From cc818b4bc79f31447722884ee42878cf3ad44b71 Mon Sep 17 00:00:00 2001
From: Anton Gorenko <anton@streamhpc.com>
Date: Mon, 28 Apr 2025 13:03:48 +0500
Subject: [PATCH] Fix strides of a/b_thread_desc

The descriptors are larger than needed (even though the compiler don't alloc registers for unused values).
---
 .../blockwise_gemm_pipeline_wmmaops_base.hpp     | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp
index 0147a890dd..6c0cc8baea 100644
--- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp
@@ -257,10 +257,10 @@ struct BlockwiseGemmWmmaops_pipeline_base
                                                 Number<A_K1>{}),
                                      make_tuple(Number<A_K1>{},
                                                 Number<KPack / A_KRow>{},
-                                                Number<KPack * A_K1>{},
-                                                Number<A_K1>{},
-                                                Number<A_K1>{},
-                                                Number<1>{}));
+                                                Number<KPack / A_KRow * MRepeat>{},
+                                                I0,
+                                                I0,
+                                                I1));
 
     static constexpr auto b_thread_desc_ =
         make_naive_tensor_descriptor(make_tuple(Number<KPack / B_K1 / B_KRow>{},
@@ -271,10 +271,10 @@ struct BlockwiseGemmWmmaops_pipeline_base
                                                 Number<B_K1>{}),
                                      make_tuple(Number<B_K1>{},
                                                 Number<KPack / B_KRow>{},
-                                                Number<KPack * B_K1>{},
-                                                Number<B_K1>{},
-                                                Number<B_K1>{},
-                                                Number<1>{}));
+                                                Number<KPack / B_KRow * NRepeat>{},
+                                                I0,
+                                                I0,
+                                                I1));
 
     // C[M, N, NumRegWmma]
     static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(