A GPU kernel developer writing compute kernels for a specific accelerator handles kernel source and performance tuning, aiming to make the kernel run correctly and efficiently on the target hardware.
Vendor-specific languages such as CUDA and existing kernel libraries are the identifiable current practice.
The public material states no specific pain; only an inference that vendor-specific languages impose learning and migration costs, with no direct evidence of user complaints or workarounds.