Hi everybody,
how about setting up a wishlist for coming releases of the SDK? I guess the Stream team already knows a lot about what we would like to see, but it might be good to have it all in one place. And it might be interesting to see what the most-wanted features are.
Here are my top 3:
1. Local arrays
I think this feature is really important to improve usability for complex kernels.
2. Better structure and derived types support
Sometimes you need complex structures and you don't want to split them into multiple streams. For example a stream of float[8][8];
3. Access to the Global Data Share
With something like sync_all could be very useful for really long shaders.
1. Expose Global Buffer to Brook+
2. Better documantation for shared memory
3. Support for local arrays inside kernels
1. PROFILER
I see so many people who come here and get frustrated because their GPU kernel code doesn't work much better than their CPU code and I honestly think that a basic profiler (doesn't have to be anything special, something similar to what CUDA offers) would really help. The SKA only tells a very small portion of what's going on and even then doesn't accurately represent some of it's data.
1. Local arrays (please!)
2. More than one scatter stream
3. Output streams with higher dimensionality than input stream(s) (so multiple outputs can be packed into a single array, like in MPI)
My two wishes ...
1. 4x 4870x2 support
2. X11-less CAL, like CUDA.
My biggest wish:
1. X11-free gpu-driver for remote use in a compute cluster !!! (or use of a dedicated second card in my workstation)
minor wishes (compared to my no. 1):
2. more GPU accelerated functions in ACML-GPU
3. OpenCL support within (not instead) CAL/Brook+
4. maybe a sparse matrix-vector library
OpenCL : PLEEEASE!!!
and i cant belive we still dont have a profiler!
...
4. Fix automatic address translation and make things like reduction of a 10000x10 stream possible.
5. OpenCL support, and if possible with Brook+ interoperability.
6. New stream domain operator is slow as it makes a intermediate copy. I think it would be useful to get some standard way (no USE_OLD_API) of selecting a stream sub-domain based only in texture coordinates like the previous one.
-explanation and direct simple example usage of each and every keyword in the language (indexoff, Normalize, etc.). Examples, if in code, should be checked thoroughly before release
-continued Firestream 9170 support
-