Legend:
-------
* = endpoint is included in range
o = endpoint is excluded in range
On intel SSE, there exists a CVTSD2SI insruction with rounding mode controllable using the MXCSR register.
k are numbers in the set of all integers.
Rounding to Nearest Integer (Symmetric Arithmetic Rounding)
C++: floor(x+0.5), except when x=0.49999... [see JDK bug below]
feature: almost ideal rounding, except for bias at k+0.5 where k is integer
feature: this is probably what you learned in elementary school
feature: symmetric about the origin
Intel x64: not possible in single instruction, they prefer to use round to closest even
ARM AArch64: FCVTAS round to nearest with ties to away
-------------------------------------
rint(x)
|
|
3.0 + *=====o
|
|
2.0 + *=====o
|
|
1.0 + *=====o
|
|
.--+--.--+--.--+--o=====o--+--.--+--.--+--. x
-3.0 -2.0 -1.0 | 1.0 2.0 3.0
|
o=====* + -1.0
|
|
o=====* + -2.0
|
|
o=====* + -3.0
|
|
Intel x87's Round to Closest (Nearest) Even Integer
Also known as Banker's Rounding
feature: default starting mode in Win32 apps
feature: ideal rounding from a numerical perspective
feature: MS DirectX uses this mode, along with single precision mode
feature: almost the same as ideal rounding except at 0.5+k points
where k is an integer.
C++: difficult to implement correctly without relying on platform-specific behavior (see notes at the end)
ARM NEON: int32x2_t vcvt_s32_f32(float32x2_t a); // VCVT.S32.F32 d0, d0
intel x64: _mm_round_sd(x, _MM_FROUND_TO_NEAREST_INT)
ARM AArch64: FCVTNS
------------------------------------------------------------------
rint_even(x)
|
|
3.0 + o====o
|
|
2.0 + *=====*
|
|
1.0 + o=====o
|
|
.--+--.--+--.--+--*=====*--+--.--+--.--+--. x
-3.0 -2.0 -1.0 | 1.0 2.0 3.0
|
o=====o + -1.0
|
|
*=====* -2.0 +
|
|
o=====o -3.0 +
|
|
Rounding to Nearest Integer, but biased towards positive infinity
feature: easy to implement using floor
feature: for integer arguments, use add and shift right
C++: rint_pos(x) = floor(x+0.5), except when x= k + 0.49999...
intel x64: _mm_round_sd(x, _MM_FROUND_TO_POS_INF)
ARM AArch64: ???
--------------------------------------------------------------------
rint_pos(x)
|
|
3.0 + *=====o
|
|
2.0 + *=====o
|
|
1.0 + *=====o
|
|
.--+--.--+--.--+--*=====o--+--.--+--.--+--. x
-3.0 -2.0 -1.0 | 1.0 2.0 3.0
|
*=====o + -1.0
|
|
*=====o + -2.0
|
|
*=====o + -3.0
|
|
Round Towards Zero (chop, fix, truncation)
feature: default rounding mode in ANSI C
C++: (int)x
feature: toward minus infinity for positive numbers,
toward positive infinity for negative numbers
feature: has a larger deadzone at zero
intel: Intel SSE3 has a FISTTP instruction for this.
intel x64: _mm_round_sd(x, _MM_FROUND_TO_ZERO)
ARM AArch64: FCVTZS
--------------------------------------------------------
int_cast(x)
|
|
3.0 +
|
|
2.0 + *=====o
|
|
1.0 + *=====o
|
|
.--+--.--+--.--o==.=====.==o--.--+--.--+--. x
-3.0 -2.0 -1.0 | 1.0 2.0 3.0
|
o=====* + -1.0
|
|
o====* -2.0 +
|
|
-3.0 +
|
|
Round towards Minus Infinity (floor)
feature: for integer or fixed point arguments, this is easy in hardware (use arithmetic shift right)
C++: floor(x)
intel x64: _mm_round_sd(x, _MM_FROUND_TO_NEG_INF)?
ARM AArch64: FCVTMS?
--------------------------------------------------------------
floor(x)
|
|
3.0 +
|
|
2.0 + *=====o
|
|
1.0 + *=====o
|
|
.--+--.--+--.--+--.--*=====o--.--+--.--+--. x
-3.0 -2.0 -1.0 | 1.0 2.0 3.0
|
*=====o -1.0
|
|
*=====o + -2.0
|
|
*=====o -3.0 +
|
|
Ceiling (ceil)
C++: ceil(x) = -floor(-x)
feature: almost the mirror image of floor(x)
intel x64: _mm_round_sd(x, _MM_FROUND_TO_POS_INF)
ARM AArch64: FCVTPS(x)
----------------------------------------
ceil(x)
|
|
3.0 + o=====*
|
|
2.0 + o=====*
|
|
1.0 o=====*
|
|
.--+--.--+--.--o==.==*--.--+--.--+--.--+--. x
-3.0 -2.0 -1.0 | 1.0 2.0 3.0
|
o=====* + -1.0
|
|
o====* -2.0 +
|
|
-3.0 +
|
|
General information on: Custom Rounding.More information on: Speed Optimizations.
Pentium rounding unsignaled overflow bug documented at: Microsoft
More about C99 rounding modes at FreeBSD mailing list
It turns out that floor(x+0.5) is not equal to round(x) when x=0.4999...
... This was the cause of this Java JDK bug.
More information on rounding bugs are at here.