Revision: 55056
http://brlcad.svn.sourceforge.net/brlcad/?rev=55056&view=rev
Author: brlcad
Date: 2013-04-05 05:17:28 +0000 (Fri, 05 Apr 2013)
Log Message:
-----------
save a few more instructions and reduce per-real-allocation overhead slightly
by not tracking per page usage. just keep track of the last page being filled
since we know the rest are full.
Modified Paths:
--------------
brlcad/trunk/src/libbu/heap.c
Modified: brlcad/trunk/src/libbu/heap.c
===================================================================
--- brlcad/trunk/src/libbu/heap.c 2013-04-05 04:20:05 UTC (rev 55055)
+++ brlcad/trunk/src/libbu/heap.c 2013-04-05 05:17:28 UTC (rev 55056)
@@ -67,14 +67,10 @@
size_t pages;
/**
- * used is an array of page sizes for tabulating how much memory
- * each page is currently using (allocate). not a counter so we
- * avoid a multiply.
+ * used tabulates how much memory the current page is using
+ * (allocated). it's not a counter so we can avoid a multiply.
*/
- size_t *used;
-
- /** keep track of our allocation counts for reporting stats */
- size_t alloc;
+ size_t used;
};
struct cpus {
@@ -89,7 +85,7 @@
* store data in a cpu-specific structure so we can avoid the need for
* mutex locking entirely. relies on static zero-initialization.
*/
-static struct cpus per_cpu[MAX_PSW] = {{{{0, 0, 0, 0}}, 0}};
+static struct cpus per_cpu[MAX_PSW] = {{{{0, 0, 0}}, 0}};
static void
@@ -97,11 +93,11 @@
{
static int printed = 0;
- size_t h, i, j;
+ size_t h, i;
+ size_t allocs = 0;
size_t misses = 0;
- size_t allocations = 0;
+ size_t got = 0;
size_t total_pages = 0;
- size_t total_alloc = 0;
size_t ncpu = bu_avail_cpus();
/* this may get registered for atexit() multiple times, so make
@@ -117,24 +113,27 @@
for (h=0; h < ncpu; h++) {
for (i=0; i < BINS; i++) {
- allocations = 0;
- for (j=0; j < per_cpu[h].bin[i].pages; j++) {
- allocations += per_cpu[h].bin[i].used[j] / (i+1);
+
+ /* capacity across all pages */
+ got = per_cpu[h].bin[i].pages * (PAGESIZE/(i+1));
+
+ if (got > 0) {
+ /* last page is partial */
+ got -= (PAGESIZE - per_cpu[h].bin[i].used)/(i+1);
+ bu_log("%04zu [%02zu] => %zu\n", i, per_cpu[h].bin[i].pages,
got);
+ allocs += got;
}
- if (allocations > 0)
- bu_log("%04zd [%02zd] => %zd\n", i, per_cpu[h].bin[i].pages,
allocations);
total_pages += per_cpu[h].bin[i].pages;
- total_alloc += allocations;
}
misses += per_cpu[h].misses;
}
bu_log("-----------------------\n"
- "size [pages] => allocs\n"
+ "size [pages] => count\n"
"Heap range: 1-%d bytes\n"
"Page size: %d bytes\n"
- "Pages: %zd (%.2lfMB)\n"
- "%zd hits, %zd misses\n"
- "=======================\n", BINS, PAGESIZE, total_pages,
(double)(total_pages * PAGESIZE) / (1024.0*1024.0), total_alloc, misses);
+ "Pages: %zu (%.2lfMB)\n"
+ "%zu allocs, %zu misses\n"
+ "=======================\n", BINS, PAGESIZE, total_pages,
(double)(total_pages * PAGESIZE) / (1024.0*1024.0), allocs, misses);
}
@@ -164,7 +163,6 @@
}
bin = &per_cpu[oncpu].bin[smo];
- bin->alloc++;
/* init */
if (bin->pages == 0) {
@@ -177,22 +175,20 @@
bin->pages++;
bin->heaps = (char **)bu_malloc(1 * sizeof(char *), "heap malloc
heaps[]");
bin->heaps[0] = (char *)bu_calloc(1, PAGESIZE, "heap calloc
heaps[][0]");
- bin->used = (size_t *)bu_malloc(1 * sizeof(size_t), "heap malloc
used[]");
- bin->used[0] = 0;
+ bin->used = 0;
}
/* grow */
- if (bin->used[bin->pages-1]+sz > PAGESIZE) {
+ if (bin->used+sz > PAGESIZE) {
bin->pages++;
bin->heaps = (char **)bu_realloc(bin->heaps, bin->pages * sizeof(char
*), "heap realloc heaps[]");
bin->heaps[bin->pages-1] = (char *)bu_calloc(1, PAGESIZE, "heap calloc
heaps[][]");
- bin->used = (size_t *)bu_realloc(bin->used, bin->pages *
sizeof(size_t), "heap realloc used[]");
- bin->used[bin->pages-1] = 0;
+ bin->used = 0;
}
/* give */
- ret = &(bin->heaps[bin->pages-1][bin->used[bin->pages-1]]);
- bin->used[bin->pages-1] += sz;
+ ret = &(bin->heaps[bin->pages-1][bin->used]);
+ bin->used += sz;
return (void *)ret;
}
This was sent by the SourceForge.net collaborative development platform, the
world's largest Open Source development site.
------------------------------------------------------------------------------
Minimize network downtime and maximize team effectiveness.
Reduce network management and security costs.Learn how to hire
the most talented Cisco Certified professionals. Visit the
Employer Resources Portal
http://www.cisco.com/web/learning/employer_resources/index.html
_______________________________________________
BRL-CAD Source Commits mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/brlcad-commits
