forked from enlightenment/efl
eina_fp: Fix eina_f32p32_mul rounding issues
Previous multiplication just truncated the fraction part to 16 bits. If we round properly we can actually increase accuracy a bit. Signed-off-by: Daniel Willmann <d.willmann@samsung.com>
This commit is contained in:
parent
7c20a90c3d
commit
5f1614e1d5
|
@ -36,10 +36,16 @@ eina_f32p32_sub(Eina_F32p32 a, Eina_F32p32 b)
|
|||
static inline Eina_F32p32
|
||||
eina_f32p32_mul(Eina_F32p32 a, Eina_F32p32 b)
|
||||
{
|
||||
/* Prevent overflow and do '(a * b) >> 32' */
|
||||
/* Basically do: Eina_F16p16 * Eina_F16p16 = Eina_F32p32 */
|
||||
/* To prevent overflow during multiplication
|
||||
* we need to reduce the precision of the fraction part
|
||||
* Shift both values to only contain 16 bit of the fraction part
|
||||
* (rounded).
|
||||
* After multiplication we again have a value with a 32-bit
|
||||
* fraction part. See also
|
||||
* http://en.wikipedia.org/wiki/Fixed-point_arithmetic#Operations
|
||||
*/
|
||||
|
||||
Eina_F32p32 up;
|
||||
Eina_F32p32 down;
|
||||
Eina_F32p32 result;
|
||||
uint64_t as, bs;
|
||||
Eina_F32p32 sign;
|
||||
|
@ -48,10 +54,13 @@ eina_f32p32_mul(Eina_F32p32 a, Eina_F32p32 b)
|
|||
as = eina_fp32p32_llabs(a);
|
||||
bs = eina_fp32p32_llabs(b);
|
||||
|
||||
up = (as >> 16) * (bs >> 16);
|
||||
down = (as & 0xFFFF) * (bs & 0xFFFF);
|
||||
/* Reduce fraction to 16-bit (rounded) */
|
||||
as = (as >> 16) + ((as >> 15) & 0x1);
|
||||
bs = (bs >> 16) + ((bs >> 15) & 0x1);
|
||||
|
||||
result = up + (down >> 32);
|
||||
up = as * bs;
|
||||
|
||||
result = up;
|
||||
|
||||
return sign < 0 ? - result : result;
|
||||
}
|
||||
|
@ -70,6 +79,8 @@ eina_f32p32_div(Eina_F32p32 a, Eina_F32p32 b)
|
|||
|
||||
sign = a ^ b;
|
||||
|
||||
/* FIXME: This should probably map to +/-inf when converting to double
|
||||
or we should abort... */
|
||||
if (b == 0)
|
||||
return sign < 0 ? (Eina_F32p32) 0x8000000000000000ull : (Eina_F32p32) 0x7FFFFFFFFFFFFFFFull;
|
||||
|
||||
|
|
Loading…
Reference in New Issue