xref: /aosp_15_r20/external/ComputeLibrary/tests/validation/NEON/ActivationLayer.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2022 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/core/utils/misc/Traits.h"
26 #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h"
27 #include "arm_compute/runtime/RuntimeContext.h"
28 #include "arm_compute/runtime/Tensor.h"
29 #include "arm_compute/runtime/TensorAllocator.h"
30 #include "src/common/cpuinfo/CpuIsaInfo.h"
31 #include "src/cpu/kernels/CpuActivationKernel.h"
32 #include "tests/NEON/Accessor.h"
33 #include "tests/PaddingCalculator.h"
34 #include "tests/datasets/ActivationFunctionsDataset.h"
35 #include "tests/datasets/ShapeDatasets.h"
36 #include "tests/framework/Asserts.h"
37 #include "tests/framework/Macros.h"
38 #include "tests/framework/datasets/Datasets.h"
39 #include "tests/validation/Validation.h"
40 #include "tests/validation/fixtures/ActivationLayerFixture.h"
41 
42 #include "arm_compute/Acl.hpp"
43 #include "support/Requires.h"
44 
45 namespace arm_compute
46 {
47 namespace test
48 {
49 namespace validation
50 {
51 namespace
52 {
53 RelativeTolerance<float> tolerance_float_sqrt(0.0001f);
54 
55 /** Define relative tolerance of the activation layer.
56  *
57  * @param[in] data_type  The data type used.
58  * @param[in] activation The activation function used.
59  *
60  * @return Relative tolerance depending on the activation function.
61  */
relative_tolerance(DataType data_type,ActivationLayerInfo::ActivationFunction activation)62 RelativeTolerance<float> relative_tolerance(DataType data_type, ActivationLayerInfo::ActivationFunction activation)
63 {
64     switch(activation)
65     {
66         case ActivationLayerInfo::ActivationFunction::LOGISTIC:
67         case ActivationLayerInfo::ActivationFunction::ELU:
68         case ActivationLayerInfo::ActivationFunction::SQRT:
69         case ActivationLayerInfo::ActivationFunction::TANH:
70         case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
71         case ActivationLayerInfo::ActivationFunction::SWISH:
72         case ActivationLayerInfo::ActivationFunction::GELU:
73             switch(data_type)
74             {
75                 case DataType::F16:
76 #if defined(ENABLE_SVE)
77                     return RelativeTolerance<float>(0.25f);
78 #else  // !defined(ENABLE_SVE)
79                     return RelativeTolerance<float>(0.1f);
80 #endif // defined(ENABLE_SVE)
81                 default:
82                     return RelativeTolerance<float>(0.05f);
83             }
84         case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
85             switch(data_type)
86             {
87                 case DataType::F16:
88 #if defined(ENABLE_SVE)
89                     return RelativeTolerance<float>(0.9f);
90 #else  // !defined(ENABLE_SVE)
91                     return RelativeTolerance<float>(0.01f);
92 #endif // defined(ENABLE_SVE)
93                 default:
94                     return RelativeTolerance<float>(0.00001f);
95             }
96         default:
97             return RelativeTolerance<float>(0.f);
98     }
99 }
100 
101 /** Define absolute tolerance of the activation layer.
102  *
103  * @param[in] data_type  The data type used.
104  * @param[in] activation The activation function used.
105  *
106  * @return Absolute tolerance depending on the activation function.
107  */
absolute_tolerance(DataType data_type,ActivationLayerInfo::ActivationFunction activation)108 AbsoluteTolerance<float> absolute_tolerance(DataType data_type, ActivationLayerInfo::ActivationFunction activation)
109 {
110     switch(activation)
111     {
112         case ActivationLayerInfo::ActivationFunction::LOGISTIC:
113         case ActivationLayerInfo::ActivationFunction::SQRT:
114         case ActivationLayerInfo::ActivationFunction::TANH:
115         case ActivationLayerInfo::ActivationFunction::SWISH:
116         case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
117             switch(data_type)
118             {
119                 case DataType::F16:
120 #if defined(ENABLE_SVE)
121                     return AbsoluteTolerance<float>(0.25f);
122 #else  // !defined(ENABLE_SVE)
123                     return AbsoluteTolerance<float>(0.01f);
124 #endif // defined(ENABLE_SVE)
125                 default:
126                     return AbsoluteTolerance<float>(0.00001f);
127             }
128         case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
129             switch(data_type)
130             {
131                 case DataType::F16:
132 #if defined(ENABLE_SVE)
133                     return AbsoluteTolerance<float>(0.9f);
134 #else  // !defined(ENABLE_SVE)
135                     return AbsoluteTolerance<float>(0.01f);
136 #endif // defined(ENABLE_SVE)
137                 default:
138                     return AbsoluteTolerance<float>(0.00001f);
139             }
140         default:
141             return AbsoluteTolerance<float>(0.f);
142     }
143 }
144 
145 /** Define absolute tolerance of the activation layer for qasymm8.
146  *
147  * @param[in] activation The activation function used.
148  *
149  * @return Absolute tolerance depending on the activation function.
150  */
tolerance_qasymm8(ActivationLayerInfo::ActivationFunction activation)151 AbsoluteTolerance<uint8_t> tolerance_qasymm8(ActivationLayerInfo::ActivationFunction activation)
152 {
153     switch(activation)
154     {
155         case ActivationLayerInfo::ActivationFunction::LOGISTIC:
156         case ActivationLayerInfo::ActivationFunction::SQRT:
157         case ActivationLayerInfo::ActivationFunction::TANH:
158         case ActivationLayerInfo::ActivationFunction::HARD_SWISH:
159         case ActivationLayerInfo::ActivationFunction::SOFT_RELU:
160         case ActivationLayerInfo::ActivationFunction::LEAKY_RELU:
161             return AbsoluteTolerance<uint8_t>(1);
162         default:
163             return AbsoluteTolerance<uint8_t>(0);
164     }
165 }
166 
167 constexpr AbsoluteTolerance<int16_t> tolerance_qsymm16(1);
168 
169 /** CNN data types */
170 const auto CNNDataTypes = framework::dataset::make("DataType",
171 {
172 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
173     DataType::F16,
174 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
175     DataType::F32,
176 });
177 
178 const auto NeonActivationFunctionsDataset = concat(datasets::ActivationFunctions(),
179                                                    framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::HARD_SWISH, ActivationLayerInfo::ActivationFunction::SWISH }));
180 
181 /** Input data sets. */
182 const auto ActivationDataset = combine(combine(framework::dataset::make("InPlace", { false, true }), NeonActivationFunctionsDataset), framework::dataset::make("AlphaBeta", { 0.5f, 1.f }));
183 
184 template <typename T, ARM_COMPUTE_REQUIRES_TA(arm_compute::utils::traits::is_floating_point<T>::value)>
test_float_sqrt_boundary_value()185 void test_float_sqrt_boundary_value()
186 {
187     constexpr auto vector_size = uint32_t{ 16 };
188 
189     auto data_type = DataType::F32;
190 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
191     data_type = std::is_same<T, half>::value ? DataType::F16 : data_type;
192 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
193 
194     const auto boundary_value_vector = std::vector<T>
195     {
196         std::numeric_limits<T>::min(),
197         T(0),
198         std::numeric_limits<T>::epsilon(),
199         std::numeric_limits<T>::max(),
200     };
201 
202     // the following size ensures that the whole logic (vector + left-over) to be tested
203     // using all boundary values iff boundary_value_vecotr.size() is smaller than vector_size.
204     auto shape = TensorShape{ vector_size + boundary_value_vector.size() };
205     auto info  = ActivationLayerInfo{ ActivationLayerInfo::ActivationFunction::SQRT };
206     auto src   = create_tensor<Tensor>(shape, data_type);
207 
208     auto act = NEActivationLayer{};
209     act.configure(&src, nullptr, info);
210     src.allocator()->allocate();
211     library->fill_static_values(Accessor(src), boundary_value_vector);
212     act.run();
213 
214     auto reference_src = SimpleTensor<T> { shape, data_type };
215     library->fill_static_values(reference_src, boundary_value_vector);
216     auto reference_dst = reference::activation_layer<T>(reference_src, info);
217 
218     validate(Accessor(src), reference_dst, tolerance_float_sqrt);
219 }
220 } // namespace
221 
222 TEST_SUITE(NEON)
TEST_SUITE(ActivationLayer)223 TEST_SUITE(ActivationLayer)
224 
225 /** Test case for memory injection in @ref cpu::CpuWinogradConv2d.
226  *
227  * Configure the operator once and inject memory at run-time in multiple executions.
228  *
229  * Checks performed in order:
230  * - Both runs compute the same output
231  */
232 TEST_CASE(ActivationAPI, framework::DatasetMode::ALL)
233 {
234     acl::StatusCode err = acl::StatusCode::Success;
235 
236     // Create context & Queue
237     acl::Context ctx(acl::Target::Cpu, &err);
238     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
239 
240     acl::Queue queue(ctx, &err);
241     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
242 
243     // Create activation operator
244     acl::TensorDescriptor src_info({ 2, 3 }, acl::DataType::Float32);
245     acl::TensorDescriptor dst_info({ 2, 3 }, acl::DataType::Float32);
246     acl::ActivationDesc   desc{ AclRelu, 6.f, 0.f, false };
247 
248     acl::Activation act(ctx, src_info, dst_info, desc, &err);
249     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
250 
251     // Create tensors and feed
252     acl::Tensor src(ctx, src_info, &err);
253     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
254     acl::Tensor dst(ctx, dst_info, &err);
255     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
256 
257     acl::TensorPack pack(ctx);
258     err = pack.add(src, ACL_SRC);
259     err = pack.add(dst, ACL_DST);
260     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
261 
262     // Execute operator
263     err = act.run(queue, pack);
264     ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success);
265 }
266 
267 // *INDENT-OFF*
268 // clang-format off
269 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
270     framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching data types
271                                             TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
272                                             TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),     // Mismatching shapes
273                                           }),
274     framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16),
275                                             TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
276                                             TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
277                                           })),
278     framework::dataset::make("ActivationInfo", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
279                                                  ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
280                                                  ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU),
281                                                })),
282     framework::dataset::make("Expected", { false, true, false})),
283     input_info, output_info, act_info, expected)
284 {
285     bool is_valid = bool(NEActivationLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), act_info));
286     ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS);
287 }
288 
289 DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, concat(concat(
290                combine(framework::dataset::make("CpuExt", std::string("NEON")),
291                        framework::dataset::make("DataType", { DataType::F32,
292                                                               DataType::F16,
293                                                               DataType::QASYMM8,
294                                                               DataType::QASYMM8_SIGNED,
295                                                               DataType::QSYMM16
296                                                             })),
297                 combine(framework::dataset::make("CpuExt", std::string("SVE")),
298                         framework::dataset::make("DataType", { DataType::F32,
299                                                                DataType::F16,
300                                                              }))),
301                 combine(framework::dataset::make("CpuExt", std::string("SVE2")),
302                         framework::dataset::make("DataType", { DataType::QASYMM8,
303                                                                DataType::QASYMM8_SIGNED,
304                                                                DataType::QSYMM16
305                                                              }))),
306                cpu_ext, data_type)
307 {
308     using namespace cpu::kernels;
309 
310     cpuinfo::CpuIsaInfo cpu_isa{};
311     cpu_isa.neon = (cpu_ext == "NEON");
312     cpu_isa.sve  = (cpu_ext == "SVE");
313     cpu_isa.sve2 = (cpu_ext == "SVE2");
314     cpu_isa.fp16 = (data_type == DataType::F16);
315 
316     const auto *selected_impl = CpuActivationKernel::get_implementation(ActivationDataTypeISASelectorData{data_type, CPUModel::GENERIC, cpu_isa,ActivationLayerInfo::ActivationFunction::BOUNDED_RELU}, cpu::KernelSelectionType::Preferred);
317 
318     ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl);
319 
320     std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_activation";
321     std::string actual   = selected_impl->name;
322 
323     ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS);
324 }
325 // clang-format on
326 // *INDENT-ON*
327 
328 template <typename T>
329 using NEActivationLayerFixture = ActivationValidationFixture<Tensor, Accessor, NEActivationLayer, T>;
330 
331 TEST_SUITE(Float)
332 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
TEST_SUITE(FP16)333 TEST_SUITE(FP16)
334 TEST_CASE(SqrtBoundaryValue, framework::DatasetMode::ALL)
335 {
336     test_float_sqrt_boundary_value<half>();
337 }
338 FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ActivationDataset),
339                                                                                                       framework::dataset::make("DataType",
340                                                                                                               DataType::F16)))
341 {
342     // Validate output
343     validate(Accessor(_target), _reference, relative_tolerance(_data_type, _function), 0.f, absolute_tolerance(_data_type, _function));
344 }
345 TEST_SUITE_END() // FP16
346 #endif           /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
347 
TEST_SUITE(FP32)348 TEST_SUITE(FP32)
349 TEST_CASE(SqrtBoundaryValue, framework::DatasetMode::ALL)
350 {
351     test_float_sqrt_boundary_value<float>();
352 }
353 FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ActivationDataset), framework::dataset::make("DataType",
354                                                                                                        DataType::F32)))
355 
356 {
357     // Validate output
358     validate(Accessor(_target), _reference, relative_tolerance(_data_type, _function), 0.f, absolute_tolerance(_data_type, _function));
359 }
360 TEST_SUITE_END() // FP32
361 TEST_SUITE_END() // Float
362 
363 template <typename T>
364 using NEActivationLayerQuantizedFixture = ActivationValidationQuantizedFixture<Tensor, Accessor, NEActivationLayer, T>;
365 
366 /** Input data sets. */
367 const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationFunction",
368 {
369     ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU,
370     ActivationLayerInfo::ActivationFunction::RELU,
371     ActivationLayerInfo::ActivationFunction::BOUNDED_RELU,
372     ActivationLayerInfo::ActivationFunction::LOGISTIC,
373     ActivationLayerInfo::ActivationFunction::TANH,
374     ActivationLayerInfo::ActivationFunction::LEAKY_RELU,
375 });
376 
377 const auto QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false }),
378                                                         concat(QuantizedActivationFunctionsDataset, framework::dataset::make("ActivationFunction", ActivationLayerInfo::ActivationFunction::HARD_SWISH))),
379                                                 framework::dataset::make("AlphaBeta", { 0.5f, 1.f }));
380 
381 TEST_SUITE(Quantized)
TEST_SUITE(QASYMM8)382 TEST_SUITE(QASYMM8)
383 FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), QuantizedActivationDataset),
384                                                                                                                   framework::dataset::make("DataType",
385                                                                                                                           DataType::QASYMM8)),
386                                                                                                                   framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
387 {
388     // Validate output
389     validate(Accessor(_target), _reference, tolerance_qasymm8(_function));
390 }
391 TEST_SUITE_END() // QASYMM8
392 
TEST_SUITE(QASYMM8_SIGNED)393 TEST_SUITE(QASYMM8_SIGNED)
394 FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), QuantizedActivationDataset),
395                                                                                                                  framework::dataset::make("DataType",
396                                                                                                                          DataType::QASYMM8_SIGNED)),
397                                                                                                                  framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10.0f) })))
398 {
399     // Validate output
400     validate(Accessor(_target), _reference, tolerance_qasymm8(_function));
401 }
402 TEST_SUITE_END() // QASYMM8_SIGNED
403 
404 /** Input data sets. */
405 const auto Int16QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationFunction",
406 {
407     ActivationLayerInfo::ActivationFunction::LOGISTIC,
408     ActivationLayerInfo::ActivationFunction::TANH,
409     ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU,
410 });
411 const auto Int16QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false }), Int16QuantizedActivationFunctionsDataset),
412                                                      framework::dataset::make("AlphaBeta", { 0.5f, 1.f }));
413 
414 TEST_SUITE(QSYMM16)
415 FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerQuantizedFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), Int16QuantizedActivationDataset),
416                                                                                                                   framework::dataset::make("DataType",
417                                                                                                                           DataType::QSYMM16)),
418                                                                                                                   framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 32768.f, 0.f) })))
419 {
420     // Validate output
421     validate(Accessor(_target), _reference, tolerance_qsymm16);
422 }
423 TEST_SUITE_END() // QSYMM16
424 TEST_SUITE_END() // Quantized
425 
426 TEST_SUITE_END() // ActivationLayer
427 TEST_SUITE_END() // Neon
428 } // namespace validation
429 } // namespace test
430 } // namespace arm_compute
431