src/core/NEON/kernels/NEBox3x3Kernel.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/NEBox3x3Kernel.h"
  25
  26 #include "arm_compute/core/Coordinates.h"
  27 #include "arm_compute/core/Helpers.h"
  28 #include "arm_compute/core/IAccessWindow.h"
  29 #include "arm_compute/core/ITensor.h"
  30 #include "arm_compute/core/NEON/INEKernel.h"
  31 #include "arm_compute/core/Validate.h"
  32 #include <arm_neon.h>
  33
  34 using namespace arm_compute;
  35
  36 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
  37 void NEBox3x3FP16Kernel::run(const Window &window, const ThreadInfo &info)
  38 {
  39     ARM_COMPUTE_UNUSED(info);
  40     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
  41     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
  42
  43     Iterator input(_input, window);
  44     Iterator output(_output, window);
  45
  46     unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-1, -1));
  47     unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
  48     unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, +1));
  49
  50     const float16x8_t oneovernine = vdupq_n_f16(1.0f / 9.0f);
  51
  52     execute_window_loop(window, [&](const Coordinates & id)
  53     {
  54         const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
  55         const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
  56         const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
  57
  58         const float16x8x2_t top_f16 =
  59         {
  60             {
  61                 vcvtq_f16_u16(vmovl_u8(vget_low_u8(top_data))),
  62                 vcvtq_f16_u16(vmovl_u8(vget_high_u8(top_data)))
  63             }
  64         };
  65
  66         const float16x8x2_t mid_f16 =
  67         {
  68             {
  69                 vcvtq_f16_u16(vmovl_u8(vget_low_u8(mid_data))),
  70                 vcvtq_f16_u16(vmovl_u8(vget_high_u8(mid_data)))
  71             }
  72         };
  73
  74         const float16x8x2_t bot_f16 =
  75         {
  76             {
  77                 vcvtq_f16_u16(vmovl_u8(vget_low_u8(bot_data))),
  78                 vcvtq_f16_u16(vmovl_u8(vget_high_u8(bot_data)))
  79             }
  80         };
  81
  82         //top left
  83         float16x8_t out = top_f16.val[0];
  84         //top mid
  85         out = vaddq_f16(out, vextq_f16(top_f16.val[0], top_f16.val[1], 1));
  86         //top right
  87         out = vaddq_f16(out, vextq_f16(top_f16.val[0], top_f16.val[1], 2));
  88         //mid left
  89         out = vaddq_f16(out, mid_f16.val[0]);
  90         //mid mid
  91         out = vaddq_f16(out, vextq_f16(mid_f16.val[0], mid_f16.val[1], 1));
  92         //mid right
  93         out = vaddq_f16(out, vextq_f16(mid_f16.val[0], mid_f16.val[1], 2));
  94         //bot left
  95         out = vaddq_f16(out, bot_f16.val[0]);
  96         //bot mid
  97         out = vaddq_f16(out, vextq_f16(bot_f16.val[0], bot_f16.val[1], 1));
  98         //bot right
  99         out = vaddq_f16(out, vextq_f16(bot_f16.val[0], bot_f16.val[1], 2));
 100
 101         out = vmulq_f16(out, oneovernine);
 102
 103         vst1_u8(output.ptr(), vqmovun_s16(vcvtq_s16_f16(out)));
 104     },
 105     input, output);
 106 }
 107 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 108
 109 BorderSize NEBox3x3Kernel::border_size() const
 110 {
 111     return BorderSize(1);
 112 }
 113
 114 void NEBox3x3Kernel::configure(const ITensor *input, ITensor *output, bool border_undefined)
 115 {
 116     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 117
 118     set_shape_if_empty(*output->info(), input->info()->tensor_shape());
 119
 120     set_format_if_unknown(*input->info(), Format::U8);
 121     set_format_if_unknown(*output->info(), Format::U8);
 122
 123     ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
 124     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
 125     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
 126     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 127
 128     _input  = input;
 129     _output = output;
 130
 131     constexpr unsigned int num_elems_processed_per_iteration = 8;
 132     constexpr unsigned int num_elems_read_per_iteration      = 16;
 133     constexpr unsigned int num_elems_written_per_iteration   = 8;
 134     constexpr unsigned int num_rows_read_per_iteration       = 3;
 135     constexpr int          rect_offset_xy                    = -1;
 136
 137     // Configure kernel window
 138     Window                 win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
 139     AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
 140
 141     update_window_and_padding(win, AccessWindowRectangle(input->info(), rect_offset_xy, rect_offset_xy, num_elems_read_per_iteration, num_rows_read_per_iteration), output_access);
 142
 143     output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
 144
 145     INEKernel::configure(win);
 146 }
 147
 148 void NEBox3x3Kernel::run(const Window &window, const ThreadInfo &info)
 149 {
 150     ARM_COMPUTE_UNUSED(info);
 151     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
 152     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
 153
 154     Iterator input(_input, window);
 155     Iterator output(_output, window);
 156
 157     unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-1, -1));
 158     unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
 159     unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, +1));
 160
 161     const float32x4_t oneovernine = vdupq_n_f32(1.0f / 9.0f);
 162
 163     execute_window_loop(window, [&](const Coordinates & id)
 164     {
 165         const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
 166         const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
 167         const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
 168
 169         const int16x8x2_t top_s16 =
 170         {
 171             {
 172                 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
 173                 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
 174             }
 175         };
 176         const int16x8x2_t mid_s16 =
 177         {
 178             {
 179                 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
 180                 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
 181             }
 182         };
 183         const int16x8x2_t bot_s16 =
 184         {
 185             {
 186                 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
 187                 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
 188             }
 189         };
 190
 191         //top left
 192         int16x8_t out = top_s16.val[0];
 193         //top mid
 194         out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 1));
 195         //top right
 196         out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
 197         //mid left
 198         out = vaddq_s16(out, mid_s16.val[0]);
 199         //mid mid
 200         out = vaddq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 1));
 201         //mid right
 202         out = vaddq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2));
 203         //bot left
 204         out = vaddq_s16(out, bot_s16.val[0]);
 205         //bot mid
 206         out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1));
 207         //bot right
 208         out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
 209
 210         float32x4_t outfloathigh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out)));
 211         float32x4_t outfloatlow  = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out)));
 212
 213         outfloathigh = vmulq_f32(outfloathigh, oneovernine);
 214         outfloatlow  = vmulq_f32(outfloatlow, oneovernine);
 215
 216         out = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(outfloatlow)),
 217                            vqmovn_s32(vcvtq_s32_f32(outfloathigh)));
 218
 219         vst1_u8(output.ptr(), vqmovun_s16(out));
 220     },
 221     input, output);
 222 }