Well, i’ve implemented the VAR thingie, and under Linux there is absolutely no performance difference to not using VAR. My code is OK, I have the same codebase for Linux and Win32 and under Win32 VAR work perfectly and give tremendeous speedup. I can get the pointer to the glXAllocateMemoryNV, and I can allocate memory, but even with 0, 0, 1 as parameters the speed is just the same!
I dont thing this option exists for the binary NVIDIA driver, but anyway i’ve activated AGP 2x and sideband addressing(added an option to NVdriver in /etc/modules.conf) and the performance impoved to the level of Win32 without VAR, and calling any VAR function does not change the speed at all!
The program even runs at the same speed with arrays in AGP memory as in system memory with disabled VAR, which is VERY strange, since normal vertex arrays from AGP memory should be slower (uncached memory).
Further suggestions are much appreciated since I have abolutely no clue why is VAr not working.