I made two patches for OpenCL (OCL) shot code. One patch refactors the existing SPH (Sphere) shot code, and the another patch implements EHY (Elliptical Hyperboloid) shot code.
||fix OCL SPH shot routine compilation errors.
||EHY shot routine in OCL.
||ELL and ARB8 shot routines in OCL.
|refactor dispatcher, shoot, optical renderer to process many rays in parallel in C when rendering an image or block. |
|grid spatial partitioning in C. |#379
|grid spatial partitioning in OCL. |#379
||HLBVH object partitioning builder in C. traversal in OCL.
||GPU side database storage of OCL implemented primitives.
||port compute intensive or critical parts of the dispatcher,
|boolean evaluation, optical renderer to OCL. |
||OCL dispatcher that performs the shot routines for a whole frame.
||OCL rasterizer that does the pixel pushing for a whole frame.
||TOR and TGC shot routines in OCL.
||REC shot routine in OCL.
||Surface normal routines for all seven OCL implemented primitives.
||BOT shot routine in OCL.
||Simple BOT shot routine in OCL that computes triangle hits and normal brute force.
||CPU HLBVH BOT shot construction with OCL traversal.
The ARB8, ARS, BOT, EHY, ELL, SPH, REC, TOR, TGC, shot routines are in SVN trunk.
SVN trunk also contains solid database device storage and a render function which given a view2model matrix, width, height, can generate an RGB8 bitmap. Diffuse and Surface Normal light models are supported. The renderer does brute force first-hit ray tracing and ignores the CSG operators. It is integrated as a render option in mged.
Week 1 : 25-31 May
- Created some example .g files in mged for the primitives to be implemented this week. The Quick Reference Card proved to be quite useful.
- Do the matrix ops for EHY (Elliptical Hyperboloid) in the OCL side.
- Made patch for ELL (Generalized Ellipsoid) and ARB8 (Arbitrary Polyhedron) OCL shots.
- M1 complete: ELL, ARB8 shot routines in OCL.
- Tried out a bunch of code browsing tools (cscope, LXR, doxygen, etc). The NetBeans IDE seems the most promising.
Week 2 : 1-7 Jun
- Read code to better understand the main rendering loop. It seems to be something like this:
do_frame() → do_run() → worker()* → do_pixel()* → rt_shootray()* → rt_*_shot()
- The code is recursive (which is problematic for OCL). I'll work on a simplified version of the rendering loop which only does the primary rays in C as a first approach. After I get the non-recursive parallel friendly C code I'll work on the OCL port.
- Updated project proposal on Google Melange.
- SVN r65153 fails to compile with a bogus error of an unused variable that's actually being used its just that GCC 4.9.1 is too dumb to figure that out.
- Made simple ray generation code in C.
- Made simple frame buffer write code in C.
- Made simple diffuse shading code in C.
Week 3 : 8-14 Jun
- Added the main boolean weaving code to our minimal renderer.
- Eliminated some gotos and made the code more thread safe.
- The simple renderer patches are in the mailing-list.
Week 4 : 15-21 Jun
- Added OpenMP compile support. Use OpenMP constructs to launch the rendering threads. This work still has some bugs in it.
- Alpha M2 patch. in mailing-list.
- Read code to better understand the main spatial partition construction routines. They seem to be something like this:
rt_prep_parallel() → rt_cut_it() → rt_nugrid_cut()
- We need something less complex that is more amenable to porting to OCL. So I will be implementing the Lagae & Dutré compact grid construction algorithm published at EGSR. First I will program in ANSI C then I will port the code to OpenCL.
- Started work on M3: grid spatial partitioning in OCL.
- ANSI C Lagae & Dutré grid construction code.
Week 5 : 22-28 Jun
- Took time off from the project to go to the CGI'15 conference.
Week 6 : 29 Jun-5 Jul
- GSoC Midterm Evaluations.
Weeks 7-8 : 6 Jul-12 Jul, 13 Jul-19 Jul
- Evaluating algorithms for grid construction.
- Selecting OCL kernels we can use to support grid construction. It seems PyOpenCL has some kernels we could use. Now the question is how to extricate the OpenCL/C from the Python...
- ANSI C grid traversal code.
- OCL grid construction on the GPU.
- M3 complete: grid spatial partitioning in OCL.
Weeks 9-10 : 20 Jul-26 Jul, 27 Jul-2 Aug
- Implemented GPU side solid database storage infrastructure.
- The OCL EHY shot code now uses the GPU solid database instead of creating the input buffers on every call.
- The code allows the primitive to decide how it is stored without imposing a convention. So one can use SoA, AoS, or whatever to store the data.
- Implemented GPU side solid database storage for SPH, ELL, ARB8.
- Extracted out some duplicated OCL code.
- M4 complete: GPU side database storage of OCL implemented primitives.
- OCL TOR (Torus) shot routine. Includes the higher order equation solver code.
- OCL TGC (Truncated General Cone) shot routine.
- Put equation solver in separate .cl file.
- M6 complete: TOR and TGC shot routines in OCL.
- General overhaul and cleanup of the OCL shot patches.
- Upgraded NVIDIA OpenCL drivers on my computer to CUDA 7.0.
- I drew up a tentative algorithm for the CSG raytrace:
# GPU execution
lengths = COMPUTE_LEN_SEGMENTS(rays, db)
segs = ALLOC_SEGMENTS(lengths)
segs = COMPUTE_SEGMENTS(rays, db, segs)
# CPU execution
waiting_segs = READ_SEGMENTS(segs)
# merge with CPU computed segments for non-accelerated primitives
finished_segs = RT_BOOL_WEAVE(waiting_segs)
partitions = RT_BOOL_FINAL(finished_segs)
pixels = VIEWSHADE(rays, db, partitions)
- This allows us to ultimately reuse the CPU code for boolean weaving, primitive normals, shaders, to have a 100% pixel accurate result. At the expense of a lot of memory traffic and CPU-side computation of some fairly maths intensive parts like the normal compute and shade. However I presently see no other way of having a 100% accurate result in the time we have available.
- Made OCL EHY shot code look exactly like the ANSI C version. Cleanups.
- Ran tests on OCL shot code to check for accuracy vs existing code:
# e.g. test for tgc
make tgc tgc
e tgc ; rt -o rt_tgc.pix # OR e tgc ; rt -o cl_tgc.pix
pixdiff rt_tgc.pix cl_tgc.pix | pix-fb
# test results:
arb8: pixdiff bytes: 777500 matching, 8932 off by 1, 0 off by many
ehy: pixdiff bytes: 760977 matching, 25443 off by 1, 12 off by many
ell: pixdiff bytes: 764588 matching, 21844 off by 1, 0 off by many
sph: pixdiff bytes: 736942 matching, 49490 off by 1, 0 off by many
tgc: pixdiff bytes: 783191 matching, 3241 off by 1, 0 off by many
tor: pixdiff bytes: 774138 matching, 12294 off by 1, 0 off by many
- The off by many problem with EHY is probably related to rounding errors with sqrt in OCL for NVIDIA using a different rounding mode than X86. It is possible to use PTX assembly, i.e.
asm("sqrt.rp.f64 %0, %1;" : "=d"(b) : "d"(a));. OCL 1.1 and over have no support for setting rounding modes without using inline assembly.
Week 11 : 3 Aug-9 Aug
- Sean applied patches #341 and #346 to SVN trunk.
- Got SVN write access.
- Applied patches #393 and #370 to SVN trunk.
- SVN trunk now has OCL shot evaluation for SPH, EHY, ELL, ARB8, TOR, TGC primitives.
- Refactored SPH (remove duplicate code, etc) and applied it to trunk.
- Move declarations to top level in order to eliminate duplicate code in trunk.
- Pass struct with primitive data to OCL as an initial step to an AoS device primitive database. Move constants into common.cl.
- Generic OCL solid shot handler. Refactored code to remove duplicates.
- Load large OCL vectors on demand to reduce stack footprint per function call.
- Fix memory leak on OCL loaded program source code.
Week 12 : 10 Aug-16 Aug
- Add inclusive and exclusive scan OCL code from PyOpenCL to trunk.
- Created a private branch for opencl.
- M2 commited to opencl branch: kludge up a simple rendering pipeline with grid spatial partitioning traversal acceleration.
- The simple ANSI C rendering pipeline only supports Lambertian reflection with a stock grey material to make things simpler. Example output for goliath.g:
- We have OCL shot tests for 278 of 288 solids. Only primitive in this database which is not ported yet is pipe.
- For future reference I get these timings for the above scene (one OCL kernel invocation per ray-primitive shot):
- SHOT: cpu = 421.568 sec, elapsed = 447.675 sec
- M4 commited to opencl branch: add device side solid database storage.
- Refactor opencl database storage on opencl branch.
- Backport solid database storage from opencl branch to trunk.
- Simple OCL rendering pipeline which brute force computes all shots and returns a silhouette of the view:
|one OCL kernel invocation per shot with BSP spatial partitioning
||OCL brute force computation of all shots (2429 solids, 929232 pixels) in one kernel call
|elapsed time: 375.68 sec
||elapsed time: 3.22 sec
- This is not an apples to apple comparison since the work done is a lot different. The brute force version ignores the CSG operators. We don't have OCL normal computation. But it's a way to gauge the possibilities here. If we implemented a BVH we could cut the iterations per pixel from 2429 to around log2(2429)*depth complexity where log2(2429)=11.25. In opencl branch.
- PS: The missile launcher tubes don't show up. The tgcs degenerate to recs. So need OCL rec shot to render this properly. Might also be an issue in other scenes.
- Add REC shot routine to opencl branch. Fixes the issues with the havoc missile launchers as seen above.
- Backport REC shot routine from opencl branch to trunk. Add checks for NULL results buffer.
- OCL normal computation for arb8, ehy, ell, rec, sph, tgc, tor.
- Backport opencl normal computation for arb8, ehy, ell, rec, sph, tgc, tor to trunk.
- Add diffuse, surface normals rendering light models to opencl branch. Screenshots:
|OCL Sphere (Surface Normals)
||OCL Sphere (Diffuse)
||OCL Havoc (Surface Normals)
|elapsed time: 0.05 sec
||elapsed time: 0.05 sec
||elapsed time: 4.20 sec
- Backport diffuse, surface normals rendering light models from opencl branch.
- On hindsight I think the grids are not a good option for BRL-CAD on the GPU. Spatial partitioning can result in duplicate shots. Shots of BRL-CAD primitives can be more expensive than simple ray-triangle shots. The alternative is to use mailboxing like the BRL-CAD ANSI C code currently does but this requires a lot of per thread memory which we can ill afford on a massively parallel architecture like a GPU. So I think we would be better served by object partitioning namely BVHs. Did another literature search to see if I could come up with some papers we could use for the boolean evaluation, BVH construction and traversal.
- As for the boolean evaluator. If we can compute this incrementally this will have a significant impact on memory loads and memory consumption.
- Retrofit HLBVH tree builder from pbrtv3 source into opencl branch.
- OCL BVH traversal in branch.
- For reference the OCL BVH can render the Havoc scene, as seen above, at elapsed time: 0.09 sec vs the 4.20 sec it took with the brute force code. i.e. it is around 45x faster for this scene. The advantage should increase for scenes with more solids.
- The HLBVH code has stabilized enough that I replaced the grids code with it.
- Added bu_pool memory pool allocator because it's useful for the HLBVH builder.
- HLBVH CPU builder code has landed on trunk.
- Integrated OCL rendering with rt (command line option "-z") and mged (diffuse and surface normals light models). Currently it is fill rate limited. Pixel pushing is done with view_pixel on a single CPU core. This should be done on the GPU. The framebuffer outputs should also support writing blocks of pixels. Currently they use line oriented output. Doing these changes would require breaking API compatibility.
Week 13 : 17 Aug-23 Aug
- Do heavy duty pixel pushing with the GPU. This speeds up rendering of Havok around 2-3x on my system. It should make even more of a difference in simpler scenes which are more fillrate than geometry performance limited. I figured out a way to do the code for this without actually breaking the API. I used a callback to get the framebuffer pointer.
- I redid the accuracy tests after reimplementing the raster parts of the code in OCL to check the accuracy. I got the same accuracy in surface normals mode as when we only computed the hit results in OCL with one kernel invocation per ray-solid intersection.
|elapsed time @ 972x956: 0.35 sec
||elapsed time @ 972x956: 0.08 sec
- This was the one primitive which had the most differences last time so I ran the test again.
ehy: pixdiff bytes: 760757 matching, 25663 off by 1, 12 off by many. I got similar results. So the pixel engine shouldn't be more innacurate than the regular one. What I did find out in surface normals mode was that the CPU code actually is showing hits with the side of the hyperboloid (see the blue dots in the figure at the left). Despite this view being top down. So maybe the GPU version is actually more accurate? The differences show a nice noisy pattern without obvious banding or moire so there don't seem to be any major issues with the hits, normals, and raster.
-z OpenCL command line option when running
- Rename table.cl to rt.cl.
- Replace branches in pixel writing with conditional moves.
- Refactor sub buffer code.
- Write depth buffer in network byte order.
- Removed scan code from PyOpenCL because of licensing issues. Good thing it wasn't being used anywhere yet.
- Remove malloc inside framebuffer grabber routine.
- Require OpenCL 1.2 or greater.
- Change OCL primitive packing routines to use memory pools.
- Initial bot, ars implementation. It just intersects all the triangles. No acceleration.
- Removed the, now unused, one kernel call per ray-primitive intersection routines.
- HLBVH bot construction and OCL traversal. Here's a screenshot:
|1 million triangles
|elapsed time @ 972x956: 0.29 sec (OCL)
|elapsed time @ 972x956: 17.49 sec (RT)
|elapsed time @ 972x956: 0.49 sec (RT bot kd-tree)
- All math operations are done in double precision FP.