fix merge bug

composable_kernel/include/utility/in_memory_operation.amd.hpp.in

@@ -149,30 +149,17 @@ struct AtomicAddData
     template <>
     __device__ void Run<AddressSpace::Vgpr, AddressSpace::Global>(const T* p_src,
                                                                   index_t src_offset,
-<<<<<<< HEAD
-                                                                  index_t /* src_range */,
-                                                                  bool src_valid T* p_dst,
-=======
                                                                   bool src_valid,
                                                                   index_t /* src_range */,
                                                                   T* p_dst,
->>>>>>> origin/master
                                                                   index_t dst_offset,
                                                                   bool dst_valid,
                                                                   index_t dst_range) const
     {
         const auto zeros = vector_t(0);
 
-<<<<<<< HEAD
-        amd_buffer_atomic_add<T, DataPerAccess>(src_valid ? &(p_src[src_offset]) : &zeros,
-                                                p_dst,
-                                                dst_offset,
-                                                dst_valid,
-                                                index_t dst_range);
-=======
         amd_buffer_atomic_add<T, DataPerAccess>(
             src_valid ? &(p_src[src_offset]) : &zeros, p_dst, dst_offset, dst_valid, dst_range);
->>>>>>> origin/master
     }
 #endif
 };

driver/include/device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw.hpp

@@ -183,7 +183,7 @@ void device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw(InDesc i
         GemmBBlockCopyDstDataPerWrite_GemmN,
         GemmCThreadCopyDstDataPerWrite_GemmN1>;
 
-    for(index_t i = 0; i < 5; ++i)
+    for(index_t i = 0; i < 1; ++i)
     {
         std::cout << "Start running " << nrepeat << " times..." << std::endl;
 

driver/src/conv_bwd_data_driver.cpp

@@ -241,7 +241,7 @@ int main(int argc, char* argv[])
 #endif
     }
 
-#if 0
+#if 1
     device_convolution_backward_data_implicit_gemm_v1r1_nchw_kcyx_nkhw
 #elif 0
     device_convolution_backward_data_implicit_gemm_v1r2_nchw_kcyx_nkhw