Jump to content

  • Log In with Google      Sign In   
  • Create Account

Speed up loop, texture lookup, and array processing in pixel shader / post process (HLSL only)


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.

  • You cannot reply to this topic
6 replies to this topic

#1 Meltac   Members   -  Reputation: 439

Like
0Likes
Like

Posted 13 November 2012 - 07:20 AM

Hi everybody

There are many situations were I need to process an array / grid / matrix of pixels in a loop in a post process, mostly for calculating some mean or maximum value or a gradient of a certain set of adjacent pixels. Examples:

  • Gaussian blur filter
  • Calculating mean luminance of a specific area on the screen
  • Getting pixel nearest to the viewer within a certain range
  • Checking if at least one pixel of a 2-dimensional array (rectangle) of pixels meets a certain condition
  • etc.

All these applications are typically processed in a 2-dimensional for-loop iterating through the x- and y-axis (rows and columns) of the specified grid, where each iteration makes at least one texture lookup. Therefor they cause heavy GPU load as soon as their dimensions exceed a certain limit, making FPS drop dramatically in many situations.

So, are there strategies to speedup execution times immensely to overcome this limitations? I've already tried these:


But that's way not enough to make my shaders run well. Are there any better approaches, or maybe some special DX10/11 instructions being of help here?

Edited by Meltac, 13 November 2012 - 07:23 AM.


Sponsor:

#2 Ashaman73   Crossbones+   -  Reputation: 7991

Like
0Likes
Like

Posted 13 November 2012 - 07:44 AM

You need to change the filter logic, e.g. use filter separation to split up your filter into two passes (instead of n*n you have only n+n texture lookups).

Then try to utilize the hardware linear filters, that is you can access with a single texture lookup the linear sum of 4 texels or atleast 2 texels, this is very useful for blur filters.

With this simple tricks you can reduce a naive blur filter implemenation (9x9=81 texture lookup) to just 10 lookups.

Edited by Ashaman73, 13 November 2012 - 07:47 AM.


#3 Meltac   Members   -  Reputation: 439

Like
0Likes
Like

Posted 13 November 2012 - 07:54 AM

Thanks for the hints. Unfortunately more than one pass is not an option in my case (engine limitation).

How do I use the hardware linear filter? Do you have an example in HLSL for me?

Edited by Meltac, 13 November 2012 - 08:09 AM.


#4 Ashaman73   Crossbones+   -  Reputation: 7991

Like
0Likes
Like

Posted 13 November 2012 - 08:27 AM

Here's a simple example.

You want to add a weighted sum of texels (gaussian blur):
w1*t1+w2*t2+...
Now you can rewrite it to
(w1/(w1+w2)*t1) + (w2/(w1+w2) t2) *(w1+w2)
Now you can rewrite
w2/(w1+w2)
<=> (w1+w2-w1)/(w1+w2)
<=>(w1+w2)/(w1+w2) - w1//(w1+w2)
<=>1 - w1//(w1+w2)
Now substitue w1//(w1+w2) with alpha and you get
((alpha*t1) + (1-alpha)* t2) *(w1+w2)
Now, (alpha*t1) + (1-alpha)* t2 is a linear combination of two texels depending on alpha which is the standard hardware linear filter when accessing a value between two texels. You just need to multiply the resulting texel with (w1+w2).
Voila, two texel with 1 access. As homework you can try to do the same with accessing 4 pixels (x and y).:D

Edited by Ashaman73, 13 November 2012 - 08:28 AM.


#5 Meltac   Members   -  Reputation: 439

Like
0Likes
Like

Posted 13 November 2012 - 08:56 AM

Urgh, I would have hoped for something like a simple function call like

float3 mean_color = tex2Dlinear(myTextureSampler, uv).rgb

or

sampler2D gSampler =
sampler_state
{
Texture = <gTexture>;
MinFilter = Linear;
MagFilter = Linear;
MipFilter = Linear;
};

So do I have to do this by hand?

#6 Ashaman73   Crossbones+   -  Reputation: 7991

Like
0Likes
Like

Posted 14 November 2012 - 12:36 AM

With newer hardware you can gather 4 texels at once (glsl gather, I'm sure there's a hlsl equivalent), but.... to really optimize a shader or filter you need to do it by hand.

So do I have to do this by hand?

All the professional engines like UDK,CE,Unity are doing crazy stuff while keeping performance.. well, this is for sure not because of using command X instead of Y Posted Image

#7 Meltac   Members   -  Reputation: 439

Like
0Likes
Like

Posted 14 November 2012 - 02:16 AM

Ok, thank you so much!




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.



PARTNERS