Current implementation calculates frame delay based on the time of
start of the loop. This inevitably causes the loop delay to be
slightly longer than the actual measurement period, thus skipping a frame
every now and then.
However, MLX90640 should ideally be working without a frame skip.
Each measurement step updates half of the pixels in the frame
(every other pixel in default "chess mode", and every other row
in "interleave mode"), while additional coefficient data (25th & 26th row)
updates every step. The compensational coefficient data only corresponds
with the pixels updated in the same step.
In short, if a frame is skipped, then half of a frame loses correction
information and becomes garbage data.
Signed-off-by: Seongyong Park <euphoriccatface@xxxxxxxxx>
---
drivers/media/i2c/video-i2c.c | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
diff --git a/drivers/media/i2c/video-i2c.c b/drivers/media/i2c/video-i2c.c
index 0465832a4..2ccb08335 100644
--- a/drivers/media/i2c/video-i2c.c
+++ b/drivers/media/i2c/video-i2c.c
@@ -445,14 +445,16 @@ static int video_i2c_thread_vid_cap(void *priv)
struct video_i2c_data *data = priv;
unsigned int delay = mult_frac(HZ, data->frame_interval.numerator,
data->frame_interval.denominator);
+ unsigned long end_jiffies = jiffies;
set_freezable();
do {
- unsigned long start_jiffies = jiffies;
struct video_i2c_buffer *vid_cap_buf = NULL;
int schedule_delay;
+ end_jiffies += delay;
+
try_to_freeze();
spin_lock(&data->slock);
@@ -477,10 +479,9 @@ static int video_i2c_thread_vid_cap(void *priv)
VB2_BUF_STATE_ERROR : VB2_BUF_STATE_DONE);
}
- schedule_delay = delay - (jiffies - start_jiffies);
-
- if (time_after(jiffies, start_jiffies + delay))
- schedule_delay = delay;
+ schedule_delay = end_jiffies - jiffies;
+ while (schedule_delay <= 0)
+ schedule_delay += delay;
schedule_timeout_interruptible(schedule_delay);
} while (!kthread_should_stop());
--
2.31.1
