[RFC] Supporting checked arithmetic for division-related operations
kimikage opened this issue · 3 comments
While there is room for some improvement and discussion, I have prepared the implementations for checked arithmetic on addition and multiplication (cf. #41 (comment)).
However, the circumstances surrounding division-related operations are more complicated.
Checked arithmetic for /
Since the result of the /
with integer arguments is a floating point number, the checked version of /
is not defined in Base.Checked
. Unlike *
and mul
, /
and div
are different operations. So, checked_div
is for div
(÷
) and not for /
.
Shall we define {checked/saturating/wrapping}_fdiv
? ("f" stands for "fractional", not "float" 😄).
Default arithmetic
Unlike addition and subtraction, the operations involved in integer division are basically "checked" arithmetic. It is not clear whether we should follow the integer style or the floating point style (which does not throw an exception). However, I personally think the checked arithmetic is a better choice for the default, since FixedPoint
numbers cannot represent Inf
or NaN
.
Division by zero in wrapping arithmetic
There are several options for the result of dividing by zero. Considering the distinction from saturating arithmetic, I think it is a good idea to return zero.
BTW, as with issue #173, the /
for Fixed
has a rounding mode of RoundNearest
.
FixedPointNumbers.jl/src/fixed.jl
Line 105 in 134646f
julia> a = Q0f7(11/128) / Q0f7(96/128)
0.109Q0f7
julia> b = Q0f7(11/96)
0.117Q0f7
julia> a - 11/96, b - 11/96
(-0.005208333333333329, 0.0026041666666666713)
Unlike multiplication, the SIMD instruction for integer division does not exist in x86 (SSE, AVX), so the method of converting to floating point numbers may be effective.
FixedPointNumbers.jl/src/normed.jl
Line 289 in 134646f
Return type of div
/cld
/fld
/fld1
Currently, div
/fld
/fld1
return a value in the integer rawtype instead of the input FixedPoint
type. This is a reasonable choice to save the bit width.
(BTW, surprisingly, support for cld
is missing. 😅)
FixedPointNumbers.jl/src/FixedPointNumbers.jl
Line 334 in f7d7bbc
However, in PR #207, we decided to promote Integer
and FixedPoint
to a floating-point type. (That was the case with Normed
even before PR #207.) That is not very convenient when used in combination with rem
and mod
, which return a FixedPoint
number.
One solution is to add widemul(x::T, y::X) where {T <: Integer, X:< FixedPoint{T}}
as a measure to avoid affecting promote_rule
. (cf. #185)
BTW, I plan to keep fld1
and mod1
as they are (without adding wrapping_fld1
etc.).
Return value of saturating_mod
and saturating_rem
.
julia> x = typemin(Q0f7); y = -eps(Q0f7);
julia> saturating_div(x, y) # 128 --> 127
127
julia> x == saturating_div(x, y) * y + saturating_rem(x, y) # should be true?
If the equation above should be always true, saturating_rem(typemin(Q0f7), -eps(Q0f7)) == -eps(Q0f7)
.
If abs(saturating_rem(x, y)) < abs(y)
should be always true, saturating_rem(typemin(Q0f7), -eps(Q0f7)) == zero(Q0f7)
.
I consider the former to be the better option.