KleidiCV Coverage Report

Directory: ./
File: kleidicv/src/resize/resize_neon.cpp
Date: 2025-09-25 14:13:34
Exec Total Coverage
Lines: 78 78 100.0%
Functions: 2 2 100.0%
Branches: 46 46 100.0%
Line Branch Exec Source
1 // SPDX-FileCopyrightText: 2023 - 2024 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
2/2
✓ Branch 0 taken 146 times.
✓ Branch 1 taken 57 times.
 349 for (; src_height >= 2; src_height -= 2, src += (src_stride * 2),
49 146 --dst_height, dst += dst_stride) {
50 146 const uint8_t *src_l = src;
51 146 uint8_t *dst_l = dst;
52 146 size_t src_width_l = src_width;
53 146 size_t dst_width_l = dst_width;
54
55
2/2
✓ Branch 0 taken 6 times.
✓ Branch 1 taken 146 times.
 152 for (; src_width_l >= 32;
56 6 src_width_l -= 32, dst_width_l -= 16, dst_l += 16, src_l += 32) {
57 6 uint8x16_t top_line_0 = vld1q_u8(src_l);
58 6 uint8x16_t top_line_1 = vld1q_u8(&src_l[16]);
59 6 uint8x16_t bottom_line_0 = vld1q_u8(&src_l[src_stride]);
60 6 uint8x16_t bottom_line_1 = vld1q_u8(&src_l[src_stride + 16]);
61
62 6 uint16x8_t top_line_pairs_summed_0 = vpaddlq_u8(top_line_0);
63 6 uint16x8_t top_line_pairs_summed_1 = vpaddlq_u8(top_line_1);
64 6 uint16x8_t bottom_line_pairs_summed_0 = vpaddlq_u8(bottom_line_0);
65 6 uint16x8_t bottom_line_pairs_summed_1 = vpaddlq_u8(bottom_line_1);
66
67 12 uint16x8_t result_before_averaging_0 =
68 6 vaddq_u16(top_line_pairs_summed_0, bottom_line_pairs_summed_0);
69 12 uint16x8_t result_before_averaging_1 =
70 6 vaddq_u16(top_line_pairs_summed_1, bottom_line_pairs_summed_1);
71
72 6 uint8x8_t result_0 = vrshrn_n_u16(result_before_averaging_0, 2);
73 6 uint8x8_t result_1 = vrshrn_n_u16(result_before_averaging_1, 2);
74
75 6 vst1_u8(&dst_l[0], result_0);
76 6 vst1_u8(&dst_l[8], result_1);
77 6 }
78
79
2/2
✓ Branch 0 taken 648 times.
✓ Branch 1 taken 146 times.
 794 for (; src_width_l > 1;
80 648 src_width_l -= 2, src_l += 2, --dst_width_l, ++dst_l) {
81 648 disable_loop_vectorization();
82 648 *dst_l = rounding_shift_right<uint16_t>(
83 1296 static_cast<uint16_t>(*src_l) + *(src_l + 1) + *(src_l + src_stride) +
84 648 *(src_l + src_stride + 1),
85 2);
86 648 }
87
88
2/2
✓ Branch 0 taken 126 times.
✓ Branch 1 taken 20 times.
 146 if (dst_width_l) {
89 20 *dst_l = rounding_shift_right<uint16_t>(
90 20 static_cast<uint16_t>(*src_l) + *(src_l + src_stride), 1);
91 20 }
92 146 }
93
94
2/2
✓ Branch 0 taken 40 times.
✓ Branch 1 taken 17 times.
 57 if (dst_height) {
95
2/2
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 17 times.
 18 for (; src_width >= 32;
96 1 src_width -= 32, dst_width -= 16, dst += 16, src += 32) {
97 1 uint8x16_t vsrc_0 = vld1q_u8(&src[0]);
98 1 uint8x16_t vsrc_1 = vld1q_u8(&src[16]);
99
100 1 uint16x8_t vsrc_line_pairs_summed_0 = vpaddlq_u8(vsrc_0);
101 1 uint16x8_t vsrc_line_pairs_summed_1 = vpaddlq_u8(vsrc_1);
102
103 1 uint8x8_t result_0 = vrshrn_n_u16(vsrc_line_pairs_summed_0, 1);
104 1 uint8x8_t result_1 = vrshrn_n_u16(vsrc_line_pairs_summed_1, 1);
105
106 1 vst1_u8(&dst[0], result_0);
107 1 vst1_u8(&dst[8], result_1);
108 1 }
109
110
2/2
✓ Branch 0 taken 53 times.
✓ Branch 1 taken 17 times.
 70 for (; src_width > 1; src_width -= 2, src += 2, --dst_width, ++dst) {
111 53 disable_loop_vectorization();
112 53 *dst = rounding_shift_right<uint16_t>(
113 53 static_cast<uint16_t>(*src) + *(src + 1), 1);
114 53 }
115
116
2/2
✓ Branch 0 taken 9 times.
✓ Branch 1 taken 8 times.
 17 if (dst_width) {
117 8 *dst = *src;
118 8 }
119 17 }
120 57 return KLEIDICV_OK;
121 65 }
122 // NOLINTEND(readability-function-cognitive-complexity)
123
124 } // namespace kleidicv::neon
125