OpenCL Image Support For Gallium3D's Clover

[oleid]

We are aiming to have a developer preview of the modified OpenCL stack (still closed source runtime but using a lot more open source code) around mid-December.

So does that mean, that in December, we can use an all open graphic stack (meaning radeonsi ; so no need to install the amdgpu-pro driver) and get openCL by simply installing one last binary blob? That would be an awesome change

[bridgman]

Sort-of... the preview will be running over the ROCM stack which includes modified kernel drivers (newer-than-upstream KFD code plus corresponding changes in amdgpu, ttm etc...). The ROCM kernel has not gone through much graphics testing other than developers booting to desktop so we aren't saying at this point you will have a do-everything system, but next step will be getting the newest amdkfd code upstream-ready so we can include it in all-open and -pro trees.

HW support in the initial preview will probably be limited to one or two high-end dGPUs.

[Veerappan]

Sort-of... the preview will be running over the ROCM stack which includes modified kernel drivers (newer-than-upstream KFD code plus corresponding changes in amdgpu, ttm etc...). The ROCM kernel has not gone through much graphics testing other than developers booting to desktop so we aren't saying at this point you will have a do-everything system, but next step will be getting the newest amdkfd code upstream-ready so we can include it in all-open and -pro trees.

HW support in the initial preview will probably be limited to one or two high-end dGPUs.

Ahh, that's what I was figuring... I'm still running a 7850 (and x6 1055t) at home, and I have kept looking at clover/libclc in my spare time because my GPU hasn't been able to run any of the kfd-based kernel driver stuff... That'll probably change when Vega and Zen both come out.

[bridgman]

Yeah, it's tough getting the HSA/ROC stack working on SI since it doesn't have the MEC block (so no AQL support and limited user queue support, also 2 compute queues on ME instead of 32+ HW scheduling on MEC).

We did hack up a partial implementation using Tahiti (79xx, which had ATC and could generate ATS/PRI requests) and Trinity (which had IOMMUv2 but not hooked up to the internal GPU) during the early days of HSA work until Kaveri silicon became available, but it required non-standard API calls because the MEC HW functionality wasn't there.

I think hacking the runtime to use kernel calls + native compiler on SI would be easier than trying to implement HSA/ROC-ish user queues.