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

added {rsqrt,rcp}d support for knc.h. test-147.ispc & test-148.ispc pass.

Browse Source
This commit is contained in:
evghenii
2014-02-05 13:55:38 +01:00
parent ecc9c88ff8
commit c59cff396d
1 changed files with 10 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
examples/intrinsics/knc.h
View File

@@ -1473,18 +1473,14 @@ static FORCEINLINE __vec16_f __rcp_varying_float(__vec16_f v) {
#endif #endif
} }
static FORCEINLINE __vec16_d __rcp_varying_double(__vec16_d x) { static FORCEINLINE __vec16_d __rcp_varying_double(__vec16_d x) {
__vec16_i64 ex = __and(__cast_bits(__vec16_i64(), x), __smear_i64<__vec16_i64>(0x7fe0000000000000)); __vec16_d y;
__vec16_d exp = __cast_bits(__vec16_d(), __sub(__smear_i64<__vec16_i64>(0x7fd0000000000000), ex)); for (int i = 0; i < 16; i++)
__vec16_f xf = __cast_fptrunc(__vec16_f(), __mul(x, exp)); __insert_element(&y, i, 1.0/__extract_element(x,i));
__vec16_f yf = __rcp_varying_float(xf); return y;
__vec16_d y = __mul(__cast_fpext(__vec16_d(), yf), exp);
y = __add(y, __mul(y, __sub(__smear_double<__vec16_d>(2.0), __mul(x, y))));
y = __add(y, __mul(y, __sub(__smear_double<__vec16_d>(2.0), __mul(x, y))));
return y;
} }
static FORCEINLINE double __rcp_uniform_double(double v) static FORCEINLINE double __rcp_uniform_double(double v)
{ {
return __extract_element(__rcp_varying_double(__smear_double<__vec16_d>(v)),0); return 1.0/v;
} }
@@ -1496,20 +1492,14 @@ static FORCEINLINE __vec16_f __rsqrt_varying_float(__vec16_f v) {
#endif #endif
} }
static FORCEINLINE __vec16_d __rsqrt_varying_double(__vec16_d x) { static FORCEINLINE __vec16_d __rsqrt_varying_double(__vec16_d x) {
__vec16_i64 ex = __and(__cast_bits(__vec16_i64(), x), __smear_i64<__vec16_i64>(0x7fe0000000000000)); __vec16_d y;
__vec16_d exp = __cast_bits(__vec16_d(), __sub(__smear_i64<__vec16_i64>(0x7fd0000000000000), ex)); for (int i = 0; i < 16; i++)
__vec16_d exph = __cast_bits(__vec16_d(), __sub(__smear_i64<__vec16_i64>(0x5fe0000000000000), __lshr(ex,1))); __insert_element(&y, i, 1.0/sqrt(__extract_element(x,i)));
__vec16_f xf = __cast_fptrunc(__vec16_f(), __mul(x, exp)); return y;
__vec16_f yf = __rsqrt_varying_float(xf);
__vec16_d y = __mul(__cast_fpext(__vec16_d(), yf), exph);
__vec16_d xh = __mul(x, __smear_double<__vec16_d>(0.5));
y = __add(y, __mul(y, __sub(__smear_double<__vec16_d>(0.5), __mul(xh, __mul(y,y)))));
y = __add(y, __mul(y, __sub(__smear_double<__vec16_d>(0.5), __mul(xh, __mul(y,y)))));
return y;
} }
static FORCEINLINE double __rsqrt_uniform_double(double v) static FORCEINLINE double __rsqrt_uniform_double(double v)
{ {
return __extract_element(__rsqrt_varying_double(__smear_double<__vec16_d>(v)),0); return 1.0/v;
} }