Cudnn (#113)

Add cudnn support

benchmarker/modules/do_cudnn.py

@@ -0,0 +1,57 @@
+"""Module contains the interface for all deep learning modules"""
+import importlib
+from pathlib import Path
+
+from benchmarker.util.abstractprocess import Process
+
+
+class Benchmark:
+    """Interface for all deep learning modules"""
+
+    def __init__(self, params, remaining_args=None):
+        path_params = f"benchmarker.modules.problems.{params['problem']['name']}.params"
+        try:
+            module_params = importlib.import_module(path_params)
+            module_params.set_extra_params(params, remaining_args)
+        except ImportError:
+            assert remaining_args == []
+        assert params["problem"]["name"] == "conv", (
+            "only conv problem is defined for this framework, "
+            f"not {params['problem']['name']}"
+        )
+        assert (
+            "nb_gpus" in params and params["nb_gpus"] == 1
+        ), "cuDNN requires exactly one GPU"
+        self.params = params
+
+    def run(self):
+        bin_path = f"problems/{self.params['problem']['name']}/main"
+        cmd_binary = Path(__file__).parent.joinpath(bin_path)
+        if not cmd_binary.is_file():
+            raise (RuntimeError(f"{cmd_binary} not found, run make manually"))
+
+        problem = self.params["problem"]
+        command = [cmd_binary]
+        command.append(self.params["gpus"][0])
+        command.append(problem["cudnn_conv_algo"])
+        command.append(problem["nb_dims"])
+        command.append(problem["cnt_samples"])
+        command.append(problem["input_channels"])
+        command.append(problem["cnt_filters"])
+        command += problem["input_size"]
+        command += problem["size_kernel"]
+        command += problem["stride"]
+        command += problem["dilation"]
+        command += problem["padding"]
+        command = list(map(str, command))
+
+        process = Process(command=command)
+        result = process.get_output()
+        print(result)
+        assert result["returncode"] == 0 and result["err"] == "", (
+            "Error from binary!\n"
+            f"retcode: {result['returncode']}\nstderr: {result['err']}"
+        )
+        elapsed_time = float(result["out"].strip())
+        self.params["time"] = elapsed_time
+        # self.params["GFLOP/sec"] = self.params["GFLOP"] / elapsed_time