Index buffer offset not being set with dynamic index buffer and indirect call

[oxheron]

Describe the bug
A clear and concise description of what the bug is.

To Reproduce
Build https://github.com/oxheron/bgfx
Run example 01-cubes
Use renderdoc or similar and look at the order of vertices emitted between the first call
and the second. You will see that it is different. If you look in the source code however you will see the vertex buffer changes between each call, and the index buffer is supposed to change to accommodate this. However, it doesn't.
Yes this is a really scuffed way to validate this, but I didn't have much time to finish this. I will try and make a better example if needed.

Expected behavior
I expected the vertices order to be the same between each call as the index buffer would change order to go along with the vertex buffer changing order.

Screenshots

[oxheron]

Sorry if the screenshot quality is bad

[oxheron]

It doesn't look like there is going to be an easy fix for this in opengl, as the functions to offset buffers don't work on instance buffers, and glMultiDrawElementsIndirect doesn't have an offset either. However Im pretty sure a pure bgfx solution would work in Vulkan and D3D

[oxheron]

And this may not be specific to dynamic buffers either, just any index buffers that are subbuffers & need to be offsetted.

[bkaradzic]

Can you fix your example to actually work as any other example with shaders binaries being in runtime instead in source directory. I just ran example you provided and it just crashes.

[oxheron]

Sorry. Thought that was a build thing not working for me. Ill fix it

[oxheron]

I cant build the hlsl or metal shaders so ill just write a script in the cubes dir to do that if you're using windows. Also, so far this issue has only occurred on OpenGL and Vulkan.

[bkaradzic]

GL shaders + SPIRV are enough, I'm looking at GL bug, SPIRV just for VK reference.

[oxheron]

Should be good now.

[bkaradzic]

It doesn't work...

If you just name your shaders vs_, fs_, cs_ it would build with provided makefile, and it would just output where it needs to be.

[oxheron]

ok. cant see why but, ill fix that now.

[bkaradzic]

I'm also on Linux, so .bat doesn't work there.

[oxheron]

yeah, that was just for windows. cause i didn't see your message about D3D till id already finished it.

[oxheron]

hopefully it works now

[bkaradzic]

Failed to load shaders/glsl/vs_bug.bin

[oxheron]

Did you build the shaders?

[oxheron]

Cause I just put the names so they should build with the system.

[bkaradzic]

diff --git a/examples/01-cubes/fs_bug.sc b/examples/01-cubes/fs_bug.sc
index 4b544fe..226f516 100644
--- a/examples/01-cubes/fs_bug.sc
+++ b/examples/01-cubes/fs_bug.sc
@@ -1,7 +1,5 @@
-vec4 vec4_splat(float input)
-{
-       return vec4(input, input, input, input);
-}
+
+#include <bgfx_shader.sh>
 
 void main()
 {
diff --git a/examples/01-cubes/vs_bug.sc b/examples/01-cubes/vs_bug.sc
index 36c3aa7..2748563 100644
--- a/examples/01-cubes/vs_bug.sc
+++ b/examples/01-cubes/vs_bug.sc
@@ -1,6 +1,8 @@
 $input a_position
 
+#include <bgfx_shader.sh>
+
 void main()
 {
-       gl_Position = vec4(a_position, 1);
+       gl_Position = vec4(a_position, 1.0);
 }

[bkaradzic]

Ok, so now when I run GL I should see red triangle. And when I run VK I should see something else?

[oxheron]

No, its a really hard bug to show

[oxheron]

But it only shows up in a rendering debugger. If you read my initial comment youll see what I mean. Also, after reading that if you have a better idea on how to confirm a bug like that then Ill implement it.

[oxheron]

Just visually i mean

[bkaradzic]

Ok, I'll try to understand what's going on... It's usually better to have repro to be obvious when something is wrong.

[oxheron]

yeah no, sorry. I 100% agree. I will say, I think my thought is that when you use dynamic index buffers bgfx packs all of the data into one index buffer, and an offset is used to make sure the right buffer is selected. But when you do a indirect call there is no way for this offset to be set for index buffers (because all index buffer offsets are part of the indirect call, which is created on the gpu). I can't say why it seems to get the vertex buffer right as I would assume this issue happens on the vertex buffers as well (but it doesn't).

[oxheron]

I implemented a potential temporary fix here, on the master branch https://github.com/oxheron/bgfx, if you want to take a look at it. It just provides an interface for users to get the index buffer offset, so it puts it on the users to add that into their indirect buffer calls. This should probably be tested more, but it worked for me on Vulkan so far (and seems that it could be a just bgfx side call anyways).

[oxheron]

I don't know if this would be a good long term fix or not however.