This code base contains the code to run OpenCL programs on Intel GPUs. This is basically the run-time code i.e. it defines the OpenCL host functions required to initialize the device, create the command queues, the kernels and the programs and run them on the GPU. The run-time does not contain the compiler. The OpenCL compiler has its own shared object and both the run-time and the compiler are interfaced with a regular C layer.
The project uses CMake with three profiles:
Basically, from the root directory of the project
> mkdir build
> ccmake ../ # to configure
Choose whatever you want for the build.
Then press 'c' to configure and 'g' to generate the code.
> make
The project depends on several external libraries:
CMake will check the dependencies and will complain if it does not find them.
Once built, the run-time produces a shared object libcl.so which basically directly implements the OpenCL API. A set of tests are also produced. They may be found in utests.
Apart from the OpenCL library itself that can be used by any OpenCL application, this code also produces various tests to ensure the compiler and the run-time consistency. This small test framework uses a simple c++ registration system to register all the unit tests.
You need to set the variable OCL_KERNEL_PATH to locate the OCL kernels. They
are with the run-time in ./kernels.
Then in utests/:
> ./run
will run all the unit tests one after the others
> ./run some_unit_test0 some_unit_test1
will only run some_unit_test0 and some_unit_test1 tests
As an important remark, the code was only tested on IVB GT2 with a rather minimal Linux distribution (ArchLinux) and a very small desktop (dwm). If you use something more sophisticated using compiz or similar stuffs, you may expect serious problems and GPU hangs.
The run-time is far from being complete. Most of the pieces have been put together to test and develop the OpenCL compiler. A partial list of things to do:
Support for samplers / textures but it should be rather easy since the low-level parts of the code already supports it
Support for events
Check that NDRangeKernels can be pushed into different queues from several threads
Support for Enqueue*Buffer. I added a straightforward extension to map /
unmap buffer. This extension clIntelMapBuffer directly maps dri_bo_map
which is really convenient
Full support for images. Today, the code just tiles everything manually which is really bad. I think the best solution to copy and create images is to use the GPU and typed writes (scatter to textures) or samplers. We would however need the vmap extension proposed by Chris Wilson to be able to map user pointers while doing to copies and the conversions.
No state tracking at all. One batch buffer is created at each "draw call" (i.e. for each NDRangeKernels). This is really inefficient since some expensive pipe controls are issued for each batch buffer
Valgrind reports some leaks in libdrm. It sounds like a false positive but it has to be checked. Idem for LLVM. There is one leak here to check
More generally, everything in the run-time that triggers the "FATAL" macro means that something that must be supported is not implemented properly (either it does not comply with the standard or it is just missing)
The code base supports a seamless integration with Fulsim i.e. you do not need to run anything else than your application to make Fulsim work with it. However, some specific step have to be completed first to make it work.
Compilation phase. You need to compile the project with fulsim enabled. You
should choose EMULATE_IVB ON in ccmake options. Actually, Haswell has not been
tested that much recently so there is a large probability it will not work
properly
Fulsim executables and DLL. Copy and paste fulsim Windows executables and DLLs into the directory where you run your code. The run-time will simply call AubLoad.exe to run Fulsim. You can get fulsim from our subversion server. We compile versions of it. They are all located in here
Run-time phase. You need to fake the machine you want to simulate. Small scripts in the root directory of the project are responsible for doing that:
> source setup_fulsim_ivb.sh 1
will run fulsim in debug mode i.e. you will be able to step into the EU code
> source setup_fulsim_ivb.sh 0
will simply run fulsim
The compiler is able to produce c++ file that simulate the behavior of the kernel. The idea is mostly to be able to gather statistics about how the kernel can run (SIMD occupancy, bank conflicts in shared local memory or cache hit/miss rates). Basically, the compiler generates a c++ file from the LLVM file (with some extra steps detailed in the OpenCL compiler documentation). Then, GCC (or ICC) is directly called to generate a shared object.
The run-time is actually able to run the simulation code directly. To enable it (and to also enable the c++ path in the compile code), a small script in the root directory has to be run:
> source setup_perfim_ivb.sh
Doing that, the complete C++ simulation path is enabled.
Right now, a modified libdrm is required to run fulsim. It completely disables the HW path (nothing will run on the HW at all) and allows to selectively dump any OpenCL buffer. Contact Ben Segovia to get the access to it.
Ben Segovia (benjamin.segovia@intel.com)