1 /*
2 * Copyright (c) 2019-2020 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/graph.h"
25
26 #include "tests/NEON/Accessor.h"
27 #include "tests/validation/Validation.h"
28 #include "tests/validation/reference/ConvolutionLayer.h"
29 #include "tests/validation/reference/Permute.h"
30
31 #include "utils/CommonGraphOptions.h"
32 #include "utils/GraphUtils.h"
33 #include "utils/Utils.h"
34
35 #include "ValidateExample.h"
36 #include "graph_validate_utils.h"
37
38 #include <utility>
39
40 using namespace arm_compute::utils;
41 using namespace arm_compute::graph::frontend;
42 using namespace arm_compute::graph_utils;
43 using namespace arm_compute::graph;
44 using namespace arm_compute;
45 using namespace arm_compute::test;
46 using namespace arm_compute::test::validation;
47
48 namespace
49 {
50 /** Convolution command line options used to configure the graph examples
51 *
52 * (Similar to common options)
53 * The options in this object get populated when "parse()" is called on the parser used to construct it.
54 * The expected workflow is:
55 *
56 * CommandLineParser parser;
57 * CommonOptions options( parser );
58 * parser.parse(argc, argv);
59 */
60 class ConvolutionOptions final : public CommonGraphValidateOptions
61 {
62 public:
ConvolutionOptions(CommandLineParser & parser)63 explicit ConvolutionOptions(CommandLineParser &parser) noexcept
64 : CommonGraphValidateOptions(parser),
65 width(parser.add_option<SimpleOption<int>>("width", 9)),
66 height(parser.add_option<SimpleOption<int>>("height", 9)),
67 channels(parser.add_option<SimpleOption<int>>("channels", 1)),
68 batch(parser.add_option<SimpleOption<int>>("batch", 1)),
69 weights_width(parser.add_option<SimpleOption<int>>("weights_width", 3)),
70 weights_height(parser.add_option<SimpleOption<int>>("weights_height", 3)),
71 OFM(parser.add_option<SimpleOption<int>>("OFM", 1)),
72 padding_top(parser.add_option<SimpleOption<int>>("padding_top", 0)),
73 padding_left(parser.add_option<SimpleOption<int>>("padding_left", 0)),
74 padding_bottom(parser.add_option<SimpleOption<int>>("padding_bottom", 0)),
75 padding_right(parser.add_option<SimpleOption<int>>("padding_right", 0)),
76 stride_x(parser.add_option<SimpleOption<int>>("stride_x", 1)),
77 stride_y(parser.add_option<SimpleOption<int>>("stride_y", 1)),
78 padding_mode(),
79 conv_mode(),
80 data_layout(),
81 scale(parser.add_option<SimpleOption<float>>("scale", 1.0f)),
82 offset(parser.add_option<SimpleOption<int>>("offset", 0)),
83 weights_scale(parser.add_option<SimpleOption<float>>("weights_scale", 1.0f)),
84 weights_offset(parser.add_option<SimpleOption<int>>("weights_offset", 0)),
85 output_scale(parser.add_option<SimpleOption<float>>("output_scale", 1.0f)),
86 output_offset(parser.add_option<SimpleOption<int>>("output_offset", 0)),
87 input_range_low(parser.add_option<SimpleOption<uint64_t>>("input_range_low")),
88 input_range_high(parser.add_option<SimpleOption<uint64_t>>("input_range_high")),
89 weights_range_low(parser.add_option<SimpleOption<uint64_t>>("weights_range_low")),
90 weights_range_high(parser.add_option<SimpleOption<uint64_t>>("weights_range_high")),
91 input_npy(parser.add_option<SimpleOption<std::string>>("input_image")),
92 output_npy(parser.add_option<SimpleOption<std::string>>("reference_image")),
93 weights_npy(parser.add_option<SimpleOption<std::string>>("weights_npy")),
94 bias_npy(parser.add_option<SimpleOption<std::string>>("bias_image"))
95 {
96 const std::set<ConvolutionPaddingMode> available_padding_modes
97 {
98 ConvolutionPaddingMode::Valid,
99 ConvolutionPaddingMode::Same
100 };
101
102 const std::set<arm_compute::graph::ConvolutionMethod> supported_convolution_methods
103 {
104 arm_compute::graph::ConvolutionMethod::Default,
105 arm_compute::graph::ConvolutionMethod::GEMM,
106 arm_compute::graph::ConvolutionMethod::Winograd,
107 arm_compute::graph::ConvolutionMethod::Direct
108 };
109
110 const std::set<DataLayout> supported_data_layouts
111 {
112 DataLayout::NHWC,
113 DataLayout::NCHW,
114 };
115
116 padding_mode = parser.add_option<EnumOption<ConvolutionPaddingMode>>("padding_mode", available_padding_modes, ConvolutionPaddingMode::Valid);
117 conv_mode = parser.add_option<EnumOption<arm_compute::graph::ConvolutionMethod>>("convolution_method", supported_convolution_methods, arm_compute::graph::ConvolutionMethod::Default);
118 data_layout = parser.add_option<EnumOption<DataLayout>>("layout", supported_data_layouts, DataLayout::NHWC);
119
120 padding_mode->set_help("Set padding mode");
121 help->set_help("Show this help message");
122 width->set_help("Set Input dimension width");
123 height->set_help("Set Input dimension height");
124 channels->set_help("Set Input dimension channels");
125 batch->set_help("Set Input dimension batch");
126 weights_width->set_help("Set weights_dimensions width");
127 weights_height->set_help("Set weights_dimensions height");
128 OFM->set_help("Set OFM");
129 padding_top->set_help("Set padding top");
130 padding_bottom->set_help("Set padding bottom");
131 padding_left->set_help("Set padding left");
132 padding_right->set_help("Set padding right");
133 stride_x->set_help("Set padding stride x");
134 stride_y->set_help("Set padding stride y");
135 conv_mode->set_help("Set convolution method");
136 scale->set_help("Quantization scale from QASYMM8");
137 offset->set_help("Quantization offset from QASYMM8");
138 weights_scale->set_help("Quantization scale from QASYMM8");
139 weights_offset->set_help("Quantization offset from QASYMM8");
140 output_scale->set_help("Quantization scale from QASYMM8");
141 output_offset->set_help("Quantization offset from QASYMM8");
142 input_npy->set_help("Use input .npy instead");
143 output_npy->set_help("Use .npy as a reference");
144 input_range_low->set_help("Lower bound for input randomization range");
145 input_range_high->set_help("Lower bound for input randomization range");
146 weights_range_low->set_help("Lower bound for input randomization range");
147 weights_range_high->set_help("Lower bound for input randomization range");
148 }
149
150 /** Fill out the supplied parameters with user supplied parameters
151 *
152 * @param[out] os Output stream.
153 * @param[in] common_params Example parameters to output
154 *
155 * @return None.
156 */
consume_parameters(ExampleParams & common_params)157 void consume_parameters(ExampleParams &common_params)
158 {
159 common_params.input.width = width->value();
160 common_params.input.height = height->value();
161 common_params.input.fm = channels->value();
162 common_params.input.batch = batch->value();
163 common_params.input.quant_info = QuantizationInfo(scale->value(), offset->value());
164 common_params.input.npy = input_npy->value();
165 common_params.input.range_low = input_range_low->value();
166 common_params.input.range_high = input_range_high->value();
167
168 common_params.weights.width = weights_width->value();
169 common_params.weights.height = weights_height->value();
170 common_params.weights.fm = OFM->value();
171 common_params.weights.npy = weights_npy->value();
172 common_params.weights.quant_info = QuantizationInfo(weights_scale->value(), weights_offset->value());
173 common_params.weights.range_low = weights_range_low->value();
174 common_params.weights.range_high = weights_range_high->value();
175
176 common_params.bias.npy = bias_npy->value();
177
178 common_params.output.quant_info = QuantizationInfo(output_scale->value(), output_offset->value());
179 common_params.output.npy = output_npy->value();
180
181 common_params.convolution.padding_mode = padding_mode->value();
182 common_params.convolution.padding_top = padding_top->value();
183 common_params.convolution.padding_bottom = padding_bottom->value();
184 common_params.convolution.padding_left = padding_left->value();
185 common_params.convolution.padding_right = padding_right->value();
186 common_params.convolution.padding_stride_x = stride_x->value();
187 common_params.convolution.padding_stride_y = stride_y->value();
188
189 common_params.data_type = data_type->value();
190 common_params.data_layout = data_layout->value();
191 common_params.convolution_method = conv_mode->value();
192 }
193
print_parameters(::std::ostream & os,const ExampleParams & common_params)194 void print_parameters(::std::ostream &os, const ExampleParams &common_params) override
195 {
196 os << "Threads : " << common_params.common_params.threads << std::endl;
197 os << "Target : " << common_params.common_params.target << std::endl;
198 os << "Data type : " << common_params.data_type << std::endl;
199 os << "Input dimensions(X,Y, Channels, Batch) : (" << common_params.input.width << "," << common_params.input.height << "," << common_params.input.fm << "," << common_params.input.batch << ")"
200 << std::endl;
201 os << "Weight dimensions(X,Y, Channels(same as input), OFM) : (" << common_params.weights.width << "," << common_params.weights.height << "," << common_params.input.fm << "," <<
202 common_params.weights.fm << ")" << std::endl;
203 os << "Padding(top, bottom, left, right) (stride x, stride y) : (" << common_params.convolution.padding_top << "," << common_params.convolution.padding_bottom << "," <<
204 common_params.convolution.padding_left << "," << common_params.convolution.padding_right << ") (" << common_params.convolution.padding_stride_x << "," << common_params.convolution.padding_stride_y <<
205 ")" << std::endl;
206 os << "Padding Mode: " << common_params.convolution.padding_mode << std::endl;
207 os << "Convolution Method: " << common_params.convolution_method << std::endl;
208 }
209
210 /** Prevent instances of this class from being copied (As this class contains pointers) */
211 ConvolutionOptions(const ConvolutionOptions &) = delete;
212 /** Prevent instances of this class from being copied (As this class contains pointers) */
213 ConvolutionOptions &operator=(const ConvolutionOptions &) = delete;
214 /** Allow instances of this class to be moved */
215 ConvolutionOptions(ConvolutionOptions &&) noexcept(true) = default;
216 /** Allow instances of this class to be moved */
217 ConvolutionOptions &operator=(ConvolutionOptions &&) noexcept(true) = default;
218 /** Default destructor */
219 ~ConvolutionOptions() override = default;
220
221 private:
222 SimpleOption<int> *width; /**< Input width */
223 SimpleOption<int> *height; /**< Input height */
224 SimpleOption<int> *channels; /**< Input channels */
225 SimpleOption<int> *batch; /**< Input batch */
226 SimpleOption<int> *weights_width; /**< weights width */
227 SimpleOption<int> *weights_height; /**< weights height */
228 SimpleOption<int> *OFM; /**< Output Feature Map */
229 SimpleOption<int> *padding_top; /**< Padding top */
230 SimpleOption<int> *padding_left; /**< Padding left */
231 SimpleOption<int> *padding_bottom; /**< Padding bottom */
232 SimpleOption<int> *padding_right; /**< Padding right */
233 SimpleOption<int> *stride_x; /**< Padding stride x */
234 SimpleOption<int> *stride_y; /**< Padding stride y */
235 EnumOption<ConvolutionPaddingMode> *padding_mode; /**< Padding mode */
236 EnumOption<arm_compute::graph::ConvolutionMethod> *conv_mode; /**< Convolution method */
237 EnumOption<arm_compute::DataLayout> *data_layout; /**< Graph data layout */
238 SimpleOption<float> *scale; /**< Input Quantization scale from QASYMM8 */
239 SimpleOption<int> *offset; /**< Input Quantization offset from QASYMM8 */
240 SimpleOption<float> *weights_scale; /**< Weights Quantization scale from QASYMM8 */
241 SimpleOption<int> *weights_offset; /**< Weights Quantization offset from QASYMM8 */
242 SimpleOption<float> *output_scale; /**< Output Quantization scale from QASYMM8 */
243 SimpleOption<int> *output_offset; /**< Output Quantization offset from QASYMM8 */
244 SimpleOption<uint64_t> *input_range_low; /**< Lower bound for input randomization range */
245 SimpleOption<uint64_t> *input_range_high; /**< Upper bound for input randomization range */
246 SimpleOption<uint64_t> *weights_range_low; /**< Lower bound for weights randomization range */
247 SimpleOption<uint64_t> *weights_range_high; /**< Upper bound for weights randomization range */
248
249 SimpleOption<std::string> *input_npy; /**< Use input .npy image */
250 SimpleOption<std::string> *output_npy; /**< Use output .npy image to verify*/
251 SimpleOption<std::string> *weights_npy; /**< Use weights .npy image */
252 SimpleOption<std::string> *bias_npy; /**< Use bias .npy image */
253 };
254
255 /** ConvolutionLayer Graph example validation accessor class */
256 template <typename D>
257 class ConvolutionVerifyAccessor final : public VerifyAccessor<D>
258 {
259 using BaseClassType = VerifyAccessor<D>;
260 using BaseClassType::BaseClassType;
261 using BaseClassType::_params;
262 using TBias = typename std::conditional<std::is_same<typename std::decay<D>::type, uint8_t>::value, int32_t, D>::type;
263
reference(SimpleTensor<D> & src,SimpleTensor<D> & weights,SimpleTensor<TBias> & bias,const TensorShape & output_shape)264 SimpleTensor<D> reference(SimpleTensor<D> &src, SimpleTensor<D> &weights, SimpleTensor<TBias> &bias, const TensorShape &output_shape) override
265 {
266 // Calculate padding information
267 const PadStrideInfo padding_info = calculate_convolution_padding(_params);
268
269 //Calculate reference
270 return reference::convolution_layer<D>(src, weights, bias, output_shape, padding_info, Size2D(1, 1),
271 1, _params.output.quant_info);
272 }
273
relative_tolerance()274 float relative_tolerance() override
275 {
276 const std::map<arm_compute::graph::Target, const std::map<DataType, float>> relative_tolerance
277 {
278 {
279 arm_compute::graph::Target::CL,
280 { { DataType::F16, 0.2f },
281 { DataType::F32, 0.5f },
282 { DataType::QASYMM8, 1.0f }
283 }
284 },
285 {
286 arm_compute::graph::Target::NEON,
287 { { DataType::F16, 0.2f },
288 { DataType::F32, 0.01f },
289 { DataType::QASYMM8, 0.0f }
290 }
291 }
292 };
293
294 if(_params.convolution_method == arm_compute::graph::ConvolutionMethod::Winograd
295 && _params.data_type == DataType::F32
296 && _params.common_params.target == arm_compute::graph::Target::NEON)
297 {
298 return 0.05f;
299 }
300 else
301 {
302 return relative_tolerance.at(_params.common_params.target).at(_params.data_type);
303 }
304 }
305
absolute_tolerance()306 float absolute_tolerance() override
307 {
308 const std::map<Target, const std::map<DataType, float>> absolute_tolerance
309 {
310 {
311 Target::CL,
312 { { DataType::F16, 0.0f },
313 { DataType::F32, 0.0001f },
314 { DataType::QASYMM8, 0.0f }
315 }
316 },
317 {
318 Target::NEON,
319 { { DataType::F16, 0.2f },
320 { DataType::F32, 0.002f },
321 { DataType::QASYMM8, 0.0f }
322 }
323 }
324 };
325
326 return absolute_tolerance.at(_params.common_params.target).at(_params.data_type);
327 }
328
tolerance_number()329 float tolerance_number() override
330 {
331 const std::map<Target, const std::map<DataType, float>> absolute_tolerance
332 {
333 {
334 Target::CL,
335 { { DataType::F16, 0.07f },
336 { DataType::F32, 0.07f },
337 { DataType::QASYMM8, 0.0f }
338 }
339 },
340 {
341 Target::NEON,
342 { { DataType::F16, 0.07f },
343 { DataType::F32, 0.0f },
344 { DataType::QASYMM8, 0.0f }
345 }
346 }
347 };
348
349 return absolute_tolerance.at(_params.common_params.target).at(_params.data_type);
350 }
351 };
352
353 } // namespace
354
355 class GraphConvolutionValidateExample final : public GraphValidateExample<ConvolutionLayer, ConvolutionOptions, ConvolutionVerifyAccessor>
356 {
357 using GraphValidateExample::graph;
358
359 public:
GraphConvolutionValidateExample()360 GraphConvolutionValidateExample()
361 : GraphValidateExample("Convolution Graph example")
362 {
363 }
364
GraphFunctionLayer(ExampleParams & params)365 ConvolutionLayer GraphFunctionLayer(ExampleParams ¶ms) override
366 {
367 const PixelValue lower = PixelValue(params.input.range_low, params.data_type, params.input.quant_info);
368 const PixelValue upper = PixelValue(params.input.range_high, params.data_type, params.input.quant_info);
369
370 const PixelValue weights_lower = PixelValue(params.weights.range_low, params.data_type, params.weights.quant_info);
371 const PixelValue weights_upper = PixelValue(params.weights.range_high, params.data_type, params.weights.quant_info);
372
373 // Calculate padding information
374 const PadStrideInfo padding_info = calculate_convolution_padding(params);
375
376 return ConvolutionLayer(params.weights.width, params.weights.height, params.weights.fm,
377 get_accessor(params.weights, weights_lower, weights_upper, 1),
378 get_accessor(params.bias, lower, upper, 2),
379 padding_info, 1, params.weights.quant_info, params.output.quant_info);
380 }
381 };
382
383 /** Main program for Graph Convolution test
384 *
385 * @param[in] argc Number of arguments
386 * @param[in] argv Arguments ( Input dimensions [width, height, channels, batch]
387 * Weights dimensions [width, height, OFM]
388 * Padding [top,bottom,left,right, Stride x, Stride y, mode [Valid / Same / Manual] )
389 * Convolution Method[ Auto/GEMM/Winograd/Direct]
390 * Verification[tolerance_number,absolute_tolerance,relative_tolerance] )
391 *
392 */
main(int argc,char ** argv)393 int main(int argc, char **argv)
394 {
395 return arm_compute::utils::run_example<GraphConvolutionValidateExample>(argc, argv);
396 }
397