\doxysection{conv\+\_\+acc\+\_\+neon\+\_\+impl.\+h}
\label{conv__acc__neon__impl_8h_source}\index{src/core/conv\_acc\_neon\_impl.h@{src/core/conv\_acc\_neon\_impl.h}}
\textbf{ Go to the documentation of this file.}
\begin{DoxyCode}{0}
\DoxyCodeLine{1 }
\DoxyCodeLine{4 \textcolor{comment}{/*}}
\DoxyCodeLine{5 \textcolor{comment}{ * (C) 2020 by sysmocom -\/ s.f.m.c. GmbH}}
\DoxyCodeLine{6 \textcolor{comment}{ * Author: Eric Wild}}
\DoxyCodeLine{7 \textcolor{comment}{ *}}
\DoxyCodeLine{8 \textcolor{comment}{ * All Rights Reserved}}
\DoxyCodeLine{9 \textcolor{comment}{ *}}
\DoxyCodeLine{10 \textcolor{comment}{ * SPDX-\/License-\/Identifier: GPL-\/2.0+}}
\DoxyCodeLine{11 \textcolor{comment}{ *}}
\DoxyCodeLine{12 \textcolor{comment}{ * This program is free software; you can redistribute it and/or modify}}
\DoxyCodeLine{13 \textcolor{comment}{ * it under the terms of the GNU General Public License as published by}}
\DoxyCodeLine{14 \textcolor{comment}{ * the Free Software Foundation; either version 2 of the License, or}}
\DoxyCodeLine{15 \textcolor{comment}{ * (at your option) any later version.}}
\DoxyCodeLine{16 \textcolor{comment}{ *}}
\DoxyCodeLine{17 \textcolor{comment}{ * This program is distributed in the hope that it will be useful,}}
\DoxyCodeLine{18 \textcolor{comment}{ * but WITHOUT ANY WARRANTY; without even the implied warranty of}}
\DoxyCodeLine{19 \textcolor{comment}{ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the}}
\DoxyCodeLine{20 \textcolor{comment}{ * GNU General Public License for more details.}}
\DoxyCodeLine{21 \textcolor{comment}{ */}}
\DoxyCodeLine{22 }
\DoxyCodeLine{23 \textcolor{comment}{/* Some distributions (notably Alpine Linux) for some strange reason}}
\DoxyCodeLine{24 \textcolor{comment}{ * don't have this \#define */}}
\DoxyCodeLine{25 \textcolor{preprocessor}{\#ifndef \_\_always\_inline}}
\DoxyCodeLine{26 \textcolor{preprocessor}{\#define \_\_always\_inline inline \_\_attribute\_\_((always\_inline))}}
\DoxyCodeLine{27 \textcolor{preprocessor}{\#endif}}
\DoxyCodeLine{28 }
\DoxyCodeLine{29 \textcolor{preprocessor}{\#define NEON\_BUTTERFLY(M0,M1,M2,M3,M4) \(\backslash\)}}
\DoxyCodeLine{30 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{31 \textcolor{preprocessor}{        M3 = vqaddq\_s16(M0, M2); \(\backslash\)}}
\DoxyCodeLine{32 \textcolor{preprocessor}{        M4 = vqsubq\_s16(M1, M2); \(\backslash\)}}
\DoxyCodeLine{33 \textcolor{preprocessor}{        M0 = vqsubq\_s16(M0, M2); \(\backslash\)}}
\DoxyCodeLine{34 \textcolor{preprocessor}{        M1 = vqaddq\_s16(M1, M2); \(\backslash\)}}
\DoxyCodeLine{35 \textcolor{preprocessor}{        M2 = vmaxq\_s16(M3, M4); \(\backslash\)}}
\DoxyCodeLine{36 \textcolor{preprocessor}{        M3 = vreinterpretq\_s16\_u16(vcgtq\_s16(M3, M4)); \(\backslash\)}}
\DoxyCodeLine{37 \textcolor{preprocessor}{        M4 = vmaxq\_s16(M0, M1); \(\backslash\)}}
\DoxyCodeLine{38 \textcolor{preprocessor}{        M1 = vreinterpretq\_s16\_u16(vcgtq\_s16(M0, M1)); \(\backslash\)}}
\DoxyCodeLine{39 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{40 }
\DoxyCodeLine{41 \textcolor{preprocessor}{\#define NEON\_DEINTERLEAVE\_K5(M0,M1,M2,M3) \(\backslash\)}}
\DoxyCodeLine{42 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{43 \textcolor{preprocessor}{        int16x8x2\_t tmp; \(\backslash\)}}
\DoxyCodeLine{44 \textcolor{preprocessor}{        tmp = vuzpq\_s16(M0, M1); \(\backslash\)}}
\DoxyCodeLine{45 \textcolor{preprocessor}{        M2 = tmp.val[0]; \(\backslash\)}}
\DoxyCodeLine{46 \textcolor{preprocessor}{        M3 = tmp.val[1]; \(\backslash\)}}
\DoxyCodeLine{47 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{48 }
\DoxyCodeLine{49 \textcolor{preprocessor}{\#define NEON\_DEINTERLEAVE\_K7(M0,M1,M2,M3,M4,M5,M6,M7,M8,M9,M10,M11,M12,M13,M14,M15) \(\backslash\)}}
\DoxyCodeLine{50 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{51 \textcolor{preprocessor}{        int16x8x2\_t tmp; \(\backslash\)}}
\DoxyCodeLine{52 \textcolor{preprocessor}{        tmp = vuzpq\_s16(M0, M1); \(\backslash\)}}
\DoxyCodeLine{53 \textcolor{preprocessor}{        M8 = tmp.val[0]; M9 = tmp.val[1]; \(\backslash\)}}
\DoxyCodeLine{54 \textcolor{preprocessor}{        tmp = vuzpq\_s16(M2, M3); \(\backslash\)}}
\DoxyCodeLine{55 \textcolor{preprocessor}{        M10 = tmp.val[0]; M11 = tmp.val[1]; \(\backslash\)}}
\DoxyCodeLine{56 \textcolor{preprocessor}{        tmp = vuzpq\_s16(M4, M5); \(\backslash\)}}
\DoxyCodeLine{57 \textcolor{preprocessor}{        M12 = tmp.val[0]; M13 = tmp.val[1]; \(\backslash\)}}
\DoxyCodeLine{58 \textcolor{preprocessor}{        tmp = vuzpq\_s16(M6, M7); \(\backslash\)}}
\DoxyCodeLine{59 \textcolor{preprocessor}{        M14 = tmp.val[0]; M15 = tmp.val[1]; \(\backslash\)}}
\DoxyCodeLine{60 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{61 }
\DoxyCodeLine{62 \textcolor{preprocessor}{\#define NEON\_BRANCH\_METRIC\_N2(M0,M1,M2,M3,M4,M6,M7) \(\backslash\)}}
\DoxyCodeLine{63 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{64 \textcolor{preprocessor}{        M0 = vmulq\_s16(M4, M0); \(\backslash\)}}
\DoxyCodeLine{65 \textcolor{preprocessor}{        M1 = vmulq\_s16(M4, M1); \(\backslash\)}}
\DoxyCodeLine{66 \textcolor{preprocessor}{        M2 = vmulq\_s16(M4, M2); \(\backslash\)}}
\DoxyCodeLine{67 \textcolor{preprocessor}{        M3 = vmulq\_s16(M4, M3); \(\backslash\)}}
\DoxyCodeLine{68 \textcolor{preprocessor}{        M6 = vcombine\_s16(vpadd\_s16(vget\_low\_s16(M0), vget\_high\_s16(M0)), vpadd\_s16(vget\_low\_s16(M1), vget\_high\_s16(M1))); \(\backslash\)}}
\DoxyCodeLine{69 \textcolor{preprocessor}{        M7 = vcombine\_s16(vpadd\_s16(vget\_low\_s16(M2), vget\_high\_s16(M2)), vpadd\_s16(vget\_low\_s16(M3), vget\_high\_s16(M3))); \(\backslash\)}}
\DoxyCodeLine{70 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{71 }
\DoxyCodeLine{72 \textcolor{preprocessor}{\#define NEON\_BRANCH\_METRIC\_N4(M0,M1,M2,M3,M4,M5) \(\backslash\)}}
\DoxyCodeLine{73 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{74 \textcolor{preprocessor}{        M0 = vmulq\_s16(M4, M0); \(\backslash\)}}
\DoxyCodeLine{75 \textcolor{preprocessor}{        M1 = vmulq\_s16(M4, M1); \(\backslash\)}}
\DoxyCodeLine{76 \textcolor{preprocessor}{        M2 = vmulq\_s16(M4, M2); \(\backslash\)}}
\DoxyCodeLine{77 \textcolor{preprocessor}{        M3 = vmulq\_s16(M4, M3); \(\backslash\)}}
\DoxyCodeLine{78 \textcolor{preprocessor}{        int16x4\_t t1 = vpadd\_s16(vpadd\_s16(vget\_low\_s16(M0), vget\_high\_s16(M0)), vpadd\_s16(vget\_low\_s16(M1), vget\_high\_s16(M1))); \(\backslash\)}}
\DoxyCodeLine{79 \textcolor{preprocessor}{        int16x4\_t t2 = vpadd\_s16(vpadd\_s16(vget\_low\_s16(M2), vget\_high\_s16(M2)), vpadd\_s16(vget\_low\_s16(M3), vget\_high\_s16(M3))); \(\backslash\)}}
\DoxyCodeLine{80 \textcolor{preprocessor}{        M5 = vcombine\_s16(t1, t2); \(\backslash\)}}
\DoxyCodeLine{81 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{82 }
\DoxyCodeLine{83 \textcolor{preprocessor}{\#define NEON\_NORMALIZE\_K5(M0,M1,M2,M3) \(\backslash\)}}
\DoxyCodeLine{84 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{85 \textcolor{preprocessor}{        M2 = vminq\_s16(M0, M1); \(\backslash\)}}
\DoxyCodeLine{86 \textcolor{preprocessor}{        int16x4\_t t = vpmin\_s16(vget\_low\_s16(M2), vget\_high\_s16(M2)); \(\backslash\)}}
\DoxyCodeLine{87 \textcolor{preprocessor}{        t = vpmin\_s16(t, t); \(\backslash\)}}
\DoxyCodeLine{88 \textcolor{preprocessor}{        t = vpmin\_s16(t, t); \(\backslash\)}}
\DoxyCodeLine{89 \textcolor{preprocessor}{        M2 = vdupq\_lane\_s16(t, 0); \(\backslash\)}}
\DoxyCodeLine{90 \textcolor{preprocessor}{        M0 = vqsubq\_s16(M0, M2); \(\backslash\)}}
\DoxyCodeLine{91 \textcolor{preprocessor}{        M1 = vqsubq\_s16(M1, M2); \(\backslash\)}}
\DoxyCodeLine{92 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{93 }
\DoxyCodeLine{94 \textcolor{preprocessor}{\#define NEON\_NORMALIZE\_K7(M0,M1,M2,M3,M4,M5,M6,M7,M8,M9,M10,M11) \(\backslash\)}}
\DoxyCodeLine{95 \textcolor{preprocessor}{\{ \(\backslash\)}}
\DoxyCodeLine{96 \textcolor{preprocessor}{        M8 = vminq\_s16(M0, M1); \(\backslash\)}}
\DoxyCodeLine{97 \textcolor{preprocessor}{        M9 = vminq\_s16(M2, M3); \(\backslash\)}}
\DoxyCodeLine{98 \textcolor{preprocessor}{        M10 = vminq\_s16(M4, M5); \(\backslash\)}}
\DoxyCodeLine{99 \textcolor{preprocessor}{        M11 = vminq\_s16(M6, M7); \(\backslash\)}}
\DoxyCodeLine{100 \textcolor{preprocessor}{        M8 = vminq\_s16(M8, M9); \(\backslash\)}}
\DoxyCodeLine{101 \textcolor{preprocessor}{        M10 = vminq\_s16(M10, M11); \(\backslash\)}}
\DoxyCodeLine{102 \textcolor{preprocessor}{        M8 = vminq\_s16(M8, M10); \(\backslash\)}}
\DoxyCodeLine{103 \textcolor{preprocessor}{        int16x4\_t t = vpmin\_s16(vget\_low\_s16(M8), vget\_high\_s16(M8)); \(\backslash\)}}
\DoxyCodeLine{104 \textcolor{preprocessor}{        t = vpmin\_s16(t, t); \(\backslash\)}}
\DoxyCodeLine{105 \textcolor{preprocessor}{        t = vpmin\_s16(t, t); \(\backslash\)}}
\DoxyCodeLine{106 \textcolor{preprocessor}{        M8 = vdupq\_lane\_s16(t, 0); \(\backslash\)}}
\DoxyCodeLine{107 \textcolor{preprocessor}{        M0 = vqsubq\_s16(M0, M8); \(\backslash\)}}
\DoxyCodeLine{108 \textcolor{preprocessor}{        M1 = vqsubq\_s16(M1, M8); \(\backslash\)}}
\DoxyCodeLine{109 \textcolor{preprocessor}{        M2 = vqsubq\_s16(M2, M8); \(\backslash\)}}
\DoxyCodeLine{110 \textcolor{preprocessor}{        M3 = vqsubq\_s16(M3, M8); \(\backslash\)}}
\DoxyCodeLine{111 \textcolor{preprocessor}{        M4 = vqsubq\_s16(M4, M8); \(\backslash\)}}
\DoxyCodeLine{112 \textcolor{preprocessor}{        M5 = vqsubq\_s16(M5, M8); \(\backslash\)}}
\DoxyCodeLine{113 \textcolor{preprocessor}{        M6 = vqsubq\_s16(M6, M8); \(\backslash\)}}
\DoxyCodeLine{114 \textcolor{preprocessor}{        M7 = vqsubq\_s16(M7, M8); \(\backslash\)}}
\DoxyCodeLine{115 \textcolor{preprocessor}{\}}}
\DoxyCodeLine{116 }
\DoxyCodeLine{117 \_\_always\_inline \textcolor{keywordtype}{void} \_neon\_metrics\_k5\_n2(\textcolor{keyword}{const} int16\_t *val, \textcolor{keyword}{const} int16\_t *outa, int16\_t *sumsa, int16\_t *paths,}
\DoxyCodeLine{118                                          \textcolor{keywordtype}{int} norm)}
\DoxyCodeLine{119 \{}
\DoxyCodeLine{120         int16\_t *\_\_restrict out = \_\_builtin\_assume\_aligned(outa, 8);}
\DoxyCodeLine{121         int16\_t *\_\_restrict sums = \_\_builtin\_assume\_aligned(sumsa, 8);}
\DoxyCodeLine{122         int16x8\_t m0, m1, m2, m3, m4, m5, m6;}
\DoxyCodeLine{123         int16x4\_t input;}
\DoxyCodeLine{124 }
\DoxyCodeLine{125         \textcolor{comment}{/* (BMU) Load and expand 8-\/bit input out to 16-\/bits */}}
\DoxyCodeLine{126         input = vld1\_s16(val);}
\DoxyCodeLine{127         m2 = vcombine\_s16(input, input);}
\DoxyCodeLine{128 }
\DoxyCodeLine{129         \textcolor{comment}{/* (BMU) Load and compute branch metrics */}}
\DoxyCodeLine{130         m0 = vld1q\_s16(\&out[0]);}
\DoxyCodeLine{131         m1 = vld1q\_s16(\&out[8]);}
\DoxyCodeLine{132 }
\DoxyCodeLine{133         m0 = vmulq\_s16(m2, m0);}
\DoxyCodeLine{134         m1 = vmulq\_s16(m2, m1);}
\DoxyCodeLine{135         m2 = vcombine\_s16(vpadd\_s16(vget\_low\_s16(m0), vget\_high\_s16(m0)),}
\DoxyCodeLine{136                           vpadd\_s16(vget\_low\_s16(m1), vget\_high\_s16(m1)));}
\DoxyCodeLine{137 }
\DoxyCodeLine{138         \textcolor{comment}{/* (PMU) Load accumulated path matrics */}}
\DoxyCodeLine{139         m0 = vld1q\_s16(\&sums[0]);}
\DoxyCodeLine{140         m1 = vld1q\_s16(\&sums[8]);}
\DoxyCodeLine{141 }
\DoxyCodeLine{142         NEON\_DEINTERLEAVE\_K5(m0, m1, m3, m4)}
\DoxyCodeLine{143 }
\DoxyCodeLine{144         \textcolor{comment}{/* (PMU) Butterflies: 0-\/7 */}}
\DoxyCodeLine{145         NEON\_BUTTERFLY(m3, m4, m2, m5, m6)}
\DoxyCodeLine{146 }
\DoxyCodeLine{147         \textcolor{keywordflow}{if} (norm)}
\DoxyCodeLine{148                 NEON\_NORMALIZE\_K5(m2, m6, m0, m1)}
\DoxyCodeLine{149 }
\DoxyCodeLine{150         vst1q\_s16(\&sums[0], m2);}
\DoxyCodeLine{151         vst1q\_s16(\&sums[8], m6);}
\DoxyCodeLine{152         vst1q\_s16(\&paths[0], m5);}
\DoxyCodeLine{153         vst1q\_s16(\&paths[8], m4);}
\DoxyCodeLine{154 \}}
\DoxyCodeLine{155 }
\DoxyCodeLine{156 \_\_always\_inline \textcolor{keywordtype}{void} \_neon\_metrics\_k5\_n4(\textcolor{keyword}{const} int16\_t *val, \textcolor{keyword}{const} int16\_t *outa, int16\_t *sumsa, int16\_t *paths,}
\DoxyCodeLine{157                                          \textcolor{keywordtype}{int} norm)}
\DoxyCodeLine{158 \{}
\DoxyCodeLine{159         int16\_t *\_\_restrict out = \_\_builtin\_assume\_aligned(outa, 8);}
\DoxyCodeLine{160         int16\_t *\_\_restrict sums = \_\_builtin\_assume\_aligned(sumsa, 8);}
\DoxyCodeLine{161         int16x8\_t m0, m1, m2, m3, m4, m5, m6;}
\DoxyCodeLine{162         int16x4\_t input;}
\DoxyCodeLine{163 }
\DoxyCodeLine{164         \textcolor{comment}{/* (BMU) Load and expand 8-\/bit input out to 16-\/bits */}}
\DoxyCodeLine{165         input = vld1\_s16(val);}
\DoxyCodeLine{166         m4 = vcombine\_s16(input, input);}
\DoxyCodeLine{167 }
\DoxyCodeLine{168         \textcolor{comment}{/* (BMU) Load and compute branch metrics */}}
\DoxyCodeLine{169         m0 = vld1q\_s16(\&out[0]);}
\DoxyCodeLine{170         m1 = vld1q\_s16(\&out[8]);}
\DoxyCodeLine{171         m2 = vld1q\_s16(\&out[16]);}
\DoxyCodeLine{172         m3 = vld1q\_s16(\&out[24]);}
\DoxyCodeLine{173 }
\DoxyCodeLine{174         NEON\_BRANCH\_METRIC\_N4(m0, m1, m2, m3, m4, m2)}
\DoxyCodeLine{175 }
\DoxyCodeLine{176         \textcolor{comment}{/* (PMU) Load accumulated path matrics */}}
\DoxyCodeLine{177         m0 = vld1q\_s16(\&sums[0]);}
\DoxyCodeLine{178         m1 = vld1q\_s16(\&sums[8]);}
\DoxyCodeLine{179 }
\DoxyCodeLine{180         NEON\_DEINTERLEAVE\_K5(m0, m1, m3, m4)}
\DoxyCodeLine{181 }
\DoxyCodeLine{182         \textcolor{comment}{/* (PMU) Butterflies: 0-\/7 */}}
\DoxyCodeLine{183         NEON\_BUTTERFLY(m3, m4, m2, m5, m6)}
\DoxyCodeLine{184 }
\DoxyCodeLine{185         \textcolor{keywordflow}{if} (norm)}
\DoxyCodeLine{186                 NEON\_NORMALIZE\_K5(m2, m6, m0, m1)}
\DoxyCodeLine{187 }
\DoxyCodeLine{188         vst1q\_s16(\&sums[0], m2);}
\DoxyCodeLine{189         vst1q\_s16(\&sums[8], m6);}
\DoxyCodeLine{190         vst1q\_s16(\&paths[0], m5);}
\DoxyCodeLine{191         vst1q\_s16(\&paths[8], m4);}
\DoxyCodeLine{192 \}}
\DoxyCodeLine{193 }
\DoxyCodeLine{194 \_\_always\_inline \textcolor{keyword}{static} \textcolor{keywordtype}{void} \_neon\_metrics\_k7\_n2(\textcolor{keyword}{const} int16\_t *val, \textcolor{keyword}{const} int16\_t *outa, int16\_t *sumsa, int16\_t *paths,}
\DoxyCodeLine{195                                                 \textcolor{keywordtype}{int} norm)}
\DoxyCodeLine{196 \{}
\DoxyCodeLine{197         int16\_t *\_\_restrict out = \_\_builtin\_assume\_aligned(outa, 8);}
\DoxyCodeLine{198         int16\_t *\_\_restrict sums = \_\_builtin\_assume\_aligned(sumsa, 8);}
\DoxyCodeLine{199         int16x8\_t m0, m1, m2, m3, m4, m5, m6, m7;}
\DoxyCodeLine{200         int16x8\_t m8, m9, m10, m11, m12, m13, m14, m15;}
\DoxyCodeLine{201         int16x4\_t input;}
\DoxyCodeLine{202 }
\DoxyCodeLine{203         \textcolor{comment}{/* (PMU) Load accumulated path matrics */}}
\DoxyCodeLine{204         m0 = vld1q\_s16(\&sums[0]);}
\DoxyCodeLine{205         m1 = vld1q\_s16(\&sums[8]);}
\DoxyCodeLine{206         m2 = vld1q\_s16(\&sums[16]);}
\DoxyCodeLine{207         m3 = vld1q\_s16(\&sums[24]);}
\DoxyCodeLine{208         m4 = vld1q\_s16(\&sums[32]);}
\DoxyCodeLine{209         m5 = vld1q\_s16(\&sums[40]);}
\DoxyCodeLine{210         m6 = vld1q\_s16(\&sums[48]);}
\DoxyCodeLine{211         m7 = vld1q\_s16(\&sums[56]);}
\DoxyCodeLine{212 }
\DoxyCodeLine{213         \textcolor{comment}{/* (PMU) Deinterleave into even and odd packed registers */}}
\DoxyCodeLine{214         NEON\_DEINTERLEAVE\_K7(m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m12, m13, m14, m15)}
\DoxyCodeLine{215 }
\DoxyCodeLine{216         \textcolor{comment}{/* (BMU) Load and expand 8-\/bit input out to 16-\/bits */}}
\DoxyCodeLine{217         input = vld1\_s16(val);}
\DoxyCodeLine{218         m7 = vcombine\_s16(input, input);}
\DoxyCodeLine{219 }
\DoxyCodeLine{220         \textcolor{comment}{/* (BMU) Load and compute branch metrics */}}
\DoxyCodeLine{221         m0 = vld1q\_s16(\&out[0]);}
\DoxyCodeLine{222         m1 = vld1q\_s16(\&out[8]);}
\DoxyCodeLine{223         m2 = vld1q\_s16(\&out[16]);}
\DoxyCodeLine{224         m3 = vld1q\_s16(\&out[24]);}
\DoxyCodeLine{225 }
\DoxyCodeLine{226         NEON\_BRANCH\_METRIC\_N2(m0, m1, m2, m3, m7, m4, m5)}
\DoxyCodeLine{227 }
\DoxyCodeLine{228         m0 = vld1q\_s16(\&out[32]);}
\DoxyCodeLine{229         m1 = vld1q\_s16(\&out[40]);}
\DoxyCodeLine{230         m2 = vld1q\_s16(\&out[48]);}
\DoxyCodeLine{231         m3 = vld1q\_s16(\&out[56]);}
\DoxyCodeLine{232 }
\DoxyCodeLine{233         NEON\_BRANCH\_METRIC\_N2(m0, m1, m2, m3, m7, m6, m7)}
\DoxyCodeLine{234 }
\DoxyCodeLine{235         \textcolor{comment}{/* (PMU) Butterflies: 0-\/15 */}}
\DoxyCodeLine{236         NEON\_BUTTERFLY(m8, m9, m4, m0, m1)}
\DoxyCodeLine{237         NEON\_BUTTERFLY(m10, m11, m5, m2, m3)}
\DoxyCodeLine{238 }
\DoxyCodeLine{239         vst1q\_s16(\&paths[0], m0);}
\DoxyCodeLine{240         vst1q\_s16(\&paths[8], m2);}
\DoxyCodeLine{241         vst1q\_s16(\&paths[32], m9);}
\DoxyCodeLine{242         vst1q\_s16(\&paths[40], m11);}
\DoxyCodeLine{243 }
\DoxyCodeLine{244         \textcolor{comment}{/* (PMU) Butterflies: 17-\/31 */}}
\DoxyCodeLine{245         NEON\_BUTTERFLY(m12, m13, m6, m0, m2)}
\DoxyCodeLine{246         NEON\_BUTTERFLY(m14, m15, m7, m9, m11)}
\DoxyCodeLine{247 }
\DoxyCodeLine{248         vst1q\_s16(\&paths[16], m0);}
\DoxyCodeLine{249         vst1q\_s16(\&paths[24], m9);}
\DoxyCodeLine{250         vst1q\_s16(\&paths[48], m13);}
\DoxyCodeLine{251         vst1q\_s16(\&paths[56], m15);}
\DoxyCodeLine{252 }
\DoxyCodeLine{253         \textcolor{keywordflow}{if} (norm)}
\DoxyCodeLine{254                 NEON\_NORMALIZE\_K7(m4, m1, m5, m3, m6, m2, m7, m11, m0, m8, m9, m10)}
\DoxyCodeLine{255 }
\DoxyCodeLine{256         vst1q\_s16(\&sums[0], m4);}
\DoxyCodeLine{257         vst1q\_s16(\&sums[8], m5);}
\DoxyCodeLine{258         vst1q\_s16(\&sums[16], m6);}
\DoxyCodeLine{259         vst1q\_s16(\&sums[24], m7);}
\DoxyCodeLine{260         vst1q\_s16(\&sums[32], m1);}
\DoxyCodeLine{261         vst1q\_s16(\&sums[40], m3);}
\DoxyCodeLine{262         vst1q\_s16(\&sums[48], m2);}
\DoxyCodeLine{263         vst1q\_s16(\&sums[56], m11);}
\DoxyCodeLine{264 \}}
\DoxyCodeLine{265 }
\DoxyCodeLine{266 \_\_always\_inline \textcolor{keyword}{static} \textcolor{keywordtype}{void} \_neon\_metrics\_k7\_n4(\textcolor{keyword}{const} int16\_t *val, \textcolor{keyword}{const} int16\_t *outa, int16\_t *sumsa, int16\_t *paths,}
\DoxyCodeLine{267                                                 \textcolor{keywordtype}{int} norm)}
\DoxyCodeLine{268 \{}
\DoxyCodeLine{269         int16\_t *\_\_restrict out = \_\_builtin\_assume\_aligned(outa, 8);}
\DoxyCodeLine{270         int16\_t *\_\_restrict sums = \_\_builtin\_assume\_aligned(sumsa, 8);}
\DoxyCodeLine{271         int16x8\_t m0, m1, m2, m3, m4, m5, m6, m7;}
\DoxyCodeLine{272         int16x8\_t m8, m9, m10, m11, m12, m13, m14, m15;}
\DoxyCodeLine{273         int16x4\_t input;}
\DoxyCodeLine{274 }
\DoxyCodeLine{275         \textcolor{comment}{/* (PMU) Load accumulated path matrics */}}
\DoxyCodeLine{276         m0 = vld1q\_s16(\&sums[0]);}
\DoxyCodeLine{277         m1 = vld1q\_s16(\&sums[8]);}
\DoxyCodeLine{278         m2 = vld1q\_s16(\&sums[16]);}
\DoxyCodeLine{279         m3 = vld1q\_s16(\&sums[24]);}
\DoxyCodeLine{280         m4 = vld1q\_s16(\&sums[32]);}
\DoxyCodeLine{281         m5 = vld1q\_s16(\&sums[40]);}
\DoxyCodeLine{282         m6 = vld1q\_s16(\&sums[48]);}
\DoxyCodeLine{283         m7 = vld1q\_s16(\&sums[56]);}
\DoxyCodeLine{284 }
\DoxyCodeLine{285         \textcolor{comment}{/* (PMU) Deinterleave into even and odd packed registers */}}
\DoxyCodeLine{286         NEON\_DEINTERLEAVE\_K7(m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m12, m13, m14, m15)}
\DoxyCodeLine{287 }
\DoxyCodeLine{288         \textcolor{comment}{/* (BMU) Load and expand 8-\/bit input out to 16-\/bits */}}
\DoxyCodeLine{289         input = vld1\_s16(val);}
\DoxyCodeLine{290         m7 = vcombine\_s16(input, input);}
\DoxyCodeLine{291 }
\DoxyCodeLine{292         \textcolor{comment}{/* (BMU) Load and compute branch metrics */}}
\DoxyCodeLine{293         m0 = vld1q\_s16(\&out[0]);}
\DoxyCodeLine{294         m1 = vld1q\_s16(\&out[8]);}
\DoxyCodeLine{295         m2 = vld1q\_s16(\&out[16]);}
\DoxyCodeLine{296         m3 = vld1q\_s16(\&out[24]);}
\DoxyCodeLine{297 }
\DoxyCodeLine{298         NEON\_BRANCH\_METRIC\_N4(m0, m1, m2, m3, m7, m4)}
\DoxyCodeLine{299 }
\DoxyCodeLine{300         m0 = vld1q\_s16(\&out[32]);}
\DoxyCodeLine{301         m1 = vld1q\_s16(\&out[40]);}
\DoxyCodeLine{302         m2 = vld1q\_s16(\&out[48]);}
\DoxyCodeLine{303         m3 = vld1q\_s16(\&out[56]);}
\DoxyCodeLine{304 }
\DoxyCodeLine{305         NEON\_BRANCH\_METRIC\_N4(m0, m1, m2, m3, m7, m5)}
\DoxyCodeLine{306 }
\DoxyCodeLine{307         m0 = vld1q\_s16(\&out[64]);}
\DoxyCodeLine{308         m1 = vld1q\_s16(\&out[72]);}
\DoxyCodeLine{309         m2 = vld1q\_s16(\&out[80]);}
\DoxyCodeLine{310         m3 = vld1q\_s16(\&out[88]);}
\DoxyCodeLine{311 }
\DoxyCodeLine{312         NEON\_BRANCH\_METRIC\_N4(m0, m1, m2, m3, m7, m6)}
\DoxyCodeLine{313 }
\DoxyCodeLine{314         m0 = vld1q\_s16(\&out[96]);}
\DoxyCodeLine{315         m1 = vld1q\_s16(\&out[104]);}
\DoxyCodeLine{316         m2 = vld1q\_s16(\&out[112]);}
\DoxyCodeLine{317         m3 = vld1q\_s16(\&out[120]);}
\DoxyCodeLine{318 }
\DoxyCodeLine{319         NEON\_BRANCH\_METRIC\_N4(m0, m1, m2, m3, m7, m7)}
\DoxyCodeLine{320 }
\DoxyCodeLine{321         \textcolor{comment}{/* (PMU) Butterflies: 0-\/15 */}}
\DoxyCodeLine{322         NEON\_BUTTERFLY(m8, m9, m4, m0, m1)}
\DoxyCodeLine{323         NEON\_BUTTERFLY(m10, m11, m5, m2, m3)}
\DoxyCodeLine{324 }
\DoxyCodeLine{325         vst1q\_s16(\&paths[0], m0);}
\DoxyCodeLine{326         vst1q\_s16(\&paths[8], m2);}
\DoxyCodeLine{327         vst1q\_s16(\&paths[32], m9);}
\DoxyCodeLine{328         vst1q\_s16(\&paths[40], m11);}
\DoxyCodeLine{329 }
\DoxyCodeLine{330         \textcolor{comment}{/* (PMU) Butterflies: 17-\/31 */}}
\DoxyCodeLine{331         NEON\_BUTTERFLY(m12, m13, m6, m0, m2)}
\DoxyCodeLine{332         NEON\_BUTTERFLY(m14, m15, m7, m9, m11)}
\DoxyCodeLine{333 }
\DoxyCodeLine{334         vst1q\_s16(\&paths[16], m0);}
\DoxyCodeLine{335         vst1q\_s16(\&paths[24], m9);}
\DoxyCodeLine{336         vst1q\_s16(\&paths[48], m13);}
\DoxyCodeLine{337         vst1q\_s16(\&paths[56], m15);}
\DoxyCodeLine{338 }
\DoxyCodeLine{339         \textcolor{keywordflow}{if} (norm)}
\DoxyCodeLine{340                 NEON\_NORMALIZE\_K7(m4, m1, m5, m3, m6, m2, m7, m11, m0, m8, m9, m10)}
\DoxyCodeLine{341 }
\DoxyCodeLine{342         vst1q\_s16(\&sums[0], m4);}
\DoxyCodeLine{343         vst1q\_s16(\&sums[8], m5);}
\DoxyCodeLine{344         vst1q\_s16(\&sums[16], m6);}
\DoxyCodeLine{345         vst1q\_s16(\&sums[24], m7);}
\DoxyCodeLine{346         vst1q\_s16(\&sums[32], m1);}
\DoxyCodeLine{347         vst1q\_s16(\&sums[40], m3);}
\DoxyCodeLine{348         vst1q\_s16(\&sums[48], m2);}
\DoxyCodeLine{349         vst1q\_s16(\&sums[56], m11);}
\DoxyCodeLine{350 \}}

\end{DoxyCode}