squiggle.c

commit 3bb2804ccf2db29dfe59ba077fe731d9697762af
parent eb1c59261072d5d9e037b66bcd51a5d183762df5
Author: NunoSempere <nuno.sempere@protonmail.com>
Date:   Sat, 13 Jan 2024 00:50:27 +0100

save squiggle_more tweaks

Diffstat:
M
squiggle_more.c
 | 
25
+++++++++++++++++++++----

1 file changed, 21 insertions(+), 4 deletions(-)
diff --git a/squiggle_more.c b/squiggle_more.c
@@ -60,13 +60,13 @@ void sampler_parallel(double (*sampler)(uint64_t* seed), double* results, int n_
     }
 
     int i;
-#pragma omp parallel private(i, quotient)
+#pragma omp parallel private(i)
     {
-#pragma omp for
+#pragma omp for 
         for (i = 0; i < n_threads; i++) {
-            int quotient = n_samples / n_threads;
             int lower_bound_inclusive = i * quotient;
             int upper_bound_not_inclusive = ((i + 1) * quotient); // note the < in the for loop below,
+
             for (int j = lower_bound_inclusive; j < upper_bound_not_inclusive; j++) {
                 results[j] = sampler(&(cache_box[i].seed));
                 // In principle, these results[j] could also result in two threads competing for the same cache line.
@@ -75,6 +75,23 @@ void sampler_parallel(double (*sampler)(uint64_t* seed), double* results, int n_
                 // b) trying to unroll loops actually makes the code slower
                 // c) 8 results[j] are 8 doubles, which fit a cache line. If n_samples/n_threads
             }
+            // Failed loop unrolling
+            /*
+            for (int j = lower_bound_inclusive; j < 4*(upper_bound_not_inclusive/4); j++) {
+                results[j+0] = sampler(&(cache_box[i].seed));
+                results[j+1] = sampler(&(cache_box[i].seed));
+                results[j+2] = sampler(&(cache_box[i].seed));
+                results[j+3] = sampler(&(cache_box[i].seed));
+                // In principle, these results[j] could also result in two threads competing for the same cache line.
+                // In practice, though,
+                // a) this would happen infrequently
+                // b) trying to unroll loops actually makes the code slower
+                // c) 8 results[j] are 8 doubles, which fit a cache line. If n_samples/n_threads
+            }
+            for (int j = 4*(upper_bound_not_inclusive/4); j < upper_bound_not_inclusive; j++) {
+                results[j] = sampler(&(cache_box[i].seed));
+            }
+            */
         }
     }
     for (int j = divisor_multiple; j < n_samples; j++) {
@@ -92,7 +109,7 @@ typedef struct ci_t {
     double high;
 } ci;
 
-static void swp(int i, int j, double xs[])
+inline static void swp(int i, int j, double xs[])
 {
     double tmp = xs[i];
     xs[i] = xs[j];