This repository contains a development version of the CUTLASS repository with experimental SYCL support enabled. The aim is to support other SYCL-enabled devices with the minimal source code modifications by using the same CUTLASS features. possible, facilitating maintenance. Given that most of the backend work happens on the CUTE implementation, the CUTLASS interface remains the same, and the SYCL support only needs changes at the atom and pipeline level.
SYCL[1] is a royalty-free, cross-platform abstraction layer that enables code for heterogeneous and offload processors to be written with modern ISO C++, and provides API and abstractions to find devices and manage resources for GPUs.
Currently, only one example works on NVIDIA SM 80.
To build CUTLASS SYCL support you need the latest version of DPC++ compiler. You can either use a recent nighly build or build the compiler from source as described in oneAPI DPC++ guidline.
Once you have your compiler installed, you need to point the
CMAKE_CUDA_HOST_COMPILER flag to the clang++ provided by it.
This enables the compilation of SYCL sources without altering the current NVCC path.
cmake -G Ninja \
-DCMAKE_CUDA_HOST_COMPILER=/path/to/clang++ \
-DCUTLASS_ENABLE_SYCL=ON \
-DDPCPP_SYCL_TARGET=nvptx64-nvidia-cuda \
-DDPCPP_SYCL_ARCH=sm_80
Currently, you can build the CuTe Tutorial using the following command:
ninja [EXAMPLE_NAME]_sycl
You can run it like this from your build directory
LD_LIBRARY_PATH=/path/to/sycl/install/lib ./examples/cute/tutorial/[EXAMPLE_NAME]_sycl
Currently, the the example 14_amper_tf32_tensorop_gemm has been implemented for SYCL on Nvidia Ampere architecture. You can build this from your build directory by running :
ninja 14_ampere_tf32_tensorop_gemm_cute
You can run it like this from your build directory
NVIDIA_TF32_OVERRIDE=1 LD_LIBRARY_PATH=/path/to/sycl/install/lib ./examples/14_ampere_tf32_tensorop_gemm/14_ampere_tf32_tensorop_gemm_cute