clamp quesiton

Accelerated Computing CUDA CUDA Programming and Performance
1

hey there

I’m using this way to read data through texture:

[codebox]texture<float,1> tex_x_float_A; // as global variable

cudaChannelFormatDesc channelDesc =

    cudaCreateChannelDesc(32, 0, 0, 0, cudaChannelFormatKindFloat);

cudaError_t err = cudaBindTexture(0,tex_x_float_A,A,channelDesc);

tex1Dfetch(tex_x_float_A, i);[/codebox]

In this way, does clamp work if variable i is too large and falls out of boundary?

Thanks!

2

Clamping should be the default behaviour in that case, yes.

3

Clamping should be the default behaviour in that case, yes.

4

Just out of curiosity, how does CUDA know the size of the array A so that it knows where the boundary is?

5

Just out of curiosity, how does CUDA know the size of the array A so that it knows where the boundary is?

6

If you’re using tex1Dfetch you don’t get clamping or filtering (see programming guide B.8.1). tex1Dfetch isn’t like a real texture lookup, it’s just a convenient way to read memory taking advantage of the texture cache and without coalescing restrictions.

BTW, you do have to specify the memory size when calling cudaBindTexture.

7

If you’re using tex1Dfetch you don’t get clamping or filtering (see programming guide B.8.1). tex1Dfetch isn’t like a real texture lookup, it’s just a convenient way to read memory taking advantage of the texture cache and without coalescing restrictions.

BTW, you do have to specify the memory size when calling cudaBindTexture.