xref: /aosp_15_r20/external/ComputeLibrary/tests/validation/NEON/QLSTMLayerNormalization.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2020-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #include "arm_compute/core/Types.h"
25 #include "arm_compute/runtime/Tensor.h"
26 #include "arm_compute/runtime/TensorAllocator.h"
27 #include "src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.h"
28 #include "tests/NEON/Accessor.h"
29 #include "tests/NEON/Helper.h"
30 #include "tests/PaddingCalculator.h"
31 #include "tests/datasets/ShapeDatasets.h"
32 #include "tests/framework/Asserts.h"
33 #include "tests/framework/Macros.h"
34 #include "tests/framework/datasets/Datasets.h"
35 #include "tests/validation/Helpers.h"
36 #include "tests/validation/Validation.h"
37 #include "tests/validation/fixtures/QLSTMLayerNormalizationFixture.h"
38 
39 namespace arm_compute
40 {
41 namespace test
42 {
43 namespace validation
44 {
45 namespace
46 {
47 constexpr uint32_t vector_size_byte = 16;
48 
49 using test::datasets::ShapeDataset;
50 using NEQLSTMLayerNormalization = NESynthetizeFunction<NEQLSTMLayerNormalizationKernel>;
51 
52 template <uint32_t num_elements_per_iter, uint32_t num_batches, uint32_t num_iteration>
53 class QLSTMLayerNormShapeDataSet : public ShapeDataset
54 {
55     static constexpr auto boundary_minus_one = num_elements_per_iter * num_iteration - 1;
56     static constexpr auto boundary           = num_elements_per_iter * num_iteration;
57     static constexpr auto boundary_plus_one  = num_elements_per_iter * num_iteration + 1;
58 
59 public:
QLSTMLayerNormShapeDataSet(std::string name)60     QLSTMLayerNormShapeDataSet(std::string name)
61         : ShapeDataset(name,
62     {
63         TensorShape{ boundary_minus_one, num_batches },
64                      TensorShape{ boundary, num_batches },
65                      TensorShape{ boundary_plus_one, num_batches }
66     })
67     {
68     }
69 };
70 
71 template <uint32_t num_elements_per_iter, uint32_t num_batches>
72 class QLSTMLayerNormShapeDataSet<num_elements_per_iter, num_batches, 0> : public ShapeDataset
73 {
74 public:
QLSTMLayerNormShapeDataSet(std::string name)75     QLSTMLayerNormShapeDataSet(std::string name)
76         : ShapeDataset(name,
77     {
78         TensorShape{ 1, num_batches },
79                      TensorShape{ 2, num_batches }
80     })
81     {
82     }
83 };
84 } // namespace
85 TEST_SUITE(NEON)
TEST_SUITE(QLSTMLayerNormalization) const86 TEST_SUITE(QLSTMLayerNormalization)
87 
88 static const TensorShape correct_input_shape{ TensorShape(15U, 2U) };
89 static const TensorShape correct_weight_shape{ TensorShape(15U) };
90 static const TensorShape correct_bias_shape{ TensorShape(15U) };
91 static const TensorShape correct_output_shape{ correct_input_shape };
92 static const DataType    correct_input_dt{ DataType::QSYMM16 };
93 static const DataType    correct_weight_dt{ DataType::QSYMM16 };
94 static const DataType    correct_bias_dt{ DataType::S32 };
95 static const DataType    correct_output_dt{ correct_input_dt };
96 static const uint32_t    tensor_num_channel{ 1 };
97 
98 // *INDENT-OFF*
99 // clang-format off
100 
101 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL,
102     zip(zip(zip(
103         framework::dataset::make("InputInfo", {
104             TensorInfo(correct_input_shape, tensor_num_channel, DataType::F16), // input supports only QSYMM16
105             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // weight supports only QSYMM16
106             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // bias supports only S32
107             TensorInfo(TensorShape(15U, 2U, 2U), tensor_num_channel, correct_input_dt), // input supports only up to 2D
108             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // weight supports only up to 1D
109             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // bias supports only up to 1D
110             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // input_shape[0] != weight_shape[0] should fail
111             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // weight_shape[0] != bias_shape[0] should fail
112             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // output shape mismatches with input shape
113             TensorInfo(correct_input_shape, tensor_num_channel, correct_input_dt), // output data type mismatches with input data type
114         }),
115         framework::dataset::make("WeightInfo", {
116             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
117             TensorInfo(correct_weight_shape, tensor_num_channel, DataType::F16),
118             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
119             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
120             TensorInfo(TensorShape(15U, 2U), tensor_num_channel, correct_weight_dt),
121             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
122             TensorInfo(TensorShape(14U), tensor_num_channel, correct_weight_dt),
123             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
124             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
125             TensorInfo(correct_weight_shape, tensor_num_channel, correct_weight_dt),
126         })
127     ),
128         framework::dataset::make("BiasInfo", {
129             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
130             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
131             TensorInfo(correct_bias_shape, tensor_num_channel, DataType::QSYMM16),
132             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
133             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
134             TensorInfo(TensorShape(15U, 2U), tensor_num_channel, correct_bias_dt),
135             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
136             TensorInfo(TensorShape(14U), tensor_num_channel, correct_bias_dt),
137             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
138             TensorInfo(correct_bias_shape, tensor_num_channel, correct_bias_dt),
139         })
140     ),
141         framework::dataset::make("OutputInfo", {
142             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
143             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
144             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
145             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
146             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
147             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
148             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
149             TensorInfo(correct_output_shape, tensor_num_channel, correct_output_dt),
150             TensorInfo(TensorShape(15, 3), tensor_num_channel, correct_output_dt),
151             TensorInfo(correct_output_shape, tensor_num_channel, DataType::S32),
152         })
153     ),
154      input_info, weight_info, bias_info, output_info)
155 {
156     const Status s = NEQLSTMLayerNormalization::validate(&input_info, &output_info, &weight_info, &bias_info);
157     ARM_COMPUTE_EXPECT(!bool(s), framework::LogLevel::ERRORS);
158 }
159 
160 // clang-format on
161 // *INDENT-ON*
162 
163 template <typename T>
164 using NEQLSTMLayerNormalizationFixture = QLSTMLayerNormalizationValidationFixture<Tensor, Accessor, NEQLSTMLayerNormalization, T>;
165 
166 TEST_SUITE(Quantized)
167 TEST_SUITE(QSYMM16)
168 
169 /** Tests will be targetting
170  * - Comparison between optimized kernel and the exact same but scalar version of reference kernel
171  * - Input shapes of 1D and 2D with the first dimension covers boundary values of 128-bit vector size (0~3 iterations)
172  * - Weight and bias 1D shape that have same size as that of input shapes
173  * - Quantization scale is greater and smaller than one.
174  * - Input values will be noted in fixture.
175  *
176  * What we can't test
177  * - Since reference kernel uses the exact the same algorithm in the same quantized domain
178  *   it is hard to fully test whether the algorithm accomplishes what it is supposed to.
179  * - The algorithm has been sensitive to quantization scale but it is hard to fully test
180  *   the sensitivity due to aforementioned reason.
181  * - Again, it is hard to fully test corner values due to the exact same algorithm of the
182  *   reference kernel and the optimized kernel.
183  */
184 
185 constexpr uint32_t qsymm16_per_vector = vector_size_byte / sizeof(int16_t);
186 
187 #define QSYMM16_DATASET_ITER(num_input_batch, num_iter)                                                              \
188     combine(combine(zip(zip(QLSTMLayerNormShapeDataSet<qsymm16_per_vector, num_input_batch, num_iter>("InputShape"), \
189                             QLSTMLayerNormShapeDataSet<qsymm16_per_vector, 1, num_iter>("WeightShape")),             \
190                         QLSTMLayerNormShapeDataSet<qsymm16_per_vector, 1, num_iter>("BiasShape")),                   \
191                     framework::dataset::make("DataType", DataType::QSYMM16)),                                        \
192             framework::dataset::make("WeightQuantizationInfo", { QuantizationInfo(1. / 8192), QuantizationInfo(8192) }))
193 
194 #define QSYMM16_DATASET_1D \
195     concat(concat(QSYMM16_DATASET_ITER(1, 0), QSYMM16_DATASET_ITER(1, 1)), QSYMM16_DATASET_ITER(1, 2))
196 
197 #define QSYMM16_DATASET_2D \
198     concat(concat(QSYMM16_DATASET_ITER(3, 0), QSYMM16_DATASET_ITER(3, 1)), QSYMM16_DATASET_ITER(3, 2))
199 
FIXTURE_DATA_TEST_CASE(RandomValue1D,NEQLSTMLayerNormalizationFixture<int16_t>,framework::DatasetMode::ALL,QSYMM16_DATASET_1D)200 FIXTURE_DATA_TEST_CASE(RandomValue1D, NEQLSTMLayerNormalizationFixture<int16_t>, framework::DatasetMode::ALL, QSYMM16_DATASET_1D)
201 {
202     // Validate output
203     validate(Accessor(_target), _reference);
204 }
205 
FIXTURE_DATA_TEST_CASE(RandomValue2D,NEQLSTMLayerNormalizationFixture<int16_t>,framework::DatasetMode::ALL,QSYMM16_DATASET_2D)206 FIXTURE_DATA_TEST_CASE(RandomValue2D, NEQLSTMLayerNormalizationFixture<int16_t>, framework::DatasetMode::ALL, QSYMM16_DATASET_2D)
207 {
208     // Validate output
209     validate(Accessor(_target), _reference);
210 }
211 
212 #undef QSYMM16_DATASET_ITER
213 #undef QSYMM16_DATASET_2D
214 #undef QSYMM16_DATASET_1D
215 
216 TEST_SUITE_END() // QSYMM16
217 TEST_SUITE_END() // Quantized
218 TEST_SUITE_END() // QLSTMLayerNormalization
219 TEST_SUITE_END() // Neon
220 
221 } // namespace validation
222 } // namespace test
223 } // namespace arm_compute
224