Re: quick comments on Matrix

On Fri, Mar 15, 2013 at 4:36 AM, Dirk Schulze <dschulze@adobe.com> wrote:

>
> On Mar 14, 2013, at 10:31 PM, Robert O'Callahan <robert@ocallahan.org>
> wrote:
>
> > On Fri, Mar 15, 2013 at 5:51 PM, Dirk Schulze <dschulze@adobe.com>
> wrote:
> >> > In rotateFromVector(By), what happens if the x or y coordinate is
> zero? Actually it's not clear to me why these can't be zero.
> >>
> >> This was taken over from SVG. The calculation is "(+/-) atan(y/x)".
> While y could be zero, x can't. I assume the editors of SVG wanted to have
> consistency between the values.
> >
> > The case where x is 0 makes geometric sense though. The only case that
> doesn't make sense is when both x *and* y are zero. I think SVG got this
> wrong.
>
> I guess what you suggest is that (x,y)^T = (0,1)^T is a rotation by 90
> degree and (1,0)^T is a rotation by 0 degree? (Note that the rotation is
> clockwise on SVG.) Maybe it is better to just specify the angle between
> (1,0)^T and the vector given by x and y. The only special case would be
> that both values are 0 (which could be defined to be a multiplication by
> the identity transform == no affect).
>
> >
> >> > What's the rationale for using double for matrix elements? Are
> implementations allowed to use 32-bit precision internally?
> >>
> >> WebIDL requires a datatype for attributes. In this case there was the
> choice between float (32-bit precision) and double (64-bit precision). The
> later seems to make more sense - especially because of Float64Array.
> >
> > I think it's worth considering whether we need to specify that the
> implementation uses double precision. On one hand, specifying the precision
> would help interop. On the other hand, requiring every Matrix to be backed
> by a 4x4 matrix of float64s might have a real performance penalty vs
> float32s.
>
> There were similar discussions on webkit-dev mailing list[1]. Usually
> 32bit maps better to the underlaying GPU. The difference is less important
> on desktop CPUs. With standards like OpenCL 1.1 full and others, GPUs are
> required to get better on double precision anyway. The decision seems to be
> more if we want to optimize for current hardware or look at the life time
> of the spec and assume that HW gets better. Changing the spec later causes
> incompatibility later (even if the difference is very small).
> WebIDL does not support optional datatypes it seems. That means I needed
> to choose between 64-bit and 32-bit. Upon on requests here on this list, it
> seems that there are use cases for double precision beyond graphics for the
> screen.
>

While GPU's are getting faster at double precision in desktop class GPUs -
even the best results are still at 50% of single precision. The issue is on
mobile where, for example, I think that Apple's PowerVR GPU cores can only
emulate double at 1/8 the performance of single - but I am not a GPU expert
and would welcome any feedback from folks who have done GPU acceleration
work for mobile.


>
> Greetings,
> Dirk
>
> [1] https://lists.webkit.org/pipermail/webkit-dev/2012-May/020764.html
>
> >
> > Rob
> > --
> > Wrfhf pnyyrq gurz gbtrgure naq fnvq, “Lbh xabj gung gur ehyref bs gur
> Tragvyrf ybeq vg bire gurz, naq gurve uvtu bssvpvnyf rkrepvfr nhgubevgl
> bire gurz. Abg fb jvgu lbh. Vafgrnq, jubrire jnagf gb orpbzr terng nzbat
> lbh zhfg or lbhe freinag, naq jubrire jnagf gb or svefg zhfg or lbhe fynir
> — whfg nf gur Fba bs Zna qvq abg pbzr gb or freirq, ohg gb freir, naq gb
> tvir uvf yvsr nf n enafbz sbe znal.” [Znggurj 20:25-28]
>
>
>