Next Move

CUDA

• Using GT 640. PTX instructions. VLIW aware kernel programming

SSE

• Any benefit using store without cache ?
• Calculate throughput.
• EB = ( Br + Bw ) / T , EB : effective bandwidth, Br ; read bytes, Bw : written bytes, T : time

HW

• Study hardware - cache, bank and so on

OpenCL

• Using NVidia OpenCL with GT 640.  Compare result with CUDA. Using buffer. Using Image.
• Using NVidia OpenCL with GT 640.  vector type 
• Using Intel OpenCL. Compare result with SSE code
• Run kernels on GPU GT 640 and CPU G2120 at the same time.
• Buy 3rd generation processor i3 - 3220 - to access GPU HD2500.

Inspection algorithm

• Image preprocessing
• Geometric Distortion corretion - radial in area camera
• Shading correction - Area, line scan. How to define the compensation ?
• periodic pattern inspection - 2 points ( hor, ver ), 4 points, 8 points, many points in horizontal and vertical
• threshold - constant, adaptive threshold, ...
• binary image handling
• morphology - open, close, kernel size, iteration count, iteration order
• segmentation - contour based, region based, bug follower, area, bounding box calculation

Distributed system

• Group communication to manage multiple inspection processor
• Join/Leave the group
• Send packet that goes out to all group member fast and reliable using broadcast or multi-cast. 
• Designated member can get the burst of REP from all the group member. 
• A processor can join more than 1 group
• Streaming images to a member - or designated member
• REQ, REP, ACK, AYA( Are you alive ? ), IAA ( I am alive ), TA( Try Again ), AU ( Address unknown )
• Make API that is flexible and easy to use ?  Allow to have different type of REQ possible ?  Refer other API for network ?