Paul Fultz II comments

Results 338 comments of


                                            Paul Fultz II

Fix: Driver --batch option sets Window Dimensions.

The driver is for developers. So the batch is not a priority to fix.

failure of test_topk<migraphx::shape::float_type, 1000, 120000> is based on the input data pattern

This can happen when the values are very close with ±1 ULP due to subtle differences in the CPU and GPU. Its not an issue that needs to be fixed.

change (cerr -> cout) for consistency across all error messages

Error messages should go to `cerr` not `cout`.

change (cerr -> cout) for consistency across all error messages

What is the motivation for this change?

change (cerr -> cout) for consistency across all error messages

> Warnings and Exceptions are being dumped on either stream as well. These should go to `cerr` unless they are being used in some kind of trace(like in `MIGRAPHX_TRACE_ELIMINATE_CONTIGUOUS`) or...

reorder_slice_add_mul matcher

Unit tests needs to be added.

reorder_slice_add_mul matcher

> Can we move the find_splits{} matcher in simplify_algebra.cpp up or have a copy of it run before find_mul_add{}? That should find the fusion opportunity over the slice instructions. No,...

reorder_slice_add_mul matcher

Also, unit tests should be added to test for the original test case.

reorder_slice_add_mul matcher

There is still a test case missing for the original issue found in the model where there is a add->dot->add being horizontally fused with one dot having an additional multiply.

Add optional fp16 rmsnorm conversion pass to fix fp16 accuracy

I dont think this is exactly the right approach. I would prefer we update the `rewrite_reduce` pass to lower `reduce_mean` to `reduce_sum`(for all cases). And then we can have a...