1 /* 2 * Copyright (c) 2017-2020 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 #ifndef ARM_COMPUTE_NESOFTMAXLAYER_H 25 #define ARM_COMPUTE_NESOFTMAXLAYER_H 26 27 #include "arm_compute/runtime/IFunction.h" 28 #include "arm_compute/runtime/MemoryGroup.h" 29 #include "arm_compute/runtime/NEON/functions/NEPermute.h" 30 #include "arm_compute/runtime/Tensor.h" 31 #include <memory> 32 33 namespace arm_compute 34 { 35 class ITensor; 36 class NELogits1DMaxKernel; 37 template <bool IS_LOG> 38 class NELogits1DSoftmaxKernel; 39 class NEFillBorderKernel; 40 41 /** Basic function to compute a SoftmaxLayer and a Log SoftmaxLayer. 42 * 43 * Softmax is calculated by : 44 * @f[ out = exp((x - max(x)) * beta) / sum(exp((x - max(x)) * beta)) @f] 45 * 46 * Log Softmax is calculated by : 47 * @f[ out = (x - max(x) * beta) - log(\sum{e^{x - max(x) * beta}}) @f] 48 * 49 * This function runs the following function/kernels: 50 * -# If axis is not 0: 51 * -# @ref NEPermute 52 * -# @ref NEFillBorderKernel 53 * -# @ref NELogits1DMaxKernel 54 * -# @ref NELogits1DSoftmaxKernel 55 */ 56 template <bool IS_LOG = false> 57 class NESoftmaxLayerGeneric : public IFunction 58 { 59 public: 60 /** Constructor */ 61 NESoftmaxLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr); 62 /** Prevent instances of this class from being copied (As this class contains pointers) */ 63 NESoftmaxLayerGeneric(const NESoftmaxLayerGeneric &) = delete; 64 /** Default move constructor */ 65 NESoftmaxLayerGeneric(NESoftmaxLayerGeneric &&) = default; 66 /** Prevent instances of this class from being copied (As this class contains pointers) */ 67 NESoftmaxLayerGeneric &operator=(const NESoftmaxLayerGeneric &) = delete; 68 /** Default move assignment operator */ 69 NESoftmaxLayerGeneric &operator=(NESoftmaxLayerGeneric &&) = default; 70 /** Default destructor */ 71 ~NESoftmaxLayerGeneric(); 72 /** Set the input and output tensors. 73 * 74 * @param[in,out] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. If the width is not a 75 * multiple of the internal processing block size, @ref NEFillBorderKernel replicates the 76 * last value of each row to the nearest multiple. 77 * @param[out] output Destination tensor. Data types supported: same as @p input. 78 * @param[in] beta (Optional) A scaling factor for the exponent. 79 * @param[in] axis (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and 80 * axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0 81 */ 82 void configure(ITensor *input, ITensor *output, float beta = 1.0f, int32_t axis = 0); 83 /** Static function to check if given info will lead to a valid configuration of @ref NESoftmaxLayer 84 * 85 * @param[in] input Source tensor info. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. 86 * @param[in] output Destination tensor info. Data types supported: same as @p input 87 * @param[in] beta (Optional) A scaling factor for the exponent. 88 * @param[in] axis (Optional) The dimension in which to apply the function. E.g. for input of shape 4x5x6 and 89 * axis=1, softmax will be applied to 4x6=24 vectors of size 5. Defaults to 0 90 * 91 * @return a status 92 */ 93 static Status validate(const ITensorInfo *input, const ITensorInfo *output, float beta = 1.0f, int32_t axis = 0); 94 95 // Inherited methods overridden: 96 void run() override; 97 98 private: 99 MemoryGroup _memory_group; 100 NEPermute _permute_input; 101 NEPermute _permute_output; 102 std::unique_ptr<NELogits1DMaxKernel> _max_kernel; 103 std::unique_ptr<NELogits1DSoftmaxKernel<IS_LOG>> _softmax_kernel; 104 std::unique_ptr<NEFillBorderKernel> _fill_border_kernel; 105 Tensor _max; 106 Tensor _tmp; 107 Tensor _input_permuted; 108 Tensor _output_permuted; 109 bool _needs_permute; 110 }; 111 112 using NESoftmaxLayer = NESoftmaxLayerGeneric<false>; 113 using NELogSoftmaxLayer = NESoftmaxLayerGeneric<true>; 114 115 } // namespace arm_compute 116 #endif /* ARM_COMPUTE_NESOFTMAXLAYER_H */ 117