do not rely on defaults for optimization ranges in unit tests as GPU dependent; and make them fully deterministic

Author: omlins

src/memopt.jl

@@ -672,19 +672,21 @@ function remove_single_point_optvars(optvars, optranges_arg, offsets, offsets_by
     return tuple((A for A in optvars if !(length(keys(offsets[A]))==1 && length(keys(offsets_by_z[A]))==1) || (!isnothing(optranges_arg) && A ∈ keys(optranges_arg)))...)
 end
 
+function score_centered_nonloop_offsets(candidate::Tuple, offsets::Tuple)
+    return sum(sum((candidate[i] - other[i])^2 for i in eachindex(candidate)) for other in offsets)
+end
+
+function select_offsets_with_max_span(offsets::AbstractDict)
+    loopspan_max = maximum(nonloop_offsets -> length(keys(offsets[nonloop_offsets])), keys(offsets))
+    offsets_with_max_span = Tuple(Tuple(nonloop_offsets) for nonloop_offsets in keys(offsets) if length(keys(offsets[nonloop_offsets])) == loopspan_max)
+    return argmin(nonloop_offsets -> (score_centered_nonloop_offsets(nonloop_offsets, offsets_with_max_span), nonloop_offsets), offsets_with_max_span)
+end
+
 function define_optranges(optranges_arg, optvars, offsets, int_type, package)
     compute_capability = get_compute_capability(package)
     optranges = Dict()
     for A in optvars
-        loopspan_max = 0
-        offsets_with_max_span = ()
-        for nonloop_offsets in keys(offsets[A])
-            loopspan = length(keys(offsets[A][nonloop_offsets]))
-            if loopspan > loopspan_max
-                loopspan_max = loopspan
-                offsets_with_max_span = nonloop_offsets
-            end
-        end
+        offsets_with_max_span = select_offsets_with_max_span(offsets[A])
         fullrange    = typemin(int_type):typemax(int_type)
         pointrange_x = offsets_with_max_span[1]:offsets_with_max_span[1]
         pointrange_y = (length(offsets_with_max_span) > 1) ? (offsets_with_max_span[2]:offsets_with_max_span[2]) : fullrange

test/test_parallel.jl

@@ -1438,6 +1438,13 @@ eval(:(
             end;
             @static if $package != $PKG_KERNELABSTRACTIONS
                 @testset "memopt" begin
+                    fullrange = typemin(Int64):typemax(Int64)
+                    tie_offsets = Dict(:B => Dict((0, 0) => Dict(0 => 1), (1, 1) => Dict(0 => 1), (1, 2) => Dict(0 => 1)))
+                    tied_optranges = ParallelStencil.define_optranges(nothing, (:B,), tie_offsets, Int64, ParallelStencil.PKG_THREADS)
+                    @test tied_optranges[:B] == (1:1, 1:1, fullrange)
+                    symmetric_tie_offsets = Dict(:B => Dict((0, 0) => Dict(0 => 1), (0, 2) => Dict(0 => 1), (2, 0) => Dict(0 => 1), (2, 2) => Dict(0 => 1)))
+                    symmetric_tie_optranges = ParallelStencil.define_optranges(nothing, (:B,), symmetric_tie_offsets, Int64, ParallelStencil.PKG_THREADS)
+                    @test symmetric_tie_optranges[:B] == (0:0, 0:0, fullrange)
                     @parallel_indices (ix, iy, iz) memopt=true loopsize=3 optvars=B optranges=(B=(0:0,0:0,0:0),) function metadata_memopt_probe!(A, B, D)
                         A[ix, iy, iz] = 2.0 * B[ix, iy, iz]
                         return
@@ -1478,24 +1485,17 @@ eval(:(
                     @test metadata_z.shmem_spans == (B = (0, 0),)
                     @test metadata_z.stencilranges == (B = (0:0, 0:0, -1:1),)
                     @test metadata_z.use_any_shmem == false
-                    @parallel_indices (ix, iy, iz) memopt=true loopsize=3 optvars=B function metadata_memopt_fullstencil_probe!(A, B, D)
+                    @parallel_indices (ix, iy, iz) memopt=true loopsize=3 optvars=B optranges=(B=(typemin(Int64):typemax(Int64), typemin(Int64):typemax(Int64), typemin(Int64):typemax(Int64)),) function metadata_memopt_fullstencil_probe!(A, B, D)
                         A[ix, iy, iz] = B[ix-1, iy-1, iz-1] + B[ix, iy, iz] + B[ix+1, iy+1, iz+1] + D[ix, iy, iz, 1]
                         return
                     end
                     metadata_full = @metadata metadata_memopt_fullstencil_probe!(A, B, D)
-                    @static if @isgpu($package)
-                        @test metadata_full.shmem_optvars == (:B,)
-                        @test metadata_full.shmem_optvars isa NTuple{1,Symbol}
-                        @test metadata_full.shmem_spans == (B = (2, 2),)
-                        @test metadata_full.stencilranges == (B = (-1:1, -1:1, -1:1),)
-                        @test metadata_full.use_any_shmem == true
-                    else
-                        @test metadata_full.shmem_optvars == ()
-                        @test metadata_full.shmem_optvars isa NTuple{0,Symbol}
-                        @test metadata_full.shmem_spans == (B = (0, 0),)
-                        @test metadata_full.stencilranges == (B = (0:0, 0:0, 0:0),)
-                        @test metadata_full.use_any_shmem == false
-                    end
+                    @test metadata_full.optranges[:B] == (fullrange, fullrange, fullrange)
+                    @test metadata_full.shmem_optvars == (:B,)
+                    @test metadata_full.shmem_optvars isa NTuple{1,Symbol}
+                    @test metadata_full.shmem_spans == (B = (2, 2),)
+                    @test metadata_full.stencilranges == (B = (-1:1, -1:1, -1:1),)
+                    @test metadata_full.use_any_shmem == true
                     @parallel_indices (ix, iy) ndims=2 memopt=true loopdim=2 loopsize=5 optvars=B optranges=(B=(0:0,-1:1,0:0),) function metadata_memopt_2d_probe!(A, B)
                         if (1 < iy < size(A, 2))
                             A[ix, iy] = B[ix, iy - 1] + B[ix, iy] + B[ix, iy + 1]