Cuda programming

[Vyper]

Is this of any use nowadays to Seti@Home:

"The CUFFT Library
=================

The Fast Fourier Transform (FFT) is a divide-and-conquer algorithm
for efficiently computing discrete Fourier transforms of complex or
real-valued data sets. The FFT is one of the most important and widely
used numerical algorithms, with applications that include
computational physics and general signal processing. This document
describes CUFFT, the "cuda" (Compute-Unified Device Architecture) FFT library. The
CUFFT library provides a simple interface for computing parallel Fast
Fourier Transforms on an NVIDIA GPU. This allows users to leverage
the floating-point power and parallelism of the GPU without having to
develop a custom, GPU-based FFT implementation.

FFT libraries typically vary in terms of supported transform sizes and
data types. For example, some libraries only implement Radix-2 FFTs,
restricting the transform size to a power of two, while other
implementations support arbitrary transform sizes. This version of the
CUFFT library supports the following features:

- 1D, 2D, and 3D transforms of complex and real-valued data.
- Batch execution for doing multiple 1D transforms in parallel.
- 2D and 3D transform sizes in the range [2, 16384] in any dimension.
- 1D transform sizes up to 8 million elements.
- In-place and Out-of-place transforms for real and complex data."

//Vyper

[Josef W. Segur]

- 1D transform sizes up to 8 million elements.

That's good, it now covers the size range needed. In earlier versions 1d had the same limitation as the 2d and 3d transforms.
                                                                                            Joe

[Vyper]

Thats REALLY splendid news, now the only thing we need is someone that can develop a Cuda client!! DevasteR! Someone,, cmon I suck at this dep.

//Vyper

[_heinz]

I read the new documentation from NVIDIA.
Its really great......maybe I can help, but at first I must end the optimized sources and study the CUDA Compiler and its stuff.
I can no make a promise.

Is the CUDA Compiler free for evaluation ??

Plan to buy a new Nvidia  maybe a XFX 7900GS 256MB AGP. At the moment I´m waiting since a week  on my ordered 22"  multisync TFT Display from LG.

regards seti_britta 

[Vyper]

Hey cool news indeed, DevasteR himself is doing a general GPU app for himself but i endorse someone too to make a Nvidia 8x generic GPU client..

You will need atleast a 8xxx based graphics card to accomplish that atm..

And yes CUDA is now free of eval.. and the API is highly understandable through documents and occupancy calculators and so on their website.

Me myself own a 8800GTX and therefor i can betatest in the highest mode available in a x64 platform atm..

Hope you will try to make a CUDA api, FFT transform is now accesible up to 8M instead of 16384 bytes as in CUDA 0.8 series..

Kind Regards Vyper

[_heinz]

found this important news ---->
Q: Does CUDA support Double Precision Floating Point arithmetic?

 A: CUDA supports the C "double" data type.  However on G80
    (e.g. GeForce 8800) GPUs, these types will get demoted to 32-bit
    floats. NVIDIA GPUs supporting double precision in hardware will
    become available in late 2007.

Think I should wait to buy a NV8800 till the new hardware is available.
regards

[Simon]

The new hardware won't be an 8800 - rather, it'll be their next model, whatever they decide to name it. In the meantime they will release a die-shrink version of GF8800 that'll probably be called the 8900 series.

Same for Cell, by the way - the next hardware refresh will be able to do double precision in hardware much more quickly (first-gen Cell BE does, too, but comparatively slowly). Pity it will probably not be inside a low-cost system like the PS3 when it does get released.

Regards,
Simon.

[_heinz]

did download the whole stuff of CUDA, read a lot and then tried to install the SDK 0.8.1 but it is not available without the hardware.... the SDK will install the CUDA Toolkit 08 automatic and it ask if hardware (8800) is installed.
Pitty to start programdevelopment without hardware is not possible. 

[Devaster]

you can use emulation mode ...

[Vyper]

Correct there are a emulation mode to atleast see that the code WOULD work but you cannot count in speed in that euqation though ;-)...

//Vyper

[Vyper]

This is a statement from a nvidia forum regarding the CUFFT library:

"CUFFT API is modelled after FFTW -- this is the only reason of such ordering."

And Complex FFT is now managing up to 8M in size while s@h only needs up to 132K atm...

Start digging expert ones.. It should be doable cause the code is made for talking with FFTW libraries from default, it should only be neccesary with some "patches" to make it to work with Nvidia Cuda API..

//Vyper

P.S I mentioned on Nvidia forums that experienced persons who has programming skill is welcome here to try to convert FFTW -> CUFFT calls to perhaps get a rollout on a Cuda client.. If this is done then Nv hardware beat Ati of beeing the first HW to support S@H .. Could be cool .. D.S

[_heinz]

Installed the SDK..... read documentation, very interesting.... 

[_heinz]

NVIDIA CUDA
Windows Release Notes
Version 1.0
is now available
New Features
Hardware Support
  o  Additional hardware support added
     - Quadro FX 5600
     - Quadro FX 4600
     - GeForce 8800 Ultra
     - GeForce 8600 GTS
     - GeForce 8600 GT
     - GeForce 8500 GT

thats cool, a 8500 GT is available at a price of 110 Euro.
-----------------------------------------------------------------------------
heinz ~seti_britta

[_heinz]

Today I compiled my first CUDA example sucessful   
---------------------------------------------------------------------------------------------------------------------------
/* Example of integrating CUDA functions into an existing
 * application / framework.
 * CPP code representing the existing application / framework.
 * Compiled with default CPP compiler.
 */
----------------------------------------------------------------------------------------------------------------------------
------ Neues Erstellen gestartet: Projekt: cppIntegration, Konfiguration: Debug Win32 ------
Die Zwischen- und Ausgabedateien für das Projekt "cppIntegration" mit der Konfiguration "Debug|Win32" werden gelöscht.
Ein benutzerdefinierter Buildschritt wird ausgeführt.
tmpxft_00000fec_00000000-11.i
Kompilieren...
main.cpp
cppIntegration_gold.cpp
Code wird generiert...
Verknüpfen...
Das Manifest wird eingebettet...
Postbuildereignis wird ausgeführt...
Das Buildprotokoll wurde unter "file://c:\Programme\NVIDIA Corporation\NVIDIA CUDA SDK\projects\cppIntegration\Debug\BuildLog.htm" gespeichert.
cppIntegration - 0 Fehler, 0 Warnung(en)
========== Alles neu erstellen: 1 erfolgreich, Fehler bei 0, 0 übersprungen ==========
heinz ~seti_britta

[Vyper]

Hey.. There you go..

Keep it up and keepem clients coming 

Kind Regards Vyper