From 66f93b6e08054e4c8a2c8999e1417fdda290c156 Mon Sep 17 00:00:00 2001
From: aska-0096 <haocwang@amd.com>
Date: Mon, 12 May 2025 03:24:06 +0000
Subject: [PATCH] tempsave

---
 .../grid/gridwise_gemm_xdl_cshuffle_v3_mx.hpp | 54 +++----------------
 1 file changed, 8 insertions(+), 46 deletions(-)

diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx.hpp
index 586d99cd64..77d554d18e 100644
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_mx.hpp
@@ -768,39 +768,20 @@ struct GridwiseGemmMX_xdl_cshuffle_v3
         // in some cases.
         else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
         {
-            constexpr index_t LdsSize = 32 * 4 / (KPerBlock * sizeof(ADataType) / APackedSize);
-            constexpr auto MLdsLayer  = LdsSize < 1 ? 1 : LdsSize;
             constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
                 make_tuple(
-                    AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
-                make_tuple(AK1Number, Number<KPerBlock * MLdsLayer>{}, I1));
+                    AK0Number, Number<MPerBlock>{}, AK1Number),
+                make_tuple(AK1Number, Number<KPerBlock>{}, I1));
 
             constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
                 a_lds_block_desc,
                 make_tuple(make_xor_with_modulo_transform(make_tuple(
-                               Number<MPerBlock / MLdsLayer>{}, Number<AK0Number * MLdsLayer>{})),
+                               Number<MPerBlock>{}, Number<AK0Number>{})),
                            make_pass_through_transform(AK1Number)),
                 make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
                 make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
 
-            constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor(
-                a_lds_block_desc_permuted,
-                make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number<MLdsLayer>{})),
-                           make_pass_through_transform(Number<MPerBlock / MLdsLayer>{}),
-                           make_pass_through_transform(AK1Number)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
-
-            constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor(
-                a_lds_block_desc_ak0_mldslayer_m_ak1,
-                make_tuple(make_pass_through_transform(AK0Number),
-                           make_merge_transform_v3_division_mod(
-                               make_tuple(Number<MPerBlock / MLdsLayer>{}, Number<MLdsLayer>{})),
-                           make_pass_through_transform(AK1Number)),
-                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-
-            return a_lds_block_desc_ak0_m_ak1;
+            return a_lds_block_desc_permuted;
         }
         else // ColumnMajor A
         {
@@ -904,39 +885,20 @@ struct GridwiseGemmMX_xdl_cshuffle_v3
         else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
         {
             // NLdsLayer * K0 as logical Bank
-            constexpr index_t LdsSize   = 32 * 4 / (KPerBlock * sizeof(BDataType) / BPackedSize);
-            constexpr index_t NLdsLayer = LdsSize < 1 ? 1 : LdsSize;
             constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
                 make_tuple(
-                    BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
-                make_tuple(BK1Number, Number<KPerBlock * NLdsLayer>{}, I1));
+                    BK0Number, Number<NPerBlock>{}, BK1Number),
+                make_tuple(BK1Number, Number<KPerBlock>{}, I1));
 
             constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
                 b_lds_block_desc,
                 make_tuple(make_xor_with_modulo_transform(make_tuple(
-                               Number<NPerBlock / NLdsLayer>{}, Number<BK0Number * NLdsLayer>{})),
+                               Number<NPerBlock>{}, Number<BK0Number>{})),
                            make_pass_through_transform(BK1Number)),
                 make_tuple(Sequence<1, 0>{}, Sequence<2>{}),
                 make_tuple(Sequence<1, 0>{}, Sequence<2>{}));
 
-            constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor(
-                b_lds_block_desc_permuted,
-                make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number<NLdsLayer>{})),
-                           make_pass_through_transform(Number<NPerBlock / NLdsLayer>{}),
-                           make_pass_through_transform(BK1Number)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
-                make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}));
-
-            constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor(
-                b_lds_block_desc_bk0_nldslayer_n_bk1,
-                make_tuple(make_pass_through_transform(BK0Number),
-                           make_merge_transform_v3_division_mod(
-                               make_tuple(Number<NPerBlock / NLdsLayer>{}, Number<NLdsLayer>{})),
-                           make_pass_through_transform(BK1Number)),
-                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-
-            return b_lds_block_desc_bk0_n_bk1;
+            return b_lds_block_desc_permuted;
         }
         else // RowMajor B
         {