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

Fix __ordered and _unordered floating point functions for C++ target.

Browse Source 
Fixes include adding "_float" and "_double" suffixes as appropriate as well
as providing a number of missing implementations.

This fixes a number of failures in the half* tests.
This commit is contained in:
Matt Pharr
2012-07-09 14:35:51 -07:00
parent 107669686c
commit bc7775aef2
6 changed files with 81 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
examples/intrinsics/knc.h
View File

@@ -955,16 +955,13 @@ static FORCEINLINE __vec16_i1 __greater_equal_float(__vec16_f a, __vec16_f b) {
return _mm512_cmpnlt_ps_mask(a, b);
}
/*
static FORCEINLINE __vec16_i1 __ordered(__vec16_f a, __vec16_f b) {
__vec16_i1 ret;
ret.v = 0;
for (int i = 0; i < 16; ++i)
ret.v |= ((a.v[i] == a.v[i]) && (b.v[i] == b.v[i])) ? (1 << i) : 0;
return ret;
static FORCEINLINE __vec16_i1 __ordered_float(__vec16_f a, __vec16_f b) {
return _mm512_cmpord_ps_mask(a, b);
}
*/
static FORCEINLINE __vec16_i1 __unordered_float(__vec16_f a, __vec16_f b) {
return _mm512_cmpunord_ps_mask(a, b);
}
static FORCEINLINE __vec16_f __select(__vec16_i1 mask, __vec16_f a, __vec16_f b) {
return _mm512_mask_mov_ps(b, mask, a);
@@ -1109,16 +1106,19 @@ static FORCEINLINE __vec16_i1 __greater_equal_double(__vec16_d a, __vec16_d b) {
return ret;
}
/*
static FORCEINLINE __vec16_i1 __ordered(__vec16_d a, __vec16_d b) {
static FORCEINLINE __vec16_i1 __ordered_double(__vec16_d a, __vec16_d b) {
__vec16_i1 ret;
ret.v = 0;
for (int i = 0; i < 16; ++i)
ret.v |= ((a.v[i] == a.v[i]) && (b.v[i] == b.v[i])) ? (1 << i) : 0;
ret.m8.m1 = _mm512_cmpord_pd_mask(a.v1, b.v1);
ret.m8.m2 = _mm512_cmpord_pd_mask(a.v2, b.v2);
return ret;
}
static FORCEINLINE __vec16_i1 __unordered_double(__vec16_d a, __vec16_d b) {
__vec16_i1 ret;
ret.m8.m1 = _mm512_cmpunord_pd_mask(a.v1, b.v1);
ret.m8.m2 = _mm512_cmpunord_pd_mask(a.v2, b.v2);
return ret;
}
*/
static FORCEINLINE __vec16_d __select(__vec16_i1 mask, __vec16_d a, __vec16_d b) {
__vec16_d ret;