Added max_di and max_dj arrays to the cache.

src/cache.h

@@ -84,35 +84,17 @@ struct cache {
 
   /* Particle z velocity. */
   float *restrict vz __attribute__((aligned(CACHE_ALIGN)));
+  
+  /* Maximum distance of pi particles into cj. */
+  float *restrict max_di __attribute__((aligned(CACHE_ALIGN)));
+
+  /* Maximum distance of pj particles into ci. */
+  float *restrict max_dj __attribute__((aligned(CACHE_ALIGN)));
 
   /* Cache size. */
   int count;
 
 #endif
-  /* Particle x position. */
-  //float *restrict rho __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle y position. */
-  //float *restrict rho_dh __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle z position. */
-  //float *restrict wcount __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle smoothing length. */
-  //float *restrict wcount_dh __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle mass. */
-  //float *restrict div_v __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle x velocity. */
-  //float *restrict curl_vx __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle y velocity. */
-  //float *restrict curl_vy __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
-  ///* Particle z velocity. */
-  //float *restrict curl_vz __attribute__((aligned(sizeof(float) * VEC_SIZE)));
-
 };
 
 #ifdef DOPAIR1_AUTO_VEC
@@ -175,6 +157,8 @@ __attribute__((always_inline)) INLINE void cache_init(struct cache *c,
     free(c->vy);
     free(c->vz);
     free(c->h);
+    free(c->max_di);
+    free(c->max_dj);
   }
 
   error += posix_memalign((void **)&c->x, alignment, sizeBytes);
@@ -185,15 +169,9 @@ __attribute__((always_inline)) INLINE void cache_init(struct cache *c,
   error += posix_memalign((void **)&c->vy, alignment, sizeBytes);
   error += posix_memalign((void **)&c->vz, alignment, sizeBytes);
   error += posix_memalign((void **)&c->h, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->rho, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->rho_dh, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->wcount, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->wcount_dh, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->div_v, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->curl_vx, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->curl_vy, alignment, sizeBytes);
-  //error += posix_memalign((void **)&c->curl_vz, alignment, sizeBytes);
-
+  error += posix_memalign((void **)&c->max_di, alignment, sizeBytes);
+  error += posix_memalign((void **)&c->max_dj, alignment, sizeBytes);
+  
   if (error != 0)
     error("Couldn't allocate cache, no. of particles: %d", (int)count);
   c->count = count;

src/runner_doiact_vec.c

@@ -972,9 +972,6 @@ __attribute__((always_inline)) INLINE void runner_doself1_density_vec_2(
 #endif /* WITH_VECTORIZATION */
 }
 
-float max_di[MAX_NO_OF_PARTS] __attribute__((aligned(sizeof(VEC_SIZE * sizeof(float))))); /* max distance into ci */
-float max_dj[MAX_NO_OF_PARTS] __attribute__((aligned(sizeof(VEC_SIZE * sizeof(float))))); /* max distance into cj */
-
 void runner_dopair1_density_vec(struct runner *r, struct cell *ci, struct cell *cj) {
 
 #ifdef WITH_VECTORIZATION
@@ -1029,7 +1026,12 @@ void runner_dopair1_density_vec(struct runner *r, struct cell *ci, struct cell *
   }
 
   int first_pi, last_pj;
-  
+  float *max_di __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+  float *max_dj __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+
+  max_di = r->par_cache.max_di;
+  max_dj = r->par_cache.max_dj;
+
   /* Find particles maximum distance into cj, max_di[] and ci, max_dj[]. */
   /* For particles in ci */  
   populate_max_d_no_cache(ci, cj, sort_i, sort_j, dx_max, rshift, max_di, max_dj, &first_pi, &last_pj);
@@ -1409,6 +1411,12 @@ void runner_dopair1_density_vec_1(struct runner *r, struct cell *ci, struct cell
 
   cache_read_two_cells_sorted(ci, cj, ci_cache, &cj_cache, sort_i, sort_j, shift);
 
+  float *max_di __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+  float *max_dj __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+
+  max_di = r->par_cache.max_di;
+  max_dj = r->par_cache.max_dj;
+
   /* Find particles maximum distance into cj, max_di[] and ci, max_dj[]. */
   /* For particles in ci */  
   populate_max_d(ci, cj, sort_i, sort_j, ci_cache, &cj_cache, dx_max, rshift, max_di, max_dj);
@@ -1833,6 +1841,12 @@ void runner_dopair1_density_vec_2(struct runner *r, struct cell *ci, struct cell
     cache_init(&cj_cache, count_j);
   }
 
+  float *max_di __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+  float *max_dj __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+
+  max_di = r->par_cache.max_di;
+  max_dj = r->par_cache.max_dj;
+
   int first_pi, last_pj;
   /* Find particles maximum distance into cj, max_di[] and ci, max_dj[]. */
   /* For particles in ci */  
@@ -2401,6 +2415,12 @@ void runner_dopair1_density_vec_3(struct runner *r, struct cell *ci, struct cell
   cache_read_two_cells(ci, cj, ci_cache, &cj_cache, shift);
   //cache_read_two_cells_sorted(ci, cj, ci_cache, &cj_cache, sort_i, sort_j, shift);
 
+  float *max_di __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+  float *max_dj __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+
+  max_di = r->par_cache.max_di;
+  max_dj = r->par_cache.max_dj;
+
   /* Find particles maximum distance into cj, max_di[] and ci, max_dj[]. */
   /* For particles in ci */  
   populate_max_d(ci, cj, sort_i, sort_j, ci_cache, &cj_cache, dx_max, rshift, max_di, max_dj);
@@ -2898,6 +2918,12 @@ void runner_dopair1_density_vec_4(struct runner *r, struct cell *ci, struct cell
 
   cache_read_cell_sorted(cj, ci_cache, sort_j, loc, shift_cj);
 
+  float *max_di __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+  float *max_dj __attribute__((aligned(sizeof(float) * VEC_SIZE)));
+
+  max_di = r->par_cache.max_di;
+  max_dj = r->par_cache.max_dj;
+
   /* Find particles maximum distance into cj, max_di[] and ci, max_dj[]. */
   /* For particles in ci */  
   float h = parts_i[sort_i[0].i].h;