From 0fdbf8a91d10df00a0635c3f2bd3959de0a126de Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Bart=C5=82omiej=20Kocot?=
 <38502616+bartekxk@users.noreply.github.com>
Date: Thu, 11 Jun 2026 15:28:21 +0000
Subject: [PATCH] [rocm-libraries] ROCm/rocm-libraries#8272 (commit 1c66ecb)

[CK] Padding on K for global load for grouped conv bwd data
 (#8272)

## Motivation

Fix incorrect results caused by lack of padding during global load in
grouped convolution backward data kernel. It is needed since there is no
OOB check for global load.

## Technical Details

Add padding needed for global load which not use OOB check.

## Test Plan

test_grouped_convnd_bwd_data*

## Test Result

Passed locally

## Submission Checklist

- [x] Look over the contributing guidelines at
https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.
---
 .../transform_conv_bwd_data_to_gemm_v1.hpp          | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
index 52d5bebcdf..50a13e297f 100644
--- a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
+++ b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
@@ -42,6 +42,7 @@ struct TransformConvBwdDataToGemm_v1
      * in convolution backward data operations.
      */
     static constexpr bool CustomTensorTransformBwdData = true;
+    static constexpr bool UsePaddingOnKDim             = std::is_same_v<IndexType, long_index_t>;
 
     template <index_t N>
     using NumberType =
@@ -694,7 +695,7 @@ struct TransformConvBwdDataToGemm_v1
                 ck::tensor_operation::device::PadTensorDescriptor(
                     out_gemmak0_gemmmraw_gemmak1_grid_desc,
                     make_tuple(AK0 * batch_k_, GemmMPerBlock, AK1),
-                    Sequence<false, DoPadGemmM, false>{});
+                    Sequence<UsePaddingOnKDim, DoPadGemmM, UsePaddingOnKDim>{});
 
             return out_gemmak0_gemmm_gemmak1_grid_desc;
         }
@@ -993,7 +994,7 @@ struct TransformConvBwdDataToGemm_v1
                 ck::tensor_operation::device::PadTensorDescriptor(
                     wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
                     make_tuple(BK0 * batch_k_, GemmNPerBlock, BK1),
-                    Sequence<false, DoPadGemmN, false>{});
+                    Sequence<UsePaddingOnKDim, DoPadGemmN, UsePaddingOnKDim>{});
 
             return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
         }
@@ -1573,7 +1574,7 @@ struct TransformConvBwdDataToGemm_v1
             return ck::tensor_operation::device::PadTensorDescriptor(
                 out_gemmk0_gemmm_gemmk1_grid_desc,
                 make_tuple(Number<GemmKPerBlock / AK1>{}, Number<GemmMPerBlock>{}, Number<AK1>{}),
-                Sequence<true, DoPadGemmM, false>{});
+                Sequence<true, DoPadGemmM, UsePaddingOnKDim>{});
         }
         else
         {
@@ -1645,7 +1646,7 @@ struct TransformConvBwdDataToGemm_v1
             return ck::tensor_operation::device::PadTensorDescriptor(
                 out_gemmk0_gemmm_gemmk1_grid_desc,
                 make_tuple(Number<GemmKPerBlock / AK1>{}, Number<GemmMPerBlock>{}, Number<AK1>{}),
-                Sequence<true, DoPadGemmM, false>{});
+                Sequence<true, DoPadGemmM, UsePaddingOnKDim>{});
         }
     }
 
@@ -1671,7 +1672,7 @@ struct TransformConvBwdDataToGemm_v1
             return ck::tensor_operation::device::PadTensorDescriptor(
                 wei_gemmk0_gemmn_gemmk1_grid_desc,
                 make_tuple(Number<GemmKPerBlock / BK1>{}, Number<GemmNPerBlock>{}, Number<BK1>{}),
-                Sequence<true, DoPadGemmN, false>{});
+                Sequence<true, DoPadGemmN, UsePaddingOnKDim>{});
         }
         else
         {
@@ -1724,7 +1725,7 @@ struct TransformConvBwdDataToGemm_v1
             return ck::tensor_operation::device::PadTensorDescriptor(
                 wei_gemmk0_gemmn_gemmk1_grid_desc,
                 make_tuple(Number<GemmKPerBlock / BK1>{}, Number<GemmNPerBlock>{}, Number<BK1>{}),
-                Sequence<true, DoPadGemmN, false>{});
+                Sequence<true, DoPadGemmN, UsePaddingOnKDim>{});
         }
     }