KleidiCV Coverage Report

Directory: ./
File: kleidicv/src/conversions/yuv420sp_to_rgb_neon.cpp
Date: 2026-01-20 20:58:59
Exec Total Coverage
Lines: 80 80 100.0%
Functions: 17 17 100.0%
Branches: 105 105 100.0%
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 <utility>
6
7 #include "kleidicv/conversions/yuv_to_rgb.h"
8 #include "kleidicv/kleidicv.h"
9 #include "kleidicv/neon.h"
10 #include "yuv420_to_rgb_neon.h"
11
12 namespace kleidicv::neon {
13 template <bool BGR, bool kAlpha>
14 class YUVSpToRGBxOrBGRx final : public YUV420XToRGBxOrBGRx<BGR, kAlpha>,
15 public UnrollOnce,
16 public TryToAvoidTailLoop {
17 public:
18 using VecTraits = neon::VecTraits<uint8_t>;
19 using ScalarType = VecTraits::ScalarType;
20 using VectorType = VecTraits::VectorType;
21 using YUV420XToRGBxOrBGRx<BGR, kAlpha>::de_interleave_indices_;
22 using YUV420XToRGBxOrBGRx<BGR, kAlpha>::yuv420x_to_rgb;
23 using YUV420XToRGBxOrBGRx<BGR, kAlpha>::v_first_;
24
25 568 explicit YUVSpToRGBxOrBGRx(bool v_first)
26 568 : YUV420XToRGBxOrBGRx<BGR, kAlpha>(v_first) {}
27
28 // Processes 2 * 16 bytes (even and odd rows) of the input YUV data, and
29 // outputs 2 * 3 (or 4) * 16 bytes of RGB (or RGBA) data per loop iteration.
30 KLEIDICV_FORCE_INLINE
31 96 void vector_path(VectorType y0, VectorType y1, VectorType uv,
32 ScalarType *rgbx_row_0, ScalarType *rgbx_row_1) {
33 // Widen U and V to 32 bits.
34 96 int32x4_t u_l = vqtbl1q_s8(uv, de_interleave_indices_.val[0]);
35 96 int32x4_t u_h = vqtbl1q_s8(uv, de_interleave_indices_.val[1]);
36
37 96 int32x4_t v_l = vqtbl1q_s8(uv, de_interleave_indices_.val[2]);
38 96 int32x4_t v_h = vqtbl1q_s8(uv, de_interleave_indices_.val[3]);
39
40 96 yuv420x_to_rgb(y0, y1, u_l, u_h, v_l, v_h, rgbx_row_0, rgbx_row_1);
41 96 }
42
43 // Processes inputs which are not long enough to fit a vector.
44 KLEIDICV_FORCE_INLINE
45 656 void scalar_path(size_t length, const ScalarType *y_row_0,
46 const ScalarType *y_row_1, const ScalarType *uv_row,
47 ScalarType *rgbx_row_0, ScalarType *rgbx_row_1) {
48 656 const uint8_t *y_rows[2] = {y_row_0, y_row_1};
49 656 uint8_t *rgbx_rows[2] = {rgbx_row_0, rgbx_row_1};
50
51 656 int32_t u_m128 = 0, v_m128 = 0;
52
8/8
✓ Branch 0 taken 164 times.
✓ Branch 1 taken 1064 times.
✓ Branch 2 taken 164 times.
✓ Branch 3 taken 1064 times.
✓ Branch 4 taken 164 times.
✓ Branch 5 taken 1064 times.
✓ Branch 6 taken 164 times.
✓ Branch 7 taken 1064 times.
 4912 for (size_t index = 0; index < length; ++index) {
53 4256 disable_loop_vectorization();
54
55 // There is one {U, V} pair for 4 Y values.
56
8/8
✓ Branch 0 taken 508 times.
✓ Branch 1 taken 556 times.
✓ Branch 2 taken 508 times.
✓ Branch 3 taken 556 times.
✓ Branch 4 taken 508 times.
✓ Branch 5 taken 556 times.
✓ Branch 6 taken 508 times.
✓ Branch 7 taken 556 times.
 4256 if ((index % 2) == 0) {
57 2224 u_m128 = uv_row[0] - 128;
58 2224 v_m128 = uv_row[1] - 128;
59 2224 uv_row += 2;
60
8/8
✓ Branch 0 taken 278 times.
✓ Branch 1 taken 278 times.
✓ Branch 2 taken 278 times.
✓ Branch 3 taken 278 times.
✓ Branch 4 taken 278 times.
✓ Branch 5 taken 278 times.
✓ Branch 6 taken 278 times.
✓ Branch 7 taken 278 times.
 2224 if (v_first_) {
61 1112 std::swap(u_m128, v_m128);
62 1112 }
63 2224 }
64
65 4256 yuv420x_to_rgb(y_rows, index, u_m128, v_m128, rgbx_rows);
66 4256 }
67 656 }
68 }; // end of class YUVSpToRGBxOrBGRx<bool, bool>
69
70 using YUVSpToRGB = YUVSpToRGBxOrBGRx<false, false>;
71 using YUVSpToRGBA = YUVSpToRGBxOrBGRx<false, true>;
72 using YUVSpToBGR = YUVSpToRGBxOrBGRx<true, false>;
73 using YUVSpToBGRA = YUVSpToRGBxOrBGRx<true, true>;
74
75 template <typename OperationType, typename ScalarType>
76 568 kleidicv_error_t yuv2rgbx_operation(
77 OperationType &operation, const ScalarType *src_y, size_t src_y_stride,
78 const ScalarType *src_uv, size_t src_uv_stride, ScalarType *dst,
79 size_t dst_stride, size_t width, size_t height) {
80
16/16
✓ Branch 0 taken 8 times.
✓ Branch 1 taken 134 times.
✓ Branch 2 taken 8 times.
✓ Branch 3 taken 134 times.
✓ Branch 4 taken 8 times.
✓ Branch 5 taken 134 times.
✓ Branch 6 taken 8 times.
✓ Branch 7 taken 134 times.
✓ Branch 8 taken 8 times.
✓ Branch 9 taken 134 times.
✓ Branch 10 taken 8 times.
✓ Branch 11 taken 134 times.
✓ Branch 12 taken 8 times.
✓ Branch 13 taken 134 times.
✓ Branch 14 taken 8 times.
✓ Branch 15 taken 134 times.
 568 CHECK_POINTER_AND_STRIDE(src_y, src_y_stride, height);
81
16/16
✓ Branch 0 taken 8 times.
✓ Branch 1 taken 126 times.
✓ Branch 2 taken 8 times.
✓ Branch 3 taken 126 times.
✓ Branch 4 taken 8 times.
✓ Branch 5 taken 126 times.
✓ Branch 6 taken 8 times.
✓ Branch 7 taken 126 times.
✓ Branch 8 taken 8 times.
✓ Branch 9 taken 126 times.
✓ Branch 10 taken 8 times.
✓ Branch 11 taken 126 times.
✓ Branch 12 taken 8 times.
✓ Branch 13 taken 126 times.
✓ Branch 14 taken 8 times.
✓ Branch 15 taken 126 times.
 536 CHECK_POINTER_AND_STRIDE(src_uv, src_uv_stride, (height + 1) / 2);
82
16/16
✓ Branch 0 taken 8 times.
✓ Branch 1 taken 118 times.
✓ Branch 2 taken 8 times.
✓ Branch 3 taken 118 times.
✓ Branch 4 taken 8 times.
✓ Branch 5 taken 118 times.
✓ Branch 6 taken 8 times.
✓ Branch 7 taken 118 times.
✓ Branch 8 taken 8 times.
✓ Branch 9 taken 118 times.
✓ Branch 10 taken 8 times.
✓ Branch 11 taken 118 times.
✓ Branch 12 taken 8 times.
✓ Branch 13 taken 118 times.
✓ Branch 14 taken 8 times.
✓ Branch 15 taken 118 times.
 504 CHECK_POINTER_AND_STRIDE(dst, dst_stride, height);
83
24/24
✓ Branch 0 taken 10 times.
✓ Branch 1 taken 108 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 98 times.
✓ Branch 4 taken 20 times.
✓ Branch 5 taken 98 times.
✓ Branch 6 taken 10 times.
✓ Branch 7 taken 108 times.
✓ Branch 8 taken 10 times.
✓ Branch 9 taken 98 times.
✓ Branch 10 taken 20 times.
✓ Branch 11 taken 98 times.
✓ Branch 12 taken 10 times.
✓ Branch 13 taken 108 times.
✓ Branch 14 taken 10 times.
✓ Branch 15 taken 98 times.
✓ Branch 16 taken 20 times.
✓ Branch 17 taken 98 times.
✓ Branch 18 taken 10 times.
✓ Branch 19 taken 108 times.
✓ Branch 20 taken 10 times.
✓ Branch 21 taken 98 times.
✓ Branch 22 taken 20 times.
✓ Branch 23 taken 98 times.
 472 CHECK_IMAGE_SIZE(width, height);
84
85 392 Rectangle rect{width, height};
86 392 ParallelRows y_rows{src_y, src_y_stride};
87 392 Rows uv_rows{src_uv, src_uv_stride};
88 392 ParallelRows rgbx_rows{dst, dst_stride, operation.output_channels()};
89
90 392 RemoveContextAdapter remove_context_adapter{operation};
91 392 OperationAdapter operation_adapter{remove_context_adapter};
92 392 RemainingPathToScalarPathAdapter remaining_path_adapter{operation_adapter};
93 392 OperationContextAdapter context_adapter{remaining_path_adapter};
94 392 ParallelRowsAdapter parallel_rows_adapter{context_adapter};
95 392 RowBasedOperation row_based_operation{parallel_rows_adapter};
96 392 zip_parallel_rows(row_based_operation, rect, y_rows, uv_rows, rgbx_rows);
97 392 return KLEIDICV_OK;
98 568 }
99
100 KLEIDICV_TARGET_FN_ATTRS
101 664 kleidicv_error_t yuv420sp_to_rgb_u8(const uint8_t *src_y, size_t src_y_stride,
102 const uint8_t *src_uv, size_t src_uv_stride,
103 uint8_t *dst, size_t dst_stride,
104 size_t width, size_t height,
105 kleidicv_color_conversion_t color_format) {
106
9/9
✓ Branch 0 taken 71 times.
✓ Branch 1 taken 71 times.
✓ Branch 2 taken 71 times.
✓ Branch 3 taken 71 times.
✓ Branch 4 taken 71 times.
✓ Branch 5 taken 71 times.
✓ Branch 6 taken 71 times.
✓ Branch 7 taken 96 times.
✓ Branch 8 taken 71 times.
 664 switch (color_format) {
107 case KLEIDICV_NV21_TO_BGR: {
108 71 YUVSpToBGR operation{true};
109 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
110 71 src_uv_stride, dst, dst_stride, width, height);
111 71 }
112
113 case KLEIDICV_NV21_TO_RGB: {
114 71 YUVSpToRGB operation{true};
115 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
116 71 src_uv_stride, dst, dst_stride, width, height);
117 71 }
118
119 case KLEIDICV_NV21_TO_BGRA: {
120 71 YUVSpToBGRA operation{true};
121 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
122 71 src_uv_stride, dst, dst_stride, width, height);
123 71 }
124
125 case KLEIDICV_NV21_TO_RGBA: {
126 71 YUVSpToRGBA operation{true};
127 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
128 71 src_uv_stride, dst, dst_stride, width, height);
129 71 }
130
131 case KLEIDICV_NV12_TO_BGR: {
132 71 YUVSpToBGR operation{false};
133 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
134 71 src_uv_stride, dst, dst_stride, width, height);
135 71 }
136
137 case KLEIDICV_NV12_TO_RGB: {
138 71 YUVSpToRGB operation{false};
139 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
140 71 src_uv_stride, dst, dst_stride, width, height);
141 71 }
142
143 case KLEIDICV_NV12_TO_BGRA: {
144 71 YUVSpToBGRA operation{false};
145 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
146 71 src_uv_stride, dst, dst_stride, width, height);
147 71 }
148
149 case KLEIDICV_NV12_TO_RGBA: {
150 71 YUVSpToRGBA operation{false};
151 142 return yuv2rgbx_operation(operation, src_y, src_y_stride, src_uv,
152 71 src_uv_stride, dst, dst_stride, width, height);
153 71 }
154
155 default:
156 96 return KLEIDICV_ERROR_NOT_IMPLEMENTED;
157 }
158
159 return KLEIDICV_ERROR_NOT_IMPLEMENTED;
160 664 }
161
162 } // namespace kleidicv::neon
163