here are three distinct compiler scopes in every Cuda project.
"Regular" compiler scope covers cpp files (compiled by the native C++
compiler). In this scope all the regular C++ tools are available but the Cuda
API is not accessible.
"Host" compiler scope resides in cu files (compiled by the Nvidia compiler).
This scope covers host-side Cuda programming. It includes memory allocation
and deallocation of device-side memory and kernel launches. Most of C++
standard library and boost library are not accessible in this compiler scope.
"Device" compiler scope resides in cu files. This scope covers device-side
multithreaded code. The most narrow set of tools is available in this scope.
Flow control operations (if,while,for) should be avoided. Function recursion
should be avoided. There is no exception throwing or handling.
Host and Device scopes operate over distinct memory domains. The data
communication between these scopes is possible only via stack variables and
Cuda API calls.
Regular and Host compiler scopes operate over the same memory domain. To
control Cuda from Regular compiler scope one needs to isolate Cuda API inside
a layer of user-made functions placed in the Host scope and then call these
functions normally from the Regular scope.
In this section we illustrate such separation of responsibilities by
implementing exception-safe and leak-safe control of device memory from
Regular scope. The source and make files are part of the PiecewisePoly project
(explained in the following sections). All the files are accessible from the
Download section.
|
