1 /*
2 * Copyright (c) 2019-2021 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/runtime/NEON/functions/NELSTMLayerQuantized.h"
25
26 #include "arm_compute/core/Utils.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
29 #include "src/common/utils/Log.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31
32 #include <cmath>
33 #include <memory>
34 #include <tuple>
35
36 namespace arm_compute
37 {
38 namespace
39 {
40 // Quantization info structures used in the LSTMQuantize layer
41 const QuantizationInfo qasymm(1.f / 128.f, 128);
42 const QuantizationInfo qsymm_3(8.f / 32768.f, 0); // qsymm16 with 3 integer bit
43 const QuantizationInfo qsymm_4(16.f / 32768.f, 0); // qsymm16 with 4 integer bit
44 const QuantizationInfo qsymm_0(1.f / 32768.f, 0); // qsymm16 with 0 integer bit
45 } // namespace
46 NELSTMLayerQuantized::~NELSTMLayerQuantized() = default;
47
NELSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager)48 NELSTMLayerQuantized::NELSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager)
49 : _memory_group(std::move(memory_manager)), _gemmlowp(), _output_stage(), _transpose_weights(), _concat_input_weights(), _concat_recurrent_weights(), _concat_weights(), _concat_inputs(),
50 _concat_bias(), _sigmoid_forget_gate(), _sigmoid_input_gate(), _sigmoid_output_gate(), _tanh_modulation_gate(), _tanh_output_state(), _add1(), _add2(), _mul1(), _mul2(), _mul3(),
51 _slice_input_tensor(), _slice_forget_tensor(), _slice_cell_tensor(), _slice_output_tensor(), _dequantize(), _quantize(), _input_to_input_weights(nullptr), _input_to_forget_weights(nullptr),
52 _input_to_cell_weights(nullptr), _input_to_output_weights(nullptr), _recurrent_to_input_weights(nullptr), _recurrent_to_forget_weights(nullptr), _recurrent_to_cell_weights(nullptr),
53 _recurrent_to_output_weights(nullptr), _input_gate_bias(nullptr), _forget_gate_bias(nullptr), _cell_bias(nullptr), _output_gate_bias(nullptr), _recurrent_weights(), _input_weights(), _weights(),
54 _input(), _weights_transposed(), _output_highp(), _output_lowp(), _bias(), _forget_gate_input(), _input_gate_input(), _output_gate_input(), _input_modulation_gate_input(), _forget_gate_output(),
55 _input_gate_output(), _output_gate_output(), _input_modulation_gate_output(), _cell_state1(), _cell_state2(), _output_state_tmp(), _output_state_out_symm(), _output_state_out_f32(),
56 _is_prepared(false)
57 {
58 }
59
configure(const ITensor * input,const ITensor * input_to_input_weights,const ITensor * input_to_forget_weights,const ITensor * input_to_cell_weights,const ITensor * input_to_output_weights,const ITensor * recurrent_to_input_weights,const ITensor * recurrent_to_forget_weights,const ITensor * recurrent_to_cell_weights,const ITensor * recurrent_to_output_weights,const ITensor * input_gate_bias,const ITensor * forget_gate_bias,const ITensor * cell_bias,const ITensor * output_gate_bias,ITensor * cell_state_in,const ITensor * output_state_in,ITensor * cell_state_out,ITensor * output_state_out)60 void NELSTMLayerQuantized::configure(const ITensor *input,
61 const ITensor *input_to_input_weights, const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights,
62 const ITensor *recurrent_to_input_weights, const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights,
63 const ITensor *input_gate_bias, const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
64 ITensor *cell_state_in, const ITensor *output_state_in,
65 ITensor *cell_state_out, ITensor *output_state_out)
66 {
67 ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
68 recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
69 input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out);
70
71 ARM_COMPUTE_ERROR_THROW_ON(NELSTMLayerQuantized::validate(input->info(), input_to_input_weights->info(), input_to_forget_weights->info(), input_to_cell_weights->info(),
72 input_to_output_weights->info(),
73 recurrent_to_input_weights->info(), recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
74 input_gate_bias->info(), forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(), cell_state_in->info(), output_state_in->info(), cell_state_out->info(), output_state_out->info()));
75
76 ARM_COMPUTE_LOG_PARAMS(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
77 recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
78 input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out);
79
80 const int input_size = input->info()->dimension(0);
81 const int batch_size = input->info()->dimension(1);
82 const int output_size = input_to_input_weights->info()->dimension(1);
83
84 const QuantizationInfo qweights = input_to_input_weights->info()->quantization_info(); // Weights quantization
85
86 auto_init_if_empty(*cell_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QSYMM16, qsymm_4));
87 auto_init_if_empty(*output_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QASYMM8, qasymm));
88
89 _input_to_input_weights = input_to_input_weights;
90 _input_to_forget_weights = input_to_forget_weights;
91 _input_to_cell_weights = input_to_cell_weights;
92 _input_to_output_weights = input_to_output_weights;
93 _recurrent_to_input_weights = recurrent_to_input_weights;
94 _recurrent_to_forget_weights = recurrent_to_forget_weights;
95 _recurrent_to_cell_weights = recurrent_to_cell_weights;
96 _recurrent_to_output_weights = recurrent_to_output_weights;
97 _input_gate_bias = input_gate_bias;
98 _forget_gate_bias = forget_gate_bias;
99 _cell_bias = cell_bias;
100 _output_gate_bias = output_gate_bias;
101
102 // Weights concatenation
103 std::vector<const ITensor *> inputs_weights_vector{ input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights };
104 std::vector<const ITensor *> recurrent_weights_vector{ recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights };
105
106 _input_weights.allocator()->init(TensorInfo(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
107 _concat_input_weights.configure(inputs_weights_vector, &_input_weights, Window::DimY);
108
109 _recurrent_weights.allocator()->init(TensorInfo(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
110 _concat_recurrent_weights.configure(recurrent_weights_vector, &_recurrent_weights, Window::DimY);
111
112 std::vector<const ITensor *> weights_vector{ &_recurrent_weights, &_input_weights };
113 _weights.allocator()->init(TensorInfo(TensorShape(output_size + input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
114 _concat_weights.configure(weights_vector, &_weights, Window::DimX);
115 _transpose_weights.configure(&_weights, &_weights_transposed);
116
117 // Input concatenation
118 std::vector<const ITensor *> input_vector{ input, output_state_in };
119 _memory_group.manage(&_input);
120 _input.allocator()->init(TensorInfo(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm));
121 _concat_inputs.configure(input_vector, &_input, Window::DimX);
122
123 // Bias concatenation
124 std::vector<const ITensor *> bias_vector{ input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias };
125 _bias.allocator()->init(TensorInfo(TensorShape(4 * output_size), 1, DataType::S32));
126 _concat_bias.configure(bias_vector, &_bias, Window::DimX);
127
128 // Invert the offset for gemmlowp
129 _input.info()->set_quantization_info(QuantizationInfo(qasymm.uniform().scale, -qasymm.uniform().offset));
130 _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
131
132 // Run gemmlowp
133 _memory_group.manage(&_output_highp);
134 _output_highp.allocator()->init(TensorInfo(TensorShape(4 * output_size, batch_size), 1, DataType::S32));
135 _gemmlowp.configure(&_input, &_weights_transposed, nullptr, &_output_highp);
136 _input.allocator()->allocate();
137
138 // Set the offset back
139 _input.info()->set_quantization_info(QuantizationInfo(qasymm.uniform().scale, qasymm.uniform().offset));
140 _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
141
142 // multiplier = (input_scale * weights_scale) / output_scale (2 ^ (-12))
143 _output_lowp.allocator()->init(TensorInfo(_output_highp.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_3));
144
145 const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
146 int32_t output_multiplier = 0;
147 int32_t output_shift = 0;
148 quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
149
150 _memory_group.manage(&_output_lowp);
151
152 GEMMLowpOutputStageInfo info;
153 info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
154 info.gemmlowp_multiplier = output_multiplier;
155 info.gemmlowp_shift = output_shift;
156 info.output_data_type = DataType::QSYMM16;
157 _output_stage.configure(&_output_highp, &_bias, &_output_lowp, info);
158 _output_highp.allocator()->allocate();
159 _bias.allocator()->allocate();
160
161 // Get the gate tensors
162 if(batch_size > 1)
163 {
164 _memory_group.manage(&_input_gate_input);
165 _slice_input_tensor.configure(&_output_lowp, &_input_gate_input, { 0, 0 }, { output_size, batch_size });
166 _memory_group.manage(&_forget_gate_input);
167 _slice_forget_tensor.configure(&_output_lowp, &_forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size });
168 _memory_group.manage(&_input_modulation_gate_input);
169 _slice_cell_tensor.configure(&_output_lowp, &_input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size });
170 _memory_group.manage(&_output_gate_input);
171 _slice_output_tensor.configure(&_output_lowp, &_output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size });
172 _output_lowp.allocator()->allocate();
173 }
174 else
175 {
176 _memory_group.manage(&_input_gate_input);
177 _slice_input_tensor.configure(&_output_lowp, &_input_gate_input, { 0 }, { output_size });
178 _memory_group.manage(&_forget_gate_input);
179 _slice_forget_tensor.configure(&_output_lowp, &_forget_gate_input, { output_size }, { 2 * output_size });
180 _memory_group.manage(&_input_modulation_gate_input);
181 _slice_cell_tensor.configure(&_output_lowp, &_input_modulation_gate_input, { 2 * output_size }, { 3 * output_size });
182 _memory_group.manage(&_output_gate_input);
183 _slice_output_tensor.configure(&_output_lowp, &_output_gate_input, { 3 * output_size }, { 4 * output_size });
184 _output_lowp.allocator()->allocate();
185 }
186
187 // Forget gate
188 _memory_group.manage(&_forget_gate_output);
189 _forget_gate_output.allocator()->init(TensorInfo(_forget_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
190 _sigmoid_forget_gate.configure(&_forget_gate_input, &_forget_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
191 _forget_gate_input.allocator()->allocate();
192
193 // Input gate
194 _memory_group.manage(&_input_gate_output);
195 _input_gate_output.allocator()->init(TensorInfo(_input_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
196 _sigmoid_input_gate.configure(&_input_gate_input, &_input_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
197 _input_gate_input.allocator()->allocate();
198
199 // Input modulation gate equation
200 _memory_group.manage(&_input_modulation_gate_output);
201 _input_modulation_gate_output.allocator()->init(TensorInfo(_input_modulation_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
202 _tanh_modulation_gate.configure(&_input_modulation_gate_input, &_input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
203 _input_modulation_gate_input.allocator()->allocate();
204
205 // Output gate
206 _memory_group.manage(&_output_gate_output);
207 _output_gate_output.allocator()->init(TensorInfo(_output_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
208 _sigmoid_output_gate.configure(&_output_gate_input, &_output_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
209 _output_gate_input.allocator()->allocate();
210
211 // Long term memory
212 _memory_group.manage(&_cell_state1);
213 _cell_state1.allocator()->init(TensorInfo(_forget_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
214 _mul1.configure(&_forget_gate_output, cell_state_in, &_cell_state1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
215 _forget_gate_output.allocator()->allocate();
216
217 _memory_group.manage(&_cell_state2);
218 _cell_state2.allocator()->init(TensorInfo(_input_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
219 _mul2.configure(&_input_gate_output, &_input_modulation_gate_output, &_cell_state2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
220 _input_modulation_gate_output.allocator()->allocate();
221 _input_gate_output.allocator()->allocate();
222
223 _add1.configure(&_cell_state1, &_cell_state2, cell_state_out, ConvertPolicy::SATURATE);
224 _cell_state1.allocator()->allocate();
225 _cell_state2.allocator()->allocate();
226
227 // Short term memory
228 _memory_group.manage(&_output_state_tmp);
229 _output_state_tmp.allocator()->init(TensorInfo(cell_state_out->info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
230 _tanh_output_state.configure(cell_state_out, &_output_state_tmp, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
231
232 _memory_group.manage(&_output_state_out_symm);
233 _output_state_out_symm.allocator()->init(TensorInfo(_output_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
234 _mul3.configure(&_output_state_tmp, &_output_gate_output, &_output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
235 _output_gate_output.allocator()->allocate();
236 _output_state_tmp.allocator()->allocate();
237
238 // Requantize the output state from QSYMM16 to QASYMM8
239 _memory_group.manage(&_output_state_out_f32);
240 _output_state_out_f32.allocator()->init(TensorInfo(_output_state_out_symm.info()->tensor_shape(), 1, DataType::F32));
241 _dequantize.configure(&_output_state_out_symm, &_output_state_out_f32);
242 _output_state_out_symm.allocator()->allocate();
243
244 _quantize.configure(&_output_state_out_f32, output_state_out);
245 _output_state_out_f32.allocator()->allocate();
246 }
247
validate(const ITensorInfo * input,const ITensorInfo * input_to_input_weights,const ITensorInfo * input_to_forget_weights,const ITensorInfo * input_to_cell_weights,const ITensorInfo * input_to_output_weights,const ITensorInfo * recurrent_to_input_weights,const ITensorInfo * recurrent_to_forget_weights,const ITensorInfo * recurrent_to_cell_weights,const ITensorInfo * recurrent_to_output_weights,const ITensorInfo * input_gate_bias,const ITensorInfo * forget_gate_bias,const ITensorInfo * cell_bias,const ITensorInfo * output_gate_bias,const ITensorInfo * cell_state_in,const ITensorInfo * output_state_in,const ITensorInfo * cell_state_out,const ITensorInfo * output_state_out)248 Status NELSTMLayerQuantized::validate(const ITensorInfo *input,
249 const ITensorInfo *input_to_input_weights, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
250 const ITensorInfo *recurrent_to_input_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
251 const ITensorInfo *input_gate_bias, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
252 const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
253 const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out)
254 {
255 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_input_weights,
256 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in,
257 output_state_in, cell_state_out, output_state_out);
258
259 const int input_size = input->dimension(0);
260 const int batch_size = input->dimension(1);
261 const int output_size = input_to_input_weights->dimension(1);
262
263 // Dimensionality checks
264 ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2);
265 ARM_COMPUTE_RETURN_ERROR_ON(input_to_input_weights->num_dimensions() > 2);
266 ARM_COMPUTE_RETURN_ERROR_ON(input_gate_bias->num_dimensions() > 1);
267 ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
268
269 TensorInfo input_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(input_size, output_size)).set_data_type(DataType::QASYMM8));
270 TensorInfo recurrent_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(output_size, output_size)).set_data_type(DataType::QASYMM8));
271 TensorInfo bias_info(input_gate_bias->clone()->set_tensor_shape(TensorShape(output_size)).set_data_type(DataType::S32));
272 TensorInfo output_state_info(cell_state_in->clone()->set_tensor_shape(TensorShape(output_size, batch_size)).set_data_type(DataType::QASYMM8).set_quantization_info(qasymm));
273 TensorInfo cell_state_info(cell_state_in->clone()->set_tensor_shape(TensorShape(output_size, batch_size)).set_data_type(DataType::QSYMM16).set_quantization_info(qsymm_4));
274
275 // Shape checks
276 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&input_weights_info, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
277 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&recurrent_weights_info, recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
278 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
279 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_in);
280 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_in);
281
282 // Data type checks
283 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input_weights_info, input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
284 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
285 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
286 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_in);
287 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_in);
288
289 // Quantization checks
290 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&input_weights_info, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
291 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
292 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_in);
293 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_in);
294
295 // Validate internal functions
296 // _concat_input_weights
297 std::vector<const ITensorInfo *> inputs_weights_vector;
298 inputs_weights_vector.emplace_back(input_to_input_weights);
299 inputs_weights_vector.emplace_back(input_to_forget_weights);
300 inputs_weights_vector.emplace_back(input_to_cell_weights);
301 inputs_weights_vector.emplace_back(input_to_output_weights);
302 const QuantizationInfo qweights = input_to_input_weights->quantization_info(); // Weights quantization
303 const TensorInfo input_weights(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
304 ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(inputs_weights_vector, &input_weights, Window::DimY));
305
306 // _concat_recurrent_weights
307 std::vector<const ITensorInfo *> recurrent_weights_vector;
308 recurrent_weights_vector.emplace_back(recurrent_to_input_weights);
309 recurrent_weights_vector.emplace_back(recurrent_to_forget_weights);
310 recurrent_weights_vector.emplace_back(recurrent_to_cell_weights);
311 recurrent_weights_vector.emplace_back(recurrent_to_output_weights);
312 const TensorInfo recurrent_weights(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
313 ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(recurrent_weights_vector, &recurrent_weights, Window::DimY));
314
315 // _concat_weights
316 std::vector<const ITensorInfo *> weights_vector;
317 weights_vector.emplace_back(&recurrent_weights);
318 weights_vector.emplace_back(&input_weights);
319 const TensorInfo weights(TensorShape(input_size + output_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
320 ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(weights_vector, &weights, Window::DimX));
321 // _transpose_weights
322 const TensorShape weights_transposed_shape(weights.tensor_shape()[1], weights.tensor_shape()[0]);
323 TensorInfo weights_transposed = weights.clone()->set_is_resizable(true).set_tensor_shape(weights_transposed_shape);
324 ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(&weights, &weights_transposed));
325
326 // _concat_inputs
327 std::vector<const ITensorInfo *> input_vector;
328 input_vector.emplace_back(input);
329 input_vector.emplace_back(output_state_in);
330 TensorInfo input_concatenated(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm);
331 ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(input_vector, &input_concatenated, Window::DimX));
332
333 // _concat_bias
334 std::vector<const ITensorInfo *> bias_vector;
335 bias_vector.emplace_back(input_gate_bias);
336 bias_vector.emplace_back(forget_gate_bias);
337 bias_vector.emplace_back(cell_bias);
338 bias_vector.emplace_back(output_gate_bias);
339
340 const TensorInfo bias_concatenated(TensorShape(4 * output_size), 1, DataType::S32);
341 ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(bias_vector, &bias_concatenated, Window::DimX));
342
343 // Invert the offset for gemmlowp
344 input_concatenated.set_quantization_info(QuantizationInfo(qasymm.uniform().scale, -qasymm.uniform().offset));
345 weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
346
347 // _gemmlowp
348 const TensorInfo output_highp(TensorShape(4 * output_size, batch_size), 1, DataType::S32);
349 ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input_concatenated, &weights_transposed, nullptr, &output_highp));
350
351 // Set the offset back
352 input_concatenated.set_quantization_info(QuantizationInfo(qasymm.uniform().scale, qasymm.uniform().offset));
353 weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
354
355 const TensorInfo output_lowp(output_highp.tensor_shape(), 1, DataType::QSYMM16, qsymm_3);
356
357 const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
358 int32_t output_multiplier = 0;
359 int32_t output_shift = 0;
360 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
361
362 // _output_stage
363 GEMMLowpOutputStageInfo info;
364 info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
365 info.gemmlowp_multiplier = output_multiplier;
366 info.gemmlowp_shift = output_shift;
367 info.output_data_type = DataType::QSYMM16;
368 ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpOutputStage::validate(&output_highp, &bias_concatenated, &output_lowp, info));
369
370 TensorInfo input_gate_input;
371 TensorInfo forget_gate_input;
372 TensorInfo input_modulation_gate_input;
373 TensorInfo output_gate_input;
374
375 if(batch_size > 1)
376 {
377 // _slice_input_tensor
378 input_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
379 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_gate_input, { 0, 0 }, { output_size, batch_size }));
380 // _slice_forget_tensor
381 forget_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
382 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size }));
383 // _slice_cell_tensor
384 input_modulation_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
385 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size }));
386 // _slice_output_tensor
387 output_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
388 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size }));
389 }
390 else
391 {
392 // _slice_input_tensor
393 input_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
394 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_gate_input, { 0 }, { output_size }));
395 // _slice_forget_tensor
396 forget_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
397 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &forget_gate_input, { output_size }, { 2 * output_size }));
398 // _slice_cell_tensor
399 input_modulation_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
400 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_modulation_gate_input, { 2 * output_size }, { 3 * output_size }));
401 // _slice_output_tensor
402 output_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
403 ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &output_gate_input, { 3 * output_size }, { 4 * output_size }));
404 }
405
406 // _sigmoid_forget_gate
407 const TensorInfo forget_gate_output(forget_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
408 ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&forget_gate_input, &forget_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
409 // _sigmoid_input_gate
410 const TensorInfo input_gate_output(input_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
411 ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&input_gate_input, &input_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
412 // _tanh_modulation_gate
413 const TensorInfo input_modulation_gate_output(input_modulation_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
414 ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&input_modulation_gate_input, &input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f)));
415 // _sigmoid_output_gate
416 const TensorInfo output_gate_output(output_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
417 ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&output_gate_input, &output_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
418
419 // _mul_forget_gate_cell_state
420 const TensorInfo cell_state_tmp1(forget_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_4);
421 ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&forget_gate_output, cell_state_in, &cell_state_tmp1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
422
423 // _mul_input_gate_input_mod_gate
424 const TensorInfo cell_state_tmp2(input_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_4);
425 ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&input_gate_output, &input_modulation_gate_output, &cell_state_tmp2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
426
427 // _add_cell_state_tmps
428 ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp1, &cell_state_tmp2, cell_state_out, ConvertPolicy::SATURATE));
429
430 // _tanh_modulation_gate
431 const TensorInfo output_state_tmp(cell_state_out->tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
432 ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(cell_state_out, &output_state_tmp, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f)));
433
434 // _mul_output_state_tmp_output_gate
435 const TensorInfo output_state_out_symm(output_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
436 ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&output_state_tmp, &output_gate_output, &output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
437
438 // _dequantize
439 const TensorInfo output_state_out_f32(output_state_out_symm.tensor_shape(), 1, DataType::F32);
440 ARM_COMPUTE_RETURN_ON_ERROR(NEDequantizationLayer::validate(&output_state_out_symm, &output_state_out_f32));
441
442 // _quantize
443 ARM_COMPUTE_RETURN_ON_ERROR(NEQuantizationLayer::validate(&output_state_out_f32, output_state_out));
444
445 if(cell_state_out->total_size() != 0)
446 {
447 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_out);
448 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_out);
449 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_out);
450 }
451
452 if(output_state_out->total_size() != 0)
453 {
454 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_out);
455 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_out);
456 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_out);
457 }
458
459 return Status{};
460 }
461
run()462 void NELSTMLayerQuantized::run()
463 {
464 prepare();
465
466 // Acquire all the temporaries
467 MemoryGroupResourceScope scope_mg(_memory_group);
468
469 // Concat and transpose the input
470 _concat_inputs.run();
471
472 // Run gemmlowp
473 _gemmlowp.run();
474 _output_stage.run();
475
476 // Slice the results
477 _slice_input_tensor.run();
478 _slice_forget_tensor.run();
479 _slice_cell_tensor.run();
480 _slice_output_tensor.run();
481
482 // Gates
483 // Forget gate
484 _sigmoid_forget_gate.run();
485
486 // Input gate
487 _sigmoid_input_gate.run();
488
489 // Input modulation gate
490 _tanh_modulation_gate.run();
491
492 // Output gate
493 _sigmoid_output_gate.run();
494
495 // Cell state (long term memory)
496 _mul1.run();
497 _mul2.run();
498 _add1.run();
499
500 // Output state (short term memory)
501 _tanh_output_state.run();
502 _mul3.run();
503
504 // Requantize output state from QSYMM16 to QASYMM8
505 _dequantize.run();
506 _quantize.run();
507 }
508
prepare()509 void NELSTMLayerQuantized::prepare()
510 {
511 if(!_is_prepared)
512 {
513 _input_weights.allocator()->allocate();
514 _concat_input_weights.run();
515
516 _input_to_input_weights->mark_as_unused();
517 _input_to_forget_weights->mark_as_unused();
518 _input_to_cell_weights->mark_as_unused();
519 _input_to_output_weights->mark_as_unused();
520
521 _recurrent_weights.allocator()->allocate();
522 _concat_recurrent_weights.run();
523 _recurrent_to_input_weights->mark_as_unused();
524 _recurrent_to_forget_weights->mark_as_unused();
525 _recurrent_to_cell_weights->mark_as_unused();
526 _recurrent_to_output_weights->mark_as_unused();
527
528 _weights.allocator()->allocate();
529 _concat_weights.run();
530
531 _input_weights.mark_as_unused();
532 _input_weights.allocator()->free();
533 _recurrent_weights.mark_as_unused();
534 _recurrent_weights.allocator()->free();
535
536 _weights_transposed.allocator()->allocate();
537 _transpose_weights.run();
538
539 _weights.mark_as_unused();
540 _weights.allocator()->free();
541
542 _bias.allocator()->allocate();
543 _concat_bias.run();
544 _input_gate_bias->mark_as_unused();
545 _forget_gate_bias->mark_as_unused();
546 _cell_bias->mark_as_unused();
547 _output_gate_bias->mark_as_unused();
548
549 _is_prepared = true;
550 }
551 }
552
553 } // namespace arm_compute
554