Michalis Spyrou | 373b407 | 2021-01-20 16:41:12 +0000 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2021 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 | #include "src/runtime/cpu/operators/CpuSoftmax.h" |
| 25 | |
| 26 | #include "arm_compute/core/Helpers.h" |
| 27 | #include "arm_compute/core/TensorInfo.h" |
| 28 | #include "arm_compute/core/Validate.h" |
| 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 30 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 31 | #include "src/core/cpu/kernels/CpuSoftmaxKernel.h" |
| 32 | #include "src/core/helpers/SoftmaxHelpers.h" |
| 33 | |
| 34 | namespace arm_compute |
| 35 | { |
| 36 | namespace cpu |
| 37 | { |
| 38 | template <bool IS_LOG> |
| 39 | CpuSoftmaxGeneric<IS_LOG>::CpuSoftmaxGeneric() |
| 40 | : _permute_input(), _permute_output(), _max_kernel(), _softmax_kernel(), _max(nullptr), _tmp(nullptr), _input_permuted(nullptr), _output_permuted(nullptr), _needs_permute(false) |
| 41 | { |
| 42 | } |
| 43 | |
| 44 | template <bool IS_LOG> |
| 45 | void CpuSoftmaxGeneric<IS_LOG>::configure(const ITensorInfo *src, ITensorInfo *dst, float beta, int32_t axis) |
| 46 | { |
| 47 | // Perform validation step |
| 48 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); |
| 49 | ARM_COMPUTE_ERROR_THROW_ON(CpuSoftmaxGeneric::validate(src, dst, beta, axis)); |
| 50 | |
| 51 | const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions()))); |
| 52 | |
| 53 | _needs_permute = actual_axis > 0; |
| 54 | |
| 55 | if(_needs_permute) |
| 56 | { |
| 57 | _input_permuted = std::make_unique<TensorInfo>(); |
| 58 | _permute_input.configure(src, _input_permuted.get(), softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis)); |
| 59 | } |
| 60 | |
| 61 | // We want to deal with a 2D input. Either it is the permuted version of the original input (4D case) |
| 62 | // or it is the original input case (2D case) |
| 63 | const ITensorInfo *tmp_input = (_needs_permute ? _input_permuted.get() : src); |
| 64 | |
| 65 | // Create intermediate tensors shapes |
| 66 | TensorShape max_sum_shape = tmp_input->tensor_shape(); |
| 67 | max_sum_shape.set(0, 1); |
| 68 | const TensorInfo input_info = tmp_input->clone()->reset_padding().set_is_resizable(true); |
| 69 | DataType tmp_data_type = is_data_type_quantized_asymmetric(tmp_input->data_type()) ? DataType::F32 : tmp_input->data_type(); |
| 70 | TensorInfo tensor_info_tmp(input_info.clone()->set_data_type(tmp_data_type)); |
| 71 | TensorInfo max_info(tmp_input->clone()->set_tensor_shape(max_sum_shape)); |
| 72 | |
| 73 | // Init intermediate tensors |
| 74 | _max = std::make_unique<TensorInfo>(max_info); |
| 75 | _tmp = std::make_unique<TensorInfo>(tensor_info_tmp); |
| 76 | |
| 77 | // Configure kernels |
| 78 | auto mk = std::make_unique<kernels::CpuLogits1DMaxKernel>(); |
| 79 | mk->configure(tmp_input, _max.get()); |
| 80 | _max_kernel = std::move(mk); |
| 81 | |
| 82 | auto sm = std::make_unique<kernels::CpuLogits1DSoftmaxKernel<IS_LOG>>(); |
| 83 | if(_needs_permute) |
| 84 | { |
| 85 | _output_permuted = std::make_unique<TensorInfo>(); |
| 86 | |
| 87 | // The normalization kernel stores the result in a permuted output tensor |
| 88 | sm->configure(tmp_input, _max.get(), _output_permuted.get(), beta, _tmp.get()); |
| 89 | |
| 90 | // Re-permute the permuted output into the requested (4D) output |
| 91 | _permute_output.configure(_output_permuted.get(), dst, softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis)); |
| 92 | } |
| 93 | else |
| 94 | { |
| 95 | // Softmax 2D case |
| 96 | sm->configure(tmp_input, _max.get(), dst, beta, _tmp.get()); |
| 97 | } |
| 98 | _softmax_kernel = std::move(sm); |
| 99 | } |
| 100 | |
| 101 | template <bool IS_LOG> |
| 102 | Status CpuSoftmaxGeneric<IS_LOG>::validate(const ITensorInfo *src, const ITensorInfo *dst, float beta, int32_t axis) |
| 103 | { |
| 104 | // Perform validation step |
| 105 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); |
| 106 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(src->num_dimensions() > 4, "Only up to 4 dimensions are supported"); |
| 107 | ARM_COMPUTE_UNUSED(beta); |
| 108 | ARM_COMPUTE_RETURN_ERROR_ON(axis < static_cast<int32_t>(-src->num_dimensions()) || static_cast<int32_t>(src->num_dimensions()) <= axis); |
| 109 | |
| 110 | // Create intermediate tensor info |
| 111 | DataType tmp_data_type = src->data_type(); |
| 112 | const TensorInfo tensor_info_tmp(src->clone()->set_data_type(tmp_data_type).set_is_resizable(true)); |
| 113 | |
| 114 | TensorShape max_sum_shape = src->tensor_shape(); |
| 115 | max_sum_shape.set(0, 1); |
| 116 | const TensorInfo tensor_info_max_sum(src->clone()->set_tensor_shape(max_sum_shape).set_data_type(tmp_data_type).set_quantization_info(src->quantization_info()).set_is_resizable(true)); |
| 117 | const TensorInfo dont_care; |
| 118 | |
| 119 | const unsigned int actual_axis = static_cast<unsigned int>(wrap_around(axis, static_cast<int32_t>(src->num_dimensions()))); |
| 120 | |
| 121 | const bool needs_permute = actual_axis > 0; |
| 122 | |
| 123 | if(needs_permute) |
| 124 | { |
| 125 | const PermutationVector permutation_vector = softmax_helpers::get_permutation_vector_from_softmax_axis(actual_axis); |
| 126 | const TensorShape permuted_shape = misc::shape_calculator::compute_permutation_output_shape(*src, permutation_vector); |
| 127 | TensorInfo input_permuted(src->clone()->set_tensor_shape(permuted_shape)); |
| 128 | ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(src, &input_permuted, permutation_vector)); |
| 129 | TensorInfo output_permuted(dst->clone()->set_tensor_shape(permuted_shape)); |
| 130 | ARM_COMPUTE_RETURN_ON_ERROR(CpuPermute::validate(&output_permuted, dst, permutation_vector)); |
| 131 | } |
| 132 | |
| 133 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DMaxKernel::validate(src, &tensor_info_max_sum)); |
| 134 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuLogits1DSoftmaxKernel<IS_LOG>::validate(&tensor_info_tmp, &tensor_info_max_sum, dst, beta, &dont_care)); |
| 135 | |
| 136 | return Status{}; |
| 137 | } |
| 138 | |
| 139 | template <bool IS_LOG> |
| 140 | void CpuSoftmaxGeneric<IS_LOG>::run(ITensorPack &tensors) |
| 141 | { |
| 142 | ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided"); |
| 143 | |
| 144 | ITensorPack max_pack; |
| 145 | ITensorPack softmax_pack; |
| 146 | |
| 147 | if(_needs_permute) |
| 148 | { |
| 149 | ITensorPack permute_in_pack; |
| 150 | permute_in_pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC)); |
| 151 | permute_in_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_INT_2)); |
| 152 | _permute_input.run(permute_in_pack); |
| 153 | |
| 154 | max_pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(ACL_INT_2)); |
| 155 | |
| 156 | softmax_pack.add_tensor(TensorType::ACL_SRC_0, tensors.get_tensor(ACL_INT_2)); |
| 157 | softmax_pack.add_tensor(TensorType::ACL_SRC_1, tensors.get_tensor(ACL_INT_1)); |
| 158 | softmax_pack.add_tensor(TensorType::ACL_DST_0, tensors.get_tensor(ACL_INT_3)); |
| 159 | softmax_pack.add_tensor(TensorType::ACL_DST_1, tensors.get_tensor(ACL_INT_0)); |
| 160 | } |
| 161 | else |
| 162 | { |
| 163 | max_pack.add_tensor(TensorType::ACL_SRC, tensors.get_const_tensor(ACL_SRC)); |
| 164 | softmax_pack.add_tensor(TensorType::ACL_SRC_0, tensors.get_const_tensor(ACL_SRC)); |
| 165 | softmax_pack.add_tensor(TensorType::ACL_SRC_1, tensors.get_tensor(ACL_INT_1)); |
| 166 | softmax_pack.add_tensor(TensorType::ACL_DST_0, tensors.get_tensor(ACL_DST)); |
| 167 | softmax_pack.add_tensor(TensorType::ACL_DST_1, tensors.get_tensor(ACL_INT_0)); |
| 168 | } |
| 169 | |
| 170 | max_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_INT_1)); |
| 171 | |
| 172 | NEScheduler::get().schedule_op(_max_kernel.get(), Window::DimY, _max_kernel->window(), max_pack); |
| 173 | NEScheduler::get().schedule_op(_softmax_kernel.get(), Window::DimY, _softmax_kernel->window(), softmax_pack); |
| 174 | |
| 175 | if(_needs_permute) |
| 176 | { |
| 177 | ITensorPack permute_out_pack; |
| 178 | permute_out_pack.add_tensor(TensorType::ACL_SRC, tensors.get_tensor(ACL_INT_3)); |
| 179 | permute_out_pack.add_tensor(TensorType::ACL_DST, tensors.get_tensor(ACL_DST)); |
| 180 | _permute_output.run(permute_out_pack); |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | template <bool IS_LOG> |
| 185 | experimental::MemoryRequirements CpuSoftmaxGeneric<IS_LOG>::workspace() const |
| 186 | { |
| 187 | experimental::MemoryRequirements req{}; |
| 188 | |
| 189 | req.push_back({ TensorType::ACL_INT_0, _tmp->total_size(), 0 }); |
| 190 | req.push_back({ TensorType::ACL_INT_1, _max->total_size(), 0 }); |
| 191 | |
| 192 | if(_needs_permute) |
| 193 | { |
| 194 | req.push_back({ TensorType::ACL_INT_2, _input_permuted->total_size(), 0 }); |
| 195 | req.push_back({ TensorType::ACL_INT_3, _output_permuted->total_size(), 0 }); |
| 196 | } |
| 197 | |
| 198 | return req; |
| 199 | } |
| 200 | |
| 201 | template class CpuSoftmaxGeneric<false>; |
| 202 | template class CpuSoftmaxGeneric<true>; |
| 203 | } // namespace cpu |
| 204 | } // namespace arm_compute |