my name is Marcus GroÃŸe and I am working in the field of 3d measurements using structured light. Our group
is evaluating the use of GPU’s for image processing tasks. In order to get to know OpenCL I have written a kernel, which
averages twentyone gray value images and writes the results into global device memory for later usage (see provided kernel code below).
The runtime of the kernel (which is measured using the clGetEventProfilingInfo) is about 113ms (GPU).
To get that fast I am using loop-unrolling as described here “http://developer.amd.com/gpu/ATIStreamSDK/ImageConvolutionOpenCL/Pages/ImageConvolutionUsingOpenCL.aspx” (about 10ms faster compared to non unrolled case).
An implementation on the CPU takes about (140ms, no loop-unrolling used and only one core used). So
for this problem there seems to be no big performance gain, when using the GPU. If have a few question related to that result.
- The problem may be that for every memory access there is only one addition made, so that the memory bandwith hinder a faster execution. Is this plausible?
- As image dimension (global buffer dimension) is a multiple of 16 memory accessed should be coalesced in my implementation. Is there a way to check this or can someone point me
to problems in my kernel-code that surpress coalesced memory access?
- Are there other options to decrease execution time?
- We use a NVIDIA-Geforce 9500GT. When switching to a more recent model (perhaps the upcoming Fermi-Cards), which
speed-up may be achieved for this presented problem (factor >10?)?
- I adresse the same problem, using image_2d and image_3d instead of the one dimensional buffer. The runtime is about the same compared to using two buffers.
Questions not related to results.
- I am also eager to see more examples written in OpenCL, which handle image processing. Perhaps someone can
point me to a link or book?
- If I do not assign the local variable avgl,…,avgl3 to the global buffer avgL the GPU seems to skip the entire
calculation of theses values which makes it difficult to track memory read/write time consumption compared to calculation time consumption.
Is there a work around?
thanks in advance,
I am using CUDA-Toolkit 3.0 + NVIDIA 9500GT
//l contains image data of one camera, r contains image data of a second camera, average values are computed for both cameras (stored into avgL and avgR)
__kernel void AverageKernel(__global float* avgL,__global float* avgR, __global float* l, __global float* r)
//get position of workitem in image
unsigned int nx = get_global_id(0);
unsigned int ny = get_global_id(1);
//variables used for loop unrolling
//average calculation of 21 images of size 640x480
//writing results to global device memory
P.S.: I posted the same topic here “http://www.khronos.org/message_boards/viewforum.php?f=37”, I hope to get more feedback by posting it here