Merge remote-tracking branch 'origin/develop' into rosenrodt/gemm-layernorm

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{});

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

+# device_normalization_instance
+set(DEVICE_NORMALIZATION_INSTANCE_SOURCE
+    device_softmax_f32_f32_instance.cpp
+    device_softmax_f16_f16_instance.cpp
+)
+
+add_library(device_normalization_instance OBJECT ${DEVICE_NORMALIZATION_INSTANCE_SOURCE})
+set_target_properties(device_normalization_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+clang_tidy_check(device_normalization_instance)

library/src/tensor_operation_instance/gpu/normalization/device_softmax_f16_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_normalization_instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f16_f16_instances = std::tuple<
+    // clang-format off
+        // InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 8, 32, 1, 8, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 4, 64, 1, 8, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 8, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 16, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 2, 128, 1, 32, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 8>,
+        DeviceSoftmax<F16, F32, F16, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 8>
+    // clang-format on
+    >;
+
+void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<3, 2>{});
+}
+
+void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 1>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 2>{});
+    add_device_operation_instances(instances, device_softmax_f16_f16_instances<4, 3>{});
+}
+
+} // namespace device_normalization_instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/normalization/device_softmax_f32_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
+#include "ck/utility/data_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_normalization_instance {
+
+using F32 = float;
+
+template <index_t Rank, index_t Reduce>
+using device_softmax_f32_f32_instances = std::tuple<
+    // clang-format off
+        // InDataType, AccDataType, OutDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, InSrcVectorDim, InSrcVectorSize, OutDstVectorSize>
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 1, 1>, // fallback kernel
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 8, 32, 1, 8, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 4, 64, 1, 8, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 8, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 16, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 2, 128, 1, 32, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 8, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 16, 1, 4, 4>,
+        DeviceSoftmax<F32, F32, F32, Rank, Reduce, 256, 1, 256, 1, 32, 1, 4, 4>
+    // clang-format on
+    >;
+
+void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<3, 2>{});
+}
+
+void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 1>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 2>{});
+    add_device_operation_instances(instances, device_softmax_f32_f32_instances<4, 3>{});
+}
+
+} // namespace device_normalization_instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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
@@ -21,27 +22,30 @@ set(PROFILER_SOURCE
     src/profile_conv_bwd_weight.cpp
     src/profile_batched_gemm_reduce.cpp
     src/profile_gemm_add_add_fastgelu.cpp
+    src/profile_normalization.cpp
 )
 
 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_normalization_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;
@@ -135,56 +91,21 @@ 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,
+                                                             CDataType,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             CElementOp>;
 
-            using ReferenceBatchedGemmInstance =
-                ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
-                                                                 BDataType,
-                                                                 CDataType,
-                                                                 AElementOp,
-                                                                 BElementOp,
-                                                                 CElementOp>;
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
 
-            auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
-            auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
 
-            auto ref_argument = ref_batched_gemm.MakeArgument(
-                a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
-
-            ref_invoker.Run(ref_argument);
-        }
+        ref_invoker.Run(ref_argument);
     }
 
     DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
@@ -195,172 +116,51 @@ 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;
+    // 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 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);
-        }
-    }
-
-    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()),
-                                          static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                          static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                          M,
-                                          N,
-                                          K,
-                                          StrideA,
-                                          StrideB,
-                                          StrideC,
-                                          ck::tensor_operation::element_wise::PassThrough{},
-                                          ck::tensor_operation::element_wise::PassThrough{},
-                                          ck::tensor_operation::element_wise::PassThrough{},
-                                          BatchCount);
-
-        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
-
-        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+            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,
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        BatchCount);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        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

@@ -21,20 +21,16 @@ namespace tensor_operation {
 namespace device {
 namespace device_gemm_instance {
 
-using F32            = float;
-using F16            = ck::half_t;
-using DPtrsGlobal    = ck::Tuple<F32*, F32*>;
-using Identity       = ck::tensor_operation::element_wise::PassThrough;
-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<
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    DInElementOps,
-    DOutElementOps>;
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Identity, Identity>;
+
+using DeviceGemmReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
 
 void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
     std::vector<DeviceGemmReduceNoOpPtr>&);
@@ -59,7 +55,7 @@ namespace profiler {
 template <typename ADataType,
           typename BDataType,
           typename CDataType,
-          typename DDataType,
+          typename ReduceDataType,
           typename ALayout,
           typename BLayout,
           typename CLayout>
@@ -99,16 +95,16 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
 
     Tensor<CDataType> c_g_m_n_host_result(
         f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
-    Tensor<DDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
+    Tensor<ReduceDataType> d0_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
         {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
+    Tensor<ReduceDataType> d1_g_m_host_result(HostTensorDescriptor(std::vector<std::size_t>(
         {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
 
     Tensor<CDataType> c_g_m_n_device_result(
         f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
-    Tensor<DDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
+    Tensor<ReduceDataType> d0_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
         {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
+    Tensor<ReduceDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
         {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
 
     std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
@@ -135,20 +131,23 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
     using AElementOp            = ck::tensor_operation::element_wise::PassThrough;
     using BElementOp            = ck::tensor_operation::element_wise::PassThrough;
     using CElementOp            = ck::tensor_operation::element_wise::PassThrough;
-    using D0ReduceOp            = ck::reduce::Add;
-    using D1ReduceOp            = ck::reduce::Add;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
     using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
     using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
-    using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
-    using DxsOutElementOps      = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
 
-    const auto a_element_op       = AElementOp{};
-    const auto b_element_op       = BElementOp{};
-    const auto c_element_op       = CElementOp{};
-    const auto dxs_in_element_op  = DxsInElementOps{};
-    const auto dxs_out_element_op = DxsOutElementOps{};
-    const auto d0_reduce_op       = D0ReduceOp{};
-    const auto d1_reduce_op       = D1ReduceOp{};
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
+
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&passthrough, &passthrough};
 
     if(do_verification)
     {
@@ -160,6 +159,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
                                                              BElementOp,
                                                              CElementOp>;
 
+        using ReduceAccDataType = ReduceDataType;
+
         auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
         auto ref_invoker      = ref_batched_gemm.MakeInvoker();
 
@@ -172,21 +173,22 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
         {
             for(int m = 0; m < M; ++m)
             {
-                float d0_acc = d0_reduce_op.GetIdentityValue<float>();
-                float d1_acc = d1_reduce_op.GetIdentityValue<float>();
+                auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+                auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
 
                 for(int n = 0; n < N; ++n)
                 {
-                    float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
-                    float d1_val;
+                    ReduceAccDataType d0_val =
+                        ck::type_convert<ReduceAccDataType>(c_g_m_n_host_result(batch, m, n));
+                    ReduceAccDataType d1_val;
 
-                    UnarySquareElementOp{}(d1_val, d0_val);
-                    d0_reduce_op(d0_acc, d0_val);
-                    d1_reduce_op(d1_acc, d1_val);
+                    square(d1_val, d0_val);
+                    reduce0_op(reduce0_acc, d0_val);
+                    reduce1_op(reduce1_acc, d1_val);
                 }
 
-                d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc);
-                d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc);
+                d0_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+                d1_g_m_host_result(batch, m) = ck::type_convert<ReduceDataType>(reduce1_acc);
             }
         }
     }
@@ -194,11 +196,13 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
     DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
     DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
     DeviceMem c_device_buf(sizeof(CDataType) * c_g_m_n_device_result.mDesc.GetElementSpace());
-    DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
-    DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 d0_g_m_device_result.mDesc.GetElementSpace());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 d1_g_m_device_result.mDesc.GetElementSpace());
 
-    auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
 
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
@@ -257,31 +261,32 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
     // profile device GEMM instances
     for(auto& gemm_ptr : gemm_ptrs)
     {
-        auto argument_ptr =
-            gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
-                                          static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                          static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                          &dxs_global,
-                                          M,
-                                          N,
-                                          K,
-                                          StrideA,
-                                          StrideB,
-                                          StrideC,
-                                          a_element_op,
-                                          b_element_op,
-                                          c_element_op,
-                                          dxs_in_element_op,
-                                          dxs_out_element_op,
-                                          BatchCount);
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          {},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {},
+                                                          gemm_element_ops,
+                                                          {},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops,
+                                                          BatchCount);
 
         auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
 
         if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             // init DO, D1 to 0
-            d0_device_buf.SetZero();
-            d1_device_buf.SetZero();
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
 
             float ave_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
@@ -311,8 +316,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
             if(do_verification)
             {
                 c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
-                d0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
-                d1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
+                reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
 
                 float c_error  = check_error(c_g_m_n_host_result, c_g_m_n_device_result);
                 float d0_error = check_error(d0_g_m_host_result, d0_g_m_device_result);

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 {
 
@@ -55,18 +33,18 @@ template <typename ADataType,
           typename D0Layout,
           typename D1Layout,
           typename ELayout>
-int profile_gemm_add_add_fastgelu_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)
+bool profile_gemm_add_add_fastgelu_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) {
@@ -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);
-        }
-    }
-
-    std::cout << "found " << device_op_ptrs.size() << " instances" << std::endl;
+    // 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 " << 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();
@@ -256,14 +207,14 @@ 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_tflops         = tflops;
-                best_ave_time       = ave_time;
-                best_gb_per_sec     = gb_per_sec;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
             }
 
             if(do_verification)
@@ -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_bias_add_reduce_impl.hpp

@@ -21,33 +21,28 @@ namespace tensor_operation {
 namespace device {
 namespace device_gemm_instance {
 
-using F32            = float;
-using F16            = ck::half_t;
-using DPtrsGlobal    = ck::Tuple<F32*, F32*>;
-using Div            = ck::tensor_operation::element_wise::UnaryDivide;
-using Identity       = ck::tensor_operation::element_wise::PassThrough;
-using Square         = ck::tensor_operation::element_wise::UnarySquare;
-using DInElementOps  = ck::Tuple<Identity, Square>;
-using DOutElementOps = ck::Tuple<Div, Div>;
-
-using DeviceGemmBiasAddReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmBiasAddReducePtr<
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    DInElementOps,
-    DOutElementOps>;
-
-void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmBiasAddReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<1, ReducePtrsGlobal::Size()>;
+
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
     std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
 
-void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
     std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
 
-void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
     std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
 
-void add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+void add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
     std::vector<DeviceGemmBiasAddReduceNoOpPtr>&);
 
 } // namespace device_gemm_instance
@@ -61,9 +56,9 @@ namespace profiler {
 template <typename ADataType,
           typename BDataType,
           typename CDataType,
-          typename C0DataType,
-          typename C1DataType,
-          typename DDataType,
+          typename BiasDataType,
+          typename D0DataType,
+          typename ReduceDataType,
           typename ALayout,
           typename BLayout,
           typename CLayout>
@@ -77,7 +72,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                                        int StrideA,
                                        int StrideB,
                                        int StrideC,
-                                       int StrideC1)
+                                       int StrideD0)
 {
     auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
         return HostTensorDescriptor(std::vector<std::size_t>({len}),
@@ -102,24 +97,24 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
     Tensor<BDataType> b_k_n(f_host_tensor_descriptor2d(K, N, StrideB, BLayout{}));
 
     Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
-    Tensor<C0DataType> bias_n(f_host_tensor_descriptor1d(N, 1));
-    Tensor<C1DataType> c1_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
-    Tensor<DDataType> d0_m_host_result(
+    Tensor<BiasDataType> bias_n(f_host_tensor_descriptor1d(N, 1));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
+    Tensor<ReduceDataType> reduce0_m_host_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_m_host_result(
+    Tensor<ReduceDataType> reduce1_m_host_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
 
     Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
-    Tensor<DDataType> d0_m_device_result(
+    Tensor<ReduceDataType> reduce0_m_device_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_m_device_result(
+    Tensor<ReduceDataType> reduce1_m_device_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
 
     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_host_result.mDesc << std::endl;
-    std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
-    std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
+    std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
+    std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
 
     std::size_t num_thread = 1;
     switch(init_method)
@@ -130,50 +125,53 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
         a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
         b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
         bias_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
-        c1_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
         break;
     default:
         std::srand(0);
         a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
         b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
         bias_n.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5}, num_thread);
-        c1_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
     }
 
     using PassThrough           = ck::tensor_operation::element_wise::PassThrough;
     using AElementOp            = PassThrough;
     using BElementOp            = PassThrough;
     using CElementOp            = PassThrough;
-    using C1ElementOp           = PassThrough;
-    using D0ReduceOp            = ck::reduce::Add;
-    using D1ReduceOp            = ck::reduce::Add;
+    using D0ElementOp           = PassThrough;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
     using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
     using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
     using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
-    using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
-    using DxsOutElementOps      = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
 
-    const auto a_element_op  = AElementOp{};
-    const auto b_element_op  = BElementOp{};
-    const auto c_element_op  = CElementOp{};
-    const auto c1_element_op = C1ElementOp{};
-    const auto d0_reduce_op  = D0ReduceOp{};
-    const auto d1_reduce_op  = D1ReduceOp{};
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
 
-    auto dxs_in_element_op  = DxsInElementOps{};
-    auto dxs_out_element_op = DxsOutElementOps{N, N};
+    auto d0_element_op    = D0ElementOp{};
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
+
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    auto div                                    = UnaryDivElementOp{N};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&div, &div};
 
     if(do_verification)
     {
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                                 BDataType,
                                                                                 CDataType,
-                                                                                DDataType,
+                                                                                ReduceDataType,
                                                                                 AElementOp,
                                                                                 BElementOp,
                                                                                 CElementOp>;
 
-        using ReduceAccDataType = DDataType;
+        using ReduceAccDataType = ReduceDataType;
 
         auto ref_gemm    = ReferenceGemmInstance{};
         auto ref_invoker = ref_gemm.MakeInvoker();
@@ -189,53 +187,53 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                 ReduceAccDataType acc = static_cast<ReduceAccDataType>(c_m_n_host_result(m, n)) +
                                         static_cast<ReduceAccDataType>(bias_n(n));
 
-                ReduceAccDataType c1 = static_cast<ReduceAccDataType>(c1_m_n(m, n));
+                ReduceAccDataType d0 = static_cast<ReduceAccDataType>(d0_m_n(m, n));
                 c_element_op(acc, acc);
-                c1_element_op(c1, c1);
-                acc += c1;
+                d0_element_op(d0, d0);
+                acc += d0;
                 c_m_n_host_result(m, n) = static_cast<CDataType>(acc);
             }
 
         for(int m = 0; m < M; ++m)
         {
-            auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
-            auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
 
             for(int n = 0; n < N; ++n)
             {
-                ReduceAccDataType c_val =
+                ReduceAccDataType d0_val =
                     ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
-                ReduceAccDataType d0_val;
                 ReduceAccDataType d1_val;
 
-                dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
-                dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
-                d0_reduce_op(d0_acc, d0_val);
-                d1_reduce_op(d1_acc, d1_val);
+                square(d1_val, d0_val);
+                reduce0_op(reduce0_acc, d0_val);
+                reduce1_op(reduce1_acc, d1_val);
             }
 
-            dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
-            dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
-            d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
-            d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
+            div(reduce0_acc, reduce0_acc);
+            div(reduce1_acc, reduce1_acc);
+            reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+            reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
         }
     }
 
     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());
-    DeviceMem bias_device_buf(sizeof(C0DataType) * bias_n.mDesc.GetElementSpace());
-    DeviceMem c1_device_buf(sizeof(C1DataType) * c1_m_n.mDesc.GetElementSpace());
-    DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
-    DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
+    DeviceMem bias_device_buf(sizeof(BiasDataType) * bias_n.mDesc.GetElementSpace());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 reduce0_m_device_result.mDesc.GetElementSpace());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 reduce1_m_device_result.mDesc.GetElementSpace());
 
-    auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());
     bias_device_buf.ToDevice(bias_n.mData.data());
-    c1_device_buf.ToDevice(c1_m_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
 
     // add device GEMM instances
     std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmBiasAddReduceNoOpPtr>
@@ -249,7 +247,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                      is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
         {
             ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_kn_mn_instances(
                     gemm_ptrs);
         }
         else if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value &&
@@ -257,7 +255,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                           is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
         {
             ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_mk_nk_mn_instances(
                     gemm_ptrs);
         }
         else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
@@ -265,7 +263,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                           is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
         {
             ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_kn_mn_instances(
                     gemm_ptrs);
         }
         else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value &&
@@ -273,7 +271,7 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                           is_same<CLayout, tensor_layout::gemm::RowMajor>::value)
         {
             ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_bias_add_reduce_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
+                add_device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f16_f16_f32_f32_km_nk_mn_instances(
                     gemm_ptrs);
         }
     }
@@ -291,34 +289,31 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
     // profile device GEMM instances
     for(auto& gemm_ptr : gemm_ptrs)
     {
-        auto argument_ptr = gemm_ptr->MakeArgumentPointer(
-            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
-            static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-            static_cast<C0DataType*>(bias_device_buf.GetDeviceBuffer()),
-            static_cast<C1DataType*>(c1_device_buf.GetDeviceBuffer()),
-            &dxs_global,
-            M,
-            N,
-            K,
-            StrideA,
-            StrideB,
-            StrideC,
-            StrideC1,
-            a_element_op,
-            b_element_op,
-            c_element_op,
-            c1_element_op,
-            dxs_in_element_op,
-            dxs_out_element_op);
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          bias_device_buf.GetDeviceBuffer(),
+                                                          {d0_device_buf.GetDeviceBuffer()},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {StrideD0},
+                                                          gemm_element_ops,
+                                                          {&d0_element_op},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops);
 
         auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
 
         if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             // init DO, D1 to 0
-            d0_device_buf.SetZero();
-            d1_device_buf.SetZero();
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
 
             float ave_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
@@ -328,9 +323,9 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
             std::size_t flop = std::size_t(2) * M * N * K + std::size_t(2) * M * N;
 
             std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
-                                   sizeof(CDataType) * M * N + sizeof(C0DataType) * M * N +
-                                   sizeof(C1DataType) * M * N + sizeof(DDataType) * M +
-                                   sizeof(DDataType) * M;
+                                   sizeof(CDataType) * M * N + sizeof(BiasDataType) * M * N +
+                                   sizeof(D0DataType) * M * N + sizeof(ReduceDataType) * M +
+                                   sizeof(ReduceDataType) * M;
 
             float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
@@ -350,12 +345,12 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
             if(do_verification)
             {
                 c_device_buf.FromDevice(c_m_n_device_result.mData.data());
-                d0_device_buf.FromDevice(d0_m_device_result.mData.data());
-                d1_device_buf.FromDevice(d1_m_device_result.mData.data());
+                reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
 
                 ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
-                ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
-                ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
+                ck::utils::check_err(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
+                ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
 
                 if(do_log)
                 {
@@ -365,13 +360,17 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                         << std::endl;
                     LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
                         << std::endl;
                 }
             }

profiler/include/profile_gemm_impl.hpp

@@ -12,112 +12,37 @@
 #include "ck/tensor_operation/gpu/device/device_gemm.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
+#include "ck/library/tensor_operation_instance/gpu/device_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_gemm.hpp"
 
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace device_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_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(
-    std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-void add_device_gemm_dl_i8_i8_i8_mk_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_i8_i8_i8_mk_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_i8_i8_i8_km_kn_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-void add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(std::vector<DeviceGemmNoOpPtr>&);
-
-} // namespace device_gemm_instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck
-
 namespace ck {
 namespace profiler {
 
 template <typename ADataType,
           typename BDataType,
-          typename CDataType,
           typename AccDataType,
+          typename CDataType,
           typename ALayout,
           typename BLayout,
           typename CLayout>
-void profile_gemm_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)
+int profile_gemm_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)
 {
+    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)
@@ -134,32 +59,25 @@ void profile_gemm_impl(int do_verification,
 
     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;
 
-    std::size_t num_thread = 1;
     switch(init_method)
     {
-    // case 0: break;
-    case 0:
-        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{}, num_thread);
-        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{}, num_thread);
-        break;
+    case 0: break;
     case 1:
-        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
-        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+        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}, num_thread);
-        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
     }
 
-    // set zero to c_device_buf
-    c_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;
     using CElementOp = ck::tensor_operation::element_wise::PassThrough;
@@ -176,303 +94,65 @@ void profile_gemm_impl(int do_verification,
     b_device_buf.ToDevice(b_k_n.mData.data());
     c_device_buf.ToDevice(c_m_n_device_result.mData.data());
 
-    // add device GEMM instances
-    std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmNoOpPtr> gemm_ptrs;
+    // add device op instances
+    const auto op_ptrs = ck::tensor_operation::device::device_gemm_instance::
+        get_device_gemm_instances<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>();
 
-    if constexpr(is_same<ADataType, float>::value && is_same<BDataType, float>::value &&
-                 is_same<CDataType, float>::value)
+    if(op_ptrs.size() <= 0)
     {
-        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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f32_f32_f32_mk_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f32_f32_f32_mk_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instances(gemm_ptrs);
-            }
-        }
+        throw std::runtime_error("wrong! no device GEMM instance found");
     }
-    else 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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
 
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_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)
-        {
-            if(KBatch > 1)
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
-            }
-            else
-            {
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
-
-                ck::tensor_operation::device::device_gemm_instance::
-                    add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(gemm_ptrs);
-            }
-        }
-    }
-    else if constexpr(is_same<ADataType, ck::bhalf_t>::value &&
-                      is_same<BDataType, ck::bhalf_t>::value &&
-                      is_same<CDataType, ck::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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instances(gemm_ptrs);
-        }
-    }
-    else if constexpr(is_same<ADataType, int8_t>::value && is_same<BDataType, int8_t>::value &&
-                      is_same<CDataType, int8_t>::value)
+    // Run reference GEMM
+    if(do_verification)
     {
-        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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instances(gemm_ptrs);
-
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_dl_i8_i8_i8_mk_kn_mn_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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instances(gemm_ptrs);
-
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_dl_i8_i8_i8_mk_nk_mn_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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instances(gemm_ptrs);
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CElementOp>;
 
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_dl_i8_i8_i8_km_kn_mn_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_gemm_instance::
-                add_device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instances(gemm_ptrs);
+        auto ref_op      = ReferenceGemmInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
 
-            ck::tensor_operation::device::device_gemm_instance::
-                add_device_gemm_dl_i8_i8_i8_km_nk_mn_instances(gemm_ptrs);
-        }
-    }
+        auto ref_argument = ref_op.MakeArgument(
+            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
 
-    if(gemm_ptrs.size() <= 0)
-    {
-        throw std::runtime_error("wrong! no device GEMM instance found");
+        ref_invoker.Run(ref_argument);
     }
 
-    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)
+    for(auto& op_ptr : op_ptrs)
     {
         auto argument_ptr =
-            gemm_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,
-                                          ck::tensor_operation::element_wise::PassThrough{},
-                                          ck::tensor_operation::element_wise::PassThrough{},
-                                          ck::tensor_operation::element_wise::PassThrough{},
-                                          KBatch);
-
-        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
-
-        if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+            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,
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{},
+                                        ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             // re-init C to zero before profiling next kernel
-            c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<CDataType>{}, num_thread);
-            c_device_buf.ToDevice(c_m_n_device_result.mData.data());
+            c_device_buf.SetZero();
 
-            std::string gemm_name = gemm_ptr->GetTypeString();
+            std::string op_name = op_ptr->GetTypeString();
 
             float ave_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
@@ -487,11 +167,11 @@ void profile_gemm_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, " << gemm_name << std::endl;
+                      << gb_per_sec << " 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;
@@ -501,86 +181,15 @@ void profile_gemm_impl(int do_verification,
             {
                 c_device_buf.FromDevice(c_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)
-                {
-                    Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
-                    Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
-                    Tensor<float> c_m_n_host_result(
-                        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-                    Tensor<float> c_m_n_device_f32_result(
-                        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-
-                    bf16_to_f32_(a_m_k, a_f32_m_k);
-                    bf16_to_f32_(b_k_n, b_f32_k_n);
-                    bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
-
-                    using ReferenceGemmInstance =
-                        ck::tensor_operation::host::ReferenceGemm<float,
-                                                                  float,
-                                                                  float,
-                                                                  float,
-                                                                  AElementOp,
-                                                                  BElementOp,
-                                                                  CElementOp>;
-
-                    auto ref_gemm    = ReferenceGemmInstance{};
-                    auto ref_invoker = ref_gemm.MakeInvoker();
-
-                    auto ref_argument = ref_gemm.MakeArgument(a_f32_m_k,
-                                                              b_f32_k_n,
-                                                              c_m_n_host_result,
-                                                              a_element_op,
-                                                              b_element_op,
-                                                              c_element_op);
-
-                    ref_invoker.Run(ref_argument);
-
-                    ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData);
-
-                    if(do_log)
-                    {
-                        LogRangeAsType<float>(
-                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
-                            << std::endl;
-                    }
-                }
-                else
-                {
-                    Tensor<CDataType> c_m_n_host_result(
-                        f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-
-                    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);
-                    ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
-
-                    if(do_log)
-                    {
-                        LogRangeAsType<float>(
-                            std::cout << "c_host  : ", c_m_n_host_result.mData, ",")
-                            << std::endl;
-                    }
-                }
+                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;
                 }
@@ -588,8 +197,7 @@ void profile_gemm_impl(int do_verification,
         }
         else
         {
-            std::cout << gemm_ptr->GetTypeString() << " does not support this GEMM problem"
-                      << std::endl;
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
         }
     }
 
@@ -631,7 +239,9 @@ void profile_gemm_impl(int do_verification,
     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_gemm_name << std::endl;
+              << best_op_name << std::endl;
+
+    return pass ? 0 : 1;
 }
 
 } // namespace profiler

profiler/include/profile_gemm_reduce_impl.hpp

@@ -21,21 +21,17 @@ namespace tensor_operation {
 namespace device {
 namespace device_gemm_instance {
 
-using F32            = float;
-using F16            = ck::half_t;
-using DPtrsGlobal    = ck::Tuple<F32*, F32*>;
-using Div            = ck::tensor_operation::element_wise::UnaryDivide;
-using Identity       = ck::tensor_operation::element_wise::PassThrough;
-using Square         = ck::tensor_operation::element_wise::UnarySquare;
-using DInElementOps  = ck::Tuple<Identity, Square>;
-using DOutElementOps = ck::Tuple<Div, Div>;
-
-using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    ck::tensor_operation::element_wise::PassThrough,
-    DInElementOps,
-    DOutElementOps>;
+using F32                 = float;
+using F16                 = ck::half_t;
+using ReducePtrsGlobal    = ck::Tuple<F32*, F32*>;
+using Div                 = ck::tensor_operation::element_wise::UnaryDivide;
+using Identity            = ck::tensor_operation::element_wise::PassThrough;
+using Square              = ck::tensor_operation::element_wise::UnarySquare;
+using ReduceInElementOps  = ck::Tuple<Identity, Square>;
+using ReduceOutElementOps = ck::Tuple<Div, Div>;
+
+using DeviceGemmReduceNoOpPtr =
+    ck::tensor_operation::device::DeviceGemmReducePtr<0, ReducePtrsGlobal::Size()>;
 
 void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
     std::vector<DeviceGemmReduceNoOpPtr>&);
@@ -60,7 +56,7 @@ namespace profiler {
 template <typename ADataType,
           typename BDataType,
           typename CDataType,
-          typename DDataType,
+          typename ReduceDataType,
           typename ALayout,
           typename BLayout,
           typename CLayout>
@@ -95,22 +91,22 @@ bool profile_gemm_reduce_impl(int do_verification,
     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<DDataType> d0_m_host_result(
+    Tensor<ReduceDataType> reduce0_m_host_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_m_host_result(
+    Tensor<ReduceDataType> reduce1_m_host_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
 
     Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-    Tensor<DDataType> d0_m_device_result(
+    Tensor<ReduceDataType> reduce0_m_device_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<DDataType> d1_m_device_result(
+    Tensor<ReduceDataType> reduce1_m_device_result(
         HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
 
     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_host_result.mDesc << std::endl;
-    std::cout << "d0_m: " << d0_m_host_result.mDesc << std::endl;
-    std::cout << "d1_m: " << d1_m_host_result.mDesc << std::endl;
+    std::cout << "reduce0_m: " << reduce0_m_host_result.mDesc << std::endl;
+    std::cout << "reduce1_m: " << reduce1_m_host_result.mDesc << std::endl;
 
     std::size_t num_thread = 1;
     switch(init_method)
@@ -130,34 +126,37 @@ bool profile_gemm_reduce_impl(int do_verification,
     using AElementOp            = ck::tensor_operation::element_wise::PassThrough;
     using BElementOp            = ck::tensor_operation::element_wise::PassThrough;
     using CElementOp            = ck::tensor_operation::element_wise::PassThrough;
-    using D0ReduceOp            = ck::reduce::Add;
-    using D1ReduceOp            = ck::reduce::Add;
-    using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+    using ReduceOp0             = ck::reduce::Add;
+    using ReduceOp1             = ck::reduce::Add;
     using UnaryIdenticElementOp = ck::tensor_operation::element_wise::PassThrough;
     using UnarySquareElementOp  = ck::tensor_operation::element_wise::UnarySquare;
-    using DxsInElementOps       = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
-    using DxsOutElementOps      = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
+    using UnaryDivElementOp     = ck::tensor_operation::element_wise::UnaryDivide;
+
+    auto a_element_op                     = AElementOp{};
+    auto b_element_op                     = BElementOp{};
+    auto c_element_op                     = CElementOp{};
+    std::array<void*, 3> gemm_element_ops = {&a_element_op, &b_element_op, &c_element_op};
 
-    const auto a_element_op = AElementOp{};
-    const auto b_element_op = BElementOp{};
-    const auto c_element_op = CElementOp{};
-    const auto d0_reduce_op = D0ReduceOp{};
-    const auto d1_reduce_op = D1ReduceOp{};
+    const auto reduce0_op = ReduceOp0{};
+    const auto reduce1_op = ReduceOp1{};
 
-    auto dxs_in_element_op  = DxsInElementOps{};
-    auto dxs_out_element_op = DxsOutElementOps{N, N};
+    auto passthrough                            = UnaryIdenticElementOp{};
+    auto square                                 = UnarySquareElementOp{};
+    auto div                                    = UnaryDivElementOp{N};
+    std::array<void*, 2> reduce_in_element_ops  = {&passthrough, &square};
+    std::array<void*, 2> reduce_out_element_ops = {&div, &div};
 
     if(do_verification)
     {
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                                 BDataType,
                                                                                 CDataType,
-                                                                                DDataType,
+                                                                                ReduceDataType,
                                                                                 AElementOp,
                                                                                 BElementOp,
                                                                                 CElementOp>;
 
-        using ReduceAccDataType = DDataType;
+        using ReduceAccDataType = ReduceDataType;
 
         auto ref_gemm    = ReferenceGemmInstance{};
         auto ref_invoker = ref_gemm.MakeInvoker();
@@ -169,37 +168,37 @@ bool profile_gemm_reduce_impl(int do_verification,
 
         for(int m = 0; m < M; ++m)
         {
-            auto d0_acc = d0_reduce_op.GetIdentityValue<ReduceAccDataType>();
-            auto d1_acc = d1_reduce_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce0_acc = reduce0_op.GetIdentityValue<ReduceAccDataType>();
+            auto reduce1_acc = reduce1_op.GetIdentityValue<ReduceAccDataType>();
 
             for(int n = 0; n < N; ++n)
             {
-                ReduceAccDataType c_val =
+                ReduceAccDataType d0_val =
                     ck::type_convert<ReduceAccDataType>(c_m_n_host_result(m, n));
-                ReduceAccDataType d0_val;
                 ReduceAccDataType d1_val;
 
-                dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
-                dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
-                d0_reduce_op(d0_acc, d0_val);
-                d1_reduce_op(d1_acc, d1_val);
+                square(d1_val, d0_val);
+                reduce0_op(reduce0_acc, d0_val);
+                reduce1_op(reduce1_acc, d1_val);
             }
 
-            dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
-            dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
-            d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
-            d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
+            div(reduce0_acc, reduce0_acc);
+            div(reduce1_acc, reduce1_acc);
+            reduce0_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce0_acc);
+            reduce1_m_host_result(m) = ck::type_convert<ReduceDataType>(reduce1_acc);
         }
     }
 
     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());
-    DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
-    DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
+    DeviceMem reduce0_device_buf(sizeof(ReduceDataType) *
+                                 reduce0_m_device_result.mDesc.GetElementSpace());
+    DeviceMem reduce1_device_buf(sizeof(ReduceDataType) *
+                                 reduce1_m_device_result.mDesc.GetElementSpace());
 
-    auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
-                                     static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
+    std::array<void*, 2> p_reduces = {reduce0_device_buf.GetDeviceBuffer(),
+                                      reduce1_device_buf.GetDeviceBuffer()};
 
     a_device_buf.ToDevice(a_m_k.mData.data());
     b_device_buf.ToDevice(b_k_n.mData.data());
@@ -258,30 +257,31 @@ bool profile_gemm_reduce_impl(int do_verification,
     // profile device GEMM instances
     for(auto& gemm_ptr : gemm_ptrs)
     {
-        auto argument_ptr =
-            gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
-                                          static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                          static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
-                                          &dxs_global,
-                                          M,
-                                          N,
-                                          K,
-                                          StrideA,
-                                          StrideB,
-                                          StrideC,
-                                          a_element_op,
-                                          b_element_op,
-                                          c_element_op,
-                                          dxs_in_element_op,
-                                          dxs_out_element_op);
+        auto argument_ptr = gemm_ptr->MakeArgumentPointer(a_device_buf.GetDeviceBuffer(),
+                                                          b_device_buf.GetDeviceBuffer(),
+                                                          nullptr,
+                                                          {},
+                                                          c_device_buf.GetDeviceBuffer(),
+                                                          p_reduces,
+                                                          M,
+                                                          N,
+                                                          K,
+                                                          StrideA,
+                                                          StrideB,
+                                                          StrideC,
+                                                          {},
+                                                          gemm_element_ops,
+                                                          {},
+                                                          reduce_in_element_ops,
+                                                          reduce_out_element_ops);
 
         auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
 
         if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             // init DO, D1 to 0
-            d0_device_buf.SetZero();
-            d1_device_buf.SetZero();
+            reduce0_device_buf.SetZero();
+            reduce1_device_buf.SetZero();
 
             float ave_time =
                 invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
@@ -311,12 +311,12 @@ bool profile_gemm_reduce_impl(int do_verification,
             if(do_verification)
             {
                 c_device_buf.FromDevice(c_m_n_device_result.mData.data());
-                d0_device_buf.FromDevice(d0_m_device_result.mData.data());
-                d1_device_buf.FromDevice(d1_m_device_result.mData.data());
+                reduce0_device_buf.FromDevice(reduce0_m_device_result.mData.data());
+                reduce1_device_buf.FromDevice(reduce1_m_device_result.mData.data());
 
                 ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
-                ck::utils::check_err(d0_m_device_result.mData, d0_m_host_result.mData);
-                ck::utils::check_err(d1_m_device_result.mData, d1_m_host_result.mData);
+                ck::utils::check_err(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
+                ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
 
                 if(do_log)
                 {
@@ -326,13 +326,17 @@ bool profile_gemm_reduce_impl(int do_verification,
                         << std::endl;
                     LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d0_host: ", d0_m_host_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d0_host: ", reduce0_m_host_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d0_device: ", d0_m_device_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d0_device: ", reduce0_m_device_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d1_host: ", d1_m_host_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d1_host: ", reduce1_m_host_result.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "d1_device: ", d1_m_device_result.mData, ",")
+                    LogRangeAsType<float>(
+                        std::cout << "d1_device: ", reduce1_m_device_result.mData, ",")
                         << std::endl;
                 }
             }

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

profiler/include/profile_grouped_gemm_impl.hpp

@@ -232,6 +232,10 @@ void profile_grouped_gemm_impl(int do_verification,
 
         auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
 
+        DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
         if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
         {
             std::string gemm_name = gemm_ptr->GetTypeString();

profiler/include/profile_normalization_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_softmax.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_softmax.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_normalization_instance {
+
+void add_device_softmax_f16_f16_rank3_instances(std::vector<DeviceNormalizationPtr>&);
+void add_device_softmax_f16_f16_rank4_instances(std::vector<DeviceNormalizationPtr>&);
+
+void add_device_softmax_f32_f32_rank3_instances(std::vector<DeviceNormalizationPtr>&);
+void add_device_softmax_f32_f32_rank4_instances(std::vector<DeviceNormalizationPtr>&);
+
+} // namespace device_normalization_instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace ck {
+namespace profiler {
+
+enum struct NormType
+{
+    LAYERNORM,
+    BATCHNORM,
+    SOFTMAX,
+};
+
+enum struct NormDataType
+{
+    F32_F32, // in, out
+    F16_F16,
+    BF16_BF16,
+    INT8_INT8,
+};
+
+// clang-format off
+template <typename NormDataType> std::string type_to_string();
+template <> std::string type_to_string<float>()   { return "f32"; }
+template <> std::string type_to_string<half_t>()  { return "f16"; }
+template <> std::string type_to_string<bhalf_t>() { return "bf16"; }
+template <> std::string type_to_string<int8_t>()  { return "int8"; }
+template <> std::string type_to_string<int32_t>() { return "int32"; }
+// clang-format on
+
+template <typename InDataType, typename AccDataType, typename OutDataType>
+void profile_normalization_impl(int do_verification,
+                                int init_method,
+                                bool do_log,
+                                bool time_kernel,
+                                std::vector<index_t> in_length,
+                                std::vector<index_t> in_strides,
+                                std::vector<index_t> reduce_dims,
+                                AccDataType alpha,
+                                AccDataType beta,
+                                NormType norm_type)
+{
+    Tensor<InDataType> in = in_strides.empty() ? Tensor<InDataType>(in_length)
+                                               : Tensor<InDataType>(in_length, in_strides);
+    Tensor<OutDataType> out(in.mDesc);
+
+    switch(init_method)
+    {
+    // case 0: break;
+    case 0:
+        in.GenerateTensorValue(GeneratorTensor_1<InDataType>{});
+        out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{});
+        break;
+    case 1:
+        in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
+        out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
+        break;
+    default:
+        in.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
+        out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
+    }
+
+    Tensor<OutDataType> out_ref(out);
+
+    DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
+    DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
+    in_dev.ToDevice(in.mData.data());
+    out_dev.ToDevice(out.mData.data());
+
+    std::vector<index_t> i_in_lengths(in.mDesc.GetLengths().begin(), in.mDesc.GetLengths().end());
+    std::vector<index_t> i_in_strides(in.mDesc.GetStrides().begin(), in.mDesc.GetStrides().end());
+
+    // add device normalization instances
+    std::vector<tensor_operation::device::DeviceNormalizationPtr> instances;
+
+    if(norm_type == NormType::SOFTMAX)
+    {
+        if constexpr(is_same<InDataType, half_t>::value && is_same<OutDataType, half_t>::value &&
+                     is_same<AccDataType, float>::value)
+        {
+            if(in_length.size() == 3)
+                tensor_operation::device::device_normalization_instance::
+                    add_device_softmax_f16_f16_rank3_instances(instances);
+
+            if(in_length.size() == 4)
+                tensor_operation::device::device_normalization_instance::
+                    add_device_softmax_f16_f16_rank4_instances(instances);
+        }
+        else if constexpr(is_same<InDataType, float>::value && is_same<OutDataType, float>::value &&
+                          is_same<AccDataType, float>::value)
+        {
+            if(in_length.size() == 3)
+                tensor_operation::device::device_normalization_instance::
+                    add_device_softmax_f32_f32_rank3_instances(instances);
+
+            if(in_length.size() == 4)
+                tensor_operation::device::device_normalization_instance::
+                    add_device_softmax_f32_f32_rank4_instances(instances);
+        }
+    }
+
+    if(instances.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device normalization instance found");
+    }
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    for(auto& inst_ptr : instances)
+    {
+        // Is this user's responsibility to check if problem mismatches kernel instance (ie. rank 3
+        // problem to rank 4 kernel) other than invoking IsSupportedArgument()?
+        if(!(inst_ptr->GetRank() == static_cast<index_t>(i_in_lengths.size()) &&
+             inst_ptr->GetNumReduceDim() == static_cast<index_t>(reduce_dims.size())))
+        {
+            continue;
+        }
+
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(i_in_lengths,
+                                                          i_in_strides,
+                                                          reduce_dims,
+                                                          &alpha,
+                                                          &beta,
+                                                          in_dev.GetDeviceBuffer(),
+                                                          out_dev.GetDeviceBuffer());
+
+        if(!inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
+            LogRange(std::cout << "input lengths = [", in_length, ", ")
+                << "], "
+                << "scaler = [" << alpha << ", " << beta << "]." << std::endl;
+            return;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes =
+            in.mDesc.GetElementSize() * sizeof(InDataType) +
+            (beta == 0.0f ? 1 : 2) * out.mDesc.GetElementSize() * sizeof(OutDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                  << inst_ptr->GetTypeString() << std::endl;
+
+        if(avg_time < best_avg_time)
+        {
+            best_instance_name = inst_ptr->GetTypeString();
+            best_avg_time      = avg_time;
+            best_gb_per_sec    = gb_per_sec;
+        }
+
+        if(do_verification)
+        {
+            // TODO: factory method to dynamically switch between different reference normalizations
+            using ReferenceFactory =
+                tensor_operation::host::ReferenceSoftmax<InDataType, OutDataType, AccDataType>;
+
+            ReferenceFactory{}.MakeInvoker().Run({in, out_ref, alpha, beta, reduce_dims});
+
+            out_dev.FromDevice(out.mData.data());
+
+            bool pass;
+            if(std::is_same<InDataType, int8_t>::value)
+            {
+                pass = ck::utils::check_err(
+                    out.mData, out_ref.mData, "Error: Incorrect results!", 0, 1);
+                if(do_log)
+                {
+                    LogRangeAsType<int>(std::cout << "in  : ", in.mData, ",") << std::endl;
+                    LogRangeAsType<int>(std::cout << "out_ref  : ", out_ref.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<int>(std::cout << "out  : ", out.mData, ",") << std::endl;
+                }
+            }
+            else
+            {
+                pass = ck::utils::check_err(out.mData, out_ref.mData);
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "in  : ", in.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "out_ref  : ", out_ref.mData, ",")
+                        << std::endl;
+                    LogRangeAsType<float>(std::cout << "out  : ", out.mData, ",") << std::endl;
+                }
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "input lengths = [", in_length, ", ")
+                    << "], "
+                    << "scaler = [" << alpha << ", " << beta << "]." << std::endl;
+            }
+        }
+    }
+    std::cout << "Best Perf for datatype = " << type_to_string<InDataType>() << "_"
+              << type_to_string<OutDataType>() << ", ";
+    LogRange(std::cout << "length = ", i_in_lengths, ",") << ", ";
+    LogRange(std::cout << "stride = ", i_in_strides, ",") << ", ";
+    LogRange(std::cout << "reduce dims ", reduce_dims, ",") << ", ";
+    std::cout << "alpha = " << alpha << ", "
+              << "beta = " << beta << ", " << best_avg_time << " ms, " << best_gb_per_sec
+              << " GB/s, " << best_instance_name << std::endl;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/src/profile_batched_gemm.cpp

@@ -15,10 +15,6 @@ enum struct GemmMatrixLayout
     MK_NK_MN, // 1
     KM_KN_MN, // 2
     KM_NK_MN, // 3
-    MK_KN_NM, // 4
-    MK_NK_NM, // 5
-    KM_KN_NM, // 6
-    KM_NK_NM, // 7
 };
 
 enum struct GemmDataType
@@ -31,7 +27,7 @@ enum struct GemmDataType
 
 int profile_batched_gemm(int argc, char* argv[])
 {
-    if(!(argc == 15))
+    if(argc != 15)
     {
         printf("arg1: tensor operation (batched_gemm: Batched GEMM)\n");
         printf("arg2: data type (0: fp32; 1: fp16, 2: bf16, 3: int8)\n");
@@ -64,330 +60,117 @@ int profile_batched_gemm(int argc, char* argv[])
 
     const int BatchCount = std::stoi(argv[14]);
 
-    if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using INT8 = int8_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto c_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType = decltype(a_type);
+        using BDataType = decltype(b_type);
+        using CDataType = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::
+            profile_batched_gemm_impl<ADataType, BDataType, CDataType, ALayout, BLayout, CLayout>(
+                do_verification,
+                init_method,
+                do_log,
+                time_kernel,
+                M,
+                N,
+                K,
+                (StrideA < 0) ? DefaultStrideA : StrideA,
+                (StrideB < 0) ? DefaultStrideB : StrideB,
+                (StrideC < 0) ? DefaultStrideC : StrideC,
+                BatchCount);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::half_t,
-                                                ck::half_t,
-                                                ck::half_t,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F32{}, F32{}, F32{}, Row{}, Row{}, Row{});
     }
-    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::half_t,
-                                                ck::half_t,
-                                                ck::half_t,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F32{}, F32{}, F32{}, Row{}, Col{}, Row{});
     }
-    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::half_t,
-                                                ck::half_t,
-                                                ck::half_t,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F32{}, F32{}, F32{}, Col{}, Row{}, Row{});
     }
-    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
+    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::half_t,
-                                                ck::half_t,
-                                                ck::half_t,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F32{}, F32{}, F32{}, Col{}, Col{}, Row{});
     }
-    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F16{}, F16{}, F16{}, Row{}, Row{}, Row{});
     }
-    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F16{}, F16{}, F16{}, Row{}, Col{}, Row{});
     }
-    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F16{}, F16{}, F16{}, Col{}, Row{}, Row{});
     }
-    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
+    else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::bhalf_t,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(F16{}, F16{}, F16{}, Col{}, Col{}, Row{});
     }
-    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<float,
-                                                float,
-                                                float,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Row{}, Row{});
     }
-    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN)
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<float,
-                                                float,
-                                                float,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(BF16{}, BF16{}, BF16{}, Row{}, Col{}, Row{});
     }
-    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN)
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<float,
-                                                float,
-                                                float,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(BF16{}, BF16{}, BF16{}, Col{}, Row{}, Row{});
     }
-    else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN)
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::KM_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<float,
-                                                float,
-                                                float,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(BF16{}, BF16{}, BF16{}, Col{}, Col{}, Row{});
     }
     else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<int8_t,
-                                                int8_t,
-                                                int8_t,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Row{}, Row{});
     }
     else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<int8_t,
-                                                int8_t,
-                                                int8_t,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? K : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(INT8{}, INT8{}, INT8{}, Row{}, Col{}, Row{});
     }
     else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_KN_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<int8_t,
-                                                int8_t,
-                                                int8_t,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? N : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Row{}, Row{});
     }
     else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::KM_NK_MN)
     {
-        ck::profiler::profile_batched_gemm_impl<int8_t,
-                                                int8_t,
-                                                int8_t,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::ColumnMajor,
-                                                ck::tensor_layout::gemm::RowMajor>(
-            do_verification,
-            init_method,
-            do_log,
-            time_kernel,
-            M,
-            N,
-            K,
-            (StrideA < 0) ? M : StrideA,
-            (StrideB < 0) ? K : StrideB,
-            (StrideC < 0) ? N : StrideC,
-            BatchCount);
+        return profile(INT8{}, INT8{}, INT8{}, Col{}, Col{}, Row{});
     }
     else
     {
-        throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");
-    }
+        std::cout << "this data_type & layout is not implemented" << std::endl;
 
-    return 0;
+        return 1;
+    }
 }

profiler/src/profile_convnd_fwd.cpp

@@ -10,11 +10,10 @@
 
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
 #include "ck/library/utility/conv_util.hpp"
 #include "ck/library/utility/fill.hpp"
 
-#include "profiler/include/profile_convnd_fwd.hpp"
-
 namespace {
 
 enum struct ConvDataType
@@ -304,7 +303,7 @@ void profile_convnd_instances(ConvDataType data_type,
 
 } // namespace
 
-int ck::profiler::profile_convnd_fwd(int argc, char* argv[])
+int profile_convnd_fwd(int argc, char* argv[])
 {
     using namespace ck::utils::conv;