Merge branch 'develop' of...

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into navi3x_md_bgemm_conv_gemmsoftmaxgemm

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+    std::vector<std::unique_ptr<
+        DeviceBatchedContractionMultipleD<1,
+                                          2,
+                                          3,
+                                          1,
+                                          F16,
+                                          F16,
+                                          F16_Tuple,
+                                          F16,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::Add>>>& instances);
+
+// Contraction + add
+template <index_t NumDimG,
+          index_t NumDimM,
+          index_t NumDimN,
+          index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename DDataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedContractionMultipleD<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        ADataType,
+        BDataType,
+        ck::Tuple<DDataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>>
+{
+    using DeviceOp =
+        DeviceBatchedContractionMultipleD<NumDimG,
+                                          NumDimM,
+                                          NumDimN,
+                                          NumDimK,
+                                          ADataType,
+                                          BDataType,
+                                          ck::Tuple<DDataType>,
+                                          EDataType,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::PassThrough,
+                                          ck::tensor_operation::element_wise::Add>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, ck::half_t> && is_same_v<BDataType, ck::half_t> &&
+                     is_same_v<DDataType, ck::half_t> && is_same_v<EDataType, ck::half_t>)
+        {
+            if constexpr(NumDimG == 1 && NumDimM == 2 && NumDimN == 3 && NumDimK == 1)
+            {
+                add_device_batched_contraction_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_mnnm_instance(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#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.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            F16,
+                                            ck::Tuple<F16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            F16,
+                                                            ck::Tuple<F16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<std::unique_ptr<
+        DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                            1,
+                                            1,
+                                            1,
+                                            1,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            BF16,
+                                            ck::Tuple<BF16>,
+                                            ck::Tuple<>,
+                                            PassThrough,
+                                            PassThrough,
+                                            ScaleAdd,
+                                            PassThrough,
+                                            PassThrough,
+                                            MaskingSpecialization::MaskOutUpperTriangle>>>&
+        instances);
+
+void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+    std::vector<
+        std::unique_ptr<DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            1,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            BF16,
+                                                            ck::Tuple<BF16>,
+                                                            ck::Tuple<>,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            ScaleAdd,
+                                                            PassThrough,
+                                                            PassThrough,
+                                                            MaskingSpecialization::MaskDisabled>>>&
+        instances);
+
+template <typename ADataType,
+          typename B0DataType,
+          typename B1DataType,
+          typename CDataType,
+          typename Acc0BiasDataType,
+          MaskingSpecialization MaskingSpec>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      1,
+                                                                      ADataType,
+                                                                      B0DataType,
+                                                                      B1DataType,
+                                                                      CDataType,
+                                                                      Acc0BiasDataType,
+                                                                      ck::Tuple<>,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      ScaleAdd,
+                                                                      PassThrough,
+                                                                      PassThrough,
+                                                                      MaskingSpec>>
+{
+    using DeviceOp = DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         1,
+                                                         ADataType,
+                                                         B0DataType,
+                                                         B1DataType,
+                                                         CDataType,
+                                                         Acc0BiasDataType,
+                                                         ck::Tuple<>,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         ScaleAdd,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         MaskingSpec>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
+                     is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
+                     Acc0BiasDataType::Size() == 1 &&
+                     is_same_v<tuple_element_t<0, Acc0BiasDataType>, half_t>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
+                          is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
+                          Acc0BiasDataType::Size() == 1 &&
+                          is_same_v<tuple_element_t<0, Acc0BiasDataType>, BF16>)
+        {
+            if constexpr(MaskingSpec == MaskingSpecialization::MaskOutUpperTriangle)
+            {
+                add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+            else if(MaskingSpec == MaskingSpecialization::MaskDisabled)
+            {
+                add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
+                    op_ptrs);
+            }
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// FP16
+void add_device_batchnorm_infer_rank_4_f16_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F16, F32, F32, F16, F16>,
+        ck::Tuple<F16>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// FP32
+void add_device_batchnorm_infer_rank_4_f32_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F32, F32, F32, F32, F32>,
+        ck::Tuple<F32>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// BF16
+void add_device_batchnorm_infer_rank_4_bf16_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<BF16, F32, F32, BF16, BF16>,
+        ck::Tuple<BF16>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+// FP64
+void add_device_batchnorm_infer_rank_4_f64_instances(
+    std::vector<std::unique_ptr<ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<F64, F64, F64, F64, F64>,
+        ck::Tuple<F64>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        4>>>&);
+
+template <typename XDataType,
+          typename YDataType,
+          typename ScaleDataType,
+          typename BiasDataType,
+          typename MeanVarDataType,
+          index_t Rank>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceElementwise<
+    ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+    ck::Tuple<YDataType>,
+    ck::tensor_operation::element_wise::NormalizeInInfer,
+    Rank>>
+{
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<
+        ck::Tuple<XDataType, MeanVarDataType, MeanVarDataType, ScaleDataType, BiasDataType>,
+        ck::Tuple<YDataType>,
+        ck::tensor_operation::element_wise::NormalizeInInfer,
+        Rank>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<XDataType, F16> && is_same_v<YDataType, F16> &&
+                     is_same_v<ScaleDataType, F16> && is_same_v<BiasDataType, F16> &&
+                     is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F32> && is_same_v<YDataType, F32> &&
+                          is_same_v<ScaleDataType, F32> && is_same_v<BiasDataType, F32> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f32_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, BF16> && is_same_v<YDataType, BF16> &&
+                          is_same_v<ScaleDataType, BF16> && is_same_v<BiasDataType, BF16> &&
+                          is_same_v<MeanVarDataType, F32>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_bf16_instances(op_ptrs);
+            }
+        }
+        else if constexpr(is_same_v<XDataType, F64> && is_same_v<YDataType, F64> &&
+                          is_same_v<ScaleDataType, F64> && is_same_v<BiasDataType, F64> &&
+                          is_same_v<MeanVarDataType, F64>)
+        {
+            if constexpr(Rank == 4)
+            {
+                add_device_batchnorm_infer_rank_4_f64_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp

@@ -7,7 +7,7 @@
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
@@ -18,11 +18,8 @@ namespace device {
 namespace instance {
 
 using Normalize                            = ck::tensor_operation::element_wise::Normalize;
-using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::DeviceElementwiseBasePtr<
-    Tuple<half_t, float, float, half_t, half_t>,
-    Tuple<half_t>,
-    Normalize,
-    2>;
+using DeviceNormalizeFromMeanMeanSquarePtr = ck::tensor_operation::device::
+    DeviceElementwisePtr<Tuple<half_t, float, float, half_t, half_t>, Tuple<half_t>, Normalize, 2>;
 
 void add_device_normalize_from_mean_squaremean_f16_f32_f32_f16_f16_instances(
     std::vector<DeviceNormalizeFromMeanMeanSquarePtr>& instances);

library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp

@@ -10,7 +10,6 @@
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
 

library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Row,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Row,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+void add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleD<Col,
+                                                    Col,
+                                                    Row_Row_Tuple,
+                                                    Row,
+                                                    F16,
+                                                    F16,
+                                                    F16_F16_Tuple,
+                                                    F16,
+                                                    PassThrough,
+                                                    PassThrough,
+                                                    AddMultiply>>>&);
+
+// GEMM + Add + Multiply
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename EDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleD<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    ELayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    EDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddMultiply>>
+{
+    using DeviceOp = DeviceGemmMultipleD<ALayout,
+                                         BLayout,
+                                         ck::Tuple<D0Layout, D1Layout>,
+                                         ELayout,
+                                         ADataType,
+                                         BDataType,
+                                         ck::Tuple<D0DataType, D1DataType>,
+                                         EDataType,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::PassThrough,
+                                         ck::tensor_operation::element_wise::AddMultiply>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<EDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<ELayout, Row>)
+            {
+                add_device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Row,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Row,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+void add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDLayernorm<Col,
+                                                             Col,
+                                                             Row_Row_Tuple,
+                                                             Row,
+                                                             F16,
+                                                             F16,
+                                                             F16_F16_Tuple,
+                                                             F16,
+                                                             F16,
+                                                             F16,
+                                                             PassThrough,
+                                                             PassThrough,
+                                                             AddReluAdd,
+                                                             PassThrough>>>&);
+
+// GEMM + Add + Relu + Add + Layernorm
+template <typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename D1Layout,
+          typename HLayout,
+          typename ADataType,
+          typename BDataType,
+          typename D0DataType,
+          typename D1DataType,
+          typename GammaDataType,
+          typename BetaDataType,
+          typename HDataType>
+struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+    ALayout,
+    BLayout,
+    ck::Tuple<D0Layout, D1Layout>,
+    HLayout,
+    ADataType,
+    BDataType,
+    ck::Tuple<D0DataType, D1DataType>,
+    GammaDataType,
+    BetaDataType,
+    HDataType,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::PassThrough,
+    ck::tensor_operation::element_wise::AddReluAdd,
+    ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmMultipleDLayernorm<ALayout,
+                                                  BLayout,
+                                                  ck::Tuple<D0Layout, D1Layout>,
+                                                  HLayout,
+                                                  ADataType,
+                                                  BDataType,
+                                                  ck::Tuple<D0DataType, D1DataType>,
+                                                  GammaDataType,
+                                                  BetaDataType,
+                                                  HDataType,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::PassThrough,
+                                                  ck::tensor_operation::element_wise::AddReluAdd,
+                                                  ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
+                     is_same_v<D0DataType, half_t> && is_same_v<D1DataType, half_t> &&
+                     is_same_v<GammaDataType, half_t> && is_same_v<BetaDataType, half_t> &&
+                     is_same_v<HDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
+                         is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                         is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+            else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
+                              is_same_v<D0Layout, Row> && is_same_v<D1Layout, Row> &&
+                              is_same_v<HLayout, Row>)
+            {
+                add_device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instances(
+                    op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp

@@ -244,6 +244,63 @@ void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(
                                                               PassThrough,
                                                               PassThrough>>>& instances);
 
+// grouped conv3d forward, NDHWGC/KZYXGC/NDHWGK
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              BF16,
+                                                              BF16,
+                                                              Empty_Tuple,
+                                                              BF16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
+                                                              NDHWGC,
+                                                              KZYXGC,
+                                                              Empty_Tuple,
+                                                              NDHWGK,
+                                                              int8_t,
+                                                              int8_t,
+                                                              Empty_Tuple,
+                                                              int8_t,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
 template <ck::index_t NumDimSpatial,
           typename InLayout,
           typename WeiLayout,
@@ -385,6 +442,31 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instances(op_ptrs);
             }
         }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
+                          is_same_v<WeiLayout, KZYXGC> && is_same_v<OutLayout, NDHWGK>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
+                         is_same_v<OutDataType, float>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
+                              is_same_v<OutDataType, half_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<WeiDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
+                              is_same_v<OutDataType, int8_t>)
+            {
+                add_device_grouped_conv3d_fwd_xdl_ndhwgc_kzyxgc_ndhwgk_int8_instances(op_ptrs);
+            }
+        }
 
         return op_ptrs;
     }

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp

@@ -76,8 +76,16 @@ template <typename InDataType,
           bool PropagateNan,
           bool OutputIndex>
 void add_device_reduce_instance_blockwise(
-    std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
-        device_op_instances)
+    std::vector<DeviceReducePtr<InDataType,
+                                AccDataType,
+                                OutDataType,
+                                Rank,
+                                NumReduceDim,
+                                ReduceOperation,
+                                InElementwiseOp,
+                                AccElementwiseOp,
+                                PropagateNan,
+                                OutputIndex>>& device_op_instances)
 {
     static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
         [&](auto i) {

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp

@@ -15,10 +15,10 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp

@@ -15,10 +15,10 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp

@@ -15,10 +15,10 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&); 
-extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp

@@ -15,14 +15,14 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
-extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<F16, F16, F16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp

@@ -15,10 +15,10 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>>&);
 // clang-format on
 
 } // namespace instance

library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp

@@ -15,10 +15,10 @@ namespace instance {
 
 // clang-format off
 // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); 
-extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&); 
+extern template void add_device_reduce_instance_blockwise<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<F16, F32, F16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
 // clang-format on
 
 } // namespace instance