src/core/NEON/kernels/NESobel3x3Kernel.cpp

   1 /*
   2  * Copyright (c) 2016, 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/NESobel3x3Kernel.h"
  25
  26 #include "arm_compute/core/Coordinates.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/Types.h"
  31 #include "arm_compute/core/Validate.h"
  32
  33 #include <arm_neon.h>
  34 #include <cstdint>
  35
  36 using namespace arm_compute;
  37
  38 NESobel3x3Kernel::NESobel3x3Kernel()
  39     : _run_sobel_x(false), _run_sobel_y(false), _input(nullptr), _output_x(nullptr), _output_y(nullptr)
  40 {
  41 }
  42
  43 BorderSize NESobel3x3Kernel::border_size() const
  44 {
  45     return BorderSize(1);
  46 }
  47
  48 void NESobel3x3Kernel::configure(const ITensor *input, ITensor *output_x, ITensor *output_y, bool border_undefined)
  49 {
  50     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
  51     ARM_COMPUTE_ERROR_ON((output_x == nullptr) && (output_y == nullptr));
  52
  53     _run_sobel_x = output_x != nullptr;
  54     _run_sobel_y = output_y != nullptr;
  55
  56     if(_run_sobel_x)
  57     {
  58         ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_x, 1, DataType::S16);
  59     }
  60
  61     if(_run_sobel_y)
  62     {
  63         ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_y, 1, DataType::S16);
  64     }
  65
  66     _input    = input;
  67     _output_x = output_x;
  68     _output_y = output_y;
  69
  70     // Configure kernel window
  71     constexpr unsigned int num_elems_processed_per_iteration = 8;
  72     constexpr unsigned int num_elems_read_per_iteration      = 16;
  73     constexpr unsigned int num_elems_written_per_iteration   = 8;
  74     constexpr unsigned int num_rows_read_per_iteration       = 3;
  75
  76     Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
  77     AccessWindowHorizontal output_x_access(output_x == nullptr ? nullptr : output_x->info(), 0, num_elems_written_per_iteration);
  78     AccessWindowHorizontal output_y_access(output_y == nullptr ? nullptr : output_y->info(), 0, num_elems_written_per_iteration);
  79
  80     update_window_and_padding(win,
  81                               AccessWindowRectangle(input->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration),
  82                               output_x_access,
  83                               output_y_access);
  84
  85     output_x_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
  86     output_y_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
  87
  88     INEKernel::configure(win);
  89 }
  90
  91 void NESobel3x3Kernel::run(const Window &window)
  92 {
  93     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
  94     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
  95
  96     const unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-1, -1));
  97     const unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
  98     const unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, 1));
  99
 100     Iterator input(_input, window);
 101     Iterator output_y;
 102     Iterator output_x;
 103
 104     if(_run_sobel_y)
 105     {
 106         output_y = Iterator(_output_y, window);
 107     }
 108
 109     if(_run_sobel_x)
 110     {
 111         output_x = Iterator(_output_x, window);
 112     }
 113
 114     static const int16x8_t two      = vdupq_n_s16(2);
 115     static const int16x8_t minustwo = vdupq_n_s16(-2);
 116
 117     if(_run_sobel_y && _run_sobel_x)
 118     {
 119         execute_window_loop(window, [&](const Coordinates & id)
 120         {
 121             const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
 122             const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
 123             const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
 124
 125             const int16x8x2_t top_s16 =
 126             {
 127                 {
 128                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
 129                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
 130                 }
 131             };
 132             const int16x8x2_t mid_s16 =
 133             {
 134                 {
 135                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
 136                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
 137                 }
 138             };
 139             const int16x8x2_t bot_s16 =
 140             {
 141                 {
 142                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
 143                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
 144                 }
 145             };
 146
 147             //SOBEL Y
 148             //top left
 149             int16x8_t out_y = vnegq_s16(top_s16.val[0]);
 150             //top mid
 151             out_y = vmlaq_s16(out_y, vextq_s16(top_s16.val[0], top_s16.val[1], 1), minustwo);
 152             //top right
 153             out_y = vsubq_s16(out_y, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
 154             //bot left
 155             out_y = vaddq_s16(out_y, bot_s16.val[0]);
 156             //bot mid
 157             out_y = vmlaq_s16(out_y, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1), two);
 158             //bot right
 159             out_y = vaddq_s16(out_y, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
 160
 161             vst1q_s16(reinterpret_cast<int16_t *>(output_y.ptr()), out_y);
 162
 163             //SOBEL X
 164             //top left
 165             int16x8_t out_x = vnegq_s16(top_s16.val[0]);
 166             //top right
 167             out_x = vaddq_s16(out_x, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
 168             //mid left
 169             out_x = vmlaq_s16(out_x, mid_s16.val[0], minustwo);
 170             //mid right
 171             out_x = vmlaq_s16(out_x, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2), two);
 172             //bot left
 173             out_x = vsubq_s16(out_x, bot_s16.val[0]);
 174             //bot right
 175             out_x = vaddq_s16(out_x, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
 176
 177             vst1q_s16(reinterpret_cast<int16_t *>(output_x.ptr()), out_x);
 178         },
 179         input, output_x, output_y);
 180     }
 181     else if(_run_sobel_x)
 182     {
 183         execute_window_loop(window, [&](const Coordinates & id)
 184         {
 185             const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
 186             const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
 187             const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
 188
 189             const int16x8x2_t top_s16 =
 190             {
 191                 {
 192                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
 193                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
 194                 }
 195             };
 196             const int16x8x2_t mid_s16 =
 197             {
 198                 {
 199                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
 200                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
 201                 }
 202             };
 203             const int16x8x2_t bot_s16 =
 204             {
 205                 {
 206                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
 207                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
 208                 }
 209             };
 210
 211             //SOBEL X
 212             //top left
 213             int16x8_t out = vnegq_s16(top_s16.val[0]);
 214             //top right
 215             out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
 216             //mid left
 217             out = vmlaq_s16(out, mid_s16.val[0], minustwo);
 218             //mid right
 219             out = vmlaq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2), two);
 220             //bot left
 221             out = vsubq_s16(out, bot_s16.val[0]);
 222             //bot right
 223             out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
 224
 225             vst1q_s16(reinterpret_cast<int16_t *>(output_x.ptr()), out);
 226         },
 227         input, output_x);
 228     }
 229     else if(_run_sobel_y)
 230     {
 231         execute_window_loop(window, [&](const Coordinates & id)
 232         {
 233             const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
 234             const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
 235
 236             const int16x8x2_t top_s16 =
 237             {
 238                 {
 239                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
 240                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
 241                 }
 242             };
 243             const int16x8x2_t bot_s16 =
 244             {
 245                 {
 246                     vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
 247                     vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
 248                 }
 249             };
 250
 251             //SOBEL Y
 252             //top left
 253             int16x8_t out = vnegq_s16(top_s16.val[0]);
 254             //top mid
 255             out = vmlaq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 1), minustwo);
 256             //top right
 257             out = vsubq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
 258             //bot left
 259             out = vaddq_s16(out, bot_s16.val[0]);
 260             //bot mid
 261             out = vmlaq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1), two);
 262             //bot right
 263             out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
 264
 265             vst1q_s16(reinterpret_cast<int16_t *>(output_y.ptr()), out);
 266         },
 267         input, output_y);
 268     }
 269 }