RISC-V Adding Kernel-Mode FPU For Linux 6.10 To Enable Recent AMD Display Support

[phoronix]

Phoronix: RISC-V Adding Kernel-Mode FPU For Linux 6.10 To Enable Recent AMD Display Support

With the upcoming Linux 6.10 kernel cycle, the RISC-V architecture code is seeing kernel-mode FPU. This kernel floating point support is needed for the AMDGPU kernel graphics driver and particular its DCN display code. In turn this should allow recent AMD Radeon graphics cards to work on RISC-V with display support using the company's open-source driver stack...

https://www.phoronix.com/news/RISC-V-Linux-6.10-Kernel-FPU

[rene]

Pretty cool that finally –over two years later– this RISCV64 RDNA and newer amdgpu work of this YouTuber who shall not be named is finally getting upstream! https://www.youtube.com/watch?v=Sv4-_a_3BKg ;-) ¯\_(ツ)_/¯

[qarium]

Pretty cool that finally –over two years later– this RISCV64 RDNA and newer amdgpu work of this YouTuber who shall not be named is finally getting upstream! https://www.youtube.com/watch?v=Sv4-_a_3BKg ;-) ¯\_(ツ)_/¯

to me it really sounds strange for over 30+ years the linux kernel people told us that the linux kernel do not need to support loating point FPU math...

but good job rene you are a hero. overcome a 30 year old stigma to not support FPU in the kernel is a bonus

[willmore]

to me it really sounds strange for over 30+ years the linux kernel people told us that the linux kernel do not need to support loating point FPU math...

but good job rene you are a hero. overcome a 30 year old stigma to not support FPU in the kernel is a bonus

Have you looked at what a mess floating point is on x86?

Why does a display driver need floating point?

[qarium]

Have you looked at what a mess floating point is on x86?
Why does a display driver need floating point?

well i can say for sure that x86 is not only a mess in case of floating point ...

its for many years clear that x86 is broken by design and a mess...

https://wccftech.com/snapdragon-x-elite-pc-cpus-over-50-percent-faster-vs-intel-core-ultra-chips/

if you see benchmarks of Snapdragon X Elite it becomes clear that in notebooks x86 is more or less obsolete.

"Why does a display driver need floating point?"

i really don't know...

[agd5f]

Why does a display driver need floating point?

The bandwidth management code needs it. It uses a software model of the display block to validate the requested display configuration.

[discordian]

The bandwidth management code needs it. It uses a software model of the display block to validate the requested display configuration.

why is fractional math not enough there? or do you need to load gpu registers with the ieee754 format?

[agd5f]

why is fractional math not enough there? or do you need to load gpu registers with the ieee754 format?

The model is also used for a number of purposes outside of the driver and it was not feasible to rewrite it and revalidate it using fixed point math.

[smitty3268]

why is fractional math not enough there? or do you need to load gpu registers with the ieee754 format?

It's a copy of their windows code.

AGX also has to do fp math for their apple gpu driver. I believe it pulls ieee values out of the firmware and needs to do some simple math on them to program registers. I think it has something to do with power management, if i recall correctly.