1 /*
2 * Copyright (c) 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/runtime/CL/CLTensor.h"
26 #include "arm_compute/runtime/CL/CLTensorAllocator.h"
27 #include "arm_compute/runtime/CL/functions/CLIndirectConvolutionLayer.h"
28 #include "tests/CL/CLAccessor.h"
29 #include "tests/datasets/ShapeDatasets.h"
30 #include "tests/framework/Macros.h"
31 #include "tests/validation/Validation.h"
32 #include "tests/validation/fixtures/DirectConvolutionLayerFixture.h"
33
34 // Note: Since the interface of indirect convolution is the same of direct convolution, we can reuse
35 // the direct convolution fixture
36
37 namespace arm_compute
38 {
39 namespace test
40 {
41 namespace validation
42 {
43 namespace
44 {
45 RelativeTolerance<half> tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */
46 RelativeTolerance<float> tolerance_fp32(0.05f); /**< Tolerance for floating point tests */
47 constexpr float abs_tolerance_f32(0.0001f); /**< Absolute tolerance for FP32 tests*/
48 constexpr float tolerance_num = 0.07f; /**< Tolerance number */
49
50 /** Activation function Dataset*/
51 const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo",
52 { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) });
53 } // namespace
54
55 TEST_SUITE(CL)
TEST_SUITE(IndirectConvolutionLayer)56 TEST_SUITE(IndirectConvolutionLayer)
57
58 /** Check whether the configuration of a indirect convolution layer with no
59 * bias leads to a successful run.
60 */
61 TEST_CASE(NoBias, framework::DatasetMode::PRECOMMIT)
62 {
63 const TensorShape src_shape_nhwc = TensorShape(8U, 27U, 13U);
64 const TensorShape wei_shape_nhwc = TensorShape(8U, 3U, 3U, 4U);
65 const TensorShape bia_shape = TensorShape(4U);
66 const TensorShape dst_shape_nhwc = TensorShape(4U, 25U, 11U);
67 constexpr DataType dt = DataType::F32;
68 constexpr DataLayout data_layout = DataLayout::NHWC;
69
70 auto src_nhwc = create_tensor<CLTensor>(src_shape_nhwc, dt, 1, QuantizationInfo(), data_layout);
71 auto wei_nhwc = create_tensor<CLTensor>(wei_shape_nhwc, dt, 1, QuantizationInfo(), data_layout);
72 auto dst_nhwc = create_tensor<CLTensor>(dst_shape_nhwc, dt, 1, QuantizationInfo(), data_layout);
73
74 TensorShape src_shape_nchw = src_shape_nhwc;
75 TensorShape wei_shape_nchw = wei_shape_nhwc;
76 TensorShape dst_shape_nchw = dst_shape_nhwc;
77
78 permute(src_shape_nchw, PermutationVector(1U, 2U, 0U));
79 permute(wei_shape_nchw, PermutationVector(1U, 2U, 0U, 3U));
80 permute(dst_shape_nchw, PermutationVector(1U, 2U, 0U));
81
82 const PadStrideInfo conv_info = PadStrideInfo(1, 1, 0, 0);
83
84 // Create indirect Convolution function
85 CLIndirectConvolutionLayer conv{};
86 conv.configure(&src_nhwc, &wei_nhwc, nullptr, &dst_nhwc, conv_info);
87
88 src_nhwc.allocator()->allocate();
89 wei_nhwc.allocator()->allocate();
90 dst_nhwc.allocator()->allocate();
91
92 library->fill_tensor_value(CLAccessor(src_nhwc), 1.f);
93 library->fill_tensor_value(CLAccessor(wei_nhwc), 1.f);
94
95 conv.run();
96
97 // Compute reference to compare
98 SimpleTensor<float> ref_src{ src_shape_nchw, dt };
99 SimpleTensor<float> ref_wei{ wei_shape_nchw, dt };
100 SimpleTensor<float> ref_bia{ bia_shape, dt };
101 library->fill_tensor_value(ref_src, 1.f);
102 library->fill_tensor_value(ref_wei, 1.f);
103 // No bias
104 library->fill_tensor_value(ref_bia, 0.f);
105 auto ref_dst = reference::convolution_layer<float>(ref_src, ref_wei, ref_bia, dst_shape_nchw, conv_info);
106
107 validate(CLAccessor(dst_nhwc), ref_dst);
108 }
109
110 /** Check whether the case of rectangle kernels i.e. when width and height of the weight_shape are not equal
111 * would lead to successful run
112 */
TEST_CASE(NonSquareKernel,framework::DatasetMode::PRECOMMIT)113 TEST_CASE(NonSquareKernel, framework::DatasetMode::PRECOMMIT)
114 {
115 const TensorShape src_shape_nhwc = TensorShape(3U, 33U, 27U);
116 const TensorShape wei_shape_nhwc = TensorShape(3U, 5U, 7U, 4U); // non-square kernel
117 const TensorShape bia_shape = TensorShape(4U);
118 const TensorShape dst_shape_nhwc = TensorShape(4U, 11U, 12U);
119 constexpr DataType dt = DataType::F32;
120 constexpr DataLayout data_layout = DataLayout::NHWC;
121
122 auto src_nhwc = create_tensor<CLTensor>(src_shape_nhwc, dt, 1, QuantizationInfo(), data_layout);
123 auto wei_nhwc = create_tensor<CLTensor>(wei_shape_nhwc, dt, 1, QuantizationInfo(), data_layout);
124 auto dst_nhwc = create_tensor<CLTensor>(dst_shape_nhwc, dt, 1, QuantizationInfo(), data_layout);
125
126 TensorShape src_shape_nchw = src_shape_nhwc;
127 TensorShape wei_shape_nchw = wei_shape_nhwc;
128 TensorShape dst_shape_nchw = dst_shape_nhwc;
129
130 permute(src_shape_nchw, PermutationVector(1U, 2U, 0U));
131 permute(wei_shape_nchw, PermutationVector(1U, 2U, 0U, 3U));
132 permute(dst_shape_nchw, PermutationVector(1U, 2U, 0U));
133
134 const PadStrideInfo conv_info = PadStrideInfo(3, 2, 1, 1, 2, 0, DimensionRoundingType::FLOOR);
135
136 // Create indirect convolution function
137 CLIndirectConvolutionLayer conv{};
138 conv.configure(&src_nhwc, &wei_nhwc, nullptr, &dst_nhwc, conv_info);
139
140 src_nhwc.allocator()->allocate();
141 wei_nhwc.allocator()->allocate();
142 dst_nhwc.allocator()->allocate();
143
144 library->fill_tensor_value(CLAccessor(src_nhwc), 1.f);
145 library->fill_tensor_value(CLAccessor(wei_nhwc), 1.f);
146
147 conv.run();
148
149 // Compute reference to compare
150 SimpleTensor<float> ref_src{ src_shape_nchw, dt };
151 SimpleTensor<float> ref_wei{ wei_shape_nchw, dt };
152 SimpleTensor<float> ref_bia{ bia_shape, dt };
153 library->fill_tensor_value(ref_src, 1.f);
154 library->fill_tensor_value(ref_wei, 1.f);
155 // No bias
156 library->fill_tensor_value(ref_bia, 0.f);
157 auto ref_dst = reference::convolution_layer<float>(ref_src, ref_wei, ref_bia, dst_shape_nchw, conv_info);
158
159 validate(CLAccessor(dst_nhwc), ref_dst);
160 }
161 // *INDENT-OFF*
162 // clang-format off
163 // Note: Since the interface of indirect convolution is the same of direct convolution, we can reuse
164 // the direct convolution fixture
165 template <typename T>
166 using CLIndirectConvolutionLayerFixture = DirectConvolutionValidationFixture<CLTensor, CLAccessor, CLIndirectConvolutionLayer, T>;
167 template <typename T>
168 using CLIndirectConvolutionLayerMixedDataLayoutFixture = DirectConvolutionValidationFixture<CLTensor, CLAccessor, CLIndirectConvolutionLayer, T, true>;
169
170 TEST_SUITE(NHWC)
TEST_SUITE(FP16)171 TEST_SUITE(FP16)
172 FIXTURE_DATA_TEST_CASE(RunSmall, CLIndirectConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT,
173 combine(combine(combine(zip(zip(zip(zip(zip(zip(
174 framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U),
175 TensorShape(19U, 5U, 16U, 4U),
176 TensorShape(13U, 5U, 17U, 2U),
177 TensorShape(32U, 37U, 13U) } ),
178 framework::dataset::make("StrideX", { 1, 3, 1, 1 })),
179 framework::dataset::make("StrideY", { 1, 3, 2, 1 })),
180 framework::dataset::make("PadX", { 1, 3, 0, 4 })),
181 framework::dataset::make("PadY", { 1, 3, 0, 4 })),
182 framework::dataset::make("KernelSize", { 3, 8, 1, 9 })),
183 framework::dataset::make("NumKernels", { 17, 3, 1, 19 })),
184 framework::dataset::make("DataType", DataType::F16)),
185 framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )),
186 framework::dataset::make("DataLayout", DataLayout::NHWC)))
187 {
188 validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num);
189 }
190
191 FIXTURE_DATA_TEST_CASE(RunLarge, CLIndirectConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY,
192 combine(combine(combine(zip(zip(zip(zip(zip(zip(
193 framework::dataset::make("InputShape", { TensorShape(800U, 800U, 3U) } ),
194 framework::dataset::make("StrideX", { 1 })),
195 framework::dataset::make("StrideY", { 1 })),
196 framework::dataset::make("PadX", { 1 })),
197 framework::dataset::make("PadY", { 1 })),
198 framework::dataset::make("KernelSize", { 9 })),
199 framework::dataset::make("NumKernels", { 3 })),
200 framework::dataset::make("DataType", DataType::F16)),
201 framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::IDENTITY) )),
202 framework::dataset::make("DataLayout", DataLayout::NHWC)))
203 {
204 validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num);
205 }
206
207 TEST_SUITE_END() // FP16
208
TEST_SUITE(FP32)209 TEST_SUITE(FP32)
210 FIXTURE_DATA_TEST_CASE(RunSmall, CLIndirectConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT,
211 combine(combine(combine(zip(zip(zip(zip(zip(zip(
212 framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U),
213 TensorShape(19U, 5U, 16U, 4U),
214 TensorShape(13U, 5U, 17U, 2U),
215 TensorShape(32U, 37U, 13U) } ),
216 framework::dataset::make("StrideX", { 1, 3, 1, 1 })),
217 framework::dataset::make("StrideY", { 1, 3, 2, 1 })),
218 framework::dataset::make("PadX", { 1, 3, 0, 4 })),
219 framework::dataset::make("PadY", { 1, 3, 0, 4 })),
220 framework::dataset::make("KernelSize", { 3, 8, 1, 9 })),
221 framework::dataset::make("NumKernels", { 17, 3, 1, 19 })),
222 framework::dataset::make("DataType", DataType::F32)),
223 framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )),
224 framework::dataset::make("DataLayout", DataLayout::NHWC)))
225 {
226 validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32);
227 }
228 FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLIndirectConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT,
229 combine(combine(combine(zip(zip(zip(zip(zip(zip(
230 framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U),
231 TensorShape(19U, 5U, 16U, 4U),
232 TensorShape(13U, 5U, 17U, 2U),
233 TensorShape(32U, 37U, 13U) } ),
234 framework::dataset::make("StrideX", { 1 })),
235 framework::dataset::make("StrideY", { 2 })),
236 framework::dataset::make("PadX", { 1 })),
237 framework::dataset::make("PadY", { 3 })),
238 framework::dataset::make("KernelSize", { 3 })),
239 framework::dataset::make("NumKernels", { 3 })),
240 framework::dataset::make("DataType", DataType::F32)),
241 framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )),
242 framework::dataset::make("DataLayout", DataLayout::NHWC)))
243 {
244 validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32);
245 }
246 FIXTURE_DATA_TEST_CASE(RunLarge, CLIndirectConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY,
247 combine(combine(combine(zip(zip(zip(zip(zip(zip(
248 framework::dataset::make("InputShape", { TensorShape(800U, 800U, 3U) } ),
249 framework::dataset::make("StrideX", { 1 })),
250 framework::dataset::make("StrideY", { 1 })),
251 framework::dataset::make("PadX", { 1 })),
252 framework::dataset::make("PadY", { 1 })),
253 framework::dataset::make("KernelSize", { 9 })),
254 framework::dataset::make("NumKernels", { 3 })),
255 framework::dataset::make("DataType", DataType::F32)),
256 framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::IDENTITY) )),
257 framework::dataset::make("DataLayout", DataLayout::NHWC)))
258 {
259 validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32);
260 }
261 TEST_SUITE_END() // FP32
262 TEST_SUITE_END() // NHWC
263 TEST_SUITE_END() // IndirectConvolutionLayer
264 TEST_SUITE_END() // CL
265
266 } // namespace validation
267 } // namespace test
268 } // namespace arm_compute
269