The OpenCL code took many months to develop - although that included dead-ends, multiple steps if refinement, and other distractions including completely unrelated work. Even with that, I'd put the effort at around 4-10x that for the NEON code.
The NEON code took a few weeks. It obviously helped immensely that the algorithm was primarily known in advance although the downscaling alterations were not. On the other hand, my total experience with NEON is far less than OpenCL and certainly C or Java in terms of hours.
One reason the NEON code was much easier to write is that because as it is so cheap to invoke, one can just concentrate on the bottlenecks, and leave the housekeeping to C. e.g. I can write a routine that processes as little as 16x16 pixels in assembly, and leave the addressing crap to C. There is also no marshalling or other api binding to worry about: the C is plain C, and the assembly is plain assembly, and even though JOCL is far far better than using the C api directly it's still quite a bit of work. As much fun as OpenCL is, it's even more fun hacking NEON because you can concentrate on the fun bits even more.
Although the OpenCL model is also based on simple kernels which should equally be simple and isolated - it isn't really quite like that in practice. All but the simplest of kernels end up turn into 64-way parallel subroutines utilising LDS, barriers, and so on. Without that you end up leaving skads of performance on the floor, so it really is necessary. Not to mention all the marshalling and boilerplate in the host-code to communicate with it. And because of the marshalling and invocation latencies pretty much everything is forced onto the GPU.
So what's the point i'm getting at?
Well after all that, the projected performance on the previously-latest-version of a popular handset is only about 5x slower than a HD7970 on a pretty beefy desktop!
Yes that 5x speedup is important enough that it is worth it and opens it up to more applications, but on a personal level i'm just totally bummed it isn't much more. It's a highly parallel and bandwidth intensive workload which should be well-suited to a GPU. Obviously opencl has the advantage that it isn't tied to a single bit of hardware. It's a pity SSE sucks so much otherwise it would be interesting to see how a desktop cpu fared on it's own.
I plan to "back port" the algorithms so it can be improved on the GPU, but I have a fairly educated feeling that another 2000% performance isn't very likely. I will also need to use some AMD proprietary extensions, so the portability will suffer too.
I'm sure I can improve it, but I just take it as a big personal slap in the face for all the effort that's gone into it so far!
Of course, the alternative view is that ARM+NEON is the bees knees - with less effort i'm getting relatively great performance. But we all knew that so it isn't such a revelation ...
The main bottleneck on ARM cpus at the moment is the memory, and if you can utilise the cache effectively it really flies. I would really like to see how a beagleboard like machine with big/little A15/A7 quad core and much faster memory would fare, all these cheap android dongles are far too constrained by their form-factor.
Update: Well I might need to eat my words here. Today I started to look at GPU optimisations based on what i'd learnt from the ARM experience and trying to reduce the bottlenecks of the GPU code.
The key word of the day: batching.
One reason I wasn't previously batching the processing is because it didn't really fit the data-flow of an earlier application. But I have now achieved something like a 50x boost in one key algorithm by a combination of batching the work more aggressively and some other significant algorithmic changes.
This is more like it. No longer bummed out ...