Line | Branch | Exec | Source |
---|---|---|---|
1 | // SPDX-FileCopyrightText: 2024 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
2 | // | ||
3 | // SPDX-License-Identifier: Apache-2.0 | ||
4 | |||
5 | #ifndef KLEIDICV_SEPARABLE_FILTER_15X15_NEON_H | ||
6 | #define KLEIDICV_SEPARABLE_FILTER_15X15_NEON_H | ||
7 | |||
8 | #include "kleidicv/neon.h" | ||
9 | #include "kleidicv/workspace/border_15x15.h" | ||
10 | |||
11 | namespace KLEIDICV_TARGET_NAMESPACE { | ||
12 | |||
13 | // Template for drivers of separable NxM filters. | ||
14 | template <typename FilterType, const size_t S> | ||
15 | class SeparableFilter; | ||
16 | |||
17 | // Driver for a separable 15x15 filter. | ||
18 | template <typename FilterType> | ||
19 | class SeparableFilter<FilterType, 15UL> { | ||
20 | public: | ||
21 | using SourceType = typename FilterType::SourceType; | ||
22 | using BufferType = typename FilterType::BufferType; | ||
23 | using DestinationType = typename FilterType::DestinationType; | ||
24 | using SourceVecTraits = typename neon::VecTraits<SourceType>; | ||
25 | using SourceVectorType = typename SourceVecTraits::VectorType; | ||
26 | using BufferVecTraits = typename neon::VecTraits<BufferType>; | ||
27 | using BufferVectorType = typename BufferVecTraits::VectorType; | ||
28 | using BorderInfoType = | ||
29 | typename ::KLEIDICV_TARGET_NAMESPACE::FixedBorderInfo15x15<SourceType>; | ||
30 | using BorderType = FixedBorderType; | ||
31 | using BorderOffsets = typename BorderInfoType::Offsets; | ||
32 | |||
33 | 32 | explicit SeparableFilter(FilterType filter) : filter_{filter} {} | |
34 | |||
35 | static constexpr size_t margin = 7UL; | ||
36 | |||
37 | 456 | void process_vertical(size_t width, Rows<const SourceType> src_rows, | |
38 | Rows<BufferType> dst_rows, | ||
39 | BorderOffsets border_offsets) const { | ||
40 | 912 | LoopUnroll2<TryToAvoidTailLoop> loop{width * src_rows.channels(), | |
41 | 456 | SourceVecTraits::num_lanes()}; | |
42 | |||
43 | 1464 | loop.unroll_once([&](size_t index) { | |
44 | 1008 | SourceVectorType src[15]; | |
45 | 1008 | src[0] = vld1q(&src_rows.at(border_offsets.c0())[index]); | |
46 | 1008 | src[1] = vld1q(&src_rows.at(border_offsets.c1())[index]); | |
47 | 1008 | src[2] = vld1q(&src_rows.at(border_offsets.c2())[index]); | |
48 | 1008 | src[3] = vld1q(&src_rows.at(border_offsets.c3())[index]); | |
49 | 1008 | src[4] = vld1q(&src_rows.at(border_offsets.c4())[index]); | |
50 | 1008 | src[5] = vld1q(&src_rows.at(border_offsets.c5())[index]); | |
51 | 1008 | src[6] = vld1q(&src_rows.at(border_offsets.c6())[index]); | |
52 | 1008 | src[7] = vld1q(&src_rows.at(border_offsets.c7())[index]); | |
53 | 1008 | src[8] = vld1q(&src_rows.at(border_offsets.c8())[index]); | |
54 | 1008 | src[9] = vld1q(&src_rows.at(border_offsets.c9())[index]); | |
55 | 1008 | src[10] = vld1q(&src_rows.at(border_offsets.c10())[index]); | |
56 | 1008 | src[11] = vld1q(&src_rows.at(border_offsets.c11())[index]); | |
57 | 1008 | src[12] = vld1q(&src_rows.at(border_offsets.c12())[index]); | |
58 | 1008 | src[13] = vld1q(&src_rows.at(border_offsets.c13())[index]); | |
59 | 1008 | src[14] = vld1q(&src_rows.at(border_offsets.c14())[index]); | |
60 | 1008 | filter_.vertical_vector_path(src, &dst_rows[index]); | |
61 | 1008 | }); | |
62 | |||
63 | 2256 | loop.tail([&](size_t index) { | |
64 | 1800 | SourceType src[15]; | |
65 | 1800 | src[0] = src_rows.at(border_offsets.c0())[index]; | |
66 | 1800 | src[1] = src_rows.at(border_offsets.c1())[index]; | |
67 | 1800 | src[2] = src_rows.at(border_offsets.c2())[index]; | |
68 | 1800 | src[3] = src_rows.at(border_offsets.c3())[index]; | |
69 | 1800 | src[4] = src_rows.at(border_offsets.c4())[index]; | |
70 | 1800 | src[5] = src_rows.at(border_offsets.c5())[index]; | |
71 | 1800 | src[6] = src_rows.at(border_offsets.c6())[index]; | |
72 | 1800 | src[7] = src_rows.at(border_offsets.c7())[index]; | |
73 | 1800 | src[8] = src_rows.at(border_offsets.c8())[index]; | |
74 | 1800 | src[9] = src_rows.at(border_offsets.c9())[index]; | |
75 | 1800 | src[10] = src_rows.at(border_offsets.c10())[index]; | |
76 | 1800 | src[11] = src_rows.at(border_offsets.c11())[index]; | |
77 | 1800 | src[12] = src_rows.at(border_offsets.c12())[index]; | |
78 | 1800 | src[13] = src_rows.at(border_offsets.c13())[index]; | |
79 | 1800 | src[14] = src_rows.at(border_offsets.c14())[index]; | |
80 | 1800 | filter_.vertical_scalar_path(src, &dst_rows[index]); | |
81 | 1800 | }); | |
82 | 456 | } | |
83 | |||
84 | 456 | void process_horizontal(size_t width, Rows<const BufferType> src_rows, | |
85 | Rows<DestinationType> dst_rows, | ||
86 | BorderOffsets border_offsets) const { | ||
87 | 912 | LoopUnroll2<TryToAvoidTailLoop> loop{width * src_rows.channels(), | |
88 | 456 | BufferVecTraits::num_lanes()}; | |
89 | |||
90 | 568 | loop.unroll_twice([&](size_t index) { | |
91 | 112 | auto src_0 = &src_rows.at(0, border_offsets.c0())[index]; | |
92 | 112 | auto src_1 = &src_rows.at(0, border_offsets.c1())[index]; | |
93 | 112 | auto src_2 = &src_rows.at(0, border_offsets.c2())[index]; | |
94 | 112 | auto src_3 = &src_rows.at(0, border_offsets.c3())[index]; | |
95 | 112 | auto src_4 = &src_rows.at(0, border_offsets.c4())[index]; | |
96 | 112 | auto src_5 = &src_rows.at(0, border_offsets.c5())[index]; | |
97 | 112 | auto src_6 = &src_rows.at(0, border_offsets.c6())[index]; | |
98 | 112 | auto src_7 = &src_rows.at(0, border_offsets.c7())[index]; | |
99 | 112 | auto src_8 = &src_rows.at(0, border_offsets.c8())[index]; | |
100 | 112 | auto src_9 = &src_rows.at(0, border_offsets.c9())[index]; | |
101 | 112 | auto src_10 = &src_rows.at(0, border_offsets.c10())[index]; | |
102 | 112 | auto src_11 = &src_rows.at(0, border_offsets.c11())[index]; | |
103 | 112 | auto src_12 = &src_rows.at(0, border_offsets.c12())[index]; | |
104 | 112 | auto src_13 = &src_rows.at(0, border_offsets.c13())[index]; | |
105 | 112 | auto src_14 = &src_rows.at(0, border_offsets.c14())[index]; | |
106 | |||
107 | 112 | BufferVectorType src_a[15], src_b[15]; | |
108 | 112 | src_a[0] = vld1q(&src_0[0]); | |
109 | 112 | src_b[0] = vld1q(&src_0[BufferVecTraits::num_lanes()]); | |
110 | 112 | src_a[1] = vld1q(&src_1[0]); | |
111 | 112 | src_b[1] = vld1q(&src_1[BufferVecTraits::num_lanes()]); | |
112 | 112 | src_a[2] = vld1q(&src_2[0]); | |
113 | 112 | src_b[2] = vld1q(&src_2[BufferVecTraits::num_lanes()]); | |
114 | 112 | src_a[3] = vld1q(&src_3[0]); | |
115 | 112 | src_b[3] = vld1q(&src_3[BufferVecTraits::num_lanes()]); | |
116 | 112 | src_a[4] = vld1q(&src_4[0]); | |
117 | 112 | src_b[4] = vld1q(&src_4[BufferVecTraits::num_lanes()]); | |
118 | 112 | src_a[5] = vld1q(&src_5[0]); | |
119 | 112 | src_b[5] = vld1q(&src_5[BufferVecTraits::num_lanes()]); | |
120 | 112 | src_a[6] = vld1q(&src_6[0]); | |
121 | 112 | src_b[6] = vld1q(&src_6[BufferVecTraits::num_lanes()]); | |
122 | 112 | src_a[7] = vld1q(&src_7[0]); | |
123 | 112 | src_b[7] = vld1q(&src_7[BufferVecTraits::num_lanes()]); | |
124 | 112 | src_a[8] = vld1q(&src_8[0]); | |
125 | 112 | src_b[8] = vld1q(&src_8[BufferVecTraits::num_lanes()]); | |
126 | 112 | src_a[9] = vld1q(&src_9[0]); | |
127 | 112 | src_b[9] = vld1q(&src_9[BufferVecTraits::num_lanes()]); | |
128 | 112 | src_a[10] = vld1q(&src_10[0]); | |
129 | 112 | src_b[10] = vld1q(&src_10[BufferVecTraits::num_lanes()]); | |
130 | 112 | src_a[11] = vld1q(&src_11[0]); | |
131 | 112 | src_b[11] = vld1q(&src_11[BufferVecTraits::num_lanes()]); | |
132 | 112 | src_a[12] = vld1q(&src_12[0]); | |
133 | 112 | src_b[12] = vld1q(&src_12[BufferVecTraits::num_lanes()]); | |
134 | 112 | src_a[13] = vld1q(&src_13[0]); | |
135 | 112 | src_b[13] = vld1q(&src_13[BufferVecTraits::num_lanes()]); | |
136 | 112 | src_a[14] = vld1q(&src_14[0]); | |
137 | 112 | src_b[14] = vld1q(&src_14[BufferVecTraits::num_lanes()]); | |
138 | |||
139 | 112 | filter_.horizontal_vector_path(src_a, &dst_rows[index]); | |
140 | 224 | filter_.horizontal_vector_path( | |
141 | 112 | src_b, &dst_rows[index + BufferVecTraits::num_lanes()]); | |
142 | 112 | }); | |
143 | |||
144 | 568 | loop.unroll_once([&](size_t index) { | |
145 | 112 | BufferVectorType src[15]; | |
146 | 112 | src[0] = vld1q(&src_rows.at(0, border_offsets.c0())[index]); | |
147 | 112 | src[1] = vld1q(&src_rows.at(0, border_offsets.c1())[index]); | |
148 | 112 | src[2] = vld1q(&src_rows.at(0, border_offsets.c2())[index]); | |
149 | 112 | src[3] = vld1q(&src_rows.at(0, border_offsets.c3())[index]); | |
150 | 112 | src[4] = vld1q(&src_rows.at(0, border_offsets.c4())[index]); | |
151 | 112 | src[5] = vld1q(&src_rows.at(0, border_offsets.c5())[index]); | |
152 | 112 | src[6] = vld1q(&src_rows.at(0, border_offsets.c6())[index]); | |
153 | 112 | src[7] = vld1q(&src_rows.at(0, border_offsets.c7())[index]); | |
154 | 112 | src[8] = vld1q(&src_rows.at(0, border_offsets.c8())[index]); | |
155 | 112 | src[9] = vld1q(&src_rows.at(0, border_offsets.c9())[index]); | |
156 | 112 | src[10] = vld1q(&src_rows.at(0, border_offsets.c10())[index]); | |
157 | 112 | src[11] = vld1q(&src_rows.at(0, border_offsets.c11())[index]); | |
158 | 112 | src[12] = vld1q(&src_rows.at(0, border_offsets.c12())[index]); | |
159 | 112 | src[13] = vld1q(&src_rows.at(0, border_offsets.c13())[index]); | |
160 | 112 | src[14] = vld1q(&src_rows.at(0, border_offsets.c14())[index]); | |
161 | 112 | filter_.horizontal_vector_path(src, &dst_rows[index]); | |
162 | 112 | }); | |
163 | |||
164 | 1136 | loop.tail([&](size_t index) { | |
165 | 680 | process_horizontal_scalar(src_rows, dst_rows, border_offsets, index); | |
166 | 680 | }); | |
167 | 456 | } | |
168 | |||
169 | 6384 | void process_horizontal_borders(Rows<const BufferType> src_rows, | |
170 | Rows<DestinationType> dst_rows, | ||
171 | BorderOffsets border_offsets) const { | ||
172 |
2/2✓ Branch 0 taken 6384 times.
✓ Branch 1 taken 11088 times.
|
17472 | for (size_t index = 0; index < src_rows.channels(); ++index) { |
173 | 11088 | disable_loop_vectorization(); | |
174 | 11088 | process_horizontal_scalar(src_rows, dst_rows, border_offsets, index); | |
175 | 11088 | } | |
176 | 6384 | } | |
177 | |||
178 | private: | ||
179 | 11768 | void process_horizontal_scalar(Rows<const BufferType> src_rows, | |
180 | Rows<DestinationType> dst_rows, | ||
181 | BorderOffsets border_offsets, | ||
182 | size_t index) const { | ||
183 | 11768 | BufferType src[15]; | |
184 | 11768 | src[0] = src_rows.at(0, border_offsets.c0())[index]; | |
185 | 11768 | src[1] = src_rows.at(0, border_offsets.c1())[index]; | |
186 | 11768 | src[2] = src_rows.at(0, border_offsets.c2())[index]; | |
187 | 11768 | src[3] = src_rows.at(0, border_offsets.c3())[index]; | |
188 | 11768 | src[4] = src_rows.at(0, border_offsets.c4())[index]; | |
189 | 11768 | src[5] = src_rows.at(0, border_offsets.c5())[index]; | |
190 | 11768 | src[6] = src_rows.at(0, border_offsets.c6())[index]; | |
191 | 11768 | src[7] = src_rows.at(0, border_offsets.c7())[index]; | |
192 | 11768 | src[8] = src_rows.at(0, border_offsets.c8())[index]; | |
193 | 11768 | src[9] = src_rows.at(0, border_offsets.c9())[index]; | |
194 | 11768 | src[10] = src_rows.at(0, border_offsets.c10())[index]; | |
195 | 11768 | src[11] = src_rows.at(0, border_offsets.c11())[index]; | |
196 | 11768 | src[12] = src_rows.at(0, border_offsets.c12())[index]; | |
197 | 11768 | src[13] = src_rows.at(0, border_offsets.c13())[index]; | |
198 | 11768 | src[14] = src_rows.at(0, border_offsets.c14())[index]; | |
199 | 11768 | filter_.horizontal_scalar_path(src, &dst_rows[index]); | |
200 | 11768 | } | |
201 | |||
202 | FilterType filter_; | ||
203 | }; // end of class SeparableFilter<FilterType, 15UL> | ||
204 | |||
205 | // Shorthand for 15x15 separable filters driver type. | ||
206 | template <class FilterType> | ||
207 | using SeparableFilter15x15 = SeparableFilter<FilterType, 15UL>; | ||
208 | |||
209 | } // namespace KLEIDICV_TARGET_NAMESPACE | ||
210 | |||
211 | #endif // KLEIDICV_SEPARABLE_FILTER_15X15_NEON_H | ||
212 |