Merge branch 'develop' into lwpck-756

library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp

@@ -17,6 +17,8 @@ void add_device_normalization_rank_2_1_f32_instances(
     add_device_operation_instances(instances,
                                    device_normalization_f32_generic_instance<Pass, 2, 1>{});
     add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 2, 1>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f32_instances<Pass, 2, 1>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp

@@ -17,6 +17,8 @@ void add_device_normalization_rank_4_3_f16_instances(
     add_device_operation_instances(instances,
                                    device_normalization_f16_generic_instance<Pass, 4, 3>{});
     add_device_operation_instances(instances, device_normalization_f16_instances<Pass, 4, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f16_instances<Pass, 4, 3>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp

@@ -17,6 +17,8 @@ void add_device_normalization_rank_4_3_f32_instances(
     add_device_operation_instances(instances,
                                    device_normalization_f32_generic_instance<Pass, 4, 3>{});
     add_device_operation_instances(instances, device_normalization_f32_instances<Pass, 4, 3>{});
+    add_device_operation_instances(instances,
+                                   device_normalization_splitk_f32_instances<Pass, 4, 3>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp

@@ -5,6 +5,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp"
 #include "ck/utility/data_type.hpp"
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -43,6 +44,32 @@ using device_normalization_f16_instances =
         // clang-format on
         >;
 
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_splitk_f16_instances =
+    // clang-format off
+    std::tuple <
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8>,
+        DeviceNormalizationSplitKImpl<F16, F16, F16, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8>
+        // clang-format on
+        >;
+
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_f16_generic_instance = std::tuple<
     // clang-format off
@@ -76,6 +103,32 @@ using device_normalization_f32_instances = std::tuple<
     // clang-format on
     >;
 
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_splitk_f32_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F32, F32, F32, F32, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_f32_generic_instance = std::tuple<
     // clang-format off
@@ -109,6 +162,32 @@ using device_normalization_f16_f32_f32_f16_instances = std::tuple<
     // clang-format on
     >;
 
+template <typename OutElementwise, index_t Rank, index_t Reduce>
+using device_normalization_splitk_f16_f32_f32_f16_instances = std::tuple<
+    // clang-format off
+        // XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorSize, BetaSrcVectorSize, YDstVectorSize>
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2>,   // irregular size
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 2, 8, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 4, 1, 4, 1, 4, 1, 4, 4>,
+        DeviceNormalizationSplitKImpl<F16, F32, F32, F32, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 4, 1, 4, 1, 4, 4>
+    // clang-format on
+    >;
+
 template <typename OutElementwise, index_t Rank, index_t Reduce>
 using device_normalization_f16_f32_f32_f16_generic_instance = std::tuple<
     // clang-format off

library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt

+set(DEVICE_POOL3D_FWD_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
+                                          device_max_pool3d_fwd_ndhwc_f16_instance.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_POOL3D_FWD_INSTANCES device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
+                                          device_max_pool3d_fwd_ndhwc_f32_instance.cpp)
+endif()
+add_instance_library(device_pool3d_fwd_instance ${DEVICE_POOL3D_FWD_INSTANCES})

library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp

@@ -11,7 +11,9 @@ namespace instance {
 static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
 
 void add_device_pool3d_fwd_ndhwc_f16_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F32, ReduceOpId, false>{});

library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp

@@ -11,7 +11,9 @@ namespace instance {
 static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
 
 void add_device_pool3d_fwd_ndhwc_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});

library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp

@@ -11,14 +11,18 @@ namespace instance {
 static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
 
 void add_device_pool3d_fwd_ndhwc_f16_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, false>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
 }
 
 void add_device_pool3d_fwd_ndhwc_index_f16_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, ReduceOpId, true>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F16, F16, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});

library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp

@@ -11,14 +11,18 @@ namespace instance {
 static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
 
 void add_device_pool3d_fwd_ndhwc_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, false>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, false>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
 }
 
 void add_device_pool3d_fwd_ndhwc_index_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, ReduceOpId, true>>>& instances)
+    std::vector<
+        std::unique_ptr<DevicePoolFwd<5, 3, F32, F32, I32, NDHWC, NDHWC, ReduceOpId, true>>>&
+        instances)
 {
     add_device_operation_instances(
         instances, device_pool3d_fwd_ndhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});

library/src/tensor_operation_instance/gpu/pool_fwd/pool_fwd_instance_common.hpp → library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp

@@ -18,21 +18,7 @@ namespace instance {
 using I32   = int32_t;
 using F16   = ck::half_t;
 using F32   = float;
-
-template <typename InDataType,
-          typename OutDataType,
-          typename IndexDataType,
-          typename ComputeDataType,
-          ReduceTensorOp ReduceOpId,
-          bool OutputIndex>
-using device_pool2d_fwd_nhwc_instances =
-    // clang-format off
-    std::tuple <
-        DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
-        DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
-        DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
-               // clang-format on
-               >;
+using NDHWC = ck::tensor_layout::convolution::NDHWC;
 
 template <typename InDataType,
           typename OutDataType,
@@ -43,9 +29,9 @@ template <typename InDataType,
 using device_pool3d_fwd_ndhwc_instances =
     // clang-format off
     std::tuple <
-        DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
-        DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
-        DevicePool3dFwd_Input_N_Di_Hi_Wi_C_Output_N_Do_Ho_Wo_C<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 1, 1, 1>,
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 2, 1, 2>,
+        DevicePool3dFwd_NDHWC_NDHWC<InDataType, OutDataType, IndexDataType, ComputeDataType, ReduceOpId, OutputIndex, 256, 256, 1, 4, 1, 4>
                // clang-format on
                >;
 

library/src/tensor_operation_instance/gpu/pool_fwd/CMakeLists.txt

-add_instance_library(device_pool_fwd_instance
-    device_avg_pool2d_fwd_nhwc_f16_instance.cpp
-    device_avg_pool2d_fwd_nhwc_f32_instance.cpp
-    device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
-    device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
-    device_max_pool2d_fwd_nhwc_f16_instance.cpp
-    device_max_pool2d_fwd_nhwc_f32_instance.cpp
-    device_max_pool3d_fwd_ndhwc_f16_instance.cpp
-    device_max_pool3d_fwd_ndhwc_f32_instance.cpp
-)

library/src/tensor_operation_instance/gpu/pool_fwd/device_avg_pool2d_fwd_nhwc_f32_instance.cpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "pool_fwd_instance_common.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
-
-void add_device_pool2d_fwd_nhwc_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, false>>>& instances)
-{
-    add_device_operation_instances(
-        instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
-}
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool2d_fwd_nhwc_f16_instance.cpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "pool_fwd_instance_common.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
-
-void add_device_pool2d_fwd_nhwc_f16_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, false>>>& instances)
-{
-    add_device_operation_instances(
-        instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F16, ReduceOpId, false>{});
-}
-
-void add_device_pool2d_fwd_nhwc_index_f16_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F16, F16, I32, ReduceOpId, true>>>& instances)
-{
-    add_device_operation_instances(
-        instances, device_pool2d_fwd_nhwc_instances<F16, F16, I32, F16, ReduceOpId, true>{});
-}
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/pool_fwd/device_max_pool2d_fwd_nhwc_f32_instance.cpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#include "pool_fwd_instance_common.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
-
-void add_device_pool2d_fwd_nhwc_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, false>>>& instances)
-{
-    add_device_operation_instances(
-        instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, false>{});
-}
-
-void add_device_pool2d_fwd_nhwc_index_f32_instances(
-    std::vector<std::unique_ptr<DevicePoolFwd<4, 2, F32, F32, I32, ReduceOpId, true>>>& instances)
-{
-    add_device_operation_instances(
-        instances, device_pool2d_fwd_nhwc_instances<F32, F32, I32, F32, ReduceOpId, true>{});
-}
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt

-set(CONV2D_PERLAYER_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
-)
-
-set(CONV2D_PERCHANNEL_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
-)
-
-set(CONV2D_BIAS_PERLAYER_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
-)
-
-set(CONV2D_BIAS_PERCHANNEL_QUANT_SRC
-    conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
-    conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
-)
+if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
 
+set(CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp)
+set(CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp)
+set(CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp)
+set(CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp)
 set(GEMM_QUANT_SRC
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
-    gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
     gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
 )
+if(DL_KERNELS)
+    list(APPEND CONV2D_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_BIAS_PERLAYER_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp)
+    list(APPEND CONV2D_BIAS_PERCHANNEL_QUANT_SRC conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp)
+    list(APPEND GEMM_QUANT_SRC
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
+        gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp)
+endif()
 
 add_instance_library(device_quantization_instance
     ${CONV2D_PERLAYER_QUANT_SRC}
@@ -36,3 +29,4 @@ add_instance_library(device_quantization_instance
     ${CONV2D_BIAS_PERCHANNEL_QUANT_SRC}
     ${GEMM_QUANT_SRC}
 )
+endif()
\ No newline at end of file

library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "conv2d_quantization_common.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 namespace ck {
 namespace tensor_operation {

library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt

-add_instance_library(device_softmax_instance
-    device_softmax_f16_f16_instance_rank3_reduce1.cpp
+set(DEVICE_SOFTMAX_INSTANCES)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f16_f16_instance_rank3_reduce1.cpp
     device_softmax_f16_f16_instance_rank3_reduce2.cpp
     device_softmax_f16_f16_instance_rank3_reduce3.cpp
     device_softmax_f16_f16_instance_rank4_reduce1.cpp
     device_softmax_f16_f16_instance_rank4_reduce2.cpp
     device_softmax_f16_f16_instance_rank4_reduce3.cpp
-    device_softmax_f16_f16_instance_rank4_reduce4.cpp
-    device_softmax_f32_f32_instance_rank3_reduce1.cpp
+    device_softmax_f16_f16_instance_rank4_reduce4.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    list(APPEND DEVICE_SOFTMAX_INSTANCES device_softmax_f32_f32_instance_rank3_reduce1.cpp
     device_softmax_f32_f32_instance_rank3_reduce2.cpp
     device_softmax_f32_f32_instance_rank3_reduce3.cpp
     device_softmax_f32_f32_instance_rank4_reduce1.cpp
     device_softmax_f32_f32_instance_rank4_reduce2.cpp
     device_softmax_f32_f32_instance_rank4_reduce3.cpp
-    device_softmax_f32_f32_instance_rank4_reduce4.cpp
-)
+    device_softmax_f32_f32_instance_rank4_reduce4.cpp)
+endif()
+add_instance_library(device_softmax_instance ${DEVICE_SOFTMAX_INSTANCES})

library/src/utility/device_memory.cpp

@@ -10,20 +10,57 @@ DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
     hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
 }
 
+void DeviceMem::Realloc(std::size_t mem_size)
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipFree(mpDeviceBuf));
+    }
+    mMemSize = mem_size;
+    hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
+}
+
 void* DeviceMem::GetDeviceBuffer() const { return mpDeviceBuf; }
 
 std::size_t DeviceMem::GetBufferSize() const { return mMemSize; }
 
 void DeviceMem::ToDevice(const void* p) const
 {
-    hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
+    if(mpDeviceBuf)
+    {
+        hip_check_error(
+            hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
+    }
+    else
+    {
+        throw std::runtime_error("ToDevice with an empty pointer");
+    }
 }
 
 void DeviceMem::FromDevice(void* p) const
 {
+    if(mpDeviceBuf)
+    {
         hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
+    }
+    else
+    {
+        throw std::runtime_error("FromDevice with an empty pointer");
+    }
 }
 
-void DeviceMem::SetZero() const { hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize)); }
+void DeviceMem::SetZero() const
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize));
+    }
+}
 
-DeviceMem::~DeviceMem() { hip_check_error(hipFree(mpDeviceBuf)); }
+DeviceMem::~DeviceMem()
+{
+    if(mpDeviceBuf)
+    {
+        hip_check_error(hipFree(mpDeviceBuf));
+    }
+}

profiler/README.md

@@ -141,3 +141,46 @@ avg_time: 0.768321
 tflops: 86.6679
 GB/s: 127.947
 ```
+
+## Profile grouped convolution backward weight kernels
+```bash
+# arg1: tensor operation (grouped_conv_bwd_weight: Grouped Convolution Backward Weight)
+# arg2: data type (0: Input fp32, Weight fp32, Output fp32
+#                  1: Input fp16, Weight fp16, Output fp16
+#                  2: Input bf16, Weight fp32, Output bf16)
+# arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, N, K, Ho, Wo]
+#                      1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, N, Ho, Wo, K]
+#                      2: Input[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Output[N, Ho, Wo, G, K]
+# arg4: verification (0: no, 1: yes)
+# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
+# arg6: print tensor value (0: no; 1: yes)
+# arg7: time kernel (0: no, 1: yes)
+# Following arguments (depending on number of spatial dims):
+#  Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
+#  G, N, K, C, 
+#  <filter spatial dimensions>, (ie Y, X for 2D)
+#  <input image spatial dimensions>, (ie Hi, Wi for 2D)
+#  <strides>, (ie Sy, Sx for 2D)
+#  <dilations>, (ie Dy, Dx for 2D)
+#  <left padding>, (ie LeftPy, LeftPx for 2D)
+#  <right padding>, (ie RightPy, RightPx for 2D)
+# SplitK
+
+ ################                   op   datatype  layout  verify  init  log  time  Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx  SplitK
+./bin/ckProfiler grouped_conv_bwd_weight          1       0       1     1    0     1      2 32 256 256 512  3  3  28  28   1   1   1   1       1       0        0        0       1
+
+ ```
+
+Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
+```
+input: dim 5, lengths {32, 512, 1024, 28, 28}, strides {411041792, 802816, 1, 28672, 1024}
+weight: dim 5, lengths {32, 512, 1024, 3, 3}, strides {4718592, 9216, 1, 3072, 1024}
+output: dim 5, lengths {32, 512, 512, 26, 26}, strides {177209344, 346112, 1, 13312, 512}
+....
+Best configuration parameters:
+name: DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
+avg_time: 68.5216
+tflops: 95.337
+GB/s: 69.2301
+```
+Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.

profiler/include/profiler/profile_gemm_multiply_add_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout>
+bool profile_gemm_multiply_add_impl(int do_verification,
+                                    int init_method,
+                                    bool /*do_log*/,
+                                    bool time_kernel,
+                                    int M,
+                                    int N,
+                                    int K,
+                                    int StrideA,
+                                    int StrideB,
+                                    int StrideD0,
+                                    int StrideD1,
+                                    int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = MultiplyAdd;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceGemmMultipleD<ALayout,
+                                                          BLayout,
+                                                          ck::Tuple<D0Layout, D1Layout>,
+                                                          ELayout,
+                                                          ADataType,
+                                                          BDataType,
+                                                          ck::Tuple<D0DataType, D1DataType>,
+                                                          EDataType,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
+                                       d1_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 2>{StrideD0, StrideD1},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck