NPP and CUDA Kernel interop

Accelerated Computing CUDA CUDA Programming and Performance
1

Hey all,

I would like to write a CUDA kernel that operates on an NPP primitive, so lets say I’ve loaded an image from file like from the examples:

[codebox] // declare a host image object for an 8-bit grayscale image

npp::ImageCPU_8u_C1 oHostSrc;

        // load gray-scale image from disk

npp::loadImage(sFilename, oHostSrc);

        // declara a device image and copy construct from the host image,

        // i.e. upload host to device

npp::ImageNPP_8u_C1 oDeviceSrc(oHostSrc);[/codebox]

Now the image is uploaded to the GPU. I would now like to have a CUDA kernel (in .cu file), that operates on the image, preferably using texture lookups (tex2D). I can’t see how I can get a relevant handle on the uploaded image from the NPP interfaces & the examples don’t appear to show anything like that. Maybe I’m missing a trick…

Thanks in advance,

James

2

Hi James,

what you’d do is you simply bind the image data (the image class has an accessor method to the GPU global memory data pointer) to a texture object:

[codebox]

texture hTexture;

size_t nTexOffset;

cudaBindTexture(&nTexOffset, hTexture, pSrc, nSrcStep * oSizeROI.height)

[/codebox]

In this example I’m binding to a 1D texture, 2D texture works similarly. The CUDA Programming Guide has information on the details.

You can then use that hTexture handle inside your kernel code for sampling from that texture.

–Frank

3

Brilliant, thanks again Frank!