[mlir-gen] Introduce basic support for quantization ops (2/n). #1089
Conversation
This patch tries to extend the mixed precision isupport and add basic infra to define or create quantization kernel as we get more clarity on kind of primitive operations required. Addition of run time test is not in scope of this patch.
|
General question, shouldn't we migrate to lighthouse's gen? |
That's indeed a good question. Might be a good example to dogfood how we facilitate downstream projects interacting with lighthouse. |
For now, I'd duplicate, because we don't know what we can build with this work. Once we're happy with the result, we should bring this up to the lighthouse and help talk about quantization upstream. |
|
I'd like to see at least two back-and-forth execution test:
|
|
Test failures are due to NFS migration and will be fixed in time |
…untime test. -Adds '-print-input' flag to print input arguments for visual inspection. -Refactored and updated the corresponding APIs.
This patch tries to extend the mixed precision isupport and add basic infra to define or create quantization kernel as we get more clarity on kind of primitive operations required.
Addition of run time test is not in scope of this patch.