Improve performance of IndexAtom a little (#615)

src/atoms/IndexAtom.jl

@@ -42,33 +42,41 @@ function evaluate(x::IndexAtom)
     return output(result)
 end
 
-function new_conic_form!(context::Context{T}, x::IndexAtom) where {T}
-    obj = conic_form!(context, only(AbstractTrees.children(x)))
-    m = length(x)
-    n = length(x.children[1])
+function _index(tape::SparseTape{T}, keep_rows::Vector{Int}) where {T}
+    A = tape.operation.matrix
+    indexed = A[keep_rows, :]
+    af = SparseAffineOperation{T}(indexed, tape.operation.vector[keep_rows])
+    return SparseTape{T}(af, tape.variables)
+end
+
+_index(tape::Vector, keep_rows::Vector{Int}) = tape[keep_rows]
+
+function _index_real(
+    obj_size::Tuple,
+    obj_tape::Union{SparseTape,SPARSE_VECTOR},
+    x::IndexAtom,
+)
     if x.inds === nothing
-        sz = length(x.cols) * length(x.rows)
-        J = Vector{Int}(undef, sz)
-        k = 1
-        num_rows = x.children[1].size[1]
-        for c in x.cols
-            for r in x.rows
-                J[k] = num_rows * (convert(Int, c) - 1) + convert(Int, r)
-                k += 1
-            end
-        end
-        index_matrix = create_sparse(T, collect(1:sz), J, one(T), m, n)
-    else
-        index_matrix = create_sparse(
-            T,
-            collect(1:length(x.inds)),
-            collect(x.inds),
-            one(T),
-            m,
-            n,
-        )
+        linear_indices = LinearIndices(CartesianIndices(obj_size))
+        return _index(obj_tape, vec(linear_indices[x.rows, x.cols]))
+    end
+    return _index(obj_tape, vec(collect(x.inds)))
+end
+
+function new_conic_form!(context::Context{T}, x::IndexAtom) where {T}
+    input = x.children[1]
+    if !iscomplex(x) # real case
+        input_tape = conic_form!(context, input)
+        return _index_real(size(input), input_tape, x)
+    end
+    input_tape = conic_form!(context, input)
+    re = _index_real(size(input), real(input_tape), x)
+    im = _index_real(size(input), imag(input_tape), x)
+    if re isa SPARSE_VECTOR
+        @assert im isa SPARSE_VECTOR
+        return ComplexStructOfVec(re, im)
     end
-    return operate(add_operation, T, sign(x), index_matrix, obj)
+    return ComplexTape(re, im)
 end
 
 function Base.getindex(

src/problem_depot/problems/affine.jl

@@ -260,6 +260,26 @@ end
     rtol,
     ::Type{T},
 ) where {T,test}
+    x = ComplexVariable(2)
+    fix!(x, [1, 2] + im * [1, 2])
+    t = Variable()
+    p = minimize(t + real(x[1]), t >= 0; numeric_type = T)
+    handle_problem!(p)
+    if test
+        @test p.optval ≈ 1 atol = atol rtol = rtol
+    end
+
+    x = Variable(4, 2)
+    y = [1:4 5:8]
+    add_constraint!(x, x == y)
+    p = minimize(dot(x[[4, 3], 2], [7, 13]); numeric_type = T)
+    handle_problem!(p)
+    if test
+        # we would get 153 if we weren't respecting the index ordering
+        @test dot(y[[4, 3], 2], [7, 13]) == 147
+        @test p.optval ≈ 147 atol = atol rtol = rtol
+    end
+
     x = Variable(2)
     p = minimize(x[1] + x[2], [x >= 1]; numeric_type = T)
 

src/problems.jl

@@ -172,13 +172,13 @@ function Context(p::Problem{T}, optimizer_factory) where {T}
     if p.head == :satisfy
         MOI.set(context.model, MOI.ObjectiveSense(), MOI.FEASIBILITY_SENSE)
     else
-        obj = _to_scalar_moi(T, cfp)
-        MOI.set(context.model, MOI.ObjectiveFunction{typeof(obj)}(), obj)
         MOI.set(
             context.model,
             MOI.ObjectiveSense(),
             p.head == :maximize ? MOI.MAX_SENSE : MOI.MIN_SENSE,
         )
+        obj = _to_scalar_moi(T, cfp)
+        MOI.set(context.model, MOI.ObjectiveFunction{typeof(obj)}(), obj)
     end
     return context
 end