1 /*
2 * Copyright (c) 2019-2020 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24
25 #include "arm_compute/runtime/CL/functions/CLLSTMLayerQuantized.h"
26
27 #include "arm_compute/core/Utils.h"
28 #include "arm_compute/core/Validate.h"
29 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
30 #include "src/core/CL/kernels/CLDepthConvertLayerKernel.h"
31 #include "src/core/CL/kernels/CLFillBorderKernel.h"
32 #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h"
33 #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h"
34 #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
35 #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h"
36 #include "src/core/CL/kernels/CLGEMMLowpReductionKernel.h"
37 #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h"
38 #include "src/core/helpers/AutoConfiguration.h"
39
40 #include <memory>
41
42 namespace arm_compute
43 {
44 namespace
45 {
46 // Quantization info structures used in the LSTMQuantize layer
47 const QuantizationInfo qasymm(1.f / 128.f, 128);
48 const QuantizationInfo qsymm_3(8.f / 32768.f, 0); // qsymm16 with 3 integer bit
49 const QuantizationInfo qsymm_4(16.f / 32768.f, 0); // qsymm16 with 4 integer bit
50 const QuantizationInfo qsymm_0(1.f / 32768.f, 0); // qsymm16 with 0 integer bit
51 } // namespace
52
CLLSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager)53 CLLSTMLayerQuantized::CLLSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager)
54 : _memory_group(std::move(memory_manager)), _gemmlowp(), _output_stage(), _transpose_weights(), _concat_input_weights(), _concat_recurrent_weights(), _concat_weights(), _concat_inputs(),
55 _concat_bias(), _sigmoid_forget_gate(), _sigmoid_input_gate(), _sigmoid_output_gate(), _tanh_modulation_gate(), _tanh_output_state(), _add_cell_state_tmps(), _add2(), _mul_forget_gate_cell_state(),
56 _mul_input_gate_input_mod_gate(), _mul_output_state_tmp_output_gate(), _slice_input_tensor(), _slice_forget_tensor(), _slice_cell_tensor(), _slice_output_tensor(), _dequantize(), _quantize(),
57 _input_to_input_weights(nullptr), _input_to_forget_weights(nullptr), _input_to_cell_weights(nullptr), _input_to_output_weights(nullptr), _recurrent_to_input_weights(nullptr),
58 _recurrent_to_forget_weights(nullptr), _recurrent_to_cell_weights(nullptr), _recurrent_to_output_weights(nullptr), _input_gate_bias(nullptr), _forget_gate_bias(nullptr), _cell_bias(nullptr),
59 _output_gate_bias(nullptr), _recurrent_weights(), _input_weights(), _weights(), _input(), _weights_transposed(), _output_highp(), _output_lowp(), _bias(), _forget_gate_input(), _input_gate_input(),
60 _output_gate_input(), _input_modulation_gate_input(), _forget_gate_output(), _input_gate_output(), _output_gate_output(), _input_modulation_gate_output(), _cell_state_tmp1(), _cell_state_tmp2(),
61 _output_state_tmp(), _output_state_out_symm(), _output_state_out_f32(), _is_prepared(false)
62 {
63 }
64
configure(const ICLTensor * input,const ICLTensor * input_to_input_weights,const ICLTensor * input_to_forget_weights,const ICLTensor * input_to_cell_weights,const ICLTensor * input_to_output_weights,const ICLTensor * recurrent_to_input_weights,const ICLTensor * recurrent_to_forget_weights,const ICLTensor * recurrent_to_cell_weights,const ICLTensor * recurrent_to_output_weights,const ICLTensor * input_gate_bias,const ICLTensor * forget_gate_bias,const ICLTensor * cell_bias,const ICLTensor * output_gate_bias,ICLTensor * cell_state_in,const ICLTensor * output_state_in,ICLTensor * cell_state_out,ICLTensor * output_state_out)65 void CLLSTMLayerQuantized::configure(const ICLTensor *input,
66 const ICLTensor *input_to_input_weights, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
67 const ICLTensor *recurrent_to_input_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
68 const ICLTensor *input_gate_bias, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
69 ICLTensor *cell_state_in, const ICLTensor *output_state_in,
70 ICLTensor *cell_state_out, ICLTensor *output_state_out)
71 {
72 configure(CLKernelLibrary::get().get_compile_context(), input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_input_weights,
73 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out,
74 output_state_out);
75 }
76
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * input_to_input_weights,const ICLTensor * input_to_forget_weights,const ICLTensor * input_to_cell_weights,const ICLTensor * input_to_output_weights,const ICLTensor * recurrent_to_input_weights,const ICLTensor * recurrent_to_forget_weights,const ICLTensor * recurrent_to_cell_weights,const ICLTensor * recurrent_to_output_weights,const ICLTensor * input_gate_bias,const ICLTensor * forget_gate_bias,const ICLTensor * cell_bias,const ICLTensor * output_gate_bias,ICLTensor * cell_state_in,const ICLTensor * output_state_in,ICLTensor * cell_state_out,ICLTensor * output_state_out)77 void CLLSTMLayerQuantized::configure(const CLCompileContext &compile_context, const ICLTensor *input,
78 const ICLTensor *input_to_input_weights, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
79 const ICLTensor *recurrent_to_input_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
80 const ICLTensor *input_gate_bias, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
81 ICLTensor *cell_state_in, const ICLTensor *output_state_in,
82 ICLTensor *cell_state_out, ICLTensor *output_state_out)
83 {
84 ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
85 recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
86 input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in, cell_state_out, output_state_out);
87
88 ARM_COMPUTE_ERROR_THROW_ON(CLLSTMLayerQuantized::validate(input->info(), input_to_input_weights->info(), input_to_forget_weights->info(), input_to_cell_weights->info(),
89 input_to_output_weights->info(),
90 recurrent_to_input_weights->info(), recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
91 input_gate_bias->info(), forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(), cell_state_in->info(), output_state_in->info(), cell_state_out->info(), output_state_out->info()));
92
93 const int input_size = input->info()->dimension(0);
94 const int batch_size = input->info()->dimension(1);
95 const int output_size = input_to_input_weights->info()->dimension(1);
96
97 const QuantizationInfo qweights = input_to_input_weights->info()->quantization_info(); // Weights quantization
98
99 auto_init_if_empty(*cell_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QSYMM16, qsymm_4));
100 auto_init_if_empty(*output_state_out->info(), TensorInfo(TensorShape(batch_size, output_size), 1, DataType::QASYMM8, qasymm));
101
102 _input_to_input_weights = input_to_input_weights;
103 _input_to_forget_weights = input_to_forget_weights;
104 _input_to_cell_weights = input_to_cell_weights;
105 _input_to_output_weights = input_to_output_weights;
106 _recurrent_to_input_weights = recurrent_to_input_weights;
107 _recurrent_to_forget_weights = recurrent_to_forget_weights;
108 _recurrent_to_cell_weights = recurrent_to_cell_weights;
109 _recurrent_to_output_weights = recurrent_to_output_weights;
110 _input_gate_bias = input_gate_bias;
111 _forget_gate_bias = forget_gate_bias;
112 _cell_bias = cell_bias;
113 _output_gate_bias = output_gate_bias;
114
115 // Weights concatenation
116 std::vector<const ICLTensor *> inputs_weights_vector;
117 inputs_weights_vector.emplace_back(input_to_input_weights);
118 inputs_weights_vector.emplace_back(input_to_forget_weights);
119 inputs_weights_vector.emplace_back(input_to_cell_weights);
120 inputs_weights_vector.emplace_back(input_to_output_weights);
121
122 std::vector<const ICLTensor *> recurrent_weights_vector;
123 recurrent_weights_vector.emplace_back(recurrent_to_input_weights);
124 recurrent_weights_vector.emplace_back(recurrent_to_forget_weights);
125 recurrent_weights_vector.emplace_back(recurrent_to_cell_weights);
126 recurrent_weights_vector.emplace_back(recurrent_to_output_weights);
127
128 _input_weights.allocator()->init(TensorInfo(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
129 _concat_input_weights.configure(compile_context, inputs_weights_vector, &_input_weights, Window::DimY);
130
131 _recurrent_weights.allocator()->init(TensorInfo(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
132 _concat_recurrent_weights.configure(compile_context, recurrent_weights_vector, &_recurrent_weights, Window::DimY);
133
134 std::vector<const ICLTensor *> weights_vector;
135 weights_vector.emplace_back(&_recurrent_weights);
136 weights_vector.emplace_back(&_input_weights);
137
138 _weights.allocator()->init(TensorInfo(TensorShape(output_size + input_size, 4 * output_size), 1, DataType::QASYMM8, qweights));
139 _concat_weights.configure(compile_context, weights_vector, &_weights, Window::DimX);
140 _transpose_weights.configure(compile_context, &_weights, &_weights_transposed);
141
142 // Input concatenation
143 std::vector<const ICLTensor *> input_vector;
144 input_vector.emplace_back(input);
145 input_vector.emplace_back(output_state_in);
146
147 _memory_group.manage(&_input);
148 _input.allocator()->init(TensorInfo(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm));
149 _concat_inputs.configure(compile_context, input_vector, &_input, Window::DimX);
150
151 // Bias concatenation
152 std::vector<const ICLTensor *> bias_vector;
153 bias_vector.emplace_back(input_gate_bias);
154 bias_vector.emplace_back(forget_gate_bias);
155 bias_vector.emplace_back(cell_bias);
156 bias_vector.emplace_back(output_gate_bias);
157
158 _bias.allocator()->init(TensorInfo(TensorShape(4 * output_size), 1, DataType::S32));
159 _concat_bias.configure(compile_context, bias_vector, &_bias, Window::DimX);
160
161 // Invert the offset for gemmlowp
162 _input.info()->set_quantization_info(QuantizationInfo(qasymm.uniform().scale, -qasymm.uniform().offset));
163 _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
164
165 // Run gemmlowp
166 _memory_group.manage(&_output_highp);
167 _output_highp.allocator()->init(TensorInfo(TensorShape(4 * output_size, batch_size), 1, DataType::S32));
168 _gemmlowp.configure(compile_context, &_input, &_weights_transposed, nullptr, &_output_highp);
169 _input.allocator()->allocate();
170
171 // Set the offset back
172 _input.info()->set_quantization_info(QuantizationInfo(qasymm.uniform().scale, qasymm.uniform().offset));
173 _weights_transposed.info()->set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
174
175 // multiplier = (input_scale * weights_scale) / output_scale (2 ^ (-12))
176 _output_lowp.allocator()->init(TensorInfo(_output_highp.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_3));
177
178 const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
179 int output_multiplier = 0;
180 int output_shift = 0;
181 quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
182
183 _memory_group.manage(&_output_lowp);
184 _output_stage.configure(compile_context, &_output_highp, &_bias, &_output_lowp, output_multiplier, output_shift);
185 _output_highp.allocator()->allocate();
186 _bias.allocator()->allocate();
187
188 // Get the gate tensors
189 if(batch_size > 1)
190 {
191 _memory_group.manage(&_input_gate_input);
192 _slice_input_tensor.configure(compile_context, &_output_lowp, &_input_gate_input, { 0, 0 }, { output_size, batch_size });
193 _memory_group.manage(&_forget_gate_input);
194 _slice_forget_tensor.configure(compile_context, &_output_lowp, &_forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size });
195 _memory_group.manage(&_input_modulation_gate_input);
196 _slice_cell_tensor.configure(compile_context, &_output_lowp, &_input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size });
197 _memory_group.manage(&_output_gate_input);
198 _slice_output_tensor.configure(compile_context, &_output_lowp, &_output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size });
199 _output_lowp.allocator()->allocate();
200 }
201 else
202 {
203 _memory_group.manage(&_input_gate_input);
204 _slice_input_tensor.configure(compile_context, &_output_lowp, &_input_gate_input, { 0 }, { output_size });
205 _memory_group.manage(&_forget_gate_input);
206 _slice_forget_tensor.configure(compile_context, &_output_lowp, &_forget_gate_input, { output_size }, { 2 * output_size });
207 _memory_group.manage(&_input_modulation_gate_input);
208 _slice_cell_tensor.configure(compile_context, &_output_lowp, &_input_modulation_gate_input, { 2 * output_size }, { 3 * output_size });
209 _memory_group.manage(&_output_gate_input);
210 _slice_output_tensor.configure(compile_context, &_output_lowp, &_output_gate_input, { 3 * output_size }, { 4 * output_size });
211 _output_lowp.allocator()->allocate();
212 }
213
214 // Forget gate
215 _memory_group.manage(&_forget_gate_output);
216 _forget_gate_output.allocator()->init(TensorInfo(_forget_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
217 _sigmoid_forget_gate.configure(compile_context, &_forget_gate_input, &_forget_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
218 _forget_gate_input.allocator()->allocate();
219
220 // Input gate
221 _memory_group.manage(&_input_gate_output);
222 _input_gate_output.allocator()->init(TensorInfo(_input_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
223 _sigmoid_input_gate.configure(compile_context, &_input_gate_input, &_input_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
224 _input_gate_input.allocator()->allocate();
225
226 // Input modulation gate equation
227 _memory_group.manage(&_input_modulation_gate_output);
228 _input_modulation_gate_output.allocator()->init(TensorInfo(_input_modulation_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
229 _tanh_modulation_gate.configure(compile_context, &_input_modulation_gate_input, &_input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
230 _input_modulation_gate_input.allocator()->allocate();
231
232 // Output gate
233 _memory_group.manage(&_output_gate_output);
234 _output_gate_output.allocator()->init(TensorInfo(_output_gate_input.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
235 _sigmoid_output_gate.configure(compile_context, &_output_gate_input, &_output_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
236 _output_gate_input.allocator()->allocate();
237
238 // Long term memory
239 _memory_group.manage(&_cell_state_tmp1);
240 _cell_state_tmp1.allocator()->init(TensorInfo(_forget_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
241 _mul_forget_gate_cell_state.configure(compile_context, &_forget_gate_output, cell_state_in, &_cell_state_tmp1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
242 _forget_gate_output.allocator()->allocate();
243
244 _memory_group.manage(&_cell_state_tmp2);
245 _cell_state_tmp2.allocator()->init(TensorInfo(_input_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_4));
246 _mul_input_gate_input_mod_gate.configure(compile_context, &_input_gate_output, &_input_modulation_gate_output, &_cell_state_tmp2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
247 _input_modulation_gate_output.allocator()->allocate();
248 _input_gate_output.allocator()->allocate();
249
250 _add_cell_state_tmps.configure(compile_context, &_cell_state_tmp1, &_cell_state_tmp2, cell_state_out, ConvertPolicy::SATURATE);
251 _cell_state_tmp1.allocator()->allocate();
252 _cell_state_tmp2.allocator()->allocate();
253
254 // Short term memory
255 _memory_group.manage(&_output_state_tmp);
256 _output_state_tmp.allocator()->init(TensorInfo(cell_state_out->info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
257 _tanh_output_state.configure(compile_context, cell_state_out, &_output_state_tmp, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f));
258
259 _memory_group.manage(&_output_state_out_symm);
260 _output_state_out_symm.allocator()->init(TensorInfo(_output_gate_output.info()->tensor_shape(), 1, DataType::QSYMM16, qsymm_0));
261 _mul_output_state_tmp_output_gate.configure(compile_context, &_output_state_tmp, &_output_gate_output, &_output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
262 _output_gate_output.allocator()->allocate();
263 _output_state_tmp.allocator()->allocate();
264
265 // Requantize the output state from QSYMM16 to QASYMM8
266 _memory_group.manage(&_output_state_out_f32);
267 _output_state_out_f32.allocator()->init(TensorInfo(_output_state_out_symm.info()->tensor_shape(), 1, DataType::F32));
268 _dequantize.configure(compile_context, &_output_state_out_symm, &_output_state_out_f32);
269 _output_state_out_symm.allocator()->allocate();
270
271 _quantize.configure(compile_context, &_output_state_out_f32, output_state_out);
272 _output_state_out_f32.allocator()->allocate();
273 }
274
validate(const ITensorInfo * input,const ITensorInfo * input_to_input_weights,const ITensorInfo * input_to_forget_weights,const ITensorInfo * input_to_cell_weights,const ITensorInfo * input_to_output_weights,const ITensorInfo * recurrent_to_input_weights,const ITensorInfo * recurrent_to_forget_weights,const ITensorInfo * recurrent_to_cell_weights,const ITensorInfo * recurrent_to_output_weights,const ITensorInfo * input_gate_bias,const ITensorInfo * forget_gate_bias,const ITensorInfo * cell_bias,const ITensorInfo * output_gate_bias,const ITensorInfo * cell_state_in,const ITensorInfo * output_state_in,const ITensorInfo * cell_state_out,const ITensorInfo * output_state_out)275 Status CLLSTMLayerQuantized::validate(const ITensorInfo *input,
276 const ITensorInfo *input_to_input_weights, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
277 const ITensorInfo *recurrent_to_input_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
278 const ITensorInfo *input_gate_bias, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
279 const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
280 const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out)
281 {
282 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_input_weights,
283 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in,
284 output_state_in, cell_state_out, output_state_out);
285 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(input, DataType::QASYMM8);
286
287 const int input_size = input->dimension(0);
288 const int batch_size = input->dimension(1);
289 const int output_size = input_to_input_weights->dimension(1);
290
291 // Dimensionality checks
292 ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 2);
293 ARM_COMPUTE_RETURN_ERROR_ON(input_to_input_weights->num_dimensions() > 2);
294 ARM_COMPUTE_RETURN_ERROR_ON(input_gate_bias->num_dimensions() > 1);
295 ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() > 2);
296
297 TensorInfo input_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(input_size, output_size)).set_data_type(DataType::QASYMM8));
298 TensorInfo recurrent_weights_info(input_to_input_weights->clone()->set_tensor_shape(TensorShape(output_size, output_size)).set_data_type(DataType::QASYMM8));
299 TensorInfo bias_info(input_gate_bias->clone()->set_tensor_shape(TensorShape(output_size)).set_data_type(DataType::S32));
300 TensorInfo output_state_info(cell_state_in->clone()->set_tensor_shape(TensorShape(output_size, batch_size)).set_data_type(DataType::QASYMM8).set_quantization_info(qasymm));
301 TensorInfo cell_state_info(cell_state_in->clone()->set_tensor_shape(TensorShape(output_size, batch_size)).set_data_type(DataType::QSYMM16).set_quantization_info(qsymm_4));
302
303 // Shape checks
304 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&input_weights_info, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
305 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&recurrent_weights_info, recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
306 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
307 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_in);
308 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_in);
309
310 // Data type checks
311 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&input_weights_info, input, input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
312 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&recurrent_weights_info, recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
313 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&bias_info, input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias);
314 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_in);
315 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_in);
316
317 // Quantization checks
318 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input_to_input_weights, input_to_forget_weights, input_to_cell_weights, input_to_output_weights);
319 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(recurrent_to_input_weights, recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
320 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_in);
321 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_in);
322
323 // Validate internal functions
324 // _concat_input_weights
325 std::vector<const ITensorInfo *> inputs_weights_vector;
326 inputs_weights_vector.emplace_back(input_to_input_weights);
327 inputs_weights_vector.emplace_back(input_to_forget_weights);
328 inputs_weights_vector.emplace_back(input_to_cell_weights);
329 inputs_weights_vector.emplace_back(input_to_output_weights);
330 const QuantizationInfo qweights = input_to_input_weights->quantization_info(); // Weights quantization
331 const TensorInfo input_weights(TensorShape(input_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
332 ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(inputs_weights_vector, &input_weights, Window::DimY));
333
334 // _concat_recurrent_weights
335 std::vector<const ITensorInfo *> recurrent_weights_vector;
336 recurrent_weights_vector.emplace_back(recurrent_to_input_weights);
337 recurrent_weights_vector.emplace_back(recurrent_to_forget_weights);
338 recurrent_weights_vector.emplace_back(recurrent_to_cell_weights);
339 recurrent_weights_vector.emplace_back(recurrent_to_output_weights);
340 const TensorInfo recurrent_weights(TensorShape(output_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
341 ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(recurrent_weights_vector, &recurrent_weights, Window::DimY));
342
343 // _concat_weights
344 std::vector<const ITensorInfo *> weights_vector;
345 weights_vector.emplace_back(&recurrent_weights);
346 weights_vector.emplace_back(&input_weights);
347 const TensorInfo weights(TensorShape(input_size + output_size, 4 * output_size), 1, DataType::QASYMM8, qweights);
348 ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(weights_vector, &weights, Window::DimX));
349 // _transpose_weights
350 const TensorShape weights_transposed_shape(weights.tensor_shape()[1], weights.tensor_shape()[0]);
351 TensorInfo weights_transposed = weights.clone()->set_is_resizable(true).set_tensor_shape(weights_transposed_shape);
352 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(&weights, &weights_transposed));
353
354 // _concat_inputs
355 std::vector<const ITensorInfo *> input_vector;
356 input_vector.emplace_back(input);
357 input_vector.emplace_back(output_state_in);
358 TensorInfo input_concatenated(TensorShape(output_size + input_size, batch_size), 1, DataType::QASYMM8, qasymm);
359 ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(input_vector, &input_concatenated, Window::DimX));
360
361 // _concat_bias
362 std::vector<const ITensorInfo *> bias_vector;
363 bias_vector.emplace_back(input_gate_bias);
364 bias_vector.emplace_back(forget_gate_bias);
365 bias_vector.emplace_back(cell_bias);
366 bias_vector.emplace_back(output_gate_bias);
367
368 const TensorInfo bias_concatenated(TensorShape(4 * output_size), 1, DataType::S32);
369 ARM_COMPUTE_RETURN_ON_ERROR(CLConcatenateLayer::validate(bias_vector, &bias_concatenated, Window::DimX));
370
371 // Invert the offset for gemmlowp
372 input_concatenated.set_quantization_info(QuantizationInfo(qasymm.uniform().scale, -qasymm.uniform().offset));
373 weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, -qweights.uniform().offset));
374
375 // _gemmlowp
376 const TensorInfo output_highp(TensorShape(4 * output_size, batch_size), 1, DataType::S32);
377 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(&input_concatenated, &weights_transposed, nullptr, &output_highp));
378
379 // Set the offset back
380 input_concatenated.set_quantization_info(QuantizationInfo(qasymm.uniform().scale, qasymm.uniform().offset));
381 weights_transposed.set_quantization_info(QuantizationInfo(qweights.uniform().scale, qweights.uniform().offset));
382
383 const TensorInfo output_lowp(output_highp.tensor_shape(), 1, DataType::QSYMM16, qsymm_3);
384
385 const float multiplier = 4096.f * qasymm.uniform().scale * qweights.uniform().scale;
386 int output_multiplier = 0;
387 int output_shift = 0;
388 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
389
390 // _output_stage
391 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&output_highp, &bias_concatenated, &output_lowp));
392
393 TensorInfo input_gate_input;
394 TensorInfo forget_gate_input;
395 TensorInfo input_modulation_gate_input;
396 TensorInfo output_gate_input;
397
398 if(batch_size > 1)
399 {
400 // _slice_input_tensor
401 input_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
402 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &input_gate_input, { 0, 0 }, { output_size, batch_size }));
403 // _slice_forget_tensor
404 forget_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
405 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &forget_gate_input, { output_size, 0 }, { 2 * output_size, batch_size }));
406 // _slice_cell_tensor
407 input_modulation_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
408 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &input_modulation_gate_input, { 2 * output_size, 0 }, { 3 * output_size, batch_size }));
409 // _slice_output_tensor
410 output_gate_input = TensorInfo(TensorShape(output_size, batch_size), 1, DataType::QSYMM16, qsymm_3);
411 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &output_gate_input, { 3 * output_size, 0 }, { 4 * output_size, batch_size }));
412 }
413 else
414 {
415 // _slice_input_tensor
416 input_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
417 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &input_gate_input, { 0 }, { output_size }));
418 // _slice_forget_tensor
419 forget_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
420 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &forget_gate_input, { output_size }, { 2 * output_size }));
421 // _slice_cell_tensor
422 input_modulation_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
423 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &input_modulation_gate_input, { 2 * output_size }, { 3 * output_size }));
424 // _slice_output_tensor
425 output_gate_input = TensorInfo(TensorShape(output_size), 1, DataType::QSYMM16, qsymm_3);
426 ARM_COMPUTE_RETURN_ON_ERROR(CLSlice::validate(&output_lowp, &output_gate_input, { 3 * output_size }, { 4 * output_size }));
427 }
428
429 // _sigmoid_forget_gate
430 const TensorInfo forget_gate_output(forget_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
431 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&forget_gate_input, &forget_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
432 // _sigmoid_input_gate
433 const TensorInfo input_gate_output(input_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
434 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_gate_input, &input_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
435 // _tanh_modulation_gate
436 const TensorInfo input_modulation_gate_output(input_modulation_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
437 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_modulation_gate_input, &input_modulation_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f)));
438 // _sigmoid_output_gate
439 const TensorInfo output_gate_output(output_gate_input.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
440 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_gate_input, &output_gate_output, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
441
442 // _mul_forget_gate_cell_state
443 const TensorInfo cell_state_tmp1(forget_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_4);
444 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&forget_gate_output, cell_state_in, &cell_state_tmp1, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
445
446 // _mul_input_gate_input_mod_gate
447 const TensorInfo cell_state_tmp2(input_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_4);
448 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&input_gate_output, &input_modulation_gate_output, &cell_state_tmp2, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
449
450 // _add_cell_state_tmps
451 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp1, &cell_state_tmp2, cell_state_out, ConvertPolicy::SATURATE));
452
453 // _tanh_modulation_gate
454 const TensorInfo output_state_tmp(cell_state_out->tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
455 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(cell_state_out, &output_state_tmp, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f)));
456
457 // _mul_output_state_tmp_output_gate
458 const TensorInfo output_state_out_symm(output_gate_output.tensor_shape(), 1, DataType::QSYMM16, qsymm_0);
459 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&output_state_tmp, &output_gate_output, &output_state_out_symm, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
460
461 // _dequantize
462 const TensorInfo output_state_out_f32(output_state_out_symm.tensor_shape(), 1, DataType::F32);
463 ARM_COMPUTE_RETURN_ON_ERROR(CLDequantizationLayer::validate(&output_state_out_symm, &output_state_out_f32));
464
465 // _quantize
466 ARM_COMPUTE_RETURN_ON_ERROR(CLQuantizationLayer::validate(&output_state_out_f32, output_state_out));
467
468 if(cell_state_out->total_size() != 0)
469 {
470 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&cell_state_info, cell_state_out);
471 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&cell_state_info, cell_state_out);
472 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&cell_state_info, cell_state_out);
473 }
474
475 if(output_state_out->total_size() != 0)
476 {
477 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(&output_state_info, output_state_out);
478 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&output_state_info, output_state_out);
479 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(&output_state_info, output_state_out);
480 }
481
482 return Status{};
483 }
484
run()485 void CLLSTMLayerQuantized::run()
486 {
487 prepare();
488
489 // Acquire all the temporaries
490 MemoryGroupResourceScope scope_mg(_memory_group);
491
492 // Concat and transpose the input
493 _concat_inputs.run();
494
495 // Run gemmlowp
496 _gemmlowp.run();
497 _output_stage.run();
498
499 // Slice the results
500 _slice_input_tensor.run();
501 _slice_forget_tensor.run();
502 _slice_cell_tensor.run();
503 _slice_output_tensor.run();
504
505 // Gates
506 // Forget gate
507 _sigmoid_forget_gate.run();
508
509 // Input gate
510 _sigmoid_input_gate.run();
511
512 // Input modulation gate
513 _tanh_modulation_gate.run();
514
515 // Output gate
516 _sigmoid_output_gate.run();
517
518 // Cell state (long term memory)
519 _mul_forget_gate_cell_state.run();
520 _mul_input_gate_input_mod_gate.run();
521 _add_cell_state_tmps.run();
522
523 // Output state (short term memory)
524 _tanh_output_state.run();
525 _mul_output_state_tmp_output_gate.run();
526
527 // Requantize output state from QSYMM16 to QASYMM8
528 _dequantize.run();
529 _quantize.run();
530 }
531
prepare()532 void CLLSTMLayerQuantized::prepare()
533 {
534 if(!_is_prepared)
535 {
536 _input_weights.allocator()->allocate();
537 _concat_input_weights.run();
538
539 _input_to_input_weights->mark_as_unused();
540 _input_to_forget_weights->mark_as_unused();
541 _input_to_cell_weights->mark_as_unused();
542 _input_to_output_weights->mark_as_unused();
543
544 _recurrent_weights.allocator()->allocate();
545 _concat_recurrent_weights.run();
546 _recurrent_to_input_weights->mark_as_unused();
547 _recurrent_to_forget_weights->mark_as_unused();
548 _recurrent_to_cell_weights->mark_as_unused();
549 _recurrent_to_output_weights->mark_as_unused();
550
551 _weights.allocator()->allocate();
552 _concat_weights.run();
553
554 _input_weights.mark_as_unused();
555 _input_weights.allocator()->free();
556 _recurrent_weights.mark_as_unused();
557 _recurrent_weights.allocator()->free();
558
559 _weights_transposed.allocator()->allocate();
560 _transpose_weights.run();
561
562 _weights.mark_as_unused();
563 _weights.allocator()->free();
564
565 _bias.allocator()->allocate();
566 _concat_bias.run();
567 _input_gate_bias->mark_as_unused();
568 _forget_gate_bias->mark_as_unused();
569 _cell_bias->mark_as_unused();
570 _output_gate_bias->mark_as_unused();
571
572 _is_prepared = true;
573 }
574 }
575
576 } // namespace arm_compute
577