IRIX Base Documentation 2001 May

home *** CD-ROM | disk | FTP | other *** search

/ IRIX Base Documentation 2001 May / SGI IRIX Base Documentation 2001 May.iso / usr / relnotes / performer_eoe / ch4.z / ch4

Wrap

Text File | 2001-04-17 | 27KB | 727 lines

- 1 - 4. _K_n_o_w_n__P_r_o_b_l_e_m_s__a_n_d__W_o_r_k_a_r_o_u_n_d_s This chapter lists the problems with the OpenGL Performer libraries and with the shared memory configurations. Note: This list is specific to OpenGL Performer for IRIX systems. Known problems with OpenGL Performer for Linux can be found in the Linux release notes, installed in /usr/doc/performer-2.4/ (Linux only). 4.1 _IIII_RRRR_IIII_XXXX _6666_...._5555 In IRIX 6.5 through IRIX 6.5.4, a change in the semantics of sginap() can cause frames to be missed if the application is running in FLOAT or LOCK phase and finishing more than a full field early. The sginap() call is used by OpenGL Performer specifically in the case where the application finishes more than a full field ahead of schedule. This does not affect applications running at field rate (ie. 60 Hz) or applications running at their requested rate. This was fixed in the IRIX 6.5.5 update. A workaround is to give the program a realtime scheduling priority using the npri program (which must be run as root or with the CAP_SCHED_MGT capability): # npri -s RR -r <any number 0-255> <program> <args> or: # npri -s FIFO -r <any number 0-255> <program> <args> To allow users to run npri without full root permissions, you can give the npri program the set-capability-on- execution attribute: # chcap CAP_SCHED_MGT+eip /usr/sbin/npri If you later want to change it back: # chcap all= /usr/sbin/npri See the _n_p_r_i(1) and _c_h_c_a_p(1) references pages for more information. - 2 - _4._2 _GGGG_eeee_nnnn_eeee_rrrr_aaaa_llll _IIII_RRRR_IIII_XXXX _IIII_ssss_ssss_uuuu_eeee_ssss +o _6._x _O_S _b_u_g _w_i_t_h _s_h_a_r_e_d _a_r_e_n_a _p_l_a_c_e_m_e_n_t: An obscure bug in IRIX 6.2-6.4 with default placement of the shared arena can cause programs to die due to lack of heap space for malloc. Typcially the program will die during X or GL initialization with a message like: PF Warning/Internal(12): pfWindow::openNewNoPort() - null visual Xlib: sequence lost (0x10002 > 0x10) in reply type 0x0 X Error: 0 Request Major Code 0 () Error Serial #2 Current Serial #16 Running the program through the unix system call tracer (par): par -s -i -SS prog options will show messages like: 966mS[ 1]perfly_ogl( 4781): brk(0x10086e58) errno = 12 (Not enough space) shortly before the death. Workaround: through the environment variable PFSHAREDBASE or in with pfSharedArenaBase() before pfInit(), explicitly set the address of the arena so that it is not too close to the heap and does not collide with other DSOs. An address that has worked with programs exhibiting the problem that are very close to perfly is: 0x18000000. ex: pfSharedArenaBase(0x18000000); or setenv PFSHAREDBASE 0x18000000 If this address is not sufficient, run again with par and the following environment variables set: setenv _RLD_PATH /usr/lib/rld.debug; setenv _RLD_ARGS -v - 3 - and redirect all output to a file. Look for the addresses of the mmap() calls and of address of DSO load actions to find a free address range in which to place the arena. Compiling with some libraries statically can make this operation easier. +o Using 7.2 Compilers you may encounter 'Unresolved symbol ___rcis'. Install patch 2483 (or its successor) - "7.2 compiler runtime environment for IRIX 6.2/6.3/6.4" to fix the problem +o The XSGIvc library on IRIX 6.2 did not support 64bit operation before patch 1808 and so functionality requiring XSGIvc, such as dynamic video resize, are currently not enabled when running 64bit. +o IRIS GL is not recommended on any OpenGL native machines such as O2, Octane, IMPACT graphics or InfiniteReality -- OpenGL operation should be used instead. +o Overflowing texture memory when using Clipmapping can cause texture corruption. This is a bug in OpenGL in InfiniteReality. If you try to run cliptex -2 mof.ct (forcing both 2 windows and 2 cliptextures), you will may get garbage texture and extremely slow (even locked up) performance. This is due to a known bug in the iR OpenGL: cliptextures can't be swapped in texture memory. It will also happen with perfly, and with cliptex -3 hl.ct if you partially fill the texture memory with other running programs first. So if you overfill your texture memory with cliptextures, you will see this behavior. +o When linking executables with Performer libraries, an alarming number of warnings about symbols that are defined in multiple .so's may be printed. These are harmless but the appropriate compiler option to disable the warning has not yet been identified. +o On RealityEngine graphics, when more than 512 textures are used the hardware and or host-side software may become confused and in some cases, may falter completely and terminate the application. This is a graphics library limitation: IRIS Performer has no limit to the number of textures. +o On other graphic systems than RealityEngine and InfiniteReality, the video rate used by IRIS Performer for frame rate control is computed at application startup on some machines and is not always exactly - 4 - correct. The problem is greatly aggravated when the system's load average is at least one process per cpu, so it is best to start a Performer app when the system is otherwise quiet. If this is a problem, the video rate can be set with _p_f_V_i_d_e_o_R_a_t_e or run locked to a processor by using _p_f_u_L_o_c_k_D_o_w_n_P_r_o_c (the -L option in perfly). +o Although IRIS Performer supports it, the OpenGL rendering model does not support different material color modes for front and back materials. When Performer encounters this case, it disables (ignores) the material color mode for the back material when using OpenGL. 4.3 _IIII_RRRR_IIII_XXXX_6666_...._2222 _aaaa_nnnn_dddd _6666_...._4444 _aaaa_nnnn_dddd _NNNN_3333_2222_////_6666_4444_bbbb_iiii_tttt +o _V_i_d_e_o _c_h_a_n_n_e_l _c_o_n_t_r_o_l _a_n_d _d_y_n_a_m_i_c _r_e_s_i_z_e: The XSGIvc X11 extension for video control did not support 64 bit operation in IRIX 6.2 and 6.4. This is fixed in IRIX 6.5. IRIS Performer 2.2.2 will support support video channel control under 64bit operation on IRIX 6.5 and later. +o _S_t_a_t_i_s_t_i_c_s: The PFSTATSHW_CPU pfStats statistics class for accumulation of statistics on CPU usage are not yet implemented for IRIX6 operation (64-bit or 32-bit). 4.4 _IIII_nnnn_ffff_iiii_nnnn_iiii_tttt_eeee_RRRR_eeee_aaaa_llll_iiii_tttt_yyyy _G_a_n_g _S_w_a_p pfChanCalligEnable() [perfly -U] will not set up GangSwap correctly in multipipe configurations. The visible result is a 1Hz frame rate, even without database [the same thing happen if you start GangSwap and you do not have the SwapReady cables linking the pipes together]. You will have to force the GangSwap using PFCHAN_SWAPBUFFERS_HW attribute in addition to pfChanCalligEnable() [perfly -G -U ...] +o _V_i_d_e_o _c_h_a_n_n_e_l _c_o_n_t_r_o_l _a_n_d _X: X primitives will not be knowledgeable of video resizing operations. This means that on a resized video channel, the X cursor will not be drawn in the proper location or of the proper size, and X menus may be unusable. +o _V_i_d_e_o _c_h_a_n_n_e_l _c_o_n_t_r_o_l _a_n_d _m_u_l_t_i_p_l_e _a_p_p_l_i_c_a_t_i_o_n_s: On InfiniteReality, only one application may be bound to a - 5 - video channel. If, for example, two copies of perfly are running, only one may do dynamic resizing of a video channel. Additionally, if a second application attempts to bind to the video channel while the first is still executing, it will incur an X error and may terminate. See the pfVideoChannel reference page for more information. +o _G_r_a_p_h_i_c_s _p_i_p_e_l_i_n_e _s_t_a_t_i_s_t_i_c_s _a_n_d _m_u_l_t_i_p_l_e _a_p_p_l_i_c_a_t_i_o_n_s: In base IRIX6.x on InfiniteReality, only one application may be doing graphics pipeline statistics at a time. If a second application attempts to collect graphics pipeline statistics simultaneously with the first, the second application may hang. If for example, there are two copies of perfly executing, one should only select graphics pipeline statistics in one at a time and disable them in one program before enabling them in another. +o _F_i_l_l _s_t_a_t_i_s_t_i_c_s _a_n_d _G_r_a_p_h_i_c_s _p_i_p_e_l_i_n_e _s_t_a_t_i_s_t_i_c_s: InfiniteReality on base IRIX 6.X does not support simultaneous usage of fill stats and graphics pipeline statistics. If both of these modes are enabled simultaneously, even through separate programs, "KONA_DMAREAD: Resource busy" error messages will be generated. IRIS Performer 2.2.2 does not allow the Application to enable both simultaneously. +o _F_r_a_m_e _b_u_f_f_e_r _r_e_a_d _b_a_c_k _a_n_d _D_V_R InfiniteReality does not support read back from the frame buffer when DVR is in use. Performer will not prevent that to happen, for example asking for fill stats when DVR is on. +o InfiniteReality does not support DVR on X for video larger than 1280. IRIS Performer will resize only in the Y direction to prevent the whole screen from being garbage. +o _V_i_s_u_a_l_s _f_o_r _O_p_e_n_G_L/_X: InfiniteReality does not support 10 bits RGB with alpha plane. Only 12 bits per component is supported when the alpha plane is enabled. 4.5 _llll_iiii_bbbb_pppp_rrrr +o _p_f_F_l_u_x_e_s: can run out of valid buffers for updating slow readers. If you are running Lock or Float Phase, and if the App is much faster than the Cull/Draw then Draw process may not be able to catch up and use non up to date data until the App slows down. For morphing - 6 - applications that require eye-frame-accurate data to be passed from the APP or CULL to the DRAW process, a workaround is to use pfCycleBuffer instead of pfFlux. +o _A_n_t_i_a_l_i_a_s_i_n_g: When it is not multisampling, _p_f_A_n_t_i_a_l_i_a_s can significantly degrade performance. Specifically, on systems other than InfiniteReality and RealityEngine, use _p_f_A_n_t_i_a_l_i_a_s only for lines and points. +o _C_o_p_l_a_n_a_r _p_o_l_y_g_o_n_s: _p_f_D_e_c_a_l works only on machines that support the _s_t_e_n_c_i_l or _d_i_s_p_l_a_c_e_p_o_l_y_g_o_n command. The default decaling mode for PFDECAL_BASE uses _d_i_s_p_l_a_c_e_p_o_l_y_g_o_n instead of _s_t_e_n_c_i_l. This can significantly improve rendering performance on some machines but can also result in visual anomalies where layer polygons incorrectly "poke" through other geometry. If you wish to use stencil then specify the PFDECAL_BASE_HIGH_QUALITY token. In addition the new reference plane mode is supported only on InfiniteReality and only with patch 1808 or equiv. +o _S_t_e_n_c_i_l: stencil is used for PFDECAL_BASE_HIGH_QUALITY and for pfStats mode PFSTATSHW_GFXPIPE_FILL. The Extreme graphics platform supports stencil with reduced depth buffer resolution so IRIS Performer will not allocate stencil bits in the default framebuffer configuration. While allocated by default, the allocation of extra stencil bits can also affect your depth resolution and number of samples for multisample on InfiniteReality and RealityEngine graphics platforms. +o _T_r_a_n_s_p_a_r_e_n_c_y: _p_f_T_r_a_n_s_p_a_r_e_n_c_y works only on machines that support either _b_l_e_n_d_f_u_n_c_t_i_o_n or _m_u_l_t_i_s_a_m_p_l_e (InfiniteReality and RealityEngine). +o Since having other applications running can impact real-time performance, it's sometimes desirable to minimize the number of daemons and other processes. If you have problems achieving real-time behavior, try the _p_f_u_L_o_c_k_C_P_U utility code in _l_i_b_p_f_u_t_i_l. You might also try turning off the desktop support and other daemons that are not crucial to your application, e.g.: % touch ~/.disableDesktop or % mkdir ~/.desktop-<machinename> % touch ~/.desktop-<machinename>/nodesktop - 7 - and for total removal do: % chkconfig desktop off % chkconfig objectserver off (gone in IRIX 6.5) % chkconfig directoryserver off % chkconfig fontserver off 4.6 _llll_iiii_bbbb_pppp_ffff +o Intersection testing of line segments (_p_f_N_o_d_e_I_s_e_c_t_S_e_g_s) against geometry in _p_f_B_i_l_l_b_o_a_r_d nodes is not yet implemented; only the bounding sphere of the entire _p_f_B_i_l_l_b_o_a_r_d is available. +o _C_h_a_n_n_e_l _f_a_d_e _L_O_D _a_t_t_r_i_b_u_t_e_s _a_n_d _m_i_x_e_d _g_r_a_p_h_i_c_s _c_o_n_f_i_g_u_r_a_t_i_o_n_s: Channel fade LOD attributes are set in the application process. The existence of multisample, required for fade LOD, is tested at the time that the attributes are set with _p_f_C_h_a_n_L_O_D_A_t_t_r. This test for multisample uses the application process window system connection (_p_f_O_p_e_n_W_S_C_o_n_n_e_c_t_i_o_n) or else the DISPLAY environment variable to select a screen for determining graphics configuration, rather than testing the current window of the pfChannel. In a multipipe environment where one graphics pipeline has multisample capability (for example, with or without sufficient Raster Manager boards for the chosen framebuffer resolution) and one does not, the application process needs to have a window system connection or DISPLAY that points to the pipeline with multisample sufficiency, in which case the non-multisample pfChannels will try to use fade LOD. In a multi-window environment, windows without multisample on a system with multisample will try to use fade LOD. +o Triple-Keyboard Option(TKO) doesn't work for IRISGL. It works fine for OpenGL. +o _C_U_L_L_o_D_R_A_W process model does not work properly. +o _p_f_E_a_r_t_h_S_k_y does not work for offset viewing frustrum. +o _p_f_A_S_D has memory leaks in the database paging. Your applicaion may run out of memory after several ASD tiles paging. +o _p_f_A_S_D has alignment timing problem. When the number of triangles in pfASD is large, there might be inconsistency in position updates of moving vehicles on terrain. - 8 - +o If you explicitly edit the _p_f_T_e_x_t_u_r_e, and then remove all reference to TexLOD, the texLOD is not destroyed. This is also consistent with our apply{Min,Max}LOD behavior. +o _N_6_4 static perfly fail to link in compilation. +o When excuting _c_l_i_p_f_l_y, on the first few frames the virtual LOD offset and other params seem to be bogus, as if the node CULL func that sets it is not getting executed or not having the proper effect. +o _M_u_l_t_i _T_h_r_e_a_d_e_d _s_t_a_g_e_s get only one configStage callback. +o Performer applications die with an X error when run on remote IMPACT display on which IRIX6.4 IMPACT Patch 1935 is installed. If you need to do remote rendering on an IMPACT display, do not install patch 1935 or upgrade to a more recent IMPACT patch (2677 or replacement). +o _p_f_C_l_i_p_T_e_x_t_u_r_e doesn't work with 8bit luminance data on IMPACT patch 1935. It might be fixed in future IMPACT patches. +o For _p_f_S_C_S (and _p_f_S_C_S which is derived from it), the value of pfGetTravMat() is not well-behaved-- sometimes it returns a matrix that includes the pfSCS transformation and sometimes not. The semantics (according to the man page) may change and become more sane in future patch releases of Performer 2.2. 4.7 _llll_iiii_bbbb_pppp_ffff_uuuu_tttt_iiii_llll +o Calling _p_f_u_N_e_w_C_l_i_p_C_e_n_t_e_r_N_o_d_e or _p_f_u_N_e_w_M_P_C_l_i_p_C_e_n_t_e_r_N_o_d_e on a pfGeode (possibly other types as well) as a child when a pfdBuilder is being used to create a scene graph (e.g. in the .im loader) can result in the callback being lost. One workaround is to create a dummy pfGroup node above the pfGeode and use that instead. +o The implementation of _p_f_u_G_e_t_C_l_o_s_e_s_t_P_o_i_n_t (and therefore _p_f_u_N_e_w_C_l_i_p_C_e_n_t_e_r_N_o_d_e which uses it) assumes that all coordinate scaling in pfSCS and pfDCS nodes in the scene are uniform, i.e. that the scale factor is the same in all three coordinate axis directions. If this is not the case, the answer returned will be valid coordinates (and texture coordinates) of a point on a triangle, but it may not be the actual closest point. - 9 - +o In the case of degenerate triangles the coordinates returned by _p_f_u_G_e_t_C_l_o_s_e_s_t_P_o_i_n_t may not be on or close to an object at all. +o _p_f_u_G_e_t_C_l_o_s_e_s_t_P_o_i_n_t does not traverse point or line primitives. +o Interaction of _p_f_u_G_e_t_C_l_o_s_e_s_t_P_o_i_n_t with automatic texture coordinate generation is very limited. Only PFTG_OBJECT_LINEAR is supported, and then only when a pfTexGen is explicitly attached to the pfGeoSet's pfGeoState, and even this has not been well tested. The safest approach is to specify texture coordinates explicitly. 4.8 _llll_iiii_bbbb_pppp_ffff_uuuu_iiii +o Libpfui has both a C API and a C++ API. The C API is actually wrappers around the C++ API and is not complete. 4.9 _llll_iiii_bbbb_pppp_ffff_dddd_bbbb +o _p_f_d_L_o_a_d_F_i_l_e__d_x_f: The DXF loader does not fully support the format. +o _p_f_d_L_o_a_d_F_i_l_e__s_g_o: The SGO loader does not support triangle strips. +o _p_f_b file format does not support ClipMap textures. 4.10 _SSSS_aaaa_mmmm_pppp_llll_eeee _PPPP_rrrr_oooo_gggg_rrrr_aaaa_mmmm_ssss +o _p_g_u_i_d_e/_l_i_b_p_u_i/_m_o_t_i_f_x_f_o_r_m_e_r: The motifxformer.c program dumps core when loading new files in N32 or 64-bit with 6.2 and 6.3. This Motif bug has been fixed in 6.4. +o _p_g_u_i_d_e/_l_i_b_p_f/_C/_l_p_s_t_a_t_e: The lpstate.c example for demonstrating pfLPointStates assumes sophisticated texturing capabilities that may not be fully supported on IMPACT, Extreme, or Indy graphics platforms. +o _p_e_r_f_l_y _n_e_e_d_s _f_u_l_l _s_c_r_e_e_n _w_i_n_d_o_w _f_o_r _D_V_R: When running perfly and using dynamic video resolution (DVR), perfly should be run with a full-screen, no-border window (ie, do not use the -W option). If run from a smaller, or bordered, window, the resizing will be very apparent as it affects the entire video channel (which is usually - 10 - the entire screen for normal workstation operation). +o _p_e_r_f_l_y _w_i_t_h _A_u_t_o_m_a_t_i_c-_D_V_R _w_i_l_l _b_e _r_e_d_r_a_w_i_n_g _G_U_I _w_h_e_n _e_n_a_b_l_e_d: When the graphical user-interface (GUI) is enabled and DVR is on, the GUI must be redrawn whenever there is a video channel resize, or whenever the mouse is in the GUI. This is very expensive and a drop in framerate will be incurred. +o _7._2 _c_o_m_p_i_l_e_r_s compilers warn about the obsolete and no longer support Olimit flag that has to be replaced with -OPT:Olimit=n in the makefile. +o The default pfGeoState applied to the textured font loaded with pfdLoadFont_type1( "anyFontName", PFDFONT_TEXTURED ) ) has a transparency setting of PFTR_ON | PFTR_NO_OCCLUDE. This causes major rendering problems for the text. WAR: manually set this value to PFTR_ON and everything works right. 4.11 _N_o_t_e_s__o_n__S_h_a_r_e_d__M_e_m_o_r_y__C_o_n_f_i_g_u_r_a_t_i_o_n_s IRIS Performer requires shared memory and uses a memory- mapped file, the location of which depends on the value of the PFTMPDIR environment variable: +o If PFTMPDIR is not set, Performer uses /_d_e_v/_z_e_r_o as the default. Running an application in this configuration: - Uses swap space - Does not require disk space until main memory is exhausted - Is faster than using a regular memory mapped file via PFTMPDIR - Causes IRIX to kill the process(es) and log an error to the console if the application runs out of space for shared memory in the swap partition. +o If PFTMPDIR is set, Performer creates files in the specified directory. Running an application in this configuration: - Requires disk space even before main memory is exhausted - Is slower than /_d_e_v/_z_e_r_o because it touches the disk - 11 - - Produces appropriately sized core dump files with no limit set by IRIS Performer - Might cause a core dump from a segmentation violation inside _p_f_M_a_l_l_o_c if the application runs out of space for shared memory in the file system containing PFTMPDIR - PFTMPDIR should be set only to a directory on a local file system.