Bart/bullet3
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

joint (non-contact constraint) solver iterations is now working on GPU, but overall slower because of data copy.

Browse Source 
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
This commit is contained in:
erwin coumans
2013-07-06 14:11:32 -07:00
parent 3ca4d68f3e
commit b8d5cecfe3
5 changed files with 96 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
src/Bullet3Dynamics/ConstraintSolver/b3SolverBody.h
View File

@@ -51,6 +51,7 @@ struct b3SimdScalar
{
__m128 m_vec128;
float m_floats[4];
float x,y,z,w;
int m_ints[4];
b3Scalar m_unusedPadding;
};
@@ -105,7 +106,7 @@ operator+(const b3SimdScalar& v1, const b3SimdScalar& v2)
#endif
///The b3SolverBody is an internal datastructure for the constraint solver. Only necessary data is packed to increase cache coherence/performance.
B3_ATTRIBUTE_ALIGNED64 (struct) b3SolverBody
B3_ATTRIBUTE_ALIGNED16 (struct) b3SolverBody
{
B3_DECLARE_ALIGNED_ALLOCATOR();
b3Transform m_worldTransform;
@@ -125,6 +126,8 @@ B3_ATTRIBUTE_ALIGNED64 (struct) b3SolverBody
int m_originalBodyIndex;
};
int padding[3];
void setWorldTransform(const b3Transform& worldTransform)
{