Compute Library
 21.02
NEQLSTMLayer.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2020-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef ARM_COMPUTE_NEQLSTMLAYER_H
25 #define ARM_COMPUTE_NEQLSTMLAYER_H
26 
27 #include "arm_compute/core/Types.h"
37 
38 #include <memory>
39 
40 namespace arm_compute
41 {
42 // Forward declarations
43 class ITensor;
44 class ITensorInfo;
45 class NEQLSTMLayerNormalizationKernel;
46 class NEGEMMLowpMatrixAReductionKernel;
47 
48 /** Basic function to run @ref NEQLSTMLayer
49  *
50  * This function calls the following Neon functions/kernels:
51  *
52  * -# @ref NEActivationLayer Activation functions (tanh and logistic)
53  * -# @ref NEArithmeticAddition Elementwise addition
54  * -# @ref NEArithmeticSubtraction Elementwise subtraction
55  * -# @ref NECopy Copy kernel for copying output_state_out to output
56  * -# @ref NEGEMMLowpMatrixMultiplyCore Quantized matrix multiplication core. Accumulators are 32-bit integers
57  * -# @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint Convert 32-bit integers into QSYMM16
58  * -# @ref NEGEMMLowpMatrixAReductionKernel For precomputing effective biases to use
59  * -# @ref NEPixelWiseMultiplication Elementwise multiplication
60  * -# @ref NETranspose Transpose function for reshaping the weights
61  * */
62 class NEQLSTMLayer : public IFunction
63 {
64 public:
65  /** Default constructor */
66  NEQLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
67  /** Prevent instances of this class from being copied (As this class contains pointers) */
68  NEQLSTMLayer(const NEQLSTMLayer &) = delete;
69  /** Prevent instances of this class from being moved (As this class contains pointers) */
70  NEQLSTMLayer(NEQLSTMLayer &&) = delete;
71  /** Prevent instances of this class from being copied (As this class contains pointers) */
72  NEQLSTMLayer &operator=(const NEQLSTMLayer &) = delete;
73  /** Prevent instances of this class from being moved (As this class contains pointers) */
74  NEQLSTMLayer &operator=(NEQLSTMLayer &&) = delete;
75  /** Default destructor */
76  ~NEQLSTMLayer();
77  /** Initialize function's tensors.
78  *
79  * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED.
80  * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
81  * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
82  * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
83  * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
84  * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
85  * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
86  * @param[in] forget_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32.
87  * @param[in] cell_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32.
88  * @param[in] output_gate_bias 1D weights tensor with dimensions [num_units]. Data type supported: S32.
89  * @param[in] cell_state_in 2D tensor with dimensions [num_units, batch_size]. Data type supported: QSYMM16.
90  * @param[in] output_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
91  * @param[out] cell_state_out Destination tensor. Output is a 2D tensor with dimensions [num_units, batch_size]. Data type supported: QSYMM16.
92  * @param[out] output_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input.
93  * @param[out] output Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input.
94  * @param[in] lstm_params Weights tensors used in peephole, CIFG and layer normalization optimizations:
95  * input_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
96  * forget_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
97  * cell_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
98  * output_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
99  * hidden_state_zero The zero point of the hidden state.
100  * hidden_state_scale The scale of the hidden state.
101  * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
102  * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
103  * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16.
104  * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
105  * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
106  * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32.
107  * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
108  * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. S32.
109  * input_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
110  * forget_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
111  * cell_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
112  * output_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
113  * cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip].
114  * If set to 0.0 then clipping is disabled.
115  * projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within
116  * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
117  */
118  void configure(const ITensor *input,
121  const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias,
122  const ITensor *cell_state_in, ITensor *output_state_in,
123  ITensor *cell_state_out, ITensor *output_state_out, ITensor *output,
124  const LSTMParams<ITensor> &lstm_params);
125 
126  /** Static function to check if given info will lead to a valid configuration of @ref NEQLSTMLayer
127  *
128  * @param[in] input Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: QASYMM8_SIGNED.
129  * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8.
130  * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8.
131  * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, num_units]. Data type supported: QSYMM8.
132  * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8.
133  * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8.
134  * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, num_units]. Data type supported: QSYMM8.
135  * @param[in] forget_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: S32.
136  * @param[in] cell_bias 1D weights tensor info with dimensions [num_units]. Data type supported: S32.
137  * @param[in] output_gate_bias 1D weights tensor info with dimensions [num_units]. Data type supported: S32.
138  * @param[in] cell_state_in 2D tensor info with dimensions [num_units, batch_size]. Data type supported: QSYMM16.
139  * @param[in] output_state_in 2D tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
140  * @param[in] cell_state_out Destination tensor info. Output is a 2D tensor info with dimensions [num_units, batch_size]. Data type supported: QSYMM16.
141  * @param[in] output_state_out Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size].Data types supported: Same as @p input.
142  * @param[in] output Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size].Data types supported: Same as @p input.
143  * @param[in] lstm_params Weights tensors info used in peephole, CIFG and layer normalization optimizations:
144  * input_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
145  * forget_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
146  * cell_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
147  * output_intermediate_scale Scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
148  * hidden_state_zero The zero point of the hidden state.
149  * hidden_state_scale The scale of the hidden state.
150  * input_to_input_weights (Optional) 2D weights tensor with dimensions [input_size, num_units]. Data type supported: QSYMM8.
151  * recurrent_to_input_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
152  * cell_to_input_weights (Optional) 1D weights tensor with dimensions [num_units]. Can be nullptr. Data type supported: QSYMM16.
153  * cell_to_forget_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
154  * cell_to_output_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
155  * input_gate_bias (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: S32.
156  * projection_weights (Optional) 2D weights tensor with dimensions [output_size, num_units]. Data type supported: QSYMM8.
157  * projection_bias (Optional) 1D weights tensor with dimensions [output_size]. S32.
158  * input_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
159  * forget_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
160  * cell_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
161  * output_layer_norm_weights (Optional) 1D weights tensor with dimensions [num_units]. Data type supported: QSYMM16.
162  * cell_threshold (Optional) The clipping threshold for the cell state, such that values are bound within [-cell_clip, cell_clip].
163  * If set to 0.0 then clipping is disabled.
164  * projection_threshold (Optional) The clipping threshold for the output from the projection layer, such that values are bound within
165  * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled.
166  * @return a status
167  */
168  static Status validate(const ITensorInfo *input,
169  const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
170  const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
171  const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
172  const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
173  const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out, const ITensorInfo *output,
174  const LSTMParams<ITensorInfo> &lstm_params);
175 
176  // Inherited methods overridden:
177  void run() override;
178  void prepare() override;
179 
180 private:
181  enum class LayerNormGate : uint8_t
182  {
183  Forget,
184  Cell,
185  Input,
186  Output,
187  Count
188  };
189  static constexpr uint8_t _layer_norm_count = static_cast<uint8_t>(LayerNormGate::Count);
190  static constexpr uint32_t _out_state_output_size_dimension_idx = 0;
191 
192  /** Internal method to configure matrix multiplication plus output stage of each gate.
193  *
194  * @param[in] mm Matrix multiplication function to use.
195  * @param[in] outstage Output stage function to use.
196  * @param[in] gemmlowp_info GEMMLowp metadata to be used by the output stage.
197  * @param[in] mm_input Input tensor to matrix multiplication function.
198  * @param[in] mm_weights Weights tensor to matrix multiplication function.
199  * @param[in] bias Bias tensor to matrix multiplication function.
200  * @param[in] outstage_res Tensor to be used for storing the result of the output stage.
201  * @param[in] gemmlowp_scale Real multiplier to be used computing multiplier and shift for requantization.
202  * @param[in] mm_res_info Tensor info to be used to initialize matrix multiplication result tensor.
203  * @param[in] mm_res_info Tensor info to be used to initialize output stage result tensor.
204  *
205  */
206  void configure_mm(NEGEMMLowpMatrixMultiplyCore &mm, NEGEMMLowpOutputStage &outstage, GEMMLowpOutputStageInfo &gemmlowp_info,
207  const ITensor *mm_input, const ITensor *mm_weights, const ITensor *bias, Tensor *mm_res,
208  Tensor *outstage_res, float gemmlowp_scale,
209  const TensorInfo &mm_res_info, const TensorInfo &outstage_tensor_info);
210 
211  MemoryGroup _memory_group;
212 
213  /** A small internel kernel do the copy between two tensors */
214  class TensorCopyKernel
215  {
216  static constexpr uint32_t max_dimension_supported = 2;
217 
218  ITensor *_src{ nullptr };
219  ITensor *_dst{ nullptr };
220  size_t _row_size{};
221  Window _window{};
222 
223  public:
224  /** Destructor */
225  ~TensorCopyKernel();
226  /** Static function to check if given info will lead to a valid configuration of @ref NEQLSTMLayer::TensorCopyKernel
227  *
228  * @param[in] src Source tensor info.
229  * @param[in] dst Destination tensor info
230  *
231  * @return a status
232  */
233  static Status validate(const ITensorInfo &src, const ITensorInfo &dst);
234  /** Set the input and output tensors.
235  *
236  * @param[in] src Source tensor
237  * @param[out] dst Destination tensor
238  */
239  void configure(ITensor &src, ITensor &dst);
240  /** run the kernel */
241  void run();
242  };
243 
244  // Functions used
245  NETranspose _transpose_input_to_forget_weights;
246  NETranspose _transpose_input_to_cell_weights;
247  NETranspose _transpose_input_to_output_weights;
248  NETranspose _transpose_input_to_input_weights;
249  NETranspose _transpose_recurrent_to_forget_weights;
250  NETranspose _transpose_recurrent_to_cell_weights;
251  NETranspose _transpose_recurrent_to_output_weights;
252  NETranspose _transpose_recurrent_to_input_weights;
253  NETranspose _transpose_projection_weights;
254  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _input_to_input_reduction;
255  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _recurrent_to_input_reduction;
256  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _input_to_forget_reduction;
257  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _recurrent_to_forget_reduction;
258  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _input_to_cell_reduction;
259  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _recurrent_to_cell_reduction;
260  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _input_to_output_reduction;
261  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _recurrent_to_output_reduction;
262  std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _projection_reduction;
263  NEArithmeticAddition _projection_bias_add;
264  NEGEMMLowpMatrixMultiplyCore _mm_input_to_forget;
265  NEGEMMLowpMatrixMultiplyCore _mm_recurrent_to_forget;
266  NEPixelWiseMultiplication _pixelwise_mul_cell_to_forget;
267  NEGEMMLowpOutputStage _input_to_forget_outstage;
268  NEGEMMLowpOutputStage _recurrent_to_forget_outstage;
269  NEGEMMLowpOutputStage _cell_to_forget_outstage;
270  NEArithmeticAddition _accumulate_input_recurrent_forget;
271  NEArithmeticAddition _accumulate_cell_forget;
272  NEActivationLayer _forget_gate_sigmoid;
273  NEGEMMLowpMatrixMultiplyCore _mm_input_to_cell;
274  NEGEMMLowpOutputStage _input_to_cell_outstage;
275  NEGEMMLowpMatrixMultiplyCore _mm_recurrent_to_cell;
276  NEGEMMLowpOutputStage _recurrent_to_cell_outstage;
277  NEArithmeticAddition _accumulate_input_recurrent_modulation;
278  NEActivationLayer _cell_gate_tanh;
279  NEArithmeticSubtraction _input_gate_sub;
280  NEGEMMLowpMatrixMultiplyCore _mm_input_to_input;
281  NEGEMMLowpOutputStage _input_to_input_outstage;
282  NEGEMMLowpMatrixMultiplyCore _mm_recurrent_to_input;
283  NEGEMMLowpOutputStage _recurrent_to_input_outstage;
284  NEArithmeticAddition _accumulate_input_recurrent_input;
285  NEPixelWiseMultiplication _pixelwise_mul_cell_to_input;
286  NEGEMMLowpOutputStage _cell_to_input_outstage;
287  NEArithmeticAddition _accumulate_cell_input;
288  NEActivationLayer _input_gate_sigmoid;
289  NEPixelWiseMultiplication _pixelwise_mul_forget_cell;
290  NEPixelWiseMultiplication _pixelwise_mul_input_cell;
291  NEArithmeticAddition _add_forget_cell;
292  NEActivationLayer _cell_clip;
293  NEGEMMLowpMatrixMultiplyCore _mm_input_to_output;
294  NEGEMMLowpOutputStage _input_to_output_outstage;
295  NEGEMMLowpMatrixMultiplyCore _mm_recurrent_to_output;
296  NEGEMMLowpOutputStage _recurrent_to_output_outstage;
297  NEArithmeticAddition _accumulate_input_recurrent_output;
298  NEPixelWiseMultiplication _pixelwise_mul_cell_to_output;
299  NEGEMMLowpOutputStage _cell_to_output_outstage;
300  NEArithmeticAddition _accumulate_cell_to_output;
301  NEActivationLayer _output_gate_sigmoid;
302  NEActivationLayer _hidden_tanh;
303  NEPixelWiseMultiplication _pixelwise_mul_hidden;
304  NEGEMMLowpOutputStage _hidden_outstage;
305  NEGEMMLowpMatrixMultiplyCore _mm_projection;
306  NEGEMMLowpOutputStage _projection_outstage;
307  NEArithmeticAddition _accumulate_projection;
308  NEActivationLayer _projection_clip;
309 
310  TensorCopyKernel _projection_bias_copy;
311  TensorCopyKernel _projection_output_to_accumulate_copy;
312  TensorCopyKernel _projection_accumulate_to_output_copy;
313  TensorCopyKernel _hidden_to_output_copy;
314 
315  std::array<std::unique_ptr<NEQLSTMLayerNormalizationKernel>, _layer_norm_count> _layer_norms;
316 
317  NECopy _copy_output;
318 
319  // Tensor pointers
320  const ITensor *_input_to_input_weights
321  {
322  nullptr
323  };
324  const ITensor *_recurrent_to_input_weights{ nullptr };
325  const ITensor *_projection_bias{ nullptr };
326  const ITensor *_input_to_forget_weights{ nullptr };
327  const ITensor *_input_to_cell_weights{ nullptr };
328  const ITensor *_input_to_output_weights{ nullptr };
329  const ITensor *_recurrent_to_forget_weights{ nullptr };
330  const ITensor *_recurrent_to_cell_weights{ nullptr };
331  const ITensor *_recurrent_to_output_weights{ nullptr };
332  const ITensor *_projection_weights{ nullptr };
333  std::array<const ITensor *, _layer_norm_count> _layer_norm_weights{};
334  std::array<const ITensor *, _layer_norm_count> _layer_norm_bias{};
335 
336  using LayerNormIndexType = typename std::underlying_type<LayerNormGate>::type;
337  inline LayerNormIndexType getGateIndex(LayerNormGate g)
338  {
339  return static_cast<LayerNormIndexType>(g);
340  }
341 
342  inline void set_layer_norm_weight(const ITensor *t, LayerNormGate g)
343  {
344  _layer_norm_weights[getGateIndex(g)] = t;
345  }
346 
347  inline void set_layer_norm_bias(const ITensor *t, LayerNormGate g)
348  {
349  _layer_norm_bias[getGateIndex(g)] = t;
350  }
351 
352  inline const ITensor *get_layer_norm_weight(LayerNormGate g)
353  {
354  return _layer_norm_weights[getGateIndex(g)];
355  }
356 
357  inline const ITensor *get_layer_norm_bias(LayerNormGate g)
358  {
359  return _layer_norm_bias[getGateIndex(g)];
360  }
361 
362  inline std::unique_ptr<NEQLSTMLayerNormalizationKernel> &get_layer_norm(LayerNormGate g)
363  {
364  return _layer_norms[getGateIndex(g)];
365  }
366 
367  void configure_layer_norm(LayerNormGate g, const ITensor *in);
368  static Status validate_layer_norm(const ITensorInfo &in, const ITensorInfo &weight, const ITensorInfo &bias);
369 
370  // Temporary tensors
371  Tensor _input_to_forget_weights_transposed{ nullptr };
372  Tensor _input_to_cell_weights_transposed{ nullptr };
373  Tensor _input_to_output_weights_transposed{ nullptr };
374  Tensor _input_to_input_weights_transposed{ nullptr };
375  Tensor _recurrent_to_forget_weights_transposed{ nullptr };
376  Tensor _recurrent_to_cell_weights_transposed{ nullptr };
377  Tensor _recurrent_to_output_weights_transposed{ nullptr };
378  Tensor _recurrent_to_input_weights_transposed{ nullptr };
379  Tensor _projection_weights_transposed{ nullptr };
380  Tensor _input_to_input_eff_bias{ nullptr };
381  Tensor _recurrent_to_input_eff_bias{ nullptr };
382  Tensor _input_to_forget_eff_bias{ nullptr };
383  Tensor _recurrent_to_forget_eff_bias{ nullptr };
384  Tensor _input_to_cell_eff_bias{ nullptr };
385  Tensor _recurrent_to_cell_eff_bias{ nullptr };
386  Tensor _input_to_output_eff_bias{ nullptr };
387  Tensor _recurrent_to_output_eff_bias{ nullptr };
388  Tensor _projection_reduction_res{ nullptr };
389  Tensor _projection_eff_bias{ nullptr };
390  Tensor _mm_input_to_forget_res{ nullptr };
391  Tensor _mm_recurrent_to_forget_res{ nullptr };
392  Tensor _mul_cell_to_forget_res{ nullptr };
393  Tensor _input_to_forget_outstage_res{ nullptr };
394  Tensor _cell_to_forget_outstage_res{ nullptr };
395  Tensor _recurrent_to_forget_outstage_res{ nullptr };
396  Tensor _forget_gate{ nullptr };
397  Tensor _mm_input_to_cell_res{ nullptr };
398  Tensor _input_to_cell_outstage_res{ nullptr };
399  Tensor _mm_recurrent_to_cell_res{ nullptr };
400  Tensor _recurrent_to_cell_outstage_res{ nullptr };
401  Tensor _cell_gate{ nullptr };
402  Tensor _mul_input_cell_res{ nullptr };
403  Tensor _mm_input_to_input_res{ nullptr };
404  Tensor _input_to_input_outstage_res{ nullptr };
405  Tensor _mm_recurrent_to_input_res{ nullptr };
406  Tensor _mul_cell_to_input_res{ nullptr };
407  Tensor _cell_to_input_outstage_res{ nullptr };
408  Tensor _recurrent_to_input_outstage_res{ nullptr };
409  Tensor _input_gate{ nullptr };
410  Tensor _mm_input_to_output_res{ nullptr };
411  Tensor _input_to_output_outstage_res{ nullptr };
412  Tensor _mm_recurrent_to_output_res{ nullptr };
413  Tensor _mul_cell_to_output_res{ nullptr };
414  Tensor _cell_to_output_outstage_res{ nullptr };
415  Tensor _recurrent_to_output_outstage_res{ nullptr };
416  Tensor _output_gate{ nullptr };
417  Tensor _hidden_mul_res{ nullptr };
418  Tensor _hidden_gate{ nullptr };
419  Tensor _mm_projection_res{ nullptr };
420  Tensor _projection_outstage_res{ nullptr };
421  Tensor _projection_out_res{ nullptr };
422  Tensor _projection_accumulate_res{ nullptr };
423  Tensor _ones{ nullptr };
424  std::array<Tensor, _layer_norm_count> _layer_norm_output{};
425 
426  inline Tensor &get_layer_norm_output(LayerNormGate g)
427  {
428  return _layer_norm_output[getGateIndex(g)];
429  }
430 
431  bool _is_prepared{ false };
432  bool _has_cifg{ false };
433  bool _has_cell_clipping{ false };
434  bool _has_projection{ false };
435  bool _has_projection_clipping{ false };
436  bool _has_peephole{ false };
437  bool _has_layer_norm{ false };
438  bool _projection_tensor_copy_required{ false };
439 };
440 } // namespace arm_compute
441 #endif /* ARM_COMPUTE_NEQLSTMLAYER_H */
NEQLSTMLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
Base class for all functions.
Definition: IFunction.h:30
Basic function to run cpu::kernels::CpuAddKernel.
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
Status class.
Definition: Error.h:52
decltype(strategy::transforms) typedef type
Interface for Neon tensor.
Definition: ITensor.h:36
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2021 Arm Limited.
Basic function to run cpu::kernels::CpuSubKernel.
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 *cell_state_in, const ITensorInfo *output_state_in, const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out, const ITensorInfo *output, const LSTMParams< ITensorInfo > &lstm_params)
Static function to check if given info will lead to a valid configuration of NEQLSTMLayer.
~NEQLSTMLayer()
Default destructor.
GEMMLowp output stage info.
Definition: Types.h:1952
Basic implementation of the tensor interface.
Definition: Tensor.h:37
Basic function to transpose a matrix on Neon.
Definition: NETranspose.h:40
void configure(const ITensor *input, const ITensor *input_to_forget_weights, const ITensor *input_to_cell_weights, const ITensor *input_to_output_weights, const ITensor *recurrent_to_forget_weights, const ITensor *recurrent_to_cell_weights, const ITensor *recurrent_to_output_weights, const ITensor *forget_gate_bias, const ITensor *cell_bias, const ITensor *output_gate_bias, const ITensor *cell_state_in, ITensor *output_state_in, ITensor *cell_state_out, ITensor *output_state_out, ITensor *output, const LSTMParams< ITensor > &lstm_params)
Initialize function&#39;s tensors.
Basic function to run NEQLSTMLayer.
Definition: NEQLSTMLayer.h:62
Basic function to run cpu::kernels::CpuActivationKernel.
Basic function to run NEPixelWiseMultiplicationKernel.
void run() override
Run the kernels contained in the function.
Basic function to execute GEMMLowpQuantizeDown kernels on Neon.
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
NEQLSTMLayer & operator=(const NEQLSTMLayer &)=delete
Prevent instances of this class from being copied (As this class contains pointers) ...
void prepare() override
Prepare the function for executing.
Basic function to execute GEMMLowpMatrixMultiplyCore on Neon.
Describe a multidimensional execution window.
Definition: Window.h:39
Basic function to run cpu::kernels::CpuCopyKernel.
Definition: NECopy.h:39