Some fixes

src/fuse_reduce.cpp

@@ -31,6 +31,7 @@
 #include <migraphx/ranges.hpp>
 #include <migraphx/check_shapes.hpp>
 #include <migraphx/matcher.hpp>
+#include <migraphx/register_op.hpp>
 #include <iterator>
 
 namespace migraphx {
@@ -49,24 +50,27 @@ struct fused_reduce
     shape compute_shape(const std::vector<shape>& inputs, std::vector<module_ref> mods) const
     {
         if(mods.size() != 1)
-        {
             MIGRAPHX_THROW("should have one submodule.");
-        }
         auto* sm = mods.front();
+        if(sm->get_output_shapes().size() != 1)
+            MIGRAPHX_THROW("Only one output supported");
         check_shapes{inputs, *this}.has(sm->get_parameter_shapes().size()).same_dims();
         auto s    = inputs.at(0);
         auto lens = s.lens();
-        for(const auto& axis : axes)
+        if (lens != sm->get_output_shapes().front().lens())
         {
-            lens[axis] = 1;
+            for(const auto& axis : axes)
+            {
+                lens[axis] = 1;
+            }
         }
-        if(sm->get_output_shapes().size() != 1)
-            MIGRAPHX_THROW("Only one output supported");
-        return inputs[0].with_lens(sm->get_output_shapes().front().type(), lens);
+
+        return shape::from_permutation(sm->get_output_shapes().front().type(), lens, find_permutation(inputs));
     }
 
     std::string name() const { return "fused_reduce"; }
 };
+MIGRAPHX_REGISTER_OP(fused_reduce);
 
 static void create_reduce_modules(module_pass_manager& mpm)
 {
@@ -87,8 +91,8 @@ static void create_reduce_modules(module_pass_manager& mpm)
         auto r  = rm->add_instruction(ins->get_operator(), x0);
         rm->add_return({r});
 
-        // TODO: Set axes
-        mpm.get_module().replace_instruction(ins, make_op("fused_reduce"), ins->inputs(), {rm});
+        auto v = ins->get_operator().to_value();
+        mpm.get_module().replace_instruction(ins, make_op("fused_reduce", {{"axes", v["axes"]}}), ins->inputs(), {rm});
     }
 }
 
@@ -130,10 +134,11 @@ struct find_reduce_pointwise
         auto ins    = r.result;
         auto reduce = r.instructions["reduce"];
 
-        auto* old_rm = reduce->module_inputs().front();
+        const auto* old_rm = reduce->module_inputs().front();
         auto* rm     = mpm.create_module(old_rm->name() + ":pointwise");
-        // Copy module
-        *rm             = *old_rm;
+        rm->set_bypass();
+        // Copy module instructions
+        rm->add_instructions(old_rm);
         auto map_ins    = get_ins_param_map(reduce->inputs(), rm);
         auto new_inputs = reduce->inputs();
         for(auto input : ins->inputs())
@@ -152,8 +157,8 @@ struct find_reduce_pointwise
             }
         }
 
-        auto out = rm->insert_instructions(std::prev(rm->end()), {ins}, map_ins);
-        rm->replace_return(out);
+        auto out = rm->add_instructions({ins}, map_ins);
+        rm->add_return(out);
         mpm.get_module().replace_instruction(ins, reduce->get_operator(), new_inputs, {rm});
     }
 };

src/targets/gpu/prefuse_ops.cpp

@@ -116,7 +116,7 @@ struct find_add_layernorm
 
 void prefuse_ops::apply(module& m) const
 {
-    match::find_matches(m, find_add_layernorm{}, find_layernorm{});
+    // match::find_matches(m, find_add_layernorm{}, find_layernorm{});
 }
 
 } // namespace gpu