Re: [PATCH] drm/amdgpu: remove distinction between explicit and implicit sync (v2)

Marek Olšák <maraeo@xxxxxxxxx> · Fri, 12 Jun 2020 18:29:34 -0400

The usage is that UMDs will no longer have to wait for idle at the end of IBs. If you have WAIT_REG_MEM or PS/CS_PARTIAL_FLUSH at the end of IBs, you can remove that. The responsibility to sync is taken over by the kernel driver.

 This has a potential to increase performance for fullscreen applications, because the kernel will sync only when the sync is required for inter-process sharing, which is never for fullscreen apps.

Also if 2 or more windowed apps are rendering, there will be no longer any sync when switching from one process to the next in the gfx ring. The sync will only happen before the compositor starts drawing the fullscreen frame. Therefore, the windowed apps running in parallel should run faster.

If the UMD syncs at the beginning of IBs (common e.g. with DCC fast clear), there will be no improvement in performance. For any improvement to be there, UMDs shouldn't sync at the beginning of IBs either, but this may not always be possible. (a fast color clear needs a sync, while a fast Z/S clear doesn't)

Marek

On Thu, Jun 11, 2020 at 8:13 AM Chunming Zhou <zhoucm1@xxxxxxx> wrote:

    I didn't check the patch details, if it is for existing implicit
      sync of shared buffer, feel free go ahead.
    But if you add some description for its usage, that will be more
      clear to others.
    -David

    在 2020/6/11 15:19, Marek Olšák 写道:

      Hi David,

        Explicit sync has nothing to do with this. This
          is for implicit sync, which is required by DRI3. This fix
          allows removing existing inefficiencies from drivers, so it's
          a good thing.

        Marek

        On Wed., Jun. 10, 2020, 03:56
          Chunming Zhou, <zhoucm1@xxxxxxx> wrote:

            在 2020/6/10 15:41, Christian König 写道:

              That's true, but for now we are stuck with the
                implicit sync for quite a number of use cases.

                My problem is rather that we already tried this and it
                backfired immediately.

                I do remember that it was your patch who introduced the
                pipeline sync flag handling and I warned that this could
                be problematic. You then came back with a QA result
                saying that this is indeed causing a huge performance
                drop in one test case and we need to do something else.
                Together we then came up with the different handling
                between implicit and explicit sync.

            Isn't pipeline sync flag to fix some issue because of
              parralel execution between jobs in one pipeline?  I really
              don't have this memory in mind why that's realted to this,
              Or do you mean extra sync hides many other potential
              issues?
            Anyway, when I go through Vulkan WSI code, the
              synchronization isn't so smooth between OS window system.
              And when I saw Jason drives explicit sync through the
              whole Linux ecosystem like Android window system does, I
              feel that's really a good direction.
            -David

                But I can't find that stupid mail thread any more. I
                knew that it was a couple of years ago when we started
                with the explicit sync for Vulkan.

                Christian.

                Am 10.06.20 um 08:29 schrieb Zhou, David(ChunMing):

                  [AMD
                      Official Use Only - Internal Distribution Only]

                  Not sue if this is right
                    direction, I think usermode wants all
                    synchronizations to be explicit. Implicit sync often
                    confuses people who don’t know its history. I
                    remember Jason from Intel  is driving explicit
                    synchronization through the Linux ecosystem, which
                    even removes implicit sync of shared buffer.

                  -David

                    From: amd-gfx <amd-gfx-bounces@xxxxxxxxxxxxxxxxxxxxx>
                      On Behalf Of Marek Olšák

                      Sent: Tuesday, June 9, 2020 6:58 PM

                      To: amd-gfx mailing list <amd-gfx@xxxxxxxxxxxxxxxxxxxxx>

                      Subject: [PATCH] drm/amdgpu: remove
                      distinction between explicit and implicit sync
                      (v2)

                      Hi,

                      This enables a full pipeline
                        sync for implicit sync. It's Christian's patch
                        with the driver version bumped. With this, user
                        mode drivers don't have to wait for idle at the
                        end of gfx IBs.

                      Any concerns?

                      Thanks,

                      Marek

                _______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx