[PATCH] drm/vc4: Flush the caches before the bin jobs, as well.

[PATCH] drm/vc4: Flush the caches before the bin jobs, as well.

[Eric Anholt]

If the frame samples from a render target that was just written, its
cache flush during the binning step may have occurred before the
previous frame's RCL was completed.  Flush the texture caches again
before starting each RCL job to make sure that the sampling of the
previous RCL's output is correct.

Fixes flickering in the top left of 3DMMES Taiji.

Signed-off-by: Eric Anholt <eric@anholt.net>
Fixes: ca26d28bbaa3 ("drm/vc4: improve throughput by pipelining binning and rendering jobs")
---
 drivers/gpu/drm/vc4/vc4_gem.c | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 6c32c89a83a9..afa7fe21b35e 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -436,6 +436,19 @@ vc4_flush_caches(struct drm_device *dev)
 		  VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
 }
 
+static void
+vc4_flush_texture_caches(struct drm_device *dev)
+{
+	struct vc4_dev *vc4 = to_vc4_dev(dev);
+
+	V3D_WRITE(V3D_L2CACTL,
+		  V3D_L2CACTL_L2CCLR);
+
+	V3D_WRITE(V3D_SLCACTL,
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
+		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC));
+}
+
 /* Sets the registers for the next job to be actually be executed in
  * the hardware.
  *
@@ -474,6 +487,14 @@ vc4_submit_next_render_job(struct drm_device *dev)
 	if (!exec)
 		return;
 
+	/* A previous RCL may have written to one of our textures, and
+	 * our full cache flush at bin time may have occurred before
+	 * that RCL completed.  Flush the texture cache now, but not
+	 * the instructions or uniforms (since we don't write those
+	 * from an RCL).
+	 */
+	vc4_flush_texture_caches(dev);
+
 	submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
 }
 
-- 
2.15.0

Re: [PATCH] drm/vc4: Flush the caches before the bin jobs, as well.

[Eric Anholt]

[-- Attachment #1: Type: text/plain, Size: 624 bytes --]

Eric Anholt <eric@anholt.net> writes:
> If the frame samples from a render target that was just written, its
> cache flush during the binning step may have occurred before the
> previous frame's RCL was completed.  Flush the texture caches again
> before starting each RCL job to make sure that the sampling of the
> previous RCL's output is correct.
>
> Fixes flickering in the top left of 3DMMES Taiji.
>
> Signed-off-by: Eric Anholt <eric@anholt.net>
> Fixes: ca26d28bbaa3 ("drm/vc4: improve throughput by pipelining binning and rendering jobs")

Whoops, in the subject, this should have been "before the *render* jobs."

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]

Re: [PATCH] drm/vc4: Flush the caches before the bin jobs, as well.

[Boris Brezillon]

On Thu, 21 Dec 2017 14:17:22 -0800
Eric Anholt <eric@anholt.net> wrote:

> If the frame samples from a render target that was just written, its
> cache flush during the binning step may have occurred before the
> previous frame's RCL was completed.  Flush the texture caches again
> before starting each RCL job to make sure that the sampling of the
> previous RCL's output is correct.
> 
> Fixes flickering in the top left of 3DMMES Taiji.
> 
> Signed-off-by: Eric Anholt <eric@anholt.net>
> Fixes: ca26d28bbaa3 ("drm/vc4: improve throughput by pipelining binning and rendering jobs")

Reviewed-by: Boris Brezillon <boris.brezillon@free-electrons.com>

> ---
>  drivers/gpu/drm/vc4/vc4_gem.c | 21 +++++++++++++++++++++
>  1 file changed, 21 insertions(+)
> 
> diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
> index 6c32c89a83a9..afa7fe21b35e 100644
> --- a/drivers/gpu/drm/vc4/vc4_gem.c
> +++ b/drivers/gpu/drm/vc4/vc4_gem.c
> @@ -436,6 +436,19 @@ vc4_flush_caches(struct drm_device *dev)
>  		  VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
>  }
>  
> +static void
> +vc4_flush_texture_caches(struct drm_device *dev)
> +{
> +	struct vc4_dev *vc4 = to_vc4_dev(dev);
> +
> +	V3D_WRITE(V3D_L2CACTL,
> +		  V3D_L2CACTL_L2CCLR);
> +
> +	V3D_WRITE(V3D_SLCACTL,
> +		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
> +		  VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC));
> +}
> +
>  /* Sets the registers for the next job to be actually be executed in
>   * the hardware.
>   *
> @@ -474,6 +487,14 @@ vc4_submit_next_render_job(struct drm_device *dev)
>  	if (!exec)
>  		return;
>  
> +	/* A previous RCL may have written to one of our textures, and
> +	 * our full cache flush at bin time may have occurred before
> +	 * that RCL completed.  Flush the texture cache now, but not
> +	 * the instructions or uniforms (since we don't write those
> +	 * from an RCL).
> +	 */
> +	vc4_flush_texture_caches(dev);
> +
>  	submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
>  }
>