Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 6124ce60b54eb5639ed19d46c79fce21cca2c83b / . / arm_compute / runtime / CL / functions / CLLSTMLayerQuantized.h
blob: 2ef7427a5a345b6a9c845c03617dad691761d3da [file] [log] [blame]
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]1/*
Teresa Charlin62687422021-04-28 10:58:49 +0100[diff] [blame]2 * Copyright (c) 2019-2021 Arm Limited.
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]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 */
Michalis Spyrouf4643372019-11-29 16:17:13 +0000[diff] [blame]24#ifndef ARM_COMPUTE_CLLSTMLAYERQUANTIZED_H
25#define ARM_COMPUTE_CLLSTMLAYERQUANTIZED_H
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]26
27#include "arm_compute/core/Types.h"
28#include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
29#include "arm_compute/runtime/CL/functions/CLConcatenateLayer.h"
30#include "arm_compute/runtime/CL/functions/CLDequantizationLayer.h"
31#include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
32#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
33#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
34#include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h"
35#include "arm_compute/runtime/CL/functions/CLQuantizationLayer.h"
36#include "arm_compute/runtime/CL/functions/CLSlice.h"
37#include "arm_compute/runtime/CL/functions/CLTranspose.h"
38
39#include "arm_compute/runtime/common/LSTMParams.h"
40
41namespace arm_compute
42{
43// Forward declarations
44class ICLTensor;
45
46/** Basic function to run @ref CLLSTMLayerQuantized
47 *
48 * This function calls the following CL functions/kernels:
49 *
50 * -# @ref CLGEMMLowpMatrixMultiplyCore Quantized matrix multiplication core. Accumulators are 32-bit integers
51 * -# @ref CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint Convert 32-bit integers into QSYMM16
52 * -# @ref CLTranspose Matrix transpose
53 * -# @ref CLConcatenateLayer Tensor concatenation
54 * -# @ref CLActivationLayer Activation functions (tanh and logistic)
55 * -# @ref CLArithmeticAddition Elementwise addition
56 * -# @ref CLPixelWiseMultiplication Elementwise multiplication
57 * -# @ref CLSlice Tensor slicing
58 * -# @ref CLDequantizationLayer Dequantize into float
59 * -# @ref CLQuantizationLayer Quantize from float
60 * */
61class CLLSTMLayerQuantized : public IFunction
62{
63public:
64 /** Default constructor */
65 CLLSTMLayerQuantized(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
66 /** Prevent instances of this class from being copied (As this class contains pointers) */
67 CLLSTMLayerQuantized(const CLLSTMLayerQuantized &) = delete;
68 /** Default move constructor */
69 CLLSTMLayerQuantized(CLLSTMLayerQuantized &&) = default;
70 /** Prevent instances of this class from being copied (As this class contains pointers) */
71 CLLSTMLayerQuantized &operator=(const CLLSTMLayerQuantized &) = delete;
72 /** Default move assignment operator */
73 CLLSTMLayerQuantized &operator=(CLLSTMLayerQuantized &&) = default;
74 /** Initialize function's tensors.
75 *
Teresa Charlin62687422021-04-28 10:58:49 +0100[diff] [blame]76 * Valid data layouts:
77 * - All
78 *
79 * Valid data type configurations:
80 * |src0 - src8 |src9 - src12 |src13 |src14 |dst0 |dst1 |
81 * |:-----------|:------------|:-------|:------|:------|:------|
82 * |QASYMM8 |S32 |QSYMM16 |QASYMM8|QSYMM16|QASYMM8|
83 *
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]84 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
85 * @param[in] input_to_input_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
86 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
87 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
88 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
89 * @param[in] recurrent_to_input_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
90 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
91 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
92 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
93 * @param[in] input_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
94 * @param[in] forget_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
95 * @param[in] cell_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
96 * @param[in] output_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
97 * @param[in] cell_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
98 * @param[in] output_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
99 * @param[out] cell_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
100 * @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.
101 */
102 void configure(const ICLTensor *input,
103 const ICLTensor *input_to_input_weights, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
104 const ICLTensor *recurrent_to_input_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
105 const ICLTensor *input_gate_bias, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
106 ICLTensor *cell_state_in, const ICLTensor *output_state_in,
107 ICLTensor *cell_state_out, ICLTensor *output_state_out);
Manuel Bottini2b84be52020-04-08 10:15:51 +0100[diff] [blame]108 /** Initialize function's tensors.
109 *
110 * @param[in] compile_context The compile context to be used.
111 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
112 * @param[in] input_to_input_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
113 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
114 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
115 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
116 * @param[in] recurrent_to_input_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
117 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
118 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
119 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
120 * @param[in] input_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
121 * @param[in] forget_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
122 * @param[in] cell_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
123 * @param[in] output_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
124 * @param[in] cell_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
125 * @param[in] output_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
126 * @param[out] cell_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
127 * @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.
128 */
129 void configure(const CLCompileContext &compile_context, const ICLTensor *input,
130 const ICLTensor *input_to_input_weights, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
131 const ICLTensor *recurrent_to_input_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
132 const ICLTensor *input_gate_bias, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
133 ICLTensor *cell_state_in, const ICLTensor *output_state_in,
134 ICLTensor *cell_state_out, ICLTensor *output_state_out);
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]135
136 /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayerQuantized
137 *
138 * @param[in] input Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
139 * @param[in] input_to_input_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
140 * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
141 * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
142 * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
143 * @param[in] recurrent_to_input_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
144 * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
145 * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
146 * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
147 * @param[in] input_gate_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
148 * @param[in] forget_gate_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
149 * @param[in] cell_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
150 * @param[in] output_gate_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
151 * @param[in] cell_state_in 2D tensor info with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
152 * @param[in] output_state_in 2D tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
153 * @param[out] cell_state_out Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
154 * @param[out] 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.
155 *
156 * @return a status
157 */
158 static Status validate(const ITensorInfo *input,
159 const ITensorInfo *input_to_input_weights, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
160 const ITensorInfo *recurrent_to_input_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
161 const ITensorInfo *input_gate_bias, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
162 const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
163 const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out);
164
165 // Inherited methods overridden:
166 void run() override;
167 void prepare() override;
168
169private:
Georgios Pinitas26014cf2019-09-09 19:00:57 +0100[diff] [blame]170 MemoryGroup _memory_group;
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]171
172 // Functions used
173 CLGEMMLowpMatrixMultiplyCore _gemmlowp;
174 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint _output_stage;
175 CLTranspose _transpose_weights;
176 CLConcatenateLayer _concat_input_weights;
177 CLConcatenateLayer _concat_recurrent_weights;
178 CLConcatenateLayer _concat_weights;
179 CLConcatenateLayer _concat_inputs;
180 CLConcatenateLayer _concat_bias;
181 CLActivationLayer _sigmoid_forget_gate;
182 CLActivationLayer _sigmoid_input_gate;
183 CLActivationLayer _sigmoid_output_gate;
184 CLActivationLayer _tanh_modulation_gate;
185 CLActivationLayer _tanh_output_state;
186 CLArithmeticAddition _add_cell_state_tmps;
187 CLArithmeticAddition _add2;
188 CLPixelWiseMultiplication _mul_forget_gate_cell_state;
189 CLPixelWiseMultiplication _mul_input_gate_input_mod_gate;
190 CLPixelWiseMultiplication _mul_output_state_tmp_output_gate;
191 CLSlice _slice_input_tensor;
192 CLSlice _slice_forget_tensor;
193 CLSlice _slice_cell_tensor;
194 CLSlice _slice_output_tensor;
195 CLDequantizationLayer _dequantize;
196 CLQuantizationLayer _quantize;
197
198 // Tensor pointers
199 const ICLTensor *_input_to_input_weights;
200 const ICLTensor *_input_to_forget_weights;
201 const ICLTensor *_input_to_cell_weights;
202 const ICLTensor *_input_to_output_weights;
203 const ICLTensor *_recurrent_to_input_weights;
204 const ICLTensor *_recurrent_to_forget_weights;
205 const ICLTensor *_recurrent_to_cell_weights;
206 const ICLTensor *_recurrent_to_output_weights;
207 const ICLTensor *_input_gate_bias;
208 const ICLTensor *_forget_gate_bias;
209 const ICLTensor *_cell_bias;
210 const ICLTensor *_output_gate_bias;
211
212 // Temporary tensors
213 CLTensor _recurrent_weights;
214 CLTensor _input_weights;
215 CLTensor _weights;
216 CLTensor _input;
217 CLTensor _weights_transposed;
218 CLTensor _output_highp;
219 CLTensor _output_lowp;
220 CLTensor _bias;
221 CLTensor _forget_gate_input;
222 CLTensor _input_gate_input;
223 CLTensor _output_gate_input;
224 CLTensor _input_modulation_gate_input;
225 CLTensor _forget_gate_output;
226 CLTensor _input_gate_output;
227 CLTensor _output_gate_output;
228 CLTensor _input_modulation_gate_output;
229 CLTensor _cell_state_tmp1;
230 CLTensor _cell_state_tmp2;
231 CLTensor _output_state_tmp;
232 CLTensor _output_state_out_symm;
233 CLTensor _output_state_out_f32;
234
235 bool _is_prepared;
236};
237} // namespace arm_compute
Michalis Spyrouf4643372019-11-29 16:17:13 +0000[diff] [blame]238#endif /* ARM_COMPUTE_CLLSTMLAYERQUANTIZED_H */