Add reduce mean and square mean

example/16_gemm_reduce/gemm_reduce_xdl_mean_squaremean_fp16.cpp

@@ -45,11 +45,13 @@ using D1ReduceOp  = ck::reduce::Add<ReduceAccDataType>;
 using DxsReduceOp = ck::Tuple<D0ReduceOp, D1ReduceOp>;
 
 using UnaryIdenticElementOp =
+    ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, false>;
+using UnaryDivElementOp =
     ck::tensor_operation::element_wise::UnaryIdentic<ReduceAccDataType, ReduceAccDataType, true>;
 using UnarySquareElementOp =
     ck::tensor_operation::element_wise::UnarySquare<ReduceAccDataType, ReduceAccDataType, false>;
 using DxsInElementOp  = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
-using DxsOutElementOp = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
+using DxsOutElementOp = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
 
 using DGlobalMemOp =
     ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,

library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp

@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using ReduceSum   = ck::reduce::Add<F32>;
 using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
 
+using Div            = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
 using Identity       = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
 using Square         = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
 using DInElementOps  = ck::Tuple<Identity, Square>;
-using DOutElementOps = ck::Tuple<Identity, Identity>;
+using DOutElementOps = ck::Tuple<Div, Div>;
 
 using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
                                                           ck::InMemoryDataOperationEnum::AtomicAdd>;

library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp

@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using ReduceSum   = ck::reduce::Add<F32>;
 using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
 
+using Div            = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
 using Identity       = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
 using Square         = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
 using DInElementOps  = ck::Tuple<Identity, Square>;
-using DOutElementOps = ck::Tuple<Identity, Identity>;
+using DOutElementOps = ck::Tuple<Div, Div>;
 
 using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
                                                           ck::InMemoryDataOperationEnum::AtomicAdd>;

library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp

@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using ReduceSum   = ck::reduce::Add<F32>;
 using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
 
+using Div            = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
 using Identity       = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
 using Square         = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
 using DInElementOps  = ck::Tuple<Identity, Square>;
-using DOutElementOps = ck::Tuple<Identity, Identity>;
+using DOutElementOps = ck::Tuple<Div, Div>;
 
 using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
                                                           ck::InMemoryDataOperationEnum::AtomicAdd>;

library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp

@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using ReduceSum   = ck::reduce::Add<F32>;
 using ReduceOps   = ck::Tuple<ReduceSum, ReduceSum>;
 
+using Div            = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
 using Identity       = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
 using Square         = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
 using DInElementOps  = ck::Tuple<Identity, Square>;
-using DOutElementOps = ck::Tuple<Identity, Identity>;
+using DOutElementOps = ck::Tuple<Div, Div>;
 
 using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
                                                           ck::InMemoryDataOperationEnum::AtomicAdd>;

profiler/include/profile_gemm_reduce_impl.hpp

@@ -19,6 +19,7 @@ namespace device_gemm_instance {
 using F32            = float;
 using F16            = ck::half_t;
 using DPtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div            = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
 using Identity       = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
 using Square         = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
 using DInElementOps  = ck::Tuple<Identity, Square>;
@@ -122,11 +123,12 @@ bool profile_gemm_reduce_impl(int do_verification,
         b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
     }
 
-    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
-    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
-    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
-    using D0ReduceOp = ck::reduce::Add<float>;
-    using D1ReduceOp = ck::reduce::Add<float>;
+    using AElementOp        = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp        = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp        = ck::tensor_operation::element_wise::PassThrough;
+    using D0ReduceOp        = ck::reduce::Add<float>;
+    using D1ReduceOp        = ck::reduce::Add<float>;
+    using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
     using UnaryIdenticElementOp =
         ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
     using UnarySquareElementOp =
@@ -138,7 +140,7 @@ bool profile_gemm_reduce_impl(int do_verification,
     const auto b_element_op       = BElementOp{};
     const auto c_element_op       = CElementOp{};
     const auto dxs_in_element_op  = DxsInElementOps{};
-    const auto dxs_out_element_op = DxsOutElementOps{};
+    const auto dxs_out_element_op = DxsOutElementOp{M, M};
     const auto d0_reduce_op       = D0ReduceOp{};
     const auto d1_reduce_op       = D1ReduceOp{};
 
@@ -170,6 +172,8 @@ bool profile_gemm_reduce_impl(int do_verification,
                 d1_reduce_op(d1_acc, d1_val);
             }
 
+            dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
+            dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
             d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
             d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
         }