IVGCVSW-5831 Add additional options to Arm NN External Delegate
* Added enable-fast-math and number-of-threads options.
* Added save-cached-network and cached-network-filepath options.
* Added external_delegate python tests for new options.
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: I7cf6522a6f895cd71ed8f369d94a5113d78594f9
diff --git a/delegate/python/test/test_external_delegate.py b/delegate/python/test/test_external_delegate.py
index 93d373d..f01a2d3 100644
--- a/delegate/python/test/test_external_delegate.py
+++ b/delegate/python/test/test_external_delegate.py
@@ -57,6 +57,115 @@
# cleanup
os.remove(tuning_file)
+@pytest.mark.GpuAccTest
+def test_external_delegate_options_gpu_cached_network(delegate_dir, test_data_folder, tmp_path):
+
+ binary_file = os.path.join(str(tmp_path), "test_binary.bin")
+ # cleanup previous test run if necessary
+ if os.path.exists(binary_file):
+ os.remove(binary_file)
+
+ # Create blank binary file to write to.
+ open(binary_file, 'a').close()
+ assert (os.path.exists(binary_file))
+ assert (os.stat(binary_file).st_size == 0)
+
+ # Run inference to save cached network.
+ armnn_delegate = tflite.load_delegate(
+ delegate_dir,
+ options={
+ "backends": "GpuAcc",
+ "save-cached-network": "1",
+ "cached-network-filepath": binary_file,
+ "logging-severity": "info"})
+
+ run_mock_model(armnn_delegate, test_data_folder)
+
+ # destroy delegate and check if file has been saved.
+ armnn_delegate.__del__()
+ assert (os.stat(binary_file).st_size != 0)
+
+ # Create second delegate to load in binary file created.
+ armnn_delegate2 = tflite.load_delegate(
+ delegate_dir,
+ options={
+ "backends": "GpuAcc",
+ "cached-network-filepath": binary_file,
+ "logging-severity": "info"})
+
+ run_mock_model(armnn_delegate2, test_data_folder)
+
+ # cleanup
+ os.remove(binary_file)
+
+@pytest.mark.GpuAccTest
+def test_external_delegate_gpu_fastmath(delegate_dir, test_data_folder):
+ # create armnn delegate with enable-fast-math
+ # fast-math is only enabled on Conv2d layer, so use conv2d model.
+ armnn_delegate = tflite.load_delegate(delegate_dir, options = {'backends': 'GpuAcc',
+ 'enable-fast-math': '1',
+ "logging-severity": "info"})
+
+ model_file_name = 'conv2d.tflite'
+
+ inputShape = [ 1, 5, 5, 1 ]
+ outputShape = [ 1, 3, 3, 1 ]
+
+ inputValues = [ 1, 5, 2, 3, 5,
+ 8, 7, 3, 6, 3,
+ 3, 3, 9, 1, 9,
+ 4, 1, 8, 1, 3,
+ 6, 8, 1, 9, 2 ]
+
+ expectedResult = [ 28, 38, 29,
+ 96, 104, 53,
+ 31, 55, 24 ]
+
+ input = np.array(inputValues, dtype=np.float32).reshape(inputShape)
+ expected_output = np.array(expectedResult, dtype=np.float32).reshape(outputShape)
+
+ # run the inference
+ armnn_outputs = run_inference(test_data_folder, model_file_name, [input], [armnn_delegate])
+
+ # check results
+ compare_outputs(armnn_outputs, [expected_output])
+
+@pytest.mark.CpuAccTest
+def test_external_delegate_cpu_options(capfd, delegate_dir, test_data_folder):
+ # create armnn delegate with enable-fast-math and number-of-threads options
+ # fast-math is only enabled on Conv2d layer, so use conv2d model.
+ armnn_delegate = tflite.load_delegate(delegate_dir, options = {'backends': 'CpuAcc',
+ 'enable-fast-math': '1',
+ 'number-of-threads': '4',
+ "logging-severity": "info"})
+
+ model_file_name = 'conv2d.tflite'
+
+ inputShape = [ 1, 5, 5, 1 ]
+ outputShape = [ 1, 3, 3, 1 ]
+
+ inputValues = [ 1, 5, 2, 3, 5,
+ 8, 7, 3, 6, 3,
+ 3, 3, 9, 1, 9,
+ 4, 1, 8, 1, 3,
+ 6, 8, 1, 9, 2 ]
+
+ expectedResult = [ 28, 38, 29,
+ 96, 104, 53,
+ 31, 55, 24 ]
+
+ input = np.array(inputValues, dtype=np.float32).reshape(inputShape)
+ expected_output = np.array(expectedResult, dtype=np.float32).reshape(outputShape)
+
+ # run the inference
+ armnn_outputs = run_inference(test_data_folder, model_file_name, [input], [armnn_delegate])
+
+ # check results
+ compare_outputs(armnn_outputs, [expected_output])
+
+ captured = capfd.readouterr()
+ assert 'Set CPPScheduler to Linear mode, with 4 threads to use' in captured.out
+
def test_external_delegate_options_wrong_logging_level(delegate_dir):
with pytest.raises(ValueError):
tflite.load_delegate(