A PIX GPU capture records all the Direct3D 12 API calls made by the game, including their parameter data. These calls can later be replayed, which enables a range of debugging and analysis features.
It is not always possible for PIX to successfully take a GPU capture if a game is calling Direct3D 12 in invalid ways. We make a best effort to be robust even in the case of incorrect usage patterns, but this is inevitably sometimes a case of garbage in, garbage out. If you are having difficulty taking GPU captures, try using the D3D12 Debug Layer and GPU-Based Validation to find and fix any bad API calls.
Windows GPU captures are not in general portable across different GPU hardware and driver versions. In most cases a capture taken on one machine will play back correctly on other similar GPUs from the same hardware family, and captures of some games may even work across GPUs from entirely different manufacturers, but it is also possible that something as trivial as a driver upgrade could break compatibility with older captures. We can only guarantee playback will succeed when the GPU and driver are exactly the same, so PIX will warn before starting analysis if there is not a perfect match. Proceed past this at your own risk!
PIX does not currently support multiple GPUs. It will always play back GPU captures on the primary adapter. If you have a hybrid machine, which contains both integrated and discrete GPUs, you can set which one is the primary in the driver control panel.
When you first load a GPU capture, data is loaded and parsed but the API calls are not yet actually played back on your GPU. Not all parts of PIX are fully functional while in this state. To enable complete functionality you must start analysis, which instructs PIX to create a Direct3D 12 device and play back the capture in the various ways necessary to extract information. The analysis Start button is found in the PIX toolbar:
The Events view shows all the API calls that were recorded into a GPU capture. There is a separate list for each Direct3D 12 queue (graphics, compute, or copy) that was used by the game:
The event list can be filtered, optionally using regular expressions. By default it only shows events that resulted in actual rendering work for the GPU hardware, as opposed to simply preparing state for use by later operations. To include non-GPU events, click the button labelled !G.
The event list works best when the game has been instrumented with the PIX event runtime.
More information about each event, such as the full set of API call parameters, is available in the Event Details view. This is not included in the default PIX layout, but can be accessed via the Views button in the upper right corner of the main PIX window.
Timing Data and Counters
The Collect Timing Data button (top right of the Events view) instructs PIX to replay the captured API calls a number of times, measuring how long each operation takes to execute on the GPU. Results from more than one replay are averaged to reduce measurement noise.
For best results, do not interact with your computer while PIX is collecting timing data, and close any other applications that might be using the GPU.
There are two sources of GPU timing information:
- Start-of-pipe timestamps report when the GPU starts processing each piece of work. Note that this data is only available when running Windows 10 build 14393 or greater with a suitable graphics driver. Timing data takes longer to collect if start-of-pipe timestamps are not available.
- End-of-pipe (EOP) timestamps report when the GPU has finished processing each piece of work.
Because GPUs are massively parallel and deeply pipelined, it is common for more than one piece of work to be executing at the same time, and for adjacent operations to overlap. PIX measures time in two different ways that can offer insight into the parallel execution model of the hardware:
- Execution Duration is measured from the start-of-pipe to the end-of-pipe timestamp of each separate piece of work. When measured in this way, GPU work items that run in parallel with prior items may appear longer than if they were run in isolation due to contention on the GPU.
- EOP Duration is measured from the previous item’s end-of-pipe (EOP) timestamp to this item’s EOP timestamp. Work items that run in parallel with prior items will therefore appear shorter than if they ran in isolation, and some items may even be reported as zero duration if they completed entirely in parallel with the previous item.
The Timeline view displays one or more lanes showing the timing of each GPU operation. There is a separate lane containing EOP Duration data for each queue (graphics, compute, or copy) used by the game, plus a single lane showing Execution Duration data (where available) combined across all the queues.
A range of counter values including Execution Duration, EOP Duration, Execution Start Time, and other values such as Samples Rendered or VS Invocations can be added to the Events view by clicking the Counters button at the top right of that view. When running on an NVIDIA GPU with a recent driver, you will also have access to ~300 NVIDIA counters, which can provide much deeper insight into the rendering and performance characteristics of your workload.
Note that PIX on Windows does not currently overlap GPU work on different queues while analyzing timing. If a game uses async compute to execute rendering and compute work simultaneously, PIX will instead measure first one and then the other. This may result in shorter reported durations for each part of the work compared to how it would execute inside the original game (due to reduced contention on the GPU) but a longer total time (due to reduced parallelization).
After selecting an event in the Events view, the State and Pipeline views (found in the Pipeline tab) will show details of the Direct3D state at the time of that event. Here you can view what resources are bound to the pipeline, shader code, inputs, outputs, and the currently bound rendertarget(s).
Right-clicking on part of the rendertarget (OM RTV) and choosing Show Pixel History will bring up a view showing all the draws that affected the selected pixel.
To debug a pixel shader: right-click on one of the entries in the Show Pixel History list, and choose Debug This Pixel.
To debug a vertex shader: right-click on one of the vertices shown under VS Output, and choose Debug This Vertex.
Shader Edit & Continue
HLSL shader code can be edited directly inside PIX, allowing you to immediately see the effect of your changes on rendering results or performance. This can be useful for prototyping and optimizing shaders, as it can greatly reduce the turnaround time when trying out different ideas.
Edit & continue is accessed from the Pipeline view. After selecting a shader you can edit its code however you like, then click Commands followed by Apply Edits:
Other views (such as OM RTV 0) will update to show the effect of your change. You may find it useful to dock more than one instance of the Pipeline view next to each other in order to view rendertarget results at the same time as editing shader code.
If you lose track of exactly what you have changed, right-click on the shader code and choose Diff with Original:
This will bring up an inline diff view, showing the original code in red with your changes in yellow:
Note that PIX on Windows (unlike Xbox) does not yet support saving modified shaders back into the .pix3 file.
Rendertarget visualizers are available in the Pipeline view when inspecting the contents of a rendertarget. The default Image visualizer just shows the image as normal:
The Wireframe visualizer highlights whatever geometry was rendered by the currently selected draw call in wireframe. Visible pixels are shaded green, while culled pixels (eg. backfacing triangles or failed depth test) are red:
The Overdraw visualizer colors the scene according to how many pixels were shaded and written to the framebuffer after passing the depth test. This is useful for understanding how well things like front-to-back sorting or a z prepass are working to reduce overdraw:
The Depth Complexity visualizer is similar to Overdraw, but disables the depth test so the results indicate only how many triangles overlapped each pixel, regardless of sorting or z prepass:
Finally, the Pixel Cost visualizer colors the scene according to the approximate rendering cost of each pixel. The brightest areas of the resulting image indicate which parts were the most expensive to render. Pixel cost is estimated by dividing the Execution Duration of each draw call by the number of resulting pixel shader invocations, and accumulating that average each time a pixel is written. This includes the cost of other pipeline stages, but it is divided evenly across all pixel shader invocations in the draw without accounting for varying vertex density, etc.
PIX provides a convenient way to analyze a capture with the D3D12 Debug Layer. Use the Run Debug Layer or Run GPU Validation buttons in the Warnings view. GPU validation is a superset of the basic debug layer, but can take longer to run.
This functionality can be useful if you want to take advantage of GPU validation to verify correctness of your Direct3D 12 API usage, but the overhead of the validation makes your game too slow to easily reach the point of interest. Rather than running GPU validation directly against the game itself, you can take a GPU capture of a single frame, then use the PIX Run GPU Validation button to validate only this one frame in isolation.