volk_32fc_index_min_32u.h

Go to the documentation of this file.

/* -*- c++ -*- */
/*
 * Copyright 2021 Free Software Foundation, Inc.
 *
 * This file is part of VOLK
 *
 * SPDX-License-Identifier: LGPL-3.0-or-later
 */
 
#ifndef INCLUDED_volk_32fc_index_min_32u_a_H
#define INCLUDED_volk_32fc_index_min_32u_a_H
 
#include <inttypes.h>
#include <stdio.h>
#include <volk/volk_common.h>
#include <volk/volk_complex.h>
 
#ifdef LV_HAVE_AVX2
#include <immintrin.h>
#include <volk/volk_avx2_intrinsics.h>
 
static inline void volk_32fc_index_min_32u_a_avx2_variant_0(uint32_t* target,
                                                            const lv_32fc_t* source,
                                                            uint32_t num_points)
{
    const __m256i indices_increment = _mm256_set1_epi32(8);
    /*
     * At the start of each loop iteration current_indices holds the indices of
     * the complex numbers loaded from memory. Explanation for odd order is given
     * in implementation of vector_32fc_index_min_variant0().
     */
    __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
 
    __m256 min_values = _mm256_set1_ps(FLT_MAX);
    __m256i min_indices = _mm256_setzero_si256();
 
    for (unsigned i = 0; i < num_points / 8u; ++i) {
        __m256 in0 = _mm256_load_ps((float*)source);
        __m256 in1 = _mm256_load_ps((float*)(source + 4));
        vector_32fc_index_min_variant0(
            in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
        source += 8;
    }
 
    // determine minimum value and index in the result of the vectorized loop
    __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
    __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
    _mm256_store_ps(min_values_buffer, min_values);
    _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
 
    float min = FLT_MAX;
    uint32_t index = 0;
    for (unsigned i = 0; i < 8; i++) {
        if (min_values_buffer[i] < min) {
            min = min_values_buffer[i];
            index = min_indices_buffer[i];
        }
    }
 
    // handle tail not processed by the vectorized loop
    for (unsigned i = num_points & (~7u); i < num_points; ++i) {
        const float abs_squared =
            lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
        if (abs_squared < min) {
            min = abs_squared;
            index = i;
        }
        ++source;
    }
 
    *target = index;
}
 
#endif /*LV_HAVE_AVX2*/
 
#ifdef LV_HAVE_AVX2
#include <immintrin.h>
#include <volk/volk_avx2_intrinsics.h>
 
static inline void volk_32fc_index_min_32u_a_avx2_variant_1(uint32_t* target,
                                                            const lv_32fc_t* source,
                                                            uint32_t num_points)
{
    const __m256i indices_increment = _mm256_set1_epi32(8);
    /*
     * At the start of each loop iteration current_indices holds the indices of
     * the complex numbers loaded from memory. Explanation for odd order is given
     * in implementation of vector_32fc_index_min_variant0().
     */
    __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
 
    __m256 min_values = _mm256_set1_ps(FLT_MAX);
    __m256i min_indices = _mm256_setzero_si256();
 
    for (unsigned i = 0; i < num_points / 8u; ++i) {
        __m256 in0 = _mm256_load_ps((float*)source);
        __m256 in1 = _mm256_load_ps((float*)(source + 4));
        vector_32fc_index_min_variant1(
            in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
        source += 8;
    }
 
    // determine minimum value and index in the result of the vectorized loop
    __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
    __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
    _mm256_store_ps(min_values_buffer, min_values);
    _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
 
    float min = FLT_MAX;
    uint32_t index = 0;
    for (unsigned i = 0; i < 8; i++) {
        if (min_values_buffer[i] < min) {
            min = min_values_buffer[i];
            index = min_indices_buffer[i];
        }
    }
 
    // handle tail not processed by the vectorized loop
    for (unsigned i = num_points & (~7u); i < num_points; ++i) {
        const float abs_squared =
            lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
        if (abs_squared < min) {
            min = abs_squared;
            index = i;
        }
        ++source;
    }
 
    *target = index;
}
 
#endif /*LV_HAVE_AVX2*/
 
#ifdef LV_HAVE_SSE3
#include <pmmintrin.h>
#include <xmmintrin.h>
 
static inline void volk_32fc_index_min_32u_a_sse3(uint32_t* target,
                                                  const lv_32fc_t* source,
                                                  uint32_t num_points)
{
    union bit128 holderf;
    union bit128 holderi;
    float sq_dist = 0.0;
 
    union bit128 xmm5, xmm4;
    __m128 xmm1, xmm2, xmm3;
    __m128i xmm8, xmm11, xmm12, xmm9, xmm10;
 
    xmm5.int_vec = _mm_setzero_si128();
    xmm4.int_vec = _mm_setzero_si128();
    holderf.int_vec = _mm_setzero_si128();
    holderi.int_vec = _mm_setzero_si128();
 
    xmm8 = _mm_setr_epi32(0, 1, 2, 3);
    xmm9 = _mm_setzero_si128();
    xmm10 = _mm_setr_epi32(4, 4, 4, 4);
    xmm3 = _mm_set_ps1(FLT_MAX);
 
    int bound = num_points >> 2;
 
    for (int i = 0; i < bound; ++i) {
        xmm1 = _mm_load_ps((float*)source);
        xmm2 = _mm_load_ps((float*)&source[2]);
 
        source += 4;
 
        xmm1 = _mm_mul_ps(xmm1, xmm1);
        xmm2 = _mm_mul_ps(xmm2, xmm2);
 
        xmm1 = _mm_hadd_ps(xmm1, xmm2);
 
        xmm3 = _mm_min_ps(xmm1, xmm3);
 
        xmm4.float_vec = _mm_cmpgt_ps(xmm1, xmm3);
        xmm5.float_vec = _mm_cmpeq_ps(xmm1, xmm3);
 
        xmm11 = _mm_and_si128(xmm8, xmm5.int_vec);
        xmm12 = _mm_and_si128(xmm9, xmm4.int_vec);
 
        xmm9 = _mm_add_epi32(xmm11, xmm12);
 
        xmm8 = _mm_add_epi32(xmm8, xmm10);
    }
 
    if (num_points >> 1 & 1) {
        xmm2 = _mm_load_ps((float*)source);
 
        xmm1 = _mm_movelh_ps(bit128_p(&xmm8)->float_vec, bit128_p(&xmm8)->float_vec);
        xmm8 = bit128_p(&xmm1)->int_vec;
 
        xmm2 = _mm_mul_ps(xmm2, xmm2);
 
        source += 2;
 
        xmm1 = _mm_hadd_ps(xmm2, xmm2);
 
        xmm3 = _mm_min_ps(xmm1, xmm3);
 
        xmm10 = _mm_setr_epi32(2, 2, 2, 2);
 
        xmm4.float_vec = _mm_cmpgt_ps(xmm1, xmm3);
        xmm5.float_vec = _mm_cmpeq_ps(xmm1, xmm3);
 
        xmm11 = _mm_and_si128(xmm8, xmm5.int_vec);
        xmm12 = _mm_and_si128(xmm9, xmm4.int_vec);
 
        xmm9 = _mm_add_epi32(xmm11, xmm12);
 
        xmm8 = _mm_add_epi32(xmm8, xmm10);
    }
 
    if (num_points & 1) {
        sq_dist = lv_creal(source[0]) * lv_creal(source[0]) +
                  lv_cimag(source[0]) * lv_cimag(source[0]);
 
        xmm2 = _mm_load1_ps(&sq_dist);
 
        xmm1 = xmm3;
 
        xmm3 = _mm_min_ss(xmm3, xmm2);
 
        xmm4.float_vec = _mm_cmpgt_ps(xmm1, xmm3);
        xmm5.float_vec = _mm_cmpeq_ps(xmm1, xmm3);
 
        xmm8 = _mm_shuffle_epi32(xmm8, 0x00);
 
        xmm11 = _mm_and_si128(xmm8, xmm4.int_vec);
        xmm12 = _mm_and_si128(xmm9, xmm5.int_vec);
 
        xmm9 = _mm_add_epi32(xmm11, xmm12);
    }
 
    _mm_store_ps((float*)&(holderf.f), xmm3);
    _mm_store_si128(&(holderi.int_vec), xmm9);
 
    target[0] = holderi.i[0];
    sq_dist = holderf.f[0];
    target[0] = (holderf.f[1] < sq_dist) ? holderi.i[1] : target[0];
    sq_dist = (holderf.f[1] < sq_dist) ? holderf.f[1] : sq_dist;
    target[0] = (holderf.f[2] < sq_dist) ? holderi.i[2] : target[0];
    sq_dist = (holderf.f[2] < sq_dist) ? holderf.f[2] : sq_dist;
    target[0] = (holderf.f[3] < sq_dist) ? holderi.i[3] : target[0];
    sq_dist = (holderf.f[3] < sq_dist) ? holderf.f[3] : sq_dist;
}
 
#endif /*LV_HAVE_SSE3*/
 
#ifdef LV_HAVE_GENERIC
static inline void volk_32fc_index_min_32u_generic(uint32_t* target,
                                                   const lv_32fc_t* source,
                                                   uint32_t num_points)
{
    float sq_dist = 0.0;
    float min = FLT_MAX;
    uint32_t index = 0;
 
    for (uint32_t i = 0; i < num_points; ++i) {
        sq_dist = lv_creal(source[i]) * lv_creal(source[i]) +
                  lv_cimag(source[i]) * lv_cimag(source[i]);
 
        if (sq_dist < min) {
            index = i;
            min = sq_dist;
        }
    }
    target[0] = index;
}
 
#endif /*LV_HAVE_GENERIC*/
 
#endif /*INCLUDED_volk_32fc_index_min_32u_a_H*/
 
#ifndef INCLUDED_volk_32fc_index_min_32u_u_H
#define INCLUDED_volk_32fc_index_min_32u_u_H
 
#include <inttypes.h>
#include <stdio.h>
#include <volk/volk_common.h>
#include <volk/volk_complex.h>
 
#ifdef LV_HAVE_AVX2
#include <immintrin.h>
#include <volk/volk_avx2_intrinsics.h>
 
static inline void volk_32fc_index_min_32u_u_avx2_variant_0(uint32_t* target,
                                                            const lv_32fc_t* source,
                                                            uint32_t num_points)
{
    const __m256i indices_increment = _mm256_set1_epi32(8);
    /*
     * At the start of each loop iteration current_indices holds the indices of
     * the complex numbers loaded from memory. Explanation for odd order is given
     * in implementation of vector_32fc_index_min_variant0().
     */
    __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
 
    __m256 min_values = _mm256_set1_ps(FLT_MAX);
    __m256i min_indices = _mm256_setzero_si256();
 
    for (unsigned i = 0; i < num_points / 8u; ++i) {
        __m256 in0 = _mm256_loadu_ps((float*)source);
        __m256 in1 = _mm256_loadu_ps((float*)(source + 4));
        vector_32fc_index_min_variant0(
            in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
        source += 8;
    }
 
    // determine minimum value and index in the result of the vectorized loop
    __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
    __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
    _mm256_store_ps(min_values_buffer, min_values);
    _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
 
    float min = FLT_MAX;
    uint32_t index = 0;
    for (unsigned i = 0; i < 8; i++) {
        if (min_values_buffer[i] < min) {
            min = min_values_buffer[i];
            index = min_indices_buffer[i];
        }
    }
 
    // handle tail not processed by the vectorized loop
    for (unsigned i = num_points & (~7u); i < num_points; ++i) {
        const float abs_squared =
            lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
        if (abs_squared < min) {
            min = abs_squared;
            index = i;
        }
        ++source;
    }
 
    *target = index;
}
 
#endif /*LV_HAVE_AVX2*/
 
#ifdef LV_HAVE_AVX2
#include <immintrin.h>
#include <volk/volk_avx2_intrinsics.h>
 
static inline void volk_32fc_index_min_32u_u_avx2_variant_1(uint32_t* target,
                                                            const lv_32fc_t* source,
                                                            uint32_t num_points)
{
    const __m256i indices_increment = _mm256_set1_epi32(8);
    /*
     * At the start of each loop iteration current_indices holds the indices of
     * the complex numbers loaded from memory. Explanation for odd order is given
     * in implementation of vector_32fc_index_min_variant0().
     */
    __m256i current_indices = _mm256_set_epi32(7, 6, 3, 2, 5, 4, 1, 0);
 
    __m256 min_values = _mm256_set1_ps(FLT_MAX);
    __m256i min_indices = _mm256_setzero_si256();
 
    for (unsigned i = 0; i < num_points / 8u; ++i) {
        __m256 in0 = _mm256_loadu_ps((float*)source);
        __m256 in1 = _mm256_loadu_ps((float*)(source + 4));
        vector_32fc_index_min_variant1(
            in0, in1, &min_values, &min_indices, &current_indices, indices_increment);
        source += 8;
    }
 
    // determine minimum value and index in the result of the vectorized loop
    __VOLK_ATTR_ALIGNED(32) float min_values_buffer[8];
    __VOLK_ATTR_ALIGNED(32) uint32_t min_indices_buffer[8];
    _mm256_store_ps(min_values_buffer, min_values);
    _mm256_store_si256((__m256i*)min_indices_buffer, min_indices);
 
    float min = FLT_MAX;
    uint32_t index = 0;
    for (unsigned i = 0; i < 8; i++) {
        if (min_values_buffer[i] < min) {
            min = min_values_buffer[i];
            index = min_indices_buffer[i];
        }
    }
 
    // handle tail not processed by the vectorized loop
    for (unsigned i = num_points & (~7u); i < num_points; ++i) {
        const float abs_squared =
            lv_creal(*source) * lv_creal(*source) + lv_cimag(*source) * lv_cimag(*source);
        if (abs_squared < min) {
            min = abs_squared;
            index = i;
        }
        ++source;
    }
 
    *target = index;
}
 
#endif /*LV_HAVE_AVX2*/
 
#ifdef LV_HAVE_NEON
#include <arm_neon.h>
#include <volk/volk_neon_intrinsics.h>
 
static inline void volk_32fc_index_min_32u_neon(uint32_t* target,
                                                const lv_32fc_t* source,
                                                uint32_t num_points)
{
    const uint32_t quarter_points = num_points / 4;
    const lv_32fc_t* sourcePtr = source;
 
    uint32_t indices[4] = { 0, 1, 2, 3 };
    const uint32x4_t vec_indices_incr = vdupq_n_u32(4);
    uint32x4_t vec_indices = vld1q_u32(indices);
    uint32x4_t vec_min_indices = vec_indices;
 
    if (num_points) {
        float min = FLT_MAX;
        uint32_t index = 0;
 
        float32x4_t vec_min = vdupq_n_f32(FLT_MAX);
 
        for (uint32_t number = 0; number < quarter_points; number++) {
            // Load complex and compute magnitude squared
            const float32x4_t vec_mag2 =
                _vmagnitudesquaredq_f32(vld2q_f32((float*)sourcePtr));
            __VOLK_PREFETCH(sourcePtr += 4);
            // a < b?
            const uint32x4_t lt_mask = vcltq_f32(vec_mag2, vec_min);
            vec_min = vbslq_f32(lt_mask, vec_mag2, vec_min);
            vec_min_indices = vbslq_u32(lt_mask, vec_indices, vec_min_indices);
            vec_indices = vaddq_u32(vec_indices, vec_indices_incr);
        }
        uint32_t tmp_min_indices[4];
        float tmp_min[4];
        vst1q_u32(tmp_min_indices, vec_min_indices);
        vst1q_f32(tmp_min, vec_min);
 
        for (int i = 0; i < 4; i++) {
            if (tmp_min[i] < min) {
                min = tmp_min[i];
                index = tmp_min_indices[i];
            }
        }
 
        // Deal with the rest
        for (uint32_t number = quarter_points * 4; number < num_points; number++) {
            const float re = lv_creal(*sourcePtr);
            const float im = lv_cimag(*sourcePtr);
            const float sq_dist = re * re + im * im;
            if (sq_dist < min) {
                min = sq_dist;
                index = number;
            }
            sourcePtr++;
        }
        *target = index;
    }
}
 
#endif /*LV_HAVE_NEON*/
 
#endif /*INCLUDED_volk_32fc_index_min_32u_u_H*/