Optimizing Iterating on Images

Chris · 03-13-2011 12:16 PM

Hi,

I have an algorithm which is split up into two parts:

Code:

REP(n) {

    if (image.lock()) { FREPA(someHugeArray) {

        someHugeArray[i] = read a set of pixels from image.color3D(...);

    } image.unlock(); }

    if (image.lock()) { FREPA(someHugeArray) {

        write a set of pixels to image.color3D(...);

    } image.unlock(); }

}

In the first part, the same pixel may get read. In the second part, the same pixel may get written. Is there any way to accelerate either/both of these parts with multi-threading (e.g. for a processor with 6 CPUs and 12 hardware-threads)? If so, what would be the best way?

Thanks

Chris · 03-14-2011 12:56 AM

1. Can I use MultiThreadedCall() to do this? and what's "ThreadMayUseGPUData();"?

I'm really not sure how to set this up properly to max out the processors - please advise - it's very important to get the best result possible for this. Thanks.

**Esenthel** · 03-14-2011 08:06 AM

Use image.data after lock to get direct memory access. Faster.
Yes you can use multi thread call
For the second function there are comments explaining it.

Chris · 03-15-2011 02:04 AM

Thanks! I found i'm getting about 3x speedup 80% load now on a quad core - this is great (this loss is caused because I have to push the writing of images to a single thread as the same pixel may get written by different threads, but the slowest part can be spread across multiple cores! - very happy), can't wait to get results from university lab hardware.

Q. Please can you expand on how to access direct memory for images of a custom resolution, in the practical example:

VecI res3d; Image voxels; voxels.createTry(res3d.x, res3d.y, res3d.z, IMAGE_B8G8R8A8, IMAGE_3D); // 32-bit image

Idealistically, for any given voxel, i'd like to read/write the first 24-bits as an unsigned int, and the last 8-bits as a Byte, but just having access to VecB4 would be lovely - or whatevers available.

Thank you so much,
Chris

**Esenthel** · 03-15-2011 10:03 AM

for b8r8g8a8 you can use vecb4:
image.lock
VecB4 *row=(VecB4*)(image.data()+y*image.pitch()+z*image.pitch2());
image.unlock;

Optimizing Iterating on Images
Author	Message