The reason for d3dx9_30 was that this DLL is available on a majority of systems.
Of course I don't know if the bug or whatever it is was fixed in any DirectX version. As it seems, it only appears on very few systems that only have XP in common. The only thing I know is that it's not a Gamestudio bug and that I can not reproduce or test it. Therefore, using a different DX version is the only thing I can do at the moment. If someone who has this bug volunteers, I can send him a special engine with the latest DirectX version for testing.
I could give it a try. The problem is very heavy on my system and I have got a cuple of easy and fast to use postprocessing shaders. My eMail adress if needed can be found in my profile.
Some good news: After spending about $1500 for new test systems we finally found a system where the problem could be reproduced.
We believe that it's a bug in the nVidia Win2K drivers. It does not happen with any other nVidia drivers, such as the Vista drivers, even on the same system. The Win2K drivers are used for XP.
Hopefully with this new information we can find either a workaround or a way to get it fixed by nVidia.
Great to hear this! I hope things go smooth in finding a solution. And thank you for purchasing new equipment to reproduce this; I really appreciate your efforts.
Update: The problem looks like a command buffer overflow in the driver, caused by certain shader commands. That's why it only happened with certain postprocessing shaders.
I think we can implement a workaround so that it does not happen anymore even with the current nVidia Win32K drivers. Unfortunately this is too late for the upcoming update that I don't want to delay further. But it will be addressed after that.