Partial Memory Uploads?

Question

Let's say I create an array.

concurrency::array<int> ar(1024);

I then fill it with data.

int count = 0;
for(int n = 0; n < /* ... */; ++n)
{
    ar_av[++count] = /*...*/;    
}

Now let's say that count < ar.extent.size(), thus I am not using the entire array ar.

If I do the following:

concurrency::parallel_for_each(concurrency::extent<1>(count), [&](const concurrency::index<1>& idx)
{
     /* ... */
}

Is there a way to hint that not the entire array "ar" needs to be uploaded to the accelerator?

Answer 1

VS 2012's C++ AMP is implemented on top of D3D 11.  In D3D 11, to execute a command (e.g. a dispatch) on GPU, all resouces needed by the command (e.g. buffers) will be swapped into the GPU memory if not there already. It's at per resource granularity (driver may or may not have some optimization for this, which could not be depended on). It means even the kernel to be dispatched only touches some portion of the resouce (even a single element), the whole resource will be swapped in.

Thanks,

Weirong

Answer 2

I understand.

How about if I first write my data to a staging array, is it possible to copy a sub range of the data to a smallar device array, i.e. copy between arrays with different extent?

Answer 3

Copy a section of an array (or the whole array) between a section of another array (or the whole array), could be achieved by first creating sections -- which yields array_view.  Then you can copy between array_view/array and array_view/array. For "section", please read:

http://blogs.msdn.com/b/nativeconcurrency/archive/2012/02/10/handling-subsections-of-data-in-c-amp.aspx

http://blogs.msdn.com/b/nativeconcurrency/archive/2012/02/14/section-example-in-c-amp.aspx

Thanks,

Weirong