gfx: overlay: Optimize filter coefficient load with sychronized flips

Without synchronized flips the driver can instruct the hardware to stop
re-loading the filter coefficients from memory once they've been loaded
by the hardware at leas once.

Bring the same optimization over to synchronized flips. To do that the
dirty state tracking must be split into software and hardware variants
due to the asynchronous nature of the synchronized flip code.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>

diff --git a/drivers/staging/mrst/drv/mdfld_overlay.c b/drivers/staging/mrst/drv/mdfld_overlay.c
index 50d75c5..8ffac3d 100644 (file)
--- a/drivers/staging/mrst/drv/mdfld_overlay.c
+++ b/drivers/staging/mrst/drv/mdfld_overlay.c
@@ -145,6 +145,7 @@ struct mfld_overlay {
        PVRSRV_KERNEL_MEM_INFO *mem_info;
        PVRSRV_KERNEL_MEM_INFO *old_mem_info;
        u32 tgid;
+       unsigned int sw_dirty;
 };
 
 #define to_mfld_overlay(plane) container_of(plane, struct mfld_overlay, base)
@@ -317,6 +318,13 @@ static void ovl_update_regs(struct mfld_overlay *ovl,
        wmb();
 }
 
+static void check_dirty_coefs(struct mfld_overlay *ovl, u32 dovsta)
+{
+       /* No pending updates, coefficients have been loaded for sure. */
+       if (dovsta & OVL_DOVSTA_OVR_UPDT)
+               ovl->dirty &= ~OVL_DIRTY_COEFS;
+}
+
 static void ovl_commit(struct mfld_overlay *ovl, unsigned int dirty)
 {
        struct drm_device *dev = ovl->dev;
@@ -325,9 +333,7 @@ static void ovl_commit(struct mfld_overlay *ovl, unsigned int dirty)
                return;
 
        if (ovl->atomic) {
-               spin_lock_irq(&ovl->regs_lock);
-               ovl->dirty |= dirty;
-               spin_unlock_irq(&ovl->regs_lock);
+               ovl->sw_dirty |= dirty;
                return;
        }
 
@@ -352,12 +358,8 @@ static void ovl_commit(struct mfld_overlay *ovl, unsigned int dirty)
 
        spin_lock_irq(&ovl->regs_lock);
 
-       /* FIXME cook up something like this when using atomic flips. */
-       if (ovl->dirty & OVL_DIRTY_COEFS && !(dirty & OVL_DIRTY_COEFS)) {
-               /* No pending updates, coefficients have been loaded for sure. */
-               if (OVL_REG_READ(ovl, OVL_DOVSTA) & OVL_DOVSTA_OVR_UPDT)
-                       ovl->dirty &= ~OVL_DIRTY_COEFS;
-       }
+       if ((ovl->dirty & OVL_DIRTY_COEFS) && !(dirty & OVL_DIRTY_COEFS))
+               check_dirty_coefs(ovl, OVL_REG_READ(ovl, OVL_DOVSTA));
 
        /* It's possible that mdfld_overlay_resume() already did this for us. */
        if (ovl->dirty & OVL_DIRTY_REGS)
@@ -1594,6 +1596,7 @@ struct mfld_overlay_flip {
        u32 tgid;
        struct mfld_overlay_regs regs;
        bool vblank_ref;
+       unsigned int dirty;
 };
 
 static void ovl_prepare(struct drm_flip *flip)
@@ -1623,8 +1626,13 @@ static bool ovl_flip(struct drm_flip *flip,
 
        ovl->curr_page = !ovl->curr_page;
 
+       ovl->dirty |= oflip->dirty;
+
        dovsta = OVL_REG_READ(ovl, OVL_DOVSTA);
 
+       if ((ovl->dirty & OVL_DIRTY_COEFS) && !(oflip->dirty & OVL_DIRTY_COEFS))
+               check_dirty_coefs(ovl, dovsta);
+
        write_ovadd(ovl);
 
        spin_unlock_irqrestore(&ovl->regs_lock, flags);
@@ -1825,19 +1833,8 @@ struct drm_flip *mdfld_overlay_atomic_flip(struct drm_plane *plane, int pipe)
        struct mfld_overlay *ovl = to_mfld_overlay(plane);
        struct mfld_overlay_flip *oflip;
 
-       if (ovl->pipe != pipe || !ovl->atomic)
-               return NULL;
-
-       spin_lock_irq(&ovl->regs_lock);
-
-       if (!(ovl->dirty & OVL_DIRTY_REGS)) {
-               spin_unlock_irq(&ovl->regs_lock);
+       if (ovl->pipe != pipe || !ovl->atomic || !ovl->sw_dirty)
                return NULL;
-       }
-
-       ovl->dirty &= ~OVL_DIRTY_REGS;
-
-       spin_unlock_irq(&ovl->regs_lock);
 
        oflip = kmalloc(sizeof *oflip, GFP_KERNEL);
        if (!oflip)
@@ -1866,5 +1863,8 @@ struct drm_flip *mdfld_overlay_atomic_flip(struct drm_plane *plane, int pipe)
        oflip->plane = plane;
        oflip->pipe = pipe;
 
+       oflip->dirty = ovl->sw_dirty;
+       ovl->sw_dirty = 0;
+
        return &oflip->base;
 }