Hello everyone!

I've been recently porting my Direct3D engine to OpenGL so it could run on platforms other than Windows, but it didn't work as intended. After a lot of debugging (using ApiTrace) I'd decided to try to run it on Ubuntu and guess what? It worked exactly like I wanted it to. Earlier this afternoon I tried to use llvmpipe (a software rasterizer) on Windows and it ran just like it did on Ubuntu.

Most objects seem to be rendered at the same place (maybe it's (0,0,0)?) and all of them have the same color so I'm thinking it may be a problem with the uniform buffers, but ApiTrace doesn't give any related error and the values seem OK.

This is how it looks like with llvmpipe:

[attachment=33261:Should.png]

This is how it looks like without it:

[attachment=33262:Windows.png]

I don't know if ApiTrace .trace files are machine independent, but I'm including one in case someone has time to take a look at it:

[attachment=33267:Test.zip]

Now, there's something very important that I should say. My laptop has Intel integrated graphics which is known to have very crappy OpenGL drivers on Windows (supports 3.1), but the MESA driver on Ubuntu is much better (supports 3.3 core). Also, when on Windows it runs on a OpenGL 3.1 context, but on Ubuntu and when using llvmpipe it runs on a OpenGL 3.3 core context. I don't know if that could make a difference.

Does someone have any idea of what it could be? Could it be a driver bug? Could MESA drivers be allowing something that shouldn't be?

Thank you for reading. Best regards.

It sounds that your Intel Chip has some unsupported extensions you are using that the Ubuntu driver support so you may need to check the extension string and shader compiler output eve if it dosent fail there may be warnings in it. At least check the error code of GetLastError() Winapi. Remember that glGetError may be resetted when running any gl function after that one that fails so you need to double check this too.

This is classic driver incompatibility. Your using things on the Ubuntu side that your intel driver can't do. Do yourself a favor and test it on a proper windows gaming pc (nvidia or amd card) to verify.

The biggest hurtle I had when porting my XNA engine to OpenGL was to remember to check glGetError after each call in DEBUG mode so I knew exactly which lines were failing. There's a lot more error checking done in OpenGL that I was used to. I found that even the same code that rendered fine on OSX and Linux, showed up wonky on windows. My issue was driver defaults. On windows, my driver defaulted some state (blend state and cull states) that weren't done on OSX and Linux. I had to write a state manager for my engine to maintain it's own state and not rely on OpenGL's.

One would think that it would be easy to take a standard api and make something cross platform. Except there are small nuances with the platforms that you have to kinda navigate with your code.

EDIT: Also, test with something basic and small, set to an identity matrix, test across all platforms and see if it's the same. A simple cube at 0,0,0 should suffice. This will allow you to start to pinpoint where the differences are.

Hi again. Sorry for the late reply. Thank you to everyone who's trying to help me.

Do you have any immediate mode code at all?

I had to google a bit to remember what that meant. No, I don't have any immediate mode code.

Are you checking for errors in Ubuntu, even though it looks correct?

I'm not checking for errors through glGetError yet (because of lazyness) but I'm using GL_KHR_debug which reports even performance warnings.

Can you compile and/or run it as OpenGL 3.1 under Ubuntu to see if you get similar results as Windows?

That's tricky. The Ubuntu driver only supports 3.2 and 3.3 when using the core context. The greatest non-core version available is 3.0, so the best I could do would be having a 3.0 context and loading 3.1 extensions. I'll try that if nothing else works.

It sounds that your Intel Chip has some unsupported extensions you are using that the Ubuntu driver support so you may need to check the extension string

That's true. The Windows driver doesn't support geometry shaders, for example. But my engine checks if all needed extensions are present, so unless I'm missing an extension (which I don't think I am) that's also not it. I'm going to take a look at it anyway.

and shader compiler output eve if it dosent fail there may be warnings in it.

My engine always does that.

At least check the error code of GetLastError() Winapi.

I don't know what GetLastError has to do with anything, but I'll take a look at that.

Remember that glGetError may be resetted when running any gl function after that one that fails so you need to double check this too.

I don't think so. According to the reference "No other errors are recorded until glGetError is called, the error code is returned, and the flag is reset to GL_NO_ERROR.". Also, ApiTrace always lists all errors and warnings.

Do yourself a favor and test it on a proper windows gaming pc (nvidia or amd card) to verify.

If I had a "proper Windows gaming PC" I wouldn't be using a laptop. Decent computers are pretty expensive where I live, so it's not that simple. Remember: just because you can afford a good computer it doesn't mean everybody can. :)

The biggest hurtle I had when porting my XNA engine to OpenGL was to remember to check glGetError after each call in DEBUG mode so I knew exactly which lines were failing.

Yeah, that's something I should do now. Even though ApiTrace verifies errors after each call, it's better to do it on my own.

My issue was driver defaults. On windows, my driver defaulted some state (blend state and cull states) that weren't done on OSX and Linux. I had to write a state manager for my engine to maintain it's own state and not rely on OpenGL's.

My engine is designed not to assume any kind of default value.

Also, test with something basic and small, set to an identity matrix, test across all platforms and see if it's the same. A simple cube at 0,0,0 should suffice. This will allow you to start to pinpoint where the differences are.

THIS. For some reason, I haven't even thought about using a simpler scene. Doing this will probably make finding the problem much easier.

Again, thank you to everyone who replied. Have a nice day.

Hi.

I made some tests with a simpler scene. First, some code (this is a very simplified version of the code):

UpdateBuffer(Mesh, &Object->WorldMatrix, Material);
glBindBuffer(GL_UNIFORM_BUFFER, <UBO>);
char* Ptr = reinterpret_cast<char*>(glMapBuffer(GL_UNIFORM_BUFFER, GL_WRITE_ONLY));
if (Ptr)
{
    std::copy(<Buffer source>, <Buffer source> + <Buffer size>, Ptr);
    glUnmapBuffer(GL_UNIFORM_BUFFER);
}
glBindVertexArray(<Object's vertex array>);
glDrawArrays(GL_TRIANGLES, <First vertex>, <Vertex count>);

The code above is executed once per object. The results seem to change if I use glBufferSubData instead of glMapBuffer, so here's a "table":

UBO update method | 1 object | 2 objects¹ | 3 objects¹
------------------+----------+-------------------------
glMapBuffer       |    OK    |     OK     | WRONG (1)
glBufferSubData   |    OK    |  WRONG (2) | WRONG (3)

¹ It didn't matter if the objects used the same VBO or not.

What I mean by "wrong" depends on the combination.

For (1), it rendered the first object ok, the second one disappeared and the third one was rendered with the wrong VBO, but with the right world matrix and material.

For (2) and (3), it rendered only the last object, with the right world matrix and material, but with the wrong VBO.

Okay, we're getting closer. The UBO update method really matters. Let's try something.

UBO update method | Run until the UBO update for the 2nd object | Run until the UBO update for the 3rd object
------------------+---------------------------------------------+---------------------------------------------
glMapBuffer       |                      OK                     |             Same as (1) above
glBufferSubData   |               Same as (2) above             |             Same as (3) above

Hmm, interesting. So glMapBuffer is erasing the second object after it's been already drawn and glBufferSubData is erasing everything.

Now, as you may know, I have no f***ing idea what the f*** is happening. I feel like even if I manage to solve this issue, another issue even worse is going to appear, so that's why I'm officially giving up support for Intel Windows OpenGL drivers. It's just not worth it. I've spent two weeks trying to solve it. But that's not a big deal since Intel GPU users on Windows can still use Direct3D 10.1.

So thanks to everyone who tried to help me and sorry that I wasted your time by not being able to solve this issue. Best regards.