MRT & Z Pre-Pass manually ?

What hardware/API are you targetting? In D3D11 with FEATURE_LEVEL_11, you can create a read-only depth stencil view for your depth buffer. That lets you keep the buffer bound for depth/stencil testing (you can't enable depth or stencil writes), while simultaneously sampling it in the pixel shader. For FEATURE_LEVEL_10_1 (DX10.1), your next-best option is to copy the contents of the depth buffer to another texture and use that for depth testing. For FEATURE_LEVEL_10 you don't even have that option, so you're pretty much stuck.

The target is D3D9 shader model 3.0, now I don't know what would be the best choice to do, because, I inevitably will draw the geometry twice because of the SSAO,
I thought that is possible to perform a depth test using the depth texture (thinking in shadow map), and discard pixels beyond the depth, but there will be z-fighting for
sure.

For dx9 you have to create an extra render target of Float32 and then you should output the depth in what ever space you need --like world space depth -- so that you can simply read it in later without reconstructing anything. In dx9 you cannot read from the depth buffer in the pixel shader. In dx10, you can create a depth buffer with the ability to sample it in the pixel shader.

Also, most early-z passes do not cause an increase in speed. You must have an extremely heavy pixel shader to get any benefit out of it.

Hi, I am about to pull the trigger with my first attempt to render to multiple targets, because I want to do SSAO,
the implementation with normals, that's why I am using MRT, the first one is for depth (D3DFMT_D24X8) the
second one is for normals (D3DFMT_A8R8G8B8).

D3DFMT_D24X8 is for use as a depth-stencil target, not as an MRT target. As mentioned above, you should use a floating-point, or ARGB8 texture here instead of a D24 one.

Or, if your card supports reading from depth-buffers, then you don't need MRT - just use the depth-stencil target.

I thought that is possible to perform a depth test using the depth texture (thinking in shadow map), and discard pixels beyond the depth, but there will be z-fighting for sure.

If the exact same geo is drawn with the exact same matrices, there won't be z-fighting.

In dx9 you cannot read from the depth buffer in the pixel shader. In dx10, you can create a depth buffer with the ability to sample it in the pixel shader.

You can read from the depth buffer in DX9 if the end-user has a DX10-level graphics card.

To do this, you make a texture with D3DFMT_INTZ instead of D3DFMT_D24X8. To check if this is supported you can use:const static D3DFORMAT D3DFMT_INTZ = (D3DFORMAT) MAKEFOURCC('I','N','T','Z');
bool supportsIntZ = d3d->CheckDeviceFormat(D3DADAPTER_DEFAULT, D3DDEVTYPE_HAL, displayMode.Format, D3DUSAGE_DEPTHSTENCIL, D3DRTYPE_TEXTURE, D3DFMT_INTZ);

Hi, Thanks

@Hodgman that DX9 secret capability is awesome, and yes, was my mistake is D3DFMT_R32F, I was thinking in depth stencil too much

Now I can read it directly, in fact the Z Pre-Pass is working now:

1 - Render to depth
2 - Clear only target, disabling Z Writting and ZFunc to LessEqual

Using 3 point lights, Z Pre Pass saves me about .1 ms on the second pass. Now that I also can read directly from depth, I was curious to make the
depth test by myself on the fragment program, first to ensure that I was using the right projective texture coordinates I did the follow specting to have
z fighting:


float2 Texcoords = 0.5 * In.Other.xy / In.Other.w + float2( 0.5, 0.5 );
Texcoords.y = 1.0f - Texcoords.y;
float dist = In.Other.z/ In.Other.w;

if(dist == tex2D( nolerpsampler, Texcoords ).r)
return float4(1.0f,0.0f,0.0f,1.0f);


And I was right:



Well, I thought it would be a matter of adding bias to the distance and do the test:


...
float dist = (In.Other.z + 0.001f) / In.Other.w;

if(dist > tex2D( nolerpsampler, Texcoords ).r)
return float4(1.0f,0.0f,0.0f,1.0f);



And, I was wrong:




Why? It is supposed to be (dist + bias), and not (dist - bias), why it works inverse to what I spected?

Thanks Again