volk_32f_x2_add_32f.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_32f_x2_add_32f_u_H
#define INCLUDED_volk_32f_x2_add_32f_u_H
 
#include <inttypes.h>
#include <stdio.h>
 
#ifdef LV_HAVE_AVX512F
#include <immintrin.h>
 
static inline void volk_32f_x2_add_32f_u_avx512f(float* cVector,
                                                 const float* aVector,
                                                 const float* bVector,
                                                 unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int sixteenthPoints = num_points / 16;
 
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
 
    __m512 aVal, bVal, cVal;
    for (; number < sixteenthPoints; number++) {
 
        aVal = _mm512_loadu_ps(aPtr);
        bVal = _mm512_loadu_ps(bPtr);
 
        cVal = _mm512_add_ps(aVal, bVal);
 
        _mm512_storeu_ps(cPtr, cVal); // Store the results back into the C container
 
        aPtr += 16;
        bPtr += 16;
        cPtr += 16;
    }
 
    number = sixteenthPoints * 16;
 
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
 
#endif /* LV_HAVE_AVX512F */
 
 
#ifdef LV_HAVE_AVX
#include <immintrin.h>
 
static inline void volk_32f_x2_add_32f_u_avx(float* cVector,
                                             const float* aVector,
                                             const float* bVector,
                                             unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int eighthPoints = num_points / 8;
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
    __m256 aVal, bVal, cVal;
    for (; number < eighthPoints; number++) {
 
        aVal = _mm256_loadu_ps(aPtr);
        bVal = _mm256_loadu_ps(bPtr);
 
        cVal = _mm256_add_ps(aVal, bVal);
 
        _mm256_storeu_ps(cPtr, cVal); // Store the results back into the C container
 
        aPtr += 8;
        bPtr += 8;
        cPtr += 8;
    }
 
    number = eighthPoints * 8;
 
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
#endif /* LV_HAVE_AVX */
 
 
#ifdef LV_HAVE_SSE
#include <xmmintrin.h>
 
static inline void volk_32f_x2_add_32f_u_sse(float* cVector,
                                             const float* aVector,
                                             const float* bVector,
                                             unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int quarterPoints = num_points / 4;
 
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
 
    __m128 aVal, bVal, cVal;
    for (; number < quarterPoints; number++) {
 
        aVal = _mm_loadu_ps(aPtr);
        bVal = _mm_loadu_ps(bPtr);
 
        cVal = _mm_add_ps(aVal, bVal);
 
        _mm_storeu_ps(cPtr, cVal); // Store the results back into the C container
 
        aPtr += 4;
        bPtr += 4;
        cPtr += 4;
    }
 
    number = quarterPoints * 4;
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
#endif /* LV_HAVE_SSE */
 
 
#ifdef LV_HAVE_GENERIC
 
static inline void volk_32f_x2_add_32f_generic(float* cVector,
                                               const float* aVector,
                                               const float* bVector,
                                               unsigned int num_points)
{
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
    unsigned int number = 0;
 
    for (number = 0; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
#endif /* LV_HAVE_GENERIC */
 
 
#endif /* INCLUDED_volk_32f_x2_add_32f_u_H */
#ifndef INCLUDED_volk_32f_x2_add_32f_a_H
#define INCLUDED_volk_32f_x2_add_32f_a_H
 
#include <inttypes.h>
#include <stdio.h>
 
#ifdef LV_HAVE_AVX512F
#include <immintrin.h>
 
static inline void volk_32f_x2_add_32f_a_avx512f(float* cVector,
                                                 const float* aVector,
                                                 const float* bVector,
                                                 unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int sixteenthPoints = num_points / 16;
 
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
 
    __m512 aVal, bVal, cVal;
    for (; number < sixteenthPoints; number++) {
 
        aVal = _mm512_load_ps(aPtr);
        bVal = _mm512_load_ps(bPtr);
 
        cVal = _mm512_add_ps(aVal, bVal);
 
        _mm512_store_ps(cPtr, cVal); // Store the results back into the C container
 
        aPtr += 16;
        bPtr += 16;
        cPtr += 16;
    }
 
    number = sixteenthPoints * 16;
 
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
 
#endif /* LV_HAVE_AVX512F */
 
 
#ifdef LV_HAVE_AVX
#include <immintrin.h>
 
static inline void volk_32f_x2_add_32f_a_avx(float* cVector,
                                             const float* aVector,
                                             const float* bVector,
                                             unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int eighthPoints = num_points / 8;
 
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
 
    __m256 aVal, bVal, cVal;
    for (; number < eighthPoints; number++) {
 
        aVal = _mm256_load_ps(aPtr);
        bVal = _mm256_load_ps(bPtr);
 
        cVal = _mm256_add_ps(aVal, bVal);
 
        _mm256_store_ps(cPtr, cVal); // Store the results back into the C container
 
        aPtr += 8;
        bPtr += 8;
        cPtr += 8;
    }
 
    number = eighthPoints * 8;
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
#endif /* LV_HAVE_AVX */
 
#ifdef LV_HAVE_SSE
#include <xmmintrin.h>
 
static inline void volk_32f_x2_add_32f_a_sse(float* cVector,
                                             const float* aVector,
                                             const float* bVector,
                                             unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int quarterPoints = num_points / 4;
 
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
 
    __m128 aVal, bVal, cVal;
    for (; number < quarterPoints; number++) {
        aVal = _mm_load_ps(aPtr);
        bVal = _mm_load_ps(bPtr);
 
        cVal = _mm_add_ps(aVal, bVal);
 
        _mm_store_ps(cPtr, cVal); // Store the results back into the C container
 
        aPtr += 4;
        bPtr += 4;
        cPtr += 4;
    }
 
    number = quarterPoints * 4;
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
#endif /* LV_HAVE_SSE */
 
 
#ifdef LV_HAVE_NEON
#include <arm_neon.h>
 
static inline void volk_32f_x2_add_32f_u_neon(float* cVector,
                                              const float* aVector,
                                              const float* bVector,
                                              unsigned int num_points)
{
    unsigned int number = 0;
    const unsigned int quarterPoints = num_points / 4;
 
    float* cPtr = cVector;
    const float* aPtr = aVector;
    const float* bPtr = bVector;
    float32x4_t aVal, bVal, cVal;
    for (number = 0; number < quarterPoints; number++) {
        // Load in to NEON registers
        aVal = vld1q_f32(aPtr);
        bVal = vld1q_f32(bPtr);
        __VOLK_PREFETCH(aPtr + 4);
        __VOLK_PREFETCH(bPtr + 4);
 
        // vector add
        cVal = vaddq_f32(aVal, bVal);
        // Store the results back into the C container
        vst1q_f32(cPtr, cVal);
 
        aPtr += 4; // q uses quadwords, 4 floats per vadd
        bPtr += 4;
        cPtr += 4;
    }
 
    number = quarterPoints * 4; // should be = num_points
    for (; number < num_points; number++) {
        *cPtr++ = (*aPtr++) + (*bPtr++);
    }
}
 
#endif /* LV_HAVE_NEON */
 
#ifdef LV_HAVE_NEONV7
extern void volk_32f_x2_add_32f_a_neonasm(float* cVector,
                                          const float* aVector,
                                          const float* bVector,
                                          unsigned int num_points);
#endif /* LV_HAVE_NEONV7 */
 
#ifdef LV_HAVE_NEONV7
extern void volk_32f_x2_add_32f_a_neonpipeline(float* cVector,
                                               const float* aVector,
                                               const float* bVector,
                                               unsigned int num_points);
#endif /* LV_HAVE_NEONV7 */
 
 
#ifdef LV_HAVE_NEONV8
#include <arm_neon.h>
 
static inline void volk_32f_x2_add_32f_neonv8(float* cVector,
                                              const float* aVector,
                                              const float* bVector,
                                              unsigned int num_points)
{
    unsigned int n = num_points;
    float* c = cVector;
    const float* a = aVector;
    const float* b = bVector;
 
    /* Process 8 floats per iteration (2x unroll) */
    while (n >= 8) {
        float32x4_t a0 = vld1q_f32(a);
        float32x4_t a1 = vld1q_f32(a + 4);
        float32x4_t b0 = vld1q_f32(b);
        float32x4_t b1 = vld1q_f32(b + 4);
        __VOLK_PREFETCH(a + 16);
        __VOLK_PREFETCH(b + 16);
 
        vst1q_f32(c, vaddq_f32(a0, b0));
        vst1q_f32(c + 4, vaddq_f32(a1, b1));
 
        a += 8;
        b += 8;
        c += 8;
        n -= 8;
    }
 
    /* Process remaining 4 floats */
    if (n >= 4) {
        vst1q_f32(c, vaddq_f32(vld1q_f32(a), vld1q_f32(b)));
        a += 4;
        b += 4;
        c += 4;
        n -= 4;
    }
 
    /* Scalar tail */
    while (n > 0) {
        *c++ = *a++ + *b++;
        n--;
    }
}
 
#endif /* LV_HAVE_NEONV8 */
 
 
#ifdef LV_HAVE_ORC
 
extern void volk_32f_x2_add_32f_a_orc_impl(float* cVector,
                                           const float* aVector,
                                           const float* bVector,
                                           int num_points);
 
static inline void volk_32f_x2_add_32f_u_orc(float* cVector,
                                             const float* aVector,
                                             const float* bVector,
                                             unsigned int num_points)
{
    volk_32f_x2_add_32f_a_orc_impl(cVector, aVector, bVector, num_points);
}
 
#endif /* LV_HAVE_ORC */
 
#ifdef LV_HAVE_RVV
#include <riscv_vector.h>
 
static inline void volk_32f_x2_add_32f_rvv(float* cVector,
                                           const float* aVector,
                                           const float* bVector,
                                           unsigned int num_points)
{
    size_t n = num_points;
    for (size_t vl; n > 0; n -= vl, aVector += vl, bVector += vl, cVector += vl) {
        vl = __riscv_vsetvl_e32m8(n);
        vfloat32m8_t va = __riscv_vle32_v_f32m8(aVector, vl);
        vfloat32m8_t vb = __riscv_vle32_v_f32m8(bVector, vl);
        __riscv_vse32(cVector, __riscv_vfadd(va, vb, vl), vl);
    }
}
#endif /*LV_HAVE_RVV*/
 
#endif /* INCLUDED_volk_32f_x2_add_32f_a_H */