Compute Library
 21.11
list.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2021 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 #ifndef SRC_CORE_NEON_KERNELS_SUB_LIST_H
25 #define SRC_CORE_NEON_KERNELS_SUB_LIST_H
26 
27 #include "arm_compute/core/Types.h"
30 
31 namespace arm_compute
32 {
33 namespace cpu
34 {
35 #define DECLARE_SUB_KERNEL(func_name) \
36  void func_name(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
37 
41 
42 #undef DECLARE_SUB_KERNEL
43 
44 template <typename T>
45 void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
46 {
47  /** SIMD vector tag type. */
49 
50  bool is_sat = policy == ConvertPolicy::SATURATE;
51 
52  // Create input windows
53  Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
54  Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
55 
56  // Clear X Dimension on execution window as we handle manually
57  Window win = window;
58  win.set(Window::DimX, Window::Dimension(0, 1, 1));
59 
60  constexpr int window_step_x = 16 / sizeof(T);
61  const auto window_start_x = static_cast<int>(window.x().start());
62  const auto window_end_x = static_cast<int>(window.x().end());
63  const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
64 
65  Iterator input1(src0, window.broadcast_if_dimension_le_one(src0->info()->tensor_shape()));
66  Iterator input2(src1, window.broadcast_if_dimension_le_one(src1->info()->tensor_shape()));
67  Iterator output(dst, window);
68 
69  if(is_broadcast_across_x)
70  {
71  const bool is_broadcast_input_2 = input2_win.x().step() == 0;
72  Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
73  Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
74  const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
75  const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
76 
77  // Clear X Dimension on execution window as we handle manually
78  non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
79 
80  Iterator broadcast_input(broadcast_tensor, broadcast_win);
81  Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
82  Iterator output(dst, win);
83 
84  execute_window_loop(win, [&](const Coordinates &)
85  {
86  const auto non_broadcast_input_ptr = reinterpret_cast<const T *>(non_broadcast_input.ptr());
87  const auto output_ptr = reinterpret_cast<T *>(output.ptr());
88 
89  const T broadcast_value = *reinterpret_cast<const T *>(broadcast_input.ptr());
90  const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
91 
92  // Compute S elements per iteration
93  int x = window_start_x;
94  for(; x <= (window_end_x - window_step_x); x += window_step_x)
95  {
96  const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
97  auto res = is_sat ? wrapper::vqsub(broadcast_value_vec, non_broadcast_v) : wrapper::vsub(broadcast_value_vec, non_broadcast_v);
98  if(is_broadcast_input_2)
99  {
100  res = wrapper::vmul(res, wrapper::vdup_n(static_cast<T>(-1), ExactTagType{}));
101  }
102  wrapper::vstore(output_ptr + x, res);
103  }
104 
105  // Compute left-over elements
106  for(; x < window_end_x; ++x)
107  {
108  const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
109  auto res = is_sat ? wrapper::sub_sat(broadcast_value, non_broadcast_v) : broadcast_value - non_broadcast_v;
110  if(is_broadcast_input_2)
111  {
112  res = static_cast<T>(-1) * res;
113  }
114 
115  *(output_ptr + x) = res;
116  }
117  },
118  broadcast_input, non_broadcast_input, output);
119  }
120  else
121  {
122  // Clear X Dimension on execution window as we handle manually
123  input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
124  input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
125 
126  Iterator input1(src0, input1_win);
127  Iterator input2(src1, input2_win);
128  Iterator output(dst, win);
129 
130  execute_window_loop(win, [&](const Coordinates &)
131  {
132  const auto input1_ptr = reinterpret_cast<const T *>(input1.ptr());
133  const auto input2_ptr = reinterpret_cast<const T *>(input2.ptr());
134  const auto output_ptr = reinterpret_cast<T *>(output.ptr());
135 
136  // Compute S elements per iteration
137  int x = window_start_x;
138  for(; x <= (window_end_x - window_step_x); x += window_step_x)
139  {
140  const auto val1 = wrapper::vloadq(input1_ptr + x);
141  const auto val2 = wrapper::vloadq(input2_ptr + x);
142  const auto res = is_sat ? wrapper::vqsub(val1, val2) : wrapper::vsub(val1, val2);
143  wrapper::vstore(output_ptr + x, res);
144  }
145 
146  // Compute left-over elements
147  for(; x < window_end_x; ++x)
148  {
149  const auto val1 = *(input1_ptr + x);
150  const auto val2 = *(input2_ptr + x);
151  *(output_ptr + x) = is_sat ? wrapper::sub_sat(val1, val2) : val1 - val2;
152  }
153  },
154  input1, input2, output);
155  }
156 }
157 } // namespace cpu
158 } // namespace arm_compute
159 #endif // SRC_CORE_NEON_KERNELS_SUB_LIST_H
void sub_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
Definition: list.h:45
constexpr int step() const
Return the step of the dimension.
Definition: Window.h:104
uint8x16_t vloadq(const uint8_t *ptr)
Definition: load.h:58
uint8x8_t vsub(const uint8x8_t &a, const uint8x8_t &b)
Definition: sub.h:39
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:77
Interface for CPU tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
typename neon_bitvector< T, BW >::tag_type neon_bitvector_tag_t
Helper type template to get the tag type of a neon vector.
Definition: traits.h:132
T x() const
Alias to access the size of the first dimension.
Definition: Dimensions.h:87
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
Coordinates of an item.
Definition: Coordinates.h:37
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
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
Window broadcast_if_dimension_le_one(const TensorShape &shape) const
Don&#39;t advance in the dimension where shape is less equal to 1.
Definition: Window.inl:120
void sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
uint8x8_t vmul(const uint8x8_t &a, const uint8x8_t &b)
Definition: mul.h:39
void sub_qsymm16_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
Definition: qsymm16.cpp:35
void sub_qasymm8_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
Definition: qasymm8.cpp:35
uint8x8_t vqsub(const uint8x8_t &a, const uint8x8_t &b)
Definition: sub.h:74
#define DECLARE_SUB_KERNEL(func_name)
Definition: list.h:35
void vstore(uint8_t *ptr, uint8x8_t val)
Definition: store.h:39
uint8x8_t vdup_n(uint8_t value, traits::vector_64_tag)
Definition: dup_n.h:41
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
uint8_t sub_sat(const uint8_t &a, const uint8_t &b)
Definition: sub.h:33
Includes all wrapper headers at once.
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:99
Iterator updated by execute_window_loop for each window element.
Definition: Helpers.h:46
constexpr int start() const
Return the start of the dimension.
Definition: Window.h:94
Describe a multidimensional execution window.
Definition: Window.h:39
ConvertPolicy
Policy to handle integer overflow.
Definition: Types.h:391
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:145