[GLTeapot] The teapot can't be controlled by a mouse

[diver]

The teapot in GLTeapot can't be controlled by a mouse if video mode depth is other than 32bit. It used to work sometime in the past.

[diver]

It seems it just ignore mouse events, keboard (alt+n) works though.

[korli]

It seems changing the priority of the direct daemon thread to less than 15 let me control the teapot with the mouse. Stefano, should I change the priority of the direct daemon thread to (B_DISPLAY_PRIORITY - 1) ?

[jackburton]

But does work well here. Especially since Axel's changes to the scheduler.

[korli]

This bug is only exposed with 1 CPU machine and VMs. Is it your test case ?

[diver]

Yes it is.

[jackburton]

Sorry, I hadn't understood what you were talking about. I thought you were talking about the GLTeapot _window_. Not the teapot itself. I can reproduce the bug. I removed the part about the black cursor because that depends on not having a hardware cursor.

[jackburton]

Replying to korli:

It seems changing the priority of the direct daemon thread to less than 15 let me control the teapot with the mouse. Stefano, should I change the priority of the direct daemon thread to (B_DISPLAY_PRIORITY - 1) ?

I find it weird that lowering the priority would fix the problem. Or does it make sense ?

[jackburton]

Replying to jackburton:

Replying to korli:

It seems changing the priority of the direct daemon thread to less than 15 let me control the teapot with the mouse. Stefano, should I change the priority of the direct daemon thread to (B_DISPLAY_PRIORITY - 1) ?

I find it weird that lowering the priority would fix the problem. Or does it make sense ?

BTW Doesn't work here. I tried lowering the priority using ProcessController but the teapot isn't controllable via mouse.

[modeenf]

Replying to korli:

This bug is only exposed with 1 CPU machine and VMs. Is it your test case ?

Replying to diver:

Yes it is.

And VMs are vmware? If that's the case and it works on reale HW perhaps it's the OS then? I have tested this on Zeta 1.21 without problems (no not in vmware).

[modeenf]

Have tested it in hrev23050 with vmware and the problem are there. GLTeapot also have problem when maximized looks like the BGLView or GLTeapot ObjectView have some problems. As the GLTeapot works in zeta and probably in BeOS i think that the problem are in BGLView?

[modeenf]

Tested to add the VMware graphic driver and now I can control the teapot with my mouse. So apparently if you don't use an accelerated driver then the teapot can't be controlled.

[jackburton]

Replying to modeenf:

Tested to add the VMware graphic driver and now I can control the teapot with my mouse. So apparently if you don't use an accelerated driver then the teapot can't be controlled.

I tested without the VMWare graphic driver, but in 32 bit depth video mode, and the teapot can be controlled. So apparently the problem appears only when using a double buffered video mode (i.e. any mode other than 32 bit depth).

[korli]

Yes, it used to work when the mesa renderer wasn't implementing DirectConnected(), and using DrawBitmap(). See hrev20797.

[stippi]

Could the problem be glReadPixels()? It is used in ObjectView to implement hit testing the objects. Where do I need to look for this function?

[jackburton]

Replying to stippi:

Could the problem be glReadPixels()? It is used in ObjectView to implement hit testing the objects. Where do I need to look for this function?

It's a mesa function. What could happen is that, since we are using a BDirectWindow, the colorspace of the front buffer (used by directwindow) is different than the one of the back buffer (used by the GL Addon), and somehow the GL addon doesn't account for this.

[stippi]

Or maybe it does, but it shouldn't in certain circumstances, since the objects are drawn as quickly as possible for the hit test (in flat shading mode with special colors that indicate their ID). Such an image never shows, so it looks like it is drawn in some back buffer. Any hints as to where I should look for this?

[jackburton]

Replying to stippi:

Or maybe it does, but it shouldn't in certain circumstances, since the objects are drawn as quickly as possible for the hit test (in flat shading mode with special colors that indicate their ID). Such an image never shows, so it looks like it is drawn in some back buffer. Any hints as to where I should look for this?

Tried the gl addon in src/add-ons/opengl/mesa_software_renderer/ ?

[stippi]

Thanks. Well the MesaSoftwareRenderer's CopyPixelsIn/Out() methods only support 32 bit colors spaces. Maybe that's the problem already. I am going to look into it later.

[korli]

Replying to stippi: Any hints as to where I should look for this?

The GL dispatcher should route the call to : src/kits/opengl/mesa/swrast/s_readpix.c function _swrast_ReadPixels() The Mesa function should then call the render buffer function fRenderBuffer->GetRow(). This one can be get_row_RGBA32(), get_row_RGB32() or get_row_RGB16().

[stippi]

It isn't what I thought (CopyPixelsIn/Out()) ... so anyone feel free to fix it.

[korli]

uses a 5 bit key value to be usable on all colorspaces

[korli]

This patch fixes the bad behavior on 16 bits colorspaces. Rendering is really done with a 16 bits colorspace buffer which is logical. Though on BeOS, it seems the rendering is done with a 32 bits colorspace buffer, which would explain why GLTeapot behaves correctly there. I think this problem is a bit implementation dependent, and GLTeapot bases its behavior on the implementation: this would be a good reason to patch GLTeapot. Thoughts ?

[stippi]

I understand the patch only vaguely, but I am pro fixing GLTeapot if it makes some wrong assumptions about pixel depths... :-)

Thanks for figuring this one out.

[korli]

I'll add a comment to this patch to explain the difference of behavior of rendering on Haiku.

[korli]

fixed in hrev23793

[phoudoin]

I think this problem is a bit implementation dependent, and GLTeapot bases its behavior on the implementation: this would be a good reason to patch GLTeapot. Thought?

You're right, GLTeapot made the assumption that rendering is always done in a 32 bits RGB pixels buffer, so no color space conversion rounding errors could occurs.

GLTeapot code is the faulty here, as such assumption in OpenGL should not be made. Notice, BTW, that BGLView don't offer a way to specify RGB colordepth. It's BGL_RGB or BGL_INDEX, period. On other platforms, there is often a way to ask for a 16, 24 or 32bits colorspace.