Sparse*Dense to Sparse Contraction yields wrong results

[ryanlevy]

The following minimal code seems to not produce correct results:

vector<int64_t> res_sizes = {1,3,2,2,8};
vector<int64_t> R_sizes   = {3,8,3};
vector<int64_t> ans_sizes = {1,2,2,3};
CTF::Tensor<> res(5,res_sizes.data()); res.fill_sp_random(0.,1.,0.25);
CTF::Tensor<> R(3,R_sizes.data()); R.fill_sp_random(0.,1.,0.25);
{
      perr<<"Dense = Dense*Dense"<<endl;
      CTF::Tensor<> ans(4,ans_sizes.data()); ans.set_name("Dense=DenseDense");
      ans["acdf"] = res["abcde"]*R["feb"];
      ans.print();
}
{
      perr<<"Sparse = Dense*Dense"<<endl;
      CTF::Tensor<> ans(4,true,ans_sizes.data()); ans.set_name("Sparse=DenseDense");
      ans["acdf"] = res["abcde"]*R["feb"];
      ans.print();
}
{
      perr<<"Sparse = Sparse*Dense"<<endl;
      res.sparsify();
      CTF::Tensor<> ans(4,true,ans_sizes.data()); ans.set_name("Sparse=SparseDense");
      ans["acdf"] = res["abcde"]*R["feb"];
      ans.print();
}

And for my particular set of random numbers

Dense = Dense*Dense
...
[0][0][0][2](8, <8.59462E-01>)
[0][1][0][2](9, <5.53183E-03>)
[0][1][1][2](11, <5.09581E-01>)
Sparse = Dense*Dense
...
[0][0][0][2](8, <8.59462E-01>)
[0][1][0][2](9, <5.53183E-03>)
[0][0][1][2](10, <0.00000E+00>)
[0][1][1][2](11, <5.09581E-01>)
Sparse=SparseDense
...
[0][0][0][2](8, <3.51847E-01>)
[0][1][0][2](9, <5.53183E-03>)
[0][1][1][2](11, <4.72762E-01>)

Where elements 8 and 11 are incorrect