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/NEBox3x3Kernel.h"
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"
34 using namespace arm_compute;
36 #ifdef ARM_COMPUTE_ENABLE_FP16
37 void NEBox3x3FP16Kernel::run(const Window &window)
39 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
40 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
42 Iterator input(_input, window);
43 Iterator output(_output, window);
45 unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-1, -1));
46 unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
47 unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, +1));
49 const float16x8_t oneovernine = vdupq_n_f16(1.0f / 9.0f);
51 execute_window_loop(window, [&](const Coordinates & id)
53 const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
54 const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
55 const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
57 const float16x8x2_t top_f16 =
60 vcvtq_f16_u16(vmovl_u8(vget_low_u8(top_data))),
61 vcvtq_f16_u16(vmovl_u8(vget_high_u8(top_data)))
65 const float16x8x2_t mid_f16 =
68 vcvtq_f16_u16(vmovl_u8(vget_low_u8(mid_data))),
69 vcvtq_f16_u16(vmovl_u8(vget_high_u8(mid_data)))
73 const float16x8x2_t bot_f16 =
76 vcvtq_f16_u16(vmovl_u8(vget_low_u8(bot_data))),
77 vcvtq_f16_u16(vmovl_u8(vget_high_u8(bot_data)))
82 float16x8_t out = top_f16.val[0];
84 out = vaddq_f16(out, vextq_f16(top_f16.val[0], top_f16.val[1], 1));
86 out = vaddq_f16(out, vextq_f16(top_f16.val[0], top_f16.val[1], 2));
88 out = vaddq_f16(out, mid_f16.val[0]);
90 out = vaddq_f16(out, vextq_f16(mid_f16.val[0], mid_f16.val[1], 1));
92 out = vaddq_f16(out, vextq_f16(mid_f16.val[0], mid_f16.val[1], 2));
94 out = vaddq_f16(out, bot_f16.val[0]);
96 out = vaddq_f16(out, vextq_f16(bot_f16.val[0], bot_f16.val[1], 1));
98 out = vaddq_f16(out, vextq_f16(bot_f16.val[0], bot_f16.val[1], 2));
100 out = vmulq_f16(out, oneovernine);
102 vst1_u8(output.ptr(), vqmovun_s16(vcvtq_s16_f16(out)));
108 BorderSize NEBox3x3Kernel::border_size() const
110 return BorderSize(1);
113 void NEBox3x3Kernel::configure(const ITensor *input, ITensor *output, bool border_undefined)
115 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
116 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8);
121 constexpr unsigned int num_elems_processed_per_iteration = 8;
122 constexpr unsigned int num_elems_read_per_iteration = 16;
123 constexpr unsigned int num_elems_written_per_iteration = 8;
124 constexpr unsigned int num_rows_read_per_iteration = 3;
125 constexpr int rect_offset_xy = -1;
127 // Configure kernel window
128 Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
129 AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
131 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);
133 output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
135 INEKernel::configure(win);
138 void NEBox3x3Kernel::run(const Window &window)
140 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
141 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INESimpleKernel::window(), window);
143 Iterator input(_input, window);
144 Iterator output(_output, window);
146 unsigned char *const input_top_ptr = _input->ptr_to_element(Coordinates(-1, -1));
147 unsigned char *const input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
148 unsigned char *const input_bot_ptr = _input->ptr_to_element(Coordinates(-1, +1));
150 const float32x4_t oneovernine = vdupq_n_f32(1.0f / 9.0f);
152 execute_window_loop(window, [&](const Coordinates & id)
154 const uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
155 const uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
156 const uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
158 const int16x8x2_t top_s16 =
161 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
162 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
165 const int16x8x2_t mid_s16 =
168 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
169 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
172 const int16x8x2_t bot_s16 =
175 vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
176 vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
181 int16x8_t out = top_s16.val[0];
183 out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 1));
185 out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
187 out = vaddq_s16(out, mid_s16.val[0]);
189 out = vaddq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 1));
191 out = vaddq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2));
193 out = vaddq_s16(out, bot_s16.val[0]);
195 out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1));
197 out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
199 float32x4_t outfloathigh = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out)));
200 float32x4_t outfloatlow = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out)));
202 outfloathigh = vmulq_f32(outfloathigh, oneovernine);
203 outfloatlow = vmulq_f32(outfloatlow, oneovernine);
205 out = vcombine_s16(vqmovn_s32(vcvtq_s32_f32(outfloatlow)),
206 vqmovn_s32(vcvtq_s32_f32(outfloathigh)));
208 vst1_u8(output.ptr(), vqmovun_s16(out));