- User Since
- Feb 2 2017, 2:24 AM (182 w, 6 d)
Landed in 1a4263d394c1a93757613bde4b1c2cf8d6a7bbb9
I took the liberty to patch and land as I plan on using this and avoid reinventing the wheel.
Please fix the build though
Mon, Aug 3
Wed, Jul 29
Much better, thanks @kumasento
Nice use of the language :) !
Tue, Jul 28
Mon, Jul 27
Make alloca default again.
Fix map capture, $ is now required for globals.
Drop CMake changes.
@chelini thanks for your contribution!
Can we please land this?
Wed, Jul 22
Thanks for the fix @maheshkhanwalkar
Looks great, thanks!
Note that a lot of this code is old and I have plans to refactor for more general usefulness between affine and linalg, but it is not highest pri for me for now.
In the meantime, exposing more things as "composable transformations" and less as passes is great, thank you for the patch!
Mon, Jul 20
Fri, Jul 17
Looks good thanks!
I've been using this for some fast/slow path lowering of vector.transfer ops, a bunch of stacked revisions are necessary to get there.
Thu, Jul 16
Landed as 22c8a08fd8a1487159564f74f24561964f6a6c97.
Very sorry for the delay, thanks Jakub!
Wed, Jul 15
Tue, Jul 14
Thank you for the cleanup @agrue!
Please let me know if you want me to land this in case you don't have push access.
Mon, Jul 13
Great, is this the root cause of the issue we were seeing on older HW ?
Fri, Jul 10
This looks good to me, there are more relaxations we can do in the future but we should be good for now.
Much nicer now, thanks for cleaning this up!
Refactoring to use the new operation seemed another logical unit.
(but I guess this is optimized by backend again anyway)
Made a quick pass but I do not think the algorithm is complete atm: in order to check lack of aliasing you need to look at whether the static intervals between read/write pairs overlap.
Here the positive example only has trivial intervals (i.e. that reduce to a single point).
Thanks for adopting what we discussed previously and special casing the impl!