[Intel-gfx] [PATCH v3 1/2] drm/i915/mtl: limit second scaler vertical scaling in ver >= 14
Coelho, Luciano
luciano.coelho at intel.com
Thu Dec 1 11:31:34 UTC 2022
On Wed, 2022-11-23 at 12:21 +0200, Luciano Coelho wrote:
> On Wed, 2022-11-23 at 11:57 +0200, Ville Syrjälä wrote:
> > On Wed, Nov 23, 2022 at 09:16:42AM +0000, Coelho, Luciano wrote:
> > > On Wed, 2022-11-23 at 08:47 +0200, Ville Syrjälä wrote:
> > > > On Tue, Nov 22, 2022 at 12:23:43PM +0200, Luca Coelho wrote:
> > > > > In newer hardware versions (i.e. display version >= 14), the second
> > > > > scaler doesn't support vertical scaling.
> > > > >
> > > > > The current implementation of the scaling limits is simplified and
> > > > > only occurs when the planes are created, so we don't know which scaler
> > > > > is being used.
> > > > >
> > > > > In order to handle separate scaling limits for horizontal and vertical
> > > > > scaling, and different limits per scaler, split the checks in two
> > > > > phases. We first do a simple check during plane creation and use the
> > > > > best-case scenario (because we don't know the scaler that may be used
> > > > > at a later point) and then do a more specific check when the scalers
> > > > > are actually being set up.
> > > > >
> > > > > Signed-off-by: Luca Coelho <luciano.coelho at intel.com>
> > > > > ---
> > >
> > >
> > > [...]
> > > > > diff --git a/drivers/gpu/drm/i915/display/intel_atomic.c b/drivers/gpu/drm/i915/display/intel_atomic.c
> > > > > index 6621aa245caf..43b1c7a227f8 100644
> > > > > --- a/drivers/gpu/drm/i915/display/intel_atomic.c
> > > > > +++ b/drivers/gpu/drm/i915/display/intel_atomic.c
> > > > > @@ -38,6 +38,7 @@
> > > > > #include "intel_atomic.h"
> > > > > #include "intel_cdclk.h"
> > > > > #include "intel_display_types.h"
> > > > > +#include "intel_fb.h"
> > > > > #include "intel_global_state.h"
> > > > > #include "intel_hdcp.h"
> > > > > #include "intel_psr.h"
> > > > > @@ -375,6 +376,52 @@ static void intel_atomic_setup_scaler(struct intel_crtc_scaler_state *scaler_sta
> > > > > mode = SKL_PS_SCALER_MODE_DYN;
> > > > > }
> > > > >
> > > > > + if (plane_state && plane_state->hw.fb) {
> > > > > + const struct drm_framebuffer *fb = plane_state->hw.fb;
> > > > > + struct drm_rect *src = &plane_state->uapi.src;
> > > > > + struct drm_rect *dst = &plane_state->uapi.dst;
> > > >
> > > > We want the scale factor checks for the pfit use case too. So this
> > > > stuff shouldn't be so tied into planes. I guess we could go with
> > > > a FIXME initially.
> > >
> > > Okay, I'll add a FIXME. It was tied to the plane checks before,
> > > though, wasn't it? So nothing should have changed in that regard here.
> > >
> > >
> > > > > + int hscale, vscale, max_vscale, max_hscale;
> > > > > +
> > > > > + if (DISPLAY_VER(dev_priv) >= 14) {
> > > > > + /*
> > > > > + * On versions 14 and up, only the first
> > > > > + * scaler supports a vertical scaling factor
> > > > > + * of more than 1.0, while a horizontal
> > > > > + * scaling factor of 3.0 is supported.
> > > > > + */
> > > > > + max_hscale = 0x30000 - 1;
> > > > > + if (*scaler_id == 0)
> > > > > + max_vscale = 0x30000 - 1;
> > > > > + else
> > > > > + max_vscale = 0x10000;
> > > >
> > > > We still have the chicken vs. egg problem here that we'd need to
> > > > consider the scale factors already when selecting the scaler.
> > > > But that could be another FIXME.
> > >
> > > Do you mean in regards to the HQ vs. non-HQ needs?
> >
> > I mean the "no downscaling on the second scaler" thing. The
> > problem is that this thing walks the scaler users in order
> > and assigns each one a scaler in turn. So if the first user
> > doesn't need downscaling but the second one does then we
> > will fail. OTOH had we just assigned the scalers in the
> > reverse order we would have succeeded.
>
> Ah, now I get it.
>
> So, I guess we can do a similar thing to what we already do earlier in
> this function to select the first scaler if HQ is needed. If
> downscaling is needed in one of the scalers but not in the other, we
> can swap them to make sure the one that needs downscaling in in the
> first one.
>
> But I agree with you in that this could be an added FIXME in this patch
> and the actual change could be made in a separate patch.
>
>
> > > > > +
> > > > > + } else if (DISPLAY_VER(dev_priv) >= 10 ||
> > > > > + !intel_format_info_is_yuv_semiplanar(fb->format, fb->modifier)) {
> > > > > + max_hscale = 0x30000 - 1;
> > > > > + max_vscale = 0x30000 - 1;
> > > > > + } else {
> > > > > + max_hscale = 0x20000 - 1;
> > > > > + max_vscale = 0x20000 - 1;
> > > > > + }
> > > >
> > > > Pre-glk hq scaler case not handled.
> > >
> > > I don't recall seen this specifically checked before. Is this the
> > > stuff I missed from g4x_sprite_check() below? Or am I missing something
> > > else?
> >
> > It was broken before in the sense that it allowed up to 3.0 scale factor
> > whether or not the HQ scaler was used or not. Now it will reject anything
> > above 2.0 even if the HQ scaler is used. So I guess technically it's a bit
> > less broken now, but more limited. Anyways, should be possible to just
> > check the scaler mode and pick the correct scaling limits based on it.
>
> I see what you mean. But the code from before had the exact same
> thing, I think? We were also checking for VER >= 10 and only then using
> 3.0. For anything else the limit was 2.0. Is your comment related to
> the FIXME I removed from below? (your last comment in this email)
>
>
> > > > > +
> > > > > + /* Check if required scaling is within limits */
> > > > > + hscale = drm_rect_calc_hscale(src, dst, 1, max_hscale);
> > > > > + vscale = drm_rect_calc_vscale(src, dst, 1, max_vscale);
> > > > > +
> > > > > + if (hscale < 0 || vscale < 0) {
> > > > > + drm_dbg_kms(&dev_priv->drm,
> > > > > + "Scaler %d doesn't support required plane scaling\n",
> > > > > + *scaler_id);
> > > > > + drm_rect_debug_print("src: ", src, true);
> > > > > + drm_rect_debug_print("dst: ", dst, false);
> > > > > +
> > > > > + scaler_state->scalers[*scaler_id].in_use = 0;
> > > > > + *scaler_id = -1;
> > > > > +
> > > > > + return;
> > > >
> > > > This would have to return an error rather than pretending that
> > > > everything is fine.
> > >
> > > We were already pretending everything is fine if a scaler if we
> > > couldn't find a free scaler, for instance, so I just kept the same
> > > logic, clearing up the scaler_id and marking the scaler as not in use
> > > as well.
> > >
> > > I can convert this to return an error, of course. But then in the "not
> > > free scaler" case we would still just ignore it or should we change the
> > > behavior and make it fail as well?
> >
> > The code is a mess, but it does look like intel_atomic_setup_scalers()
> > should fail correctly if we can't find enough scalers.
>
> Okay, so I'll change this function to return errors in both cases.
>
>
> > > [...]
> > > > > diff --git a/drivers/gpu/drm/i915/display/intel_atomic_plane.c b/drivers/gpu/drm/i915/display/intel_atomic_plane.c
> > > > > index 10e1fc9d0698..9100f328df60 100644
> > > > > --- a/drivers/gpu/drm/i915/display/intel_atomic_plane.c
> > > > > +++ b/drivers/gpu/drm/i915/display/intel_atomic_plane.c
> > > > > @@ -887,7 +887,7 @@ void intel_crtc_planes_update_arm(struct intel_atomic_state *state,
> > > > >
> > > > > int intel_atomic_plane_check_clipping(struct intel_plane_state *plane_state,
> > > > > struct intel_crtc_state *crtc_state,
> > > > > - int min_scale, int max_scale,
> > > > > + bool allow_scaling,
> > > > > bool can_position)
> > > > > {
> > > > > struct drm_i915_private *i915 = to_i915(plane_state->uapi.plane->dev);
> > > > > @@ -897,19 +897,50 @@ int intel_atomic_plane_check_clipping(struct intel_plane_state *plane_state,
> > > > > const struct drm_rect *clip = &crtc_state->pipe_src;
> > > > > unsigned int rotation = plane_state->hw.rotation;
> > > > > int hscale, vscale;
> > > > > + int max_hscale, min_hscale, max_vscale, min_vscale;
> > > > >
> > > > > if (!fb) {
> > > > > plane_state->uapi.visible = false;
> > > > > return 0;
> > > > > }
> > > > >
> > > > > + /*
> > > > > + * At this point we don't really know the HW limitations, so
> > > > > + * we just sanitize the values against the maximum supported
> > > > > + * scaling.
> > > > > + */
> > > > > + if (allow_scaling) {
> > > > > + min_vscale = 1;
> > > > > + min_hscale = 1;
> > > > > +
> > > > > + if (DISPLAY_VER(i915) < 10 ||
> > > > > + intel_format_info_is_yuv_semiplanar(fb->format,
> > > > > + fb->modifier)) {
> > > > > + max_vscale = 0x20000 - 1;
> > > > > + max_hscale = 0x20000 - 1;
> > > > > + } else {
> > > > > + max_vscale = 0x30000 - 1;
> > > > > + max_hscale = 0x30000 - 1;
> > > > > + }
> > > > > + } else {
> > > > > + min_hscale = DRM_PLANE_NO_SCALING;
> > > > > + max_hscale = DRM_PLANE_NO_SCALING;
> > > > > + min_vscale = DRM_PLANE_NO_SCALING;
> > > > > + max_vscale = DRM_PLANE_NO_SCALING;
> > > > > + }
> > > >
> > > > I still don't see the point in moving this hw specific knowledge
> > > > from the more hw specific files into the hw agnostic file.
> > >
> > > Is this file really that HW agnostic? I see lots of version checks with
> > > "if (DISPLAY_VER(x))" all over the place.
> >
> > It's hw agnostic in the sense that most of it applies
> > to all hw generations. And this function in particular is
> > pretty much entirely hw agnostic currently.
> >
>
>
>
>
> > > As we discussed before, I think this kind of rules should be in HW-
> > > specific configurations, but we don't have that yet. And I thought it
> > > would be better to keep these decisions in a single place rather than
> > > just calling functions in other files...
> > >
> > > If you prefer, I can move this back to skl_universal_plane.c or some
> > > other of the skl_*.c files, but TBH they don't seem to be the right
> > > place for this to me either...
> >
> > The current place knows exactly what kind of hardware/plane its
> > dealing with, and thus knows its limits. Seems perfectly fine to me.
>
> By "current place" you mean before this patch? I.e. in
> skl_universal_plane.c?
>
>
> > > [...]
> > > > > diff --git a/drivers/gpu/drm/i915/display/intel_sprite.c b/drivers/gpu/drm/i915/display/intel_sprite.c
> > > > > index e6b4d24b9cd0..9ad1173a0551 100644
> > > > > --- a/drivers/gpu/drm/i915/display/intel_sprite.c
> > > > > +++ b/drivers/gpu/drm/i915/display/intel_sprite.c
> > > > > @@ -1355,22 +1355,11 @@ g4x_sprite_check(struct intel_crtc_state *crtc_state,
> > > > > {
> > > > > struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
> > > > > struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
> > > > > - int min_scale = DRM_PLANE_NO_SCALING;
> > > > > - int max_scale = DRM_PLANE_NO_SCALING;
> > > > > int ret;
> > > > >
> > > > > - if (g4x_fb_scalable(plane_state->hw.fb)) {
> > > > > - if (DISPLAY_VER(dev_priv) < 7) {
> > > > > - min_scale = 1;
> > > > > - max_scale = 16 << 16;
> > > > > - } else if (IS_IVYBRIDGE(dev_priv)) {
> > > > > - min_scale = 1;
> > > > > - max_scale = 2 << 16;
> > > > > - }
> > > > > - }
> > > >
> > > > So what happened to these limits?
> > >
> > > Oh, it seems that I lost them. I guess they should be moved to the
> > > intel_atomic_plane_check_clipping() function. Again, to keep it all in
> > > a single place. But this seems to be only required in the sprite code,
> > > so I'm not sure what I can do.
> > >
> > > It's a problem to have this kinds of checks everywhere. 😞
I can't wrap my head around this. With my patch, the IVIBRIDGE part is
okay, because IVIBRIDGE is version 7 and for anything below version 10
we use 2.0 as the limit.
But I don't understand why anything below version 7 has a limit of 16.0
and only for this sprite related check. Is that really correct? Why is
it only related to the sprite checks?
> > >
> > > [...]
> > > > > diff --git a/drivers/gpu/drm/i915/display/skl_universal_plane.c b/drivers/gpu/drm/i915/display/skl_universal_plane.c
> > > > > index 76490cc59d8f..e2ae6624378f 100644
> > > > > --- a/drivers/gpu/drm/i915/display/skl_universal_plane.c
> > > > > +++ b/drivers/gpu/drm/i915/display/skl_universal_plane.c
> > > > > @@ -1463,22 +1463,6 @@ static int skl_plane_check_nv12_rotation(const struct intel_plane_state *plane_s
> > > > > return 0;
> > > > > }
> > > > >
> > > > > -static int skl_plane_max_scale(struct drm_i915_private *dev_priv,
> > > > > - const struct drm_framebuffer *fb)
> > > > > -{
> > > > > - /*
> > > > > - * We don't yet know the final source width nor
> > > > > - * whether we can use the HQ scaler mode. Assume
> > > > > - * the best case.
> > > > > - * FIXME need to properly check this later.
> > > > > - */
> > > >
> > > > Doesn't look like that FIXME has been dealt with as far
> > > > as the hq scaler is concerned.
> > >
> > > We now check the limits _after_ having decided whether HQ mode is used.
> > > So that should be covered, right?
> >
> > If we actaully add the hq mode check when determining the scaling
> > limits.
>
> Right, so the new FIXME I'll add as discussed above should cover this.
>
> --
> Cheers,
> Luca.
More information about the Intel-gfx
mailing list