xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/NEReverseKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2018-2021 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 #include "src/core/NEON/kernels/NEReverseKernel.h"
25 
26 #include "arm_compute/core/TensorInfo.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/core/Window.h"
29 #include "src/core/NEON/wrapper/wrapper.h"
30 #include "src/core/helpers/AutoConfiguration.h"
31 #include "src/core/helpers/WindowHelpers.h"
32 
33 namespace arm_compute
34 {
35 namespace
36 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * axis)37 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
38 {
39     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output, axis);
40     //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use CPU FP16 instructions.
41     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
42     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(axis, 1, DataType::U32);
43     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->num_dimensions() > 1, "Axis must be a 1D tensor");
44     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->dimension(0) > 4, "Only up to 4 dimensions can be reversed");
45 
46     // Checks performed when output is configured
47     if(output->total_size() != 0)
48     {
49         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
50         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
51         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
52     }
53 
54     return Status{};
55 }
56 } // namespace
57 
NEReverseKernel()58 NEReverseKernel::NEReverseKernel()
59     : _input(nullptr), _output(nullptr), _axis(nullptr)
60 {
61 }
62 
configure(const ITensor * input,ITensor * output,const ITensor * axis)63 void NEReverseKernel::configure(const ITensor *input, ITensor *output, const ITensor *axis)
64 {
65     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, axis);
66 
67     _input  = input;
68     _output = output;
69     _axis   = axis;
70 
71     // Output tensor auto initialization if not yet initialized
72     auto_init_if_empty(*output->info(), *input->info()->clone());
73 
74     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis->info()));
75 
76     // Configure kernel window
77     INEKernel::configure(calculate_max_window(*output->info()));
78 }
79 
validate(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * axis)80 Status NEReverseKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
81 {
82     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis));
83 
84     return Status{};
85 }
86 
87 template <typename T>
run_reverse(const Window & window,const ITensor * input,const ITensor * axis,ITensor * output)88 void run_reverse(const Window &window, const ITensor *input, const ITensor *axis, ITensor *output)
89 {
90     int axis_bit = 0;
91     for(unsigned int i = 0; i < axis->info()->dimension(0); ++i)
92     {
93         const int axis_i = *(reinterpret_cast<const int *>(axis->buffer()) + i);
94         axis_bit |= 1 << axis_i;
95     }
96 
97     // Check if we need a left-over loop for the y dimension
98     const int window_step_x  = 16 / input->info()->element_size();
99     const int window_start_x = window.x().start();
100     const int window_end_x   = window.x().end();
101 
102     Window win(window);
103     win.set(Window::DimX, Window::Dimension(0, 1, 1));
104 
105     Iterator input_it(input, win);
106     execute_window_loop(win, [&](const Coordinates & id)
107     {
108         int x = window_start_x;
109         for(; x <= (window_end_x - window_step_x); x += window_step_x)
110         {
111             auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr()) + x);
112 
113             // Reverse 0 axis
114             if(axis_bit & 0x1)
115             {
116                 in = wrapper::vrev64(in);
117                 in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in));
118             }
119 
120             const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - window_step_x : x;
121             const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
122             const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
123             const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
124 
125             auto out_ptr = reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w)));
126             wrapper::vstore(out_ptr, in);
127         }
128 
129         // Compute left-over elements
130         for(; x < window_end_x; ++x)
131         {
132             const auto in = *(reinterpret_cast<T *>(input_it.ptr()) + x);
133 
134             const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - 1 : x;
135             const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
136             const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
137             const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
138 
139             *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) = in;
140         }
141     },
142     input_it);
143 }
144 
run(const Window & window,const ThreadInfo & info)145 void NEReverseKernel::run(const Window &window, const ThreadInfo &info)
146 {
147     ARM_COMPUTE_UNUSED(info);
148     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
149     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
150 
151     switch(_input->info()->element_size())
152     {
153         case 4:
154             run_reverse<uint32_t>(window, _input, _axis, _output);
155             break;
156         case 2:
157             run_reverse<uint16_t>(window, _input, _axis, _output);
158             break;
159         case 1:
160             run_reverse<uint8_t>(window, _input, _axis, _output);
161             break;
162         default:
163             ARM_COMPUTE_ERROR("Element size not supported");
164     }
165 }
166 } // namespace arm_compute
167