Re: [PATCH v4 3/4] media: venus: Update to bitrate based clock scaling

[amasule]

Hi,

On 2019-07-15 21:28, Stanimir Varbanov wrote:
> Hi,
>
> On 7/2/19 5:46 PM, Aniket Masule wrote:
> > Introduced clock scaling using bitrate, preavious
> > calculations consider only the cycles per mb.
> > Also, clock scaling is now triggered before every
> > buffer being queued to the device. This helps in
> > deciding precise clock cycles required.
> >
> > Signed-off-by: Aniket Masule <amasule@xxxxxxxxxxxxxx>
> > ---
> >  drivers/media/platform/qcom/venus/helpers.c | 31 
> > +++++++++++++++++++++++++----
> >  1 file changed, 27 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/media/platform/qcom/venus/helpers.c 
> > b/drivers/media/platform/qcom/venus/helpers.c
> > index 2c976e4..5726d86 100644
> > --- a/drivers/media/platform/qcom/venus/helpers.c
> > +++ b/drivers/media/platform/qcom/venus/helpers.c
> > @@ -399,17 +399,26 @@ static int scale_clocks(struct venus_inst *inst)
> >  	return ret;
> >  }
> >
> > -static unsigned long calculate_vpp_freq(struct venus_inst *inst)
> > +static unsigned long calculate_inst_freq(struct venus_inst *inst,
> > +					 unsigned long filled_len)
> >  {
> > -	unsigned long vpp_freq = 0;
> > +	unsigned long vpp_freq = 0, vsp_freq = 0;
> > +	u64 fps = inst->fps;
> >  	u32 mbs_per_sec;
> >
> >  	mbs_per_sec = load_per_instance(inst);
> >  	vpp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vpp_freq;
> >  	/* 21 / 20 is overhead factor */
> >  	vpp_freq += vpp_freq / 20;
> > +	vsp_freq = mbs_per_sec * inst->clk_data.codec_freq_data->vsp_freq;
> >
> > -	return vpp_freq;
> > +	/* 10 / 7 is overhead factor */
> > +	if (inst->session_type == VIDC_SESSION_TYPE_ENC)
> > +		vsp_freq += (inst->controls.enc.bitrate * 10) / 7;
> > +	else
> > +		vsp_freq += ((fps * filled_len * 8) * 10) / 7;
> > +
> > +	return max(vpp_freq, vsp_freq);
> >  }
> >
> >  static int scale_clocks_v4(struct venus_inst *inst)
> > @@ -417,13 +426,27 @@ static int scale_clocks_v4(struct venus_inst 
> > *inst)
> >  	struct venus_core *core = inst->core;
> >  	const struct freq_tbl *table = core->res->freq_tbl;
> >  	unsigned int num_rows = core->res->freq_tbl_size;
> > +	struct v4l2_m2m_ctx *m2m_ctx = inst->m2m_ctx;
> >  	struct clk *clk = core->clks[0];
> >  	struct device *dev = core->dev;
> >  	unsigned int i;
> >  	unsigned long freq = 0, freq_core1 = 0, freq_core2 = 0;
> > +	unsigned long filled_len = 0;
> > +	struct venus_buffer *buf, *n;
> > +	struct vb2_buffer *vb;
> >  	int ret;
> >
> > -	freq = calculate_vpp_freq(inst);
> > +	mutex_lock(&inst->lock);
> > +	v4l2_m2m_for_each_src_buf_safe(m2m_ctx, buf, n) {
> > +		vb = &buf->vb.vb2_buf;
> > +		filled_len = max(filled_len, vb2_get_plane_payload(vb, 0));
> > +	}
> > +	mutex_unlock(&inst->lock);
> > +
> > +	if (inst->session_type == VIDC_SESSION_TYPE_DEC && !filled_len)
> > +		return 0;
> > +
> > +	freq = calculate_inst_freq(inst, filled_len);
> >
> >  	if (freq > table[0].freq)
> >  		dev_warn(dev, "HW is overloaded, needed: %lu max: %lu\n",
>
> The original patch has a call to load_scale_clocks from
> venus_helper_vb2_buf_queue, why it is not included here?

Yes, I need to include it for before every buffer being queued to the 
device
to have accurate precise clock.

Regards,
Aniket