runner_doiact_vec.c

/*******************************************************************************
 * This file is part of SWIFT.
 * Copyright (c) 2016 James Willis (james.s.willis@durham.ac.uk)
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as published
 * by the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 ******************************************************************************/

/* Config parameters. */
#include "../config.h"

/* This object's header. */
#include "runner_doiact_vec.h"

/* Local headers. */
#include "active.h"

#if defined(WITH_VECTORIZATION) && defined(GADGET2_SPH)

static const vector kernel_gamma2_vec = FILL_VEC(kernel_gamma2);

/**
 * @brief Compute the vector remainder interactions from the secondary cache.
 *
 * @param int_cache (return) secondary #cache of interactions between two
 * particles.
 * @param icount Interaction count.
 * @param v_rhoSum (return) #vector holding the cumulative sum of the density
 * update on pi.
 * @param v_rho_dhSum (return) #vector holding the cumulative sum of the density
 * gradient update on pi.
 * @param v_wcountSum (return) #vector holding the cumulative sum of the wcount
 * update on pi.
 * @param v_wcount_dhSum (return) #vector holding the cumulative sum of the
 * wcount
 * gradient update on pi.
 * @param v_div_vSum (return) #vector holding the cumulative sum of the
 * divergence
 * update on pi.
 * @param v_curlvxSum (return) #vector holding the cumulative sum of the curl of
 * vx update on pi.
 * @param v_curlvySum (return) #vector holding the cumulative sum of the curl of
 * vy update on pi.
 * @param v_curlvzSum (return) #vector holding the cumulative sum of the curl of
 * vz update on pi.
 * @param v_hi_inv #vector of 1/h for pi.
 * @param v_vix #vector of x velocity of pi.
 * @param v_viy #vector of y velocity of pi.
 * @param v_viz #vector of z velocity of pi.
 * @param icount_align (return) Interaction count after the remainder
 * interactions have been performed, should be a multiple of the vector length.
 */
__attribute__((always_inline)) INLINE static void calcRemInteractions(
    struct c2_cache *const int_cache, const int icount, vector *v_rhoSum,
    vector *v_rho_dhSum, vector *v_wcountSum, vector *v_wcount_dhSum,
    vector *v_div_vSum, vector *v_curlvxSum, vector *v_curlvySum,
    vector *v_curlvzSum, vector v_hi_inv, vector v_vix, vector v_viy,
    vector v_viz, int *icount_align) {

  mask_t int_mask, int_mask2;

  /* Work out the number of remainder interactions and pad secondary cache. */
  *icount_align = icount;
  int rem = icount % (NUM_VEC_PROC * VEC_SIZE);
  if (rem != 0) {
    int pad = (NUM_VEC_PROC * VEC_SIZE) - rem;
    *icount_align += pad;

    /* Initialise masks to true. */
    vec_init_mask_true(int_mask);
    vec_init_mask_true(int_mask2);

    /* Pad secondary cache so that there are no contributions in the interaction
     * function. */
    for (int i = icount; i < *icount_align; i++) {
      int_cache->mq[i] = 0.f;
      int_cache->r2q[i] = 1.f;
      int_cache->dxq[i] = 0.f;
      int_cache->dyq[i] = 0.f;
      int_cache->dzq[i] = 0.f;
      int_cache->vxq[i] = 0.f;
      int_cache->vyq[i] = 0.f;
      int_cache->vzq[i] = 0.f;
    }

    /* Zero parts of mask that represent the padded values.*/
    if (pad < VEC_SIZE) {
      vec_pad_mask(int_mask2, pad);
    } else {
      vec_pad_mask(int_mask, VEC_SIZE - rem);
      vec_zero_mask(int_mask2);
    }

    /* Perform remainder interaction and remove remainder from aligned
     * interaction count. */
    *icount_align = icount - rem;
    runner_iact_nonsym_2_vec_density(
        &int_cache->r2q[*icount_align], &int_cache->dxq[*icount_align],
        &int_cache->dyq[*icount_align], &int_cache->dzq[*icount_align],
        v_hi_inv, v_vix, v_viy, v_viz, &int_cache->vxq[*icount_align],
        &int_cache->vyq[*icount_align], &int_cache->vzq[*icount_align],
        &int_cache->mq[*icount_align], v_rhoSum, v_rho_dhSum, v_wcountSum,
        v_wcount_dhSum, v_div_vSum, v_curlvxSum, v_curlvySum, v_curlvzSum,
        int_mask, int_mask2, 1);
  }
}

/**
 * @brief Left-packs the values needed by an interaction into the secondary
 * cache (Supports AVX, AVX2 and AVX512 instruction sets).
 *
 * @param mask Contains which particles need to interact.
 * @param pjd Index of the particle to store into.
 * @param v_r2 #vector of the separation between two particles squared.
 * @param v_dx #vector of the x separation between two particles.
 * @param v_dy #vector of the y separation between two particles.
 * @param v_dz #vector of the z separation between two particles.
 * @param cell_cache #cache of all particles in the cell.
 * @param int_cache (return) secondary #cache of interactions between two
 * particles.
 * @param icount Interaction count.
 * @param v_rhoSum #vector holding the cumulative sum of the density update on
 * pi.
 * @param v_rho_dhSum #vector holding the cumulative sum of the density gradient
 * update on pi.
 * @param v_wcountSum #vector holding the cumulative sum of the wcount update on
 * pi.
 * @param v_wcount_dhSum #vector holding the cumulative sum of the wcount
 * gradient
 * update on pi.
 * @param v_div_vSum #vector holding the cumulative sum of the divergence update
 * on pi.
 * @param v_curlvxSum #vector holding the cumulative sum of the curl of vx
 * update
 * on pi.
 * @param v_curlvySum #vector holding the cumulative sum of the curl of vy
 * update
 * on pi.
 * @param v_curlvzSum #vector holding the cumulative sum of the curl of vz
 * update
 * on pi.
 * @param v_hi_inv #vector of 1/h for pi.
 * @param v_vix #vector of x velocity of pi.
 * @param v_viy #vector of y velocity of pi.
 * @param v_viz #vector of z velocity of pi.
 */
__attribute__((always_inline)) INLINE static void storeInteractions(
    const int mask, const int pjd, vector *v_r2, vector *v_dx, vector *v_dy,
    vector *v_dz, const struct cache *const cell_cache,
    struct c2_cache *const int_cache, int *icount, vector *v_rhoSum,
    vector *v_rho_dhSum, vector *v_wcountSum, vector *v_wcount_dhSum,
    vector *v_div_vSum, vector *v_curlvxSum, vector *v_curlvySum,
    vector *v_curlvzSum, vector v_hi_inv, vector v_vix, vector v_viy,
    vector v_viz) {

/* Left-pack values needed into the secondary cache using the interaction mask.
 */
#if defined(HAVE_AVX2) || defined(HAVE_AVX512_F)
  mask_t packed_mask;
  VEC_FORM_PACKED_MASK(mask, packed_mask);

  VEC_LEFT_PACK(v_r2->v, packed_mask, &int_cache->r2q[*icount]);
  VEC_LEFT_PACK(v_dx->v, packed_mask, &int_cache->dxq[*icount]);
  VEC_LEFT_PACK(v_dy->v, packed_mask, &int_cache->dyq[*icount]);
  VEC_LEFT_PACK(v_dz->v, packed_mask, &int_cache->dzq[*icount]);
  VEC_LEFT_PACK(vec_load(&cell_cache->m[pjd]), packed_mask,
                &int_cache->mq[*icount]);
  VEC_LEFT_PACK(vec_load(&cell_cache->vx[pjd]), packed_mask,
                &int_cache->vxq[*icount]);
  VEC_LEFT_PACK(vec_load(&cell_cache->vy[pjd]), packed_mask,
                &int_cache->vyq[*icount]);
  VEC_LEFT_PACK(vec_load(&cell_cache->vz[pjd]), packed_mask,
                &int_cache->vzq[*icount]);

  /* Increment interaction count by number of bits set in mask. */
  (*icount) += __builtin_popcount(mask);
#else
  /* Quicker to do it serially in AVX rather than use intrinsics. */
  for (int bit_index = 0; bit_index < VEC_SIZE; bit_index++) {
    if (mask & (1 << bit_index)) {
      /* Add this interaction to the queue. */
      int_cache->r2q[*icount] = v_r2->f[bit_index];
      int_cache->dxq[*icount] = v_dx->f[bit_index];
      int_cache->dyq[*icount] = v_dy->f[bit_index];
      int_cache->dzq[*icount] = v_dz->f[bit_index];
      int_cache->mq[*icount] = cell_cache->m[pjd + bit_index];
      int_cache->vxq[*icount] = cell_cache->vx[pjd + bit_index];
      int_cache->vyq[*icount] = cell_cache->vy[pjd + bit_index];
      int_cache->vzq[*icount] = cell_cache->vz[pjd + bit_index];

      (*icount)++;
    }
  }

#endif /* defined(HAVE_AVX2) || defined(HAVE_AVX512_F) */

  /* Flush the c2 cache if it has reached capacity. */
  if (*icount >= (C2_CACHE_SIZE - (NUM_VEC_PROC * VEC_SIZE))) {

    int icount_align = *icount;

    /* Peform remainder interactions. */
    calcRemInteractions(int_cache, *icount, v_rhoSum, v_rho_dhSum, v_wcountSum,
                        v_wcount_dhSum, v_div_vSum, v_curlvxSum, v_curlvySum,
                        v_curlvzSum, v_hi_inv, v_vix, v_viy, v_viz,
                        &icount_align);

    mask_t int_mask, int_mask2;
    vec_init_mask_true(int_mask);
    vec_init_mask_true(int_mask2);

    /* Perform interactions. */
    for (int j = 0; j < icount_align; j += (NUM_VEC_PROC * VEC_SIZE)) {
      runner_iact_nonsym_2_vec_density(
          &int_cache->r2q[j], &int_cache->dxq[j], &int_cache->dyq[j],
          &int_cache->dzq[j], v_hi_inv, v_vix, v_viy, v_viz, &int_cache->vxq[j],
          &int_cache->vyq[j], &int_cache->vzq[j], &int_cache->mq[j], v_rhoSum,
          v_rho_dhSum, v_wcountSum, v_wcount_dhSum, v_div_vSum, v_curlvxSum,
          v_curlvySum, v_curlvzSum, int_mask, int_mask2, 0);
    }

    /* Reset interaction count. */
    *icount = 0;
  }
}

/**
 * @brief Populates the arrays max_index_i and max_index_j with the maximum
 * indices of
 * particles into their neighbouring cells. Also finds the first pi that
 * interacts with any particle in cj and the last pj that interacts with any
 * particle in ci.
 *
 * @param ci #cell pointer to ci
 * @param cj #cell pointer to cj
 * @param sort_i #entry array for particle distance in ci
 * @param sort_j #entry array for particle distance in cj
 * @param dx_max maximum particle movement allowed in cell
 * @param rshift cutoff shift
 * @param hi_max Maximal smoothing length in cell ci
 * @param hj_max Maximal smoothing length in cell cj
 * @param di_max Maximal position on the axis that can interact in cell ci
 * @param dj_min Minimal position on the axis that can interact in cell ci
 * @param max_index_i array to hold the maximum distances of pi particles into
 * #cell cj
 * @param max_index_j array to hold the maximum distances of pj particles into
 * #cell cj
 * @param init_pi first pi to interact with a pj particle
 * @param init_pj last pj to interact with a pi particle
 * @param max_active_bin The largest time-bin active during this step.
 */
__attribute__((always_inline)) INLINE static void populate_max_index_no_cache(
    const struct cell *ci, const struct cell *cj,
    const struct entry *restrict sort_i, const struct entry *restrict sort_j,
    const float dx_max, const float rshift, const double hi_max,
    const double hj_max, const double di_max, const double dj_min,
    int *max_index_i, int *max_index_j, int *init_pi, int *init_pj,
    const timebin_t max_active_bin) {

  const struct part *restrict parts_i = ci->parts;
  const struct part *restrict parts_j = cj->parts;

  int first_pi = 0, last_pj = cj->count - 1;
  int temp;

  /* Find the leftmost active particle in cell i that interacts with any
   * particle in cell j. */
  first_pi = ci->count;
  int active_id = first_pi - 1;
  while (first_pi > 0 && sort_i[first_pi - 1].d + dx_max + hi_max > dj_min) {
    first_pi--;
    /* Store the index of the particle if it is active. */
    if (part_is_active_no_debug(&parts_i[sort_i[first_pi].i], max_active_bin))
      active_id = first_pi;
  }

  /* Set the first active pi in range of any particle in cell j. */
  first_pi = active_id;

  /* Find the maximum index into cell j for each particle in range in cell i. */
  if (first_pi < ci->count) {

    /* Start from the first particle in cell j. */
    temp = 0;

    const struct part *pi = &parts_i[sort_i[first_pi].i];
    const float first_di =
        sort_i[first_pi].d + pi->h * kernel_gamma + dx_max - rshift;

    /* Loop through particles in cell j until they are not in range of pi.
     * Make sure that temp stays between 0 and cj->count - 1.*/
    while (temp < cj->count - 1 && first_di > sort_j[temp].d) temp++;

    max_index_i[first_pi] = temp;

    /* Populate max_index_i for remaining particles that are within range. */
    for (int i = first_pi + 1; i < ci->count; i++) {
      temp = max_index_i[i - 1];
      pi = &parts_i[sort_i[i].i];

      const float di = sort_i[i].d + pi->h * kernel_gamma + dx_max - rshift;

      /* Make sure that temp stays between 0 and cj->count - 1.*/
      while (temp < cj->count - 1 && di > sort_j[temp].d) temp++;

      max_index_i[i] = temp;
    }
  } else {
    /* Make sure that max index is set to first particle in cj.*/
    max_index_i[ci->count - 1] = 0;
  }

  /* Find the rightmost active particle in cell j that interacts with any
   * particle in cell i. */
  last_pj = -1;
  active_id = last_pj;
  while (last_pj < cj->count &&
         sort_j[last_pj + 1].d - hj_max - dx_max < di_max) {
    last_pj++;
    /* Store the index of the particle if it is active. */
    if (part_is_active_no_debug(&parts_j[sort_j[last_pj].i], max_active_bin))
      active_id = last_pj;
  }

  /* Set the last active pj in range of any particle in cell i. */
  last_pj = active_id;

  /* Find the maximum index into cell i for each particle in range in cell j. */
  if (last_pj >= 0) {

    /* Start from the last particle in cell i. */
    temp = ci->count - 1;

    const struct part *pj = &parts_j[sort_j[last_pj].i];
    const float last_dj =
        sort_j[last_pj].d - dx_max - pj->h * kernel_gamma + rshift;

    /* Loop through particles in cell i until they are not in range of pj. */
    while (temp > 0 && last_dj < sort_i[temp].d) temp--;

    max_index_j[last_pj] = temp;

    /* Populate max_index_j for remaining particles that are within range. */
    for (int i = last_pj - 1; i >= 0; i--) {
      temp = max_index_j[i + 1];
      pj = &parts_j[sort_j[i].i];
      const float dj = sort_j[i].d - dx_max - (pj->h * kernel_gamma) + rshift;

      while (temp > 0 && dj < sort_i[temp].d) temp--;

      max_index_j[i] = temp;
    }
  } else {
    /* Make sure that max index is set to last particle in ci.*/
    max_index_j[0] = ci->count - 1;
  }

  *init_pi = first_pi;
  *init_pj = last_pj;
}

/**
 * @brief Populates the arrays max_index_i and max_index_j with the maximum
 * indices of
 * particles into their neighbouring cells. Also finds the first pi that
 * interacts with any particle in cj and the last pj that interacts with any
 * particle in ci.
 *
 * @param ci #cell pointer to ci
 * @param cj #cell pointer to cj
 * @param sort_i #entry array for particle distance in ci
 * @param sort_j #entry array for particle distance in cj
 * @param dx_max maximum particle movement allowed in cell
 * @param rshift cutoff shift
 * @param hi_max_raw Maximal smoothing length in cell ci
 * @param hj_max_raw Maximal smoothing length in cell cj
 * @param h_max Maximal smoothing length in both cells scaled by kernel_gamma
 * @param di_max Maximal position on the axis that can interact in cell ci
 * @param dj_min Minimal position on the axis that can interact in cell ci
 * @param max_index_i array to hold the maximum distances of pi particles into
 * #cell cj
 * @param max_index_j array to hold the maximum distances of pj particles into
 * #cell cj
 * @param init_pi first pi to interact with a pj particle
 * @param init_pj last pj to interact with a pi particle
 * @param max_active_bin The largest time-bin active during this step.
 */
__attribute__((always_inline)) INLINE static void
populate_max_index_no_cache_force(const struct cell *ci, const struct cell *cj,
                                  const struct entry *restrict sort_i,
                                  const struct entry *restrict sort_j,
                                  const float dx_max, const float rshift,
                                  const double hi_max_raw,
                                  const double hj_max_raw, const double h_max,
                                  const double di_max, const double dj_min, int *max_index_i,
                                  int *max_index_j, int *init_pi, int *init_pj,
                                  const timebin_t max_active_bin) {

  const struct part *restrict parts_i = ci->parts;
  const struct part *restrict parts_j = cj->parts;

  int first_pi = 0, last_pj = cj->count - 1;
  int temp;

  /* Find the leftmost active particle in cell i that interacts with any
   * particle in cell j. */
  first_pi = ci->count;
  int active_id = first_pi - 1;
  while (first_pi > 0 &&
         sort_i[first_pi - 1].d + dx_max + h_max - rshift > dj_min) {
    first_pi--;
    /* Store the index of the particle if it is active. */
    if (part_is_active_no_debug(&parts_i[sort_i[first_pi].i], max_active_bin))
      active_id = first_pi;
  }

  /* Set the first active pi in range of any particle in cell j. */
  first_pi = active_id;

  /* Find the maximum index into cell j for each particle in range in cell i. */
  if (first_pi < ci->count) {

    /* Start from the first particle in cell j. */
    temp = 0;

    const struct part *pi = &parts_i[sort_i[first_pi].i];
    const float first_di = sort_i[first_pi].d +
                           max(pi->h, hj_max_raw) * kernel_gamma + dx_max -
                           rshift;

    /* Loop through particles in cell j until they are not in range of pi.
     * Make sure that temp stays between 0 and cj->count - 1.*/
    while (temp < cj->count - 1 && first_di > sort_j[temp].d) temp++;

    max_index_i[first_pi] = temp;

    /* Populate max_index_i for remaining particles that are within range. */
    for (int i = first_pi + 1; i < ci->count; i++) {
      temp = max_index_i[i - 1];
      pi = &parts_i[sort_i[i].i];

      const float di =
          sort_i[i].d + max(pi->h, hj_max_raw) * kernel_gamma + dx_max - rshift;

      /* Make sure that temp stays between 0 and cj->count - 1.*/
      while (temp < cj->count - 1 && di > sort_j[temp].d) temp++;

      max_index_i[i] = temp;
    }
  } else {
    /* Make sure that max index is set to first particle in cj.*/
    max_index_i[ci->count - 1] = 0;
  }

  /* Find the rightmost active particle in cell j that interacts with any
   * particle in cell i. */
  last_pj = -1;
  active_id = last_pj;
  while (last_pj < cj->count &&
      sort_j[last_pj + 1].d - h_max - dx_max < di_max) {
    last_pj++;
    /* Store the index of the particle if it is active. */
    if (part_is_active_no_debug(&parts_j[sort_j[last_pj].i], max_active_bin))
      active_id = last_pj;
  }

  /* Set the last active pj in range of any particle in cell i. */
  last_pj = active_id;

  /* Find the maximum index into cell i for each particle in range in cell j. */
  if (last_pj >= 0) {

    /* Start from the last particle in cell i. */
    temp = ci->count - 1;

    const struct part *pj = &parts_j[sort_j[last_pj].i];
    const float last_dj = sort_j[last_pj].d - dx_max -
                          max(pj->h, hi_max_raw) * kernel_gamma + rshift;

    /* Loop through particles in cell i until they are not in range of pj. */
    while (temp > 0 && last_dj < sort_i[temp].d) temp--;

    max_index_j[last_pj] = temp;

    /* Populate max_index_j for remaining particles that are within range. */
    for (int i = last_pj - 1; i >= 0; i--) {
      temp = max_index_j[i + 1];
      pj = &parts_j[sort_j[i].i];

      const float dj = sort_j[i].d - dx_max -
                       (max(pj->h, hi_max_raw) * kernel_gamma) + rshift;

      while (temp > 0 && dj < sort_i[temp].d) temp--;

      max_index_j[i] = temp;
    }
  } else {
    /* Make sure that max index is set to last particle in ci.*/
    max_index_j[0] = ci->count - 1;
  }

  *init_pi = first_pi;
  *init_pj = last_pj;
}

#endif /* WITH_VECTORIZATION && GADGET2_SPH */

/**
 * @brief Compute the cell self-interaction (non-symmetric) using vector
 * intrinsics with one particle pi at a time.
 *
 * @param r The #runner.
 * @param c The #cell.
 */
__attribute__((always_inline)) INLINE void runner_doself1_density_vec(
    struct runner *r, struct cell *restrict c) {

#if defined(WITH_VECTORIZATION) && defined(GADGET2_SPH)

  /* Get some local variables */
  const struct engine *e = r->e;
  const timebin_t max_active_bin = e->max_active_bin;
  struct part *restrict parts = c->parts;
  const int count = c->count;

  TIMER_TIC;

  /* Anything to do here? */
  if (!cell_is_active(c, e)) return;

  if (!cell_are_part_drifted(c, e)) error("Interacting undrifted cell.");

#ifdef SWIFT_DEBUG_CHECKS
  for (int i = 0; i < count; i++) {
    /* Check that particles have been drifted to the current time */
    if (parts[i].ti_drift != e->ti_current)
      error("Particle pi not drifted to current time");
  }
#endif

  /* Get the particle cache from the runner and re-allocate
   * the cache if it is not big enough for the cell. */
  struct cache *restrict cell_cache = &r->ci_cache;

  if (cell_cache->count < count) cache_init(cell_cache, count);

  /* Read the particles from the cell and store them locally in the cache. */
  cache_read_particles(c, cell_cache);

  /* Create secondary cache to store particle interactions. */
  struct c2_cache int_cache;

  /* Loop over the particles in the cell. */
  for (int pid = 0; pid < count; pid++) {

    /* Get a pointer to the ith particle. */
    struct part *restrict pi = &parts[pid];

    /* Is the ith particle active? */
    if (!part_is_active_no_debug(pi, max_active_bin)) continue;

    const float hi = cell_cache->h[pid];

    /* Fill particle pi vectors. */
    const vector v_pix = vector_set1(cell_cache->x[pid]);
    const vector v_piy = vector_set1(cell_cache->y[pid]);
    const vector v_piz = vector_set1(cell_cache->z[pid]);
    const vector v_hi = vector_set1(hi);
    const vector v_vix = vector_set1(cell_cache->vx[pid]);
    const vector v_viy = vector_set1(cell_cache->vy[pid]);
    const vector v_viz = vector_set1(cell_cache->vz[pid]);

    const float hig2 = hi * hi * kernel_gamma2;
    const vector v_hig2 = vector_set1(hig2);

    /* Get the inverse of hi. */
    vector v_hi_inv = vec_reciprocal(v_hi);

    /* Reset cumulative sums of update vectors. */
    vector v_rhoSum = vector_setzero();
    vector v_rho_dhSum = vector_setzero();
    vector v_wcountSum = vector_setzero();
    vector v_wcount_dhSum = vector_setzero();
    vector v_div_vSum = vector_setzero();
    vector v_curlvxSum = vector_setzero();
    vector v_curlvySum = vector_setzero();
    vector v_curlvzSum = vector_setzero();

    /* Pad cache if there is a serial remainder. */
    int count_align = count;
    const int rem = count % (NUM_VEC_PROC * VEC_SIZE);
    if (rem != 0) {
      count_align += (NUM_VEC_PROC * VEC_SIZE) - rem;

      /* Set positions to the same as particle pi so when the r2 > 0 mask is
       * applied these extra contributions are masked out.*/
      for (int i = count; i < count_align; i++) {
        cell_cache->x[i] = v_pix.f[0];
        cell_cache->y[i] = v_piy.f[0];
        cell_cache->z[i] = v_piz.f[0];
      }
    }

    /* The number of interactions for pi and the padded version of it to
     * make it a multiple of VEC_SIZE. */
    int icount = 0, icount_align = 0;

    /* Find all of particle pi's interacions and store needed values in the
     * secondary cache.*/
    for (int pjd = 0; pjd < count_align; pjd += (NUM_VEC_PROC * VEC_SIZE)) {

      /* Load 2 sets of vectors from the particle cache. */
      const vector v_pjx = vector_load(&cell_cache->x[pjd]);
      const vector v_pjy = vector_load(&cell_cache->y[pjd]);
      const vector v_pjz = vector_load(&cell_cache->z[pjd]);

      const vector v_pjx2 = vector_load(&cell_cache->x[pjd + VEC_SIZE]);
      const vector v_pjy2 = vector_load(&cell_cache->y[pjd + VEC_SIZE]);
      const vector v_pjz2 = vector_load(&cell_cache->z[pjd + VEC_SIZE]);

      /* Compute the pairwise distance. */
      vector v_dx, v_dy, v_dz, v_r2;
      vector v_dx_2, v_dy_2, v_dz_2, v_r2_2;

      v_dx.v = vec_sub(v_pix.v, v_pjx.v);
      v_dx_2.v = vec_sub(v_pix.v, v_pjx2.v);
      v_dy.v = vec_sub(v_piy.v, v_pjy.v);
      v_dy_2.v = vec_sub(v_piy.v, v_pjy2.v);
      v_dz.v = vec_sub(v_piz.v, v_pjz.v);
      v_dz_2.v = vec_sub(v_piz.v, v_pjz2.v);

      v_r2.v = vec_mul(v_dx.v, v_dx.v);
      v_r2_2.v = vec_mul(v_dx_2.v, v_dx_2.v);
      v_r2.v = vec_fma(v_dy.v, v_dy.v, v_r2.v);
      v_r2_2.v = vec_fma(v_dy_2.v, v_dy_2.v, v_r2_2.v);
      v_r2.v = vec_fma(v_dz.v, v_dz.v, v_r2.v);
      v_r2_2.v = vec_fma(v_dz_2.v, v_dz_2.v, v_r2_2.v);

      /* Form a mask from r2 < hig2 and r2 > 0.*/
      mask_t v_doi_mask, v_doi_mask_self_check, v_doi_mask2,
          v_doi_mask2_self_check;

      /* Form r2 > 0 mask and r2 < hig2 mask. */
      vec_create_mask(v_doi_mask_self_check, vec_cmp_gt(v_r2.v, vec_setzero()));
      vec_create_mask(v_doi_mask, vec_cmp_lt(v_r2.v, v_hig2.v));

      /* Form r2 > 0 mask and r2 < hig2 mask. */
      vec_create_mask(v_doi_mask2_self_check,
                      vec_cmp_gt(v_r2_2.v, vec_setzero()));
      vec_create_mask(v_doi_mask2, vec_cmp_lt(v_r2_2.v, v_hig2.v));

      /* Combine two masks and form integer masks. */
      const int doi_mask = vec_is_mask_true(v_doi_mask) &
                           vec_is_mask_true(v_doi_mask_self_check);
      const int doi_mask2 = vec_is_mask_true(v_doi_mask2) &
                            vec_is_mask_true(v_doi_mask2_self_check);

#ifdef DEBUG_INTERACTIONS_SPH
      for (int bit_index = 0; bit_index < VEC_SIZE; bit_index++) {
        if (doi_mask & (1 << bit_index)) { 
          if(pi->num_ngb_density < MAX_NUM_OF_NEIGHBOURS)
            pi->ids_ngbs_density[pi->num_ngb_density] = parts[pjd + bit_index].id;
          ++pi->num_ngb_density;
        }

        if (doi_mask2 & (1 << bit_index)) { 
          if(pi->num_ngb_density < MAX_NUM_OF_NEIGHBOURS)
            pi->ids_ngbs_density[pi->num_ngb_density] = parts[pjd + VEC_SIZE + bit_index].id;
          ++pi->num_ngb_density;
        }
      }
#endif

      /* If there are any interactions left pack interaction values into c2
       * cache. */
      if (doi_mask) {
        storeInteractions(doi_mask, pjd, &v_r2, &v_dx, &v_dy, &v_dz, cell_cache,
                          &int_cache, &icount, &v_rhoSum, &v_rho_dhSum,
                          &v_wcountSum, &v_wcount_dhSum, &v_div_vSum,
                          &v_curlvxSum, &v_curlvySum, &v_curlvzSum, v_hi_inv,
                          v_vix, v_viy, v_viz);
      }
      if (doi_mask2) {
        storeInteractions(doi_mask2, pjd + VEC_SIZE, &v_r2_2, &v_dx_2, &v_dy_2,
                          &v_dz_2, cell_cache, &int_cache, &icount, &v_rhoSum,
                          &v_rho_dhSum, &v_wcountSum, &v_wcount_dhSum,
                          &v_div_vSum, &v_curlvxSum, &v_curlvySum, &v_curlvzSum,
                          v_hi_inv, v_vix, v_viy, v_viz);
      }
    }

    /* Perform padded vector remainder interactions if any are present. */
    calcRemInteractions(&int_cache, icount, &v_rhoSum, &v_rho_dhSum,
                        &v_wcountSum, &v_wcount_dhSum, &v_div_vSum,
                        &v_curlvxSum, &v_curlvySum, &v_curlvzSum, v_hi_inv,
                        v_vix, v_viy, v_viz, &icount_align);

    /* Initialise masks to true in case remainder interactions have been
     * performed. */
    mask_t int_mask, int_mask2;
    vec_init_mask_true(int_mask);
    vec_init_mask_true(int_mask2);

    /* Perform interaction with 2 vectors. */
    for (int pjd = 0; pjd < icount_align; pjd += (NUM_VEC_PROC * VEC_SIZE)) {
      runner_iact_nonsym_2_vec_density(
          &int_cache.r2q[pjd], &int_cache.dxq[pjd], &int_cache.dyq[pjd],
          &int_cache.dzq[pjd], v_hi_inv, v_vix, v_viy, v_viz,
          &int_cache.vxq[pjd], &int_cache.vyq[pjd], &int_cache.vzq[pjd],
          &int_cache.mq[pjd], &v_rhoSum, &v_rho_dhSum, &v_wcountSum,
          &v_wcount_dhSum, &v_div_vSum, &v_curlvxSum, &v_curlvySum,
          &v_curlvzSum, int_mask, int_mask2, 0);
    }

    /* Perform horizontal adds on vector sums and store result in particle pi.
     */
    VEC_HADD(v_rhoSum, pi->rho);
    VEC_HADD(v_rho_dhSum, pi->density.rho_dh);
    VEC_HADD(v_wcountSum, pi->density.wcount);
    VEC_HADD(v_wcount_dhSum, pi->density.wcount_dh);
    VEC_HADD(v_div_vSum, pi->density.div_v);
    VEC_HADD(v_curlvxSum, pi->density.rot_v[0]);
    VEC_HADD(v_curlvySum, pi->density.rot_v[1]);
    VEC_HADD(v_curlvzSum, pi->density.rot_v[2]);

    /* Reset interaction count. */
    icount = 0;
  } /* loop over all particles. */

  TIMER_TOC(timer_doself_density);

#else

  error("Incorrectly calling vectorized Gadget-2 functions!");

#endif /* WITH_VECTORIZATION */
}

/**
 * @brief Compute the interactions between a cell pair, but only for the
 *      given indices in ci. (Vectorised)
 *
 * @param r The #runner.
 * @param c The first #cell.
 * @param parts The #part to interact.
 * @param ind The list of indices of particles in @c c to interact with.
 * @param pi_count The number of particles in @c ind.
 */
__attribute__((always_inline)) INLINE void runner_doself_subset_density_vec(
    struct runner *r, struct cell *restrict c, struct part *restrict parts,
    int *restrict ind, int pi_count) {

#if defined(WITH_VECTORIZATION) && defined(GADGET2_SPH)

  const int count = c->count;

  TIMER_TIC;

  /* Get the particle cache from the runner and re-allocate
   * the cache if it is not big enough for the cell. */
  struct cache *restrict cell_cache = &r->ci_cache;

  if (cell_cache->count < count) cache_init(cell_cache, count);

  /* Read the particles from the cell and store them locally in the cache. */
  cache_read_particles(c, cell_cache);

  /* Create secondary cache to store particle interactions. */
  struct c2_cache int_cache;

  /* Loop over the subset of particles in the parts that need updating. */
  for (int pid = 0; pid < pi_count; pid++) {

    /* Get a pointer to the ith particle. */
    struct part *pi = &parts[ind[pid]];

#ifdef SWIFT_DEBUG_CHECKS
    const struct engine *e = r->e;
    if (!part_is_active(pi, e)) error("Inactive particle in subset function!");
#endif

    const float hi = pi->h;

    /* Fill particle pi vectors. */
    const vector v_pix = vector_set1(pi->x[0] - c->loc[0]);
    const vector v_piy = vector_set1(pi->x[1] - c->loc[1]);
    const vector v_piz = vector_set1(pi->x[2] - c->loc[2]);
    const vector v_hi = vector_set1(hi);
    const vector v_vix = vector_set1(pi->v[0]);
    const vector v_viy = vector_set1(pi->v[1]);
    const vector v_viz = vector_set1(pi->v[2]);

    const float hig2 = hi * hi * kernel_gamma2;
    const vector v_hig2 = vector_set1(hig2);

    /* Get the inverse of hi. */
    vector v_hi_inv = vec_reciprocal(v_hi);

    /* Reset cumulative sums of update vectors. */
    vector v_rhoSum = vector_setzero();
    vector v_rho_dhSum = vector_setzero();
    vector v_wcountSum = vector_setzero();
    vector v_wcount_dhSum = vector_setzero();
    vector v_div_vSum = vector_setzero();
    vector v_curlvxSum = vector_setzero();
    vector v_curlvySum = vector_setzero();
    vector v_curlvzSum = vector_setzero();

    /* Pad cache if there is a serial remainder. */
    int count_align = count;
    const int rem = count % (NUM_VEC_PROC * VEC_SIZE);
    if (rem != 0) {
      const int pad = (NUM_VEC_PROC * VEC_SIZE) - rem;

      count_align += pad;

      /* Set positions to the same as particle pi so when the r2 > 0 mask is
       * applied these extra contributions are masked out.*/
      for (int i = count; i < count_align; i++) {
        cell_cache->x[i] = v_pix.f[0];
        cell_cache->y[i] = v_piy.f[0];
        cell_cache->z[i] = v_piz.f[0];
      }
    }

    /* The number of interactions for pi and the padded version of it to
     * make it a multiple of VEC_SIZE. */
    int icount = 0, icount_align = 0;

    /* Find all of particle pi's interacions and store needed values in the
     * secondary cache.*/
    for (int pjd = 0; pjd < count_align; pjd += (NUM_VEC_PROC * VEC_SIZE)) {

      /* Load 2 sets of vectors from the particle cache. */
      const vector v_pjx = vector_load(&cell_cache->x[pjd]);
      const vector v_pjy = vector_load(&cell_cache->y[pjd]);
      const vector v_pjz = vector_load(&cell_cache->z[pjd]);

      const vector v_pjx2 = vector_load(&cell_cache->x[pjd + VEC_SIZE]);
      const vector v_pjy2 = vector_load(&cell_cache->y[pjd + VEC_SIZE]);
      const vector v_pjz2 = vector_load(&cell_cache->z[pjd + VEC_SIZE]);

      /* Compute the pairwise distance. */
      vector v_dx, v_dy, v_dz, v_r2;
      vector v_dx_2, v_dy_2, v_dz_2, v_r2_2;

      /* p_i - p_j */
      v_dx.v = vec_sub(v_pix.v, v_pjx.v);
      v_dx_2.v = vec_sub(v_pix.v, v_pjx2.v);
      v_dy.v = vec_sub(v_piy.v, v_pjy.v);
      v_dy_2.v = vec_sub(v_piy.v, v_pjy2.v);
      v_dz.v = vec_sub(v_piz.v, v_pjz.v);
      v_dz_2.v = vec_sub(v_piz.v, v_pjz2.v);

      /* r2 = dx^2 + dy^2 + dz^2 */
      v_r2.v = vec_mul(v_dx.v, v_dx.v);
      v_r2_2.v = vec_mul(v_dx_2.v, v_dx_2.v);
      v_r2.v = vec_fma(v_dy.v, v_dy.v, v_r2.v);
      v_r2_2.v = vec_fma(v_dy_2.v, v_dy_2.v, v_r2_2.v);
      v_r2.v = vec_fma(v_dz.v, v_dz.v, v_r2.v);
      v_r2_2.v = vec_fma(v_dz_2.v, v_dz_2.v, v_r2_2.v);

      /* Form a mask from r2 < hig2 and r2 > 0.*/
      mask_t v_doi_mask, v_doi_mask_self_check, v_doi_mask2,
          v_doi_mask2_self_check;

      /* Form r2 > 0 mask and r2 < hig2 mask. */
      vec_create_mask(v_doi_mask_self_check, vec_cmp_gt(v_r2.v, vec_setzero()));
      vec_create_mask(v_doi_mask, vec_cmp_lt(v_r2.v, v_hig2.v));

      /* Form r2 > 0 mask and r2 < hig2 mask. */
      vec_create_mask(v_doi_mask2_self_check,
                      vec_cmp_gt(v_r2_2.v, vec_setzero()));
      vec_create_mask(v_doi_mask2, vec_cmp_lt(v_r2_2.v, v_hig2.v));

      /* Combine two masks and form integer masks. */
      const int doi_mask = vec_is_mask_true(v_doi_mask) &
                           vec_is_mask_true(v_doi_mask_self_check);
      const int doi_mask2 = vec_is_mask_true(v_doi_mask2) &
                            vec_is_mask_true(v_doi_mask2_self_check);

#ifdef DEBUG_INTERACTIONS_SPH
      for (int bit_index = 0; bit_index < VEC_SIZE; bit_index++) {