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

Due diligence tweaks.

Browse Source
This commit is contained in:
james.brodman
2013-10-01 12:17:57 -04:00
parent dc8895352a
commit d45c5767d8
1 changed files with 12 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
examples/intrinsics/knc-i1x16.h
View File

@@ -1021,12 +1021,22 @@ static FORCEINLINE __vec16_i64 __mul(const __vec16_i32 &a, const __vec16_i64 &_b
_mm512_mulhi_epi32(a.v, b.v_lo))).cvt2zmm(); _mm512_mulhi_epi32(a.v, b.v_lo))).cvt2zmm();
} }
#if __ICC_VERSION == 1400 #if __ICC_VERSION >= 1400
static FORCEINLINE __vec16_i64 __mul(__vec16_i64 a, __vec16_i64 b) { static FORCEINLINE __vec16_i64 __mul(__vec16_i64 a, __vec16_i64 b) {
return __vec16_i64(_mm512_mullox_epi64(a.v1, b.v1), _mm512_mullox_epi64(a.v2,b.v2)); return __vec16_i64(_mm512_mullox_epi64(a.v1, b.v1), _mm512_mullox_epi64(a.v2,b.v2));
} }
#else #else
BINARY_OP(__vec16_i64, __mul, *) static FORCEINLINE __vec16_i64 __mul(const __vec16_i64 &a, const __vec16_i64 &b)
{
__vec16_i32 lo = _mm512_mullo_epi32(a.v_lo,b.v_lo);
__vec16_i32 hi_m1 = _mm512_mulhi_epi32(a.v_lo, b.v_lo);
__vec16_i32 hi_m2 = _mm512_mullo_epi32(a.v_hi, b.v_lo);
__vec16_i32 hi_m3 = _mm512_mullo_epi32(a.v_lo, b.v_hi);
__mmask16 carry = 0;
__vec16_i32 hi_p23 = _mm512_addsetc_epi32(hi_m2, hi_m1, &carry);
__vec16_i32 hi = _mm512_adc_epi32(hi_m3, carry, hi_p23, &carry);
return __vec16_i64(lo, hi);
}
#endif #endif
#endif #endif