jmarsh wrote on 2020-02-19, 21:42:

Do you read the output of configure to make sure everything is being detected correctly?

I haven't been building the 0.74-3 exe, just testing against the official download.

Out of interest, I performed a 32-bit MinGW build of SVN though as I hadn't tested that in a while. I didn't see anything untoward in the configure output but performance of this executable is better than I used to get: 4550000 parrots, 1687.1 XT. Not as fast as a VS build, but significantly faster than stock 0.74-3 in Speedtst (3033333 parrots, 1124.7 XT). VS build is still 11% faster though. Maybe something has changed because I seem to recall getting performance similar to stock with my SVN builds in the past.

To be fair, the performance delta between 32-bit builds in Quake at 320x200 doesn't seem as dramatic as that shown in the Speedtst scores. 0.74-3 gets ~212fps, 32-bit MinGW SVN build gets ~220fps, 32-bit VS build gets ~245. The VS build is about 15% faster than 0.74-3 and is 11% faster than the MinGW SVN build (same as Speedtst result). No reason to use any of these over a 64-bit build though, as that's significantly faster than all of them.

EDIT: Changed MinGW 32-bit result as the score I previously got was a low outlier.

I ran this on my i9-9900KS @ 5.2 GHz and the framerate is actually LOWER (about 74 FPS) than my i7-7700K 5 GHz runs. Weird. I got 83 FPS in 800x600 in timedemo 1 in fullscreen mode, though.

MrEWhite wrote on 2020-02-22, 02:21:

I ran this on my i9-9900KS @ 5.2 GHz and the framerate is actually LOWER (about 74 FPS) than my i7-7700K 5 GHz runs. Weird. I got 83 FPS in 800x600 in timedemo 1 in fullscreen mode, though.

2/76 = ~3% is in borders of measurement accuracy, according to what I saw. There was no significant per Hz improvement between those CPU series and the clocks 5.0 and 5.2 are very close. In totally identical environment and making several tests runs you are expected to notice 2-4% average improvement (for fine difference of CPUs is better to use 320x200 mode).
For human perception up 10% fps difference is doubtful to be noticable.

Silanda wrote on 2020-02-19, 22:34:

To be fair, the performance delta between 32-bit builds in Quake at 320x200 doesn't seem as dramatic as that shown in the Speedtst scores. 0.74-3 gets ~212fps, 32-bit MinGW SVN build gets ~220fps, 32-bit VS build gets ~245. The VS build is about 15% faster than 0.74-3 and is 11% faster than the MinGW SVN build (same as Speedtst result). No reason to use any of these over a 64-bit build though, as that's significantly faster than all of them.

According to jmarsh, official 64-bit build has no dynamic core. So this variant has noticed ones or is tested for compatibility issues still.
When 64-bit dynamic mode will be inputed in the official release, it may become useful to add 64-bit version of DOSBox in the benchmark theme.

In the 1st theme's message I've added the testing procedure with bat files.
It would be interesting to see a direct comparision between different DOSBox builds of today.
official 0.74 (32-bit), official 0.74-3 (32-bit), SVN MinGW (32-bit), SVN MinGW (64-bit), SVN VS (32-bit), SVN VS (64-bit)
6 numbers of results. The used SVN version should be said.
Emulator's files (.exe and .dll) should be changed between the testing runs.
The comparision may be limited to Quake 320x200 dynamic core testing (with single demo1 run). It needs to place '#' before the line "quake.exe +vid_mode 13 +timedemo demo1" in quake_test.conf. The testing is run by tests2.bat, press "Space", fps is written, build's files are changed and the testing is repeated.
In case you'll create those 4 SVN builds, it would be useful to place them on a filehosting like mediafire, so other CPUs could be tested with them also. For example, to check what happens on modern Intel CPUs.

Official (0.74-3 and older) 64-bit builds have a dynamic core, but it's not the 64-bit dynamic_x86 core.

jmarsh wrote on 2020-03-01, 02:05:

Official (0.74-3 and older) 64-bit builds have a dynamic core, but it's not the 64-bit dynamic_x86 core.

then those are not 64-bit completely builds 😀

Ryzen 9 3900x, no PBO., opengl output, cycle max

speedtest:
core (simple/auto/dynamic/normal/full)

dosbox daum : 263,6/306,7/1297,7/306,7/173,9
dosbox 0,74-3 : 312,4/248,1/803,4/248,1/248,1
dosbox ECE: 231,1/312,4/1054,4/306,7/312,4

1) Why is there huge differences (31%) between dosbox 0.74-3 performance and dosbox ECE performance with dynamic core ? Dosbox daum is even the fastest.

2) I I launched speedtest and relaunched a second pass of speedtest wuithout close dosbox session, I have with core dynamic 1687,1XT with dosbox ece (1054,4 in the first pass), dosbox daum(1297.7 in the first pass). With dosbox 0.74-3 I have 1124.7 instead of 803.4 with the first pass.
Why these differences ?

edit:
Doom dynamic (realtics/FPS) / normal (realtics/FPS)
dosbox daum : (269/278fps)/ (688/109FPS)
dosbox 0,74-3 : (300/249fps)/ (1109/67FPS)
dosbox ECE: (281/266fps)/ (655/144FPS)

Quake 320x200 (demo1,2,3) /800x600 (demo1,2,3)
dosbox daum : (224/228/211)/ (67/74/66)
dosbox 0,74-3 : (182/188/174)/ (60/67/59)
dosbox ECE: (200/202/189)/ (65/72/64)

willow wrote on 2020-05-23, 08:28:

With dosbox 0.74-3 I have 1124.7 instead of 803.4 with the first pass.
Why these differences ?

It's better to try a real game as Doom or Quake and in case you'll get significant (>10%) instable benchmark results there, - to ask about strange emulation behavior in a specific theme or development section. The reasons mb on your hardware, OS and background apps, not in the emulator only.

The first pass has to translate the code then run it, subsequent runs don't need to translate if the memory hasn't been changed.

jmarsh wrote on 2020-06-02, 06:32:

The first pass has to translate the code then run it, subsequent runs don't need to translate if the memory hasn't been changed.

Speedtest by 2 runs (2 identical strings in a batch file) by the recommended procedure gives results differing <5% on my machine/OS and 0.74. This % is similar to random fps differences (to both sides) when running games benchmarks.
He got significant 40% difference. If same happens also with games benchmarks, then may happen with common games. In this case may be needed to run a game 2 times to get the max performance, what would be not appropriate and be fixed.

Tertz wrote on 2020-06-03, 21:01:

In this case may be needed to run a game 2 times to get the max performance, what would be not appropriate and be fixed.

You can't magically cache stuff without loading it first.

jmarsh wrote on 2020-06-03, 23:22:

You can't magically cache stuff without loading it first.

There can be different reasons for the jump of the performance.

Must have been a Windows update or my RAM upgrade, because I beat my previous scores at a LOWER clock speed.

i9 9900KS @ 5.1 GHz
32GB DDR4 @ 3600 MHz
Windows 10 Pro 64-bit

Using DOSBox 0.74-3:

Quake:
78.4 88.0 77.1 / 242.9 247.2 230.3

MrEWhite
The important is the comparision between CPUs, so this needs to use the same test - 0.74 (32 bit)

Pentium N5000 - 2585mhz win10-64
I am use Preconfigured benchmark tests pack from first page oh this thread.
Quake results - 17, 20, 17 / 61, 63, 58

i5-12400F - 4400mhz win10-64
I am use Preconfigured benchmark tests pack from first page oh this thread.
Quake results - 89, 103, 90 / 276, 284, 261

Ryzen 5600 win10-64
I am use Preconfigured benchmark tests pack from first page oh this thread.
Quake results - 73, 81, 71