Extented GenVector library for multi-target execution.
A basic CUDA configuration looks like this:
cmake .. -Dcuda=ON -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++ -DCMAKE_CUDA_COMPILER=/path/to/cuda/bin/nvcc -DCMAKE_CUDA_ARCHITECTURES=xx -DCMAKE_CUDA_FLAGS="-arch=sm_XX", where `XX` indicates the value of target CUDA capability.
SYCL offers different memory management strategies. Here, we implement buffers+accessors or device pointers. Default memory management is handled with device pointers. In order to enable buffers+accessors, it is sufficient to set -Dsycl_buffers=ON in the cmake configuration.
A basic SYCL configuration with AdaptiveCpp looks like this:
cmake .. -Dadaptivecpp=ON -DAdaptiveCpp_DIR=/path/to/AdaptiveCpp/install/lib/cmake/AdaptiveCpp -DACPP_TARGETS="<targets>"
AdaptiveCpp targets specification defines which compilation flows AdaptiveCpp should enable, and which devices from a compilation flow AdaptiveCpp should target during compilation. A CUDA example lookd like this: -DACPP_TARGETS="cuda:sm_86". A HIP exaple looks like this: -DACPP_TARGETS="hip:gfx90a"
A basic SYCL configuration with oneAPI targeting CUDA backends looks like this:
cmake .. -Doneapi=ON -Dsyclcuda=ON -DCMAKE_CUDA_ARCHITECTURES=XX -DCUDA_TOOLKIT_ROOT_DIR=/path/to/cuda/
where XX indicates the value of target CUDA capability.
A basic SYCL configuration with oneAPI targeting HIP backends looks like this:
cmake .. -Doneapi=ON -Dsyclamd=ON -DCMAKE_OFFLOAD_ARCHITECTURES=gfxXXX
where XXX is set accordingly to the AMD GPU model.
In order to compile test targets, set -Dtesting=ON in the cmake configuration. Set -Dsingle_precision=ON for compiling single precision test targets.
In order to enable time measurements (and their print to stdout), set -Dtiming=ON in the cmake configuration.