It depends a lot on the representation. Bitmap will not get a lot better than what we've got at the moment with JPeg or Gif. Vector representations can vary wildly in compactness and rendering speed. Eg. SVG bad, CSS gradients good. SVG bad, EPS/PDF good.
Obviously drawing from CPU can be much quicker than reading from disk. And if you draw into the GPU buffer first you get all of the advantages of GPU acceleration later.
And it can be much more bandwidth efficient.
Take streaming a bitmap over http. You have to take the penalty of the PNG header, the penalty of assembling a data:url and a 33% fixed penalty for Base64 encoding the binary bits.
So to stream a 1x256px gradient sprite there is a shedload of wasted bits.
If you're in vector land you can just send the gradient endpoints.
If you're not in WebKit land or with pre-sent vector algorithms (rare) you also have to stream the vector drawing code. In that case you can ammortize the cost of the drawing code with a simpler/smaller wire representation.
Blitting a cached PNG will always be faster than CPU based drawing code and I think the app size tradeoff is worth it.
Reading your PNG from the disk in the first place will be slower than reading a compact vector, rendering it and caching.
Re-rendering from CPU can often be faster than GPU bus access times on PCs.
Anyways, it depends on what you're optimising for. I was just illustrating how vector can be much better than bitmap in many situations.
Got benchmarks for that?
Reference: https://developer.apple.com/library/mac/#documentation/graph... (it works on iOS too)
(Note that CGLayer is not related to CALayer or CGTransparencyLayer -- they are absolutely separate things. CALayers are incredibly useful in practice; CGLayer was kind of a dead end.)
There's also no reason you can't build your images the first time they're used, and save them out to PNG files for successive uses (if profiling determines that it is indeed more optimal). Even if you end up using graphics at runtime, you'll save work by using code to generate your graphics.
For instance, simply write utility programs to generate all of your graphics with code -- this is definitely faster than photoshop when you have buttons in lots of different languages. It's also useful if you want to use a font where it's not legal to embed it. (You can generate all of your button imagery using code and localization files, then remove the embedded fonts and use the generated PNG files in the release version.)
"While some of the apps have received additional features, it seems likely that the increase in size is mainly down to the huge graphics needed to fill the new iPad's 2048 x 1536 Retina Display. It's worth remembering that these are only download sizes, and once installed the apps may be even larger. Regular apps will likely receive a similar bump in size once developers update them with hi-def graphics"
"iPad apps up to five times larger..."
However, current PNG compressors are really good at optimizing flat color areas, gradients, and lines. It seems to me like most images would get only a relatively small increase in size due to this upsizing (1.5x-2x).
And I don't get why Apple bothers with CgBi PNGs, since endian swap on ARM is a single cycle instruction, and it's probably free anyway while moving the bitmap to the GPU memory.
The SoCs in iOS devices are shared-memory systems, so system memory is GPU memory.[1] But you're right, the endian swap could easily be done in the PNG decoder.
[1] I have to admit I'm not sure it's implemented as zero-copy through to userspace, though. The normal, 20-year-old GL texture APIs can certainly only be implemented via copying, and I'm not sure if OpenGL ES has anything like mappable texture pixel buffers. It would make a lot of sense though, considering how memory-starved embedded systems are.
Anyway sometimes you need drawing code -- it's much more animatable, for example.
Drawing your UI is never ever going to be as cheap as loading a converted PNG (which Apple's modified pngcrush converts for you). I think a lot of devs have the draw/vector vs. precomposed bitmap tradeoff the wrong way round. Drawing all of your gradated UIButtons with CoreGraphics methods is a false economy compared to just loading a stretchable PNG. Almost all of Apple's UI system imagery is bitmap based, and for a good reason.