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 */
24 #include "arm_compute/runtime/NEON/functions/NEConvolution.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/ITensor.h"
28 #include "arm_compute/core/PixelValue.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Utils.h"
31 #include "arm_compute/core/Validate.h"
32 #include "arm_compute/runtime/NEON/NEScheduler.h"
33 #include "arm_compute/runtime/TensorAllocator.h"
34 #include "src/core/NEON/kernels/NEConvolutionKernel.h"
35 #include "src/core/NEON/kernels/NEConvolutionKernel.h"
36 #include "src/core/NEON/kernels/NEFillBorderKernel.h"
37 #include "support/MemorySupport.h"
38
39 #include <array>
40 #include <utility>
41
42 namespace arm_compute
43 {
44 NEConvolution3x3::~NEConvolution3x3() = default;
45
configure(ITensor * input,ITensor * output,const int16_t * conv,uint32_t scale,BorderMode border_mode,uint8_t constant_border_value)46 void NEConvolution3x3::configure(ITensor *input, ITensor *output, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value)
47 {
48 auto k = arm_compute::support::cpp14::make_unique<NEConvolution3x3Kernel>();
49 k->configure(input, output, conv, scale, border_mode == BorderMode::UNDEFINED);
50 _kernel = std::move(k);
51
52 auto b = arm_compute::support::cpp14::make_unique<NEFillBorderKernel>();
53 b->configure(input, _kernel->border_size(), border_mode, PixelValue(constant_border_value));
54 _border_handler = std::move(b);
55 }
56
57 template <unsigned int matrix_size>
58 NEConvolutionSquare<matrix_size>::~NEConvolutionSquare() = default;
59
60 template <unsigned int matrix_size>
NEConvolutionSquare(std::shared_ptr<IMemoryManager> memory_manager)61 NEConvolutionSquare<matrix_size>::NEConvolutionSquare(std::shared_ptr<IMemoryManager> memory_manager)
62 : _memory_group(std::move(memory_manager)), _tmp(), _is_separable(false), _kernel_hor(), _kernel_vert(), _kernel(), _border_handler()
63 {
64 }
65
66 template <unsigned int matrix_size>
configure(ITensor * input,ITensor * output,const int16_t * conv,uint32_t scale,BorderMode border_mode,uint8_t constant_border_value)67 void NEConvolutionSquare<matrix_size>::configure(ITensor *input, ITensor *output, const int16_t *conv, uint32_t scale, BorderMode border_mode,
68 uint8_t constant_border_value)
69 {
70 ARM_COMPUTE_ERROR_ON(conv == nullptr);
71 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8);
72 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16);
73
74 std::array<int16_t, matrix_size> conv_col{ { 0 } };
75 std::array<int16_t, matrix_size> conv_row{ { 0 } };
76
77 _is_separable = separate_matrix(conv, conv_col.data(), conv_row.data(), matrix_size);
78
79 auto b = arm_compute::support::cpp14::make_unique<NEFillBorderKernel>();
80 if(_is_separable)
81 {
82 DataType intermediate_type = DataType::UNKNOWN;
83 std::tie(std::ignore, intermediate_type) = data_type_for_convolution(conv_col.data(), conv_row.data(), matrix_size);
84
85 _tmp.allocator()->init(TensorInfo(input->info()->tensor_shape(), 1, intermediate_type));
86
87 // Manage intermediate buffers
88 _memory_group.manage(&_tmp);
89
90 // Calculate scale
91 if(scale == 0)
92 {
93 scale = calculate_matrix_scale(conv, matrix_size);
94 }
95
96 _kernel_hor = arm_compute::support::cpp14::make_unique<NESeparableConvolutionHorKernel<matrix_size>>();
97 _kernel_vert = arm_compute::support::cpp14::make_unique<NESeparableConvolutionVertKernel<matrix_size>>();
98
99 _kernel_hor->configure(input, &_tmp, conv_row.data(), border_mode == BorderMode::UNDEFINED);
100 _kernel_vert->configure(&_tmp, output, conv_col.data(), scale, border_mode == BorderMode::UNDEFINED);
101
102 _tmp.allocator()->allocate();
103
104 b->configure(input, _kernel_hor->border_size(), border_mode, PixelValue(constant_border_value));
105 }
106 else
107 {
108 _kernel = arm_compute::support::cpp14::make_unique<NEConvolutionKernel<matrix_size>>();
109 _kernel->configure(input, output, conv, scale, border_mode == BorderMode::UNDEFINED);
110 b->configure(input, _kernel->border_size(), border_mode, PixelValue(constant_border_value));
111 }
112 _border_handler = std::move(b);
113 }
114
115 template <unsigned int matrix_size>
run()116 void NEConvolutionSquare<matrix_size>::run()
117 {
118 NEScheduler::get().schedule(_border_handler.get(), Window::DimZ);
119
120 if(_is_separable)
121 {
122 MemoryGroupResourceScope scope_mg(_memory_group);
123
124 NEScheduler::get().schedule(_kernel_hor.get(), Window::DimY);
125 NEScheduler::get().schedule(_kernel_vert.get(), Window::DimY);
126 }
127 else
128 {
129 NEScheduler::get().schedule(_kernel.get(), Window::DimY);
130 }
131 }
132
133 template class arm_compute::NEConvolutionSquare<5>;
134 template class arm_compute::NEConvolutionSquare<7>;
135 template class arm_compute::NEConvolutionSquare<9>;
136
137 NEConvolutionRectangle::~NEConvolutionRectangle() = default;
138
configure(ITensor * input,ITensor * output,const int16_t * conv,uint32_t rows,uint32_t cols,uint32_t scale,BorderMode border_mode,uint8_t constant_border_value)139 void NEConvolutionRectangle::configure(ITensor *input, ITensor *output, const int16_t *conv, uint32_t rows, uint32_t cols, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value)
140 {
141 auto k = arm_compute::support::cpp14::make_unique<NEConvolutionRectangleKernel>();
142 k->configure(input, output, conv, rows, cols, scale, border_mode == BorderMode::UNDEFINED);
143 _kernel = std::move(k);
144
145 auto b = arm_compute::support::cpp14::make_unique<NEFillBorderKernel>();
146 b->configure(input, _kernel->border_size(), border_mode, PixelValue(constant_border_value));
147 _border_handler = std::move(b);
148 }
149 } // namespace arm_compute
150