Compute Library
 21.02
NEGaussian3x3Kernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2016-2020 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  */
25 
33 
34 #include <arm_neon.h>
35 
36 using namespace arm_compute;
37 
39 {
40  return BorderSize(1);
41 }
42 
43 void NEGaussian3x3Kernel::configure(const ITensor *input, ITensor *output, bool border_undefined)
44 {
47 
48  _input = input;
49  _output = output;
50 
51  // Configure kernel window
52  constexpr unsigned int num_elems_processed_per_iteration = 8;
53  constexpr unsigned int num_elems_read_per_iteration = 16;
54  constexpr unsigned int num_elems_written_per_iteration = 8;
55  constexpr unsigned int num_rows_read_per_iteration = 3;
56 
57  Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size());
58  AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration);
59 
61  AccessWindowRectangle(input->info(), -border_size().left, -border_size().top, num_elems_read_per_iteration, num_rows_read_per_iteration),
62  output_access);
63 
64  output_access.set_valid_region(win, input->info()->valid_region(), border_undefined, border_size());
65 
66  INEKernel::configure(win);
67 }
68 
70 {
71  ARM_COMPUTE_UNUSED(info);
74 
75  Iterator input(_input, window);
76  Iterator output(_output, window);
77 
78  const uint8_t *input_bot_ptr = _input->ptr_to_element(Coordinates(-1, -1));
79  const uint8_t *input_mid_ptr = _input->ptr_to_element(Coordinates(-1, 0));
80  const uint8_t *input_top_ptr = _input->ptr_to_element(Coordinates(-1, +1));
81 
82  static const int16x8_t two = vdupq_n_s16(2);
83  static const int16x8_t four = vdupq_n_s16(4);
84 
85  execute_window_loop(window, [&](const Coordinates &)
86  {
87  uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());
88  uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());
89  uint8x16_t bot_data = vld1q_u8(input_bot_ptr + input.offset());
90 
91  const int16x8x2_t top_s16 =
92  {
93  {
94  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(top_data))),
95  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(top_data)))
96  }
97  };
98  const int16x8x2_t mid_s16 =
99  {
100  {
101  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mid_data))),
102  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mid_data)))
103  }
104  };
105  const int16x8x2_t bot_s16 =
106  {
107  {
108  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(bot_data))),
109  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(bot_data)))
110  }
111  };
112 
113  //top left
114  int16x8_t out = top_s16.val[0];
115  //top mid
116  out = vmlaq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 1), two);
117  //top right
118  out = vaddq_s16(out, vextq_s16(top_s16.val[0], top_s16.val[1], 2));
119  //mid left
120  out = vmlaq_s16(out, mid_s16.val[0], two);
121  //mid mid
122  out = vmlaq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 1), four);
123  //mid right
124  out = vmlaq_s16(out, vextq_s16(mid_s16.val[0], mid_s16.val[1], 2), two);
125  //bot left
126  out = vaddq_s16(out, bot_s16.val[0]);
127  //bot mid
128  out = vmlaq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 1), two);
129  //bot right
130  out = vaddq_s16(out, vextq_s16(bot_s16.val[0], bot_s16.val[1], 2));
131 
132  vst1_u8(output.ptr(), vqshrun_n_s16(out, 4));
133  },
134  input, output);
135 }
unsigned int top
top of the border
Definition: Types.h:375
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
void configure(const ITensor *input, ITensor *output, bool border_undefined)
Set the source, destination and border mode of the kernel.
Container for 2D border size.
Definition: Types.h:273
1 channel, 1 U8 per channel
Interface for Neon tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
virtual ValidRegion valid_region() const =0
Valid region of the tensor.
Implementation of a rectangular access pattern.
bool update_window_and_padding(Window &win, Ts &&... patterns)
Update window and padding size for each of the access patterns.
Definition: WindowHelpers.h:46
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
Coordinates of an item.
Definition: Coordinates.h:37
Implementation of a row access pattern.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
Definition: Helpers.inl:139
unsigned int left
left of the border
Definition: Types.h:378
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
#define ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:790
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
Information about executing thread and CPU.
Definition: CPPTypes.h:235
unsigned int num_elems_processed_per_iteration
void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators)
Iterate through the passed window, automatically adjusting the iterators and calling the lambda_funct...
Definition: Helpers.inl:77
BorderSize border_size() const override
The size of the border for that kernel.
void run(const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
constexpr size_t offset() const
Return the offset in bytes from the first element to the current position of the iterator.
Definition: Helpers.inl:134
Iterator updated by execute_window_loop for each window element.
Definition: Helpers.h:46
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205