fix parallel scaling experiment

scripts/run.jl

@@ -14,4 +14,6 @@ ctx(experiment_surrogate_outerinner_ref, "surrogate/outerinner_ref")
 
 ctx(experiment_global_basic, "global/basic")
 
+# note: run these with multiple worker processes
+# don't forget `@everywhere Pkg.activate(".")`
 ctx(experiment_scaling_opticalflow, "scaling/opticalflow")

scripts/run_experiments.jl

@@ -270,7 +270,7 @@ function experiment_scaling_opticalflow(ctx)
     β = 0.001
 
     ninner = 300
-    Mdir = 2 * floor(Int, sqrt(nworkers()))
+    Mdir = 2 * floor(Int, sqrt(nworkers())) # to have enough workers available
     M = (Mdir, Mdir)
     overlap = (5, 5)
 
@@ -301,10 +301,12 @@ function experiment_scaling_opticalflow(ctx)
 
     tg = timeit(galg)
 
+    nparallel = prod(M) ÷ 2^2
     ws = workers()
+    @assert nparallel <= length(ws)
     df = DataFrame()
-    for np in 0:Mdir÷2
-        nw = 2^np
+    for nw in 1:nparallel
+        nparallel % nw == 0 || continue
         push!(df, (
             nworkers = nw,
             time = timeit(dalg(ws[1:nw]))))

src/problems.jl

@@ -20,11 +20,12 @@ DualTVL1ROFOpProblem(g, B, λ::Real) = DualTVL1ROFOpProblem(g, B, fill!(similar(
 function energy(p, prob::DualTVL1ROFOpProblem)
     d = ndims(p)
 
-    @inline kfΛ(w) = @inbounds divergence(w) + prob.g[w.position]
+    @inline kfΛ(w) = @inbounds divergence(w)
     kΛ = Kernel{ntuple(_->-1:1, d)}(kfΛ)
 
     # v = div(p) + g
     v = map(kΛ, extend(p, ExtensionNothing()))
+    v .+= prob.g
 
     # |v|_B^2
     u = prob.B * vec(v)