Re: [RFC][PATCH 5/5] drm/i915/display: Add Nearest-neighbor based integer scaling support
From: Ville Syrjälä
Date: Thu Mar 12 2020 - 09:54:47 EST
On Thu, Mar 12, 2020 at 09:13:24AM +0000, Laxminarayan Bharadiya, Pankaj wrote:
>
>
> > -----Original Message-----
> > From: Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx>
> > Sent: 10 March 2020 21:47
> > To: Laxminarayan Bharadiya, Pankaj
> > <pankaj.laxminarayan.bharadiya@xxxxxxxxx>
> > Cc: jani.nikula@xxxxxxxxxxxxxxx; daniel@xxxxxxxx; intel-
> > gfx@xxxxxxxxxxxxxxxxxxxxx; dri-devel@xxxxxxxxxxxxxxxxxxxxx; airlied@xxxxxxxx;
> > maarten.lankhorst@xxxxxxxxxxxxxxx; tzimmermann@xxxxxxx;
> > mripard@xxxxxxxxxx; mihail.atanassov@xxxxxxx; Joonas Lahtinen
> > <joonas.lahtinen@xxxxxxxxxxxxxxx>; Vivi, Rodrigo <rodrigo.vivi@xxxxxxxxx>;
> > Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>; Souza, Jose
> > <jose.souza@xxxxxxxxx>; De Marchi, Lucas <lucas.demarchi@xxxxxxxxx>;
> > Roper, Matthew D <matthew.d.roper@xxxxxxxxx>; Deak, Imre
> > <imre.deak@xxxxxxxxx>; Shankar, Uma <uma.shankar@xxxxxxxxx>; linux-
> > kernel@xxxxxxxxxxxxxxx; Nautiyal, Ankit K <ankit.k.nautiyal@xxxxxxxxx>
> > Subject: Re: [RFC][PATCH 5/5] drm/i915/display: Add Nearest-neighbor
> > based integer scaling support
> >
> > On Tue, Feb 25, 2020 at 12:35:45PM +0530, Pankaj Bharadiya wrote:
> > > Integer scaling (IS) is a nearest-neighbor upscaling technique that
> > > simply scales up the existing pixels by an integer (i.e., whole
> > > number) multiplier.Nearest-neighbor (NN) interpolation works by
> > > filling in the missing color values in the upscaled image with that of
> > > the coordinate-mapped nearest source pixel value.
> > >
> > > Both IS and NN preserve the clarity of the original image. Integer
> > > scaling is particularly useful for pixel art games that rely on sharp,
> > > blocky images to deliver their distinctive look.
> > >
> > > Program the scaler filter coefficients to enable the NN filter if
> > > scaling filter property is set to DRM_SCALING_FILTER_NEAREST_NEIGHBOR
> > > and enable integer scaling.
> > >
> > > Bspec: 49247
> > >
> > > Signed-off-by: Pankaj Bharadiya
> > > <pankaj.laxminarayan.bharadiya@xxxxxxxxx>
> > > Signed-off-by: Ankit Nautiyal <ankit.k.nautiyal@xxxxxxxxx>
> > > ---
> > > drivers/gpu/drm/i915/display/intel_display.c | 83
> > > +++++++++++++++++++- drivers/gpu/drm/i915/display/intel_display.h |
> > > 2 + drivers/gpu/drm/i915/display/intel_sprite.c | 20 +++--
> > > 3 files changed, 97 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/display/intel_display.c
> > > b/drivers/gpu/drm/i915/display/intel_display.c
> > > index b5903ef3c5a0..6d5f59203258 100644
> > > --- a/drivers/gpu/drm/i915/display/intel_display.c
> > > +++ b/drivers/gpu/drm/i915/display/intel_display.c
> > > @@ -6237,6 +6237,73 @@ void skl_scaler_disable(const struct
> > intel_crtc_state *old_crtc_state)
> > > skl_detach_scaler(crtc, i);
> > > }
> > >
> > > +/**
> > > + * Theory behind setting nearest-neighbor integer scaling:
> > > + *
> > > + * 17 phase of 7 taps requires 119 coefficients in 60 dwords per set.
> > > + * The letter represents the filter tap (D is the center tap) and
> > > +the number
> > > + * represents the coefficient set for a phase (0-16).
> > > + *
> > > + * +------------+------------------------+------------------------+
> > > + * |Index value | Data value coeffient 1 | Data value coeffient 2 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 00h | B0 | A0 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 01h | D0 | C0 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 02h | F0 | E0 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 03h | A1 | G0 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 04h | C1 | B1 |
> > > + * +------------+------------------------+------------------------+
> > > + * | ... | ... | ... |
> > > + * +------------+------------------------+------------------------+
> > > + * | 38h | B16 | A16 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 39h | D16 | C16 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 3Ah | F16 | C16 |
> > > + * +------------+------------------------+------------------------+
> > > + * | 3Bh | Reserved | G16 |
> > > + * +------------+------------------------+------------------------+
> > > + *
> > > + * To enable nearest-neighbor scaling: program scaler coefficents
> > > +with
> > > + * the center tap (Dxx) values set to 1 and all other values set to
> > > +0 as per
> > > + * SCALER_COEFFICIENT_FORMAT
> > > + *
> > > + */
> > > +void skl_setup_nearest_neighbor_filter(struct drm_i915_private
> > *dev_priv,
> > > + enum pipe pipe, int scaler_id)
> >
> > skl_scaler_...
> >
> > > +{
> > > +
> > > + int coeff = 0;
> > > + int phase = 0;
> > > + int tap;
> > > + int val = 0;
> >
> > Needlessly wide scope for most of these.
> >
> > > +
> > > + /*enable the index auto increment.*/
> > > + intel_de_write_fw(dev_priv, SKL_PS_COEF_INDEX_SET0(pipe,
> > scaler_id),
> > > + _PS_COEE_INDEX_AUTO_INC);
> > > +
> > > + for (phase = 0; phase < 17; phase++) {
> > > + for (tap = 0; tap < 7; tap++) {
> > > + coeff++;
> >
> > Can be part of the % check.
>
> OK.
>
> >
> > > + if (tap == 3)
> > > + val = (phase % 2) ? (0x800) : (0x800 << 16);
> >
> > Parens overload.
>
> OK. Will remove.
> >
> > > +
> > > + if (coeff % 2 == 0) {
> > > + intel_de_write_fw(dev_priv,
> > SKL_PS_COEF_DATA_SET0(pipe, scaler_id), val);
> > > + val = 0;
> >
> > Can drop this val=0 if you move the variable into tight scope and initialize
> > there.
>
> Moving val=0 initialization to the tight scope will not work here as we need
> to retain "val" and write only when 2 coefficients are ready (since 2
> coefficients are packed in 1 dword).
>
> e.g. for (12th , 11th) coefficients, coefficient reg value should be ( (0 << 16) | 0x800).
> If we initialize val = 0 in tight loop, 0 will be written to coefficient register.
Hmm, right. I guess I'd try to rearrange this to iterate the
registers directly instead of the phases and taps. Something
like this perhaps:
static int cnl_coef_tap(int i)
{
return i % 7;
}
static u16 cnl_coef(int t)
{
return t == 3 ? 0x0800 : 0x3000;
}
static void cnl_program_nearest_filter_coefs(void)
{
int i;
for (i = 0; i < 17 * 7; i += 2) {
uint32_t tmp;
int t;
t = cnl_coef_tap(i);
tmp = cnl_nearest_filter_coef(t);
t = cnl_coef_tap(i + 1);
tmp |= cnl_nearest_filter_coef(t) << 16;
intel_de_write_fw(tmp);
}
}
More readable I think. The downside being all those modulo operations
but hopefully that's all in the noise when it comes to performance.
--
Ville Syrjälä
Intel