Difference between revisions of "User:Catchchaos/GSoC2016/Proposal"
(GSoC Proposal for GPGPU Boolean Evaluation)
Latest revision as of 00:59, 5 May 2016
GPGPU Support for Boolean Evaluation
Name: Param Hanji
Email: email@example.com, firstname.lastname@example.org
IRC Handle: catchchaos
I’m a second year undergraduate student at the National Institute of Technology, Karnataka, pursuing a degree from The Department of Information Technology. I recently started getting myself involved in the wonderful world of open source and am looking to use this opportunity to really immerse myself in it. An AI enthusiast at heart, I’ve been teaching myself GPU programming in my freetime.
BRL-CAD has an excellent parallel Ray Tracing implementation, but use of the GPU to achieve it was introduced by Jon Engbert and then, recently developed by Vasco Alexandre da Silva Costa during GSOC 15. The purpose of this project is to build on this and introduce GPU parallelism for Constructive Solid Geometry (CSG) boolean operations. Existing GPU implementations are in OpenCL and I will continue to use the same due to its similarity to C (extension of C99), platform independence and extensive documentation.
Currently, OpenCL support exists only for a handful of primitives. My project will start by accelerating a few more (such as eto, hrt, part etc.) to get used to the existing codebase and give me time to learn more about ray tracing and CSG. Also the code has a few bugs as it is relatively new. Constantly looking for and fixing these bugs will also improve OpenCL support.
The next part of the project involves looking at the code responsible for boolean evaluation (largely in librt/bool.c) and identifying the functions that can and should be parallelized, including but not limited to rt_bool_eval, rt_boolweave, rt_boolfinal and others. Currently, they cannot be ported to OpenCL directly because of the presence of structure pointers, linked lists, goto statements and dynamic allocation of memory, features which are simply absent in OpenCL.
The boolean weave function used by BRL-CAD accepts lists of segments and partitions as input. These are internally stored as doubly linked lists. However, only arrays can be passed to OpenCL kernel functions. Several operations including traversing the lists, inserting at the front and back, deleting elements, etc. need to be altered if these lists are to be replaced with arrays. Also, several structures need to be redefined inside the OpenCL file so that they can be received and instantiated. Parallelizing weave can start with bool_weave0(), which is a special case of boolweave(), with similar inputs and similar operations, without the extensive branching that is present in the latter. Several macros used in both the functions like GET_PT_INIT, BU_LIST_FIRST, BU_LIST_DEQUEUE, INSERT_PT need to be either rewritten or eliminated.
Function boolfinal() uses gotos in its body. The first step will be to restructure the code to get rid of the goto statements. Very much like weave, this function too extensively uses linked lists. Again, the task is to represent all the relevant information in the form of arrays to be used by OpenCL. Boolfinal() is called by other functions in shoot.c and bundle.c. Once .cl file is generated, it needs to be linked, compiled and, run from these files, if and when OpenCL is detected on a particular system (much like it is done with the primitives currently).
Before doing that, miscellaneous functions like bool_eval(), partition_eligible(), bool_equal_overlap_tables() etc. That are used by boolfinal() need to be re-written in OpenCL. Some of these require a little more effort than others. For example, bool_eval() has a number of goto statements, while rt_bool_growstack() performs dynamic memory allocation. The report linked below maybe could be of help for the memory allocation. Otherwise, other options need to be explored.
- Roy Spliet, “A comprehensive study of Dynamic Memory Management in OpenCL kernels”, Master thesis report, Delft University of Technology, June 5th, 2013
- Patch containing .cl files for accelerating some primitives(eto, hrt)
- Patch introducing OpenCL support to rt_boolweave()
- Patch to extend OpenCl support for rt_boolfinal() by including rt_boolfinal() as well
- Patch to complete bool.c OpenCL implementation by including other miscellaneous functions
- Community Bonding Period: Learn up about general ray tracing concepts and CSG. Get a working knowledge of the various custom BRL-CAD structures (seg, partition, soltab, rt_i, region to name a few) relevant to boolean evaluation. Also enhance current knowledge of OpenCL with the intention of applying it to this project - like dynamic allocation, representation of structure pointers etc. Lastly, thoroughly read HACKING plain text file in the root folder of BRL-CAD.
- Week 1: Accelerate 1 among eto, hrt, part (or some other primitive) to learn codebase. Also, read up about how to represent relevant data as arrays for bool.c functions.
- Week 2: Complete acceleration of 1 or 2 more primitives. Submit patches incorporating the inclusion.
- Week 3: Start work on boolweave(). A good starting point is bool_weave0() as it is a simpler function, yet quite similar. In the meanwhile, profile the code to decide which functions need to be parallelized.
- Week 4: Finish up generating OpenCL code for weave and submit a patch.
- Week 5: Restructure bool_eval() to get rid of the goto statements and decide how to perform dynamic memory allocation for rt_bool_growstack().
- Week 6: Complete OpenCL implementations of the 2 functions. Also, profile the code to check for any other functions which need special attention.
- Week 7: Buffer week to finish any pending tasks. Test the code to make sure it works as required.
- Week 8: Start work on boolfinal() by eliminating goto statements and redefining macros.
- Week 9: Based on the earlier learning (for boolweave()), restructure the data to be processed as arrays instead of doubly linked lists.
- Week 10: Finish up accelerating boolfinal() and test it to ensure the new code works well with the rest of BRL-CAD code.
- Weeks 11 & 12: Compare the OpenCL code against the serial version. Check and measure the performance increase obtained. See how it matches up with the theoretical estimate. Perform code clean up as well.
My exams will get over in April. Hence, I might be a little inactive during the beginning of the community bonding period. I'm free and willing to put in 45-50 hour shifts (or more if needed) during the summer. I have vacations from the end of April to late July, which fits in nicely with GSOC timeline. Even after that, my first month of college is bound to be less stressful and I'll have plenty of time to work on this project. I'm extremely flexible with timings and am willing to adjust to my mentor's schedule if required. The only period of inactivity would be a week long vacation sometime in late April.
I first came across BRL-CAD when I was looking for GPGPU projects using OpenCL. I then started interacting with the community which I found really interactive and welcoming. My lack of computer graphics concepts was overlooked and I was directed to relevant resources whenever I required any. I submitted my first ever patch with BRL-CAD! The documentation for BRL-CAD is also extensive and helpful. I'm looking forward to continue my open source experience with BRL-CAD and have listed it as my one and only choice of organization for GSOC '16. I've learned a lot in such a short period and am looking forward to continue doing so.
I learned that a 20-25% performance improvement can be achieved using OpenCL for Boolean evaluation, which is significant. I've spent the last couple of weeks making a patch, getting familiar with the code and will continue to do so. My patch to accelerate epa was recently accepted. I'm also working on two other projects for my course in OpenCL on prime number generation and FFT.
The project on prime number generation is still being formulated. The basic idea is to implement the Sieve of Eratosthenes algorithm in a parallel manner. With a small set of predefined primes (less than 100), we hope to recursively call the function and assign the task of “crossing-off” multiples of a single prime to a single GPU. This will be written in OpenCL, using the PyOpenCL API within the next couple of weeks. The parallel implementation of Cooley-Tukey algorithm for FFT is almost complete and due for submission in a week.
I believe that I will be well equipped to tackle this project right in time for the start of the GSoC coding period. I've been constantly in touch with members of the community and will continue to do so. I'd love to use this opportunity to really start development with the core BRL-CAD team.