16-byte aligned matrix in DirectX9 and D3DX with Havok Physics 2012

Question

I want to ask a question about D3DXMATRIX and D3DXMATRIXA16, and how to use them with Havok Physics.

I'm using Visual Studio 2012 on Windows 8.

I successfully created an instance of hkpWorld, added an hkpRigidbody instance with hkpSphereShape.

The class representing the sphere object has:

hkpRigidBody m_RigidBody;  // Rigidbody
D3DXMATRIX m_Transform;    // Rendering matrix

Judging from the values of m_RigidBody->getPosition(), the simulation is running without problem.
So now obviously I want to render the sphere according to its position and orientation calculated by havok.

This is how I tried to extract the matrix for rendering:

const hkTransform& t = m_RigidBody->getTransform();
t.get4x4ColumnMajor(&m_Transform._11);

Then I happily ran the program, which would break in get4x4ColumnMajor and give me this message through my error report function:

C:\NightlyJobs\09-07-Sun-01\Source\Common/Base/Math/Vector/Sse/hkSseVector4.inl( 1831): [0x64211C2F] Assert: pointer must be aligned for SIMD.

I took this to mean that the matrix had to be 16-byte aligned, so I changed

D3DXMATRIX m_Transform;

to

D3DXMATRIXA16 m_Transform;

This time, the problem is still there but isn't 100% reproducible.
Sometimes it runs without problem, sometimes get4x4ColumnMajor fails with above message.

Changing the matrix declaration to

_declspec(align(16)) D3DXMATRIX m_Transform;

did not help either.

So, to summarize my question:

• How would I go about extracting the matrix suitable for rendering with DirectX9?
• Why do D3DXMATRIXA16 and _declspec(align(16)) seem to be inconsistent about being 16-byte aligned or not, every time I run the program ?

Answer 1

The main thing to remember is that _declspec(align(16)) only affects the struct/class layout, is respected for local or global variables, but has no affect on heap allocated memory.

In other words, it doesn't matter if the member variable of a class is marked with alignment if the class is allocated with new. Heap allocations by default with x86 (32-bit) are only 8-byte aligned, so you have a 50/50 shot of getting the required alignment that way. By default x64 native (64-bit) alignment is 16-byte.

Otherwise you need to make use of _aligned_malloc to get other alignment behavior from heap allocation. See MSDN for some more notes on this.