[BUG] Data race when use PACKET3_DMA_DATA?

[Chen Lei]

Thanks very much. I get it now.


> -----Original Messages-----
> From: "Christian König" <ckoenig.leichtzumerken at gmail.com>
> Sent Time: 2021-06-04 15:40:08 (Friday)
> To: "Chen Lei" <chenlei18s at ict.ac.cn>, "Alex Deucher" <alexdeucher at gmail.com>
> Cc: "amd-gfx list" <amd-gfx at lists.freedesktop.org>
> Subject: Re: [BUG] Data race when use PACKET3_DMA_DATA?
> 
> Hi,
> 
> I think your problem comes from the missing understanding that the 
> hardware is heavily pipelined.
> 
> In other words commands you send to the hardware just kick of 
> asynchronously processing, e.g. a CP DMA command just kicks a copy 
> operation but the CP then continue executing commands.
> 
> Same thing for a RELEASE_MEM packet, it just kicks of an operation to 
> write a value to an address when all Compute or 3D rendering is completed.
> 
> But if you want to synchronize execution of the CP commands you still 
> need to block for that valuie to be written or otherwise the CP will 
> just keep going with the next command.
> 
> Regards,
> Christian.
> 
> Am 03.06.21 um 05:37 schrieb Chen Lei:
> > I had seperated the dispatch packet and DMA DATA packet into 2 IBs, and called the amdgpu_cs_ioctl twice.
> > If I was not mistaken, the `amdgpu_ib_schedule` would emit fence packets for each amdgpu_cs_ioctl call.
> > Did I need to insert the fence packet explicitly after the dispatch packet?
> >
> >
> > > -----Original Messages-----
> > > From: "Alex Deucher" <alexdeucher at gmail.com>
> > > Sent Time: 2021-06-03 10:11:46 (Thursday)
> > > To: "Chen Lei" <chenlei18s at ict.ac.cn>
> > > Cc: "amd-gfx list" <amd-gfx at lists.freedesktop.org>
> > > Subject: Re: Re: [BUG] Data race when use PACKET3_DMA_DATA?
> > >
> > > On Wed, Jun 2, 2021 at 8:29 PM Chen Lei <chenlei18s at ict.ac.cn> wrote:
> > > >
> > > > Hi Alex. Thanks for your quick reply.
> > > > I first submit the OpenCL kernel packet and then submit the DMA DATA packet. And the OpenCL kernel reads the value written by the DMA DATA packet.
> > > > If I understand you correctly, that is because the CP engine continues on to process the DMA DATA packet after launching the OpenCL kernel. If so, is there any way to sync the CP engine until the OpenCL kernel is complete?
> > > >
> > >
> > > Once the kernel has been dispatched to the shader cores, the CP will
> > > continue to execute packets in the queue.  If you want it to wait for
> > > the pipeline to drain you'll need to insert a fence packet (e.g.,
> > > RELEASE_MEM).
> > >
> > > Alex
> > >
> > > >
> > > > > -----Original Messages-----
> > > > > From: "Alex Deucher" <alexdeucher at gmail.com>
> > > > > Sent Time: 2021-06-02 21:37:51 (Wednesday)
> > > > > To: "Chen Lei" <chenlei18s at ict.ac.cn>
> > > > > Cc: "amd-gfx list" <amd-gfx at lists.freedesktop.org>
> > > > > Subject: Re: [BUG] Data race when use PACKET3_DMA_DATA?
> > > > >
> > > > > On Wed, Jun 2, 2021 at 8:44 AM Chen Lei <chenlei18s at ict.ac.cn> wrote:
> > > > > >
> > > > > > Hi, I noticed that there are two ways to do DMA for amd gpu: the SDMA copy packet and the PM4 dma packet.
> > > > > >
> > > > > > I had tested the PM4 dma packet:  PACKET3_DMA_DATA. In most of time, it works.
> > > > > >
> > > > > > But when I launch an OpenCL kernel followed by a host-to-gpu DMA packet, it seems that the OpenCL kernel read the new value written by the following DMA packet.
> > > > > >
> > > > > > Both the OpenCL kernel and the PM4 dma packet are submitted using the amdgpu_cs_ioctl, and they are submitted to the same ring.
> > > > > >
> > > > > > I was not family with the hardware details. According to my understanding, because the ring is FIFO, there is no need for any explicit synchronization between the OpenCL kernel launch packet and the dma packet. So the result looked weird. And when I add the synchronization(i.e. amdgpu_cs_wait_ioctl) before the dma packet, everything is OK.
> > > > > >
> > > > > > Was it a hardware bug or did I miss something?
> > > > > >
> > > > >
> > > > > The CP DMA engine is separate from the various CP micro engines.  When
> > > > > there is a DMA DATA packet, the DMA operation is offloaded to the CP
> > > > > DMA engine and the CP engine that processed the packet continues on to
> > > > > the next packet.  You need to use the ENGINE_SEL and CP_SYNC bits in
> > > > > the DMA DATA packet to specify the behavior you want.  The ENGINE_SEL
> > > > > bit selects which CP engine processes the packet (PFP or ME) and the
> > > > > CP_SYNC bit stops further packet processing on the selected engine
> > > > > until the DMA is complete.
> > > > >
> > > > > Alex
> > > > </chenlei18s at ict.ac.cn></amd-gfx at lists.freedesktop.org></chenlei18s at ict.ac.cn></alexdeucher at gmail.com>
> > > > _______________________________________________
> > > > amd-gfx mailing list
> > > > amd-gfx at lists.freedesktop.org
> > > > https://lists.freedesktop.org/mailman/listinfo/amd-gfx
> > </chenlei18s at ict.ac.cn></amd-gfx at lists.freedesktop.org></chenlei18s at ict.ac.cn></alexdeucher at gmail.com>
> > _______________________________________________
> > amd-gfx mailing list
> > amd-gfx at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/amd-gfx
</amd-gfx at lists.freedesktop.org></alexdeucher at gmail.com></chenlei18s at ict.ac.cn></ckoenig.leichtzumerken at gmail.com>