Make creation of output buffer much faster

Author: meggart

src/buffers.jl

@@ -43,7 +43,7 @@ is_init_callable(::Union{DataType,UnionAll}) = Val{true}()
 array_from_init(::Nothing, buftype, bufsize) = zeros(buftype, bufsize)
 array_from_init(init, buftype, bufsize) = array_from_init(init, is_init_callable(init), buftype, bufsize)
 array_from_init(init, ::Val{true}, buftype, bufsize) = buftype[init() for _ in CartesianIndices(bufsize)]
-array_from_init(init, ::Val{false}, buftype, bufsize) = buftype[init for _ in CartesianIndices(bufsize)]
+array_from_init(init, ::Val{false}, buftype, bufsize) = fill(convert(buftype, init), bufsize)
 
 #buftype_from_init(_,ia) =
 

src/distribute.jl

@@ -46,7 +46,7 @@ function split_dim_reasons(op,lr,outars)
     for (spec,ar) in zip(op.outspecs,outars)
         foreach(1:ndims(lr)) do idim
             if is_output_chunk_overlap(spec,ar,idim,lr)
-                @warn "Overlapping output chunks in dimension $idim"
+                op.f.red === nothing && @warn "Overlapping output chunks in dimension $idim"
                 push!(ret[idim],:output_chunk)
             end
             if is_output_reducedim(spec,ar,idim)

src/graphs.jl

@@ -308,8 +308,8 @@ end
 function graph_fuse_ready(g::MwopGraph)
     label = zeros(Int, maximum(keys(g.nodes)))
     Graphs.connected_components!(label, g)
-    _, d = Graphs.components(label)
-    length(g.connections) == count(>(0), keys(d))
+    comps, _ = Graphs.components(label)
+    length(g.connections) == count(i -> length(i) > 1, comps)
 end
 
 

src/runner.jl

@@ -171,8 +171,8 @@ end
 function default_loopbody(inow, op, inbuffers_pure, outbuffers, threaded, outars, cb, runfilter, piddir)
     @debug "inow = ", inow
     if all(c -> need_run(inow, c), runfilter)
-        inbuffers_wrapped = read_range.((inow,),op.inars,inbuffers_pure);
-        outbuffers_now = extract_outbuffer.((inow,),op.outspecs,op.f.init,op.f.buftype,outbuffers)
+        inbuffers_wrapped = read_range.((inow,), op.inars, inbuffers_pure)
+        outbuffers_now = extract_outbuffer.((inow,), op.outspecs, op.f.init, op.f.buftype, outbuffers)
         run_block(op,inow,inbuffers_wrapped,outbuffers_now,threaded)
         put_buffer.((inow,),outbuffers_now,outars,(piddir,))
         clean_aggregator.(outbuffers)