D3D10 Shadow Mapping

For the weights, just use the Gaussian distribution function I posted link earlier in this thread. I use variance 0.75 myself, which seems to work ok. Also, normalize the weights so that they add up to 1.

Ok i made also gaussian blur with this results:

I computed Gaussian distribution using the formula found on wikipedia.

float GaussianWidth(float Sigma, float Displace)
{
	return	(1/(sqrt(2*D3DX_PI) * Sigma)) * (pow(exp(1.0f),-(Displace*Displace)/(2 * Sigma * Sigma)));
}

What do you mean "normalize" weights?

With Sigma = 1 and 9 samples (from -4 to 4), and 128x128 texture





What do you think about it? Should i blur more? Changing what? Samples number? Sigma?

I do not know why, but saving image using PIX (like last one) is different than saving images using STAMP:



Looks like miss anisotropic filtering (that is enabled)

For future reference if you hit "alt-printscreen" it will grab the current window's contents to the clipboard. Probably the best way to do screenshots and get it representitive of what you are seeing, IMO.

Your screenshots look, to me, like anisotropic filtering is turned OFF.

Well your blur could definitely do with looking a tad better. For a 9x9 blur you should get MUCH better results than that ...

Originally Posted by Goz

For future reference if you hit "alt-printscreen" it will grab the current window's contents to the clipboard. Probably the best way to do screenshots and get it representitive of what you are seeing, IMO.

Your screenshots look, to me, like anisotropic filtering is turned OFF.

But it's enabled, as you can see from screenshot grabbed from Pix...i will check better this, anyway.

Originally Posted by Goz

Well your blur could definitely do with looking a tad better. For a 9x9 blur you should get MUCH better results than that ...

I thought same thing but for now i do not have got ideas...mabye gaussian parameters?

Originally Posted by XVincentX

What do you mean "normalize" weights?

When you sum the weights, make sure they equal to 1.

could you post up the code for your 2 passes as well?

Sure.
Gaussian coefficients are computed using this forumula

float GaussianWidth(float Sigma, float Displace)
{
	return	(1/(sqrt(2*D3DX_PI) * Sigma)) * (pow(exp(1.0f),-(Displace*Displace)/(2 * Sigma * Sigma)));
}

Due to packing HLSL rules, coefficient are grouped in float4, in this way

cbuffer cTwiceFrame
{
	float4 SampleOffsets[SAMPLE_COUNT/2];
	float4  SampleWeights[SAMPLE_COUNT/4];
};

SAMPLE_COUNT = 8 (the nineth sample is taken directly in Pixel Shader).

Data is filled in this way

cTwiceFrame.SampleWeights[0].x = GaussianWidth(1.0f,-4.0f);
	cTwiceFrame.SampleWeights[0].y = GaussianWidth(1.0f,-3.0f);
	cTwiceFrame.SampleWeights[0].z = GaussianWidth(1.0f,-2.0f);
	cTwiceFrame.SampleWeights[0].w = GaussianWidth(1.0f,-1.0f);
	cTwiceFrame.SampleWeights[1].x = GaussianWidth(1.0f, 1.0f);
	cTwiceFrame.SampleWeights[1].y = GaussianWidth(1.0f, 2.0f);
	cTwiceFrame.SampleWeights[1].z = GaussianWidth(1.0f, 3.0f);
	cTwiceFrame.SampleWeights[1].w = GaussianWidth(1.0f, 4.0f);

While texture displacement:

cTwiceFrame.SampleOffsets[0] = D3DXVECTOR4(0,(-4.0f * 1/SMX),0,(-3.0f * 1/SMX));
		cTwiceFrame.SampleOffsets[1] = D3DXVECTOR4(0,(-2.0f * 1/SMX),0,(-1.0f * 1/SMX));
		cTwiceFrame.SampleOffsets[2] = D3DXVECTOR4(0,( 1.0f * 1/SMX),0,( 2.0f * 1/SMX));
		cTwiceFrame.SampleOffsets[4] = D3DXVECTOR4(0,( 3.0f * 1/SMX),0,( 4.0f * 1/SMX));

And make Pixel Shader. So reuse the same target and make same pixel shader with these

cTwiceFrame.SampleOffsets[0] = D3DXVECTOR4((-4.0f * 1/SMX),0,(-3.0f * 1/SMX),0);
	cTwiceFrame.SampleOffsets[1] = D3DXVECTOR4((-2.0f * 1/SMX),0,(-1.0f * 1/SMX),0);
	cTwiceFrame.SampleOffsets[2] = D3DXVECTOR4(( 1.0f * 1/SMX),0,( 2.0f * 1/SMX),0);
	cTwiceFrame.SampleOffsets[4] = D3DXVECTOR4(( 3.0f * 1/SMX),0,( 4.0f * 1/SMX),0);

Finally, the pixel shader:

float4 Blur(VSB_OUTPUT Inp) : SV_TARGET
{
    float4 c = 0;
    
    float4 SampleOnZero = SpotText.Sample(SamplText,Inp.Tex);

    for (int i = 0; i < SAMPLE_COUNT / 4; i++)
    {
        int i2 = i * 2;

        c += SpotText.Sample( SamplText, Inp.Tex + SampleOffsets[i2].xy ) * SampleWeights[i].x;
        c += SpotText.Sample( SamplText, Inp.Tex + SampleOffsets[i2].zw ) * SampleWeights[i].y;
        c += SpotText.Sample( SamplText, Inp.Tex + SampleOffsets[i2 + 1].xy ) * SampleWeights[i].z;
        c += SpotText.Sample( SamplText, Inp.Tex + SampleOffsets[i2 + 1].zw ) * SampleWeights[i].w;
    }
            
    return c + (SampleOnZero*0.39); //0.39 is GaussianWidth with 1.0f,0.0f.
}

So?

Try:

float GaussianWidth(float Sigma, float Displace)
{
	return	(1/(sqrt(2*D3DX_PI) * Sigma)) * exp( -(Displace*Displace)/(2 * Sigma * Sigma)));
}

exp( 1.0 ) IS 1.0

What you are saying is raise 1 to the power you should be rasing e too ...

Originally Posted by Goz

exp( 1.0 ) IS 1.0

Actually exp(0) = 1.0, and exp(1.0) = e.

Originally Posted by Reedbeta

Actually exp(0) = 1.0, and exp(1.0) = e.

goddammit reedbeta would you stop correcting my completely retarded ramblings hehehe

Still not convinced by thaty GaussianWidth function though.

You know btw that value for "-x" is the same as for "x"? ie you don't need to run it for every value.

Can you post up the values returned by GaussianWidth?

Originally Posted by Reedbeta

Actually exp(0) = 1.0, and exp(1.0) = e.

I just used it to avoid writing costans by my own.
Gaussian values coming soon.

Values for you

GaussianWidth(1.0f,-4.0f); 0.00013383027
GaussianWidth(1.0f,-3.0f); 0.0044318493
GaussianWidth(1.0f,-2.0f); 0.053990975
GaussianWidth(1.0f,-1.0f); 0.24197073

Your function version return quite same values, except for last 2 numbers.

Originally Posted by XVincentX

Values for you

GaussianWidth(1.0f,-4.0f); 0.00013383027
GaussianWidth(1.0f,-3.0f); 0.0044318493
GaussianWidth(1.0f,-2.0f); 0.053990975
GaussianWidth(1.0f,-1.0f); 0.24197073

Your function version return quite same values, except for last 2 numbers.

It should return identical values ...

I can't see anything wrong in your code .. the only thing that strikes me as being a potential problem (and I'm grasping at straws here) is your value for "SMX" ...

The variance value you are using is way too small. The Gaussian weight values are way too small for more distant samples thus they barely contribute to the value. Use the value 0.75 that I originally said for x=[-1, 1].

I just tried, but i did not find any noticeable improvement.




I was also forced to mul by 2 all spotlight becouse it was too dark

So what are the new values?

Sorry i forgot to paste them
4.6268983e-008
5.8531972e-005
0.0096201422
0.20542553

How about trying to do what I have been saying for quite some time here? Spoon feeding follows:

float weights[9];
float weight_sum=0.0f;
for(int i=0; i<9; ++i)
{
  float weight=GaussianWeight(0.75f, float(i-4)/4.0f);
  weight_sum+=weight;
  weights[i]=weight;
}
for(unsigned i=0; i<9; ++i)
  weight[i]/=weight_sum;

I'm sorry you should understand me.
When you work so much and too many time on the same project, sometimes happens to me to have fog and clouds on my head and forget all suggestion. I will try also this other.
Thank you still for the help.

With normalization the darkening is vanished, and i can see a better blur (even if little).

Anisotropic filtering, 128x128 variance shadow map (Gauss 0.75). The ligher is due to 2*light i used before

No anisotropic, 128x128 variance shadow map (Gauss 0.75).

Lit factor. Looks like perfect.


Shadow Map.

Thank you all for your help.