Re: 945GM(E) screen rotation and slow GL.
I've done tests in Second Life where the framerate in Linux is about 2x worse than in Windows. On Wed, Jan 14, 2009 at 11:30 AM, Mateusz Kaduk wrote: > 2009/1/14 Vedran Rodic >> >> Probably unrelated, but the old 915tex driver had about 50% better >> framerate in glxgears than either the old Mesa driver, or the current. >> Of course, this driver is now obsolete. Do you have any idea why was >> that driver so much faster? If I remember correctly, the performance >> improvement was also visible in quake3. > > afaik glxgears now is vblank synced by default and someone even wrote its > not a benchmarking tool. You can disable vblank syncing with dri-conf. When > it comes to openarena, regnum-online, nexuiz without GLSL it preformes fine > and is pretty playable on GM965. Windows does not preform that much faster > and no tests were done to show real difference. > > ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 945GM(E) screen rotation and slow GL.
2009/1/14 Vedran Rodic > Probably unrelated, but the old 915tex driver had about 50% better > framerate in glxgears than either the old Mesa driver, or the current. > Of course, this driver is now obsolete. Do you have any idea why was > that driver so much faster? If I remember correctly, the performance > improvement was also visible in quake3. > afaik glxgears now is vblank synced by default and someone even wrote its not a benchmarking tool. You can disable vblank syncing with dri-conf. When it comes to openarena, regnum-online, nexuiz without GLSL it preformes fine and is pretty playable on GM965. Windows does not preform that much faster and no tests were done to show real difference. ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 945GM(E) screen rotation and slow GL.
On Wed, Jan 14, 2009 at 6:27 AM, Keith Packard wrote: > You're fairly seriously memory bandwidth limited with that chipset, and > at that resolution. Probably unrelated, but the old 915tex driver had about 50% better framerate in glxgears than either the old Mesa driver, or the current. Of course, this driver is now obsolete. Do you have any idea why was that driver so much faster? If I remember correctly, the performance improvement was also visible in quake3. Vedran ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 945GM(E) screen rotation and slow GL.
Keith Packard wrote: >> Do you think the 945 chipset is just too slow to rotate a 1920x1200 >> screen? > > One thing that may improve performance a bit is when we get tiling > support for the displayed buffer. That work requires some additional > kernel infrastructure to manage the 'fence' registers which are > responsible for linearizing tiled regions for software rendering code. > > You might try 16bpp instead of 32bpp; it won't look as good, but it > should go faster. Tried that: 85 fps in normal mode, 27 fps for left-rotation. Would a board with a 965GM chipset fare much better, you think? I haven't seen many small form-factor embedded boards with that chipset yet, unfortunately. Cheers, Rob ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 945GM(E) screen rotation and slow GL.
On Wed, 2009-01-14 at 13:18 +0800, Rob Kramer wrote: > Cool, thanks. That's a lot better at 18 fps, but still a lot worse than > normal landscape operation. Is that expected? I thought with modern GL > setups, rotation basically came for free -- but I'm a GL newbie :) The rotation is done by drawing to an upright frame buffer and then copying that to the rotated frame buffer for presentation. That's a lot of pixels to copy around each frame. > Do you think the 945 chipset is just too slow to rotate a 1920x1200 > screen? You're fairly seriously memory bandwidth limited with that chipset, and at that resolution. One thing that may improve performance a bit is when we get tiling support for the displayed buffer. That work requires some additional kernel infrastructure to manage the 'fence' registers which are responsible for linearizing tiled regions for software rendering code. You might try 16bpp instead of 32bpp; it won't look as good, but it should go faster. -- keith.pack...@intel.com signature.asc Description: This is a digitally signed message part ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 945GM(E) screen rotation and slow GL.
Keith Packard wrote: > On Wed, 2009-01-14 at 09:51 +0800, Rob Kramer wrote: > >> When testing screen rotation to portrait mode, the glxgears framerate >> drops from 51 fps (not too great to start with) to 1 fps. > > Thanks! We lost our rotation acceleration when we switched to dri_bo's > for pixmaps. Cool, thanks. That's a lot better at 18 fps, but still a lot worse than normal landscape operation. Is that expected? I thought with modern GL setups, rotation basically came for free -- but I'm a GL newbie :) Do you think the 945 chipset is just too slow to rotate a 1920x1200 screen? Cheers! Rob ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
Re: 945GM(E) screen rotation and slow GL.
On Wed, 2009-01-14 at 09:51 +0800, Rob Kramer wrote: > When testing screen rotation to portrait mode, the glxgears framerate drops > from 51 fps (not too great to start with) to 1 fps. Thanks! We lost our rotation acceleration when we switched to dri_bo's for pixmaps. Try this patch: diff --git a/src/i830_display.c b/src/i830_display.c index 50fbc4d..4fba1fb 100644 --- a/src/i830_display.c +++ b/src/i830_display.c @@ -1601,6 +1601,7 @@ static PixmapPtr i830_crtc_shadow_create(xf86CrtcPtr crtc, void *data, int width, int height) { ScrnInfoPtr pScrn = crtc->scrn; +I830CrtcPrivatePtr intel_crtc = crtc->driver_private; I830Ptr pI830 = I830PTR(pScrn); int rotate_pitch; PixmapPtr rotate_pixmap; @@ -1621,6 +1622,8 @@ i830_crtc_shadow_create(xf86CrtcPtr crtc, void *data, int width, int height) xf86DrvMsg(pScrn->scrnIndex, X_ERROR, "Couldn't allocate shadow pixmap for rotated CRTC\n"); } +if (intel_crtc->rotate_mem && intel_crtc->rotate_mem->bo) + i830_set_pixmap_bo(rotate_pixmap, intel_crtc->rotate_mem->bo); return rotate_pixmap; } -- keith.pack...@intel.com signature.asc Description: This is a digitally signed message part ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg
945GM(E) screen rotation and slow GL.
Hi, I have an embedded board with 945GM chipset (1.8GHz Celeron M) connected via DVI to a 1920x1200 screen. I'm trying to benchmark GL performance a bit with glxgears, which should be good enough to give a rough estimate of the device's capabilities. (Is there a better generic GL benchmark tool?) When testing screen rotation to portrait mode, the glxgears framerate drops from 51 fps (not too great to start with) to 1 fps. I remember in the past there were always dire performance warnings regarding XAA and rotation; impact of shadow FBs and all that. Is that still a valid concern for a modern setup, i.e. is GL expected to be sluggish/unusable in rotated mode? Should I not use the 'rotate' feature at all, but rotate GL geometries/textures myself, in my application? (that would be painful, I guess). Using kernel 2.6.28.3. Xorg, drm, mesa are from Fedora rawhide. The intel driver is git HEAD. For the driver, it doesn't seem to make a difference whether I specify UXA or not. Cheers, Rob ___ xorg mailing list xorg@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/xorg