diff mbox

[06/35] media: ti-vpe: vpe: Do not perform job transaction atomically

Message ID 20160928212040.26547-1-bparrot@ti.com (mailing list archive)
State New, archived
Headers show

Commit Message

Benoit Parrot Sept. 28, 2016, 9:20 p.m. UTC
From: Nikhil Devshatwar <nikhil.nd@ti.com>

Current VPE driver does not start the job until all the buffers for
a transaction are not queued. When running in multiple context, this might
increase the processing latency.

Alternate solution would be to try to continue the same context as long as
buffers for the transaction are ready; else switch the context. This may
increase number of context switches but it reduces latency significantly.

In this approach, the job_ready always succeeds as long as there are
buffers on the CAPTURE and OUTPUT stream. Processing may start immediately
as the first 2 iterations don't need extra source buffers. Shift all the
source buffers after each iteration and remove the oldest buffer.

Also, with this removes the constraint of pre buffering 3 buffers before
call to STREAMON in case of de-interlacing.

Signed-off-by: Nikhil Devshatwar <nikhil.nd@ti.com>
Signed-off-by: Benoit Parrot <bparrot@ti.com>
---
 drivers/media/platform/ti-vpe/vpe.c | 32 ++++++++++++++++----------------
 1 file changed, 16 insertions(+), 16 deletions(-)

Comments

Hans Verkuil Oct. 17, 2016, 2:17 p.m. UTC | #1
On 09/28/2016 11:20 PM, Benoit Parrot wrote:
> From: Nikhil Devshatwar <nikhil.nd@ti.com>
> 
> Current VPE driver does not start the job until all the buffers for
> a transaction are not queued. When running in multiple context, this might

I think this should be: s/not queued/queued/, right?

> increase the processing latency.
> 
> Alternate solution would be to try to continue the same context as long as
> buffers for the transaction are ready; else switch the context. This may
> increase number of context switches but it reduces latency significantly.
> 
> In this approach, the job_ready always succeeds as long as there are
> buffers on the CAPTURE and OUTPUT stream. Processing may start immediately
> as the first 2 iterations don't need extra source buffers. Shift all the
> source buffers after each iteration and remove the oldest buffer.
> 
> Also, with this removes the constraint of pre buffering 3 buffers before
> call to STREAMON in case of de-interlacing.
> 
> Signed-off-by: Nikhil Devshatwar <nikhil.nd@ti.com>
> Signed-off-by: Benoit Parrot <bparrot@ti.com>
> ---
>  drivers/media/platform/ti-vpe/vpe.c | 32 ++++++++++++++++----------------
>  1 file changed, 16 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/media/platform/ti-vpe/vpe.c b/drivers/media/platform/ti-vpe/vpe.c
> index a0b29685fb69..9c38eff5df46 100644
> --- a/drivers/media/platform/ti-vpe/vpe.c
> +++ b/drivers/media/platform/ti-vpe/vpe.c
> @@ -898,15 +898,14 @@ static struct vpe_ctx *file2ctx(struct file *file)
>  static int job_ready(void *priv)
>  {
>  	struct vpe_ctx *ctx = priv;
> -	int needed = ctx->bufs_per_job;
>  
> -	if (ctx->deinterlacing && ctx->src_vbs[2] == NULL)
> -		needed += 2;	/* need additional two most recent fields */
> -
> -	if (v4l2_m2m_num_src_bufs_ready(ctx->fh.m2m_ctx) < needed)
> -		return 0;
> -
> -	if (v4l2_m2m_num_dst_bufs_ready(ctx->fh.m2m_ctx) < needed)
> +	/*
> +	 * This check is needed as this might be called directly from driver
> +	 * When called by m2m framework, this will always satisy, but when

typo: satisfy

> +	 * called from vpe_irq, this might fail. (src stream with zero buffers)
> +	 */
> +	if (v4l2_m2m_num_src_bufs_ready(ctx->fh.m2m_ctx) <= 0 ||
> +		v4l2_m2m_num_dst_bufs_ready(ctx->fh.m2m_ctx) <= 0)
>  		return 0;
>  
>  	return 1;
> @@ -1116,19 +1115,20 @@ static void device_run(void *priv)
>  	struct sc_data *sc = ctx->dev->sc;
>  	struct vpe_q_data *d_q_data = &ctx->q_data[Q_DATA_DST];
>  
> -	if (ctx->deinterlacing && ctx->src_vbs[2] == NULL) {
> -		ctx->src_vbs[2] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
> -		WARN_ON(ctx->src_vbs[2] == NULL);
> -		ctx->src_vbs[1] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
> -		WARN_ON(ctx->src_vbs[1] == NULL);
> -	}
> -
>  	ctx->src_vbs[0] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
>  	WARN_ON(ctx->src_vbs[0] == NULL);
>  	ctx->dst_vb = v4l2_m2m_dst_buf_remove(ctx->fh.m2m_ctx);
>  	WARN_ON(ctx->dst_vb == NULL);
>  
>  	if (ctx->deinterlacing) {
> +
> +		if (ctx->src_vbs[2] == NULL) {
> +			ctx->src_vbs[2] = ctx->src_vbs[0];
> +			WARN_ON(ctx->src_vbs[2] == NULL);
> +			ctx->src_vbs[1] = ctx->src_vbs[0];
> +			WARN_ON(ctx->src_vbs[1] == NULL);
> +		}
> +
>  		/*
>  		 * we have output the first 2 frames through line average, we
>  		 * now switch to EDI de-interlacer
> @@ -1349,7 +1349,7 @@ static irqreturn_t vpe_irq(int irq_vpe, void *data)
>  	}
>  
>  	ctx->bufs_completed++;
> -	if (ctx->bufs_completed < ctx->bufs_per_job) {
> +	if (ctx->bufs_completed < ctx->bufs_per_job && job_ready(ctx)) {
>  		device_run(ctx);
>  		goto handled;
>  	}
> 

Regards,

	Hans
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Benoit Parrot Oct. 24, 2016, 4:57 p.m. UTC | #2
Hans Verkuil <hverkuil@xs4all.nl> wrote on Mon [2016-Oct-17 16:17:15 +0200]:
> On 09/28/2016 11:20 PM, Benoit Parrot wrote:
> > From: Nikhil Devshatwar <nikhil.nd@ti.com>
> > 
> > Current VPE driver does not start the job until all the buffers for
> > a transaction are not queued. When running in multiple context, this might
> 
> I think this should be: s/not queued/queued/, right?

Yep, will fix that.

> 
> > increase the processing latency.
> > 
> > Alternate solution would be to try to continue the same context as long as
> > buffers for the transaction are ready; else switch the context. This may
> > increase number of context switches but it reduces latency significantly.
> > 
> > In this approach, the job_ready always succeeds as long as there are
> > buffers on the CAPTURE and OUTPUT stream. Processing may start immediately
> > as the first 2 iterations don't need extra source buffers. Shift all the
> > source buffers after each iteration and remove the oldest buffer.
> > 
> > Also, with this removes the constraint of pre buffering 3 buffers before
> > call to STREAMON in case of de-interlacing.
> > 
> > Signed-off-by: Nikhil Devshatwar <nikhil.nd@ti.com>
> > Signed-off-by: Benoit Parrot <bparrot@ti.com>
> > ---
> >  drivers/media/platform/ti-vpe/vpe.c | 32 ++++++++++++++++----------------
> >  1 file changed, 16 insertions(+), 16 deletions(-)
> > 
> > diff --git a/drivers/media/platform/ti-vpe/vpe.c b/drivers/media/platform/ti-vpe/vpe.c
> > index a0b29685fb69..9c38eff5df46 100644
> > --- a/drivers/media/platform/ti-vpe/vpe.c
> > +++ b/drivers/media/platform/ti-vpe/vpe.c
> > @@ -898,15 +898,14 @@ static struct vpe_ctx *file2ctx(struct file *file)
> >  static int job_ready(void *priv)
> >  {
> >  	struct vpe_ctx *ctx = priv;
> > -	int needed = ctx->bufs_per_job;
> >  
> > -	if (ctx->deinterlacing && ctx->src_vbs[2] == NULL)
> > -		needed += 2;	/* need additional two most recent fields */
> > -
> > -	if (v4l2_m2m_num_src_bufs_ready(ctx->fh.m2m_ctx) < needed)
> > -		return 0;
> > -
> > -	if (v4l2_m2m_num_dst_bufs_ready(ctx->fh.m2m_ctx) < needed)
> > +	/*
> > +	 * This check is needed as this might be called directly from driver
> > +	 * When called by m2m framework, this will always satisy, but when
> 
> typo: satisfy

Will fix.

> 
> > +	 * called from vpe_irq, this might fail. (src stream with zero buffers)
> > +	 */
> > +	if (v4l2_m2m_num_src_bufs_ready(ctx->fh.m2m_ctx) <= 0 ||
> > +		v4l2_m2m_num_dst_bufs_ready(ctx->fh.m2m_ctx) <= 0)
> >  		return 0;
> >  
> >  	return 1;
> > @@ -1116,19 +1115,20 @@ static void device_run(void *priv)
> >  	struct sc_data *sc = ctx->dev->sc;
> >  	struct vpe_q_data *d_q_data = &ctx->q_data[Q_DATA_DST];
> >  
> > -	if (ctx->deinterlacing && ctx->src_vbs[2] == NULL) {
> > -		ctx->src_vbs[2] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
> > -		WARN_ON(ctx->src_vbs[2] == NULL);
> > -		ctx->src_vbs[1] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
> > -		WARN_ON(ctx->src_vbs[1] == NULL);
> > -	}
> > -
> >  	ctx->src_vbs[0] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
> >  	WARN_ON(ctx->src_vbs[0] == NULL);
> >  	ctx->dst_vb = v4l2_m2m_dst_buf_remove(ctx->fh.m2m_ctx);
> >  	WARN_ON(ctx->dst_vb == NULL);
> >  
> >  	if (ctx->deinterlacing) {
> > +
> > +		if (ctx->src_vbs[2] == NULL) {
> > +			ctx->src_vbs[2] = ctx->src_vbs[0];
> > +			WARN_ON(ctx->src_vbs[2] == NULL);
> > +			ctx->src_vbs[1] = ctx->src_vbs[0];
> > +			WARN_ON(ctx->src_vbs[1] == NULL);
> > +		}
> > +
> >  		/*
> >  		 * we have output the first 2 frames through line average, we
> >  		 * now switch to EDI de-interlacer
> > @@ -1349,7 +1349,7 @@ static irqreturn_t vpe_irq(int irq_vpe, void *data)
> >  	}
> >  
> >  	ctx->bufs_completed++;
> > -	if (ctx->bufs_completed < ctx->bufs_per_job) {
> > +	if (ctx->bufs_completed < ctx->bufs_per_job && job_ready(ctx)) {
> >  		device_run(ctx);
> >  		goto handled;
> >  	}
> > 
> 
> Regards,
> 
> 	Hans
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/media/platform/ti-vpe/vpe.c b/drivers/media/platform/ti-vpe/vpe.c
index a0b29685fb69..9c38eff5df46 100644
--- a/drivers/media/platform/ti-vpe/vpe.c
+++ b/drivers/media/platform/ti-vpe/vpe.c
@@ -898,15 +898,14 @@  static struct vpe_ctx *file2ctx(struct file *file)
 static int job_ready(void *priv)
 {
 	struct vpe_ctx *ctx = priv;
-	int needed = ctx->bufs_per_job;
 
-	if (ctx->deinterlacing && ctx->src_vbs[2] == NULL)
-		needed += 2;	/* need additional two most recent fields */
-
-	if (v4l2_m2m_num_src_bufs_ready(ctx->fh.m2m_ctx) < needed)
-		return 0;
-
-	if (v4l2_m2m_num_dst_bufs_ready(ctx->fh.m2m_ctx) < needed)
+	/*
+	 * This check is needed as this might be called directly from driver
+	 * When called by m2m framework, this will always satisy, but when
+	 * called from vpe_irq, this might fail. (src stream with zero buffers)
+	 */
+	if (v4l2_m2m_num_src_bufs_ready(ctx->fh.m2m_ctx) <= 0 ||
+		v4l2_m2m_num_dst_bufs_ready(ctx->fh.m2m_ctx) <= 0)
 		return 0;
 
 	return 1;
@@ -1116,19 +1115,20 @@  static void device_run(void *priv)
 	struct sc_data *sc = ctx->dev->sc;
 	struct vpe_q_data *d_q_data = &ctx->q_data[Q_DATA_DST];
 
-	if (ctx->deinterlacing && ctx->src_vbs[2] == NULL) {
-		ctx->src_vbs[2] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
-		WARN_ON(ctx->src_vbs[2] == NULL);
-		ctx->src_vbs[1] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
-		WARN_ON(ctx->src_vbs[1] == NULL);
-	}
-
 	ctx->src_vbs[0] = v4l2_m2m_src_buf_remove(ctx->fh.m2m_ctx);
 	WARN_ON(ctx->src_vbs[0] == NULL);
 	ctx->dst_vb = v4l2_m2m_dst_buf_remove(ctx->fh.m2m_ctx);
 	WARN_ON(ctx->dst_vb == NULL);
 
 	if (ctx->deinterlacing) {
+
+		if (ctx->src_vbs[2] == NULL) {
+			ctx->src_vbs[2] = ctx->src_vbs[0];
+			WARN_ON(ctx->src_vbs[2] == NULL);
+			ctx->src_vbs[1] = ctx->src_vbs[0];
+			WARN_ON(ctx->src_vbs[1] == NULL);
+		}
+
 		/*
 		 * we have output the first 2 frames through line average, we
 		 * now switch to EDI de-interlacer
@@ -1349,7 +1349,7 @@  static irqreturn_t vpe_irq(int irq_vpe, void *data)
 	}
 
 	ctx->bufs_completed++;
-	if (ctx->bufs_completed < ctx->bufs_per_job) {
+	if (ctx->bufs_completed < ctx->bufs_per_job && job_ready(ctx)) {
 		device_run(ctx);
 		goto handled;
 	}