Command line argument parser. TexuteCube now illustrates dynamic binding

[gleblebedev]

Dynamic binding seems to be broken in Direct3D11 and works fine in OpenGL and Vulkan.

Please run the TexturedCube.Desktop sample with one of the following arguments to test the issue:

-g Direct3D11 -g Vulkan -g OpenGL

[mellinoe]

The sample is working fine on D3D11 for me. What issue are you hitting?

[gleblebedev]

Do you see 3 boxes or just one?

[gleblebedev]

I've changed the sample to use dynamic binding

[mellinoe]

I see 3 boxes -- same as the other backends.

[gleblebedev]

Now that becomes interesting. Are you on Nvidia or AMD hardware? Mine is AMD.

[mellinoe]

I'm running on Nvidia right now. I can test AMD later today to see if that exhibits any issue. I'm assuming the problem is that only one cube is rendered?

[gleblebedev]

Yes. It looks like all use the same matrix for me as if offset isn't applied. Where is the method that applies the offset? I'll try to debug it now.

[mellinoe]

I can confirm that it's not working correctly on my AMD machine, so there's some problem there. I don't have all of my dev tools installed on that machine, so I'll need to set some of that up to get more info.

[mellinoe]

The code that actually applies the offset would be here: https://github.com/mellinoe/veldrid/blob/master/src/Veldrid/D3D11/D3D11CommandList.cs#L880

[gleblebedev]

In renderdoc I see vssetconstantbuffers1 is used to set the offset. The offset is 16 and 32 which is correct (16*16 = 256).

[gleblebedev]

What tool would you recommend to use to debug it?

[gleblebedev]

This looks strange. For the first box: Then the second one: And the third:

[gleblebedev]

It looks like pNumConstants is incorrect - it should be 16 or at least 48-firstContant position. Otherwise the end of the buffer exceeds the size of the buffer.

[gleblebedev]

Also the first box is binding to the whole buffer. Should it use VSSetConstantBuffers1 anyway, even with offset = 0?

public bool IsFullRange => Offset == 0 && Size == Buffer.SizeInBytes;

How Buffer.SizeInBytes is set? I don't see a way to supply size of the buffer.

[gleblebedev]

I see if I set DeviceBufferRange(_worldBuffer, 0, 64) to resource set it has now correct size and use VSSetConstantBuffers1 for the first box. But the result is still the same - 1 box on the screen :(

[gleblebedev]

If I set new DeviceBufferRange(_worldBuffer, 256, 64) is uses the second matrix as expected... Kinda...

[gleblebedev]

I found one more interesting thing. If i change first value in the render loop for (int i = "THISVALUE"; i < _worldMatrixOffsets.Length; ++i) then the corresponding box becomes visible....

[gleblebedev]

this looks like something related :)

https://docs.microsoft.com/de-de/windows/win32/api/d3d11_1/nf-d3d11_1-id3d11devicecontext1-vssetconstantbuffers1#calling-vssetconstantbuffers1-with-command-list-emulation

[gleblebedev]

I don't think I can simulate this without changing the Veldrid Direct3D11 code. I hope setting constant buffer to null will work :) May I ask you to test it?

[gleblebedev]

Starting the command list for every box definitely helps. So I'm 99% sure that's the offset not been reset for emulated command list.

[mellinoe]

Thanks for that MSDN link. Adding that workaround seems to resolve the issue on my AMD machine. Preview version 4.7.0-ge6647cf4a7 includes the fix for you to test locally. I will push out a stable version of 4.7.0 sometime soon, since that includes a number of other misc. fixes.

[mellinoe]

4.7.0 is uploaded now, so you can update this branch and test that out.

[gleblebedev]

Thanks a million!

[gleblebedev]

I can confirm - it works now. I've updated the PR.

[gleblebedev]

Ok, now the next issue. OpenGL uniform binding is broken at Veldrid.SPIRV version 1.0.12. I've updated the version to illustrate the problem.

It looks very much connected to https://github.com/KhronosGroup/SPIRV-Cross/issues/1107

Uniform binding can't find a uniform:

Generated OpenGL shader code:

version 330

layout(std140) uniform vdspv_1_0 { mat4 World; } _14;

layout(std140) uniform vdspv_0_1 { mat4 View; } _33;

layout(std140) uniform vdspv_0_0 { mat4 Projection; } _41;

layout(location = 0) in vec3 Position; out vec2 vdspv_fsin0; layout(location = 1) in vec2 TexCoords;

void main() { vec4 worldPosition = _14.World vec4(Position, 1.0); vec4 viewPosition = _33.View worldPosition; vec4 clipPosition = _41.Projection * viewPosition; gl_Position = clipPosition; vdspv_fsin0 = TexCoords; }

As you can see the names are broken now.

[gleblebedev]

I guess that's the reason for the regression: https://github.com/mellinoe/veldrid-spirv/commit/05ac7865f12f6fb345e96bcadc8b4617b4c7c442

[gleblebedev]

I guess there suppose to be code for OpenGL that binds uniforms based on location but it's missing in 4.7 release :(