Intel® C for Metal Development Package
The Intel® C for Metal Development Package is a software development package for Intel® Graphics Technology.
The Intel C for Metal Development Package is distributed under the MIT license. You may obtain a copy of the License at: https://opensource.org/licenses/MIT
Ubuntu* 16.04 with Linux* kernel 4.10
For kernel versions other than 4.10, the full functionality can't be guaranteed.
Dependencies (Intel Components)
Intel® Media Driver for VAAPI Intel® C for Metal Runtime (included in Media Driver) Intel® C for Metal Compiler Intel® Graphics Compiler
Install the Intel C for Metal Development Package
Folder structure after installation
<package_root> ├── compiler │ ├── bin │ └── include_llvm ├── documents │ ├── cmrtlib │ └── compiler ├── drivers │ ├── media_driver │ │ ├── release │ │ └── release-internal │ └── IGC ├── examples ├── readme.txt └── runtime ├── include └── lib
Details of folder structure
compiler: the compiler binaries and libraries used to build CM language source codes to GPU kernels running on Render engine
compiler/bin: The front-end and offline back-end compiler binaries to build GPU kernels
compiler/include_llvm: The header files needed during compiling GPU kernels
runtime: the libraries and headers defines a set of CM APIs used by CPU applications, through which GPU kernels can be loaded onto render engine and be executed
runtime/include: The header files defining the CM APIs
runtime/lib: The libraries implementing the CM APIs
documents/compiler: The documents about the CM compiler usage and the CM language to write GPU kernels
documents/runtime: The documents about the CM APIs to load and execute GPU kernels
drivers: the media driver and IGC as low-level supported drivers to load and run GPU kernels. Release version means stripped driver binaries, while release-internal means driver binaries with symbols and other debugging functionalities, like results dumping.
examples: Each folder under examples is an example, including host source code xxx.cpp and kernel source code xxx_genx.cpp.
Install Both Intel Media Driver for VAAPI and Intel Graphics Compiler by Script
cd drivers sudo ./install.sh
Install the Intel Media Driver for VAAPI Alone
Choose release or release-internal driver package
dpkg -i intel-media-2.xxx.deb
Note the following:
This command needs to be run in privileged mode
If it has this error “trying to overwrite xxx, which is also in package xxx”, it means you may have installed part of the libs before. Please add ‘--force-overwrite’ option in the installing command
The libva and libdrm libraries and headers, needed by building and executing CM applications, will also be installed together with media driver.
The media driver is now a open source project in https://github.com/intel/media-driver. You can also build from source to get the media driver libraries.
Install Intel Graphics Compiler Alone
Get the IGC package from drivers/IGC/
dpkg -i IGC-xxx.deb
Note that this command needs to be run in privileged mode
Build and Execute
Layout of an Intel C for Metal application
Usually an Intel C for Metal application contains two parts: host code running on CPU and kernel code running on GPU. Take /examples/linear_walker as an example:
linear_walker.cpp is the host application source code. It is standard C++ source code calling the Intel C for Metal APIs to load the GPU kernels. It is compiled by g++ and executed on CPU.
linear_walker_genx.cpp is the GPU kernel source code. It is written using the Intel C for Metal language. It is compiled by the Intel C for Metal compiler, cmc in compiler/bin, and is loaded by the host application and executed on GPU.
Build an example using provided make file
Change directory to any subfolder in examples, e.g. examples/linear_walker, then run the make command:
make -f ../Makefile.linux
Two binaries will be generated:
hw_x64.linear_walker: the executable on CPU, generated from linear_walker.cpp
linear_walker_genx.isa: the GPU kernel loaded by the CPU executable and executed on render engine on GPU, generated from linear_walker_genx.cpp
Run the host application code
Make sure both the host application binary and the GPU kernel binary are generated, and they are in the same directory.
Run the host application binary directly.
Build an example using command lines (advanced)
Use two command lines below to build host code
g++ linear_walker.cpp -w -g -std=gnu++0x -c -rdynamic -fPIC -O0 -msse4.1 -ffloat-store -I"<package_root>/runtime/include" -o linear_walker.o g++ linear_walker.o -fPIC <package_root>/runtime/lib/x64/igfxcmrt64.so -o linear_walker -lva -ldl -lva-drm -lpthread -m64
Use one command line below to build kernel code
<package_root>/compiler/bin/cmc linear_walker_genx.cpp -isystem <package_root>/compiler/include
Build and run all the examples in the package
A shell script is provided to build and run all the examples in the examples folder:
Contribu To This Package
The source code of all the components in the package are maintained on github:
Intel Media Driver for VAAPI and Intel C for Metal Runtime: https://github.com/intel/media-driver
Intel C for Metal compiler and examples: https://github.com/intel/cm-compiler
Intel Graphics Compiler: https://github.com/intel/intel-graphics-compiler
We welcome your contributions including bug fixes, optimizations, and new features. You may create a new pull request on github for your changes. We will review it, test it, and give you our feedback. When the change is finalized, we will push it to the project repository.
For any new example using CM, as long as you have the entire application that can be built and run on either Linux or Windows, you can email it to the maintainer of Intel C for Metal compiler. We can help you customize it, and publish it in our example folder.
Intel® microarchitecture code name Broadwell, Intel® microarchitecture code name Skylake, Apollo Lake, Kaby Lake, Coffee Lake, and Cannon Lake.
Known Issues and Limitations
(*) Other names and brands may be claimed as property of others.