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

added fast approximate rcp(double) accurate to 15 digits

Browse Source
This commit is contained in:
Evghenii
2014-02-04 15:23:34 +01:00
parent eb1a495a7a
commit fe98fe8cdc
3 changed files with 34 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
stdlib.ispc
View File

@@ -1391,11 +1391,32 @@ static inline uniform float rcp(uniform float v) {
return __rcp_uniform_float(v);
}
#define RCPD(QUAL) \
__declspec(safe) \
static inline QUAL double __rcp_iterate_##QUAL##_double(QUAL double v, QUAL double iv) \
{ \
iv = iv * (2.0d - v*iv); \
iv = iv * (2.0d - v*iv); \
return iv; \
} \
__declspec(safe) \
static inline QUAL double __rcp_safe_##QUAL##_double(QUAL double x) \
{ \
if (x <= 1.0d+33 && x >= 1.0d-33) \
return __rcp_iterate_##QUAL##_double(x, rcp((QUAL float)x)); \
QUAL int64 ex = intbits(x) & 0x7fe0000000000000; \
QUAL double exp = doublebits( 0x7fd0000000000000 + ~ex ); \
QUAL double y = rcp((QUAL float)(x*exp)); \
return __rcp_iterate_##QUAL##_double(x, y*exp); \
}
RCPD(varying)
__declspec(safe)
static inline double rcp(double v) {
return __rcp_varying_double(v, (IntMaskType)__mask);
}
RCPD(uniform)
__declspec(safe)
static inline uniform double rcp(uniform double v) {
return __rcp_uniform_double(v, (IntMaskType)__mask);
@@ -3529,7 +3550,7 @@ static inline uniform double sqrt(uniform double v) {
#define RSQRTD(QUAL) \
__declspec(safe) \
static inline QUAL double __rsqrt_iterate_double(QUAL double x, QUAL double y) \
static inline QUAL double __rsqrt_iterate_##QUAL##_double(QUAL double x, QUAL double y) \
{ \
QUAL double xh = x*0.5d; \
y += y*(0.5d0 - xh*y*y); \
@@ -3540,12 +3561,12 @@ __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)); \
return __rsqrt_iterate_##QUAL##_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); \
return __rsqrt_iterate_##QUAL##_double(x, y*exph); \
}
RSQRTD(varying)