Michele Di Giorgio | 9175392 | 2019-06-13 10:56:59 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 2 | * Copyright (c) 2019-2020 Arm Limited. |
Michele Di Giorgio | 9175392 | 2019-06-13 10:56:59 +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 Spyrou | ebcebf1 | 2020-10-21 00:04:14 +0100 | [diff] [blame] | 24 | #include "src/core/NEON/kernels/NEMeanStdDevNormalizationKernel.h" |
Michele Di Giorgio | 9175392 | 2019-06-13 10:56:59 +0100 | [diff] [blame] | 25 | |
Michele Di Giorgio | 9175392 | 2019-06-13 10:56:59 +0100 | [diff] [blame] | 26 | #include "arm_compute/core/Helpers.h" |
| 27 | #include "arm_compute/core/ITensor.h" |
Michele Di Giorgio | 9175392 | 2019-06-13 10:56:59 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/TensorInfo.h" |
| 29 | #include "arm_compute/core/Types.h" |
| 30 | #include "arm_compute/core/Window.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 31 | #include "src/core/CPP/Validate.h" |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 32 | #include "src/core/NEON/NEMath.h" |
| 33 | #include "src/core/NEON/wrapper/wrapper.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 34 | #include "src/core/helpers/AutoConfiguration.h" |
| 35 | #include "src/core/helpers/WindowHelpers.h" |
Michele Di Giorgio | 9175392 | 2019-06-13 10:56:59 +0100 | [diff] [blame] | 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | namespace |
| 40 | { |
| 41 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, float epsilon) |
| 42 | { |
| 43 | ARM_COMPUTE_UNUSED(epsilon); |
| 44 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); |
| 45 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); |
| 46 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() > 2, "Input tensor cannot have more than 2 dimensions"); |
| 47 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); |
| 48 | |
| 49 | // Checks performed when output is configured |
| 50 | if((output != nullptr) && (output->total_size() != 0)) |
| 51 | { |
| 52 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 53 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 54 | } |
| 55 | return Status{}; |
| 56 | } |
| 57 | |
| 58 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) |
| 59 | { |
| 60 | if(output != nullptr) |
| 61 | { |
| 62 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 63 | // Output auto inizialitation if not yet initialized |
| 64 | auto_init_if_empty(*output, *input); |
| 65 | } |
| 66 | |
| 67 | // This kernel doesn't need padding. A left-over for loop on dimension X, we cannot have any read or write out of memory |
| 68 | // For this reason num_elems_processed_per_iteration is set to 1 |
| 69 | Window win = calculate_max_window(*input, Steps()); |
| 70 | if(output != nullptr) |
| 71 | { |
| 72 | output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); |
| 73 | } |
| 74 | |
| 75 | return std::make_pair(Status{}, win); |
| 76 | } |
| 77 | } // namespace |
| 78 | |
| 79 | template <typename ScalarType, int size> |
| 80 | void NEMeanStdDevNormalizationKernel::mean_stddev_normalization(const Window &window) |
| 81 | { |
| 82 | using ExactTagType = typename wrapper::traits::neon_vector<ScalarType, size>::tag_type; |
| 83 | |
| 84 | // Set build options |
| 85 | Window win = window; |
| 86 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 87 | |
| 88 | const int window_step_x = size; |
| 89 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 90 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 91 | |
| 92 | Iterator input(_input, win); |
| 93 | Iterator output(_output, win); |
| 94 | |
| 95 | execute_window_loop(win, [&](const Coordinates &) |
| 96 | { |
| 97 | int x = window_start_x; |
| 98 | auto in_ptr = reinterpret_cast<const ScalarType *>(input.ptr()); |
| 99 | auto out_ptr = reinterpret_cast<ScalarType *>(output.ptr()); |
| 100 | |
| 101 | auto sum_vec = wrapper::vdup_n(static_cast<ScalarType>(0.f), ExactTagType{}); |
| 102 | auto sum_sq_vec = wrapper::vdup_n(static_cast<ScalarType>(0.f), ExactTagType{}); |
| 103 | |
| 104 | for(; x <= (window_end_x - window_step_x); x += window_step_x) |
| 105 | { |
| 106 | auto data = wrapper::vloadq(in_ptr + x); |
| 107 | sum_vec = wrapper::vadd(sum_vec, data); |
| 108 | sum_sq_vec = wrapper::vadd(sum_sq_vec, wrapper::vmul(data, data)); |
| 109 | } |
| 110 | |
| 111 | auto sum_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_vec), wrapper::vgetlow(sum_vec)); |
| 112 | auto sum_sq_carry_res = wrapper::vpadd(wrapper::vgethigh(sum_sq_vec), wrapper::vgetlow(sum_sq_vec)); |
| 113 | for(int i = 0; i < size / 4; ++i) |
| 114 | { |
| 115 | sum_carry_res = wrapper::vpadd(sum_carry_res, sum_carry_res); |
| 116 | sum_sq_carry_res = wrapper::vpadd(sum_sq_carry_res, sum_sq_carry_res); |
| 117 | } |
| 118 | |
| 119 | auto sum = wrapper::vgetlane(sum_carry_res, 0); |
| 120 | auto sum_sq = wrapper::vgetlane(sum_sq_carry_res, 0); |
| 121 | |
| 122 | // Compute left-over elements |
| 123 | for(; x < window_end_x; ++x) |
| 124 | { |
| 125 | ScalarType data = *(in_ptr + x); |
| 126 | sum += data; |
| 127 | sum_sq += data * data; |
| 128 | } |
| 129 | |
| 130 | ScalarType mean = sum / _input->info()->dimension(0); |
| 131 | ScalarType var = (sum_sq / _input->info()->dimension(0)) - (mean * mean); |
| 132 | ScalarType stddev_inv = 1.f / sqrt(var + _epsilon); |
| 133 | |
| 134 | auto mean_vec = wrapper::vdup_n(mean, ExactTagType{}); |
| 135 | auto stddev_inv_vec = wrapper::vdup_n(stddev_inv, ExactTagType{}); |
| 136 | for(x = window_start_x; x <= (window_end_x - window_step_x); x += window_step_x) |
| 137 | { |
| 138 | auto data = wrapper::vloadq(in_ptr + x); |
| 139 | auto res = wrapper::vmul(wrapper::vsub(data, mean_vec), stddev_inv_vec); |
| 140 | // Store results |
| 141 | wrapper::vstore(out_ptr + x, res); |
| 142 | } |
| 143 | for(; x < window_end_x; ++x) |
| 144 | { |
| 145 | *(out_ptr + x) = (*(in_ptr + x) - mean) * stddev_inv; |
| 146 | } |
| 147 | }, |
| 148 | input, output); |
| 149 | } |
| 150 | |
| 151 | NEMeanStdDevNormalizationKernel::NEMeanStdDevNormalizationKernel() |
| 152 | : _input(nullptr), _output(nullptr), _epsilon(1e-8f), _func(nullptr) |
| 153 | { |
| 154 | } |
| 155 | |
| 156 | void NEMeanStdDevNormalizationKernel::configure(ITensor *input, ITensor *output, float epsilon) |
| 157 | { |
| 158 | ARM_COMPUTE_ERROR_ON_NULLPTR(input); |
| 159 | |
| 160 | ARM_COMPUTE_ERROR_THROW_ON(NEMeanStdDevNormalizationKernel::validate(input->info(), (output != nullptr) ? output->info() : nullptr, epsilon)); |
| 161 | |
| 162 | _input = input; |
| 163 | _output = (output == nullptr) ? input : output; |
| 164 | _epsilon = epsilon; |
| 165 | |
| 166 | // Configure kernel window |
| 167 | auto win_config = validate_and_configure_window(input->info(), (output == nullptr) ? nullptr : output->info()); |
| 168 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 169 | ICPPKernel::configure(win_config.second); |
| 170 | |
| 171 | // Configure function to run based on different data types |
| 172 | const DataType data_type = input->info()->data_type(); |
| 173 | switch(data_type) |
| 174 | { |
| 175 | case DataType::F32: |
| 176 | _func = &NEMeanStdDevNormalizationKernel::mean_stddev_normalization<float, 4>; |
| 177 | break; |
| 178 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 179 | case DataType::F16: |
| 180 | _func = &NEMeanStdDevNormalizationKernel::mean_stddev_normalization<float16_t, 8>; |
| 181 | break; |
| 182 | #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 183 | default: |
| 184 | ARM_COMPUTE_ERROR("Not Supported"); |
| 185 | break; |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | Status NEMeanStdDevNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, float epsilon) |
| 190 | { |
| 191 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, epsilon)); |
| 192 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (output != nullptr) ? output->clone().get() : nullptr).first); |
| 193 | return Status{}; |
| 194 | } |
| 195 | |
| 196 | void NEMeanStdDevNormalizationKernel::run(const Window &window, const ThreadInfo &info) |
| 197 | { |
| 198 | ARM_COMPUTE_UNUSED(info); |
| 199 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 200 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 201 | ARM_COMPUTE_ERROR_ON(_func == nullptr); |
| 202 | |
| 203 | (this->*_func)(window); |
| 204 | } |
| 205 | } // namespace arm_compute |