commit fe4403e076f4179c360dccbd330e304b9c0593eb
author Lei Zhang <antiagainst@google.com> Tue Jun 01 15:09:43 2021 -0400
committer GitHub <noreply@github.com> Tue Jun 01 12:09:43 2021 -0700
tree 1e17d7424bcb44811799a5f530ffa559a87b95b7
parent 155e91312fe6a612fe8af876cc0330853d89f0f5

[cmake] Support defining benchmark suites from MLIR input modules (#6024)

We have seen several benchmark artifacts generation issues in
the benchmark pipeline thus far. It's because the benchmark pipeline
is totally detached from the normal CI testing flow and it's not
run pre-submit.

This commit adds a CMake rule, `iree_mlir_benchmark_suite`, to
support defining benchmark suites from MLIR input modules. With it,
we can have benchmark artifact generated by building a normal CMake
target: `cmake --target iree-benchmark-suites`.

Concretely, this rule downloads imported model MLIR files, invokes
`iree-translate` to convert them according to the given translation
flags in the rule, and then generates the corresponding `flagfile`
for `iree-benchmark-module` to use.

With this commit, we should be able to make benchmark artifact
generation a step in the normal CI testing flow. Additionally,
we have the additional benefit that now we have a unified and simple
way for bots and developers to generate the full benchmark suite.

Ideally, we shouldn't be downloading with CMake; we should directly
compile from the initial source (i.e., TensorFlow Python models).
It avoids the manual step for uploading imported model somewhere for
every model we need to benchmark and avoids out-of-date inputs.
But that is tangled with the pending Python testing infrastructure
revamp so we'd prefer to not do that right now.