diff --git a/src/engine.c b/src/engine.c
index eb936f61b296c94d933855de874cc3949dce89bc..121d1419c337a2644f349cb137169031aab83e4d 100644
--- a/src/engine.c
+++ b/src/engine.c
@@ -3903,61 +3903,61 @@ void engine_split(struct engine *e, struct partition *initial_partition) {
* @param e The #engine.
*/
void engine_dump_snapshot(struct engine *e) {
-
- struct clocks_time time1, time2;
- clocks_gettime(&time1);
-
-#ifdef SWIFT_DEBUG_CHECKS
- /* Check that all cells have been drifted to the current time.
- * That can include cells that have not
- * previously been active on this rank. */
- space_check_drift_point(e->s, e->ti_current, /* check_mpole=*/0);
-
- /* Be verbose about this */
- if (e->nodeID == 0) {
- if (e->policy & engine_policy_cosmology)
- message("Dumping snapshot at a=%e",
- exp(e->ti_current * e->time_base) * e->cosmology->a_begin);
- else
- message("Dumping snapshot at t=%e",
- e->ti_current * e->time_base + e->time_begin);
- }
-#else
- if (e->verbose) {
- if (e->policy & engine_policy_cosmology)
- message("Dumping snapshot at a=%e",
- exp(e->ti_current * e->time_base) * e->cosmology->a_begin);
- else
- message("Dumping snapshot at t=%e",
- e->ti_current * e->time_base + e->time_begin);
- }
-#endif
-
-/* Dump... */
-#if defined(HAVE_HDF5)
-#if defined(WITH_MPI)
-#if defined(HAVE_PARALLEL_HDF5)
- write_output_parallel(e, e->snapshot_base_name, e->internal_units,
- e->snapshot_units, e->nodeID, e->nr_nodes,
- MPI_COMM_WORLD, MPI_INFO_NULL);
-#else
- write_output_serial(e, e->snapshot_base_name, e->internal_units,
- e->snapshot_units, e->nodeID, e->nr_nodes, MPI_COMM_WORLD,
- MPI_INFO_NULL);
-#endif
-#else
- write_output_single(e, e->snapshot_base_name, e->internal_units,
- e->snapshot_units);
-#endif
-#endif
-
- /* Flag that we dumped a snapshot */
- e->step_props |= engine_step_prop_snapshot;
-
- clocks_gettime(&time2);
- if (e->verbose)
- message("writing particle properties took %.3f %s.",
- (float)clocks_diff(&time1, &time2), clocks_getunit());
+//
+// struct clocks_time time1, time2;
+// clocks_gettime(&time1);
+//
+//#ifdef SWIFT_DEBUG_CHECKS
+// /* Check that all cells have been drifted to the current time.
+// * That can include cells that have not
+// * previously been active on this rank. */
+// space_check_drift_point(e->s, e->ti_current, /* check_mpole=*/0);
+//
+// /* Be verbose about this */
+// if (e->nodeID == 0) {
+// if (e->policy & engine_policy_cosmology)
+// message("Dumping snapshot at a=%e",
+// exp(e->ti_current * e->time_base) * e->cosmology->a_begin);
+// else
+// message("Dumping snapshot at t=%e",
+// e->ti_current * e->time_base + e->time_begin);
+// }
+//#else
+// if (e->verbose) {
+// if (e->policy & engine_policy_cosmology)
+// message("Dumping snapshot at a=%e",
+// exp(e->ti_current * e->time_base) * e->cosmology->a_begin);
+// else
+// message("Dumping snapshot at t=%e",
+// e->ti_current * e->time_base + e->time_begin);
+// }
+//#endif
+//
+///* Dump... */
+//#if defined(HAVE_HDF5)
+//#if defined(WITH_MPI)
+//#if defined(HAVE_PARALLEL_HDF5)
+// write_output_parallel(e, e->snapshot_base_name, e->internal_units,
+// e->snapshot_units, e->nodeID, e->nr_nodes,
+// MPI_COMM_WORLD, MPI_INFO_NULL);
+//#else
+// write_output_serial(e, e->snapshot_base_name, e->internal_units,
+// e->snapshot_units, e->nodeID, e->nr_nodes, MPI_COMM_WORLD,
+// MPI_INFO_NULL);
+//#endif
+//#else
+// write_output_single(e, e->snapshot_base_name, e->internal_units,
+// e->snapshot_units);
+//#endif
+//#endif
+//
+// /* Flag that we dumped a snapshot */
+// e->step_props |= engine_step_prop_snapshot;
+//
+// clocks_gettime(&time2);
+// if (e->verbose)
+// message("writing particle properties took %.3f %s.",
+// (float)clocks_diff(&time1, &time2), clocks_getunit());
}
/**
diff --git a/src/kernel_hydro.h b/src/kernel_hydro.h
index 3d5ec6ac84a77941739f5d3b57ed0340c831c061..bff83fa569a78b638decdf486ad4852538f05aa2 100644
--- a/src/kernel_hydro.h
+++ b/src/kernel_hydro.h
@@ -517,6 +517,7 @@ __attribute__((always_inline)) INLINE static void kernel_deval_1_vec(
* when the mask is 0.*/
w->v = vec_blend(mask_reg, w->v, w2.v);
dw_dx->v = vec_blend(mask_reg, dw_dx->v, dw_dx2.v);
+ print_vector(*w);
#else
#error "Vectorisation not supported for this kernel!!!"
diff --git a/src/runner_doiact_vec.c b/src/runner_doiact_vec.c
index 182e81e99c442cf5e27405ea71321e22e7f374e3..cb5c96cc3776f01b0903074266778188a0a76d9e 100644
--- a/src/runner_doiact_vec.c
+++ b/src/runner_doiact_vec.c
@@ -1491,6 +1491,7 @@ void runner_dopair1_density_vec(struct runner *r, struct cell *ci,
v_r2.v = vec_mul(v_dx.v, v_dx.v);
v_r2.v = vec_fma(v_dy.v, v_dy.v, v_r2.v);
v_r2.v = vec_fma(v_dz.v, v_dz.v, v_r2.v);
+ print_vector(v_r2);
mask_t v_doi_mask;
diff --git a/src/vector.h b/src/vector.h
index 9e56e9e2e3545f8877df0c9098287dfdf1a30112..60798af8e59382cc40c4f0e811ee5e92c692eb3c 100644
--- a/src/vector.h
+++ b/src/vector.h
@@ -19,6 +19,7 @@
******************************************************************************/
#ifndef SWIFT_VECTOR_H
#define SWIFT_VECTOR_H
+#include <stdio.h>
/* Config parameters. */
#include "../config.h"
@@ -405,24 +406,26 @@
#include <arm_neon.h>
#define VEC_SIZE 4
#define VEC_FLOAT float32x4_t
-#define VEC_DB float64x2
+#define VEC_DBL float64x2_t
#define VEC_INT int32x4_t
+#define VEC_UINT uint32x4_t
#define vec_load(a) vld1q_f32(a)
#define vec_store(a, adds) vst1q_f32(addr,a)
-#define vec_set_zero() vmovq_n_f32(0.0f)
-#define vec_set1(a) vld1q_dup_f32(a)
+#define vec_setzero() vmovq_n_f32(0.0f)
+#define vec_set1(a) vmovq_n_f32(a)
+#define vec_setint1(a) vmovq_n_s32(a)
#define vec_add(a, b) vaddq_f32(a,b)
-#define vec_add_mask(a, b, mask) vec_add(a, ((vector)vec_and(((vector)b).m,mask)).v)
+#define vec_mask_add(a, b, mask) vec_add(a, ((vector)vec_and(((vector)b).m,mask)).v)
#define vec_sub(a, b) vsubq_f32(a, b)
#define vec_mask_sub(a, b, mask) vec_sub(a, ((vector)vec_and(((vector)b).m,mask)).v)
#define vec_mul(a, b) vmulq_f32(a, b)
#define vec_div(a, b) vdivq_f32(a, b)
-#define vec_sqrt(a) vsqrt_f32(a)
+#define vec_sqrt(a) vsqrtq_f32(a)
#define vec_rcp(a) vrecpeq_f32(a)
-#define vec_rsqrt(a) vrsqrte_f32(a)
+#define vec_rsqrt(a) vrsqrteq_f32(a)
#define vec_ftoi(a) vcvtq_s32_f32(a)
-#define vec_fmin(a, b) vpmin_f32(a, b)
-#define vec_fmax(a, b) vpmax_f32(q, b)
+#define vec_fmin(a, b) vpminq_f32(a, b)
+#define vec_fmax(a, b) vpmaxq_f32(a, b)
#define vec_fabs(a) vabsq_f32(a)
#define vec_floor(a) vcvtq_f32_s32(vcvtmq_s32_f32(a))
#define vec_cmp_gt(a, b) vcgtzq_f32(vec_sub(a,b))
@@ -431,27 +434,30 @@
#define vec_cmp_lt(a, b) vcgtzq_f32(vec_sub(b,a))
#define vec_cmp_lte(a, b) vcgezq_f32(vec_sub(b,a))
#define vec_cmp_result(a) vec_not(vceqzq_f32(a))
-#define vec_is_mask_true(a) vec_not(vceqzq_f32(a.v))
-#define vec_and(a, b) vandq_s32(a, b)
+#define vec_is_mask_true(a) vaddvq_s32(a.m)
+// Write vector conversion function?
+#define vec_create_mask(mask, cond) mask.m = ((vector)cond).m
+#define vec_and(a, b) vandq_s32(((vector)a).m, ((vector)b).m)
#define vec_mask_and(a, b) vec_and(a.v, b.v)
-#define vec_and_mask(a, mask) vec_and( ((vector)a).v, mask )
+#define vec_and_mask(a, mask) vcvtq_f32_s32(vec_and(a, mask.v))
#define vec_init_mask_true(mask) mask.m = vec_setint1(0xFFFFFFFF)
#define vec_combine_masks(mask1, mask2) \
- ({ mask1.v = vec_mask_and(mask1,mask2); })
+ ({ mask1.v = vcvtq_f32_s32(vec_mask_and(mask1,mask2)); })
#define vec_zero_mask(mask) mask.v = vec_setzero()
#define vec_pad_mask(mask, pad) \
for(int i = VEC_SIZE - (pad); i < VEC_SIZE; i++) mask.i[i] = 0
+// Change to vector conversion function?
#define vec_blend(mask, a, b) \
- ((vector)vec_or( vec_and(mask.m, ((vector)b).m ), vec_and(vec_not(mask.v), ((vector)a).m ).v)
-#define vec_or(a, b) vorrq_u32(a, b)
-#define vec_not(a) vmvnq_u32(a)
+ ((vector)vec_or( vec_and(mask.m, ((vector)b).m ), vcvtq_s32_f32(((vector)vec_and(vec_not(mask.m), ((vector)a).m )).v))).v
+#define vec_or(a, b) vorrq_s32(((vector)a).m, ((vector)b).m)
+#define vec_not(a) vmvnq_s32(((vector)a).m)
#define FILL_VEC(a) \
{ .f[0] = a, .f[1] = a, .f[2] = a, .f[3] = a}
-#define VEC_HADD(a,b) b += vaddvq_f32(a)
+#define VEC_HADD(a,b) b += vaddvq_f32(((vector)a).v)
-#define VEC_HMAX(a,b) b = max(b,vmaxvq_f32(a))
+#define VEC_HMAX(a,b) b = max(b,vmaxvq_f32(((vector)a).v))
#define vec_fma(a, b, c) vec_add(vec_mul(a,b),c)
#define vec_fnma(a,b,c) vec_sub(c, vec_mul(a,b))
@@ -465,11 +471,16 @@ typedef union {
VEC_FLOAT v;
VEC_DBL vd;
VEC_INT m;
+ VEC_UINT um;
float f[VEC_SIZE];
double d[VEC_SIZE / 2];
int i[VEC_SIZE];
} vector;
+inline void print_vector(vector v) {
+ printf("vector %.5e %.5e %.5e %.5e\n", v.f[0], v.f[1], v.f[2], v.f[3]);
+ fflush(stdout);
+}
/* Define the mask type depending on the instruction set used. */
#ifdef HAVE_AVX512_F
typedef __mmask16 mask_t;
diff --git a/tests/test27cells.c b/tests/test27cells.c
index cc34f503304feb56799a2d31baa3416b940202d3..82ec184fafb711bc6ce3f710529a89ac0e52d148 100644
--- a/tests/test27cells.c
+++ b/tests/test27cells.c
@@ -385,13 +385,13 @@ int main(int argc, char *argv[]) {
/* Choke on FP-exceptions */
#ifdef HAVE_FE_ENABLE_EXCEPT
- feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);
+ //feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);
#endif
/* Get some randomness going */
srand(0);
- char c;
+ signed char c;
while ((c = getopt(argc, argv, "m:s:h:p:n:r:t:d:f:v:")) != -1) {
switch (c) {
case 'h':
@@ -594,11 +594,11 @@ int main(int argc, char *argv[]) {
ticks face_time = timings[4] + timings[10] + timings[12] + timings[14] +
timings[16] + timings[22];
- message("Corner calculations took : %15lli ticks.", corner_time / runs);
- message("Edge calculations took : %15lli ticks.", edge_time / runs);
- message("Face calculations took : %15lli ticks.", face_time / runs);
- message("Self calculations took : %15lli ticks.", timings[13] / runs);
- message("SWIFT calculation took : %15lli ticks.", time / runs);
+ message("Corner calculations took : %15lu ticks.", corner_time / runs);
+ message("Edge calculations took : %15lu ticks.", edge_time / runs);
+ message("Face calculations took : %15lu ticks.", face_time / runs);
+ message("Self calculations took : %15lu ticks.", timings[13] / runs);
+ message("SWIFT calculation took : %15lu ticks.", time / runs);
/* Now perform a brute-force version for accuracy tests */
@@ -624,7 +624,7 @@ int main(int argc, char *argv[]) {
dump_particle_fields(outputFileName, main_cell, cells);
/* Output timing */
- message("Brute force calculation took : %15lli ticks.", toc - tic);
+ message("Brute force calculation took : %15lu ticks.", toc - tic);
/* Clean things to make the sanitizer happy ... */
for (int i = 0; i < 27; ++i) clean_up(cells[i]);
diff --git a/tests/testInteractions.c b/tests/testInteractions.c
index e14fddd640764c7e22a217fb483791494ba4fae0..8046b734d2d38a64514c67df925f00d9cb06710e 100644
--- a/tests/testInteractions.c
+++ b/tests/testInteractions.c
@@ -405,9 +405,9 @@ void test_interactions(struct part test_part, struct part *parts, size_t count,
if (check_results(pi_serial, pj_serial, pi_vec, pj_vec, count))
message("Differences found...");
- message("The serial interactions took : %15lli ticks.",
+ message("The serial interactions took : %15lu ticks.",
serial_time / runs);
- message("The vectorised interactions took : %15lli ticks.", vec_time / runs);
+ message("The vectorised interactions took : %15lu ticks.", vec_time / runs);
message("Speed up: %15fx.", (double)(serial_time) / vec_time);
}
@@ -677,9 +677,9 @@ void test_force_interactions(struct part test_part, struct part *parts,
if (check_results(pi_serial, pj_serial, pi_vec, pj_vec, count))
message("Differences found...");
- message("The serial interactions took : %15lli ticks.",
+ message("The serial interactions took : %15lu ticks.",
serial_time / runs);
- message("The vectorised interactions took : %15lli ticks.", vec_time / runs);
+ message("The vectorised interactions took : %15lu ticks.", vec_time / runs);
message("Speed up: %15fx.", (double)(serial_time) / vec_time);
}
@@ -693,7 +693,7 @@ int main(int argc, char *argv[]) {
/* Get some randomness going */
srand(0);
- char c;
+ signed char c;
while ((c = getopt(argc, argv, "h:s:n:r:")) != -1) {
switch (c) {
case 'h':