MLBEDSW-4022: support PAD followed by pool operator
PAD followed by max/average pool is run on NPU if NPU
padding can be used. Average pool is converted to depthwise.
Change-Id: Icc3652e6d9ecff5ac3dc7d92080313d90c245404
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
diff --git a/ethosu/vela/test/test_graph_optimiser.py b/ethosu/vela/test/test_graph_optimiser.py
index 40b8cd5..285b3ac 100644
--- a/ethosu/vela/test/test_graph_optimiser.py
+++ b/ethosu/vela/test/test_graph_optimiser.py
@@ -1,4 +1,4 @@
-# Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved.
+# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0
#
@@ -157,6 +157,53 @@
assert pad_op not in op.ifm.ops
+def test_optimise_pad_followed_by_avg_pool():
+ """
+ Tests that the PAD operator is bypassed when followed by a average pool operator,
+ and that the average pool is converted to a depthwise
+ """
+ # Create Pad operation followed by AvgPool
+ quant = testutil.default_quant_params()
+ in_tens = Tensor([1, 76, 75, 64], DataType.uint8, "input")
+ in_tens.quantization = quant
+ pad_input = create_const_tensor("pad_input", [4, 2], DataType.int32, [[0, 0], [2, 1], [1, 1], [0, 0]])
+ temp_tens = Tensor([1, 79, 77, 64], DataType.uint8, "pad_out")
+ temp_tens.quantization = quant.clone()
+ out_tens = Tensor([1, 76, 75, 64], DataType.uint8, "output")
+ out_tens.quantization = quant.clone()
+
+ pad_op = testutil.create_op(Op.Pad, [in_tens, pad_input], temp_tens)
+ attrs = {
+ "padding": Padding.VALID,
+ "ksize": [1, 5, 3, 1],
+ "stride_w": 2,
+ "stride_h": 2,
+ "dilation_w_factor": 1,
+ "dilation_h_factor": 1,
+ }
+ attrs["strides"] = (1, attrs["stride_h"], attrs["stride_w"], 1)
+ pad_op.run_on_npu = True
+ conv2d_op = testutil.create_op(Op.AvgPool, [temp_tens], out_tens, attrs)
+ conv2d_op.run_on_npu = True
+ nng = Graph()
+ sg = testutil.create_subgraph([pad_op, conv2d_op])
+ nng.subgraphs.append(sg)
+ arch = testutil.create_arch()
+
+ optimise_pad(conv2d_op, nng, arch)
+
+ op = sg.output_tensors[0].ops[0]
+ assert op.type == Op.DepthwiseConv2DBias
+ assert op.attrs["padding"] == Padding.EXPLICIT
+ assert op.attrs["explicit_padding"] == (2, 1, 1, 1)
+ assert op.ifm.shape == [1, 76, 75, 64]
+ assert pad_op not in op.ifm.ops
+ # Check that bias and weight tensors have been added
+ assert op.bias.shape == [64]
+ print("op.weights:", op.weights)
+ assert op.weights.shape == [5, 3, 1, 64]
+
+
def test_remove_reshape():
"""
Tests that the expected reshape are removed in graph_optimisation
diff --git a/ethosu/vela/test/test_supported_operators.py b/ethosu/vela/test/test_supported_operators.py
index 3e9724d..6401d29 100644
--- a/ethosu/vela/test/test_supported_operators.py
+++ b/ethosu/vela/test/test_supported_operators.py
@@ -609,14 +609,7 @@
op_consumer = testutil.create_op_with_quant_tensors(Op.ConcatTFLite, [1, 1, 1, 4], [1, 1, 1, 8])
op.ofm.consumer_list = [op_consumer]
assert not support.is_operator_supported(op)
- op_consumer = testutil.create_op_with_quant_tensors(Op.AvgPool, [1, 8, 8, 8], [1, 8, 8, 8])
- op_consumer.attrs = {
- "stride_w": 2,
- "stride_h": 2,
- "filter_width": 2,
- "filter_height": 2,
- "padding": Padding.VALID,
- }
+ op_consumer = testutil.create_elemwise_op(Op.Add, "op", [1, 3, 3, 1], [1, 3, 3, 1], [1, 3, 3, 1])
op.ofm.consumer_list = [op_consumer]
assert not support.is_operator_supported(op)
@@ -655,6 +648,55 @@
assert support.is_operator_supported(op) == expected
+pad_avg_pool_test_data = [
+ ((3, 3), (1, 1, 1, 1), True),
+ ((2, 4), (1, 2, 1, 2), True),
+ ((5, 3), (2, 1, 2, 1), True),
+ ((5, 3), (0, 1, 2, 1), True),
+ ((5, 3), (2, 0, 2, 1), True),
+ ((5, 3), (2, 1, 0, 1), True),
+ ((5, 3), (2, 1, 0, 1), True),
+ ((4, 4), (2, 2, 2, 2), True),
+ ((4, 4), (1, 2, 2, 2), False),
+ ((4, 4), (2, 1, 2, 2), False),
+ ((4, 4), (2, 2, 1, 2), False),
+ ((4, 4), (2, 2, 2, 1), False),
+]
+
+
+@pytest.mark.parametrize("k_size, padding, expected", pad_avg_pool_test_data)
+def test_pad_followed_by_avg_pool(k_size, padding, expected):
+ # Tests PAD followed by AvgPool
+ k_w, k_h = k_size
+ top, left, bottom, right = padding
+ pad_values = [[0, 0], [top, bottom], [left, right], [0, 0]]
+ dtype = DataType.int8
+ qp = testutil.default_quant_params()
+ in_shape = [1, 15, 17, 8]
+ out_shape = [1, in_shape[1] + top + bottom, in_shape[2] + left + right, in_shape[3]]
+ in0 = Tensor(in_shape, dtype, "in")
+ in0.quantization = qp
+ pad_tensor = create_const_tensor(
+ name="pad", shape=list(np.shape(pad_values)), values=pad_values, dtype=DataType.int32
+ )
+ out = Tensor(out_shape, dtype, "out")
+ out.quantization = qp.clone()
+ op = testutil.create_op(Op.Pad, [in0, pad_tensor], out)
+ pool_out_tens = Tensor(in_shape, dtype, "output")
+ pool_out_tens.quantization = qp.clone()
+ attrs = {
+ "padding": Padding.VALID,
+ "ksize": [1, k_w, k_h, 1],
+ "stride_w": 1,
+ "stride_h": 1,
+ "dilation_w_factor": 1,
+ "dilation_h_factor": 1,
+ }
+ pool_op = testutil.create_op(Op.AvgPool, [out], pool_out_tens, attrs)
+ pool_op.add_input_tensor(out)
+ assert support.is_operator_supported(op) == expected
+
+
def create_strided_slice():
# Creates a valid strided slice operator with some valid inputs/outputs
op = create_strided_slice_op([1, 10, 10, 10], [1, 5, 5, 10], [127, 2, 2, 0], [0, 7, -3, 0])