OpenCL: large graph support through RAM->VRAM streaming

[EgorOrachyov]

Requirements:

• Async streaming of chunks of graph from RAM into VRAM
• Dynamic chunk size adaptation
• Partial operations evaluation support

[gsvgit]

What do you think about SVM ? Can it be helpful in this task?

[EgorOrachyov]

@gsvgit By def, shared virtual memory is a virtual address space for multiple processors, devices which can access the same address in seamles fashion.

The question is, how it is implemented for a gpu.

When and how gpu loads pages? What happens if there not enough GPU space? What is the maximum allocation size for a svm allocation?

Streaming solves the problem of large vram usage and overcomes limit of single allocation, which also is a limit of maximum range, which can be addressed inside a kernel.

Svm seems like a gimmick. But how it is implemented across vendors and which overhead has?

All in all, seems svm wont solve problem of large data.