From 850b9adbf977ba795e239e10d9aa04783e2e49aa Mon Sep 17 00:00:00 2001
From: joye <joye@amd.com>
Date: Tue, 3 Jun 2025 12:56:38 +0800
Subject: [PATCH] current perf best kernel

---
 ...evice_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

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 024ea6645b..27e5c4af3f 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
@@ -1058,7 +1058,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                 auto launch_kernel = [&]() {
                     // constexpr bool has_main_loop = has_main_k_block_loop.value;
                     constexpr index_t GroupPerBlock = 64;
-                    constexpr index_t BatchPerBlock = 64;
+                    constexpr index_t BatchPerBlock = 8;
                     const auto kernel = kernel_grouped_conv_bwd_data_optimized<ADataType,
                                                                                EDataType,
                                                                                GroupPerBlock,
@@ -1066,7 +1066,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
                     return launch_and_time_kernel(stream_config,
                                                   kernel,
-                                                  dim3(1344 / GroupPerBlock * 2, 1, 1),
+                                                  dim3(1344 / GroupPerBlock * 16, 1, 1),
                                                   dim3(512),
                                                   0,
                                                   p_a_grid,