yum-archive/TaSTT-Whisper

High-performance GPGPU inference of OpenAI's Whisper automatic speech recognition (ASR) model

git clone https://git.yummers.dev/yum-archive/TaSTT-Whisper

KonstantinPerformance results for version 1.5993cfa5

master
2.8 KiB47 linesraw
1    CPU Tasks
2LoadModel       2.88964 seconds
3RunComplete     140.747 seconds
4Run     140.661 seconds
5Callbacks       20.302 milliseconds, 44 calls, 461.409 microseconds average
6Spectrogram     468.419 milliseconds, 41 calls, 11.4249 milliseconds average
7Sample  139.558 milliseconds, 527 calls, 264.815 microseconds average
8Encode  87.5396 seconds, 9 calls, 9.72662 seconds average
9Decode  53.0971 seconds, 9 calls, 5.89968 seconds average
10DecodeStep      52.9566 seconds, 527 calls, 100.487 milliseconds average
11    GPU Tasks
12LoadModel       1.86694 seconds
13Run     140.175 seconds
14Encode  88.7441 seconds, 9 calls, 9.86046 seconds average
15EncodeLayer     75.809 seconds, 288 calls, 263.226 milliseconds average
16Decode  51.4306 seconds, 9 calls, 5.71451 seconds average
17DecodeStep      51.43 seconds, 527 calls, 97.5901 milliseconds average
18DecodeLayer     48.1822 seconds, 16864 calls, 2.85711 milliseconds average
19    Compute Shaders
20mulMatTiledEx   69.1011 seconds, 2880 calls, 23.9934 milliseconds average
21mulMatTiled     21.009 seconds, 3465 calls, 6.06321 milliseconds average
22mulMatByRowTiled        20.0965 seconds, 166278 calls, 120.861 microseconds average
23mulMatByRowTiledEx      9.61326 seconds, 33152 calls, 289.975 microseconds average
24softMaxFixed    3.7631 seconds, 17152 calls, 219.397 microseconds average
25norm    2.23806 seconds, 51704 calls, 43.2859 microseconds average
26convolutionMain2Fixed   2.12825 seconds, 9 calls, 236.472 milliseconds average
27matReshapePanels        2.0333 seconds, 1737 calls, 1.17058 milliseconds average
28addRepeatGelu   1.5491 seconds, 17170 calls, 90.2211 microseconds average
29scaleInPlace    1.32928 seconds, 17152 calls, 77.5001 microseconds average
30copyConvert     1.23135 seconds, 34880 calls, 35.3026 microseconds average
31fmaRepeat1      1.10337 seconds, 51704 calls, 21.3401 microseconds average
32addRepeatEx     1.00095 seconds, 51168 calls, 19.562 microseconds average
33copyTranspose   846.807 milliseconds, 34304 calls, 24.6854 microseconds average
34addRepeat       704.028 milliseconds, 17728 calls, 39.7128 microseconds average
35softMaxLong     608.58 milliseconds, 527 calls, 1.1548 milliseconds average
36convolutionMain 522.249 milliseconds, 9 calls, 58.0277 milliseconds average
37addRepeatScale  500.937 milliseconds, 33728 calls, 14.8523 microseconds average
38softMax 236.054 milliseconds, 16864 calls, 13.9975 microseconds average
39diagMaskInf     171.964 milliseconds, 16864 calls, 10.1971 microseconds average
40convolutionPrep1        60.7331 milliseconds, 18 calls, 3.37406 milliseconds average
41convolutionPrep2        33.441 milliseconds, 18 calls, 1.85783 milliseconds average
42add     12.0883 milliseconds, 9 calls, 1.34314 milliseconds average
43addRows 1.9724 milliseconds, 527 calls, 3.74269 microseconds average
44    Memory Usage
45Model   892.591 KB RAM, 2.8815 GB VRAM
46Context 92.2612 MB RAM, 1.19934 GB VRAM
47Total   93.1329 MB RAM, 4.08084 GB VRAM