Only apply to bs>1; wip fuse dot add

src/simplify_algebra.cpp

@@ -802,6 +802,9 @@ struct find_conv_dot_horiz_fusion
             int axis        = 1;
             int concat_axis = 0;
             if(name == "dot")
+            {
+                std::cout << ins->get_shape().lens().front() << std::endl;
+                if(ins->get_shape().lens().front() > 1)
                 {
                     axis        = int(args.front()->get_shape().lens().size() - 1);
                     concat_axis = axis - 1;
@@ -822,8 +825,15 @@ struct find_conv_dot_horiz_fusion
                         input);
                     auto fused     = m.insert_instruction(std::next(reshape), op, reshape, concat);
                     int64_t offset = 0;
+                    std::vector<instruction_ref> add_args;
+                    instruction_ref next_ins;
                     for(auto arg : range(start, last))
                     {
+                        auto aarg = std::next(std::next(arg));
+                        while(aarg->name() != "add")
+                            aarg = std::next(aarg);
+                        aarg->debug_print();
+                        add_args.push_back(aarg->inputs().front());
                         fused = m.insert_instruction(
                             std::next(fused),
                             make_op("reshape",
@@ -837,9 +847,16 @@ struct find_conv_dot_horiz_fusion
                             fused);
                         offset += len;
                     }
+                    m.debug_print();
 
                     return;
                 }
+                else
+                {
+                    axis        = int(args.front()->get_shape().lens().size() - 1);
+                    concat_axis = axis;
+                }
+            }
 
             for(auto arg : args)
                 m.move_instructions(arg, input);