[PATCH 1/2] drm/amdgpu: Enable scatter gather display support

[Alex Deucher]

On Mon, Mar 19, 2018 at 4:29 PM, Li, Samuel <Samuel.Li at amd.com> wrote:
>>to my earlier point, there may be cases where it is advantageous to put
>> display buffers in vram even if s/g display is supported
>
> Agreed. That is also why the patch has the options to let user select where
> to put display buffers.
>
> As whether to put the option in Mesa or kernel, it seems the difference is
> not much. Also, since amdgpufb can request even without mesa, kernel might
> be a better choice. In addition, putting in the kernel can save client’s
> duplicate work(mesa, ogl, vulkan, 2d, kernel…)

Why do we even expose different memory pools to the UMDs in the first
place ;)  Each pool has performance characteristics that may be
relevant for a particular work load.  Only the UMDs really know the
finer points of those workloads. In general, you don't want the kernel
dictating policy if you can avoid it.  The kernel exposes
functionality and userspace sets the policy.  With the location set in
userspace, each app/user can have whatever policy makes sense for
their use case all at the same time without needing to tweak their
kernel for every use case.

Alex

>
>
>
> Regards,
>
> Samuel Li
>
>
>
> From: Marek Olšák [mailto:maraeo at gmail.com]
> Sent: Monday, March 19, 2018 4:27 PM
> To: Deucher, Alexander <Alexander.Deucher at amd.com>
> Cc: Li, Samuel <Samuel.Li at amd.com>; Koenig, Christian
> <Christian.Koenig at amd.com>; Alex Deucher <alexdeucher at gmail.com>; Michel
> Dänzer <michel at daenzer.net>; amd-gfx list <amd-gfx at lists.freedesktop.org>
>
>
> Subject: Re: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> When Mesa wants a buffer in VRAM, it always sets VRAM. It relies on BO move
> throttling to prevent unnecessary BO moves.
>
>
>
> My questions are:
>
> - what should Mesa do differently for tiny VRAM?
>
> - what is a tiny VRAM?
>
> - if VRAM is tiny, which allocations should we put there?
>
>
>
> Marek
>
>
>
> On Mon, Mar 19, 2018 at 4:15 PM, Deucher, Alexander
> <Alexander.Deucher at amd.com> wrote:
>
> s/not/now/.  I meant to say, “we have NOW excluded the possibility of ever
> setting displays anywhere else without a kernel update”.
>
>
>
> Alex
>
>
>
> From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On Behalf Of
> Deucher, Alexander
> Sent: Monday, March 19, 2018 4:13 PM
>
>
> To: Li, Samuel <Samuel.Li at amd.com>; Koenig, Christian
> <Christian.Koenig at amd.com>; Marek Olšák <maraeo at gmail.com>
> Cc: Alex Deucher <alexdeucher at gmail.com>; Michel Dänzer
> <michel at daenzer.net>; amd-gfx list <amd-gfx at lists.freedesktop.org>
> Subject: Re: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> I'm not sure what you mean by the 3 scenarios.  Generally userspace selects
> what domains it wants a buffer to be in, vram, gtt, or both (don't care).
> I'd rather not have the kernel second guess the UMDs if we can help it.  I'd
> rather leave the kernel for cases where we have to force things due to hw
> bugs, or hw restrictions, etc.  If we force all display buffers to be in gtt
> in the kernel, we have not excluded the possibility of ever setting displays
> anywhere else without a kernel update.  E.g., to my earlier point, there may
> be cases where it is advantageous to put display buffers in vram even if s/g
> display is supported.  That was the point I was trying to make about user
> mode selecting the domain (vram of gtt or vram|gtt).  Say you have a board
> with 2 GB of ram and 1 GB is carved out for "vram".  In that case, it would
> make sense to put the buffer in vram because otherwise you are wasting a
> comparatively scarce resource.
>
>
>
> Alex
>
> ________________________________
>
> From: Li, Samuel
> Sent: Monday, March 19, 2018 3:58:52 PM
> To: Deucher, Alexander; Koenig, Christian; Marek Olšák
> Cc: Alex Deucher; Michel Dänzer; amd-gfx list
> Subject: RE: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> Alex,
>
>
>
> I assume you are talking the three scenarios here, 1)VRAM, 2)GTT,
> 3)VRAM/GTT.
>
> But kernel will need the decision too(amdgpufb). I think it shall be better
> to do it in kernel, instead of different clients(mesa, ddx, kernel …)
>
>
>
> Regards,
>
> Samuel Li
>
>
>
> From: Deucher, Alexander
> Sent: Monday, March 19, 2018 3:54 PM
> To: Li, Samuel <Samuel.Li at amd.com>; Koenig, Christian
> <Christian.Koenig at amd.com>; Marek Olšák <maraeo at gmail.com>
> Cc: Alex Deucher <alexdeucher at gmail.com>; Michel Dänzer
> <michel at daenzer.net>; amd-gfx list <amd-gfx at lists.freedesktop.org>
> Subject: Re: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> My personal preference is still to plumb this through to mesa rather than
> forcing it in the kernel.
>
>
>
> Alex
>
> ________________________________
>
> From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> on behalf of Li,
> Samuel <Samuel.Li at amd.com>
> Sent: Monday, March 19, 2018 3:50:34 PM
> To: Koenig, Christian; Marek Olšák
> Cc: Alex Deucher; Michel Dänzer; amd-gfx list
> Subject: RE: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> Christian,
>
>
>
> You misunderstood Alex’s comments,
>
>
>
>>Regardless of which scenarios we need to support, I think we also need
>
>>to really plumb this through to mesa however since user space is who
>
>>ultimately requests the location.  Overriding it in the kernel gets
>
>>tricky and can lead to ping-ponging as others have noted.  Better to
>
>
>
> Here Alex mentioned the scenarios is 1)VRAM, 2)GTT, 3)VRAM/GTT.
>
> His concern is this might cause ping-pong, not about preferred domain. Since
> preferred domain can solve the ping-pong issue, it shall address his concern
> here.
>
>
>
> Regards,
>
> Samuel Li
>
>
>
> From: Christian König [mailto:ckoenig.leichtzumerken at gmail.com]
> Sent: Monday, March 19, 2018 3:45 PM
> To: Li, Samuel <Samuel.Li at amd.com>; Marek Olšák <maraeo at gmail.com>; Koenig,
> Christian <Christian.Koenig at amd.com>
> Cc: Alex Deucher <alexdeucher at gmail.com>; Michel Dänzer
> <michel at daenzer.net>; amd-gfx list <amd-gfx at lists.freedesktop.org>
> Subject: Re: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> Quoting Alex:
>
> Regardless of which scenarios we need to support, I think we also need
>
> to really plumb this through to mesa however since user space is who
>
> ultimately requests the location.  Overriding it in the kernel gets
>
> tricky and can lead to ping-ponging as others have noted.  Better to
>
> have user space know what chips support it or not and request display
>
> buffers in GTT or VRAM from the start.
>
> And I completely agree with Alex here. So overriding the domain in the
> kernel is a serious NAK from my side as well.
>
> Please implement the necessary bits in Mesa, shouldn't be more than a few
> lines of code anyway.
>
> Regards,
> Christian.
>
> Am 19.03.2018 um 20:42 schrieb Li, Samuel:
>
> Agreed.
>
>
>
>>I think that the consensus with Alex and me is that we should avoid exactly
>> that.
> Christian, Alex’s concern is about ping-pong, not about the preferred
> domain.
>
> Regards,
>
> Samuel Li
>
>
>
> From: Marek Olšák [mailto:maraeo at gmail.com]
> Sent: Monday, March 19, 2018 3:39 PM
> To: Koenig, Christian <Christian.Koenig at amd.com>
> Cc: Li, Samuel <Samuel.Li at amd.com>; Michel Dänzer <michel at daenzer.net>; Alex
> Deucher <alexdeucher at gmail.com>; amd-gfx list
> <amd-gfx at lists.freedesktop.org>
> Subject: Re: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
>
>
> On Mon, Mar 19, 2018 at 3:27 PM, Christian König <christian.koenig at amd.com>
> wrote:
>
> I think that the consensus with Alex and me is that we should avoid exactly
> that.
>
> Overriding the preferred domain in the kernel is a no-go for that patch set,
> so please implement the discussed changes in Mesa.
>
>
>
> I don't see how Mesa can make a smarter decision than the kernel. If you
> overwrite the preferred domain of the buffer in the kernel, there will be no
> ping-ponging between domains. Mesa never changes the initial preferred
> domain.
>
>
>
> Marek
>
>
>
>
>
>
> Regards,
> Christian.
>
>
>
> Am 19.03.2018 um 20:22 schrieb Li, Samuel:
>
> I agree with Marek/Michel: since kernel sets the domain before scanning out,
> it shall update the preferred domain here too.
>
> Regards,
> Samuel Li
>
> -----Original Message-----
> From: Koenig, Christian
> Sent: Thursday, March 08, 2018 4:07 AM
> To: Michel Dänzer <michel at daenzer.net>; Li, Samuel
> <Samuel.Li at amd.com>; Alex Deucher <alexdeucher at gmail.com>
> Cc: amd-gfx list <amd-gfx at lists.freedesktop.org>
> Subject: Re: [PATCH 1/2] drm/amdgpu: Enable scatter gather display support
>
> Am 08.03.2018 um 09:35 schrieb Michel Dänzer:
>
> On 2018-03-07 10:47 AM, Christian König wrote:
>
> Am 07.03.2018 um 09:42 schrieb Michel Dänzer:
>
> On 2018-03-06 07:23 PM, Christian König wrote:
>
> E.g. the last time I tested it placing things into GTT still
> resulted in quite a performance penalty for rendering.
>
> FWIW, I think the penalty is most likely IOMMU related. Last time I
> tested, I couldn't measure a big difference with IOMMU disabled.
>
> No, the penalty I'm talking about came from the ping/pong we did with
> the scanout buffers.
>
> See when I tested this the DDX and Mesa where unmodified, so both
> still assumed VRAM as placement for scanout BOs, but the kernel
> forced scanout BOs into GTT for testing.
>
> So what happened was that on scanout we moved the VRAM BO to GTT
>
> and
>
> after unpinning it on the first command submission which used the BO
> we moved it back to VRAM again.
>
> In the meantime, I've had the same idea as Marek: Can't the kernel
> driver simply change the BO's preferred domain to GTT when scanning
> out from it? Then it won't move back to VRAM.
>
> Yes, I've considered this as well.
>
> But I think making the decision in Mesa is the cleaner approach.
>
> E.g. so far we only override the placement decision of userspace for two
> reasons:
> 1. We where running out of memory in VRAM.
> 2. We have a hardware restriction which makes VRAM usage mandatory.
>
> And even then we never adjust the placement permanently, we just
> temporary moved the buffer where it was needed and moved it back after
> the operation completed.
>
> Additional to that Mesa might want to set even more flags and/or changes
> it's behavior. E.g. use a tilling mode which both importer and export in an
> A+A laptop understands etc...
>
> Regards,
> Christian.
>