marko10_000
/
openclonk
forked from Mirrors/openclonk
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix C4Real modulo operator

floating-point
Julius Michaelis 2012-05-03 00:15:20 +02:00
parent 633107b292
commit 0d805d5985
1 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
src/lib/C4Real.h
View File

@ -92,15 +92,18 @@ public:
inline C4Real &operator /= (const C4Real &rhs) { value = _mm_div_ss(value, rhs.value); return *this; }
private: static inline __m128 float_hexconst(int32_t c) { union {float f;int i;} fi; fi.i=c; return _mm_set_ps1(fi.f); } public: // can't use a reinterpret_cast here because the parameter is type-punned
inline C4Real &operator %= (const C4Real &rhs) {
__m128 sign = _mm_and_ps(value, float_hexconst(0x80000000));
// unfortunately, there is no _mm_mod_ss. The trick is to abuse the 23 bit significancy by adding (or substracting) a value, which will leave the result truncated. We can work with that
__m128 div = _mm_div_ss(value, rhs.value);
// round
__m128 sig = _mm_or_ps(_mm_and_ps(div, float_hexconst(0x80000000)), float_hexconst(0x4b000000)); // generate (sign?+:-)2^23
__m128 sig = _mm_or_ps(sign, float_hexconst(0x4b000000)); // generate (sign?+:-)2^23
__m128 trunc = _mm_sub_ps(_mm_add_ps(div, sig), sig);
// unfortunately, we have rounded towards 0: the hack add -1 when trunc is < 0 (fortunately, < returns 0xffffffff when true)
trunc = _mm_add_ps(trunc, _mm_and_ps(_mm_cmplt_ps(trunc, _mm_set_ps1(0.0)), float_hexconst(0xbf800000)));
// rounding hacks
trunc = _mm_add_ps(trunc, _mm_and_ps(_mm_cmpgt_ps(trunc, div), float_hexconst(0xbf800000))); // conditionally substract 1
// finish calc
value = _mm_sub_ps(value, _mm_mul_ps(trunc, rhs.value));
// sign fix: TODO: find out where the implementation is broken and do it properly
value = _mm_or_ps(sign, _mm_and_ps(value, float_hexconst(0x7fffffff)));
return *this;
}