KleidiCV Coverage Report

Directory:	./
File:	kleidicv/include/kleidicv/filters/separable_filter_15x15_sc.h
Date:	2025-11-25 17:23:32

	Exec	Total	Coverage
Lines:	183	183	100.0%
Functions:	26	26	100.0%
Branches:	2	2	100.0%

  
      Line
      Branch
      Exec
      Source
    
      // SPDX-FileCopyrightText: 2024 - 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
    
      //
    
      // SPDX-License-Identifier: Apache-2.0
    
      #ifndef KLEIDICV_SEPARABLE_FILTER_15X15_SC_H
    
      #define KLEIDICV_SEPARABLE_FILTER_15X15_SC_H
    
      #include "kleidicv/sve2.h"
    
      #include "kleidicv/workspace/border_15x15.h"
    
      // It is used by SVE2 and SME, the actual namespace will reflect it.
    
      namespace KLEIDICV_TARGET_NAMESPACE {
    
      // Template for drivers of separable NxM filters.
    
      template <typename FilterType, const size_t S>
    
      class SeparableFilter;
    
      // Driver for a separable 15x15 filter.
    
      template <typename FilterType>
    
      class SeparableFilter<FilterType, 15UL> {
    
       public:
    
        using SourceType = typename FilterType::SourceType;
    
        using BufferType = typename FilterType::BufferType;
    
        using DestinationType = typename FilterType::DestinationType;
    
        using SourceVecTraits =
    
            typename ::KLEIDICV_TARGET_NAMESPACE::VecTraits<SourceType>;
    
        using SourceVectorType = typename SourceVecTraits::VectorType;
    
        using BufferVecTraits =
    
            typename ::KLEIDICV_TARGET_NAMESPACE::VecTraits<BufferType>;
    
        using BufferVectorType = typename BufferVecTraits::VectorType;
    
        using BorderInfoType =
    
            typename ::KLEIDICV_TARGET_NAMESPACE::FixedBorderInfo15x15<SourceType>;
    
        using BorderType = FixedBorderType;
    
        using BorderOffsets = typename BorderInfoType::Offsets;
    
      96
        explicit SeparableFilter(FilterType filter) KLEIDICV_STREAMING
    
      96
            : filter_{filter} {}
    
        static constexpr size_t margin = 7UL;
    
      1368
        void process_vertical(size_t width, Rows<const SourceType> src_rows,
    
                              Rows<BufferType> dst_rows,
    
                              BorderOffsets border_offsets) const KLEIDICV_STREAMING {
    
      1368
          LoopUnroll2 loop{width * src_rows.channels(), SourceVecTraits::num_lanes()};
    
      2712
          loop.unroll_once([&](size_t index) KLEIDICV_STREAMING {
    
      1344
            svbool_t pg_all = SourceVecTraits::svptrue();
    
      1344
            vertical_vector_path(pg_all, src_rows, dst_rows, border_offsets, index);
    
      1344
          });
    
      2736
          loop.remaining([&](size_t index, size_t length) KLEIDICV_STREAMING {
    
      1368
            svbool_t pg = SourceVecTraits::svwhilelt(index, length);
    
      1368
            vertical_vector_path(pg, src_rows, dst_rows, border_offsets, index);
    
      1368
          });
    
      1368
        }
    
      1368
        void process_horizontal(size_t width, Rows<const BufferType> src_rows,
    
                                Rows<DestinationType> dst_rows,
    
                                BorderOffsets border_offsets) const
    
            KLEIDICV_STREAMING {
    
      1368
          svbool_t pg_all = BufferVecTraits::svptrue();
    
      1368
          LoopUnroll2 loop{width * src_rows.channels(), BufferVecTraits::num_lanes()};
    
      1704
          loop.unroll_twice([&](size_t index) KLEIDICV_STREAMING {
    
      672
            horizontal_vector_path_2x(pg_all, src_rows, dst_rows, border_offsets,
    
      336
                                      index);
    
      336
          });
    
      1704
          loop.unroll_once([&](size_t index) KLEIDICV_STREAMING {
    
      336
            horizontal_vector_path(pg_all, src_rows, dst_rows, border_offsets, index);
    
      336
          });
    
      2400
          loop.remaining([&](size_t index, size_t length) KLEIDICV_STREAMING {
    
      1032
            svbool_t pg = BufferVecTraits::svwhilelt(index, length);
    
      1032
            horizontal_vector_path(pg, src_rows, dst_rows, border_offsets, index);
    
      1032
          });
    
      1368
        }
    
        // Processing of horizontal borders is always scalar because border offsets
    
        // change for each and every element in the border.
    
      19152
        void process_horizontal_borders(
    
            Rows<const BufferType> src_rows, Rows<DestinationType> dst_rows,
    
            BorderOffsets border_offsets) const KLEIDICV_STREAMING {
    
        2/2✓ Branch 0 taken 19152 times.
✓ Branch 1 taken 33264 times.

      52416
          for (size_t index = 0; index < src_rows.channels(); ++index) {
    
      33264
            disable_loop_vectorization();
    
      33264
            process_horizontal_border(src_rows, dst_rows, border_offsets, index);
    
      33264
          }
    
      19152
        }
    
       private:
    
      2712
        void vertical_vector_path(svbool_t pg, Rows<const SourceType> src_rows,
    
                                  Rows<BufferType> dst_rows,
    
                                  BorderOffsets border_offsets,
    
                                  size_t index) const KLEIDICV_STREAMING {
    
      5424
          SourceVectorType src_0 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c0())[index]);
    
      5424
          SourceVectorType src_1 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c1())[index]);
    
      5424
          SourceVectorType src_2 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c2())[index]);
    
      5424
          SourceVectorType src_3 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c3())[index]);
    
      5424
          SourceVectorType src_4 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c4())[index]);
    
      5424
          SourceVectorType src_5 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c5())[index]);
    
      5424
          SourceVectorType src_6 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c6())[index]);
    
      5424
          SourceVectorType src_7 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c7())[index]);
    
      5424
          SourceVectorType src_8 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c8())[index]);
    
      5424
          SourceVectorType src_9 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c9())[index]);
    
      5424
          SourceVectorType src_10 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c10())[index]);
    
      5424
          SourceVectorType src_11 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c11())[index]);
    
      5424
          SourceVectorType src_12 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c12())[index]);
    
      5424
          SourceVectorType src_13 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c13())[index]);
    
      5424
          SourceVectorType src_14 =
    
      2712
              svld1(pg, &src_rows.at(border_offsets.c14())[index]);
    
      40680
          std::reference_wrapper<SourceVectorType> sources[15] = {
    
      21696
              src_0, src_1, src_2,  src_3,  src_4,  src_5,  src_6, src_7,
    
      18984
              src_8, src_9, src_10, src_11, src_12, src_13, src_14};
    
      2712
          filter_.vertical_vector_path(pg, sources, &dst_rows[index]);
    
      2712
        }
    
      336
        void horizontal_vector_path_2x(svbool_t pg, Rows<const BufferType> src_rows,
    
                                       Rows<DestinationType> dst_rows,
    
                                       BorderOffsets border_offsets,
    
                                       size_t index) const KLEIDICV_STREAMING {
    
      336
          auto src_0 = &src_rows.at(0, border_offsets.c0())[index];
    
      336
          auto src_1 = &src_rows.at(0, border_offsets.c1())[index];
    
      336
          auto src_2 = &src_rows.at(0, border_offsets.c2())[index];
    
      336
          auto src_3 = &src_rows.at(0, border_offsets.c3())[index];
    
      336
          auto src_4 = &src_rows.at(0, border_offsets.c4())[index];
    
      336
          auto src_5 = &src_rows.at(0, border_offsets.c5())[index];
    
      336
          auto src_6 = &src_rows.at(0, border_offsets.c6())[index];
    
      336
          auto src_7 = &src_rows.at(0, border_offsets.c7())[index];
    
      336
          auto src_8 = &src_rows.at(0, border_offsets.c8())[index];
    
      336
          auto src_9 = &src_rows.at(0, border_offsets.c9())[index];
    
      336
          auto src_10 = &src_rows.at(0, border_offsets.c10())[index];
    
      336
          auto src_11 = &src_rows.at(0, border_offsets.c11())[index];
    
      336
          auto src_12 = &src_rows.at(0, border_offsets.c12())[index];
    
      336
          auto src_13 = &src_rows.at(0, border_offsets.c13())[index];
    
      336
          auto src_14 = &src_rows.at(0, border_offsets.c14())[index];
    
      336
          BufferVectorType src_0_0 = svld1(pg, &src_0[0]);
    
      336
          BufferVectorType src_1_0 = svld1_vnum(pg, &src_0[0], 1);
    
      336
          BufferVectorType src_0_1 = svld1(pg, &src_1[0]);
    
      336
          BufferVectorType src_1_1 = svld1_vnum(pg, &src_1[0], 1);
    
      336
          BufferVectorType src_0_2 = svld1(pg, &src_2[0]);
    
      336
          BufferVectorType src_1_2 = svld1_vnum(pg, &src_2[0], 1);
    
      336
          BufferVectorType src_0_3 = svld1(pg, &src_3[0]);
    
      336
          BufferVectorType src_1_3 = svld1_vnum(pg, &src_3[0], 1);
    
      336
          BufferVectorType src_0_4 = svld1(pg, &src_4[0]);
    
      336
          BufferVectorType src_1_4 = svld1_vnum(pg, &src_4[0], 1);
    
      336
          BufferVectorType src_0_5 = svld1(pg, &src_5[0]);
    
      336
          BufferVectorType src_1_5 = svld1_vnum(pg, &src_5[0], 1);
    
      336
          BufferVectorType src_0_6 = svld1(pg, &src_6[0]);
    
      336
          BufferVectorType src_1_6 = svld1_vnum(pg, &src_6[0], 1);
    
      336
          BufferVectorType src_0_7 = svld1(pg, &src_7[0]);
    
      336
          BufferVectorType src_1_7 = svld1_vnum(pg, &src_7[0], 1);
    
      336
          BufferVectorType src_0_8 = svld1(pg, &src_8[0]);
    
      336
          BufferVectorType src_1_8 = svld1_vnum(pg, &src_8[0], 1);
    
      336
          BufferVectorType src_0_9 = svld1(pg, &src_9[0]);
    
      336
          BufferVectorType src_1_9 = svld1_vnum(pg, &src_9[0], 1);
    
      336
          BufferVectorType src_0_10 = svld1(pg, &src_10[0]);
    
      336
          BufferVectorType src_1_10 = svld1_vnum(pg, &src_10[0], 1);
    
      336
          BufferVectorType src_0_11 = svld1(pg, &src_11[0]);
    
      336
          BufferVectorType src_1_11 = svld1_vnum(pg, &src_11[0], 1);
    
      336
          BufferVectorType src_0_12 = svld1(pg, &src_12[0]);
    
      336
          BufferVectorType src_1_12 = svld1_vnum(pg, &src_12[0], 1);
    
      336
          BufferVectorType src_0_13 = svld1(pg, &src_13[0]);
    
      336
          BufferVectorType src_1_13 = svld1_vnum(pg, &src_13[0], 1);
    
      336
          BufferVectorType src_0_14 = svld1(pg, &src_14[0]);
    
      336
          BufferVectorType src_1_14 = svld1_vnum(pg, &src_14[0], 1);
    
      5040
          std::reference_wrapper<BufferVectorType> sources_0[15] = {
    
      1680
              src_0_0,  src_0_1,  src_0_2,  src_0_3,  src_0_4,
    
      1680
              src_0_5,  src_0_6,  src_0_7,  src_0_8,  src_0_9,
    
      1680
              src_0_10, src_0_11, src_0_12, src_0_13, src_0_14};
    
      336
          filter_.horizontal_vector_path(pg, sources_0, &dst_rows[index]);
    
      5040
          std::reference_wrapper<BufferVectorType> sources_1[15] = {
    
      1680
              src_1_0,  src_1_1,  src_1_2,  src_1_3,  src_1_4,
    
      1680
              src_1_5,  src_1_6,  src_1_7,  src_1_8,  src_1_9,
    
      1680
              src_1_10, src_1_11, src_1_12, src_1_13, src_1_14};
    
      672
          filter_.horizontal_vector_path(
    
      336
              pg, sources_1, &dst_rows[index + BufferVecTraits::num_lanes()]);
    
      336
        }
    
      1368
        void horizontal_vector_path(svbool_t pg, Rows<const BufferType> src_rows,
    
                                    Rows<DestinationType> dst_rows,
    
                                    BorderOffsets border_offsets,
    
                                    size_t index) const KLEIDICV_STREAMING {
    
      2736
          BufferVectorType src_0 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c0())[index]);
    
      2736
          BufferVectorType src_1 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c1())[index]);
    
      2736
          BufferVectorType src_2 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c2())[index]);
    
      2736
          BufferVectorType src_3 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c3())[index]);
    
      2736
          BufferVectorType src_4 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c4())[index]);
    
      2736
          BufferVectorType src_5 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c5())[index]);
    
      2736
          BufferVectorType src_6 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c6())[index]);
    
      2736
          BufferVectorType src_7 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c7())[index]);
    
      2736
          BufferVectorType src_8 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c8())[index]);
    
      2736
          BufferVectorType src_9 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c9())[index]);
    
      2736
          BufferVectorType src_10 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c10())[index]);
    
      2736
          BufferVectorType src_11 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c11())[index]);
    
      2736
          BufferVectorType src_12 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c12())[index]);
    
      2736
          BufferVectorType src_13 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c13())[index]);
    
      2736
          BufferVectorType src_14 =
    
      1368
              svld1(pg, &src_rows.at(0, border_offsets.c14())[index]);
    
      20520
          std::reference_wrapper<BufferVectorType> sources[15] = {
    
      10944
              src_0, src_1, src_2,  src_3,  src_4,  src_5,  src_6, src_7,
    
      9576
              src_8, src_9, src_10, src_11, src_12, src_13, src_14};
    
      1368
          filter_.horizontal_vector_path(pg, sources, &dst_rows[index]);
    
      1368
        }
    
      33264
        void process_horizontal_border(Rows<const BufferType> src_rows,
    
                                       Rows<DestinationType> dst_rows,
    
                                       BorderOffsets border_offsets,
    
                                       size_t index) const KLEIDICV_STREAMING {
    
      33264
          BufferType src[15];
    
      33264
          src[0] = src_rows.at(0, border_offsets.c0())[index];
    
      33264
          src[1] = src_rows.at(0, border_offsets.c1())[index];
    
      33264
          src[2] = src_rows.at(0, border_offsets.c2())[index];
    
      33264
          src[3] = src_rows.at(0, border_offsets.c3())[index];
    
      33264
          src[4] = src_rows.at(0, border_offsets.c4())[index];
    
      33264
          src[5] = src_rows.at(0, border_offsets.c5())[index];
    
      33264
          src[6] = src_rows.at(0, border_offsets.c6())[index];
    
      33264
          src[7] = src_rows.at(0, border_offsets.c7())[index];
    
      33264
          src[8] = src_rows.at(0, border_offsets.c8())[index];
    
      33264
          src[9] = src_rows.at(0, border_offsets.c9())[index];
    
      33264
          src[10] = src_rows.at(0, border_offsets.c10())[index];
    
      33264
          src[11] = src_rows.at(0, border_offsets.c11())[index];
    
      33264
          src[12] = src_rows.at(0, border_offsets.c12())[index];
    
      33264
          src[13] = src_rows.at(0, border_offsets.c13())[index];
    
      33264
          src[14] = src_rows.at(0, border_offsets.c14())[index];
    
      33264
          filter_.horizontal_scalar_path(src, &dst_rows[index]);
    
      33264
        }
    
        FilterType filter_;
    
      };  // end of class SeparableFilter<FilterType, 15UL>
    
      // Shorthand for 15x15 separable filters driver type.
    
      template <class FilterType>
    
      using SeparableFilter15x15 = SeparableFilter<FilterType, 15UL>;
    
      }  // namespace KLEIDICV_TARGET_NAMESPACE
    
      #endif  // KLEIDICV_SEPARABLE_FILTER_15X15_SC_H

Line	Branch	Exec	Source
1			// SPDX-FileCopyrightText: 2024 - 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
2			//
3			// SPDX-License-Identifier: Apache-2.0
4
5			#ifndef KLEIDICV_SEPARABLE_FILTER_15X15_SC_H
6			#define KLEIDICV_SEPARABLE_FILTER_15X15_SC_H
7
8			#include "kleidicv/sve2.h"
9			#include "kleidicv/workspace/border_15x15.h"
10
11			// It is used by SVE2 and SME, the actual namespace will reflect it.
12			namespace KLEIDICV_TARGET_NAMESPACE {
13
14			// Template for drivers of separable NxM filters.
15			template <typename FilterType, const size_t S>
16			class SeparableFilter;
17
18			// Driver for a separable 15x15 filter.
19			template <typename FilterType>
20			class SeparableFilter<FilterType, 15UL> {
21			public:
22			using SourceType = typename FilterType::SourceType;
23			using BufferType = typename FilterType::BufferType;
24			using DestinationType = typename FilterType::DestinationType;
25			using SourceVecTraits =
26			typename ::KLEIDICV_TARGET_NAMESPACE::VecTraits<SourceType>;
27			using SourceVectorType = typename SourceVecTraits::VectorType;
28			using BufferVecTraits =
29			typename ::KLEIDICV_TARGET_NAMESPACE::VecTraits<BufferType>;
30			using BufferVectorType = typename BufferVecTraits::VectorType;
31			using BorderInfoType =
32			typename ::KLEIDICV_TARGET_NAMESPACE::FixedBorderInfo15x15<SourceType>;
33			using BorderType = FixedBorderType;
34			using BorderOffsets = typename BorderInfoType::Offsets;
35
36		96	explicit SeparableFilter(FilterType filter) KLEIDICV_STREAMING
37		96	: filter_{filter} {}
38
39			static constexpr size_t margin = 7UL;
40
41		1368	void process_vertical(size_t width, Rows<const SourceType> src_rows,
42			Rows<BufferType> dst_rows,
43			BorderOffsets border_offsets) const KLEIDICV_STREAMING {
44		1368	LoopUnroll2 loop{width * src_rows.channels(), SourceVecTraits::num_lanes()};
45
46		2712	loop.unroll_once([&](size_t index) KLEIDICV_STREAMING {
47		1344	svbool_t pg_all = SourceVecTraits::svptrue();
48		1344	vertical_vector_path(pg_all, src_rows, dst_rows, border_offsets, index);
49		1344	});
50
51		2736	loop.remaining([&](size_t index, size_t length) KLEIDICV_STREAMING {
52		1368	svbool_t pg = SourceVecTraits::svwhilelt(index, length);
53		1368	vertical_vector_path(pg, src_rows, dst_rows, border_offsets, index);
54		1368	});
55		1368	}
56
57		1368	void process_horizontal(size_t width, Rows<const BufferType> src_rows,
58			Rows<DestinationType> dst_rows,
59			BorderOffsets border_offsets) const
60			KLEIDICV_STREAMING {
61		1368	svbool_t pg_all = BufferVecTraits::svptrue();
62		1368	LoopUnroll2 loop{width * src_rows.channels(), BufferVecTraits::num_lanes()};
63
64		1704	loop.unroll_twice([&](size_t index) KLEIDICV_STREAMING {
65		672	horizontal_vector_path_2x(pg_all, src_rows, dst_rows, border_offsets,
66		336	index);
67		336	});
68
69		1704	loop.unroll_once([&](size_t index) KLEIDICV_STREAMING {
70		336	horizontal_vector_path(pg_all, src_rows, dst_rows, border_offsets, index);
71		336	});
72
73		2400	loop.remaining([&](size_t index, size_t length) KLEIDICV_STREAMING {
74		1032	svbool_t pg = BufferVecTraits::svwhilelt(index, length);
75		1032	horizontal_vector_path(pg, src_rows, dst_rows, border_offsets, index);
76		1032	});
77		1368	}
78
79			// Processing of horizontal borders is always scalar because border offsets
80			// change for each and every element in the border.
81		19152	void process_horizontal_borders(
82			Rows<const BufferType> src_rows, Rows<DestinationType> dst_rows,
83			BorderOffsets border_offsets) const KLEIDICV_STREAMING {
84	2/2 ✓ Branch 0 taken 19152 times. ✓ Branch 1 taken 33264 times.	52416	for (size_t index = 0; index < src_rows.channels(); ++index) {
85		33264	disable_loop_vectorization();
86		33264	process_horizontal_border(src_rows, dst_rows, border_offsets, index);
87		33264	}
88		19152	}
89
90			private:
91		2712	void vertical_vector_path(svbool_t pg, Rows<const SourceType> src_rows,
92			Rows<BufferType> dst_rows,
93			BorderOffsets border_offsets,
94			size_t index) const KLEIDICV_STREAMING {
95		5424	SourceVectorType src_0 =
96		2712	svld1(pg, &src_rows.at(border_offsets.c0())[index]);
97		5424	SourceVectorType src_1 =
98		2712	svld1(pg, &src_rows.at(border_offsets.c1())[index]);
99		5424	SourceVectorType src_2 =
100		2712	svld1(pg, &src_rows.at(border_offsets.c2())[index]);
101		5424	SourceVectorType src_3 =
102		2712	svld1(pg, &src_rows.at(border_offsets.c3())[index]);
103		5424	SourceVectorType src_4 =
104		2712	svld1(pg, &src_rows.at(border_offsets.c4())[index]);
105		5424	SourceVectorType src_5 =
106		2712	svld1(pg, &src_rows.at(border_offsets.c5())[index]);
107		5424	SourceVectorType src_6 =
108		2712	svld1(pg, &src_rows.at(border_offsets.c6())[index]);
109		5424	SourceVectorType src_7 =
110		2712	svld1(pg, &src_rows.at(border_offsets.c7())[index]);
111		5424	SourceVectorType src_8 =
112		2712	svld1(pg, &src_rows.at(border_offsets.c8())[index]);
113		5424	SourceVectorType src_9 =
114		2712	svld1(pg, &src_rows.at(border_offsets.c9())[index]);
115		5424	SourceVectorType src_10 =
116		2712	svld1(pg, &src_rows.at(border_offsets.c10())[index]);
117		5424	SourceVectorType src_11 =
118		2712	svld1(pg, &src_rows.at(border_offsets.c11())[index]);
119		5424	SourceVectorType src_12 =
120		2712	svld1(pg, &src_rows.at(border_offsets.c12())[index]);
121		5424	SourceVectorType src_13 =
122		2712	svld1(pg, &src_rows.at(border_offsets.c13())[index]);
123		5424	SourceVectorType src_14 =
124		2712	svld1(pg, &src_rows.at(border_offsets.c14())[index]);
125
126		40680	std::reference_wrapper<SourceVectorType> sources[15] = {
127		21696	src_0, src_1, src_2, src_3, src_4, src_5, src_6, src_7,
128		18984	src_8, src_9, src_10, src_11, src_12, src_13, src_14};
129		2712	filter_.vertical_vector_path(pg, sources, &dst_rows[index]);
130		2712	}
131
132		336	void horizontal_vector_path_2x(svbool_t pg, Rows<const BufferType> src_rows,
133			Rows<DestinationType> dst_rows,
134			BorderOffsets border_offsets,
135			size_t index) const KLEIDICV_STREAMING {
136		336	auto src_0 = &src_rows.at(0, border_offsets.c0())[index];
137		336	auto src_1 = &src_rows.at(0, border_offsets.c1())[index];
138		336	auto src_2 = &src_rows.at(0, border_offsets.c2())[index];
139		336	auto src_3 = &src_rows.at(0, border_offsets.c3())[index];
140		336	auto src_4 = &src_rows.at(0, border_offsets.c4())[index];
141		336	auto src_5 = &src_rows.at(0, border_offsets.c5())[index];
142		336	auto src_6 = &src_rows.at(0, border_offsets.c6())[index];
143		336	auto src_7 = &src_rows.at(0, border_offsets.c7())[index];
144		336	auto src_8 = &src_rows.at(0, border_offsets.c8())[index];
145		336	auto src_9 = &src_rows.at(0, border_offsets.c9())[index];
146		336	auto src_10 = &src_rows.at(0, border_offsets.c10())[index];
147		336	auto src_11 = &src_rows.at(0, border_offsets.c11())[index];
148		336	auto src_12 = &src_rows.at(0, border_offsets.c12())[index];
149		336	auto src_13 = &src_rows.at(0, border_offsets.c13())[index];
150		336	auto src_14 = &src_rows.at(0, border_offsets.c14())[index];
151
152		336	BufferVectorType src_0_0 = svld1(pg, &src_0[0]);
153		336	BufferVectorType src_1_0 = svld1_vnum(pg, &src_0[0], 1);
154		336	BufferVectorType src_0_1 = svld1(pg, &src_1[0]);
155		336	BufferVectorType src_1_1 = svld1_vnum(pg, &src_1[0], 1);
156		336	BufferVectorType src_0_2 = svld1(pg, &src_2[0]);
157		336	BufferVectorType src_1_2 = svld1_vnum(pg, &src_2[0], 1);
158		336	BufferVectorType src_0_3 = svld1(pg, &src_3[0]);
159		336	BufferVectorType src_1_3 = svld1_vnum(pg, &src_3[0], 1);
160		336	BufferVectorType src_0_4 = svld1(pg, &src_4[0]);
161		336	BufferVectorType src_1_4 = svld1_vnum(pg, &src_4[0], 1);
162		336	BufferVectorType src_0_5 = svld1(pg, &src_5[0]);
163		336	BufferVectorType src_1_5 = svld1_vnum(pg, &src_5[0], 1);
164		336	BufferVectorType src_0_6 = svld1(pg, &src_6[0]);
165		336	BufferVectorType src_1_6 = svld1_vnum(pg, &src_6[0], 1);
166		336	BufferVectorType src_0_7 = svld1(pg, &src_7[0]);
167		336	BufferVectorType src_1_7 = svld1_vnum(pg, &src_7[0], 1);
168		336	BufferVectorType src_0_8 = svld1(pg, &src_8[0]);
169		336	BufferVectorType src_1_8 = svld1_vnum(pg, &src_8[0], 1);
170		336	BufferVectorType src_0_9 = svld1(pg, &src_9[0]);
171		336	BufferVectorType src_1_9 = svld1_vnum(pg, &src_9[0], 1);
172		336	BufferVectorType src_0_10 = svld1(pg, &src_10[0]);
173		336	BufferVectorType src_1_10 = svld1_vnum(pg, &src_10[0], 1);
174		336	BufferVectorType src_0_11 = svld1(pg, &src_11[0]);
175		336	BufferVectorType src_1_11 = svld1_vnum(pg, &src_11[0], 1);
176		336	BufferVectorType src_0_12 = svld1(pg, &src_12[0]);
177		336	BufferVectorType src_1_12 = svld1_vnum(pg, &src_12[0], 1);
178		336	BufferVectorType src_0_13 = svld1(pg, &src_13[0]);
179		336	BufferVectorType src_1_13 = svld1_vnum(pg, &src_13[0], 1);
180		336	BufferVectorType src_0_14 = svld1(pg, &src_14[0]);
181		336	BufferVectorType src_1_14 = svld1_vnum(pg, &src_14[0], 1);
182
183		5040	std::reference_wrapper<BufferVectorType> sources_0[15] = {
184		1680	src_0_0, src_0_1, src_0_2, src_0_3, src_0_4,
185		1680	src_0_5, src_0_6, src_0_7, src_0_8, src_0_9,
186		1680	src_0_10, src_0_11, src_0_12, src_0_13, src_0_14};
187		336	filter_.horizontal_vector_path(pg, sources_0, &dst_rows[index]);
188		5040	std::reference_wrapper<BufferVectorType> sources_1[15] = {
189		1680	src_1_0, src_1_1, src_1_2, src_1_3, src_1_4,
190		1680	src_1_5, src_1_6, src_1_7, src_1_8, src_1_9,
191		1680	src_1_10, src_1_11, src_1_12, src_1_13, src_1_14};
192		672	filter_.horizontal_vector_path(
193		336	pg, sources_1, &dst_rows[index + BufferVecTraits::num_lanes()]);
194		336	}
195
196		1368	void horizontal_vector_path(svbool_t pg, Rows<const BufferType> src_rows,
197			Rows<DestinationType> dst_rows,
198			BorderOffsets border_offsets,
199			size_t index) const KLEIDICV_STREAMING {
200		2736	BufferVectorType src_0 =
201		1368	svld1(pg, &src_rows.at(0, border_offsets.c0())[index]);
202		2736	BufferVectorType src_1 =
203		1368	svld1(pg, &src_rows.at(0, border_offsets.c1())[index]);
204		2736	BufferVectorType src_2 =
205		1368	svld1(pg, &src_rows.at(0, border_offsets.c2())[index]);
206		2736	BufferVectorType src_3 =
207		1368	svld1(pg, &src_rows.at(0, border_offsets.c3())[index]);
208		2736	BufferVectorType src_4 =
209		1368	svld1(pg, &src_rows.at(0, border_offsets.c4())[index]);
210		2736	BufferVectorType src_5 =
211		1368	svld1(pg, &src_rows.at(0, border_offsets.c5())[index]);
212		2736	BufferVectorType src_6 =
213		1368	svld1(pg, &src_rows.at(0, border_offsets.c6())[index]);
214		2736	BufferVectorType src_7 =
215		1368	svld1(pg, &src_rows.at(0, border_offsets.c7())[index]);
216		2736	BufferVectorType src_8 =
217		1368	svld1(pg, &src_rows.at(0, border_offsets.c8())[index]);
218		2736	BufferVectorType src_9 =
219		1368	svld1(pg, &src_rows.at(0, border_offsets.c9())[index]);
220		2736	BufferVectorType src_10 =
221		1368	svld1(pg, &src_rows.at(0, border_offsets.c10())[index]);
222		2736	BufferVectorType src_11 =
223		1368	svld1(pg, &src_rows.at(0, border_offsets.c11())[index]);
224		2736	BufferVectorType src_12 =
225		1368	svld1(pg, &src_rows.at(0, border_offsets.c12())[index]);
226		2736	BufferVectorType src_13 =
227		1368	svld1(pg, &src_rows.at(0, border_offsets.c13())[index]);
228		2736	BufferVectorType src_14 =
229		1368	svld1(pg, &src_rows.at(0, border_offsets.c14())[index]);
230		20520	std::reference_wrapper<BufferVectorType> sources[15] = {
231		10944	src_0, src_1, src_2, src_3, src_4, src_5, src_6, src_7,
232		9576	src_8, src_9, src_10, src_11, src_12, src_13, src_14};
233		1368	filter_.horizontal_vector_path(pg, sources, &dst_rows[index]);
234		1368	}
235
236		33264	void process_horizontal_border(Rows<const BufferType> src_rows,
237			Rows<DestinationType> dst_rows,
238			BorderOffsets border_offsets,
239			size_t index) const KLEIDICV_STREAMING {
240		33264	BufferType src[15];
241		33264	src[0] = src_rows.at(0, border_offsets.c0())[index];
242		33264	src[1] = src_rows.at(0, border_offsets.c1())[index];
243		33264	src[2] = src_rows.at(0, border_offsets.c2())[index];
244		33264	src[3] = src_rows.at(0, border_offsets.c3())[index];
245		33264	src[4] = src_rows.at(0, border_offsets.c4())[index];
246		33264	src[5] = src_rows.at(0, border_offsets.c5())[index];
247		33264	src[6] = src_rows.at(0, border_offsets.c6())[index];
248		33264	src[7] = src_rows.at(0, border_offsets.c7())[index];
249		33264	src[8] = src_rows.at(0, border_offsets.c8())[index];
250		33264	src[9] = src_rows.at(0, border_offsets.c9())[index];
251		33264	src[10] = src_rows.at(0, border_offsets.c10())[index];
252		33264	src[11] = src_rows.at(0, border_offsets.c11())[index];
253		33264	src[12] = src_rows.at(0, border_offsets.c12())[index];
254		33264	src[13] = src_rows.at(0, border_offsets.c13())[index];
255		33264	src[14] = src_rows.at(0, border_offsets.c14())[index];
256		33264	filter_.horizontal_scalar_path(src, &dst_rows[index]);
257		33264	}
258
259			FilterType filter_;
260			}; // end of class SeparableFilter<FilterType, 15UL>
261
262			// Shorthand for 15x15 separable filters driver type.
263			template <class FilterType>
264			using SeparableFilter15x15 = SeparableFilter<FilterType, 15UL>;
265
266			} // namespace KLEIDICV_TARGET_NAMESPACE
267
268			#endif // KLEIDICV_SEPARABLE_FILTER_15X15_SC_H
269