Compute Library
 22.05
impl.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2020-2021 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 
29 namespace arm_compute
30 {
31 namespace cpu
32 {
33 template <typename ScalarType>
34 void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
35 {
36  /** SIMD vector tag type. */
38 
39  // Create input windows
40  Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
41  Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
42 
43  // Clear X Dimension on execution window as we handle manually
44  Window win = window;
45  win.set(Window::DimX, Window::Dimension(0, 1, 1));
46 
47  constexpr int window_step_x = 16 / sizeof(ScalarType);
48  const auto window_start_x = static_cast<int>(window.x().start());
49  const auto window_end_x = static_cast<int>(window.x().end());
50  const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
51 
52  if(is_broadcast_across_x)
53  {
54  const bool is_broadcast_input_2 = input2_win.x().step() == 0;
55  Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
56  Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
57  const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
58  const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
59 
60  // Clear X Dimension on execution window as we handle manually
61  non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
62 
63  Iterator broadcast_input(broadcast_tensor, broadcast_win);
64  Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
65  Iterator output(dst, win);
66 
67  execute_window_loop(win, [&](const Coordinates &)
68  {
69  const auto non_broadcast_input_ptr = reinterpret_cast<const ScalarType *>(non_broadcast_input.ptr());
70  const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
71 
72  const ScalarType broadcast_value = *reinterpret_cast<const ScalarType *>(broadcast_input.ptr());
73  const auto broadcast_value_vec = wrapper::vdup_n(broadcast_value, ExactTagType{});
74 
75  // Compute S elements per iteration
76  int x = window_start_x;
77  for(; x <= (window_end_x - window_step_x); x += window_step_x)
78  {
79  const auto non_broadcast_v = wrapper::vloadq(non_broadcast_input_ptr + x);
80  const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(broadcast_value_vec, non_broadcast_v) : wrapper::vadd(broadcast_value_vec, non_broadcast_v);
81  wrapper::vstore(output_ptr + x, res);
82  }
83 
84  // Compute left-over elements
85  for(; x < window_end_x; ++x)
86  {
87  const auto non_broadcast_v = *(non_broadcast_input_ptr + x);
88  *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(broadcast_value, non_broadcast_v) : broadcast_value + non_broadcast_v;
89  }
90  },
91  broadcast_input, non_broadcast_input, output);
92  }
93  else
94  {
95  // Clear X Dimension on execution window as we handle manually
96  input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
97  input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
98 
99  Iterator input1(src0, input1_win);
100  Iterator input2(src1, input2_win);
101  Iterator output(dst, win);
102 
103  execute_window_loop(win, [&](const Coordinates &)
104  {
105  const auto input1_ptr = reinterpret_cast<const ScalarType *>(input1.ptr());
106  const auto input2_ptr = reinterpret_cast<const ScalarType *>(input2.ptr());
107  const auto output_ptr = reinterpret_cast<ScalarType *>(output.ptr());
108 
109  // Compute S elements per iteration
110  int x = window_start_x;
111  for(; x <= (window_end_x - window_step_x); x += window_step_x)
112  {
113  const auto val1 = wrapper::vloadq(input1_ptr + x);
114  const auto val2 = wrapper::vloadq(input2_ptr + x);
115  const auto res = (policy == ConvertPolicy::SATURATE) ? wrapper::vqadd(val1, val2) : wrapper::vadd(val1, val2);
116  wrapper::vstore(output_ptr + x, res);
117  }
118 
119  // Compute left-over elements
120  for(; x < window_end_x; ++x)
121  {
122  const auto val1 = *(input1_ptr + x);
123  const auto val2 = *(input2_ptr + x);
124  *(output_ptr + x) = (policy == ConvertPolicy::SATURATE) ? wrapper::add_sat(val1, val2) : val1 + val2;
125  }
126  },
127  input1, input2, output);
128  }
129 }
130 
131 template void add_same_neon<float>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
132 template void add_same_neon<uint8_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
133 template void add_same_neon<int32_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
134 template void add_same_neon<int16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
135 
136 #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
137 template void add_same_neon<float16_t>(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window);
138 #endif /* (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
139 
140 } // namespace cpu
141 } // namespace arm_compute
constexpr int step() const
Return the step of the dimension.
Definition: Window.h:106
uint8x16_t vloadq(const uint8_t *ptr)
Definition: load.h:58
uint8x8_t vadd(const uint8x8_t &a, const uint8x8_t &b)
Definition: add.h:39
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:79
Interface for CPU tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2022 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
uint8_t add_sat(const uint8_t &a, const uint8_t &b)
Definition: add.h:33
template void add_same_neon< float >(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
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
uint8x8_t vqadd(const uint8x8_t &a, const uint8x8_t &b)
Definition: add.h:73
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
template void add_same_neon< int16_t >(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
template void add_same_neon< uint8_t >(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
void vstore(uint8_t *ptr, uint8x8_t val)
Definition: store.h:39
void add_same_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
Definition: impl.cpp:34
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
Includes all wrapper headers at once.
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:101
template void add_same_neon< int32_t >(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
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:96
Describe a multidimensional execution window.
Definition: Window.h:39
ConvertPolicy
Policy to handle integer overflow.
Definition: Types.h:404
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:158