2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #ifndef __ARM_COMPUTE_CLSOFTMAXLAYERKERNEL_H__
25 #define __ARM_COMPUTE_CLSOFTMAXLAYERKERNEL_H__
27 #include "arm_compute/core/CL/ICLSimple3DKernel.h"
35 /** Interface for the identifying the max value of 1D Logits */
36 class CLLogits1DMaxKernel : public ICLSimple3DKernel
39 /** Set the input and output tensors.
41 * @param[in] input Source tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32
42 * @param[out] output Destination tensor. Data types supported: same as @p input
44 void configure(const ICLTensor *input, ICLTensor *output);
45 /** Static function to check if given info will lead to a valid configuration of @ref CLLogits1DMaxKernel
47 * @param[in] input Source tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32
48 * @param[in] output Destination tensor. Data types supported: same as @p input
52 static Status validate(const ITensorInfo *input, const ITensorInfo *output);
55 /** Interface for shifting, exponentiating and summing the logits */
56 class CLLogits1DShiftExpSumKernel : public ICLKernel
59 /** Default constructor */
60 CLLogits1DShiftExpSumKernel();
61 /** Prevent instances of this class from being copied (As this class contains pointers) */
62 CLLogits1DShiftExpSumKernel(const CLLogits1DShiftExpSumKernel &) = delete;
63 /** Prevent instances of this class from being copied (As this class contains pointers) */
64 CLLogits1DShiftExpSumKernel &operator=(const CLLogits1DShiftExpSumKernel &) = delete;
65 /** Allow instances of this class to be moved */
66 CLLogits1DShiftExpSumKernel(CLLogits1DShiftExpSumKernel &&) = default;
67 /** Allow instances of this class to be moved */
68 CLLogits1DShiftExpSumKernel &operator=(CLLogits1DShiftExpSumKernel &&) = default;
69 /** Set the input and output tensors.
71 * @param[in] input Source tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32
72 * @param[in] max Max values tensor. Data types supported: same as @p input
73 * @param[out] output Destination tensor. Data types supported: S32 for QASYMM8 @p input, or same as @p input
74 * @param[out] sum Sum of 1D logits tensor. Data types supported: S32 for QASYMM8 @p input, or same as @p input
75 * @param[in] beta (Optional) A scaling factor for the exponent. Defaults to 1.0
77 void configure(const ICLTensor *input, const ICLTensor *max, ICLTensor *output, ICLTensor *sum, float beta = 1.0f);
78 /** Static function to check if given info will lead to a valid configuration of @ref CLLogits1DShiftExpSumKernel
80 * @param[in] input Source tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32
81 * @param[in] max Max values tensor. Data types supported: same as @p input
82 * @param[in] output Destination tensor. Data types supported: S32 for QASYMM8 @p input, or same as @p input
83 * @param[in] sum Sum of 1D logits tensor. Data types supported: S32 for QASYMM8 @p input, or same as @p input
87 static Status validate(const ITensorInfo *input, const ITensorInfo *max, const ITensorInfo *output, const ITensorInfo *sum);
89 // Inherited methods overridden:
90 void run(const Window &window, cl::CommandQueue &queue) override;
93 const ICLTensor *_input;
94 const ICLTensor *_max;
99 /** Interface for max, shifting, exponentiating and summing the logits */
100 class CLLogits1DMaxShiftExpSumKernel : public ICLKernel
103 /** Info for whether a parallel reduction will be run and the vector size of the execution. */
104 using ParallelReductionInfo = std::tuple<bool, unsigned int>;
107 /** Default constructor */
108 CLLogits1DMaxShiftExpSumKernel();
109 /** Prevent instances of this class from being copied (As this class contains pointers) */
110 CLLogits1DMaxShiftExpSumKernel(const CLLogits1DMaxShiftExpSumKernel &) = delete;
111 /** Prevent instances of this class from being copied (As this class contains pointers) */
112 CLLogits1DMaxShiftExpSumKernel &operator=(const CLLogits1DMaxShiftExpSumKernel &) = delete;
113 /** Allow instances of this class to be moved */
114 CLLogits1DMaxShiftExpSumKernel(CLLogits1DMaxShiftExpSumKernel &&) = default;
115 /** Allow instances of this class to be moved */
116 CLLogits1DMaxShiftExpSumKernel &operator=(CLLogits1DMaxShiftExpSumKernel &&) = default;
117 /** Set the input and output tensors.
119 * @param[in] input Source tensor. Data types supported: QS8/QS16/F16/F32
120 * @param[in,out] max Max values tensor. Data types supported: same as @p input
121 * @param[out] output Destination tensor. Data types supported: same as @p input
122 * @param[out] sum Sum of 1D logits tensor. Data types supported: same as @p input
123 * @param[in] beta (Optional) A scaling factor for the exponent. Defaults to 1.f
125 void configure(const ICLTensor *input, ICLTensor *max, ICLTensor *output, ICLTensor *sum, float beta = 1.0f);
126 /** Static function to check if given info will lead to a valid configuration of @ref CLLogits1DMaxShiftExpSumKernel
128 * @param[in] input Source tensor. Data types supported: QS8/QS16/F16/F32
129 * @param[in] max Max values tensor. Data types supported: same as @p input
130 * @param[in] output Destination tensor. Data types supported: same as @p input
131 * @param[in] sum Sum of 1D logits tensor. Data types supported: same as @p input
135 static Status validate(const ITensorInfo *input, const ITensorInfo *max, const ITensorInfo *output, const ITensorInfo *sum);
136 /** Checks if the given size is eligible for parallel reduction
138 * @note Serial reduction is launched for width < (_grid_size * _serial_vector_size).
139 * @note Parallel reduction is launched for width >= (_grid_size * _serial_vector_size) and vector_size is forced to 4.
141 * @param[in] size Size to check
143 * @return A two-element tuple where the first element is a boolean specifying if a parallel reduction will be run,
144 * while the second element is the vector size of the execution.
146 static ParallelReductionInfo is_parallel_reduction(size_t size);
148 // Inherited methods overridden:
149 void run(const Window &window, cl::CommandQueue &queue) override;
152 const ICLTensor *_input;
158 static const unsigned int _grid_size;
159 static const unsigned int _serial_vector_size;
160 static const unsigned int _parallel_vector_size;
162 /** Interface for calculating the final step of the Softmax Layer where each logit value is multiplied by the inverse of the sum of the logits. */
163 class CLLogits1DNormKernel : public ICLKernel
166 /** Default constructor */
167 CLLogits1DNormKernel();
168 /** Prevent instances of this class from being copied (As this class contains pointers) */
169 CLLogits1DNormKernel(const CLLogits1DNormKernel &) = delete;
170 /** Prevent instances of this class from being copied (As this class contains pointers) */
171 CLLogits1DNormKernel &operator=(const CLLogits1DNormKernel &) = delete;
172 /** Allow instances of this class to be moved */
173 CLLogits1DNormKernel(CLLogits1DNormKernel &&) = default;
174 /** Allow instances of this class to be moved */
175 CLLogits1DNormKernel &operator=(CLLogits1DNormKernel &&) = default;
176 /** Set the input and output tensors.
178 * @param[in] input Source tensor. Data types supported: QS8/QS16/S32/F16/F32
179 * @param[in] sum Sum tensor. Dimensions should be dim(input)-1. Data types supported: same as @p input
180 * @param[out] output Destination tensor. Data types supported: QASYMM8 for S32 @p input, or same as @p input
181 * @param[in] beta (Optional) A scaling factor for the exponent. (Default = 1.0)
183 void configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, float beta = 1.0f);
184 /** Static function to check if given info will lead to a valid configuration of @ref CLLogits1DNormKernel
186 * @param[in] input Source tensor. Data types supported: QS8/QS16/S32/F16/F32
187 * @param[in] sum Sum tensor. Dimensions should be dim(input)-1. Data types supported: same as @p input
188 * @param[in] output Destination tensor. Data types supported: QASYMM8 for S32 @p input, or same as @p input
192 static Status validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output);
194 // Inherited methods overridden:
195 void run(const Window &window, cl::CommandQueue &queue) override;
198 const ICLTensor *_input;
199 const ICLTensor *_sum;
202 } // namespace arm_compute
203 #endif /*__ARM_COMPUTE_CLSOFTMAXLAYERKERNEL_H__ */