Thanks @Laurawly for the proposal. I agree this is something that is good to have as they brings the best of all worlds.
Looking at the MNK=2048 benchmark. It seems to suggest a limitation in the auto-scheduler's search space construction (the hand written [topi recipie](https://github.com/apache/tvm/tree/main/apps/topi_recipe/gemm) gets to closer to peak about 80-90% of cublas on titan X). In complementary to this effort, we will also explore the possibility to utilize the primitives in cutlass, including those for the FMA computation and memory move via tensorization. So see if we can leverage these sub-functions in different finer-grained levels, eventually bringing the insights from cutlass to auto-scheduling. --- [Visit Topic](https://discuss.tvm.apache.org/t/rfc-byoc-nvidia-cutlass-integration/9147/2) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.apache.org/email/unsubscribe/dd39482a42cbf57a999fa6fb3d1f46befcfe7866081c6b6fe3047ba8ea399fae).
