CUTLASS 2.11 - November 2022

CUTLASS is a collection of CUDA C++ template abstractions for implementing high-performance matrix-multiplication (GEMM) and related computations at all levels and scales within CUDA. It incorporates strategies for hierarchical decomposition and data movement similar to those used to implement cuBLAS and cuDNN. CUTLASS decomposes these "moving parts" into reusable, modular software components abstracted by C++ template classes. These thread-wide, warp-wide, block-wide, and device-wide primitives can be specialized and tuned via custom tiling sizes, data types, and other algorithmic policy. The resulting flexibility simplifies their use as building blocks within custom kernels and applications.

To support a wide variety of applications, CUTLASS provides extensive support for mixed-precision computations, providing specialized data-movement and multiply-accumulate abstractions for half-precision floating point (FP16), BFloat16 (BF16), Tensor Float 32 (TF32), single-precision floating point (FP32), FP32 emulation via tensor core instruction, double-precision floating point (FP64) types, integer data types (4b and 8b), and binary data types (1b). CUTLASS demonstrates warp-synchronous matrix multiply operations targeting the programmable, high-throughput Tensor Cores implemented by NVIDIA's Volta, Turing, and Ampere architectures.

CUTLASS implements high-performance Convolution via the implicit GEMM algorithm. Implicit GEMM is the formulation of a convolution operation as a GEMM thereby taking advantage of CUTLASS's modular GEMM pipeline. This allows CUTLASS to build convolutions by reusing highly optimized warp-wide GEMM components and below.

See the Quick Start Guide to get started quickly.

See the functionality listing for the list of operations supported at each level of the execution model hierarchy.

CUTLASS 2.11 is an update to CUTLASS adding:

• Stream-K, which is a new general way to do split-K. It can not only improve performance, but can also significantly reduce the number of tile sizes that need to be profiled to find the best one.
• Fused multi-head attention kernel. It has two variants: one for fixed sequence lengths, and another for variable sequence lengths.
• Dual GEMM. It can run two GEMMs that share the same left input matrix in one kernel.
• Hopper improves double precision matrix multiplication by 2x compared to Ampere at iso-clocks. It is supported since CUDA 11.8.
• BLAS3 functions with Hoppers new double precision matrix multiplication instructions.
• ELL Block Sparse GEMM.
• Optimized Group Conv.
• Optimized DepthWise Conv.
• Scripts to fuse multiple back-to-back GEMM.
• FP8 data type definition and conversion routines.
• Updates and bugfixes from the community (thanks!). Big shout out to Meta's xFormers.
• Deprecation announcement: CUTLASS plans to deprecate the following in the next major release:
  • Maxwell and Pascal GPU architectures
  • Ubuntu 16.04
  • CUDA 10.2

See the CHANGELOG for a detailed listing of releases and updates.