Note
The published documentation is available at rocThrust in an organized, easy-to-read format, with search and a table of contents. The documentation source files reside in the docs
folder of this repository. As with all ROCm projects, the documentation is open source. For more information on contributing to the documentation, see Contribute to ROCm documentation.
Thrust is a parallel algorithm library. It has been ported to HIP and ROCm, which use the rocPRIM library. The HIP-ported library works on HIP and ROCm software. Currently there is no CUDA backend in place.
Software requirements include:
- CMake (3.10.2 or later)
- AMD ROCm Software (1.8.0 or later)
- Including the HipCC compiler, which must be set as your C++ compiler for ROCm
- rocPRIM library
- This is automatically downloaded and built by the CMake script
- Python 3.6 or higher (for HIP on Windows; only required for install scripts)
- Visual Studio 2019 with Clang support (for HIP on Windows)
- Strawberry Perl (for HIP on Windows)
Optional:
- GoogleTest
- Required only for tests; building tests is enabled by default
- This is automatically downloaded and built by the CMake script
- doxygen
- Required for building the documentation
For ROCm hardware requirements, refer to:
Documentation for rocThrust available at https://rocm.docs.amd.com/projects/rocThrust/en/latest/.
You can build our documentation locally using the following commands:
# Go to rocThrust docs directory
cd rocThrust; cd docs
# Install Python dependencies
python3 -m pip install -r sphinx/requirements.txt
# Build the documentation
python3 -m sphinx -T -E -b html -d _build/doctrees -D language=en . _build/html
# For e.g. serve the HTML docs locally
cd _build/html
python3 -m http.server
git clone https://github.com/ROCm/rocThrust
# Go to rocThrust directory, create and go to the build directory.
cd rocThrust; mkdir build; cd build
# Configure rocThrust, setup options for your system.
# Build options:
# DISABLE_WERROR - ON by default, This flag disable the -Werror compiler flag
# BUILD_TEST - OFF by default,
# BUILD_HIPSTDPAR_TEST - OFF by default,
# BUILD_EXAMPLES - OFF by default,
# BUILD_BENCHMARKS - OFF by default,
# DOWNLOAD_ROCPRIM - OFF by default, when ON rocPRIM will be downloaded to the build folder,
# RNG_SEED_COUNT - 0 by default, controls non-repeatable random dataset count
# PRNG_SEEDS - 1 by default, reproducible seeds to generate random data
#
# ! IMPORTANT !
# On ROCm platform set C++ compiler to HipCC. You can do it by adding 'CXX=<path-to-hipcc>'
# before 'cmake' or setting cmake option 'CMAKE_CXX_COMPILER' with the path to the HipCC compiler.
#
[CXX=hipcc] cmake ../. # or cmake-gui ../.
# Build
make -j4
# Optionally, run tests if they're enabled.
ctest --output-on-failure
# Package
make package
# Install
[sudo] make install
We've added initial support for HIP on Windows. To install, use the provided rmake.py
Python script:
git clone https://github.com/ROCm/rocThrust.git
cd rocThrust
# the -i option will install rocPRIM to C:\hipSDK by default
python rmake.py -i
# the -c option will build all clients including unit tests
python rmake.py -c
# Performance improvement option. If you define THRUST_HIP_PRINTF_ENABLED before
# thrust includes to 0, you can disable printfs on device side and improve
# performance. The default value is 1
#define THRUST_HIP_PRINTF_ENABLED 0
We recommended including rocThrust into a CMake project by using its package configuration files.
# On ROCm rocThrust requires rocPRIM
find_package(rocprim REQUIRED CONFIG PATHS "/opt/rocm/rocprim")
# "/opt/rocm" - default install prefix
find_package(rocthrust REQUIRED CONFIG PATHS "/opt/rocm/rocthrust")
...
includes rocThrust headers and roc::rocprim_hip target
target_link_libraries(<your_target> roc::rocthrust)
# Go to rocThrust build directory
cd rocThrust; cd build
# Configure with examples flag on
CXX=hipcc cmake -DBUILD_TEST=ON ..
# Build tests
make -j4
# To run all tests
ctest
# To run unit tests for rocThrust
./test/<unit-test-name>
This feature requires CMake 3.16+ to be used for building and testing. (Prior versions of CMake can't assign IDs to tests when running in parallel. Assigning tests to distinct devices could only be done at the cost of extreme complexity.)
Unit tests can make use of the
CTest Resource Allocation feature, which enables distributing tests across multiple GPUs in an intelligent manner. This feature can
accelerate testing when multiple GPUs of the same family are in a system. It can also test multiple
product families from one invocation without having to use the HIP_VISIBLE_DEVICES
environment
variable. CTest Resource Allocation requires a resource spec file.
Using `RESOURCE_GROUPS` and `--resource-spec-file` with CMake and CTest, respectively for versions
prior to 3.16 omits the feature silently. Therefore, you must ensure that the `cmake` and `ctest` you
invoke are sufficiently recent.
There is a utility script in the repo that may be called independently:
# Go to rocThrust build directory
cd rocThrust; cd build
# Invoke directly or use CMake script mode via cmake -P
../cmake/GenerateResourceSpec.cmake
# Assuming you have 2 compatible GPUs in the system
ctest --resource-spec-file ./resources.json --parallel 2
Assuming you have two GPUs from the gfx900 family and they are the first devices enumerated by the
system, you can specify -D AMDGPU_TEST_TARGETS=gfx900
during configuration to specify that you
want only one family to be tested. If you leave this var empty (default), the default device in the system
is targeted. To specify that there are two GPUs that should be targeted, you must feed a JSON file to
CTest using the --resource-spec-file <path_to_file>
flag. For example:
{
"version": {
"major": 1,
"minor": 0
},
"local": [
{
"gfx900": [
{
"id": "0"
},
{
"id": "1"
}
]
}
]
}
There are two CMake configuration-time options that control random data fed to unit tests.
-
RNG_SEED_COUNT
: 0 by default, controls non-repeatable random dataset count.- Draws values from a default constructed
std::random_device
. - Should tests fail, the actual seed producing the failure is reported by Googletest, which allows for reproducibility.
- Draws values from a default constructed
-
PRNG_SEEDS
: 1 by default, controls repeatable dataset seeds.- This is a CMake formatted (semicolon delimited) array of 32-bit unsigned integrals. Note that
semicolons often collide with shell command parsing. We advise escaping the entire CMake CLI
argument to avoid having the variable pick up quotation marks. For example, pass
cmake "-DPRNG_SEEDS=1;2;3;4"
instead ofcmake -DPRNG_SEEDS="1;2;3;4"
(these cases differ in how the CMake executable receives arguments from the operating system).
- This is a CMake formatted (semicolon delimited) array of 32-bit unsigned integrals. Note that
semicolons often collide with shell command parsing. We advise escaping the entire CMake CLI
argument to avoid having the variable pick up quotation marks. For example, pass
# Go to rocThrust build directory
cd rocThrust; cd build
# Configure with examples flag on
CXX=hipcc cmake -DBUILD_EXAMPLES=ON ..
# Build examples
make -j4
# Run the example you want to run
# ./examples/example_thrust_<example-name>
# For example:
./examples/example_thrust_version
# Example for linking with cpp files
./examples/cpp_integration/example_thrust_linking
# Go to rocThrust build directory
cd rocThrust; cd build
# Configure with benchmarks flag on
CXX=hipcc cmake -DBUILD_BENCHMARKS=ON ..
# Build benchmarks
make -j4
# Run the benchmarks
./benchmarks/benchmark_thrust_bench
rocThrust also hosts the header files for HIPSTDPAR.
Within these headers, a great part of the C++ Standard Library parallel algorithms are overloaded so that rocThrust's and rocPRIM's implementations of those algorithms are used when they are invoked with the parallel_unsequenced_policy
policy.
When compiling with the proper flags (see LLVM (AMD's fork) docs1 for the complete list), the HIPSTDPAR headers are implicitly included by the compiler, and therefore the execution of these parallel algorithms will be offloaded to AMD devices.
HIPSTDPAR is currently packaged along rocThrust. The hipstdpar
package is set up as a virtual package provided by rocthrust
, so the latter needs to be installed entirely for getting HIPSTDPAR's headers. Conversely, installing the rocthrust
package will also include HIPSTDPAR's headers in the system.
rocThrust also includes some tests for checking the correct building of HIPSTDPAR implementations. These are located under the tests/hipstdpar folder. When configuring the project with the BUILD_TEST
option on, these tests will also be enabled. Additionally, one can configure only HIPSTDPAR's tests by disabling BUILD_TEST
and enabling BUILD_HIPSTDPAR_TEST
. In general, the following steps can be followed for building and running the tests:
git clone https://github.com/ROCm/rocThrust
# Go to rocThrust directory, create and go to the build directory.
cd rocThrust; mkdir build; cd build
# Configure rocThrust.
[CXX=hipcc] cmake ../. -D BUILD_TEST=ON # Configure rocThrust's and HIPSTDPAR's tests.
[CXX=hipcc] cmake ../. -D BUILD_TEST=OFF -D BUILD_HIPSTDPAR_TEST=ON # Only configure HIPSTDPAR's tests.
# Build
make -j4
# Run tests.
ctest --output-on-failure
- rocPRIM and rocThrust libraries
- TBB library
- Notice that oneTBB (oneAPI TBB) may fail to compile when libstdc++-9 or -10 is used, due to them using legacy TBB interfaces that are incompatible with the oneTBB ones (see the release notes).
- CMake (3.10.2 or later)
You can report bugs and feature requests through the GitHub issue tracker.
rocThrust is distributed under the Apache 2.0 LICENSE.
Footnotes
-
Altough currently only AMD's fork of LLVM contains the docs for the C++ Standard Parallelism Offload Support, both of them (the upstream LLVM and AMD's fork) do support it. ↩