[PATCH v2 4/4] ipa: rkisp1: use active state in setControls

[Mikhail Rudenko]

The results of AGC algorithm currently travel a long way before being
actually applied. Let's consider the common case when 4 requests are
used and frames 0-3 are queued before start(). AGC is first run when
frame 0 stats are ready, and exposure/gain are saved in
activeState. Then, when frame 4 is queued, they are stored in
corresponding frameContext. Four frames later, frame 4 is captured,
and after processing the stats buffer, exposure/gain are extracted
from frameContext and queued to delayedControls in setControls(). On
frame 5, delayedControls apply them. Assuming a control delay of 2, the
settings will become effective for frame 7. So, it takes 7 frames for
the AGC algorithm to get feedback. This results in suboptimal
convergence rate.

If we instead use just computed exposure/gain from the activeState in
setControls, the delay is reduced from 7 frames to 3 frames. Tests
on OV4689 sensor show faster convergence and no unwanted side effects.
---
 src/ipa/rkisp1/rkisp1.cpp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/src/ipa/rkisp1/rkisp1.cpp b/src/ipa/rkisp1/rkisp1.cpp
index 680a7eee..8793d59f 100644
--- a/src/ipa/rkisp1/rkisp1.cpp
+++ b/src/ipa/rkisp1/rkisp1.cpp
@@ -458,9 +458,11 @@ void IPARkISP1::setControls(unsigned int frame)
 	 * internal sensor delays and other timing parameters into account.
 	 */

-	IPAFrameContext &frameContext = context_.frameContexts.get(frame);
-	uint32_t exposure = frameContext.agc.exposure;
-	uint32_t gain = context_.camHelper->gainCode(frameContext.agc.gain);
+	const auto &agc = context_.activeState.agc;
+	uint32_t exposure = agc.autoEnabled ?
+			    agc.automatic.exposure : agc.manual.exposure;
+	uint32_t gain = context_.camHelper->gainCode(agc.autoEnabled ?
+						     agc.automatic.gain : agc.manual.gain);

 	ControlList ctrls(sensorControls_);
 	ctrls.set(V4L2_CID_EXPOSURE, static_cast<int32_t>(exposure));
--
2.46.0