Project

General

Profile

Task #1456

remove the use of nbat->alloc/free pointers

Added by Szilárd Páll over 5 years ago. Updated over 4 years ago.

Status:
New
Priority:
Low
Category:
core library
Target version:
-
Difficulty:
uncategorized
Close

Description

Tto use page locked memory with CUDA, the nbat->alloc/free function pointers are used. With the required CUDA version bumped to 4.0, we can now (unconditionally) switch to page-locking memory using cudaHostRegister() instead of calling the CUDA API to allocate.

The solution should involve page-locking all host memory used in transfers to or from a GPU (when using GPUs).

The use of the write-combined flag could be considered again.


Related issues

Related to GROMACS - Task #1382: drop old CUDA support (3.2 and/or 4.0)Closed11/13/2013

History

#1 Updated by Szilárd Páll over 5 years ago

  • Related to Task #1382: drop old CUDA support (3.2 and/or 4.0) added

#2 Updated by Szilárd Páll over 4 years ago

  • Priority changed from Normal to Low

In the light of other accelerated code-paths based on OpenCL, we need to re-assess whether it's best to keep the custom allocators or switch to page-locking memory managed by mdrun.

Also available in: Atom PDF