D3D interop RELOADED isn't supposed to be better than OpenGL...?

Accelerated Computing CUDA CUDA Programming and Performance
1

I’ve read posts from lot of people that have asked similar stuff before. <img src=‘http://hqnveipbwb20/public/style_emoticons/<#EMO_DIR#>/crying.gif’ class=‘bbc_emoticon’ alt=‘:’(’ /> But I haven’t found a real good explanation about this.

( and yes… used the search button lot of times…)

I’ll really appreciate any useful explanation or link.

[codebox]void HiddenRender()

{

  // RENDER MODEL GEOMETRY

  g_renderToSurface1->BeginScene( targetTexture , &g_viewport1 );

… lights, matrices, geometry …

g_renderToSurface1->EndScene( D3DTEXF_LINEAR );

if ( g_applyCuda )

  {

       res = cudaD3D9MapResources(1, (IDirect3DResource9**)& targetTexture );

// MY OWN BIZARRE MACHINE-VISION STUFF

   // RunKernels();

res = cudaD3D9UnmapResources( 1 , (IDirect3DResource9**)& targetTexture );

  }

}[/codebox]

ahhh!!, i forget something:

g_pd3dDevice->CreateTexture( Width , Height , 1 , D3DUSAGE_RENDERTARGET , D3DFMT_X8R8G8B8 , D3DPOOL_DEFAULT , & targetTexture , NULL )

Now, what annoys me:

(on the same machine, Cuda 2.1, Vs2008 express edition, core duo + 8500GT…)

ONLY RENDERING -->> 8000 fps

ONLY MAPPING - UNMAPPING -->> 700 fps

I run the ‘SobelFilter’ example (opengl…) -->> 14000 fps

I run ‘simpleD3D9Texture’ example , JUST MAPPING-UNMAPPING, similar texture sizes than me —>> about 700 fps

Is D3D INTEROP not supposed to be faster than OpenGl at the moment … ? (read a post about that…)

Why all that time to map a texture that is already supposed to be at the device?

Furthermore, why do i need to map that texture every cycle? ( i alway work on the same one…, it should be enough with ‘locking’ it…)

Is there other way to access the rendered surface/texture without that penalty…?

( so I can success…, receive congratulations from my boss, get fame and women… External Image

Thank you guys for your time.

2

Direct3D interoperability requires the driver to convert from the special texture format to linear format by doing a copy behind the scenes, so it has many of the same issues as OpenGL interop.

3

thanks for the explanation. Master Green.

I imagined something like that. But the existence of ‘tex1D(…)’ ‘tex2D(…)’ functions made me think that Cuda was able to access directly those textures.

Anyway. I`m sure that the gurus at Nvidia will make some improvements with that in future versions.

As a developer of machine vision systems, I have great expectations and great plans about Cuda :ph34r: