New accuracy for knl.h
This commit is contained in:
@@ -2615,7 +2615,7 @@ static FORCEINLINE __vec16_f __sqrt_varying_float(__vec16_f v) {
|
||||
|
||||
static FORCEINLINE __vec16_f __rcp_varying_float(__vec16_f v) {
|
||||
#ifdef ISPC_FAST_MATH
|
||||
return _mm512_rcp23_ps(v); // Approximation with 23 bits of accuracy.
|
||||
return _mm512_rcp28_ps(v); // Approximation with 28 bits of accuracy.
|
||||
#else
|
||||
return _mm512_recip_ps(v);
|
||||
#endif
|
||||
@@ -2634,7 +2634,7 @@ static FORCEINLINE double __rcp_uniform_double(double v)
|
||||
|
||||
static FORCEINLINE __vec16_f __rsqrt_varying_float(__vec16_f v) {
|
||||
#ifdef ISPC_FAST_MATH
|
||||
return _mm512_rsqrt23_ps(v); // Approximation with 0.775ULP accuracy
|
||||
return _mm512_rsqrt28_ps(v); // Approximation with 28 bits of accuracy
|
||||
#else
|
||||
return _mm512_invsqrt_ps(v);
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user