Add the contact point from MPR, in addition to SAT/clipping contacts. Added a new kernel to clear/reset the number of contacts in pairs (stored in the z component)
Always sample unit sphere directions, if there are more edge-edge combinations than unit sphere directions (162 by default)
Remember last running demo for Bullet 3 (and save it in a text file, Bullet
Enable the testFileFracture.bullet in the Bullet2FileDemo
Break up clipHullHullConcaveConvexKernel into multiple stages, so it might 'fit' in Apple's OpenCL implementation
Implemented bvhTraversalKernel and findConcaveSeparatingAxis on CPU (debugging, possible future CPU version)
integrateSingleTransform
fix bug in registerRigidBody (could lead to random crashes, especially when performing picking/adding rigid bodies afterwards)
Remove constructor from b3Vector3, to make it a POD type, so it can go into a union (and more compatible with OpenCL float4)
Use b3MakeVector3 instead of constructor
Share some code between C++ and GPU in a shared file: see b3TransformAabb2 in src/Bullet3Collision/BroadPhaseCollision/shared/b3Aabb.h
Improve PairBench a bit, show timings and #overlapping pairs.
Increase shadowmap default size to 8192x8192 (hope the GPU supports it)
remove duplicate data in b3Contact4 (now in btContact4Data shared between CPU/C++ and OpenCL)
OpenCL kernels use #include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
Increase number of batches back to 250 (from 50), need to fix this hard coded number (see https://github.com/erwincoumans/bullet3/issues/12)
Work towards GJK/EPA, in addition to SAT/clipping (early on)
and embed the included files directly in the stringified files.
We need this, because we start sharing struct definitions and code between C/C++ and OpenCL (and potentially other languages)
preprocessor is from http://github.com/willsteel/lcpp
Use tetrahedra instead of barrel for convex demo (until performance for edge-edge is improved)
Increased #overlapping pair capacity from 12 to 16 / objec
improved parallel batching, don't try to write for static objects,
this fixed a bug, when the hash of a static object was identical with hash of dynamic objects, causing it to be assigned a bogus 100+i batching number
The parallel batching is still not enabled, because we need to measure the batching size (todo)
Will move joint setup to GPU, and then some benefit should be visible.
Don't use 64 alignment, it causes data structures size mismatch between cpu and gpu
implement CPU version of narrowphase convex collision, for comparison/debug purposes
start towards cpu/gpu sync, for adding/removing bodies (work in progress)
