KleidiCV Coverage Report

Directory: ./
File: kleidicv/src/arithmetics/in_range_sc.h
Date: 2025-09-25 14:13:34
Exec Total Coverage
Lines: 90 90 100.0%
Functions: 20 20 100.0%
Branches: 30 30 100.0%
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 #ifndef KLEIDICV_IN_RANGE_SC_H
6 #define KLEIDICV_IN_RANGE_SC_H
7
8 #include "kleidicv/kleidicv.h"
9 #include "kleidicv/sve2.h"
10
11 namespace KLEIDICV_TARGET_NAMESPACE {
12
13 template <typename ScalarType>
14 class InRange;
15
16 template <>
17 class InRange<uint8_t> : public UnrollTwice {
18 public:
19 using ContextType = Context;
20 using VecTraits = KLEIDICV_TARGET_NAMESPACE::VecTraits<uint8_t>;
21 using VectorType = typename VecTraits::VectorType;
22 using SignedScalarType = typename std::make_signed<uint8_t>::type;
23 using SignedVecTraits =
24 KLEIDICV_TARGET_NAMESPACE::VecTraits<SignedScalarType>;
25 using SignedVectorType = typename SignedVecTraits::VectorType;
26
27 150 InRange(VectorType &vec_lower_bound,
28 VectorType &vec_upper_bound) KLEIDICV_STREAMING
29 150 : vec_lower_bound_(vec_lower_bound),
30 150 vec_upper_bound_(vec_upper_bound) {}
31
32 // NOLINTBEGIN(readability-make-member-function-const)
33 693 VectorType vector_path(ContextType ctx, VectorType src) KLEIDICV_STREAMING {
34 693 svbool_t pg = ctx.predicate();
35
36 693 VectorType diff_low = svsub_x(pg, src, vec_lower_bound_);
37 // Shift subtraction result 7 bits to the right, i.e. divide by 2^7 to keep
38 // sign bit only.
39 1386 VectorType result_within_low =
40 693 VecTraits::svreinterpret(SignedVecTraits::svasr_n(
41 693 pg, SignedVecTraits::svreinterpret(diff_low), 7));
42
43 693 VectorType diff_up = svsub_x(pg, vec_upper_bound_, src);
44 1386 VectorType result_within_up =
45 693 VecTraits::svreinterpret(SignedVecTraits::svasr_n(
46 693 pg, SignedVecTraits::svreinterpret(diff_up), 7));
47
48 // src[i] < lower_bound OR src[i] > upper_bound
49 693 VectorType out_of_range = svorr_x(pg, result_within_low, result_within_up);
50 // NOT(out_of_range) to set within elements to 1 and the rest to 0.
51 693 VectorType within_range = svcnot_x(pg, out_of_range);
52 // Negate to set elements within to 0xFF (all 1s).
53 1386 return VecTraits::svreinterpret(
54 693 svqneg_x(pg, SignedVecTraits::svreinterpret(within_range)));
55 693 }
56 // NOLINTEND(readability-make-member-function-const)
57
58 private:
59 VectorType &vec_lower_bound_;
60 VectorType &vec_upper_bound_;
61 }; // end of class InRange<uint8_t>
62
63 template <>
64 class InRange<float> {
65 public:
66 using SrcVecTraits = KLEIDICV_TARGET_NAMESPACE::VecTraits<float>;
67 using SrcVectorType = typename SrcVecTraits::VectorType;
68 using DstVecTraits = KLEIDICV_TARGET_NAMESPACE::VecTraits<uint8_t>;
69 using DstVectorType = typename DstVecTraits::VectorType;
70
71 150 InRange(float lower_bound, float upper_bound) KLEIDICV_STREAMING
72 150 : lower_bound_(lower_bound),
73 150 upper_bound_(upper_bound) {}
74
75 179 void process_row(size_t width, Columns<const float> src,
76 Columns<uint8_t> dst) KLEIDICV_STREAMING {
77 358 LoopUnroll{width, SrcVecTraits::num_lanes()}
78 739 .unroll_n_times<4>([&](size_t step) KLEIDICV_STREAMING {
79 560 svbool_t pg_src = SrcVecTraits::svptrue();
80 560 SrcVectorType src_v0 = svld1(pg_src, &src[0]);
81 560 SrcVectorType src_v1 = svld1_vnum(pg_src, &src[0], 1);
82 560 SrcVectorType src_v2 = svld1_vnum(pg_src, &src[0], 2);
83 560 SrcVectorType src_v3 = svld1_vnum(pg_src, &src[0], 3);
84 1120 DstVectorType res0 =
85 560 vector_path(pg_src, src_v0, src_v1, src_v2, src_v3);
86 560 svbool_t pg_dst = DstVecTraits::svptrue();
87 560 svst1(pg_dst, &dst[0], res0);
88 560 src += ptrdiff_t(step);
89 560 dst += ptrdiff_t(step);
90 560 })
91 295 .remaining([&](size_t length, size_t) KLEIDICV_STREAMING {
92 116 size_t index = 0;
93 116 svbool_t pg = SrcVecTraits::svwhilelt(index, length);
94
2/2
✓ Branch 0 taken 225 times.
✓ Branch 1 taken 116 times.
 341 while (svptest_first(SrcVecTraits::svptrue(), pg)) {
95 225 SrcVectorType src_vector = svld1(pg, &src[ptrdiff_t(index)]);
96 225 DstVectorType result_vector = remaining_path(pg, src_vector);
97 450 svst1b(pg, &dst[ptrdiff_t(index)],
98 225 svreinterpret_u32(result_vector));
99 // Update loop counter and calculate the next governing predicate.
100 225 index += SrcVecTraits::num_lanes();
101 225 pg = SrcVecTraits::svwhilelt(index, length);
102 225 }
103 116 });
104 179 }
105
106 private:
107 // NOLINTBEGIN(readability-make-member-function-const)
108 560 DstVectorType vector_path(svbool_t full_pg, SrcVectorType fsrc0,
109 SrcVectorType fsrc1, SrcVectorType fsrc2,
110 SrcVectorType fsrc3) KLEIDICV_STREAMING {
111 1120 svbool_t pred0 = svand_z(full_pg, svcmpge(full_pg, fsrc0, lower_bound_),
112 560 svcmple(full_pg, fsrc0, upper_bound_));
113 560 auto res00 = svsel(pred0, svdup_u32(0xFF), svdup_u32(0));
114
115 1120 svbool_t pred1 = svand_z(full_pg, svcmpge(full_pg, fsrc1, lower_bound_),
116 560 svcmple(full_pg, fsrc1, upper_bound_));
117 560 auto res01 = svsel(pred1, svdup_u32(0xFF), svdup_u32(0));
118
119 1120 svbool_t pred2 = svand_z(full_pg, svcmpge(full_pg, fsrc2, lower_bound_),
120 560 svcmple(full_pg, fsrc2, upper_bound_));
121 560 auto res10 = svsel(pred2, svdup_u32(0xFF), svdup_u32(0));
122
123 1120 svbool_t pred3 = svand_z(full_pg, svcmpge(full_pg, fsrc3, lower_bound_),
124 560 svcmple(full_pg, fsrc3, upper_bound_));
125 560 auto res11 = svsel(pred3, svdup_u32(0xFF), svdup_u32(0));
126
127 1120 auto res0 =
128 560 svuzp1(svreinterpret_u16_u32(res00), svreinterpret_u16_u32(res01));
129 1120 auto res1 =
130 560 svuzp1(svreinterpret_u16_u32(res10), svreinterpret_u16_u32(res11));
131 1120 return svuzp1(svreinterpret_u8_u16(res0), svreinterpret_u8_u16(res1));
132 560 }
133 // NOLINTEND(readability-make-member-function-const)
134
135 // NOLINTBEGIN(readability-make-member-function-const)
136 225 DstVectorType remaining_path(svbool_t &pg,
137 SrcVectorType src) KLEIDICV_STREAMING {
138 450 svbool_t predicate = svand_z(pg, svcmpge(pg, src, lower_bound_),
139 225 svcmple(pg, src, upper_bound_));
140 450 return svsel(predicate, DstVecTraits::svdup(0xFF), DstVecTraits::svdup(0));
141 225 }
142 // NOLINTEND(readability-make-member-function-const)
143
144 float lower_bound_;
145 float upper_bound_;
146 }; // end of class InRange<float>
147
148 template <typename T>
149 318 kleidicv_error_t in_range_sc(const T *src, size_t src_stride, uint8_t *dst,
150 size_t dst_stride, size_t width, size_t height,
151 T lower_bound, T upper_bound) KLEIDICV_STREAMING {
152
8/8
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 156 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 156 times.
✓ Branch 4 taken 4 times.
✓ Branch 5 taken 156 times.
✓ Branch 6 taken 4 times.
✓ Branch 7 taken 156 times.
 318 CHECK_POINTER_AND_STRIDE(src, src_stride, height);
153
8/8
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 154 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 154 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 154 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 154 times.
 312 CHECK_POINTER_AND_STRIDE(dst, dst_stride, height);
154
12/12
✓ Branch 0 taken 2 times.
✓ Branch 1 taken 152 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 150 times.
✓ Branch 4 taken 4 times.
✓ Branch 5 taken 150 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 152 times.
✓ Branch 8 taken 2 times.
✓ Branch 9 taken 150 times.
✓ Branch 10 taken 4 times.
✓ Branch 11 taken 150 times.
 308 CHECK_IMAGE_SIZE(width, height);
155
156 300 Rectangle rect{width, height};
157 300 Rows<const T> src_rows{src, src_stride};
158 300 Rows<uint8_t> dst_rows{dst, dst_stride};
159
160 using VecTraits = KLEIDICV_TARGET_NAMESPACE::VecTraits<T>;
161 using VectorType = typename VecTraits::VectorType;
162
163 if constexpr (std::is_same_v<T, uint8_t>) {
164 150 VectorType vec_lower_bound = VecTraits::svdup(lower_bound);
165 150 VectorType vec_upper_bound = VecTraits::svdup(upper_bound);
166 150 InRange<T> operation{vec_lower_bound, vec_upper_bound};
167 150 apply_operation_by_rows(operation, rect, src_rows, dst_rows);
168 150 } else {
169 150 InRange<T> operation{lower_bound, upper_bound};
170 150 zip_rows(operation, rect, src_rows, dst_rows);
171 150 }
172
173 300 return KLEIDICV_OK;
174 318 }
175
176 } // namespace KLEIDICV_TARGET_NAMESPACE
177
178 #endif // KLEIDICV_IN_RANGE_SC_H
179