Stephen Neuendorffer

With the --localize-locks pass converting locks into integer indexes, we no longer need to be restricted to useLock operations existing primarily in CoreOps. Instead, we can simply pass around the...

Make aie-vectorize robust to unimplementable vector sizes

1

Currently the vectorizer will simply assert if an incompatible vector size is created with the super-vectorizer.

These magic register values should get factored out somewhere, probably runtime_lib/test_library.cpp

This needs to get moved to a proper build location.

Build direct-to-binary backend target

2

Current backend targets use the LibXAIE API to configure the AIE array through memory-mapped accesses. This has a number of downsides: 1) LibXAIE has significant memory usage and we want...

Currently, we target a device where Tile 0,0 is always in the lower left corner of the device. However, in many cases we want to implement code onto a portion...

Currently, we consider the location where code runs to be physical locations in specific tiles. We should create a logical model of processes and channels (possibly based on the CIRCT...

Currently locks are accessed absolutely, i.e. with a specific tile location and lock number. In addition, host code has explicit libXAIE calls. We should create a local model for locks...

The MLIR assume_aligment op can annotate alignment information on memrefs. We currently generate an assume_alignment op for all AIE buffers: `memref.assume_alignment %lineOut, 32 : memref ` This will enable successful...