4 ms·
Very interesting. This could be applied recursively to AI compilers to accelerate the training pipeline. After all, GPU code is just a bunch of kernels expresse
by haltist 3y ago
Very interesting. This could be applied recursively to AI compilers to accelerate the training pipeline. After all, GPU code is just a bunch of kernels expressed in some syntax that is very similar to LLVM low level representation so the same strategy that works for LLVM should work for other intermediate representations of GPU kernels.