From e61f956fc2def843366fe6320809eeffd15ba002 Mon Sep 17 00:00:00 2001
From: "assistant-librarian[bot]"
 <assistant-librarian[bot]@users.noreply.github.com>
Date: Fri, 4 Jul 2025 14:07:12 +0000
Subject: [PATCH] Merge commit '3d70c638d1a217869fe0d90636232d239786b4e5' into
 develop

---
 ...uped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
index efb91bd13d..89a304fda4 100644
--- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
@@ -611,7 +611,19 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                 // If stride is larger than windows size then we will have some empty places
                 image_covered_strides &= conv_filter_strides[d] <= b_g_k_c_xs_lengths[d + I3];
             }
+            bool if_d_is_output_mem  = false;
+            const void* out_mem_void = static_cast<const void*>(p_e);
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                if(p_ds[i] == out_mem_void)
+                {
+                    if_d_is_output_mem = true;
+                }
+            });
+
             bwd_needs_zero_out = k_batch_ > 1 || !image_covered_dilation || !image_covered_strides;
+
+            // Temporary workaround untill prove/fix above conditions.
+            bwd_needs_zero_out = !if_d_is_output_mem;
             e_space_size_bytes =
                 ck::accumulate_n<long_index_t>(
                     e_g_n_c_wis_lengths_.begin(), NDimSpatial + I3, 1, std::multiplies<>()) *