Resolve 10 - Poor R3D performance

[Christopher*Seguine]

Problem: CPU based R3d decode usage is very poor.


First system 2x e5-2687w - 32 cores, seconds system has 4x E5-4650 - 64 cores
64 cores system performance is no better than 32 core system.

GPU = 2x Titan, GPU utilization is 40% or less
Storage = 32 drive raid, 1.25GB/sec



CPU Utilization goes DOWN when increasing to Full Resolution decode,
something is broken.



Source: Red R3d 5120x2700

Timeline: 3996x2160

Decklink 4k to 4k Projector




Half Res Good 24fps 20-30% CPU Utilization 8x DPdecoder.exe 5%
Half Res Prem 20-24fps 35-50% CPU 8x DPdecoder.exe 5%
Full Res Prem 5-8fps 15-25% CPU 8x DPdecoder.exe 2%




I have tried adjusting the config.dat to:


Local.Resource.Processors.1 = 4
Local.DtManger.1.Threads = 16
Local.DtManger.1.Red.Threads = 16
DpDecoder.1.Instances = 16


Increasing Threads and Instances to 16 has no increase in performance or CPU utilization, Each instance just uses less - 1%.
Increasing beyond 16 makes it worse - black frames



Something must be broken with only 25% CPU utilization for 5k full resolution decode.
CPU utilization should not go down when quality is increased from Half to Full.


We could use 2x RedRockets, however we do no use the Rocket for final renders due to the decode quality.
Switching them on for playback and off for render is problematic as there is a significant difference so it hard to make accurate grades.


CPU decode seems to be artificially limited or broken.

 

[Alan Gordon]

I have the same experience and it's terrible. I've brought it up with DaVinci support but have gotten no response on why this is.

My guess is the Software debayer isn't multi-threaded very well and each instance of DPDecoder debayers a single frame of R3D, so the performance becomes very CPU frequency dependent. But I haven't had the resources to test this theory fully.

I really hope RED releases the GPU decode to the SDK soon and that Blackmagic incorporates it into the next build of version 10.

 

[Christopher*Seguine]

GPU decode performance is better, but there is still some major problem between HalfHigh and FULL.

Full Decode makes poor use of GPU or CPU

system has 128gb ram, disk transfer is 1.25GB/sec avg, GPU utilization is less than 70%, there are no bottlenecks I can find besides r3d decoding.


CPU/Half High: 1 thread! barely using 5% cpu,

CPU/ Full Resolution: only using 3 cores! CPU usage less than half resolution decode

GPU/ Half High Resolution: (GTX 680/ -decodethreads 64) Better, but CPU utilization is still low.

GPU/ FULL Resolution: and worse again, not even using half the cores, and CPU and GPU utilization is much less than Half Resolution.

 

[Paul Russell]

I'm hoping that they can improve this so that I can spend more money on Nvidia cards and not have to buy a Rocket-X

 

[Christopher*Seguine]

"Number of Red Decoder decoders configured (64) is more than the number licensed(32)"

What is this nonsense?

RED is purposefully limiting the number of R3D decoder threads in Resolve?

 

[Alan Gordon]

It's probably limited in Resolve. Saved for linux folk perhaps?