refactor

driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw.hpp

@@ -487,6 +487,8 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(
                                                        in_left_pads,
                                                        in_right_pads);
 
+    for(index_t i = 0; i < 5; ++i)
+    {
         float ave_time = launch_kernel_dynamic_gemm_v1<
             BlockSize,
             typename vector_type<TInWei, InWeiVectorSize>::type,
@@ -552,6 +554,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nchw_kcyx_nkhw(
                      (std::size_t(1000) * 1000 * 1000) / ave_time;
 
         std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
+    }
 
     // copy result back to host
     out_n_k_ho_wo_device_buf.FromDevice(out_n_k_ho_wo.mData.data());

driver/include/device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk.hpp

@@ -393,6 +393,8 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(
                                                        in_left_pads,
                                                        in_right_pads);
 
+    for(index_t i = 0; i < 5; ++i)
+    {
         float ave_time = launch_kernel_dynamic_gemm_v1<
             BlockSize,
             typename vector_type<TInWei, InWeiVectorSize>::type,
@@ -457,6 +459,7 @@ void device_dynamic_convolution_forward_implicit_gemm_v4r4_nhwc_kyxc_nhwk(
                      (std::size_t(1000) * 1000 * 1000) / ave_time;
 
         std::cout << "Average time : " << ave_time << " ms, " << perf << " TFlop/s" << std::endl;
+    }
 
     // copy result back to host
     out_n_ho_wo_k_device_buf.FromDevice(out_n_ho_wo_k.mData.data());