This is mirror of the CLRadeonExtender project.
Original site is here https://clrx.nativeboinc.org or http://clrx.nativeboinc.org if SSL certificate doesn't work.
DONATION: If you want to donate this project just send Bitcoins to following address: 3M6z45UGXgimFs3QGQgbmqVZzK9s3RnfLL
CLRadeonExtender provides tools to develop software in low-level for the Radeon GPU's compatible with GCN 1.0/1.1/1.2/1.4 (AMD VEGA) architecture. Since version 0.1.8 also AMD VEGA Deep Learning extensions has been supported. Currently, this project have two tools to develop that software:
- clrxasm - the GCN assembler
- clrxdisasm - the GCN disassembler
Both tools can operate on four binary formats:
- the AMD Catalyst OpenCL program binaries
- the GalliumCompute (Mesa) program binaries
- the AMD Catalyst OpenCL 2.0 (new AMD format) program binaries
- the ROCm (RadeonOpenCompute) program binaries
CLRadeonExtender not only provides basic tools to low-level development, but also
allow to embed own assembler with AMD Catalyst driver through CLwrapper.
An embedded assembler can be called from clBuildProgram
OpenCL call
with specified option -xasm
. Refer to README and INSTALL to learn about CLRXWrapper.
CLRadeonExtender requires:
- C++11 compliant compiler (Clang++ or GCC 4.7 or later, MSVC 2015 or later)
- any build system supported by CMake (GNU make, NMake, Visual Studio, ...)
- CMake system (2.8 or later)
- Threads support (for Linux, recommended NPTL)
- Unix-like (Linux or BSD) system or Windows system
Optionally, CLRXWrapper requires:
- libOpenCL.so or OpenCL.dll
- OpenCL headers
- OpenGL headers (to 0.1.5 version)
- OpenCL ICD (for example from AMD Catalyst driver)
- AMD Catalyst driver or AMDGPU-PRO driver.
and documentation requires:
- pod2man utility for Unix manuals
- markdown_py utility for CLRX Documentation
- Doxygen for CLRX API Documentation
To build system you should create a build directory in source code package:
mkdir build
and run:
cmake .. [cmake options]
Optional CMake configuration options for build:
- CMAKE_BUILD_TYPE - type of build (Release, Debug, GCCSan, GCCSSP).
- CMAKE_INSTALL_PREFIX - prefix for installation (for example '/usr/local')
- BUILD_32BIT - build 32-bit binaries (works only in the Unix/Linux 64-bit environment)
- BUILD_TESTS - build all tests
- BUILD_SAMPLES - build OpenCL samples
- BUILD_DOCUMENTATION - build project documentation (doxygen, unix manuals, user doc)
- BUILD_DOXYGEN - build doxygen documentation
- BUILD_MANUAL - build Unix manual pages
- BUILD_CLRXDOC - build CLRX user documentation
- BUILD_STATIC_EXE - build with statically linked executables
- GCC5CXX11NEWABI - build with new GCC5 C++11 ABI (required if GCC uses old ABI by default)
- NO_STATIC - no static libraries
- NO_CLWRAPPER - do not build CLRXWrapper
- CPU_ARCH - target CPU architecture (in GCC parameter to -march, for MSVC parameter to /arch:)
- OPENCL_DIST_DIR - an OpenCL directory distribution installation (optional)
You can just add one or many of these options to cmake command:
cmake .. -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTS=ON
or for Microsoft Visual C++ (for NMake):
cmake .. -G "NMake Makefiles" [cmake configuration options]
or for Microsoft Visual C++ (for same Visual Studio)
cmake .. -G "Visual Studio XXXXXX [arch]" [cmake configuration options]
where XXXX - version of Visual Studio (7, 14 2015, ...). arch - architecture (Win64, ARM) (optional).
After creating Makefiles scripts you can compile project:
make
or make -jX
- where X is number of processors.
or (for NMake)
nmake
or just execute build option in Visual Studio.
After building you can check whether project is working (if you will build tests):
ctest
Creating documentation will be done by this command (if you will enable a building documentation, required for version 0.1):
make Docs
or (for NMake)
nmake Docs
Due to unknown reasons, the compilation under clang++ will be failed. We recommend to use
gcc compiler to build the CLRadeonExtender. You should prepend cmake command by CXX=g++
:
CXX=g++ cmake .. ....
.
Installation is easy. Just run command:
make install
or (for NMake):
nmake install
The default (without '-gcc5' in name) binary libraries for Linux are compiled for C++11 old ABI, hence you must add option -D_GLIBCXX_USE_CXX11_ABI=0 to compiler commands if you are using GCC 5.0 or higher or compiler that by default uses new C++11 ABI.
Usage of the clrxasm is easy:
clrxasm [-o outputFile] [options] [file ...]
If no file specified clrxasm read source from standard input.
Useful options:
- -g DEVICETYPE - device type ('pitcairn', 'bonaire'...)
- -A ARCH - architecture ('gcn1.0', 'gcn1.1', 'gcn1.2' or 'gcn1.4')
- -b BINFMT - binary format ('amd', 'amdcl2', 'gallium', 'rocm', 'rawcode')
- -t VERSION - driver version for which a binary will be generated
- -w - suppress warnings
Usage of the clrxdisasm:
clrxdisasm [options] [file ...]
and clrxdisasm will print a disassembled code to standard output.
Useful options for clrxdisasm:
- -a - print everything (not only code, but also kernels and their metadatas)
- -f - print floating points
- -h - print hexadecimal instruction codes
- -C - print configuration dump instead metadatas, CALnotes and setup data
- -g DEVICETYPE - device type ('pitcairn', 'bonaire'...)
- -A ARCH - architecture ('gcn1.0', 'gcn1.1', 'gcn1.2' or 'gcn1.4')
- -t VERSION - driver version for which a binary will be generated
A CLRX assembler accepts source from disassembler.