Compute Library
 21.02
CLLSTMLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2021 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  */
25 
26 #include "arm_compute/core/Utils.h"
45 
46 namespace arm_compute
47 {
49 using namespace arm_compute::utils::info_helpers;
50 
51 CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
52  : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(),
53  _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(),
54  _transpose_cell_state(std::make_unique<CLTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
55  _pixelwise_mul_cell_state2(), _fully_connected_output(), _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(),
56  _fully_connected_output_state(), _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(),
57  _concat_weights_input_gate(), _concat_weights_output(), _ones_fill(), _mean_std_norm_input_gate(), _pixelwise_mul_input_gate_coeff(), _accum_input_gate_bias(), _mean_std_norm_forget_gate(),
58  _pixelwise_mul_forget_gate_coeff(), _accum_forget_gate_bias(), _mean_std_norm_cell_gate(), _pixelwise_mul_cell_gate_coeff(), _accum_cell_gate_bias(), _mean_std_norm_output_gate(),
59  _pixelwise_mul_output_gate_coeff(), _accum_output_gate_bias(), _input_gate_out1(), _input_gate_out2(), _input_gate_out3(), _input_gate_out4(), _forget_gate_out1(), _forget_gate_out2(),
60  _forget_gate_out3(), _forget_gate_out4(), _forget_gate_out5(), _forget_gate_out6(), _cell_state_out1(), _cell_state_out2(), _cell_state_out3(), _cell_state_out4(), _cell_state_out5(), _output1(),
61  _output2(), _output3(), _output4(), _cell_state_activation(), _output_state1(), _ones(), _input_layer_norm_out1(), _input_layer_norm_out2(), _forget_layer_norm_out1(), _forget_layer_norm_out2(),
62  _cell_layer_norm_out1(), _cell_layer_norm_out2(), _output_layer_norm_out1(), _output_layer_norm_out2(), _run_peephole_opt(false), _run_cifg_opt(false), _perform_cell_clipping(false),
63  _has_projection_weights(false), _perform_projection_clipping(false), _is_prepared(false), _is_layer_norm_lstm(false)
64 {
65 }
66 
67 CLLSTMLayer::~CLLSTMLayer() = default;
68 
72  const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
73  const ICLTensor *output_state_in, ICLTensor *cell_state_in,
74  ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
75  const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
76 {
77  configure(CLKernelLibrary::get().get_compile_context(), input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights,
78  recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, output_state_in, cell_state_in, scratch_buffer, output_state_out, cell_state_out, output, lstm_params, activation_info,
79  cell_threshold, projection_threshold);
80 }
81 
82 void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input,
85  const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
86  const ICLTensor *output_state_in, ICLTensor *cell_state_in,
87  ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
88  const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
89 {
91  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
92  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
93  forget_gate_bias, cell_bias, output_gate_bias,
94  output_state_in, cell_state_in,
95  scratch_buffer, output_state_out, cell_state_out, output);
96 
97  _is_layer_norm_lstm = lstm_params.use_layer_norm();
98 
99  // Set lstm parameters
100  LSTMParams<ITensorInfo> lstm_params_info{};
101  build_lstm_params_tensor_info(lstm_params, &lstm_params_info);
102 
103  // Validate
104  ARM_COMPUTE_ERROR_THROW_ON(CLLSTMLayer::validate(input->info(), input_to_forget_weights->info(),
105  input_to_cell_weights->info(), input_to_output_weights->info(),
106  recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
107  forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(),
108  output_state_in->info(), cell_state_in->info(),
109  scratch_buffer->info(), output_state_out->info(), cell_state_out->info(), output->info(),
110  lstm_params_info, activation_info, cell_threshold, projection_threshold));
111 
112  const TensorShape cell_state_shape = cell_state_in->info()->tensor_shape();
113  // Configure block that calculates the forget gate
114  // forget_gate = Activation(input * input_to_forget_weights + output_state_in * recurrent_to_forget_weights + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias)
115  // We optimize this as follows:
116  // forget_gate = Activation( (input,output_state_in) * (input_to_forget_weights,recurrent_to_forget_weights) + PixelWiseMul(cell_state, cell_to_forget_weights) + forget_gate_bias
117  _forget_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
118  _forget_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
119  _forget_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
120 
121  std::vector<const ICLTensor *> inputs_vector;
122  inputs_vector.emplace_back(input);
123  inputs_vector.emplace_back(output_state_in);
125  _forget_gate_out2.allocator()->init(TensorInfo(concat_shape, 1, input->info()->data_type()));
126 
127  _memory_group.manage(&_forget_gate_out2);
128  _concat_inputs_forget_gate.configure(compile_context, inputs_vector, &_forget_gate_out2, Window::DimX);
129 
130  std::vector<const ICLTensor *> weights_vector;
131 
132  weights_vector.emplace_back(input_to_forget_weights);
133  weights_vector.emplace_back(recurrent_to_forget_weights);
134  const TensorShape weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(weights_vector, 0);
135  _forget_gate_out6.allocator()->init(TensorInfo(weights_concat_shape, 1, input->info()->data_type()));
136 
137  _concat_weights_forget_gate.configure(compile_context, weights_vector, &_forget_gate_out6, Window::DimX);
138 
139  _memory_group.manage(&_forget_gate_out5);
140  _fully_connected_forget_gate.configure(compile_context, &_forget_gate_out2, &_forget_gate_out6, (_is_layer_norm_lstm) ? nullptr : forget_gate_bias, &_forget_gate_out5);
141  _memory_group.manage(&_forget_gate_out1);
142  _memory_group.manage(&_forget_gate_out3);
143  _forget_gate_out6.allocator()->allocate();
144 
145  CLTensor *forget_gate_out = &_forget_gate_out5;
146  if(lstm_params.has_peephole_opt())
147  {
148  _forget_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
149 
150  _run_peephole_opt = true;
151  _memory_group.manage(&_forget_gate_out4);
152  _pixelwise_mul_forget_gate.configure(compile_context, cell_state_in, lstm_params.cell_to_forget_weights(), &_forget_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
153  _accum_forget_gate1.configure(compile_context, &_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE);
154  _forget_gate_out4.allocator()->allocate();
155  _forget_gate_out5.allocator()->allocate();
156  forget_gate_out = &_forget_gate_out3;
157  }
158  else
159  {
160  _forget_gate_out3.allocator()->allocate();
161  }
162  if(_is_layer_norm_lstm)
163  {
164  _forget_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
165  _forget_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
166  _memory_group.manage(&_forget_layer_norm_out1);
167  _memory_group.manage(&_forget_layer_norm_out2);
168  _mean_std_norm_forget_gate.configure(compile_context, forget_gate_out);
169  _pixelwise_mul_forget_gate_coeff.configure(compile_context, forget_gate_out, lstm_params.forget_layer_norm_weights(), &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE,
171  // forget_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
172  forget_gate_out->allocator()->allocate();
173  _accum_forget_gate_bias.configure(compile_context, &_forget_layer_norm_out1, forget_gate_bias, &_forget_layer_norm_out2, ConvertPolicy::SATURATE);
174  _forget_layer_norm_out1.allocator()->allocate();
175  forget_gate_out = &_forget_layer_norm_out2;
176  }
177  _activation_forget_gate.configure(compile_context, forget_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
178 
179  // Configure block that calculates the input gate
180  // input_gate = Activation(input * input_to_input_weights + output_state * recurrent_to_input_weights + PixelWiseMul(cell_state, cell_to_input_weights) + input_gate_bias), without CIFG
181  // input_gate = 1 - forget_gate, with CIFG
182  // We optimize this as follows:
183  // input_gate = Activation((input,output_state) * (input_to_input_weights,recurrent_to_input_weights) + PixelWiseMul(cell_state, cell_to_input_weights) + input_gate_bias), without CIFG
184  _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
185  CLTensor *input_gate_out = &_input_gate_out1;
186  if(lstm_params.has_cifg_opt())
187  {
188  _memory_group.manage(&_input_gate_out1);
189  _ones.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
190  _ones_fill.configure(compile_context, &_ones, PixelValue(1, _ones.info()->data_type()));
191  _subtract_input_gate.configure(compile_context, &_ones, forget_gate_out, &_input_gate_out1, ConvertPolicy::SATURATE);
192  _ones.allocator()->allocate();
193  _run_cifg_opt = true;
194  }
195  else
196  {
197  _input_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
198  _input_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
199 
200  std::vector<const ICLTensor *> lstm_weights;
201  lstm_weights.emplace_back(lstm_params.input_to_input_weights());
202  lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
203  TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
204  _input_gate_out2.allocator()->init(TensorInfo(lstm_weights_concat_shape, 1, input->info()->data_type()));
205 
206  _concat_weights_input_gate.configure(compile_context, lstm_weights, &_input_gate_out2, Window::DimX);
207 
208  _memory_group.manage(&_input_gate_out1);
209 
210  _memory_group.manage(&_input_gate_out3);
211  _fully_connected_input_gate.configure(compile_context, &_forget_gate_out2, &_input_gate_out2, (_is_layer_norm_lstm) ? nullptr : lstm_params.input_gate_bias(), &_input_gate_out3);
212  _input_gate_out2.allocator()->allocate();
213 
214  input_gate_out = &_input_gate_out3;
215  if(_run_peephole_opt)
216  {
217  _memory_group.manage(&_input_gate_out4);
218  _pixelwise_mul_input_gate.configure(compile_context, cell_state_in, lstm_params.cell_to_input_weights(), &_input_gate_out4, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
219  _accum_input_gate1.configure(compile_context, &_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE);
220  _input_gate_out3.allocator()->allocate();
221  _input_gate_out4.allocator()->allocate();
222  input_gate_out = &_input_gate_out1;
223  }
224  else
225  {
226  _input_gate_out1.allocator()->allocate();
227  }
228 
229  if(_is_layer_norm_lstm)
230  {
231  _input_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
232  _input_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
233  _memory_group.manage(&_input_layer_norm_out1);
234  _memory_group.manage(&_input_layer_norm_out2);
235  _mean_std_norm_input_gate.configure(compile_context, input_gate_out);
236  _pixelwise_mul_input_gate_coeff.configure(compile_context, input_gate_out, lstm_params.input_layer_norm_weights(), &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE,
238  // input_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
239  input_gate_out->allocator()->allocate();
240  _accum_input_gate_bias.configure(compile_context, &_input_layer_norm_out1, lstm_params.input_gate_bias(), &_input_layer_norm_out2, ConvertPolicy::SATURATE);
241  _input_layer_norm_out1.allocator()->allocate();
242  input_gate_out = &_input_layer_norm_out2;
243  }
244  _activation_input_gate.configure(compile_context, input_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
245  }
246 
247  // Configure block that calculates the cell state
248  // cell_state = Clip((PixelwiseMul(input_gate, Activation(input * input_to_cell_weights + output_state_in * recurrent_to_cell_weights + cell_bias)) + PixelwiseMul(forget_gate, cell_state)), cell_threshold)
249  TensorShape cell_state1_shape = compute_transposed_shape(*recurrent_to_output_weights->info());
250  _cell_state_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
251  _cell_state_out2.allocator()->init(TensorInfo(cell_state1_shape, 1, input->info()->data_type()));
252  _cell_state_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
253  _cell_state_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
254  _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
255 
256  _memory_group.manage(&_cell_state_out1);
257  _fully_connected_cell_state.configure(compile_context, input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1);
258  _memory_group.manage(&_cell_state_out2);
259  _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights, &_cell_state_out2);
260  _memory_group.manage(&_cell_state_out3);
261  _gemm_cell_state1.configure(compile_context, output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
262  _cell_state_out2.allocator()->allocate();
263  _memory_group.manage(&_cell_state_out4);
264  _accum_cell_state1.configure(compile_context, &_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE);
265  CLTensor *cell_state_out_ptr = &_cell_state_out4;
266  if(_is_layer_norm_lstm)
267  {
268  _cell_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
269  _cell_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
270  _memory_group.manage(&_cell_layer_norm_out1);
271  _memory_group.manage(&_cell_layer_norm_out2);
272  _mean_std_norm_cell_gate.configure(compile_context, cell_state_out_ptr);
273  _pixelwise_mul_cell_gate_coeff.configure(compile_context, cell_state_out_ptr, lstm_params.cell_layer_norm_weights(), &_cell_layer_norm_out1, 1, ConvertPolicy::SATURATE,
275  // cell_state_out_ptr is going to be reassigned, so allocate the tensor that it was assigned to before
276  cell_state_out_ptr->allocator()->allocate();
277  _accum_cell_gate_bias.configure(compile_context, &_cell_layer_norm_out1, cell_bias, &_cell_layer_norm_out2, ConvertPolicy::SATURATE);
278  _cell_layer_norm_out1.allocator()->allocate();
279  cell_state_out_ptr = &_cell_layer_norm_out2;
280  }
281  _activation_cell_state.configure(compile_context, cell_state_out_ptr, nullptr, activation_info);
282  _memory_group.manage(&_cell_state_out5);
283  _pixelwise_mul_cell_state1.configure(compile_context, cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
284  cell_state_out_ptr->allocator()->allocate();
285  _pixelwise_mul_cell_state2.configure(compile_context, forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
286  _accum_cell_state2.configure(compile_context, &_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE);
287  _cell_state_out3.allocator()->allocate();
288  _cell_state_out5.allocator()->allocate();
289  // Perform clipping
290  if(cell_threshold != 0.f)
291  {
292  _perform_cell_clipping = true;
293  _cell_clip.configure(compile_context, &_cell_state_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold));
294  }
295 
296  // Configure block that calculates the output
297  // output_state_out = Activation(input * input_to_output_weights + output_state_in * recurrent_to_output_weights + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias)
298  // We optimize this as follows:
299  // output_state_out = Activation( (input,output_state_in) * (input_to_output_weights, recurrent_to_output_weights) + PixelWiseMul(cell_state, cell_to_output_weights) + output_gate_bias)
300  _output1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
301  _output4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
302  std::vector<const ICLTensor *> in_out_weights;
303  in_out_weights.emplace_back(input_to_output_weights);
304  in_out_weights.emplace_back(recurrent_to_output_weights);
305  TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
306  _output2.allocator()->init(TensorInfo(in_out_weights_concat_shape, 1, input->info()->data_type()));
307 
308  _concat_weights_output.configure(compile_context, in_out_weights, &_output2, Window::DimX);
309 
310  _memory_group.manage(&_output1);
311  _memory_group.manage(&_output4);
312 
313  _fully_connected_output.configure(compile_context, &_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, &_output4);
314 
315  _output2.allocator()->allocate();
316  _forget_gate_out2.allocator()->allocate();
317 
318  CLTensor *output_gate_out = &_output4;
319  if(lstm_params.has_peephole_opt())
320  {
321  _output3.allocator()->init(TensorInfo(_cell_state_out1.info()->tensor_shape(), 1, input->info()->data_type()));
322 
323  _memory_group.manage(&_output3);
324  _pixelwise_mul_output_state1.configure(compile_context, &_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
325  _accum_output1.configure(compile_context, &_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
326  _output4.allocator()->allocate();
327  output_gate_out = &_output1;
328 
329  // Allocate intermediate buffers
330  _output3.allocator()->allocate();
331  }
332  else
333  {
334  _output1.allocator()->allocate();
335  }
336  if(_is_layer_norm_lstm)
337  {
338  _output_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
339  _output_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
340  _memory_group.manage(&_output_layer_norm_out1);
341  _memory_group.manage(&_output_layer_norm_out2);
342  _mean_std_norm_output_gate.configure(compile_context, output_gate_out);
343  _pixelwise_mul_output_gate_coeff.configure(compile_context, output_gate_out, lstm_params.output_layer_norm_weights(), &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE,
345  // output_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
346  output_gate_out->allocator()->allocate();
347  _accum_output_gate_bias.configure(compile_context, &_output_layer_norm_out1, output_gate_bias, &_output_layer_norm_out2, ConvertPolicy::SATURATE);
348  _output_layer_norm_out1.allocator()->allocate();
349  output_gate_out = &_output_layer_norm_out2;
350  }
351  _activation_output.configure(compile_context, output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
352 
353  // Configure block that calculates the output state
354  /** lstm_res = PixelwiseMul(output, Activation(cell_state))
355  *
356  * -- Clip(lstm_res * projection_weights + projection_bias, projection_threshold) , if there is a projection
357  * /
358  * output_state = --
359  * \
360  * -- lstm_res , otherwise
361  */
362  ICLTensor *output_state_out_tmp = lstm_params.has_projection() ? &_output_state1 : output_state_out;
363  _cell_state_activation.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
364  _output_state1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
365 
366  _memory_group.manage(&_cell_state_activation);
367  _activation_output_state.configure(compile_context, &_cell_state_out1, &_cell_state_activation, activation_info);
368  _pixelwise_mul_output_state2.configure(compile_context, &_cell_state_activation, output_gate_out, output_state_out_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
369  _cell_state_activation.allocator()->allocate();
370 
371  if(lstm_params.has_projection())
372  {
373  _has_projection_weights = true;
374  _fully_connected_output_state.configure(compile_context, output_state_out_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out);
375  _output_state1.allocator()->allocate();
376  // Perform clipping
377  if(projection_threshold != 0.f)
378  {
379  _perform_projection_clipping = true;
380  _projection_clip.configure(compile_context, output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold));
381  }
382  }
383 
384  // Copy cell state and output
385  _copy_cell_state.configure(compile_context, &_cell_state_out1, cell_state_out);
386  _copy_output.configure(compile_context, output_state_out, output);
387 
388  // Vector for holding the tensors to store in scratch buffer
389  std::vector<const ICLTensor *> scratch_inputs;
390  if(!lstm_params.has_cifg_opt())
391  {
392  scratch_inputs.emplace_back(input_gate_out);
393  }
394  scratch_inputs.emplace_back(&_cell_state_out1);
395  scratch_inputs.emplace_back(forget_gate_out);
396  scratch_inputs.emplace_back(output_gate_out);
397  _concat_scratch_buffer.configure(compile_context, scratch_inputs, scratch_buffer, Window::DimX);
398  input_gate_out->allocator()->allocate();
399  _cell_state_out1.allocator()->allocate();
400  forget_gate_out->allocator()->allocate();
401  output_gate_out->allocator()->allocate();
402 }
403 
407  const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
408  const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
409  const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output,
410  const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
411 {
413  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
414  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
415  forget_gate_bias, cell_bias, output_gate_bias,
416  output_state_in, cell_state_in,
417  scratch_buffer, output_state_out, cell_state_out, output);
418 
419  // Check data types
422  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
423  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
424  forget_gate_bias, cell_bias, output_gate_bias,
425  output_state_in, cell_state_in,
426  scratch_buffer, output_state_out, cell_state_out, output);
427 
428  // Check dimensions
430  ARM_COMPUTE_RETURN_ERROR_ON(input_to_forget_weights->num_dimensions() > 2);
431  ARM_COMPUTE_RETURN_ERROR_ON(input_to_cell_weights->num_dimensions() > 2);
432  ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->num_dimensions() > 2);
433  ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_forget_weights->num_dimensions() > 2);
434  ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_cell_weights->num_dimensions() > 2);
435  ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->num_dimensions() > 2);
436  ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() > 1);
438  ARM_COMPUTE_RETURN_ERROR_ON(output_gate_bias->num_dimensions() > 1);
439  ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
440  ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() > 2);
441  ARM_COMPUTE_RETURN_ERROR_ON(scratch_buffer->num_dimensions() > 2);
442  ARM_COMPUTE_RETURN_ERROR_ON(output_state_out->num_dimensions() > 2);
443  ARM_COMPUTE_RETURN_ERROR_ON(cell_state_out->num_dimensions() > 2);
445  ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->dimension(0) * 4 != scratch_buffer->dimension(0)
446  && cell_bias->dimension(0) * 3 != scratch_buffer->dimension(0));
447 
448  const unsigned int num_batches = input->dimension(1);
449  const unsigned int num_cells = input_to_output_weights->dimension(1);
450 
451  if(lstm_params.use_layer_norm())
452  {
453  // If CIFG is used, input layer normalization weights tensor is omitted
454  if(lstm_params.has_cifg_opt())
455  {
456  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights() != nullptr);
457  }
458  else
459  {
461  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->num_dimensions() > 1);
462  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->dimension(0) != num_cells);
464  }
465 
468  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->num_dimensions() > 1);
469  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->num_dimensions() > 1);
470  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->num_dimensions() > 1);
471  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->dimension(0) != num_cells);
472  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->dimension(0) != num_cells);
473  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->dimension(0) != num_cells);
474  }
475 
476  // Check peephole optimization
477  if(lstm_params.has_peephole_opt())
478  {
480  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->num_dimensions() > 1);
481  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_output_weights()->num_dimensions() > 1);
482  }
483 
484  TensorShape units_out_transposed_shape = compute_transposed_shape(*recurrent_to_output_weights);
485  TensorShape num_units_transposed_shape = compute_transposed_shape(*forget_gate_bias);
486  const TensorInfo units_out_transposed_info = TensorInfo(units_out_transposed_shape, 1, input->data_type());
487  const TensorInfo num_units_transposed_info = TensorInfo(num_units_transposed_shape, 1, input->data_type());
488 
489  TensorInfo input_gate = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
490  TensorInfo forget_gate = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
491  TensorInfo output_gate_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
492  TensorInfo cell_state_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
493 
494  // Validate forget gate
495  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate));
496 
497  std::vector<const ITensorInfo *> inputs_vector;
498  inputs_vector.emplace_back(input);
499  inputs_vector.emplace_back(output_state_in);
501  TensorInfo forget_gate_concat = TensorInfo(concat_shape, 1, input->data_type());
502 
503  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector, &forget_gate_concat, Window::DimX));
504 
505  if(lstm_params.has_peephole_opt())
506  {
508  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
509  }
510  if(lstm_params.use_layer_norm())
511  {
515  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE));
516  }
518 
519  // Validate input gate
520  if(!lstm_params.has_cifg_opt())
521  {
523  lstm_params.recurrent_to_input_weights(),
524  lstm_params.input_gate_bias());
525  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_to_input_weights()->num_dimensions() > 2);
526  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.recurrent_to_input_weights()->num_dimensions() > 2);
527  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_gate_bias()->num_dimensions() > 1);
528 
529  std::vector<const ITensorInfo *> lstm_weights;
530  lstm_weights.emplace_back(lstm_params.input_to_input_weights());
531  lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
532  TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
533  TensorInfo lstm_gate_concat = TensorInfo(lstm_weights_concat_shape, 1, input->data_type());
534  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(lstm_weights, &lstm_gate_concat, Window::DimX));
535 
536  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, lstm_params.input_to_input_weights(), (lstm_params.use_layer_norm()) ? nullptr : lstm_params.input_gate_bias(), &input_gate));
537 
538  if(lstm_params.has_peephole_opt())
539  {
541  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1);
543  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
544  }
545 
546  if(lstm_params.use_layer_norm())
547  {
551  }
553  }
554  else
555  {
557  }
558 
559  // Validate cell state
560  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp));
561  ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo()));
562  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
563  if(lstm_params.use_layer_norm())
564  {
568  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE));
569  }
570  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, activation_info));
573  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
574  if(cell_threshold != 0.f)
575  {
577  cell_threshold)));
578  }
579 
580  std::vector<const ITensorInfo *> in_out_weights;
581  in_out_weights.emplace_back(input_to_output_weights);
582  in_out_weights.emplace_back(recurrent_to_output_weights);
583  TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
584  TensorInfo in_out_gate_concat = TensorInfo(in_out_weights_concat_shape, 1, input->data_type());
585  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(in_out_weights, &in_out_gate_concat, Window::DimX));
586  // Validate output gate tmp
587  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp));
588 
589  if(lstm_params.has_peephole_opt())
590  {
593  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
594  }
595  if(lstm_params.use_layer_norm())
596  {
600  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE));
601  }
603 
604  // Validate output state
605  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info));
607  if(lstm_params.has_projection())
608  {
609  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out));
610  if(projection_threshold != 0.f)
611  {
612  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output_state_out, output_state_out,
613  ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)));
614  }
615  }
616 
617  // Validate copy kernel
618  ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(&cell_state_tmp, cell_state_out));
619  ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(output_state_out, output));
620 
621  // Validate scratch concatenation
622  std::vector<const ITensorInfo *> inputs_vector_info_raw;
623  if(!lstm_params.has_cifg_opt())
624  {
625  inputs_vector_info_raw.push_back(&input_gate);
626  }
627  inputs_vector_info_raw.push_back(&cell_state_tmp);
628  inputs_vector_info_raw.push_back(&forget_gate);
629  inputs_vector_info_raw.push_back(&output_gate_tmp);
630 
631  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector_info_raw, scratch_buffer, Window::DimX));
632  return Status{};
633 }
634 
636 {
637  prepare();
638 
639  MemoryGroupResourceScope scope_mg(_memory_group);
640 
641  _concat_inputs_forget_gate.run();
642 
643  _fully_connected_forget_gate.run();
644 
645  if(_run_peephole_opt)
646  {
647  _pixelwise_mul_forget_gate.run();
648  _accum_forget_gate1.run();
649  }
650  if(_is_layer_norm_lstm)
651  {
652  _mean_std_norm_forget_gate.run();
653  _pixelwise_mul_forget_gate_coeff.run();
654  _accum_forget_gate_bias.run();
655  }
656  _activation_forget_gate.run();
657 
658  if(_run_cifg_opt)
659  {
660  _ones_fill.run();
661  _subtract_input_gate.run();
662  }
663  else
664  {
665  _fully_connected_input_gate.run();
666 
667  if(_run_peephole_opt)
668  {
669  _pixelwise_mul_input_gate.run();
670  _accum_input_gate1.run();
671  }
672 
673  if(_is_layer_norm_lstm)
674  {
675  _mean_std_norm_input_gate.run();
676  _pixelwise_mul_input_gate_coeff.run();
677  _accum_input_gate_bias.run();
678  }
679  _activation_input_gate.run();
680  }
681 
682  _fully_connected_cell_state.run();
683  CLScheduler::get().enqueue(*_transpose_cell_state);
684  _gemm_cell_state1.run();
685  _accum_cell_state1.run();
686  if(_is_layer_norm_lstm)
687  {
688  _mean_std_norm_cell_gate.run();
689  _pixelwise_mul_cell_gate_coeff.run();
690  _accum_cell_gate_bias.run();
691  }
692  _activation_cell_state.run();
693  _pixelwise_mul_cell_state1.run();
694  _pixelwise_mul_cell_state2.run();
695  _accum_cell_state2.run();
696 
697  if(_perform_cell_clipping)
698  {
699  _cell_clip.run();
700  }
701 
702  _fully_connected_output.run();
703 
704  if(_run_peephole_opt)
705  {
706  _pixelwise_mul_output_state1.run();
707  _accum_output1.run();
708  }
709  if(_is_layer_norm_lstm)
710  {
711  _mean_std_norm_output_gate.run();
712  _pixelwise_mul_output_gate_coeff.run();
713  _accum_output_gate_bias.run();
714  }
715  _activation_output.run();
716 
717  _activation_output_state.run();
718  _pixelwise_mul_output_state2.run();
719 
720  if(_has_projection_weights)
721  {
722  _fully_connected_output_state.run();
723  if(_perform_projection_clipping)
724  {
725  _projection_clip.run();
726  }
727  }
728 
729  _copy_cell_state.run();
730  _copy_output.run();
731 
732  _concat_scratch_buffer.run();
733 }
734 
736 {
737  if(!_is_prepared)
738  {
739  _concat_weights_forget_gate.run();
740  if(!_run_cifg_opt)
741  {
742  _concat_weights_input_gate.run();
743  }
744  _concat_weights_output.run();
745  _is_prepared = true;
746  }
747 }
748 } // namespace arm_compute
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
Static function to check if given info will lead to a valid configuration of CLActivationLayer.
const T * projection_weights() const
Definition: LSTMParams.h:227
const T * input_to_input_weights() const
Definition: LSTMParams.h:197
Shape of a tensor.
Definition: TensorShape.h:39
void configure(const ICLTensor *input, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams< ICLTensor > &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold=0.f, float projection_threshold=0.f)
Initialize function&#39;s tensors.
Definition: CLLSTMLayer.cpp:69
bool use_layer_norm() const
Definition: LSTMParams.h:312
TensorInfo * info() const override
Interface to be implemented by the child class to return the tensor&#39;s metadata.
Definition: CLTensor.cpp:41
void run() override
Run the kernels contained in the function.
static Status validate(const ITensorInfo *input, const ITensorInfo *output=nullptr, float epsilon=1e-8f)
Static function to check if given info will lead to a valid configuration of CLMeanStdDevNormalizatio...
void run() override
Run the kernels contained in the function.
Definition: CLGEMM.cpp:778
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
bool has_peephole_opt() const
Definition: LSTMParams.h:297
static CLScheduler & get()
Access the scheduler singleton.
T * forget_layer_norm_weights() const
Definition: LSTMParams.h:242
void build_lstm_params_tensor_info(const LSTMParams< T > &lstm_params, LSTMParams< ITensorInfo > *lstm_params_info)
Build LSTMParams<ITensorInfo> object by extracting the metadata from each tensor. ...
Definition: InfoHelpers.h:71
void run() override
Run the kernels contained in the function.
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
void run() override
Run the kernels contained in the function.
Definition: CLCopy.cpp:73
1 channel, 1 F32 per channel
void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel&#39;s inputs, output and convertion policy.
bool has_cifg_opt() const
Definition: LSTMParams.h:307
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
void run() override
Run the kernels contained in the function.
CLTensorAllocator * allocator()
Return a pointer to the tensor&#39;s allocator.
Definition: CLTensor.cpp:61
void configure(ICLTensor *input, ICLTensor *output=nullptr, float epsilon=1e-8f)
Initialise the function&#39;s input and outputs.
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
CLLSTMLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
Definition: CLLSTMLayer.cpp:51
T * cell_to_input_weights() const
Definition: LSTMParams.h:207
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
void configure(std::vector< const ICLTensor *> &inputs_vector, ICLTensor *output, size_t axis)
Initialise the kernel&#39;s inputs vector and output.
void run() override
Run the kernels contained in the function.
void init(const TensorInfo &input, size_t alignment=0)
Initialize a tensor based on the passed TensorInfo.
Copyright (c) 2017-2021 Arm Limited.
~CLLSTMLayer()
Default destructor.
1 channel, 1 F16 per channel
TensorShape compute_transposed_shape(const ITensorInfo &input)
Calculate the transposed shape of a tensor.
DataType data_type() const override
Data type used for each element of the tensor.
Definition: TensorInfo.h:270
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
const T * recurrent_to_input_weights() const
Definition: LSTMParams.h:202
const T * projection_bias() const
Definition: LSTMParams.h:232
Interface to enqueue OpenCL kernels and get/set the OpenCL CommandQueue and ICLTuner.
void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Set the input and output tensors.
void run() override
Run the kernels contained in the function.
Definition: CLFill.cpp:72
T * output_layer_norm_weights() const
Definition: LSTMParams.h:252
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
void run() override final
Run the kernels contained in the function.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
void configure(ICLTensor *input, ICLTensor *output, Window *dst_window=nullptr)
Initialise the function&#39;s source and destination.
Definition: CLCopy.cpp:52
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Static function to check if given info will lead to a valid configuration of CLFullyConnectedLayer.
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 opencl::kernels::ClSatur...
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
void configure(ICLTensor *input1, ICLTensor *input2, ICLTensor *output, ConvertPolicy policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel&#39;s inputs, output and conversion policy.
static Status validate(const ITensorInfo *input, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias, const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in, const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output, const LSTMParams< ITensorInfo > &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold=0.f, float projection_threshold=0.f)
Static function to check if given info will lead to a valid configuration of CLLSTMLayer.
void run() override
Run the kernels contained in the function.
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 opencl::kernels::ClSatur...
void enqueue(ICLKernel &kernel, bool flush=true)
Schedule the execution of the passed kernel if possible.
void configure(ICLTensor *tensor, const PixelValue &constant_value, Window *window=nullptr)
Initialize the kernel&#39;s tensor and filling value.
Definition: CLFill.cpp:52
CLCompileContext class.
T * cell_to_forget_weights() const
Definition: LSTMParams.h:217
static Status validate(const ITensorInfo *input, const ITensorInfo *output, Window *dst_window=nullptr)
Static function to check if given info will lead to a valid configuration of CLCopy.
Definition: CLCopy.cpp:68
bool has_projection() const
Definition: LSTMParams.h:302
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 CLConcatenateLayer.
void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, ConvertPolicy policy, const ActivationLayerInfo &act_info=ActivationLayerInfo())
Initialise the kernel&#39;s inputs, output and conversion policy.
void allocate() override
Allocate size specified by TensorInfo of OpenCL memory.
T * cell_to_output_weights() const
Definition: LSTMParams.h:222
Rounds to nearest value; half rounds to nearest even.
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
T * input_layer_norm_weights() const
Definition: LSTMParams.h:237
void configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel&#39;s inputs and output.
Definition: CLGEMM.cpp:666
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
const T * input_gate_bias() const
Definition: LSTMParams.h:212
void configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
Set the input and output tensor.
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of CLGEMM.
Definition: CLGEMM.cpp:727
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
GEMM information class.
Definition: Types.h:2003
T * cell_layer_norm_weights() const
Definition: LSTMParams.h:247
OpenCL kernel which transposes the elements of a matrix.
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:262
void run() override
Run the kernels contained in the function.
TensorShape calculate_concatenate_shape(const std::vector< T *> &input, size_t axis)
Calculate the concatenate output shape of the concatenate operation along a single axis...
void run() override
Run the kernels contained in the function.
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 CLPixelWiseMultiplicatio...
void prepare() override
Prepare the function for executing.
Basic implementation of the OpenCL tensor interface.
Definition: CLTensor.h:41