NVIDIA Releases DALI Library & nvJPEG GPU-Accelerated Library For JPEG Decode

[phoronix]

Phoronix: NVIDIA Releases DALI Library & nvJPEG GPU-Accelerated Library For JPEG Decode

For coinciding with the start of the Computer Vision and Patern Recognition conference starting this week in Utah, NVIDIA has a slew of new software announcements...

http://www.phoronix.com/scan.php?page=news_item&px=NVIDIA-DALI-nvJPEG

[AsuMagic]

What is GPU-accelerated JPEG decompression useful for? Processing big batches of images?

[czz0]

Seems useless for desktop use, even for batch editing images. Maybe some supercomputer/server working with JPG images will benefit.

[Ikaris]

Processing big batches of images?

Seems like it. https://blogs.nvidia.com/blog/2018/0...e-datacenters/ You don't have to read the whole thing, just look at images, it will be very obvious.

I can see the CPU and PCIe becoming bottlenecks when feeding several GPUs with pictures in datacenters optimized for data read throughput.

[wizard69]

What is GPU-accelerated JPEG decompression useful for? Processing big batches of images?

My guess would be that it is helpful for images stored on local GPU memory. Otherwise i cant recall of a decompression delay in recent times.

[nh2_]

My guess would be that it is helpful for images stored on local GPU memory. Otherwise i cant recall of a decompression delay in recent times.

Large jpeg images (e.g. from high-resolution DSLRs) can still take ~0.5 to 1 seconds to decode; you can see it when switching through them quickly with an image viewer like `eog` and the arrow keys. It can be very annoying.

[carewolf]

Large jpeg images (e.g. from high-resolution DSLRs) can still take ~0.5 to 1 seconds to decode; you can see it when switching through them quickly with an image viewer like `eog` and the arrow keys. It can be very annoying.

Get a better image viewer with a better decoder.

[coder]

What is GPU-accelerated JPEG decompression useful for? Processing big batches of images?

As with most of Nvidia's recent activities, this is all about deep learning.

Imagine training sets comprised of hundreds of thousands or millions of images. The benefits of reading these from storage and sending them to the GPU in the compressed domain, then having the GPU decode them should be self-evident. Even if it doesn't relieve a throughput bottleneck, having the GPU decode them is certainly more power-efficient.

The talk of hybrid decoding suggests they're probably running the Huffman decoder on the CPU. As that's branch-heavy, serial code, it wouldn't make sense to waste GPU cycles on it.

[coder]

Get a better image viewer with a better decoder.

Better yet, install libjpeg-turbo - a drop-in replacement for libjpeg:

https://libjpeg-turbo.org/

[Beerbaron23]

Get a better image viewer with a better decoder.

He was most likely referring to Raw images taken with a DSLR camera, they get big, 40MB per picture or more. If the JPEG's are doing it, it's most likely the programs default adjutsments for viewing to liven them up a little.

But obviously this tech is for where you would need to decode massive amounts of jpeg data, like searching for a specific person via real-time face detection or whatever else they are cooking up...