www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - any BSGP implementation available?

reply davidl <davidl 126.com> writes:
I've seen someone commented BSGP implementation looks pretty modern. So I
suppose someone
already got its source code?
Looking for some research on GPU computing
Dec 16 2008
next sibling parent reply sandeepk <a b.com> writes:
davidl Wrote:

 I've seen someone commented BSGP implementation looks pretty modern. So I
suppose someone
 already got its source code?
 Looking for some research on GPU computing

I attended the talk on BSGP at SIGGRAPH Asia and I found BSGP to be very high level and simple to read and write (though I haven't really done any programming in it yet). Maybe Kun Zhou ( http://www.kunzhou.net/ ) can help you regarding the source code. I also wondered about implementing it for D. Currently the number of GPU computing "languages" is almost ridiculous, add in marketing and most developers don't really seem know what's really right for them or what they should be excited about.
Dec 16 2008
parent davidl <davidl 126.com> writes:
在 Wed, 17 Dec 2008 13:18:44 +0800,sandeepk <a b.com> 写道:

 davidl Wrote:

 I've seen someone commented BSGP implementation looks pretty modern. So I
suppose someone
 already got its source code?
 Looking for some research on GPU computing

I attended the talk on BSGP at SIGGRAPH Asia and I found BSGP to be very high level and simple to read and write (though I haven't really done any programming in it yet). Maybe Kun Zhou ( http://www.kunzhou.net/ ) can help you regarding the source code. I also wondered about implementing it for D. Currently the number of GPU computing "languages" is almost ridiculous, add in marketing and most developers don't really seem know what's really right for them or what they should be excited about.

But BSGP introduced several different keyword, I dislike some of them, yet the compiler implementation is rather advanced. The small binary executable provides you a full feature of compile time exec, and several passes of optimization. I've heard that the source was bought by NV corp, so I just want to know if there's some previous source available.
Dec 16 2008
prev sibling next sibling parent "Robert Jacques" <sandford jhu.edu> writes:
On Tue, 16 Dec 2008 20:40:05 -0800, davidl <davidl 126.com> wrote:

 I've seen someone commented BSGP implementation looks pretty modern. So  
 I suppose someone
 already got its source code?

Nope. BSGP is a Microsoft Research project so while a compiler is available, I doubt the source will be released. However, BSP (http://en.wikipedia.org/wiki/Bulk_Synchronous_Parallel) stuff is available (i.e. BSFC++ http://f.loulergue.free.fr/research/bsfcpp/main.html)
 Looking for some research on GPU computing

I've some GPGPU papers written using CUDA and D. (http://ciis.lcsr.jhu.edu/Research/Radiation_Oncology) CUDA just got (as of 2.1) the ability for the display driver to compile code (it looks like it only ptx, which is NVIDIA's portable GPU assembly language, right now). Which means you could convert D code to ptx using single assignment and then let NVIDIA optimize it. I have D bindings and a high-level api for CUDA if you'd like (~sigh~ Someday I'll actually get around to open sourcing them)
Dec 16 2008
prev sibling next sibling parent davidl <davidl 126.com> writes:
在 Wed, 17 Dec 2008 13:59:46 +0800,Robert Jacques <sandford jhu.edu> 写道:

 On Tue, 16 Dec 2008 20:40:05 -0800, davidl <davidl 126.com> wrote:

 I've seen someone commented BSGP implementation looks pretty modern. So
 I suppose someone
 already got its source code?

Nope. BSGP is a Microsoft Research project so while a compiler is available, I doubt the source will be released. However, BSP (http://en.wikipedia.org/wiki/Bulk_Synchronous_Parallel) stuff is available (i.e. BSFC++ http://f.loulergue.free.fr/research/bsfcpp/main.html)

I thought you've seen the implementation of the compiler :D I doubt if he would be that kind to give me the source :) since rumours say it's been sold
 Looking for some research on GPU computing

I've some GPGPU papers written using CUDA and D. (http://ciis.lcsr.jhu.edu/Research/Radiation_Oncology) CUDA just got (as of 2.1) the ability for the display driver to compile code (it looks like it only ptx, which is NVIDIA's portable GPU assembly language, right now). Which means you could convert D code to ptx using single assignment and then let NVIDIA optimize it. I have D bindings and a high-level api for CUDA if you'd like (~sigh~ Someday I'll actually get around to open sourcing them)

That's cool!
Dec 16 2008
prev sibling parent "Robert Jacques" <sandford jhu.edu> writes:
On Tue, 16 Dec 2008 22:54:21 -0800, davidl <davidl 126.com> wrote:

 I thought you've seen the implementation of the compiler :D I doubt if he
 would be that kind to give me the source :) since rumours say it's been  
 sold

Well, if it's true that it's been sold to NVIDIA, then you can expect it to be open sourced as the CUDA compiler is based on Open64 and is therefore open sourced.
Dec 16 2008