update

Author: Egor-Krivov

dl_bench/cli/launcher.py

@@ -194,6 +194,7 @@ def main():
                 / (10**12)
             )
         )
+        print("FPS: {:.1f}".format(results.get("samples_per_s", 0)))
         pprint.pprint(report)
 
     if args.output is not None:

dl_bench/utils.py

@@ -145,6 +145,7 @@ def __init__(self, device, compiler, dtype="float32") -> None:
         self.dtype = str_to_dtype(dtype)
 
     def to_device(self, x: torch.Tensor):
+        x = x.contiguous(memory_format=torch.channels_last)
         if self.device_name in ("cuda", "xpu"):
             return x.to(self.device)
         elif self.device_name == "cpu":
@@ -230,6 +231,7 @@ def _compile_model(compile_mode: str, device, model: Module, sample_input, dtype
             # enable oneDNN graph fusion globally
             torch.jit.enable_onednn_fusion(True)
             compiled_model = torch.jit.trace(model, sample_input)
+
             compiled_model = torch.jit.freeze(compiled_model)
             print("Compiled with torchscript onednn")
         elif compile_mode == "ipex":

scripts/margin.sh

@@ -18,7 +18,7 @@ do
 	      for name in "${CNNS[@]}"
 	      do
 		  echo "Benchmark $name with BS=$BS and DTYPE=$DTYPE"
-		  numactl -N 1 benchmark-run -b cnn -p "name='${name}',batch_size='$BS'" --dtype "${DTYPE}" --benchmark_desc "${name}_bs$BS" --host "${HOST}" -c "${COMPILER}"
+		  numactl -N 1 benchmark-run -b cnn -p "name='${name}',batch_size='$BS'" --dtype "${DTYPE}" --benchmark_desc "${name}_bs$BS" --host "${HOST}" -c "${COMPILER}" --verbose --skip_verification
 	      done
 	  done
 	done