diff --git a/src/cache.h b/src/cache.h
index e66534ebc4859317724cf5d65eada4e850af253c..49ed478f91c5cb0b028c19d052621a05c02ff0da 100644
--- a/src/cache.h
+++ b/src/cache.h
@@ -23,6 +23,7 @@
#include "../config.h"
/* Local headers */
+#include "align.h"
#include "cell.h"
#include "error.h"
#include "part.h"
@@ -30,9 +31,7 @@
#include "vector.h"
#define NUM_VEC_PROC 2
-#define CACHE_ALIGN 64
#define C2_CACHE_SIZE (NUM_VEC_PROC * VEC_SIZE * 6) + (NUM_VEC_PROC * VEC_SIZE)
-#define C2_CACHE_ALIGN sizeof(float) * VEC_SIZE
#ifdef WITH_VECTORIZATION
/* Cache struct to hold a local copy of a cells' particle
@@ -40,31 +39,31 @@
struct cache {
/* Particle x position. */
- float *restrict x __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict x SWIFT_CACHE_ALIGN;
/* Particle y position. */
- float *restrict y __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict y SWIFT_CACHE_ALIGN;
/* Particle z position. */
- float *restrict z __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict z SWIFT_CACHE_ALIGN;
/* Particle smoothing length. */
- float *restrict h __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict h SWIFT_CACHE_ALIGN;
/* Particle mass. */
- float *restrict m __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict m SWIFT_CACHE_ALIGN;
/* Particle x velocity. */
- float *restrict vx __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict vx SWIFT_CACHE_ALIGN;
/* Particle y velocity. */
- float *restrict vy __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict vy SWIFT_CACHE_ALIGN;
/* Particle z velocity. */
- float *restrict vz __attribute__((aligned(CACHE_ALIGN)));
+ float *restrict vz SWIFT_CACHE_ALIGN;
/* Maximum index into neighbouring cell for particles that are in range. */
- int *restrict max_index __attribute__((aligned(CACHE_ALIGN)));
+ int *restrict max_index SWIFT_CACHE_ALIGN;
/* Cache size. */
int count;
@@ -75,28 +74,28 @@ struct cache {
struct c2_cache {
/* Separation between two particles squared. */
- float r2q[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float r2q[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* x separation between two particles. */
- float dxq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float dxq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* y separation between two particles. */
- float dyq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float dyq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* z separation between two particles. */
- float dzq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float dzq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* Mass of particle pj. */
- float mq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float mq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* x velocity of particle pj. */
- float vxq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float vxq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* y velocity of particle pj. */
- float vyq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float vyq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
/* z velocity of particle pj. */
- float vzq[C2_CACHE_SIZE] __attribute__((aligned(C2_CACHE_ALIGN)));
+ float vzq[C2_CACHE_SIZE] SWIFT_CACHE_ALIGN;
};
/**
@@ -130,15 +129,15 @@ __attribute__((always_inline)) INLINE void cache_init(struct cache *c,
free(c->max_index);
}
- error += posix_memalign((void **)&c->x, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->y, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->z, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->m, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->vx, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->vy, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->vz, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->h, CACHE_ALIGN, sizeBytes);
- error += posix_memalign((void **)&c->max_index, CACHE_ALIGN, sizeIntBytes);
+ error += posix_memalign((void **)&c->x, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->y, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->z, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->m, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->vx, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->vy, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->vz, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->h, SWIFT_CACHE_ALIGNMENT, sizeBytes);
+ error += posix_memalign((void **)&c->max_index, SWIFT_CACHE_ALIGNMENT, sizeIntBytes);
if (error != 0)
error("Couldn't allocate cache, no. of particles: %d", (int)count);
@@ -152,25 +151,39 @@ __attribute__((always_inline)) INLINE void cache_init(struct cache *c,
* @param ci_cache The cache.
*/
__attribute__((always_inline)) INLINE void cache_read_particles(
- const struct cell *const ci, struct cache *const ci_cache) {
+ const struct cell *restrict const ci, struct cache *restrict const ci_cache) {
#if defined(GADGET2_SPH)
+/* Let the compiler know that the data is aligned and create pointers to the
+ * arrays inside the cache. */
+swift_align_and_restrict_information(x, ci_cache->x, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(y, ci_cache->y, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(z, ci_cache->z, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(h, ci_cache->h, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(m, ci_cache->m, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(vx, ci_cache->vx, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(vy, ci_cache->vy, float, SWIFT_CACHE_ALIGNMENT);
+swift_align_and_restrict_information(vz, ci_cache->vz, float, SWIFT_CACHE_ALIGNMENT);
+
+const struct part *restrict parts = ci->parts;
+double loc[3];
+loc[0] = ci->loc[0];
+loc[1] = ci->loc[1];
+loc[2] = ci->loc[2];
+
/* Shift the particles positions to a local frame so single precision can be
* used instead of double precision. */
-#if defined(WITH_VECTORIZATION) && defined(__ICC)
-#pragma vector aligned
-#endif
for (int i = 0; i < ci->count; i++) {
- ci_cache->x[i] = ci->parts[i].x[0] - ci->loc[0];
- ci_cache->y[i] = ci->parts[i].x[1] - ci->loc[1];
- ci_cache->z[i] = ci->parts[i].x[2] - ci->loc[2];
- ci_cache->h[i] = ci->parts[i].h;
-
- ci_cache->m[i] = ci->parts[i].mass;
- ci_cache->vx[i] = ci->parts[i].v[0];
- ci_cache->vy[i] = ci->parts[i].v[1];
- ci_cache->vz[i] = ci->parts[i].v[2];
+ x[i] = (float)(parts[i].x[0] - loc[0]);
+ y[i] = (float)(parts[i].x[1] - loc[1]);
+ z[i] = (float)(parts[i].x[2] - loc[2]);
+ h[i] = parts[i].h;
+
+ m[i] = parts[i].mass;
+ vx[i] = parts[i].v[0];
+ vy[i] = parts[i].v[1];
+ vz[i] = parts[i].v[2];
}
#endif
@@ -322,13 +335,13 @@ __attribute__((always_inline)) INLINE void cache_read_two_cells_sorted(
* interaction.
*/
__attribute__((always_inline)) INLINE void cache_read_two_partial_cells_sorted(
- const struct cell *const ci, const struct cell *const cj,
- struct cache *const ci_cache, struct cache *const cj_cache,
+ const struct cell *restrict const ci, const struct cell *restrict const cj,
+ struct cache *restrict const ci_cache, struct cache *restrict const cj_cache,
const struct entry *restrict sort_i, const struct entry *restrict sort_j,
- const double *const shift, int *first_pi, int *last_pj,
+ const double *restrict const shift, int *first_pi, int *last_pj,
const int num_vec_proc) {
- int idx, ci_cache_idx;
+ int idx;
/* Pad number of particles read to the vector size. */
int rem = (ci->count - *first_pi) % (num_vec_proc * VEC_SIZE);
if (rem != 0) {
@@ -346,74 +359,95 @@ __attribute__((always_inline)) INLINE void cache_read_two_partial_cells_sorted(
int first_pi_align = *first_pi;
int last_pj_align = *last_pj;
-
-/* Shift the particles positions to a local frame (ci frame) so single precision
- * can be
- * used instead of double precision. Also shift the cell ci, particles positions
- * due to BCs but leave cell cj. */
-#if defined(WITH_VECTORIZATION) && defined(__ICC)
-#pragma vector aligned
-#endif
- for (int i = first_pi_align; i < ci->count; i++) {
- /* Make sure ci_cache is filled from the first element. */
- ci_cache_idx = i - first_pi_align;
- idx = sort_i[i].i;
- ci_cache->x[ci_cache_idx] = ci->parts[idx].x[0] - ci->loc[0] - shift[0];
- ci_cache->y[ci_cache_idx] = ci->parts[idx].x[1] - ci->loc[1] - shift[1];
- ci_cache->z[ci_cache_idx] = ci->parts[idx].x[2] - ci->loc[2] - shift[2];
- ci_cache->h[ci_cache_idx] = ci->parts[idx].h;
-
- ci_cache->m[ci_cache_idx] = ci->parts[idx].mass;
- ci_cache->vx[ci_cache_idx] = ci->parts[idx].v[0];
- ci_cache->vy[ci_cache_idx] = ci->parts[idx].v[1];
- ci_cache->vz[ci_cache_idx] = ci->parts[idx].v[2];
+ const struct part *restrict parts_i = ci->parts;
+ const struct part *restrict parts_j = cj->parts;
+ double loc[3];
+ loc[0] = ci->loc[0];
+ loc[1] = ci->loc[1];
+ loc[2] = ci->loc[2];
+
+/* Let the compiler know that the data is aligned and create pointers to the
+ * arrays inside the cache. */
+ swift_align_and_restrict_information(x, ci_cache->x, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(y, ci_cache->y, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(z, ci_cache->z, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(h, ci_cache->h, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(m, ci_cache->m, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(vx, ci_cache->vx, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(vy, ci_cache->vy, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(vz, ci_cache->vz, float, SWIFT_CACHE_ALIGNMENT);
+
+ int ci_cache_count = ci->count - first_pi_align;
+ /* Shift the particles positions to a local frame (ci frame) so single precision
+ * can be
+ * used instead of double precision. Also shift the cell ci, particles positions
+ * due to BCs but leave cell cj. */
+ for (int i = 0; i < ci_cache_count; i++) {
+ idx = sort_i[i + first_pi_align].i;
+ x[i] = (float)(parts_i[idx].x[0] - loc[0] - shift[0]);
+ y[i] = (float)(parts_i[idx].x[1] - loc[1] - shift[1]);
+ z[i] = (float)(parts_i[idx].x[2] - loc[2] - shift[2]);
+ h[i] = parts_i[idx].h;
+
+ m[i] = parts_i[idx].mass;
+ vx[i] = parts_i[idx].v[0];
+ vy[i] = parts_i[idx].v[1];
+ vz[i] = parts_i[idx].v[2];
}
/* Pad cache with fake particles that exist outside the cell so will not
* interact.*/
- float fake_pix = 2.0f * ci->parts[sort_i[ci->count - 1].i].x[0];
+ float fake_pix = 2.0f * parts_i[sort_i[ci->count - 1].i].x[0];
for (int i = ci->count - first_pi_align;
i < ci->count - first_pi_align + VEC_SIZE; i++) {
- ci_cache->x[i] = fake_pix;
- ci_cache->y[i] = 1.f;
- ci_cache->z[i] = 1.f;
- ci_cache->h[i] = 1.f;
-
- ci_cache->m[i] = 1.f;
- ci_cache->vx[i] = 1.f;
- ci_cache->vy[i] = 1.f;
- ci_cache->vz[i] = 1.f;
+ x[i] = fake_pix;
+ y[i] = 1.f;
+ z[i] = 1.f;
+ h[i] = 1.f;
+
+ m[i] = 1.f;
+ vx[i] = 1.f;
+ vy[i] = 1.f;
+ vz[i] = 1.f;
}
+
+/* Let the compiler know that the data is aligned and create pointers to the
+ * arrays inside the cache. */
+ swift_align_and_restrict_information(xj, cj_cache->x, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(yj, cj_cache->y, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(zj, cj_cache->z, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(hj, cj_cache->h, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(mj, cj_cache->m, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(vxj, cj_cache->vx, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(vyj, cj_cache->vy, float, SWIFT_CACHE_ALIGNMENT);
+ swift_align_and_restrict_information(vzj, cj_cache->vz, float, SWIFT_CACHE_ALIGNMENT);
-#if defined(WITH_VECTORIZATION) && defined(__ICC)
-#pragma vector aligned
-#endif
for (int i = 0; i <= last_pj_align; i++) {
idx = sort_j[i].i;
- cj_cache->x[i] = cj->parts[idx].x[0] - ci->loc[0];
- cj_cache->y[i] = cj->parts[idx].x[1] - ci->loc[1];
- cj_cache->z[i] = cj->parts[idx].x[2] - ci->loc[2];
- cj_cache->h[i] = cj->parts[idx].h;
-
- cj_cache->m[i] = cj->parts[idx].mass;
- cj_cache->vx[i] = cj->parts[idx].v[0];
- cj_cache->vy[i] = cj->parts[idx].v[1];
- cj_cache->vz[i] = cj->parts[idx].v[2];
+ xj[i] = (float)(parts_j[idx].x[0] - loc[0]);
+ yj[i] = (float)(parts_j[idx].x[1] - loc[1]);
+ zj[i] = (float)(parts_j[idx].x[2] - loc[2]);
+ hj[i] = parts_j[idx].h;
+
+ mj[i] = parts_j[idx].mass;
+ vxj[i] = parts_j[idx].v[0];
+ vyj[i] = parts_j[idx].v[1];
+ vzj[i] = parts_j[idx].v[2];
}
/* Pad cache with fake particles that exist outside the cell so will not
* interact.*/
float fake_pjx = 2.0f * cj->parts[sort_j[cj->count - 1].i].x[0];
for (int i = last_pj_align + 1; i < last_pj_align + 1 + VEC_SIZE; i++) {
- cj_cache->x[i] = fake_pjx;
- cj_cache->y[i] = 1.f;
- cj_cache->z[i] = 1.f;
- cj_cache->h[i] = 1.f;
-
- cj_cache->m[i] = 1.f;
- cj_cache->vx[i] = 1.f;
- cj_cache->vy[i] = 1.f;
- cj_cache->vz[i] = 1.f;
+ xj[i] = fake_pjx;
+ yj[i] = 1.f;
+ zj[i] = 1.f;
+ hj[i] = 1.f;
+
+ mj[i] = 1.f;
+ vxj[i] = 1.f;
+ vyj[i] = 1.f;
+ vzj[i] = 1.f;
}
}