aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Use native float/half conversion instructions with the AVX2 target.

Browse Source
This commit is contained in:
Matt Pharr
2012-01-24 15:33:38 -08:00
parent a5b7fca7e0
commit 1867b5b317
12 changed files with 453 additions and 256 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
builtins/target-avx2.ll
View File

@@ -63,6 +63,44 @@ define <8 x i32> @__max_varying_uint32(<8 x i32>, <8 x i32>) nounwind readonly a
ret <8 x i32> %m
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; float/half conversions
declare <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16>) nounwind readnone
; 0 is round nearest even
declare <8 x i16> @llvm.x86.vcvtps2ph.256(<8 x float>, i32) nounwind readnone
define <8 x float> @__half_to_float_varying(<8 x i16> %v) nounwind readnone {
%r = call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %v)
ret <8 x float> %r
}
define <8 x i16> @__float_to_half_varying(<8 x float> %v) nounwind readnone {
%r = call <8 x i16> @llvm.x86.vcvtps2ph.256(<8 x float> %v, i32 0)
ret <8 x i16> %r
}
define float @__half_to_float_uniform(i16 %v) nounwind readnone {
%v1 = bitcast i16 %v to <1 x i16>
%vv = shufflevector <1 x i16> %v1, <1 x i16> undef,
<8 x i32> <i32 0, i32 undef, i32 undef, i32 undef,
i32 undef, i32 undef, i32 undef, i32 undef>
%rv = call <8 x float> @llvm.x86.vcvtph2ps.256(<8 x i16> %vv)
%r = extractelement <8 x float> %rv, i32 0
ret float %r
}
define i16 @__float_to_half_uniform(float %v) nounwind readnone {
%v1 = bitcast float %v to <1 x float>
%vv = shufflevector <1 x float> %v1, <1 x float> undef,
<8 x i32> <i32 0, i32 undef, i32 undef, i32 undef,
i32 undef, i32 undef, i32 undef, i32 undef>
; round to nearest even
%rv = call <8 x i16> @llvm.x86.vcvtps2ph.256(<8 x float> %vv, i32 0)
%r = extractelement <8 x i16> %rv, i32 0
ret i16 %r
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; gather
@@ -70,6 +108,3 @@ gen_gather(8, i8)
gen_gather(8, i16)
gen_gather(8, i32)
gen_gather(8, i64)