Rewrite broadcast/broadcast! over a pair of (input) sparse matrices with new (JuliaLang#19372) semantics / less pessimistic allocation behavior and optimize a few inefficiently handled common cases. Also provide a broadcast/broadcast! implementation capable of (relatively) efficiently handling an arbitrary number of (input) sparse matrices. Integrate with the map/map! implementation, and touch up a few rough edges of the map/map! implementation. Test.

Author: Sacha0

base/sparse/linalg.jl

@@ -507,7 +507,7 @@ function norm(A::SparseMatrixCSC,p::Real=2)
             nA::Tsum = 0
             for j=1:n
                 colSum::Tsum = 0
-                for i = A.colptr[j]:A.colptr[j+1]-1
+                for i in nzrange(A, j)
                     colSum += abs(A.nzval[i])
                 end
                 nA = max(nA, colSum)
@@ -517,7 +517,7 @@ function norm(A::SparseMatrixCSC,p::Real=2)
             throw(ArgumentError("2-norm not yet implemented for sparse matrices. Try norm(Array(A)) or norm(A, p) where p=1 or Inf."))
         elseif p==Inf
             rowSum = zeros(Tsum,m)
-            for i=1:length(A.nzval)
+            for i in 1:nnz(A)
                 rowSum[A.rowval[i]] += abs(A.nzval[i])
             end
             return convert(Tnorm, maximum(rowSum))