Added sync for nvidia backend (#84)

Author: Egor-Krivov

dl_bench/utils.py

@@ -153,6 +153,10 @@ def to_device(self, x: torch.Tensor):
         else:
             raise ValueError("Unknown device")
 
+    def sync(self):
+        if self.device_name == 'cuda':
+            torch.cuda.synchronize()
+
     def prepare_eval_transformer(self, model):
         model = model.to(memory_format=torch.channels_last)
 
@@ -394,6 +398,7 @@ def inference(self, backend: Backend):
             # Duration is inconsistent now
             with tm.timeit("duration_s"):
                 for i, x in enumerate(test_loader):
+                    backend.sync()
                     s = get_time()
                     x = backend.to_device(x)
                     if backend.dtype != torch.float32:
@@ -409,6 +414,7 @@ def inference(self, backend: Backend):
                         start = time.perf_counter()
                         continue
 
+                    backend.sync()
                     fw_times.append(get_time() - s)
                     n_items += len(x)
                     outputs.append(y)