External Interface (#304)

* add client example

* clean

* clean

* reorg

* clean up profiler

* reorg

* clea

* fix profiler

* function for getinstances

* update client example

* update client example

* update client example

* update

* update example

* update Jenkins file

* update cmake

* update Jenkins

library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp

@@ -67,9 +67,11 @@ using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_in
         >;
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
-    std::vector<
-        DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
-        instances)
+    std::vector<DeviceBatchedGemmReducePtr<PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           DInElementOps,
+                                           DOutElementOps>>& instances)
 {
     add_device_operation_instances(
         instances,

library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp

@@ -67,9 +67,11 @@ using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_in
         >;
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
-    std::vector<
-        DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
-        instances)
+    std::vector<DeviceBatchedGemmReducePtr<PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           DInElementOps,
+                                           DOutElementOps>>& instances)
 {
     add_device_operation_instances(
         instances,

library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp

@@ -64,9 +64,11 @@ using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_in
         >;
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
-    std::vector<
-        DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, DInElementOps, DOutElementOps>>&
-        instances)
+    std::vector<DeviceBatchedGemmReducePtr<PassThrough,
+                                           PassThrough,
+                                           PassThrough,
+                                           DInElementOps,
+                                           DOutElementOps>>& instances)
 {
     add_device_operation_instances(
         instances,

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

@@ -28,14 +28,6 @@ set(DEVICE_GEMM_INSTANCE_SOURCE
    device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp;
    device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp;
    device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp;
-   device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp;
-   device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp;
-   device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp;
-   device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp;
-   device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp;
-   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp;
-   device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp;
-   device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp;
    device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp;
    device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp;
    device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp;

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

+set(DEVICE_GEMM_SPLITK_INSTANCE_SOURCE
+   device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp;
+   device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp;
+   device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp;
+   device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp;
+   device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp;
+   device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp;
+   device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp;
+   device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp;
+)
+
+add_library(device_gemm_splitk_instance OBJECT ${DEVICE_GEMM_SPLITK_INSTANCE_SOURCE})
+
+target_compile_features(device_gemm_splitk_instance PUBLIC)
+set_target_properties(device_gemm_splitk_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp

@@ -46,7 +46,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp

@@ -46,7 +46,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp

@@ -46,7 +46,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp

@@ -83,7 +83,7 @@ using device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances = std::tuple<
 //     >;
 
 void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp

@@ -46,7 +46,7 @@ using device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp

@@ -46,7 +46,7 @@ using device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp

@@ -51,7 +51,7 @@ using device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances{});

library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp → library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp

@@ -51,7 +51,7 @@ using device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances = std::tuple<
     >;
 
 void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(
-    std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
+    std::vector<DeviceGemmSplitKPtr<PassThrough, PassThrough, PassThrough>>& instances)
 {
     add_device_operation_instances(instances,
                                    device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances{});

profiler/CMakeLists.txt

@@ -6,6 +6,7 @@ include_directories(BEFORE
 set(PROFILER_SOURCE
     src/profiler.cpp
     src/profile_gemm.cpp
+    src/profile_gemm_splitk.cpp
     src/profile_gemm_bias_2d.cpp
     src/profile_gemm_bias_relu.cpp
     src/profile_gemm_bias_relu_add.cpp
@@ -27,21 +28,22 @@ add_executable(ckProfiler ${PROFILER_SOURCE})
 
 target_link_libraries(ckProfiler PRIVATE host_tensor)
 target_link_libraries(ckProfiler PRIVATE conv_util)
-target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
-target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_splitk_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_bias2d_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_instance)
 target_link_libraries(ckProfiler PRIVATE device_gemm_bias_relu_add_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
 target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
+target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv1d_fwd_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv3d_fwd_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
 target_link_libraries(ckProfiler PRIVATE device_convnd_bwd_data_instance)
-target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
-target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
-target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
-target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
+target_link_libraries(ckProfiler PRIVATE device_reduce_instance)

profiler/include/profile_batched_gemm_impl.hpp

@@ -7,56 +7,17 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
+#include "ck/library/tensor_operation_instance/gpu/device_batched_gemm_instance.hpp"
+
 #include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/conv_util.hpp"
 #include "ck/library/host_tensor/device_memory.hpp"
 #include "ck/library/host_tensor/host_tensor.hpp"
 #include "ck/library/host_tensor/host_tensor_generator.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace device_batched_gemm_instance {
-
-using DeviceGemmNoOpPtr =
-    ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
-                                                ck::tensor_operation::element_wise::PassThrough,
-                                                ck::tensor_operation::element_wise::PassThrough>;
-
-void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-
-} // namespace device_batched_gemm_instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck
-
 namespace ck {
 namespace profiler {
 
@@ -103,27 +64,22 @@ bool profile_batched_gemm_impl(int do_verification,
         f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
     Tensor<CDataType> c_g_m_n_device_result(
         f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
-    std::unique_ptr<Tensor<float>> c_f32_g_m_n_host_result   = nullptr;
-    std::unique_ptr<Tensor<float>> c_f32_g_m_n_device_result = nullptr;
 
     std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
     std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
     std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
 
-    std::size_t num_thread = 1;
     switch(init_method)
     {
     case 0: break;
     case 1:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
-        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
         break;
     default:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
-        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
     }
-    // set zero to c_device_buf
-    c_g_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
 
     using AElementOp = ck::tensor_operation::element_wise::PassThrough;
     using BElementOp = ck::tensor_operation::element_wise::PassThrough;
@@ -135,40 +91,6 @@ bool profile_batched_gemm_impl(int do_verification,
 
     if(do_verification)
     {
-        if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
-                     is_same<BDataType, ck::bhalf_t>::value &&
-                     is_same<CDataType, ck::bhalf_t>::value)
-        {
-            Tensor<float> a_f32_g_m_k(
-                f_host_tensor_descriptor(BatchCount, M, K, StrideA, ALayout{}));
-            Tensor<float> b_f32_g_k_n(
-                f_host_tensor_descriptor(BatchCount, K, N, StrideB, BLayout{}));
-            c_f32_g_m_n_host_result = std::make_unique<Tensor<float>>(
-                f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
-            c_f32_g_m_n_device_result = std::make_unique<Tensor<float>>(
-                f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
-
-            bf16_to_f32_(a_g_m_k, a_f32_g_m_k);
-            bf16_to_f32_(b_g_k_n, b_f32_g_k_n);
-
-            using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
-                ReferenceBatchedGemm<float, float, float, AElementOp, BElementOp, CElementOp>;
-
-            auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
-            auto ref_invoker      = ref_batched_gemm.MakeInvoker();
-
-            auto ref_argument = ref_batched_gemm.MakeArgument(a_f32_g_m_k,
-                                                              b_f32_g_k_n,
-                                                              *c_f32_g_m_n_host_result,
-                                                              a_element_op,
-                                                              b_element_op,
-                                                              c_element_op);
-
-            ref_invoker.Run(ref_argument);
-        }
-        else
-        {
-
         using ReferenceBatchedGemmInstance =
             ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
                                                              BDataType,
@@ -185,7 +107,6 @@ bool profile_batched_gemm_impl(int do_verification,
 
         ref_invoker.Run(ref_argument);
     }
-    }
 
     DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
     DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
@@ -195,154 +116,30 @@ bool profile_batched_gemm_impl(int do_verification,
     b_device_buf.ToDevice(b_g_k_n.mData.data());
     c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
 
-    // add device GEMM instances
-    std::vector<ck::tensor_operation::device::device_batched_gemm_instance::DeviceGemmNoOpPtr>
-        gemm_ptrs;
-
-    if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&
-                 is_same<CDataType, half_t>::value)
-    {
-        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances(gemm_ptrs);
-        }
-    }
-    else if constexpr(is_same<ADataType, bhalf_t>::value && is_same<BDataType, bhalf_t>::value &&
-                      is_same<CDataType, bhalf_t>::value)
-    {
-        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instances(gemm_ptrs);
-        }
-    }
-    else if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
-                      is_same<CDataType, float>::value)
-    {
-        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instances(gemm_ptrs);
-        }
-    }
-    else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
-                      is_same<CDataType, int8_t>::value)
-    {
-        if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                     is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::RowMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(gemm_ptrs);
-        }
-        else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value &&
-                          is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
-        {
-            ck::tensor_operation::device::device_batched_gemm_instance::
-                add_device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instances(gemm_ptrs);
-        }
-    }
+    // add device op instances
+    const auto op_ptrs = ck::tensor_operation::device::device_batched_gemm_instance::
+        get_device_batched_gemm_instances<ADataType,
+                                          BDataType,
+                                          CDataType,
+                                          ALayout,
+                                          BLayout,
+                                          CLayout>();
 
-    if(gemm_ptrs.size() <= 0)
+    if(op_ptrs.size() <= 0)
     {
         throw std::runtime_error("wrong! no device GEMM instance found");
     }
 
-    std::string best_gemm_name;
+    std::string best_op_name;
     float best_ave_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
 
-    // profile device GEMM instances
-    for(auto& gemm_ptr : gemm_ptrs)
+    // profile device op instances
+    for(auto& op_ptr : op_ptrs)
     {
         auto argument_ptr =
-            gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                                         static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
                                         static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
                                         M,
@@ -356,11 +153,14 @@ bool profile_batched_gemm_impl(int do_verification,
                                         ck::tensor_operation::element_wise::PassThrough{},
                                         BatchCount);
 
-        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
 
-        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
-            std::string gemm_name = gemm_ptr->GetTypeString();
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
 
             float ave_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
@@ -376,11 +176,11 @@ bool profile_batched_gemm_impl(int do_verification,
             float gb_per_sec = num_btype / 1.E6 / ave_time;
 
             std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
-                      << " GB/s, " << gemm_name << std::endl;
+                      << " GB/s, " << op_name << std::endl;
 
             if(tflops > best_tflops)
             {
-                best_gemm_name  = gemm_name;
+                best_op_name    = op_name;
                 best_tflops     = tflops;
                 best_ave_time   = ave_time;
                 best_gb_per_sec = gb_per_sec;
@@ -390,20 +190,8 @@ bool profile_batched_gemm_impl(int do_verification,
             {
                 c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
 
-                if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
-                             is_same<BDataType, ck::bhalf_t>::value &&
-                             is_same<CDataType, ck::bhalf_t>::value)
-                {
-
-                    bf16_to_f32_(c_g_m_n_device_result, *c_f32_g_m_n_device_result);
-                    float err = check_error(*c_f32_g_m_n_host_result, *c_f32_g_m_n_device_result);
-                    pass      = pass && (err < 1E-6);
-                }
-                else
-                {
-                    float err = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
-                    pass      = pass && (err < 1E-6);
-                }
+                pass = pass &
+                       ck::utils::check_err(c_g_m_n_device_result.mData, c_g_m_n_host_result.mData);
 
                 if(do_log)
                 {
@@ -419,13 +207,12 @@ bool profile_batched_gemm_impl(int do_verification,
         }
         else
         {
-            std::cout << "this device GEMM instance does not support this GEMM problem"
-                      << std::endl;
+            std::cout << op_ptr->GetTypeString() << " 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_gemm_name << std::endl;
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
     return pass;
 }

profiler/include/profile_batched_gemm_reduce_impl.hpp

@@ -6,7 +6,7 @@
 #include "ck/ck.hpp"
 #include "ck/utility/reduction_operator.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_gemm_reduce.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_reduce.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
@@ -29,7 +29,7 @@ using Square         = ck::tensor_operation::element_wise::UnarySquare;
 using DInElementOps  = ck::Tuple<Identity, Square>;
 using DOutElementOps = ck::Tuple<Identity, Identity>;
 
-using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
+using DeviceBatchedGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceBatchedGemmReducePtr<
     ck::tensor_operation::element_wise::PassThrough,
     ck::tensor_operation::element_wise::PassThrough,
     ck::tensor_operation::element_wise::PassThrough,
@@ -37,16 +37,16 @@ using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePt
     DOutElementOps>;
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
-    std::vector<DeviceGemmReduceNoOpPtr>&);
+    std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
-    std::vector<DeviceGemmReduceNoOpPtr>&);
+    std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
-    std::vector<DeviceGemmReduceNoOpPtr>&);
+    std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
-    std::vector<DeviceGemmReduceNoOpPtr>&);
+    std::vector<DeviceBatchedGemmReduceNoOpPtr>&);
 
 } // namespace device_gemm_instance
 } // namespace device
@@ -204,7 +204,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
     b_device_buf.ToDevice(b_g_k_n.mData.data());
 
     // add device GEMM instances
-    std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmReduceNoOpPtr>
+    std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceBatchedGemmReduceNoOpPtr>
         gemm_ptrs;
 
     if constexpr(is_same<ADataType, half_t>::value && is_same<BDataType, half_t>::value &&

profiler/include/profile_convnd_fwd.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-namespace ck {
-namespace profiler {
-
-int profile_convnd_fwd(int argc, char* argv[]);
-
-} // namespace profiler
-} // namespace ck

profiler/include/profile_gemm_add_add_fastgelu_impl.hpp

@@ -9,6 +9,9 @@
 #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/device_gemm_add_add_fastgelu_instance.hpp"
+
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/host_tensor/device_memory.hpp"
 #include "ck/library/host_tensor/host_tensor.hpp"
@@ -16,31 +19,6 @@
 #include "ck/library/host_tensor/host_conv.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace device_gemm_instance {
-
-using DeviceGemmAddAddFastGeluPtr = ck::tensor_operation::device::DeviceGemmMultipleDPtr<
-    2,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::AddAddFastGelu>;
-
-void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
-    std::vector<DeviceGemmAddAddFastGeluPtr>&);
-void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
-    std::vector<DeviceGemmAddAddFastGeluPtr>&);
-void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
-    std::vector<DeviceGemmAddAddFastGeluPtr>&);
-void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
-    std::vector<DeviceGemmAddAddFastGeluPtr>&);
-
-} // namespace device_gemm_instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck
-
 namespace ck {
 namespace profiler {
 
@@ -55,7 +33,7 @@ template <typename ADataType,
           typename D0Layout,
           typename D1Layout,
           typename ELayout>
-int profile_gemm_add_add_fastgelu_impl(int do_verification,
+bool profile_gemm_add_add_fastgelu_impl(int do_verification,
                                         int init_method,
                                         bool /*do_log*/,
                                         bool time_kernel,
@@ -122,48 +100,21 @@ int profile_gemm_add_add_fastgelu_impl(int do_verification,
     const auto b_element_op   = BElementOp{};
     const auto cde_element_op = CDEElementOp{};
 
-    // add device GEMM instances
-    std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmAddAddFastGeluPtr>
-        device_op_ptrs;
-
-    if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
-                 is_same_v<EDataType, half_t>)
-    {
-        if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
-                     is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
-                     is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
-        {
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
-                    device_op_ptrs);
-        }
-        else if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
-                          is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
-                          is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
-        {
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
-                    device_op_ptrs);
-        }
-        else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
-                          is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
-                          is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
-        {
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
-                    device_op_ptrs);
-        }
-        else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
-                          is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
-                          is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
-        {
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
-                    device_op_ptrs);
-        }
-    }
+    // add device op instances
+    const auto op_ptrs = ck::tensor_operation::device::device_gemm_instance::
+        get_device_gemm_add_add_fastgelu_instances<ADataType,
+                                                   BDataType,
+                                                   AccDataType,
+                                                   D0DataType,
+                                                   D1DataType,
+                                                   EDataType,
+                                                   ALayout,
+                                                   BLayout,
+                                                   D0Layout,
+                                                   D1Layout,
+                                                   ELayout>();
 
-    std::cout << "found " << device_op_ptrs.size() << " instances" << std::endl;
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
 
     // run reference
     if(do_verification)
@@ -207,7 +158,7 @@ int profile_gemm_add_add_fastgelu_impl(int do_verification,
     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_device_op_name;
+    std::string best_op_name;
     float best_ave_time   = 0;
     float best_tflops     = 0;
     float best_gb_per_sec = 0;
@@ -215,14 +166,14 @@ int profile_gemm_add_add_fastgelu_impl(int do_verification,
     bool pass = true;
 
     // profile device operation instances
-    for(auto& device_op_ptr : device_op_ptrs)
+    for(auto& op_ptr : op_ptrs)
     {
-        auto argument_ptr = device_op_ptr->MakeArgumentPointer(
+        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()},
-            static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
+            e_device_buf.GetDeviceBuffer(),
             M,
             N,
             K,
@@ -234,11 +185,11 @@ int profile_gemm_add_add_fastgelu_impl(int do_verification,
             b_element_op,
             cde_element_op);
 
-        auto invoker_ptr = device_op_ptr->MakeInvokerPointer();
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
 
-        std::string device_op_name = device_op_ptr->GetTypeString();
+        std::string op_name = op_ptr->GetTypeString();
 
-        if(device_op_ptr->IsSupportedArgument(argument_ptr.get()))
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             // re-init E to zero before profiling a kernel
             e_device_buf.SetZero();
@@ -256,11 +207,11 @@ int profile_gemm_add_add_fastgelu_impl(int do_verification,
             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, " << device_op_name << std::endl;
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
 
             if(tflops > best_tflops)
             {
-                best_device_op_name = device_op_name;
+                best_op_name    = op_name;
                 best_tflops     = tflops;
                 best_ave_time   = ave_time;
                 best_gb_per_sec = gb_per_sec;
@@ -276,14 +227,14 @@ int profile_gemm_add_add_fastgelu_impl(int do_verification,
         }
         else
         {
-            std::cout << device_op_name << " does not support this problem" << std::endl;
+            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_device_op_name << std::endl;
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
 
-    return pass ? 0 : 1;
+    return pass;
 }
 
 } // namespace profiler

profiler/include/profile_gemm_splitk_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <iostream>
+#include <typeinfo>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/device_gemm_splitk_instance.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/host_tensor/device_memory.hpp"
+#include "ck/library/host_tensor/host_tensor.hpp"
+#include "ck/library/host_tensor/host_tensor_generator.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_gemm_splitk_impl(int do_verification,
+                              int init_method,
+                              bool do_log,
+                              bool time_kernel,
+                              int M,
+                              int N,
+                              int K,
+                              int StrideA,
+                              int StrideB,
+                              int StrideC,
+                              int KBatch)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, 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<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_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});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    c_device_buf.ToDevice(c_m_n_device_result.mData.data());
+
+    // add device op instances
+    const auto op_ptrs =
+        ck::tensor_operation::device::device_gemm_instance::get_device_gemm_splitk_instances<
+            ADataType,
+            BDataType,
+            CDataType,
+            ALayout,
+            BLayout,
+            CLayout>();
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device operation instance found");
+    }
+
+    // Run reference GEMM
+    if(do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
+
+        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_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+    }
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device GEMM instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+                                        static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
+                                        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                        M,
+                                        N,
+                                        K,
+                                        StrideA,
+                                        StrideB,
+                                        StrideC,
+                                        a_element_op,
+                                        b_element_op,
+                                        c_element_op,
+                                        KBatch);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init C to zero before profiling next kernel
+            c_device_buf.SetZero();
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            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(CDataType) * 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)
+            {
+                c_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+                pass =
+                    pass & ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
+                        << std::endl;
+                }
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+
+    if constexpr(is_same<CDataType, float>::value)
+    {
+        std::cout << "Best Perf for datatype = f32";
+    }
+    else if constexpr(is_same<CDataType, half_t>::value)
+    {
+        std::cout << "Best Perf for datatype = f16";
+    }
+    else if constexpr(is_same<CDataType, bhalf_t>::value)
+    {
+        std::cout << "Best Perf for datatype = bf16";
+    }
+    else if constexpr(is_same<CDataType, int8_t>::value)
+    {
+        std::cout << "Best Perf for datatype = int8";
+    }
+
+    if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " ALayout =  RowMajor";
+    }
+    else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " ALayout =  ColumnMajor";
+    }
+
+    if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
+    {
+        std::cout << " BLayout =  RowMajor";
+    }
+    else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
+    {
+        std::cout << " BLayout =  ColumnMajor";
+    }
+
+    std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
+              << " StrideB = " << StrideB << " StrideC = " << StrideC << " : " << best_ave_time
+              << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck