Testing The Open-Source "RADV" Radeon Vulkan Driver vs. AMDGPU-PRO

[lethalwp]

Btw. an article about the state of HEVC Main 10 decode would be interesting as vdpauinfo/vainfo show support for it with Polaris GPUs.

For what i've tested, hardware decoding of hevc in main10 in polaris works like a charm, in vaapi. For that i used ffmpeg-git + this homemade patch: http://ffmpeg.org/pipermail/ffmpeg-d...st/197623.html then use mpv with -hwdec=vaapi and opengl renderer.

But i've also noticed some problems when hardware decoding a simple mpeg2 from a dvd on a rx480. That i don't know where the problem is
(bad looking blocks, sometimes even flickering blocks)

[jrch2k8]

irrelevant, intel does not use llvmlol, that compiler will generate code for intel gpus

Nir is an IR and in theory SPIR-V should in theory be portable to the mesa compiler that intel uses. Since AMD Gallium driver don't use NIR i had doubts which parts they used. but whatever B already answered that NIR->LLVM.

My point was on SPIR-V to NIR IR, since both intel and RADV use it, it should help having more eyes optimizing this stage then easying the job of their respective compilers and layers after it.

[grigi]

Wow. This is impressive. Well done David & Bas!
Do you have plans to backport it to CI and SI?

[Kano]

RADV is a very interesting project, maybe even better than the Intel Vulkan implementation already. I hope there will be much more Vulkan games in the near future. Btw. an article about the state of HEVC Main 10 decode would be interesting as vdpauinfo/vainfo show support for it with Polaris GPUs. A working implementation would be very cool. Kaby Lake based laptops/NUCs should be available soon too, so this could be compared then. Kaby Lake desktop CPUs take a bit more time.

[pal666]

Well i think using NIR is actually good idea B, after all intel uses it too for their anv driver.

irrelevant, intel does not use llvm

reuse mesa shader compiler that intel uses??

lol, that compiler will generate code for intel gpus

[chimpy]

BNieuwenhuizen Well hopefully AMD and their people (Bridgman) will reach out to you and David and lend their resources. If Radv is this far along, AMD should might as well use it.

[pal666]

and nor week

[pal666]

btw, it was neither full nor from scratch

[pal666]

The funny thing is that the legal review is taking more than 6 months while writing a full driver from scratch took just a week.

it's not just legal review. somebody has to turn windows vuklan driver to opensourceable form. they have no resources for that, so do not expect they will magically find resources to maintain another driver

[pal666]

Holy shit i expected it to be way less than this, Holy Moly AMDGPU-PRO took a beating this week.

not really. radv won on cpu-bound tests because it does less work on cpu. on gpu-bound tests radv lost because less work on cpu means less optimal gpu setup