Re: [osg-users] StandardShadowMap on ATI
Hi Everyone browsing this thread. I just downoaded Catalyst 10.8 for Windows. Its a pleasure to report that problems (broken TexMat base Textures) with osgShadow example on recent Catalyst versions 10.4-10.7 seem to be fixed on 10.8. Thats the way to go AMD ;-). I tested on OSG trunk 32 bit / Windows 7 64 bit / Radeon 5980 . Wojtek -- From: Bradley Baker Searles bsear...@carsim.com Sent: Wednesday, August 04, 2010 6:13 PM To: osg-users@lists.openscenegraph.org Subject: Re: [osg-users] StandardShadowMap on ATI Hi Wojtek, Ahh, good catch with the improper derivation! :) Thank you. I'm attaching the corrected st.h file. I just tried to get the osgShadow example working (-4 --lispsm), and it didn't work with any of the workarounds that I'd used in my code (not using variables to index light sources, don't use derived light values such as gl_FrontLightModelProduct, skip empty program add, etc). Now the standard LispSM does its lighting in the vertex shader, and all of my Programs do their lighting in the fragment programs... so there may be additional tricks for vert shaders that I'm not yet aware of? I haven't tried any of the other techniques yet, but perhaps I'll see if those can be patched up. But... even if we could patch them, the fixes are ugly enough that I don't think they'd belong in the osg source. Perhaps I should post some of these issues on the ATI forums. Thank you! Baker -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30533#30533 Attachments: http://forum.openscenegraph.org//files/st_191.h ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
Re: [osg-users] StandardShadowMap on ATI
Perfect ! One thing got me intrigued, though. You derive internal ViewData class from MinimalShadowMap::ViewData instead of VsvCustomShadowTechniqueClass::ViewData. Is it deliberate ? I suspect this works because VsvCustomShadowTechniqueClass is actually MinimalCullBoundsShadowMap and MinimalCullBoundsShadowMap::ViewData does not differ from MinimalShadowMap::ViewData. But if one day you will try to change VsvCustomShadowTechniqueClass typedef to other ViewDependentTechnique class some issues may appear. PS. In the meantime I managed to install Catalsyst 10.7. Textures on island scene are still broken (as in 10.5/10.6) no matter which shadow technique I use. They are broken even with the simplest ShadowMap. Perhaps, you have found a workaround for this problem ? Wojtek -- From: Bradley Baker Searles bsear...@carsim.com Sent: Tuesday, August 03, 2010 10:47 PM To: osg-users@lists.openscenegraph.org Subject: Re: [osg-users] StandardShadowMap on ATI Alright, so just wanted to post the code for the overrides I did, just in case anyone stumbles upon this forum entry and wants to see precisely what I did. It's a bit different than the example that Wojtek posted in this thread (as referenced above): http://forum.openscenegraph.org/viewtopic.php?t=4178 (Wojtek's MyViewDependentShadowMap example) :D Seems to be working very well. Thanks again- Baker Searles -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30504#30504 Attachments: http://forum.openscenegraph.org//files/osgforum_ati_107_drivers_266.jpg http://forum.openscenegraph.org//files/st_174.cpp http://forum.openscenegraph.org//files/st_482.h ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
Re: [osg-users] StandardShadowMap on ATI
Hi Wojtek, Ahh, good catch with the improper derivation! :) Thank you. I'm attaching the corrected st.h file. I just tried to get the osgShadow example working (-4 --lispsm), and it didn't work with any of the workarounds that I'd used in my code (not using variables to index light sources, don't use derived light values such as gl_FrontLightModelProduct, skip empty program add, etc). Now the standard LispSM does its lighting in the vertex shader, and all of my Programs do their lighting in the fragment programs... so there may be additional tricks for vert shaders that I'm not yet aware of? I haven't tried any of the other techniques yet, but perhaps I'll see if those can be patched up. But... even if we could patch them, the fixes are ugly enough that I don't think they'd belong in the osg source. Perhaps I should post some of these issues on the ATI forums. Thank you! Baker -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30533#30533 Attachments: http://forum.openscenegraph.org//files/st_191.h ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
Re: [osg-users] StandardShadowMap on ATI
Alright, so just wanted to post the code for the overrides I did, just in case anyone stumbles upon this forum entry and wants to see precisely what I did. It's a bit different than the example that Wojtek posted in this thread (as referenced above): http://forum.openscenegraph.org/viewtopic.php?t=4178 (Wojtek's MyViewDependentShadowMap example) :D Seems to be working very well. Thanks again- Baker Searles -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30504#30504 Attachments: http://forum.openscenegraph.org//files/osgforum_ati_107_drivers_266.jpg http://forum.openscenegraph.org//files/st_174.cpp http://forum.openscenegraph.org//files/st_482.h ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
Re: [osg-users] StandardShadowMap on ATI
Hi Baker, My responses below. Just for reference, the modifications I've made in my glsl shaders to get it running on ATI were mainly the following: - only index texture coordinates with constants, variables (even const) don't seem to work. - ensure all variables are initialized (sloppy on my part) - hardcode to light 0, the accumulation loop did not work. - don't use gl_FrontMaterial / gl_BackMaterial, as the values didn't work as they did with nVidia, and produced unexpected values. - don't use gl_FrontLightModelProduct, same reason as previous. Thats the ATI. I stay with Catalyst 10.4. With 10.5 I had the issues you show on osgshadow screenshots. 10.6 and 10.7 does not start on my machine. Bombs with blue screen on startup. Thats maybe related to the fact that I switched to dual GPU Radeon 5970. I saw on ATI forums that blue screens happen on dual GPU Radeons. I'm not using alpha in a nonstandard way (at least not intentionally), and I don't use discard for anything. That's good information though, and I will fiddle around with my shaders to see if they have any impact (although with the NULL program, they shouldn't be used, eh?). So having said that, the use of a NULL program indeed does seem like it should be valid, as specified in the OpenGL 21.20061201 spec (p. 75), and indeed OSG does call glUseProgram(0), which should revert to fixed function. So this is a driver bug, right? Still, I agree it might be useful to let the user provide a Program to use, or whether to attach the NULL program. I agree. However, its should be more generic to allow for modification of other StateAttributes as well. Some callback maybe... Currently the only option is overriding technique its internal ViewData class (which is bit tricky but I think I have posted 'how to do it' on the forum). I ran osgshadow on the ATI 4670 with the 10.7 drivers, and I'll post the results below. I'm also attaching a screenshot of the vehicle in osgviewerWX, which seems to be having some texture coordinate issues (again, fine on nVidia). Hmm, even the osg cow seems wrong. I don't know precisely what the Texture Matrix fix is that you're speaking about, but I can poke around the forums and see if I find the issue. There is a line in Main Vertex Shader where tex coords at index zero 0 are computed: gl_TexCoord[0] = gl_TextureMatrix[0] * gl_MultiTexCoord0; gl_TextureMatrix seems to be initialized to 0 on ATI Catalyst 10.4 and resulting tex coords are always 0. Workaround for this is to either comment out gl_TextureMatrix[0] multiplication in this shader or set up identity osg::TexMat at the scene root state set. With above change I am fairly happy on 10.4. So I recommend you try these drivers. For my laptop with really old ATi XPress 1200 I use Catalyst 10.2 and these drivers seem to be even better because osgshadow works without any modifications. I must be honest with you though. I actually have two cards on my Windows 7 one NVidia 8800 and Radeon 5970. And NVidia is most of the time the card of choice for OpenGL. So I maybe shoud not say I do lots of extensive testing on the Radeon. I switch to it when I want to check code compatibilty. As I wrote earlier I am not able to get 10.6 and 10.7 to start up. But I have earlier tried 10.5 and they showed the same problems with texture that looked like your screenshots. We discussed this on the osg-forum couple weeks ago (http://forum.openscenegraph.org/viewtopic.php?p=29501#29501) So you may consider giving 10.4 or earlier drivers second chance ;-) And by the way, thanks for the lispsm submission, it's really well done! I am happy when people use it ;-) Cheers, Wojtek ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
Re: [osg-users] StandardShadowMap on ATI
Hi Wojtek- Thanks again for the thoughtful response. I believe I have enough information to work around this issue now. I just wanted to post my last set of findings for anyone else who might be helped by the thread. I installed catalyst 10.4 and 10.2, and I never quite got to a point where everything worked as expected. I will post some screenshots below. I am thinking I will probably just go with the latest drivers and work around the NULL program driver bug, either by hacking our local OSG copy, or doing the proper ViewData override as you suggested (haven't looked at those details yet). Ahh, yes the texture matrix issue in the glsl. I had indeed already done that (just omitted the matrix altogether in the shader as you referenced). I didn't think to attach an osg TexMat identity matrix at the root, so I was confused by your reference to the osg TexMat class :) But the TexMat solution is probably a more appropriate workaround. Take care- Baker Searles P.S. Yes I too had a dual-gpu ATI setup at home last year (dual 4850 in crossfire) and they were crazy fast, but I had render issues in Fallout 3 (HDR effects would blink on and off) so I had to drop down to using just one of the boards. I guess you can't do that though with the 5970 having both chips on one board :) -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30405#30405 Attachments: http://forum.openscenegraph.org//files/ati_102_osgshadow_4_sm_211.jpg http://forum.openscenegraph.org//files/ati_102_osgshadow_4_lispsm_299.jpg http://forum.openscenegraph.org//files/ati_102_osgshadow_4_stsm_112.jpg http://forum.openscenegraph.org//files/ati_102_osgviewer_coupe_106.jpg http://forum.openscenegraph.org//files/ati_102_coupe_143.jpg ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
Re: [osg-users] StandardShadowMap on ATI
Hi Bradley, I have not tested StandardShadowMap with Catalyst 10.7. 10.5 10.6 did not work for me so I have been staying at 10.4. With 10.4 the only issue I had was TexMat not initialized to identity and mainVertexShader in Standard Shadow Map was producing 0,0 base texture coords for all frgments. I guess you have a bit different situation, though. You write that your scene already contains programs with shaders. When NULL program is applied with override, it means that your shaders are blocked and fixed pipeline draws the shadow map. So you may have the problems if your shaders compute alpha with some non standard (non repeatable in fixed pipeline) way or discard fragments. I must admit that for above case setting NULL program might be seen as a bug. However, it was done in purpose to make shadow map rendering faster because color does not need to be computed and thus there is no need to use existing usually more complex shaders. But if they discard fragments or compute depth it may be not such good idea after all... We may try to come up with a new method to let user select if he/she wants to set this NULL program or not. Out of curiosity does the same problem appear in osgshadow example ? The example works on 10.4 (with minor TexMat modification). Try: osgshadow -4 --lispsm or osgshadow -4 --stsm Cheers, Wojtek -- From: Bradley Baker Searles bsear...@carsim.com Sent: Wednesday, July 28, 2010 12:48 AM To: osg-users@lists.openscenegraph.org Subject: [osg-users] StandardShadowMap on ATI Hi, I have been working for a while on compatibility issues with nVidia and ATI in our OpenSceneGraph application. Aside from a fair number of GLSL shader source issues, one of the last remaining problems was that textured objects were not casting shadows (please see first attached screenshot). After a little detective work, it seems that the empty osg Program attached to the StateSet belonging to the _camera in StandardShadowMap::ViewData::init() is causing the problem: // optimization attributes osg::Program* program = new osg::Program; stateset-setAttribute( program, osg::StateAttribute::OVERRIDE | osg::StateAttribute::ON ); stateset-setMode ( GL_LIGHTING, osg::StateAttribute::OVERRIDE | osg::StateAttribute::OFF ); stateset-setMode ( GL_BLEND, osg::StateAttribute::OVERRIDE | osg::StateAttribute::OFF ); #if 0 // fixed pipeline seems faster (at least on my 7800) program-addShader( new osg::Shader( osg::Shader::FRAGMENT, uniform sampler2D texture; \n void main(void) \n {\n gl_FragColor = texture2D( texture, gl_TexCoord[0].xy ); \n } \n ) ); // program-addShader Fragment program-addShader( new osg::Shader( osg::Shader::VERTEX, void main(void) \n {\n gl_Position = ftransform(); \n gl_TexCoord[0] = gl_MultiTexCoord0; \n } \n ) ); // program-addShader Vertex #endif The setting of the fragment and vertex shaders to the stateset are #if 0'd out, but the osg Program setting is not. Anyway, if I skip adding the empty program, things seem fine. Oddly, if I keep the empty program and skip the AlphaFunc above it that allows texture alpha to affect the depth buffer render, it casts shadows properly (well, without the alpha penetration of course). Everything works fine with nVidia drivers, even with the empty program. My scene has a mixture of shader programs depending on how the objects are setup, but they all include the shadow shader source (except the flat shaded objects like the force arrows shown below). So it seems to be an easy fix, but I was curious if anyone else had run into this problem? I did many searches on the forum and came up empty. And if anyone has deeper illumination into how the empty program could affect the statesets (seems like a bug to me since nVidia works fine), or if there is a more appropriate fix I'm not thinking of, I'd appreciate the feedback. Thanks- Baker Searles Hardware Versions: nVidia GTS 250 ATI Radeon HD4670 Software Versions: OSG @ 2.8.3 ATI Display Drivers @ 10.7 nVidia Display Drivers @ 258.96 -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30329#30329 Attachments: http://forum.openscenegraph.org//files/ati_emptyglslprogram_noalphafunc_163.jpg http://forum.openscenegraph.org//files/ati_noemptyglslprogram_192.jpg http://forum.openscenegraph.org//files/ati_emptyglslprogram_debughud_103.jpg http://forum.openscenegraph.org//files/ati_emptyglslprogram_167.jpg ___ osg-users mailing list osg-users@lists.openscenegraph.org http
Re: [osg-users] StandardShadowMap on ATI
Hi Wojtek, Thank you so much for the response. Just for reference, the modifications I've made in my glsl shaders to get it running on ATI were mainly the following: - only index texture coordinates with constants, variables (even const) don't seem to work. - ensure all variables are initialized (sloppy on my part) - hardcode to light 0, the accumulation loop did not work. - don't use gl_FrontMaterial / gl_BackMaterial, as the values didn't work as they did with nVidia, and produced unexpected values. - don't use gl_FrontLightModelProduct, same reason as previous. I'm not using alpha in a nonstandard way (at least not intentionally), and I don't use discard for anything. That's good information though, and I will fiddle around with my shaders to see if they have any impact (although with the NULL program, they shouldn't be used, eh?). So having said that, the use of a NULL program indeed does seem like it should be valid, as specified in the OpenGL 21.20061201 spec (p. 75), and indeed OSG does call glUseProgram(0), which should revert to fixed function. So this is a driver bug, right? Still, I agree it might be useful to let the user provide a Program to use, or whether to attach the NULL program. I ran osgshadow on the ATI 4670 with the 10.7 drivers, and I'll post the results below. I'm also attaching a screenshot of the vehicle in osgviewerWX, which seems to be having some texture coordinate issues (again, fine on nVidia). Hmm, even the osg cow seems wrong. I don't know precisely what the Texture Matrix fix is that you're speaking about, but I can poke around the forums and see if I find the issue. And by the way, thanks for the lispsm submission, it's really well done! Thanks Baker Searles -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30355#30355 Attachments: http://forum.openscenegraph.org//files/osgcow_504.jpg http://forum.openscenegraph.org//files/osgviewerwx_coupe_325.jpg http://forum.openscenegraph.org//files/osgshadow_4_stsm_369.jpg http://forum.openscenegraph.org//files/osgshadow_4_lispsm_133.jpg ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
[osg-users] StandardShadowMap on ATI
Hi, I have been working for a while on compatibility issues with nVidia and ATI in our OpenSceneGraph application. Aside from a fair number of GLSL shader source issues, one of the last remaining problems was that textured objects were not casting shadows (please see first attached screenshot). After a little detective work, it seems that the empty osg Program attached to the StateSet belonging to the _camera in StandardShadowMap::ViewData::init() is causing the problem: // optimization attributes osg::Program* program = new osg::Program; stateset-setAttribute( program, osg::StateAttribute::OVERRIDE | osg::StateAttribute::ON ); stateset-setMode ( GL_LIGHTING, osg::StateAttribute::OVERRIDE | osg::StateAttribute::OFF ); stateset-setMode ( GL_BLEND, osg::StateAttribute::OVERRIDE | osg::StateAttribute::OFF ); #if 0 // fixed pipeline seems faster (at least on my 7800) program-addShader( new osg::Shader( osg::Shader::FRAGMENT, uniform sampler2D texture; \n void main(void) \n { \n gl_FragColor = texture2D( texture, gl_TexCoord[0].xy ); \n } \n ) ); // program-addShader Fragment program-addShader( new osg::Shader( osg::Shader::VERTEX, void main(void) \n { \n gl_Position = ftransform(); \n gl_TexCoord[0] = gl_MultiTexCoord0; \n } \n ) ); // program-addShader Vertex #endif The setting of the fragment and vertex shaders to the stateset are #if 0'd out, but the osg Program setting is not. Anyway, if I skip adding the empty program, things seem fine. Oddly, if I keep the empty program and skip the AlphaFunc above it that allows texture alpha to affect the depth buffer render, it casts shadows properly (well, without the alpha penetration of course). Everything works fine with nVidia drivers, even with the empty program. My scene has a mixture of shader programs depending on how the objects are setup, but they all include the shadow shader source (except the flat shaded objects like the force arrows shown below). So it seems to be an easy fix, but I was curious if anyone else had run into this problem? I did many searches on the forum and came up empty. And if anyone has deeper illumination into how the empty program could affect the statesets (seems like a bug to me since nVidia works fine), or if there is a more appropriate fix I'm not thinking of, I'd appreciate the feedback. Thanks- Baker Searles Hardware Versions: nVidia GTS 250 ATI Radeon HD4670 Software Versions: OSG @ 2.8.3 ATI Display Drivers @ 10.7 nVidia Display Drivers @ 258.96 -- Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=30329#30329 Attachments: http://forum.openscenegraph.org//files/ati_emptyglslprogram_noalphafunc_163.jpg http://forum.openscenegraph.org//files/ati_noemptyglslprogram_192.jpg http://forum.openscenegraph.org//files/ati_emptyglslprogram_debughud_103.jpg http://forum.openscenegraph.org//files/ati_emptyglslprogram_167.jpg ___ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org