Sign in to follow this  

Normalmapping Performance

This topic is 1591 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

Is there any solution to the performance slowdown issue with normalmapping?

 

It's not even related to the specular calculation but just the texture fill itself.

 

Meaning if I have something like this:

 

vec4 colour = texture2D( F_color_texture, f_TexCoord );

 

vec3 normal = texture2D( F_bumpmap_texture, f_TexCoord );

 

final = colour * normal;

 

 

Then I experience the slowdown issue. Drops from 260 to 130 :(

Share this post


Link to post
Share on other sites
What units are those numbers in? I'm guessing they're frames per second?
To get some more intuitive numbers, you can calculate milliseconds per frame with ms = 1000/fps.
Which gives: 3.85ms per frame, increased up to 7.69ms per frame, or an increase of 3.85 (a doubling in cost).
 
If you look at two different shaders:
vec4 colour = texture2D( F_color_texture, f_TexCoord );
final = colour;
vec4 colour = texture2D( F_color_texture, f_TexCoord );
vec3 normal = texture2D( F_bumpmap_texture, f_TexCoord );
final = colour * normal;
Then it's obvious that the second one does twice as much work as the first one... so you should expect this pixel shader to be double the cost of the first one.

If that's hypothesis is backed up by actual measurements, then it means that the pixel shader is the bottleneck in your program.

What kind of GPU do you have? ~3ms to fill the screen with a basic shader (one texture fetch) is quite slow.
What kind of objects are you drawing? How many vertices/triangles are being drawn? How many pixels are covered? What format are the textures? Do the textures have mipmaps, and do you have mipmapping sampling states enabled?

Share this post


Link to post
Share on other sites

Yes, units in frames per second..

Well... here's the thing though..

 

Yeah it may seem like it does "twice as much work" - but this isn't exactly true.

 

For example, I also have a texture2D() call to look up my lightmap - *but it uses different texcoords* - so - you don't have the entire lightmap texture repeated many times everywhere (like the bumpmap), instead you just have small portions of the lightmap applied everywhere, so performance is actually negligent. (Basically the same as without a lightmap lookup)

 

But, like I just mentioned, the bumpmap is applied in full everywhere the regular diffuse color texturing is, since they share the same texcoords.

 

Not that it's extremely important, but I do have somewhat of a slower card.. Radeon HD 5450. Drawing the environment (3D FPS map). The map is only 2158/4336 vertices/triangles, and I don't know how many are on screen when I drop to 130 FPS, but it does depend on what's on screen. For example if I just face a wall I get 300+ FPS.. But if I go to one wide open section with a lot of long-distance wall exposed, then I get the drop to 130 (from 260 in the same section without the bumpmap). Resolution is a low 800x600. Textures are GL_UNSIGNED_BYTE GL_RGBA. No mipmaps, MIN/MAG == GL_LINEAR.

Share this post


Link to post
Share on other sites

I think maybe your GPU is stalling on things in other places in the rendering

There's no way for 2200 vertices to do that unless your vertex program is gigantic

I really don't think your GPU is that slow, simply because it's not an onboard GPU... tongue.png

 

Check your code for glGet*, occlusion result queries that aren't completed yet or in the same frame etc.

Anything that might be forcing GPU to flush

 

Keep in mind that mipmaps may actually increase fragment shader throughput :)

Edited by Kaptein

Share this post


Link to post
Share on other sites

It's definitely the bumpmapping... If I simply comment out "final*=normal;" then I get the 260 FPS :x

Share this post


Link to post
Share on other sites

Textures are GL_UNSIGNED_BYTE GL_RGBA. No mipmaps, MIN/MAG == GL_LINEAR.

 

This is your problem, just as Hodgman and Kaptein already said. Without mipmapping you are sampling texels from all over the place, whenever the texture is "scaled down" as in the far reaches of your "long-distance wall". This renders your texture caches ineffective and increases bandwidth.

 

This is not a bumpmapping problem. You have the same problem with your color texture and you will also experience it when you try to render particles.

Share this post


Link to post
Share on other sites
Sign in to follow this