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

Add separate set of builtins for AVX2.

Browse Source 
(i.e., stop just reusing the ones for AVX1).

For now the only difference is that the int/uint min/max
functions call the new intrinsic for that.  Once gather is
available from LLVM, that will go here as well.
This commit is contained in:
Matt Pharr
2012-01-13 14:39:33 -08:00
parent 0f8eee9809
commit 58a0b4a20d
9 changed files with 374 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
builtins/target-avx-x2.ll
View File

@@ -170,33 +170,6 @@ define <16 x float> @__min_varying_float(<16 x float>,
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; int min/max
define <16 x i32> @__min_varying_int32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
binary4to16(ret, i32, @llvm.x86.sse41.pminsd, %0, %1)
ret <16 x i32> %ret
}
define <16 x i32> @__max_varying_int32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
binary4to16(ret, i32, @llvm.x86.sse41.pmaxsd, %0, %1)
ret <16 x i32> %ret
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; unsigned int min/max
define <16 x i32> @__min_varying_uint32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
binary4to16(ret, i32, @llvm.x86.sse41.pminud, %0, %1)
ret <16 x i32> %ret
}
define <16 x i32> @__max_varying_uint32(<16 x i32>, <16 x i32>) nounwind readonly alwaysinline {
binary4to16(ret, i32, @llvm.x86.sse41.pmaxud, %0, %1)
ret <16 x i32> %ret
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
; horizontal ops
@@ -622,12 +595,7 @@ define void @__masked_store_blend_64(<16 x i64>* nocapture %ptr, <16 x i64> %new
}
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; gather/scatter
gen_gather(16, i8)
gen_gather(16, i16)
gen_gather(16, i32)
gen_gather(16, i64)
;; scatter
gen_scatter(16, i8)
gen_scatter(16, i16)