Jerry Wu

Add files to describe benchmarks in python

7

Files to describe the benchmarks in python. The python scripts to define benchmarks will go under `build_tools/benchmarks/configs`

**Backgrounds** Currently under `build_tools` we have executable python scripts in multiple directories (`benchmarks`, `docker`, ...). They contain some duplicate utility functions, which can be factored out into the common python...

Currently when building the benchmark suites, we fetch the models from the network as one of the build dependencies. This makes the build error confusing when there is a network...

Currently `build_tools/scripts/download_file.py` doesn't retry when there is a network issue, which causes the benchmark suites to fail to build sometimes. We should either add the retry mechanism to `download_file.py` or...

This is the tracking bug for bufferization issues related to the multi-use/multi-results fusion created by the new fusion heuristics. - [ ] Fix inplace bufferization check in DispatchLinalgOnTensors.

[mlir] tileConsumerAndFuseProducerGreedilyUsingSCFForOp incorrectly fuses on reduction dim

3

When using `scf::tileConsumerAndFuseProducerGreedilyUsingSCFForOp` to tile and fuse `linalg.matmul` with `linalg.fill`, it doesn't keep `linalg.fill` outside the reduction loop, which results in incorrect result. This can be reproduced with https://github.com/pzread/llvm-project/commit/9ea44bdee2f0b47289f80061248c88223c0c99ab. In...

## Situation The set of tests can be stopped early if any tests failed. ## Expectation Reduce the usage of time on fail tests.

## Situation The default scheduler of Linux is not suitable for judge workloads. ## Expectation Config the scheduler to improve the judge performance.