|
| 1 | +# Copyright 2026 Arm Limited and/or its affiliates. |
| 2 | +# |
| 3 | +# This source code is licensed under the BSD-style license found in the |
| 4 | +# LICENSE file in the root directory of this source tree. |
| 5 | + |
| 6 | +import executorch.backends.arm.tosa.dialect # noqa: F401 |
| 7 | +import pytest |
| 8 | +import torch |
| 9 | +from executorch.backends.arm.tosa.dialect.lib import TosaValueError |
| 10 | +from executorch.backends.arm.tosa.specification import ( |
| 11 | + TosaLoweringContext, |
| 12 | + TosaSpecification, |
| 13 | +) |
| 14 | +from executorch.exir.dialects._ops import ops as exir_ops |
| 15 | +from torch._subclasses.fake_tensor import FakeTensorMode |
| 16 | + |
| 17 | + |
| 18 | +def _fake_tensor(dtype: torch.dtype, mode: FakeTensorMode) -> torch.Tensor: |
| 19 | + return mode.from_tensor(torch.empty((2, 3), dtype=dtype)) |
| 20 | + |
| 21 | + |
| 22 | +_DATA_LAYOUT_OPS = [ |
| 23 | + pytest.param( |
| 24 | + lambda x: exir_ops.backend.tosa.CONCAT.default([x, x], axis=0), |
| 25 | + (4, 3), |
| 26 | + id="concat", |
| 27 | + ), |
| 28 | + pytest.param( |
| 29 | + lambda x: exir_ops.backend.tosa.PAD.default(x, [1, 2, 3, 4], value=0), |
| 30 | + (5, 10), |
| 31 | + id="pad", |
| 32 | + ), |
| 33 | + pytest.param( |
| 34 | + lambda x: exir_ops.backend.tosa.RESHAPE.default(x, [3, 2]), |
| 35 | + (3, 2), |
| 36 | + id="reshape", |
| 37 | + ), |
| 38 | + pytest.param( |
| 39 | + lambda x: exir_ops.backend.tosa.REVERSE.default(x, axis=0), |
| 40 | + (2, 3), |
| 41 | + id="reverse", |
| 42 | + ), |
| 43 | + pytest.param( |
| 44 | + lambda x: exir_ops.backend.tosa.SLICE.default(x, [0, 1], [2, 2]), |
| 45 | + (2, 2), |
| 46 | + id="slice", |
| 47 | + ), |
| 48 | + pytest.param( |
| 49 | + lambda x: exir_ops.backend.tosa.TILE.default(x, [1, 2]), |
| 50 | + (2, 6), |
| 51 | + id="tile", |
| 52 | + ), |
| 53 | + pytest.param( |
| 54 | + lambda x: exir_ops.backend.tosa.TRANSPOSE.default(x, [1, 0]), |
| 55 | + (3, 2), |
| 56 | + id="transpose", |
| 57 | + ), |
| 58 | +] |
| 59 | + |
| 60 | +_POSITIVE_DTYPES = [ |
| 61 | + pytest.param("TOSA-1.1+FP", torch.float32, id="fp32"), |
| 62 | + pytest.param("TOSA-1.1+INT", torch.int32, id="int32"), |
| 63 | + pytest.param("TOSA-1.1+FP", torch.bool, id="bool"), |
| 64 | + pytest.param("TOSA-1.1+INT+int64", torch.int64, id="int64"), |
| 65 | + pytest.param("TOSA-1.1+FP+bf16", torch.bfloat16, id="bf16"), |
| 66 | + pytest.param("TOSA-1.1+FP+fp8e4m3", torch.float8_e4m3fn, id="fp8e4m3"), |
| 67 | + pytest.param("TOSA-1.1+FP+fp8e5m2", torch.float8_e5m2, id="fp8e5m2"), |
| 68 | +] |
| 69 | + |
| 70 | + |
| 71 | +@pytest.mark.parametrize("spec,dtype", _POSITIVE_DTYPES) |
| 72 | +@pytest.mark.parametrize("op,expected_shape", _DATA_LAYOUT_OPS) |
| 73 | +def test_data_layout_ops_positive(op, expected_shape, spec, dtype) -> None: |
| 74 | + with TosaLoweringContext( |
| 75 | + TosaSpecification.create_from_string(spec) |
| 76 | + ), FakeTensorMode() as mode: |
| 77 | + output = op(_fake_tensor(dtype, mode)) |
| 78 | + |
| 79 | + assert output.dtype == dtype |
| 80 | + assert tuple(output.shape) == expected_shape |
| 81 | + |
| 82 | + |
| 83 | +@pytest.mark.parametrize( |
| 84 | + "op,error_match", |
| 85 | + [ |
| 86 | + pytest.param( |
| 87 | + lambda x: exir_ops.backend.tosa.CONCAT.default([x, x], axis=2), |
| 88 | + "out of range", |
| 89 | + id="concat", |
| 90 | + ), |
| 91 | + pytest.param( |
| 92 | + lambda x: exir_ops.backend.tosa.PAD.default(x, [0, -1, 0, 0], value=0), |
| 93 | + "non-negative", |
| 94 | + id="pad", |
| 95 | + ), |
| 96 | + pytest.param( |
| 97 | + lambda x: exir_ops.backend.tosa.RESHAPE.default(x, [-2, -3]), |
| 98 | + "Negative dimension", |
| 99 | + id="reshape", |
| 100 | + ), |
| 101 | + pytest.param( |
| 102 | + lambda x: exir_ops.backend.tosa.REVERSE.default(x, axis=2), |
| 103 | + "out of range", |
| 104 | + id="reverse", |
| 105 | + ), |
| 106 | + pytest.param( |
| 107 | + lambda x: exir_ops.backend.tosa.SLICE.default(x, [0, 0], [2, 0]), |
| 108 | + r"Expected start \+ size", |
| 109 | + id="slice", |
| 110 | + ), |
| 111 | + pytest.param( |
| 112 | + lambda x: exir_ops.backend.tosa.TILE.default(x, [0, 1]), |
| 113 | + "TILE multiples must be positive", |
| 114 | + id="tile", |
| 115 | + ), |
| 116 | + pytest.param( |
| 117 | + lambda x: exir_ops.backend.tosa.TRANSPOSE.default(x, [0, 0]), |
| 118 | + "Invalid permutation", |
| 119 | + id="transpose", |
| 120 | + ), |
| 121 | + ], |
| 122 | +) |
| 123 | +def test_data_layout_ops_reject_invalid_arguments(op, error_match) -> None: |
| 124 | + with TosaLoweringContext( |
| 125 | + TosaSpecification.create_from_string("TOSA-1.1+FP") |
| 126 | + ), FakeTensorMode() as mode: |
| 127 | + with pytest.raises(TosaValueError, match=error_match): |
| 128 | + op(_fake_tensor(torch.float32, mode)) |
| 129 | + |
| 130 | + |
| 131 | +@pytest.mark.parametrize("op,expected_shape", _DATA_LAYOUT_OPS) |
| 132 | +def test_data_layout_ops_reject_int64_without_extension(op, expected_shape) -> None: |
| 133 | + with TosaLoweringContext( |
| 134 | + TosaSpecification.create_from_string("TOSA-1.1+FP") |
| 135 | + ), FakeTensorMode() as mode: |
| 136 | + with pytest.raises(TosaValueError, match="Unsupported dtype"): |
| 137 | + op(_fake_tensor(torch.int64, mode)) |
| 138 | + |
| 139 | + |
| 140 | +def test_int16_data_layout_dtype_support_follows_tosa_spec() -> None: |
| 141 | + with TosaLoweringContext( |
| 142 | + TosaSpecification.create_from_string("TOSA-1.0+INT") |
| 143 | + ), FakeTensorMode() as mode: |
| 144 | + x = _fake_tensor(torch.int16, mode) |
| 145 | + |
| 146 | + assert exir_ops.backend.tosa.RESHAPE.default(x, [3, 2]).dtype == torch.int16 |
| 147 | + assert exir_ops.backend.tosa.REVERSE.default(x, axis=0).dtype == torch.int16 |
| 148 | + assert exir_ops.backend.tosa.TILE.default(x, [1, 1]).dtype == torch.int16 |
| 149 | + |
| 150 | + with pytest.raises(TosaValueError, match="Unsupported dtype"): |
| 151 | + exir_ops.backend.tosa.CONCAT.default([x, x], axis=0) |
| 152 | + |
| 153 | + with TosaLoweringContext( |
| 154 | + TosaSpecification.create_from_string("TOSA-1.0+INT+int16") |
| 155 | + ), FakeTensorMode() as mode: |
| 156 | + x = _fake_tensor(torch.int16, mode) |
| 157 | + assert exir_ops.backend.tosa.CONCAT.default([x, x], axis=0).dtype == torch.int16 |
| 158 | + |
| 159 | + |
| 160 | +def test_pad_rejects_wrong_padding_length() -> None: |
| 161 | + with TosaLoweringContext( |
| 162 | + TosaSpecification.create_from_string("TOSA-1.0+FP") |
| 163 | + ), FakeTensorMode() as mode: |
| 164 | + with pytest.raises(TosaValueError, match="Padding length"): |
| 165 | + exir_ops.backend.tosa.PAD.default( |
| 166 | + mode.from_tensor(torch.randn((2, 3), dtype=torch.float32)), |
| 167 | + [1, 2], |
| 168 | + value=0.0, |
| 169 | + ) |
| 170 | + |
| 171 | + |
| 172 | +def test_reshape_rejects_size_change(): |
| 173 | + with TosaLoweringContext( |
| 174 | + TosaSpecification.create_from_string("TOSA-1.1+FP") |
| 175 | + ), FakeTensorMode() as mode: |
| 176 | + with pytest.raises(TosaValueError, match="same number of elements"): |
| 177 | + exir_ops.backend.tosa.RESHAPE.default( |
| 178 | + mode.from_tensor(torch.randn((2, 3), dtype=torch.float32)), |
| 179 | + [5], |
| 180 | + ) |
0 commit comments