KleidiCV Coverage Report

Directory: ./
File: kleidicv/src/resize/resize_neon.cpp
Date: 2025-11-25 17:23:32
Exec Total Coverage
Lines: 101 101 100.0%
Functions: 2 2 100.0%
Branches: 46 46 100.0%
Line Branch Exec Source
1 // SPDX-FileCopyrightText: 2023 - 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
2 //
3 // SPDX-License-Identifier: Apache-2.0
4
5 #include "kleidicv/kleidicv.h"
6 #include "kleidicv/neon.h"
7 #include "kleidicv/resize/resize.h"
8
9 namespace kleidicv::neon {
10
11 KLEIDICV_TARGET_FN_ATTRS
12 120 static kleidicv_error_t check_dimensions(size_t src_dim, size_t dst_dim) {
13 120 size_t half_src_dim = src_dim / 2;
14
15
2/2
✓ Branch 0 taken 58 times.
✓ Branch 1 taken 62 times.
 120 if ((src_dim % 2) == 0) {
16
2/2
✓ Branch 0 taken 60 times.
✓ Branch 1 taken 2 times.
 62 if (dst_dim == half_src_dim) {
17 60 return KLEIDICV_OK;
18 }
19 2 } else {
20
4/4
✓ Branch 0 taken 34 times.
✓ Branch 1 taken 24 times.
✓ Branch 2 taken 32 times.
✓ Branch 3 taken 2 times.
 58 if (dst_dim == half_src_dim || dst_dim == (half_src_dim + 1)) {
21 56 return KLEIDICV_OK;
22 }
23 }
24
25 4 return KLEIDICV_ERROR_RANGE;
26 120 }
27
28 // Disable the warning, as the complexity is just above the threshold, it's
29 // better to leave it in one piece.
30 // NOLINTBEGIN(readability-function-cognitive-complexity)
31 KLEIDICV_TARGET_FN_ATTRS
32 65 kleidicv_error_t resize_to_quarter_u8(const uint8_t *src, size_t src_stride,
33 size_t src_width, size_t src_height,
34 uint8_t *dst, size_t dst_stride,
35 size_t dst_width, size_t dst_height) {
36
4/4
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 64 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 64 times.
 65 CHECK_POINTER_AND_STRIDE(src, src_stride, src_height);
37
4/4
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 63 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 63 times.
 64 CHECK_POINTER_AND_STRIDE(dst, dst_stride, dst_height);
38
6/6
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 62 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 61 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 61 times.
 63 CHECK_IMAGE_SIZE(src_width, src_height);
39
40
4/4
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 59 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 59 times.
 63 if (kleidicv_error_t ret = check_dimensions(src_width, dst_width)) {
41 2 return ret;
42 }
43
44
4/4
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 57 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 57 times.
 61 if (kleidicv_error_t ret = check_dimensions(src_height, dst_height)) {
45 2 return ret;
46 }
47
48 using VecTraits = neon::VecTraits<uint8_t>;
49 57 constexpr size_t kVectorLengthX2 = kVectorLength * 2;
50 57 constexpr size_t kVectorLengthX4 = kVectorLength * 4;
51
52
2/2
✓ Branch 0 taken 146 times.
✓ Branch 1 taken 57 times.
 349 for (; src_height >= 2; src_height -= 2, src += (src_stride * 2),
53 146 --dst_height, dst += dst_stride) {
54 146 const uint8_t *src_l = src;
55 146 uint8_t *dst_l = dst;
56 146 size_t src_width_l = src_width;
57 146 size_t dst_width_l = dst_width;
58
59
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 146 times.
 152 for (; src_width_l >= kVectorLengthX4;
60 6 src_width_l -= kVectorLengthX4, dst_width_l -= kVectorLengthX2,
61 6 dst_l += kVectorLengthX2, src_l += kVectorLengthX4) {
62 6 KLEIDICV_PREFETCH(src_l + 1024);
63 6 KLEIDICV_PREFETCH(src_l + src_stride + 1024);
64
65 6 uint8x16x4_t top_line, bottom_line;
66 6 uint16x8_t top_line_pairs_summed[4];
67 6 uint16x8_t bottom_line_pairs_summed[4];
68 6 uint16x8_t result_before_averaging[4];
69 6 uint8x16x2_t result;
70
71 6 VecTraits::load(src_l, top_line);
72 6 VecTraits::load(&src_l[src_stride], bottom_line);
73
74 6 top_line_pairs_summed[0] = vpaddlq_u8(top_line.val[0]);
75 6 top_line_pairs_summed[1] = vpaddlq_u8(top_line.val[1]);
76 6 top_line_pairs_summed[2] = vpaddlq_u8(top_line.val[2]);
77 6 top_line_pairs_summed[3] = vpaddlq_u8(top_line.val[3]);
78
79 6 bottom_line_pairs_summed[0] = vpaddlq_u8(bottom_line.val[0]);
80 6 bottom_line_pairs_summed[1] = vpaddlq_u8(bottom_line.val[1]);
81 6 bottom_line_pairs_summed[2] = vpaddlq_u8(bottom_line.val[2]);
82 6 bottom_line_pairs_summed[3] = vpaddlq_u8(bottom_line.val[3]);
83
84 6 result_before_averaging[0] =
85 6 vaddq_u16(top_line_pairs_summed[0], bottom_line_pairs_summed[0]);
86 6 result_before_averaging[1] =
87 6 vaddq_u16(top_line_pairs_summed[1], bottom_line_pairs_summed[1]);
88 6 result_before_averaging[2] =
89 6 vaddq_u16(top_line_pairs_summed[2], bottom_line_pairs_summed[2]);
90 6 result_before_averaging[3] =
91 6 vaddq_u16(top_line_pairs_summed[3], bottom_line_pairs_summed[3]);
92
93 6 result.val[0] =
94 6 vrshrn_high_n_u16(vrshrn_n_u16(result_before_averaging[0], 2),
95 result_before_averaging[1], 2);
96 6 result.val[1] =
97 6 vrshrn_high_n_u16(vrshrn_n_u16(result_before_averaging[2], 2),
98 result_before_averaging[3], 2);
99
100 6 VecTraits::store(result, dst_l);
101 6 }
102
103
2/2
✓ Branch 0 taken 840 times.
✓ Branch 1 taken 146 times.
 986 for (; src_width_l > 1;
104 840 src_width_l -= 2, src_l += 2, --dst_width_l, ++dst_l) {
105 840 disable_loop_vectorization();
106 840 *dst_l = rounding_shift_right<uint16_t>(
107 1680 static_cast<uint16_t>(*src_l) + *(src_l + 1) + *(src_l + src_stride) +
108 840 *(src_l + src_stride + 1),
109 2);
110 840 }
111
112
2/2
✓ Branch 0 taken 126 times.
✓ Branch 1 taken 20 times.
 146 if (dst_width_l) {
113 20 *dst_l = rounding_shift_right<uint16_t>(
114 20 static_cast<uint16_t>(*src_l) + *(src_l + src_stride), 1);
115 20 }
116 146 }
117
118
2/2
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 17 times.
 57 if (dst_height) {
119
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 17 times.
 18 for (; src_width >= kVectorLengthX4;
120 1 src_width -= kVectorLengthX4, dst_width -= kVectorLengthX2,
121 1 dst += kVectorLengthX2, src += kVectorLengthX4) {
122 1 uint8x16x4_t vsrc;
123 1 uint16x8_t vsrc_line_pairs_summed[4];
124 1 uint8x16x2_t result;
125 1 VecTraits::load(&src[0], vsrc);
126
127 1 vsrc_line_pairs_summed[0] = vpaddlq_u8(vsrc.val[0]);
128 1 vsrc_line_pairs_summed[1] = vpaddlq_u8(vsrc.val[1]);
129 1 vsrc_line_pairs_summed[2] = vpaddlq_u8(vsrc.val[2]);
130 1 vsrc_line_pairs_summed[3] = vpaddlq_u8(vsrc.val[3]);
131
132 1 result.val[0] =
133 1 vrshrn_high_n_u16(vrshrn_n_u16(vsrc_line_pairs_summed[0], 1),
134 vsrc_line_pairs_summed[1], 1);
135 1 result.val[1] =
136 1 vrshrn_high_n_u16(vrshrn_n_u16(vsrc_line_pairs_summed[2], 1),
137 vsrc_line_pairs_summed[3], 1);
138
139 1 VecTraits::store(result, dst);
140 1 }
141
142
2/2
✓ Branch 0 taken 101 times.
✓ Branch 1 taken 17 times.
 118 for (; src_width > 1; src_width -= 2, src += 2, --dst_width, ++dst) {
143 101 disable_loop_vectorization();
144 101 *dst = rounding_shift_right<uint16_t>(
145 101 static_cast<uint16_t>(*src) + *(src + 1), 1);
146 101 }
147
148
2/2
✓ Branch 0 taken 9 times.
✓ Branch 1 taken 8 times.
 17 if (dst_width) {
149 8 *dst = *src;
150 8 }
151 17 }
152 57 return KLEIDICV_OK;
153 65 }
154 // NOLINTEND(readability-function-cognitive-complexity)
155
156 } // namespace kleidicv::neon
157