RadeonSI Driver Adding Displayable DCC For Raven Ridge

[phoronix]

Phoronix: RadeonSI Driver Adding Displayable DCC For Raven Ridge

Delta Color Compression (DCC) support for scan-out surfaces will soon be supported by AMD Raven Ridge hardware in conjunction with the latest AMDGPU Linux kernel code and the RadeonSI OpenGL driver...

http://www.phoronix.com/scan.php?page=news_item&px=RadeonSI-Raven-Scanout-DCC

[Venemo]

Can someone explain what exactly DCC is, and what kind of benefit does enabling it give to the end user?

[dlq84]

Can someone explain what exactly DCC is, and what kind of benefit does enabling it give to the end user?

Maybe this will answer your questions https://gpuopen.com/dcc-overview/

[schmidtbag]

Maybe this will answer your questions https://gpuopen.com/dcc-overview/

Pretty interesting read. The part that stood out to me the most was the part about using FP32. Really gets me to wonder though:
GPUs and software for nearly 20 years now have supported 24-bit color, never really going any higher, and there doesn't seem to be much of a push in the industry to change that (yes, I know there are true 32-bit programs and displays, but they're rare and expensive). So, why don't most gaming GPUs use "true" FP24? Shaving off all those "useless" transistors seems like they could offer a MAJOR improvement in terms of transistor density, overall performance and stability, and possibly decrease overall bandwidth consumption.
Meanwhile, for compute tasks, how many of them really need all of the registers in FP64? I'm sure FP48 would be sufficient in the vast majority of use-cases.

[andrei_me]

Michael a follow-up article of the current situation of the your raven ridge laptop would be nice

[Guest]

schmidtbag - You're forgetting the alpha channel.

[schmidtbag]

schmidtbag - You're forgetting the alpha channel.

According to the link provided by dlq84, the extra 8 bits are discarded. To me, that means they're effectively not used. Am I misunderstanding something?

Bandwidth. FP24 (weird) requires up to 3 times the bandwidth of 24-bit. Note that by 24-bit color they don't mean 24 bits per channel, but rather 8 bits per channel (3 channels).

I don't think you understood my question. Like I said to sandy8925, that link says 24-bit color doesn't allocate all 32 bits. So, if the vast majority of calculations on a GPU don't need 32 bit hardware, why not make it 24 bit hardware? In doing so, that would reduce bandwidth.

[slotdime]

@andrei_me , +1

[Guest]

schmidtbag - It says that's done for depth, not colour.

[Marc Driftmeyer]

According to the link provided by dlq84, the extra 8 bits are discarded. To me, that means they're effectively not used. Am I misunderstanding something?


I don't think you understood my question. Like I said to sandy8925, that link says 24-bit color doesn't allocate all 32 bits. So, if the vast majority of calculations on a GPU don't need 32 bit hardware, why not make it 24 bit hardware? In doing so, that would reduce bandwidth.

You're not writing a device driver to target video games, and all other major application spaces be damned. The #3 explanation alludes to much more than just `games,' and how FP32 is a better solution than FP16, which would require two operations and then the third to toss the 8 bit alpha when a specific application does not leverage it.