On Mon, 14 May 2012 15:45:54 -0500, Gor Gyolchanyan <[email protected]> wrote:
That's why CUDA is completely pointless: you're gonna wrap it anyway. Instead of wrapping a past-age vendor-locked pre-OpenCL technology, you could wrap modern portable technology. The richness of the language is completely irrelevant, as it's gonna be wrapped anyway. Yes, CUDA is completely pointless.
CUDA, or more specifically the PTX byte code is an open standard with backends for x86, ATI and NVIDIA. And although we are providing a mapping from D to an intermediate representation, that representation should be able to express as many of the higher level concepts as possible in order for the JIT to generate the most optimal code. Technologically, OpenCL/DirectCompute are at a level below CUDA 1.0 (we are now at 4.x) since both standards were written with the lowest common denominator in mind. Furthermore, at least on the HPC side, until another manufacturer decides to make a GPU with ECC support, NVIDIA is the only game in town.
