target/linux/bcm27xx/patches-5.4/950-0546-drm-vc4-plane-Improve-LBM-usage.patch

   1 From 81072e19a85bfa3f80c23acdff6156522d945efa Mon Sep 17 00:00:00 2001
   2 From: Dave Stevenson <dave.stevenson@raspberrypi.com>
   3 Date: Tue, 11 Feb 2020 16:55:02 +0000
   4 Subject: [PATCH] drm/vc4: plane: Improve LBM usage
   5
   6 LBM allocations were always taking the worst case sizing of
   7 max(src_width, dst_width) * 16. This is significantly over
   8 the required sizing, and stops us rendering multiple 4k images
   9 to the screen.
  10
  11 Add some of the additional constraints to more accurately
  12 describe the LBM requirements.
  13
  14 Signed-off-by: Dave Stevenson <dave.stevenson@raspberrypi.com>
  15 Signed-off-by: Maxime Ripard <maxime@cerno.tech>
  16 ---
  17  drivers/gpu/drm/vc4/vc4_plane.c | 31 ++++++++++++++++++++-----------
  18  1 file changed, 20 insertions(+), 11 deletions(-)
  19
  20 --- a/drivers/gpu/drm/vc4/vc4_plane.c
  21 +++ b/drivers/gpu/drm/vc4/vc4_plane.c
  22 @@ -142,9 +142,10 @@ static const struct hvs_format *vc4_get_
  23         return NULL;
  24  }
  25
  26 -static enum vc4_scaling_mode vc4_get_scaling_mode(u32 src, u32 dst)
  27 +static enum vc4_scaling_mode vc4_get_scaling_mode(u32 src, u32 dst,
  28 +                                                 bool chroma_vrep)
  29  {
  30 -       if (dst == src)
  31 +       if (dst == src && !chroma_vrep)
  32                 return VC4_SCALING_NONE;
  33         if (3 * dst >= 2 * src)
  34                 return VC4_SCALING_PPF;
  35 @@ -369,9 +370,11 @@ static int vc4_plane_setup_clipping_and_
  36                 return ret;
  37
  38         vc4_state->x_scaling[0] = vc4_get_scaling_mode(vc4_state->src_w[0],
  39 -                                                      vc4_state->crtc_w);
  40 +                                                      vc4_state->crtc_w,
  41 +                                                      false);
  42         vc4_state->y_scaling[0] = vc4_get_scaling_mode(vc4_state->src_h[0],
  43 -                                                      vc4_state->crtc_h);
  44 +                                                      vc4_state->crtc_h,
  45 +                                                      false);
  46
  47         vc4_state->is_unity = (vc4_state->x_scaling[0] == VC4_SCALING_NONE &&
  48                                vc4_state->y_scaling[0] == VC4_SCALING_NONE);
  49 @@ -384,10 +387,12 @@ static int vc4_plane_setup_clipping_and_
  50
  51                 vc4_state->x_scaling[1] =
  52                         vc4_get_scaling_mode(vc4_state->src_w[1],
  53 -                                            vc4_state->crtc_w);
  54 +                                            vc4_state->crtc_w,
  55 +                                            v_subsample == 2);
  56                 vc4_state->y_scaling[1] =
  57                         vc4_get_scaling_mode(vc4_state->src_h[1],
  58 -                                            vc4_state->crtc_h);
  59 +                                            vc4_state->crtc_h,
  60 +                                            v_subsample == 2);
  61
  62                 /* YUV conversion requires that horizontal scaling be enabled
  63                  * on the UV plane even if vc4_get_scaling_mode() returned
  64 @@ -437,10 +442,7 @@ static void vc4_write_ppf(struct vc4_pla
  65  static u32 vc4_lbm_size(struct drm_plane_state *state)
  66  {
  67         struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
  68 -       /* This is the worst case number.  One of the two sizes will
  69 -        * be used depending on the scaling configuration.
  70 -        */
  71 -       u32 pix_per_line = max(vc4_state->src_w[0], (u32)vc4_state->crtc_w);
  72 +       u32 pix_per_line;
  73         u32 lbm;
  74
  75         /* LBM is not needed when there's no vertical scaling. */
  76 @@ -448,6 +450,11 @@ static u32 vc4_lbm_size(struct drm_plane
  77             vc4_state->y_scaling[1] == VC4_SCALING_NONE)
  78                 return 0;
  79
  80 +       if (vc4_state->x_scaling[0] == VC4_SCALING_TPZ)
  81 +               pix_per_line = vc4_state->crtc_w;
  82 +       else
  83 +               pix_per_line = vc4_state->src_w[0];
  84 +
  85         if (!vc4_state->is_yuv) {
  86                 if (vc4_state->y_scaling[0] == VC4_SCALING_TPZ)
  87                         lbm = pix_per_line * 8;
  88 @@ -583,7 +590,9 @@ static int vc4_plane_allocate_lbm(struct
  89                 spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
  90                 ret = drm_mm_insert_node_generic(&vc4->hvs->lbm_mm,
  91                                                  &vc4_state->lbm,
  92 -                                                lbm_size, 32, 0, 0);
  93 +                                                lbm_size,
  94 +                                                vc4->hvs->hvs5 ? 64 : 32,
  95 +                                                0, 0);
  96                 spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
  97
  98                 if (ret)