Cuda convolution reverb (Reverberate LE)

VST, AU, AAX, CLAP, etc. Plugin Virtual Effects Discussion
RELATED
PRODUCTS

Post

Hi

I've posted a beta up of a new convolution reverb that uses Cuda (there's also a native CPU build there too) that I've created and hope to get a little community feedback before I release it properly, see http://www.liquidsonics.com/publicbeta/

Some more details are in a post I put up in the dev section, but I am re-posting here to see if I can reach a wider audience, http://www.kvraudio.com/forum/viewtopic.php?t=240824

The plugin is for Windows VST and needs an 8 series and above NVIDIA card for the Cuda build. It's similar to that designed by Nils Schneider but has a fuller GUI, paragraphic EQ, ADSHR envelopes, support for flac files, pre-delay, 50-150% stretch and start/end crop.

Any feedback regarding installing/running/performance issues would be much appreciated.

Regards
Matt
Last edited by liquidsonics on Thu Sep 17, 2009 10:26 pm, edited 4 times in total.

Post

Great - looks interesting! Don't have NIVDIA here but will check out the native version. Can you tell me whether the plugin introduces any latency? I've love to see a convolution plug without latency... don't want to have to pay out for SIR 2 might be have to :)

Post

Zero dB wrote:Great - looks interesting! Don't have NIVDIA here but will check out the native version. Can you tell me whether the plugin introduces any latency? I've love to see a convolution plug without latency... don't want to have to pay out for SIR 2 might be have to :)
The Cuda and Native versions of Reverberate LE (which are going to be released for free or be very low cost) have a minimum of 512 samples latency. Getting the Cuda version down to zero latency isn't practical currently and I intend the native version to mirror the Cuda version functionally. I'm interested to know what kind of CPU hit people get with the lowest Cuda settings (Cuda is much more efficient the more you process at once).

Reverberate LE is a pre-cursor to the full Reverberate which is almost done, and this plugin will have a zero input/output latency mode (regardless of host/soundcard buffer setting) along with a number of other improvements. It's probably going to be native CPU only. I'll be charging for the enhanced plugin, but I'm going to try and keep the price low as by now I feel like there is space in the market for a lower priced convolution reverb option.

Regards
Matt

Post

NVGF9500GT :party:

off to test cuda vershion now ;)

Subz

Post

all seems to be working after i updated my gfx drivers :)

is there any way i can see the amount of GPU its using?

the sound is very good but the silent gap you get when it re calculates settings after moving a parameter is a bit annoying,

Subz

PS, do you plan on doing any other CUDA FX?

Post

loading 2 instances in XT2.5 causes sound to crap out at 256 buffer in host, (upping host buffer to 512 then i can load more than 1 instance)

Subz

Post

djsubject wrote:all seems to be working after i updated my gfx drivers :)

is there any way i can see the amount of GPU its using?

the sound is very good but the silent gap you get when it re calculates settings after moving a parameter is a bit annoying,

Subz

PS, do you plan on doing any other CUDA FX?
I do not plan to implement a GPU usage figure right now because the NVIDIA Cuda API does not currently expose such information. The only way to do it is to make an educated guess based on the time taken for Cuda to finish processing and I am not yet satisfied with the results (I don't want to put misleading info up and I find it's not sufficiently accurate with multiple instances).

I'll have a think about the silence during reprocessing issue. I may drop down to a dry signal rather than silence perhaps. I'd be interested if anybody else finds it really annoying.

While I'm experimenting with some other Cuda effects, but am concentrating on the reverb for now.

Thanks for your input :)

Post

djsubject wrote:loading 2 instances in XT2.5 causes sound to crap out at 256 buffer in host, (upping host buffer to 512 then i can load more than 1 instance)

Subz
Does upping the plugin latency help with the lower buffer size in XT?

Post

Hi,

I've just seen this thread. Is there any limitations/restrictions with the Reverberate LE public beta?

Post

dickiefunk wrote:Hi,

I've just seen this thread. Is there any limitations/restrictions with the Reverberate LE public beta?
Oops just read it expires in May!

Post

liquidsonics wrote:
djsubject wrote:loading 2 instances in XT2.5 causes sound to crap out at 256 buffer in host, (upping host buffer to 512 then i can load more than 1 instance)

Subz
Does upping the plugin latency help with the lower buffer size in XT?
i did try that first & it did not help,

Subz

Post

I posted a new version that drops down to dry signal during recompute.

I also found that energyXT seems not to call VST plugin destructors (seems to be a known bug looking at http://www.kvraudio.com/forum/viewtopic.php?p=3394932 that I don't think is fixed yet, I tried the v2.5 demo). After consuming all the GPU memory over many plugin instances it is never released by closing the plugin (as usually happens automatically by calling the plugin destructor). This is probably not such a noticable issue when using CPU based plugins with a ton of system ram but GPU ram is quickly used up in this plugin so it really needs to be freed when not in use.

I have not been able to replicate the energyXT 256 buffer size issue on my Core 2 Duo 2.4GHz system, so can't work out what the issue there is. Does it do the same on other sequencers on your platform, such as reaper?

I'd be interested if any other hosts have issues. Also I am keen to hear from people who it does work for and what systems they use so I can move towards the full release with more confidence.

Post

djsubject wrote:loading 2 instances in XT2.5 causes sound to crap out at 256 buffer in host, (upping host buffer to 512 then i can load more than 1 instance)

Subz
After a little investigation I've posted an update which should improve performance of the GPU version at 512 samples latency when using small soundcard buffer sizes. I'd be interested to know if this has improved the issue for you (you may still find lots of cascaded instances on the same channel will eventually clog up at low buffer sizes, but it should be ok in more normal usage models with multiple instances on different channels). I believe the CPU version was unaffected by the issue.

Post

I've upped a new version, which is an update for the CPU version to provide a more seamless transition between IR settings using double-buffering. I've also gone back to dropping back to silence for the GPU version during IR transitions to reduce audible artifacts.

Post

Updated GPU version now to drop to in-sync dry signal during recompute.

Post Reply

Return to “Effects”