Incomplete sensitivities for 1D+ output_variables (#52)

* Iterate correctly over sensitivities for multidimensional variables

* Use reference rather than writing directly to yS

* Cache matrix properties

* use size_t for nnz_out constants

* Add comments

---------

Co-authored-by: Agriya Khetarpal <[email protected]>

Author: pipliggins

src/pybammsolvers/idaklu_source/IDAKLUSolverOpenMP.inl

@@ -855,32 +855,69 @@ void IDAKLUSolverOpenMP<ExprSet>::SetStepOutputSensitivities(
   sunrealtype &tval,
   sunrealtype *y_val,
   const vector<sunrealtype*>& yS_val,
-    int &i_save
+  int &i_save
   ) {
   DEBUG("IDAKLUSolver::SetStepOutputSensitivities");
-  // Calculate sensitivities
-  vector<sunrealtype> dens_dvar_dp = vector<sunrealtype>(number_of_parameters, 0);
-  for (size_t dvar_k=0; dvar_k<functions->dvar_dy_fcns.size(); dvar_k++) {
-    // Isolate functions
+
+  // Running index over the flattened outputs
+  size_t global_out_idx = 0;
+
+  // Loop over each variable
+  for (size_t dvar_k = 0; dvar_k < functions->var_fcns.size(); ++dvar_k) {
     Expression* dvar_dy = functions->dvar_dy_fcns[dvar_k];
     Expression* dvar_dp = functions->dvar_dp_fcns[dvar_k];
+
     // Calculate dvar/dy
     (*dvar_dy)({&tval, y_val, functions->inputs.data()}, {&res_dvar_dy[0]});
-    // Calculate dvar/dp and convert to dense array for indexing
+    // Calculate dvar/dp
     (*dvar_dp)({&tval, y_val, functions->inputs.data()}, {&res_dvar_dp[0]});
-    for (int k=0; k<number_of_parameters; k++) {
-      dens_dvar_dp[k]=0;
-    }
-    for (int k=0; k<dvar_dp->nnz_out(); k++) {
-      dens_dvar_dp[dvar_dp->get_row()[k]] = res_dvar_dp[k];
-    }
-    // Calculate sensitivities
-    for (int paramk=0; paramk<number_of_parameters; paramk++) {
-      auto &yS_back_paramk = yS[i_save][paramk];
-      yS_back_paramk[dvar_k] = dens_dvar_dp[paramk];
 
-      for (int spk=0; spk<dvar_dy->nnz_out(); spk++) {
-        yS_back_paramk[dvar_k] += res_dvar_dy[spk] * yS_val[paramk][dvar_dy->get_col()[spk]];
+    // Get number of output components for this function (e.g., scalar → 1; vector → >1)
+    const size_t n_rows = functions->var_fcns[dvar_k]->nnz_out();
+
+    // Number of nonzeros in the sparse Jacobians (for dvar/dy and dvar/dp)
+    const size_t dvar_dy_nnz = dvar_dy->nnz_out();
+    const size_t dvar_dp_nnz = dvar_dp->nnz_out();
+
+    // Row/column indices of nonzero entries (compressed sparse row format)
+    const auto& dvar_dy_row = dvar_dy->get_row();  // output component (row) for y
+    const auto& dvar_dy_col = dvar_dy->get_col();  // state variable index (column) for y
+    const auto& dvar_dp_row = dvar_dp->get_row();  // output component (row) for p
+    const auto& dvar_dp_col = dvar_dp->get_col();  // parameter index (column) for p
+
+    // Temporary dense vector to hold doutput_row/dp_k for each parameter
+    vector<sunrealtype> dvar_dp_dense(number_of_parameters, 0.0);
+
+    // Loop over each scalar component (row) of the output function
+    for (size_t row = 0; row < n_rows; ++row, ++global_out_idx) {
+      // Dense dvar_row/dp_k vector (reset to zero)
+      std::fill(dvar_dp_dense.begin(), dvar_dp_dense.end(), 0.0);
+
+      // Fill in dvar_row/dp_k from sparse structure
+      for (size_t nz = 0; nz < dvar_dp_nnz; ++nz) {
+        if (dvar_dp_row[nz] == static_cast<int>(row)) {
+          // dvar_dp_col[nz] is parameter index k
+          dvar_dp_dense[dvar_dp_col[nz]] = res_dvar_dp[nz];  // direct derivative
+        }
+      }
+
+      // For each parameter p_k, compute total d(output_row)/d(p_k)
+      for (int paramk = 0; paramk < number_of_parameters; paramk++) {
+        auto &yS_back_paramk = yS[i_save][paramk];  // Sensitivity vector for p_k at save step i_save
+
+        // Start with direct contribution doutput/dp_k
+        sunrealtype sens = dvar_dp_dense[paramk];
+
+        // Add chain rule term
+        for (size_t nz = 0; nz < dvar_dy_nnz; ++nz) {
+          if (dvar_dy_row[nz] == static_cast<int>(row)) {
+            // dvar_dy_col[nz] = j (state index)
+            // yS_val[paramk][j] = dy_j/dp_k
+            sens += res_dvar_dy[nz] * yS_val[paramk][dvar_dy_col[nz]];
+          }
+        }
+
+        yS_back_paramk[global_out_idx] = sens;
       }
     }
   }