Curious why no hybrid MSAA/SSAA?

arjan de lumens · Jul 8, 2004

Does alpha-to-coverage work correctly when there are multiple layers of alpha-blended/tested polygons drawn on top of each other?

Xmas · Jul 8, 2004

arjan de lumens said:
Does alpha-to-coverage work correctly when there are multiple layers of alpha-blended/tested polygons drawn on top of each other?

Not quite, but that situation doesn't happen very often in those cases where alpha test is used currently. However, alpha to coverage has the same problem as alpha blending with magnification. If you get close to the edges, they will blur (and with a2c, you'll have banding, as you only have a small number of different coverage levels).

Scali · Jul 8, 2004

By keeping the test and the write in the same place we can enforce read-write coherency with relatively simple control and some tweaks to the required Z cache logic.

This was the vital part of information about hardware implementations that I was missing

I wasn't aware of the fact that the rasterizer could start on the next polygon before shading was complete. So I didn't know a problem could even exist there.

Anyway, if you would have some kind of 'dirty regions', where z was not updated yet, you could still draw (parts of) polygons that do not cover dirty regions... Perhaps the hierarchical zbuffers will incorporate this one day?
And why do I have the feeling that we'd be moving more and more towards tile-based (deferred) rendering

K.I.L.E.R · Jul 8, 2004

Couldn't you mystically generate a formula which will apply SS to all pixels while they are compressed?

This way they get bigger but not as big as if you were to uncompress everything.

OpenGL guy · Jul 8, 2004

Xmas said:
OpenGL guy said:

But if your compressed value depends on sample location...

Click to expand...

I don't get it. If you adjust the sample positions, then you're trying to put sample No. x at exactly the same location where it is in multisampling operation.

And even if that's not possible, it's not too bad. Leave the sampling positions where they are. Ok, your texture and geometry sampling points are not in the same location then, but this is the case with multisampling, too. It may give slight overlap artifacts if polygons with alpha test are connected to polygons without alpha test, but considering how alpha test edges look like, I think it is a worthy tradeoff for many games.

The problem is far worse than you think. If you think it's trivial, write a program that does everything you've described, give me the source code and I'll make it fail catastrophically.

Rendering with multisample mask certainly isn't the best thing for performance because it makes compression almost useless for those tiles affected, but it's better than supersampling for the whole frame.

Click to expand...

But, as I mentioned, you'd have to decompress the whole buffer, meaning very poor performance... worse than supersampling since supersampling would benefit from compression.

Click to expand...

You don't have to decompress the Z buffer if you don't mess with the sample positions, but only use multisample mask, do you?

If you don't move the sample positions, then probably not, however you're going to get weird artifacts on the edges on some objects. These weird artifacts can be very ugly.

Dark Helmet · Jul 9, 2004

Cryect said:
Also wouldn't any wise programmer batch his alpha tested polys all together as well as all of his alpha blended polys at the end?

Not necessarily. Consider ALPHA_TO_COVERAGE.

Xmas · Jul 9, 2004

Dark Helmet said:
Not necessarily. Consider ALPHA_TO_COVERAGE.

More correctly, it's GL_SAMPLE_ALPHA_TO_COVERAGE_ARB, and the name shows probably the biggest problem with a2c...

Xmas · Jul 9, 2004

OpenGL guy said:
The problem is far worse than you think. If you think it's trivial, write a program that does everything you've described, give me the source code and I'll make it fail catastrophically.

Is C#/MDX fine with you?

K.I.L.E.R · Jul 9, 2004

The problem is far worse than you think. If you think it's trivial, write a program that does everything you've described, give me the source code and I'll make it fail catastrophically.

Are you serious?

Why not make it a competition? See if anyone can come up with a way to make it work.

OpenGL guy · Jul 9, 2004

Xmas said:
OpenGL guy said:

The problem is far worse than you think. If you think it's trivial, write a program that does everything you've described, give me the source code and I'll make it fail catastrophically.

Click to expand...

Is C#/MDX fine with you?

C or C++ please... I don't want to have to learn a new language just for this

Dark Helmet · Jul 9, 2004

Xmas said:
Dark Helmet said:

Not necessarily. Consider ALPHA_TO_COVERAGE.

Click to expand...

More correctly, it's GL_SAMPLE_ALPHA_TO_COVERAGE_ARB, and the name shows probably the biggest problem with a2c...

How so? The thought running through my mind is it mandates the slow path (frag shader -> alpha value -> now we know which samples, so we can: stencil/alpha/depth test).

Xmas · Jul 9, 2004

OpenGL guy said:
C or C++ please... I don't want to have to learn a new language just for this

Well, then maybe you tell me what you need and I will write it. After learning MDX, I really don't want to go back to those clumsy COM interfaces.

Dark Helmet said:
How so? The thought running through my mind is it mandates the slow path (frag shader -> alpha value -> now we know which samples, so we can: stencil/alpha/depth test).

I meant the "GL_" part. There is no a2c in DirectX.

OpenGL guy · Jul 9, 2004

Xmas said:
OpenGL guy said:

C or C++ please... I don't want to have to learn a new language just for this

Click to expand...

Well, then maybe you tell me what you need and I will write it. After learning MDX, I really don't want to go back to those clumsy COM interfaces.

You should be able to just grab one of the MS DX9 SDK samples, rip out the rendering path and add in everything you've described above for alpha tested textures. Then put the source someplace and I'll show you how to break it

Xmas · Jul 10, 2004

OpenGL guy said:
You should be able to just grab one of the MS DX9 SDK samples, rip out the rendering path and add in everything you've described above for alpha tested textures. Then put the source someplace and I'll show you how to break it

Ok, just a very quick hack:
Take the "Billboard" sample and replace CMyD3DApplication:

rawTrees with this one:

Code:

HRESULT CMyD3DApplication::DrawTrees()
{
	bool bUseSuperSampling = true;
	D3DXMATRIX pMatrix, tMatrix, pnMatrix[4];
	float samplepos[4][2] = /*{ { -0.375f, -0.125 }, { 0.125f, -0.375f }, 
							  { -0.125f, 0.375f }, { 0.375f, 0.125f } };*/ // RG
	{ { 0, 0 }, { 0, 0.5f }, { 0.5f, 0 }, { 0.5f, 0.5f } } ;	// OG
	D3DVIEWPORT9 viewport;

	if(bUseSuperSampling)
	{
		m_pd3dDevice->SetRenderState(D3DRS_MULTISAMPLEANTIALIAS, TRUE);
		m_pd3dDevice->GetTransform(D3DTS_PROJECTION, &amp;pMatrix);
		m_pd3dDevice->GetViewport(&amp;viewport);
		// prepare shifted projection matrices for sample positions
		for(int j = 0; j &lt; 4; ++j)
		{
			D3DXMatrixTranslation(&amp;tMatrix,  2 * samplepos[j][0] / viewport.Width, 
				2 * samplepos[j][1] / viewport.Height, 0);
			D3DXMatrixMultiply(&amp;pnMatrix[j], &amp;pMatrix, &amp;tMatrix);			
		}
	}

    // Enable alpha testing (skips pixels with less than a certain alpha.)
    if( m_d3dCaps.AlphaCmpCaps &amp; D3DPCMPCAPS_GREATEREQUAL )
    {
        m_pd3dDevice->SetRenderState( D3DRS_ALPHATESTENABLE, TRUE );
        m_pd3dDevice->SetRenderState( D3DRS_ALPHAREF,        0x7f );//*** 
        m_pd3dDevice->SetRenderState( D3DRS_ALPHAFUNC, D3DCMP_GREATEREQUAL );
    }

    // Loop through and render all trees
    m_pd3dDevice->SetStreamSource( 0, m_pTreeVB, 0, sizeof(TREEVERTEX) );
    m_pd3dDevice->SetFVF( TREEVERTEX::FVF );
    for( DWORD i=0; i&lt;NUM_TREES; i++ )
    {
        // Quick culling for trees behind the camera
        // This calculates the tree position relative to the camera, and
        // projects that vector against the camera's direction vector. A
        // negative dot product indicates a non-visible tree.
        if( 0 > ( m_Trees[i].vPos.x - m_vEyePt.x ) * g_vDir.x + 
                ( m_Trees[i].vPos.z - m_vEyePt.z ) * g_vDir.z )
        {
            break;
        }
        
        // Set the tree texture
        m_pd3dDevice->SetTexture( 0, m_pTreeTextures[m_Trees[i].dwTreeTexture] );

        // Translate the billboard into place
        m_matBillboardMatrix._41 = m_Trees[i].vPos.x;
        m_matBillboardMatrix._42 = m_Trees[i].vPos.y;
        m_matBillboardMatrix._43 = m_Trees[i].vPos.z;
        m_pd3dDevice->SetTransform( D3DTS_WORLD, &amp;m_matBillboardMatrix );

		//***
		
		if(bUseSuperSampling)
		{
			for(int j = 0; j &lt; 4; ++j)
			{
				m_pd3dDevice->SetRenderState(D3DRS_MULTISAMPLEMASK, 1 &lt;&lt; j);

				m_pd3dDevice->SetTransform(D3DTS_PROJECTION, &amp;pnMatrix[j]);

				// Render the billboard
				m_pd3dDevice->DrawPrimitive( D3DPT_TRIANGLESTRIP, m_Trees[i].dwOffset, 2 );
			}
		}
		else
		{
			// Render the billboard
			m_pd3dDevice->DrawPrimitive( D3DPT_TRIANGLESTRIP, m_Trees[i].dwOffset, 2 );
		}
		//***
    }

    // Restore state
    D3DXMATRIXA16  matWorld;
    D3DXMatrixIdentity( &amp;matWorld );
    m_pd3dDevice->SetTransform( D3DTS_WORLD, &amp;matWorld );
    m_pd3dDevice->SetRenderState( D3DRS_ALPHATESTENABLE,    FALSE );

	//***
	if(bUseSuperSampling)
	{
		m_pd3dDevice->SetRenderState(D3DRS_MULTISAMPLEMASK, 0xffff);
		m_pd3dDevice->SetRenderState(D3DRS_MULTISAMPLEANTIALIAS, TRUE);
		m_pd3dDevice->SetTransform(D3DTS_PROJECTION, &amp;pMatrix);
	}
	//***

    return S_OK;
}

You have to set multisample type to 4 samples for this to work.

nelg · Jul 15, 2004

So, can it be broken?

OpenGL guy · Jul 15, 2004

nelg said:
So, can it be broken?

Haven't had time to look, but I've not forgotten about it.

Colourless · Jul 16, 2004

Doing that breaks gamma corrected AA. 'Nuf said

DemoCoder · Jul 16, 2004

Not on all HW. On some HW, GCAA can be implemented via pixel shaders.

OpenGL guy · Jul 16, 2004

DemoCoder said:
Not on all HW. On some HW, GCAA can be implemented via pixel shaders.

It doesn't even have to break when GCAA is done by the AA HW

Mintmaster · Jul 16, 2004

Xmas said:
And even if that's not possible, it's not too bad. Leave the sampling positions where they are. Ok, your texture and geometry sampling points are not in the same location then, but this is the case with multisampling, too. It may give slight overlap artifacts if polygons with alpha test are connected to polygons without alpha test, but considering how alpha test edges look like, I think it is a worthy tradeoff for many games.

This seems like a very serious source of artifacts to me. I think the best case scenario would be setting it up so that all samples in a pixel have the same Z value (i.e. the translation matrix offsets match the subsample offsets). Then you get just plain aliased intersections between alpha-tested and non alpha-tested polys. Better than no alpha-test AA, but it may not be good enough considering the perf hit of your solution, Xmas. Still, we all love having IQ options, right?

OpenGL guy, how does centroid sampling work when the multisample mask is set? Is the texture sampling location equal to the centroid of the enabled samples or all samples in the poly for this pixel?

Curious why no hybrid MSAA/SSAA?

arjan de lumens

Xmas

Porous

Scali

K.I.L.E.R

Retarded moron

OpenGL guy

Dark Helmet

Xmas

Porous

Xmas

Porous

K.I.L.E.R

Retarded moron

OpenGL guy

Dark Helmet

Xmas

Porous

OpenGL guy

Xmas

Porous

nelg

OpenGL guy

Colourless

Monochrome wench

DemoCoder

OpenGL guy

Mintmaster

Similar threads