Re: [PATCH v3 1/8] media: hantro: Trace hevc hw cycles performance register

[Benjamin Gaignard]

Le 18/06/2021 à 20:58, Ezequiel Garcia a écrit :
> Hi Benjamin,
>
> On Fri, 2021-06-18 at 15:15 +0200, Benjamin Gaignard wrote:
> > After each hevc decoded frame trace the hardware performance.
> > It provides the number of hw cycles spend per decoded macroblock.
> Please add some documentation about how these are supposed
> to be used. It will be easier to discuss after seeing
> things in actiion.
>
> A good place for the documentation would be:
>
> https://www.kernel.org/doc/html/latest/admin-guide/media/v4l-drivers.html

Ok that will be in v4

> [..]
> > @@ -22,6 +23,21 @@ static inline void hantro_write_addr(struct hantro_dev *vpu,
> >          vdpu_write(vpu, addr & 0xffffffff, offset);
> >   }
> >   
> > +void hantro_g2_hevc_dec_done(struct hantro_ctx *ctx)
> I'm worried about the runtime cost this would have.
>
> I see other drivers (i915, panfrost) seem to have an ioctl
> to enable the perf counters.
>
> Perhaps we don't need an ioctl, but a module param would be enough
> for now.

You can be reassured the overhead of traces points is very low has explained
in this article: https://lwn.net/Articles/330402/

> > +{
> > +       const struct hantro_hevc_dec_ctrls *ctrls = &ctx->hevc_dec.ctrls;
> > +       const struct v4l2_ctrl_hevc_sps *sps = ctrls->sps;
> > +       struct hantro_dev *vpu = ctx->dev;
> > +       u32 hw_cycles = 0;
> > +       u32 mbs = (sps->pic_width_in_luma_samples *
> > +                  sps->pic_height_in_luma_samples) >> 8;
> > +
> > +       if (mbs)
> > +               hw_cycles = vdpu_read(vpu, G2_HW_PERFORMANCE) / mbs;
> > +
> > +       trace_hantro_hevc_perf(ctx, hw_cycles);
> > +}
> > +
> [..]
> > +
> > +TRACE_EVENT(hantro_hevc_perf,
> > +       TP_PROTO(struct hantro_ctx *ctx, u32 hw_cycles),
> > +
> > +       TP_ARGS(ctx, hw_cycles),
> > +
> > +       TP_STRUCT__entry(
> > +               __field(int, minor)
> > +               __field(u32, hw_cycles)
> > +       ),
> > +
> > +       TP_fast_assign(
> > +               __entry->minor = ctx->fh.vdev->minor;
> Tracking performance per minor doesn't seem useful,
> we'd like to track per-fd (i.e. per context).

This part of the driver doesn't know for which fd the decoding job is done
so impossible to add it there.

Regards,
Benjamin

> Thanks,
> Ezequiel