Add the logic to tighten-up the inputs.

lib/nnc/ccv_cnnp_model_gradient_checkpointing.c

@@ -201,6 +201,31 @@ void ccv_cnnp_model_apply_gradient_checkpoints(ccv_cnnp_compiled_data_t* const c
 		}
 		if (input_execs->rnum <= 0 || output_execs->rnum <= 0)
 			continue;
+		ccv_nnc_graph_visit_t* const reverse_visit = ccv_nnc_graph_visit_new(graph, reversed_nodes, exec_rnum, (ccv_nnc_graph_exec_symbol_t*)ccv_array_get(output_execs, 0), output_execs->rnum, (ccv_nnc_graph_exec_symbol_t*)ccv_array_get(input_execs, 0), input_execs->rnum, 1);
+		ccv_nnc_graph_visit_for(reverse_visit, exec_info, node, idx) {
+			if (idx < exec_rnum && !CCV_NNC_GRAPH_EXEC_IS_DEAD(node->flags))
+				maskbit[idx >> 5] |= (1u << (idx & 0x1f));
+		} ccv_nnc_graph_visit_endfor
+		// Check if any of the items in input_execs is not marked, if it is not, there is no connection from it to output_execs, no need to visit.
+		for (j = 0; j < input_execs->rnum;)
+		{
+			const int idx = ((ccv_nnc_tensor_symbol_t*)ccv_array_get(input_execs, j))->d;
+			if (maskbit[idx >> 5] & (1u << (idx & 0x1f)))
+			{
+				++j;
+				continue;
+			}
+			// Not marked, remove this one.
+			if (j < input_execs->rnum - 1)
+				*(ccv_nnc_tensor_symbol_t*)ccv_array_get(input_execs, j) = *(ccv_nnc_tensor_symbol_t*)ccv_array_get(input_execs, input_execs->rnum - 1);
+			--input_execs->rnum;
+		}
+		// Reset maskbit back.
+		ccv_nnc_graph_visit_for(reverse_visit, exec_info, node, idx) {
+			if (idx < exec_rnum && !CCV_NNC_GRAPH_EXEC_IS_DEAD(node->flags))
+				maskbit[idx >> 5] &= ~(1u << (idx & 0x1f));
+		} ccv_nnc_graph_visit_endfor
+		ccv_nnc_graph_visit_free(reverse_visit);
 		// Fill in blanks (i.e. the backward ops that are not showing in above, but should be included to avoid excluding necessary ones). This is done by flowing gradients from outputs back all the way to inputs.
 		ccv_array_clear(input_gradient_execs);
 		ccv_array_clear(output_gradient_execs);