On Mon, 2015-11-30 at 15:36 +0000, Daniel Thompson wrote:
> On 30 November 2015 at 11:43, andrew-ct chen
> <andrew-ct.chen@xxxxxxxxxxxx> wrote:
> > On Fri, 2015-11-27 at 12:21 +0000, Daniel Thompson wrote:
> >> On 27/11/15 12:10, andrew-ct chen wrote:
> >> >>> +
> >> >>> > >+ memcpy((void *)send_obj->share_buf, buf, len);
> >> >>> > >+ send_obj->len = len;
> >> >>> > >+ send_obj->id = id;
> >> >>> > >+ vpu_cfg_writel(vpu, 0x1, HOST_TO_VPU);
> >> >>> > >+
> >> >>> > >+ /* Wait until VPU receives the command */
> >> >>> > >+ timeout = jiffies + msecs_to_jiffies(IPI_TIMEOUT_MS);
> >> >>> > >+ do {
> >> >>> > >+ if (time_after(jiffies, timeout)) {
> >> >>> > >+ dev_err(vpu->dev, "vpu_ipi_send: IPI timeout!\n");
> >> >>> > >+ return -EIO;
> >> >>> > >+ }
> >> >>> > >+ } while (vpu_cfg_readl(vpu, HOST_TO_VPU));
> >> >> >
> >> >> >Do we need to busy wait every time we communicate with the co-processor?
> >> >> >Couldn't we put this wait*before* we write to HOST_TO_VPU above.
> >> >> >
> >> >> >That way we only spin when there is a need to.
> >> >> >
> >> > Since the hardware VPU only allows that one client sends the command to
> >> > it each time.
> >> > We need the wait to make sure VPU accepted the command and cleared the
> >> > interrupt and then the next command would be served.
> >>
> >> I understand that the VPU can only have on message outstanding at once.
> >>
> >> I just wonder why we busy wait *after* sending the first command rather
> >> than *before* sending the second one.
> >
> > No other special reasons. Just send one command and wait until VPU gets
> > the command. Then, I think this wait also can be put before we write to
> > HOST_TO_VPU.Is this better than former? May I know the reason?
>
> Busy waiting is bad; it is a waste of host CPU processor time and/or power.
>
> When the busy wait occurs after queuing then we will busy wait for
> every command we send.
>
> If busy wait occurs before next queuing then we will wait for a
> shorter time in total because we have the chance to do something
> useful on the host before we have to wait.
>
Got it. Thanks a lot for the explanation.
I'll put the busy wait before next queuing in next version.
>
> >> Streamed decode/encode typically ends up being rate controlled by
> >> capture or display meaning that in these cases we don't need to busy
> >> wait at all (because by the time we send the next frame the VPU has
> >> already accepted the previous message).
> >
> > For now, only one device "encoder" exists, it is true.
> > But, we'll have encoder and decoder devices, the decode and encode
> > requested to VPU are simultaneous.
>
> Sure, I accept that lock and busy-wait is an appropriate way to ensure
> safety (assuming the VPU is fairly quick clearing the HOST_TO_VPU
> flag).
>
The busy wait time is less than 1ms in average.
>
> > Is this supposed to be removed for this patches and we can add it back
> > if the another device(decoder) is ready for review?
>
> No I'm not suggesting that.
>
> I only recommend moving the busy wait *before* end sending of command
> (for efficiency reasons mentioned above).
Ok. Thanks.
>
>
> Daniel.
--
To unsubscribe from this list: send the line "unsubscribe devicetree" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
