The rocRAND project provides functions that generate pseudo-random and quasi-random numbers.
The rocRAND library is implemented in the HIP programming language and optimised for AMD's latest discrete GPUs. It is designed to run on top of AMD's Radeon Open Compute ROCm runtime, but it also works on CUDA enabled GPUs.
Additionally, the project includes a wrapper library called hipRAND which allows user to easily port CUDA applications that use cuRAND library to the HIP layer. In ROCm environment hipRAND uses rocRAND, however in CUDA environment cuRAND is used instead.
- XORWOW
- MRG32k3a
- Mersenne Twister for Graphic Processors (MTGP32)
- Philox (4x32, 10 rounds)
- Sobol32
- Git
- cmake (3.0.2 or later)
- C++ compiler with C++11 support
- For AMD platforms:
- For CUDA platforms:
- HIP (hcc is not required)
- Latest CUDA SDK
Optional:
- GTest (required only for tests; building tests is enabled by default)
- Use
GTEST_ROOT
to specify GTest location (also see FindGTest) - Note: If GTest is not already installed, it will be automatically downloaded and built
- Use
- Fortran compiler (required only for Fortran wrapper)
gfortran
is recommended.
- Python 2.7+ or 3.5+ (required only for Python wrapper)
If some dependencies are missing, cmake script automatically downloads, builds and
installs them. Setting DEPENDENCIES_FORCE_DOWNLOAD
option ON
forces script to
not to use system-installed libraries, and to download all dependencies.
git clone https://github.com/ROCmSoftwarePlatform/rocRAND.git
# Go to rocRAND directory, create and go to build directory
cd rocRAND; mkdir build; cd build
# Configure rocRAND, setup options for your system
# Build options: BUILD_TEST (off by default), BUILD_BENCHMARK (off by default), BUILD_SHARED_LIBS (on by default)
#
# ! IMPORTANT !
# Set C++ compiler to HCC or HIP-clang. You can do it by adding 'CXX=<path-to-compiler>'
# before 'cmake' or setting cmake option 'CMAKE_CXX_COMPILER' to path to the compiler.
#
# The python interface do not work with static library.
#
[CXX=hipcc] cmake -DBUILD_BENCHMARK=ON ../. # or cmake-gui ../.
# To configure rocRAND for Nvidia platforms, 'CXX=<path-to-nvcc>', `CXX=nvcc` or omitting the flag
# entirely before 'cmake' is sufficient
[CXX=nvcc] cmake -DBUILD_BENCHMARK=ON ../. # or cmake-gui ../.
# or
cmake -DBUILD_BENCHMARK=ON ../. # or cmake-gui ../.
# Build
# For ROCM-1.6, if a HCC runtime error is caught, consider setting
# HCC_AMDGPU_TARGET=<arch> in front of make as a workaround
make -j4
# Optionally, run tests if they're enabled
ctest --output-on-failure
# Install
[sudo] make install
Note: Existing gtest library in the system (especially static gtest libraries built with other compilers)
may cause build failure; if errors are encountered with existing gtest library or other dependencies,
DEPENDENCIES_FORCE_DOWNLOAD
flag can be passed to cmake, as mentioned before, to help solve the problem.
Note: To disable inline assembly optimisations in rocRAND (for both the host library and
the device functions provided in rocrand_kernel.h
) set cmake option ENABLE_INLINE_ASM
to OFF
.
# Go to rocRAND build directory
cd rocRAND; cd build
# To run all tests
ctest
# To run unit tests
./test/<unit-test-name>
# Go to rocRAND build directory
cd rocRAND; cd build
# To run benchmark for generate functions:
# engine -> all, xorwow, mrg32k3a, mtgp32, philox, sobol32
# distribution -> all, uniform-uint, uniform-float, uniform-double, normal-float, normal-double,
# log-normal-float, log-normal-double, poisson
# Further option can be found using --help
./benchmark/benchmark_rocrand_generate --engine <engine> --dis <distribution>
# To run benchmark for device kernel functions:
# engine -> all, xorwow, mrg32k3a, mtgp32, philox, sobol32
# distribution -> all, uniform-uint, uniform-float, uniform-double, normal-float, normal-double,
# log-normal-float, log-normal-double, poisson, discrete-poisson, discrete-custom
# further option can be found using --help
./benchmark/benchmark_rocrand_kernel --engine <engine> --dis <distribution>
# To compare against cuRAND (cuRAND must be supported):
./benchmark/benchmark_curand_generate --engine <engine> --dis <distribution>
./benchmark/benchmark_curand_kernel --engine <engine> --dis <distribution>
# Go to rocRAND build directory
cd rocRAND; cd build
# To run Pearson Chi-squared and Anderson-Darling tests, which verify
# that distribution of random number agrees with the requested distribution:
# engine -> all, xorwow, mrg32k3a, mtgp32, philox, sobol32
# distribution -> all, uniform-float, uniform-double, normal-float, normal-double,
# log-normal-float, log-normal-double, poisson
./test/stat_test_rocrand_generate --engine <engine> --dis <distribution>
The latest rocRAND documentation and API description can be found here.
It can also be build using the following commands
# go to rocRAND doc directory
cd rocRAND; cd doc
# run doxygen
doxygen Doxyfile
# open html/index.html
- C++ wrappers for host API of rocRAND and hipRAND are in files
rocrand.hpp
andhiprand.hpp
. - Fortran wrappers.
- Python wrappers: rocRAND and hipRAND.
Bugs and feature requests can be reported through the issue tracker.
Contributions of any kind are most welcome! More details are found at CONTRIBUTING and LICENSE. Please note that statistical tests link to TestU01 library distributed under GNU General Public License (GPL) version 3, thus GPL version 3 license applies to that part of the project.