Work in progress: this is a wiki post, everyone is welcome to modify it directly
See the previous published edition.
Welcome to the sixteenth issue of the MLIR (bi)Weekly, a newsletter (published on Friday) covering developments in MLIR, and related projects in the ecosystem. MLIR (bi)Weekly is brought to you by a collective effort of contributors, we welcome your contributions!
- A small tutorial page was written to help understanding better the C++ class associated to the IR structure, and how to traverse it.
- The global dialect registry will be removed in the next two weeks, please update if you haven’t already!
- The C API development is making progress with StringRef, AffineMap,
- Multiple discussion and design on the Python bindings, e.g. Revisiting ownership and lifetime in the python bindings
- TableGen now emits directly the namespace nesting in the generated file, and fully qualify references to symbols with the entire namespace path. This makes it more robust overall to ambiguous name resolution.
- We have added buffer allocation support for
shape.assuming, which was required for the kernel generator project (TensorFlow).
shape.shape_ofnow lowers via the newly added
dynamic_tensor_from_elements, avoiding stack allocated memrefs while the IR is still at tensor level.
- We added shape.cstr_require to model arbitrary constraints that are expressed outside of the shape dialect.
- A first pass to lower
shape.cstr_*operations to side-effecting assert operations is underway. This will allow us to actually check shape related constraints in generated code.
Optimizations and Code Generation
- Now that matvec runs correctly in XLA:CPU, focus has shifted to performance. Currently better than the hand-written emitter for AVX2, still investigating regressions for customers that cannot use AVX.
- After some discussion we decided to enable optimizations that assume 32-bit indices by default for the Vector dialect, since this yields the best performance (it is unclear if vectors really need full 64-bit index space; clients can still opt-out though)
- Ongoing brainstorming and prototyping with sparse tensor lowering
- SPIR-V target environment resource limits are enhanced to include more fields like subgroup size, max shared memory size, vendor/product id, etc.
- Recursive struct support is coming to SPIR-V
- Unification work on Linalg tensors / buffers has started.