NVTune

1
NVTune

• Kenneth Hurley

2
NVTuneOverview

• What issues are we trying to solve?
• Games and applications need to have high frame
  rates
• Answer questions for developers
• How well is my code interfacing with the driver?
• Is it possible to speed up the process?
• Why is the hardware not doing what I expect?
• Are there logic errors in my use of Direct X?
• How can we solve these issues?
• Have the driver tell us where it is spending its
  time
• Have the driver output any problems it may see.
• Example Using a texture stage in the wrong way
• Example Locking a front buffer causing the
  triangle pipeline to stall.

3
NVTuneOverview (continued)

• How NVTune help you
• Captures CPU cycles spent in NVIDIAS driver
  code.
• Simple and easy to use
• Enhances other development tools such as VTUNE,
  in that it gives more of a breakdown of what is
  happening in the driver.
• Low overhead. Display refresh is specified by the
  user. Logging display is at lowest priority for
  a thread.
• Newest code and driver will allow more of an
  inside view of what is going on inside the
  GeFORCE.
• Directly tied into NVIDIAS Direct X Driver

4
NVTuneOverview (continued)

• Architecture
• Time stamp counter used (RDTSC)
• VXD/SYS Interface for application and driver
• Logging output is directly from the driver

Log Global Memory
Statistics Application
D3D Driver
Data
Statistics VXD
Hardware
5
NVTuneOverview (continued)

• What it captures (examples)
• Draw Primitives (indexed, non-indexed, wireframe,
  etc.)
• Blits
• Flips
• Render State changes
• Resource Lock/Unlock
• Textures
• Vertex buffers
• Surfaces
• All function calls that reach the driver

6
NVTune - Toolbar

• Toolbar
• New - Opens three new windows, if no windows are
  open.
• Save - Saves active window in comma delimited
  format
• Cut - No action currently
• Copy - No action currently
• Paste - No action currently
• Print - No action currently, will print current
  window
• View D3D function window
• View frames per second window
• View log window
• Start statistics capture
• Stop statistics capture
• Erase values in current window
• Connect to remote statistics machine
• Set configuration

7
NVTune - Function Split Window
8
NVTune - Function Split WindowContinued

• Function split window
• Allows you to choose which functions are
  displayed during statistical capture
• Functions are not persistent if Active Only is
  checked in the configuration
• Data can be saved to comma delimited text file
• Updates at specific time intervals
• Displays function, class, number of hits, CPU
  clocks, of total functions and of CPU usage.
• Sorting by column is enabled. Sorting
  automatically updates every refresh.

9
NVTune - Frames Per Second Graph
10
NVTune - Frames Per Second Graph Continued

• Frames per second graph
• Cursor allows you to look at a particular frame
  and what the average frame rate is.
• Visual cues, on spikes/lows for 3d functions

11
NVTune - Logging Window
12
NVTune - Logging WindowContinued

• Logging Window
• Visual output for logging data.
• Can be saved to comma delimited text file. This
  allows you to import into excel and make your own
  graphs.

13
NVTune - Configuration

• Easy Configuration
• Click on toolbar or menu

14
NVTune General Options
15
NVTune General OptionsContinued

• Logging Level
• Decimal/Hexadecimal display for programmers
• Display only active functions
• Allow automatic updating of only functions that
  are active
• Hotkey Sets the hotkey that turns the capture
  on/off.
• Allows full screen applications to capture data
  inside of game
• Any key combination can be used
• When setting press one key at a time then dialog
  will skip to next edit box
• Remote Access
• Small stub on target machine
• Across TCP/IP through use of IP address or
  machine name
• Still in Development!

16
NVTune - Capture Options
17
NVTune - Capture OptionsContinued

• Frame End Signaling
• Sets how frames are ended
• Flip, blit to front buffer or Begin/End Scene
  Pairing
• Range Capture Enable
• Enable start and stop between successive frames
• Enable checkmark, then set start and ending frame
  to capture between
• Refresh interval in milliseconds
• How often visual updates are made in function
  window

18
NVTune - Logging Options
19
NVTune - Logging OptionsConfiguration (continued)

• Logging Options
• None
• Minimal logging
• Only logs function calls
• Medium logging
• Logs function calls and lt 1 DWORD parameters
• Maximum logging
• Logs function calls, all parameters and vertices
• Performance Warnings
• Shows red highlighted performance problems
  reported by driver
• Logging Memory
• Sets how much memory for logging file
• Doesnt utilize Virtual Memory for performance
  reasons

20
NVTuneAdditional Features

• Context Sensitive Help
• Online FAQ built into statistics application
• Highlight function and press F1.
• Some function havent been documented yet. If
  you find one let us know and we will document it
• Search function for log files
• Use Ctrl-F to find
• Use F3 to find again

21
NVTuneFuture Enhancements

• GeFORCE Specific features
• Ability to track additional features such as
  pixels per clock, pipeline stalls and almost
  anything that can happen inside of the hardware.
• This will be made available through performance
  monitor registers.
• Transform and Lighting CPU usage inside of
  drivers
• DX7/DX8 additional RenderStates
• Windows 2000/Windows 9x Support

22
NVTuneWhere to get it

• Available to Registered Developers
• http//nvdeveloper.nvidia.com
• Report Bugs/Feature Requests statdriver_at_nvidia.co
  m