[Direct3D] Handling Lost Devices

Hello and welcome to another topic about 3D programming.

Let's suppose we're programming in Direct3D9.

Today I'll be talking about one of the most annoying problems found in 3D applications: lost devices.

Lost Devices
First things first: what's a lost device anyway? When a device goes in a lost state it means it can't place it's results anywhere. It then isn't able to put it's buffers to any place on screen. This might happen when a user Alt Tabs out of a fullscreen game for example: the GPU then can't place it's frames anywhere anymore, so it becomes lost.

Lost devices can't be fully accessed by the Direct3D API anymore. For example, draw calls will return D3DERR_INVALIDCALL instead of D3D_OK (success). This means you can't tell the GPU to do anything useful anymore. To get it working again, you need to reset it.


Resetting your GPU Card
Just to assure you, you don't need to reset your PC or anything, you only need to empty it's memory manually by software (API). Remember, most of Direct3D's API isn't working anymore. There's just about five calls we can make:

• devicepointer->TestCooperativeLevel(). Think of this one as a doorbell. No one coming at the door (not returning D3D_OK)? Your GPU is lost.
• devicepointer->Reset(). This one will wipe GPU RAM and reset all States you've set (like SetRenderState and SetSamplerState).
• resourceinheritingfromIUnknown->Release(). You need to tell the GPU you don't need resources in GPU RAM anymore with this function. You can't even forget a single resource pointer: it will make your app crash when reseting.
• resourcebackedupincpuram->OnLostDevice(). Call this on any object that has a backup in CPU RAM.
• resourcebackedupincpuram->OnResetDevice(). Call this on any object that has a backup in CPU RAM.

As I've said, you need to empty the GPU's memory before you can reset it. One minor note first though: as any experienced PC user does *cough* you make backups *cough*:

• Resources put in D3DPOOL_DEFAULT are no-backup resources and can be found in the best RAM possible (GPU RAM). If we run out of GPU RAM, we put it in CPU RAM. If we can't store it there, leave it on the drive, crash, or put it in Page File. Clear enough.
• But we've also got D3DPOOL_MANAGED. These resources are copied to CPU RAM, and only when needed they get copied (not moved) to GPU RAM. This means there's always a backup available of these resources.

Let's put together a TODO-list of our reset then. What needs to be done is (in this order):

1. Release() any resources that are stored in GPU RAM (D3DPOOL_DEFAULT).
2. Put any resources that are stored both in GPU RAM and in CPU RAM (D3DPOOL_MANAGED) on hold and remove them from GPU RAM. Do this by calling OnLostDevice() on them.
3. Reset the device with devicepointer->Reset(D3DPRESENT_PARAMETERS).
4. Then tell all resources backed up in CPU RAM to copy back to GPU RAM by calling OnResetDevice() on them.
5. Recreate your resources that were put in D3DPOOL_DEFAULT.

An example

Let's say we've got ourselves a basic engine with the following resources:

• A GPU font, called ID3DXFont, used to draw our tooltip text on the GPU.
• The FX Framework, called ID3DXEffect, used to modify shader parameters.
• A shadowmap (color + depth), created by CreateTexture and CreateDepthStencilSurface. These are IDirect3D9Surface's and a IDirect3D9Texture.

In this case, the reset sequence will look like this:

void D3D::resetD3D() {

// Saveable resources in D3DPOOL_MANAGED

font->OnLostDevice();

FX->OnLostDevice();

// Unsavable resources that don't have a backup

ShadowTex->Release();

ShadowTexTopSurface->Release();

ShadowDepthTopSurface->Release();

// Let's throw everything away

d3ddev->Reset(&d3dpp);

// And copy the CPU RAM resources to GPU RAM

font->OnResetDevice();

FX->OnResetDevice();

// Recreate D3DPOOL_DEFAULT stuff

initD3D(1);

}

I hope this will be clear enough, and if not, tell me!

[Direct3D HLSL] Normal Mapping tutorial

For an explanation about why to use tangent space, read this tidbit of text.

Let's assume we're using Direct3D and HLSL.

Converting to Tangent (or texture) space

Normals stored in the texture are surface orientation dependent and are stored in what's called Tangent Space. But all the other lighting components such as view direction are supplied in world space. Because we can't use world space, why not convert every lighting component we need to compare the normal with, to this format called tangent space? Why not compare apples to apples?

Changing coordinate systems requires transformation. I'll just skip the hardcore math, but what I do want to explain here is that we need a matrix to transform world to tangent space. Just like we need a matrix to get world space from object space, we need a matrix to convert to tangent space. Remember this:

• We need the surface orientation, because that's where the texture normals depend on.
• We know everything about our surface (a triangle).
• Any lighting component we need in PS (lightdir,viewdir,surfacedir) needs to be multiplied by the resulting matrix.

/* We need 3 triangle corner positions, 3 triangle texture coordinates and a normal. Tangent and bitangent are the variables we're constructing */


// Determine surface orientation by calculating triangles edges

D3DXVECTOR3 edge1 = pos2 - pos1;

D3DXVECTOR3 edge2 = pos3 - pos1;

D3DXVec3Normalize(&edge1, &edge1);

D3DXVec3Normalize(&edge2, &edge2);

// Do the same in texture space

D3DXVECTOR2 texEdge1 = tex2 - tex1;

D3DXVECTOR2 texEdge2 = tex3 - tex1;

D3DXVec2Normalize(&texEdge1, &texEdge1);

D3DXVec2Normalize(&texEdge2, &texEdge2);

// A determinant returns the orientation of the surface

float det = (texEdge1.x * texEdge2.y) - (texEdge1.y * texEdge2.x);

// Account for imprecision

D3DXVECTOR3 bitangenttest;

if(fabsf(det) < 1e-6f) {

// Equal to zero (almost) means the surface lies flat on its back

tangent.x = 1.0f;

tangent.y = 0.0f;

tangent.z = 0.0f;

bitangenttest.x = 0.0f;

bitangenttest.y = 0.0f;

bitangenttest.z = 1.0f;

} else {

det = 1.0f / det;

tangent.x = (texEdge2.y * edge1.x - texEdge1.y * edge2.x) * det;

tangent.y = (texEdge2.y * edge1.y - texEdge1.y * edge2.y) * det;

tangent.z = (texEdge2.y * edge1.z - texEdge1.y * edge2.z) * det;

bitangenttest.x = (-texEdge2.x * edge1.x + texEdge1.x * edge2.x) * det;

bitangenttest.y = (-texEdge2.x * edge1.y + texEdge1.x * edge2.y) * det;

bitangenttest.z = (-texEdge2.x * edge1.z + texEdge1.x * edge2.z) * det;

D3DXVec3Normalize(&tangent, &tangent);

D3DXVec3Normalize(&bitangenttest, &bitangenttest);

}

// As the bitangent equals to the cross product between the normal and the tangent running along the surface, calculate it

D3DXVec3Cross(&bitangent, &normal, &tangent);

// Since we don't know if we must negate it, compare it with our computed one above

float crossinv = (D3DXVec3Dot(&bitangent, &bitangenttest) < 0.0f) ? -1.0f : 1.0f;

bitangent *= crossinv;

/* and add it to our model buffers */

We need to create a 3x3 matrix to be able to use it to convert object normals to surface-relative ones. This matrix should be built by adding the three components up in a matrix, and then transposing it in de Vertex Shader:

// tangentin, binormalin and normalin are 3D vectors supplied by the CPU

float3x3 tbnmatrix = transpose(float3x3(tangentin,binormalin,normalin));

// then multiply any vector we need in tangent space (the ones to be compared to

// the normal in the texture). For example, the light direction:

float3 lightdirtangent = mul(lightdir,tbnmatrix);

Then we're almost done. The only thing we need to do now is pass all the converted stuff to the Pixel Shader. Inside the same Pixel Shader retrieve the normal from the texture. Now you're supposed to end up with for example the light direction in tangent space. Then do your lighting calculations as you would always do, with the only exception being the source of the normal:

// we're inside a Pixel Shader now
// texture coordinates are equal to the ones used for the diffuse color map
float3 normal = tex2D(normalmapsampler,coordin);


// color is stored in the [0,1] range (0 - 255), but we want our normals to be
// in the range op [-1,1].
// solution: multiply them by 2 (yields [0,2]) and substract one (yields [-1,1]).
normal = 2.0f*normal-1.0f;


// now that we've got our normal to work with, obtain (for example) lightdir 
// for Phong shading
// lightdirtangentin is the same vector as lightdir in the VS around 
// 20 lines above
float3 lightdir = normalize(lightdirtangentin);


/* use the variables as you would always do with your favourite lighting model */