[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Public WebGL] Should GLfloat and GLclampf be typedefs for "unrestricted float"?



On 12-11-19 09:03 PM, Mark Callow wrote:

On 2012/11/20 0:35, Benoit Jacob wrote:
For many WebGL functions taking GLfloat, it is actually very important
to _not_ throw, because these floats are often the result of nontrivial
arithmetic, and nontrivial floating-point arithmetic does produce
non-finite values, and good floating-point API design is to handle such
non-finite values gracefully.

So it's actually very important to make GLfloat unrestricted.
+1
For GLclampf, I don't know. Since GLfloat must be unrestricted, and IIUC
current browsers implement unrestricted anyways, it may be simpler to
keep it unrestricted. That may also be a good idea as GLclampf's too may
be the result of nontrivial floating point operations.
If GLfloat is unrestricted GLclampf should also be.

But note that the ES 2.0 spec says using NaN or infinity will produce unspecified results; it also says that these values should not result in "GL interruption or termination."

It's OK, in fact inevitable, to say that rendering results are undefined. Keep in mind that whether or not a non-finite value is produced at any stage in a shader, is undefined in the first place, as the shader compiler is allowed (AFAIK) to perform optimizations that can change the results of floating point arithmetic, such as a*(b*c) -> (a*b)*c.

Also note that the question of whether non-finite issues occur cannot be reduced to the question of whether non-finite values are passed to any GL functions. For example, if I set a uniform float to 1e20, that's a finite value, but if I square it in my shader, it will become infinite even with highp. So, sad as it may seem, a GL spec can only say very little about non-finite values. It can say in a non-normative way that non-finite values should only be produced at any stage if that is a reasonable result given the inputs and the shaders; and it should say that non-finite values only result in undefined rendering for the specific triangles that involve them.

Since it's probable that current GL implementations behave like that anyway, it should be safe to bet on that, add a note to that effect in the WebGL spec, and maybe have that conversation with other GL working groups if needed (I don't know that that is needed).

Cheers,
Benoit


The WebGL spec. does not specify what the results should be either so I would hardly say that non-finite values are handled "gracefully." Graceful handling would probably require that the values be propagated as specified in IEEE 754. But I think throwing an exception would be akin to "GL interruption" so I come down on the side of unrestricted.

Regards

    -Mark

--
注意:この電子メールには、株式会社エイチアイの機密情報が含まれている場合が有ります。正式なメール受信者では無い場合はメール複製、 再配信または情報の使用を固く禁じております。エラー、手違いでこのメールを受け取られましたら削除を行い配信者にご連絡をお願いい たし ます.

NOTE: This electronic mail message may contain confidential and privileged information from HI Corporation. If you are not the intended recipient, any disclosure, photocopying, distribution or use of the contents of the received information is prohibited. If you have received this e-mail in error, please notify the sender immediately and permanently delete this message and all related copies.