0
101 Sep 18, 2006 at 08:26

I wrote C++ code with D3D to do some general purpose computation by GPU in these days.I can not find efficient and accurate method when I want to compute the average value of all texels in a Quad-texture by GPU.
So do you have good idea ? Please teach me, thanks a lot.

#### 4 Replies

0
104 Sep 18, 2006 at 08:55

thats like a min filter, or box filter, im not sure if there is a quick way to do that, thats why mips were invented.
maybe someone else can help… if you would want to do it, you have to sample the input texture into the shader, once per each texel in the quad then average and output, thats the only way i can think of.

0
101 Sep 18, 2006 at 13:40

Thanks for reply. In fact, I got a method which can handle this question, but not efficent. My method is :

split the quad texture into 4 pieces like that

+

+

• |__1st_|__2nd_|
|__1st_|__2nd_|
+

+

• |__3rd_|__4th_|
|__3rd_|__4th_|
+

+

+

And I write a shader to render a quad which equal to the 1st pieces, with sampling 4 times for each 4 pieces. Output the average of the 4 sampling values into another quad texture which size is half of the origin one ( area is quater 1/4). After some steps, the sizes of texture are 1x1. So I get the result.

But this method is too slow to be not suit for my need. I wish that someone can teach me some more efficient method. So I post this thread.

0
104 Sep 18, 2006 at 14:09

the only thing i could think of to speed it up is to make it less accurate, if you want to do it completely accurately and quick im not sure if you can get both.
heres one way, but its way not perfect.
just reduce the size of the texture to 8x8, then just average that normally, it wont output perfectly, thats how i read they did average luminence checks in scenes.

0
165 Sep 18, 2006 at 15:45

Try using automatic mipmap generation and then reading back the value from the highest mipmap. However, your method of reading 4 pixels at a time in a shader will work better if you read 4 pixels that are right next to each other rather than 4 pixels widely separated in the image. For instance you would resample the image to half its dimensions by averaging over each 2x2 box of pixels. This is pretty much how automatic mipmap generation works anyway, but you can do it yourself if you need more control over the process.