Compute Library
 21.08
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"
34 
35 namespace arm_compute
36 {
38 using namespace arm_compute::utils::info_helpers;
39 
40 CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
41  : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(),
42  _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(),
43  _transpose_cell_state(std::make_unique<opencl::kernels::ClTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
44  _pixelwise_mul_cell_state2(), _fully_connected_output(), _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(),
45  _fully_connected_output_state(), _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(),
46  _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(),
47  _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(),
48  _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(),
49  _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(),
50  _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(),
51  _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),
52  _has_projection_weights(false), _perform_projection_clipping(false), _is_prepared(false), _is_layer_norm_lstm(false)
53 {
54 }
55 
56 CLLSTMLayer::~CLLSTMLayer() = default;
57 
61  const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
62  const ICLTensor *output_state_in, ICLTensor *cell_state_in,
63  ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
64  const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
65 {
66  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,
67  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,
68  cell_threshold, projection_threshold);
69 }
70 
71 void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input,
74  const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
75  const ICLTensor *output_state_in, ICLTensor *cell_state_in,
76  ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
77  const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
78 {
80  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
81  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
82  forget_gate_bias, cell_bias, output_gate_bias,
83  output_state_in, cell_state_in,
84  scratch_buffer, output_state_out, cell_state_out, output);
85 
86  _is_layer_norm_lstm = lstm_params.use_layer_norm();
87 
88  // Set lstm parameters
89  LSTMParams<ITensorInfo> lstm_params_info{};
90  build_lstm_params_tensor_info(lstm_params, &lstm_params_info);
91 
92  // Validate
93  ARM_COMPUTE_ERROR_THROW_ON(CLLSTMLayer::validate(input->info(), input_to_forget_weights->info(),
94  input_to_cell_weights->info(), input_to_output_weights->info(),
95  recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
96  forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(),
97  output_state_in->info(), cell_state_in->info(),
98  scratch_buffer->info(), output_state_out->info(), cell_state_out->info(), output->info(),
99  lstm_params_info, activation_info, cell_threshold, projection_threshold));
100 
101  const TensorShape cell_state_shape = cell_state_in->info()->tensor_shape();
102  // Configure block that calculates the forget gate
103  // 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)
104  // We optimize this as follows:
105  // 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
106  _forget_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
107  _forget_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
108  _forget_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
109 
110  std::vector<const ICLTensor *> inputs_vector;
111  inputs_vector.emplace_back(input);
112  inputs_vector.emplace_back(output_state_in);
114  _forget_gate_out2.allocator()->init(TensorInfo(concat_shape, 1, input->info()->data_type()));
115 
116  _memory_group.manage(&_forget_gate_out2);
117  _concat_inputs_forget_gate.configure(compile_context, inputs_vector, &_forget_gate_out2, Window::DimX);
118 
119  std::vector<const ICLTensor *> weights_vector;
120 
121  weights_vector.emplace_back(input_to_forget_weights);
122  weights_vector.emplace_back(recurrent_to_forget_weights);
123  const TensorShape weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(weights_vector, 0);
124  _forget_gate_out6.allocator()->init(TensorInfo(weights_concat_shape, 1, input->info()->data_type()));
125 
126  _concat_weights_forget_gate.configure(compile_context, weights_vector, &_forget_gate_out6, Window::DimX);
127 
128  _memory_group.manage(&_forget_gate_out5);
129  _fully_connected_forget_gate.configure(compile_context, &_forget_gate_out2, &_forget_gate_out6, (_is_layer_norm_lstm) ? nullptr : forget_gate_bias, &_forget_gate_out5);
130  _memory_group.manage(&_forget_gate_out1);
131  _memory_group.manage(&_forget_gate_out3);
132  _forget_gate_out6.allocator()->allocate();
133 
134  CLTensor *forget_gate_out = &_forget_gate_out5;
135  if(lstm_params.has_peephole_opt())
136  {
137  _forget_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
138 
139  _run_peephole_opt = true;
140  _memory_group.manage(&_forget_gate_out4);
141  _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);
142  _accum_forget_gate1.configure(compile_context, &_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE);
143  _forget_gate_out4.allocator()->allocate();
144  _forget_gate_out5.allocator()->allocate();
145  forget_gate_out = &_forget_gate_out3;
146  }
147  else
148  {
149  _forget_gate_out3.allocator()->allocate();
150  }
151  if(_is_layer_norm_lstm)
152  {
153  _forget_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
154  _forget_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
155  _memory_group.manage(&_forget_layer_norm_out1);
156  _memory_group.manage(&_forget_layer_norm_out2);
157  _mean_std_norm_forget_gate.configure(compile_context, forget_gate_out);
158  _pixelwise_mul_forget_gate_coeff.configure(compile_context, forget_gate_out, lstm_params.forget_layer_norm_weights(), &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE,
160  // forget_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
161  forget_gate_out->allocator()->allocate();
162  _accum_forget_gate_bias.configure(compile_context, &_forget_layer_norm_out1, forget_gate_bias, &_forget_layer_norm_out2, ConvertPolicy::SATURATE);
163  _forget_layer_norm_out1.allocator()->allocate();
164  forget_gate_out = &_forget_layer_norm_out2;
165  }
166  _activation_forget_gate.configure(compile_context, forget_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
167 
168  // Configure block that calculates the input gate
169  // 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
170  // input_gate = 1 - forget_gate, with CIFG
171  // We optimize this as follows:
172  // 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
173  _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
174  CLTensor *input_gate_out = &_input_gate_out1;
175  if(lstm_params.has_cifg_opt())
176  {
177  _memory_group.manage(&_input_gate_out1);
178  _ones.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
179  _ones_fill.configure(compile_context, &_ones, PixelValue(1, _ones.info()->data_type()));
180  _subtract_input_gate.configure(compile_context, &_ones, forget_gate_out, &_input_gate_out1, ConvertPolicy::SATURATE);
181  _ones.allocator()->allocate();
182  _run_cifg_opt = true;
183  }
184  else
185  {
186  _input_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
187  _input_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
188 
189  std::vector<const ICLTensor *> lstm_weights;
190  lstm_weights.emplace_back(lstm_params.input_to_input_weights());
191  lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
192  TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
193  _input_gate_out2.allocator()->init(TensorInfo(lstm_weights_concat_shape, 1, input->info()->data_type()));
194 
195  _concat_weights_input_gate.configure(compile_context, lstm_weights, &_input_gate_out2, Window::DimX);
196 
197  _memory_group.manage(&_input_gate_out1);
198 
199  _memory_group.manage(&_input_gate_out3);
200  _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);
201  _input_gate_out2.allocator()->allocate();
202 
203  input_gate_out = &_input_gate_out3;
204  if(_run_peephole_opt)
205  {
206  _memory_group.manage(&_input_gate_out4);
207  _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);
208  _accum_input_gate1.configure(compile_context, &_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE);
209  _input_gate_out3.allocator()->allocate();
210  _input_gate_out4.allocator()->allocate();
211  input_gate_out = &_input_gate_out1;
212  }
213  else
214  {
215  _input_gate_out1.allocator()->allocate();
216  }
217 
218  if(_is_layer_norm_lstm)
219  {
220  _input_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
221  _input_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
222  _memory_group.manage(&_input_layer_norm_out1);
223  _memory_group.manage(&_input_layer_norm_out2);
224  _mean_std_norm_input_gate.configure(compile_context, input_gate_out);
225  _pixelwise_mul_input_gate_coeff.configure(compile_context, input_gate_out, lstm_params.input_layer_norm_weights(), &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE,
227  // input_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
228  input_gate_out->allocator()->allocate();
229  _accum_input_gate_bias.configure(compile_context, &_input_layer_norm_out1, lstm_params.input_gate_bias(), &_input_layer_norm_out2, ConvertPolicy::SATURATE);
230  _input_layer_norm_out1.allocator()->allocate();
231  input_gate_out = &_input_layer_norm_out2;
232  }
233  _activation_input_gate.configure(compile_context, input_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
234  }
235 
236  // Configure block that calculates the cell state
237  // 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)
238  TensorShape cell_state1_shape = compute_transposed_shape(*recurrent_to_output_weights->info());
239  _cell_state_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
240  _cell_state_out2.allocator()->init(TensorInfo(cell_state1_shape, 1, input->info()->data_type()));
241  _cell_state_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
242  _cell_state_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
243  _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
244 
245  _memory_group.manage(&_cell_state_out1);
246  _fully_connected_cell_state.configure(compile_context, input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1);
247  _memory_group.manage(&_cell_state_out2);
248  _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights->info(), _cell_state_out2.info());
249  _recurrent_to_cell_weights = recurrent_to_cell_weights;
250  _memory_group.manage(&_cell_state_out3);
251  _gemm_cell_state1.configure(compile_context, output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
252  _cell_state_out2.allocator()->allocate();
253  _memory_group.manage(&_cell_state_out4);
254  _accum_cell_state1.configure(compile_context, &_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE);
255  CLTensor *cell_state_out_ptr = &_cell_state_out4;
256  if(_is_layer_norm_lstm)
257  {
258  _cell_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
259  _cell_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
260  _memory_group.manage(&_cell_layer_norm_out1);
261  _memory_group.manage(&_cell_layer_norm_out2);
262  _mean_std_norm_cell_gate.configure(compile_context, cell_state_out_ptr);
263  _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,
265  // cell_state_out_ptr is going to be reassigned, so allocate the tensor that it was assigned to before
266  cell_state_out_ptr->allocator()->allocate();
267  _accum_cell_gate_bias.configure(compile_context, &_cell_layer_norm_out1, cell_bias, &_cell_layer_norm_out2, ConvertPolicy::SATURATE);
268  _cell_layer_norm_out1.allocator()->allocate();
269  cell_state_out_ptr = &_cell_layer_norm_out2;
270  }
271  _activation_cell_state.configure(compile_context, cell_state_out_ptr, nullptr, activation_info);
272  _memory_group.manage(&_cell_state_out5);
273  _pixelwise_mul_cell_state1.configure(compile_context, cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
274  cell_state_out_ptr->allocator()->allocate();
275  _pixelwise_mul_cell_state2.configure(compile_context, forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
276  _accum_cell_state2.configure(compile_context, &_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE);
277  _cell_state_out3.allocator()->allocate();
278  _cell_state_out5.allocator()->allocate();
279  // Perform clipping
280  if(cell_threshold != 0.f)
281  {
282  _perform_cell_clipping = true;
283  _cell_clip.configure(compile_context, &_cell_state_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold));
284  }
285 
286  // Configure block that calculates the output
287  // 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)
288  // We optimize this as follows:
289  // 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)
290  _output1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
291  _output4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
292  std::vector<const ICLTensor *> in_out_weights;
293  in_out_weights.emplace_back(input_to_output_weights);
294  in_out_weights.emplace_back(recurrent_to_output_weights);
295  TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
296  _output2.allocator()->init(TensorInfo(in_out_weights_concat_shape, 1, input->info()->data_type()));
297 
298  _concat_weights_output.configure(compile_context, in_out_weights, &_output2, Window::DimX);
299 
300  _memory_group.manage(&_output1);
301  _memory_group.manage(&_output4);
302 
303  _fully_connected_output.configure(compile_context, &_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, &_output4);
304 
305  _output2.allocator()->allocate();
306  _forget_gate_out2.allocator()->allocate();
307 
308  CLTensor *output_gate_out = &_output4;
309  if(lstm_params.has_peephole_opt())
310  {
311  _output3.allocator()->init(TensorInfo(_cell_state_out1.info()->tensor_shape(), 1, input->info()->data_type()));
312 
313  _memory_group.manage(&_output3);
314  _pixelwise_mul_output_state1.configure(compile_context, &_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
315  _accum_output1.configure(compile_context, &_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
316  _output4.allocator()->allocate();
317  output_gate_out = &_output1;
318 
319  // Allocate intermediate buffers
320  _output3.allocator()->allocate();
321  }
322  else
323  {
324  _output1.allocator()->allocate();
325  }
326  if(_is_layer_norm_lstm)
327  {
328  _output_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
329  _output_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
330  _memory_group.manage(&_output_layer_norm_out1);
331  _memory_group.manage(&_output_layer_norm_out2);
332  _mean_std_norm_output_gate.configure(compile_context, output_gate_out);
333  _pixelwise_mul_output_gate_coeff.configure(compile_context, output_gate_out, lstm_params.output_layer_norm_weights(), &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE,
335  // output_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
336  output_gate_out->allocator()->allocate();
337  _accum_output_gate_bias.configure(compile_context, &_output_layer_norm_out1, output_gate_bias, &_output_layer_norm_out2, ConvertPolicy::SATURATE);
338  _output_layer_norm_out1.allocator()->allocate();
339  output_gate_out = &_output_layer_norm_out2;
340  }
341  _activation_output.configure(compile_context, output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
342 
343  // Configure block that calculates the output state
344  /** lstm_res = PixelwiseMul(output, Activation(cell_state))
345  *
346  * -- Clip(lstm_res * projection_weights + projection_bias, projection_threshold) , if there is a projection
347  * /
348  * output_state = --
349  * \
350  * -- lstm_res , otherwise
351  */
352  ICLTensor *output_state_out_tmp = lstm_params.has_projection() ? &_output_state1 : output_state_out;
353  _cell_state_activation.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
354  _output_state1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
355 
356  _memory_group.manage(&_cell_state_activation);
357  _activation_output_state.configure(compile_context, &_cell_state_out1, &_cell_state_activation, activation_info);
358  _pixelwise_mul_output_state2.configure(compile_context, &_cell_state_activation, output_gate_out, output_state_out_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
359  _cell_state_activation.allocator()->allocate();
360 
361  if(lstm_params.has_projection())
362  {
363  _has_projection_weights = true;
364  _fully_connected_output_state.configure(compile_context, output_state_out_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out);
365  _output_state1.allocator()->allocate();
366  // Perform clipping
367  if(projection_threshold != 0.f)
368  {
369  _perform_projection_clipping = true;
370  _projection_clip.configure(compile_context, output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold));
371  }
372  }
373 
374  // Copy cell state and output
375  _copy_cell_state.configure(compile_context, &_cell_state_out1, cell_state_out);
376  _copy_output.configure(compile_context, output_state_out, output);
377 
378  // Vector for holding the tensors to store in scratch buffer
379  std::vector<const ICLTensor *> scratch_inputs;
380  if(!lstm_params.has_cifg_opt())
381  {
382  scratch_inputs.emplace_back(input_gate_out);
383  }
384  scratch_inputs.emplace_back(&_cell_state_out1);
385  scratch_inputs.emplace_back(forget_gate_out);
386  scratch_inputs.emplace_back(output_gate_out);
387  _concat_scratch_buffer.configure(compile_context, scratch_inputs, scratch_buffer, Window::DimX);
388  input_gate_out->allocator()->allocate();
389  _cell_state_out1.allocator()->allocate();
390  forget_gate_out->allocator()->allocate();
391  output_gate_out->allocator()->allocate();
392 }
393 
397  const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
398  const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
399  const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output,
400  const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
401 {
403  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
404  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
405  forget_gate_bias, cell_bias, output_gate_bias,
406  output_state_in, cell_state_in,
407  scratch_buffer, output_state_out, cell_state_out, output);
408 
409  // Check data types
412  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
413  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
414  forget_gate_bias, cell_bias, output_gate_bias,
415  output_state_in, cell_state_in,
416  scratch_buffer, output_state_out, cell_state_out, output);
417 
418  // Check dimensions
420  ARM_COMPUTE_RETURN_ERROR_ON(input_to_forget_weights->num_dimensions() > 2);
421  ARM_COMPUTE_RETURN_ERROR_ON(input_to_cell_weights->num_dimensions() > 2);
422  ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->num_dimensions() > 2);
423  ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_forget_weights->num_dimensions() > 2);
424  ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_cell_weights->num_dimensions() > 2);
425  ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->num_dimensions() > 2);
426  ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() > 1);
428  ARM_COMPUTE_RETURN_ERROR_ON(output_gate_bias->num_dimensions() > 1);
429  ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
430  ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() > 2);
431  ARM_COMPUTE_RETURN_ERROR_ON(scratch_buffer->num_dimensions() > 2);
432  ARM_COMPUTE_RETURN_ERROR_ON(output_state_out->num_dimensions() > 2);
433  ARM_COMPUTE_RETURN_ERROR_ON(cell_state_out->num_dimensions() > 2);
435  ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->dimension(0) * 4 != scratch_buffer->dimension(0)
436  && cell_bias->dimension(0) * 3 != scratch_buffer->dimension(0));
437 
438  const unsigned int num_batches = input->dimension(1);
439  const unsigned int num_cells = input_to_output_weights->dimension(1);
440 
441  if(lstm_params.use_layer_norm())
442  {
443  // If CIFG is used, input layer normalization weights tensor is omitted
444  if(lstm_params.has_cifg_opt())
445  {
446  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights() != nullptr);
447  }
448  else
449  {
451  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->num_dimensions() > 1);
452  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->dimension(0) != num_cells);
454  }
455 
458  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->num_dimensions() > 1);
459  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->num_dimensions() > 1);
460  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->num_dimensions() > 1);
461  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->dimension(0) != num_cells);
462  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->dimension(0) != num_cells);
463  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->dimension(0) != num_cells);
464  }
465 
466  // Check peephole optimization
467  if(lstm_params.has_peephole_opt())
468  {
470  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->num_dimensions() > 1);
471  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_output_weights()->num_dimensions() > 1);
472  }
473 
474  TensorShape units_out_transposed_shape = compute_transposed_shape(*recurrent_to_output_weights);
475  TensorShape num_units_transposed_shape = compute_transposed_shape(*forget_gate_bias);
476  const TensorInfo units_out_transposed_info = TensorInfo(units_out_transposed_shape, 1, input->data_type());
477  const TensorInfo num_units_transposed_info = TensorInfo(num_units_transposed_shape, 1, input->data_type());
478 
479  TensorInfo input_gate = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
480  TensorInfo forget_gate = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
481  TensorInfo output_gate_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
482  TensorInfo cell_state_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
483 
484  // Validate forget gate
485  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate));
486 
487  std::vector<const ITensorInfo *> inputs_vector;
488  inputs_vector.emplace_back(input);
489  inputs_vector.emplace_back(output_state_in);
491  TensorInfo forget_gate_concat = TensorInfo(concat_shape, 1, input->data_type());
492 
493  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector, &forget_gate_concat, Window::DimX));
494 
495  if(lstm_params.has_peephole_opt())
496  {
498  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
499  }
500  if(lstm_params.use_layer_norm())
501  {
505  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE));
506  }
508 
509  // Validate input gate
510  if(!lstm_params.has_cifg_opt())
511  {
513  lstm_params.recurrent_to_input_weights(),
514  lstm_params.input_gate_bias());
515  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_to_input_weights()->num_dimensions() > 2);
516  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.recurrent_to_input_weights()->num_dimensions() > 2);
517  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_gate_bias()->num_dimensions() > 1);
518 
519  std::vector<const ITensorInfo *> lstm_weights;
520  lstm_weights.emplace_back(lstm_params.input_to_input_weights());
521  lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
522  TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
523  TensorInfo lstm_gate_concat = TensorInfo(lstm_weights_concat_shape, 1, input->data_type());
524  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(lstm_weights, &lstm_gate_concat, Window::DimX));
525 
526  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));
527 
528  if(lstm_params.has_peephole_opt())
529  {
531  ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1);
533  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
534  }
535 
536  if(lstm_params.use_layer_norm())
537  {
541  }
543  }
544  else
545  {
547  }
548 
549  // Validate cell state
550  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp));
551  ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo()));
552  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
553  if(lstm_params.use_layer_norm())
554  {
558  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE));
559  }
560  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, activation_info));
563  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
564  if(cell_threshold != 0.f)
565  {
567  cell_threshold)));
568  }
569 
570  std::vector<const ITensorInfo *> in_out_weights;
571  in_out_weights.emplace_back(input_to_output_weights);
572  in_out_weights.emplace_back(recurrent_to_output_weights);
573  TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
574  TensorInfo in_out_gate_concat = TensorInfo(in_out_weights_concat_shape, 1, input->data_type());
575  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(in_out_weights, &in_out_gate_concat, Window::DimX));
576  // Validate output gate tmp
577  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp));
578 
579  if(lstm_params.has_peephole_opt())
580  {
583  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
584  }
585  if(lstm_params.use_layer_norm())
586  {
590  ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE));
591  }
593 
594  // Validate output state
595  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info));
597  if(lstm_params.has_projection())
598  {
599  ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out));
600  if(projection_threshold != 0.f)
601  {
602  ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output_state_out, output_state_out,
603  ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)));
604  }
605  }
606 
607  // Validate copy kernel
608  ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(&cell_state_tmp, cell_state_out));
609  ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(output_state_out, output));
610 
611  // Validate scratch concatenation
612  std::vector<const ITensorInfo *> inputs_vector_info_raw;
613  if(!lstm_params.has_cifg_opt())
614  {
615  inputs_vector_info_raw.push_back(&input_gate);
616  }
617  inputs_vector_info_raw.push_back(&cell_state_tmp);
618  inputs_vector_info_raw.push_back(&forget_gate);
619  inputs_vector_info_raw.push_back(&output_gate_tmp);
620 
621  ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector_info_raw, scratch_buffer, Window::DimX));
622  return Status{};
623 }
624 
626 {
627  prepare();
628 
629  MemoryGroupResourceScope scope_mg(_memory_group);
630 
631  _concat_inputs_forget_gate.run();
632 
633  _fully_connected_forget_gate.run();
634 
635  if(_run_peephole_opt)
636  {
637  _pixelwise_mul_forget_gate.run();
638  _accum_forget_gate1.run();
639  }
640  if(_is_layer_norm_lstm)
641  {
642  _mean_std_norm_forget_gate.run();
643  _pixelwise_mul_forget_gate_coeff.run();
644  _accum_forget_gate_bias.run();
645  }
646  _activation_forget_gate.run();
647 
648  if(_run_cifg_opt)
649  {
650  _ones_fill.run();
651  _subtract_input_gate.run();
652  }
653  else
654  {
655  _fully_connected_input_gate.run();
656 
657  if(_run_peephole_opt)
658  {
659  _pixelwise_mul_input_gate.run();
660  _accum_input_gate1.run();
661  }
662 
663  if(_is_layer_norm_lstm)
664  {
665  _mean_std_norm_input_gate.run();
666  _pixelwise_mul_input_gate_coeff.run();
667  _accum_input_gate_bias.run();
668  }
669  _activation_input_gate.run();
670  }
671 
672  _fully_connected_cell_state.run();
674  pack.add_tensor(TensorType::ACL_SRC, _recurrent_to_cell_weights);
675  pack.add_tensor(TensorType::ACL_DST, &_cell_state_out2);
676  CLScheduler::get().enqueue_op(*_transpose_cell_state,
677  pack,
678  false);
679  _gemm_cell_state1.run();
680  _accum_cell_state1.run();
681  if(_is_layer_norm_lstm)
682  {
683  _mean_std_norm_cell_gate.run();
684  _pixelwise_mul_cell_gate_coeff.run();
685  _accum_cell_gate_bias.run();
686  }
687  _activation_cell_state.run();
688  _pixelwise_mul_cell_state1.run();
689  _pixelwise_mul_cell_state2.run();
690  _accum_cell_state2.run();
691 
692  if(_perform_cell_clipping)
693  {
694  _cell_clip.run();
695  }
696 
697  _fully_connected_output.run();
698 
699  if(_run_peephole_opt)
700  {
701  _pixelwise_mul_output_state1.run();
702  _accum_output1.run();
703  }
704  if(_is_layer_norm_lstm)
705  {
706  _mean_std_norm_output_gate.run();
707  _pixelwise_mul_output_gate_coeff.run();
708  _accum_output_gate_bias.run();
709  }
710  _activation_output.run();
711 
712  _activation_output_state.run();
713  _pixelwise_mul_output_state2.run();
714 
715  if(_has_projection_weights)
716  {
717  _fully_connected_output_state.run();
718  if(_perform_projection_clipping)
719  {
720  _projection_clip.run();
721  }
722  }
723 
724  _copy_cell_state.run();
725  _copy_output.run();
726 
727  _concat_scratch_buffer.run();
728 }
729 
731 {
732  if(!_is_prepared)
733  {
734  _concat_weights_forget_gate.run();
735  if(!_run_cifg_opt)
736  {
737  _concat_weights_input_gate.run();
738  }
739  _concat_weights_output.run();
740  _is_prepared = true;
741  }
742 }
743 } // namespace arm_compute
OpenCL kernel to transpose a 2D tensor.
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:225
const T * input_to_input_weights() const
Definition: LSTMParams.h:195
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:58
bool use_layer_norm() const
Definition: LSTMParams.h:310
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:100
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
bool has_peephole_opt() const
Definition: LSTMParams.h:295
static CLScheduler & get()
Access the scheduler singleton.
T * forget_layer_norm_weights() const
Definition: LSTMParams.h:240
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:305
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:40
T * cell_to_input_weights() const
Definition: LSTMParams.h:205
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:1475
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:242
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
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:200
const T * projection_bias() const
Definition: LSTMParams.h:230
Interface to enqueue OpenCL kernels and get/set the OpenCL CommandQueue and ICLTuner.
void run() override
Run the kernels contained in the function.
Definition: CLFill.cpp:72
T * output_layer_norm_weights() const
Definition: LSTMParams.h:250
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
void configure(const CLCompileContext &compile_context, 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:69
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...
void enqueue_op(ICLKernel &kernel, ITensorPack &tensors, bool flush=true)
Schedule the execution of the passed kernel if possible.
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 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:215
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
void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Set the input and output tensors.
bool has_projection() const
Definition: LSTMParams.h:300
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:220
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:235
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
const T * input_gate_bias() const
Definition: LSTMParams.h:210
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:95
Tensor packing service.
Definition: ITensorPack.h:39
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:43
GEMM information class.
Definition: Types.h:1939
T * cell_layer_norm_weights() const
Definition: LSTMParams.h:245
const TensorShape & tensor_shape() const override
Size for each dimension of the tensor.
Definition: TensorInfo.h:234
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 add_tensor(int id, ITensor *tensor)
Add tensor to the pack.
Definition: ITensorPack.cpp:39
void prepare() override
Prepare the function for executing.
Basic implementation of the OpenCL tensor interface.
Definition: CLTensor.h:41