volk_16i_max_star_16i.h

Go to the documentation of this file.

/* -*- c++ -*- */
/*
 * Copyright 2012, 2014 Free Software Foundation, Inc.
 *
 * This file is part of VOLK
 *
 * SPDX-License-Identifier: LGPL-3.0-or-later
 */
 
#ifndef INCLUDED_volk_16i_max_star_16i_a_H
#define INCLUDED_volk_16i_max_star_16i_a_H
 
#include <inttypes.h>
#include <stdio.h>
 
#ifdef LV_HAVE_SSSE3
 
#include <emmintrin.h>
#include <tmmintrin.h>
#include <xmmintrin.h>
 
static inline void
volk_16i_max_star_16i_a_ssse3(short* target, short* src0, unsigned int num_points)
{
    const unsigned int num_bytes = num_points * 2;
 
    short candidate = src0[0];
    short cands[8];
    __m128i xmm0, xmm1, xmm3, xmm4, xmm5, xmm6;
 
    __m128i* p_src0;
 
    p_src0 = (__m128i*)src0;
 
    int bound = num_bytes >> 4;
    int leftovers = (num_bytes >> 1) & 7;
 
    int i = 0;
 
    xmm1 = _mm_setzero_si128();
    xmm0 = _mm_setzero_si128();
    //_mm_insert_epi16(xmm0, candidate, 0);
 
    xmm0 = _mm_shuffle_epi8(xmm0, xmm1);
 
    for (i = 0; i < bound; ++i) {
        xmm1 = _mm_load_si128(p_src0);
        p_src0 += 1;
        // xmm2 = _mm_sub_epi16(xmm1, xmm0);
 
        xmm3 = _mm_cmpgt_epi16(xmm0, xmm1);
        xmm4 = _mm_cmpeq_epi16(xmm0, xmm1);
        xmm5 = _mm_cmpgt_epi16(xmm1, xmm0);
 
        xmm6 = _mm_xor_si128(xmm4, xmm5);
 
        xmm3 = _mm_and_si128(xmm3, xmm0);
        xmm4 = _mm_and_si128(xmm6, xmm1);
 
        xmm0 = _mm_add_epi16(xmm3, xmm4);
    }
 
    _mm_store_si128((__m128i*)cands, xmm0);
 
    for (i = 0; i < 8; ++i) {
        candidate = ((short)(candidate - cands[i]) > 0) ? candidate : cands[i];
    }
 
    for (i = 0; i < leftovers; ++i) {
        candidate = ((short)(candidate - src0[(bound << 3) + i]) > 0)
                        ? candidate
                        : src0[(bound << 3) + i];
    }
 
    target[0] = candidate;
}
 
#endif /*LV_HAVE_SSSE3*/
 
#ifdef LV_HAVE_NEON
#include <arm_neon.h>
 
static inline void
volk_16i_max_star_16i_neon(short* target, short* src0, unsigned int num_points)
{
    const unsigned int eighth_points = num_points / 8;
    unsigned number;
    int16x8_t input_vec;
    int16x8_t diff, zeros;
    uint16x8_t comp1, comp2;
    zeros = vdupq_n_s16(0);
 
    int16x8x2_t tmpvec;
 
    int16x8_t candidate_vec = vld1q_dup_s16(src0);
    short candidate;
    ++src0;
 
    for (number = 0; number < eighth_points; ++number) {
        input_vec = vld1q_s16(src0);
        __VOLK_PREFETCH(src0 + 16);
        diff = vsubq_s16(candidate_vec, input_vec);
        comp1 = vcgeq_s16(diff, zeros);
        comp2 = vcltq_s16(diff, zeros);
 
        tmpvec.val[0] = vandq_s16(candidate_vec, (int16x8_t)comp1);
        tmpvec.val[1] = vandq_s16(input_vec, (int16x8_t)comp2);
 
        candidate_vec = vaddq_s16(tmpvec.val[0], tmpvec.val[1]);
        src0 += 8;
    }
    vst1q_s16(&candidate, candidate_vec);
 
    for (number = 0; number < num_points % 8; number++) {
        candidate = ((int16_t)(candidate - src0[number]) > 0) ? candidate : src0[number];
    }
    target[0] = candidate;
}
#endif /*LV_HAVE_NEON*/
 
#ifdef LV_HAVE_GENERIC
 
static inline void
volk_16i_max_star_16i_generic(short* target, short* src0, unsigned int num_points)
{
    const unsigned int num_bytes = num_points * 2;
 
    int i = 0;
 
    int bound = num_bytes >> 1;
 
    short candidate = src0[0];
    for (i = 1; i < bound; ++i) {
        candidate = ((short)(candidate - src0[i]) > 0) ? candidate : src0[i];
    }
    target[0] = candidate;
}
 
#endif /*LV_HAVE_GENERIC*/
 
 
#endif /*INCLUDED_volk_16i_max_star_16i_a_H*/