Archives Discussions

[jean-claude]

Implicit tream operators

Hi,

Again a basic question:

Just to want to understand if there is any performance penalty (timing,..)
in taking advantage of kernel auto up-sampling/down-sampling...

for instance in this vector %2 downsampling :

kernel move_to(float src<>, out float dest<> {
  dest = src;
}

main :

float A_str<100>;
float B_str<50>;
move_to(A,B);         // B gets all values at even positions in A

So is upsampling/downsampling a pure Brook software feature, or is it de facto supported by the GPU hardware?

In any case what is the performance price to pay versus using aligned-sized vectors??

Thanks

Jean-Claude

[udeepta]

There are a lot of things going on inside the hood, so it is not as black and white as pure software or totally hardware. To keep things simple (under the hood), you may want to use a gather stream and do the resizing yourself.