Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 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 | #ifndef ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE |
| 25 | #define ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE |
| 26 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 27 | #include "arm_compute/core/KernelDescriptors.h" |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/TensorShape.h" |
| 29 | #include "arm_compute/core/Types.h" |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 30 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 31 | #include "tests/AssetsLibrary.h" |
| 32 | #include "tests/Globals.h" |
| 33 | #include "tests/IAccessor.h" |
| 34 | #include "tests/framework/Asserts.h" |
| 35 | #include "tests/framework/Fixture.h" |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 36 | #include "tests/validation/Helpers.h" |
Georgios Pinitas | 5a7e776 | 2017-12-01 16:27:29 +0000 | [diff] [blame] | 37 | #include "tests/validation/reference/GEMMLowp.h" |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 38 | |
| 39 | #include <random> |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
| 43 | namespace test |
| 44 | { |
| 45 | namespace validation |
| 46 | { |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 47 | namespace |
| 48 | { |
| 49 | template <typename U> |
| 50 | void fill(U &&tensor, int i) |
| 51 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 52 | switch(tensor.data_type()) |
| 53 | { |
| 54 | case DataType::QSYMM8_PER_CHANNEL: |
| 55 | { |
| 56 | int min_bound = 128; |
| 57 | int max_bound = -127; |
| 58 | for(size_t j = 0; j < tensor.quantization_info().scale().size(); j++) |
| 59 | { |
| 60 | std::pair<int, int> bounds = get_symm_quantized_per_channel_bounds(tensor.quantization_info(), -1.0f, 1.0f, i); |
| 61 | if(bounds.first < min_bound) |
| 62 | { |
| 63 | min_bound = bounds.first; |
| 64 | } |
| 65 | if(bounds.second > max_bound) |
| 66 | { |
| 67 | max_bound = bounds.second; |
| 68 | } |
| 69 | } |
| 70 | std::uniform_int_distribution<int8_t> distribution(min_bound, max_bound); |
| 71 | library->fill(tensor, distribution, i); |
| 72 | break; |
| 73 | } |
| 74 | case DataType::QASYMM8: |
| 75 | { |
| 76 | std::uniform_int_distribution<uint8_t> distribution(1, 254); |
| 77 | library->fill(tensor, distribution, i); |
| 78 | break; |
| 79 | } |
| 80 | case DataType::F16: |
Giorgio Arena | 6aeb217 | 2020-12-15 15:45:43 +0000 | [diff] [blame^] | 81 | { |
| 82 | arm_compute::utils::uniform_real_distribution_fp16 distribution{ half(-1.0f), half(1.0f) }; |
| 83 | library->fill(tensor, distribution, i); |
| 84 | break; |
| 85 | } |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 86 | case DataType::F32: |
| 87 | { |
Giorgio Arena | 6aeb217 | 2020-12-15 15:45:43 +0000 | [diff] [blame^] | 88 | std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 89 | library->fill(tensor, distribution, i); |
| 90 | break; |
| 91 | } |
| 92 | default: |
| 93 | library->fill_tensor_uniform(tensor, i); |
| 94 | } |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 95 | } |
| 96 | |
| 97 | template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d, bool reinterpret_output_as_3d, typename OutputType, bool is_fused = false> |
| 98 | TensorType compute_gemmlowp_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 99 | GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo(), DataType data_type_a = DataType::QASYMM8, DataType data_type_b = DataType::QASYMM8, |
| 100 | QuantizationInfo b_qinfo = QuantizationInfo()) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 101 | { |
| 102 | // Create tensors |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 103 | DataType data_type_output = output_stage.type == GEMMLowpOutputStageType::NONE ? DataType::S32 : data_type_a; |
| 104 | |
| 105 | TensorType a = create_tensor<TensorType>(shape_a, data_type_a, 1); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 106 | TensorType b = create_tensor<TensorType>(shape_b, data_type_b, 1); // gemm output before output stage mismatch if i pass data_layout_output here. to be investigated |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 107 | TensorType output = create_tensor<TensorType>(shape_output, data_type_output, 1); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 108 | |
| 109 | a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset)); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 110 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 111 | if(data_type_b == DataType::QSYMM8_PER_CHANNEL) |
| 112 | { |
| 113 | b.info()->set_quantization_info(b_qinfo); |
| 114 | } |
| 115 | else |
| 116 | { |
| 117 | b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset)); |
| 118 | } |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 119 | TensorType bias; |
| 120 | if(is_fused) |
| 121 | { |
| 122 | TensorShape bias_shape(shape_b[0]); |
| 123 | bias = create_tensor<TensorType>(bias_shape, DataType::S32, 1); |
| 124 | } |
| 125 | |
| 126 | // Create and configure function |
| 127 | // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output |
| 128 | FunctionType gemmlowp; |
| 129 | // TODO (COMPMID-1672) - Extending the test to validate add bias in offset contribution |
| 130 | gemmlowp.configure(&a, &b, is_fused ? &bias : nullptr, &output, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_output[2] : 0), reinterpret_input_as_3d, false, output_stage)); |
| 131 | |
| 132 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 133 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 134 | ARM_COMPUTE_EXPECT(output.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 135 | |
| 136 | // Allocate tensors |
| 137 | a.allocator()->allocate(); |
| 138 | b.allocator()->allocate(); |
| 139 | output.allocator()->allocate(); |
| 140 | |
| 141 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 142 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 143 | ARM_COMPUTE_EXPECT(!output.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 144 | |
| 145 | // Fill tensors |
| 146 | fill(AccessorType(a), 0); |
| 147 | fill(AccessorType(b), 1); |
| 148 | |
| 149 | if(is_fused) |
| 150 | { |
| 151 | ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 152 | bias.allocator()->allocate(); |
| 153 | ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 154 | fill(AccessorType(bias), 2); |
| 155 | } |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 156 | // Compute GEMM function |
| 157 | gemmlowp.run(); |
| 158 | return output; |
| 159 | } |
| 160 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 161 | template <bool reinterpret_input_as_3d, typename TI = uint8_t, typename TW = uint8_t> |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 162 | SimpleTensor<int32_t> compute_gemmlowp_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 163 | DataType data_type_a = DataType::QASYMM8, DataType data_type_b = DataType::QASYMM8, QuantizationInfo b_qinfo = QuantizationInfo()) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 164 | { |
| 165 | TensorShape shape_a_to_use = shape_a; |
| 166 | if(reinterpret_input_as_3d) |
| 167 | { |
| 168 | // Collapse the second and third dimension if the input is 3D |
| 169 | shape_a_to_use.collapse(2U, 1U); |
| 170 | } |
| 171 | |
| 172 | // Create reference |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 173 | SimpleTensor<TI> a{ shape_a_to_use, data_type_a, 1 }; |
| 174 | SimpleTensor<TW> b{ shape_b, data_type_b, 1, data_type_b == DataType::QSYMM8_PER_CHANNEL ? b_qinfo : QuantizationInfo(1.0f / 255, b_offset) }; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 175 | |
| 176 | // Fill reference |
| 177 | fill(a, 0); |
| 178 | fill(b, 1); |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 179 | return reference::gemmlowp_matrix_multiply_core<int32_t, TI, TW>(a, b, shape_output, a_offset, b_offset); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 180 | } |
| 181 | } |
| 182 | |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 183 | template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false> |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 184 | class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 185 | { |
| 186 | public: |
| 187 | template <typename...> |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 188 | void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 189 | { |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 190 | _target = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset); |
| 191 | _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset); |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 192 | } |
| 193 | |
| 194 | protected: |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 195 | TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 196 | { |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 197 | return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, int32_t>(shape_a, shape_b, shape_output, a_offset, b_offset); |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 198 | } |
| 199 | |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 200 | SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 201 | { |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 202 | return compute_gemmlowp_reference<reinterpret_input_as_3d>(shape_a, shape_b, shape_output, a_offset, b_offset); |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 203 | } |
| 204 | |
Pablo Tello | 6ff12a0 | 2017-11-02 16:09:35 +0000 | [diff] [blame] | 205 | TensorType _target{}; |
| 206 | SimpleTensor<int32_t> _reference{}; |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 207 | }; |
| 208 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 209 | template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, typename TI = uint8_t, typename TW = uint8_t> |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 210 | class GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture : public framework::Fixture |
| 211 | { |
| 212 | public: |
| 213 | template <typename...> |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 214 | void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage, DataType data_type_b) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 215 | { |
| 216 | ARM_COMPUTE_EXPECT(output_stage.type != GEMMLowpOutputStageType::NONE, framework::LogLevel::ERRORS); |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 217 | DataType data_type_a = data_type_b == DataType::QASYMM8_SIGNED ? DataType::QASYMM8_SIGNED : DataType::QASYMM8; |
| 218 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 219 | if(data_type_b == DataType::QSYMM8_PER_CHANNEL) |
| 220 | { |
| 221 | output_stage.is_quantized_per_channel = true; |
| 222 | const size_t num_channels = shape_b[0]; |
| 223 | std::vector<float> scales(num_channels); |
| 224 | std::uniform_real_distribution<> distribution(0, 1); |
| 225 | library->fill(scales, distribution, 0); |
| 226 | output_stage.gemmlowp_multipliers.resize(num_channels); |
| 227 | output_stage.gemmlowp_shifts.resize(num_channels); |
| 228 | for(size_t i = 0; i < num_channels; ++i) |
| 229 | { |
Michele Di Giorgio | f29d1b7 | 2019-10-29 10:58:13 +0000 | [diff] [blame] | 230 | quantization::calculate_quantized_multiplier(scales[i], &output_stage.gemmlowp_multipliers[i], &output_stage.gemmlowp_shifts[i]); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 231 | } |
| 232 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 233 | _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, 0, output_stage, data_type_a, data_type_b, QuantizationInfo(scales)); |
| 234 | _target = compute_target(shape_a, shape_b, shape_output, a_offset, 0, output_stage, data_type_a, data_type_b, QuantizationInfo(scales)); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 235 | } |
| 236 | else |
| 237 | { |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 238 | _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage, data_type_a, data_type_b, QuantizationInfo()); |
| 239 | _target = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage, data_type_a, data_type_b, QuantizationInfo()); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 240 | } |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 241 | } |
| 242 | |
| 243 | protected: |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 244 | TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage, |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 245 | DataType data_type_a, DataType data_type_b, QuantizationInfo b_qinfo) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 246 | { |
| 247 | return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, qasymm8_t, true>(shape_a, shape_b, shape_output, a_offset, b_offset, |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 248 | output_stage, data_type_a, data_type_b, b_qinfo); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 249 | } |
| 250 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 251 | SimpleTensor<TI> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, |
| 252 | GEMMLowpOutputStageInfo output_stage, DataType data_type_a, DataType data_type_b, QuantizationInfo b_qinfo) |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 253 | { |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 254 | SimpleTensor<int32_t> output = compute_gemmlowp_reference<reinterpret_input_as_3d, TI, TW>(shape_a, shape_b, shape_output, a_offset, b_offset, data_type_a, data_type_b, b_qinfo); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 255 | |
| 256 | TensorShape bias_shape(shape_b[0]); |
| 257 | SimpleTensor<int32_t> bias{ bias_shape, DataType::S32, 1 }; |
| 258 | fill(bias, 2); |
| 259 | |
| 260 | switch(output_stage.type) |
| 261 | { |
| 262 | case GEMMLowpOutputStageType::QUANTIZE_DOWN: |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 263 | return reference::gemmlowp_quantize_down_scale<int32_t, TW>(output, bias, |
| 264 | output_stage.gemmlowp_offset, output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 265 | break; |
| 266 | case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT: |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 267 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, TW>(output, bias, |
| 268 | output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 269 | break; |
| 270 | default: |
| 271 | ARM_COMPUTE_ERROR("Not Supported!"); |
| 272 | } |
| 273 | } |
| 274 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 275 | TensorType _target{}; |
| 276 | SimpleTensor<TI> _reference{}; |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 277 | }; |
| 278 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 279 | template <typename TensorType, typename AccessorType, typename FunctionType> |
| 280 | class GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture : public framework::Fixture |
| 281 | { |
| 282 | public: |
| 283 | template <typename...> |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 284 | void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 285 | { |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 286 | _target = compute_target(shape, result_offset, result_mult_int, result_shift, min, max, add_bias); |
| 287 | _reference = compute_reference(shape, result_offset, result_mult_int, result_shift, min, max, add_bias); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 288 | } |
| 289 | |
| 290 | protected: |
| 291 | template <typename U> |
| 292 | void fill(U &&tensor, int i) |
| 293 | { |
| 294 | std::uniform_int_distribution<> distribution(-6000, 6000); |
| 295 | library->fill(tensor, distribution, i); |
| 296 | } |
| 297 | |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 298 | TensorType compute_target(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 299 | { |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 300 | TensorShape shape_bias(shape[0]); |
| 301 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 302 | // Create tensors |
| 303 | TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 304 | TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1); |
| 305 | TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8, 1); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 306 | |
| 307 | // Create and configure function |
Luca Foschiani | 4b86953 | 2020-02-13 15:07:36 +0000 | [diff] [blame] | 308 | FunctionType output_stage; |
| 309 | GEMMLowpOutputStageInfo output_stage_info = GEMMLowpOutputStageInfo(); |
| 310 | output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN; |
| 311 | output_stage_info.gemmlowp_offset = result_offset; |
| 312 | output_stage_info.gemmlowp_multiplier = result_mult_int; |
| 313 | output_stage_info.gemmlowp_shift = result_shift; |
| 314 | output_stage_info.gemmlowp_min_bound = min; |
| 315 | output_stage_info.gemmlowp_max_bound = max; |
| 316 | output_stage_info.output_data_type = DataType::QASYMM8; |
| 317 | output_stage.configure(&a, add_bias ? &b : nullptr, &c, output_stage_info); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 318 | |
| 319 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 320 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 321 | |
| 322 | // Allocate tensors |
| 323 | a.allocator()->allocate(); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 324 | c.allocator()->allocate(); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 325 | |
| 326 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 327 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 328 | |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 329 | // Fill tensor |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 330 | fill(AccessorType(a), 0); |
| 331 | |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 332 | if(add_bias) |
| 333 | { |
| 334 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 335 | |
| 336 | // Allocate bias tensor |
| 337 | b.allocator()->allocate(); |
| 338 | |
| 339 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 340 | |
| 341 | // Fill tensor |
| 342 | fill(AccessorType(b), 1); |
| 343 | } |
| 344 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 345 | // Compute GEMM function |
| 346 | output_stage.run(); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 347 | return c; |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 348 | } |
| 349 | |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 350 | SimpleTensor<uint8_t> compute_reference(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 351 | { |
| 352 | // Create reference |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 353 | TensorShape shape_bias(shape[0]); |
| 354 | |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 355 | SimpleTensor<int32_t> a{ shape, DataType::S32, 1 }; |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 356 | SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 }; |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 357 | |
| 358 | // Fill reference |
| 359 | fill(a, 0); |
| 360 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 361 | const std::vector<int32_t> result_mult_int_vec = { result_mult_int }; |
| 362 | const std::vector<int32_t> result_shift_vec = { result_shift }; |
| 363 | |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 364 | if(add_bias) |
| 365 | { |
| 366 | // Fill bias |
| 367 | fill(b, 1); |
| 368 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 369 | return reference::gemmlowp_quantize_down_scale<int32_t, uint8_t>(a, b, result_offset, result_mult_int_vec, result_shift_vec, min, max); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 370 | } |
| 371 | else |
| 372 | { |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 373 | return reference::gemmlowp_quantize_down_scale<int32_t, uint8_t>(a, result_offset, result_mult_int_vec, result_shift_vec, min, max); |
Gian Marco | 6b77e91 | 2017-11-17 09:27:57 +0000 | [diff] [blame] | 374 | } |
Gian Marco | e75a02b | 2017-11-08 12:24:09 +0000 | [diff] [blame] | 375 | } |
| 376 | |
| 377 | TensorType _target{}; |
| 378 | SimpleTensor<uint8_t> _reference{}; |
| 379 | }; |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 380 | |
| 381 | template <typename TensorType, typename AccessorType, typename FunctionType> |
Luca Foschiani | 4b86953 | 2020-02-13 15:07:36 +0000 | [diff] [blame] | 382 | class GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture : public framework::Fixture |
| 383 | { |
| 384 | public: |
| 385 | template <typename...> |
| 386 | void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
| 387 | { |
| 388 | _target = compute_target(shape, result_offset, result_mult_int, result_shift, min, max, add_bias); |
| 389 | _reference = compute_reference(shape, result_offset, result_mult_int, result_shift, min, max, add_bias); |
| 390 | } |
| 391 | |
| 392 | protected: |
| 393 | template <typename U> |
| 394 | void fill(U &&tensor, int i) |
| 395 | { |
| 396 | std::uniform_int_distribution<> distribution(-6000, 6000); |
| 397 | library->fill(tensor, distribution, i); |
| 398 | } |
| 399 | |
| 400 | TensorType compute_target(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
| 401 | { |
| 402 | TensorShape shape_bias(shape[0]); |
| 403 | |
| 404 | // Create tensors |
| 405 | TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1); |
| 406 | TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1); |
| 407 | TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8_SIGNED, 1); |
| 408 | |
| 409 | // Create and configure function |
| 410 | FunctionType output_stage; |
| 411 | GEMMLowpOutputStageInfo output_stage_info = GEMMLowpOutputStageInfo(); |
| 412 | output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN; |
| 413 | output_stage_info.gemmlowp_offset = result_offset; |
| 414 | output_stage_info.gemmlowp_multiplier = result_mult_int; |
| 415 | output_stage_info.gemmlowp_shift = result_shift; |
| 416 | output_stage_info.gemmlowp_min_bound = min; |
| 417 | output_stage_info.gemmlowp_max_bound = max; |
| 418 | output_stage_info.output_data_type = DataType::QASYMM8_SIGNED; |
| 419 | output_stage.configure(&a, add_bias ? &b : nullptr, &c, output_stage_info); |
| 420 | |
| 421 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 422 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 423 | |
| 424 | // Allocate tensors |
| 425 | a.allocator()->allocate(); |
| 426 | c.allocator()->allocate(); |
| 427 | |
| 428 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 429 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 430 | |
| 431 | // Fill tensor |
| 432 | fill(AccessorType(a), 0); |
| 433 | |
| 434 | if(add_bias) |
| 435 | { |
| 436 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 437 | |
| 438 | // Allocate bias tensor |
| 439 | b.allocator()->allocate(); |
| 440 | |
| 441 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 442 | |
| 443 | // Fill tensor |
| 444 | fill(AccessorType(b), 1); |
| 445 | } |
| 446 | |
| 447 | // Compute GEMM function |
| 448 | output_stage.run(); |
| 449 | return c; |
| 450 | } |
| 451 | |
| 452 | SimpleTensor<int8_t> compute_reference(const TensorShape &shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
| 453 | { |
| 454 | // Create reference |
| 455 | TensorShape shape_bias(shape[0]); |
| 456 | |
| 457 | SimpleTensor<int32_t> a{ shape, DataType::S32, 1 }; |
| 458 | SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 }; |
| 459 | |
| 460 | // Fill reference |
| 461 | fill(a, 0); |
| 462 | |
| 463 | const std::vector<int32_t> result_mult_int_vec = { result_mult_int }; |
| 464 | const std::vector<int32_t> result_shift_vec = { result_shift }; |
| 465 | |
| 466 | if(add_bias) |
| 467 | { |
| 468 | // Fill bias |
| 469 | fill(b, 1); |
| 470 | |
| 471 | return reference::gemmlowp_quantize_down_scale<int32_t, int8_t>(a, b, result_offset, result_mult_int_vec, result_shift_vec, min, max); |
| 472 | } |
| 473 | else |
| 474 | { |
| 475 | return reference::gemmlowp_quantize_down_scale<int32_t, int8_t>(a, result_offset, result_mult_int_vec, result_shift_vec, min, max); |
| 476 | } |
| 477 | } |
| 478 | |
| 479 | TensorType _target{}; |
| 480 | SimpleTensor<int8_t> _reference{}; |
| 481 | }; |
| 482 | |
| 483 | template <typename TensorType, typename AccessorType, typename FunctionType> |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame] | 484 | class GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture : public framework::Fixture |
| 485 | { |
| 486 | public: |
| 487 | template <typename...> |
| 488 | void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias) |
| 489 | { |
| 490 | _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias); |
| 491 | _reference = compute_reference(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias); |
| 492 | } |
| 493 | |
| 494 | protected: |
| 495 | template <typename U> |
| 496 | void fill(U &&tensor, int i) |
| 497 | { |
| 498 | std::uniform_int_distribution<> distribution(-6000, 6000); |
| 499 | library->fill(tensor, distribution, i); |
| 500 | } |
| 501 | |
| 502 | TensorType compute_target(const TensorShape &shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias) |
| 503 | { |
| 504 | TensorShape shape_bias(shape[0]); |
| 505 | |
| 506 | // Create tensors |
| 507 | TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1); |
| 508 | TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1); |
| 509 | TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8_SIGNED, 1); |
| 510 | |
| 511 | // Create and configure function |
| 512 | FunctionType output_stage; |
| 513 | output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max); |
| 514 | |
| 515 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 516 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 517 | |
| 518 | // Allocate tensors |
| 519 | a.allocator()->allocate(); |
| 520 | c.allocator()->allocate(); |
| 521 | |
| 522 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 523 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 524 | |
| 525 | // Fill tensor |
| 526 | fill(AccessorType(a), 0); |
| 527 | |
| 528 | if(add_bias) |
| 529 | { |
| 530 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 531 | |
| 532 | // Allocate bias tensor |
| 533 | b.allocator()->allocate(); |
| 534 | |
| 535 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 536 | |
| 537 | // Fill tensor |
| 538 | fill(AccessorType(b), 1); |
| 539 | } |
| 540 | |
| 541 | // Compute GEMM function |
| 542 | output_stage.run(); |
| 543 | return c; |
| 544 | } |
| 545 | |
| 546 | SimpleTensor<int8_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, |
| 547 | bool add_bias) |
| 548 | { |
| 549 | // Create reference |
| 550 | TensorShape shape_bias(shape[0]); |
| 551 | |
| 552 | SimpleTensor<int32_t> a{ shape, DataType::S32, 1 }; |
| 553 | SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 }; |
| 554 | |
| 555 | // Fill reference |
| 556 | fill(a, 0); |
| 557 | |
| 558 | const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier }; |
| 559 | const std::vector<int32_t> result_shift_vec = { result_shift }; |
| 560 | |
| 561 | if(add_bias) |
| 562 | { |
| 563 | // Fill bias |
| 564 | fill(b, 1); |
| 565 | |
| 566 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int8_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max); |
| 567 | } |
| 568 | else |
| 569 | { |
| 570 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int8_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max); |
| 571 | } |
| 572 | } |
| 573 | |
| 574 | TensorType _target{}; |
| 575 | SimpleTensor<int8_t> _reference{}; |
| 576 | }; |
| 577 | |
| 578 | template <typename TensorType, typename AccessorType, typename FunctionType> |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 579 | class GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture : public framework::Fixture |
| 580 | { |
| 581 | public: |
| 582 | template <typename...> |
| 583 | void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias) |
| 584 | { |
| 585 | _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias); |
| 586 | _reference = compute_reference(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias); |
| 587 | } |
| 588 | |
| 589 | protected: |
| 590 | template <typename U> |
| 591 | void fill(U &&tensor, int i) |
| 592 | { |
| 593 | std::uniform_int_distribution<> distribution(-6000, 6000); |
| 594 | library->fill(tensor, distribution, i); |
| 595 | } |
| 596 | |
| 597 | TensorType compute_target(const TensorShape &shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias) |
| 598 | { |
| 599 | TensorShape shape_bias(shape[0]); |
| 600 | |
| 601 | // Create tensors |
| 602 | TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1); |
| 603 | TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1); |
| 604 | TensorType c = create_tensor<TensorType>(shape, DataType::QASYMM8, 1); |
| 605 | |
| 606 | // Create and configure function |
| 607 | FunctionType output_stage; |
| 608 | output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max); |
| 609 | |
| 610 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 611 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 612 | |
| 613 | // Allocate tensors |
| 614 | a.allocator()->allocate(); |
| 615 | c.allocator()->allocate(); |
| 616 | |
| 617 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 618 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 619 | |
| 620 | // Fill tensor |
| 621 | fill(AccessorType(a), 0); |
| 622 | |
| 623 | if(add_bias) |
| 624 | { |
| 625 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 626 | |
| 627 | // Allocate bias tensor |
| 628 | b.allocator()->allocate(); |
| 629 | |
| 630 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 631 | |
| 632 | // Fill tensor |
| 633 | fill(AccessorType(b), 1); |
| 634 | } |
| 635 | |
| 636 | // Compute GEMM function |
| 637 | output_stage.run(); |
| 638 | return c; |
| 639 | } |
| 640 | |
| 641 | SimpleTensor<uint8_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, |
| 642 | bool add_bias) |
| 643 | { |
| 644 | // Create reference |
| 645 | TensorShape shape_bias(shape[0]); |
| 646 | |
| 647 | SimpleTensor<int32_t> a{ shape, DataType::S32, 1 }; |
| 648 | SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 }; |
| 649 | |
| 650 | // Fill reference |
| 651 | fill(a, 0); |
| 652 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 653 | const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier }; |
| 654 | const std::vector<int32_t> result_shift_vec = { result_shift }; |
| 655 | |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 656 | if(add_bias) |
| 657 | { |
| 658 | // Fill bias |
| 659 | fill(b, 1); |
| 660 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame] | 661 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 662 | } |
| 663 | else |
| 664 | { |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame] | 665 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, uint8_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, result_offset_after_shift, min, max); |
Gian Marco | 58c5794 | 2017-11-28 09:10:03 +0000 | [diff] [blame] | 666 | } |
| 667 | } |
| 668 | |
| 669 | TensorType _target{}; |
| 670 | SimpleTensor<uint8_t> _reference{}; |
| 671 | }; |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 672 | |
Sheri Zhang | 1b14c75 | 2020-03-09 14:29:52 +0000 | [diff] [blame] | 673 | template <typename TensorType, typename AccessorType, typename FunctionType, typename T> |
| 674 | class GEMMLowpQuantizeDownInt32ScaleByFloatValidationFixture : public framework::Fixture |
| 675 | { |
| 676 | public: |
| 677 | template <typename...> |
| 678 | void setup(DataType data_type, TensorShape shape, float result_real_multiplier, int32_t result_offset, int32_t min, int32_t max, bool add_bias) |
| 679 | { |
| 680 | _target = compute_target(data_type, shape, result_real_multiplier, result_offset, min, max, add_bias); |
| 681 | _reference = compute_reference(shape, result_real_multiplier, result_offset, min, max, add_bias); |
| 682 | } |
| 683 | |
| 684 | protected: |
| 685 | template <typename U> |
| 686 | void fill(U &&tensor, int i) |
| 687 | { |
| 688 | // To avoid data all being clampped |
| 689 | std::uniform_int_distribution<> distribution(-500, 500); |
| 690 | library->fill(tensor, distribution, i); |
| 691 | } |
| 692 | |
| 693 | TensorType compute_target(DataType data_type, const TensorShape &shape, float result_multiplier, int32_t result_offset, int32_t min, int32_t max, bool add_bias) |
| 694 | { |
| 695 | TensorShape shape_bias(shape[0]); |
| 696 | |
| 697 | // Create tensors |
| 698 | TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1); |
| 699 | TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1); |
| 700 | TensorType c = create_tensor<TensorType>(shape, data_type, 1); |
| 701 | |
| 702 | // create output stage info |
| 703 | GEMMLowpOutputStageInfo info; |
| 704 | info.gemmlowp_max_bound = max; |
| 705 | info.gemmlowp_min_bound = min; |
| 706 | info.gemmlowp_real_multiplier = result_multiplier; |
| 707 | info.gemmlowp_offset = result_offset; |
| 708 | info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FLOAT; |
| 709 | info.output_data_type = data_type; |
| 710 | |
| 711 | // Create and configure function |
| 712 | FunctionType output_stage; |
| 713 | output_stage.configure(&a, add_bias ? &b : nullptr, &c, info); |
| 714 | |
| 715 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 716 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 717 | |
| 718 | // Allocate tensors |
| 719 | a.allocator()->allocate(); |
| 720 | c.allocator()->allocate(); |
| 721 | |
| 722 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 723 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 724 | |
| 725 | // Fill tensor |
| 726 | fill(AccessorType(a), 0); |
| 727 | |
| 728 | if(add_bias) |
| 729 | { |
| 730 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 731 | |
| 732 | // Allocate bias tensor |
| 733 | b.allocator()->allocate(); |
| 734 | |
| 735 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 736 | |
| 737 | // Fill tensor |
| 738 | fill(AccessorType(b), 1); |
| 739 | } |
| 740 | |
| 741 | // Compute GEMM function |
| 742 | output_stage.run(); |
| 743 | return c; |
| 744 | } |
| 745 | |
| 746 | SimpleTensor<T> compute_reference(const TensorShape &shape, float_t result_real_multiplier, int32_t result_offset, int32_t min, int32_t max, bool add_bias) |
| 747 | { |
| 748 | // Create reference |
| 749 | TensorShape shape_bias(shape[0]); |
| 750 | |
| 751 | SimpleTensor<int32_t> a{ shape, DataType::S32, 1 }; |
| 752 | SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 }; |
| 753 | |
| 754 | // Fill reference |
| 755 | fill(a, 0); |
| 756 | |
| 757 | const std::vector<float_t> result_float_multiplier_vec = { result_real_multiplier }; |
| 758 | |
| 759 | if(add_bias) |
| 760 | { |
| 761 | // Fill bias |
| 762 | fill(b, 1); |
| 763 | |
| 764 | return reference::gemmlowp_quantize_down_scale_by_float<int32_t, T>(a, b, result_float_multiplier_vec, result_offset, min, max); |
| 765 | } |
| 766 | else |
| 767 | { |
| 768 | return reference::gemmlowp_quantize_down_scale_by_float<int32_t, T>(a, result_float_multiplier_vec, result_offset, min, max); |
| 769 | } |
| 770 | } |
| 771 | |
| 772 | TensorType _target{}; |
| 773 | SimpleTensor<T> _reference{}; |
| 774 | }; |
| 775 | |
Gian Marco Iodice | bc415af | 2019-06-13 15:58:32 +0100 | [diff] [blame] | 776 | template <typename TensorType, typename AccessorType, typename FunctionType> |
| 777 | class GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture : public framework::Fixture |
| 778 | { |
| 779 | public: |
| 780 | template <typename...> |
| 781 | void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
| 782 | { |
| 783 | _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias); |
| 784 | _reference = compute_reference(shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias); |
| 785 | } |
| 786 | |
| 787 | protected: |
| 788 | template <typename U> |
| 789 | void fill(U &&tensor, int i) |
| 790 | { |
| 791 | std::uniform_int_distribution<> distribution(-6000, 6000); |
| 792 | library->fill(tensor, distribution, i); |
| 793 | } |
| 794 | |
| 795 | TensorType compute_target(const TensorShape &shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min, int32_t max, bool add_bias) |
| 796 | { |
| 797 | TensorShape shape_bias(shape[0]); |
| 798 | |
| 799 | // Create tensors |
| 800 | TensorType a = create_tensor<TensorType>(shape, DataType::S32, 1); |
| 801 | TensorType b = create_tensor<TensorType>(shape_bias, DataType::S32, 1); |
| 802 | TensorType c = create_tensor<TensorType>(shape, DataType::QSYMM16, 1); |
| 803 | |
| 804 | // Create and configure function |
| 805 | FunctionType output_stage; |
| 806 | output_stage.configure(&a, add_bias ? &b : nullptr, &c, result_fixedpoint_multiplier, result_shift, min, max); |
| 807 | |
| 808 | ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 809 | ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 810 | |
| 811 | // Allocate tensors |
| 812 | a.allocator()->allocate(); |
| 813 | c.allocator()->allocate(); |
| 814 | |
| 815 | ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 816 | ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 817 | |
| 818 | // Fill tensor |
| 819 | fill(AccessorType(a), 0); |
| 820 | |
| 821 | if(add_bias) |
| 822 | { |
| 823 | ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 824 | |
| 825 | // Allocate bias tensor |
| 826 | b.allocator()->allocate(); |
| 827 | |
| 828 | ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 829 | |
| 830 | // Fill tensor |
| 831 | fill(AccessorType(b), 1); |
| 832 | } |
| 833 | |
| 834 | // Compute GEMM function |
| 835 | output_stage.run(); |
| 836 | return c; |
| 837 | } |
| 838 | |
| 839 | SimpleTensor<int16_t> compute_reference(const TensorShape &shape, int32_t result_fixed_point_multiplier, int32_t result_shift, int32_t min, int32_t max, |
| 840 | bool add_bias) |
| 841 | { |
| 842 | // Create reference |
| 843 | TensorShape shape_bias(shape[0]); |
| 844 | |
| 845 | SimpleTensor<int32_t> a{ shape, DataType::S32, 1 }; |
| 846 | SimpleTensor<int32_t> b{ shape_bias, DataType::S32, 1 }; |
| 847 | |
| 848 | // Fill reference |
| 849 | fill(a, 0); |
| 850 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame] | 851 | const std::vector<int32_t> result_fixed_point_multiplier_vec = { result_fixed_point_multiplier }; |
| 852 | const std::vector<int32_t> result_shift_vec = { result_shift }; |
| 853 | |
Gian Marco Iodice | bc415af | 2019-06-13 15:58:32 +0100 | [diff] [blame] | 854 | if(add_bias) |
| 855 | { |
| 856 | // Fill bias |
| 857 | fill(b, 1); |
| 858 | |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame] | 859 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int16_t>(a, b, result_fixed_point_multiplier_vec, result_shift_vec, 0, min, max); |
Gian Marco Iodice | bc415af | 2019-06-13 15:58:32 +0100 | [diff] [blame] | 860 | } |
| 861 | else |
| 862 | { |
Georgios Pinitas | 448a81f | 2019-11-21 14:10:25 +0000 | [diff] [blame] | 863 | return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, int16_t>(a, result_fixed_point_multiplier_vec, result_shift_vec, 0, min, max); |
Gian Marco Iodice | bc415af | 2019-06-13 15:58:32 +0100 | [diff] [blame] | 864 | } |
| 865 | } |
| 866 | |
| 867 | TensorType _target{}; |
| 868 | SimpleTensor<int16_t> _reference{}; |
| 869 | }; |
| 870 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 871 | template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 872 | class GEMMLowpMatrixMultiplyReshapedValidationFixture : public framework::Fixture |
| 873 | { |
| 874 | public: |
| 875 | template <typename...> |
| 876 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs, |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 877 | bool interleave_rhs, DataType data_type) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 878 | { |
| 879 | GEMMLHSMatrixInfo lhs_info; |
| 880 | lhs_info.m0 = m0; |
| 881 | lhs_info.k0 = k0; |
| 882 | lhs_info.v0 = v0; |
| 883 | lhs_info.interleave = interleave_lhs; |
| 884 | lhs_info.transpose = false; |
| 885 | |
| 886 | GEMMRHSMatrixInfo rhs_info; |
| 887 | rhs_info.n0 = n0; |
| 888 | rhs_info.k0 = k0; |
| 889 | rhs_info.h0 = h0; |
| 890 | rhs_info.interleave = interleave_rhs; |
| 891 | rhs_info.transpose = true; |
| 892 | |
| 893 | // Set the tensor shapes for LHS and RHS matrices |
| 894 | const TensorShape lhs_shape(k, m, batch_size); |
| 895 | const TensorShape rhs_shape(n, k, batch_size); |
| 896 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 897 | _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type); |
| 898 | _reference = compute_reference(lhs_shape, rhs_shape, data_type); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 899 | } |
| 900 | |
| 901 | protected: |
| 902 | template <typename U> |
| 903 | void fill(U &&tensor, int i) |
| 904 | { |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 905 | switch(tensor.data_type()) |
| 906 | { |
| 907 | case DataType::QASYMM8: |
| 908 | { |
| 909 | // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path |
| 910 | std::uniform_int_distribution<> distribution(1, 254); |
| 911 | library->fill(tensor, distribution, i); |
| 912 | } |
| 913 | break; |
| 914 | case DataType::QASYMM8_SIGNED: |
| 915 | { |
| 916 | std::uniform_int_distribution<> distribution(-127, 126); |
| 917 | library->fill(tensor, distribution, i); |
| 918 | } |
| 919 | break; |
| 920 | default: |
| 921 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 922 | } |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 923 | } |
| 924 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 925 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, DataType data_type) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 926 | { |
| 927 | // Create tensors |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 928 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 929 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 930 | TensorType lhs_reshaped; |
| 931 | TensorType rhs_reshaped; |
| 932 | TensorType dst; |
| 933 | |
| 934 | const unsigned int M = lhs_shape[1]; |
| 935 | const unsigned int N = rhs_shape[0]; |
| 936 | const unsigned int K = lhs_shape[0]; |
| 937 | |
| 938 | // The output tensor will be auto-initialized within the function |
| 939 | |
| 940 | // Create and configure function |
| 941 | ReshapeLHSFunctionType reshape_lhs; |
| 942 | ReshapeRHSFunctionType reshape_rhs; |
| 943 | GEMMFunctionType gemm; |
| 944 | reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); |
| 945 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
| 946 | gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K)); |
| 947 | |
| 948 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 949 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 950 | |
| 951 | // Allocate tensors |
| 952 | lhs.allocator()->allocate(); |
| 953 | rhs.allocator()->allocate(); |
| 954 | lhs_reshaped.allocator()->allocate(); |
| 955 | rhs_reshaped.allocator()->allocate(); |
| 956 | dst.allocator()->allocate(); |
| 957 | |
| 958 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 959 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 960 | ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 961 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 962 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 963 | |
| 964 | // Fill tensors |
| 965 | fill(AccessorType(lhs), 0); |
| 966 | fill(AccessorType(rhs), 1); |
| 967 | |
| 968 | // Compute GEMM |
| 969 | reshape_lhs.run(); |
| 970 | reshape_rhs.run(); |
| 971 | gemm.run(); |
| 972 | |
| 973 | return dst; |
| 974 | } |
| 975 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 976 | SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 977 | { |
| 978 | TensorShape dst_shape = lhs_shape; |
| 979 | dst_shape[0] = rhs_shape[0]; |
| 980 | dst_shape[1] = lhs_shape[1]; |
| 981 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 982 | switch(data_type) |
| 983 | { |
| 984 | case DataType::QASYMM8: |
| 985 | { |
| 986 | // Create reference |
| 987 | SimpleTensor<uint8_t> lhs{ lhs_shape, data_type, 1 }; |
| 988 | SimpleTensor<uint8_t> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 989 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 990 | // Fill reference |
| 991 | fill(lhs, 0); |
| 992 | fill(rhs, 1); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 993 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 994 | return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); |
| 995 | } |
| 996 | case DataType::QASYMM8_SIGNED: |
| 997 | { |
| 998 | // Create reference |
| 999 | SimpleTensor<int8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1000 | SimpleTensor<int8_t> rhs{ rhs_shape, data_type, 1 }; |
| 1001 | |
| 1002 | // Fill reference |
| 1003 | fill(lhs, 0); |
| 1004 | fill(rhs, 1); |
| 1005 | |
| 1006 | return reference::gemmlowp_matrix_multiply_core<int32_t, int8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1007 | } |
| 1008 | default: |
| 1009 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 1010 | } |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1011 | } |
| 1012 | |
| 1013 | TensorType _target{}; |
| 1014 | SimpleTensor<int32_t> _reference{}; |
| 1015 | }; |
| 1016 | |
| 1017 | template <typename TensorType, typename AccessorType, typename ReshapeLHSFunctionType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 1018 | class GEMMLowpMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture |
| 1019 | { |
| 1020 | public: |
| 1021 | template <typename...> |
| 1022 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1023 | bool interleave_lhs, bool interleave_rhs, DataType data_type) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1024 | { |
| 1025 | GEMMLHSMatrixInfo lhs_info; |
| 1026 | lhs_info.m0 = m0; |
| 1027 | lhs_info.k0 = k0; |
| 1028 | lhs_info.v0 = v0; |
| 1029 | lhs_info.interleave = interleave_lhs; |
| 1030 | lhs_info.transpose = false; |
| 1031 | |
| 1032 | GEMMRHSMatrixInfo rhs_info; |
| 1033 | rhs_info.n0 = n0; |
| 1034 | rhs_info.k0 = k0; |
| 1035 | rhs_info.h0 = h0; |
| 1036 | rhs_info.interleave = interleave_rhs; |
| 1037 | rhs_info.transpose = true; |
| 1038 | |
| 1039 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1040 | const unsigned int m = m_w * m_h; |
| 1041 | |
| 1042 | // Set the tensor shapes for LHS and RHS matrices |
| 1043 | const TensorShape lhs_shape(k, m, batch_size); |
| 1044 | const TensorShape rhs_shape(n, k, batch_size); |
| 1045 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1046 | _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, m_h, data_type); |
| 1047 | _reference = compute_reference(lhs_shape, rhs_shape, m_h, data_type); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1048 | } |
| 1049 | |
| 1050 | protected: |
| 1051 | template <typename U> |
| 1052 | void fill(U &&tensor, int i) |
| 1053 | { |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1054 | switch(tensor.data_type()) |
| 1055 | { |
| 1056 | case DataType::QASYMM8: |
| 1057 | { |
| 1058 | // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path |
| 1059 | std::uniform_int_distribution<> distribution(1, 254); |
| 1060 | library->fill(tensor, distribution, i); |
| 1061 | } |
| 1062 | break; |
| 1063 | case DataType::QASYMM8_SIGNED: |
| 1064 | { |
| 1065 | std::uniform_int_distribution<> distribution(-127, 126); |
| 1066 | library->fill(tensor, distribution, i); |
| 1067 | } |
| 1068 | break; |
| 1069 | default: |
| 1070 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 1071 | } |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1072 | } |
| 1073 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1074 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, unsigned int m_h, |
| 1075 | DataType data_type) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1076 | { |
| 1077 | // Create tensors |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1078 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1079 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1080 | TensorType lhs_reshaped; |
| 1081 | TensorType rhs_reshaped; |
| 1082 | TensorType dst; |
| 1083 | |
| 1084 | const unsigned int M = lhs_shape[1]; |
| 1085 | const unsigned int N = rhs_shape[0]; |
| 1086 | const unsigned int K = lhs_shape[0]; |
| 1087 | |
| 1088 | // The output tensor will be auto-initialized within the function |
| 1089 | |
| 1090 | // Create and configure function |
| 1091 | ReshapeLHSFunctionType reshape_lhs; |
| 1092 | ReshapeRHSFunctionType reshape_rhs; |
| 1093 | GEMMFunctionType gemm; |
| 1094 | reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info); |
| 1095 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
| 1096 | gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h)); |
| 1097 | |
| 1098 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1099 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1100 | |
| 1101 | // Allocate tensors |
| 1102 | lhs.allocator()->allocate(); |
| 1103 | rhs.allocator()->allocate(); |
| 1104 | lhs_reshaped.allocator()->allocate(); |
| 1105 | rhs_reshaped.allocator()->allocate(); |
| 1106 | dst.allocator()->allocate(); |
| 1107 | |
| 1108 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1109 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1110 | ARM_COMPUTE_EXPECT(!lhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1111 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1112 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1113 | |
| 1114 | // Fill tensors |
| 1115 | fill(AccessorType(lhs), 0); |
| 1116 | fill(AccessorType(rhs), 1); |
| 1117 | |
| 1118 | // Compute GEMM |
| 1119 | reshape_lhs.run(); |
| 1120 | reshape_rhs.run(); |
| 1121 | gemm.run(); |
| 1122 | |
| 1123 | return dst; |
| 1124 | } |
| 1125 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1126 | SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, unsigned int m_h, DataType data_type) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1127 | { |
| 1128 | TensorShape dst_shape = lhs_shape; |
| 1129 | dst_shape.set(0, rhs_shape[0]); |
| 1130 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1131 | dst_shape.set(2, m_h); |
| 1132 | dst_shape.set(3, lhs_shape[2]); |
| 1133 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1134 | switch(data_type) |
| 1135 | { |
| 1136 | case DataType::QASYMM8: |
| 1137 | { |
| 1138 | // Create reference |
| 1139 | SimpleTensor<uint8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1140 | SimpleTensor<uint8_t> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1141 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1142 | // Fill reference |
| 1143 | fill(lhs, 0); |
| 1144 | fill(rhs, 1); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1145 | |
Sheri Zhang | 28287af | 2020-02-25 14:13:54 +0000 | [diff] [blame] | 1146 | return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1147 | } |
| 1148 | case DataType::QASYMM8_SIGNED: |
| 1149 | { |
| 1150 | // Create reference |
| 1151 | SimpleTensor<int8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1152 | SimpleTensor<int8_t> rhs{ rhs_shape, data_type, 1 }; |
| 1153 | |
| 1154 | // Fill reference |
| 1155 | fill(lhs, 0); |
| 1156 | fill(rhs, 1); |
| 1157 | |
| 1158 | return reference::gemmlowp_matrix_multiply_core<int32_t, int8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1159 | } |
| 1160 | default: |
| 1161 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 1162 | } |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 1163 | } |
| 1164 | |
| 1165 | TensorType _target{}; |
| 1166 | SimpleTensor<int32_t> _reference{}; |
| 1167 | }; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1168 | |
| 1169 | template <typename TensorType, typename AccessorType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 1170 | class GEMMLowpMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture |
| 1171 | { |
| 1172 | public: |
| 1173 | template <typename...> |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1174 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, |
| 1175 | unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, DataType data_type) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1176 | { |
| 1177 | GEMMLHSMatrixInfo lhs_info; |
| 1178 | lhs_info.m0 = m0; |
| 1179 | lhs_info.k0 = k0; |
| 1180 | |
| 1181 | GEMMRHSMatrixInfo rhs_info; |
| 1182 | rhs_info.n0 = n0; |
| 1183 | rhs_info.k0 = k0; |
| 1184 | rhs_info.h0 = h0; |
| 1185 | rhs_info.interleave = interleave_rhs; |
| 1186 | rhs_info.transpose = transpose_rhs; |
| 1187 | |
| 1188 | // Set the tensor shapes for LHS and RHS matrices |
| 1189 | const TensorShape lhs_shape(k, m, batch_size); |
| 1190 | const TensorShape rhs_shape(n, k, batch_size); |
| 1191 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1192 | _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type); |
| 1193 | _reference = compute_reference(lhs_shape, rhs_shape, data_type); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1194 | } |
| 1195 | |
| 1196 | protected: |
| 1197 | template <typename U> |
| 1198 | void fill(U &&tensor, int i) |
| 1199 | { |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1200 | switch(tensor.data_type()) |
| 1201 | { |
| 1202 | case DataType::QASYMM8: |
| 1203 | { |
| 1204 | // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path |
| 1205 | std::uniform_int_distribution<> distribution(1, 254); |
| 1206 | library->fill(tensor, distribution, i); |
| 1207 | } |
| 1208 | break; |
| 1209 | case DataType::QASYMM8_SIGNED: |
| 1210 | { |
| 1211 | std::uniform_int_distribution<> distribution(-127, 126); |
| 1212 | library->fill(tensor, distribution, i); |
| 1213 | } |
| 1214 | break; |
| 1215 | default: |
| 1216 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 1217 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1218 | } |
| 1219 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1220 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, |
| 1221 | const GEMMRHSMatrixInfo &rhs_info, DataType data_type) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1222 | { |
| 1223 | // Create tensors |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1224 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1225 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1226 | TensorType rhs_reshaped; |
| 1227 | TensorType dst; |
| 1228 | |
| 1229 | const unsigned int M = lhs_shape[1]; |
| 1230 | const unsigned int N = rhs_shape[0]; |
| 1231 | const unsigned int K = lhs_shape[0]; |
| 1232 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 1233 | GEMMKernelInfo gemm_info; |
| 1234 | gemm_info.m = M; |
| 1235 | gemm_info.n = N; |
| 1236 | gemm_info.k = K; |
| 1237 | gemm_info.lhs_info = lhs_info; |
| 1238 | gemm_info.rhs_info = rhs_info; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1239 | // The output tensor will be auto-initialized within the function |
| 1240 | |
| 1241 | // Create and configure function |
| 1242 | ReshapeRHSFunctionType reshape_rhs; |
| 1243 | GEMMFunctionType gemm; |
| 1244 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 1245 | gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1246 | |
| 1247 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1248 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1249 | |
| 1250 | // Allocate tensors |
| 1251 | lhs.allocator()->allocate(); |
| 1252 | rhs.allocator()->allocate(); |
| 1253 | rhs_reshaped.allocator()->allocate(); |
| 1254 | dst.allocator()->allocate(); |
| 1255 | |
| 1256 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1257 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1258 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1259 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1260 | |
| 1261 | // Fill tensors |
| 1262 | fill(AccessorType(lhs), 0); |
| 1263 | fill(AccessorType(rhs), 1); |
| 1264 | |
| 1265 | // Compute GEMM |
| 1266 | reshape_rhs.run(); |
| 1267 | gemm.run(); |
| 1268 | |
| 1269 | return dst; |
| 1270 | } |
| 1271 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1272 | SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1273 | { |
| 1274 | TensorShape dst_shape = lhs_shape; |
| 1275 | dst_shape[0] = rhs_shape[0]; |
| 1276 | dst_shape[1] = lhs_shape[1]; |
| 1277 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1278 | if(data_type == DataType::QASYMM8) |
| 1279 | { |
| 1280 | // Create reference |
| 1281 | SimpleTensor<uint8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1282 | SimpleTensor<uint8_t> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1283 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1284 | // Fill reference |
| 1285 | fill(lhs, 0); |
| 1286 | fill(rhs, 1); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1287 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1288 | return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1289 | } |
| 1290 | else |
| 1291 | { |
| 1292 | // Create reference |
| 1293 | SimpleTensor<int8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1294 | SimpleTensor<int8_t> rhs{ rhs_shape, data_type, 1 }; |
| 1295 | |
| 1296 | // Fill reference |
| 1297 | fill(lhs, 0); |
| 1298 | fill(rhs, 1); |
| 1299 | |
| 1300 | return reference::gemmlowp_matrix_multiply_core<int32_t, int8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1301 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1302 | } |
| 1303 | |
| 1304 | TensorType _target{}; |
| 1305 | SimpleTensor<int32_t> _reference{}; |
| 1306 | }; |
| 1307 | |
| 1308 | template <typename TensorType, typename AccessorType, typename ReshapeRHSFunctionType, typename GEMMFunctionType> |
| 1309 | class GEMMLowpMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture |
| 1310 | { |
| 1311 | public: |
| 1312 | template <typename...> |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1313 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, |
| 1314 | unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, DataType data_type) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1315 | { |
| 1316 | GEMMLHSMatrixInfo lhs_info; |
| 1317 | lhs_info.m0 = m0; |
| 1318 | lhs_info.k0 = k0; |
| 1319 | |
| 1320 | GEMMRHSMatrixInfo rhs_info; |
| 1321 | rhs_info.n0 = n0; |
| 1322 | rhs_info.k0 = k0; |
| 1323 | rhs_info.h0 = h0; |
| 1324 | rhs_info.interleave = interleave_rhs; |
| 1325 | rhs_info.transpose = transpose_rhs; |
| 1326 | |
| 1327 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1328 | const unsigned int m = m_w * m_h; |
| 1329 | |
| 1330 | // Set the tensor shapes for LHS and RHS matrices |
| 1331 | const TensorShape lhs_shape(k, m, batch_size); |
| 1332 | const TensorShape rhs_shape(n, k, batch_size); |
| 1333 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1334 | _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, m_h, data_type); |
| 1335 | _reference = compute_reference(lhs_shape, rhs_shape, m_h, data_type); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1336 | } |
| 1337 | |
| 1338 | protected: |
| 1339 | template <typename U> |
| 1340 | void fill(U &&tensor, int i) |
| 1341 | { |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1342 | switch(tensor.data_type()) |
| 1343 | { |
| 1344 | case DataType::QASYMM8: |
| 1345 | { |
| 1346 | // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path |
| 1347 | std::uniform_int_distribution<> distribution(1, 254); |
| 1348 | library->fill(tensor, distribution, i); |
| 1349 | } |
| 1350 | break; |
| 1351 | case DataType::QASYMM8_SIGNED: |
| 1352 | { |
| 1353 | std::uniform_int_distribution<> distribution(-127, 126); |
| 1354 | library->fill(tensor, distribution, i); |
| 1355 | } |
| 1356 | break; |
| 1357 | default: |
| 1358 | ARM_COMPUTE_ERROR("Unsupported data type"); |
| 1359 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1360 | } |
| 1361 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1362 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, |
| 1363 | const GEMMRHSMatrixInfo &rhs_info, unsigned int m_h, DataType data_type) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1364 | { |
| 1365 | // Create tensors |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1366 | TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); |
| 1367 | TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1368 | TensorType rhs_reshaped; |
| 1369 | TensorType dst; |
| 1370 | |
| 1371 | const unsigned int M = lhs_shape[1]; |
| 1372 | const unsigned int N = rhs_shape[0]; |
| 1373 | const unsigned int K = lhs_shape[0]; |
| 1374 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 1375 | GEMMKernelInfo gemm_info; |
| 1376 | gemm_info.m = M; |
| 1377 | gemm_info.n = N; |
| 1378 | gemm_info.k = K; |
| 1379 | gemm_info.depth_output_gemm3d = m_h; |
| 1380 | gemm_info.lhs_info = lhs_info; |
| 1381 | gemm_info.rhs_info = rhs_info; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1382 | // The output tensor will be auto-initialized within the function |
| 1383 | |
| 1384 | // Create and configure function |
| 1385 | ReshapeRHSFunctionType reshape_rhs; |
| 1386 | GEMMFunctionType gemm; |
| 1387 | reshape_rhs.configure(&rhs, &rhs_reshaped, rhs_info); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 1388 | gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1389 | |
| 1390 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1391 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1392 | |
| 1393 | // Allocate tensors |
| 1394 | lhs.allocator()->allocate(); |
| 1395 | rhs.allocator()->allocate(); |
| 1396 | rhs_reshaped.allocator()->allocate(); |
| 1397 | dst.allocator()->allocate(); |
| 1398 | |
| 1399 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1400 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1401 | ARM_COMPUTE_EXPECT(!rhs_reshaped.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1402 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1403 | |
| 1404 | // Fill tensors |
| 1405 | fill(AccessorType(lhs), 0); |
| 1406 | fill(AccessorType(rhs), 1); |
| 1407 | |
| 1408 | // Compute GEMM |
| 1409 | reshape_rhs.run(); |
| 1410 | gemm.run(); |
| 1411 | |
| 1412 | return dst; |
| 1413 | } |
| 1414 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1415 | SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, unsigned int m_h, DataType data_type) |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1416 | { |
| 1417 | TensorShape dst_shape = lhs_shape; |
| 1418 | dst_shape.set(0, rhs_shape[0]); |
| 1419 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1420 | dst_shape.set(2, m_h); |
| 1421 | dst_shape.set(3, lhs_shape[2]); |
| 1422 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1423 | if(data_type == DataType::QASYMM8) |
| 1424 | { |
| 1425 | // Create reference |
| 1426 | SimpleTensor<uint8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1427 | SimpleTensor<uint8_t> rhs{ rhs_shape, data_type, 1 }; |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1428 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1429 | // Fill reference |
| 1430 | fill(lhs, 0); |
| 1431 | fill(rhs, 1); |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1432 | |
Michele Di Giorgio | f9179d3 | 2019-11-27 16:17:30 +0000 | [diff] [blame] | 1433 | return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1434 | } |
| 1435 | else |
| 1436 | { |
| 1437 | // Create reference |
| 1438 | SimpleTensor<int8_t> lhs{ lhs_shape, data_type, 1 }; |
| 1439 | SimpleTensor<int8_t> rhs{ rhs_shape, data_type, 1 }; |
| 1440 | |
| 1441 | // Fill reference |
| 1442 | fill(lhs, 0); |
| 1443 | fill(rhs, 1); |
| 1444 | |
| 1445 | return reference::gemmlowp_matrix_multiply_core<int32_t, int8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1446 | } |
Gian Marco Iodice | 62251f7 | 2019-03-11 16:07:12 +0000 | [diff] [blame] | 1447 | } |
| 1448 | |
| 1449 | TensorType _target{}; |
| 1450 | SimpleTensor<int32_t> _reference{}; |
| 1451 | }; |
Gian Marco Iodice | e751062 | 2019-06-03 17:28:17 +0100 | [diff] [blame] | 1452 | |
| 1453 | template <typename TensorType, typename AccessorType, typename GEMMFunctionType> |
| 1454 | class GEMMLowpMatrixMultiplyNativeValidationFixture : public framework::Fixture |
| 1455 | { |
| 1456 | public: |
| 1457 | template <typename...> |
| 1458 | void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0) |
| 1459 | { |
| 1460 | GEMMLHSMatrixInfo lhs_info; |
| 1461 | lhs_info.m0 = m0; |
| 1462 | lhs_info.k0 = k0; |
| 1463 | |
| 1464 | GEMMRHSMatrixInfo rhs_info; |
| 1465 | rhs_info.n0 = n0; |
| 1466 | rhs_info.k0 = k0; |
| 1467 | |
| 1468 | // Set the tensor shapes for LHS and RHS matrices |
| 1469 | const TensorShape lhs_shape(k, m, batch_size); |
| 1470 | const TensorShape rhs_shape(n, k, batch_size); |
| 1471 | |
| 1472 | _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info); |
| 1473 | _reference = compute_reference(lhs_shape, rhs_shape); |
| 1474 | } |
| 1475 | |
| 1476 | protected: |
| 1477 | template <typename U> |
| 1478 | void fill(U &&tensor, int i) |
| 1479 | { |
| 1480 | // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path |
| 1481 | std::uniform_int_distribution<> distribution(1, 254); |
| 1482 | library->fill(tensor, distribution, i); |
| 1483 | } |
| 1484 | |
| 1485 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info) |
| 1486 | { |
| 1487 | // Create tensors |
| 1488 | TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::QASYMM8, 1); |
| 1489 | TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::QASYMM8, 1); |
| 1490 | TensorType dst; |
| 1491 | |
| 1492 | const unsigned int M = lhs_shape[1]; |
| 1493 | const unsigned int N = rhs_shape[0]; |
| 1494 | const unsigned int K = lhs_shape[0]; |
| 1495 | |
| 1496 | // The output tensor will be auto-initialized within the function |
| 1497 | |
| 1498 | // Create and configure function |
| 1499 | GEMMFunctionType gemm; |
| 1500 | gemm.configure(&lhs, &rhs, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K)); |
| 1501 | |
| 1502 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1503 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1504 | |
| 1505 | // Allocate tensors |
| 1506 | lhs.allocator()->allocate(); |
| 1507 | rhs.allocator()->allocate(); |
| 1508 | dst.allocator()->allocate(); |
| 1509 | |
| 1510 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1511 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1512 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1513 | |
| 1514 | // Fill tensors |
| 1515 | fill(AccessorType(lhs), 0); |
| 1516 | fill(AccessorType(rhs), 1); |
| 1517 | |
| 1518 | // Compute GEMM |
| 1519 | gemm.run(); |
| 1520 | |
| 1521 | return dst; |
| 1522 | } |
| 1523 | |
| 1524 | SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape) |
| 1525 | { |
| 1526 | TensorShape dst_shape = lhs_shape; |
| 1527 | dst_shape[0] = rhs_shape[0]; |
| 1528 | dst_shape[1] = lhs_shape[1]; |
| 1529 | |
| 1530 | // Create reference |
| 1531 | SimpleTensor<uint8_t> lhs{ lhs_shape, DataType::QASYMM8, 1 }; |
| 1532 | SimpleTensor<uint8_t> rhs{ rhs_shape, DataType::QASYMM8, 1 }; |
| 1533 | |
| 1534 | // Fill reference |
| 1535 | fill(lhs, 0); |
| 1536 | fill(rhs, 1); |
| 1537 | |
| 1538 | return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1539 | } |
| 1540 | |
| 1541 | TensorType _target{}; |
| 1542 | SimpleTensor<int32_t> _reference{}; |
| 1543 | }; |
| 1544 | |
| 1545 | template <typename TensorType, typename AccessorType, typename GEMMFunctionType> |
| 1546 | class GEMMLowpMatrixMultiplyNative3DValidationFixture : public framework::Fixture |
| 1547 | { |
| 1548 | public: |
| 1549 | template <typename...> |
| 1550 | void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0) |
| 1551 | { |
| 1552 | GEMMLHSMatrixInfo lhs_info; |
| 1553 | lhs_info.m0 = m0; |
| 1554 | lhs_info.k0 = k0; |
| 1555 | |
| 1556 | GEMMRHSMatrixInfo rhs_info; |
| 1557 | rhs_info.n0 = n0; |
| 1558 | rhs_info.k0 = k0; |
| 1559 | |
| 1560 | // In case of GEMM3D, m is the product between m_w and m_h |
| 1561 | const unsigned int m = m_w * m_h; |
| 1562 | |
| 1563 | // Set the tensor shapes for LHS and RHS matrices |
| 1564 | const TensorShape lhs_shape(k, m, batch_size); |
| 1565 | const TensorShape rhs_shape(n, k, batch_size); |
| 1566 | |
| 1567 | _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, m_h); |
| 1568 | _reference = compute_reference(lhs_shape, rhs_shape, m_h); |
| 1569 | } |
| 1570 | |
| 1571 | protected: |
| 1572 | template <typename U> |
| 1573 | void fill(U &&tensor, int i) |
| 1574 | { |
| 1575 | // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path |
| 1576 | std::uniform_int_distribution<> distribution(1, 254); |
| 1577 | library->fill(tensor, distribution, i); |
| 1578 | } |
| 1579 | |
| 1580 | TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, unsigned int m_h) |
| 1581 | { |
| 1582 | // Create tensors |
| 1583 | TensorType lhs = create_tensor<TensorType>(lhs_shape, DataType::QASYMM8, 1); |
| 1584 | TensorType rhs = create_tensor<TensorType>(rhs_shape, DataType::QASYMM8, 1); |
| 1585 | TensorType dst; |
| 1586 | |
| 1587 | const unsigned int M = lhs_shape[1]; |
| 1588 | const unsigned int N = rhs_shape[0]; |
| 1589 | const unsigned int K = lhs_shape[0]; |
| 1590 | |
| 1591 | // The output tensor will be auto-initialized within the function |
| 1592 | |
| 1593 | // Create and configure function |
| 1594 | GEMMFunctionType gemm; |
| 1595 | gemm.configure(&lhs, &rhs, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h)); |
| 1596 | |
| 1597 | ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1598 | ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1599 | |
| 1600 | // Allocate tensors |
| 1601 | lhs.allocator()->allocate(); |
| 1602 | rhs.allocator()->allocate(); |
| 1603 | dst.allocator()->allocate(); |
| 1604 | |
| 1605 | ARM_COMPUTE_EXPECT(!lhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1606 | ARM_COMPUTE_EXPECT(!rhs.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1607 | ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); |
| 1608 | |
| 1609 | // Fill tensors |
| 1610 | fill(AccessorType(lhs), 0); |
| 1611 | fill(AccessorType(rhs), 1); |
| 1612 | |
| 1613 | // Compute GEMM |
| 1614 | gemm.run(); |
| 1615 | |
| 1616 | return dst; |
| 1617 | } |
| 1618 | |
| 1619 | SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, unsigned int m_h) |
| 1620 | { |
| 1621 | TensorShape dst_shape = lhs_shape; |
| 1622 | dst_shape.set(0, rhs_shape[0]); |
| 1623 | dst_shape.set(1, lhs_shape[1] / m_h); |
| 1624 | dst_shape.set(2, m_h); |
| 1625 | dst_shape.set(3, lhs_shape[2]); |
| 1626 | |
| 1627 | // Create reference |
| 1628 | SimpleTensor<uint8_t> lhs{ lhs_shape, DataType::QASYMM8, 1 }; |
| 1629 | SimpleTensor<uint8_t> rhs{ rhs_shape, DataType::QASYMM8, 1 }; |
| 1630 | |
| 1631 | // Fill reference |
| 1632 | fill(lhs, 0); |
| 1633 | fill(rhs, 1); |
| 1634 | |
| 1635 | return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); |
| 1636 | } |
| 1637 | |
| 1638 | TensorType _target{}; |
| 1639 | SimpleTensor<int32_t> _reference{}; |
| 1640 | }; |
Pablo Tello | 299025a | 2017-09-29 11:30:12 +0100 | [diff] [blame] | 1641 | } // namespace validation |
| 1642 | } // namespace test |
| 1643 | } // namespace arm_compute |
George Wort | 2d7e683 | 2019-02-22 16:37:41 +0000 | [diff] [blame] | 1644 | #endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */ |