Parallel Programming in Native Code

Parallel programming using C++ AMP, PPL and Agents libraries.

Start GPU Debugging in Visual Studio 2012

Start GPU Debugging in Visual Studio 2012

  • Comments 11

There is fantastic debugging support for C++ AMP in Visual Studio 2012, and in this post I’ll show you how easy it is to get started.

1. Create a new "Hello World" project

Follow the steps described in an earlier post to create the "Hello World" project. Do not try this on one of your projects yet, try it with our "Hello World" first, building under the default Debug configuration.

2. Add breakpoints

Add three breakpoints on lines 8, 11, and 14. See the following screenshot.

clip_image002

3. Hit F5 to test CPU debugging

Hit the F5 key (or Debug->Start Debugging menu) to start debugging as normal, and you should see something like the following screenshot

clip_image003

Notice that the breakpoint on line 11, the one in restrict(amp) code, is shown as an unbound breakpoint and it will not be hit. If you continue execution (e.g. F5 again) you will break again on line 14.

That is what the hollow breakpoint icon is trying to convey, which if you zoom in on with the editor you will also see the yellow warning triangle.

clip_image005

In fact if you hover over it you will see this tooltip:

clip_image006

The clue is in the last part of the sentence: "the target architecture of this line is not supported by the current debugger code type". What it is saying is that your selected debug engine is not capable of debugging restrict(amp) code. You select a debug engine by selecting the “Debugger Type” in the project properties. If you do not select one, a default is used. By default, we expect you prefer debugging your CPU code.

So if you want to debug your restrict(amp) code, you need to change the Debugger Type - on to the next section.

4. Configure project properties

Stop debugging (Shift+F5) and change the selected debug engine as discussed previously. I.e. from the "Project" menu select "Properties" and under "Configuration Properties" select the "Debugging" node.

Background Info:

Notice how the Debugger Type is set to "Auto". Other options include "Native Only" (to use the native debug engine to debug native code on the CPU) and "Managed Only" (to use the managed debug engine for debugging code under the .NET CLR). The default "Auto" option means it will select the right CPU debug engine based on the contents of the binary. In the C++ AMP case, it is valid that you may want to debug using the native CPU debug engine or you may want to debug using the GPU engine, so we chose the default to be the native debug engine. As hinted previously when discussing the attach scenario, you cannot select them both. In some future release, we hope to be able to have the two debug engines (native and GPU) co-exist side-by-side in the same debugging session, and that is an aspiration rather than a commitment. So for Visual Studio 11 the story will remain the same as it is for the Beta.

Now, change "Debugger Type" from "Auto" to "GPU Only" and you will see this screenshot

clip_image008

Notice that when you choose "GPU only" two additional property options appear. We'll talk about both of them in future posts, but for now suffice to keep the defaults, and notice that the "Debugging Accelerator Type" is set to "GPU - Software Emulator".

5. Hit F5 to test GPU debugging

With that single project configuration change, you can now hit F5 and your breakpoints in restrict(amp) code will be hit, as the following screenshot shows.

clip_image009

Of course, now you cannot debug your restrict(cpu) code in the same debugging session, and those breakpoints become hollow, and carry the same tooltip message. C'est la vie.

So, at this point you are debugging against the REF emulator, which is great - you don't even need capable hardware for debugging!

If you are not running Windows 8 on your main C++ AMP development machine, continue to the next step.

6. Dealing with error on Windows 7

If you try the above with VS 2012 on Windows 7, you'll be able to achieve all steps except at step 5 when you hit F5 you might see the following error:

clip_image011

We have added support for C++ AMP debugging on Windows 7. To try out GPU debugging on Windows 7, please refer to C++ AMP GPU debugging now available on Windows 7 blog post.

To try out GPU debugging on Windows 7, you need to apply the latest Windows updates. If for some reason you cannot update your machine, a workaround that may be good for some scenarios is to continue with Visual Studio 2012 installed on your main development machine on Windows 7 and then remote debug to Windows 8 on a remote machine (or a VM that you run on your local machine). We describe that process in the blog post on Remote GPU Debugging in Visual Studio 11.

Potential debugging gotcha for your own code

So now that you can debug the "Hello World" project with the GPU debugger, you are ready to try GPU debugging with your own code.

You will have noticed that the code above does not explicitly specify an accelerator, which means that it is using the default accelerator picked by the runtime. We have done some magic so when you select the GPU debugger, we switch the default accelerator to the “Debugging accelerator type” selected in the project properties. When you are running in release mode or not using the GPU debugger, the default will of course be what we consider the best on your system, as described in the default accelerator post.

However, if you are not relying on the default accelerator, and instead you are explicitly specifying an accelerator_view directly in your code, then you will probably not be running on REF, so your breakpoints will not be hit (unless you have GPU hardware debugging support, which is a topic for another day). So in that case you should change your C++ AMP code to explicitly use the REF accelerator, and then the GPU debugger will work for you again. For example if you were passing your own accelerator_view to parallel_for_each, you would change it in debug mode to be like this:

accelerator_view acc_view = accelerator(accelerator::direct3d_ref).default_view

parallel_for_each(acc_view, av.extent, …

Next steps

The next steps for us in the GPU debugging story is to talk about the two additional options in the project properties I mentioned above (including debugging on GPU hardware), and to talk about the very rich GPU and parallel debugging experience in Visual Studio 2012: hitting breakpoints, stepping and Run To Cursor, inspecting variables and call stacks, Debug Location Toolbar, GPU Threads, Parallel Stacks, Thread Markers, Parallel Watch, Flagging threads, Freezing threads, Run Current Tile To Cursor, inspecting Registers, raw Memory and bytecode through Disassembly debugging, and.... the C++ AMP race detection tool through GPU Memory Access Exceptions! Stay tuned.

For now, please try the steps above, and I have already attached the "Hello World" project for you to download. Beyond that, I recommend following the more complete GPU debugging walkthrough on MSDN. If you have questions, please leave them below in the comments.

Blog - Comment List MSDN TechNet
  • Loading...
Leave a Comment
  • Please add 2 and 2 and type the answer here:
  • Post