Archives Discussions

[tugrul_512bit]

I'm developing an open-source API that uses multiple OpenCL devices to compute kernels, adding features like pipelining and load balancing, but now I need to add a device-to-device pipeline such that a GPU will send data to another GPU without touching  system RAM, only through inter-pci-e connection. What is the OpenCL API command for that? I guess this command

clEnqueueMigrateMemObjects

should have that but its documentation in Khronos site doesn't tell anything about it.

Details about my project:

wiki: Home · tugrul512bit/Cekirdekler Wiki · GitHub

tutorial: Easy OpenCL Multiple Device Load Balancing and Pipelining For C#: Cekirdekler API - CodeProject

thank you for your time.

[dipak · Accepted solution]

It's depends on the implementation. AFAIK, on AMD platform, MigrateMemObjects currently still uses system memory.

cl_amd_bus_addressable_memory​ extension is there which supports direct transfer,  however I've doubt whether it would be applicable in this scenario or not. Also this extension available for FirePro boards only.

Regards,

[tugrul_512bit]

Then  a HD7870 has to visit RAM before talking to R7-240.