It's 80 chars right now, so that can't be done without sacrificing some golden buzz tech. And without that buzz tech, you never would have seen this...
Language creator here (didn't expect to see this!). Spot on about WebCL not being widely deployed yet, which means code generating it is mostly for kiosk etc. purposes.
For more immediate deployment, I have most of an alternative codegen working that emits multicore + GPU code. Basically, CSS selectors + layout are multicore via asm.js + web workers + typed arrays, and GPU rendering via WebGL.
If anyone is interested in an alternative GPU backend for layout + CSS selectors, say via Rivertrail or old school GPGPU, drop a line :)
I'd love some kind of posting on FTL: on why it was introduced, what constraints it has, what it enables Superconductor to do that couldn't be done with a general purpose language.
Did you give any thought to trying to intake JavaScript itself as the target language?
We originally built FTL to help formalize and automatically parallelize CSS. (Results of that work are going into Mozilla's Servo browser.) That includes constructs for both declarative layout and declarative parallelization. Our PPOPP 2013 paper is the best place to look right now, and a video of my talk at SFJS last week should be posted soon.
We already have JS as a target language ;-) You can try an old version here: http://jsfiddle.net/lmeyerov/Awubf/ . It's similar to the current lang version except without macros. I'm hoping to have a full release with tutorials and a fast non-WebCL version by September, but we'll see :)
Another approach that uses WebGL only and scales to billions of linked points at 50fps is the inMems project by the Stanford Visualization Group: http://vis.stanford.edu/papers/immens. The key idea there is binning and data tiles.
Is there a hard dependency on WebCL? I believe the Chrome team recently said they have no existing plans to adopt it.
I talked to Jeff recently -- they're doing cool stuff with immens. AFAICT, Superconductor does a couple of things differently. First, it is DSLs for making new visualizations, not running analytics within ones they've prebuilt. Second, Immens uses server preprocessing -- we do not require it, and only use it to drop data loading from 5-10s to < 1s. The idea of a render cloud is great (others do this as well), so I've been chatting to Spark folks (http://spark-project.org/) about a backend to compile our DSL into real-time graph processing on the cloud. A shorter-term option may be to extend our CUDA/OpenCL backend to use new multi-GPU extensions.
RE:WebCL, WebCL is simply one of our compilation targets. We can also compile to multithreaded C++, JavaScript (naive or typed arrays), OpenCL, and CUDA as well. For the web space, I'm working on an accelerated alternative of (workers+asm.js)+WebGL, and going from WebCL to RiverTrail isn't too big of a leap.
Found no live-examples to run. WebCL's not supported by my browser. WebCL's not integrating with the WebGL rendering pipeline (no buffer sharing, no atomic synchroniczation primitives etc.)
I'd suggest solving a few fundamental problems with the technology first, sorry to rain on the parade.
Yeah, WebCL->WebGL buffer sharing is a big deal: having to copy the data from GPU->CPU->GPU for the layout->rendering hand-off is a ~5-10ms hit on the animation loop.
My benchmarks show that asm.js + workers + webgl will have that 5-10ms hit, but otherwise performs at similar scales. If we figure out streaming data transfers to overlap communication with computation, it would be even less.
Compute shaders solve essentially all these issues, and more.
The graphics pipeline consists of programmable and semi-fixed components and it is quite often useful to interact with it. For that purpose compute shaders introduce compute kernels, atomic counters, atomic locks, sharable buffers and some hookups to the pipeline stages.
This is functionality gleaned from Cuda, but repackaged in a more or less portable way. Of course compute shader kernels aren't as flexible as either Cuda or OpenCL programs. However OpenCL is by now quite far behind "the state of the art".
Ah, interesting. OpenCL/WebCL solve this problem, so I think the question is portability. Will compute shaders be landing in most WebGL implementations? WebGL seems generally limited to OpenGL ES, so I had only explored BrookGPU-style alternatives.
Targeting CUDA/OpenCL/WebCL was so that we can rely upon intermediate GPGPU compilers. Our DSL compiler outputs a bunch of for-loops, which traditional compilers love to optimize (the JIT for JS, LLVM for OpenCL, etc.). If we targeted a low-level language such as nvidia's PTX, I suspect we'd still want something like LLVM inbetween. Is there a good preprocessor for compute shaders?
At least in the OpenGL world, compute shaders use more or less the same syntax as other kinds of shaders (GLSL).
It's true that WebGL does not have compute shaders (and neither does OpenGL ES), yet.
However OpenGL ES roughly tracks OpenGL's development with OpenGL ES 2.0 == OGL 2.0, OpenGL ES 3.0 == OGL 3.0, in each case with a bunch of features stripped but with the essential advances in place. OpenGL ES 4.0 which should come about in the next 2-3 years is very likely to include compute shaders.
15 comments
[ 2.6 ms ] story [ 40.0 ms ] threadFor more immediate deployment, I have most of an alternative codegen working that emits multicore + GPU code. Basically, CSS selectors + layout are multicore via asm.js + web workers + typed arrays, and GPU rendering via WebGL.
If anyone is interested in an alternative GPU backend for layout + CSS selectors, say via Rivertrail or old school GPGPU, drop a line :)
I'd love to hear if anyone would be interested in:
-- web-based data visualization on mobile that is 50X+ faster than libraries such as D3
-- massive visualizations on the laptop with ~1MM data points real-time
Did you give any thought to trying to intake JavaScript itself as the target language?
We already have JS as a target language ;-) You can try an old version here: http://jsfiddle.net/lmeyerov/Awubf/ . It's similar to the current lang version except without macros. I'm hoping to have a full release with tutorials and a fast non-WebCL version by September, but we'll see :)
Another approach that uses WebGL only and scales to billions of linked points at 50fps is the inMems project by the Stanford Visualization Group: http://vis.stanford.edu/papers/immens. The key idea there is binning and data tiles.
Is there a hard dependency on WebCL? I believe the Chrome team recently said they have no existing plans to adopt it.
RE:WebCL, WebCL is simply one of our compilation targets. We can also compile to multithreaded C++, JavaScript (naive or typed arrays), OpenCL, and CUDA as well. For the web space, I'm working on an accelerated alternative of (workers+asm.js)+WebGL, and going from WebCL to RiverTrail isn't too big of a leap.
I'd suggest solving a few fundamental problems with the technology first, sorry to rain on the parade.
Luckily, the Safari version of WebCL we use has integrated buffer sharing (https://github.com/SRA-SiliconValley/webkit-webcl). I'd expect that to be in the Chrome port being developed. The Nokia port for Firefox a couple years back didn't have it, though I don't know what's going on with the new one @ https://www.khronos.org/bugzilla/show_bug.cgi?id=792 .
My benchmarks show that asm.js + workers + webgl will have that 5-10ms hit, but otherwise performs at similar scales. If we figure out streaming data transfers to overlap communication with computation, it would be even less.
The graphics pipeline consists of programmable and semi-fixed components and it is quite often useful to interact with it. For that purpose compute shaders introduce compute kernels, atomic counters, atomic locks, sharable buffers and some hookups to the pipeline stages.
This is functionality gleaned from Cuda, but repackaged in a more or less portable way. Of course compute shader kernels aren't as flexible as either Cuda or OpenCL programs. However OpenCL is by now quite far behind "the state of the art".
Targeting CUDA/OpenCL/WebCL was so that we can rely upon intermediate GPGPU compilers. Our DSL compiler outputs a bunch of for-loops, which traditional compilers love to optimize (the JIT for JS, LLVM for OpenCL, etc.). If we targeted a low-level language such as nvidia's PTX, I suspect we'd still want something like LLVM inbetween. Is there a good preprocessor for compute shaders?
It's true that WebGL does not have compute shaders (and neither does OpenGL ES), yet.
However OpenGL ES roughly tracks OpenGL's development with OpenGL ES 2.0 == OGL 2.0, OpenGL ES 3.0 == OGL 3.0, in each case with a bunch of features stripped but with the essential advances in place. OpenGL ES 4.0 which should come about in the next 2-3 years is very likely to include compute shaders.