Project

General

Profile

Feature #3020

modernize DeviceBuffer and GPU memory management

Added by Artem Zhmurov 4 months ago. Updated 2 months ago.

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

Description

Currently, the management of the allocated and actual sizes of the buffers on the GPU is handled by the consumers of the allocate/reallocate/freeDeviceBuffer(...). This should be change in favor of the proper DeviceBuffer class that will have the integer/size_t fields to manage the size of the buffer as well as the pointer to the buffer.

Along these changes, we should clarify and modernize the "buffered" memory allocation. Options:
  • make allocation size match the host-side vectors' reserved size; pro: consistency; con: will significantly limit max simulation size (due to running out of GPU memory)
  • use custom extra buffer reservation (possibly with additional customization features to set this to zero for memory that does not need it).

History

#1 Updated by Szilárd Páll 3 months ago

  • Project changed from Support Platforms to GROMACS
  • Difficulty uncategorized added

#2 Updated by Szilárd Páll 2 months ago

  • Subject changed from Make DeviceBuffer into a proper class to modernize DeviceBuffer and GPU memory management
  • Description updated (diff)

Also available in: Atom PDF