After research, I find the F16C feature is only supported on
IVB later platforms and needs at least GCC 4.6 support.
This will cause some compatible issues. The beignet project
may be cross-compiled on some old platform using old version
GCC, and some one may use other compiler to build it.
According to this, I still prefer to use the software imitation
for half float. It's verbose, but it will not have any side effect
for runtime performance.
On 2015年05月22日 14:28, He Junyan wrote:
Thanks for your information.
I will do some research for it.
On 2015年05月22日 05:51, Matt Turner wrote:
On Thu, May 21, 2015 at 1:25 AM, <junyan...@inbox.com> wrote:
From: Junyan He <junyan...@linux.intel.com>
Because the CPU of X86 does not support half float
instructions, there is no support for half float operations.
So we introduce the half class to handle the operations for
half float using llvm's APFloat utility.
Ivybridge and newer have the F16C instruction set
(http://en.wikipedia.org/wiki/F16C) which offers instructions to
convert half-precision <-> single-precision floats.
I don't know if it's valuable to use it, but it's there.
_______________________________________________
Beignet mailing list
Beignet@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/beignet
_______________________________________________
Beignet mailing list
Beignet@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/beignet
_______________________________________________
Beignet mailing list
Beignet@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/beignet
