Fixed splitk gemm fp32 (#384)

* add scripts

* fixed splitK_gemm_fp32

* clean

* clean

include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp

@@ -93,9 +93,9 @@ struct DeviceGemmXdlSplitK : public DeviceGemmSplitK<ALayout,
 
         const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
             a_grid_desc_m_k,
-            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(M)),
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
-            make_tuple(Sequence<0>{}, Sequence<1>{}));
+            make_tuple(Sequence<1>{}, Sequence<0>{}));
 
         if constexpr(GemmSpec == GemmSpecialization::MNPadding)
         {

script/run_full_performance_tests.sh

@@ -127,10 +127,10 @@ print_log_header $reduction_log $env_type $branch $host_name
 export splitK_gemm_log="perf_splitK_gemm_${gpu_arch}.log"
 print_log_header $splitK_gemm_log $env_type $branch $host_name
 
-#../script/profile_splitK_gemm.sh gemm_splitk 0 0 $verify 1 0 1 4 | tee -a $splitK_gemm_log
-#../script/profile_splitK_gemm.sh gemm_splitk 0 1 $verify 1 0 1 4 | tee -a $splitK_gemm_log
-#../script/profile_splitK_gemm.sh gemm_splitk 0 2 $verify 1 0 1 4 | tee -a $splitK_gemm_log
-#../script/profile_splitK_gemm.sh gemm_splitk 0 3 $verify 1 0 1 4 | tee -a $splitK_gemm_log
+../script/profile_splitK_gemm.sh gemm_splitk 0 0 $verify 1 0 1 4 | tee -a $splitK_gemm_log
+../script/profile_splitK_gemm.sh gemm_splitk 0 1 $verify 1 0 1 4 | tee -a $splitK_gemm_log
+../script/profile_splitK_gemm.sh gemm_splitk 0 2 $verify 1 0 1 4 | tee -a $splitK_gemm_log
+../script/profile_splitK_gemm.sh gemm_splitk 0 3 $verify 1 0 1 4 | tee -a $splitK_gemm_log
 
 ../script/profile_splitK_gemm.sh gemm_splitk 1 0 $verify 1 0 1 4 | tee -a $splitK_gemm_log
 ../script/profile_splitK_gemm.sh gemm_splitk 1 1 $verify 1 0 1 4 | tee -a $splitK_gemm_log