On 24/07/17 11:47 AM, Alex Deucher wrote: > On Mon, Jul 24, 2017 at 11:25 AM, Tom St Denis <tom.stdenis at amd.com> wrote: >> This patch series reflects a slew of changes that fell out of >> team meetings with mesa/kmd members over the last week. >> >> We remove the "risky" option and the --print command since it's more >> annoying than helpful. >> >> We add the ability to specify the VMID in hex with --vram which >> is handy with the last patch since you can specify which VM hub to use >> in hex easier than in decimal. >> >> The third patch includes various small cleanups found by static analysis >> and code review. >> >> The last patch includes two new features which were written at the same >> time and were a bit tricky to separate (more effort than worth). First, >> we add the ability to halt waves when reading them which should help >> reading consistent data. Next we add the ability to decode multimedia >> IBs from the UVD/VCN rings. Next we fix using the 'mmMC_VM_FB_LOCATION' >> register which was traditionally zero on VI (and below) but now has >> non-zero values which broke VM decoding. Tested on Carrizo/Polaris/Raven. > > Note that the fb_location is potentially non-0 on all asics going > forward. We just leave it as programmed via the vbios now. umr treated it as zero previously and I was able to read IBs/etc on my various VI hardware. In testing the patches this morning I discovered it didn't work and my offsets were off by a huge margin which aligned with what I was reading from the register :-). As of this morning I can read gfx/uvd IBs on VI APU/dGPU and my only AI+ hardware which is an APU. I assume vega10 will be the same though hopefully. If someone with a vega10 could run umr -O follow_ib,verbose -R vcn_dec[0:79] (with /dev/fmem or STRICT_MEM disabled...) that could easily confirm if VMID 0 decodings are working. Thanks, Tom
