portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-14 13:57:07 +01:00
Code Releases Activity

If no f16c instructions supported must use software half precision conversion.

Browse Source 
This will also become useful on BG/Q, so will move out from SSE4 into a general area.
Lifted the Eigen half precision from web. Looks sensible, but not extensively regressed
against the intrinsics implementation yet.
This commit is contained in:
paboyle
2017-04-20 15:30:52 +01:00
parent e1a2319d01
commit 3844bcf800
2 changed files with 113 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
configure.ac
View File

@ -84,16 +84,15 @@ case ${ac_LAPACK} in
esac
############### FP16 conversions
AC_ARG_ENABLE([fp16],
[AC_HELP_STRING([--enable-fp16=yes|no], [enable fp16 comms])],
[ac_FP16=${enable_fp16}], [ac_FP16=no])
case ${ac_FP16} in
no)
;;
AC_ARG_ENABLE([sfw-fp16],
[AC_HELP_STRING([--enable-sfw-fp16=yes|no], [enable software fp16 comms])],
[ac_SFW_FP16=${enable_sfw_fp16}], [ac_SFW_FP16=yes])
case ${ac_SFW_FP16} in
yes)
AC_DEFINE([USE_FP16],[1],[conversion to fp16]);;
AC_DEFINE([SFW_FP16],[1],[software conversion to fp16]);;
no);;
*)
;;
AC_MSG_ERROR(["SFW FP16 option not supported ${ac_SFW_FP16}"]);;
esac
############### MKL
@ -189,7 +188,14 @@ case ${ax_cv_cxx_compiler_vendor} in
case ${ac_SIMD} in
SSE4)
AC_DEFINE([SSE4],[1],[SSE4 intrinsics])
SIMD_FLAGS='-msse4.2';;
case ${ac_SFW_FP16} in
yes)
SIMD_FLAGS='-msse4.2';;
no)
SIMD_FLAGS='-msse4.2 -mf16c';;
*)
AC_MSG_ERROR(["SFW_FP16 must be either yes or no value ${ac_SFW_FP16} "]);;
esac;;
AVX)
AC_DEFINE([AVX1],[1],[AVX intrinsics])
SIMD_FLAGS='-mavx -mf16c';;
@ -423,7 +429,6 @@ AC_OUTPUT
echo "~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Summary of configuration for $PACKAGE v$VERSION
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
----- PLATFORM ----------------------------------------
architecture (build) : $build_cpu
os (build) : $build_os
@ -436,6 +441,7 @@ SIMD : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
Threading : ${ac_openmp}
Communications type : ${comms_type}
Default precision : ${ac_PRECISION}
Software FP16 conversion : ${ac_SFW_FP16}
RNG choice : ${ac_RNG}
GMP : `if test "x$have_gmp" = xtrue; then echo yes; else echo no; fi`
LAPACK : ${ac_LAPACK}