Non-local Means Acceleration With Cuda

Performance testing usage:

11 input arguments (inlcuding executable) ./nlm_cuda Seed(0 for random) kernelsToRun(e.g. 1010) N M patch(3/5/..) patchSigma filtSigma blockSize timesToRun printStats

kernelsToRun: Specifies which kernels to run in binary, numbered from 1 to 4. 1 to run 0 to not.

• 1: Using Global memory
• 2: Using Global memory coalesced
• 3: Using Shared memory
• 4: Using Shared memory coalesced

N/M: Array dimensions.
patch: Patch window side size, must be odd.
patchSigma: Variable for the Gauss patch intensity.
filtSigma: Variable for the filter's intensity.
blockSize: How to break up pixels into thread blocks. Array x dimension MUST be multiple of this for shared algoriths (3-4)
timesToRun: How many times to run the kernels. Usefull for stats and averaging.
printStats: If 1, prints some usefull stats from the time measurements.

File filtering usage:

9 input arguments (inlcuding executable)
./nlm_cuda csvToRead N M patch(3/5/..) patchSigma filtSigma blockSize csvToWrite

csvToRead: Path to the input csv array
csvToWrite: Path to the output csv array

For usage with actual images

Use the readApplyNoiseAndNLM matlab script.