Re: [geometry] Replace isIdentity()?

On Sat, Jun 7, 2014 at 1:17 AM, Kristopher Giesing <kris.giesing@gmail.com>
wrote:

> There are already plenty of examples where theoretically idempotent
> operations aren't idempotent due to rounding error.  Off the top of my
> head, "Math.sqrt(x)*Math.sqrt(x) === x" generally evaluates to false if x
> isn't an integer.
>
> Therefore, I don't think we need to treat matrix identity in any
> particularly special regard here.  I would vote to leave isIdentity() in
> its most basic and obvious form, that is, testing that the matrix truly is
> the identity matrix, and not try to get tricky here about tracking what
> transforms have been applied.
>
> I personally think it is actually valuable for programmers to hit these
> kinds of problems so that they understand the implications of finite
> precision math.
>

Are you saying we should provide this API so programmers learn that they
shouldn't use it?


> On Fri, Jun 6, 2014 at 11:33 PM, Dirk Schulze <dschulze@adobe.com> wrote:
>
>> Hi,
>>
>> During the discussion on dev-platform@lists.mozilla.org[1] some
>> engineers raised concerns to support isIdentity.
>>
>> To quote Rik:
>>
>> “"
>> […] as matrices get computed, they are going to jump unpredicably
>> between being exactly identity and not. People using isIdentity() to jump
>> between code paths are going to get unexpected jumps between code paths
>> i.e. typically performance cliffs, or worse if they start asserting that a
>> matrix should or should not be exactly identity. For that reason, I would
>> remove the isIdentity method.
>> “”
>>
>> So while the following is stable on a matrix that is already identity:
>>
>>         matrix.translateBy(10).translateBy(-10);
>>
>> Certain functions are likely to cause not identity matrices:
>>
>>         matrix.rotateBy(54.4, 20.5, 20.5).rotateBy(-54.4, 20.5, 20.5);
>>
>> Note that it rotates by 54.4 degrees and then rotates back by -54.4
>> degrees. While both operations should eliminate themself, it could happen
>> that the matrix is not identity anymore. The strange part: on further
>> transformations it could be identity again. This is because of different
>> computational precision and rounding deviations. This is usually worst if
>> trigonometric functions are involved. All rotate functions use sin() or
>> cos() and are more likely to cause the described behavior.
>>
>

Received on Saturday, 7 June 2014 21:33:16 UTC