From f21428e20f2dc9ea6a95bf0efe8a6d717ff85978 Mon Sep 17 00:00:00 2001 From: Romain Goyet Date: Thu, 3 Aug 2017 14:59:34 +0200 Subject: [PATCH] [liba] More detailed README for the externals Change-Id: I9c5c4d9e86933aa81e3f45048bb004acafed41bc --- liba/src/external/README.txt | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/liba/src/external/README.txt b/liba/src/external/README.txt index b4f6256ad..b4c29ee07 100644 --- a/liba/src/external/README.txt +++ b/liba/src/external/README.txt @@ -1,2 +1,15 @@ +"softfloat": Unaltered files from SoftFloat v3c. +http://www.jhauser.us/arithmetic/SoftFloat.html + "openbsd": Unaltered files from OpenBSD 6.0. Original path is "lib/libm/src" "openbsd/include": Compatibility headers needed to build files from OpenBSD. +http://www.openbsd.org + +Why use OpenBSD's libm? First and foremost, nearly everyone use fdlibm in a way +or another, and OpenBSD is no exception. That being said, OpenBSD comes with a +single-precision variant (which fdlibm itself lacks). And more interestingly, +this variant does all its computation in single-precision (it never upgrades to +double-precision). In our case, this is very interesting because we're doing +single-precision computation when we better performance at the cost of accuracy, +and in our cse our hardware has a single-precision FPU so switching to double +can yield much lower performances.