Skip to content

[update][feature][discuss] GPU-accelerated detection with ROCm #7

Description

@mrsrina

So. It is an issue, but also a discussion about how to implement it.

I know we can handle detection on the CPU side using collision chunks, but that is not our objective in VokeGPU. We really want to accelerate detection using the giant power of stream processors. Async accurate data transfer to mapped-memory body region. This is possible with some good work.

Basically, fixed-region memory should be used in favor of async data transfer. A giant mapped host memory region indexed with body references, this is next to a streaming screencasting. Syncing GPU and CPU is the major question itself, since we need to wait for callbacks from the device to the host to be sure that the detection is properly working, no lie information or noise.

Metadata

Metadata

Assignees

Labels

ROCmRelated to the AMD ROCm model implementation.documentationImprovements or additions to documentationenhancementNew feature or request

Type

Fields

No fields configured for Task.

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions