2 * Copyright (c) 2016, 2017 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #include "arm_compute/core/NEON/kernels/NEGaussian3x3Kernel.h"
26 #include "arm_compute/core/Coordinates.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/NEON/INEKernel.h"
30 #include "arm_compute/core/Validate.h"
34 using namespace arm_compute;
36 BorderSize NEGaussian3x3Kernel::border_size() const
41 void NEGaussian3x3Kernel::configure(const ITensor *input, ITensor *output, bool border_undefined)
43 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
44 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
49 // Configure kernel window
50 constexpr unsigned int num_elems_processed_per_iteration = 8;
51 constexpr unsigned int num_elems_read_per_iteration = 16;
52 constexpr unsigned int num_elems_written_per_iteration = 8;
53 constexpr unsigned int num_rows_read_per_iteration = 3;
55 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
56 AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
58 update_window_and_padding(win,
59 AccessWindowRectangle(input->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration),
62 output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
64 INEKernel::configure(win);
67 void NEGaussian3x3Kernel::run(const Window &window)
69 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
70 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
72 Iterator input(_input, window);
73 Iterator output(_output, window);
75 const uint8_t *input_bot_ptr = _input->ptr_to_element(Coordinates(-1, -1));
76 const uint8_t *input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
77 const uint8_t *input_top_ptr = _input->ptr_to_element(Coordinates(-1, +1));
79 static const int16x8_t two = vdupq_n_s16(2);
80 static const int16x8_t four = vdupq_n_s16(4);
82 execute_window_loop(window, [&](const Coordinates & id)
84 uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
85 uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
86 uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
88 const int16x8x2_t top_s16 =
91 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
92 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
95 const int16x8x2_t mid_s16 =
98 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
99 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
102 const int16x8x2_t bot_s16 =
105 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
106 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
111 int16x8_t out = top_s16.val[0];
113 out = vmlaq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 1), two);
115 out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
117 out = vmlaq_s16(out, mid_s16.val[0], two);
119 out = vmlaq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 1), four);
121 out = vmlaq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2), two);
123 out = vaddq_s16(out, bot_s16.val[0]);
125 out = vmlaq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1), two);
127 out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
129 vst1_u8(output.ptr(), vqshrun_n_s16(out, 4));