[ROCm] Specialized binary elementwise broadcast kernel for mixed dtypes with float/bfloat16/half (#2795)

cherry-pick of pytorch#167233

Fixes #SWDEV-551924

Author: jerrymannil

aten/src/ATen/native/cuda/CUDALoops.cuh

@@ -893,6 +893,69 @@ struct type_specialized_kernel_launcher {
   }
 };
 
+template <int arg_index>
+struct type_specialized_broadcast_kernel_launcher {
+  template <
+      typename func_t,
+      typename array_t,
+      typename dtypes_t,
+      typename calc_t>
+  static void apply(
+      int64_t numel,
+      func_t f,
+      array_t data,
+      dtypes_t dtypes,
+      calc_t offset_calc) {
+        using traits = function_traits<func_t>;
+        using ret_t = typename traits::result_type;
+        using arg0_t = typename traits::template arg<0>::type;
+        using arg1_t = typename traits::template arg<1>::type;
+        if (dtypes[0] == rt_binary_specializations[arg_index][0] &&
+          dtypes[1] == rt_binary_specializations[arg_index][1] &&
+          dtypes[2] == rt_binary_specializations[arg_index][2]) {
+            using ret_cpp_t = c10::impl::ScalarTypeToCPPTypeT<rt_binary_specializations[arg_index][0]>;
+            using arg0_cpp_t = c10::impl::ScalarTypeToCPPTypeT<rt_binary_specializations[arg_index][1]>;
+            using arg1_cpp_t = c10::impl::ScalarTypeToCPPTypeT<rt_binary_specializations[arg_index][2]>;
+            constexpr int grp_sz = 128;
+            launch_legacy_kernel_manual_unroll<grp_sz, 4>(numel, [=] GPU_LAMBDA(int idx, bool unrl) {
+              if (unrl) {
+                auto offsets0 = offset_calc.get(idx);
+                auto offsets1 = offset_calc.get(idx + grp_sz);
+                auto offsets2 = offset_calc.get(idx + grp_sz * 2);
+                auto offsets3 = offset_calc.get(idx + grp_sz * 3);
+                void* out0 = data[0] + offsets0[0];
+                void* out1 = data[0] + offsets1[0];
+                void* out2 = data[0] + offsets2[0];
+                void* out3 = data[0] + offsets3[0];
+                auto u = c10::load<arg0_cpp_t>(data[1] + offsets0[1]);
+                auto v = c10::load<arg1_cpp_t>(data[2] + offsets0[2]);
+                ret_t result0 = f(c10::convert<arg0_t>(u), c10::convert<arg1_t>(v));
+                auto u1 = c10::load<arg0_cpp_t>(data[1] + offsets1[1]);
+                auto v1 = c10::load<arg1_cpp_t>(data[2]+ offsets1[2]);
+                ret_t result1 = f(c10::convert<arg0_t>(u1), c10::convert<arg1_t>(v1));
+                auto u2 = c10::load<arg0_cpp_t>(data[1] + offsets2[1]);
+                auto v2 = c10::load<arg1_cpp_t>(data[2] + offsets2[2]);
+                ret_t result2 = f(c10::convert<arg0_t>(u2), c10::convert<arg1_t>(v2));
+                auto u3 = c10::load<arg0_cpp_t>(data[1] + offsets3[1]);
+                auto v3 = c10::load<arg1_cpp_t>(data[2] + offsets3[2]);
+                ret_t result3 = f(c10::convert<arg0_t>(u3), c10::convert<arg1_t>(v3));
+                *(ret_cpp_t*)out0 = c10::convert<ret_cpp_t>(result0);
+                *(ret_cpp_t*)out1 = c10::convert<ret_cpp_t>(result1);
+                *(ret_cpp_t*)out2 = c10::convert<ret_cpp_t>(result2);
+                *(ret_cpp_t*)out3 = c10::convert<ret_cpp_t>(result3);
+              } else {
+                auto offsets = offset_calc.get(idx);
+                void* out = data[0] + offsets[0];
+                auto u = c10::load<arg0_cpp_t>(data[1] + offsets[1]);
+                auto v = c10::load<arg1_cpp_t>(data[2] + offsets[2]);
+                ret_t result = f(c10::convert<arg0_t>(u), c10::convert<arg1_t>(v));
+                *(ret_cpp_t*)out = c10::convert<ret_cpp_t>(result);
+              }
+            });
+        }
+      }
+};
+
 } // namespace
 #endif
 
@@ -995,6 +1058,32 @@ void gpu_kernel_impl(TensorIteratorBase& iter, const func_t& f) {
     }
     auto offset_calc = ::make_offset_calculator<traits::arity + 1>(iter);
 #ifdef USE_ROCM
+    if (check_binary_rt_types_for_specialization(iter)) {
+      // constexpr to reduce the amount of kernels generated for
+      // broadcast elementwise with mexed dtypes and limit which functors are actually
+      // applied to the load and store at compile time.
+      using func_tuple = typename traits::ArgsTuple;
+      if constexpr (
+        std::is_same_v<float, arg0_t> && traits::arity == 2 &&
+        check_binary_functor_types_for_specialization<
+          func_tuple,
+          float,
+          float,
+          traits::arity,
+          /*arg_num=*/0>::check()) {
+            memory::detail::static_unroll<
+              type_specialized_broadcast_kernel_launcher,
+              rt_binary_specializations.size()>::with_args(
+                numel,
+                f,
+                data,
+                dtypes,
+                offset_calc
+            );
+            return;
+      }
+    }
+
     constexpr int grp_sz = 128;
     launch_legacy_kernel_manual_unroll<grp_sz, 4>(numel, [=] GPU_LAMBDA(int idx, bool unrl) {
       if (unrl) {