leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

CUTLASS 3.4.0 (#1286)

Browse Source 
* CUTLASS 3.4.0

* Update CHANGELOG.md

---------

Co-authored-by: Pradeep Ramani <prramani@nvidia.com>
This commit is contained in:
Pradeep Ramani
2023-12-29 15:21:31 -05:00
committed by GitHub
co-authored by Pradeep Ramani
parent b7508e3379
commit 8236f30675
211 changed files with 11409 additions and 2763 deletions
Download Patch File Download Diff File Expand all files Collapse all files
include/cutlass/gemm/dispatch_policy.hpp
+21
View File
@@ -53,6 +53,8 @@ struct KernelTma { };
struct KernelTmaWarpSpecialized { };
struct KernelTmaWarpSpecializedPingpong { };
struct KernelTmaWarpSpecializedCooperative { };
struct KernelArrayTmaWarpSpecializedCooperative { };
struct KernelGroupTmaWarpSpecializedCooperative { };
//////////////////////////////////////////////////////////////////////////////
@@ -65,6 +67,8 @@ struct KernelTmaWarpSpecializedCooperative { };
struct KernelTmaWarpSpecializedFP8FastAccum : KernelTmaWarpSpecialized { };
struct KernelTmaWarpSpecializedPingpongFP8FastAccum : KernelTmaWarpSpecializedPingpong { };
struct KernelTmaWarpSpecializedCooperativeFP8FastAccum: KernelTmaWarpSpecializedCooperative { };
struct KernelArrayTmaWarpSpecializedCooperativeFP8FastAccum : KernelArrayTmaWarpSpecializedCooperative { };
struct KernelGroupTmaWarpSpecializedCooperativeFP8FastAccum : KernelGroupTmaWarpSpecializedCooperative { };
// Policies to opt into mixed type GEMMs
struct KernelTmaWarpSpecializedMixedInput : KernelTmaWarpSpecialized { };
@@ -225,6 +229,23 @@ struct MainloopSm90TmaGmmaWarpSpecializedFP8
"KernelSchedule must be one of the warp specialized policies");
};
// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp specialized dynamic schedule for Ptr-Array and Grouped Gemm
template<
int Stages_,
class ClusterShape_ = Shape<_1,_1,_1>,
class KernelSchedule = KernelGroupTmaWarpSpecializedCooperative
>
struct MainloopSm90ArrayTmaGmmaWarpSpecialized {
constexpr static int Stages = Stages_;
using ClusterShape = ClusterShape_;
using ArchTag = arch::Sm90;
using Schedule = KernelSchedule;
static_assert(
cute::is_base_of_v<KernelArrayTmaWarpSpecializedCooperative, KernelSchedule> ||
cute::is_base_of_v<KernelGroupTmaWarpSpecializedCooperative, KernelSchedule>,
"KernelSchedule must be one of the Ptr-Array or Grouped Gemm TMA Warp Specialized Cooperative policies");
};
//////////////////////////////////////////////////////////////////////////////
} // namespace cutlass::gemm