| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* -*- c++ -*- */ | ||
| 2 | /* | ||
| 3 | * Copyright 2012, 2014 Free Software Foundation, Inc. | ||
| 4 | * | ||
| 5 | * This file is part of VOLK | ||
| 6 | * | ||
| 7 | * SPDX-License-Identifier: LGPL-3.0-or-later | ||
| 8 | */ | ||
| 9 | |||
| 10 | /*! | ||
| 11 | * \page volk_16i_max_star_16i | ||
| 12 | * | ||
| 13 | * \b Overview | ||
| 14 | * | ||
| 15 | * <FIXME> | ||
| 16 | * | ||
| 17 | * <b>Dispatcher Prototype</b> | ||
| 18 | * \code | ||
| 19 | * void volk_16i_max_star_16i(short* target, short* src0, unsigned int num_points); | ||
| 20 | * \endcode | ||
| 21 | * | ||
| 22 | * \b Inputs | ||
| 23 | * \li src0: The input vector. | ||
| 24 | * \li num_points: The number of complex data points. | ||
| 25 | * | ||
| 26 | * \b Outputs | ||
| 27 | * \li target: The output value of the max* operation. | ||
| 28 | * | ||
| 29 | * \b Example | ||
| 30 | * \code | ||
| 31 | * int N = 10000; | ||
| 32 | * | ||
| 33 | * volk_16i_max_star_16i(); | ||
| 34 | * | ||
| 35 | * volk_free(x); | ||
| 36 | * volk_free(t); | ||
| 37 | * \endcode | ||
| 38 | */ | ||
| 39 | |||
| 40 | #ifndef INCLUDED_volk_16i_max_star_16i_a_H | ||
| 41 | #define INCLUDED_volk_16i_max_star_16i_a_H | ||
| 42 | |||
| 43 | #include <inttypes.h> | ||
| 44 | #include <stdio.h> | ||
| 45 | |||
| 46 | #ifdef LV_HAVE_SSSE3 | ||
| 47 | |||
| 48 | #include <emmintrin.h> | ||
| 49 | #include <tmmintrin.h> | ||
| 50 | #include <xmmintrin.h> | ||
| 51 | |||
| 52 | static inline void | ||
| 53 | ✗ | volk_16i_max_star_16i_a_ssse3(short* target, short* src0, unsigned int num_points) | |
| 54 | { | ||
| 55 | ✗ | const unsigned int num_bytes = num_points * 2; | |
| 56 | |||
| 57 | ✗ | short candidate = src0[0]; | |
| 58 | short cands[8]; | ||
| 59 | __m128i xmm0, xmm1, xmm3, xmm4, xmm5, xmm6; | ||
| 60 | |||
| 61 | __m128i* p_src0; | ||
| 62 | |||
| 63 | ✗ | p_src0 = (__m128i*)src0; | |
| 64 | |||
| 65 | ✗ | int bound = num_bytes >> 4; | |
| 66 | ✗ | int leftovers = (num_bytes >> 1) & 7; | |
| 67 | |||
| 68 | ✗ | int i = 0; | |
| 69 | |||
| 70 | ✗ | xmm1 = _mm_setzero_si128(); | |
| 71 | ✗ | xmm0 = _mm_setzero_si128(); | |
| 72 | //_mm_insert_epi16(xmm0, candidate, 0); | ||
| 73 | |||
| 74 | ✗ | xmm0 = _mm_shuffle_epi8(xmm0, xmm1); | |
| 75 | |||
| 76 | ✗ | for (i = 0; i < bound; ++i) { | |
| 77 | ✗ | xmm1 = _mm_load_si128(p_src0); | |
| 78 | ✗ | p_src0 += 1; | |
| 79 | // xmm2 = _mm_sub_epi16(xmm1, xmm0); | ||
| 80 | |||
| 81 | ✗ | xmm3 = _mm_cmpgt_epi16(xmm0, xmm1); | |
| 82 | ✗ | xmm4 = _mm_cmpeq_epi16(xmm0, xmm1); | |
| 83 | ✗ | xmm5 = _mm_cmpgt_epi16(xmm1, xmm0); | |
| 84 | |||
| 85 | ✗ | xmm6 = _mm_xor_si128(xmm4, xmm5); | |
| 86 | |||
| 87 | ✗ | xmm3 = _mm_and_si128(xmm3, xmm0); | |
| 88 | ✗ | xmm4 = _mm_and_si128(xmm6, xmm1); | |
| 89 | |||
| 90 | ✗ | xmm0 = _mm_add_epi16(xmm3, xmm4); | |
| 91 | } | ||
| 92 | |||
| 93 | _mm_store_si128((__m128i*)cands, xmm0); | ||
| 94 | |||
| 95 | ✗ | for (i = 0; i < 8; ++i) { | |
| 96 | ✗ | candidate = ((short)(candidate - cands[i]) > 0) ? candidate : cands[i]; | |
| 97 | } | ||
| 98 | |||
| 99 | ✗ | for (i = 0; i < leftovers; ++i) { | |
| 100 | ✗ | candidate = ((short)(candidate - src0[(bound << 3) + i]) > 0) | |
| 101 | ? candidate | ||
| 102 | ✗ | : src0[(bound << 3) + i]; | |
| 103 | } | ||
| 104 | |||
| 105 | ✗ | target[0] = candidate; | |
| 106 | ✗ | } | |
| 107 | |||
| 108 | #endif /*LV_HAVE_SSSE3*/ | ||
| 109 | |||
| 110 | #ifdef LV_HAVE_NEON | ||
| 111 | #include <arm_neon.h> | ||
| 112 | |||
| 113 | static inline void | ||
| 114 | volk_16i_max_star_16i_neon(short* target, short* src0, unsigned int num_points) | ||
| 115 | { | ||
| 116 | const unsigned int eighth_points = num_points / 8; | ||
| 117 | unsigned number; | ||
| 118 | int16x8_t input_vec; | ||
| 119 | int16x8_t diff, zeros; | ||
| 120 | uint16x8_t comp1, comp2; | ||
| 121 | zeros = vdupq_n_s16(0); | ||
| 122 | |||
| 123 | int16x8x2_t tmpvec; | ||
| 124 | |||
| 125 | int16x8_t candidate_vec = vld1q_dup_s16(src0); | ||
| 126 | short candidate; | ||
| 127 | ++src0; | ||
| 128 | |||
| 129 | for (number = 0; number < eighth_points; ++number) { | ||
| 130 | input_vec = vld1q_s16(src0); | ||
| 131 | __VOLK_PREFETCH(src0 + 16); | ||
| 132 | diff = vsubq_s16(candidate_vec, input_vec); | ||
| 133 | comp1 = vcgeq_s16(diff, zeros); | ||
| 134 | comp2 = vcltq_s16(diff, zeros); | ||
| 135 | |||
| 136 | tmpvec.val[0] = vandq_s16(candidate_vec, (int16x8_t)comp1); | ||
| 137 | tmpvec.val[1] = vandq_s16(input_vec, (int16x8_t)comp2); | ||
| 138 | |||
| 139 | candidate_vec = vaddq_s16(tmpvec.val[0], tmpvec.val[1]); | ||
| 140 | src0 += 8; | ||
| 141 | } | ||
| 142 | vst1q_s16(&candidate, candidate_vec); | ||
| 143 | |||
| 144 | for (number = 0; number < num_points % 8; number++) { | ||
| 145 | candidate = ((int16_t)(candidate - src0[number]) > 0) ? candidate : src0[number]; | ||
| 146 | } | ||
| 147 | target[0] = candidate; | ||
| 148 | } | ||
| 149 | #endif /*LV_HAVE_NEON*/ | ||
| 150 | |||
| 151 | #ifdef LV_HAVE_GENERIC | ||
| 152 | |||
| 153 | static inline void | ||
| 154 | ✗ | volk_16i_max_star_16i_generic(short* target, short* src0, unsigned int num_points) | |
| 155 | { | ||
| 156 | ✗ | const unsigned int num_bytes = num_points * 2; | |
| 157 | |||
| 158 | ✗ | int i = 0; | |
| 159 | |||
| 160 | ✗ | int bound = num_bytes >> 1; | |
| 161 | |||
| 162 | ✗ | short candidate = src0[0]; | |
| 163 | ✗ | for (i = 1; i < bound; ++i) { | |
| 164 | ✗ | candidate = ((short)(candidate - src0[i]) > 0) ? candidate : src0[i]; | |
| 165 | } | ||
| 166 | ✗ | target[0] = candidate; | |
| 167 | ✗ | } | |
| 168 | |||
| 169 | #endif /*LV_HAVE_GENERIC*/ | ||
| 170 | |||
| 171 | |||
| 172 | #endif /*INCLUDED_volk_16i_max_star_16i_a_H*/ | ||
| 173 |