Skip to content

Latest commit

 

History

History
100 lines (83 loc) · 6.94 KB

README_sycl.md

File metadata and controls

100 lines (83 loc) · 6.94 KB
Raw

RELION SYCL/DPC++ version

This is SYCL/DPC++ version of RELION

Build & Running

$ mkdir build_sycl; cd build_sycl
$ {Load Intel oneAPI toolkit and SYCL/Level Zero/OpenCL runtime environment}
$ sycl-ls                        # This will display available SYCL devices
$ cmake \
-DCMAKE_C_COMPILER=mpiicx \
-DCMAKE_CXX_COMPILER=mpiicpx \
-DMPI_C_COMPILER=mpiicx \
-DMPI_CXX_COMPILER=mpiicpx \
-DMKLFFT=ON \
-DSYCL=ON \
-DCUDA=OFF \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_C_FLAGS="-O3" \
-DCMAKE_CXX_FLAGS="-O3 -march=native" \
..

$ #### This is Intel GPU Level Zero backend specific #####
$ export ZE_AFFINITY_MASK=0         # Use only the first available Level Zero device. This can be replaced by --gpu 0 syntax.
$ export ZEX_NUMBER_OF_CCS=0:4,1:4  # Set this only if you are putting more than one MPI ranks per GPU. 0:4 means 4 MPI ranks running on card 0
$ #### End of Intel GPU Level Zero backend specific  #####
$ # For finer control of SYCL devcices, please see the above descrpition on ONEAPI_DEVICE_SELECTOR
$ {Run 2D/3D/refinement application by replacing --gpu/--cpu with --gpu/--sycl/--sycl-opencl/--sycl-cpu/--sycl-cuda/--sycl-hip}

Optional runtime environment variables

  • The below shell environment variables can be tested for more potential SYCL specific tuning. Setting it to "1" or "on" will enable these features.
    • relionSyclUseCuda: This with --gpu have the same meaning as --sycl-cuda. --sycl-cuda will be used even if --gpu/--sycl is specified in command lines.
    • relionSyclUseHip: This with --gpu have the same meaning as --sycl-hip. --sycl-hip will be used even if --gpu/--sycl is specified in command lines
    • relionSyclUseInOrderQueue: Use in-order SYCL queue. Without this, out-of-order SYCL queue is used by default. (experimental)
    • relionSyclUseAsyncSubmission: Remove wait() for each SYCL kernel submission. (experimental)
    • relionSyclUseStream: Create new in-order SYCL queue for each cudaStream. (experimental)
    • relionSyclUseSubSubDevice: Create separate SYCL queue for each CCS. (experimental)
    • relionSyclBlockSize: SYCL memory pool block size. This takes precedence over relionSyclHostBlockSize and relionSyclDeviceBlockSize.
    • relionSyclHostBlockSize: SYCL memory pool block size for sycl::malloc_host. Default is 256MB
    • relionSyclDeviceBlockSize: SYCL memory pool block size for sycl::malloc_device. Default is 256MB
    • MAX_MPI_BLOCK: Maximum MPI message size per single MPI API call for point-to-point and collective communication. This takes precedence over MAX_MPI_P2P_BLOCK and MAX_MPI_COLL_BLOCK.
    • MAX_MPI_P2P_BLOCK: Maximum MPI message size per single MPI API call for point-to-point communication. Default is 4GB.
    • MAX_MPI_COLL_BLOCK: Maximum MPI message size per single MPI API call for collective communication. Default is 64MB.

Added macros

  • For CMake configuration

    • SYCL(=ON/OFF): Enable SYCL based acceleration build
    • SYCL_CUDA_COMPILE(=ON/OFF): Enable SYCL compilation for CUDA target
    • SYCL_CUDA_TARGET: SYCL CUDA arch target (i.e. 80)
    • SYCL_HIP_COMPILE(=ON/OFF): Enable SYCL compilation for HIP target
    • SYCL_HIP_TARGET: SYCL HIP arch target (i.e gfx90a)
    • SyclForceOneDPL(=ON/OFF): Use oneDPL(https://github.com/oneapi-src/oneDPL) if it can be used. This has the same effect as setting "-DUSE_ONEDPL" for CMAKE_CXX_FLAGS below. (experimental)
    • SYCL_AOT_COMPILE(=ON/OFF): Enable AOT(Ahead-Of-Time) compilation for SPIR64 target. Default target is pvc. (for future use)
    • SYCL_AOT_TARGET(=ON/OFF): Specify AOT(Ahead-Of-Time) SPIR64 target. Possible list can be checked using "ocloc compile --help" command. (for future use)
    • SYCL_HOST_FLAGS(=list of flags with space as separator): Additional flags for host compiler (for future use)
    • SYCL_COMPILER_NAME: SYCL compiler command name (for future use)
    • SYCL_COMPILE_FLAGS: Additional SYCL compile flags (for future use)
    • SYCL_LINK_FLAGS: SYCL link flags except "-lsycl -lOpenCL" if needed (for future use)

  • Others for testing purpose (just defining with -D* is needed in cmake -DCMAKE_CXX_FLAGS)

    • USE_ONEDPL: Use SYCL kernel for weight sorting routines. If this is set, oneDPL(https://github.com/oneapi-src/oneDPL) is used when it is beneficial. (experimental)
      • USE_LESS_ONEDPL: If this is set, oneDPL is used only when there is no other implementation.
      • USE_MORE_ONEDPL: If this is set, oneDPL is used everywhere when it is applicable. If this is set, you SHOULD NOT SET DisableIndirectAccess=1.
    • SYCL_OFFLOAD_FFT: Use SYCL kernel for the current FFTW routines. (Not implemented)
    • INTEL_EXPLICIT_SIMD: Use Explicit SIMD extension for SYCL kernels. (Not implemented)
    • INTEL_SG_SIZE: Used for Intel sub-group size in SYCL kernel. 32 is recommended for PVC and 16 is for ATS. (Not tested well)
    • USE_IPP: Use Intel IPP library's RadixSort for sortOnDevice instead of std::sort. Enabled by default if IPP library exists.
    • USE_MPI_COLLECTIVE: Use MPI collective whenever possible. Enabled by default for ALTCPU and SYCL.
    • USE_EXISTING_SYCL_DEVICE: This will copy and use created SYCL device pointer instead of creating new SYCL device for each thread. Not recommended.
    • USE_SINCOS_TABLE: Pre-calculate sine/cosine table before main loop in some kernels (Not implemented)

  • Macros defined in source code (No need to specify in CMake configuration and it is defined by -DSYCL=ON above)

    • _SYCL_ENABLED: source codes with SYCL which may be compiled with any SYCL compiler
    • _DPCPP_ENABLED: source codes with DPC++ which should be compiled with Intel oneAPI compiler only

  • Prefined macros by Intel(R) oneAPI compiler

    • __INTEL_LLVM_COMPILER: For icx and icpx