Create an arm64 build compatible with apple silicon M-series macs

[newadventure079]

Right now, the app is x86_64. Code Type: X86-64 (Translated)

Adding compiling and building steps to create an arm64 build that runs natively on apple silicon macs would be great.

[theophanemayaud]

That would be pretty great, I've been wanting to do this as soon as M1 came out. QT is now working with arm, but only on Mac, not windows. As this app is cross platform, I need 1 work machine to be able to build for both windows and macOS. So for now, I'll wait on qt.

[theophanemayaud]

I found a way to build as universal app for apple arm and intel machines. This involved building ffmpeg and opencv as universal libraries as well. OpenCV is simple, as it is build with CMake which directly supports a parameter building as universal. ffmep however involves building separately for each architecture (it can still both be done on M1 simply cross building for x86), then using apple command line tools to merge both libraries into a universal library. This is keeping ffmpeg at the same minor version (from 4.4.0 to 4.4.4) but updating opencv to 4.8.0 from 4.5.1.

Now I need to test if it still works as expected, and performance.

Testing performance (M1 MacBook Air 8gb ram, external SSD; 2018 i5 quad core MB Pro 16gb ram, external SSD): Careful, file discovery might need to be redone after intensive long load, it seems initial run is being a lot faster than it throttles. Starting with universal, I tried to first run it once, then immediately after run it a second time so it doesn't cool down.

File discovery

• Intel on M1 no cache: 40/s (about 9000 files in 2 minutes, and 1 min 15 s from 7000 to 1000... thermal throttling ?!) Process takes 100% but only on 1 core.
• Intel on M1 with cache: 36/s and drops to 36/s (55s for 6 000 but drops to 2min 46s for 6000 to 12000... thermal throttling ?!) 100% of only 1 core
• Universal on M1 no cache: 46/s (107/s: 1min33s for first 10 000 after preheating with a first complete run before; then 46/s: 43s for 2 000). 100% 1 core
• Universal on M1 with cache: 45/s (113/s: 1:28 for first 10 000; then 45/s: 44s from 10 000 to 12 000)
• Universal on Intel no cache: 14/s (33/s: 5:05 for first 10 000; then 14/s: 2:23 from 10 000 to 12 000)
• Universal on Intel with cache: 14/s (33/s: 5:03 for first 10 000; then 14/s: 2:21 from 10 000 to 12 000)
• Intel on Intel no cache: 14/s (33/s: 5:02 for 10 000; then 14/s: 2:23 from 10 000 to 12 000). Takes 100% of 1 core
• Intel on Intel with cache: 13/s (32/s: 5:14 for 10 000; then 13/s 2:34 from 10 000 to 12 000). Takes 100% of 1 core

Video processing

• Intel on M1 no cache: 14/s (3000 in 3min 33s). Process takes all cores to 100%.
• Intel on M1 with cache : 179/s (400 to 7400 in 39s). Process takes all cores to 100%
• Universal on M1 no cache: 28/s (20/s: 1:59 for first 2 400; then 28/s: 4:05 from 5 000 to 12000). Process takes all cores to 100%
• Universal on M1 with cache: 196/s (196/s: 51s for 10 000; then 285/s: 7s from 10 000 to 12 000)
• Universal on Intel no cache: 11/s (11/s: 15:31 for first 10 000; then 18/s: 1:49 from 10 000 to 12 000)
• Universal on Intel with cache: 113/s (113/s: 1:02 from 1000 to 8000)
• Intel on Intel no cache: 9/s (13/s: 14:30 for 11 500; then 9/s: 1:55 from 11 500 to 12 500). Takes all cores to 100%
• Intel on Intel with cache: 103/s (103/s 1:37 for 10 000; then 181/s: 11s from 10 000 to 12 000). Takes all cores to 100%

Overall time to 1st pair (270GB folder of 12 505 videos)

• Intel on M1 no cache: 16min 15s
• Intel on M1 with cache: 5min 5s
• Universal on M1 no cache: 11:35
• Universal on M1 with cache: 3:25
• Universal on Intel no cache: 26:36
• Universal on Intel with cache: 10:12
• Intel on Intel no cache: 24:33
• Intel on Intel with cache: 10:31

Intact videos

• Intel on M1 no cache: 12 328
• Intel on M1 with cache: 12 330
• Universal on M1 no cache: 12 330
• Universal on M1 with cache: 12 330
• Universal on Intel no cache:
• Universal on Intel with cache: 12 330
• Intel on Intel no cache: 12 328 intact video(s) out of 12505 total
• Intel on Intel with cache: 12 330

Found duplicates

• Intel on M1 no cache: 6 557 at 97.1 GB
• Intel on M1 with cache: 6 542 at 96.9 GB
• Universal on M1 no cache: 6 568 at 97.2 GB
• Universal on M1 with cache: 6 555 at 97.1 GB
• Universal on Intel no cache:
• Universal on Intel with cache: 6 545 at 96.9GB
• Intel on Intel no cache: 6558 at 97.1 GB
• Intel on Intel with cache: 6543 at 96.9 GB

Recap	Machine	File discovery	Video processing	Overall time to 1st pair	Intact videos	Found duplicates.
Intel on M1 (no cache)	40/s	14/s	16min 15s	12 328	6 557 at 97.1 GB
Intel on M1 (with cache)	36/s	179/s	5min 5s	12 330	6 542 at 96.9 GB
Universal on M1 (no cache)	46/s	28/s	11min 35s	12 330	6 568 at 97.2 GB
Universal on M1 (with cache)	45/s	196/s	3min 25s	12 330	6 555 at 97.1 GB
Universal on Intel (no cache)	14/s	11/s	26min 36s
Universal on Intel (with cache)	14/s	113/s	10min 12s	12 330	6 545 at 96.9GB
Intel on Intel (no cache)	14/s	9/s	24min 33s	12 328	6 558 at 97.1 GB
Intel on Intel (with cache)	13/s	103/s	10min 31s	12 330	6 543 at 96.9 GB

So overall, the new performance is good :

• On arm macs, it finds a few more duplicates than before but in 30% less time
• On intel macs, it finds the same number of duplicates in about the same time

[tompalmer]

Any ETA for completion ?

[theophanemayaud]

Hi, happy new year ! Sorry for the looong delay. I've just submitted to the apple App Store the new universal binary app build natively now also for arm (M1/M2/M3) with version 1.10.0 that I just !!! It should be out in a few days once apple approves it. I hope you like it and it helps you out :) It's about 30% faster than the rosetta translated version build only for intel x86.