resolved conflicts

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(