makortel/pixel-standalone

Pixel raw2digi test program

The purpose of this test program is to experiment with various "performance portability" frameworks and libraries.

Overall structure

The test programs are divided in three units

• main_*.cc: contains main(), reads input, prints total timing. Plays the role of the "experiment framework".
• analyzer_*.cc: plays the role of a framework module (even though the "event loop" is there). Calls the memory tranfers (if necessary), and the computational kernel.
• rawtodigi_*.cc: contains the computational kernel (which is mostly just shuffling bytes around memory)

Current implementations

Implementation	make target	Executable (also make target)	#ifdef macros
Naive CPU	naive	test-naive	DIGI_NAIVE
CUDA	cuda	test-cuda	DIGI_CUDA
Alpaka	alpaka	test-alpaka	DIGI_ALPAKA, ALPAKA_ACC_*
- only on CPU		test-alpaka-ser (sync)	DIGI_ALPAKA, ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED
		test-alpaka-tbb (async)	DIGI_ALPAKA, ALPAKA_ACC_CPU_B_TBB_T_SEQ_ENABLED
		test-alpaka-omp2 (async)	DIGI_ALPAKA, ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED
- only on GPU		test-alpaka-gpu (async)	DIGI_ALPAKA, ALPAKA_ACC_GPU_CUDA_ENABLED
Cupla	cupla	test-cupla	DIGI_CUPLA, ALPAKA_ACC_*
- only on CPU		test-cupla-seq-seq-async	DIGI_CUPLA, CUPLA_STREAM_ASYNC_ENABLED=1, ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED
		test-cupla-seq-seq-sync	DIGI_CUPLA, CUPLA_STREAM_ASYNC_ENABLED=0, ALPAKA_ACC_CPU_B_SEQ_T_SEQ_ENABLED
		test-cupla-tbb-seq-async	DIGI_CUPLA, CUPLA_STREAM_ASYNC_ENABLED=1, ALPAKA_ACC_CPU_B_TBB_T_SEQ_ENABLED
		test-cupla-opm2-seq-async	DIGI_CUPLA, CUPLA_STREAM_ASYNC_ENABLED=1, ALPAKA_ACC_CPU_B_OMP2_T_SEQ_ENABLED
- only on GPU		test-cupla-cuda-async	DIGI_CUPLA, CUPLA_STREAM_ASYNC_ENABLED=1, ALPAKA_ACC_GPU_CUDA_ENABLED
Kokkos on CPU	kokkos	test-kokkos-serial	DIGI_KOKKOS, DIGI_KOKKOS_SERIAL
		test-kokkos-openmp	DIGI_KOKKOS, DIGI_KOKKOS_OPENMP
		test-kokkosview-serial	DIGI_KOKKOS, DIGI_KOKKOS_SERIAL, DIGI_KOKKOSVIEW
		test-kokkosview-openmp	DIGI_KOKKOS, DIGI_KOKKOS_OPENMP DIGI_KOKKOSVIEW
Kokkos on GPU		test-kokkos-cuda	DIGI_KOKKOS, DIGI_KOKKOS_CUDA
		test-kokkosview-cuda	DIGI_KOKKOS, DIGI_KOKKOS_CUDA DIGI_KOKKOSVIEW
Intel oneAPI	oneapi	test-oneapi	DIGI_ONEAPI
- OpenCL		test-oneapi-opencl	DIGI_ONEAPI
- CUDA		test-oneapi-cuda	DIGI_ONEAPI

The per-technology targets build all the executables of that technology. For finer-grained compilation, use the executable names directly as make targets.

Naive CPU

The only requirements for "naive CPU" are g++ supporting C++17 in the $PATH.

CUDA

The CUDA test program requires a recent CUDA version (nvcc supporting C++14 and --expt-relaxed-constexpr) and a machine with an NVIDIA GPU. By deafult, the binaries target SM 3.5, 5.0, 6.0 and 7.0. Different targets can be added in the Makefile.

Alpaka release 0.4.0

The Alpaka test program can be compiled for different backends; so far it has been tested with the CUDA, serial, and TBB backends. The CUDA backend requires CUDA 9.2 through 10.2, and has been tested with gcc 7.x and gcc 8.x.

The backend is chosen at compile time setting one of the ALPAKA_ACC preprocessor symbols. The test-alpaka binary tries to exercise all available backends.

See the instructions on the Patatrack Wiki for installing Alpaka and Cupla.

Cupla release 0.2.0

The Cupla test program can be compiled for different backends; so far it has been tested with the CUDA, serial, TBB and OpenMP backends. The CUDA backend requires CUDA 9.2 through 10.2, and has been tested with gcc 7.x and gcc 8.x.

The backend is chosen at compile time setting one of the ALPAKA_ACC preprocessor symbols. The test-cupla binary tries to exercise all available backends.

See the instructions on the Patatrack Wiki for installing Alpaka and Cupla.

Kokkos release 3.0.00

The Kokkos test programs require Kokkos' source. Run something along the following before compiling any of the test programs

# In some directory
git clone --branch 3.0.00 https://github.com/kokkos/kokkos.git
export KOKKOS_BASE=$PWD/kokkos

If CUDA is enabled (i.e. $CUDA_BASE points to an existing directory), the $CUDA_BASE/bin should be put in $PATH before compilation. In addition, all Kokkos test programs need to be run on a machine with GPU.

If Kokkos is enabled (i.e.$KOKKOS_BASE is set), everything will be compiled with nvcc_wrapper (also other targets than Kokkos test programs).

Note that Kokkos' runtime library gets built on the fly, and the libkokkos.a and the intermediate object files are placed to the working directory.

Intel oneAPI

The test program relies on the In-Order Queues and Unified Shared Memory extensions to SYCL 1.2.1, which are currently available in Intel oneAPI toolchain and in the LLVM SYCL branch.

The beta version of Intel oneAPI can be obtained from https://software.intel.com/en-us/oneapi .

The in-development version of the LLVM compiler with support for SYCL, Intel's extensions, and Codeplay's CUDA backend is available on GitHub at https://github.com/intel/llvm/ . See the instructions on the Patatrack Wiki for building the SYCL toolchain.

The test program should run on any available SYCL device, and can select it at runtime based on the command line options.

How to add a new implementation?

• Copy of (e.g.) the rawtodigi_naive.h for the new thing (with name of the thing after the underscore)
• Enclose all specifics under #ifdef DIGI_THING
• Add new build rules to Makefile
• Update README.md, including the requirements to compile and run with the thing