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

added support for fast approximate rsqrt(double). Provide 16 digit accurancy but is over 3x faster than 1/sqrt(double)

Browse Source
This commit is contained in:
Evghenii
2014-02-04 14:44:54 +01:00
parent b0753dc93d
commit eb1a495a7a
5 changed files with 48 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
stdlib.ispc
View File

@@ -1393,12 +1393,12 @@ static inline uniform float rcp(uniform float v) {
__declspec(safe)
static inline double rcp(double v) {
return __rcp_varying_double(v);
return __rcp_varying_double(v, (IntMaskType)__mask);
}
__declspec(safe)
static inline uniform double rcp(uniform double v) {
return __rcp_uniform_double(v);
return __rcp_uniform_double(v, (IntMaskType)__mask);
}
///////////////////////////////////////////////////////////////////////////
@@ -3527,14 +3527,37 @@ static inline uniform double sqrt(uniform double v) {
return __sqrt_uniform_double(v);
}
__declspec(safe)
static inline double rsqrt(double v) {
return __rsqrt_varying_double(v);
#define RSQRTD(QUAL) \
__declspec(safe) \
static inline QUAL double __rsqrt_iterate_double(QUAL double x, QUAL double y) \
{ \
QUAL double xh = x*0.5d; \
y += y*(0.5d0 - xh*y*y); \
y += y*(0.5d0 - xh*y*y); \
return y; \
} \
__declspec(safe) \
static inline QUAL double __rsqrt_safe_##QUAL##_double (QUAL double x) \
{ \
if (x <= 1.0d+33 && x >= 1.0d-33) \
return __rsqrt_iterate_double(x, rsqrt((QUAL float)x)); \
QUAL int64 ex = intbits(x) & 0x7fe0000000000000; \
QUAL double exp = doublebits( 0x7fd0000000000000 - ex ); /* 1.0d/exponent */ \
QUAL double exph = doublebits( 0x5fe0000000000000 - (ex >> 1)); /* 1.0d/sqrt(exponent) */ \
QUAL double y = rsqrt((QUAL float)(x*exp)); \
return __rsqrt_iterate_double(x, y*exph); \
}
RSQRTD(varying)
__declspec(safe)
static inline double rsqrt(double v) {
return __rsqrt_varying_double(v, (IntMaskType)__mask);
}
RSQRTD(uniform)
__declspec(safe)
static inline uniform double rsqrt(uniform double v) {
return __rsqrt_uniform_double(v);
return __rsqrt_uniform_double(v, (IntMaskType)__mask);
}
__declspec(safe)
static inline double ldexp(double x, int n) {