Questions on ti.sync under taichi scope and python scope

Yuhai · 2021 年10 月 31 日 05:51

1.When call ti.sync on python scope,it is equal to cudaDeviceSynchronize or cudaStreamSynchronize ?
2.I can call ti.sync in kernel scope. Does it equal to __syncthreads() in CUDA?

Thanks!

Velaciela · 2021 年11 月 3 日 04:27

Currently ti.sync() corresponds to cudaStreamSynchronize() on the CUDA backend
I find interfaces in these two places:
taichi/program/program.cpp
taichi/llvm/llvm_program.cpp
Taichi uses CUDA Driver API, so it’s actually cuStreamSynchronize().

Calling ti.sync() within @ti.kernel is undefined behavior, not sure if there are side effects.
I initialized like this ti.init(print_kernel_nvptx=True,arch=ti.cuda), and didn’t see relevant sync barrier statement in the PTX kernel