Cory Bloor
Cory Bloor
I'm reopening this PR for two reasons: 1. The compiler feature to enable reusing gfx1030 code objects on gfx1031 has been delayed. 2. The `HSA_OVERRIDE_GFX_VERSION` workaround does not work on...
> [The approach in this pull request] just doesn't scale very well if we want to enable gfx1030, gfx1031, gfx1032, gfx1033, gfx1034, and gfx1035. Instead, we are working on a...
> I was digging deeper on the gtx1010 compatibiliy, has far i investigate it's way more complicated that expected, the rocm tensile libraries are missing for that gpu since years...
> Interesting going to take a look at rocm5.5 however i want to run llamacpp with my rx5700 and i think it only supports rocm5.6 and 5.7 I don't think...
@littlewu2508, I'm closing because https://github.com/llvm/llvm-project/pull/76955 has been accepted into LLVM. The best path forward for enabling gfx1031 and other RDNA 2 architectures will be to target the gfx10.3-generic ISA in...
> @cgmb im using ubuntu 22.04 and rocm6 do you think can work? going to try to compare with vulkan performance. No. Your driver will work, but https://github.com/ROCm/Tensile/issues/1757 will prevent...
Does ed25a0b570b6cabe06ae97b4f16edf943d96d540 help?
I've submitted a fix for this. I'm not an expert on the OpenCL development model, but I believe it is on track for ROCm 5.3.
@Maxzor, 4b579bf25a0e7e0262f2836d03720a974bd2e866 is now on the develop branch.
> This has been sitting around for a while. Should it be resurrected? Maybe, but I don't have the bandwidth to do that quite yet. The official Linux and Windows...