XNA Math and Access Violation

Warning! Some information on this page is older than 5 years now. I keep it for reference, but it probably doesn't reflect my current knowledge and beliefs.

# XNA Math and Access Violation

Wed
16
Feb 2011

XNA Math is a great math library bundled with DirectX SDK. I've written about it here: Introduction to XNA Math, XNA Math Cheat Sheet. Recently I've started coding some small raytracer using this library (especially XMVECTOR type) and I came across an error:

"Access violation reading location 0xffffffff"

Google didn't display many results when asked about this error in connection with "XNA Math" or "XMVECTOR", but I found a topic on our Polish gamdev forum: xnamath i dziwne crashe. Directed to other web places from there, I discovered that, just as I thought, this bug comes from misaligned address of the vector. XMVECTOR on PC is an alias for SSE __m128 type, which has to be always aligned to 16-byte boundary. Compiler seems to know about this alignment requirement and respects it whenever XMVECTOR is defined on the stack, passed as function parameter or returned from a function, but not when the vector is a member of some dynamically allocated class or struct. So this won't work:

struct Scene {
    XMVECTOR dir_to_light;
};

Scene *scene = new Scene();
scene->dir_to_light = XMVector3Normalize(
    XMVectorSet(1.5f, 1.f, -2.f, 0.f)); // Crash!

Adding special alignment declaration supported by Visual C++ - __declspec(align(16)) - doesn't help. it looks like operator new always aligns allocated objects to 8 bytes and there is no way to change this behavior other than redefining it to use your own custom memory allocator.

So I suppose the most simple and convenient way to walk around this problem is to use more "casual" types for storing vectors - the ones that do not use SSE and do not require alignment, like XMFLAOT3 (which is just struct with float x, y, z). Anyway XMVECTOR is designed to optimize computations when compiler is able to place it in the CPU registers. So my solution looks like this:

struct Scene {
    XMFLOAT3 dir_to_light;
};

Scene *scene = new Scene();
XMVECTOR dir_to_light = XMVector3Normalize(
    XMVectorSet(1.5f, 1.f, -2.f, 0.f));
XMStoreFloat3(&scene->dir_to_light, dir_to_light);

...

XMVECTOR dir_to_light = XMLoadFloat3(&scene->dir_to_light);
XMVECTOR n_dot_l = XMVectorSaturate(XMVector3Dot(normal, dir_to_light));

Comments | #directx #math Share

Comments

STAT NO AD
[Stat] [STAT NO AD] [Download] [Dropbox] [pub] [Mirror] [Privacy policy]
Copyright © 2004-2019