Grouped convolution backward weight special vector size loads (#1772)

* Grouped convolution backward weight special vector size loads

* Instnaces and tests

* Fixes

* Add 7 and 13 special cases

* fix comments

* Fix

* Fix2

* fixes

* fix atomic add bf16

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -1558,14 +1558,23 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
             }
         }
 
-        if(!(arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0 &&
+        const bool is_w_pad_zero = arg.input_left_pads_[NDimSpatial - 1] == 0 &&
+                                   arg.input_right_pads_[NDimSpatial - 1] == 0;
+        const auto X                 = arg.filter_spatial_lengths_[NDimSpatial - 1];
+        const bool XC_access_allowed = arg.Conv_G_ == 1 &&
+                                       (arg.Conv_C_ * X) % BBlockTransferSrcScalarPerVector == 0 &&
+                                       is_w_pad_zero;
+
+        if(!((arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0 || XC_access_allowed) &&
              arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0))
         {
-            if(!(arg.Conv_K_ == 1 && arg.compute_ptr_offset_of_batch_.BatchStrideA_ == 1))
+            if(!(arg.Conv_K_ == 1 && arg.compute_ptr_offset_of_batch_.BatchStrideA_ == 1 &&
+                 NumGroupsToMerge > 1))
             {
                 return false;
             }
-            if(!(arg.Conv_C_ == 1 && arg.compute_ptr_offset_of_batch_.BatchStrideB_ == 1))
+            if(!(arg.Conv_C_ == 1 && arg.compute_ptr_offset_of_batch_.BatchStrideB_ == 1 &&
+                 NumGroupsToMerge > 1))
             {
                 return false;
             }

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -584,6 +584,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         {
             return false;
         }
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t>)
+        {
+            return false;
+        }
         if constexpr(NDimSpatial == 1)
         {
             if constexpr(!is_GNWC_GKXC_GNWK<InLayout, WeiLayout, OutLayout>())

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -54,6 +54,19 @@ struct ThreadwiseTensorSliceTransfer_v3r1
     using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
 
     static constexpr auto I0  = Number<0>{};
+    static constexpr auto I1  = Number<1>{};
+    static constexpr auto I2  = Number<2>{};
+    static constexpr auto I3  = Number<3>{};
+    static constexpr auto I4  = Number<4>{};
+    static constexpr auto I5  = Number<5>{};
+    static constexpr auto I6  = Number<6>{};
+    static constexpr auto I7  = Number<7>{};
+    static constexpr auto I8  = Number<8>{};
+    static constexpr auto I10 = Number<10>{};
+    static constexpr auto I12 = Number<12>{};
+    static constexpr auto I13 = Number<13>{};
+    static constexpr auto I14 = Number<14>{};
+    static constexpr auto I16 = Number<16>{};
 
     static constexpr index_t PackedSize = []() {
         if constexpr(is_same_v<remove_cvref_t<SrcData>, pk_i4_t>)
@@ -198,9 +211,6 @@ struct ThreadwiseTensorSliceTransfer_v3r1
             src_oob_thread_scratch_tuple_(thread_scratch_id)
                 .template SetAsType<bool>(src_data_idx_seq, is_src_valid);
 
-            using src_vector_type = vector_type_maker_t<SrcData, SrcScalarPerVector>;
-            using src_vector_t    = typename src_vector_type::type;
-
             using dst_vector_type = vector_type_maker_t<DstData, SrcScalarPerVector>;
             using dst_vector_t    = typename dst_vector_type::type;
             dst_vector_type op_r_v;
@@ -234,13 +244,62 @@ struct ThreadwiseTensorSliceTransfer_v3r1
             using src_elem_op_vec_t = typename vector_type<SrcData, elem_op_vec_len>::type;
             using dst_elem_op_vec_t = typename vector_type<DstData, elem_op_vec_len>::type;
 
-            auto src_vector_container = src_vector_type{
-                src_buf.template Get<src_vector_t>(src_coord_.GetOffset() / PackedSize, true)};
-
-            static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto idx) {
-                // apply the src elementwise op and convert to DstData under the hood if needed
-                src_element_op_(op_r_v.template AsType<dst_elem_op_vec_t>()(idx),
-                                src_vector_container.template AsType<src_elem_op_vec_t>()[idx]);
+            using VectorSizeLookupTable    = Tuple<Sequence<>,
+                                                Sequence<I1>,
+                                                Sequence<I2>,
+                                                Sequence<I2, I1>,
+                                                Sequence<I4>,
+                                                Sequence<I4, I1>,
+                                                Sequence<I4, I2>,
+                                                Sequence<I4, I2, I1>,
+                                                Sequence<I8>,
+                                                Sequence<I8, I1>,
+                                                Sequence<I8, I2>,
+                                                Sequence<I8, I2, I1>,
+                                                Sequence<I8, I4>,
+                                                Sequence<I8, I4, I1>,
+                                                Sequence<I8, I4, I2>,
+                                                Sequence<I8, I4, I2, I1>,
+                                                Sequence<I16>>;
+            using VectorOffsetsLookupTable = Tuple<Sequence<>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0, I2>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0, I4>,
+                                                   Sequence<I0, I4>,
+                                                   Sequence<I0, I4, I6>,
+                                                   Sequence<I0>,
+                                                   Sequence<I0, I8>,
+                                                   Sequence<I0, I8>,
+                                                   Sequence<I0, I8, I10>,
+                                                   Sequence<I0, I8>,
+                                                   Sequence<I0, I8, I12>,
+                                                   Sequence<I0, I8, I12>,
+                                                   Sequence<I0, I8, I12, I14>,
+                                                   Sequence<I0>>;
+
+            static_for<0, tuple_element_t<SrcScalarPerVector, VectorSizeLookupTable>::Size(), 1>{}(
+                [&](auto v_idx) {
+                    constexpr auto VectorLoadSize =
+                        tuple_element_t<SrcScalarPerVector, VectorSizeLookupTable>::At(v_idx);
+                    constexpr auto LoadOffset =
+                        tuple_element_t<SrcScalarPerVector, VectorOffsetsLookupTable>::At(v_idx);
+
+                    using src_vector_container   = vector_type_maker_t<SrcData, VectorLoadSize>;
+                    using src_vector_container_t = typename src_vector_container::type;
+
+                    src_vector_container src_vector =
+                        src_vector_container{src_buf.template Get<src_vector_container_t>(
+                            src_coord_.GetOffset() / PackedSize + LoadOffset, true)};
+
+                    static_for<0, VectorLoadSize / elem_op_vec_len, 1>{}([&](auto idx) {
+                        // apply the src elementwise op and convert to DstData under the hood if
+                        // needed
+                        src_element_op_(
+                            op_r_v.template AsType<dst_elem_op_vec_t>()(idx + LoadOffset),
+                            src_vector.template AsType<src_elem_op_vec_t>()[idx]);
+                    });
                 });
 
             // copy data from src_vector_container into src_thread_scratch_

include/ck/utility/data_type.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -314,6 +314,76 @@ struct vector_type<T, 2, typename std::enable_if_t<is_native_type<T>()>>
     }
 };
 
+template <typename T>
+struct vector_type<T, 3, typename std::enable_if_t<is_native_type<T>()>>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d3_t __attribute__((ext_vector_type(3)));
+
+    using type = d3_t;
+
+    union
+    {
+        d3_t d3_;
+        StaticallyIndexedArray<d1_t, 3> d1x3_;
+        StaticallyIndexedArray<d2_t, 1> d2x1_;
+        StaticallyIndexedArray<d3_t, 1> d3x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d3_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x3_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+        else if constexpr(is_same<X, d3_t>::value)
+        {
+            return data_.d3x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d3_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x3_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+        else if constexpr(is_same<X, d3_t>::value)
+        {
+            return data_.d3x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
 template <typename T>
 struct vector_type<T, 4, typename std::enable_if_t<is_native_type<T>()>>
 {
@@ -384,6 +454,158 @@ struct vector_type<T, 4, typename std::enable_if_t<is_native_type<T>()>>
     }
 };
 
+template <typename T>
+struct vector_type<T, 5, typename std::enable_if_t<is_native_type<T>()>>
+{
+    using d1_t = T;
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d5_t __attribute__((ext_vector_type(5)));
+
+    using type = d5_t;
+
+    union
+    {
+        d5_t d5_;
+        StaticallyIndexedArray<d1_t, 5> d1x5_;
+        StaticallyIndexedArray<d4_t, 1> d4x1_;
+        StaticallyIndexedArray<d5_t, 1> d5x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d4_t>::value || is_same<X, d5_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x5_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+        else if constexpr(is_same<X, d5_t>::value)
+        {
+            return data_.d5x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d4_t>::value || is_same<X, d5_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x5_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+        else if constexpr(is_same<X, d5_t>::value)
+        {
+            return data_.d5x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 7, typename std::enable_if_t<is_native_type<T>()>>
+{
+    using d1_t = T;
+    typedef T d2_t __attribute__((ext_vector_type(2)));
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d7_t __attribute__((ext_vector_type(7)));
+
+    using type = d7_t;
+
+    union
+    {
+        d7_t d7_;
+        StaticallyIndexedArray<d1_t, 7> d1x7_;
+        StaticallyIndexedArray<d2_t, 3> d2x3_;
+        StaticallyIndexedArray<d4_t, 1> d4x1_;
+        StaticallyIndexedArray<d7_t, 1> d7x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d7_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x7_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x3_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+        else if constexpr(is_same<X, d7_t>::value)
+        {
+            return data_.d7x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d7_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x7_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x3_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+        else if constexpr(is_same<X, d7_t>::value)
+        {
+            return data_.d7x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
 template <typename T>
 struct vector_type<T, 8, typename std::enable_if_t<is_native_type<T>()>>
 {
@@ -466,6 +688,88 @@ struct vector_type<T, 8, typename std::enable_if_t<is_native_type<T>()>>
     }
 };
 
+template <typename T>
+struct vector_type<T, 13, typename std::enable_if_t<is_native_type<T>()>>
+{
+    using d1_t = T;
+    typedef T d4_t __attribute__((ext_vector_type(4)));
+    typedef T d8_t __attribute__((ext_vector_type(8)));
+    typedef T d13_t __attribute__((ext_vector_type(13)));
+
+    using type = d13_t;
+
+    union
+    {
+        d13_t d13_;
+        StaticallyIndexedArray<d1_t, 13> d1x13_;
+        StaticallyIndexedArray<d4_t, 3> d4x3_;
+        StaticallyIndexedArray<d8_t, 1> d8x1_;
+        StaticallyIndexedArray<d13_t, 1> d13x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{0}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d4_t>::value ||
+                          is_same<X, d8_t>::value || is_same<X, d13_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x13_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x3_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+        else if constexpr(is_same<X, d13_t>::value)
+        {
+            return data_.d13x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d4_t>::value ||
+                          is_same<X, d8_t>::value || is_same<X, d13_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x13_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x3_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+        else if constexpr(is_same<X, d13_t>::value)
+        {
+            return data_.d13x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
 template <typename T>
 struct vector_type<T, 16, typename std::enable_if_t<is_native_type<T>()>>
 {

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -75,6 +75,28 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances
     // clang-format on
     >;
 
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    48,    64,     32,   8,   16,   16,    3,    4,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              3,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    48,     32,   8,   16,   16,    4,    3,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              3,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    80,     32,   8,   16,   16,    4,    5,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              5,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,   112,     32,   8,   16,   16,    4,    7,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              7,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,   208,     32,   8,   16,   16,    4,   13,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,             13,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
+        // clang-format on
+        >;
+
 template <ck::index_t NDimSpatial,
           typename ALayout,
           typename BLayout,
@@ -118,6 +140,28 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance
     // clang-format on
     >;
 
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregular_instances =
+    std::tuple<
+        // clang-format off
+        //#########################################|              Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
+        //#########################################|              Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|   ToMerge|
+        //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
+        //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     BF16,     BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    48,    64,     32,   8,   16,   16,    3,    4,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              3,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     BF16,     BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    48,     32,   8,   16,   16,    4,    3,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              3,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     BF16,     BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    80,     32,   8,   16,   16,    4,    5,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              5,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     BF16,     BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,   112,     32,   8,   16,   16,    4,    7,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              7,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,     BF16,     BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,   208,     32,   8,   16,   16,    4,   13,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,             13,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
+        // clang-format on
+        >;
+
 template <ck::index_t NDimSpatial,
           typename ALayout,
           typename BLayout,

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -358,6 +358,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                         op_ptrs);
                     add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_instances(
                         op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_irregular_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_irregular_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_BF16
@@ -383,6 +387,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                         op_ptrs);
                     add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instances(
                         op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_irregular_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_irregular_instances(
+                        op_ptrs);
                 }
 #endif
             }
@@ -478,6 +486,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                         op_ptrs);
                     add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instances(
                         op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_BF16
@@ -503,6 +515,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                         op_ptrs);
                     add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_instances(
                         op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_irregular_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_irregular_instances(
+                        op_ptrs);
                 }
 #endif
 #if defined CK_ENABLE_FP16 && defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_xdl.inc

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -149,6 +149,30 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_p
                                                            PassThrough,
                                                            PassThrough>>>& instances);
 
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
 void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
                                                            NGCHW,
@@ -234,6 +258,30 @@ void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pi
                                                            PassThrough,
                                                            PassThrough>>>& instances);
 
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
 void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
                                                            NGCHW,
@@ -384,6 +432,30 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf1
                                                            PassThrough,
                                                            PassThrough>>>& instances);
 
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
 void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
                                                            NGCDHW,
@@ -469,6 +541,30 @@ void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16
                                                            PassThrough,
                                                            PassThrough>>>& instances);
 
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances);
+
 void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
                                                            NGCDHW,

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt

@@ -19,6 +19,10 @@ set(GROUPED_CONV2D_BWD_WEIGHT
     xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp
     xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instance.cpp
     xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_irregular_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_irregular_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_irregular_instance.cpp
+    xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_irregular_instance.cpp
     )
 
 if(DL_KERNELS)

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregular_instances<
+            2,
+            NHWGC,
+            GKYXC,
+            NHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregular_instances<
+            2,
+            NHWGC,
+            GKYXC,
+            NHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v5>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances<
+            2,
+            NHWGC,
+            GKYXC,
+            NHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
+                                                           NHWGC,
+                                                           GKYXC,
+                                                           NHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances<
+            2,
+            NHWGC,
+            GKYXC,
+            NHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v5>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt

@@ -19,6 +19,10 @@ set(GROUPED_CONV3D_BWD_WEIGHT
     xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp
     xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instance.cpp
     xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp
+    xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instance.cpp
+    xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instance.cpp
+    xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_irregular_instance.cpp
+    xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_irregular_instance.cpp
     )
 
 if(DL_KERNELS)

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregular_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           BF16,
+                                                           BF16,
+                                                           BF16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregular_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v5>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
+void add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_irregular_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
+                                                           NDHWGC,
+                                                           GKZYXC,
+                                                           NDHWGK,
+                                                           F16,
+                                                           F16,
+                                                           F16,
+                                                           PassThrough,
+                                                           PassThrough,
+                                                           PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular_instances<
+            3,
+            NDHWGC,
+            GKZYXC,
+            NDHWGK,
+            ConvBwdWeightDefault,
+            BlockGemmPipelineScheduler::Intrawave,
+            BlockGemmPipelineVersion::v5>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -89,6 +89,7 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
     in_device_buf.ToDevice(input.mData.data());
     out_device_buf.ToDevice(output.mData.data());
 
+    float max_accumulated_value = 0;
     if(do_verification)
     {
         auto ref_conv     = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
@@ -114,6 +115,8 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
                                                   {});
 
         ref_invoker.Run(ref_argument);
+        max_accumulated_value =
+            *std::max_element(weight_host_result.mData.begin(), weight_host_result.mData.end());
     }
 
     using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
@@ -237,7 +240,39 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
                 {
                     wei_device_buf.FromDevice(weight_device_result.mData.data());
 
-                    bool pass = ck::utils::check_err(weight_device_result, weight_host_result);
+                    using ComputeType =
+                        std::conditional_t<sizeof(ComputeTypeA) < sizeof(ComputeTypeB),
+                                           ComputeTypeA,
+                                           ComputeTypeB>;
+                    using AccDataType =
+                        std::conditional_t<std::is_same_v<ComputeType, int8_t>, int32_t, float>;
+                    const index_t num_accums         = output.GetElementSize() / conv_param.K_;
+                    const index_t num_accums_split_k = split_k_list[split_k_id];
+                    // Calculate thresholds
+                    auto rtol =
+                        ck::utils::get_relative_threshold<ComputeType, WeiDataType, AccDataType>(
+                            num_accums / num_accums_split_k);
+                    auto atol =
+                        ck::utils::get_absolute_threshold<ComputeType, WeiDataType, AccDataType>(
+                            max_accumulated_value / num_accums_split_k,
+                            num_accums / num_accums_split_k);
+                    // Calculate error due to split_k accumulation
+                    auto rtol_split_k =
+                        ck::utils::get_relative_threshold<WeiDataType, WeiDataType, WeiDataType>(
+                            num_accums_split_k);
+                    auto atol_split_k =
+                        ck::utils::get_absolute_threshold<WeiDataType, WeiDataType, WeiDataType>(
+                            max_accumulated_value, num_accums_split_k);
+                    // Use higher threshold
+                    rtol      = std::max(rtol, rtol_split_k);
+                    atol      = std::max(atol, atol_split_k);
+                    bool pass = ck::utils::check_err(weight_device_result,
+                                                     weight_host_result,
+                                                     "Error: Incorrect results!",
+                                                     rtol,
+                                                     atol);
+                    std::cout << "Relative error threshold: " << rtol
+                              << " Absolute error threshold: " << atol << std::endl;
 
                     if(!pass)
                     {

test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <cstdlib>
 #include <iostream>
@@ -43,7 +43,6 @@ class TestGroupedConvndBwdWeight : public ::testing::Test
                 return true;
             }
         }
-
         if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             // on gfx11 only support for 3d is implemented
@@ -143,19 +142,23 @@ using KernelTypes2d = ::testing::Types<
     std::tuple<float, float, float, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
     std::tuple<ck::half_t, ck::half_t, ck::half_t, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
     std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNHWC, GKYXC, GNHWK, ck::Number<2>>,
+    std::tuple<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
     std::tuple<float, float, float, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
     std::tuple<ck::half_t, ck::half_t, ck::half_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
     std::tuple<ck::bhalf_t, float, ck::bhalf_t, NHWGC, GKYXC, NHWGK, ck::Number<2>>,
+    std::tuple<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t, NGCHW, GKYXC, NGKHW, ck::Number<2>>,
     std::tuple<ck::half_t, ck::half_t, ck::half_t, NGCHW, GKYXC, NGKHW, ck::Number<2>>>;
 using KernelTypes3d = ::testing::Types<
     std::tuple<float, float, float, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
     std::tuple<ck::half_t, ck::half_t, ck::half_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
     std::tuple<ck::bhalf_t, float, ck::bhalf_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
     std::tuple<int8_t, int8_t, int8_t, GNDHWC, GKZYXC, GNDHWK, ck::Number<3>>,
+    std::tuple<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
     std::tuple<float, float, float, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
     std::tuple<ck::half_t, ck::half_t, ck::half_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
     std::tuple<ck::bhalf_t, float, ck::bhalf_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
     std::tuple<int8_t, int8_t, int8_t, NDHWGC, GKZYXC, NDHWGK, ck::Number<3>>,
+    std::tuple<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t, NGCDHW, GKZYXC, NGKDHW, ck::Number<3>>,
     std::tuple<ck::half_t, ck::half_t, ck::half_t, NGCDHW, GKZYXC, NGKDHW, ck::Number<3>>>;
 
 TYPED_TEST_SUITE(TestGroupedConvndBwdWeight1d, KernelTypes1d);
@@ -179,6 +182,8 @@ TYPED_TEST(TestGroupedConvndBwdWeight2d, Test2D)
     this->conv_params.clear();
     this->conv_params.push_back(
         {2, 2, 64, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back({2, 2, 64, 3, 3, {1, 1}, {7, 7}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
+    this->conv_params.push_back({2, 2, 64, 5, 5, {1, 1}, {7, 7}, {1, 1}, {1, 1}, {0, 0}, {0, 0}});
     this->conv_params.push_back(
         {2, 2, 4, 128, 256, {3, 3}, {14, 14}, {1, 1}, {1, 1}, {1, 1}, {1, 1}});
     this->conv_params.push_back(