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/runtime/CL/functions/CLLSTMLayerQuantized.h"
25 #include "tests/CL/CLAccessor.h"
26 #include "tests/PaddingCalculator.h"
27 #include "tests/Utils.h"
28 #include "tests/datasets/LSTMLayerDataset.h"
29 #include "tests/framework/Asserts.h"
30 #include "tests/framework/Macros.h"
31 #include "tests/framework/datasets/Datasets.h"
32 #include "tests/validation/Validation.h"
33
34 #include <vector>
35
36 namespace arm_compute
37 {
38 namespace test
39 {
40 namespace validation
41 {
42 namespace
43 {
44 template <typename T>
fill_tensor(CLTensor & tensor,const std::vector<T> & v)45 inline void fill_tensor(CLTensor &tensor, const std::vector<T> &v)
46 {
47 tensor.map(true);
48 // Import memory accounting for padding
49 TensorShape t_shape = tensor.info()->tensor_shape();
50 Window window;
51 window.use_tensor_dimensions(t_shape);
52 Iterator out(&tensor, window);
53 execute_window_loop(window, [&](const Coordinates & id)
54 {
55 *reinterpret_cast<T *>(out.ptr()) = v[coord2index(t_shape, id)];
56 },
57 out);
58 tensor.unmap();
59 }
60
61 template <typename T>
fill_tensor(SimpleTensor<T> & tensor,const std::vector<T> & v)62 inline void fill_tensor(SimpleTensor<T> &tensor, const std::vector<T> &v)
63 {
64 std::memcpy(tensor.data(), v.data(), sizeof(T) * v.size());
65 }
66
67 } // namespace
68
69 TEST_SUITE(CL)
TEST_SUITE(LSTMLayerQuantized)70 TEST_SUITE(LSTMLayerQuantized)
71
72 // *INDENT-OFF*
73 // clang-format off
74 TEST_SUITE(IntegrationTestCase)
75 TEST_SUITE(MultSmallerEq1)
76 TEST_CASE(RunSmall, framework::DatasetMode::PRECOMMIT)
77 {
78 const int batch_size = 2;
79 const int input_size = 2;
80 const int output_size = 4;
81
82 QuantizationInfo qasymm(1.f / 128.f, 128);
83 QuantizationInfo qweights(1.f / 128.f, 128);
84 QuantizationInfo qsymm_3(8.f / 32768.f, 0);
85 QuantizationInfo qsymm_4(16.f / 32768.f, 0);
86
87 TensorShape input_shape{ input_size, batch_size };
88 TensorShape input_weights_shape{ input_size, output_size };
89 TensorShape recurrent_weights_shape{ output_size, output_size };
90 TensorShape output_shape{ output_size, batch_size};
91 TensorShape bias_shape{ output_size };
92
93 auto input_to_input_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
94 auto input_to_forget_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
95 auto input_to_cell_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
96 auto input_to_output_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
97 auto recurrent_to_input_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
98 auto recurrent_to_forget_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
99 auto recurrent_to_cell_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
100 auto recurrent_to_output_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
101 auto input_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
102 auto forget_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
103 auto cell_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
104 auto output_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
105
106 // LSTM input
107 auto input = create_tensor<CLTensor>(input_shape, DataType::QASYMM8, 1, qasymm);
108
109 // LSTM output state
110 auto output_state = create_tensor<CLTensor>(output_shape, DataType::QASYMM8, 1, qasymm);
111
112 // LSTM cell state
113 auto cell_state = create_tensor<CLTensor>(output_shape, DataType::QSYMM16, 1, qsymm_4);
114
115 CLLSTMLayerQuantized lstmq;
116
117 lstmq.configure(&input, &input_to_input_weights, &input_to_forget_weights, &input_to_cell_weights, &input_to_output_weights,
118 &recurrent_to_input_weights, &recurrent_to_forget_weights, &recurrent_to_cell_weights, &recurrent_to_output_weights,
119 &input_gate_bias, &forget_gate_bias, &cell_gate_bias, &output_gate_bias, &cell_state, &output_state, &cell_state, &output_state);
120
121 input.allocator()->allocate();
122 input_to_input_weights.allocator()->allocate();
123 input_to_forget_weights.allocator()->allocate();
124 input_to_cell_weights.allocator()->allocate();
125 input_to_output_weights.allocator()->allocate();
126 recurrent_to_input_weights.allocator()->allocate();
127 recurrent_to_forget_weights.allocator()->allocate();
128 recurrent_to_cell_weights.allocator()->allocate();
129 recurrent_to_output_weights.allocator()->allocate();
130 input_gate_bias.allocator()->allocate();
131 forget_gate_bias.allocator()->allocate();
132 cell_gate_bias.allocator()->allocate();
133 output_gate_bias.allocator()->allocate();
134 cell_state.allocator()->allocate();
135 output_state.allocator()->allocate();
136
137 // Fill weights and biases
138 fill_tensor(input_to_input_weights, std::vector<uint8_t>{ 47, 168,
139 66, 239,
140 6, 42,
141 237, 236 });
142
143 fill_tensor(input_to_forget_weights, std::vector<uint8_t> { 204, 193,
144 148, 59,
145 113, 17,
146 66, 197 });
147
148 fill_tensor(input_to_cell_weights, std::vector<uint8_t> { 172, 101,
149 184, 209,
150 165, 82,
151 108, 209 });
152
153 fill_tensor(input_to_output_weights, std::vector<uint8_t> { 203, 244,
154 219, 114,
155 130, 16,
156 163, 222 });
157
158 fill_tensor(recurrent_to_input_weights, std::vector<uint8_t> { 162, 168, 7, 95,
159 91, 155, 108, 216,
160 255, 100, 48, 188,
161 58, 37, 186, 147 });
162
163 fill_tensor(recurrent_to_forget_weights, std::vector<uint8_t> { 46, 58, 47, 170,
164 246, 96, 12, 99,
165 68, 23, 186, 161,
166 237, 164, 89, 6 });
167
168 fill_tensor(recurrent_to_cell_weights, std::vector<uint8_t> { 234, 99, 71, 206,
169 205, 159, 64, 253,
170 191, 148, 116, 8,
171 209, 136, 59, 138 });
172
173 fill_tensor(recurrent_to_output_weights, std::vector<uint8_t> { 23, 241, 137, 36,
174 206, 5, 227, 56,
175 254, 176, 231, 47,
176 18, 201, 161, 11 });
177
178 fill_tensor(input_gate_bias, std::vector<int> {-103038, 30525, 115255, -38154 });
179 fill_tensor(forget_gate_bias, std::vector<int> { -23428, 126970, 116806, 46307 });
180 fill_tensor(cell_gate_bias, std::vector<int> { 128006, 69949, -42808, 42568 });
181 fill_tensor(output_gate_bias, std::vector<int> { -67066, -53607, 47233, 7300 });
182
183 SimpleTensor<uint8_t> expected_output(output_shape, DataType::QASYMM8, 1, qasymm);
184
185 // Initialize state
186 fill_tensor(output_state, std::vector<uint8_t> { 128, 128, 128, 128,
187 128, 128, 128, 128 });
188 fill_tensor(cell_state, std::vector<int16_t> { 0, 0, 0, 0,
189 0, 0, 0, 0 });
190
191 // First input
192 fill_tensor(input, std::vector<uint8_t> { 106, 193,
193 155, 150 });
194
195 fill_tensor(expected_output, std::vector<uint8_t> { 128, 130, 36, 134,
196 128, 131, 35, 133 });
197
198 lstmq.run();
199 validate(CLAccessor(output_state), expected_output);
200
201 // Second input
202 fill_tensor(expected_output, std::vector<uint8_t> { 128, 129, 12, 137,
203 128, 131, 10, 136 });
204 lstmq.run();
205 validate(CLAccessor(output_state), expected_output);
206
207 // Third input
208 fill_tensor(expected_output, std::vector<uint8_t> { 128, 129, 8, 140,
209 128, 130, 6, 138 });
210 lstmq.run();
211 validate(CLAccessor(output_state), expected_output);
212 }
213
TEST_CASE(RunLarge,framework::DatasetMode::PRECOMMIT)214 TEST_CASE(RunLarge, framework::DatasetMode::PRECOMMIT)
215 {
216 const int batch_size = 16;
217 const int input_size = 8;
218 const int output_size = 8;
219
220
221 QuantizationInfo qasymm(1.f / 128.f, 128);
222 QuantizationInfo qweights(1.f / 128.f, 128);
223 QuantizationInfo qsymm_3(8.f / 32768.f, 0);
224 QuantizationInfo qsymm_4(16.f / 32768.f, 0);
225
226 TensorShape input_shape{ input_size, batch_size };
227 TensorShape input_weights_shape{ input_size, output_size };
228 TensorShape recurrent_weights_shape{ output_size, output_size };
229 TensorShape output_shape{ output_size, batch_size};
230 TensorShape bias_shape{ output_size };
231
232 auto input_to_input_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
233 auto input_to_forget_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
234 auto input_to_cell_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
235 auto input_to_output_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
236 auto recurrent_to_input_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
237 auto recurrent_to_forget_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
238 auto recurrent_to_cell_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
239 auto recurrent_to_output_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
240 auto input_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
241 auto forget_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
242 auto cell_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
243 auto output_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
244
245 // LSTM input
246 auto input = create_tensor<CLTensor>(input_shape, DataType::QASYMM8, 1, qasymm);
247
248 // LSTM output state
249 auto output_state = create_tensor<CLTensor>(output_shape, DataType::QASYMM8, 1, qasymm);
250
251 // LSTM cell state
252 auto cell_state = create_tensor<CLTensor>(output_shape, DataType::QSYMM16, 1, qsymm_4);
253
254 CLLSTMLayerQuantized lstmq;
255
256 lstmq.configure(&input, &input_to_input_weights, &input_to_forget_weights, &input_to_cell_weights, &input_to_output_weights,
257 &recurrent_to_input_weights, &recurrent_to_forget_weights, &recurrent_to_cell_weights, &recurrent_to_output_weights,
258 &input_gate_bias, &forget_gate_bias, &cell_gate_bias, &output_gate_bias, &cell_state, &output_state, &cell_state, &output_state);
259
260 input.allocator()->allocate();
261 input_to_input_weights.allocator()->allocate();
262 input_to_forget_weights.allocator()->allocate();
263 input_to_cell_weights.allocator()->allocate();
264 input_to_output_weights.allocator()->allocate();
265 recurrent_to_input_weights.allocator()->allocate();
266 recurrent_to_forget_weights.allocator()->allocate();
267 recurrent_to_cell_weights.allocator()->allocate();
268 recurrent_to_output_weights.allocator()->allocate();
269 input_gate_bias.allocator()->allocate();
270 forget_gate_bias.allocator()->allocate();
271 cell_gate_bias.allocator()->allocate();
272 output_gate_bias.allocator()->allocate();
273 cell_state.allocator()->allocate();
274 output_state.allocator()->allocate();
275
276 // Fill weights and biases
277 fill_tensor(input_to_input_weights, std::vector<uint8_t>{ 141, 89, 200, 180, 46, 50, 87, 128,
278 149, 227, 177, 187, 212, 229, 54, 111,
279 131, 116, 3, 58, 196, 26, 131, 255,
280 22, 106, 216, 69, 239, 12, 232, 207,
281 184, 56, 236, 172, 28, 143, 161, 124,
282 255, 33, 197, 122, 47, 197, 26, 229,
283 91, 79, 11, 160, 26, 80, 100, 36,
284 248, 186, 97, 61, 125, 46, 14, 100, });
285
286 fill_tensor(input_to_forget_weights, std::vector<uint8_t> { 237, 165, 141, 249, 72, 116, 36 , 115,
287 234, 213, 85, 84, 59, 62, 150, 246,
288 182, 102, 158, 214, 182, 183, 94, 11,
289 158, 192, 92, 189, 160, 219, 206, 249,
290 88, 213, 193, 244, 151, 72, 129, 49,
291 239, 83, 106, 9, 169, 187, 125, 171,
292 32, 141, 126, 92, 13, 36, 224, 150,
293 187, 250, 178, 169, 89, 214, 91, 173 });
294
295 fill_tensor(input_to_cell_weights, std::vector<uint8_t> { 93, 103, 226, 139, 185, 252, 129, 171,
296 159, 32, 25, 175, 224, 183, 165, 35,
297 207, 69, 238, 228, 149, 214, 79, 6,
298 5, 66, 102, 14, 19, 111, 36, 143,
299 22, 85, 13, 78, 236, 121, 122, 77,
300 249, 39, 88, 12, 205, 143, 93, 240,
301 167, 89, 188, 50, 73, 69, 201, 251,
302 59, 32, 203, 184, 139, 191, 199, 74});
303
304 fill_tensor(input_to_output_weights, std::vector<uint8_t> { 205, 7, 95, 104, 252, 143, 226, 73,
305 229, 114, 152, 171, 221, 153, 73, 229,
306 153, 165, 223, 239, 100, 38, 172, 211,
307 226, 133, 239, 207, 116, 230, 170, 100,
308 241, 95, 171, 124, 63, 115, 32, 127,
309 141, 239, 53, 193, 201, 53, 104, 178,
310 186, 212, 167, 107, 226, 230, 71, 213,
311 148, 217, 19, 248, 233, 195, 183, 156 });
312
313 fill_tensor(recurrent_to_input_weights, std::vector<uint8_t> { 147, 112, 140, 103, 3, 255, 17, 49,
314 84, 112, 144, 213, 138, 142, 112, 66,
315 117, 30, 101, 35, 25, 132, 211, 229,
316 183, 208, 102, 16, 38, 85, 101, 152,
317 226, 83, 132, 22, 161, 110, 157, 129,
318 184, 63, 168, 42, 220, 126, 209, 157,
319 5, 88, 243, 83, 249, 19, 226, 209,
320 173, 96, 185, 77, 146, 227, 238, 136 });
321
322
323 fill_tensor(recurrent_to_forget_weights, std::vector<uint8_t> { 52, 132, 92, 200, 213, 32, 213, 37,
324 116, 142, 116, 180, 4, 172, 158, 143,
325 110, 40, 99, 28, 221, 153, 133, 2,
326 247, 144, 198, 100, 20, 15, 221, 196,
327 159, 178, 188, 151, 171, 15, 25, 217,
328 178, 109, 110, 118, 128, 39, 232, 234,
329 184, 214, 177, 13, 56, 6, 28, 252,
330 89, 187, 242, 59, 146, 111, 132, 129});
331
332 fill_tensor(recurrent_to_cell_weights, std::vector<uint8_t> { 70, 44, 137, 29, 36, 127, 1, 241,
333 26, 241, 142, 114, 67, 181, 49, 57,
334 131, 152, 175, 77, 23, 63, 37, 124,
335 150, 113, 95, 103, 110, 201, 69, 97,
336 196, 242, 62, 214, 66, 19, 45, 135,
337 22, 168, 149, 104, 77, 101, 36, 68,
338 170, 116, 222, 100, 109, 1, 154, 18,
339 133, 215, 105, 93, 31, 57, 231, 112 });
340
341
342 fill_tensor(recurrent_to_output_weights, std::vector<uint8_t> { 45 , 181 , 220 , 219 , 49 , 63 , 49 , 129,
343 7 , 166 , 104 , 114 , 83 , 40 , 1 , 195,
344 245 , 142 , 82 , 232 , 104 , 245 , 82 , 196,
345 111 , 56 , 156 , 9 , 141 , 240 , 180 , 148,
346 247 , 198 , 234 , 137 , 13 , 210 , 161 , 192,
347 196 , 59 , 233 , 184 , 142 , 187 , 140 , 166,
348 2 , 95 , 152 , 46 , 71 , 46 , 113 , 32,
349 175 , 229 , 86 , 87 , 62 , 93 , 74 , 130});
350
351 fill_tensor(input_gate_bias, std::vector<int> { -40040, -106916, -92315, -79123, 45160, -17954, 50962, -63758 });
352 fill_tensor(forget_gate_bias, std::vector<int> { -128514, 8463, -57831, 116977, 106547, -28132, -124557, 44941 });
353 fill_tensor(cell_gate_bias, std::vector<int> { 88388 , 123601, -116148, -13022, 21619, 48926, 57523, 39332 });
354 fill_tensor(output_gate_bias, std::vector<int> { 59485 , -33070, 21386, -100633, -115959, 125768, -56407, 24897 });
355
356 SimpleTensor<uint8_t> expected_output(output_shape, DataType::QASYMM8, 1, qasymm);
357
358 // Initialize state
359 fill_tensor(output_state, std::vector<uint8_t> { 128, 128, 128, 128, 128, 128, 128, 128,
360 128, 128, 128, 128, 128, 128, 128, 128,
361 128, 128, 128, 128, 128, 128, 128, 128,
362 128, 128, 128, 128, 128, 128, 128, 128,
363 128, 128, 128, 128, 128, 128, 128, 128,
364 128, 128, 128, 128, 128, 128, 128, 128,
365 128, 128, 128, 128, 128, 128, 128, 128,
366 128, 128, 128, 128, 128, 128, 128, 128,
367 128, 128, 128, 128, 128, 128, 128, 128,
368 128, 128, 128, 128, 128, 128, 128, 128,
369 128, 128, 128, 128, 128, 128, 128, 128,
370 128, 128, 128, 128, 128, 128, 128, 128,
371 128, 128, 128, 128, 128, 128, 128, 128,
372 128, 128, 128, 128, 128, 128, 128, 128,
373 128, 128, 128, 128, 128, 128, 128, 128,
374 128, 128, 128, 128, 128, 128, 128, 128 });
375
376 fill_tensor(cell_state, std::vector<int16_t> { 0, 0, 0, 0, 0, 0, 0, 0,
377 0, 0, 0, 0, 0, 0, 0, 0,
378 0, 0, 0, 0, 0, 0, 0, 0,
379 0, 0, 0, 0, 0, 0, 0, 0,
380 0, 0, 0, 0, 0, 0, 0, 0,
381 0, 0, 0, 0, 0, 0, 0, 0,
382 0, 0, 0, 0, 0, 0, 0, 0,
383 0, 0, 0, 0, 0, 0, 0, 0,
384 0, 0, 0, 0, 0, 0, 0, 0,
385 0, 0, 0, 0, 0, 0, 0, 0,
386 0, 0, 0, 0, 0, 0, 0, 0,
387 0, 0, 0, 0, 0, 0, 0, 0,
388 0, 0, 0, 0, 0, 0, 0, 0,
389 0, 0, 0, 0, 0, 0, 0, 0,
390 0, 0, 0, 0, 0, 0, 0, 0,
391 0, 0, 0, 0, 0, 0, 0, 0});
392
393 // First input
394 fill_tensor(input, std::vector<uint8_t> { 247, 203, 159, 131, 182, 114, 207, 195,
395 48 , 61 , 154, 16, 80, 101, 116, 255,
396 50 , 115 , 45, 186, 75, 212, 98, 48,
397 88 , 146 , 24, 143, 218, 174, 203, 200,
398 239 , 16 , 66, 136, 234, 54, 94, 51,
399 101 , 128 , 220, 213, 164, 82, 137, 255,
400 70 , 165 , 234, 220, 66, 35, 183, 206,
401 39 , 57 , 180, 202, 23, 172, 224, 109,
402 102 , 215 , 186, 82, 215, 147, 85, 187,
403 96 , 249 , 59, 116, 150, 44, 167, 128,
404 34 , 217 , 148, 193, 243, 38, 250, 208,
405 112 , 130 , 208, 29, 16, 122, 20, 92,
406 24 , 72 , 104, 29, 150, 233, 151, 19,
407 158 , 192 , 254, 70, 73, 142, 106, 152,
408 3 , 61 , 24, 135, 212, 9, 80, 234,
409 147 , 246 , 83, 249, 49, 14, 68, 50});
410
411 fill_tensor(expected_output, std::vector<uint8_t> {131, 128, 128, 128, 128, 180, 129, 133,
412 136, 128, 126, 128, 128, 173, 135, 130,
413 160, 128, 128, 128, 128, 138, 132, 129,
414 131, 128, 127, 128, 128, 169, 129, 131,
415 133, 128, 128, 128, 128, 182, 130, 129,
416 131, 128, 128, 128, 128, 163, 129, 130,
417 131, 128, 128, 128, 128, 149, 132, 129,
418 143, 128, 127, 128, 128, 150, 134, 131,
419 134, 128, 128, 128, 128, 167, 130, 130,
420 131, 128, 128, 128, 128, 152, 132, 129,
421 128, 128, 128, 128, 128, 169, 130, 130,
422 173, 128, 128, 128, 128, 148, 139, 130,
423 152, 128, 128, 128, 128, 168, 139, 132,
424 147, 128, 128, 128, 128, 161, 131, 132,
425 130, 128, 128, 128, 128, 159, 134, 128,
426 140, 128, 128, 128, 128, 133, 132, 128 });
427
428 lstmq.run();
429 validate(CLAccessor(output_state), expected_output);
430
431 // Second input
432 fill_tensor(expected_output, std::vector<uint8_t> { 130, 128, 128, 128, 128, 205, 129, 137,
433 135, 128, 127, 128, 128, 190, 137, 132,
434 160, 128, 128, 128, 128, 142, 133, 131,
435 130, 128, 128, 128, 128, 185, 129, 133,
436 132, 128, 128, 128, 128, 198, 131, 130,
437 130, 128, 128, 128, 128, 178, 130, 131,
438 131, 128, 128, 128, 128, 158, 132, 131,
439 142, 128, 127, 128, 128, 158, 135, 134,
440 133, 128, 128, 128, 128, 178, 131, 132,
441 131, 128, 128, 128, 128, 160, 132, 130,
442 128, 128, 128, 128, 128, 190, 131, 131,
443 170, 128, 128, 128, 128, 157, 142, 131,
444 149, 128, 128, 128, 128, 178, 142, 135,
445 145, 128, 128, 128, 129, 173, 132, 135,
446 129, 128, 128, 128, 128, 171, 134, 129,
447 140, 128, 128, 128, 128, 135, 132, 129});
448 lstmq.run();
449 validate(CLAccessor(output_state), expected_output);
450 }
451 TEST_SUITE_END() // MultSmallerEq1
452
TEST_SUITE(MultGreater1)453 TEST_SUITE(MultGreater1)
454 TEST_CASE(RunSmall, framework::DatasetMode::PRECOMMIT)
455 {
456 //Input sequence length is 1
457 const int batch_size = 2;
458 const int input_size = 2;
459 const int output_size = 4;
460
461 QuantizationInfo qasymm(1.f / 128.f, 128);
462 QuantizationInfo qweights(1.f / 16.f, 16);
463 QuantizationInfo qsymm_3(8.f / 32768.f, 0);
464 QuantizationInfo qsymm_4(16.f / 32768.f, 0);
465
466 TensorShape input_shape{ input_size, batch_size };
467 TensorShape input_weights_shape{ input_size, output_size };
468 TensorShape recurrent_weights_shape{ output_size, output_size };
469 TensorShape output_shape{ output_size, batch_size};
470 TensorShape bias_shape{ output_size };
471
472 auto input_to_input_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
473 auto input_to_forget_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
474 auto input_to_cell_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
475 auto input_to_output_weights = create_tensor<CLTensor>(input_weights_shape, DataType::QASYMM8, 1, qweights);
476 auto recurrent_to_input_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
477 auto recurrent_to_forget_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
478 auto recurrent_to_cell_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
479 auto recurrent_to_output_weights = create_tensor<CLTensor>(recurrent_weights_shape, DataType::QASYMM8, 1, qweights);
480 auto input_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
481 auto forget_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
482 auto cell_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
483 auto output_gate_bias = create_tensor<CLTensor>(bias_shape, DataType::S32);
484
485 // LSTM input
486 auto input = create_tensor<CLTensor>(input_shape, DataType::QASYMM8, 1, qasymm);
487
488 // LSTM output state
489 auto output_state = create_tensor<CLTensor>(output_shape, DataType::QASYMM8, 1, qasymm);
490
491 // LSTM cell state
492 auto cell_state = create_tensor<CLTensor>(output_shape, DataType::QSYMM16, 1, qsymm_4);
493
494 CLLSTMLayerQuantized lstmq;
495
496 lstmq.configure(&input, &input_to_input_weights, &input_to_forget_weights, &input_to_cell_weights, &input_to_output_weights,
497 &recurrent_to_input_weights, &recurrent_to_forget_weights, &recurrent_to_cell_weights, &recurrent_to_output_weights,
498 &input_gate_bias, &forget_gate_bias, &cell_gate_bias, &output_gate_bias, &cell_state, &output_state, &cell_state, &output_state);
499
500 input.allocator()->allocate();
501 input_to_input_weights.allocator()->allocate();
502 input_to_forget_weights.allocator()->allocate();
503 input_to_cell_weights.allocator()->allocate();
504 input_to_output_weights.allocator()->allocate();
505 recurrent_to_input_weights.allocator()->allocate();
506 recurrent_to_forget_weights.allocator()->allocate();
507 recurrent_to_cell_weights.allocator()->allocate();
508 recurrent_to_output_weights.allocator()->allocate();
509 input_gate_bias.allocator()->allocate();
510 forget_gate_bias.allocator()->allocate();
511 cell_gate_bias.allocator()->allocate();
512 output_gate_bias.allocator()->allocate();
513 cell_state.allocator()->allocate();
514 output_state.allocator()->allocate();
515
516 // Fill weights and biases
517 fill_tensor(input_to_input_weights, std::vector<uint8_t>{ 122, 130,
518 124, 134,
519 120, 122,
520 134, 134 });
521
522 fill_tensor(input_to_forget_weights, std::vector<uint8_t> { 204, 193,
523 148, 59,
524 113, 17,
525 66, 197 });
526
527 fill_tensor(input_to_cell_weights, std::vector<uint8_t> { 172, 101,
528 184, 209,
529 165, 82,
530 108, 209 });
531
532 fill_tensor(input_to_output_weights, std::vector<uint8_t> { 203, 244,
533 219, 114,
534 130, 16,
535 163, 222 });
536
537 fill_tensor(recurrent_to_input_weights, std::vector<uint8_t> { 162, 168, 7, 95,
538 91, 155, 108, 216,
539 255, 100, 48, 188,
540 58, 37, 186, 147 });
541
542 fill_tensor(recurrent_to_forget_weights, std::vector<uint8_t> { 46, 58, 47, 170,
543 246, 96, 12, 99,
544 68, 23, 186, 161,
545 237, 164, 89, 6 });
546
547 fill_tensor(recurrent_to_cell_weights, std::vector<uint8_t> { 234, 99, 71, 206,
548 205, 159, 64, 253,
549 191, 148, 116, 8,
550 209, 136, 59, 138 });
551
552 fill_tensor(recurrent_to_output_weights, std::vector<uint8_t> { 23, 241, 137, 36,
553 206, 5, 227, 56,
554 254, 176, 231, 47,
555 18, 201, 161, 11 });
556
557 fill_tensor(input_gate_bias, std::vector<int> {-103038, 30525, 115255, -38154 });
558 fill_tensor(forget_gate_bias, std::vector<int> { -23428, 126970, 116806, 46307 });
559 fill_tensor(cell_gate_bias, std::vector<int> { 128006, 69949, -42808, 42568 });
560 fill_tensor(output_gate_bias, std::vector<int> { -67066, -53607, 47233, 7300 });
561
562 SimpleTensor<uint8_t> expected_output(output_shape, DataType::QASYMM8, 1, qasymm);
563
564 // Initialize state
565 fill_tensor(output_state, std::vector<uint8_t> { 128, 128, 128, 128,
566 128, 128, 128, 128 });
567 fill_tensor(cell_state, std::vector<int16_t> { 0, 0, 0, 0,
568 0, 0, 0, 0 });
569
570 // First input
571 fill_tensor(input, std::vector<uint8_t> { 106, 193,
572 155, 150 });
573
574 fill_tensor(expected_output, std::vector<uint8_t> { 128, 128, 31, 128,
575 128, 128, 31, 128 });
576
577 lstmq.run();
578 validate(CLAccessor(output_state), expected_output);
579
580 // Second input
581 fill_tensor(expected_output, std::vector<uint8_t> { 128, 128, 5, 128,
582 128, 128, 5, 128 });
583 lstmq.run();
584 validate(CLAccessor(output_state), expected_output);
585
586 // Third input
587 fill_tensor(expected_output, std::vector<uint8_t> { 128, 128, 1, 128,
588 128, 128, 1, 128, });
589 lstmq.run();
590 validate(CLAccessor(output_state), expected_output);
591 }
592 TEST_SUITE_END() // MultGreater1
593 TEST_SUITE_END() // IntegrationTestCase
594 // clang-format on
595 // *INDENT-ON*
596
597 TEST_SUITE_END() // LSTMLayerQuantized
598 TEST_SUITE_END() // CL
599 } // namespace validation
600 } // namespace test
601 } // namespace arm_compute
602