Use CUDA to speed up encoding > 8bit color formats

Vouk

Playing with CUDA right now and already got a speed increase of 69% when using bit depths greater than 8 bit on my development machine.

Still working on it.

Vouk

Exporting a project with 2048 x 1152 and float32 color precision (per channel).

1 pixel per CUDA thread processing 4 floats per FMA in a 32 x 32 block

Unfortunately the cudaMemCopy's are like 10-15ms each. Processing more frames per call would speed this up even more.

System One

AVX2 (i7-7700K)

[11:26:35] Frame #277: vRender: 39 us, vProcess: 41738 us, vEncoding: 19984 us, aRender: 70 us, aEncoding: 268 us, Latency: 64781 us

[11:26:35] Frame #278: vRender: 37 us, vProcess: 44431 us, vEncoding: 19885 us, aRender: 61 us, aEncoding: 12 us, Latency: 66833 us

[11:26:35] Frame #279: vRender: 33 us, vProcess: 41376 us, vEncoding: 18816 us, aRender: 62 us, aEncoding: 310 us, Latency: 62777 us

[11:26:35] Frame #280: vRender: 39 us, vProcess: 43909 us, vEncoding: 18867 us, aRender: 55 us, aEncoding: 218 us, Latency: 65696 us

[11:26:35] Frame #281: vRender: 35 us, vProcess: 43756 us, vEncoding: 20499 us, aRender: 53 us, aEncoding: 231 us, Latency: 66800 us

[11:26:35] Frame #282: vRender: 31 us, vProcess: 43390 us, vEncoding: 20808 us, aRender: 65 us, aEncoding: 315 us, Latency: 66789 us

CUDA (Quadro P2000)

[09:54:33] Frame #288: vRender: 40 us, vProcess: 18746 us, vEncoding: 22134 us, aRender: 68 us, aEncoding: 19 us, Latency: 44891 us

[09:54:33] Frame #289: vRender: 36 us, vProcess: 21211 us, vEncoding: 18160 us, aRender: 72 us, aEncoding: 322 us, Latency: 42256 us

[09:54:34] Frame #290: vRender: 36 us, vProcess: 18531 us, vEncoding: 20253 us, aRender: 59 us, aEncoding: 214 us, Latency: 41408 us

[09:54:34] Frame #291: vRender: 35 us, vProcess: 18369 us, vEncoding: 22336 us, aRender: 70 us, aEncoding: 327 us, Latency: 43288 us

[09:54:34] Frame #292: vRender: 40 us, vProcess: 17668 us, vEncoding: 18668 us, aRender: 63 us, aEncoding: 17 us, Latency: 38560 us

[09:54:34] Frame #293: vRender: 36 us, vProcess: 17704 us, vEncoding: 19705 us, aRender: 71 us, aEncoding: 327 us, Latency: 40145 us

System Two

AVX2 (i7-8700K)

[11:30:00] Frame #70: vRender: 31 us, vProcess: 36414 us, vEncoding: 16255 us, aRender: 1083 us, aEncoding: 10 us, Latency: 55197 us

[11:30:00] Frame #71: vRender: 83 us, vProcess: 40397 us, vEncoding: 15759 us, aRender: 577 us, aEncoding: 244 us, Latency: 59374 us

[11:30:00] Frame #72: vRender: 30 us, vProcess: 36319 us, vEncoding: 15735 us, aRender: 930 us, aEncoding: 357 us, Latency: 54855 us

[11:30:00] Frame #73: vRender: 1774 us, vProcess: 47668 us, vEncoding: 70102 us, aRender: 13 us, aEncoding: 245 us, Latency: 121703 us

[11:30:00] Frame #74: vRender: 34 us, vProcess: 40626 us, vEncoding: 15824 us, aRender: 610 us, aEncoding: 8 us, Latency: 58531 us

[11:30:00] Frame #75: vRender: 35 us, vProcess: 40386 us, vEncoding: 15860 us, aRender: 565 us, aEncoding: 234 us, Latency: 58775 us

CUDA (GeForce RTX 2080 TI)

[09:28:47] Frame #1720: vRender: 30 us, vProcess: 12659 us, vEncoding: 13886 us, aRender: 893 us, aEncoding: 332 us, Latency: 29435 us

[09:28:47] Frame #1721: vRender: 38 us, vProcess: 13909 us, vEncoding: 17943 us, aRender: 894 us, aEncoding: 421 us, Latency: 35310 us

[09:28:47] Frame #1722: vRender: 39 us, vProcess: 13063 us, vEncoding: 14418 us, aRender: 558 us, aEncoding: 8 us, Latency: 30184 us

[09:28:47] Frame #1723: vRender: 32 us, vProcess: 13319 us, vEncoding: 14304 us, aRender: 12 us, aEncoding: 343 us, Latency: 29725 us

[09:28:47] Frame #1724: vRender: 51 us, vProcess: 14712 us, vEncoding: 15048 us, aRender: 653 us, aEncoding: 244 us, Latency: 33087 us

[09:28:47] Frame #1725: vRender: 30 us, vProcess: 13147 us, vEncoding: 15400 us, aRender: 570 us, aEncoding: 7 us, Latency: 30813 us

Vouk

NOT FOR PRODUCTION USE

Test release: https://github.com/Vouk/voukoder-….0.msi?raw=true

(Use the latest NVIDIA driver version)

Vouk

Please report if ...

It works
It crashes
it is faster than the version before
it is slower than before
...

Kleinrotti

Stats for 1080p 10Bit nvenc Hevc

Connector 1.5.0:

- Average vProcess 4700us

- FPS 47

Connector 1.6.0:

- Average vProcess 9200us

- FPS 37

Hardware:

AMD TR1950x

GTX 1080ti with newest driver

32GB RAM

Vouk

Yes, please use i.e. YUV 4:2:0 10 bit.

Kleinrotti Seems the high core count works pretty good with AVX2 code.

@All Other experiences?

Kleinrotti

Zitat

@Kleinrotti Seems the high core count works pretty good with AVX2 code.

Maybe it would be nice to have a checkbox in the future, to use AVX2 or CUDA. To prevent a bottleneck for high core count CPUS.

Vouk

I'm not sure how CUDA will be finally implemented. Currently I'd like to get some feedback first. But yes, I won't force you to use slower functionality.

Kleinrotti

Is there a AMD GPU equivalent to CUDA to use this speed improvement on AMD cards too?

Vouk

AMD has a CUDA equivalent called ROC but it isn't that popular.

kobol

Want to test it, but cant find the settings to set YUV 4:2:0 10 bit in Premiere?! Please help, where is it?

Vouk

Use the x264 or x265 encoder. You will find it in the "Pixel Format" dropdown.

kobol

Der Inhalt kann nicht angezeigt werden, da er nicht mehr verfügbar ist.

No such settings ...

Oops, sorry. Found it in X264.

Tested it. Its about 4 times slower than using NVENC X264.

hctsly

CRF encoding is in use, it cannot be accelerated by GPU, can it be added?

Vouk

No, not possible. I really wish other encoders would support it, too.