From 2b5b496c25177b070c247e977d338bb0eea20704 Mon Sep 17 00:00:00 2001
From: NunoSempere <nuno.sempere@protonmail.com>
Date: Fri, 12 Jan 2024 00:23:01 +0100
Subject: [PATCH] seed cache box name, make "quotient" private var

---
 squiggle_more.c | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)

diff --git a/squiggle_more.c b/squiggle_more.c
index 0146707..766a66f 100644
--- a/squiggle_more.c
+++ b/squiggle_more.c
@@ -10,10 +10,11 @@
 
 /* Parallel sampler */
 #define CACHE_LINE_SIZE 64
-typedef struct padded_seed_t {
+typedef struct seed_cache_box_t {
     uint64_t* seed;
     char padding[CACHE_LINE_SIZE - sizeof(uint64_t*)];
-} padded_seed;
+} seed_cache_box;
+// This avoid false sharing. Dealing with this shaves ~2ms.
 
 void sampler_parallel(double (*sampler)(uint64_t* seed), double* results, int n_threads, int n_samples)
 {
@@ -36,14 +37,14 @@ void sampler_parallel(double (*sampler)(uint64_t* seed), double* results, int n_
     int divisor_multiple = quotient * n_threads;
 
     // uint64_t** seeds = malloc((size_t)n_threads * sizeof(uint64_t*));
-    padded_seed* seeds = (padded_seed*) malloc(sizeof(padded_seed) * (size_t)n_threads);
+    seed_cache_box* cache_box = (seed_cache_box*) malloc(sizeof(seed_cache_box) * (size_t)n_threads);
     srand(1);
     for (int i = 0; i < n_threads; i++) {
-        seeds[i].seed = malloc(sizeof(uint64_t*));
+        cache_box[i].seed = malloc(sizeof(uint64_t*));
         // Constraints:
         // - xorshift can't start with 0
         // - the seeds should be reasonably separated and not correlated
-        *(seeds[i].seed) = (uint64_t)rand() * (UINT64_MAX / RAND_MAX);
+        *(cache_box[i].seed) = (uint64_t)rand() * (UINT64_MAX / RAND_MAX);
         // printf("#%ld: %lu\n",i, *seeds[i]);
 
         // Other initializations tried:
@@ -53,26 +54,26 @@ void sampler_parallel(double (*sampler)(uint64_t* seed), double* results, int n_
     }
 
     int i;
-#pragma omp parallel private(i)
+#pragma omp parallel private(i, quotient)
     {
 #pragma omp for
         for (i = 0; i < n_threads; i++) {
             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(seeds[i].seed);
+                results[j] = sampler(cache_box[i].seed);
             }
         }
     }
     for (int j = divisor_multiple; j < n_samples; j++) {
-        results[j] = sampler(seeds[0].seed);
+        results[j] = sampler(cache_box[0].seed);
         // we can just reuse a seed, this isn't problematic because we are not doing multithreading
     }
 
     for (int i = 0; i < n_threads; i++) {
-        free(seeds[i].seed);
+        free(cache_box[i].seed);
     }
-    free(seeds);
+    free(cache_box);
 }
 
 /* Get confidence intervals, given a sampler */