• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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  */
24 #include "arm_compute/runtime/CL/functions/CLLSTMLayer.h"
25 
26 #include "arm_compute/core/Utils.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/core/utils/misc/InfoHelpers.h"
29 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
30 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
31 #include "arm_compute/runtime/CL/CLScheduler.h"
32 #include "src/core/CL/kernels/CLFillBorderKernel.h"
33 #include "src/gpu/cl/kernels/ClTransposeKernel.h"
34 
35 #include "src/common/utils/Log.h"
36 
37 namespace arm_compute
38 {
39 using namespace arm_compute::misc::shape_calculator;
40 using namespace arm_compute::utils::info_helpers;
41 
CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)42 CLLSTMLayer::CLLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
43     : _memory_group(std::move(memory_manager)), _fully_connected_input_gate(), _accum_input_gate1(), _subtract_input_gate(), _pixelwise_mul_input_gate(), _activation_input_gate(),
44       _fully_connected_forget_gate(), _accum_forget_gate1(), _pixelwise_mul_forget_gate(), _activation_forget_gate(), _fully_connected_cell_state(), _gemm_cell_state1(),
45       _transpose_cell_state(std::make_unique<opencl::kernels::ClTransposeKernel>()), _accum_cell_state1(), _accum_cell_state2(), _pixelwise_mul_cell_state1(), _activation_cell_state(), _cell_clip(),
46       _pixelwise_mul_cell_state2(), _fully_connected_output(), _pixelwise_mul_output_state1(), _accum_output1(), _activation_output(), _activation_output_state(), _pixelwise_mul_output_state2(),
47       _fully_connected_output_state(), _projection_clip(), _copy_cell_state(), _copy_output(), _concat_scratch_buffer(), _concat_inputs_forget_gate(), _concat_weights_forget_gate(),
48       _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(),
49       _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(),
50       _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(),
51       _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(),
52       _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(),
53       _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),
54       _has_projection_weights(false), _perform_projection_clipping(false), _is_prepared(false), _is_layer_norm_lstm(false)
55 {
56 }
57 
58 CLLSTMLayer::~CLLSTMLayer() = default;
59 
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,float projection_threshold)60 void CLLSTMLayer::configure(const ICLTensor *input,
61                             const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
62                             const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
63                             const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
64                             const ICLTensor *output_state_in, ICLTensor *cell_state_in,
65                             ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
66                             const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
67 {
68     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,
69               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,
70               cell_threshold, projection_threshold);
71 }
72 
configure(const CLCompileContext & compile_context,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,float projection_threshold)73 void CLLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input,
74                             const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
75                             const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
76                             const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
77                             const ICLTensor *output_state_in, ICLTensor *cell_state_in,
78                             ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output,
79                             const LSTMParams<ICLTensor> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
80 {
81     ARM_COMPUTE_ERROR_ON_NULLPTR(input,
82                                  input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
83                                  recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
84                                  forget_gate_bias, cell_bias, output_gate_bias,
85                                  output_state_in, cell_state_in,
86                                  scratch_buffer, output_state_out, cell_state_out, output);
87 
88     ARM_COMPUTE_LOG_PARAMS(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights,
89                            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,
90                            output, lstm_params, activation_info, cell_threshold, projection_threshold);
91 
92     _is_layer_norm_lstm = lstm_params.use_layer_norm();
93 
94     // Set lstm parameters
95     LSTMParams<ITensorInfo> lstm_params_info{};
96     build_lstm_params_tensor_info(lstm_params, &lstm_params_info);
97 
98     // Validate
99     ARM_COMPUTE_ERROR_THROW_ON(CLLSTMLayer::validate(input->info(), input_to_forget_weights->info(),
100                                                      input_to_cell_weights->info(), input_to_output_weights->info(),
101                                                      recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
102                                                      forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(),
103                                                      output_state_in->info(), cell_state_in->info(),
104                                                      scratch_buffer->info(), output_state_out->info(), cell_state_out->info(), output->info(),
105                                                      lstm_params_info, activation_info, cell_threshold, projection_threshold));
106 
107     const TensorShape cell_state_shape = cell_state_in->info()->tensor_shape();
108     // Configure block that calculates the forget gate
109     // 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)
110     // We optimize this as follows:
111     // 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
112     _forget_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
113     _forget_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
114     _forget_gate_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
115 
116     std::vector<const ICLTensor *> inputs_vector;
117     inputs_vector.emplace_back(input);
118     inputs_vector.emplace_back(output_state_in);
119     const TensorShape concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, 0);
120     _forget_gate_out2.allocator()->init(TensorInfo(concat_shape, 1, input->info()->data_type()));
121 
122     _memory_group.manage(&_forget_gate_out2);
123     _concat_inputs_forget_gate.configure(compile_context, inputs_vector, &_forget_gate_out2, Window::DimX);
124 
125     std::vector<const ICLTensor *> weights_vector;
126 
127     weights_vector.emplace_back(input_to_forget_weights);
128     weights_vector.emplace_back(recurrent_to_forget_weights);
129     const TensorShape weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(weights_vector, 0);
130     _forget_gate_out6.allocator()->init(TensorInfo(weights_concat_shape, 1, input->info()->data_type()));
131 
132     _concat_weights_forget_gate.configure(compile_context, weights_vector, &_forget_gate_out6, Window::DimX);
133 
134     _memory_group.manage(&_forget_gate_out5);
135     _fully_connected_forget_gate.configure(compile_context, &_forget_gate_out2, &_forget_gate_out6, (_is_layer_norm_lstm) ? nullptr : forget_gate_bias, &_forget_gate_out5);
136     _memory_group.manage(&_forget_gate_out1);
137     _memory_group.manage(&_forget_gate_out3);
138     _forget_gate_out6.allocator()->allocate();
139 
140     CLTensor *forget_gate_out = &_forget_gate_out5;
141     if(lstm_params.has_peephole_opt())
142     {
143         _forget_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
144 
145         _run_peephole_opt = true;
146         _memory_group.manage(&_forget_gate_out4);
147         _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);
148         _accum_forget_gate1.configure(compile_context, &_forget_gate_out5, &_forget_gate_out4, &_forget_gate_out3, ConvertPolicy::SATURATE);
149         _forget_gate_out4.allocator()->allocate();
150         _forget_gate_out5.allocator()->allocate();
151         forget_gate_out = &_forget_gate_out3;
152     }
153     else
154     {
155         _forget_gate_out3.allocator()->allocate();
156     }
157     if(_is_layer_norm_lstm)
158     {
159         _forget_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
160         _forget_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
161         _memory_group.manage(&_forget_layer_norm_out1);
162         _memory_group.manage(&_forget_layer_norm_out2);
163         _mean_std_norm_forget_gate.configure(compile_context, forget_gate_out);
164         _pixelwise_mul_forget_gate_coeff.configure(compile_context, forget_gate_out, lstm_params.forget_layer_norm_weights(), &_forget_layer_norm_out1, 1, ConvertPolicy::SATURATE,
165                                                    RoundingPolicy::TO_NEAREST_EVEN);
166         // forget_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
167         forget_gate_out->allocator()->allocate();
168         _accum_forget_gate_bias.configure(compile_context, &_forget_layer_norm_out1, forget_gate_bias, &_forget_layer_norm_out2, ConvertPolicy::SATURATE);
169         _forget_layer_norm_out1.allocator()->allocate();
170         forget_gate_out = &_forget_layer_norm_out2;
171     }
172     _activation_forget_gate.configure(compile_context, forget_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
173 
174     // Configure block that calculates the input gate
175     // 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
176     // input_gate = 1 - forget_gate, with CIFG
177     // We optimize this as follows:
178     // 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
179     _input_gate_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
180     CLTensor *input_gate_out = &_input_gate_out1;
181     if(lstm_params.has_cifg_opt())
182     {
183         _memory_group.manage(&_input_gate_out1);
184         _ones.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
185         _ones_fill.configure(compile_context, &_ones, PixelValue(1, _ones.info()->data_type()));
186         _subtract_input_gate.configure(compile_context, &_ones, forget_gate_out, &_input_gate_out1, ConvertPolicy::SATURATE);
187         _ones.allocator()->allocate();
188         _run_cifg_opt = true;
189     }
190     else
191     {
192         _input_gate_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
193         _input_gate_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
194 
195         std::vector<const ICLTensor *> lstm_weights;
196         lstm_weights.emplace_back(lstm_params.input_to_input_weights());
197         lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
198         TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
199         _input_gate_out2.allocator()->init(TensorInfo(lstm_weights_concat_shape, 1, input->info()->data_type()));
200 
201         _concat_weights_input_gate.configure(compile_context, lstm_weights, &_input_gate_out2, Window::DimX);
202 
203         _memory_group.manage(&_input_gate_out1);
204 
205         _memory_group.manage(&_input_gate_out3);
206         _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);
207         _input_gate_out2.allocator()->allocate();
208 
209         input_gate_out = &_input_gate_out3;
210         if(_run_peephole_opt)
211         {
212             _memory_group.manage(&_input_gate_out4);
213             _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);
214             _accum_input_gate1.configure(compile_context, &_input_gate_out3, &_input_gate_out4, &_input_gate_out1, ConvertPolicy::SATURATE);
215             _input_gate_out3.allocator()->allocate();
216             _input_gate_out4.allocator()->allocate();
217             input_gate_out = &_input_gate_out1;
218         }
219         else
220         {
221             _input_gate_out1.allocator()->allocate();
222         }
223 
224         if(_is_layer_norm_lstm)
225         {
226             _input_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
227             _input_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
228             _memory_group.manage(&_input_layer_norm_out1);
229             _memory_group.manage(&_input_layer_norm_out2);
230             _mean_std_norm_input_gate.configure(compile_context, input_gate_out);
231             _pixelwise_mul_input_gate_coeff.configure(compile_context, input_gate_out, lstm_params.input_layer_norm_weights(), &_input_layer_norm_out1, 1, ConvertPolicy::SATURATE,
232                                                       RoundingPolicy::TO_NEAREST_EVEN);
233             // input_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
234             input_gate_out->allocator()->allocate();
235             _accum_input_gate_bias.configure(compile_context, &_input_layer_norm_out1, lstm_params.input_gate_bias(), &_input_layer_norm_out2, ConvertPolicy::SATURATE);
236             _input_layer_norm_out1.allocator()->allocate();
237             input_gate_out = &_input_layer_norm_out2;
238         }
239         _activation_input_gate.configure(compile_context, input_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
240     }
241 
242     // Configure block that calculates the cell state
243     // 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)
244     TensorShape cell_state1_shape = compute_transposed_shape(*recurrent_to_output_weights->info());
245     _cell_state_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
246     _cell_state_out2.allocator()->init(TensorInfo(cell_state1_shape, 1, input->info()->data_type()));
247     _cell_state_out3.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
248     _cell_state_out4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
249     _cell_state_out5.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
250 
251     _memory_group.manage(&_cell_state_out1);
252     _fully_connected_cell_state.configure(compile_context, input, input_to_cell_weights, (_is_layer_norm_lstm) ? nullptr : cell_bias, &_cell_state_out1);
253     _memory_group.manage(&_cell_state_out2);
254     _transpose_cell_state->configure(compile_context, recurrent_to_cell_weights->info(), _cell_state_out2.info());
255     _recurrent_to_cell_weights = recurrent_to_cell_weights;
256     _memory_group.manage(&_cell_state_out3);
257     _gemm_cell_state1.configure(compile_context, output_state_in, &_cell_state_out2, nullptr, &_cell_state_out3, 1.f, 0.f);
258     _cell_state_out2.allocator()->allocate();
259     _memory_group.manage(&_cell_state_out4);
260     _accum_cell_state1.configure(compile_context, &_cell_state_out1, &_cell_state_out3, &_cell_state_out4, ConvertPolicy::SATURATE);
261     CLTensor *cell_state_out_ptr = &_cell_state_out4;
262     if(_is_layer_norm_lstm)
263     {
264         _cell_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
265         _cell_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
266         _memory_group.manage(&_cell_layer_norm_out1);
267         _memory_group.manage(&_cell_layer_norm_out2);
268         _mean_std_norm_cell_gate.configure(compile_context, cell_state_out_ptr);
269         _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,
270                                                  RoundingPolicy::TO_NEAREST_EVEN);
271         // cell_state_out_ptr is going to be reassigned, so allocate the tensor that it was assigned to before
272         cell_state_out_ptr->allocator()->allocate();
273         _accum_cell_gate_bias.configure(compile_context, &_cell_layer_norm_out1, cell_bias, &_cell_layer_norm_out2, ConvertPolicy::SATURATE);
274         _cell_layer_norm_out1.allocator()->allocate();
275         cell_state_out_ptr = &_cell_layer_norm_out2;
276     }
277     _activation_cell_state.configure(compile_context, cell_state_out_ptr, nullptr, activation_info);
278     _memory_group.manage(&_cell_state_out5);
279     _pixelwise_mul_cell_state1.configure(compile_context, cell_state_out_ptr, input_gate_out, &_cell_state_out5, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
280     cell_state_out_ptr->allocator()->allocate();
281     _pixelwise_mul_cell_state2.configure(compile_context, forget_gate_out, cell_state_in, &_cell_state_out3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
282     _accum_cell_state2.configure(compile_context, &_cell_state_out5, &_cell_state_out3, &_cell_state_out1, ConvertPolicy::SATURATE);
283     _cell_state_out3.allocator()->allocate();
284     _cell_state_out5.allocator()->allocate();
285     // Perform clipping
286     if(cell_threshold != 0.f)
287     {
288         _perform_cell_clipping = true;
289         _cell_clip.configure(compile_context, &_cell_state_out1, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, cell_threshold, -cell_threshold));
290     }
291 
292     // Configure block that calculates the output
293     // 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)
294     // We optimize this as follows:
295     // 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)
296     _output1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
297     _output4.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
298     std::vector<const ICLTensor *> in_out_weights;
299     in_out_weights.emplace_back(input_to_output_weights);
300     in_out_weights.emplace_back(recurrent_to_output_weights);
301     TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
302     _output2.allocator()->init(TensorInfo(in_out_weights_concat_shape, 1, input->info()->data_type()));
303 
304     _concat_weights_output.configure(compile_context, in_out_weights, &_output2, Window::DimX);
305 
306     _memory_group.manage(&_output1);
307     _memory_group.manage(&_output4);
308 
309     _fully_connected_output.configure(compile_context, &_forget_gate_out2, &_output2, (_is_layer_norm_lstm) ? nullptr : output_gate_bias, &_output4);
310 
311     _output2.allocator()->allocate();
312     _forget_gate_out2.allocator()->allocate();
313 
314     CLTensor *output_gate_out = &_output4;
315     if(lstm_params.has_peephole_opt())
316     {
317         _output3.allocator()->init(TensorInfo(_cell_state_out1.info()->tensor_shape(), 1, input->info()->data_type()));
318 
319         _memory_group.manage(&_output3);
320         _pixelwise_mul_output_state1.configure(compile_context, &_cell_state_out1, lstm_params.cell_to_output_weights(), &_output3, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
321         _accum_output1.configure(compile_context, &_output4, &_output3, &_output1, ConvertPolicy::SATURATE);
322         _output4.allocator()->allocate();
323         output_gate_out = &_output1;
324 
325         // Allocate intermediate buffers
326         _output3.allocator()->allocate();
327     }
328     else
329     {
330         _output1.allocator()->allocate();
331     }
332     if(_is_layer_norm_lstm)
333     {
334         _output_layer_norm_out1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
335         _output_layer_norm_out2.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
336         _memory_group.manage(&_output_layer_norm_out1);
337         _memory_group.manage(&_output_layer_norm_out2);
338         _mean_std_norm_output_gate.configure(compile_context, output_gate_out);
339         _pixelwise_mul_output_gate_coeff.configure(compile_context, output_gate_out, lstm_params.output_layer_norm_weights(), &_output_layer_norm_out1, 1, ConvertPolicy::SATURATE,
340                                                    RoundingPolicy::TO_NEAREST_EVEN);
341         // output_gate_out is going to be reassigned, so allocate the tensor that it was assigned to before
342         output_gate_out->allocator()->allocate();
343         _accum_output_gate_bias.configure(compile_context, &_output_layer_norm_out1, output_gate_bias, &_output_layer_norm_out2, ConvertPolicy::SATURATE);
344         _output_layer_norm_out1.allocator()->allocate();
345         output_gate_out = &_output_layer_norm_out2;
346     }
347     _activation_output.configure(compile_context, output_gate_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
348 
349     // Configure block that calculates the output state
350     /** lstm_res = PixelwiseMul(output, Activation(cell_state))
351      *
352      *                      -- Clip(lstm_res * projection_weights + projection_bias, projection_threshold) , if there is a projection
353      *                     /
354      *  output_state =  --
355      *                     \
356      *                      -- lstm_res , otherwise
357      */
358     ICLTensor *output_state_out_tmp = lstm_params.has_projection() ? &_output_state1 : output_state_out;
359     _cell_state_activation.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
360     _output_state1.allocator()->init(TensorInfo(cell_state_shape, 1, input->info()->data_type()));
361 
362     _memory_group.manage(&_cell_state_activation);
363     _activation_output_state.configure(compile_context, &_cell_state_out1, &_cell_state_activation, activation_info);
364     _pixelwise_mul_output_state2.configure(compile_context, &_cell_state_activation, output_gate_out, output_state_out_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN);
365     _cell_state_activation.allocator()->allocate();
366 
367     if(lstm_params.has_projection())
368     {
369         _has_projection_weights = true;
370         _fully_connected_output_state.configure(compile_context, output_state_out_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out);
371         _output_state1.allocator()->allocate();
372         // Perform clipping
373         if(projection_threshold != 0.f)
374         {
375             _perform_projection_clipping = true;
376             _projection_clip.configure(compile_context, output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold));
377         }
378     }
379 
380     // Copy cell state and output
381     _copy_cell_state.configure(compile_context, &_cell_state_out1, cell_state_out);
382     _copy_output.configure(compile_context, output_state_out, output);
383 
384     // Vector for holding the tensors to store in scratch buffer
385     std::vector<const ICLTensor *> scratch_inputs;
386     if(!lstm_params.has_cifg_opt())
387     {
388         scratch_inputs.emplace_back(input_gate_out);
389     }
390     scratch_inputs.emplace_back(&_cell_state_out1);
391     scratch_inputs.emplace_back(forget_gate_out);
392     scratch_inputs.emplace_back(output_gate_out);
393     _concat_scratch_buffer.configure(compile_context, scratch_inputs, scratch_buffer, Window::DimX);
394     input_gate_out->allocator()->allocate();
395     _cell_state_out1.allocator()->allocate();
396     forget_gate_out->allocator()->allocate();
397     output_gate_out->allocator()->allocate();
398 }
399 
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,float projection_threshold)400 Status CLLSTMLayer::validate(const ITensorInfo *input,
401                              const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
402                              const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
403                              const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
404                              const ITensorInfo *output_state_in, const ITensorInfo *cell_state_in,
405                              const ITensorInfo *scratch_buffer, const ITensorInfo *output_state_out, const ITensorInfo *cell_state_out, const ITensorInfo *output,
406                              const LSTMParams<ITensorInfo> &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold, float projection_threshold)
407 {
408     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input,
409                                         input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
410                                         recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
411                                         forget_gate_bias, cell_bias, output_gate_bias,
412                                         output_state_in, cell_state_in,
413                                         scratch_buffer, output_state_out, cell_state_out, output);
414 
415     // Check data types
416     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
417     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input,
418                                                        input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
419                                                        recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
420                                                        forget_gate_bias, cell_bias, output_gate_bias,
421                                                        output_state_in, cell_state_in,
422                                                        scratch_buffer, output_state_out, cell_state_out, output);
423 
424     // Check dimensions
425     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2);
426     ARM_COMPUTE_RETURN_ERROR_ON(input_to_forget_weights->num_dimensions() > 2);
427     ARM_COMPUTE_RETURN_ERROR_ON(input_to_cell_weights->num_dimensions() > 2);
428     ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->num_dimensions() > 2);
429     ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_forget_weights->num_dimensions() > 2);
430     ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_cell_weights->num_dimensions() > 2);
431     ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->num_dimensions() > 2);
432     ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() > 1);
433     ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->num_dimensions() > 1);
434     ARM_COMPUTE_RETURN_ERROR_ON(output_gate_bias->num_dimensions() > 1);
435     ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
436     ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() > 2);
437     ARM_COMPUTE_RETURN_ERROR_ON(scratch_buffer->num_dimensions() > 2);
438     ARM_COMPUTE_RETURN_ERROR_ON(output_state_out->num_dimensions() > 2);
439     ARM_COMPUTE_RETURN_ERROR_ON(cell_state_out->num_dimensions() > 2);
440     ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 2);
441     ARM_COMPUTE_RETURN_ERROR_ON(cell_bias->dimension(0) * 4 != scratch_buffer->dimension(0)
442                                 && cell_bias->dimension(0) * 3 != scratch_buffer->dimension(0));
443 
444     const unsigned int num_batches = input->dimension(1);
445     const unsigned int num_cells   = input_to_output_weights->dimension(1);
446 
447     if(lstm_params.use_layer_norm())
448     {
449         // If CIFG is used, input layer normalization weights tensor is omitted
450         if(lstm_params.has_cifg_opt())
451         {
452             ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights() != nullptr);
453         }
454         else
455         {
456             ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_layer_norm_weights());
457             ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->num_dimensions() > 1);
458             ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_layer_norm_weights()->dimension(0) != num_cells);
459             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, lstm_params.input_layer_norm_weights());
460         }
461 
462         ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.forget_layer_norm_weights(), lstm_params.cell_layer_norm_weights(), lstm_params.output_layer_norm_weights());
463         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, lstm_params.forget_layer_norm_weights(), lstm_params.cell_layer_norm_weights(), lstm_params.output_layer_norm_weights());
464         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->num_dimensions() > 1);
465         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->num_dimensions() > 1);
466         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->num_dimensions() > 1);
467         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_layer_norm_weights()->dimension(0) != num_cells);
468         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_layer_norm_weights()->dimension(0) != num_cells);
469         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_layer_norm_weights()->dimension(0) != num_cells);
470     }
471 
472     // Check peephole optimization
473     if(lstm_params.has_peephole_opt())
474     {
475         ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_output_weights(), lstm_params.cell_to_forget_weights());
476         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->num_dimensions() > 1);
477         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_output_weights()->num_dimensions() > 1);
478     }
479 
480     TensorShape      units_out_transposed_shape = compute_transposed_shape(*recurrent_to_output_weights);
481     TensorShape      num_units_transposed_shape = compute_transposed_shape(*forget_gate_bias);
482     const TensorInfo units_out_transposed_info  = TensorInfo(units_out_transposed_shape, 1, input->data_type());
483     const TensorInfo num_units_transposed_info  = TensorInfo(num_units_transposed_shape, 1, input->data_type());
484 
485     TensorInfo input_gate      = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
486     TensorInfo forget_gate     = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
487     TensorInfo output_gate_tmp = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
488     TensorInfo cell_state_tmp  = TensorInfo(TensorShape(num_cells, num_batches), 1, input->data_type());
489 
490     // Validate forget gate
491     ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_forget_weights, (lstm_params.use_layer_norm()) ? nullptr : forget_gate_bias, &forget_gate));
492 
493     std::vector<const ITensorInfo *> inputs_vector;
494     inputs_vector.emplace_back(input);
495     inputs_vector.emplace_back(output_state_in);
496     const TensorShape concat_shape       = arm_compute::misc::shape_calculator::calculate_concatenate_shape(inputs_vector, 0);
497     TensorInfo        forget_gate_concat = TensorInfo(concat_shape, 1, input->data_type());
498 
499     ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector, &forget_gate_concat, Window::DimX));
500 
501     if(lstm_params.has_peephole_opt())
502     {
503         ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &forget_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
504         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
505     }
506     if(lstm_params.use_layer_norm())
507     {
508         ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&forget_gate));
509         ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&forget_gate, lstm_params.forget_layer_norm_weights(), &forget_gate, 1, ConvertPolicy::SATURATE,
510                                                                         RoundingPolicy::TO_NEAREST_EVEN));
511         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE));
512     }
513     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
514 
515     // Validate input gate
516     if(!lstm_params.has_cifg_opt())
517     {
518         ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(),
519                                             lstm_params.recurrent_to_input_weights(),
520                                             lstm_params.input_gate_bias());
521         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_to_input_weights()->num_dimensions() > 2);
522         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.recurrent_to_input_weights()->num_dimensions() > 2);
523         ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_gate_bias()->num_dimensions() > 1);
524 
525         std::vector<const ITensorInfo *> lstm_weights;
526         lstm_weights.emplace_back(lstm_params.input_to_input_weights());
527         lstm_weights.emplace_back(lstm_params.recurrent_to_input_weights());
528         TensorShape lstm_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(lstm_weights, 0);
529         TensorInfo  lstm_gate_concat          = TensorInfo(lstm_weights_concat_shape, 1, input->data_type());
530         ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(lstm_weights, &lstm_gate_concat, Window::DimX));
531 
532         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));
533 
534         if(lstm_params.has_peephole_opt())
535         {
536             ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_input_weights());
537             ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_input_weights()->num_dimensions() > 1);
538             ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
539             ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, &input_gate, &input_gate, ConvertPolicy::SATURATE));
540         }
541 
542         if(lstm_params.use_layer_norm())
543         {
544             ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&input_gate));
545             ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
546             ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE));
547         }
548         ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
549     }
550     else
551     {
552         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticSubtraction::validate(&forget_gate, &forget_gate, &forget_gate, ConvertPolicy::SATURATE));
553     }
554 
555     // Validate cell state
556     ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_cell_weights, (lstm_params.use_layer_norm()) ? nullptr : cell_bias, &cell_state_tmp));
557     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(output_state_in, &units_out_transposed_info, nullptr, &cell_state_tmp, 1.f, 0.f, GEMMInfo()));
558     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
559     if(lstm_params.use_layer_norm())
560     {
561         ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&cell_state_tmp));
562         ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_layer_norm_weights(), &cell_state_tmp, 1, ConvertPolicy::SATURATE,
563                                                                         RoundingPolicy::TO_NEAREST_EVEN));
564         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE));
565     }
566     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, activation_info));
567     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
568     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
569     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
570     if(cell_threshold != 0.f)
571     {
572         ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, cell_threshold,
573                                                                                                               -cell_threshold)));
574     }
575 
576     std::vector<const ITensorInfo *> in_out_weights;
577     in_out_weights.emplace_back(input_to_output_weights);
578     in_out_weights.emplace_back(recurrent_to_output_weights);
579     TensorShape in_out_weights_concat_shape = arm_compute::misc::shape_calculator::calculate_concatenate_shape(in_out_weights, 0);
580     TensorInfo  in_out_gate_concat          = TensorInfo(in_out_weights_concat_shape, 1, input->data_type());
581     ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(in_out_weights, &in_out_gate_concat, Window::DimX));
582     // Validate output gate tmp
583     ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, input_to_output_weights, (lstm_params.use_layer_norm()) ? nullptr : output_gate_bias, &output_gate_tmp));
584 
585     if(lstm_params.has_peephole_opt())
586     {
587         ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, lstm_params.cell_to_output_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE,
588                                                                         RoundingPolicy::TO_NEAREST_EVEN));
589         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, &output_gate_tmp, &output_gate_tmp, ConvertPolicy::SATURATE));
590     }
591     if(lstm_params.use_layer_norm())
592     {
593         ARM_COMPUTE_RETURN_ON_ERROR(CLMeanStdDevNormalizationLayer::validate(&output_gate_tmp));
594         ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&output_gate_tmp, lstm_params.output_layer_norm_weights(), &output_gate_tmp, 1, ConvertPolicy::SATURATE,
595                                                                         RoundingPolicy::TO_NEAREST_EVEN));
596         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE));
597     }
598     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
599 
600     // Validate output state
601     ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info));
602     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
603     if(lstm_params.has_projection())
604     {
605         ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out));
606         if(projection_threshold != 0.f)
607         {
608             ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output_state_out, output_state_out,
609                                                                     ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)));
610         }
611     }
612 
613     // Validate copy kernel
614     ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(&cell_state_tmp, cell_state_out));
615     ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(output_state_out, output));
616 
617     // Validate scratch concatenation
618     std::vector<const ITensorInfo *> inputs_vector_info_raw;
619     if(!lstm_params.has_cifg_opt())
620     {
621         inputs_vector_info_raw.push_back(&input_gate);
622     }
623     inputs_vector_info_raw.push_back(&cell_state_tmp);
624     inputs_vector_info_raw.push_back(&forget_gate);
625     inputs_vector_info_raw.push_back(&output_gate_tmp);
626 
627     ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_vector_info_raw, scratch_buffer, Window::DimX));
628     return Status{};
629 }
630 
run()631 void CLLSTMLayer::run()
632 {
633     prepare();
634 
635     MemoryGroupResourceScope scope_mg(_memory_group);
636 
637     _concat_inputs_forget_gate.run();
638 
639     _fully_connected_forget_gate.run();
640 
641     if(_run_peephole_opt)
642     {
643         _pixelwise_mul_forget_gate.run();
644         _accum_forget_gate1.run();
645     }
646     if(_is_layer_norm_lstm)
647     {
648         _mean_std_norm_forget_gate.run();
649         _pixelwise_mul_forget_gate_coeff.run();
650         _accum_forget_gate_bias.run();
651     }
652     _activation_forget_gate.run();
653 
654     if(_run_cifg_opt)
655     {
656         _ones_fill.run();
657         _subtract_input_gate.run();
658     }
659     else
660     {
661         _fully_connected_input_gate.run();
662 
663         if(_run_peephole_opt)
664         {
665             _pixelwise_mul_input_gate.run();
666             _accum_input_gate1.run();
667         }
668 
669         if(_is_layer_norm_lstm)
670         {
671             _mean_std_norm_input_gate.run();
672             _pixelwise_mul_input_gate_coeff.run();
673             _accum_input_gate_bias.run();
674         }
675         _activation_input_gate.run();
676     }
677 
678     _fully_connected_cell_state.run();
679     ITensorPack pack;
680     pack.add_tensor(TensorType::ACL_SRC, _recurrent_to_cell_weights);
681     pack.add_tensor(TensorType::ACL_DST, &_cell_state_out2);
682     CLScheduler::get().enqueue_op(*_transpose_cell_state,
683                                   pack,
684                                   false);
685     _gemm_cell_state1.run();
686     _accum_cell_state1.run();
687     if(_is_layer_norm_lstm)
688     {
689         _mean_std_norm_cell_gate.run();
690         _pixelwise_mul_cell_gate_coeff.run();
691         _accum_cell_gate_bias.run();
692     }
693     _activation_cell_state.run();
694     _pixelwise_mul_cell_state1.run();
695     _pixelwise_mul_cell_state2.run();
696     _accum_cell_state2.run();
697 
698     if(_perform_cell_clipping)
699     {
700         _cell_clip.run();
701     }
702 
703     _fully_connected_output.run();
704 
705     if(_run_peephole_opt)
706     {
707         _pixelwise_mul_output_state1.run();
708         _accum_output1.run();
709     }
710     if(_is_layer_norm_lstm)
711     {
712         _mean_std_norm_output_gate.run();
713         _pixelwise_mul_output_gate_coeff.run();
714         _accum_output_gate_bias.run();
715     }
716     _activation_output.run();
717 
718     _activation_output_state.run();
719     _pixelwise_mul_output_state2.run();
720 
721     if(_has_projection_weights)
722     {
723         _fully_connected_output_state.run();
724         if(_perform_projection_clipping)
725         {
726             _projection_clip.run();
727         }
728     }
729 
730     _copy_cell_state.run();
731     _copy_output.run();
732 
733     _concat_scratch_buffer.run();
734 }
735 
prepare()736 void CLLSTMLayer::prepare()
737 {
738     if(!_is_prepared)
739     {
740         _concat_weights_forget_gate.run();
741         if(!_run_cifg_opt)
742         {
743             _concat_weights_input_gate.run();
744         }
745         _concat_weights_output.run();
746         _is_prepared = true;
747     }
748 }
749 } // namespace arm_compute
750