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
2.9 KiB47 linesraw
1    CPU Tasks
2LoadModel       1.49776 seconds
3RunComplete     110.474 seconds
4Run     110.407 seconds
5Callbacks       14.0412 milliseconds, 44 calls, 319.118 microseconds average
6Spectrogram     201.605 milliseconds, 41 calls, 4.91719 milliseconds average
7Sample  65.5117 milliseconds, 527 calls, 124.311 microseconds average
8Encode  64.8806 seconds, 9 calls, 7.20896 seconds average
9Decode  45.5097 seconds, 9 calls, 5.05663 seconds average
10DecodeStep      45.4439 seconds, 527 calls, 86.2313 milliseconds average
11    GPU Tasks
12LoadModel       951.086 milliseconds
13Run     110.123 seconds
14Encode  65.7998 seconds, 9 calls, 7.31109 seconds average
15EncodeLayer     56.2581 seconds, 288 calls, 195.341 milliseconds average
16Decode  44.3232 seconds, 9 calls, 4.9248 seconds average
17DecodeStep      44.3227 seconds, 527 calls, 84.1039 milliseconds average
18DecodeLayer     41.6477 seconds, 16864 calls, 2.46962 milliseconds average
19    Compute Shaders
20mulMatTiledEx   51.4882 seconds, 2880 calls, 17.8778 milliseconds average
21mulMatByRowTiled        17.0626 seconds, 166278 calls, 102.615 microseconds average
22mulMatTiled     15.2984 seconds, 3465 calls, 4.41513 milliseconds average
23mulMatByRowTiledEx      9.04106 seconds, 33152 calls, 272.715 microseconds average
24softMaxFixed    3.04176 seconds, 17152 calls, 177.342 microseconds average
25norm    1.98157 seconds, 51704 calls, 38.3254 microseconds average
26convolutionMain2Fixed   1.8145 seconds, 9 calls, 201.611 milliseconds average
27matReshapePanels        1.526 seconds, 1737 calls, 878.526 microseconds average
28addRepeatGelu   1.24631 seconds, 17170 calls, 72.5866 microseconds average
29scaleInPlace    1.23743 seconds, 17152 calls, 72.1448 microseconds average
30copyConvert     1.02044 seconds, 34880 calls, 29.2557 microseconds average
31fmaRepeat1      993.664 milliseconds, 51704 calls, 19.2183 microseconds average
32addRepeatEx     953.85 milliseconds, 51168 calls, 18.6415 microseconds average
33copyTranspose   705.073 milliseconds, 34304 calls, 20.5537 microseconds average
34addRepeat       581.089 milliseconds, 17728 calls, 32.778 microseconds average
35addRepeatScale  553.89 milliseconds, 33728 calls, 16.4222 microseconds average
36softMaxLong     387.949 milliseconds, 527 calls, 736.147 microseconds average
37convolutionMain 363.351 milliseconds, 9 calls, 40.3723 milliseconds average
38softMax 242.956 milliseconds, 16864 calls, 14.4068 microseconds average
39diagMaskInf     179.046 milliseconds, 16864 calls, 10.617 microseconds average
40convolutionPrep1        45.2096 milliseconds, 18 calls, 2.51164 milliseconds average
41convolutionPrep2        28.6853 milliseconds, 18 calls, 1.59363 milliseconds average
42add     8.1107 milliseconds, 9 calls, 901.189 microseconds average
43addRows 1.542 milliseconds, 527 calls, 2.926 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