4 ms·
So between this and gdev (https://github.com/shinpei0208/gdev https://github.com/shinpei0208/gdev), is it now (or soon to be) possible to write CUDA in Linux wi
by tsuraan 14y ago
So between this and gdev (https://github.com/shinpei0208/gdev https://github.com/shinpei0208/gdev), is it now (or soon to be) possible to write CUDA in Linux with both an open toolchain and without the nVidia binary blob for video cards? Or am I misunderstanding what's going on here?
- keveman 14y agoYou are partly right. The frontend of the compiler that translates the CUDA dialect of C++ to LLVM IR is not open source.
- ajross 14y agoI'm trying to figure that out too. But my guess is no: this compiler spits out a blob that needs to be fed to the proprietary NVIDIA driver. It exposes (but doesn't really "document") the functionality of the compute engine (and maybe the texture units?). It doesn't touch the DMA engine for getting the data on and off the card, nor the interface to the hardware schedule to make it run.
- codedivine 14y agoNo. It does not generate GPU binaries. It generates "PTX", which is a pseudo-assembler format. You then feed the PTX to Nvidia driver to generate the actual binary. AFAIK the actual ISA and binary formats are not openly (officially) documented.