Hi Andre,

I'm not familiar with the sample c message optimization. Probably Ken can comment.

However I could check the internal spec here and I saw this bit only exists with this meaning in Haswell.

For all the other platforms, including Broadwell it got re-purposed with a different meaning and a programming note: "This bit should be programmed to zero (0h) at all times."

Also, I could not find any workaround documented anywhere recommending this bit to be set.

So, I would not recommend to use it in any product, even downstream. Regardless the state of sample c message optimization in later platforms.

Thanks, Rodrigo.

On Tue, May 04, 2021 at 08:07:14PM -0300, André Almeida wrote:

Hi there,

While browsing an old downstream kernel, I found a patch[0] that enables sample_c optimizations at Broadwell GPUs. The message from the upstream commit that enables it for Haswell[1] (and presumably where the code at[0] was copied from) states that "[..] later platforms remove this bit, and apparently always enable the optimization".

Could you confirm that Broadwell and following architectures enable this optimization by default (and thus, patch[0] is a no-op), or should I upstream it?

Thanks, André

[0] https://github.com/ValveSoftware/steamos_kernel/commit/198990f13e1d9429864c1...

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?i...

Hello,

Yes, that bit only exists on Haswell. On Haswell, sample_c operations were processed at 1 pixel/clock unless you set that bit, in which case they get processed at 4 pixels/clock. The downside is that it breaks some obscure media feature that apparently no one used.

Broadwell and later alway process sample_c operations at the fast speed, and there is no bit to control it. I would say that downstream patch is incorrect.

--Ken

On Tuesday, May 4, 2021 4:07:14 PM PDT André Almeida wrote:

Hi there,

While browsing an old downstream kernel, I found a patch[0] that enables sample_c optimizations at Broadwell GPUs. The message from the upstream commit that enables it for Haswell[1] (and presumably where the code at[0] was copied from) states that "[..] later platforms remove this bit, and apparently always enable the optimization".

Could you confirm that Broadwell and following architectures enable this optimization by default (and thus, patch[0] is a no-op), or should I upstream it?

Thanks, André

[0] https://github.com/ValveSoftware/steamos_kernel/commit/198990f13e1d9429864c1...

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?i...