diff options
| author | S. Gilles <sgilles@math.umd.edu> | 2018-03-21 23:21:23 -0400 |
|---|---|---|
| committer | S. Gilles <sgilles@math.umd.edu> | 2018-03-22 10:59:33 -0400 |
| commit | d27469804f73bf0874250eda520532875ecf5645 (patch) | |
| tree | 3b46f6f7a4a91cd5d831d832798cbb08cc682ec6 /mbld | |
| parent | 31356a321f79afc2dd2cc5a6b7c9c72f77cf3394 (diff) | |
| download | mc-d27469804f73bf0874250eda520532875ecf5645.tar.gz | |
Implement assembly version of fused multiply-add
The feature flag for fma includes OSXSave and AVX detection, as
these are prerequisites for using the xmm/ymm registers. We do not,
however, check the result of XGETBV (as recommended in the Intel
software developer manual, vol 1, 14.5.3) because this seems to be
a check that should be performed at runtime (see the note in section
14.3).
This is slightly uncomfortable -- perhaps libmath should come with
an __init__ that bails if the user is trying to use AVX stuff when
the OS isn't using XSAVE.
Diffstat (limited to 'mbld')
| -rw-r--r-- | mbld/opts.myr | 7 | ||||
| -rw-r--r-- | mbld/syssel.myr | 3 |
2 files changed, 10 insertions, 0 deletions
diff --git a/mbld/opts.myr b/mbld/opts.myr index b82921f..8956338 100644 --- a/mbld/opts.myr +++ b/mbld/opts.myr @@ -35,6 +35,13 @@ pkg bld = /* not exactly portable, but good enough for now */ const CpuidSSE4 : uint64= 0x180000 + + /* + Intel manuals (vol 1, 14.5.3) say AVX, OSXSAVE also + needed. For full portability, XGETBV also needs to be + checked, though it isn't right now. + */ + const CpuidFMA : uint64= 0x18001000 extern const cpufeatures : (-> uint64) ;; diff --git a/mbld/syssel.myr b/mbld/syssel.myr index 756cf4d..f0223ca 100644 --- a/mbld/syssel.myr +++ b/mbld/syssel.myr @@ -165,6 +165,9 @@ const addsysattrs = {b, tags if opt_cpufeatures & CpuidSSE4 == CpuidSSE4 tag(b, "sse4") ;; + if opt_cpufeatures & CpuidFMA == CpuidFMA + tag(b, "fma") + ;; | unknown: std.fatal("unknown architecture {}\n", unknown) ;; |