about the FAQ "POV-Ray/raytracing on GPU" I am actually quite curious 
how many systems would actually support double precession floating point 
types (as needed by POV-Ray) and how *fast* they are.

If you are not using Windows(tm) there is no need to read on because
Apple's OpenCL implementation does currently not support 64bit floats 
anyway and the same is AFAIK true for any Linux system. This might and 
should change in the future but this is the current situation.

I've written a simple Windows program (a kind of "realtime" Mandelbrot 
explorer) that is meant to check the capabilities of your system.

You can download it here (please check the included readme.txt)

http://www.lilysoft.org/Download/YAMF/yamf.zip

and send an email to me (ive[at]lilysoft[dot]org) or maybe simply reply 
to this post. My guess is that both the availability and performance 
gain will be quite disappointing.


thanks a lot
-Ive

Post a reply to this message

Am 16.08.2011 13:58, schrieb Ive:
> about the FAQ "POV-Ray/raytracing on GPU" I am actually quite curious
> how many systems would actually support double precession floating point
> types (as needed by POV-Ray) and how *fast* they are.
>
> If you are not using Windows(tm) there is no need to read on because
> Apple's OpenCL implementation does currently not support 64bit floats
> anyway and the same is AFAIK true for any Linux system. This might and
> should change in the future but this is the current situation.
>
> I've written a simple Windows program (a kind of "realtime" Mandelbrot
> explorer) that is meant to check the capabilities of your system.
>
> You can download it here (please check the included readme.txt)
>
> http://www.lilysoft.org/Download/YAMF/yamf.zip
>
> and send an email to me (ive[at]lilysoft[dot]org) or maybe simply reply
> to this post. My guess is that both the availability and performance
> gain will be quite disappointing.

Benchmark result:
---------------------------
Window size 540 x 540 pixel
Maximal 1024 iterations


CPU with 8 threads
fps 6.82527

OpenCL GPU with 30 worker units
GeForce GTX 285
fps 26.41509
---------------------------
(CPU is an Intel i7 QuadCore)

OpenCL information:
---------------------------
Platform 0
   OpenCL 1.0 CUDA 4.0.1
   NVIDIA CUDA
   NVIDIA Corporation

   Device 0 - GPU  30 computing units active
     JIT compiler installed
     Data types with preferred and native vector width
       8bit integer: 1, 0
     16bit integer: 1, 0
     32bit integer: 1, 0
     64bit integer: 1, 0
     16bit float: not supported
     32bit float: 1, 0
     64bit float: 1, 0
---------------------------

Post a reply to this message

Am 18.08.2011 11:53, schrieb clipka:

>
> CPU with 8 threads
> fps 6.82527
>
> OpenCL GPU with 30 worker units
> GeForce GTX 285
> fps 26.41509
> ---------------------------
> (CPU is an Intel i7 QuadCore)
>

Thats what I'd have guessed and for anything with more *logic* involved 
there is less speedup to be expected.
You can also install the Intel OpenCL driver from

http://software.intel.com/en-us/articles/download-intel-opencl-sdk

only the runtime at the bottom is needed. Intel's JIT compiler does a 
pretty good job in optimizing. It uses SSE4 instructions when available 
and is good for a speedup of about 15% compared to 'native' CPU threads.

And people with an AMD CPU but with a NVidea graphics card can download 
and install the AMD Stream-SDK to add OpenCL support for the CPU. 
Currently without support for double precession floats but this is 
announced to change with the next release. No need to do this when also 
a AMD graphics card is used as OpenCL is part of the Catalyst driver.


> OpenCL information:
> ---------------------------
> Platform 0
> OpenCL 1.0 CUDA 4.0.1
> NVIDIA CUDA
> NVIDIA Corporation
>
> Device 0 - GPU 30 computing units active
> JIT compiler installed
> Data types with preferred and native vector width
> 8bit integer: 1, 0
> 16bit integer: 1, 0
> 32bit integer: 1, 0
> 64bit integer: 1, 0
> 16bit float: not supported
> 32bit float: 1, 0
> 64bit float: 1, 0
> ---------------------------

This isn't the latest NVidea driver or is it?

Anyway related to POV-Ray: OpenCL for 3.7 does IMHO not make any sense 
as it would require major changes within the multi-threading framework.

On the other hand it looks quite straightforward to implement for MCPov
and as MCPov is currently just single-threaded a 10x speedup when 
running on the GPU seems realistic. Well, most likely I do not see some 
pitfalls there but as time permits I will give it a try.

-Ive

Post a reply to this message

Am 22.08.2011 09:11, schrieb Ive:

>> OpenCL information:
>> ---------------------------
>> Platform 0
>> OpenCL 1.0 CUDA 4.0.1
>> NVIDIA CUDA
>> NVIDIA Corporation
>>
>> Device 0 - GPU 30 computing units active
>> JIT compiler installed
>> Data types with preferred and native vector width
>> 8bit integer: 1, 0
>> 16bit integer: 1, 0
>> 32bit integer: 1, 0
>> 64bit integer: 1, 0
>> 16bit float: not supported
>> 32bit float: 1, 0
>> 64bit float: 1, 0
>> ---------------------------
>
> This isn't the latest NVidea driver or is it?

It's 6.14.12.7533, to which I upgraded in June this year. Don't know if 
there's anything newer by now.

> Anyway related to POV-Ray: OpenCL for 3.7 does IMHO not make any sense
> as it would require major changes within the multi-threading framework.
>
> On the other hand it looks quite straightforward to implement for MCPov
> and as MCPov is currently just single-threaded a 10x speedup when
> running on the GPU seems realistic. Well, most likely I do not see some
> pitfalls there but as time permits I will give it a try.

Mind the recursion.

Post a reply to this message

Am 22.08.2011 14:25, schrieb clipka:

>>> OpenCL 1.0 CUDA 4.0.1

>> This isn't the latest NVidea driver or is it?

> It's 6.14.12.7533, to which I upgraded in June this year. Don't know if
> there's anything newer by now.

Strange. I was wondering about OpenCL version 1.0 and also about the 
missing support for 'half' floats as they are NVideas 'own' invention.


> Mind the recursion.

Sure, but every recursion can also be expressed as an iteration ;)

-Ive

Post a reply to this message

>> Mind the recursion.
>
> Sure, but every recursion can also be expressed as an iteration ;)

I'm fairly sure that's incorrect, but I can't find a reference...

-- 
http://blog.orphi.me.uk/
http://www.zazzle.com/MathematicalOrchid*

Post a reply to this message

Am 22.08.2011 19:56, schrieb Orchid XP v8:
>>> Mind the recursion.
>>
>> Sure, but every recursion can also be expressed as an iteration ;)
>
> I'm fairly sure that's incorrect, but I can't find a reference...
>

...and I'm fairly sure it is true as long there is proof that the 
iterative calculus is Turing-complete as all Turing-complete calculi are 
idempotent.
But I'm much more interested in the practical aspect and expressing 
recursive raytracing with iterations is trivial in this case as the 
OpenCL kernel-language is close to C99 (just without recursion).

-Ive

Post a reply to this message

Window size 540 x 540 pixel
Maximal 1024 iterations


CPU with 4 threads
fps 70.70707

I don't know why I don't get OpenCL options enabled , I have 11.8 
Catalyst and a a couple ATI RADEON 4670 in CrossFire with a AMD Phenom 
II X4 3.4Ghz. ANy clues?

Post a reply to this message

Am 14.10.2011 06:06, schrieb Saul Luizaga:
> Window size 540 x 540 pixel
> Maximal 1024 iterations
>
>
> CPU with 4 threads
> fps 70.70707
>
> I don't know why I don't get OpenCL options enabled , I have 11.8
> Catalyst and a a couple ATI RADEON 4670 in CrossFire with a AMD Phenom
> II X4 3.4Ghz. ANy clues?

AFAIK only ATI FirePro and Radeon 5xxx/6xxx graphics cards do support 
OpenCL. The Radeon 4xxxx series does definitely not support it. Sorry.

-Ive

Post a reply to this message

Actually they do but only partially, yes the 5000 series and above are 
fully compliant, too bad my 4670s are double precision only, they don't 
handle single precision floats and have some power to offer & guess I 
got screwed over by AMD, oh well.

Post a reply to this message