• Create Account

Single big shader buffer for hardware instancing

Old topic!

Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

17 replies to this topic

#1lipsryme  Members

Posted 28 March 2013 - 06:51 AM

I saw this part from the battlefield3 presentation (http://dice.se/wp-content/uploads/GDC11_DX11inBF3_Public.pdf) on page 30 where they show how they transfer transformations of each instance to the shader using a single big buffer and a const buffer with indices to index this buffer at the right position using the SV_InstanceID semantic. Can someone tell me what exactly this buffer is ?

This part here:

Buffer<float4> instanceVectorBuffer : register(t0);


I'm not too familiar with buffers in directx11 beyond constant buffers...

Isn't register(t0) for texture slots ?  Also they seem to use 3x4 floats for their transformations, so does that mean they construct their transformation matrix inside the shader using translation, rotation and scale vectors ?

And last how would I set such a buffer in the DX11 API ?

#2Hodgman  Moderators

Posted 28 March 2013 - 07:27 AM

The MSDN documentation is very lacking in explaining these concepts...

IIRC:

The t registers are similar to the c registers, but they're designed for random memory access patterns (like texture lookups) rather than constant memory access patterns (e.g. it's assumed that every pixel will read the same constants, but may read different texels). You can bind both textures and DX11 buffer objects to the t registers and then load values from them in the shader.

You bind values to the t registers using *SSetShaderResources, which takes "shader resource views". You can create a view of a texture, which is the common case, but you can also create a view of a buffer.

As mentioned above, you should prefer this method of binding buffers to t registers when you're going to be performing random lookups into the buffer. If every pixel/vertex/etc is going to read the same values from the buffer, then you should bind it to a c register using the usual method.

Also they seem to use 3x4 floats for their transformations, so does that mean they construct their transformation matrix inside the shader using translation, rotation and scale vectors ?

No, in a regular transformation matrix that's been constructed from a traslation + rotation + scale, the 4th row/column (depending on your conventions) will always be [0,0,0,1] so you can hard-code that value in the shader to save some space in the buffer.

I'm still not as experienced with DX11 as I am with DX9, so I hope that's correct

Edited by Hodgman, 28 March 2013 - 07:32 AM.

#3lipsryme  Members

Posted 28 March 2013 - 07:35 AM

Yes I meant 3 x float4 (sorry for the confusing syntax   )

Alright thanks for the info.

EDIT: By the way this buffer then should be dynamic, correct? Since I need to put in or remove elements during rendering.

Edited by lipsryme, 28 March 2013 - 07:40 AM.

#4Hodgman  Moderators

Posted 28 March 2013 - 08:45 AM

Yeah by the looks of it, they'd potentially be updating the entire buffer every frame.

#5lipsryme  Members

Posted 28 March 2013 - 08:49 AM

EDIT: Nevermind got it fixed...my shader wasn't up to date when I ran it

Edited by lipsryme, 28 March 2013 - 09:30 AM.

#6kauna  Members

Posted 28 March 2013 - 09:45 AM

The advantage of a buffer over cbuffer is the size (128mb vs 64kb, although in typical scenario 2-4mb should be enough) which allows you to store a frame worth of data to the buffer, also it allows you to fill the buffer once so you'll minimize buffer updates to minimum. The minor inconvenience is that you'll be able to store only float data (probably you can get around this with some bit manipulation method).

The buffer is pratically a texture and you'd bind it as vertex shader resource.

float4x4 GetInstanceMatrix(uint InstID,uint Offset)
{
uint BufferOffset = InstID * ElementsPerInstance + StartIndex + Offset;

// InstID = InstanceID
// ElementsPerInstance, how many float4's there are per drawn instance (typically 3 for a single mesh if 4x3 matrices are used)
// StartIndex = location of the first float4 inside the buffer for the first instance
// Offset (used to retrieve bone matrices)

float4 r0 = InstanceData.Load(BufferOffset + 0);
float4 r1 = InstanceData.Load(BufferOffset + 1);
float4 r2 = InstanceData.Load(BufferOffset + 2);<span> </span>
float4 r3 = float4(0.0f,0.0f,0.0f,1.0f);

return float4x4(r0,r1,r2,r3);
}



Inside the vertex shader you may use the above code to retrieve the transform matrix for each vertex.

Cheers!

#7Matias Goldberg  Members

Posted 28 March 2013 - 10:16 AM

The advantage of a buffer over cbuffer is the size (128mb vs 64kb, although in typical scenario 2-4mb should be enough)

Another important advantage from a tbuffer over a cbuffer is that cbuffer suffer from constant waterfalling, which make them a horrible fit for scalable hw accelerated vertex skinning (indexing the constant buffer with a different index in each vertex).

When you're not suffering constant waterfalling, cbuffers can be faster though, for well.... constant data.

#8Sock5  Members

Posted 28 March 2013 - 02:32 PM

Ok this may sound stupid, but what about just using regular instancing?Has this method been benchmarked against regular instancing with a second vertex buffer?

>removed<

#9kauna  Members

Posted 29 March 2013 - 10:29 AM

Ok this may sound stupid, but what about just using regular instancing?Has this method been benchmarked against regular instancing with a second vertex buffer?

If you mean that regular instancing means using a second vertex buffer, I remember reading some years ago that one of the big companies wrote that shaders using a second vertex stream will take a performance hit. It had something to do with the fact that the vertex declaration structure gets pretty big.

I think that the problem with classic instancing is also, that you'll need to define a vertex declaration for each different vertex shader type where the parameters are different. Normally of course you would just have a transform matrix as instance data for the vertex shader, but what if you need also some other per instance parameters, such as color, inverse world matrix or some other parameter which makes your instance different from the others. When using second vertex stream you'll need a different vertex declaration for each case.

In the case of the generic buffer, you'll need just one vertex declaration since the vertex data won't change. After it is the job of the vertex shader to extract the desired data from the generic buffer. Of course you'll need to make sure that the buffer contains the expected data. Also, since the buffer can be pretty big you'll be updating data much less often to the graphic card, which reduces API calls. Also, the program side logic get simpler when you don't need to worry so much about "will the data fit to the buffer or not, do I have to make several draw calls instead of one".

Of course, under D3D11 the instancing just another parameter that you can access in the vertex shader (InstanceID) and after that you may read or generate the instancing data in the best way you can imagine, though vertex streams can't be randomly accessed.

Cheers!

#10lipsryme  Members

Posted 01 April 2013 - 08:20 AM

So..I've created the buffer like this:

	D3D11_BUFFER_DESC instanceBufferDesc = {};

instanceBufferDesc.Usage = D3D11_USAGE_DYNAMIC;
instanceBufferDesc.ByteWidth = sizeof(XMFLOAT4);
instanceBufferDesc.CPUAccessFlags = D3D11_CPU_ACCESS_WRITE;
instanceBufferDesc.MiscFlags = 0;
instanceBufferDesc.StructureByteStride = 0;

hr = this->device->CreateBuffer(&instanceBufferDesc, NULL, &this->instanceTransformBuffer);


is that correct so far ?

Now I'm a little clueless on how to create the shader resource view from this.
Any ideas ? Do I set the shader resource view description to NULL ?

This seems to work without errors:
	D3D11_SHADER_RESOURCE_VIEW_DESC shaderResourceDesc;



Now if the above is correct, how can I basically push float4's onto the buffer ?

Edited by lipsryme, 01 April 2013 - 10:53 AM.

#11kauna  Members

Posted 01 April 2013 - 06:01 PM

Use DeviceContext->Map(...) with D3D11_MAP_WRITE_DISCARD to get a pointer to an empty the buffer.

Fill the data (with memcpy for example) and the call Unmap. Don't write outside of the buffer! It will cause memory corruption and may even hang your computer.

It doesn't really make sense to use the buffer for one float, but I assume it to be a test. I used DXGI_FORMAT_R32G32B32A32_FLOAT as the buffer format since handling things as float4 makes maybe a bit more sense. You'll need at least 3x4 floats to store a typical matrix, but of course probably you'll want to allocate some megabytes of storage (within reason of course).

Cheers!

Edited by kauna, 01 April 2013 - 06:10 PM.

#12lipsryme  Members

Posted 02 April 2013 - 06:47 AM

Got it working perfectly thanks !

So the only way is to define a constant max size for this buffer inside the code...?

Worst case would be 3 x numInstances which could add up to quite a lot.

By the way I'm currently assigning each instance the 3 float4's inside a loop. I assume it would be more efficient to add them inside a structure which I then copy once using memcopy or is that neglectable?

I'm only doing map / unmap once, but I'm filling the data like this:

// Update buffer
for(int u = 0; u < 3; u++)
{
XMStoreFloat4(&pInstanceData[(numInstances * 3) + u], worldTransform.r[u]);
}


basically in the loop that goes through every instance and counts them to do the instanced draw call at the end.

Also am I correct in thinking that I am restricted to one specific material (texture and properties from const buffer) for every unique instance group ?

update: Just read about putting them inside a Texture2DArray...so for specific material properties would it make sense to also store them inside a buffer like I did with the transforms ? Might be overkill for e.g. booleans but still...

Edited by lipsryme, 02 April 2013 - 02:35 PM.

#13mrheisenberg  Members

Posted 03 April 2013 - 08:16 PM

Got it working perfectly thanks !

So the only way is to define a constant max size for this buffer inside the code...?

Worst case would be 3 x numInstances which could add up to quite a lot.

By the way I'm currently assigning each instance the 3 float4's inside a loop. I assume it would be more efficient to add them inside a structure which I then copy once using memcopy or is that neglectable?

I'm only doing map / unmap once, but I'm filling the data like this:

// Update buffer
for(int u = 0; u < 3; u++)
{
XMStoreFloat4(&pInstanceData[(numInstances * 3) + u], worldTransform.r[u]);
}


basically in the loop that goes through every instance and counts them to do the instanced draw call at the end.

Also am I correct in thinking that I am restricted to one specific material (texture and properties from const buffer) for every unique instance group ?

update: Just read about putting them inside a Texture2DArray...so for specific material properties would it make sense to also store them inside a buffer like I did with the transforms ? Might be overkill for e.g. booleans but still...

Wait, can I ask something - how does your draw call look like?I mean, do you call DrawIndexedInstanced with only 1 buffer?Or with 1 vertex buffer and 1 empty instance buffer/or having the ID inside the instance buffer data?

#14lipsryme  Members

Posted 04 April 2013 - 04:14 AM

I only use a vertex buffer yes, not using instance buffer at all but giving it per instance transformations via a float4 buffer that holds these.
They're then accessed inside the shader using the instanceID.

Edited by lipsryme, 04 April 2013 - 04:15 AM.

#15mrheisenberg  Members

Posted 15 April 2013 - 11:54 PM

I only use a vertex buffer yes, not using instance buffer at all but giving it per instance transformations via a float4 buffer that holds these.
They're then accessed inside the shader using the instanceID.

Hm I tried to do the same as you to see how it compares to normal instancing, however nothing leaves the vertex shader.The Shader Debugger detects that the sphere goes in:
However nothing gets rasterized.It can't be the view matrix, since it works on normal rendering, so it has to be something to do with the large buffer object.Did you ever experience such an issue?

#16lipsryme  Members

Posted 16 April 2013 - 04:28 AM

Hmm I do remember having some issues that no pixel shader was being executed....can't remember exactly what it was though...

Should look something like this:

	D3D11_INPUT_ELEMENT_DESC lo[] =
{
{"POSITION", 0, DXGI_FORMAT_R32G32B32_FLOAT, 0, 0, D3D11_INPUT_PER_VERTEX_DATA, 0 },
{"TEXCOORD", 0, DXGI_FORMAT_R32G32_FLOAT, 1, D3D11_APPEND_ALIGNED_ELEMENT, D3D11_INPUT_PER_VERTEX_DATA, 0},
{"NORMAL",   0, DXGI_FORMAT_R32G32B32_FLOAT, 2, D3D11_APPEND_ALIGNED_ELEMENT, D3D11_INPUT_PER_VERTEX_DATA, 0},
{"TANGENT",  0, DXGI_FORMAT_R32G32B32_FLOAT, 3, D3D11_APPEND_ALIGNED_ELEMENT, D3D11_INPUT_PER_VERTEX_DATA, 0},
};


fourth parameter is important here as this is the vertex buffer slot...

Also try to make sure the float4s are combined in the correct order.

I've had a lot of problems with this. In the end I had to transpose the matrix first and then pass them to the cbuffer because otherwise you'd loose the 4th row because

if you pass your 3 rows something like:

cbufferData.data[i] = worldMatrix.r[i];


you're passing only 3 vectors of this matrix and you will loose important data if you pass the rows instead of the columns.

That's why I changed it over to columns and it worked:

float4x4 GetInstanceTransform(uint instID, uint offset)
{
uint BufferOffset = instID * elementsPerInstance + startIndex + offset;

float4 c0 = InstanceDataBuffer.Load(BufferOffset + 0);
float4 c1 = InstanceDataBuffer.Load(BufferOffset + 1);
float4 c2 = InstanceDataBuffer.Load(BufferOffset + 2);
float4 c3 = float4(0.0f, 0.0f, 0.0f, 1.0f);

float4x4 _World = { c0.x, c1.x, c2.x, c3.x,
c0.y, c1.y, c2.y, c3.y,
c0.z, c1.z, c2.z, c3.z,
c0.w, c1.w, c2.w, c3.w };

return _World;
}

I don't see a way around that...but maybe someone does ?

Edited by lipsryme, 16 April 2013 - 04:29 AM.

#17mrheisenberg  Members

Posted 16 April 2013 - 07:59 AM

That's why I changed it over to columns and it worked

oh I see, cause in your other post you specify #pragma pack_matrix( row_major )

I'll try to get it to work tonight when i can to see what happens

#18lipsryme  Members

Posted 16 April 2013 - 08:04 AM

Yes but that only applies to float4x4 values coming from the cbuffer. It doesn't apply to what you use inside the vertex shader and also not for float4's.

Old topic!

Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.