diff options
Diffstat (limited to 'ethosu/vela/test')
| -rw-r--r-- | ethosu/vela/test/test_graph_optimiser.py | 49 | ||||
| -rw-r--r-- | ethosu/vela/test/test_supported_operators.py | 58 |
2 files changed, 98 insertions, 9 deletions
diff --git a/ethosu/vela/test/test_graph_optimiser.py b/ethosu/vela/test/test_graph_optimiser.py index 40b8cd5d..285b3ac5 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 @@ def test_optimise_pad(): 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 3e9724d3..6401d29d 100644 --- a/ethosu/vela/test/test_supported_operators.py +++ b/ethosu/vela/test/test_supported_operators.py @@ -609,14 +609,7 @@ def test_constraint_pad_consumer(): 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 @@ def test_constraint_leading_pad_size(top, left, kernel_size, expected): 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]) |