Can I use a kernel to generate data in global memory and then lanch second kernel using the result of first kernel in texture memory?
if answer is yes,how can I do it?
Are there chance to avoid data transference or allocate data?
If I need transfer to host to use texture memory, is just use global without reallocation faster?
Sorry for my english.
You can copy from linear memory to a texture without copying back to the host.
As far as i know, you do need to copy it (from device to device), you cant just say “this is now a texture”.
Isn’t that just for CUDA Arrays, while linear memory mapped as a texture can be copy-free?