test y2 ≈ A * xbit
 
         B = rand(R, N, N);
         G1 = Matrix{TC}(undef, N, 1);
         G2 = similar(G1);
         # G3 = similar(G1);
-        AtmulvB!(G1,B,1)
-        AtmulvBavx1!(G2,B,1)
+        AtmulvB!(G1,B,1);
+        AtmulvBavx1!(G2,B,1);
         @test G1 ≈ G2
         fill!(G2, TC(NaN)); AtmulvBavx2!(G2,B,1);