volk_32fc_x2_divide_32fc.h

Go to the documentation of this file.

/* -*- c++ -*- */
/*
 * Copyright 2016 Free Software Foundation, Inc.
 *
 * This file is part of VOLK
 *
 * SPDX-License-Identifier: LGPL-3.0-or-later
 */
 
#ifndef INCLUDED_volk_32fc_x2_divide_32fc_u_H
#define INCLUDED_volk_32fc_x2_divide_32fc_u_H
 
#include <float.h>
#include <inttypes.h>
#include <volk/volk_complex.h>
 
 
#ifdef LV_HAVE_GENERIC
 
static inline void volk_32fc_x2_divide_32fc_generic(lv_32fc_t* cVector,
                                                    const lv_32fc_t* aVector,
                                                    const lv_32fc_t* bVector,
                                                    unsigned int num_points)
{
    lv_32fc_t* cPtr = cVector;
    const lv_32fc_t* aPtr = aVector;
    const lv_32fc_t* bPtr = bVector;
 
    for (unsigned int number = 0; number < num_points; number++) {
        *cPtr++ = (*aPtr++) / (*bPtr++);
    }
}
#endif /* LV_HAVE_GENERIC */
 
 
#ifdef LV_HAVE_SSE3
#include <pmmintrin.h>
#include <volk/volk_sse3_intrinsics.h>
 
static inline void volk_32fc_x2_divide_32fc_u_sse3(lv_32fc_t* cVector,
                                                   const lv_32fc_t* numeratorVector,
                                                   const lv_32fc_t* denumeratorVector,
                                                   unsigned int num_points)
{
    /*
     * we'll do the "classical"
     *  a      a b*
     * --- = -------
     *  b     |b|^2
     * */
    unsigned int number = 0;
    const unsigned int quarterPoints = num_points / 4;
 
    __m128 num01, num23, den01, den23, norm, result;
    lv_32fc_t* c = cVector;
    const lv_32fc_t* a = numeratorVector;
    const lv_32fc_t* b = denumeratorVector;
 
    for (; number < quarterPoints; number++) {
        num01 = _mm_loadu_ps((float*)a);                  // first pair
        den01 = _mm_loadu_ps((float*)b);                  // first pair
        num01 = _mm_complexconjugatemul_ps(num01, den01); // a conj(b)
        a += 2;
        b += 2;
 
        num23 = _mm_loadu_ps((float*)a);                  // second pair
        den23 = _mm_loadu_ps((float*)b);                  // second pair
        num23 = _mm_complexconjugatemul_ps(num23, den23); // a conj(b)
        a += 2;
        b += 2;
 
        norm = _mm_magnitudesquared_ps_sse3(den01, den23);
        den01 = _mm_unpacklo_ps(norm, norm);
        den23 = _mm_unpackhi_ps(norm, norm);
 
        result = _mm_div_ps(num01, den01);
        _mm_storeu_ps((float*)c, result); // Store the results back into the C container
        c += 2;
        result = _mm_div_ps(num23, den23);
        _mm_storeu_ps((float*)c, result); // Store the results back into the C container
        c += 2;
    }
 
    number *= 4;
    for (; number < num_points; number++) {
        *c = (*a) / (*b);
        a++;
        b++;
        c++;
    }
}
#endif /* LV_HAVE_SSE3 */
 
 
#ifdef LV_HAVE_AVX
#include <immintrin.h>
#include <volk/volk_avx_intrinsics.h>
 
static inline void volk_32fc_x2_divide_32fc_u_avx(lv_32fc_t* cVector,
                                                  const lv_32fc_t* numeratorVector,
                                                  const lv_32fc_t* denumeratorVector,
                                                  unsigned int num_points)
{
    /*
     * we'll do the "classical"
     *  a      a b*
     * --- = -------
     *  b     |b|^2
     * */
    unsigned int number = 0;
    const unsigned int quarterPoints = num_points / 4;
 
    __m256 num, denum, mul_conj, sq, mag_sq, mag_sq_un, div;
    lv_32fc_t* c = cVector;
    const lv_32fc_t* a = numeratorVector;
    const lv_32fc_t* b = denumeratorVector;
 
    for (; number < quarterPoints; number++) {
        num = _mm256_loadu_ps(
            (float*)a); // Load the ar + ai, br + bi ... as ar,ai,br,bi ...
        denum = _mm256_loadu_ps(
            (float*)b); // Load the cr + ci, dr + di ... as cr,ci,dr,di ...
        mul_conj = _mm256_complexconjugatemul_ps(num, denum);
        sq = _mm256_mul_ps(denum, denum); // Square the values
        mag_sq_un = _mm256_hadd_ps(
            sq, sq); // obtain the actual squared magnitude, although out of order
        mag_sq = _mm256_permute_ps(mag_sq_un, 0xd8); // I order them
        // best guide I found on using these functions:
        // https://software.intel.com/sites/landingpage/IntrinsicsGuide/#expand=2738,2059,2738,2738,3875,3874,3875,2738,3870
        div = _mm256_div_ps(mul_conj, mag_sq);
 
        _mm256_storeu_ps((float*)c, div); // Store the results back into the C container
 
        a += 4;
        b += 4;
        c += 4;
    }
 
    number = quarterPoints * 4;
 
    for (; number < num_points; number++) {
        *c++ = (*a++) / (*b++);
    }
}
#endif /* LV_HAVE_AVX */
 
 
#endif /* INCLUDED_volk_32fc_x2_divide_32fc_u_H */
 
 
#ifndef INCLUDED_volk_32fc_x2_divide_32fc_a_H
#define INCLUDED_volk_32fc_x2_divide_32fc_a_H
 
#include <float.h>
#include <inttypes.h>
#include <stdio.h>
#include <volk/volk_complex.h>
 
#ifdef LV_HAVE_SSE3
#include <pmmintrin.h>
#include <volk/volk_sse3_intrinsics.h>
 
static inline void volk_32fc_x2_divide_32fc_a_sse3(lv_32fc_t* cVector,
                                                   const lv_32fc_t* numeratorVector,
                                                   const lv_32fc_t* denumeratorVector,
                                                   unsigned int num_points)
{
    /*
     * we'll do the "classical"
     *  a      a b*
     * --- = -------
     *  b     |b|^2
     * */
    unsigned int number = 0;
    const unsigned int quarterPoints = num_points / 4;
 
    __m128 num01, num23, den01, den23, norm, result;
    lv_32fc_t* c = cVector;
    const lv_32fc_t* a = numeratorVector;
    const lv_32fc_t* b = denumeratorVector;
 
    for (; number < quarterPoints; number++) {
        num01 = _mm_load_ps((float*)a);                   // first pair
        den01 = _mm_load_ps((float*)b);                   // first pair
        num01 = _mm_complexconjugatemul_ps(num01, den01); // a conj(b)
        a += 2;
        b += 2;
 
        num23 = _mm_load_ps((float*)a);                   // second pair
        den23 = _mm_load_ps((float*)b);                   // second pair
        num23 = _mm_complexconjugatemul_ps(num23, den23); // a conj(b)
        a += 2;
        b += 2;
 
        norm = _mm_magnitudesquared_ps_sse3(den01, den23);
 
        den01 = _mm_unpacklo_ps(norm, norm); // select the lower floats twice
        den23 = _mm_unpackhi_ps(norm, norm); // select the upper floats twice
 
        result = _mm_div_ps(num01, den01);
        _mm_store_ps((float*)c, result); // Store the results back into the C container
        c += 2;
        result = _mm_div_ps(num23, den23);
        _mm_store_ps((float*)c, result); // Store the results back into the C container
        c += 2;
    }
 
    number *= 4;
    for (; number < num_points; number++) {
        *c = (*a) / (*b);
        a++;
        b++;
        c++;
    }
}
#endif /* LV_HAVE_SSE */
 
#ifdef LV_HAVE_AVX
#include <immintrin.h>
#include <volk/volk_avx_intrinsics.h>
 
static inline void volk_32fc_x2_divide_32fc_a_avx(lv_32fc_t* cVector,
                                                  const lv_32fc_t* numeratorVector,
                                                  const lv_32fc_t* denumeratorVector,
                                                  unsigned int num_points)
{
    /*
     * Guide to AVX intrisics:
     * https://software.intel.com/sites/landingpage/IntrinsicsGuide/#
     *
     * we'll do the "classical"
     *  a      a b*
     * --- = -------
     *  b     |b|^2
     *
     */
    lv_32fc_t* c = cVector;
    const lv_32fc_t* a = numeratorVector;
    const lv_32fc_t* b = denumeratorVector;
 
    const unsigned int eigthPoints = num_points / 8;
 
    __m256 num01, num23, denum01, denum23, complex_result, result0, result1;
 
    for (unsigned int number = 0; number < eigthPoints; number++) {
        // Load the ar + ai, br + bi ... as ar,ai,br,bi ...
        num01 = _mm256_load_ps((float*)a);
        denum01 = _mm256_load_ps((float*)b);
 
        num01 = _mm256_complexconjugatemul_ps(num01, denum01);
        a += 4;
        b += 4;
 
        num23 = _mm256_load_ps((float*)a);
        denum23 = _mm256_load_ps((float*)b);
        num23 = _mm256_complexconjugatemul_ps(num23, denum23);
        a += 4;
        b += 4;
 
        complex_result = _mm256_hadd_ps(_mm256_mul_ps(denum01, denum01),
                                        _mm256_mul_ps(denum23, denum23));
 
        denum01 = _mm256_shuffle_ps(complex_result, complex_result, 0x50);
        denum23 = _mm256_shuffle_ps(complex_result, complex_result, 0xfa);
 
        result0 = _mm256_div_ps(num01, denum01);
        result1 = _mm256_div_ps(num23, denum23);
 
        _mm256_store_ps((float*)c, result0);
        c += 4;
        _mm256_store_ps((float*)c, result1);
        c += 4;
    }
 
    volk_32fc_x2_divide_32fc_generic(c, a, b, num_points - eigthPoints * 8);
}
#endif /* LV_HAVE_AVX */
 
 
#ifdef LV_HAVE_NEON
#include <arm_neon.h>
 
static inline void volk_32fc_x2_divide_32fc_neon(lv_32fc_t* cVector,
                                                 const lv_32fc_t* aVector,
                                                 const lv_32fc_t* bVector,
                                                 unsigned int num_points)
{
    lv_32fc_t* cPtr = cVector;
    const lv_32fc_t* aPtr = aVector;
    const lv_32fc_t* bPtr = bVector;
 
    float32x4x2_t aVal, bVal, cVal;
    float32x4_t bAbs, bAbsInv;
 
    const unsigned int quarterPoints = num_points / 4;
    unsigned int number = 0;
    for (; number < quarterPoints; number++) {
        aVal = vld2q_f32((const float*)(aPtr));
        bVal = vld2q_f32((const float*)(bPtr));
        aPtr += 4;
        bPtr += 4;
        __VOLK_PREFETCH(aPtr + 4);
        __VOLK_PREFETCH(bPtr + 4);
 
        bAbs = vmulq_f32(bVal.val[0], bVal.val[0]);
        bAbs = vmlaq_f32(bAbs, bVal.val[1], bVal.val[1]);
 
        bAbsInv = vrecpeq_f32(bAbs);
        bAbsInv = vmulq_f32(bAbsInv, vrecpsq_f32(bAbsInv, bAbs));
        bAbsInv = vmulq_f32(bAbsInv, vrecpsq_f32(bAbsInv, bAbs));
 
        cVal.val[0] = vmulq_f32(aVal.val[0], bVal.val[0]);
        cVal.val[0] = vmlaq_f32(cVal.val[0], aVal.val[1], bVal.val[1]);
        cVal.val[0] = vmulq_f32(cVal.val[0], bAbsInv);
 
        cVal.val[1] = vmulq_f32(aVal.val[1], bVal.val[0]);
        cVal.val[1] = vmlsq_f32(cVal.val[1], aVal.val[0], bVal.val[1]);
        cVal.val[1] = vmulq_f32(cVal.val[1], bAbsInv);
 
        vst2q_f32((float*)(cPtr), cVal);
        cPtr += 4;
    }
 
    for (number = quarterPoints * 4; number < num_points; number++) {
        *cPtr++ = (*aPtr++) / (*bPtr++);
    }
}
#endif /* LV_HAVE_NEON */
 
 
#endif /* INCLUDED_volk_32fc_x2_divide_32fc_a_H */