Massively Parallel Huffman Decoding on GPUs

[SolidComp]

I don't remember if I posted this before, but a lot of threads were deleted anyway:

https://github.com/weissenberger/gpuhd

https://www.researchgate.net/publica...coding_on_GPUs

It's interesting that they compare it to the Zstd CPU-based Huffman decoder.

The need for CUDA seems like a serious constraint in the middle of AMD's new dawn... I wonder what it buys you over using OpenCL or whatever the Windows equivalent is (DirectX Math?)

[Bulat Ziganshin]

CUDA is easier to program than OpenCL 1.2 and has more features. OpenCL 2.0 is still unsupported by NVidia cards

It's why I prefer to use CUDA for research, and for production - either OpenCL 1.2 (for simpler algorithms) or double implementation OpenCL2+CUDA .

[thorfdbg]

NVIDIA managed to establish a whole development community around CUDA, they offer support, courses, training, compilers and a toolchain around it. AMD does pretty much... nothing. When I visited the IBC trade show last September, people from NVIDIA came along asking whether we would be interested in this or that feature, for training etc.... Nobody from AMD came. Hence, no miracle CUDA is popular, it is pushed by its vendor.