Make copy_parameters_t user-facing and beef it up #466
Assignees
Comments
eyalroz
added a commit
that referenced
this issue
Feb 17, 2023
* `copy_parameters_t<N>` is now out of `detail_` - and in it's own file * The `endpoint_t` enum is now also out of `detail_` (and in `copy_parameters.hpp`) * Our array-to-raw-mem copy function now determines the destination address' context * Add an array-to-raw-mem copy function which takes the destination context explicitly
eyalroz
added a commit
that referenced
this issue
Feb 18, 2023
* Added a memory::type_t enum value, `non_cuda`, for memory not allocated by CUDA * `memory::type_of()` will no longer fail for non-CUDA-allocated memory * `copy_parameters_t<N>` is not out of `detail_`. * The `endpoint_t` enum is now also out of `detail_` * Beefed up `copy_parameters_t<N>` with additional convenience methods *`copy_parameters_t<N>` now "recognizes" non-CUDA-allocated memory, treating it like CUDA host memory * Our array-to-raw-mem copy function now determines the destination address' context * Add an array-to-raw-mem copy function which takes the destination context explicitly
eyalroz
added a commit
that referenced
this issue
Feb 18, 2023
* Added a memory::type_t enum value, `non_cuda`, for memory not allocated by CUDA * `memory::type_of()` will no longer fail for non-CUDA-allocated memory * `copy_parameters_t<N>` is not out of `detail_`. * The `endpoint_t` enum is now also out of `detail_` * Beefed up `copy_parameters_t<N>` with additional convenience methods *`copy_parameters_t<N>` now "recognizes" non-CUDA-allocated memory, treating it like CUDA host memory * Our array-to-raw-mem copy function now determines the destination address' context * Add an array-to-raw-mem copy function which takes the destination context explicitly
eyalroz
added a commit
that referenced
this issue
Mar 9, 2023
* Added a memory::type_t enum value, `non_cuda`, for memory not allocated by CUDA * `memory::type_of()` will no longer fail for non-CUDA-allocated memory * `copy_parameters_t<N>` is not out of `detail_`. * The `endpoint_t` enum is now also out of `detail_` * Beefed up `copy_parameters_t<N>` with additional convenience methods *`copy_parameters_t<N>` now "recognizes" non-CUDA-allocated memory, treating it like CUDA host memory * Our array-to-raw-mem copy function now determines the destination address' context * Add an array-to-raw-mem copy function which takes the destination context explicitly
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Our
copy_parameters_ttemplate, inheriting from the CUDA driver's 2D and 3D copy parameters structs respectively, is currently only used for copying arrays. But - it is, in fact, the general-purpose copying primitive. Its use could, in theory, replace all other driver copying APIs; and in fact, with CUDA graphs, it does so.It therefore makes sense that we offer some kind of "general copy parameters" structure for the user to play with, then pass on to
cuda::memory::copy().We'll begin by exposing the structure to the user and adding some convenience methods; but later on we should probably also protect the user from passing the structure without proper initialization, and generally make this into more of a builder class than an inheritor of the raw structure.
The text was updated successfully, but these errors were encountered: