Thursday 6 November 2014

Developing OpenCL with Code::Blocks on the Parallella

Part 1: Compiling and Linking OpenCL with Code::Blocks


In this post I'll go through the steps you need to set up Code::Blocks IDE to compile and link both host code and accelerator (OpenCL) code using GNU GCC and the Brown Deer offline compiler, clcc and linker clld.

While Brown Deer provide an OpenCL JIT compiler, using it during development can be tedious given that you can only tell if your OpenCL code is compiling correctly by running your application. Using the offline compiler from within Code::Blocks allows you to stay in one environment for both host and accelerator code.

I've been learning Code::Blocks because I was finding Eclipse tediously slow. This might improve in future but for now Code::Blocks performs better on the Parallella and critically, the compile time is as good as you would get if you were using the command line. I found that Eclipse was a little better organised but once I'd figured out my way around Code::Blocks it was fine. Most of the oddities are in the compiler and project setup and you don't tend to spend much time there once these are set.

I have to acknowledge the contribution of AndyC firstly for bringing my attention to Code::Blocks and then for helping me along the way when I got stuck. His original post is here: http://forums.parallella.org/viewtopic.php?f=13&t=1658.

I'm using the Ubuntu 14.04 distro with Brown Deer OpenCL ver. 1.6.0 and Code::Blocks 13.12.






Prerequisites


1. Download Code::Blocks using: sudo apt-get install codeblocks

2. Rebuild the Brown Deer compiler! Yes I was surprised too. The installation in the standard image has a problem with the linker. The getting started guide is here: http://www.browndeertechnology.com/docs/Parallella_quick_start_guide.pdf. It was written for version 1.5 but section 2.3 is relevant to version 1.6 as well.

Get the sources: git clone https://github.com/browndeer/coprthr.git and pay careful attention to the prerequisites, libconfig, libelf and libevent (GCC will be installed already). When I ran configure it could not find event.h. If this happens to you, first check if you can find the library with whereis:

linaro-nano:~> whereis libevent
libevent: /usr/lib/libevent.so /usr/lib/libevent.la /usr/lib/libevent.a

This shows that libevent was installed but I had to look into configure to discover that it was expecting event.h to be in .../include/event2. The ... turned out to be /usr so I had to use the --with-libevent=/usr. So my final configure command was:

./configure --enable-epiphany --with-libevent=/usr

As with anything on Linux, if it does not work first time you have to sniff around a bit.

The make and sudo make install worked fine for me.


Code::Blocks


Code::Blocks has various ways of organising your work. In this post I'm going to concentrate on the setup you would use if you are primarily writing in C++ and OpenCL. Therefore I'm going to set the global settings to be those use for OpenCL development. These will be inherited by all subsequent projects. 

I'm also developing a command line project. Other types of project may need variations on these settings.

Setting up the Compilers


The first thing to do is to set up the compilers that you are going to use. You can do this before creating a project.

The object of this section is to set up a C++ compiler and an OpenCL compiler. If you only have one file with your accelerator code and want to keep your life a little simpler, read the sections on setting up GNU GCC and the basic build targets (Debug and Release) and use the method described in the last Project Setup section entitled Aside: Single File Compilation.

GNU GCC


The "standard" compiler on the Parallella is the GNU GCC compiler and we'll use this for the host code.

To get to the compiler settings go to the Settings menu and choose Compiler... which will bring up this box:


GNU GCC is the default selected compiler and there are some standard flags set. These can stay where they are.

Go to the Linker settings tab and add the Brown Deer libraries that you will be using. Here I've only added stdcl and ocl, but for example, if you are using COPRTHR then add coprthr.




Next go to the Search directories tab and add:



and similarly add /usr/local/browndeer/lib to the Linker sub-tab.

The Toolchain executables and Custom variables tabs can be alone. In the next tab along, Build options I set as following but these are not terribly important:



[Note: if you only want to compile one file with accelerator code in it, now is the time to skip forward to the section on Creating a Project and then the section titled: Aside: Single File Compilation.]

Brown Deer OpenCL


Next we have to set up the tool chain to compile the cl code. This is a little more complicated because the offline compiler has to be invoked with root privileges. We'll get to that at the very end.

First, create a copy of the GNU GCC compiler settings using the Copy button in the top section of the Global compiler settings box. This will first ask you for the name of the new compiler which I have called BD OpenCL:



This will create an exact copy of everything that we had for the GNU GCC compiler. From here on, make sure you select the BD OpenCL compiler otherwise you can get into a real mess.

First, switch off the standard flags:


Next, remove the settings from the Linker settings and Search directory settings that we set for GNU GCC. These are not needed.

The next tab to look at is the Toolchain executables. This is critical. Ignoring the Debugger entry for now it must look like this:


The next tab, Custom variables is a way of setting "environment like" variables from within Code::Blocks. I found that I had to be explicit with all environment variables because my Linux user environment got a little mangled on the way in. I defined the following three:


These are merely local copies of your bash environment variables $PATH, $LD_LIBRARY_PATH and $EPIPHANY_HDF.

This next bit brings it all together.

Code::Blocks puts all of the settings together using a template and then passes the resulting string to a shell process. The template is found at the very end of the compiler settings - on the Other settings tab. Click the Advanced options button and don't be put off by the scary warning box:


Once you have pressed the Yes button with confidence, you get:


See, not that scary after all.

The template is the text in the Command line macro box. Because it is chopped off the whole compiler string is:

sudo EPIPHANY_HDF=$EPIPHANY_HDF LD_LIBRARY_PATH=$LD_LIBRARY_PATH PATH=$PATH bash -c "$compiler $options $includes $file -o $object"

Take note of the Command macros: list on the right hand side of the box. If the Compiler complains of something being missing (e.g. if you are want to use a resource compiler) you will have to find the macro from the list and include it in the appropriate place in the Command line macro.

Also, a quick word about sudo. Sudo is a funny little thing that gives you super user powers and little else. If you type sudo env at the command prompt you will see that sudo's environment is very sparse. The -E switch is supposed to preserve the callers environment but I found that the environment that Code::Blocks stripped out LD_LIBRARY_PATH and even by explicitly setting it I found that -E did not seem to pass it on. Therefore, the most reliable way of creating the correct environment for clcc was to set it explicitly at invocation.

Similarly, the linker template screen is:


The full string is:

sudo EPIPHANY_HDF=$EPIPHANY_HDF LD_LIBRARY_PATH=$LD_LIBRARY_PATH PATH=$PATH:/bin bash -c "$linker $link_options $libdirs $link_objects $libs  -o $exe_output $object"

Note that the Command: I've selected is Link object files to console executable. There are a number of different options here. You need to figure out which one is invoked based on your build target (see below).

Project Setup


We have not quite gotten to the stage of being able to compile an application yet. There are some project specific steps that are needed first. Unfortunately they are a little circular and difficult to lay out in a nice linear fashion so you might have to go through this process a couple of times before everything is in place.

Creating a Project


The File | New > Project... will do the job without any tricks. I've only used Console Applications so far and I generally assume that I'm writing in C++. You can leave the default compiler as GNU GCC.


Adding Files

At this point is is a good idea to add some files.


Adding a C++ file for the host code is also straight forward. Use the File | New > File... menu. When you get to the screen below, just add the file name with path and select the C++ build targets (more on this below). 



Add a CL file is essentially the same. I just choose an Empty C/C++ file and give it a name ending in .cl. Leave the build targets unchecked.



Once added it will appear in the Others folder in your project.

Build Targets


I think of a build target as an output that you want to produce in your project. For the purposes of this post, I want to produce a Linux console application containing the host code and an elf (executable loadable file) containing the accelerator code.

You have already created at least one build target when you created your project. The default targets are called Debug and Release and are created as a result of this step in the new project sequence:




To add, change or remove a build target you can either choose the Project | Properties menu or right click on the project and choose Properties. Either way you will get this screen:




The only changes I've made here are the Platforms (remove Windows and Mac) and Object names generation because I like to call my main cpp file and cl file the same name.

The default build targets will be pretty much created correctly. The only little wrinkle is that the project creation dialog creates a "Hello World!" called main.cpp and this is set as the build target file. As I said above, I like to call my main file after my project so I have to create a new cpp file in the project and assign to the Debug and Release build target and then delete main.cpp.

[This next section is only relevant if you have set up the Brown Deer OpenCL compiler as a separate build step. If you are going to use the single .cl file option then skip forward to the section titled: Aside: Single File Compilation.]

To set up the build target for your cl files click on the Build targets tab and then click the Add button. In my example below, I've deleted the default Release build target and called my new target CL. Fill out the fields as shown here and select you cl file(s) in the Build target files box at the bottom:



Now that you have a build target for your accelerator code, you can select that when you create a new cl source file.

Project Build Options


Notice the Build options... button. This is where you choose which compiler is used for the selected build target and the default compiler switches for the project. Click Build options... or select Project | Build options from the main screen.



Two things to change here. With the CL build target selected, first set the Selected compiler to the OpenCL compiler we set up above, in this case we called it BD OpenCL. Second, set the Policy to be "Use target options only". The reason for this is that common compiler options (e.g. -Wall) are set globally and this setting overrides these global settings. To change the global settings click on the project name (in this case "ndfork") in the left hand panel.

This dialog is where you set all of your project specific settings. If there are include files, libraries, compiler variables that are only relevant to this project and not to all projects then dig around here and find where to set them. The Pre/post build steps tab can be useful if you want to do some behind-the-scenes shuffling before or after compilation. The output will appear on the compile log.

Final Step (the bit I always forget...)


The last thing you need to do is to set the build target and build steps for your cl file(s). In the project panel right-click on the cl file and choose Properties and then click on the Build  tab:





Select both Compile file and Link file and CL in the Belongs in targets box. For the cl file that you added before setting up the CL build target none of these boxes will be checked. Once you have added the CL build target you can select it as the build target when you add the file. The Compile file and Link file check boxes are not selected by default and so, if you wish to compile and link them, you must check these for every file you add.


Aside: Single File Compilation


By now you are probably quite astounded about the amount of complexity there is getting the compilers and project set up. I was. If you are only going to have one cl file for your accelerator code and want to avoid some of that complexity then there is a way to compile your accelerator code without setting up a separate compiler and build targets. It is "back door" way of compiling that by-passes the selected compiler.

If you right-click on the cl file, click Properties and select the Advanced tab you can define a Custom build. With the GNU GCC Compiler selected check the Use custom command to build this file: box. Then in the text box below it you need to put the whole compile string:



Note that the compiler has a full path and the Options (-c and -v) are now explicit. This is because we are now using the settings for the GNU GCC compiler and are in its environment. Therefore the $options variable refer the the GNU GCC options. The $file and $output will still work because they refer to this file. 

I've left in the $EPIPHANY_HDF, $LD_LIBRARY_PATH and $PATH variables to be passed to sudo. If you want to do this you still have to set them up as custom variables (it would be better to set them using Project | Build options... rather than as a global setting associated with the compiler). If you include /usr/local/browndeer/bin in the $PATH then you can leave it out of the bash call.



For those who skipped the OpenCL compiler setup, the values of the variables are those from your bash environment.

Compiling


Now that everything is in place we can compile a target.



From the main screen I use the tool bar but the Build menu provides the same functions or you can use Ctrl-F9 which must be the dumbest short-cut ever devised.

There are a couple of quirks to the OpenCL compilation. 

The offline compiler, clcc creates an intermediate file which is then passed to the epiphany compiler in the standard SDK. This means that the file name associated with the error will look something like UfLa0a.cpp. This means that double-clicking on the error will not highlight it in the editor. Oddly enough, if you #include another cl file which has an error, the file name associated with that error is correct and double-clicking works.

The intermediate stages also means that any error causes a cascade of errors after it. Just start from the top of the list recompiling as you go... the actual errors should be pretty obvious.

Adding a Virtual Target


If we leave things as they are now, we will have to build twice, once for each build target. If this is too tedious, Code::Blocks allows the creation of virtual targets which are combinations of "real" targets.

To create a virtual target go to the build targets tab in the project properties box (Project | Properties > Build targets). From here click the Virtual targets... button:


All pretty simple but there is one trick here. Click on your cl target first (CL in my example) and close. Click Virtual targets... again and click on your host code target (Debug in the example). This is to get the cl target to compile first which (if it compiles correctly) will produce the object file ready for linking (see linking below).

One word of warning here... check your target files after you do this. My cl file in the CL target was switched off and switched on in the Debug target.

Now you can choose your virtual target on the button bar and have both targets compile sequentially.

Checking your ELF file


Brown Deer have included a version of the nm utility called clnm. This will show you what the content of the elf file is:


If everything worked properly you should see your kernels in the list and some entries showing your source code if you have not used the -b (only binaries) compile option. (I'm not exactly sure why the kernels appear twice.)

Using Compiled Accelerator Object Code


At this point we have generated a .elf file and confirmed its contents. I believe that there is a way of using the .elf file directly but I can't find a description of how to do it. If anyone can give me a hand on this I'd appreciate it. While working on the debugging I had a thought, "Maybe I can link the elf file just like the object file?". After all, an elf seems to be analogous to a DLL on Windows. Turns out that this is the case. So the method described below works for either the object file or the elf file.

Linking


The method described in the Brown Deer documentation uses the object file containing your accelerator code and links it into your executable. To do this you need to add the accelerator code object file onto the end of the link stage of the main executable. Go to the Project | Build options... click on the options for your main executable and then the Linker settings tab:


The Other linker options string must be the path to the cl object code. This will be added onto the end of the list of object files that are passed to the linker.

Calling


Calling your kernel from your host code is a tiny change. Your call to clopen using the JIT compiler would look like this:

void* clhandle = clopen(context, (char*)string_to_cl_file, CLLD_NOW);

With the code linked in:

void* clhandle = clopen(context, 0, CLLD_NOW); 

The path to your accelerator code has been replace with a ZERO! That's it... nothing more to do. The short pause that happens as the JIT compiler goes about its business will disappear and everything else will look the same.

Final Thoughts


I hope that this guide has saved you some time and I hope Code::Blocks suits your way of working. As always, please let me know if I've missed anything or if any bits are confusing in any way.

Up next in Part 2 I'm going to get into some Debugging action. I split the two up because this guide is big enough as it is. I have not gotten into it yet so I hope that there will not be any back tracking to do on the compiler setup. 

Cheers.

No comments:

Post a Comment