cutlass 3.9 update (#2255)

* cutlass 3.9 update

* rebase

* fixes out of shared memory for blockwise Blackwell

* doc format

* fix issue 2253

* disable host ref by default

* fix sm120 smem capacity

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Yujia Zhai
2025-04-24 15:42:40 -04:00
committed by GitHub
co-authored by yuzhai Haicheng Wu
parent 8e345c5c5b
commit 331a1f5b3f
143 changed files with 18089 additions and 5935 deletions
@@ -437,9 +437,11 @@ void BlockScaledGemmOperationProfiler::GemmProblem::initialize_result(
set_argument(result, "k", problem_space, k);
set_argument(result, "cluster_m", problem_space, cluster_m);
set_argument(result, "cluster_n", problem_space, cluster_n);
set_argument(result, "cluster_k", problem_space, cluster_k);
auto cluster_shape = operation_desc.tile_description.cluster_shape;
auto is_dynamic = cluster_shape.m() == 0 || cluster_shape.n() == 0 || cluster_shape.k() == 0;
set_argument(result, "cluster_m", problem_space, is_dynamic ? this->cluster_m : cluster_shape.m());
set_argument(result, "cluster_n", problem_space, is_dynamic ? this->cluster_n : cluster_shape.n());
set_argument(result, "cluster_k", problem_space, is_dynamic ? this->cluster_k : cluster_shape.k());
set_argument(result, "cluster_m_fallback", problem_space, cluster_m_fallback);
set_argument(result, "cluster_n_fallback", problem_space, cluster_n_fallback);
set_argument(result, "cluster_k_fallback", problem_space, cluster_k_fallback);