Home

PVRTune User Manual Developer

1. Total triangle no Texture uploads Texture modifications Revision PowerVR SDK REL_3 5 3649239a 36 User Manual og Imagination Public Imagination Technologies Software counters Scissor calls Viewport calls Frame buffer accesses Vertex shader compiles Fragment shader compiles Program links Framebuffer access bytes Texture uploads bytes Texture modifications bytes Buffer object uploads bytes Buffer object modifications bytes Uniform uploads Context binds Shader proportion pixel Shader proportion vertex Shader proportion compute Vertices per batch Shader slot scheduled count compute Shader slot scheduled count pixel Shader slot scheduled count vertex Shader slot size pixel Shader slot size vertex 3 6 6 Counter Properties PVRTune GUI displays of a wide range of information associated with counters By exploring the properties of a counter it becomes possible to learn more about e g its description and implications when a high value is recorded amongst others The most recently clicked counter is displayed in the Counter Properties window Figure 28 The window captures some basic details including the counter name associated colour and the maximum value of the Y axis to be used on the graph view PVRTune 37 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Pu
2. qat N Commands PVRPerfServer to auto quit after a specified number of seconds 2 5 5 Run Time Options PVRPerfServer supports several key presses at run time This is summarised in Table 2 Revision PowerVR SDK REL_3 5 3649239a 8 User Manual og Imagination Public Imagination Technologies Table 2 Run time hotkeys Key Effect H Show help text M Send a mark Useful for placing simple markers into the data stream annotated with a number that increments for each mark Q Quit PVRPerfServer PVRTune 9 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 3 PVRTune GUI 3 1 The Basics 3 1 1 Overview The PVRTune GUI captures and presents the real time information sent by PVRPerfServer and also provides a means of saving and loading PVRTUNE files for analysis 3 1 2 Installation PVRTune GUI can be installed from the PowerVR SDK installer 1 Navigate to the PowerVR Insider SDK webpage powervrinsider com and download the SDK 2 Launch the installer and follow the on screen instructions 3 Once the SDK has been successfully installed PVRTune GUI will be available in lt InstallDir gt PVRTuneDeveloper PVRTune lt PLATFORM gt 3 1 3 User Interface Layout The default analysis interface is displayed after a connection has been established or after a PVRTUNE file has been loaded see Section 3 3 for more information on how to connect The PVRTune GUI displays
3. Force 2x2 textures Forces textures to be 2 pixels by 2 pixels in size Force flat colour frag shader Forces the fragment shader to output a single colour Disable stencil test Disables the stencil test Disable depth test Disables the depth test Disable scissor test Disables the scissor test Disable draw calls Disables all g1Draw function calls Disable flush and finish Disables all g1Flush and glFinish function calls Disable texture modifications Disables g1TexSubImage2D and glTexSubImage3D This disables the updating of texture sub regions Disable texture Disables texture filters All texture filters are set to the nearest point filtering Disable discard Disables the alpha test alpha test Disable frame buffer Disables gl1ReadPixels function calls accesses Culling mode App defined Forces the culling mode to perform culling as defined in the user application Culling mode None Disables culling Culling mode Back Forces the culling mode to perform back face culling Culling mode Front Forces the culling mode to perform front face culling Revision PowerVR SDK REL_3 5 3649239a 40 User Manual og Imagination Public Imagination Technologies 3 9 1 OpenGL ES Timing Data In some situations in addition to the Tiler and Renderer processing times it is also useful to know more about how the API graphics cal
4. eccccceeccceeeeeeeneeceeeeeceeeeeeaeeseeneeseaeeeseaeeseeeteaes 23 Figure 13 Example of Tiler and Renderer timing data A 26 Figure 14 TranSten taSkS ecann Aa S AAE ERA AA ARAE A 27 Figure 15 Example of task colour COGING sissi sainnaiann kaain a a AR A aA 28 Figure 16 Example of driver timing data 28 Figure 17 Example of active counters change 29 Figure 18 Example of event ordinal reset 0 cc eeeececeeeeene cece eee ee erences ee eaaeeeeeeaaeeeeesaaeeeeeeaeeeeeeaeeeeneaa 29 Figure 19 Example of custom mark eee ee eenee eee eeee ee eee aaa EANA ae ee AEAEE 30 Figure 20 Example of power off Geo 30 Figure 21 Example of data loss Period 0 20 0 eeececeeeeeeeeeeeeeee eset eae eeeeeaaeeeeeeaaeeeeeeaaeeeeeeaaeeeeeeaeeeeeeaeeeeneaa 31 Figure 22 Example of hardware reset Gertod 31 Figure 23 Example of SPM mode nenn nsen nnen nenn 31 Figure 24 Counter Table window ssssssssssssssssenssenssrnssrnsstnnsttnstnnsttnstnnttnnttnnntnnntnnnennnennntnnonnnenneen nenn 32 Figure 25 Dialog box for selecting and deselecting Columns 33 Figure 26 Adding a Counter to a graph view eeeeeececeeeneeeeeeneeeeeeaaeeeeeeaaeeeeeeaaeeeeeeaaeeeeesaeeeeeenaeeeeeeaa 34 Figure 27 Creating a new graph by adding a counter to a graph placebolder 34 Figure 28 Counter Properties window unuttu nnn nennnnn nnna 38 Foue 29 PID WNdON ca TETS 39 Figure 30 Remote Editor window essisieresisi iseia aa aa a E Na ARA E a A 39 Figure 31 Renderstate Over
5. 2 Click the plot corresponding to the counter This will highlight the plot 3 Press Delete to remove the plotted counter data from the graph The same action can be achieved by right clicking the counter plot and selecting the Remove Selected Counter option from the action menu 3 6 4 Counter Legend in the Timeline Area The Counter Legend is located at the right hand side of the Timeline area and lists the counters that are only relevant to the plotted graphs see Figure 12 A number of options are available to interact with the information displayed in the Counter Legend and these are next discussed Highlight a Plot Corresponding to a Counter In order to highlight a plot for a counter displayed in the Counter Legend mouse over the name of the counter Add a Column of Data to the Counter Legend The data sets displayed in the Counter Legend can be customized on demand In order to do this perform the following steps 1 Right click anywhere in the Counter Legend This will open an action menu with several options 2 Click the Select Column option This will open a dialog box to aid the selection of columns see Figure 25 Note For further information about the description of the various columns and their functions see Table 5 3 Select the desired column to display by using its corresponding checkbox Multiple columns can be chosen if required 4 Click the Accept button to apply the changes The selected column
6. Figure 30 Remote Editor window Note For more information on PVRScope refer to the PVRScope User Manual 3 9 Renderstate Override Window The renderstate override functionality in PVRTune permits the user to remotely control the renderstate of a target device This provides a useful means of prototyping render changes as well as enabling the quick identification of obscure bottlenecks The renderstate override functionality is accessible from the Renderstate Override window and in order to use it PVRTrace profiling mode should be enabled PVRTune 39 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public QJ imagination Render State Override ax Connection 17 OGLES2Water PID 26223 D Force zero viewport Disable blending Force 2x2 textures Force flat colour frag shader Disable stencil test Disable depth test Disable scissor test Disable draw calls Disable flush and finish Disable texture modifications Disable texture filtering Disable discard alpha test Disable frame buffer accesses Culling mode App defined e Figure 31 Renderstate Override window Table 7 provides a list of the various renderstate override options as well as their description Table 7 Explanation of renderstate override options Option Description Force zero viewport Modifies the viewport to have zero sized dimensions Disable blending Disables alpha blending
7. In addition to the Tiler and Renderer timelines there can also be others depending on the hardware being profiled A summary of hardware specific terms is provided in Appendix A Also consult the PowerVR Hardware Architecture Guide for a more detailed understanding of the PowerVR hardware architecture Each block displayed in the representation of the timing data corresponds to a given task or activity within a frame and is colour coded to make the frame process ID and render target easily identifiable In addition to Tiler and Renderer timing data there can also exist other sets of data for transfer tasks 2D core time for chips with dedicated 2D cores compute time for PowerVR Series6 onwards and custom timing data sent using PVRScope Renderer task 3 12 PID 13274 iframe 145 UI 3 529ms Renderer busy D Renderer 22 775ms 12 Renderer tasks MT Tiler 4 354ms 6 Tiler tasks a 24 753ms latency pent Lei lil N N View Details of Timing Data A wealth of information can be gained from the timing data by holding the mouse pointer over a task in the timeline The information displayed this way details the process ID the frame number the total number of tasks on each timeline Such as Tiler and Renderer that can be attributed to the same frame the time spent on each task the time spent processing a set of tasks and the total time spent processing the fram
8. e PVRTune This is the client side element of the combined package It provides a graphical representation of the real time information being sent by PVRPerfServer as well as providing a means to save and load PVRTUNE files Note On Android PVRHub is an application that has a user interface On Linux it is a set of folders and scripts For more information regarding PVRHub see the PVRHub User Manual PVRTune 5 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og Imagination PVRTune PVRPerfServer PVRScope Enabled of PVRHub Application PVRScope PVRScopeServices Device Driver PVRTune Figure 1 Structure of PVRTune Revision PowerVR SDK REL_3 5 3649239a 6 User Manual og Imagination Public Imagination Technologies 2 PVRPerfServer 2 1 Overview PVRPerfServer is a console application for Android Linux Neutrino and Windows Its purpose is to read counters and registers from the PowerVR hardware on a device and either save the data to a PVRTUNE file or transmit that data to PVRTune over a network 2 2 Requirements In order for PVRPerfServer to function correctly the driver must have performance profiling enabled In many cases as the overhead is so small this functionality is available by default This can be checked by confirming the existence of PVRScopeServices dll libPVRScopeServices soor an equivalent If PVRPerfServer fails to initialise it is likely that performanc
9. 2 1 File Menu Figure 3 illustrates the File menu which provides options for opening and saving files managing tabs and exiting PVRTune GUI Revision PowerVR SDK REL_3 5 3649239a 12 User Manual og Imagination Public Imagination Technologies File Edit View Connection New Tab Ctrl T Close Tab Ctrl W Rename Tab F2 PF Open fa Save Ctrl S Exit Figure 3 File menu Open a New Tab To open a new tab in the Timeline area click File gt New Tab Figure 3 This is a useful feature that allows multiple graph views to be opened at the same time Close a Currently Selected Tab Closing the currently selected tab in the Timeline area can be achieved by clicking File gt Close Tab Figure 3 Rename a Tab Tabs present in the Timeline area can be renamed on demand To rename a tab select it and click File gt Rename Tab Figure 3 Open a File To open a PVRTUNE file click File gt Open Figure 3 This will open a dialog box for browsing to the required file Save a File To save a PVRTUNE file click File gt Save Figure 3 This will open a dialog box in which the name of the file can be input and the file saved Exit PVRTune To close PVRTune GUI click File gt Exit Figure 3 3 2 2 Edit Menu Figure 4 illustrates the Edit menu which provides options for editing user preferences View Connectior SS Preferences Figure 4 Edit menu Preferences User preference
10. 32s column gives the average over the last 32 seconds Selected By selecting a time range see Select a Time Range and using this checkbox it is possible to view the average over the selected region Line This column responds to mouse hover It shows the values of counters at the time associated with the position of the mouse in the graph view Task This column responds to mouse hover It shows the average value of counters over the time frame of a single Tiler or Renderer task Frame This column responds to mouse hover It shows the average value of given counters over the timeframe of a single frame View This column responds to mouse hover It shows the average value of counters over the timeframe of the graph for which the cursor is currently over Add a Column of Data to the Counter Table An additional column of data can be included in the Counter Table window by selecting it from the participating dialog box Figure 25 Multiple columns can also be chosen if necessary Remove a Column of Data from the Counter Table A column of data can be removed from the Counter Table window by selecting it from the participating dialog box Figure 25 Multiple columns can also be chosen if necessary PVRTune 33 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 3 6 2 Add a Counter to the Graph View During connected analysis in PVRTune GUI counters are updated in rea
11. d frame time in seconds o n a ms 3353 S58 Similar to the frame time counter this value is not calculated per CPU process This counter represents the number of frames the GPU has been able to d g render within the selected period of time regardless of which process Tier 2 111ms 1 Tier task submitted the rendering task H Gi GK On most platforms V sync will prevent the number of frames per second fo ep ET exceeding the refresh rate of the display What does a high value mean Lol u fmooth 8 Time period 19 648ms Zoom 32 g Monitor Remote Editor ep W a ane PVR a Q d d O O P P P 3 1 4 Getting Started with Analysing an Application There are several benefits to using PVRTune to analyse an application and the achievement of effective analysis is largely driven by the ability to understand how to identify application bottlenecks see Section 4 By analysing an application it is possible to obtain improved frame rate to increase user enjoyment and application responsiveness Furthermore using PVRTune allows reduced render time to be obtained without increased frame rate in order to gain more idle time thereby saving power Analysing using PVRTune can also help increase visual quality without sacrificing frame rate There are two main approaches to analysing an application in PVRTune namely through connected analysis which displays real time data and offline analysis from a saved PVRTUNE file These main steps invo
12. is then added to the Counter Legend Remove a Counter from the Counter Legend A counter can be removed from the Counter Legend when it is no longer required To achieve this click the counter and hit the Delete key Alternatively right clicking the counter will open an action menu from which the Delete Selection option can be chosen Removing a counter from the Counter Legend also removes its corresponding plot from the graph view Remove all Counters from the Counter Legend All counters displayed in the Counter Legend can be removed on demand To achieve this right click anywhere in the Counter Legend to open an action menu from which the Delete A11 option can be chosen Removing all counters from the Counter Legend also removes their corresponding plots from the graph view PVRTune 35 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Remove a Column of Data from the Counter Legend To remove a column of data from the Counter Legend perform the following steps 1 Right click anywhere in the Counter Legend This will open an action menu with several options 2 Click the Select Column option This will open a dialog box to aid the deselection of columns see Figure 25 3 Uncheck the desired column by using its corresponding checkboxes Multiple columns can be deselected if required 4 Click the Accept button to apply the changes The deselected column is then removed from the Counter Leg
13. receives counter updates much quicker than otherwise This option is enabled by default Timing Data Enable When this option is ticked PVRTune receives Tiler or Renderer timing data This option is enabled by default Operating System Name and version details of the operating system of the target device Driver name Version of the PowerVR driver being used Device Variant Variant specific details of the device Device Series Series specific details of the device Server Description PVRPerfServer name and version number Server Build Release and build version of PVRPerfServer Server Connection Time local The local time at which PVRPerfServer became connected Server Connection Time UTC The Coordinated Universal Time UTC at which PVRPerfServer became connected Note Clicking the Connection gt PVRPerfServer gt Send Quit Message option sends a message to PVRPerfServer requesting the process to exit 3 3 9 Close a Connection A connection can be closed on demand when it is no longer required To close a connection click Connection gt Close If real time analysis is being performed then clicking the Close option once will first stop connection to PVRPerfServer allowing the data to be saved as a PVRTUNE file or for continuing the analysis as is Clicking the Close option a second time will exit the analysis and return the user to the connection form On the other hand in
14. the case of offline analysis from a previously saved PVRTUNE file clicking the Close option simply exits the analysis and returns the user to the connection form PVRTune 21 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 3 4 Monitor Window Performance at a Glance PVRTune GUI comes with a monitoring functionality which allows the user to view a high level overview of graphics core workloads thereby giving insight on how to isolate application bottlenecks The Monitor window displays and categorises the various loads being tracked by PVRTune as shown in Figure 11 The dropdown box for the time period Figure 11a facilitates choosing over how long the loads should be averaged In the example this time period is set to the default 0 5s value Note Further information on how to analyse bottlenecks can be found in Section 4 Monitor Gx Tiler active Renderer active 2D active Shader load CPU load by core CPUload byPID B System memory load Time period 0 5s en a Monitor Remote Editor PID Figure 11 Monitor window Revision PowerVR SDK REL_3 5 3649239a 22 User Manual og Imagination Public Imagination Technologies 3 5 Working with the Timeline The captured data is plotted in the Timeline area of PVRTune GUI and displayed as a graphical representation in order to facilitate analysis Figure 12 illustrates an example of static data displayed after a sav
15. 1 Identify the counter that is of interest either from the Counter Table window see Figure 24 or from the Counter Legendin the Timeline area 2 Select the counter by clicking it This will display the properties of the counter in the Counter Properties window Figure 28 3 Click the colour icon associated with the counter Figure 28a This will open a colour picker from which a suitable substitute colour can be chosen 4 Select the desired colour from the colour picker 5 Click OK to complete the procedure Change the Y Axis Scaling of a Counter The ability to change the Y axis scaling of a counter is a handy feature that allows small and large value plots to be scaled for visibility and interpretation This can be achieved by using the Y Axis field provided in the Counter Properties window Figure 28b For particularly large value plots such as transformations per frame it is necessary to enter a large scale On the other hand for very small value plots Such as frame time it is necessary to enter a fractional value View the Description of a Counter Counter description is very important as it helps data interpretation and analysis To view the description of a counter perform the following step 1 Identify the counter that is of interest either from the Counter Table window see Figure 24 or from the Counter Legendin the Timeline area 2 Select the counter by clicking it This will display the properties of the counter in th
16. O localhost Ctrl L Recent Connections PVRPerfServer gt S i Close Ctrl C Figure 7 Connection menu 3 2 5 Help Menu The Help menu is opened by selecting the appropriate option from the Menu bar Figure 8 It provides options for accessing PVRTune help assets sending feedback viewing general PVRTune release information and checking for software updates Revision PowerVR SDK REL_3 5 3649239a 16 User Manual og Imagination Public Imagination Technologies PVRTune Help H Feedback Check for Updates About PVRTune Figure 8 Help menu View PVRTune User Manual To view the PVRTune User Manual click Help gt PVRTune Help Figure 8 Submit Feedback To provide feedback click Help gt Feedback This will open a dialog box where instructions are displayed on how to post feedback and request for support Figure 8 Check for Updates As of SDK release 3 0 PVRTune is able to auto update However to force check for software updates click Help gt Check for Updates Figure 8 About PVRTune To view basic information about PVRTune release information such as versioning and contact details click Help gt About Figure 8 PVRTune 17 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 3 3 Connection Management Establishing a connection is a fundamental step prior to being able to visualize hardware performance data in PVRTune GUI The connecti
17. a large amount of information Figure 2 split into the following sections e Menu bar Figure 2a The section enables access to several options related to file edit view connection and help e Timeline area Figure 2b This section displays counter data as a function of time Counter Table window Figure 2c This section lists all the hardware and software counters that are in use e Counter Properties window Figure 2d This section is used to view counter specific information when a counter is chosen from the Counter Table e Find window Figure 2e This section allows the user to search for a variety of items such as counters and markers e Renderstate Override window Figure 2f This section allows the user to interact with the renderstate of the target device e Monitor window Figure 2g This section displays various loads recorded by the utility e Remote Editor window Figure 2h This section details data from PVRScope e PID window Figure 2i This section displays the connections being used in the recording e Status bar Figure 2j This section provides general information pertaining to the running of PVRTune Revision PowerVR SDK REL_3 5 3649239a 10 User Manual og Imagination Public Imagination Technologies E filev c Shadowljun pyrtune PVRTuneDeveloper v14 104 3 L N Sax ile Edit View Connection Help Counter Ta
18. al og Imagination Public Imagination Technologies g active counters changed Figure 17 Example of active counters changed Event Ordinal Reset This event is usually hidden by active counters changed The event represents a change in the ordering of the counters or counter sources read by PVRPerfServer and appears as a vertical green line Figure 18 event ordinal reset Figure 18 Example of event ordinal reset Custom Mark Custom marks are marks that have been sent to PVRTune either by a PVRScope enabled application or by pressing the M key in PVRPerfServer A custom mark appears as a vertical red line Figure 19 PVRTune 29 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination PID 4311 frame 8700 Figure 19 Example of custom mark Power off Period Power off periods are represented by vertical grey blocks in the graph view Figure 20 These events occur when the hardware has gone to sleep or powering down to save power due to a lack of work wooo AN e Figure 20 Example of power off period Data Loss Period A data loss period is one during which PVRTune has lost data that was sent by PVRPerfServer A data loss period is represented as a vertical green block Figure 21 There are several possibilities to minimize data loss such as e Decrease device CPU load e Alter the PVRPerfServer data read rate using the c command line option
19. an slow all operations on the hardware This is platform specific and as such there is no counter to record it As a rule of thumb action should always be taken to reduce bandwidth use whenever possible through the correct use of texture compression mesh optimisation and by avoiding unnecessary texture reads etc Note Bandwidth in System on Chip SoC devices is shared amongst all components of the chip Non graphics processor areas of the chip using large amounts of bandwidth may still cause application graphics to be bandwidth limited Revision PowerVR SDK REL_3 5 3649239a 44 User Manual og Imagination Public Imagination Technologies 5 Contact Details For further support visit our forum http forum imgtec com Or file a ticket in our support system https ovrsupport imgtec com To learn more about our PowerVR Graphics SDK and Insider programme please visit http www powervrinsider com For general enquiries please visit our website http imgtec com corporate contactus asp PVRTune 45 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Appendix A Hardware Terms Quick Reference Table 8 identifies a list of hardware terms in order to make it easier for the user to interpret the statistics displayed in PVRTune For a more detailed interpretation refer to the help function in the software itself as the statistics will change depending on the hardware Table 8 Quick reference of ha
20. anges cccccceeseeeeeeeeceeeeeeaeeeeeeeceaeeeeaaeseeaaeseeeeeseaeeseaaesteneeees 7 2 5 SAS TA E A E E E E E A A E E E 8 2 5 1 alte lee HE 8 2 5 2 BI E E souwecwbanehs E E E E E E T cuctvaweteenteavancveneetuantandeatbens 8 2 5 3 Neutrino and Klee 8 2 5 4 Command Line EES ssssass inuina ansi aaka aaee ai paaa ada iaai a aiaa 8 2 5 5 R n Nme QOPtONS E 8 3 PV FR CU CT 10 3 1 The BASICS iisstecesnateecbhan dees dE dE ESA 10 3 1 1 OVEIVIOW ech feavs aks esti Eege alee 10 3 1 2 nStallation EE 10 3 1 3 User Interface Layout ac casio sates aisieiined nite A 10 3 1 4 Getting Started with Analysing an Application c ccccceeeeeeeseseeeeeseeeeeneeeeenees 11 3 2 Men TEE 12 3 2 1 SEI 12 3 2 2 Edit Mentgen geseebge eerste ETA AEA EAA A 13 3 2 3 VIE EN 14 3 2 4 Connection MENU sericese ege age EENS 16 3 2 5 Hop MOM i sariaren cites A AAAA 16 3 3 Connection Management sicctevevsccvessaccecersaacdida te cvedons AER aana aaa edian aAa aE aaa 18 3 3 1 Connect to a Target DeVICC ccecceeececeeeeeceeeeeeseeeeeeeeceaeeesaaeseeneeseeeesaeeeenaeennees 18 3 3 2 Connect Using a Broadcasting Genera 18 3 3 3 Connect via LOCAaINOStiseccsscciestecectetvas ccclistecvedans ng ac geed aaaea E aaa 18 3 3 4 Connect Using a Recent Connection cccccceseeeceeeeeceeeeeeeaeeeeneeseeeeeseaeeeeeeeeeeees 19 3 3 5 Connect Using a Saved File 19 3 3 6 Alternative Ways of Connecting ceccceceeeeeeeeee
21. ary gt where lt binary gt is the desired application for profiling 2 5 3 Neutrino and Windows From a command line interface run the PVRPerfServer executable Note Only one client at a time may be connected to an instance of PVRPerfServer 2 5 4 Command Line Options PVRPerfServer supports several command line options as identified next Table 1 Command line options Option h Effect Show help text Show help text disable hwperf Disables the use of PVRScope s hardware performance functionality group N On start up switch the hardware to the specified counter group number port N Network port to use Default is 6520 sendto myfile Instead of using the network record data directly to the specified file t N Time in milliseconds between counter updates The default value is 2 This value can be increased to reduce the CPU usage of PVRPerfServer c N Time in milliseconds between CPU Load updates The default value is 200 pid Gather data for CPU usage memory usage and PID executable name of relevant programs pid N M Gather data for CPU usage memory usage and PID executable name of relevant programs In addition track the specified PIDs periodic 1 0 Enables disables periodic timing tasks for use when recording to a file graphics 1 0 Enables disables graphics timing tasks for use when recording to a file
22. aseceeesaeeceaesaaeeaeesaeeaeesneeaeeeenneaeenenees 39 Revision PowerVR SDK REL_3 5 3649239a 2 User Manual og Imagination Public Imagination Technologies 3 9 Renderstate Override WiNdoOW cccccccccesssecececsneeeeesseeeeeseeeeeesaeeeeesaeeesessaeeeesssneeeeneea 39 3 9 1 OpenGL ES Timing Data 41 3 10 kind Olive Te 41 4 Identifying Bottlenecke annsscnnnnnanannnnn 42 4 1 CPU EMAC E E E E A E 42 4 2 Vertex Linte E 43 4 3 WES VM LUMO DEE 43 4 4 Fragment LIMMO ie ean aa Ea aA a Ea aS ES 44 4 5 Bandwidth Limited EE 44 5 COntact Details isisi aae ee 45 Appendix A Hardware Terms Quick Reference c ccssecceseceesseeeseeesesneeeeseeeseeesesnaeseseenenseees 46 List of Figures Figure 1 Structure of DVD Tune 6 Figure 2 General layout of PYURTUNG GUI 0 0 cece cee ee ene cece ee ceeeeeeeaeeeeeeeeeeeeecaaeeeeaaeseeeeeseaeeesaeeseneeeeaes 11 SIKOR 13 Figure 4 GI MOM genee geegent E ES E E regeert 13 Figure 5 Preferences dialog D sinnsearan anisan innn RAE SENAERA AAEN ARAA eee 14 FIGUIG 6 VICW MOMU A E E A A E A A E A nnd A needed 15 Figure 7 Connection MENUciusrisiininiisainaii aaaea aaia aa debiedannte reste 16 Figure 8 Help Menu EE 17 Figure 9 Connection interface displayed after launching PVRTune GU 18 Figure 10 PVRPerfServer Details dialog box 20 Figure 11 Monitor wimdow seesiacssrn iu easi R T enna ae eee naaeeeeeeaaeee eee 22 Figure 12 Graphical representation of Captured data
23. ble X vewo E views Find ax Name X Le Overview Name 1s Selected View YAis go KE v Overview s e H 20 active 00 ou 00 1000 Bike BR CPU load 465 0 0 366 100 0 H Frame time 2159m 0 0 3351m 100 0 Il Frames per second FPS 463 ou 28 1000 Il GPU clock speed 23647M 0 0 173 96M 100 0 Ge es Il Renderer active 28 1 00 429 1000 7 System memory load 604 ou 59 7 100 0 UR Tiler active 69 ou 107 100 0 GPU Advanced ICH We 20 time per treme osu 00 o 1000 I Renderer time per frame 620m 00 1436m 100 0 x Force 2x2 textures RB Tiler time per frame 148m 00 360m 1000 Force fia frag shades v Renderer able stencil te f UR HSR efficiency 543 og 59 4 100 0 Disable depth te Shade z IR Cycles per pixel 17 00 20 100 cles per vertex l 0 0 f IR cyc 5 p 432 100 1 Processing load pixel 114 00 279 100 0 Il Processing load vertex 05 on 14 1000 RB Shaded pixels per frame 422M 00 955M 100m PR Shaded nivel ner second __ 195 S1M AN 284 m9uM 190 0 __ 635 310 Counter Properties ax jj EE PID 8x Name Frames per second FPS fi Source Overview 11 232ms latency x Index D Name Y Avis frames 20 E aa PVRPerfs What does this counter show H 2 3267 Renderertaskif4 Es LB 2465 This counter represents the average number of frames per second processed sPID 13274 Bi 3714 q 2 by the GPU over the selected period of time This value is the reciprocal of the Renderer frame 501 8 796ms 4 Renderer tasks i
24. blic og imagination Counter Properties ax Name CPU load a gt Source Overview Y Axis 100 000 6 b What does this counter show This counter shows the combined load of all CPU cores C gt What does a high value mean If the CPU load is very high and there are large gaps between the Tiler and Renderer timing blocks then the system is most likely CPU limited If this is the case the load of the Tiler Texturing and Shader Processor units will also be low e g lt 50 In a system where there are multiple CPU cores a value of 100 would indicate that all CPU cores are 100 busy On a platform with a dual core CPU a value of 50 could indicate that one CPU core is 100 busy and the other is idle 0 busy or it could be caused by the cores being at any other combination of loads that result in a 50 average If for example a single threaded application is running on a dual core CPU and the load is 50 the application might be maxing out one of the cores If the value of this counter is very high you should do the following Run a CPU profiler Investigate the issue further with a dedicated CPU profiler Enable API profiling PVRTune can capture timing data for API calls For more information please refer to the PVRTune documentation Figure 28 Counter Properties window Change the Colour of a Counter The user is able to customize the colour of a counter on demand by following these steps
25. der State Override PID Find Remote Editor LISSI se stl E 3 fea Monitor Figure 6 View menu Find To initiate a search click View gt Find Figure 6 This will open the Find window or jump to it for carrying subsequent steps in searching for information see Section 3 9 1 Use Default Counter Y Axis Scales Selecting the Use Default Counter Y Axis Scales option Figure 6 informs PVRTune to use its default Y axis values for counters The information appears in the Y Axis field in the Counter Properties window Insert New Graph Pane This option Figure 6 allows adding a new graph within the selected view inthe Timeline area Restore Default Graph Panes This option Figure 6 causes only the timing data to display in the Timeline area hiding additional graph panes The timing data then occupies the full space within the Timeline area Toggle Single Window Layout To switch to a single window layout for the graph displayed in the Timeline area see Section 3 5 1 toggle the option View gt Single window Layout Figure 6 Toggle Split Window Layout To switch to a split window layout for the graph displayed in the Timeline area toggle View gt Split Window Layout Figure 6 This organises two graph views within the same tab Customize Counter Table Columns To customize the display of columns in the Counter Table window see Section 3 6 1 click View gt Select Columns Figure 6 This will
26. e Revision PowerVR SDK REL_3 5 3649239a 26 User Manual og Imagination Public Imagination Technologies Transfer Tasks Transfer tasks represent the time spent processing tasks related to moving large amounts of memory such as blitting or texture uploading Figure 14 shows an example where a series of transfer tasks are displayed as timing data These tasks are depicted as a series of grey blocks with pink tips Figure 14 Transfer tasks Task Colour Coding The displayed timing data is colour coded for convenience of interpretation Blocks of the same colour represent a single frame see Figure 13 where the colours are recycled every sixteen frames In addition to the general colour of a block its top and bottom tips are also coloured The top tips represent the process ID where different colours indicate different process IDs On the other hand the bottom tips represent the render target where different colours indicate different render targets Figure 15 provides an example of task colour coding used in the display of timing data The core colour of a task represents a single frame Each frame is given a core colour Figure 15a an associated process Figure 15b and a render target Figure 15c Repeated top tip colours refer to the same process ID In Figure 15 the tasks were generated by one process which is indicated through the use of a single colour pale blue across the top tips Repeated bottom tip colours refe
27. e Counter Properties window Figure 28 3 Interpret the detailed information provided Figure 28c Revision PowerVR SDK REL_3 5 3649239a 38 User Manual og Imagination Public Imagination Technologies 3 7 Process ID PID Window The PID window in PVRTune GUI displays a list of the individual connections being used in the recording It only displays the known PIDs relevant to the operation and not all the PIDs running on the system are listed PID Gx Index ID Name Ho 13341 data data com powervr PVRHub bi Bi 13274 2 3267 P 2465 D 3714 H 5 3353 Figure 29 PID window Note If no name can be depicted in the Name column inspect the pid command line option from PVRPerfServer see Section 2 5 4 3 8 Remote Editor Window This functionality allows for real time remote editing of shaders and various numerical value adjustments PVRScope enabled applications can register data as being editable or readable by PVRTune The value of this data can then be passed by PVRScope via PVRPerfServer to PVRTune Any value that has been registered appears in the Remote Editor window in PVRTune GUI Figure 30 Values can be edited using the options provided in the window and the changes are automatically sent back to the application Remote Editor ax SH 0 4616 Item Name Edit Controls EA FragShader Edit VertShader v Edit max variatio 100 min thickne 100
28. e There could be congestion on the network the data is being transmitted on This can be alleviated by using a less congested network or by attempting ad hoc networking e Save data to a file rather than send it over the network see Section 2 5 4 for the sendto command line option for PVRPerfServer e Reduce the data quantity This can be achieved by disabling the periodic or graphics option via the PVRPerfServer command line see Section 2 5 4 Note PVRPerfServer command line options can also be set remotely from PVRTune see Section 2 5 4 Revision PowerVR SDK REL_3 5 3649239a 30 User Manual og Imagination Public Imagination Technologies Figure 21 Example of data loss period Hardware Reset Period Hardware reset periods occur when the hardware locks up and stops responding possibly due to firmware or driver error This usually happens due to the buffer filling up because of the limited storage available A hardware reset can be identified as a vertical pink block Figure 22 V Figure 22 Example of hardware reset period Smart Parameter Management Mode The Smart Parameter Management SPM mode is used to manage parameters when it buffer overflows A Renderer task labelled Renderer SPM Task appears when this mode is activated Figure 23 Renderer SPM task 33 37 PID 97 iFrame 4702 0 788ms Renderer busy Li Le E ice Figure 23 Example of SPM mode PVRTune 31 Revi
29. e profiling support has been removed from the device drivers 2 3 Installation 2 3 1 Package Installation PVRTrace can be installed from the PowerVR SDK installer 1 Navigate to the PowerVR Insider SDK webpage powervrinsider com and download the SDK 2 Launch the installer and follow the on screen instructions 3 Once the SDK has been successfully installed PVRPerfServer can be found within the PVRTune folder in the install directory lt InstallDir gt PVRTuneDeveloper PVRPerfServer lt PLATFORM gt 2 3 2 Android PVRPerfServer functionality is obtained by installing PVRHub for Android To install this application runadb install PVRHub apk on the local machine 2 3 3 Linux With the package successfully installed copy the PVRHub folder to the target device and follow the instructions found in the PVRHub User Manual 2 3 4 Neutrino and Windows With the package successfully installed copy the PVRPerfServer binary to the target device 2 4 Driver Compatibility 2 4 1 Compatible DDK Ranges On PowerVR hardware devices PVRPerfServer is limited to supporting the following driver versions e 1 3 13 1589 onwards e 1 4 14 593 onwards PVRTune 7 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies 2 5 Usage 2 5 1 Android Public QJ imagination With the Android APK installed open the application menu and run PVRHub 2 5 2 Linux From a command line interface run pvr profile lt bin
30. ed PVRTUNE file is opened using the default graph view The information is split into three distinct graphs which capture render timing data and counter data Note If the analysis were based on real time profiling which is dynamic in nature the displayed timing data would appear to be continuously changing with time Name X gt Overview Renderer r ii Smooth 8 Time period 45 7590s Zoom 47616 b c a Figure 12 Graphical representation of captured data 3 5 1 Using the Graph View A number of user actions can be performed when interacting with the graph view These are documented next Zoom in and out of a Graph Zooming in and out of a graph is a useful feature for being able to visualize details of the captured data as well as to have a more general perspective of the trend of the data as a whole One method of zooming in and out is to use the scroll button of a mouse in the graph view Alternatively a zoom value can be specified in the Zoom field Figure 12a The lower the zoom value the bigger the magnification and vice versa PVRTune 23 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Note By default scrolling up decreases the size of the graph whereas scrolling down zooms into it The mouse scroll order can be inverted by using the Preferences dialog box see Section Preferences Smooth out Counters Counters displayed in
31. eeeeeeeeeeecaeeeeaaeseeeeeseaeeesaeeneneeeeaes 19 3 3 7 View Connection Status cccccccccceeceeseceeeeeceaeeeeaaeeeeeeeceaeeesaaesseaeeseeeeescaeeseaeeeenees 19 3 3 8 View and Modify PVRPerfServer Settings 0 c ccccececeeeeeeeeeeeeeeeseeeeeseeeeeeeeeeeees 20 3 3 9 Close a Connection cccceeccecececeeeeeeeeeeeeeeeeeceaeeeeaaeseeeeseaeeesaaeseeaeeseeeesaeeseaeesenes 21 3 4 Monitor Window Performance at a Glance 22 3 5 Working with the Timeline 00 000 eee eececeeeeneeeeeeeeaeeeeeeaaeeeeeeaaeeeeesaaeeeeeeaaeeeseeaeeeeeeaeeeeeeaaes 23 3 5 1 Using the Graph View ccccccecccceeseceeeeeceeeeeaaeeeeneeeeaeeeseaeeesaaeseeeeeseaeeesaeeneaaeseeeees 23 3 5 2 TUPI GD teh EE 26 3 5 3 EV UIC EE 28 3 6 Working with Counters c ccceccceceeeeeeeeeeeeeeeceeeeeceaeeeeaaeseeeeeseaeeecaaeseeaaeseeeeeseaeeesaeeseaeeenaes 32 3 6 1 Select Columns to Display cccccceeeceeeeeeeceeeeeeeeeeeeeeeceaeeesaaeseeneeeeeeeessaeeeeeeesenees 32 3 6 2 Add a Counter to the Graph View 34 3 6 3 Remove a Counter from the Graph View 35 3 6 4 Counter Legend in the Timeline Area ccccccceeececeeneeceeeeeceaeeeeneeseneeescaeeeeeesennees 35 3 6 5 PVRTrace Software Counters nenn nnt 36 3 6 6 Counter Properties 0 ccceccceceeeeeeeeeeeeeeeceeeeeseaeeeeaaesaeeeceaeeesaaeseeaeeseeeesaeseenaeesenees 37 3 7 Process ID PID Wimdow AANEREN 39 3 8 Remote Editor Window ccccccesccccecsecceeeneeceeeeeeeceees
32. end 3 6 5 PVRTrace Software Counters An application is able to send custom software counters to PVRTune via PVRScope When an application is being profiled specifically by PVRTrace the corresponding software counters sent to PVRTune using PVRScope provide data on certain statistics within the running application Enable PVRTrace Software Counters Software counters can be enabled on different operating systems The approach to enabling these counters on Android Linux Neutrino and Windows is as follows e Android PVRHub should be launched On the Options screen ensure that the setting called Enable Software Counters is enabled PVRTrace Recording Libraries must be installed on the device for this to work e Linux With PVRHub installed run Gur profile lt binary gt where lt binary gt is the application that is required to be profiled with software counters active e Neutrino and Windows In the pvrtraceconfig json file on the device ensure that for Profiling the SoftwareCounters and Enabled options are set to true List of PVRTrace Software Counters Table 6 lists the software counters that are available from the PVRTrace libraries Table 6 List of PVRTrace software counters Software counters Indexed draw calls Non indexed draw calls Points no Line no list Line no loop Line no strip Total line no Triangle no list Triangle no strip Triangle no fan
33. ents also known as marks are displayed in the timeline in addition to all of the other information These marks are signified by bars running from the top of the graph to the bottom Figure 12e A full list of the supported events can be found in Section 3 5 3 Use the Counter Legend The Timeline area also contains a Counter Legend section which can be used for several purposes Figure 12f See Section 3 6 4 for more details PVRTune 25 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 3 5 2 Timing Data Timing data can be made visible on a graph by choosing the Render Timing Data option from the action menu displayed when right clicking the graph see Section Change the Graph Rendering Options By default in a multi graph view e g using default view settings the timing data is always displayed in the bottommost graph Figure 13 depicts an example of captured timing data The timing data is arranged on several timelines each with its own label The timeline labelled Tiler represents the Tile Accelerator TA core time which is a measure of the time spent in tiling culling the frame and running vertex shaders The timeline labelled Renderer represents Renderer time and is a measure of how much time is spent fetching textures processing fragment shaders and other fragment processing tasks These timelines represent the two main stages in the Tile Based Deferred Rendering TBDR process Note
34. er When launched and periodically after launch PVRPerfServer broadcasts its existence to the subnet to which it is connected PVRTune GUI lists all the instances of PVRPerfServer broadcasting on the user s subnet in the area of the connection form identified in Figure 9b Connection can be established by selecting the desired target IP address in the list 3 3 3 Connect via Localhost Connecting via localhost implies the establishment of a connection between PVRTune and an instance of PVRPerfServer running on the same device that PVRTune is running on The option for connecting via localhost will be displayed in the connection form if it can be reached Revision PowerVR SDK REL_3 5 3649239a 18 User Manual og Imagination Public Imagination Technologies Note PVRTune uses TCP port 6520 to send data to and from Android devices In order to setup a localhost connection adb forward must be used to forward TCP data across the USB connection To run adb forward enter the following command in a command prompt ddbirorvardi tep EE 3 3 4 Connect Using a Recent Connection Recently used connections and PVRTUNE file paths are listed in the area of the connection form identified in Figure 9c This area also lists recently accessed PVRTUNE files if any To connect using a recently used connection select the desired connection from the list 3 3 5 Connect Using a Saved File A saved PVRTUNE file can be loaded by clicking File gt Open and b
35. g graphs can be added and removed on demand This procedure is highlighted in Section 3 2 1 Save a Graph as an Image PVRTune GUI comes with options to allow graphs to be saved as images There are different modes for saving a graph as an image and the procedure is next highlighted 1 Right click a graph in the Timeline area This will open an action menu 2 From the menu select the desired save settings which are one of Save Image with HUD saves foreground content when mouse hovering over a graph Save Image saves image in the same location as the binary and Save Image To Saves image to a user defined location Pause the View during Connected Analysis During real time profiling the graph view continuously changes with time since the collected data is dynamic To pause the view at a given time perform the following steps 1 Right click a graph in the Timeline area This will open an action menu 2 Select the option Pause View Add and Remove Counters to Graph Counters can be added on demand to the timeline by selecting them from the Counter Table window and dragging them to the graph This is covered in more detail in Section 3 6 2 Note Some counters when added to a graph view may not be immediately visible as a result of very low or very high Y axis values This can be overcome by adjusting the Y axis scaling of the counter see Section Change the Y Axis Scaling of a Counter View Events Several ev
36. itecture Guide for a more detailed understanding of the PowerVR hardware architecture Revision PowerVR SDK REL_3 5 3649239a 46 User Manual og imagination Public Imagination Technologies Imagination Technologies the Imagination Technologies logo AMA Codescape Ensigma IMGworks 12P PowerVR PURE PURE Digital MeOS Meta MBX MTX PDP SGX UCC USSE VXD and VXE are trademarks or registered trademarks of Imagination Technologies Limited All other logos products trademarks and registered trademarks are the property of their respective owners PVRTune 47 Revision PowerVR SDK REL_3 5 3649239a
37. l time and can be plotted against the timeline to give a graphical representation of their change over time Counters can also be added during offline analysis To add a counter to a graph view perform the following steps 1 Identify the counter that is of interest from the Counter Table window Figure 24 2 Drag and drop the counter from the Counter Table window to an existing graph present in the Timeline area This results in the addition of new plotted data on the graph Figure 26 shows an example where the Frames per second counter represented by the line graph has been added to a graph view with rendered timing data 2D Renderer Figure 26 Adding a counter to a graph view Note A new graph can be created by dragging and dropping a counter to a graph placeholder available in the Timeline area A graph placeholder is the narrow area of space located above or below a graph view An example is shown in Figure 27 a gt 2D Renderer Figure 27 Creating a new graph by adding a counter to a graph placeholder Revision PowerVR SDK REL_3 5 3649239a 34 User Manual og Imagination Public Imagination Technologies 3 6 3 Remove a Counter from the Graph View Counters plotted in the graph view can be removed when they are no longer required To remove a counter from the graph view perform the following steps 1 Inthe Timeline area identify the counter that needs to be removed
38. ls are processed by the driver This mode of operation shows timing data being pulled out of PVRTrace recording libraries visualised as a graph displaying the draw calls as well as EGL swap buffer operations This is configurable which means that it is possible to see all or specific EGL calls By being able to record OpenGL ES timing data this way you can gain further insight into how for example work is cascading up through the graphics pipeline from the driver to the CPU This can consequently help you isolate any bottlenecks you might be experiencing 3 10 Find Window PVRTune GUI comes with a search capability which is helpful for quickly finding information The Find window Figure 32 allows the user to input search terms and the retrieved results are narrowed down using an incremental search approach A number of items can be searched this way including activities eg Renderer marks e g abc counters eg triangles per frames e g 40000 PIDs eg 4242 time as per the X axis in the graph view e g 1200 and time range by separating two numbers with a pair of dots e g 1200 1 1200 2 Find gx Shader v Wei 9 matches View all Counter Overview Shader processing load a Counter GPU Advanced Shader Cycles per pixel _ Counter GPU Advanced Shader Cycles per vertex i Counter GPU Advanced Shader Processing load Counter GPU Advanced Shader Processing load Counter GPU Advanced Shader Shaded pixels
39. lved in these two approaches are summarised next Connected Analysis Connected analysis is the assessment of real time data from a target device connected using PVRPerfServer The core steps to performing a connected analysis are as follows Ensure that the target device is using the PowerVR device drivers Boot the target device Install PVRPerfServer on the target device If necessary initialise the PowerVR device drivers Run PVRPerfServer If successful PVRPerfServer outputs the server name IP address and port number 6 Run PVRTune GUI on the host machine It is important that both PVRTune and PVRPerfServer have matching versions 7 If PVRPerfServer is on the same subnet as PVRTune the target device will be broadcast in PVRTune GUI see Section 3 3 2 If this is not the case then manually enter the IP address of the server see Section 3 3 1 OV ee N PVRTune 11 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 8 Once successfully connected identify the bottleneck in the application by analysing the output of PVRTune for more information see Section 4 9 Attempt to resolve the bottleneck For more information on optimization techniques see the document entitled PowerVR Performance Recommendations 10 Repeat the necessary steps with the newly optimized application until performance is at the required level or no further bottlenecks can be identified Offline Analy
40. n viewing averaged values over the range see Section 3 6 1 Change the Graph View Layout The layout of the graph view can be quickly changed to a single or split view layout This is carried out by using the Single View Layout and Split View Layout options from the View menu see Section 0 Change the Graph Rendering Options During an analysis task certain types of data can be shown or hidden on demand in the graph view To change the data render options perform the following i Right click a graph in the Timeline area where the change needs to be made This will open an action menu 2 From the menu select or remove the options Render 25 50 75 Quarters Render Timing Data or Render Marks as appropriate The change will be reflected on the graph Revision PowerVR SDK REL_3 5 3649239a 24 User Manual og Imagination Public Imagination Technologies Restore the Default Graph View After changes to the graph view have been made it is possible to revert to default settings if required To achieve this perform the following Right click a graph in the Timeline area This will open an action menu 2 Select the option Restore Default View Close a Graph in a Multi Graph View In a multi graph view as in Figure 12 it is possible to close one of the graphs by clicking its corresponding button to close it Figure 12d To display the graph again restore the graph view to default Add and Remove Tabs Tabs displayin
41. ne has been receiving data or was connected if the current status is disconnected Count The number of times PVRTune has performed an action related to its current status e g if PVRTune is receiving data then the count indicates the number of times it has received data from PVRPerfServer Count per second The rate at which the count value moves per second Gap The average time period between actions as measured by the count value Receive rate The rate at which PVRTune is receiving data PVRTune 19 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Data Description DCS The percentage progress towards automatic disconnect due to memory usage after PVRTune has recorded 1GB of data it disconnects from PVRPerfServer Group The currently active counter group 3 3 8 View and Modify PVRPerfServer Settings Figure 10 illustrates the PVRPerfServer Details dialog box which is accessed by selecting Connection gt PVRPerfServer gt Details from the Menu bar The dialog box can be used to view the current PVRPerfServer information and modify certain settings In other words when PVRTune is connected with PVRPerfServer the dialog box allows PVRTune to remotely control the counter groups that are currently active on PVRPerfServer This in turn then activates or deactivates certain counters on the PowerVR hardware p E PVRPerfServer Details Lin
42. nual for PVRTune and PVRPerfServer It includes installation instructions functionality explanations and useful tips on how to make the most out of the applications 1 2 Software Overview PVRTune is a performance analysis tool for the PowerVR graphics cores It uses driver level counters and hardware registers to provide real time data on the performance of an application running on a PowerVR graphics core PVRTune receives performance data from either PVRPerfServer or PVRHub running on the target device Figure 1 illustrates the conceptual structure of PVRTune in relationship to the interactions taking place at the client software device software and device hardware levels The following concepts are relevant to the figure e PVRPerfServer This is an application which runs on the target device and reads counters and registers from the driver and also optionally from any application which uses the PVRScope library and either saves the data to a PVRTUNE file or transmits the data to PVRTune over a network e PVRHub This is an application which runs on Android or Linux and allows an instance of PVRPerfServer or PVRTrace to run on the target device For the remainder of this document all PVRPerfServer functionality can be accomplished using PVRHub unless otherwise stated e PVRScope This is a performance analysis library that retrieves graphics core counter data and works alongside PVRTune to augment it with custom markers and counters
43. og Imagination Public Imagination Technologies PowerVR by imagination PVRTune User Manual Copyright Imagination Technologies Limited All Rights Reserved This publication contains proprietary information which is subject to change without notice and is supplied as is without warranty of any kind Imagination Technologies and the Imagination Technologies logo are trademarks or registered trademarks of Imagination Technologies Limited All other logos products trademarks and registered trademarks are the property of their respective owners Filename PVRTune User Manual Version PowerVR SDK REL_3 5 3649239a External Issue Issue Date 22 Jul 2015 Author H Imagination Technologies Limited PVRTune 1 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Contents 1 Ge TT e iness dete ces soca deie eh ce cea debe ceed shee ene donee ed ch ance donecsd dhebesnvobedenesbencssviebedests 5 1 1 Document OVEIVICW cccceceeeceeeeeeeeaeeeeaaeceeeeeceaeeecaaeeeeaeeseaeeesaaeeesaaeedeaeeseeeeseaeessaeeeeneeteaes 5 1 2 Software Cvermlew tatt tnnt tnn tnnt nnntn ntun Ennen unnn en nennen nenn 5 2 PV RPGH ET E 7 2 1 EEN H 2 2 IR ln 7 2 3 Installation ene eave dee needed ees eee ele ce ede 7 2 3 1 Package Installation EE 7 2 3 2 PANNING EE 7 2 3 3 BI E TE 7 2 3 4 Neutrino and WinGOWwS EE 7 2 4 Driver Compatibility actui iniisa oaea aaee a aaa paa adaa SEET 7 2 4 1 Compatible DDK R
44. on management capability of PVRTune provides a range of options to facilitate the task of connecting Upon launching PVRTune the connection form is displayed as shown in Figure 9 Notice that the various windows and options used for analysis purposes are disabled at that point until a valid connection is made or a PVRTUNE file is loaded E PvatuneDeveloper 14 104 So File Edit View Connection Help Counter Table ax Find ax Enter a server name or IP address a _ ov Te Broadcasting servers on subnet Shortcuts i Android USB device E pvrtune focalhost Render State Override ax Counter Properties ax Note in order to connect to an Android device over USB you wil led to use adb forward tcp 6520 tcp 6520 Monitor Remote Editor PID Not connected b c Figure 9 Connection interface displayed after launching PVRTune GUI Note The connection form can be customized as a pop window and can also be assigned an individual tab in the Timeline area see Section Preferences 3 3 1 Connect to a Target Device To connect to a target device perform the following steps 1 Enter an IP address or IP resolvable name of the target device into the Connect to box Figure 9a 2 Click the Go button to start the connection If an instance of PVRPerfServer is found PVRTune will connect to the target device 3 3 2 Connect Using a Broadcasting Serv
45. open a dialog box to allow the selection of the required columns see Section 3 6 1 PVRTune 15 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Refresh PVRTune GUI To refresh PVRTune GUI click View gt Refresh Figure 6 Show or Hide Counter Table Window To show or hide the Counter Table window toggle View gt Counter Table Figure 6 Show or Hide Counter Properties Window To show or hide the Counter Properties window see Section 3 6 6 toggle View gt Counter Properties Figure 6 Show or Hide Renderstate Override Window To show or hide the Renderstate Override window see Section 3 9 toggle View gt Renderstate Override Figure 6 Show or Hide PID Window To show or hide the PID window see Section 3 7 toggle View gt PID Figure 6 Show or Hide Find Window To show or hide the Find window see Section 3 9 1 toggle View gt Search Figure 6 How or Hide Remote Editor Window To show or hide the Remote Editor window see Section 3 8 toggle View gt Remote Editor Figure 6 Show or Hide Monitor Window To show or hide the Monitor window see Section 3 4 toggle View gt Monitor Figure 6 3 2 4 Connection Menu Figure 7 illustrates the Connection menu which is used for connection management purposes A number of options are available from this menu and these are covered in further detail in Section 3 3 Connection Help New Ctri
46. pe 4 T r Figure 32 Find window Note Counters can be dragged and dropped from the Find window to the graph view in order to generate new counter plots PVRTune 41 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 4 Identifying Bottlenecks The ability to identify bottlenecks is vital when analysing the performance of an application Section 3 4 and Section 3 9 have previously explained two of the features of PVRTune for monitoring graphics core workloads and renderstate respectively in the effort to help identify some types of bottlenecks This section explains how to identify bottlenecks using the graph view The bottlenecks usually fall into one of five categories e CPU limited e Vertex limited e V Sync limited e Fragment limited e Bandwidth limited 4 1 CPU Limited A CPU limited application is often identifiable as an application suffering from poor performance or frame rate even though the graphics core usage is not high In PVRTune this can be very easily identified since CPU limited applications have a CPU load that is either at or near one hundred percent Figure 33a or is widely variable c a b Figure 33 Identifying CPU limited applications Other identifying factors include gaps in the Shader load caused by the PowerVR hardware going to sleep while waiting for further instructions Figure 33b and regular visible gaps between frame
47. r to the same render targets In the example three render targets are present and these are indicated by the three colours used for the bottom tips of the tasks In double or triple buffered situations the render targets are different for each frame alternating between each of the back buffer targets PVRTune 27 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination b Renderer c Figure 15 Example of task colour coding Driver Timing Data Driver timing data becomes relevant when PVRTune is used in combination with PVRTrace The data is displayed in its own timeline and is represented as a series of tasks labelled with the thread ID that called the driver An example of driver timing data is provided in Figure 16 Thread 15411 Figure 16 Example of driver timing data Note By default only the most expensive calls are displayed e g g1DrawElements glReadPixels shader compilation and texture uploads For a more verbose output open PVRHub select Options and setAPI Function Timing Events to Verbose Warning Setting API Function Timing Events to Verbose may affect performance 3 5 3 Event List Active Counters Changed This event represents the point at which the active counter group has been changed using some custom hardware counter group The event appears as a vertical grey line Figure 17 Revision PowerVR SDK REL_3 5 3649239a 28 User Manu
48. rdware terms Term Description Tiler The Tiler is responsible for processing the vertices and tiling This includes transforming vertices culling clipping and storing the data The timing data contains the TA and Shader processing operations Renderer The Renderer is responsible for processing the fragments within a given tile and the timing data contains all the ISP TSP and fragment processing operations TA The Tile Accelerator TA takes geometric data that has been transformed by the shader processor as input and clips projects and culls it The TA Core timing data contains vertex processing and tile setup TSP The Texture and Shading Processor TSP performs interpolation and schedules tasks for the shader processor and texture data pre fetches ISP The Image Synthesis Processor ISP is responsible for per tile Hidden Surface Removal HSR to ensure that the fragments processed by the TSP are only those that will affect the rendered image Shader The PowerVR shader processor is a flexible multi threaded processor capable of executing vertex fragment graphics core instructions and issuing memory access operations such as texture fetches Compute Series6 only The Compute Core is dedicated to compute tasks issued via the Compute Data Master CDM Compute timing data includes the programmable arithmetic handled by the shader processor Note Refer to the PowerVR Hardware Arch
49. ride WINdOW ccccccceeeeceeseeeeeeeeceaeeeeaaeeeeneeseaeeesaaeeesaaeseeeeeseaeeseaeeseaeeeeaes 40 Figure 32 Find Kittel VE 41 Figure 33 Identifying CPU limited applications ccecceeeseceeeeeeeeeeeeeeeceeeeeseaeeeeaaeseeeeeseaeeesaeeseeeeeaes 42 Figure 34 Identifying vertex limited applications 20 2 2 ee eeeeeeeeeeeeeeeene ee ee eae eset eaaeeeeeeaeeeeeeaeeeeeeaeeeeneea 43 PVRTune 3 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Figure 35 Identifying fragment limited applications ee eeceeeeeeeeeeeeeeeeeeeeeaeeeeeeaaeeeeesaeeeeeeaeeeeeeaa 44 List of Tables Table 1 Command line ele TE 8 Table 2 Run time NOtk6yS ive cscs isa a N E sacar taeedineeeesteeeehiaetleds 9 Table 3 Connection status information in the Status bar 19 Table 4 Information captured in the PVRPerfServer Details dialog box 20 Table 5 Summary of columns and their descripilon 33 Table 6 List of PVRTrace Software counters eee cceeeecneeeeeenneeeeeeaeeeeeeaaeeeeeeaaeeeeeeaaeeesesaaeeeeeeaeeeeneaaes 36 Table 7 Explanation of renderstate Override Options ceeececeeeeeeeeeeeeeeeeeeeaeeeeeeaaeeeeeeaeeeeeeaeeeeneaaes 40 Table 8 Quick reference of hardware tems nnu nnnnnntnnnnn nnan nenn 46 Revision PowerVR SDK REL_3 5 3649239a 4 User Manual og Imagination Public Imagination Technologies 1 Introduction 1 1 Document Overview The purpose of this document is to serve as a complete user ma
50. rowsing to the required file see Section Open a File 3 3 6 Alternative Ways of Connecting It is possible using the Connection menu see Figure 7 to control some aspects of connection management These are Establish a new connection Select the New option in the Connection menu to open a dialog box for inputting the IP address or IP resolvable name This is essentially the same procedure discussed in Section 3 3 1 e Connect via localhost Select the localhost option in the Connection menu to connect via localhost This is essentially the same procedure discussed in Section 3 3 3 e Connect using a recent connection Select the Recent Connections option in the Connection menu to view a list of previously accessed connections and choose the desired connection from the list This is essentially the same procedure discussed in Section 3 3 4 3 3 7 View Connection Status Data related to the connection status of PVRTune is displayed in the Status bar area of PVRTune GUI see Figure 2j Table 3 lists the details of the data that is displayed from left to right in the Status bar An example of connection status data is Receiving 11 2640s 56607 4922 58ps 0 000sp 337 9KB s 0 7 0 Table 3 Connection status information in the Status bar Data Description Status This indicates the current status of PVRTune The options available are Connecting Receiving and Disconnected Time This indicates for how long PVRTu
51. s are set by clicking Edit gt Preferences inthe Menu bar This opens the Preferences dialog box Figure 5 which displays various options to enable the customization of PVRTune GUI The Preferences dialog box contains several options e Background This option is available under the Graph section of the Preferences dialog box and allows the background colour of graph inthe Timeline area to be changed from the default setting to either light or dark in order to make it easier to depict lines PVRTune 13 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination Draw Marks across all timelines This option is turned off by default but can be toggled if necessary to show up marks in the Timeline area e Hide empty timelines Toggle this option to show or hide empty timelines in the Timeline area e Zoom control invert mouse wheel This option is available under the Graph section of the Preferences dialog box The mouse wheel can be used to zoom in and out of a graph in the Timeline area With the option unchecked i e set to default scrolling down zooms into the area while scrolling up zooms out of it This default setting can be inverted by ticking the Zoom control invert mouse wheel checkbox e Integrate connection form This option is available under the User Interface section of the Preferences dialog box Unchecking this option enables the user to make the connection form see Section 3 3 a pop up
52. s when displaying timing data Figure 33c Revision PowerVR SDK REL_3 5 3649239a 42 User Manual og imagination Public Imagination Technologies 4 2 Vertex Limited Vertex limited applications are applications where the bottleneck comes from processing either large amounts of vertices per frame or from the use of a complex vertex shader or both This can be identified by large gaps between Renderer tasks Figure 34a while there is little or no gap between Tiler tasks Figure 34b a b Figure 34 Identifying vertex limited applications Further information can be gained from the Processing load Vertex andthe Tiler load counters If Tiler load is high Figure 34c but Processing load Vertex is not then the scene has too many vertices in it and the cost is coming from the tiling process On the other hand if Processing load Vertex is high Figure 34d but Tiler load is not then the bottleneck is likely to be in the vertex shader 4 3 V Sync Limited Vertical Synchronisation V Sync is a display option that forces an application to synchronize graphical updates with the update rate of the screen This causes some frames to be slightly delayed and enforces a maximum refresh rate but reduces screen tearing and can save power V Sync limited applications are often characterized by intermittent gaps between frames in the graph view and the frame rate appears to be limited at a set maximum value If possible V s
53. sion PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 3 6 Working with Counters Counter information is listed in the Counter Table window of PVRTune GUI An example of listed counters is shown in Figure 24 The counters are organised under different headings eg Overview GPU Advanced etc which can be expanded or collapsed as needed During connected analysis the table also displays the counters available within specific counter groups In the example the Group dropdown box is disabled Figure 24a which is normal during offline analysis Note For a full list of default PVRTune counters and their description refer to the document entitled PVRTune Counter List and Description Counter Table a x a gt Group Name is Selected View Y Axis v Overview H 2D active 0 0 0 0 0 0 100 0 E CPU load 46 5 0 0 36 9 100 0 BW Frame time 21 59m 0 0 39 65m 100 0 IB Frames per second FPS 46 3 0 0 25 2 100 0 E GPU clock speed 236 47M 0 0 205 48M 100 0 BB Renderer active 28 7 0 0 29 9 100 0 I System memory load 60 4 0 0 60 0 100 0 I Tiler active 6 9 0 0 64 100 0 v GPU Advanced I 2D time per frame 0 99u 0 0 41 89p 100 0 P Renderer time per frame 6 20m 0 0 11 85m 100 0 RB Tiler time per frame 148m 0 0 252m 100 0 v Renderer BB HSR efficiency 54 3 0 0 49 3 100 0 v Shader BB Cycles per pixel 17 0 0 2 2 100 0 BB Cycles per vertex 42 5 0 0 43 8 100 0 _ Proce
54. sis It is possible to analyse an application without being directly connected to PVRPerfServer This requires a PVRTUNE file of the target application to have been created prior to analysis This approach is particularly useful when large amounts of data are being lost due to network load or high CPU usage on the client machine A PVRTUNE file can be created either through the use of the sendto command line parameter for PVRPerfServer or by saving the file of an existing tune using PVRTune GUI see Section Save a File Note PVRPerfServer only saves counter data for a single counter group when using the sendto parameter By default the group is set to zero but this can be changed using the group command line parameter With a PVRTUNE file created and copied across to an accessible location perform the following steps i Identify the version of PVRTune and PVRPerfServer used to create the file 2 Run the version of PVRTune that matches the identified PVRPerfServer version 3 Open the file in PVRTune GUI see Section 3 3 5 PVRTune GUI will then display the tuning data 4 Identify the bottleneck 5 Attempt to resolve the bottleneck For more information on optimization techniques see the document entitled PowerVR Performance Recommendations 6 Repeat the necessary steps with the newly optimized application until performance is at the required level or no further bottlenecks can be identified 3 2 Menu Bar 3
55. ssing load pixel 114 0 0 178 100 0 id Processing load vertex 0 5 0 0 0 8 100 0 BB Shaded pixels per frame 422M 0 0 789M 1 00m I Shaded pixels per second 195 51M 0 0 198 94M 100 0 Shaded vertices per frame 671K 0 0 18 89K 100 0 I Shaded vertices per second 311 03K 0 0 476 51K 100 0 v Processes v PID 13274 0 H 20 active 0 0 0 0 0 0 100 0 BB 2D time per frame 0 0 0 0 0 0 100 0 BW Frames per second FPS 23 4 0 0 24 6 1000 Figure 24 Counter Table window 3 6 1 Select Columns to Display By default counter data are tabulated following a similar format as shown in Figure 24 Additional data sets can be customized for display by right clicking the Counter Table window and choosing the Select Columns option as shown in Figure 25 Revision PowerVR SDK REL_3 5 3649239a 32 User Manual og Imagination Public Imagination Technologies a E PvetuneDeveloperGUI Select which columns will be visible in the counter list Source 0 1s 0 5s dis 2s 4s 8s 10s 16s 32s Highlight V Selected Line Task Frame V Y Axis e Figure 25 Dialog box for selecting and deselecting columns Table 5 identifies a summary of the additional columns that can be selected and deselected on demand Table 5 Summary of columns and their description Column Description 0 1s to 32s These values represent time frames The columns show the average value of counters over the most recent timespan e g the
56. the timeline can be smoothed out by specifying a value in the Smooth field Figure 12b The maximum value is 80 Increasing the value will smooth out the counter and the counter plots become easier to interpret View Captured Data over the Entire Time Period To adjust the graph view such that the captured data is displayed for the entire time period perform the following 1 Right click a graph in the Timeline area This will open an action menu 2 From the menu select the option called View Selection All Note The value of the total time period is specified at the bottom of the graph view Figure 12c View the Start of the Captured Data To adjust the graph view such that the start of the captured data is shown perform the following 1 Right click a graph in the Timeline area This will open an action menu 2 From the menu select the option called View Earliest View the End of the Captured Data To adjust the graph view such that the end of the captured data is shown perform the following Right click a graph in the Timeline area This will open an action menu 2 From the menu select the option called View Latest Select a Time Range A specific region can be selected on a graph during analysis This can be achieved by using the Ctr1 key while left clicking and dragging on a graph To deselect an already selected region press Ctr1 and left click anywhere on the graph The selection of a time range is relevant whe
57. ux version 3 4 5 997834 se infra 50210 04 gcc version 4 6 x google 20120106 prerelease GCC 1 SMP PREEMPT Mon Jul 1 18 17 12 KST 2013 Driver Name 1 10 220 9507 Device Variant PowerVR SGX544MP3 Device Series PowerVR SeriesSXT Operating System Server Description PVRPerfServerDeveloper v14 72 Server Build 3 3 2797549 Server Connection Time local Tue Mar 11 11 15 44 2014 Server Connection Time UTC Tue Mar 11 11 15 44 2014 PVRPerfServer and PVRTune version mismatch Server version PVRPerfServerDeveloper v14 72 Expected PVRPerfServerDeveloper v14 104 Figure 10 PVRPerfServer Details dialog box Note Any warning encountered in the running of PVRPerfServer and PVRTune is also captured in the dialog box Table 4 provides a description of the fields present in the PVRPerfServer Details dialog box Table 4 Information captured in the PVRPerfServer Details dialog box Field Description Group Active counter group on PVRPerfServer Different counter groups can be selected as appropriate to change their settings Sample Time ms Time in milliseconds between the hardware samples taken by PVRPerfServer CPU Load sample time ms Time in milliseconds between the CPU Load samples taken by PVRPerfServer Revision PowerVR SDK REL_3 5 3649239a 20 User Manual og Imagination Public Imagination Technologies Field Description Periodic Data Enable When this option is ticked PVRTune
58. window e Hide unnecessary tabs This option is available under the User Interface section of the Preferences dialog box and if unchecked an individual tab is assigned for the connection form inthe Timeline area This facilitates quick access to the connection form after a connection is made or with a file loaded a SY Preferences Graph views Background Default Draw Marks across all timelines v Hide empty timelines Graph controls Zoom control invert mouse wheel User Interface V Integrate connection form V Hide unnecessary tabs _ Cancel Figure 5 Preferences dialog box 3 2 3 View Menu Figure 6 illustrates the View menu The menu provides options for finding information as well as showing or hiding the various windows present in the PVRTune GUI thereby allowing for workspace customization Note The workspace can also be customized by dragging and dropping the individual dockable sections to either the right or left hand sides of the interface or as standalone windows Revision PowerVR SDK REL_3 5 3649239a 14 User Manual og Imagination Public Imagination Technologies Connection Help Q Find Ctrl F Use Default Counter Colours Use Default Counter Y Axis Scales Insert New Graph Pane Restore Default Graph Panes Single Window Layout D D Split Window Layout Select Columns Ctrl N si Refresh F5 Counter Table Counter Properties Ren
59. ync should be disabled when profiling an application as it adds noise to the PVRTune output and this makes it more difficult to diagnose where optimization work could be beneficial or if completed optimization has been successful PVRTune 43 Revision PowerVR SDK REL_3 5 3649239a Imagination Technologies Public og imagination 4 4 Fragment Limited Fragment limited applications are very common and occur in most scenes that have fewer vertices than the number of pixels in the framebuffer Fragment limited applications can be identified when there is the presence of no gaps between Renderer tasks Figure 35a large gaps between Tiler tasks Figure 35b a high value of Processing load Pixel Figure 35c or high Shader clock cycles per pixel Figure 35d if c d NN sis if d vi I CARRA a b Figure 35 Identifying fragment limited applications 4 5 Bandwidth Limited Cases of bandwidth limited applications are both hard to visualize and identify as they may appear as other bottlenecks Programs may be bandwidth limited if e Timeline shows the application to be fragment limited but the Processing load Pixel is low e Timeline shows the application to be vertex limited but the Processing load Vertex and Tiler load are low Other instances of bandwidth limitation may occur For example if many units are accessing memory simultaneously then the available system memory bandwidth limits c

PVRTune User Manual Developer

Contents

Download Pdf Manuals

Related Search

Related Contents