I built it using this (setting some variables before):

  meson --cross-file build-wine64.txt --prefix ${dest_dir} "$dxvk_build"
  cd "$dxvk_build"
  meson configure -Dbuildtype=release
  ninja

Does release enable compiler optimizations?

@shmerl

Does release enable compiler optimizations?

buildtype=release enable -O3 optimization level. The flags enabled at this level can vary between GCC versions, see e.g. enabled by default at -O3 ("General Improvements" section):

• https://gcc.gnu.org/gcc-7/changes.html
• https://gcc.gnu.org/gcc-8/changes.html

You can:

• check compiler versions used for build (Meson usually prints both, during MinGW build)
• modify your build-wine64.txt to use -O2, and try e.g. -Dbuildtype=plain
  or (for newer Meson versions) -Dbuildtype=custom -Doptimization=2

Such steps will help to find out is regression caused by -O3, by the specific flag, or by different GCC version.

The other question is: Did you build the DLL version yourself or did you use the pre-compiled one? Because you didn't mention how you built the other version...

Otherwise, comparing both versions is meaningless: The compiler is probably different, and versions may differ.

I built both myself. I'll run some tests using suggestions to lower optimization level for winelib case.

Compilers used (according to Meson):

• Winelib build:

Native C compiler: cc (gcc 8.2.0 "cc (Debian 8.2.0-9) 8.2.0")
Cross C compiler: winegcc (gcc 8.2.0)
Native C++ compiler: c++ (gcc 8.2.0 "c++ (Debian 8.2.0-9) 8.2.0")
Cross C++ compiler: wineg++ (gcc 8.2.0)

• dll build:

Native C compiler: cc (gcc 8.2.0 "cc (Debian 8.2.0-9) 8.2.0")
Cross C compiler: x86_64-w64-mingw32-gcc (gcc 7.3.0)
Native C++ compiler: c++ (gcc 8.2.0 "c++ (Debian 8.2.0-9) 8.2.0")
Cross C++ compiler: x86_64-w64-mingw32-g++ (gcc 7.3.0)

So there is that difference already.

Strange, I can't reproduce the issue now (no performance regression with either -O2 or default (-O3) when using winelib. May be it's something with the kernel since I was switching between 4.19.x and 4.20.0. Let me compare that.

Nope, the kernel doesn't make a difference. Not sure what changed really. This can probably be closed, unless it comes up again.

@shmerl Maybe you are experiencing the same pitfall I did and during tests something was eating your VRAM, e.g. an open browser with multiple tabs which wasn't open during on of your test cases (or memory usage was different)? The effect was about half the performance here...

Good point, I didn't monitor my VRAM usage. I'm actually not sure how to do that for amdgpu. Are there some good methods?

Ah, this produces something, but probably means pages:

sudo cat /sys/kernel/debug/dri/0/amdgpu_vram_mm | pcregrep 'total|usage'
total: c, used 152311 free 1940745
man size:2093056 pages, ram usage:594MB, vis usage:42MB

It makes sense, since I have 8 GB VRAM. So 8,573,157,376 B / 2,093,056 = 4,096 B. I.e. one page is 4 KB.

Created a crude AMD GPU load / VRAM monitor if it's useful for anyone.

Created a crude AMD GPU load / VRAM monitor

https://github.com/clbr/radeontop ?

I tried radeontop, it produces no data for my Vega.

Is it crashing?

Hm, just for the record, I'm using 4.0~rc1, but never experienced such crash with earlier versions as well. The build is from WineHQ.

@doitsujin Does "unreadable" mean "very tiny fonts"? Both keywords ("unreadable" and "infinite") ring a bell here that I was able to fix.

@shmerl I started the game about at the same place as you with a winelib and the standard dll and I don't have perfomance differences:

dll release 0.94 : 52.3 FPS
winelib 0.94 (-march=native -O2) : 52.5 FPS

my setup is very similar to yours, if I remember well (just smaller):
AMD Ryzen 2600
AMD RX 470 (POLARIS 10)

@edmondo: I don't see it now either, so I can't explain what happened before when it was halved.

@shmerl Did you set CFLAGS or something similar? Or update meson lately? While debugging a build issue here, I noticed that the winelib build no longer respected my CFLAGS set through a makefile (or maybe it never did and I didn't notice). Upon inspection, I found that there was a fix to meson to ignore CFLAGS for cross-builds. Since winelib is implemented as a meson cross-build, that could explain something...

Meson was updated not so long ago in Debian testing to 0.49: https://tracker.debian.org/pkg/meson

But I encountered this before that update happened, so I was still using 0.48 then.