qazal
qazal
cool! I'll take some time to review this today.
This doesn't look right: https://github.com/tinygrad/tinygrad/actions/runs/10621574069/job/29444432870?pr=5976#step:24:480 In general, it's good practice to keep your branch up-to-date and make sure the process replay diff matches what the reviewer would expect. Like this...
feel free to mark as ready for review when the upcast issues are fixed.
benchmarks red https://github.com/tinygrad/tinygrad/actions/runs/10714523120/job/29708335366
PTX looks wrong https://github.com/tinygrad/tinygrad/actions/runs/10717084318/job/29716009654?pr=5976#step:6:1
I'd merge a diff that has ASTs in test_linearizer passing verify_lazyop. That'd help make progress on Lowerer changes independently. Also, we're near the line count limit but don't worry too...
How's this going?
sg, I think we can independently make some progress on the UOpGraph toposort.
cool, you know we can't merge big diffs; But it's worth exploring the full scope and incrementally ship in clean, small diffs.
yea this will work right now because we always have both scalar and vec dtype UOps in a single kernel. But you should consider the prefix import in both paths...