Re: Adding Opencl On Dev C
Achmat Das
I've started Xcode and created an empty project; created a main.c and a .cl kernel file, pasting what is on the Apple developer site, and I am not able to get anything to build, even after adding a target.
If you are using Visual Studio then you would need to add the include and lib folders to your project as well as adding OpenCL.lib in order for everything to compile correctly. Also you would need to use the #include"CL/cl.h" command in your code.
i. first i need to build an OpenCL ICD loader which will load shared library (libOpenCL.so) whose name is on first line of .icd file. Is that enough to find new hardware device in OpenCL?
ii. secondly, i have to modify CL/cl.h file by adding a new device type say for example XYX such as
#define CL_DEVICE_TYPE_XYZ (1
I too was confounded by this perplexing problem until I found a series of forum posts bout Virtual Lighttable in which the participants do some debugging. It turns out that the NVIDIA drivers need a custom modprobe rule that is not installed by default with either nvidia-331 or the the opencl drivers. These rules are provided in the nvidia-modprobe package.
If we follow the post, we can see that we need:
- LLVM with '--enable-experimental-targets=R600', but the stable repo already provides this so we are fine.
- Libclc can be found in AUR: libclc-svn
- Mesa with '--enable-opencl'. This has to be built manually with ABS adding the flag to the PKGBUILD.
For the first case you could rebuild imagemagick and other programs using libcl through abs with --disable-opencl and remove the dependency on libcl / opencl-headers .
Then you can rebuild the specific programs you want to test with libclc against that.
You can save yourself some work by using Lordheavhy's unofficial mesa-git repo at lcarlier/mesa-git/ .
It builds mesa with opencl enabled, and llvm/clang/ libclc dev versions are also in that repo.
But the libOpenCL.so* came from libcl, and I had the conflict when I tried to install my compiled version of mesa. So I think that maybe "my mesa" tries to install the headers and the libOpenCL.so* files, which in reality came from opencl-headers and libcl.
All I did was adding OpenCL to pkgconfig. I never used OpenCL on ubuntu. how did you install it? Is it included in the repositories? Does pkg-config --libs OpenCL work? I suspect that your system is not correctly configured to find all the paths to your OpenCL installation.
In summary, Cycles could have support added for some older GPUs (like the RX 570). However, adding support is either out of control of the Blender foundation, a low priority, or would introduce a lot of maintenance cost for the Cycles code base depending on the approach taken. And as such, support for these older GPUs may not be re-implemented.
Then I have started the application with TI_OCL_DEBUG=ccs as it is suggested in the webside. It dumps some information as in the website and waits for the key. I have copied the opencl*****.out file to the development environment from the target, I have loaded the symbols from it added the symbols from dsp.out. When the time comes to adding a breakpoint as it is suggested by the application ( some OpenMP generated function is suggested. ) CCS could not enable it. CCS adds it to the list of breakpoints, it is still gray, even if I tried to enable it. Furthermore I could not add any breakpoints to the offloaded code fragments in the CCS.
As you see no debugging symbols added, adding breakpoint is failed. In your reply, you have requested for the logs. Which logs do you want? I appreciate if you can give some information regarding the logs I should collect.
I don't know -g parameter should be in the --tc and --hc compiler options or not, with the makefile above, debug symbols added to opencl******.out files. (I have checked with nm6x command on the target.)
There are announcements from Intel about adding support for their GPUs: OneAPI/L0, OpenVINO and OpenCL coming to the Windows Subsystem for Linux for Intel GPUs - Windows Command Line, but I could not find anything about the AMD support
This was a new/clean install of Ubu18. It is quite possible that I did fail to notice that "amdgpu-install -y" was installed. But i do know that "amdgpu-pro-install -y --opencl=legacy" was an absolute requirement for me to get Ubu 18 to recognize my GPU
AMDGPU is already installed in the kernel so no need to modify it (actually if you install the one in amdgpu-pro you end up with a different one, that is exactly the point of this thread: how to install OpenCL without modifying the already installed open source AMDGPU wich we want to keep) but as you rightly said, OpenCL is not and must be installed manually. So you only need to run "amdgpu-pro-install -y --opencl=legacy" to install OpenCL and you are fine.
The host program controls the execution of kernels on the compute devices. The host program is written in C, but bindings for other languages like C++ and Python exists. The OpenCL API is defined in the cl.h (or opencl.h for apple) header file. Below is the code for the host program that executes the kernel above on compute device. I will not go into details on each step as this is supposed to be an introductory article. The main steps of a host program is as follows:
i am using gpu of type nvidia,i am using opencl but when i run the program using ctrl+F5(start without debugging )then i get result in which gpu takes more time than cpu but when i run the program cpu takes more time than gpu and result is also i am giving
start without debugging -> cpu time=6127 ms gpu time= 6240 ms
start with debug-> cpu time= 18354 ms gpu time= 9125 ms
So far, i have installed the latest Pro Radeon drivers (version 21.10) using./amdgpu-pro-install -y --opencl=pal,legacy, i have installed OpenCL headers using sudo apt-get install opencl-headers and i have installed ocl-icd-opencl-dev as well as cl-info and glmark.
Further information that may be useful: I have also installed pocl-opencl-icd which produced a working OpenCL environment and the Intel drivers for my CPU, which doesn't work at the moment (but might get fixed by fixing the AMD platform?)
The arguments for --opencl listed in the AMD installation instructions do not apply (it says for hardware newer than Vega 10, i.e. RX Vega 56 and RX Vega 64 and Vega Frontier Edition, --opencl=pal should be used).
So, after installing AMD FirePro and NVIDIA GeForce to a single motherboard(with size ATX or bigger with at least two PCIe3.0), one can install AMD driver and NVIDIA driver to a single OS(for example, Windows 10) and use OpenCL ICD(maybe one needs to install AMD OpenCL SDK and NVIDIA OpenCL SDK prior to compile OpenCL code?) by adding "#pragma OPENCL EXTENSION cl_khr_icd : enable" below the line "#include (or corresponding header file)"?
Not getting a work unit will be due to the number of people folding exceeding the limitations of the current system (They are working on adding more capacity). I don't know anything about linux but I have gotten some bad work units in the past. will start a wu sometimes manage to get a few % in and then stops and goes back to downloading. prob best to let it run for a while and see if the next work unit dose the same thing or if it folds correctly.
For this post we are going to focus on turning the sub problem of adding a list of numbers together in a parallel way. It is much like the previous problem, however we only care about the final sum, not the intermediate sums. You may be asking why I am explaining such a simple task, however as it turns out it is quite complex to do utilising parallel computation and we can learn a lot. When we all first started programming, we dealt with simple problems and learnt much from them, we are doing the same for parallel computing.
For the astute reader will notice that this isn't particularly useful and we would need to get the result of all the numbers added together. So the next step would either to be keep going in this adding process, or to output the immediate steps and then compute the sum somewhere else. As we're outputting the immediate results to local memory, you want to have each kernel wait until the other kernels are done, and then continue to add the numbers together.
I want to quickly introduce the idea of Single Instruction Multiple Data (SIMD) here as it plays an important part in creating a mental model of how data is mapped into memory and then operated on. SIMD is a CPU (GPUs have it as well) instruction set which allows you to apply the same operation on multiple pieces of data. It does this by putting a chunk of data into a large register (128b, 256b, 512b, etc) and then operating on it. In source code this often looks like applying operations (add, multiple, swap, shift, etc) on fixed sized arrays. The following code is a simplified version of adding two sets of numbers together.
The import part to realise here is that the adding operation will happen on all pairs of data at once, and won't happen separately. This quick aside is meant to quickly demonstrate that the time complexity of programs is often very different to the Big O algorithmic complexity and we need to pay special attention to how we do this in OpenCL as things are not as they seem.
Alternatively the internal header opencl-c.h containing the declarationscan be included manually using -include or -I followed by the pathto the header location. The header can be found in the clang source tree orinstallation directory.
aa06259810