53 #ifndef INCLUDED_volk_16i_max_star_16i_a_H 54 #define INCLUDED_volk_16i_max_star_16i_a_H 68 const unsigned int num_bytes = num_points*2;
70 short candidate = src0[0];
72 __m128i xmm0, xmm1, xmm3, xmm4, xmm5, xmm6;
76 p_src0 = (__m128i*)src0;
78 int bound = num_bytes >> 4;
79 int leftovers = (num_bytes >> 1) & 7;
83 xmm1 = _mm_setzero_si128();
84 xmm0 = _mm_setzero_si128();
87 xmm0 = _mm_shuffle_epi8(xmm0, xmm1);
89 for(i = 0; i < bound; ++
i) {
90 xmm1 = _mm_load_si128(p_src0);
94 xmm3 = _mm_cmpgt_epi16(xmm0, xmm1);
95 xmm4 = _mm_cmpeq_epi16(xmm0, xmm1);
96 xmm5 = _mm_cmpgt_epi16(xmm1, xmm0);
98 xmm6 = _mm_xor_si128(xmm4, xmm5);
100 xmm3 = _mm_and_si128(xmm3, xmm0);
101 xmm4 = _mm_and_si128(xmm6, xmm1);
103 xmm0 = _mm_add_epi16(xmm3, xmm4);
106 _mm_store_si128((__m128i*)cands, xmm0);
108 for(i = 0; i < 8; ++
i) {
109 candidate = ((short)(candidate - cands[i]) > 0) ? candidate : cands[i];
112 for(i = 0; i < leftovers; ++
i) {
113 candidate = ((short)(candidate - src0[(bound << 3) +
i]) > 0) ? candidate : src0[(bound << 3) + i];
116 target[0] = candidate;
122 #include <arm_neon.h> 127 const unsigned int eighth_points = num_points / 8;
130 int16x8_t diff, zeros;
131 uint16x8_t comp1, comp2;
132 zeros = vdupq_n_s16(0);
136 int16x8_t candidate_vec = vld1q_dup_s16(src0 );
140 for(number=0; number < eighth_points; ++number) {
141 input_vec = vld1q_s16(src0);
143 diff = vsubq_s16(candidate_vec, input_vec);
144 comp1 = vcgeq_s16(diff, zeros);
145 comp2 = vcltq_s16(diff, zeros);
147 tmpvec.val[0] = vandq_s16(candidate_vec, (int16x8_t)comp1);
148 tmpvec.val[1] = vandq_s16(input_vec, (int16x8_t)comp2);
150 candidate_vec = vaddq_s16(tmpvec.val[0], tmpvec.val[1]);
153 vst1q_s16(&candidate, candidate_vec);
155 for(number=0; number < num_points%8; number++) {
156 candidate = ((int16_t)(candidate - src0[number]) > 0) ? candidate : src0[number];
158 target[0] = candidate;
162 #ifdef LV_HAVE_GENERIC 167 const unsigned int num_bytes = num_points*2;
171 int bound = num_bytes >> 1;
173 short candidate = src0[0];
174 for(i = 1; i < bound; ++
i) {
175 candidate = ((short)(candidate - src0[i]) > 0) ? candidate : src0[i];
177 target[0] = candidate;
static void volk_16i_max_star_16i_generic(short *target, short *src0, unsigned int num_points)
Definition: volk_16i_max_star_16i.h:165
#define __VOLK_PREFETCH(addr)
Definition: volk_common.h:39
for i
Definition: volk_config_fixed.tmpl.h:25
static void volk_16i_max_star_16i_neon(short *target, short *src0, unsigned int num_points)
Definition: volk_16i_max_star_16i.h:125
static void volk_16i_max_star_16i_a_ssse3(short *target, short *src0, unsigned int num_points)
Definition: volk_16i_max_star_16i.h:66