AMD Releases HSA Linux kernel driver v0.8 at HSAFoundation/HSA-Drivers-Linux-AMD · GitHub.
It is really nice to know HSA apps may have the ability to allocate the GPU local memory. However, it seems the provisional HSA Runtime specification did not specifically touch on how this is exposed in the API, though I guess it is exposed as a memory region (with like WriteCombine pages?) and making use of the hsa_memory_* API. May you please provide some insight on this? Thanks.
What is AMD's plan for releasing OpenCL stack that utilizes HSA features on Kaveri? According to HC-4017, HSA Compilers Technology, by Debyendu Das (page 17), OpenCL 1.2 with AMD extensions was scheduled for Q2 this year. It looks like all the components mentioned in the presentation have been released except for OpenCL run-time that uses HSA. Can you tell us when you are target to release new OpenCL run-time?
Sorry for late answer, just noticed this.
The SDMA hardware block (and SDMA queues) are used for transferring memory around without tying up compute resources. Transfers can also happen in parallel with execution of compute kernels.