Gitiles
Code Review Sign In
review.mlplatform.org / ml / ComputeLibrary / 70d33bdfd36e1b44b0573189dca67ed7c63dd59e / . / arm_compute / runtime / CL / functions / CLLSTMLayerQuantized.h
blob: 1d390600885b20b184e091be0c615565ce79a4d4 [file] [log] [blame]
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]1/*
2 * Copyright (c) 2019 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 */
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 *
76 * @param[in] input Source tensor. Input is a 2D tensor with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
77 * @param[in] input_to_input_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
78 * @param[in] input_to_forget_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
79 * @param[in] input_to_cell_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
80 * @param[in] input_to_output_weights 2D weights tensor with dimensions [input_size, output_size]. Data type supported: Same as @p input.
81 * @param[in] recurrent_to_input_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
82 * @param[in] recurrent_to_forget_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
83 * @param[in] recurrent_to_cell_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
84 * @param[in] recurrent_to_output_weights 2D weights tensor with dimensions [output_size, output_size]. Data type supported: Same as @p input.
85 * @param[in] input_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
86 * @param[in] forget_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
87 * @param[in] cell_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
88 * @param[in] output_gate_bias 1D weights tensor with dimensions [output_size]. Data type supported: S32.
89 * @param[in] cell_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
90 * @param[in] output_state_in 2D tensor with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
91 * @param[out] cell_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
92 * @param[out] output_state_out Destination tensor. Output is a 2D tensor with dimensions [output_size, batch_size].Data types supported: Same as @p input.
93 */
94 void configure(const ICLTensor *input,
95 const ICLTensor *input_to_input_weights, const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
96 const ICLTensor *recurrent_to_input_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
97 const ICLTensor *input_gate_bias, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
98 ICLTensor *cell_state_in, const ICLTensor *output_state_in,
99 ICLTensor *cell_state_out, ICLTensor *output_state_out);
100
101 /** Static function to check if given info will lead to a valid configuration of @ref CLLSTMLayerQuantized
102 *
103 * @param[in] input Source tensor info. Input is a 2D tensor info with dimensions [input_size, batch_size]. Data types supported: QASYMM8.
104 * @param[in] input_to_input_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
105 * @param[in] input_to_forget_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
106 * @param[in] input_to_cell_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
107 * @param[in] input_to_output_weights 2D weights tensor info with dimensions [input_size, output_size]. Data type supported: Same as @p input.
108 * @param[in] recurrent_to_input_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
109 * @param[in] recurrent_to_forget_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
110 * @param[in] recurrent_to_cell_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
111 * @param[in] recurrent_to_output_weights 2D weights tensor info with dimensions [output_size, output_size]. Data type supported: Same as @p input.
112 * @param[in] input_gate_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
113 * @param[in] forget_gate_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
114 * @param[in] cell_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
115 * @param[in] output_gate_bias 1D weights tensor info with dimensions [output_size]. Data type supported: S32.
116 * @param[in] cell_state_in 2D tensor info with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
117 * @param[in] output_state_in 2D tensor info with dimensions [output_size, batch_size]. Data type supported: Same as @p input.
118 * @param[out] cell_state_out Destination tensor info. Output is a 2D tensor info with dimensions [output_size, batch_size]. Data type supported: QSYMM16.
119 * @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.
120 *
121 * @return a status
122 */
123 static Status validate(const ITensorInfo *input,
124 const ITensorInfo *input_to_input_weights, const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
125 const ITensorInfo *recurrent_to_input_weights, const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
126 const ITensorInfo *input_gate_bias, const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
127 const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
128 const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out);
129
130 // Inherited methods overridden:
131 void run() override;
132 void prepare() override;
133
134private:
Georgios Pinitas26014cf2019-09-09 19:00:57 +0100[diff] [blame]135 MemoryGroup _memory_group;
Manuel Bottini10c53f12019-07-17 16:11:53 +0100[diff] [blame]136
137 // Functions used
138 CLGEMMLowpMatrixMultiplyCore _gemmlowp;
139 CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint _output_stage;
140 CLTranspose _transpose_weights;
141 CLConcatenateLayer _concat_input_weights;
142 CLConcatenateLayer _concat_recurrent_weights;
143 CLConcatenateLayer _concat_weights;
144 CLConcatenateLayer _concat_inputs;
145 CLConcatenateLayer _concat_bias;
146 CLActivationLayer _sigmoid_forget_gate;
147 CLActivationLayer _sigmoid_input_gate;
148 CLActivationLayer _sigmoid_output_gate;
149 CLActivationLayer _tanh_modulation_gate;
150 CLActivationLayer _tanh_output_state;
151 CLArithmeticAddition _add_cell_state_tmps;
152 CLArithmeticAddition _add2;
153 CLPixelWiseMultiplication _mul_forget_gate_cell_state;
154 CLPixelWiseMultiplication _mul_input_gate_input_mod_gate;
155 CLPixelWiseMultiplication _mul_output_state_tmp_output_gate;
156 CLSlice _slice_input_tensor;
157 CLSlice _slice_forget_tensor;
158 CLSlice _slice_cell_tensor;
159 CLSlice _slice_output_tensor;
160 CLDequantizationLayer _dequantize;
161 CLQuantizationLayer _quantize;
162
163 // Tensor pointers
164 const ICLTensor *_input_to_input_weights;
165 const ICLTensor *_input_to_forget_weights;
166 const ICLTensor *_input_to_cell_weights;
167 const ICLTensor *_input_to_output_weights;
168 const ICLTensor *_recurrent_to_input_weights;
169 const ICLTensor *_recurrent_to_forget_weights;
170 const ICLTensor *_recurrent_to_cell_weights;
171 const ICLTensor *_recurrent_to_output_weights;
172 const ICLTensor *_input_gate_bias;
173 const ICLTensor *_forget_gate_bias;
174 const ICLTensor *_cell_bias;
175 const ICLTensor *_output_gate_bias;
176
177 // Temporary tensors
178 CLTensor _recurrent_weights;
179 CLTensor _input_weights;
180 CLTensor _weights;
181 CLTensor _input;
182 CLTensor _weights_transposed;
183 CLTensor _output_highp;
184 CLTensor _output_lowp;
185 CLTensor _bias;
186 CLTensor _forget_gate_input;
187 CLTensor _input_gate_input;
188 CLTensor _output_gate_input;
189 CLTensor _input_modulation_gate_input;
190 CLTensor _forget_gate_output;
191 CLTensor _input_gate_output;
192 CLTensor _output_gate_output;
193 CLTensor _input_modulation_gate_output;
194 CLTensor _cell_state_tmp1;
195 CLTensor _cell_state_tmp2;
196 CLTensor _output_state_tmp;
197 CLTensor _output_state_out_symm;
198 CLTensor _output_state_out_f32;
199
200 bool _is_prepared;
201};
202} // namespace arm_compute
Michalis Spyrouf4643372019-11-29 16:17:13 +0000[diff] [blame]203#endif /* ARM_COMPUTE_CLLSTMLAYERQUANTIZED_H */