CUTLASS 3.8 Release (#2059)

* CUTLASS 3.8 Release

* update

* Update README.md

* Revert "Update README.md"

This reverts commit b353e36fe83e0815f99b44e46c0c95494c44726b.

* update

* update

---------

Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
mihir-awatramani
2025-01-25 02:44:06 -05:00
committed by GitHub
co-authored by Haicheng Wu Haicheng Wu
parent 9eb01fa0b0
commit 389e493055
290 changed files with 91222 additions and 291 deletions
+5 -1
View File
@@ -219,7 +219,11 @@ which has to happen at the end among the participating warps.
This is because each warp computes using only a "slice" of CtaTileK,
so each warp only has a partial sum before the reduction.
### Warp Specialization
### Hopper Warp Specialization
Note: the following section on warp-specialization contains details that are specific
to the Hopper kernel design. Blackwell SM100 kernels have a substantially different warp-specialization structure,
however, the concept of separating out producer and consumer agents still applies.
Starting with Hopper, CUTLASS 3.0 incorporates the concept of [Warp Specialization](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#spatial-partitioning-also-known-as-warp-specialization)
as part of the kernel design. A thread block is partitioned into two sets of warps, [*producer* warp group](../../include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized.hpp) and [*consumer* warp group](../../include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized.hpp). The *producer* warp group loads data from global memory into shared memory buffers using the new [Tensor Memory Accelerator (TMA)](https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/).