Ettore Tiotto

MLIR has passe that allow detection of buffer allocation and injection of deallocations. The following documentation https://mlir.llvm.org/docs/BufferDeallocationInternals advertises the existence of a buffer hoisting pass.

cmake build broken on Mac

4

The latest version of `onnx-mlir` can no longer build cleanly on Mac using `ninja`. This is the error: ``` CMake Error: Running '/usr/local/bin/ninja' '-C' '/Users/[email protected]/Source/onnx/onnx-mlir/build' '-t' 'recompact' failed with: ninja:...

Possible memory leak in getDataLayout(const Location &loc)

1

Potential memory leaks reported by `valgrind`: ``` ==95294== 25 bytes in 1 blocks are possibly lost in loss record 4 of 91 ==95294== at 0x51AEE63: operator new(unsigned long) (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)...

Threat all warnings as errors.

9

Given that `onnx-mlir` is a project that leverages LLVM and MLIR, and that the LLVM code policy is to consider warnings as if they were errors (https://llvm.org/docs/CodingStandards.html#treat-compiler-warnings-like-errors), I think we...

Choice of placeholder type during ONNX translation to onnx mlir dialect

2

During code review of https://github.com/onnx/onnx-mlir/pull/1110 there was a discussion on the meaning of `tensor`. Currently that type is used as a `placeholder` for the output type of the onnx mlir...

Fix CoalescePass

2

The `setCoalescedEncoding` function can handle operations that have a 'mem access ptr' with type `RankedTensorType`: ``` void setCoalescedEncoding(ModuleAxisInfoAnalysis &axisInfoAnalysis, Operation *op, int numWarps, int threadsPerWarp, llvm::MapVector &layoutMap) { Value ptr...