The addition of AMD ray-tracing in Cycles has a long history. Its important to remember those that purchased RDNA2 near release with ā(Patch to be public soon)ā on T82557 have been let down.
Put yourself in their shoes, would you consider 18+ months or after the release of RDNA3 āsoonā in late 2020?
HIP on Linux has been enabled in the daily builds on March the 16th, it has been switched on by differential D14360. The HIP SDK for Windows so far hasnāt been publicly available for Windows, but it is available for Linux. It needs the hipcc compiler, but this can be installed from the ROCm package. Because of this, itās possible to compile Blender on Linux. This version will recognise the GPU, and try to run HIP on it. Unfortunately it looks like there are issues on Linux as well. AMD has just released ROCm 5.1.2, but this hasnāt solved the issues as far as I can see.
Because Vega support in Cycles HIP is currently too buggy at the moment to be officially enabled in the Windows or Linux builds of Cycles.
Also during the testing of the enabling of Cycles HIP on Linux, some issues were experienced related to Vega. So Vega support was disabled pre-maturely for Linux Cycles HIP to ensure Cycles HIP was enabled on Linux in a usable form for some users in a timely manner.
Just tested 5.1.3 on classroom with a RX6800, similar performance to 5.1.2.
Might get some insights on Vega support at the fortnightly cycles meeting, however May 11th never got posted so who knows if the upcoming May 25th will be posted. Pretty poor communication overall.
5.1.3 did nothing to fix the hipTexObjectCreate crash on RDNA1 GPUs, at least. Looking at Github, I donāt know if HIP was even updated - there have been no changes to the 5.1.x branch in a month.
When you compile the current code, you can enable Vega on Linux by changing just 2 lines. Commit rBc46e58817cd7 has been reverted later because Vega is crashing when using HIP, but you can redo these changes.
You need to install ROCm with use cases hip, and hiplibsdk, but this does contain everything thatās needed to support running HIP on Linux. Itās working, you get the HIP panel in the System Preferences, and it will list the Vega 64. When you then render a scene, it will start, and it will invoke hippc, but this will unfortunately throws an error a couple of times, and the kernel will fail.
However, I understand this may well be fixed by a future version of ROCm. I had high hopes for 5.1.2, but the same error still is thrown. I havenāt tried 5.1.3 yet, but I somehow doubt it will fix it right now. Maybe ROCm 5.2.0?
For my Linux setup the difference was about 1%, I do not believe it is worthy to annoy the devs with a bug at this stage. Feel free to open one for Windows however.
So Iāve applied those changes back and tried to build with Rocm 5.1.3 and Vega enabled, and it fails with these errors:
src/blender/intern/cycles/kernel/../kernel/device/gpu/image.h:34:18: note: candidate function not viable: call to __device__ function from __host__ function
ccl_device float cubic_g0(float a)
^
fatal error: too many errors emitted, stopping now [-ferror-limit=]
2 warnings and 20 errors generated when compiling for gfx1010.
It doesnāt fail at building for Vega, it fails in general for any arch.
UPD: I was building the latest alpha from git source, and in fact it fails to build without Vega enabled, so this may be some unrelated bug:
Iāve installed ROCm 5.1.3 today, updated Blender, and then applied the patch to add Vega to the list of HIP Devices. Iāve done this a number of times, and I havenāt had issues compiling the code so far, and had in fact no issues today. Iāve run Debian, but have mainly been using Ubuntu.
Iāve been compiling the code cloned from the Blender git according to the building instructions on the Blender Wiki for Ubuntu found here. So far the whole process of updating and compiling the code has gone remarkably smoothly.
All that needs to be done to add Vega is to add gfx900 to line 448 of CMakeLists.txt, and to change line 54 of file util.h to return (major >= 9). So far this change has compiled successfully every time.
Only discrete GPUs are supported at the moment, not integrated GPUs. I donāt have a timeline for when they will be supported, or if that will still happen at all.
Hello thanks for the reply. Shouldnāt the rx 6700s gpu be fine since itās a dedicated graphics card on the new asus g14. The radeon 680m is the integrated gpu of the g14 which I understand not working but the 6700s is one of the switchable dedicated amd gpu that can be used by the computer similar to how nvidia has optimus since it has two graphics cards.
edit: Found out rx 6700s and rx 6800s gpu are gfx1032 but some reviewers had issues with hip cycles involving a kernel error. Wonder if this is related to the igpu 680m in the laptop they tested. Anyways just wanted to add thank you the blender software is awesome with awesome developer and I appreciate the work you guys do