NVIDIA PerfKit 5: New Tools for 3D Developers

Performance analysis in Frame Profiler

PerfHUD mode, called Frame Profiler, that appeared in the previous version of this utility is intended to help you optimize performance of 3D applications. Frame Profiler uses hardware GPU units and instrumented code of the driver to measure the load of separate GPU units in an application, it helps detect bottlenecks in the graphics pipeline. When PerfHUD goes to this mode, it runs several experiments to analyze performance of the current frame and gives a detailed report on draw calls.

Frame Profiler mode is an efficient method to determine problems in some scenes. It displays the most exacting draw calls in a frame as well as provides detailed info on each draw call. It allows to determine those that require your special attention. The mode consists in rendering the current frame several times and measuring GPU usage for each draw call separately. Proceeding from this information, draw calls are grouped into state buckets. All calls in a single bucket have similar characteristics, so removing a bottleneck in one of draw calls will increase performance of other draw calls in this bucket.

GPU usage diagram is displayed in the bottom by default. The upper bar of the graph shows an entire frame, the bars below show the load of separate GPU units (vertex assembly unit, geometry unit, unified shader units, texturing unit, ROP unit, and frame buffer unit) during the current draw call and all calls in a given state bucket. The yellow section of each bar shows the total time of a frame in milliseconds, the red one - the time GPU units are used by all state buckets calls, the orange one - the time a unit is used in the current draw call.

The GPU usage graph resembles the one displayed in Performance Dashboard mode. But the graph in Frame Profiler mode is based on the analysis of a single frame and displays GPU usage for each draw call separately. You can also see the following graphs: draw call durations in milliseconds - the time each draw call lasts in a frame; the time the FastZ depth/stencil unit is used - it shows activity time of this unit in milliseconds and a counter of processed pixels in each draw call.

As in the previous mode, Frame Profiler has the second interface view - Advanced, to access state inspectors. This Frame Profiler view allows to see details (PerfHUD 5 also allows to modify) of a given draw call (geometry data, pixel/geometry/vertex shaders used, ROPs) and see the time spent for each draw call in a selected frame. The mode is implemented in the same style as the Advanced mode in Frame Debugger. The color diagram resembles the bars in the GPU usage graphs, described above.

Minor Defects

Advantages of PerfKit 5 for 3D developers are apparent. These tools make programmers' lives easier. They get a convenient tool to analyze performance and debug their programs. Now it's time to speak about a fly in the ointment. Like in any new product, there are some defects in PerfKit 5, which are rapidly fixed in updates (so make sure you use the latest version by visiting this NVIDIA web site). For example, some users face installation problems, instrumented drivers do not always work correctly on all operating systems, there may be a runtime error, when you exit from a Direct3D application running under PerfHUD. For some strange reason, we couldn't use PDH in Windows Vista, while it worked fine in XP.

There is another minor problem - a performance drop due to the instrumented driver in Windows Vista may be higher than that in Windows XP. But NVIDIA are working on it. They promise to fix this problem in future. Besides, there is still no updated documentation for PerfKit. Only part of the User Guide was updated at the time this article was written.

In addition, it's not quite clear how to enable/disable special features of the instrumented drivers. This may come in handy, because special features reduce performance by several percents. NVIDIA Panel does not open in Windows XP tray, it crashes with an initialization error, and NVIDIA Developer Control Panel does not seem to have any options to disable driver instrumentation. That's another bug, although it does not ruin the main task of PerfKit. It copes with it brilliantly.

Conclusions

PerfKit offers very convenient tools to detect performance bottlenecks in 3D applications, optimize them, and eliminate render errors. You can use PerfKit to evaluate the effect of each unit in the graphics pipeline on the overall performance, and then eliminate bottlenecks.

Features of the updated PerfKit 5 make it much easier to determine problems in code. Software and hardware performance counters offer useful information about NVIDIA GPU operations. They can be used to evaluate the load on main GPU units in various scenes. There are several ways to use counters: you can use NVPerfAPI from your application or analyze their data in PerfMon and PIX. If you need a visual and convenient tool to analyze a Direct3D application, you may use PerfHUD, which will provide full information about a given application and GPU performance in a convenient form in real time as the application runs.

The latest version of PerfKit has several new features, which 3D developers have been waiting for a long time: support for the NVIDIA G8x architecture, full support for Windows Vista and Direct3D 10 applications, more experiments, hardware and software counters, convenient customizable interface of PerfHUD 5, editable vertex/geometry/pixel shaders and raster operations in real time, and many others. Better compatibility and stability were not forgotten either. Some minor bugs from previous versions were fixed. There have appeared new bugs, of course, but they are being fixed in frequent updates.

In our opinion, the most interesting new feature of PerfHUD is editing shaders and render states in real time. The utility allows to modify shader code on-the-fly and immediately see the results, which makes it much easier to test new ideas and optimize old ones.

So, PerfKit 5 will be extremely useful for all developers of modern 3D applications that use Direct3D or OpenGL. It will help solve most complex development problems: locating performance bottlenecks and render errors, full use of GPU resources, reaching balance between the load on various GPU units, etc.