Cyberpunk 2077 does not leverage simultaneous multi-threading (SMT) on AMD Ryzen processors, according to multiple technical reviews of the game that tested it with various processors. The game does leverage the analogous HyperThreading feature on rival Intel Core processors.
…
According to them, Cyberpunk 2077 reuses AMD GPUOpen pseudo-code to optimize its scheduler for the processor. It was originally designed to let an application use more threads when an AMD "Bulldozer" processor is used; but has the opposite effect when a non-Bulldozer AMD processor is detected. The game looks for "AuthenticAMD" processor brand, and "family = 0x15" (AMD K15 or Bulldozer/derivative), and only then engages "logical processors" (as identified by Windows OS scheduler as part of its Bulldozer-optimization). When any other, including a newer AMD processor is detected, the code makes the game scheduler only send traffic to the physical cores, and not to their logical processors.
This problem is really just about detecting the correct thread count,
not about leaveraging SMT. Even if you spin up 8 threads on a 8 core/16 thread CPU, you don't control whether they run on separate cores or not, scheduling is up to the OS kernel, not the game.
I don't think this qualifies as optimization at all, it's just hardware detection. Software isn't "optimized" for SMT.
There were some comments of Intel's Compiler being involved in these shenanigans, again.
Can't help but think they're onto something.
Of course, those conspiracy theories never dies.
Rest assured those theories are nonsense, and this problem have
nothing to do with compilers.
How in the earth this AAA game has still not been updated for the new ryzen cpus?
This is certainly not excusable for a studio of this size and a game with this kind of budget.
But it's still very understandable how it happened; this is not a bug that is glaringly obvious to a normal game tester, and the core development team may only have a very limited set of hardware they test continously throughout development.
the code says something like:
if AuthenticAMD than
then use this
if (Bulldozer than Threadcount=logical)
if not (than threadcount=corecount)
yeah it´s from 2017 when Win10 didn´t have its threading optimizations, so eventually it´s very dull to make anything like this, every game should get all logical threads, in some very old code or rare cornercases it can be manually fixed with affinity in taskmanager or SMT turned off, but in general that part of the 2017 code is idiotic now.
Should only take a couple of minutes/seconds to fix.
This is about reading the
CPUID instruction to detect CPU features, it has nothing to do with Windows.
The same instruction is used to detect CPU features on all platforms, inlcuding the cpuinfo tool on Linux.
Unfortunately, some extensions to CPUID is not standardized across CPU makers (and even microarchitectures), so it may require some extra interpretation to detect the correct core and thread count on various AMD, Intel and VIA CPUs.
But nevertheless, this "problem" is well known and only takes a few lines of code to detect correctly (at least for known CPUs so far, I can't guarantee for future CPUs).
This still makes little sense to me, most threading frameworks used everywhere typically only expose logical threads with no real way of targeting cores and the reason for that is very simple, trying to take SMT into the way equation is a terrible idea with mostly negative results. The OS will typically prioritize the distribution of threads across cores first anyway…
You are mostly right.
The bug here is only about detecting the correct amount of available threads, not whether a specific thread is using SMT or not, that's normally not controlled by the application. So the article claiming SMT is not leveraged is technically incorrect.
AMDs technical dev support is rather like a good looking single guy at a pub perving at all the hotties thinking damn I could smash that 9 times on a sunday then going home to jerk off in a sock or worse taking the drunk fattatoe in the corner home and shashin that!
Didn't AMD boast about having hundreds of engineers dedicated to this?
At least this bug comes down to reading a single instruction, not something that requires any major effort from anyone.
-----
One additional thought;
If this is hard, just wait until Alder Lake with different cores arrives. There may be a lot of software relying on the CPUID instruction to fetch capabilities, but this instruction returns the features of the core, not the whole CPU, which can get messy if different cores have different ISA features.
