> why won't everyone migrate to Vulkan compute shaders? 1. No libraries. cudnn, cublas, cufft, are all the fastest available (except maybe magma sometimes), pl

> why won't everyone migrate to Vulkan compute shaders? 1. No libraries. cudnn, cublas, cufft, are all the fastest available (except maybe magma sometimes), plus no one writing an actual application wants to reinvent a fast gemm. Also cutlass, cub, thrust, ... 2. No c++. The "standard" seems to be glsl, and a "prototype" opencl c -> spir-v compiler doesn't give me much confidence in that approach. 3. No one wants to use vulkan apis directly, and there are approximately 5 billion different "vulkan compute" wrapper utility libraries. i.e. No consistent platform.