AMDMIGraphX
AMDMIGraphX copied to clipboard
AMD's graph optimization engine.
We were not propagating the dominators for single input instructions, which caused us to miss some dominators.
Fix a bug that caused an exception in torch-migraphx benchmarks. Closes Issue https://github.com/ROCm/AMDMIGraphX/issues/3009
* Need to sort the inputs into `select_module` alphabetically by name otherwise there's a possibility that they don't match later because of the `order` attribute of the parameters in the...
Model `speech_transformer` in the DLM pytorch performance test causes an exception which is traced to an infinite recursion in the MIGRAPHX method `module::sort()`. The error is only seen in branch...
rocgdb is not supported on 22.04. This works around the issue by installing the older python version used on ubuntu 20.04(the newer python is still install in the dockerfile).
This fixes a bug with inner_broadcast and improves the handling for more cases. For the simple cases with the same broadcasts and dimension, it will just insert the broadcasting afterwards....
Updates the `find_pointwise_reduce` matcher in the `fuse_reduce` compiler pass, to not attempt to fuse a `multibroadcast` instruction if the shape has an incompatible number of dimensions (rank). This cures the...
Allows squeeze to work on an axis with dynamic_dimension that intersects with `{1, 1}`. Needed to work with fully unknown dimensions `{0, max_int}`.