Re: Notes about ROCm on RDNA2
I test HSA_OVERRIDE_GFX_ VERION=10.3.0 on my RX5700XT and ROCm-5.2,
both tensorflow-rocm and pytorch-1.12.0 can run mnist properly.
So, do we have to support navi10 separately? Looks like we can use
the navi21 codes.
Interesting. The RX 5700 XT is gfx1010. I suppose the gfx1010 and
gfx1030 ISAs share a common subset of instructions. I suspect that you
will eventually find something that doesn't work.
I haven't gone through the documentation [1] in detail, but gfx1010 is
an RDNA1 ISA and gfx1030 is an RDNA2 ISA. If the code object uses any
instructions that were added in RDNA2, you're going to have a problem
(e.g., dot product operations such as V_DOT2_F32_F16).
In short, no. You cannot safely use HSA_OVERRIDE_GFX_VERSION=10.3.0 with
the RX 5700 XT. However, it should be possible to compile all of
pytorch's dependencies from source for gfx1010. Though, you might need
to make a few minor patches to fix the occasional #ifdef __gfx1030__ check.
Sincerely,
Cory Bloor
[1]: https://gpuopen.com/amd-isa-documentation/
Reply to: