KleidiCV Coverage Report

Directory: ./
File: kleidicv/src/arithmetics/add_abs_with_threshold_neon.cpp
Date: 2025-09-25 14:13:34
Exec Total Coverage
Lines: 32 32 100.0%
Functions: 5 5 100.0%
Branches: 24 24 100.0%
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 <limits>
6
7 #include "kleidicv/kleidicv.h"
8 #include "kleidicv/neon.h"
9
10 namespace kleidicv::neon {
11
12 template <typename ScalarType>
13 class SaturatingAddAbsWithThreshold final : public UnrollOnce,
14 public UnrollTwice,
15 public TryToAvoidTailLoop {
16 public:
17 using VecTraits = neon::VecTraits<ScalarType>;
18 using VectorType = typename VecTraits::VectorType;
19
20 78 explicit SaturatingAddAbsWithThreshold(ScalarType threshold)
21 78 : threshold_{threshold}, threshold_vec_{vdupq_n_s16(threshold)} {}
22
23 988 VectorType vector_path(VectorType src_a, VectorType src_b) {
24 988 VectorType add_abs = vqaddq_s16(vqabsq_s16(src_a), vqabsq_s16(src_b));
25 1976 return vandq_s16(add_abs, vcgtq_s16(add_abs, threshold_vec_));
26 988 }
27
28 166 ScalarType scalar_path(ScalarType src_a, ScalarType src_b) {
29 166 ScalarType add_abs = 0;
30
31
2/2
✓ Branch 0 taken 145 times.
✓ Branch 1 taken 21 times.
 166 if (__builtin_add_overflow(saturate_abs(src_a), saturate_abs(src_b),
32 &add_abs)) {
33 21 add_abs = std::numeric_limits<ScalarType>::max();
34 21 }
35
2/2
✓ Branch 0 taken 86 times.
✓ Branch 1 taken 80 times.
 166 return add_abs > threshold_ ? add_abs : 0;
36 166 }
37
38 private:
39 332 ScalarType saturate_abs(ScalarType input) {
40
2/2
✓ Branch 0 taken 5 times.
✓ Branch 1 taken 327 times.
 332 if (std::numeric_limits<ScalarType>::is_signed &&
41 332 input == std::numeric_limits<ScalarType>::lowest()) {
42 5 return std::numeric_limits<ScalarType>::max();
43 }
44 327 return std::abs(input);
45 332 }
46
47 ScalarType threshold_;
48 VectorType threshold_vec_;
49 }; // end of class SaturatingAddAbsWithThreshold<ScalarType>
50
51 template <typename T>
52 86 kleidicv_error_t saturating_add_abs_with_threshold(
53 const T *src_a, size_t src_a_stride, const T *src_b, size_t src_b_stride,
54 T *dst, size_t dst_stride, size_t width, size_t height, T threshold) {
55
4/4
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 84 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 84 times.
 86 CHECK_POINTER_AND_STRIDE(src_a, src_a_stride, height);
56
4/4
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 82 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 82 times.
 84 CHECK_POINTER_AND_STRIDE(src_b, src_b_stride, height);
57
4/4
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 80 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 80 times.
 82 CHECK_POINTER_AND_STRIDE(dst, dst_stride, height);
58
6/6
✓ Branch 0 taken 1 times.
✓ Branch 1 taken 79 times.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 78 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 78 times.
 80 CHECK_IMAGE_SIZE(width, height);
59
60 78 SaturatingAddAbsWithThreshold<T> operation{threshold};
61 78 Rectangle rect{width, height};
62 78 Rows<const T> src_a_rows{src_a, src_a_stride};
63 78 Rows<const T> src_b_rows{src_b, src_b_stride};
64 78 Rows<T> dst_rows{dst, dst_stride};
65 78 apply_operation_by_rows(operation, rect, src_a_rows, src_b_rows, dst_rows);
66 78 return KLEIDICV_OK;
67 86 }
68
69 #define KLEIDICV_INSTANTIATE_TEMPLATE(type) \
70 template KLEIDICV_TARGET_FN_ATTRS kleidicv_error_t \
71 saturating_add_abs_with_threshold<type>( \
72 const type *src_a, size_t src_a_stride, const type *src_b, \
73 size_t src_b_stride, type *dst, size_t dst_stride, size_t width, \
74 size_t height, type threshold)
75
76 KLEIDICV_INSTANTIATE_TEMPLATE(int16_t);
77
78 } // namespace kleidicv::neon
79