summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorMathis Rosenhauer <rosenhauer@dkrz.de>2014-10-27 14:48:20 (GMT)
committerMathis Rosenhauer <rosenhauer@dkrz.de>2014-10-27 14:48:20 (GMT)
commit1763545e66b3612c0bf6d13b5b1b28e7485d429e (patch)
treee77d384483ad64430838acdda0c600f0dfc41df4
parenta63f50c1a97b7c92aad1d3e641b94c6172f675ad (diff)
downloadlibaec-1763545e66b3612c0bf6d13b5b1b28e7485d429e.zip
libaec-1763545e66b3612c0bf6d13b5b1b28e7485d429e.tar.gz
libaec-1763545e66b3612c0bf6d13b5b1b28e7485d429e.tar.bz2
Blurb about why 32bit is slow.
-rw-r--r--doc/README13
1 files changed, 13 insertions, 0 deletions
diff --git a/doc/README b/doc/README
new file mode 100644
index 0000000..4a96d8e
--- /dev/null
+++ b/doc/README
@@ -0,0 +1,13 @@
+---------------------
+ 32 bits vs. 64 bits
+---------------------
+
+libaec compiles and runs both in 32 and 64 bits. For encoding and
+decoding I see a performance degradation with 32 bit binaries of 38%
+and 33% resp.
+
+The main reason for this is that encoding and decoding use a buffer of
+type uin64_t to align a potentially 32 bit wide sample with the output
+(encoding) or input (decoding) stream. Avoiding the use of this
+uin64_t buffer variable would degrade 64 bit performance for a modest
+gain in 32 bit performance so I decided against it.