shithub: mc

Info  •  Files  •  Log  •  Branches

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.

--- a/lib/math/bld.sub

+++ b/lib/math/bld.sub

@@ -9,6 +9,7 @@

 	fpmath-sum-impl.myr

 	# fused-multiply-add

+	fpmath-fma-impl+posixy-x64-fma.s

 	fpmath-fma-impl.myr

 	lib ../std:std

--- /dev/null

+++ b/lib/math/fpmath-fma-impl+posixy-x64-fma.s

@@ -1,0 +1,13 @@

+.globl math$fma32

+.globl math$_fma32

+math$fma32:

+math$_fma32:

+	vfmadd132ss %xmm1, %xmm2, %xmm0

+	ret

+

+.globl math$fma64

+.globl math$_fma64

+math$fma64:

+math$_fma64:

+	vfmadd132sd %xmm1, %xmm2, %xmm0

+	ret

--- a/mbld/opts.myr

+++ b/mbld/opts.myr

@@ -35,6 +35,13 @@

 	/* 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)

 ;;

--- a/mbld/syssel.myr

+++ b/mbld/syssel.myr

@@ -165,6 +165,9 @@

 			if opt_cpufeatures & CpuidSSE4 == CpuidSSE4

 				tag(b, "sse4")

 			;;

+			if opt_cpufeatures & CpuidFMA == CpuidFMA

+				tag(b, "fma")

+			;;

 		| unknown:

 			std.fatal("unknown architecture {}\n", unknown)

 		;;