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
2.7 KiB44 linesraw
1    CPU Tasks
2LoadModel       950.578 milliseconds
3RunComplete     27.5329 seconds
4Run     27.434 seconds
5Callbacks       10.6484 milliseconds, 44 calls, 242.009 microseconds average
6Spectrogram     199.106 milliseconds, 41 calls, 4.85624 milliseconds average
7Sample  58.7404 milliseconds, 527 calls, 111.462 microseconds average
8Encode  11.3813 seconds, 9 calls, 1.26459 seconds average
9Decode  16.0418 seconds, 9 calls, 1.78242 seconds average
10DecodeStep      15.9829 seconds, 527 calls, 30.3281 milliseconds average
11    GPU Tasks
12LoadModel       805.211 milliseconds
13Run     27.338 seconds
14Encode  11.3967 seconds, 9 calls, 1.2663 seconds average
15EncodeLayer     9.78685 seconds, 288 calls, 33.9821 milliseconds average
16Decode  15.9412 seconds, 9 calls, 1.77125 seconds average
17DecodeStep      15.9412 seconds, 527 calls, 30.249 milliseconds average
18DecodeLayer     15.0511 seconds, 16864 calls, 892.499 microseconds average
19    Compute Shaders
20mulMatTiled     12.0503 seconds, 6345 calls, 1.89919 milliseconds average
21mulMatByRowTiled        9.45404 seconds, 199430 calls, 47.4053 microseconds average
22norm    1.32432 seconds, 51704 calls, 25.6135 microseconds average
23fmaRepeat1      583.884 milliseconds, 51704 calls, 11.2928 microseconds average
24addRepeatEx     536.551 milliseconds, 51168 calls, 10.4861 microseconds average
25softMaxFixed    534.105 milliseconds, 17152 calls, 31.1395 microseconds average
26copyConvert     500.4 milliseconds, 34880 calls, 14.3463 microseconds average
27copyTranspose   377.38 milliseconds, 34304 calls, 11.001 microseconds average
28addRepeatScale  315.294 milliseconds, 33728 calls, 9.34814 microseconds average
29addRepeatGelu   283.978 milliseconds, 17170 calls, 16.5392 microseconds average
30softMaxLong     245.57 milliseconds, 527 calls, 465.976 microseconds average
31scaleInPlace    226.545 milliseconds, 17152 calls, 13.2081 microseconds average
32softMax 212.206 milliseconds, 16864 calls, 12.5834 microseconds average
33addRepeat       209.397 milliseconds, 17728 calls, 11.8117 microseconds average
34convolutionMain2Fixed   184.615 milliseconds, 9 calls, 20.5128 milliseconds average
35diagMaskInf     107.423 milliseconds, 16864 calls, 6.36998 microseconds average
36convolutionMain 74.7954 milliseconds, 9 calls, 8.3106 milliseconds average
37convolutionPrep1        20.9316 milliseconds, 18 calls, 1.16287 milliseconds average
38convolutionPrep2        3.8103 milliseconds, 18 calls, 211.683 microseconds average
39addRows 3.7939 milliseconds, 527 calls, 7.19905 microseconds average
40add     1.0895 milliseconds, 9 calls, 121.056 microseconds average
41    Memory Usage
42Model   892.591 KB RAM, 2.8815 GB VRAM
43Context 92.2612 MB RAM, 1.14026 GB VRAM
44Total   93.1329 MB RAM, 4.02176 GB VRAM