Gunes Bayir | ae72a46 | 2023-01-29 13:24:24 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2023 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 | |
| 25 | #ifndef TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE |
| 26 | #define TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE |
| 27 | |
| 28 | #include "arm_compute/core/TensorShape.h" |
| 29 | #include "arm_compute/core/Types.h" |
| 30 | #include "tests/AssetsLibrary.h" |
| 31 | #include "tests/Globals.h" |
| 32 | #include "tests/IAccessor.h" |
| 33 | #include "tests/framework/Asserts.h" |
| 34 | #include "tests/framework/Fixture.h" |
| 35 | #include "tests/validation/Helpers.h" |
| 36 | #include "tests/validation/reference/ActivationLayer.h" |
| 37 | #include "tests/validation/reference/ArithmeticOperations.h" |
| 38 | #include "tests/validation/reference/DequantizationLayer.h" |
| 39 | #include "tests/validation/reference/PixelWiseMultiplication.h" |
| 40 | #include "tests/validation/reference/QuantizationLayer.h" |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | namespace test |
| 45 | { |
| 46 | namespace validation |
| 47 | { |
| 48 | template <typename TensorType, typename AccessorType, typename FunctionType, typename T> |
| 49 | class AddMulAddGenericFixture : public framework::Fixture |
| 50 | { |
| 51 | public: |
| 52 | template <typename...> |
| 53 | void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info, bool interm_out) |
| 54 | { |
| 55 | compute_target(shape, data_type, act_info, interm_out); |
| 56 | } |
| 57 | |
| 58 | protected: |
| 59 | template <typename U> |
| 60 | void fill(U &&tensor, int i, DataType data_type) |
| 61 | { |
| 62 | switch(data_type) |
| 63 | { |
| 64 | case DataType::F32: |
| 65 | library->fill_tensor_uniform(tensor, i, -10.f, 10.f); |
| 66 | break; |
| 67 | case DataType::F16: |
| 68 | library->fill_tensor_uniform(tensor, i, -1.f, 1.f); |
| 69 | break; |
| 70 | default: |
| 71 | library->fill_tensor_uniform(tensor, i); |
| 72 | break; |
| 73 | } |
| 74 | } |
| 75 | |
| 76 | void compute_target(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info, bool interm_out) |
| 77 | { |
| 78 | TensorShape b_shape(shape.x()); |
| 79 | |
| 80 | // Create tensors |
| 81 | TensorType input1 = create_tensor<TensorType>(shape, data_type, 1, _input1_qinfo); |
| 82 | TensorType input2 = create_tensor<TensorType>(shape, data_type, 1, _input2_qinfo); |
| 83 | TensorType bn_mul = create_tensor<TensorType>(b_shape, data_type, 1, _bn_mul_qinfo); |
| 84 | TensorType bn_add = create_tensor<TensorType>(b_shape, data_type, 1, _bn_add_qinfo); |
| 85 | TensorType add_output = create_tensor<TensorType>(shape, data_type, 1, _add_output_qinfo); |
| 86 | TensorType final_output = create_tensor<TensorType>(shape, data_type, 1, _final_output_qinfo); |
| 87 | |
| 88 | // Create and configure function |
| 89 | FunctionType add_mul_add; |
| 90 | add_mul_add.configure(&input1, &input2, &bn_mul, &bn_add, interm_out ? &add_output : nullptr, &final_output, ConvertPolicy::SATURATE, act_info); |
| 91 | |
| 92 | // Allocate tensors |
| 93 | input1.allocator()->allocate(); |
| 94 | input2.allocator()->allocate(); |
| 95 | bn_mul.allocator()->allocate(); |
| 96 | bn_add.allocator()->allocate(); |
| 97 | |
| 98 | if(interm_out) |
| 99 | { |
| 100 | add_output.allocator()->allocate(); |
| 101 | } |
| 102 | |
| 103 | final_output.allocator()->allocate(); |
| 104 | |
| 105 | // Fill tensors |
| 106 | fill(AccessorType(input1), 0, data_type); |
| 107 | fill(AccessorType(input2), 1, data_type); |
| 108 | fill(AccessorType(bn_mul), 2, data_type); |
| 109 | fill(AccessorType(bn_add), 3, data_type); |
| 110 | |
| 111 | // // Compute function |
| 112 | add_mul_add.run(); |
| 113 | |
| 114 | _target = std::move(final_output); |
| 115 | |
| 116 | if(interm_out) |
| 117 | { |
| 118 | _interm_target = std::move(add_output); |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | TensorType _target{}; |
| 123 | TensorType _interm_target{}; |
| 124 | SimpleTensor<T> _reference{}; |
| 125 | SimpleTensor<T> _interm_reference{}; |
| 126 | |
| 127 | QuantizationInfo _input1_qinfo{}; |
| 128 | QuantizationInfo _input2_qinfo{}; |
| 129 | QuantizationInfo _bn_mul_qinfo{}; |
| 130 | QuantizationInfo _bn_add_qinfo{}; |
| 131 | QuantizationInfo _add_output_qinfo{}; |
| 132 | QuantizationInfo _final_output_qinfo{}; |
| 133 | }; |
| 134 | |
| 135 | template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool interm_out> |
| 136 | class AddMulAddFloatValidationFixture : public AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T> |
| 137 | { |
| 138 | public: |
| 139 | using Parent = AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>; |
| 140 | |
| 141 | template <typename...> |
| 142 | void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo act_info) |
| 143 | { |
| 144 | Parent::setup(shape, data_type, act_info, interm_out); |
| 145 | compute_reference(shape, data_type, act_info); |
| 146 | } |
| 147 | |
| 148 | // Compute Reference is moved outside of the generic fixture because with the quantized data types, |
| 149 | // it becomes a very different implementation with intermediate tensors' data types being always float. |
| 150 | // This way the reference calculations are more readable and the size of the classes will be smaller |
| 151 | // due to unrepeated fill() and target() methods. |
| 152 | void compute_reference(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info) |
| 153 | { |
| 154 | TensorShape b_shape(shape.x()); |
| 155 | |
| 156 | // Create reference |
| 157 | SimpleTensor<T> input1{ shape, data_type }; |
| 158 | SimpleTensor<T> input2{ shape, data_type }; |
| 159 | SimpleTensor<T> bn_mul{ b_shape, data_type }; |
| 160 | SimpleTensor<T> bn_add{ b_shape, data_type }; |
| 161 | SimpleTensor<T> add_output{ shape, data_type, 1 }; |
| 162 | |
| 163 | SimpleTensor<T> bn_mul_out{ shape, data_type }; |
| 164 | SimpleTensor<T> bn_add_out{ shape, data_type }; |
| 165 | |
| 166 | // Fill reference |
| 167 | Parent::fill(input1, 0, data_type); |
| 168 | Parent::fill(input2, 1, data_type); |
| 169 | Parent::fill(bn_mul, 2, data_type); |
| 170 | Parent::fill(bn_add, 3, data_type); |
| 171 | |
| 172 | reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, input1, input2, add_output, ConvertPolicy::SATURATE); |
| 173 | bn_mul_out = reference::pixel_wise_multiplication<T, T, T>(add_output, bn_mul, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_UP, data_type); |
| 174 | reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, bn_mul_out, bn_add, bn_add_out, ConvertPolicy::SATURATE); |
| 175 | |
| 176 | if(interm_out) |
| 177 | { |
| 178 | Parent::_interm_reference = std::move(add_output); |
| 179 | } |
| 180 | |
| 181 | if(act_info.enabled() && act_info.activation() != ActivationLayerInfo::ActivationFunction::IDENTITY) |
| 182 | { |
| 183 | Parent::_reference = reference::activation_layer(bn_add_out, act_info); |
| 184 | } |
| 185 | else |
| 186 | { |
| 187 | Parent::_reference = std::move(bn_add_out); |
| 188 | } |
| 189 | } |
| 190 | }; |
| 191 | |
| 192 | template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool interm_out> |
| 193 | class AddMulAddQuantizedValidationFixture : public AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T> |
| 194 | { |
| 195 | public: |
| 196 | using Parent = AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>; |
| 197 | |
| 198 | template <typename...> |
| 199 | void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo act_info, |
| 200 | QuantizationInfo input1_qinfo, QuantizationInfo input2_qinfo, QuantizationInfo bn_mul_qinfo, |
| 201 | QuantizationInfo bn_add_qinfo, QuantizationInfo add_output_qinfo, QuantizationInfo final_output_qinfo) |
| 202 | { |
| 203 | // Quantization arguments moved to class attributes to prevent long function declerations |
| 204 | Parent::_input1_qinfo = input1_qinfo; |
| 205 | Parent::_input2_qinfo = input2_qinfo; |
| 206 | Parent::_bn_mul_qinfo = bn_mul_qinfo; |
| 207 | Parent::_bn_add_qinfo = bn_add_qinfo; |
| 208 | Parent::_add_output_qinfo = add_output_qinfo; |
| 209 | Parent::_final_output_qinfo = final_output_qinfo; |
| 210 | |
| 211 | Parent::setup(shape, data_type, act_info, interm_out); |
| 212 | compute_reference(shape, data_type, act_info); |
| 213 | } |
| 214 | |
| 215 | // Compute Reference is moved outside of the generic fixture because with the quantized data types, |
| 216 | // it becomes a very different implementation with intermediate tensors' data types being always float. |
| 217 | // This way the reference calculations are more readable and the size of the classes will be smaller |
| 218 | // due to unrepeated fill() and target() methods. |
| 219 | void compute_reference(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info) |
| 220 | { |
| 221 | TensorShape b_shape(shape.x()); |
| 222 | |
| 223 | // Create reference |
| 224 | SimpleTensor<T> input1{ shape, data_type, 1, Parent::_input1_qinfo }; |
| 225 | SimpleTensor<T> input2{ shape, data_type, 1, Parent::_input2_qinfo }; |
| 226 | SimpleTensor<T> bn_mul{ b_shape, data_type, 1, Parent::_bn_mul_qinfo }; |
| 227 | SimpleTensor<T> bn_add{ b_shape, data_type, 1, Parent::_bn_add_qinfo }; |
| 228 | |
| 229 | // Fill input tensors |
| 230 | Parent::fill(input1, 0, data_type); |
| 231 | Parent::fill(input2, 1, data_type); |
| 232 | Parent::fill(bn_mul, 2, data_type); |
| 233 | Parent::fill(bn_add, 3, data_type); |
| 234 | |
| 235 | SimpleTensor<float> input1_dequantized = reference::dequantization_layer<float>(input1); |
| 236 | SimpleTensor<float> input2_dequantized = reference::dequantization_layer<float>(input2); |
| 237 | SimpleTensor<float> bn_mul_dequantized = reference::dequantization_layer<float>(bn_mul); |
| 238 | SimpleTensor<float> bn_add_dequantized = reference::dequantization_layer<float>(bn_add); |
| 239 | |
| 240 | SimpleTensor<float> add_output_dequantized{ shape, DataType::F32 }; |
| 241 | SimpleTensor<float> bn_add_out_dequantized{ shape, DataType::F32 }; |
| 242 | |
| 243 | reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, input1_dequantized, input2_dequantized, add_output_dequantized, ConvertPolicy::SATURATE); |
| 244 | SimpleTensor<float> bn_mul_out_dequantized = reference::pixel_wise_multiplication<float, float, float>(add_output_dequantized, bn_mul_dequantized, 1.f, ConvertPolicy::SATURATE, |
| 245 | RoundingPolicy::TO_NEAREST_UP, DataType::F32); |
| 246 | reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, bn_mul_out_dequantized, bn_add_dequantized, bn_add_out_dequantized, ConvertPolicy::SATURATE); |
| 247 | |
| 248 | if(interm_out) |
| 249 | { |
| 250 | Parent::_interm_reference = reference::quantization_layer<float, T>(add_output_dequantized, data_type, Parent::_add_output_qinfo); |
| 251 | } |
| 252 | |
| 253 | if(act_info.enabled() && act_info.activation() != ActivationLayerInfo::ActivationFunction::IDENTITY) |
| 254 | { |
| 255 | SimpleTensor<T> ref = reference::quantization_layer<float, T>(bn_add_out_dequantized, data_type, Parent::_final_output_qinfo); |
| 256 | Parent::_reference = reference::activation_layer(ref, act_info); |
| 257 | } |
| 258 | else |
| 259 | { |
| 260 | Parent::_reference = reference::quantization_layer<float, T>(bn_add_out_dequantized, data_type, Parent::_final_output_qinfo); |
| 261 | } |
| 262 | } |
| 263 | }; |
| 264 | } // namespace validation |
| 265 | } // namespace test |
| 266 | } // namespace arm_compute |
| 267 | |
| 268 | #endif /* TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE */ |