Add perf grouping (#841)

Co-authored-by: mvermeulen <5479696+mvermeulen@users.noreply.github.com>

src/program.cpp

@@ -484,6 +484,14 @@ double common_average(const std::vector<double>& v)
     return total / std::distance(v.begin() + n, v.end() - n);
 }
 
+std::string perf_group(const operation& op)
+{
+    auto attr = op.attributes();
+    if(attr.contains("group"))
+        return attr.at("group").to<std::string>();
+    return op.name();
+}
+
 void program::perf_report(std::ostream& os, std::size_t n, parameter_map params) const
 {
     auto& ctx = this->impl->ctx;
@@ -538,7 +546,7 @@ void program::perf_report(std::ostream& os, std::size_t n, parameter_map params)
     for(auto&& p : ins_vec)
     {
         double avg = common_average(p.second);
-        op_times[p.first->name()] += avg;
+        op_times[perf_group(p.first->get_operator())] += avg;
         total_instruction_time += avg;
     }
     double calculate_overhead_time    = total_time - total_instruction_time;

src/targets/cpu/binary.cpp

@@ -14,6 +14,8 @@ struct dnnl_binary : dnnl_op<dnnl_binary, dnnl::binary>
         return pack_join(self.reflect_base(self, f), pack(f(self.algo, "algo")));
     }
 
+    std::string group() const { return this->name() + "::" + algo; }
+
     std::string name() const { return "dnnl::binary"; }
 
     shape compute_shape(std::vector<shape> inputs) const

src/targets/cpu/eltwise.cpp

@@ -17,6 +17,8 @@ struct dnnl_eltwise : dnnl_op<dnnl_eltwise, dnnl::eltwise_forward>
                          pack(f(self.algo, "algo"), f(self.alpha, "alpha"), f(self.beta, "beta")));
     }
 
+    std::string group() const { return this->name() + "::" + algo; }
+
     std::string name() const { return "dnnl::eltwise"; }
 
     shape compute_shape(std::vector<shape> inputs) const

src/targets/cpu/include/migraphx/cpu/dnnl.hpp

@@ -74,6 +74,25 @@ struct dnnl_op : auto_register_op<Derived>
         return reflect_base(self, f);
     }
 
+    std::string group() const
+    {
+        const auto& self = static_cast<const Derived&>(*this);
+        return self.name();
+    }
+
+    value attributes() const
+    {
+        std::vector<std::string> names;
+        std::transform(post_ops.begin(), post_ops.end(), std::back_inserter(names), [](auto&& op) {
+            return op.algo;
+        });
+        const auto& self = static_cast<const Derived&>(*this);
+        auto g           = self.group();
+        if(not names.empty())
+            g += "<" + join_strings(names, ",") + ">";
+        return {{"group", g}};
+    }
+
     std::size_t get_extra_post_op_args() const
     {
         return std::count_if(post_ops.begin(), post_ops.end(), [](const auto& po) {