Compute Library
 21.02
NELSTMLayerQuantized.cpp
Go to the documentation of this file.
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  */
25 
26 #include "arm_compute/core/Utils.h"
40 
41 #include <cmath>
42 #include <memory>
43 #include <tuple>
44 
45 namespace arm_compute
46 {
47 namespace
48 {
49 // Quantization info structures used in the LSTMQuantize layer
50 const QuantizationInfo qasymm(1.f / 128.f, 128);
51 const QuantizationInfo qsymm_3(8.f / 32768.f, 0); // qsymm16 with 3 integer bit
52 const QuantizationInfo qsymm_4(16.f / 32768.f, 0); // qsymm16 with 4 integer bit
53 const QuantizationInfo qsymm_0(1.f / 32768.f, 0); // qsymm16 with 0 integer bit
54 } // namespace
56 
57 NELSTMLayerQuantized::NELSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager)
58  : _memory_group(std::move(memory_manager)), _gemmlowp(), _output_stage(), _transpose_weights(), _concat_input_weights(), _concat_recurrent_weights(), _concat_weights(), _concat_inputs(),
59  _concat_bias(), _sigmoid_forget_gate(), _sigmoid_input_gate(), _sigmoid_output_gate(), _tanh_modulation_gate(), _tanh_output_state(), _add1(), _add2(), _mul1(), _mul2(), _mul3(),
60  _slice_input_tensor(), _slice_forget_tensor(), _slice_cell_tensor(), _slice_output_tensor(), _dequantize(), _quantize(), _input_to_input_weights(nullptr), _input_to_forget_weights(nullptr),
61  _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),
62  _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(),
63  _input(), _weights_transposed(), _output_highp(), _output_lowp(), _bias(), _forget_gate_input(), _input_gate_input(), _output_gate_input(), _input_modulation_gate_input(), _forget_gate_output(),
64  _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(),
65  _is_prepared(false)
66 {
67 }
68 
72  const ITensor *input_gate_bias, const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
73  ITensor *cell_state_in, const ITensor *output_state_in,
74  ITensor *cell_state_out, ITensor *output_state_out)
75 {
76  ARM_COMPUTE_ERROR_ON_NULLPTR(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  ARM_COMPUTE_ERROR_THROW_ON(NELSTMLayerQuantized::validate(input->info(), input_to_input_weights->info(), input_to_forget_weights->info(), input_to_cell_weights->info(),
81  input_to_output_weights->info(),
82  recurrent_to_input_weights->info(), recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
83  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()));
84 
85  const int input_size = input->info()->dimension(0);
86  const int batch_size = input->info()->dimension(1);
87  const int output_size = input_to_input_weights->info()->dimension(1);
88 
89  const QuantizationInfo qweights = input_to_input_weights->info()->quantization_info(); // Weights quantization
90 
91  auto_init_if_empty(*cell_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QSYMM16, qsymm_4));
92  auto_init_if_empty(*output_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QASYMM8, qasymm));
93 
94  _input_to_input_weights = input_to_input_weights;
95  _input_to_forget_weights = input_to_forget_weights;
96  _input_to_cell_weights = input_to_cell_weights;
97  _input_to_output_weights = input_to_output_weights;
98  _recurrent_to_input_weights = recurrent_to_input_weights;
99  _recurrent_to_forget_weights = recurrent_to_forget_weights;
100  _recurrent_to_cell_weights = recurrent_to_cell_weights;
101  _recurrent_to_output_weights = recurrent_to_output_weights;
102  _input_gate_bias = input_gate_bias;
103  _forget_gate_bias = forget_gate_bias;
104  _cell_bias = cell_bias;
105  _output_gate_bias = output_gate_bias;
106 
107  // Weights concatenation
108  std::vector<const ITensor *> inputs_weights_vector{ input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights };
109  std::vector<const ITensor *> recurrent_weights_vector{ recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights };
110 
111  _input_weights.allocator()->init(TensorInfo(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
112  _concat_input_weights.configure(inputs_weights_vector, &_input_weights, Window::DimY);
113 
114  _recurrent_weights.allocator()->init(TensorInfo(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
115  _concat_recurrent_weights.configure(recurrent_weights_vector, &_recurrent_weights, Window::DimY);
116 
117  std::vector<const ITensor *> weights_vector{ &_recurrent_weights, &_input_weights };
118  _weights.allocator()->init(TensorInfo(TensorShape(output_size + input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
119  _concat_weights.configure(weights_vector, &_weights, Window::DimX);
120  _transpose_weights.configure(&_weights, &_weights_transposed);
121 
122  // Input concatenation
123  std::vector<const ITensor *> input_vector{ input, output_state_in };
124  _memory_group.manage(&_input);
125  _input.allocator()->init(TensorInfo(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm));
126  _concat_inputs.configure(input_vector, &_input, Window::DimX);
127 
128  // Bias concatenation
129  std::vector<const ITensor *> bias_vector{ input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias };
130  _bias.allocator()->init(TensorInfo(TensorShape(4 * output_size), 1, DataType::S32));
131  _concat_bias.configure(bias_vector, &_bias, Window::DimX);
132 
133  // Invert the offset for gemmlowp
135  _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
136 
137  // Run gemmlowp
138  _memory_group.manage(&_output_highp);
139  _output_highp.allocator()->init(TensorInfo(TensorShape(4 * output_size, batch_size), 1, DataType::S32));
140  _gemmlowp.configure(&_input, &_weights_transposed, nullptr, &_output_highp);
141  _input.allocator()->allocate();
142 
143  // Set the offset back
145  _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
146 
147  // multiplier = (input_scale * weights_scale) / output_scale (2 ^ (-12))
148  _output_lowp.allocator()->init(TensorInfo(_output_highp.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_3));
149 
150  const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
151  int32_t output_multiplier = 0;
152  int32_t output_shift = 0;
153  quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
154 
155  _memory_group.manage(&_output_lowp);
156  _output_stage.configure(&_output_highp, &_bias, &_output_lowp, output_multiplier, output_shift);
157  _output_highp.allocator()->allocate();
158  _bias.allocator()->allocate();
159 
160  // Get the gate tensors
161  if(batch_size > 1)
162  {
163  _memory_group.manage(&_input_gate_input);
164  _slice_input_tensor.configure(&_output_lowp, &_input_gate_input, { 0, 0 }, { output_size, batch_size });
165  _memory_group.manage(&_forget_gate_input);
166  _slice_forget_tensor.configure(&_output_lowp, &_forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size });
167  _memory_group.manage(&_input_modulation_gate_input);
168  _slice_cell_tensor.configure(&_output_lowp, &_input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size });
169  _memory_group.manage(&_output_gate_input);
170  _slice_output_tensor.configure(&_output_lowp, &_output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size });
171  _output_lowp.allocator()->allocate();
172  }
173  else
174  {
175  _memory_group.manage(&_input_gate_input);
176  _slice_input_tensor.configure(&_output_lowp, &_input_gate_input, { 0 }, { output_size });
177  _memory_group.manage(&_forget_gate_input);
178  _slice_forget_tensor.configure(&_output_lowp, &_forget_gate_input, { output_size }, { 2 * output_size });
179  _memory_group.manage(&_input_modulation_gate_input);
180  _slice_cell_tensor.configure(&_output_lowp, &_input_modulation_gate_input, { 2 * output_size }, { 3 * output_size });
181  _memory_group.manage(&_output_gate_input);
182  _slice_output_tensor.configure(&_output_lowp, &_output_gate_input, { 3 * output_size }, { 4 * output_size });
183  _output_lowp.allocator()->allocate();
184  }
185 
186  // Forget gate
187  _memory_group.manage(&_forget_gate_output);
188  _forget_gate_output.allocator()->init(TensorInfo(_forget_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
189  _sigmoid_forget_gate.configure(&_forget_gate_input, &_forget_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
190  _forget_gate_input.allocator()->allocate();
191 
192  // Input gate
193  _memory_group.manage(&_input_gate_output);
194  _input_gate_output.allocator()->init(TensorInfo(_input_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
195  _sigmoid_input_gate.configure(&_input_gate_input, &_input_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
196  _input_gate_input.allocator()->allocate();
197 
198  // Input modulation gate equation
199  _memory_group.manage(&_input_modulation_gate_output);
200  _input_modulation_gate_output.allocator()->init(TensorInfo(_input_modulation_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
201  _tanh_modulation_gate.configure(&_input_modulation_gate_input, &_input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
202  _input_modulation_gate_input.allocator()->allocate();
203 
204  // Output gate
205  _memory_group.manage(&_output_gate_output);
206  _output_gate_output.allocator()->init(TensorInfo(_output_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
207  _sigmoid_output_gate.configure(&_output_gate_input, &_output_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
208  _output_gate_input.allocator()->allocate();
209 
210  // Long term memory
211  _memory_group.manage(&_cell_state1);
212  _cell_state1.allocator()->init(TensorInfo(_forget_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
213  _mul1.configure(&_forget_gate_output, cell_state_in, &_cell_state1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
214  _forget_gate_output.allocator()->allocate();
215 
216  _memory_group.manage(&_cell_state2);
217  _cell_state2.allocator()->init(TensorInfo(_input_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
218  _mul2.configure(&_input_gate_output, &_input_modulation_gate_output, &_cell_state2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
219  _input_modulation_gate_output.allocator()->allocate();
220  _input_gate_output.allocator()->allocate();
221 
222  _add1.configure(&_cell_state1, &_cell_state2, cell_state_out, ConvertPolicy::SATURATE);
223  _cell_state1.allocator()->allocate();
224  _cell_state2.allocator()->allocate();
225 
226  // Short term memory
227  _memory_group.manage(&_output_state_tmp);
228  _output_state_tmp.allocator()->init(TensorInfo(cell_state_out->info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
229  _tanh_output_state.configure(cell_state_out, &_output_state_tmp, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
230 
231  _memory_group.manage(&_output_state_out_symm);
232  _output_state_out_symm.allocator()->init(TensorInfo(_output_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
233  _mul3.configure(&_output_state_tmp, &_output_gate_output, &_output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
234  _output_gate_output.allocator()->allocate();
235  _output_state_tmp.allocator()->allocate();
236 
237  // Requantize the output state from QSYMM16 to QASYMM8
238  _memory_group.manage(&_output_state_out_f32);
239  _output_state_out_f32.allocator()->init(TensorInfo(_output_state_out_symm.info()->tensor_shape(), 1, DataType::F32));
240  _dequantize.configure(&_output_state_out_symm, &_output_state_out_f32);
241  _output_state_out_symm.allocator()->allocate();
242 
243  _quantize.configure(&_output_state_out_f32, output_state_out);
244  _output_state_out_f32.allocator()->allocate();
245 }
246 
251  const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
252  const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out)
253 {
254  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,
255  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,
256  output_state_in, cell_state_out, output_state_out);
257 
258  const int input_size = input->dimension(0);
259  const int batch_size = input->dimension(1);
260  const int output_size = input_to_input_weights->dimension(1);
261 
262  // Dimensionality checks
264  ARM_COMPUTE_RETURN_ERROR_ON(input_to_input_weights->num_dimensions() > 2);
265  ARM_COMPUTE_RETURN_ERROR_ON(input_gate_bias->num_dimensions() > 1);
266  ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
267 
268  TensorInfo input_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(input_size, output_size)).set_data_type(DataType::QASYMM8));
269  TensorInfo recurrent_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(output_size, output_size)).set_data_type(DataType::QASYMM8));
270  TensorInfo bias_info(input_gate_bias->clone()->set_tensor_shape(TensorShape(output_size)).set_data_type(DataType::S32));
271  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));
272  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));
273 
274  // Shape checks
275  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);
276  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);
277  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
278  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_in);
279  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_in);
280 
281  // Data type checks
282  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);
283  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
284  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
285  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_in);
286  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_in);
287 
288  // Quantization checks
289  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);
290  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
291  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_in);
292  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_in);
293 
294  // Validate internal functions
295  // _concat_input_weights
296  std::vector<const ITensorInfo *> inputs_weights_vector;
297  inputs_weights_vector.emplace_back(input_to_input_weights);
298  inputs_weights_vector.emplace_back(input_to_forget_weights);
299  inputs_weights_vector.emplace_back(input_to_cell_weights);
300  inputs_weights_vector.emplace_back(input_to_output_weights);
301  const QuantizationInfo qweights = input_to_input_weights->quantization_info(); // Weights quantization
302  const TensorInfo input_weights(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
303  ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(inputs_weights_vector, &input_weights, Window::DimY));
304 
305  // _concat_recurrent_weights
306  std::vector<const ITensorInfo *> recurrent_weights_vector;
307  recurrent_weights_vector.emplace_back(recurrent_to_input_weights);
308  recurrent_weights_vector.emplace_back(recurrent_to_forget_weights);
309  recurrent_weights_vector.emplace_back(recurrent_to_cell_weights);
310  recurrent_weights_vector.emplace_back(recurrent_to_output_weights);
311  const TensorInfo recurrent_weights(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
312  ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(recurrent_weights_vector, &recurrent_weights, Window::DimY));
313 
314  // _concat_weights
315  std::vector<const ITensorInfo *> weights_vector;
316  weights_vector.emplace_back(&recurrent_weights);
317  weights_vector.emplace_back(&input_weights);
318  const TensorInfo weights(TensorShape(input_size + output_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
320  // _transpose_weights
321  const TensorShape weights_transposed_shape(weights.tensor_shape()[1], weights.tensor_shape()[0]);
322  TensorInfo weights_transposed = weights.clone()->set_is_resizable(true).set_tensor_shape(weights_transposed_shape);
323  ARM_COMPUTE_RETURN_ON_ERROR(NETranspose::validate(&weights, &weights_transposed));
324 
325  // _concat_inputs
326  std::vector<const ITensorInfo *> input_vector;
327  input_vector.emplace_back(input);
328  input_vector.emplace_back(output_state_in);
329  TensorInfo input_concatenated(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm);
330  ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(input_vector, &input_concatenated, Window::DimX));
331 
332  // _concat_bias
333  std::vector<const ITensorInfo *> bias_vector;
334  bias_vector.emplace_back(input_gate_bias);
335  bias_vector.emplace_back(forget_gate_bias);
336  bias_vector.emplace_back(cell_bias);
337  bias_vector.emplace_back(output_gate_bias);
338 
339  const TensorInfo bias_concatenated(TensorShape(4 * output_size), 1, DataType::S32);
340  ARM_COMPUTE_RETURN_ON_ERROR(NEConcatenateLayer::validate(bias_vector, &bias_concatenated, Window::DimX));
341 
342  // Invert the offset for gemmlowp
344  weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
345 
346  // _gemmlowp
347  const TensorInfo output_highp(TensorShape(4 * output_size, batch_size), 1, DataType::S32);
348  ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(&input_concatenated, &weights_transposed, nullptr, &output_highp));
349 
350  // Set the offset back
352  weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
353 
354  const TensorInfo output_lowp(output_highp.tensor_shape(), 1, DataType::QSYMM16, qsymm_3);
355 
356  const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
357  int32_t output_multiplier = 0;
358  int32_t output_shift = 0;
359  ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
360 
361  // _output_stage
362  ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&output_highp, &bias_concatenated, &output_lowp));
363 
364  TensorInfo input_gate_input;
365  TensorInfo forget_gate_input;
366  TensorInfo input_modulation_gate_input;
367  TensorInfo output_gate_input;
368 
369  if(batch_size > 1)
370  {
371  // _slice_input_tensor
372  input_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
373  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_gate_input, { 0, 0 }, { output_size, batch_size }));
374  // _slice_forget_tensor
375  forget_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
376  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size }));
377  // _slice_cell_tensor
378  input_modulation_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
379  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size }));
380  // _slice_output_tensor
381  output_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
382  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size }));
383  }
384  else
385  {
386  // _slice_input_tensor
387  input_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
388  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_gate_input, { 0 }, { output_size }));
389  // _slice_forget_tensor
390  forget_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
391  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &forget_gate_input, { output_size }, { 2 * output_size }));
392  // _slice_cell_tensor
393  input_modulation_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
394  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &input_modulation_gate_input, { 2 * output_size }, { 3 * output_size }));
395  // _slice_output_tensor
396  output_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
397  ARM_COMPUTE_RETURN_ON_ERROR(NESlice::validate(&output_lowp, &output_gate_input, { 3 * output_size }, { 4 * output_size }));
398  }
399 
400  // _sigmoid_forget_gate
401  const TensorInfo forget_gate_output(forget_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
403  // _sigmoid_input_gate
404  const TensorInfo input_gate_output(input_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
406  // _tanh_modulation_gate
407  const TensorInfo input_modulation_gate_output(input_modulation_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
408  ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(&input_modulation_gate_input, &input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f)));
409  // _sigmoid_output_gate
410  const TensorInfo output_gate_output(output_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
412 
413  // _mul_forget_gate_cell_state
414  const TensorInfo cell_state_tmp1(forget_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_4);
416 
417  // _mul_input_gate_input_mod_gate
418  const TensorInfo cell_state_tmp2(input_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_4);
419  ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&input_gate_output, &input_modulation_gate_output, &cell_state_tmp2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
420 
421  // _add_cell_state_tmps
422  ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAddition::validate(&cell_state_tmp1, &cell_state_tmp2, cell_state_out, ConvertPolicy::SATURATE));
423 
424  // _tanh_modulation_gate
425  const TensorInfo output_state_tmp(cell_state_out->tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
427 
428  // _mul_output_state_tmp_output_gate
429  const TensorInfo output_state_out_symm(output_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
430  ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplication::validate(&output_state_tmp, &output_gate_output, &output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
431 
432  // _dequantize
433  const TensorInfo output_state_out_f32(output_state_out_symm.tensor_shape(), 1, DataType::F32);
434  ARM_COMPUTE_RETURN_ON_ERROR(NEDequantizationLayer::validate(&output_state_out_symm, &output_state_out_f32));
435 
436  // _quantize
437  ARM_COMPUTE_RETURN_ON_ERROR(NEQuantizationLayer::validate(&output_state_out_f32, output_state_out));
438 
439  if(cell_state_out->total_size() != 0)
440  {
441  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_out);
442  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_out);
443  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_out);
444  }
445 
446  if(output_state_out->total_size() != 0)
447  {
448  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_out);
449  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_out);
450  ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_out);
451  }
452 
453  return Status{};
454 }
455 
457 {
458  prepare();
459 
460  // Acquire all the temporaries
461  MemoryGroupResourceScope scope_mg(_memory_group);
462 
463  // Concat and transpose the input
464  _concat_inputs.run();
465 
466  // Run gemmlowp
467  _gemmlowp.run();
468  _output_stage.run();
469 
470  // Slice the results
471  _slice_input_tensor.run();
472  _slice_forget_tensor.run();
473  _slice_cell_tensor.run();
474  _slice_output_tensor.run();
475 
476  // Gates
477  // Forget gate
478  _sigmoid_forget_gate.run();
479 
480  // Input gate
481  _sigmoid_input_gate.run();
482 
483  // Input modulation gate
484  _tanh_modulation_gate.run();
485 
486  // Output gate
487  _sigmoid_output_gate.run();
488 
489  // Cell state (long term memory)
490  _mul1.run();
491  _mul2.run();
492  _add1.run();
493 
494  // Output state (short term memory)
495  _tanh_output_state.run();
496  _mul3.run();
497 
498  // Requantize output state from QSYMM16 to QASYMM8
499  _dequantize.run();
500  _quantize.run();
501 }
502 
504 {
505  if(!_is_prepared)
506  {
507  _input_weights.allocator()->allocate();
508  _concat_input_weights.run();
509 
510  _input_to_input_weights->mark_as_unused();
511  _input_to_forget_weights->mark_as_unused();
512  _input_to_cell_weights->mark_as_unused();
513  _input_to_output_weights->mark_as_unused();
514 
515  _recurrent_weights.allocator()->allocate();
516  _concat_recurrent_weights.run();
517  _recurrent_to_input_weights->mark_as_unused();
518  _recurrent_to_forget_weights->mark_as_unused();
519  _recurrent_to_cell_weights->mark_as_unused();
520  _recurrent_to_output_weights->mark_as_unused();
521 
522  _weights.allocator()->allocate();
523  _concat_weights.run();
524 
525  _input_weights.mark_as_unused();
526  _input_weights.allocator()->free();
527  _recurrent_weights.mark_as_unused();
528  _recurrent_weights.allocator()->free();
529 
530  _weights_transposed.allocator()->allocate();
531  _transpose_weights.run();
532 
533  _weights.mark_as_unused();
534  _weights.allocator()->free();
535 
536  _bias.allocator()->allocate();
537  _concat_bias.run();
538  _input_gate_bias->mark_as_unused();
539  _forget_gate_bias->mark_as_unused();
540  _cell_bias->mark_as_unused();
541  _output_gate_bias->mark_as_unused();
542 
543  _is_prepared = true;
544  }
545 }
546 
547 } // namespace arm_compute
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
void run() override
Run the kernels contained in the function.
void run() override
Run the kernels contained in the function.
~NELSTMLayerQuantized()
Default destructor.
Shape of a tensor.
Definition: TensorShape.h:39
void run() override final
Run the kernels contained in the function.
quantized, symmetric fixed-point 16-bit number
std::unique_ptr< ITensorInfo > clone() const override
Provide a clone of the current object of class T.
Definition: TensorInfo.cpp:316
static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, ConvertPolicy policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Static function to check if given info will lead to a valid configuration of NEArithmeticAddition.
void init(const TensorAllocator &allocator, const Coordinates &coords, TensorInfo &sub_info)
Shares the same backing memory with another tensor allocator, while the tensor info might be differen...
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(...)
Definition: Validate.h:610
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
QuantizationInfo qweights(1.f/16.f, 16)
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
[NEActivationLayer snippet]
void run() override
Run the kernels contained in the function.
void configure(const ITensor *input, ITensor *output)
Configure the kernel.
1 channel, 1 F32 per channel
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Activation Layer Information class.
Definition: Types.h:1550
Interface for Neon tensor.
Definition: ITensor.h:36
void configure(const ITensor *input1, const ITensor *input2, ITensor *output, ConvertPolicy policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel&#39;s inputs, output and conversion policy.
Copyright (c) 2017-2021 Arm Limited.
QuantizationInfo qsymm_3(8.f/32768.f, 0)
ITensorInfo & set_quantization_info(const QuantizationInfo &quantization_info) override
Set the quantization settings (scale and offset) of the tensor.
Definition: TensorInfo.cpp:380
TensorAllocator * allocator()
Return a pointer to the tensor&#39;s allocator.
Definition: Tensor.cpp:48
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor&#39;s metadata.
Definition: Tensor.cpp:33
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
void mark_as_unused() const
Marks a tensor as unused.
Definition: ITensor.cpp:168
NELSTMLayerQuantized(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
1 channel, 1 S32 per channel
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
void configure(const ITensor *input, ITensor *output)
Initialise the kernel&#39;s inputs and output.
Definition: NETranspose.cpp:32
Quantization information.
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
void configure(const ITensor *input, ITensor *output, const Coordinates &starts, const Coordinates &ends)
Configure kernel.
Definition: NESlice.cpp:85
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NEDequantizationLayer.
void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *output, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel&#39;s inputs, output.
void run() override
Run the kernels contained in the function.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
void configure(const ITensor *input1, const ITensor *input2, ITensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel&#39;s inputs, output and convertion policy.
void run() override
Run the kernels contained in the function.
quantized, asymmetric fixed-point 8-bit number unsigned
void allocate() override
Allocate size specified by TensorInfo of CPU memory.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
virtual ITensorInfo & set_quantization_info(const QuantizationInfo &quantization_info)=0
Set the quantization settings (scale and offset) of the tensor.
void prepare() override
Prepare the function for executing.
void free() override
Free allocated CPU memory.
void configure(const ITensor *input, ITensor *output)
Set the input and output tensors.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Static function to check if given info will lead to a valid configuration of NEPixelWiseMultiplicatio...
void run() override
Run the kernels contained in the function.
virtual size_t total_size() const =0
Returns the total size of the tensor in bytes.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:443
void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min=std::numeric_limits< int32_t >::lowest(), int max=std::numeric_limits< int32_t >::max())
Initialise the kernel&#39;s inputs, output.
void 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)
Initialize function&#39;s tensors.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
void configure(ITensor *input, ITensor *output, ActivationLayerInfo activation_info)
[NEActivationLayer snippet]
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
QuantizationInfo qasymm(1.f/128.f, 128)
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of NEGEMMLowpMatrixMultiply...
static Status validate(const std::vector< const ITensorInfo *> &inputs_vector, const ITensorInfo *output, size_t axis)
Static function to check if given info will lead to a valid configuration of NEConcatenateLayer.
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NEQuantizationLayer.
QuantizationInfo qsymm_4(16.f/32768.f, 0)
void configure(std::vector< const ITensor *> inputs_vector, ITensor *output, size_t axis)
Initialise the kernel&#39;s inputs vector and output.
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:262
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Coordinates &starts, const Coordinates &ends)
Static function to check if given info will lead to a valid configuration of NESlice.
Definition: NESlice.cpp:80
Truncates the least significant values that are lost in operations.
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NETranspose.
Definition: NETranspose.cpp:39
static Status 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)
Static function to check if given info will lead to a valid configuration of NELSTMLayer.
void run() override
Run the kernels contained in the function.
Definition: NESlice.cpp:93
static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min=std::numeric_limits< int32_t >::lowest(), int max=std::numeric_limits< int32_t >::max())
Static function to check if given info will lead to a valid configuration of NEGEMMLowpQuantizeDownIn...