Merge branch 'gravity_speedup' into 'master'

Use particle caches for the gravity P-P interactions. Ewald summation for the gravity force checks



See merge request !398

examples/EAGLE_12/eagle_12.yml

@@ -26,7 +26,7 @@ Statistics:
 # Parameters for the self-gravity scheme
 Gravity:
   eta:                   0.025    # Constant dimensionless multiplier for time integration.
-  epsilon:               0.0001   # Softening length (in internal units).
+  epsilon:               0.001    # Softening length (in internal units).
   theta:                 0.7      # Opening angle (Multipole acceptance criterion)
   
 # Parameters for the hydrodynamics scheme

examples/EAGLE_6/eagle_6.yml

@@ -12,6 +12,9 @@ TimeIntegration:
   time_end:   1e-2  # The end time of the simulation (in internal units).
   dt_min:     1e-10 # The minimal time-step size of the simulation (in internal units).
   dt_max:     1e-4  # The maximal time-step size of the simulation (in internal units).
+
+Scheduler:
+  cell_split_size: 64
   
 # Parameters governing the snapshots
 Snapshots:

examples/main.c

@@ -569,6 +569,11 @@ int main(int argc, char *argv[]) {
     message("nr of cells at depth %i is %i.", data[0], data[1]);
   }
 
+/* Initialise the table of Ewald corrections for the gravity checks */
+#ifdef SWIFT_GRAVITY_FORCE_CHECKS
+  if (periodic) gravity_exact_force_ewald_init(dim[0]);
+#endif
+
   /* Initialise the external potential properties */
   struct external_potential potential;
   if (with_external_gravity)

src/align.h

@@ -23,9 +23,55 @@
  * @brief The default struct alignment in SWIFT.
  */
 #define SWIFT_STRUCT_ALIGNMENT 32
+
 /**
  * @brief Defines alignment of structures
  */
 #define SWIFT_STRUCT_ALIGN __attribute__((aligned(SWIFT_STRUCT_ALIGNMENT)))
 
+/**
+ * @brief The default cache alignment in SWIFT.
+ */
+#define SWIFT_CACHE_ALIGNMENT 64
+
+/**
+ * @brief Defines alignment of caches
+ */
+#define SWIFT_CACHE_ALIGN __attribute__((aligned(SWIFT_CACHE_ALIGNMENT)))
+
+/**
+ * @brief Macro to tell the compiler that a given array has the specified
+ * alignment.
+ *
+ * Note that this turns into a no-op but gives information to the compiler.
+ *
+ * @param array The array.
+ * @param alignment The alignment in bytes of the array.
+ */
+#if defined(__ICC)
+#define swift_align_information(array, alignment) \
+  __assume_aligned(array, alignment);
+#elif defined(__GNUC__)
+#define swift_align_information(array, alignment) \
+  array = __builtin_assume_aligned(array, alignment);
+#else
+#define swift_align_information(array, alignment) ;
+#endif
+
+/**
+ * @brief Macro to tell the compiler that a given number is 0 modulo a given
+ * size.
+ *
+ * Note that this turns into a no-op but gives information to the compiler.
+ * GCC does not have the equivalent built-in so defaults to nothing.
+ *
+ * @param var The variable
+ * @param size The modulo of interest.
+ */
+#if defined(__ICC)
+#define swift_assume_size(var, size) __assume(var % size == 0);
+#else
+#define swift_assume_size(var, size) ;
+#endif
+
 #endif /* SWIFT_ALIGN_H */

src/engine.c

@@ -1742,7 +1742,7 @@ void engine_make_self_gravity_tasks_mapper(void *map_data, int num_elements,
     /* If the cells is local build a self-interaction */
     scheduler_addtask(sched, task_type_self, task_subtype_grav, 0, 0, ci, NULL);
 
-    /* Deal with periodicity dependencies */
+    /* Deal with periodicity FFT task dependencies */
     const int ghost_id = cell_getid(cdim_ghost, i / 4, j / 4, k / 4);
     if (ghost_id > n_ghosts) error("Invalid ghost_id");
     if (periodic) {
@@ -1750,6 +1750,10 @@ void engine_make_self_gravity_tasks_mapper(void *map_data, int num_elements,
       ci->grav_ghost[1] = ghosts[2 * ghost_id + 1];
     }
 
+    /* Recover the multipole information */
+    struct gravity_tensors *const multi_i = ci->multipole;
+    const double CoM_i[3] = {multi_i->CoM[0], multi_i->CoM[1], multi_i->CoM[2]};
+
     /* Loop over every other cell */
     for (int ii = 0; ii < cdim[0]; ii++) {
       for (int jj = 0; jj < cdim[1]; jj++) {
@@ -1768,11 +1772,27 @@ void engine_make_self_gravity_tasks_mapper(void *map_data, int num_elements,
           /* Is that neighbour local ? */
           if (cj->nodeID != nodeID) continue;  // MATTHIEU
 
-          /* Are the cells to close for a MM interaction ? */
-          if (!gravity_multipole_accept_rebuild(ci->multipole, cj->multipole,
-                                                theta_crit_inv, periodic,
-                                                dim)) {
+          /* Recover the multipole information */
+          struct gravity_tensors *const multi_j = cj->multipole;
+
+          /* Get the distance between the CoMs */
+          double dx = CoM_i[0] - multi_j->CoM[0];
+          double dy = CoM_i[1] - multi_j->CoM[1];
+          double dz = CoM_i[2] - multi_j->CoM[2];
+
+          /* Apply BC */
+          if (periodic) {
+            dx = nearest(dx, dim[0]);
+            dy = nearest(dy, dim[1]);
+            dz = nearest(dz, dim[2]);
+          }
+          const double r2 = dx * dx + dy * dy + dz * dz;
+
+          /* Are the cells too close for a MM interaction ? */
+          if (!gravity_multipole_accept_rebuild(multi_i, multi_j,
+                                                theta_crit_inv, r2)) {
 
+            /* Ok, we need to add a direct pair calculation */
             scheduler_addtask(sched, task_type_pair, task_subtype_grav, 0, 0,
                               ci, cj);
           }
@@ -4534,6 +4554,10 @@ void engine_init(struct engine *e, struct space *s,
     e->runners[k].cj_cache.count = 0;
     cache_init(&e->runners[k].ci_cache, CACHE_SIZE);
     cache_init(&e->runners[k].cj_cache, CACHE_SIZE);
+    e->runners[k].ci_gravity_cache.count = 0;
+    e->runners[k].cj_gravity_cache.count = 0;
+    gravity_cache_init(&e->runners[k].ci_gravity_cache, space_splitsize);
+    gravity_cache_init(&e->runners[k].cj_gravity_cache, space_splitsize);
 #endif
 
     if (verbose) {
@@ -4620,8 +4644,12 @@ void engine_compute_next_snapshot_time(struct engine *e) {
 void engine_clean(struct engine *e) {
 
 #ifdef WITH_VECTORIZATION
-  for (int i = 0; i < e->nr_threads; ++i) cache_clean(&e->runners[i].ci_cache);
-  for (int i = 0; i < e->nr_threads; ++i) cache_clean(&e->runners[i].cj_cache);
+  for (int i = 0; i < e->nr_threads; ++i) {
+    cache_clean(&e->runners[i].ci_cache);
+    cache_clean(&e->runners[i].cj_cache);
+    gravity_cache_clean(&e->runners[i].ci_gravity_cache);
+    gravity_cache_clean(&e->runners[i].cj_gravity_cache);
+  }
 #endif
   free(e->runners);
   free(e->snapshotUnits);

src/gravity.c

@@ -21,9 +21,15 @@
 #include "../config.h"
 
 /* Some standard headers. */
+#include <float.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <unistd.h>
 
+#ifdef HAVE_HDF5
+#include <hdf5.h>
+#endif
+
 /* This object's header. */
 #include "gravity.h"
 
@@ -39,6 +45,256 @@ struct exact_force_data {
   double const_G;
 };
 
+#ifdef SWIFT_GRAVITY_FORCE_CHECKS
+
+/* Size of the Ewald table */
+#define Newald 64
+
+/* Components of the Ewald correction */
+static float fewald_x[Newald + 1][Newald + 1][Newald + 1];
+static float fewald_y[Newald + 1][Newald + 1][Newald + 1];
+static float fewald_z[Newald + 1][Newald + 1][Newald + 1];
+
+/* Factor used to normalize the access to the Ewald table */
+float ewald_fac;
+#endif
+
+/**
+ * @brief Allocates the memory and computes one octant of the
+ * Ewald correction table.
+ *
+ * We follow Hernquist, Bouchet & Suto, 1991, ApJS, Volume 75, p.231-240,
+ * equations (2.14a) and (2.14b) with alpha = 2. We consider all terms with
+ * |x - nL| < 4L and |h|^2 < 16.
+ *
+ * @param boxSize The side-length (L) of the volume.
+ */
+void gravity_exact_force_ewald_init(double boxSize) {
+
+#ifdef SWIFT_GRAVITY_FORCE_CHECKS
+  const ticks tic = getticks();
+  message("Computing Ewald correction table...");
+
+  /* Level of correction  (Hernquist et al. 1991)*/
+  const float alpha = 2.f;
+
+  /* some useful constants */
+  const float alpha2 = alpha * alpha;
+  const float factor_exp1 = 2.f * alpha / sqrt(M_PI);
+  const float factor_exp2 = -M_PI * M_PI / alpha2;
+  const float factor_sin = 2.f * M_PI;
+  const float boxSize_inv2 = 1.f / (boxSize * boxSize);
+
+  /* Ewald factor to access the table */
+  ewald_fac = (double)(2 * Newald) / boxSize;
+
+  /* Zero everything */
+  bzero(fewald_x, (Newald + 1) * (Newald + 1) * (Newald + 1) * sizeof(float));
+  bzero(fewald_y, (Newald + 1) * (Newald + 1) * (Newald + 1) * sizeof(float));
+  bzero(fewald_z, (Newald + 1) * (Newald + 1) * (Newald + 1) * sizeof(float));
+
+  /* Compute the values in one of the octants */
+  for (int i = 0; i <= Newald; ++i) {
+    for (int j = 0; j <= Newald; ++j) {
+      for (int k = 0; k <= Newald; ++k) {
+
+        if (i == 0 && j == 0 && k == 0) continue;
+
+        /* Distance vector */
+        const float r_x = 0.5f * ((float)i) / Newald;
+        const float r_y = 0.5f * ((float)j) / Newald;
+        const float r_z = 0.5f * ((float)k) / Newald;
+
+        /* Norm of distance vector */
+        const float r2 = r_x * r_x + r_y * r_y + r_z * r_z;
+        const float r_inv = 1.f / sqrtf(r2);
+        const float r_inv3 = r_inv * r_inv * r_inv;
+
+        /* Normal gravity potential term */
+        float f_x = r_x * r_inv3;
+        float f_y = r_y * r_inv3;
+        float f_z = r_z * r_inv3;
+
+        for (int n_i = -4; n_i <= 4; ++n_i) {
+          for (int n_j = -4; n_j <= 4; ++n_j) {
+            for (int n_k = -4; n_k <= 4; ++n_k) {
+
+              const float d_x = r_x - n_i;
+              const float d_y = r_y - n_j;
+              const float d_z = r_z - n_k;
+
+              /* Discretised distance */
+              const float r_tilde2 = d_x * d_x + d_y * d_y + d_z * d_z;
+              const float r_tilde_inv = 1.f / sqrtf(r_tilde2);
+              const float r_tilde = r_tilde_inv * r_tilde2;
+              const float r_tilde_inv3 =
+                  r_tilde_inv * r_tilde_inv * r_tilde_inv;
+
+              const float val =
+                  erfcf(alpha * r_tilde) +
+                  factor_exp1 * r_tilde * expf(-alpha2 * r_tilde2);
+
+              /* First correction term */
+              const float f = val * r_tilde_inv3;
+              f_x -= f * d_x;
+              f_y -= f * d_y;
+              f_z -= f * d_z;
+            }
+          }
+        }
+
+        for (int h_i = -4; h_i <= 4; ++h_i) {
+          for (int h_j = -4; h_j <= 4; ++h_j) {
+            for (int h_k = -4; h_k <= 4; ++h_k) {
+
+              const float h2 = h_i * h_i + h_j * h_j + h_k * h_k;
+
+              const float h2_inv = 1.f / (h2 + FLT_MIN);
+              const float h_dot_x = h_i * r_x + h_j * r_y + h_k * r_z;
+
+              const float val = 2.f * h2_inv * expf(h2 * factor_exp2) *
+                                sinf(factor_sin * h_dot_x);
+
+              /* Second correction term */
+              f_x -= val * h_i;
+              f_y -= val * h_j;
+              f_z -= val * h_k;
+            }
+          }
+        }
+
+        /* Save back to memory */
+        fewald_x[i][j][k] = f_x;
+        fewald_y[i][j][k] = f_y;
+        fewald_z[i][j][k] = f_z;
+      }
+    }
+  }
+
+/* Dump the Ewald table to a file */
+#ifdef HAVE_HDF5
+  hid_t h_file =
+      H5Fcreate("Ewald.hdf5", H5F_ACC_TRUNC, H5P_DEFAULT, H5P_DEFAULT);
+  if (h_file < 0) error("Error while opening file for Ewald dump.");
+
+  /* Create dataspace */
+  hsize_t dim[3] = {Newald + 1, Newald + 1, Newald + 1};
+  hid_t h_space = H5Screate_simple(3, dim, NULL);
+  hid_t h_data;
+  h_data = H5Dcreate(h_file, "Ewald_x", H5T_NATIVE_FLOAT, h_space, H5P_DEFAULT,
+                     H5P_DEFAULT, H5P_DEFAULT);
+  H5Dwrite(h_data, H5T_NATIVE_FLOAT, h_space, H5S_ALL, H5P_DEFAULT,
+           &(fewald_x[0][0][0]));
+  H5Dclose(h_data);
+  h_data = H5Dcreate(h_file, "Ewald_y", H5T_NATIVE_FLOAT, h_space, H5P_DEFAULT,
+                     H5P_DEFAULT, H5P_DEFAULT);
+  H5Dwrite(h_data, H5T_NATIVE_FLOAT, h_space, H5S_ALL, H5P_DEFAULT,
+           &(fewald_y[0][0][0]));
+  H5Dclose(h_data);
+  h_data = H5Dcreate(h_file, "Ewald_z", H5T_NATIVE_FLOAT, h_space, H5P_DEFAULT,
+                     H5P_DEFAULT, H5P_DEFAULT);
+  H5Dwrite(h_data, H5T_NATIVE_FLOAT, h_space, H5S_ALL, H5P_DEFAULT,
+           &(fewald_z[0][0][0]));
+  H5Dclose(h_data);
+  H5Sclose(h_space);
+  H5Fclose(h_file);
+#endif
+
+  /* Apply the box-size correction */
+  for (int i = 0; i <= Newald; ++i) {
+    for (int j = 0; j <= Newald; ++j) {
+      for (int k = 0; k <= Newald; ++k) {
+        fewald_x[i][j][k] *= boxSize_inv2;
+        fewald_y[i][j][k] *= boxSize_inv2;
+        fewald_z[i][j][k] *= boxSize_inv2;
+      }
+    }
+  }
+
+  /* Say goodbye */
+  message("Ewald correction table computed (took %.3f %s). ",
+          clocks_from_ticks(getticks() - tic), clocks_getunit());
+#else
+  error("Gravity checking function called without the corresponding flag.");
+#endif
+}
+
+#ifdef SWIFT_GRAVITY_FORCE_CHECKS
+/**
+ * @brief Compute the Ewald correction for a given distance vector r.
+ *
+ * We interpolate the Ewald correction tables using a tri-linear interpolation
+ * similar to a CIC.
+ *
+ * @param rx x-coordinate of distance vector.
+ * @param ry y-coordinate of distance vector.
+ * @param rz z-coordinate of distance vector.
+ * @param corr (return) The Ewald correction.
+ */
+__attribute__((always_inline)) INLINE static void
+gravity_exact_force_ewald_evaluate(double rx, double ry, double rz,
+                                   double corr[3]) {
+
+  const double s_x = (rx < 0.) ? 1. : -1.;
+  const double s_y = (ry < 0.) ? 1. : -1.;
+  const double s_z = (rz < 0.) ? 1. : -1.;
+  rx = fabs(rx);
+  ry = fabs(ry);
+  rz = fabs(rz);
+
+  int i = (int)(rx * ewald_fac);
+  if (i >= Newald) i = Newald - 1;
+  const double dx = rx * ewald_fac - i;
+  const double tx = 1. - dx;
+
+  int j = (int)(ry * ewald_fac);
+  if (j >= Newald) j = Newald - 1;
+  const double dy = ry * ewald_fac - j;
+  const double ty = 1. - dy;
+
+  int k = (int)(rz * ewald_fac);
+  if (k >= Newald) k = Newald - 1;
+  const double dz = rz * ewald_fac - k;
+  const double tz = 1. - dz;
+
+  /* Interpolation in X */
+  corr[0] = 0.;
+  corr[0] += fewald_x[i + 0][j + 0][k + 0] * tx * ty * tz;
+  corr[0] += fewald_x[i + 0][j + 0][k + 1] * tx * ty * dz;
+  corr[0] += fewald_x[i + 0][j + 1][k + 0] * tx * dy * tz;
+  corr[0] += fewald_x[i + 0][j + 1][k + 1] * tx * dy * dz;
+  corr[0] += fewald_x[i + 1][j + 0][k + 0] * dx * ty * tz;
+  corr[0] += fewald_x[i + 1][j + 0][k + 1] * dx * ty * dz;
+  corr[0] += fewald_x[i + 1][j + 1][k + 0] * dx * dy * tz;
+  corr[0] += fewald_x[i + 1][j + 1][k + 1] * dx * dy * dz;
+  corr[0] *= s_x;
+
+  /* Interpolation in Y */
+  corr[1] = 0.;
+  corr[1] += fewald_y[i + 0][j + 0][k + 0] * tx * ty * tz;
+  corr[1] += fewald_y[i + 0][j + 0][k + 1] * tx * ty * dz;
+  corr[1] += fewald_y[i + 0][j + 1][k + 0] * tx * dy * tz;
+  corr[1] += fewald_y[i + 0][j + 1][k + 1] * tx * dy * dz;
+  corr[1] += fewald_y[i + 1][j + 0][k + 0] * dx * ty * tz;
+  corr[1] += fewald_y[i + 1][j + 0][k + 1] * dx * ty * dz;
+  corr[1] += fewald_y[i + 1][j + 1][k + 0] * dx * dy * tz;
+  corr[1] += fewald_y[i + 1][j + 1][k + 1] * dx * dy * dz;
+  corr[1] *= s_y;
+
+  /* Interpolation in Z */
+  corr[2] = 0.;
+  corr[2] += fewald_z[i + 0][j + 0][k + 0] * tx * ty * tz;
+  corr[2] += fewald_z[i + 0][j + 0][k + 1] * tx * ty * dz;
+  corr[2] += fewald_z[i + 0][j + 1][k + 0] * tx * dy * tz;
+  corr[2] += fewald_z[i + 0][j + 1][k + 1] * tx * dy * dz;
+  corr[2] += fewald_z[i + 1][j + 0][k + 0] * dx * ty * tz;
+  corr[2] += fewald_z[i + 1][j + 0][k + 1] * dx * ty * dz;
+  corr[2] += fewald_z[i + 1][j + 1][k + 0] * dx * dy * tz;
+  corr[2] += fewald_z[i + 1][j + 1][k + 1] * dx * dy * dz;
+  corr[2] *= s_z;
+}
+#endif
+
 /**
  * @brief Checks whether the file containing the exact accelerations for
  * the current choice of parameters already exists.
@@ -116,6 +372,12 @@ void gravity_exact_force_compute_mapper(void *map_data, int nr_gparts,
     if (gpi->id_or_neg_offset % SWIFT_GRAVITY_FORCE_CHECKS == 0 &&
         gpart_is_active(gpi, e)) {
 
+      /* Get some information about the particle */
+      const double pix[3] = {gpi->x[0], gpi->x[1], gpi->x[2]};
+      const double hi = gpi->epsilon;
+      const double hi_inv = 1. / hi;
+      const double hi_inv3 = hi_inv * hi_inv * hi_inv;
+
       /* Be ready for the calculation */
       double a_grav[3] = {0., 0., 0.};
 
@@ -128,9 +390,9 @@ void gravity_exact_force_compute_mapper(void *map_data, int nr_gparts,
         if (gpi == gpj) continue;
 
         /* Compute the pairwise distance. */
-        double dx = gpi->x[0] - gpj->x[0];
-        double dy = gpi->x[1] - gpj->x[1];
-        double dz = gpi->x[2] - gpj->x[2];
+        double dx = gpj->x[0] - pix[0];
+        double dy = gpj->x[1] - pix[1];
+        double dz = gpj->x[2] - pix[2];
 
         /* Now apply periodic BC */
         if (periodic) {
@@ -140,34 +402,41 @@ void gravity_exact_force_compute_mapper(void *map_data, int nr_gparts,
         }
 
         const double r2 = dx * dx + dy * dy + dz * dz;
-        const double r = sqrt(r2);
-        const double ir = 1. / r;
+        const double r_inv = 1. / sqrt(r2);
+        const double r = r2 * r_inv;
         const double mj = gpj->mass;
-        const double hi = gpi->epsilon;
         double f;
-        const double f_lr = 1.;
 
         if (r >= hi) {
 
           /* Get Newtonian gravity */
-          f = mj * ir * ir * ir * f_lr;
+          f = mj * r_inv * r_inv * r_inv;
 
         } else {
 
-          const double hi_inv = 1. / hi;
-          const double hi_inv3 = hi_inv * hi_inv * hi_inv;
           const double ui = r * hi_inv;
           double W;
 
           kernel_grav_eval_double(ui, &W);
 
           /* Get softened gravity */
-          f = mj * hi_inv3 * W * f_lr;
+          f = mj * hi_inv3 * W;
         }
 
-        a_grav[0] -= f * dx;
-        a_grav[1] -= f * dy;
-        a_grav[2] -= f * dz;
+        a_grav[0] += f * dx;
+        a_grav[1] += f * dy;
+        a_grav[2] += f * dz;
+
+        /* Apply Ewald correction for periodic BC */
+        if (periodic && r > 1e-5 * hi) {
+
+          double corr[3];
+          gravity_exact_force_ewald_evaluate(dx, dy, dz, corr);
+
+          a_grav[0] += mj * corr[0];
+          a_grav[1] += mj * corr[1];
+          a_grav[2] += mj * corr[2];
+        }
       }
 
       /* Store the exact answer */

src/gravity.h

@@ -34,6 +34,8 @@
 #include "./gravity/Default/gravity.h"
 #include "./gravity/Default/gravity_iact.h"
 
+void gravity_exact_force_ewald_init(double boxSize);
+void gravity_exact_force_ewald_free();
 void gravity_exact_force_compute(struct space *s, const struct engine *e);
 void gravity_exact_force_check(struct space *s, const struct engine *e,
                                float rel_tol);

src/gravity_cache.h

+/*******************************************************************************
+ * This file is part of SWIFT.
+ * Copyright (c) 2016 Matthieu Schaller (matthieu.schaller@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/>.
+ *
+ ******************************************************************************/
+#ifndef SWIFT_GRAVITY_CACHE_H
+#define SWIFT_GRAVITY_CACHE_H
+
+/* Config parameters. */
+#include "../config.h"
+
+/* Local headers */
+#include "align.h"
+#include "error.h"
+#include "gravity.h"
+#include "vector.h"
+
+/**
+ * @brief A SoA object for the #gpart of a cell.
+ *
+ * This is used to help vectorize the leaf-leaf gravity interactions.
+ */
+struct gravity_cache {
+
+  /*! #gpart x position. */
+  float *restrict x SWIFT_CACHE_ALIGN;
+
+  /*! #gpart y position. */
+  float *restrict y SWIFT_CACHE_ALIGN;
+
+  /*! #gpart z position. */
+  float *restrict z SWIFT_CACHE_ALIGN;
+
+  /*! #gpart softening length. */
+  float *restrict epsilon SWIFT_CACHE_ALIGN;
+
+  /*! #gpart mass. */
+  float *restrict m SWIFT_CACHE_ALIGN;
+
+  /*! #gpart x acceleration. */
+  float *restrict a_x SWIFT_CACHE_ALIGN;
+
+  /*! #gpart y acceleration. */
+  float *restrict a_y SWIFT_CACHE_ALIGN;
+
+  /*! #gpart z acceleration. */
+  float *restrict a_z SWIFT_CACHE_ALIGN;
+
+  /*! Cache size */
+  int count;
+};
+
+/**
+ * @brief Frees the memory allocated in a #gravity_cache
+ *
+ * @param c The #gravity_cache to free.
+ */
+static INLINE void gravity_cache_clean(struct gravity_cache *c) {
+
+  if (c->count > 0) {
+    free(c->x);
+    free(c->y);
+    free(c->z);
+    free(c->epsilon);
+    free(c->m);
+    free(c->a_x);
+    free(c->a_y);
+    free(c->a_z);
+  }
+  c->count = 0;
+}