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

Fix for unsigned cmp for int64

Browse Source
This commit is contained in:
Vsevolod Livinskiy
2014-11-27 19:00:38 +04:00
parent 7e63862dff
commit 70adb5d283
1 changed files with 61 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
examples/intrinsics/knc.h
View File

@@ -1028,6 +1028,67 @@ static FORCEINLINE __vec16_i1 __not_equal_i64_and_mask(const __vec16_i64 &a, con
}
static FORCEINLINE __vec16_i1 __unsigned_less_equal_i64(__vec16_i64 a, __vec16_i64 b) {
__vec16_i1 ret_hi = __unsigned_less_equal_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_less_equal_i32(a.v_lo, b.v_lo);
return ret_hi | ret_lo;
}
static FORCEINLINE __vec16_i1 __unsigned_less_equal_i64_and_mask(__vec16_i64 a, __vec16_i64 b, __vec16_i1 m)
{
__vec16_i1 ret_hi = __unsigned_less_equal_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_less_equal_i32(a.v_lo, b.v_lo);
return (ret_hi | ret_lo) & m;
}
static FORCEINLINE __vec16_i1 __unsigned_greater_equal_i64(__vec16_i64 a, __vec16_i64 b) {
__vec16_i1 ret_hi = __unsigned_greater_equal_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_greater_equal_i32(a.v_lo, b.v_lo);
return ret_hi | ret_lo;
}
static FORCEINLINE __vec16_i1 __unsigned_greater_equal_i64_and_mask(__vec16_i64 a, __vec16_i64 b, __vec16_i1 m)
{
__vec16_i1 ret_hi = __unsigned_greater_equal_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_greater_equal_i32(a.v_lo, b.v_lo);
return (ret_hi | ret_lo) & m;
}
static FORCEINLINE __vec16_i1 __unsigned_less_than_i64(__vec16_i64 a, __vec16_i64 b) {
__vec16_i1 ret_hi = __unsigned_less_than_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_less_than_i32(a.v_lo, b.v_lo);
return ret_hi | ret_lo;
}
static FORCEINLINE __vec16_i1 __unsigned_less_than_i64_and_mask(__vec16_i64 a, __vec16_i64 b, __vec16_i1 m)
{
__vec16_i1 ret_hi = __unsigned_less_than_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_less_than_i32(a.v_lo, b.v_lo);
return (ret_hi | ret_lo) & m;
}
static FORCEINLINE __vec16_i1 __unsigned_greater_than_i64(__vec16_i64 a, __vec16_i64 b) {
__vec16_i1 ret_hi = __unsigned_greater_than_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_greater_than_i32(a.v_lo, b.v_lo);
return ret_hi | ret_lo;
}
static FORCEINLINE __vec16_i1 __unsigned_greater_than_i64_and_mask(__vec16_i64 a, __vec16_i64 b, __vec16_i1 m)
{
__vec16_i1 ret_hi = __unsigned_greater_than_i32(a.v_hi, b.v_hi);
__vec16_i1 ret_lo = __unsigned_greater_than_i32(a.v_lo, b.v_lo);
return (ret_hi | ret_lo) & m;
}
static FORCEINLINE __vec16_i64 __select(__vec16_i1 mask,
__vec16_i64 a, __vec16_i64 b) {
__vec16_i64 ret;
ret.v_hi = _mm512_mask_mov_epi32(b.v_hi, mask, a.v_hi);
ret.v_lo = _mm512_mask_mov_epi32(b.v_lo, mask, a.v_lo);
return ret;
}
template <class RetVecType> static RetVecType __smear_i64(const int64_t &l);
template <> FORCEINLINE __vec16_i64 __smear_i64<__vec16_i64>(const int64_t &l) {