After #123, apply the general to kernel bench, so that we can start evaluating more kernels from the library. This may generate a lot of _specialized pipelines_ for each kernel / target pairs, which is ok for this stage. A later stage shall combine / common-up the multiple pipelines, with possible upstreaming of new passes/transforms, etc.
After #123, apply the general to kernel bench, so that we can start evaluating more kernels from the library.
This may generate a lot of specialized pipelines for each kernel / target pairs, which is ok for this stage.
A later stage shall combine / common-up the multiple pipelines, with possible upstreaming of new passes/transforms, etc.