CuTe GEMM

Layout

main.cu: a simple test file taken from the CUTLASS repository for testing a correct build setup. tiled_copy.cu: a CuTe copy kernel for practicing layouts, tile management, and copying. simple_gemm.cu: a naive matmul kernel using CuTe. utils.cuh: Benchamrking and cuBLAS utilities print_layouts.cu: Scratchspace for printing and examining CuTe layouts

Build

Clone the CUTLASS git submodule with git submodule update --init --recursive.

Then, select a compute capability in the CMakeLists.txt for your GPU:

if(NOT DEFINED CUTLASS_NVCC_ARCHS)
    # 80 is for a100
    set(CUTLASS_NVCC_ARCHS 80 CACHE STRING "CUDA architectures to compile for")
endif()

Change the number after CUTLASS_NVCC_ARCHS to do so, and see the CMakeLists.txt for some other options.

Then, build the project with Cmake:

mkdir build
cd build
cmake ..
make

The GEMM kernel and benchmarking is run with

cd build
./simple_gemm