"Faster" texture internal format

OpenGL Development JOGL Development
#1

Hi all,

I’m trying to speed up my application which makes heavy use of texture loading, and I was experimenting with texture internal format. At first, for loading a 256x256 INT_ARGB BufferedImage to a texture using GL_RGBA internal format, GL_BRGA format and GL_UNSIGNED_INT_8_8_8_8_REV type, I got times around 10 ms (in a MobilityRadeon 9000). Since transparency wasn’t the issue for me, I used 256x256 INT_RGB buffered images and textures using GL_RGB internal format, GL_BRGA format and GL_UNSIGNED_BYTE type. By using this configuration, texture loading time was reduced to 5ms. (In all cases, the time measured was the time of the single glTexImage2D call).

Since I’m not really an expert concerning the various OpenGL internal formats and data types, is there a better image type/internal format/format/data type configuration for faster texture loading times? What is the configuration you people use most for loading textures?

Also, all the above tests were conducted with the old texture handling way, without using the TextureIO related API intruduced in latest JSR231 versions. Is using this API going to improve texture loading times in any way?

TIA
N

#2

There’s no one hard and fast rule. Every graphics card has different optimised paths. What is fastest on one card may be slowest on another, even within cards produced by the same company.

The TextureIO classes should make things a lot faster because they avoid the need to bring the texture contents up into Java.

#3

Does your application load a lot of textures or reload a lot of textures? There are fundamental performance differences between glTexImage2D and glTexSubImage2D and if you don’t need to change the size of a texture but need to continually reload data into it then you should be using glTexSubImage2D repeatedly and not glTexImage2D.

The TextureIO classes have fairly efficient code paths but are more focused on correctness. The TGA loading path through that code is probably the fastest.

#4

Speaking of TGA, which is very OpenGL-friendly
because of its flipped Y-axis… However, suppose
one has to work with JPGs, what is the preferred
way of flipping the loaded image since JPGs are
not flipped on the Y-axis like TGAs.

I have to work with many large JPGs all the time
so I avoid creating an extra copy of BufferedImage
and flipping it in code by flipping the V texture coord
instead but this must be handled carefully…

Any better ideas?

.rex

#5

The TextureIO classes perform the flip scanline-by-scanline so avoid having twice the storage allocated at any given time. This isn’t as fast as just flipping the texture coordinates vertically but makes things much easier on the application.