add DEBUG_LOG

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_kloop_v1.hpp

@@ -1047,7 +1047,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Kloop_Xdl_CShuffle_V1
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-#if 0
+#if DEBUG_LOG
         arg.Print();
 #endif
 

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_kloop_v2.hpp

@@ -1048,7 +1048,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Kloop_Xdl_CShuffle_V2
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-#if 0
+#if DEBUG_LOG
         arg.Print();
 #endif
 

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_phased_v1.hpp

@@ -1041,7 +1041,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Phased_Xdl_CShuffle_V1
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-#if 0
+#if DEBUG_LOG
         arg.Print();
 #endif
 

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v1.hpp

@@ -1020,6 +1020,9 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+#if DEBUG_LOG
+        arg.Print();
+#endif
 
         if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
              ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||

include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v2.hpp

@@ -1047,6 +1047,9 @@ struct DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+#if DEBUG_LOG
+        arg.Print();
+#endif
 
         if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
              ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||