src/core/NEON/kernels/NEPoolingLayerKernel.cpp

   1 /*
   2  * Copyright (c) 2017 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 "arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h"
  25
  26 #include "arm_compute/core/AccessWindowStatic.h"
  27 #include "arm_compute/core/Error.h"
  28 #include "arm_compute/core/Helpers.h"
  29 #include "arm_compute/core/ITensor.h"
  30 #include "arm_compute/core/TensorInfo.h"
  31 #include "arm_compute/core/Utils.h"
  32 #include "arm_compute/core/Validate.h"
  33 #include "arm_compute/core/Window.h"
  34
  35 #include <algorithm>
  36 #include <arm_neon.h>
  37 #include <limits>
  38 #include <string>
  39 #include <tuple>
  40
  41 using namespace arm_compute;
  42
  43 namespace
  44 {
  45 inline float calculate_avg_scale(const Coordinates &id, const int pool_size, const int upper_bound_w, const int upper_bound_h,
  46                                  const int pad_x, const int pad_y, const int stride_x, const int stride_y)
  47 {
  48     int start_x = id.x() * stride_x - pad_x;
  49     int start_y = id.y() * stride_y - pad_y;
  50     int end_x   = std::min(start_x + pool_size, upper_bound_w);
  51     int end_y   = std::min(start_y + pool_size, upper_bound_h);
  52     return 1.f / ((end_y - start_y) * (end_x - start_x));
  53 }
  54 } // namespace
  55
  56 NEPoolingLayerKernel::NEPoolingLayerKernel()
  57     : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _border_size(0)
  58 {
  59 }
  60
  61 BorderSize NEPoolingLayerKernel::border_size() const
  62 {
  63     return _border_size;
  64 }
  65
  66 void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info)
  67 {
  68     int                   pool_pad_x      = 0;
  69     int                   pool_pad_y      = 0;
  70     int                   pool_stride_x   = 0;
  71     int                   pool_stride_y   = 0;
  72     unsigned int          pooled_w        = 0;
  73     unsigned int          pooled_h        = 0;
  74     PoolingType           pool_type       = pool_info.pool_type();
  75     int                   pool_size       = pool_info.pool_size();
  76     const PadStrideInfo   pad_stride_info = pool_info.pad_stride_info();
  77     DimensionRoundingType pool_round      = pad_stride_info.round();
  78     std::tie(pool_pad_x, pool_pad_y)       = pad_stride_info.pad();
  79     std::tie(pool_stride_x, pool_stride_y) = pad_stride_info.stride();
  80
  81     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
  82     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
  83     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
  84     ARM_COMPUTE_ERROR_ON(2 != pool_size && 3 != pool_size);
  85     ARM_COMPUTE_ERROR_ON(pool_pad_x >= pool_size || pool_pad_y >= pool_size);
  86
  87     // Check output dimensions
  88     std::tie(pooled_w, pooled_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1),
  89                                                      pool_size, pool_stride_x, pool_stride_y,
  90                                                      pool_pad_x, pool_pad_y, pool_round);
  91     ARM_COMPUTE_UNUSED(pooled_w);
  92     ARM_COMPUTE_UNUSED(pooled_h);
  93     ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pooled_w) || (output->info()->dimension(1) != pooled_h));
  94
  95     const int num_elems_read_per_iteration = (pool_size == 2) ? 2 : 4; // We use vload4 for pooling3
  96     const int input_width                  = input->info()->dimension(0);
  97     const int input_height                 = input->info()->dimension(1);
  98     const int upper_bound_w                = ((pooled_w - 1) * pool_stride_x - pool_pad_x + num_elems_read_per_iteration) - input_width;
  99     const int upper_bound_h                = ((pooled_h - 1) * pool_stride_y - pool_pad_y + pool_size) - input_height;
 100
 101     // Set instance variables
 102     _input              = input;
 103     _output             = output;
 104     _pool_info          = pool_info;
 105     _border_size        = BorderSize(pool_pad_y, pool_pad_x);
 106     _border_size.right  = std::max(upper_bound_w, pool_pad_x);
 107     _border_size.bottom = std::max(upper_bound_h, pool_pad_y);
 108
 109     // Select appropriate function
 110     switch(pool_size)
 111     {
 112         case 2:
 113             _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling2<PoolingType::AVG> : &NEPoolingLayerKernel::pooling2<PoolingType::MAX>;
 114             break;
 115         case 3:
 116             _func = (PoolingType::AVG == pool_type) ? &NEPoolingLayerKernel::pooling3<PoolingType::AVG> : &NEPoolingLayerKernel::pooling3<PoolingType::MAX>;
 117             break;
 118         default:
 119             ARM_COMPUTE_ERROR("Unsupported pooling size");
 120             break;
 121     }
 122
 123     // Configure kernel window
 124     constexpr unsigned int num_elems_processed_per_iteration = 1;
 125
 126     Window                 win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration));
 127     AccessWindowStatic     input_access(input->info(), -pool_pad_x, -pool_pad_y, input_width + _border_size.right, input_height + _border_size.bottom);
 128     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
 129     update_window_and_padding(win, input_access, output_access);
 130     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
 131     INEKernel::configure(win);
 132 }
 133
 134 template <PoolingType pooling_type>
 135 void NEPoolingLayerKernel::pooling2(const Window &window_input, const Window &window)
 136 {
 137     Iterator input(_input, window_input);
 138     Iterator output(_output, window);
 139
 140     constexpr int pool_size = 2;
 141     int           pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y = 0;
 142     std::tie(pool_pad_x, pool_pad_y)       = _pool_info.pad_stride_info().pad();
 143     std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride();
 144     const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x;
 145     const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y;
 146
 147     const unsigned char *const input_top_ptr    = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y)));
 148     const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1));
 149
 150     execute_window_loop(window, [&](const Coordinates & id)
 151     {
 152         const float32x2_t top_data    = vld1_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset()));
 153         const float32x2_t bottom_data = vld1_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset()));
 154         float32x2_t       res         = {};
 155         if(pooling_type == PoolingType::AVG)
 156         {
 157             // Calculate scale
 158             float             scale   = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y);
 159             const float32x2_t scale_v = vdup_n_f32(scale);
 160
 161             // Perform pooling
 162             const float32x2_t sum_data = vadd_f32(top_data, bottom_data);
 163             res                        = vmul_f32(vpadd_f32(sum_data, sum_data), scale_v);
 164         }
 165         else
 166         {
 167             const float32x2_t max_data = vmax_f32(top_data, bottom_data);
 168             res                        = vpmax_f32(max_data, max_data);
 169         }
 170         *(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(res, 0);
 171     },
 172     input, output);
 173 }
 174
 175 template <PoolingType pooling_type>
 176 void NEPoolingLayerKernel::pooling3(const Window &window_input, const Window &window)
 177 {
 178     Iterator input(_input, window_input);
 179     Iterator output(_output, window);
 180
 181     constexpr const int pool_size = 3;
 182     int                 pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y = 0;
 183     std::tie(pool_pad_x, pool_pad_y)       = _pool_info.pad_stride_info().pad();
 184     std::tie(pool_stride_x, pool_stride_y) = _pool_info.pad_stride_info().stride();
 185     const int upper_bound_w = _input->info()->dimension(0) + pool_pad_x;
 186     const int upper_bound_h = _input->info()->dimension(1) + pool_pad_y;
 187
 188     const unsigned char *const input_top_ptr    = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y)));
 189     const unsigned char *const input_middle_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 1));
 190     const unsigned char *const input_bottom_ptr = _input->ptr_to_element(Coordinates(-static_cast<int>(pool_pad_x), -static_cast<int>(pool_pad_y) + 2));
 191
 192     execute_window_loop(window, [&](const Coordinates & id)
 193     {
 194         const float32x4_t top_data    = vld1q_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset()));
 195         const float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(input_middle_ptr + input.offset()));
 196         const float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset()));
 197         float32x2_t       res         = {};
 198         if(pooling_type == PoolingType::AVG)
 199         {
 200             // Calculate scale
 201             float             scale   = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y);
 202             const float32x2_t scale_v = vdup_n_f32(scale);
 203
 204             // Perform pooling
 205             const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data);
 206             res                        = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data));
 207             res                        = vmul_f32(vpadd_f32(res, res), scale_v);
 208         }
 209         else
 210         {
 211             const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data);
 212             res                        = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data));
 213             res                        = vpmax_f32(res, res);
 214         }
 215         *(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(res, 0);
 216     },
 217     input, output);
 218 }
 219
 220 void NEPoolingLayerKernel::run(const Window &window)
 221 {
 222     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 223     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 224     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 225
 226     unsigned int pool_stride_x, pool_stride_y = 0;
 227     std::tie(pool_stride_x, pool_stride_y)    = _pool_info.pad_stride_info().stride();
 228
 229     // Set step for input in x and y direction for the input
 230     Window window_input(window);
 231     window_input.set(Window::DimX, Window::Dimension(window.x().start() * pool_stride_x, window.x().end() * pool_stride_x, pool_stride_x));
 232     window_input.set(Window::DimY, Window::Dimension(window.y().start() * pool_stride_y, window.y().end() * pool_stride_y, pool_stride_y));
 233
 234     // Run function
 235     (this->*_func)(window_input, window);
 236 }