Merge pull request buddy-compiler#2 from axmat/onnxruntime

[examples][conv-opt] Add conv2d benchmark for ONNX Runtime

Author: zhanghb97

.gitignore

@@ -2,4 +2,7 @@
 /build*
 
 # vscode configurations
-/.vscode
+/.vscode
+
+# ONNX models
+*.onnx

examples/conv-opt/comparison/README.md

@@ -43,6 +43,14 @@ $ # n kernel size, default 3, alternative options: 3, 5, 7, 9
 $ # target, default 'llvm -mcpu=skylake-avx512', alternative options: 'cuda'
 ```
 
+- ONNX Runtime
+
+```
+$ cd buddy-mlir/examples/conv-opt/comparison/
+$ python3 gen-conv-models.py
+$ python3 onnxruntime-conv2d.py
+```
+
 ## Difference and Performance Comparison
 
 We proposed a coefficients broadcasting algorithm with strip mining strategy (CB-SM) to accelerate 2D convolutions, 

examples/conv-opt/comparison/gen-conv-models.py

@@ -0,0 +1,89 @@
+import onnx
+from onnx import numpy_helper
+import numpy as np
+
+# Filter
+sobel = {
+  3: np.array([[1, 0, -1],
+    [2, 0, -2],
+    [1, 0, -1]], dtype='float32'),
+  5: np.array([[2, 1, 0, -1, -2],
+    [3, 2, 0, -2, -3],
+    [4, 3, 0, -3, -4],
+    [3, 2, 0, -2, -3],
+    [2, 1, 0, -1, -2]], dtype='float32'),
+  7: np.array([[3, 2, 1, 0, -1, -2, -3],
+    [4, 3, 2, 0, -2, -3, -4],
+    [5, 4, 3, 0, -3, -4, -5],
+    [6, 5, 4, 0, -4, -5, -6],
+    [5, 4, 3, 0, -3, -4, -5],
+    [4, 3, 2, 0, -2, -3, -4],
+    [3, 2, 1, 0, -1, -2, -3]], dtype='float32'),
+  9: np.array([[4, 3, 2, 1, 0, -1, -2, -3, -4],
+    [5, 4, 3, 2, 0, -2, -3, -4, -5],
+    [6, 5, 4, 3, 0, -3, -4, -5, -6],
+    [7, 6, 5, 4, 0, -4, -5, -6, -7],
+    [8, 7, 6, 5, 0, -5, -6, -7, -8],
+    [7, 6, 5, 4, 0, -4, -5, -6, -7],
+    [6, 5, 4, 3, 0, -3, -4, -5, -6],
+    [5, 4, 3, 2, 0, -2, -3, -4, -5],
+    [4, 3, 2, 1, 0, -1, -2, -3, -4]], dtype='float32')
+}
+
+def get_output_shape(i):
+  if i == 3:
+    return [1, 1, 2046, 2046]
+  elif i == 5:
+    return [1, 1, 2044, 2044]
+  elif i == 7:
+    return [1, 1, 2042, 2042]
+  elif i == 9:
+    return [1, 1, 2040, 2040]
+
+def main():
+  for i in range(3, 10, 2):
+    # Filter
+    w = sobel[i].reshape((1, 1, i, i))
+
+    # Input
+    x = np.random.rand(1, 1, 2048, 2048).astype('float32')
+
+    # Initializer of the weight
+    initializer_w = numpy_helper.from_array(w, 'w')
+
+    tensor_w = onnx.helper.make_tensor_value_info('w', onnx.TensorProto.FLOAT, [1, 1, i, i])
+    tensor_x = onnx.helper.make_tensor_value_info('x', onnx.TensorProto.FLOAT, [1, 1, 2048, 2048])
+    tensor_y = onnx.helper.make_tensor_value_info('y', onnx.TensorProto.FLOAT, get_output_shape(i))
+
+    # Create a node
+    node_def = onnx.helper.make_node(
+      'Conv',
+      inputs=['x', 'w'],
+      outputs=['y'],
+      kernel_shape=[i, i]
+    )
+
+    # Create the graph
+    graph_def = onnx.helper.make_graph(
+      [node_def],
+      f'conv_{i}x{i}',
+      [tensor_x],
+      [tensor_y],
+      [initializer_w]
+    )
+
+    # Create the model
+    model_def = onnx.helper.make_model(graph_def,
+      producer_name='python_script',
+      ir_version=6
+    )
+    model_def.opset_import[0].version = 10
+
+    # Check the model
+    onnx.checker.check_model(model_def)
+
+    # Save the model
+    onnx.save(model_def, f'conv_{i}x{i}.onnx')
+
+if __name__ == "__main__":
+  main()

examples/conv-opt/comparison/onnxruntime-conv2d.py

@@ -0,0 +1,33 @@
+import numpy as np
+import cv2
+import onnxruntime
+import time
+
+def test_conv2d(img, filter_size):
+  start = time.time()
+  # Load the model
+  model_path = f'conv_{filter_size}x{filter_size}.onnx'
+  ort_session = onnxruntime.InferenceSession(model_path)
+  # Run inference
+  ort_inputs = {ort_session.get_inputs()[0].name: img}
+  ort_outs = ort_session.run(None, ort_inputs)
+  edge_detect = ort_outs[0]
+  edge_detect = edge_detect.squeeze()
+  end = time.time()
+  print(f'conv {filter_size}x{filter_size} : {end - start}')
+  return edge_detect
+
+def main():
+  img = cv2.imread('../images/YuTu2048.png',cv2.IMREAD_GRAYSCALE)
+  # Convert the image to numpy array.
+  img = np.array(img, dtype='float32')
+  img = img.reshape((1, 1, img.shape[0], img.shape[1]))
+  '''
+  Perform the edget detection.
+  '''
+  for i in range(3, 10, 2):
+    edge_detect = test_conv2d(img, i)
+    cv2.imwrite(f'./onnxruntime-conv2d_{i}.png', edge_detect)
+
+if __name__ == "__main__":
+  main()