For anyone wants to understand the codebase

[Lurkrazy]

I would recommend to check out the HPCRL/tvm.

This implementation is well-organized and entirely Python-based, making it easier to follow and understand the search strategy logic. If you're looking for additional clarity or insights into how the code works, this implementation could be helpful.

[lqq-lqq]

Why it is invalid when "up_tb < reg1 * reg2" ? The code location is following:

ics24tvm/src/auto_scheduler/search_policy/sketch_policy.cc

Line 672 in 0558769

if (up_tb >= reg1 * reg2) {

Looking forward to your answer.

[lqq-lqq]

Sorry, I was missing a line before, I got it!

[Lurkrazy]

Why it is invalid when "up_tb < reg1 * reg2" ? The code location is following:

ics24tvm/src/auto_scheduler/search_policy/sketch_policy.cc

Line 672 in 0558769

if (up_tb >= reg1 * reg2) {

Looking forward to your answer.

We mutated in the shared memory view, which means, the TB_tile = thread block size * reg1 * reg2. So up_tb < reg1 * reg2 is not valid here.

[Lurkrazy]

Sorry, I was missing a line before, I got it!

Thank you for your interest in our work. I'd be happy to assist with any questions you have regarding the code.