commit 5feef4868aa2b4a088fa13c316b0e2a9f5924d01
author bjacob <benoitjacob@google.com> Thu Oct 21 16:53:58 2021 -0400
committer GitHub <noreply@github.com> Thu Oct 21 16:53:58 2021 -0400
tree 8f7589374a62e6d34f399e0d984dd78fbf3919a8
parent a25c12cdeedce472f5550cfe285dd1dbe956ae2f

end-to-end matmul tests (#7347)

end-to-end matmul tests, meaning: we generate input matrices, perform linalg.matmul's, check that the results are correct.

In an earlier attempt (PR #7154) the test was entirely generated code. That approach was abandoned because it potentially amounted to comparing the compiler's output with itself.

The new approach here is to have a custom C test runner that has its own C reference implementation of matmul and uses it to compare the output of the generated code, which it drives by means of replay-trace. So now each test is not just a .mlir source, but a (.mlir, .yaml) pair where the yaml file is a trace, determining the runtime values of the inputs of the code being tested. Both .mlir and .yaml files are co-generated by a Python script.

A new CMake function iree_trace_runner_test is added. It is a variant of iree_check_test, taking a custom trace runner and trace file, instead of iree_check_test's behavior of just running the module in iree-check-module.

A new CMake function iree_generated_trace_runner_test is added to call the Python script to co-generate the .yaml and the .mlirsource and call iree_trace_runner_test on the resulting files.

One aspect of PR #7154 is retained here: all the above-mentioned CMake functions are generalized also to allow running iree-opt (or an alternate tool) as a pre-processing step on the source .mlir file, before running iree-translate. That is because by default, new MLIR transformations are not necessarily known to iree-translate. In our case, that's the mmt4d-related transformations.

The tests added in this PR share the prefix e2e_matmul (that can be used with ctest -R).