[Intel-gfx] [PATCH v3 4/5] drm/i915: Adjust CDCLK accordingly to our DBuf bw needs
Lisovskiy, Stanislav
stanislav.lisovskiy at intel.com
Mon Mar 30 18:16:49 UTC 2020
On Mon, Mar 30, 2020 at 08:07:31PM +0300, Ville Syrjälä wrote:
> On Mon, Mar 30, 2020 at 03:23:53PM +0300, Stanislav Lisovskiy wrote:
> > According to BSpec max BW per slice is calculated using formula
> > Max BW = CDCLK * 64. Currently when calculating min CDCLK we
> > account only per plane requirements, however in order to avoid
> > FIFO underruns we need to estimate accumulated BW consumed by
> > all planes(ddb entries basically) residing on that particular
> > DBuf slice. This will allow us to put CDCLK lower and save power
> > when we don't need that much bandwidth or gain additional
> > performance once plane consumption grows.
> >
> > v2: - Fix long line warning
> > - Limited new DBuf bw checks to only gens >= 11
> >
> > v3: - Lets track used Dbuf bw per slice and per crtc in bw state
> > (or may be in DBuf state in future), that way we don't need
> > to have all crtcs in state and those only if we detect if
> > are actually going to change cdclk, just same way as we
> > do with other stuff, i.e intel_atomic_serialize_global_state
> > and co. Just as per Ville's paradigm.
> > - Made dbuf bw calculation procedure look nicer by introducing
> > for_each_dbuf_slice_in_mask - we often will now need to iterate
> > slices using mask.
> > - According to experimental results CDCLK * 64 accounts for
> > overall bandwidth across all dbufs, not per dbuf.
> >
> > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy at intel.com>
> > ---
> > drivers/gpu/drm/i915/display/intel_bw.c | 61 ++++++++++++++++++-
> > drivers/gpu/drm/i915/display/intel_bw.h | 8 +++
> > drivers/gpu/drm/i915/display/intel_cdclk.c | 25 ++++++++
> > drivers/gpu/drm/i915/display/intel_display.c | 8 +++
> > .../drm/i915/display/intel_display_power.h | 2 +
> > drivers/gpu/drm/i915/intel_pm.c | 34 ++++++++++-
> > drivers/gpu/drm/i915/intel_pm.h | 3 +
> > 7 files changed, 138 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/display/intel_bw.c b/drivers/gpu/drm/i915/display/intel_bw.c
> > index 573a1c206b60..e9d65820fb76 100644
> > --- a/drivers/gpu/drm/i915/display/intel_bw.c
> > +++ b/drivers/gpu/drm/i915/display/intel_bw.c
> > @@ -6,6 +6,7 @@
> > #include <drm/drm_atomic_state_helper.h>
> >
> > #include "intel_bw.h"
> > +#include "intel_pm.h"
> > #include "intel_display_types.h"
> > #include "intel_sideband.h"
> > #include "intel_atomic.h"
> > @@ -338,7 +339,6 @@ static unsigned int intel_bw_crtc_data_rate(const struct intel_crtc_state *crtc_
> >
> > return data_rate;
> > }
> > -
> > void intel_bw_crtc_update(struct intel_bw_state *bw_state,
> > const struct intel_crtc_state *crtc_state)
> > {
> > @@ -419,6 +419,65 @@ intel_atomic_bw_get_state(struct intel_atomic_state *state)
> > return to_intel_bw_state(bw_state);
> > }
> >
> > +int intel_bw_calc_min_cdclk(struct intel_atomic_state *state)
> > +{
> > + struct drm_i915_private *dev_priv = to_i915(state->base.dev);
> > + int i = 0;
> > + enum plane_id plane_id;
> > + struct intel_crtc_state *crtc_state;
>
> const
Why? There are lots of places where we use it that way.
Even in intel_compute_min_cdclk:
struct intel_crtc_state *crtc_state;
int min_cdclk, i;
enum pipe pipe;
for_each_new_intel_crtc_in_state(state, crtc, crtc_state, i) {
Also I grepped - almost every place in intel_display.c
which has for_each_new_intel_crtc_in_state uses crtc_state as non-const.
It is called that way intel_calc_active_pipes, intel_atomic_check_crtcs,...
almost everywhere in intel_display.c
So can you please explain what do you mean?..
>
> > + struct intel_crtc *crtc;
> > + int max_bw = 0;
> > + int min_cdclk;
> > + enum pipe pipe;
> > + struct intel_bw_state *bw_state;
> > + int slice_id = 0;
>
> Bunch of needless intiialization, needlessly wide scope, etc.
max_bw needs to be accumulated so I guess it should be init to 0.
Agree regarding slice_id - just a leftover after I started to
use for_each_dbuf_slice macro.
>
> > +
> > + bw_state = intel_atomic_bw_get_state(state);
> > +
>
> Spurious whitespace.
>
> > + if (IS_ERR(bw_state))
> > + return PTR_ERR(bw_state);
> > +
> > + for_each_new_intel_crtc_in_state(state, crtc, crtc_state, i) {
> > + struct intel_crtc_bw *crtc_bw = &bw_state->dbuf_bw_used[crtc->pipe];
> > +
> > + memset(&crtc_bw->dbuf_bw, 0, sizeof(crtc_bw->dbuf_bw));
> > +
> > + for_each_plane_id_on_crtc(crtc, plane_id) {
> > + struct skl_ddb_entry *plane_alloc =
> > + &crtc_state->wm.skl.plane_ddb_y[plane_id];
> > + struct skl_ddb_entry *uv_plane_alloc =
> > + &crtc_state->wm.skl.plane_ddb_uv[plane_id];
>
> const
>
> > + unsigned int data_rate = crtc_state->data_rate[plane_id];
> > +
>
> more strange whitespace
>
> > + unsigned int dbuf_mask = skl_ddb_dbuf_slice_mask(dev_priv, plane_alloc);
> > +
> > + dbuf_mask |= skl_ddb_dbuf_slice_mask(dev_priv, uv_plane_alloc);
>
> Looks bad when you initialize part of it in declaration and the rest
> later.
Didn't want to have a long time here, will have to make a separate function then.
>
> > +
> > + DRM_DEBUG_KMS("Got dbuf mask %x for pipe %c ddb %d-%d plane %d data rate %d\n",
> > + dbuf_mask, pipe_name(crtc->pipe), plane_alloc->start,
> > + plane_alloc->end, plane_id, data_rate);
> > +
> > + for_each_dbuf_slice_in_mask(slice_id, dbuf_mask)
> > + crtc_bw->dbuf_bw[slice_id] += data_rate;
>
> This doesn't feel quite right for planar formats.
>
> For pre-icl it works by accident since we only have the one slice so
> we don't end up accounting for the full bandwidth from both color planes
> to multiple slices. If we had multiple slices and chroma and luma had
> been allocated to different slices we'd count the same thing multiple
> times.
>
> For icl+ we seem to assign the full data rate to the UV plane's slice(s)
> since only the UV plane has data_rate[] != 0.
This sounds interesting, so do you mean that we need something like data_rate
and data_rate_uv?
My logic here is that we just get a set of slices used by uv and y planes
in total and then account for data_rate for those.
We probably then need another patch to split data rate then.
I was aware of that actually but anyway being pessimistic here is better.
Also same about if plane would be sitting on 2 slices at the same time.
>
> > + }
> > + }
> > +
> > + for_each_dbuf_slice(slice_id) {
> > + int total_bw_per_slice = 0;
> > +
> > + for_each_pipe(dev_priv, pipe) {
> > + struct intel_crtc_bw *crtc_bw = &bw_state->dbuf_bw_used[pipe];
> > +
> > + total_bw_per_slice += crtc_bw->dbuf_bw[slice_id];
> > + }
> > + max_bw += total_bw_per_slice;
>
> So we're aggregating all the bw instead of per-slice? Is this based on
> the other mail you sent? Deserves a comment explaining why we do such
> odd things.
Agree need to add a comment. Currently it seems the only way to make it
work without either having underruns or 8K not usable with bumped CDCLK.
>
> > + }
> > +
> > + min_cdclk = max_bw / 64;
> > +
> > + return min_cdclk;
> > +}
> > +
> > int intel_bw_atomic_check(struct intel_atomic_state *state)
> > {
> > struct drm_i915_private *dev_priv = to_i915(state->base.dev);
> > diff --git a/drivers/gpu/drm/i915/display/intel_bw.h b/drivers/gpu/drm/i915/display/intel_bw.h
> > index 9a5627be6876..d2b5f32b0791 100644
> > --- a/drivers/gpu/drm/i915/display/intel_bw.h
> > +++ b/drivers/gpu/drm/i915/display/intel_bw.h
> > @@ -10,11 +10,16 @@
> >
> > #include "intel_display.h"
> > #include "intel_global_state.h"
> > +#include "intel_display_power.h"
> >
> > struct drm_i915_private;
> > struct intel_atomic_state;
> > struct intel_crtc_state;
> >
> > +struct intel_crtc_bw {
> > + int dbuf_bw[I915_MAX_DBUF_SLICES];
> > +};
> > +
> > struct intel_bw_state {
> > struct intel_global_state base;
> >
> > @@ -31,6 +36,8 @@ struct intel_bw_state {
> > */
> > u8 qgv_points_mask;
> >
> > + struct intel_crtc_bw dbuf_bw_used[I915_MAX_PIPES];
>
> The name of the struct isn't very good if it just contains the
> dbuf bw numbers.
>
> > +
> > unsigned int data_rate[I915_MAX_PIPES];
> > u8 num_active_planes[I915_MAX_PIPES];
>
> Maybe collect all this to the per-crtc struct?
>
> > };
> > @@ -53,5 +60,6 @@ void intel_bw_crtc_update(struct intel_bw_state *bw_state,
> > const struct intel_crtc_state *crtc_state);
> > int icl_pcode_restrict_qgv_points(struct drm_i915_private *dev_priv,
> > u32 points_mask);
> > +int intel_bw_calc_min_cdclk(struct intel_atomic_state *state);
> >
> > #endif /* __INTEL_BW_H__ */
> > diff --git a/drivers/gpu/drm/i915/display/intel_cdclk.c b/drivers/gpu/drm/i915/display/intel_cdclk.c
> > index 979a0241fdcb..036774e7f3ec 100644
> > --- a/drivers/gpu/drm/i915/display/intel_cdclk.c
> > +++ b/drivers/gpu/drm/i915/display/intel_cdclk.c
> > @@ -25,6 +25,7 @@
> > #include "intel_cdclk.h"
> > #include "intel_display_types.h"
> > #include "intel_sideband.h"
> > +#include "intel_bw.h"
> >
> > /**
> > * DOC: CDCLK / RAWCLK
> > @@ -2001,11 +2002,19 @@ int intel_crtc_compute_min_cdclk(const struct intel_crtc_state *crtc_state)
> > {
> > struct drm_i915_private *dev_priv =
> > to_i915(crtc_state->uapi.crtc->dev);
> > + struct intel_atomic_state *state = NULL;
> > int min_cdclk;
> >
> > if (!crtc_state->hw.enable)
> > return 0;
> >
> > + /*
> > + * FIXME: Unfortunately when this gets called from intel_modeset_setup_hw_state
> > + * there is no intel_atomic_state at all. So lets not then use it.
> > + */
> > + if (crtc_state->uapi.state)
> > + state = to_intel_atomic_state(crtc_state->uapi.state);
>
> This still indicates that either this isn't the right place to call this
> or we have the state stored in the wrong place.
>
> I think I'd just move the thing into intel_compute_min_cdclk() as a start.
Was thinking about this, why I didn't put it here is because just wanted
to keep all cdclk regulating logic in this function, because intel_compute_min_cdclk
is kind of higher level already which is supposed to call that one per crtc_state.
Otherwise CDCLK regulating logic will be kind of spread across those two.
>
> > +
> > min_cdclk = intel_pixel_rate_to_cdclk(crtc_state);
> >
> > /* pixel rate mustn't exceed 95% of cdclk with IPS on BDW */
> > @@ -2080,6 +2089,22 @@ int intel_crtc_compute_min_cdclk(const struct intel_crtc_state *crtc_state)
> > if (IS_TIGERLAKE(dev_priv))
> > min_cdclk = max(min_cdclk, (int)crtc_state->pixel_rate);
> >
> > + /*
> > + * Similar story as with skl_write_plane_wm and intel_enable_sagv
> > + * - in some certain driver parts, we don't have any guarantee that
> > + * parent exists. So we might be having a crtc_state without
> > + * parent state.
> > + */
> > + if (INTEL_GEN(dev_priv) >= 11) {
> > + if (state) {
> > + int dbuf_bw_cdclk = intel_bw_calc_min_cdclk(state);
> > +
> > + DRM_DEBUG_KMS("DBuf bw min cdclk %d current min_cdclk %d\n",
> > + dbuf_bw_cdclk, min_cdclk);
> > + min_cdclk = max(min_cdclk, dbuf_bw_cdclk);
> > + }
> > + }
> > +
> > if (min_cdclk > dev_priv->max_cdclk_freq) {
> > drm_dbg_kms(&dev_priv->drm,
> > "required cdclk (%d kHz) exceeds max (%d kHz)\n",
> > diff --git a/drivers/gpu/drm/i915/display/intel_display.c b/drivers/gpu/drm/i915/display/intel_display.c
> > index 9fd32d61ebfe..fa2870c0d7fd 100644
> > --- a/drivers/gpu/drm/i915/display/intel_display.c
> > +++ b/drivers/gpu/drm/i915/display/intel_display.c
> > @@ -14678,6 +14678,14 @@ static int intel_atomic_check_planes(struct intel_atomic_state *state,
> > if (old_active_planes == new_active_planes)
> > continue;
> >
> > + /*
> > + * active_planes bitmask has been updated, whenever amount
> > + * of active planes had changed we need to recalculate CDCLK
> > + * as it depends on total bandwidth now, not only min_cdclk
> > + * per plane.
> > + */
> > + *need_cdclk_calc = true;
> > +
> > ret = intel_crtc_add_planes_to_state(state, crtc, new_active_planes);
> > if (ret)
> > return ret;
> > diff --git a/drivers/gpu/drm/i915/display/intel_display_power.h b/drivers/gpu/drm/i915/display/intel_display_power.h
> > index 468e8fb0203a..9e33fb90422f 100644
> > --- a/drivers/gpu/drm/i915/display/intel_display_power.h
> > +++ b/drivers/gpu/drm/i915/display/intel_display_power.h
> > @@ -308,6 +308,8 @@ intel_display_power_put_async(struct drm_i915_private *i915,
> > }
> > #endif
> >
> > +#define I915_MAX_DBUF_SLICES 2
> > +
> > enum dbuf_slice {
> > DBUF_S1,
> > DBUF_S2,
> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > index 551933e3f7da..5dcd1cd09ad7 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > @@ -4055,10 +4055,9 @@ icl_get_first_dbuf_slice_offset(u32 dbuf_slice_mask,
> > return offset;
> > }
> >
> > -static u16 intel_get_ddb_size(struct drm_i915_private *dev_priv)
> > +u16 intel_get_ddb_size(struct drm_i915_private *dev_priv)
> > {
> > u16 ddb_size = INTEL_INFO(dev_priv)->ddb_size;
> > -
> > drm_WARN_ON(&dev_priv->drm, ddb_size == 0);
> >
> > if (INTEL_GEN(dev_priv) < 11)
> > @@ -4067,6 +4066,37 @@ static u16 intel_get_ddb_size(struct drm_i915_private *dev_priv)
> > return ddb_size;
> > }
> >
> > +u32 skl_ddb_dbuf_slice_mask(struct drm_i915_private *dev_priv,
> > + const struct skl_ddb_entry *entry)
> > +{
> > + u32 slice_mask = 0;
> > + u16 ddb_size = intel_get_ddb_size(dev_priv);
> > + u16 num_supported_slices = INTEL_INFO(dev_priv)->num_supported_dbuf_slices;
> > + u16 slice_size = ddb_size / num_supported_slices;
> > + u16 start_slice;
> > + u16 end_slice;
> > +
> > + if (!skl_ddb_entry_size(entry))
> > + return 0;
> > +
> > + start_slice = entry->start / slice_size;
> > + end_slice = (entry->end - 1) / slice_size;
> > +
> > + DRM_DEBUG_KMS("ddb size %d slices %d slice size %d start slice %d end slice %d\n",
> > + ddb_size, num_supported_slices, slice_size, start_slice, end_slice);
> > +
> > + /*
> > + * Per plane DDB entry can in a really worst case be on multiple slices
> > + * but single entry is anyway contigious.
> > + */
> > + while (start_slice <= end_slice) {
> > + slice_mask |= 1 << start_slice;
> > + start_slice++;
> > + }
>
> GENMASK() or somehting?
Will fix, thanks for proposal.
>
> > +
> > + return slice_mask;
> > +}
> > +
> > static u8 skl_compute_dbuf_slices(const struct intel_crtc_state *crtc_state,
> > u8 active_pipes);
> >
> > diff --git a/drivers/gpu/drm/i915/intel_pm.h b/drivers/gpu/drm/i915/intel_pm.h
> > index 069515f04170..41c61ad71ce6 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.h
> > +++ b/drivers/gpu/drm/i915/intel_pm.h
> > @@ -37,6 +37,9 @@ void skl_pipe_ddb_get_hw_state(struct intel_crtc *crtc,
> > struct skl_ddb_entry *ddb_y,
> > struct skl_ddb_entry *ddb_uv);
> > void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv);
> > +u16 intel_get_ddb_size(struct drm_i915_private *dev_priv);
> > +u32 skl_ddb_dbuf_slice_mask(struct drm_i915_private *dev_priv,
> > + const struct skl_ddb_entry *entry);
> > void skl_pipe_wm_get_hw_state(struct intel_crtc *crtc,
> > struct skl_pipe_wm *out);
> > void g4x_wm_sanitize(struct drm_i915_private *dev_priv);
> > --
> > 2.24.1.485.gad05a3d8e5
>
> --
> Ville Syrjälä
> Intel
More information about the Intel-gfx
mailing list