Sheri Zhang | 6d9c982 | 2021-09-24 16:02:57 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/Helpers.h" |
| 25 | #include "arm_compute/core/Types.h" |
| 26 | #include "arm_compute/runtime/NEON/functions/NEConv3D.h" |
| 27 | #include "arm_compute/runtime/Tensor.h" |
| 28 | #include "arm_compute/runtime/TensorAllocator.h" |
| 29 | #include "tests/NEON/Accessor.h" |
| 30 | #include "tests/PaddingCalculator.h" |
| 31 | #include "tests/datasets/ShapeDatasets.h" |
| 32 | #include "tests/framework/Asserts.h" |
| 33 | #include "tests/framework/Macros.h" |
| 34 | #include "tests/framework/datasets/Datasets.h" |
| 35 | #include "tests/validation/Validation.h" |
| 36 | #include "tests/validation/fixtures/DirectConvolution3DFixture.h" |
| 37 | |
| 38 | namespace arm_compute |
| 39 | { |
| 40 | namespace test |
| 41 | { |
| 42 | namespace validation |
| 43 | { |
| 44 | namespace |
| 45 | { |
| 46 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 47 | const RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2f)); /**< Relative tolerance value for FP16 types */ |
| 48 | const AbsoluteTolerance<float> abs_tolerance_f16(0.2f); /**< Absolute tolerance for FP16 types */ |
| 49 | constexpr float tolerance_num = 0.07f; /**< Tolerance number for the FP16 implementation */ |
| 50 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 51 | constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */ |
| 52 | |
| 53 | /* The following tests are from real use-case that made DirectConvolution |
| 54 | * overflows in terms of its tensor indexing. This test case is using |
| 55 | * a separate tolerance due to the following reason. |
| 56 | * - It has shown that it requires generally larger absolute tolerance |
| 57 | * for large numbers or larger relative tolerance for small numbers. |
| 58 | * - With the first reason, since it is mainly testing index overflow, |
| 59 | * a value with a margin is used to avoid uninteded test failures |
| 60 | * during nightly. |
| 61 | */ |
| 62 | constexpr AbsoluteTolerance<float> usecase_tolerance_fp32(0.05f); |
| 63 | |
| 64 | /** Activation function Dataset*/ |
| 65 | const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", |
| 66 | { |
| 67 | ActivationLayerInfo(), |
| 68 | ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) |
| 69 | }); |
| 70 | |
| 71 | const auto data_precommit = combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( |
| 72 | datasets::SmallDirectConv3DShapes(), |
| 73 | framework::dataset::make("StrideX", { 1, 5, 8 })), |
| 74 | framework::dataset::make("StrideY", { 1, 2, 3 })), |
| 75 | framework::dataset::make("StrideZ", { 1, 2, 1 })), |
| 76 | framework::dataset::make("PadX", { 0, 1, 2 })), |
| 77 | framework::dataset::make("PadY", { 0, 2, 1 })), |
| 78 | framework::dataset::make("PadZ", { 0, 3, 5 })), |
| 79 | framework::dataset::make("KernelWidth", { 3, 5, 9 })), |
| 80 | framework::dataset::make("KernelHeight", { 2, 1, 3 })), |
| 81 | framework::dataset::make("KernelDepth", { 1, 2, 3 })), |
| 82 | framework::dataset::make("NumKernels", { 2, 3, 8 })), |
| 83 | framework::dataset::make("HasBias", { true, false })), |
| 84 | ActivationFunctionsDataset); |
| 85 | } // namespace |
| 86 | |
| 87 | TEST_SUITE(NEON) |
| 88 | TEST_SUITE(Convolution3D) |
| 89 | |
| 90 | // *INDENT-OFF* |
| 91 | // clang-format off |
| 92 | DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( |
| 93 | framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Mismatching data type input/weights |
| 94 | TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Mismatching input feature maps |
| 95 | TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid weights dimensions |
| 96 | TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NHWC), // Invalid data layout |
| 97 | TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid biases size |
| 98 | TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid biases dimensions |
| 99 | TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid output size |
| 100 | }), |
| 101 | framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F16), |
| 102 | TensorInfo(TensorShape(4U, 3U, 3U, 3U, 3U), 1U, DataType::F32), |
| 103 | TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U, 3U), 1U, DataType::F32), |
| 104 | TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), |
| 105 | TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), |
| 106 | TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), |
| 107 | TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), |
| 108 | })), |
| 109 | framework::dataset::make("BiasesInfo",{ TensorInfo(TensorShape(4U), 1U, DataType::F32), |
| 110 | TensorInfo(TensorShape(4U), 1U, DataType::F32), |
| 111 | TensorInfo(TensorShape(4U), 1U, DataType::F32), |
| 112 | TensorInfo(TensorShape(4U), 1U, DataType::F32), |
| 113 | TensorInfo(TensorShape(3U), 1U, DataType::F32), |
| 114 | TensorInfo(TensorShape(4U, 2U), 1U, DataType::F32), |
| 115 | TensorInfo(TensorShape(4U), 1U, DataType::F32), |
| 116 | })), |
| 117 | framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), |
| 118 | TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), |
| 119 | TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), |
| 120 | TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), |
| 121 | TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), |
| 122 | TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), |
| 123 | TensorInfo(TensorShape(26U, 11U, 4U), 1U, DataType::F32), |
| 124 | })), |
| 125 | framework::dataset::make("Expected", { false, false, false, false, false, false, false })), |
| 126 | input_info, weights_info, biases_info, output_info, expected) |
| 127 | { |
| 128 | const Conv3dInfo conv3d_info(Size3D(1, 1, 1), Padding3D(0, 0, 0), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false); |
| 129 | bool is_valid = bool(NEConv3D::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv3d_info)); |
| 130 | ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); |
| 131 | } |
| 132 | // clang-format on |
| 133 | // *INDENT-ON* |
| 134 | |
| 135 | template <typename T> |
| 136 | using NEDirectConvolution3DFixture = DirectConvolution3DValidationFixture<Tensor, Accessor, NEConv3D, T>; |
| 137 | |
| 138 | TEST_SUITE(Float) |
| 139 | TEST_SUITE(FP32) |
| 140 | FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolution3DFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(data_precommit, |
| 141 | framework::dataset::make("DataType", DataType::F32)), |
| 142 | framework::dataset::make("DataLayout", { DataLayout::NDHWC }))) |
| 143 | { |
| 144 | // Validate output |
| 145 | validate(Accessor(_target), _reference, tolerance_fp32); |
| 146 | } |
| 147 | TEST_SUITE_END() // FP32 |
| 148 | |
| 149 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 150 | TEST_SUITE(FP16) |
| 151 | FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolution3DFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(data_precommit, |
| 152 | framework::dataset::make("DataType", DataType::F16)), |
| 153 | framework::dataset::make("DataLayout", { DataLayout::NDHWC }))) |
| 154 | { |
| 155 | // Validate output |
| 156 | validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_f16); |
| 157 | } |
| 158 | TEST_SUITE_END() // FP16 |
| 159 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 160 | |
| 161 | TEST_SUITE_END() // Float |
| 162 | TEST_SUITE_END() // Convolution3D |
| 163 | TEST_SUITE_END() // Neon |
| 164 | } // namespace validation |
| 165 | } // namespace test |
| 166 | } // namespace arm_compute |