***Random Ball Cover (RBC) v0.2.6*** Lawrence Cayton work@lcayton.com (C) Copyright 2010, Lawrence Cayton [work@lcayton.com] This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program. If not, see . --------------------------------------------------------------------- SUMMARY This is a C and CUDA implementation of the Random Ball Cover data structure for fast nearest neighbor search on a GPU. The code implements the one-shot algorithm. See the following papers for a detailed description of the search algorithm and the theory behind it. * L. Cayton, A nearest neighbor data structure for graphics hardware. ADMS, 2010. * L. Cayton, Accelerating nearest neighbor search on manycore systems. Twenty-Sixth IEEE International Parallel and Distributed Processing Symposium (IPDPS), 2012. --------------------------------------------------------------------- COMPILATION Type make in a shell. Requires GCC and NVCC (CUDA). The code has been developed under GCC 4.4 and CUDA 3.1. --------------------------------------------------------------------- USE A sample driver is provided for the RBC. To try it out, type $ testRBC at the prompt and a list of options will be displayed. The sample driver can be used with either text or binary input. The text input format is one database element per line, with features separated by spaces. The binary input format is in floats, but can be changed to doubles. The output file format is a list of the queries' NNs, followed by a list of the distances to those NNs. Note that by default, all input and output is stored in single-precision (float) format. Basic functionality is provided through this driver, but I recommend integrating the RBC code directly into your code for the best results. For many applications, the RBC needs to be built only once, and then can be queried many times. The method requires a single parameter, the number of representatives. This parameter allows you to trade-off between search quality and search speed. The best way to set this parameter is to try a few different values out; a good starting point is generally 5*sqrt(n), where n is the number of database points. Use the eval option (-e) to print out the error rate. See the paper (Cayton, 2012) for detailed information on this parameter. The sample_input directory contains examples in both binary and text. The sample_db set contains 1024 points, each of which has 16 dimensions. The sample_query set contains 128 sample queries (which of course also have 16 dimensions). To try it out, run $ testRBC -X sample_input/sample_db.txt -Q sample_input/sample_queries.txt -n 1024 -m 128 -d 16 -r 128 or to try it out with the binary files, run $ testRBC -x sample_input/sample_db.bin -q sample_input/sample_queries.bin -n 1024 -m 128 -d 16 -r 128 Note that the -r 128 descibes the number of representatives, which controls the accuracy of the search. You might try varying this parameter to see the effects (there is nothing special about 128). You can print out the accuracy using by adding the -e switch; this will say the average number of the 32 nearest neighbors that were actually returned. --------------------------------------------------------------------- FILES * brute.{h,cu} -- implementation of brute force search (CPU and GPU versions) * defs.h -- definitions of constants and macros, including the distance metric. * driver.cu -- example code for using the RBC data structure. * kernels.{h,cu} -- implementation of all the (device) kernel functions, except those related to the scan (see sKernels below) * kernelWrap.{h,cu} -- CPU wrapper code around the kernels. * rbc.{h,cu} -- the core of the RBC data structure. Includes the implementation of build and search algorithms. * rbc_include.h -- header file to include in your driver. * sKernel.{h,cu} -- implementation of the kernel functions related to the parallel scan algorithm (used within the build method). * sKernelWrap.{h,cu} -- wrappers for the kernels in sKernel. * utils.{h,cu} -- misc utilities used in the code. * utilsGPU.{h,cu} -- misc utilities related to the GPU. --------------------------------------------------------------------- MISC NOTES ON THE CODE * The code currently computes distance using the L_2 (Euclidean) metric. If you wish to use a different notion of distance, you must modify defs.h. It is quite simple to switch to any metric that operates alongs the coordinates independently (eg, any L_p metric), but more complex metrics will require some aditional work. The L_1 metric (manhatten distance) is already defined in defs.h. * The k-NN code is currently hard-coded for k=32. It is hard-coded because it uses a manually implemented sorting network. This design allows all sorting to take place in on-chip (shared) memory, and is highly efficient. Note that the NNs are returned in sorted order, so that if one wants only, say, 5 NNs, one can simply ignore the last 27 returned indices. For k>32, contact the author. * The code requires that the entire DB and query set fit into the device memory. * Currently the software works in single precision. If you wish to switch to double precision, you must edit the defs.h file. Simply uncomment the lines typedef double real; #define MAX_REAL DBL_MAX and comment out the lines typedef float real; #define MAX_REAL FLT_MAX Then, you must add the compiler flag -arch=sm_20 to the NVCCFLAGS line of the Makefile (or sm_13 for older GPUs). Finally, do a $ make clean followed by another make. * For the most part, device variables (ie arrays residing on the GPU) begin with a lowercase d. For example, the device version of the DB variable x is dx. * The computePlan code is a bit more complex than is needed for the version of the RBC search algorithm described in the paper. The search algorithm described in the paper has two steps: (1) Find the closest representative to the query. (2) Explore the points owned by that representative (ie the s-closest points to the representative in the DB). The computePlan code is more complex to make it easy to try out other options. For example, one could search the points owned by the *two* closest representatives to the query instead. This would require only minor changes to the code, though is currently untested. * This software has been tested on the following graphics cards: NVIDIA GTX 285, GT 430, GTX 480, GeForce 320M, Tesla c2050 * This sotware has been developed under the following software setup: Ubuntu 10.04 (linux) gcc 4.4 cuda 3.2 It has also been tested under Mac OSX. Please share your experience getting it to work under Windows! * If you are running this code on a GPU which is also driving your display: A well-known issue with CUDA code in this situation is that a process within the operating system will automatically kill kernels that have been running for more than 5-10 seconds or so. You can get around this in Linux by switching out of X-Windows (often CTRL-ALT-F1 does the trick) and running the code directly from the terminal.