Merge branch 'develop' into gemm_layernorm_welford

library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce3.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_i8_i8_rank4_reduce3_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_rank4_reduce4.hpp"
+#include "ck/library/tensor_operation_instance/gpu/softmax/device_softmax_i8_i8_instance_type.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+static constexpr index_t RANK = 4;
+
+void add_device_softmax_i8_i8_rank4_reduce4_instances(
+    std::vector<DeviceSoftmaxPtr<I8, F32, I8, PassThrough, PassThrough, RANK>>& instances)
+{
+    add_device_operation_instances(instances, device_softmax_i8_i8_instances<RANK, 4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/CMakeLists.txt

@@ -20,12 +20,12 @@ set(PROFILER_SOURCE
     src/profile_conv_fwd_bias_relu.cpp
     src/profile_conv_fwd_bias_relu_add.cpp
     src/profile_conv_bwd_data.cpp
-    src/profile_conv_bwd_weight.cpp
     src/profile_grouped_conv_fwd.cpp
+    src/profile_grouped_conv_bwd_weight.cpp
     src/profile_reduce.cpp
     src/profile_groupnorm.cpp
     src/profile_layernorm.cpp
-    src/profile_normalization.cpp
+    src/profile_softmax.cpp
 )
 
 add_executable(ckProfiler ${PROFILER_SOURCE})
@@ -49,10 +49,13 @@ target_link_libraries(ckProfiler PRIVATE device_grouped_conv3d_fwd_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv1d_bwd_data_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
 target_link_libraries(ckProfiler PRIVATE device_conv3d_bwd_data_instance)
-target_link_libraries(ckProfiler PRIVATE device_conv1d_bwd_weight_instance)
-target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
-target_link_libraries(ckProfiler PRIVATE device_conv3d_bwd_weight_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv1d_bwd_weight_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv2d_bwd_weight_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv3d_bwd_weight_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_normalization_instance)
+target_link_libraries(ckProfiler PRIVATE device_softmax_instance)
 target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
+
+rocm_install(TARGETS ckProfiler COMPONENT profiler)

profiler/include/profile_batched_gemm_add_relu_gemm_add_impl.hpp

@@ -14,6 +14,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 
 namespace ck {
@@ -111,15 +112,15 @@ bool profile_batched_gemm_add_relu_gemm_add_impl(bool do_verification,
                                        std::size_t stride,
                                        std::size_t batch_stride,
                                        auto layout) {
+        using namespace ck::literals;
+
         if(std::is_same<decltype(layout), Row>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
         }
     };
 
@@ -330,8 +331,7 @@ bool profile_batched_gemm_add_relu_gemm_add_impl(bool do_verification,
             {
                 e1_g_m_o_device_buf.FromDevice(e1_g_m_o_device_result.mData.data());
 
-                pass = pass & ck::utils::check_err(e1_g_m_o_device_result.mData,
-                                                   e1_g_m_o_host_result.mData);
+                pass = pass & ck::utils::check_err(e1_g_m_o_device_result, e1_g_m_o_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_batched_gemm_gemm_impl.hpp

@@ -16,6 +16,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 
 namespace ck {
@@ -105,15 +106,15 @@ bool profile_batched_gemm_gemm_impl(bool do_verification,
                                        std::size_t stride,
                                        std::size_t batch_stride,
                                        auto layout) {
+        using namespace ck::literals;
+
         if(std::is_same<decltype(layout), Row>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
         }
     };
 
@@ -283,8 +284,7 @@ bool profile_batched_gemm_gemm_impl(bool do_verification,
             {
                 c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
 
-                pass = pass &
-                       ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData);
+                pass = pass & ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_batched_gemm_impl.hpp

@@ -16,6 +16,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 
 namespace ck {
@@ -50,15 +51,15 @@ bool profile_batched_gemm_impl(int do_verification,
                                        std::size_t stride,
                                        std::size_t batch_stride,
                                        auto layout) {
+        using namespace ck::literals;
+
         if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
         }
     };
 
@@ -202,8 +203,7 @@ bool profile_batched_gemm_impl(int do_verification,
             {
                 c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
 
-                pass = pass &
-                       ck::utils::check_err(c_g_m_n_device_result.mData, c_g_m_n_host_result.mData);
+                pass = pass & ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_batched_gemm_reduce_impl.hpp

@@ -14,6 +14,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 
 namespace ck {
@@ -78,15 +79,15 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
                                        std::size_t col,
                                        std::size_t stride,
                                        auto layout) {
+        using namespace ck::literals;
+
         if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({row * stride, stride, 1}));
+            return HostTensorDescriptor({batch_count, row, col}, {row * stride, stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({col * stride, 1, stride}));
+            return HostTensorDescriptor({batch_count, row, col}, {col * stride, 1_uz, stride});
         }
     };
 
@@ -95,17 +96,13 @@ 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<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<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<ReduceDataType> d0_g_m_host_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_host_result({BatchCount, M});
 
     Tensor<CDataType> c_g_m_n_device_result(
         f_host_tensor_descriptor(BatchCount, M, N, StrideC, CLayout{}));
-    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<ReduceDataType> d1_g_m_device_result(HostTensorDescriptor(std::vector<std::size_t>(
-        {static_cast<std::size_t>(BatchCount), static_cast<std::size_t>(M)})));
+    Tensor<ReduceDataType> d0_g_m_device_result({BatchCount, M});
+    Tensor<ReduceDataType> d1_g_m_device_result({BatchCount, M});
 
     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;
@@ -319,12 +316,9 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
                 reduce0_device_buf.FromDevice(d0_g_m_device_result.mData.data());
                 reduce1_device_buf.FromDevice(d1_g_m_device_result.mData.data());
 
-                bool c_error =
-                    ck::utils::check_err(c_g_m_n_device_result.mData, c_g_m_n_host_result.mData);
-                bool d0_error =
-                    ck::utils::check_err(d0_g_m_device_result.mData, d0_g_m_host_result.mData);
-                bool d1_error =
-                    ck::utils::check_err(d1_g_m_device_result.mData, d1_g_m_host_result.mData);
+                bool c_error  = ck::utils::check_err(c_g_m_n_device_result, c_g_m_n_host_result);
+                bool d0_error = ck::utils::check_err(d0_g_m_device_result, d0_g_m_host_result);
+                bool d1_error = ck::utils::check_err(d1_g_m_device_result, d1_g_m_host_result);
 
                 pass = pass && (c_error == true);
                 pass = pass && (d0_error == true);

profiler/include/profile_batched_gemm_softmax_gemm_impl.hpp

@@ -16,6 +16,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
 
@@ -29,7 +30,8 @@ template <typename ADataType,
           typename ALayout,
           typename B0Layout,
           typename B1Layout,
-          typename CLayout>
+          typename CLayout,
+          bool MaskOutUpperTriangle>
 bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
                                             int init_method,
                                             bool do_log,
@@ -46,16 +48,18 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
                                             int BatchStrideA  = -1,
                                             int BatchStrideB0 = -1,
                                             int BatchStrideB1 = -1,
-                                            int BatchStrideC  = -1)
+                                            int BatchStrideC  = -1,
+                                            float alpha       = 1.f)
 
 {
 
     using Row           = tensor_layout::gemm::RowMajor;
     using Col           = tensor_layout::gemm::ColumnMajor;
     using PassThrough   = tensor_operation::element_wise::PassThrough;
+    using Scale         = tensor_operation::element_wise::Scale;
     using AElementOp    = PassThrough;
     using B0ElementOp   = PassThrough;
-    using Acc0ElementOp = PassThrough;
+    using Acc0ElementOp = Scale;
     using B1ElementOp   = PassThrough;
     using CElementOp    = PassThrough;
     using AccDataType   = float;
@@ -67,7 +71,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
                                                                                 AccDataType,
                                                                                 AElementOp,
                                                                                 B0ElementOp,
-                                                                                CElementOp>;
+                                                                                Acc0ElementOp>;
 
     // Ref Softmax: fp32 in, various type out
     using ReferenceSoftmaxInstance =
@@ -110,15 +114,15 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
                                        std::size_t stride,
                                        std::size_t batch_stride,
                                        auto layout) {
+        using namespace ck::literals;
+
         if(std::is_same<decltype(layout), Row>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+            return HostTensorDescriptor({batch_count, row, col}, {batch_stride, 1_uz, stride});
         }
     };
 
@@ -185,7 +189,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
 
     auto a_element_op    = AElementOp{};
     auto b0_element_op   = B0ElementOp{};
-    auto acc0_element_op = Acc0ElementOp{};
+    auto acc0_element_op = Acc0ElementOp{alpha};
     auto b1_element_op   = B1ElementOp{};
     auto c_element_op    = CElementOp{};
 
@@ -201,7 +205,8 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
                                                                             B0ElementOp,
                                                                             Acc0ElementOp,
                                                                             B1ElementOp,
-                                                                            CElementOp>;
+                                                                            CElementOp,
+                                                                            MaskOutUpperTriangle>;
 
     // get device op instances
     const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
@@ -214,10 +219,16 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
         auto ref_gemm0          = ReferenceGemm0Instance{};
         auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
         auto ref_gemm0_argument = ref_gemm0.MakeArgument(
-            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, PassThrough{});
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, Scale{alpha});
 
         ref_gemm0_invoker.Run(ref_gemm0_argument);
 
+        // mask out upper triangle
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(MaskOutUpperTriangle && idx[1] < idx[2])
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
         auto ref_softmax          = ReferenceSoftmaxInstance{};
         auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
         auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
@@ -297,8 +308,7 @@ bool profile_batched_gemm_softmax_gemm_impl(bool do_verification,
             {
                 c_g_m_o_device_buf.FromDevice(c_g_m_o_device_result.mData.data());
 
-                pass = pass &
-                       ck::utils::check_err(c_g_m_o_device_result.mData, c_g_m_o_host_result.mData);
+                pass = pass & ck::utils::check_err(c_g_m_o_device_result, c_g_m_o_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_batched_gemm_masking_scale_softmax_gemm_permute_impl.hpp → profiler/include/profile_batched_gemm_softmax_gemm_permute_impl.hpp

@@ -7,51 +7,48 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
-#include "ck/library/tensor_operation_instance/gpu/batched_gemm_masking_scale_softmax_gemm_permute.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp"
 
 #include "ck/library/utility/check_err.hpp"
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
 
 namespace ck {
 namespace profiler {
 
-template <typename ADataType,
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          index_t NumDimO,
+          typename ADataType,
           typename B0DataType,
           typename B1DataType,
           typename CDataType,
-          typename ALayout,
-          typename B0Layout,
-          typename B1Layout,
-          typename CPermuteNumDims_G_M_O>
-bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verification,
-                                                                  int init_method,
-                                                                  bool do_log,
-                                                                  bool time_kernel,
-                                                                  int M,
-                                                                  int N,
-                                                                  int K,
-                                                                  int O,
-                                                                  int G0,
-                                                                  int G1,
-                                                                  int StrideA       = -1,
-                                                                  int StrideB0      = -1,
-                                                                  int StrideB1      = -1,
-                                                                  int BatchStrideA  = -1,
-                                                                  int BatchStrideB0 = -1,
-                                                                  int BatchStrideB1 = -1,
-                                                                  float alpha       = 1.f)
+          typename Acc0BiasesDataType,
+          typename Acc1BiasesDataType,
+          tensor_operation::device::MaskingSpecialization MaskingSpec>
+bool profile_batched_gemm_softmax_gemm_permute_impl(bool do_verification,
+                                                    int init_method,
+                                                    bool do_log,
+                                                    bool time_kernel,
+                                                    int M,
+                                                    int N,
+                                                    int K,
+                                                    int O,
+                                                    int G0,
+                                                    int G1,
+                                                    float alpha = 1.f)
 
 {
 
-    using Row           = tensor_layout::gemm::RowMajor;
-    using Col           = tensor_layout::gemm::ColumnMajor;
     using PassThrough   = tensor_operation::element_wise::PassThrough;
     using Scale         = tensor_operation::element_wise::Scale;
     using AElementOp    = PassThrough;
@@ -60,6 +57,7 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
     using B1ElementOp   = PassThrough;
     using CElementOp    = PassThrough;
     using AccDataType   = float;
+    using tensor_operation::device::MaskingSpecialization;
 
     // Ref Gemm0: various type in, fp32 out
     using ReferenceGemm0Instance = tensor_operation::host::ReferenceBatchedGemm<ADataType,
@@ -85,67 +83,33 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
 
     bool pass = true;
 
-    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
-    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
-
-    const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
-    const int DefaultStrideB0 = ck::is_same_v<B0Layout, Row> ? N : K;
-    const int DefaultStrideB1 = ck::is_same_v<B1Layout, Row> ? O : N;
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
 
-    StrideA  = (StrideA < 0) ? DefaultStrideA : StrideA;
-    StrideB0 = (StrideB0 < 0) ? DefaultStrideB0 : StrideB0;
-    StrideB1 = (StrideB1 < 0) ? DefaultStrideB1 : StrideB1;
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
 
-    const int DefaultBatchStrideA  = (ck::is_same_v<ALayout, Col> ? K : M) * StrideA;
-    const int DefaultBatchStrideB0 = (ck::is_same_v<B0Layout, Col> ? N : K) * StrideB0;
-    const int DefaultBatchStrideB1 = (ck::is_same_v<B1Layout, Col> ? O : N) * StrideB1;
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
 
-    BatchStrideA  = BatchStrideA < 0 ? DefaultBatchStrideA : BatchStrideA;
-    BatchStrideB0 = BatchStrideB0 < 0 ? DefaultBatchStrideB0 : BatchStrideB0;
-    BatchStrideB1 = BatchStrideB1 < 0 ? DefaultBatchStrideB1 : BatchStrideB1;
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
 
     const int BatchCount = G0 * G1;
 
-    auto f_host_tensor_descriptor = [](std::size_t batch_count,
-                                       std::size_t row,
-                                       std::size_t col,
-                                       std::size_t stride,
-                                       std::size_t batch_stride,
-                                       auto layout) {
-        if(std::is_same<decltype(layout), Row>::value)
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, stride, 1}));
-        }
-        else
-        {
-            return HostTensorDescriptor(std::vector<std::size_t>({batch_count, row, col}),
-                                        std::vector<std::size_t>({batch_stride, 1, stride}));
-        }
-    };
-
-    // C_m_o = A_m_k * B0_k_n * B1_n_o
-    Tensor<ADataType> a_g_m_k(
-        f_host_tensor_descriptor(BatchCount, M, K, StrideA, BatchStrideA, ALayout{}));
-    Tensor<B0DataType> b0_g_k_n(
-        f_host_tensor_descriptor(BatchCount, K, N, StrideB0, BatchStrideB0, B0Layout{}));
-    Tensor<B1DataType> b1_g_n_o(
-        f_host_tensor_descriptor(BatchCount, N, O, StrideB1, BatchStrideB1, B1Layout{}));
-    Tensor<CDataType> c_gs_ms_os_host_result(
-        std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-        std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-    Tensor<CDataType> c_gs_ms_os_device_result(
-        std::vector<std::size_t>(c_gs_ms_os_lengths.begin(), c_gs_ms_os_lengths.end()),
-        std::vector<std::size_t>(c_gs_ms_os_strides.begin(), c_gs_ms_os_strides.end()));
-    // Host verification: Output of Gemm0 is input A of Gemm1
-    Tensor<AccDataType> acc0_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-    Tensor<ADataType> a1_g_m_n(f_host_tensor_descriptor(BatchCount, M, N, N, M * N, Row{}));
-    Tensor<CDataType> c_g_m_o_host_result(std::vector<int>{BatchCount, M, O},
-                                          std::vector<int>{M * O, O, 1});
-
-    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
-    std::cout << "b0_g_k_n: " << b0_g_k_n.mDesc << std::endl;
-    std::cout << "b1_g_n_o: " << b1_g_n_o.mDesc << std::endl;
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
     std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
 
     std::srand(1); // work around test flakiness
@@ -157,38 +121,38 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
         // or not. May want to try exact same approach as the GPU kernel in the host reference
         // GEMM+Softmax+GEMM function to see if the accuracy discrepancy goes away. Until then,
         // shrink the input value range as it is less likely to produce errors of around ~1e-3.
-        // a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
-        // b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
-        // b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        // a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        // b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        // b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-5, 5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
         break;
     case 2:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
         break;
     case 3:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
         break;
     default:
-        a_g_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
-        b0_g_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
-        b1_g_n_o.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
     }
 
-    DeviceMem a_g_m_k_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSize());
-    DeviceMem b0_g_k_n_device_buf(sizeof(B0DataType) * b0_g_k_n.mDesc.GetElementSize());
-    DeviceMem b1_g_n_o_device_buf(sizeof(B1DataType) * b1_g_n_o.mDesc.GetElementSize());
-    DeviceMem c_gs_ms_os_device_buf(sizeof(CDataType) *
-                                    c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
+    DeviceMem c_device_buf(sizeof(CDataType) *
+                           c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
 
-    a_g_m_k_device_buf.ToDevice(a_g_m_k.mData.data());
-    b0_g_k_n_device_buf.ToDevice(b0_g_k_n.mData.data());
-    b1_g_n_o_device_buf.ToDevice(b1_g_n_o.mData.data());
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
 
     auto a_element_op    = AElementOp{};
     auto b0_element_op   = B0ElementOp{};
@@ -196,20 +160,23 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
     auto b1_element_op   = B1ElementOp{};
     auto c_element_op    = CElementOp{};
 
-    using DeviceOp =
-        tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<ALayout,
-                                                                      B0Layout,
-                                                                      B1Layout,
-                                                                      CPermuteNumDims_G_M_O,
-                                                                      ADataType,
-                                                                      B0DataType,
-                                                                      B1DataType,
-                                                                      CDataType,
-                                                                      AElementOp,
-                                                                      B0ElementOp,
-                                                                      Acc0ElementOp,
-                                                                      B1ElementOp,
-                                                                      CElementOp>;
+    using DeviceOp = tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                                   1,
+                                                                                   1,
+                                                                                   1,
+                                                                                   1,
+                                                                                   ADataType,
+                                                                                   B0DataType,
+                                                                                   B1DataType,
+                                                                                   CDataType,
+                                                                                   ck::Tuple<>,
+                                                                                   ck::Tuple<>,
+                                                                                   AElementOp,
+                                                                                   B0ElementOp,
+                                                                                   Acc0ElementOp,
+                                                                                   B1ElementOp,
+                                                                                   CElementOp,
+                                                                                   MaskingSpec>;
 
     // get device op instances
     const auto op_ptrs = tensor_operation::device::instance::DeviceOperationInstanceFactory<
@@ -219,6 +186,26 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
 
     if(do_verification)
     {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+
         auto ref_gemm0          = ReferenceGemm0Instance{};
         auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
         auto ref_gemm0_argument = ref_gemm0.MakeArgument(
@@ -228,7 +215,7 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
 
         // mask out upper triangle
         acc0_g_m_n.ForEach([&](auto& self, auto idx) {
-            if(idx[1] < idx[2])
+            if(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle && idx[1] < idx[2])
                 self(idx) = -ck::NumericLimits<float>::Infinity();
         });
 
@@ -265,23 +252,24 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
     for(auto& op_ptr : op_ptrs)
     {
         auto argument_ptr = op_ptr->MakeArgumentPointer(
-            static_cast<ADataType*>(a_g_m_k_device_buf.GetDeviceBuffer()),
-            static_cast<B0DataType*>(b0_g_k_n_device_buf.GetDeviceBuffer()),
-            static_cast<B1DataType*>(b1_g_n_o_device_buf.GetDeviceBuffer()),
-            static_cast<CDataType*>(c_gs_ms_os_device_buf.GetDeviceBuffer()),
-            M,
-            N,
-            K,
-            O,
-            BatchCount,
+            static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
+            static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+            static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+            {}, // std::array<void*, 1> p_acc0_biases;
+            {}, // std::array<void*, 1> p_acc1_biases;
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
             c_gs_ms_os_lengths,
             c_gs_ms_os_strides,
-            StrideA,
-            StrideB0,
-            StrideB1,
-            BatchStrideA,
-            BatchStrideB0,
-            BatchStrideB1,
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
+            {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
             a_element_op,
             b0_element_op,
             acc0_element_op,
@@ -319,18 +307,18 @@ bool profile_batched_gemm_masking_scale_softmax_gemm_permute_impl(bool do_verifi
 
             if(do_verification)
             {
-                c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+                c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
 
-                pass = pass & ck::utils::check_err(c_gs_ms_os_device_result.mData,
-                                                   c_gs_ms_os_host_result.mData);
+                pass =
+                    pass & ck::utils::check_err(c_gs_ms_os_device_result, c_gs_ms_os_host_result);
 
                 if(do_log)
                 {
-                    LogRangeAsType<float>(std::cout << "a_g_m_k: ", a_g_m_k.mData, ",")
+                    LogRangeAsType<float>(std::cout << "a_gs_ms_ks: ", a_gs_ms_ks.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "b0_g_k_n : ", b0_g_k_n.mData, ",")
+                    LogRangeAsType<float>(std::cout << "b0_gs_ns_ks : ", b0_gs_ns_ks.mData, ",")
                         << std::endl;
-                    LogRangeAsType<float>(std::cout << "b1_g_n_o : ", b1_g_n_o.mData, ",")
+                    LogRangeAsType<float>(std::cout << "b1_gs_os_ns : ", b1_gs_os_ns.mData, ",")
                         << std::endl;
                     LogRangeAsType<float>(
                         std::cout << "c_gs_ms_os_host_result : ", c_gs_ms_os_host_result.mData, ",")

profiler/include/profile_conv_bwd_data_impl.hpp

@@ -209,8 +209,7 @@ bool profile_conv_bwd_data_impl(int do_verification,
             {
                 in_device_buf.FromDevice(input_device_result.mData.data());
 
-                pass =
-                    pass & ck::utils::check_err(input_device_result.mData, input_host_result.mData);
+                pass = pass & ck::utils::check_err(input_device_result, input_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp

@@ -12,6 +12,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp"
 
 namespace ck {
@@ -68,19 +69,19 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
             if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
                          is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
                          is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
             }
             else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
                               is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
                               is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
             }
         };
 
@@ -92,8 +93,7 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
         f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
 
     // bias: assume contiguous 1d vector
-    Tensor<OutDataType> bias_k(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(K)})));
+    Tensor<OutDataType> bias_k({K});
 
     // residual: assume same layout as output tensor
     Tensor<OutDataType> resi_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
@@ -251,8 +251,7 @@ void profile_conv_fwd_bias_relu_add_impl(int do_verification,
             {
                 out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
 
-                ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
-                                     out_n_k_ho_wo_host_result.mData);
+                ck::utils::check_err(out_n_k_ho_wo_device_result, out_n_k_ho_wo_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_conv_fwd_bias_relu_impl.hpp

@@ -12,6 +12,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp"
 
 namespace ck {
@@ -68,19 +69,19 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
+            using namespace ck::literals;
+
             if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
                          is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
                          is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, H * W, W, 1_uz});
             }
             else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
                               is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
                               is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
-                                            std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
+                return HostTensorDescriptor({N_, C_, H, W}, {C_ * H * W, 1_uz, W * C_, C_});
             }
         };
 
@@ -92,8 +93,7 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
         f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
 
     // bias: assume contiguous 1d vector
-    Tensor<OutDataType> bias_k(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(K)})));
+    Tensor<OutDataType> bias_k({K});
 
     std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
     std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
@@ -239,8 +239,7 @@ void profile_conv_fwd_bias_relu_impl(int do_verification,
             {
                 out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
 
-                ck::utils::check_err(out_n_k_ho_wo_device_result.mData,
-                                     out_n_k_ho_wo_host_result.mData);
+                ck::utils::check_err(out_n_k_ho_wo_device_result, out_n_k_ho_wo_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_conv_fwd_impl.hpp

@@ -191,7 +191,7 @@ bool profile_conv_fwd_impl(int do_verification,
             {
                 out_device_buf.FromDevice(device_output.mData.data());
 
-                pass = pass & ck::utils::check_err(device_output.mData, host_output.mData);
+                pass = pass & ck::utils::check_err(device_output, host_output);
 
                 if(do_log)
                 {

profiler/include/profile_convnd_bwd_data_impl.hpp

@@ -453,7 +453,7 @@ bool profile_convnd_bwd_data_impl(int do_verification,
                     std::cout << "Pass Info: " << conv_ptr->GetTypeString() << std::endl;
                 }
 
-                success = ck::utils::check_err(input_host_result.mData, input_device_result.mData);
+                success = ck::utils::check_err(input_host_result, input_device_result);
 
                 if(do_log)
                 {

profiler/include/profile_convnd_bwd_weight_impl.hpp

@@ -433,7 +433,7 @@ bool profile_convnd_bwd_weight_impl(int do_verification,
         {
             wei_device_buf.FromDevice(weights_device_result.mData.data());
 
-            success = ck::utils::check_err(weights_host_result.mData, weights_device_result.mData);
+            success = ck::utils::check_err(weights_host_result, weights_device_result);
 
             if(success == false)
             {

profiler/include/profile_elementwise_layernorm_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/library/tensor_operation_instance/gpu/elementwise_normalization.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename HostTensorA, typename HostTensorB, typename HostTensorC, typename Functor>
+void host_elementwise2D(HostTensorC& C,
+                        const HostTensorA& A,
+                        const HostTensorB& B,
+                        const std::vector<std::size_t>& shape,
+                        Functor functor)
+{
+    using ctype = ck::remove_reference_t<decltype(C(0, 0))>;
+
+    for(std::size_t m = 0; m < shape[0]; ++m)
+        for(std::size_t n = 0; n < shape[1]; ++n)
+        {
+            auto a_val  = A(m, n);
+            auto b_val  = B(m, n);
+            ctype c_val = 0;
+            functor(c_val, a_val, b_val);
+            C(m, n) = c_val;
+        }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename AccDataType,
+          typename YDataType>
+bool profile_elementwise_layernorm_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        std::vector<index_t> length)
+{
+    using Add         = ck::tensor_operation::element_wise::Add;
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+    if(length.size() != 2)
+        return false;
+
+    index_t M      = length[0];
+    index_t N      = length[1];
+    index_t Stride = N;
+
+    constexpr int Rank         = 2;
+    constexpr int NumReduceDim = 1;
+
+    std::vector<index_t> reduce_dim      = {1};
+    std::vector<index_t> gammaBetaLength = {N};
+    std::vector<index_t> gammaBetaStride = {0, 1};
+
+    auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
+        using namespace ck::literals;
+
+        return HostTensorDescriptor({row, col}, {stride, 1_uz});
+    };
+
+    Tensor<ADataType> a(length);
+    Tensor<BDataType> b(length);
+    Tensor<GammaDataType> gamma(gammaBetaLength);
+    Tensor<BetaDataType> beta(gammaBetaLength);
+    Tensor<YDataType> y(length);
+    Tensor<YDataType> host_y(length);
+
+    switch(init_method)
+    {
+    case 0:
+        a.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
+        b.GenerateTensorValue(GeneratorTensor_1<BDataType>{});
+        gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
+        beta.GenerateTensorValue(GeneratorTensor_1<BetaDataType>{});
+        break;
+    case 1:
+        a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
+        beta.GenerateTensorValue(GeneratorTensor_2<BetaDataType>{-5, 5});
+        break;
+    default:
+        a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0, 1});
+        b.GenerateTensorValue(GeneratorTensor_3<BDataType>{0, 1});
+        gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
+        beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_dev(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_dev(sizeof(ADataType) * b.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
+    DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+
+    a_dev.ToDevice(a.mData.data());
+    b_dev.ToDevice(b.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    beta_dev.ToDevice(beta.mData.data());
+
+    std::array<const void*, 2> input = {a_dev.GetDeviceBuffer(), b_dev.GetDeviceBuffer()};
+
+    // add device normalization instances
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwiseNormalization<
+        ck::Tuple<ADataType, BDataType>,
+        GammaDataType,
+        BetaDataType,
+        AccDataType,
+        YDataType,
+        Add,
+        PassThrough,
+        2,
+        1>;
+
+    // get device op instances
+    const auto instance_ptrs =
+        ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+            DeviceOp>::GetInstances();
+
+    std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    if(do_verification)
+    {
+        using XDataType             = ADataType;
+        std::vector<std::size_t> mn = {static_cast<unsigned long>(M),
+                                       static_cast<unsigned long>(N)};
+        Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
+        host_elementwise2D<Tensor<ADataType>, Tensor<BDataType>, Tensor<XDataType>, Add>(
+            x, a, b, mn, Add{});
+
+        using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
+                                                                                 GammaDataType,
+                                                                                 BetaDataType,
+                                                                                 YDataType,
+                                                                                 AccDataType,
+                                                                                 PassThrough,
+                                                                                 Rank,
+                                                                                 NumReduceDim>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+    }
+
+    int num_kernel = 0;
+
+    for(auto& inst_ptr : instance_ptrs)
+    {
+        auto argument_ptr = inst_ptr->MakeArgumentPointer(
+            length,
+            {
+                std::vector<ck::index_t>{a.mDesc.GetStrides().begin(), a.mDesc.GetStrides().end()},
+                std::vector<ck::index_t>{b.mDesc.GetStrides().begin(), b.mDesc.GetStrides().end()},
+            },
+            gammaBetaStride,
+            gammaBetaStride,
+            std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
+            reduce_dim,
+            1e-4,
+            input,
+            gamma_dev.GetDeviceBuffer(),
+            beta_dev.GetDeviceBuffer(),
+            y_dev.GetDeviceBuffer(),
+            Add{},
+            PassThrough{});
+
+        if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            ++num_kernel;
+        }
+        else
+        {
+            continue;
+        }
+
+        auto invoker_ptr = inst_ptr->MakeInvokerPointer();
+
+        float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+        std::size_t num_bytes = a.mDesc.GetElementSize() * sizeof(ADataType) +
+                                b.mDesc.GetElementSize() * sizeof(BDataType) +
+                                gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
+                                beta.mDesc.GetElementSize() * sizeof(BetaDataType) +
+                                y.mDesc.GetElementSize() * sizeof(YDataType);
+
+        float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+        if(time_kernel)
+            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)
+        {
+            y_dev.FromDevice(y.mData.data());
+
+            bool pass =
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+
+            if(do_log)
+            {
+                LogRangeAsType<float>(std::cout << "a  : ", a.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "b  : ", b.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "host_y  : ", host_y.mData, ",") << std::endl;
+                LogRangeAsType<float>(std::cout << "y  : ", y.mData, ",") << std::endl;
+            }
+
+            if(!pass)
+            {
+                std::cout << inst_ptr->GetTypeString() << " failed verification: ";
+                LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
+                return false;
+            }
+            else
+            {
+                if(time_kernel)
+                    std::cout << "pass" << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", length, ",") << ", ";
+        std::cout << "num_kernel = " << num_kernel << ", best perf = " << best_avg_time << " ms, "
+                  << best_gb_per_sec << " GB/s, " << best_instance_name << std::endl;
+    }
+
+    if(num_kernel == 0)
+    {
+        std::cout << "Error: No kernel is tested" << std::endl;
+        return false;
+    }
+
+    return true;
+}
+
+} // namespace profiler
+} // namespace ck

profiler/include/profile_gemm_add_add_fastgelu_impl.hpp

@@ -16,6 +16,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -47,15 +48,15 @@ bool profile_gemm_add_add_fastgelu_impl(int do_verification,
 {
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -121,8 +122,7 @@ bool profile_gemm_add_add_fastgelu_impl(int do_verification,
     // run reference
     if(do_verification)
     {
-        Tensor<AccDataType> c_m_n(HostTensorDescriptor(
-            std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
+        Tensor<AccDataType> c_m_n({M, N});
 
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                                 BDataType,
@@ -223,8 +223,7 @@ bool profile_gemm_add_add_fastgelu_impl(int do_verification,
             {
                 e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
-                pass = pass &&
-                       ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
             }
         }
         else

profiler/include/profile_gemm_bias_add_reduce_impl.hpp

@@ -14,6 +14,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -75,21 +76,20 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                                        int StrideD0)
 {
     auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
-        return HostTensorDescriptor(std::vector<std::size_t>({len}),
-                                    std::vector<std::size_t>({stride}));
+        return HostTensorDescriptor({len}, {stride});
     };
 
     auto f_host_tensor_descriptor2d =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -99,16 +99,12 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
     Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
     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<ReduceDataType> reduce1_m_host_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<ReduceDataType> reduce0_m_host_result({M});
+    Tensor<ReduceDataType> reduce1_m_host_result({M});
 
     Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor2d(M, N, StrideC, CLayout{}));
-    Tensor<ReduceDataType> reduce0_m_device_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
-    Tensor<ReduceDataType> reduce1_m_device_result(
-        HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(M)})));
+    Tensor<ReduceDataType> reduce0_m_device_result({M});
+    Tensor<ReduceDataType> reduce1_m_device_result({M});
 
     std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
     std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
@@ -347,9 +343,9 @@ void profile_gemm_bias_add_reduce_impl(int do_verification,
                 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(reduce0_m_device_result.mData, reduce0_m_host_result.mData);
-                ck::utils::check_err(reduce1_m_device_result.mData, reduce1_m_host_result.mData);
+                ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
+                ck::utils::check_err(reduce0_m_device_result, reduce0_m_host_result);
+                ck::utils::check_err(reduce1_m_device_result, reduce1_m_host_result);
 
                 if(do_log)
                 {

profiler/include/profile_gemm_bilinear_impl.hpp

@@ -16,6 +16,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -46,15 +47,15 @@ bool profile_gemm_bilinear_impl(int do_verification,
 {
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -116,8 +117,7 @@ bool profile_gemm_bilinear_impl(int do_verification,
     // run reference
     if(do_verification)
     {
-        Tensor<AccDataType> c_m_n(HostTensorDescriptor(
-            std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
+        Tensor<AccDataType> c_m_n({M, N});
 
         using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                                 BDataType,
@@ -215,8 +215,7 @@ bool profile_gemm_bilinear_impl(int do_verification,
             {
                 e_device_buf.FromDevice(e_m_n_device_result.mData.data());
 
-                pass = pass &&
-                       ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
             }
         }
         else

profiler/include/profile_gemm_impl.hpp

@@ -18,6 +18,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 
 namespace ck {
@@ -45,15 +46,15 @@ int profile_gemm_impl(int do_verification,
 
     auto f_host_tensor_descriptor =
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
             if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }
             else
             {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
             }
         };
 
@@ -187,8 +188,7 @@ int profile_gemm_impl(int 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);
+                pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
 
                 if(do_log)
                 {