Benchmarks of 2019-06-14 to 2019-06-17

bench_notes.txt

system information:
    OS: Fedora 30
    uname -r: 5.1.7-300.fc30.x86_64
    CPU: Intel® Core™ i7-6700K CPU @ 4.00GHz × 8
    15.6 GiB of RAM

# Benchmark

workloads:
    GTA V
    Strange Brigade
    Dawn of War III
    Mad Max
    Rise of the Tomb Raider
    Thrones of Britannia
    Total War: Warhammer II
    The Talos Principle
    F1 2017
    Three Kingdoms
    Doom (2016) (with and without the fast path)
        running under Wayland again
        AMDGPU-PRO has artifacts with hands and zombies
gpu: Vega 64
run in 4K and use highest settings
software versions:
    LLVM rL363268
    RADV/ACO a092f01be8a (based on Mesa c58dff753c2)
    RADV/LLVM c58dff753c2
    AMDVLK 2019.Q2.5
    AMDGPU-PRO 19.10-785425
    Proton 4.2-7

set /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor to "performance"
set /sys/devices/system/cpu/intel_pstate/no_turbo to "1"
set /sys/class/drm/card0/device/power_dpm_force_performance_level to "profile_standard"

# pipeline-db

time pipeline-db runs comparing AMDVLK, RADV/LLVM, RADV/ACO and AMDGPU-PRO
    use a ACO compiled with IR validation disabled (now disabled by default in release builds)

compiletime.txt

  RADV/ACO (vkpipeline-db run): 410.09 seconds, 405.73 seconds (avg 407.91 seconds)
 RADV/LLVM (vkpipeline-db run): 745.23 seconds, 748.83 seconds (avg 747.03 seconds)
AMDGPU-PRO (vkpipeline-db run): 217.39 seconds, 214.94 seconds (avg 216.16 seconds)
    AMDVLK (vkpipeline-db run): 623.23 seconds, 620.74 seconds (avg 621.99 seconds)

   RADV/ACO w/o ACO (vkpipeline-db run): 355.71 seconds
  RADV/ACO w/o LLVM (vkpipeline-db run): 227.72 seconds

approximate breakdown of ACO run:
   52.20 seconds spent in ACO (13% of the run)
   180.19 seconds spent in LLVM (44% of the run)
   175.52 seconds spent in NIR and other areas (43% of the run)

estimated performance of ACO compared to LLVM:
   391.32 seconds spent compiling FS/CS in LLVM run
   so ACO is 7.5x faster than than LLVM (52.20 seconds)

information from RADV_DEBUG=compiletime
   LLVM spends 384.72 CPU seconds and ACO spends 50.28 CPU seconds
   ACO's CPU time is 7.7x lower than LLVM

VS/GS/TCS/TES guesstimates:
   since 180.19 seconds in an ACO run is spent in LLVM
   and ACO is 7.5x faster for FS/CS
   if ACO's performance relative to LLVM is the same for other stages
   implementing VS/GS/TCS/TES might remove about ~156 seconds from ACO runs
   and bring down the time to about ~252 seconds

game_benchmarks.txt

The kernel release during the Rise of the Tomb Raider runs was "5.1.8-300.fc30.x86_64".

Strange brigade
RADV/LLVM: 48.33 FPS
RADV/ACO: 51.07 FPS
AMDVLK: 50.33 FPS
AMDGPU-PRO: 53.20 FPS
RADV/ACO is 5.7% faster than RADV/LLVM

Total War Saga: Thrones of Britannia
RADV/LLVM: 14.50 FPS
RADV/ACO: 15.33 FPS
AMDVLK: 14.57 FPS
AMDGPU-PRO: 11.97 FPS
RADV/ACO is 5.7% faster than RADV/LLVM

Warhammer 40,000: Dawn of War III
RADV/LLVM: 35.57 FPS
RADV/ACO: 35.88 FPS
AMDVLK: 33.62 FPS
AMDGPU-PRO: 32.37 FPS
RADV/ACO is 0.9% faster than RADV/LLVM

F1 2017
RADV/LLVM: 54.00 FPS
RADV/ACO: 54.00 FPS
AMDVLK: 56.00 FPS
AMDGPU-PRO: 55.00 FPS
RADV/ACO is 0.0% faster than RADV/LLVM

Grand Theft Auto V ("scene-0" scene)
RADV/LLVM: 16.84 FPS
RADV/ACO: 16.54 FPS
AMDVLK: 15.71 FPS
AMDGPU-PRO: 16.09 FPS
RADV/ACO is -1.8% faster than RADV/LLVM

Grand Theft Auto V ("scene-1" scene)
RADV/LLVM: 16.27 FPS
RADV/ACO: 15.99 FPS
AMDVLK: 15.61 FPS
AMDGPU-PRO: 15.82 FPS
RADV/ACO is -1.7% faster than RADV/LLVM

Grand Theft Auto V ("scene-2" scene)
RADV/LLVM: 16.02 FPS
RADV/ACO: 15.71 FPS
AMDVLK: 15.53 FPS
AMDGPU-PRO: 15.84 FPS
RADV/ACO is -2.0% faster than RADV/LLVM

Grand Theft Auto V ("scene-3" scene)
RADV/LLVM: 18.00 FPS
RADV/ACO: 17.73 FPS
AMDVLK: 19.10 FPS
AMDGPU-PRO: 19.60 FPS
RADV/ACO is -1.5% faster than RADV/LLVM

Grand Theft Auto V ("scene-4" scene)
RADV/LLVM: 17.67 FPS
RADV/ACO: 17.35 FPS
AMDVLK: 17.71 FPS
AMDGPU-PRO: 18.00 FPS
RADV/ACO is -1.8% faster than RADV/LLVM

Mad Max ("camp-hollow" scene)
RADV/LLVM: 45.80 FPS
RADV/ACO: 46.78 FPS
AMDVLK: 45.56 FPS
AMDGPU-PRO: 44.82 FPS
RADV/ACO is 2.1% faster than RADV/LLVM

Mad Max ("stronghold" scene)
RADV/LLVM: 42.00 FPS
RADV/ACO: 42.47 FPS
AMDVLK: 45.76 FPS
AMDGPU-PRO: 45.39 FPS
RADV/ACO is 1.1% faster than RADV/LLVM

Mad Max ("hope-glory" scene)
RADV/LLVM: 45.49 FPS
RADV/ACO: 47.15 FPS
AMDVLK: 44.88 FPS
AMDGPU-PRO: 43.98 FPS
RADV/ACO is 3.7% faster than RADV/LLVM

Mad Max ("landmover" scene)
RADV/LLVM: 52.13 FPS
RADV/ACO: 53.57 FPS
AMDVLK: 51.42 FPS
AMDGPU-PRO: 50.10 FPS
RADV/ACO is 2.7% faster than RADV/LLVM

Rise of the Tomb Raider ("mountain" scene)
RADV/LLVM: 48.33 FPS
RADV/ACO: 49.03 FPS
AMDVLK: 47.67 FPS
AMDGPU-PRO: 48.43 FPS
RADV/ACO is 1.4% faster than RADV/LLVM

Rise of the Tomb Raider ("tomb" scene)
RADV/LLVM: 36.50 FPS
RADV/ACO: 38.53 FPS
AMDVLK: 36.27 FPS
AMDGPU-PRO: 38.07 FPS
RADV/ACO is 5.6% faster than RADV/LLVM

Rise of the Tomb Raider ("valley" scene)
RADV/LLVM: 36.30 FPS
RADV/ACO: 37.13 FPS
AMDVLK: 35.50 FPS
AMDGPU-PRO: 36.30 FPS
RADV/ACO is 2.3% faster than RADV/LLVM

The Talos Principle
RADV/LLVM: 70.10 FPS
RADV/ACO: 70.17 FPS
AMDVLK: 68.97 FPS
AMDGPU-PRO: 70.03 FPS
RADV/ACO is 0.1% faster than RADV/LLVM

Total War: THREE KINGDOMS
RADV/LLVM: 15.80 FPS
RADV/ACO: 16.00 FPS
AMDVLK: 14.17 FPS
AMDGPU-PRO: 14.90 FPS
RADV/ACO is 1.3% faster than RADV/LLVM

Total War: WARHAMMER II ("fallen" scene)
RADV/LLVM: 29.07 FPS
RADV/ACO: 29.93 FPS
AMDVLK: 28.53 FPS
AMDGPU-PRO: 27.00 FPS
RADV/ACO is 3.0% faster than RADV/LLVM

Total War: WARHAMMER II ("skaven" scene)
RADV/LLVM: 29.60 FPS
RADV/ACO: 29.63 FPS
AMDVLK: 31.13 FPS
AMDGPU-PRO: 29.37 FPS
RADV/ACO is 0.1% faster than RADV/LLVM

Doom 2016 (generic path)
RADV/ACO: 50.46 FPS
RADV/LLVM: 47.17 FPS
RADV/ACO is 7.0% faster than RADV/LLVM

Doom 2016 (AMD path)
RADV/ACO: 54.85 FPS
RADV/LLVM: 45.34 FPS
AMDVLK: 49.66 FPS
AMDGPU-PRO: 57.54 FPS
RADV/ACO is 21.0% faster than RADV/LLVM

Frankly, these benchmarks miss to some extend the most important contribution of ACO. 1 FPS or a few FPS compared to for RADV/LLVM or AMDVLK extra is nice but not so impactful. When you upgrade to a new graphics card with 2.5X the performance of your current one then you literally won't notice the difference in performance between the two software-stacks.

However, ACO also changed some games from unplayable to great, that is where it shines. A concrete example which I experienced: Borderlands2 native has a very bad performance like many games from that era. Kudos to the developers for supporting Linux back then but the performance is unplayable for me due to both the low FPS and the irregular framerendertimes, this on a 3700X and 5700 XT around the year 2020.
The Proton version of Borderlands2 also was broken back then, think about 10-20 FPS, something like that, maybe even worse.
Simply enabling ACO changed the performance from that to near-native, >400 FPS. That is how impactful it was in certain situations. I can only speculate why, maybe because the graphics are quite light (DX9, relatively low-complexity graphics from a technical point of view) but the compile-time got in the way and that was fixed by using ACO. I am pointing this all out because the Archwiki about amdgpu refers to this Github page. I think that this nuance matters.