Debugging GPU Code

Applies to: Visual Studio Visual Studio for Mac

You can debug C++ code that is running on the graphics processing unit (GPU). GPU debugging support in Visual Studio includes race detection, launching processes and attaching to them, and integration into the debugging windows.

Supported Platforms

Debugging is supported on Windows 7, Windows 8, Windows 10, Windows 11, Windows Server 2008 R2, Windows Server 2012 and Windows Server 2016. For debugging on the software emulator, Windows 8, Windows 10, Windows 11, or Windows Server 2012, Windows Server 2016 is required. For debugging on the hardware, you must install the drivers for your graphics card. Not all hardware vendors implement all debugger features. See the vendor documentation for limitations.

Configuring GPU Debugging

The debugger cannot break on both CPU code and GPU code in the same app execution. By default, the debugger breaks on CPU code. To debug GPU code, use one of these two steps:

• In the Debug Type list on the Standard toolbar, choose GPU Only.

• In Solution Explorer, on the shortcut menu for the project, choose Properties. In the Property Pages dialog box, select Debugging, and then select GPU Only in the Debugger Type list.

Launching and Attaching to Applications

You can use the Visual Studio debugging commands to start and stop GPU debugging. For more information, see Navigating through Code with the Debugger. You can also attach the GPU debugger to a running process, but only if that process executes GPU code. For more information, see Attach to Running Processes.

Run Current Tile to Cursor and Run to Cursor

When you are debugging on the GPU, you have two options for running to the cursor location. The commands for both options are available on the shortcut menu of the code editor.

1. The Run to Cursor command runs your app until it reaches the cursor location and then breaks. This does not imply that the current thread runs to the cursor; rather, it means that the first thread that reaches the cursor point triggers the break. See Navigating through Code with the Debugger

2. The Run Current Tile to Cursor command runs your app until all of the threads in the current tile reach the cursor and then breaks.

Debugging Windows

By using certain debugging windows, you can examine, flag, and freeze GPU threads. For more information, see:

• Using the Parallel Stacks Window

• Using the Tasks Window

• How to: Use the Parallel Watch Window

• Debug Threads and Processes (Debug Location toolbar)

• How to: Use the GPU Threads Window

Data Synchronization Exceptions

The debugger can identify several data synchronization conditions during execution. When a condition is detected, the debugger enters the break state. You have two options—Break or Continue. By using the Exceptions dialog box, you can configure whether the debugger detects these conditions and also which conditions it will break for. For more information, see Managing Exceptions with the Debugger. You can also use the Options dialog box to specify that the debugger should ignore exceptions if the data that's written doesn't change the value of the data. For more information, see General, Debugging, Options Dialog Box.

Troubleshooting

Specifying an accelerator

Breakpoints in GPU code are only hit if the code is running on the accelerator::direct3d_ref (REF) accelerator. If you do not specify an accelerator in your code, the REF accelerator is automatically selected as the Debugging Accelerator Type in the project properties. If your code explicitly selects an accelerator, then the REF accelerator will not be used during debugging and the breakpoints will not be hit unless your GPU hardware has debugging support. You can remedy this by writing your code so that it uses the REF accelerator during debugging. For more information, see project properties and Using accelerator and accelerator_view Objects and Project Settings for a C++ Debug Configuration.

Conditional Breakpoints

Conditional breakpoints in GPU code are supported, but not every expression can be evaluated on the device. When an expression can't be evaluated on the device, it is evaluated on the debugger. The debugger is likely to run more slowly than the device.

Error: There is a configuration issue with the selected Debugging Accelerator Type.

This error happens when there is an inconsistency between the project settings and the configuration of the PC that you are debugging on. For more information, see Project Settings for a C++ Debug Configuration.

Error: The debug driver for the selected Debugging Accelerator Type is not installed on the target machine.

This error happens if you are debugging on a remote PC. The debugger cannot determine until run time whether the drivers are installed on the remote PC. The drivers are available from the manufacturer of the graphics card.

Error: Timeout Detection and Recovery (TDR) must be disabled at the remote site.

It is possible for C++ AMP computations to exceed the default time interval that's set by the Windows timeout detection and recovery process (TDR). When that happens, the computation is canceled and the data is lost. For more information, see Handling TDRs in C++ AMP.

See also

• Walkthrough: Debugging a C++ AMP Application
• Project Settings for a C++ Debug Configuration
• Start GPU Debugging in Visual Studio