Quote from: Porcus on 2023-04-12 11:50:36

To the "showcases" division, a file downloaded from <stupidlyhi-rez site

According to @bennetng , it is not only the resolution itself that is dumb, it is the combination of resolution and "crappy quality junior MIDI sequencing stuff", so take it with a grain of salt.

@bennetng

What?
you are a malicious person.

By the way, @Case - you did provide some compile that in some situations performed well. Care to post an exe of 1.4.3 too?

Very slow figures recorded for the 23 June build posted by https://hydrogenaud.io/index.php/topic,124356.msg1029259.html#msg1029259

I don't know what has been done to the code but MSVC compiles are super slow nowadays.

<Project>
  <PropertyGroup>
    <LLVMInstallDir>C:\Program Files\LLVM\</LLVMInstallDir>
    <LLVMToolsVersion>16</LLVMToolsVersion>
  </PropertyGroup>
</Project>

Left a computer running unattended for some days on CDDA material to compare Win-x64 1.4.3 builds, taken from https://hydrogenaud.io/index.php/topic,124356.0.html , at various settings from -8pr7 (heaviest) to -0 --no-md5-sum. I included official 1.4.2 for a baseline.

As readers of the thread probably know already, results do depend on CPU and should be taken with that grain of salt.

Figures quoted in the table are speed relative to realtime - higher is better this post! (Edit3, got that wrong first edit!)

.	-8pr7	-8	-5	-2e	-0	'00'		00 is -0 --no-md5-sum. Remarks per build:
1.4.2 official	50	180	351	401	524	626		1.4.2 is second-to-slowest except -8pr7 (near-tied to Wombat-Clang)
1.4.3 official	58	209	451	468	609	751		Always #3.
Wombat-Clang	50	215	447	426	568	700		-8 was the 'least consistent timing' in the pack
Wombat-GCC	59	224	494	540	667	826		Fastest on LPC predictors, second-fastest (and close!) on fixed
Rarewares	54	174	355	410	592	737		Over AVX2: Penalty in seconds "not much bigger" for -8pr7 compared to -8
RW-AVX2	59	217	488	542	681	849		Fastest on fixed predictors, second-fastest (and close!) on LPC
Ozz	34	69	193	369	500	619		Ooof. Not equally disastrous on fixed predictors.

Fastest builds were always one of Wombat-GCC and Rarewares with AVX2 optimizations. Both improved over official at every preset tested.
Differences between the two fastest: Using Wombat-GCC over Rarewares on -5 & -8 & -8pr7 would save you about a minute per hour - and using Rarewares over Wombat-GCC on the three other modes would save you about a minute per hour as well.

Computer: HP Prodesk with i5-7500T @ 2.70 GHz.
Corpus: the 38 CDs in my signature. One file per CD.
"Method": "One run" := encode the 38 CD .wav images AND two re-encode from FLAC images. Did first a run to get the CPU to stable heat, discarded that, did three more runs and recorded the median of those three. For sanity-checking variability: also computed speeds using the fastest of those three runs; had I used fastest-of-three, the numbers in the table would have been 0 to 2 higher except Wombat-Clang at -8, which would have gone up from 215 to 222.

Left a computer running unattended for some days on CDDA material to compare Win-x64 1.4.3 builds, taken from https://hydrogenaud.io/index.php/topic,124356.0.html , at various settings
[...]
Computer: HP Prodesk with i5-7500T @ 2.70 GHz.

Meanwhile, it seems that fixed predictors aren't Clang's best friends either ... on this computer, at least.

-2er7, here is where the fastest Clang takes fifty percent more time than the winner
5:36 for WombatSomeflagsGCCdisableasm. I have no idea why this one makes ten percent faster than anything else - I re-ran and got pretty the same figures over again.

No idea when the point is reached it becomes silly