[NHervé]

Thanks, but I think accuracy is not the only thing in emulation. There is also the speed, and this is not the strong point of DeSmuME.
I've remarked that the MMU I/O funcs contain a lot of "proc==ARMCPU_ARMx" assertions. I think rewriting that with a group of funcs should speed up a little.
(Dualis uses a lot of funcs for I/O handling, each for a memory region, and a group for each proc).

PS : Shash : I enabled GL_NORMALIZE that forces the normals to be normalized. This avoids a conflict when using glScale.

I compared my emu to the real DS, it's exactly identical, unless my DS is buggy. In your screen the water looks too gray.