From patchwork Wed Nov 13 13:12:50 2024 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Paul Elder X-Patchwork-Id: 21885 Return-Path: X-Original-To: parsemail@patchwork.libcamera.org Delivered-To: parsemail@patchwork.libcamera.org Received: from lancelot.ideasonboard.com (lancelot.ideasonboard.com [92.243.16.209]) by patchwork.libcamera.org (Postfix) with ESMTPS id 43A2EBE173 for ; Wed, 13 Nov 2024 13:13:25 +0000 (UTC) Received: from lancelot.ideasonboard.com (localhost [IPv6:::1]) by lancelot.ideasonboard.com (Postfix) with ESMTP id EEED66581E; Wed, 13 Nov 2024 14:13:24 +0100 (CET) Authentication-Results: lancelot.ideasonboard.com; dkim=pass (1024-bit key; unprotected) header.d=ideasonboard.com header.i=@ideasonboard.com header.b="TNOtL0+V"; dkim-atps=neutral Received: from perceval.ideasonboard.com (perceval.ideasonboard.com [213.167.242.64]) by lancelot.ideasonboard.com (Postfix) with ESMTPS id 7A36065816 for ; Wed, 13 Nov 2024 14:13:23 +0100 (CET) Received: from neptunite.flets-east.jp (unknown [IPv6:2404:7a81:160:2100:eb25:7aa0:7da7:eb24]) by perceval.ideasonboard.com (Postfix) with ESMTPSA id 343FF1306; Wed, 13 Nov 2024 14:13:07 +0100 (CET) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=ideasonboard.com; s=mail; t=1731503590; bh=HkBAWMufoWK21R/4mNX+Yn9tKT9evsJ9f/yELTctYGY=; h=From:To:Cc:Subject:Date:In-Reply-To:References:From; b=TNOtL0+VAdBOuVApDX2c8SRzUaNQdOXIlvt4PjTYgHOfjyphQvHq6P8hJyj/+to8n 10mw3kd9st1deut0Ds5VmC0dsjkdpOBIkK3k9cEFBkrLIszIL9tf1rHK4AajWgefBk XYCQH9lGmc0+34CLKNx2MMFoM7Ia8OlA4QpT7o8w= From: Paul Elder To: libcamera-devel@lists.libcamera.org Cc: laurent.pinchart@ideasonboard.com, jacopo.mondi@ideasonboard.com, naush@raspberrypi.com, david.plowman@raspberrypi.com, Paul Elder , Jacopo Mondi Subject: [PATCH v3 2/8] Documentation: design: ae: Document the design for AE controls Date: Wed, 13 Nov 2024 22:12:50 +0900 Message-Id: <20241113131256.3170817-3-paul.elder@ideasonboard.com> X-Mailer: git-send-email 2.39.2 In-Reply-To: <20241113131256.3170817-1-paul.elder@ideasonboard.com> References: <20241113131256.3170817-1-paul.elder@ideasonboard.com> MIME-Version: 1.0 X-BeenThere: libcamera-devel@lists.libcamera.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: libcamera-devel-bounces@lists.libcamera.org Sender: "libcamera-devel" Document the design and rationale for the AE-related controls. Also add documentation for the controls. Signed-off-by: Paul Elder Signed-off-by: Jacopo Mondi Reviewed-by: Stefan Klug --- Changes in v3: - merge the control documentation into the same document (including the patch) - this is because it was a bit unwieldy to put it in control_ids.cpp.in, now that it's used generically to generate control ids of all namespaces --- Documentation/design/ae.rst | 348 ++++++++++++++++++++++++++++++++++++ 1 file changed, 348 insertions(+) create mode 100644 Documentation/design/ae.rst diff --git a/Documentation/design/ae.rst b/Documentation/design/ae.rst new file mode 100644 index 000000000000..f121afecea5a --- /dev/null +++ b/Documentation/design/ae.rst @@ -0,0 +1,348 @@ +.. SPDX-License-Identifier: CC-BY-SA-4.0 + +Design of Exposure and Gain controls +==================================== + +This document explains the design and rationale of the controls related to +exposure and gain. This includes the all-encompassing auto-exposure (AE), the +manual exposure control, and the manual gain control. + +Description of the problem +-------------------------- + +Sub controls +^^^^^^^^^^^^ + +There are more than one control that make up exposure: exposure, gain, and +aperture (though for now we will not consider aperture). We already had +individual controls for setting the values of manual exposure and manual gain, +but for switching between auto mode and manual mode we only had a high-level +boolean AeEnable control that would set *both* exposure and gain to auto mode +or manual mode; we had no way to set one to auto and the other to manual. + +So, we need to introduce two new controls to act as "levers" to indicate +individually for exposure and gain if the value would come from AEGC or if it +would come from the manual control value. + +Aperture priority +^^^^^^^^^^^^^^^^^ + +We eventually will support aperture, and so whatever our solution is for having +only some controls on auto and the others on manual needs to be extensible. + +Flickering when going from auto to manual +^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + +When a manual exposure or gain value is requested by the application, it costs +a few frames worth of time for them to take effect. This means that during a +transition from auto to manual, there would be flickering in the control values +and the transition won't be smooth. + +Take for instance the following flow, where we start on auto exposure (which +for the purposes of the example increments by 1 each frame) and we want to +switch seamlessly to manual exposure, which involves copying the exposure value +computed by the auto exposure algorithm: + +:: + + +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ + | N | | N+1 | | N+2 | | N+3 | | N+4 | | N+5 | | N+6 | + +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ + + Mode requested: Auto Auto Auto Manual Manual Manual Manual + Exp requested: N/A N/A N/A 2 2 2 2 + Set in Frame: N+2 N+3 N+4 N+5 N+6 N+7 N+8 + + Mode used: Auto Auto Auto Auto Auto Manual Manual + Exp used: 0 1 2 3 4 2 2 + +As we can see, after frame N+2 completes, we copy the exposure value that was +used for frame N+2 (which was computed by AE algorithm), and queue that value +into request N+3 with manual mode on. However, as it takes two frames for the +exposure to be set, the exposure still changes since it is set by AE, and we +get a flicker in the exposure during the switch from auto to manual. + +A solution is to *not submit* any exposure value when manual mode is enabled, +and wait until the manual mode as been "applied" before copying the exposure +value: + +:: + + +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ + | N | | N+1 | | N+2 | | N+3 | | N+4 | | N+5 | | N+6 | + +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ + + Mode requested: Auto Auto Auto Manual Manual Manual Manual + Exp requested: N/A N/A N/A None None None 5 + Set in Frame: N+2 N+3 N+4 N+5 N+6 N+7 N+8 + + Mode used: Auto Auto Auto Auto Auto Manual Manual + Exp used: 0 1 2 3 4 5 5 + +In practice, this works. However, libcamera has a policy where once a control +is submitted, its value is saved and does not need to be resubmitted (even +though this isn't implemented yet). So if the manual exposure value was set +while auto mode was on, in theory the value would be saved, so when manual mode +is enabled, the exposure value that was previously set would immediately be +used. Clearly this solution isn't correct, but it can serve as the basis for a +proper solution, with some more rigorous rules. + +Existing solutions +------------------ + +Raspberry Pi +^^^^^^^^^^^^ + +The raspberry pi IPA got around the lack of individual AeEnable controls for +exposure and gain by using magic values. When AeEnable was false, if one of the +manual control values was set to 0 then the value computed by AEGC would be +used for just that control. This solution isn't desirable, as it prevents +that magic value from being used as a valid value. + +To get around the flickering issue, when AeEnable was false, the raspberry pi +AEGC would simply stop updating the values to be set. As mentioned above, since +the value retention mechanism hasn't actually been implemented yet, this +worked. But, it's not a proper solution. + +Android +^^^^^^^ + +The Android HAL specification requires that exposure and gain (sensitivity) +must both be manual or both be auto. It cannot be that one is manual while the +other is auto, so they simply don't support sub controls. + +For the flickering issue, the Android HAL has an AeLock control. To transition +from auto to manual, the application would keep AE on auto, and turn on the +lock. Once the lock has propagated through, then the value can be copied from +the result into the request and the lock disabled and the mode set to manual. + +The problem with this solution is, besides the extra complexity, that it is +ambiguous what happens if there is a state transition from manual to locked +(even though it's a state transition that doesn't make sense). If locked is +defined to "use the last automatically computed values" then it could use the +values from the last time it AE was set to auto, or it would be undefined if AE +was never auto (eg. it started out as manual), or if AE is implemented to run +in the background it could just use the current values that are computed. If +locked is defined to "use the last value that was set" there would be less +ambiguity. Still, it's better if we can make it impossible to execute this +nonsensical state transition, and if we can reduce the complexity of having +this extra control or extra setting on a lever. + +Summary of goals +---------------- + + - We need a lock of some sort, to instruct the AEGC to not update output + results + + - We need manual modes, to override the values computed by the AEGC + + - We need to support seamless transitions from auto to manual, and do so + without flickering + + - We need custom minimum values for the manual controls; that is, no magic + values for enabling/disabling auto + + - All of these need to be done with AE sub-controls (exposure time, analogue + gain) and be extensible to aperture in the future + +Our solution +------------ + +A diagram of our solution: + +:: + + +----------------------------+-------------+------------------+-----------------+ + | INPUT | ALGORITHM | RESULT | OUTPUT | + +----------------------------+-------------+------------------+-----------------+ + + ExposureTimeMode ExposureTimeMode + ---------------------+----------------------------------------+-----------------> + 0: Auto | | + 1: Manual | V + | |\ + | | \ + | /----------------------------------> | 1| ExposureTime + | | +-------------+ exposure time | | --------------> + \--)--> | | --------------> | 0| + ExposureTime | | | | / + ------------------------+--> | | |/ + | | AeState + | AEGC | -----------------------------------> + AnalogueGain | | + ------------------------+--> | | |\ + | | | | \ + /--)--> | | --------------> | 0| AnalogueGain + | | +-------------+ analogue gain | | --------------> + | \----------------------------------> | 1| + | | / + | |/ + | ^ + AnalogueGainMode | | AnalogueGainMode + ---------------------+----------------------------------------+-----------------> + 0: Auto + 1: Manual + + +The diagram is divided in four sections horizontally: + + - Input: The values received from the request controls + + - Algorithm: The algorithm itself + + - Result: The values calculated by the algorithm + + - Output: The values that sent in result metadata and applied to the device + +The four input controls are divided between manual values (ExposureTime and +AnalogueGain), and operation modes (ExposureTimeMode and AnalogueGainMode). The +former are the manual values, the latter control how they're applied. The two +modes are independent from each other, and each can take one of two values: + + - Auto (0): The AGC computes the value normally. The AGC result is applied + to the output. The manual value is ignored *and is not retained*. + + - Manual (1): The AGC uses the manual value internally. The corresponding + manual control from the request is applied to the output. The AGC result + is ignored. + +The AeState control reports the state of the unified AEGC block. If both +ExposureTimeMode and AnalogueGainMode are set to disabled then it will report +Idle. If at least one of the two is set to auto, then AeState will report +if the AEGC has Converged or not (Searching). This control replaces the old +AeLocked control, as it was insufficient for reporting the AE state. + +There is a caveat to the disabled mode: the manual control value is not +retained if it is set during auto mode. This means that if the disabled mode is +entered without also setting the manual value, then it will enter a state +similar to "locked", where the last automatically computed value while the mode +was auto will be used. Once the manual value is set, then that will be used and +retained as usual. + +This simulates an auto -> locked -> manual or auto -> manual state transition, +and makes it impossible to do the nonsensical manual -> locked state +transition. + +We specifically do not have a "master AE control" like the old AeEnable. This +is because we have the individual mode controls, and if we had a master AE +control it would be a "control that sets other controls", which could easily +get out of control. + +With this solution, the earlier example would become: + +:: + + +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ + | N+2 | | N+3 | | N+4 | | N+5 | | N+6 | | N+7 | | N+8 | | N+9 | | N+10| + +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ +-----+ + Mode requested: Auto Disab Disab Disab Disab Disab Disab Disab Disab + Exp requested: N/A None None None None 10 None 10 10 + Set in Frame: N+4 N+5 N+6 N+7 N+8 N+9 N+10 N+10 N+10 + + Mode used: Auto Auto Auto Disab Disab Disab Disab Disab Disab + Exp used: 2 3 4 5 5 5 5 10 10 + +This example is extended by a few frames to exhibit the simulated "locked" +state. At frame N+5 the application has confirmed that the auto mode has been +disabled, but does not provide a manual value until request N+7. Thus, the +value that is used in requests N+5 and N+6 (where the mode is disabled), comes +from the last value that was used when the mode was auto, which comes from +frame N+4. + +Then, in N+7, a manual value of 10 is supplied. It takes until frame N+9 for +the exposure to be applied. N+8 does not supply a manual value, but the last +supplied value is retained, so a manual value of 10 is still used and set in +frame N+10. + +Although this behavior is the same as what we had with waiting for the manual +mode to propagate (in the section "Description of the problem"), this time it +is correct as we have defined specifically that if a manual value was specified +while the mode was auto, it will not be retained. + +Description of the controls +=========================== + + libcamera offers the following controls related to exposure and gain: + + - AnalogueGain + + - AnalogueGainMode + + - ExposureTime + + - ExposureTimeMode + + - AeState + + Auto-exposure and auto-gain can be enabled and disabled separately using the + ExposureTimeMode and AnalogueGainMode controls respectively. There is no + overarching AeEnable control. + + For each of exposure and gain, we can model it with three states: auto, + locked, and manual. Note that AnalogueGainMode and ExposureTimeMode only + have two values, as the locked state is simulated. + + :: + + /---------------------------------\ + | | + V | + +--------+ +--------+ + | | ---------------------> | | + | Auto | +--------+ | Manual | + | | ----> | Locked | ----> | | + +--------+ +--------+ +--------+ + ^ | + | | + \----------------/ + + Notice from the state diagram that locked to manual is a one-way state + change, as the reverse direction is nonsensical (see the design document for + more details on this topic). + + The exposure/gain is in the Auto state when + ExposureTimeMode/AnalogueGainMode is set to Auto. In this state, the value + that is computed by the AE algorithm is applied to the image sensor. Any + value that is supplied in the ExposureTime/AnalogueGain control is ignored + and is not retained. + + If ExposureTimeMode/AnalogueGainMode is set to Manual, it can put us in + either the Locked or Manual state. The difference is in if + ExposureTime/AnalogueGain has been supplied. If it has not yet been + supplied, then we are in the Locked state. If it has been supplied, then we + are in the Manual state. + + In both the Locked state and the Manual state the exposure/gain value does + not come from the AE algorithm. In the Locked state the value comes from the + last value computed by the AE algorithm while the state was Auto, or if the + state was never Auto (e.g. we started in Locked, or the camera doesn't + support Auto), then the value should be a best-effort default value. In the + Manual state the value comes from the value supplied in the + ExposureTime/AnalogueGain control. + + To transition from the Locked state to the Manual state, a value needs to be + submitted in ExposureTime/AnalogueGain. Once the state has transitioned to + Manual, then this value will be retained, and doesn't need to be resubmitted + if it doesn't change. + + To transition to the Auto state, simply set + ExposureTimeMode/AnalogueGainMode to Auto. + + + The AeState metadata reports the state of the AE algorithm. As AE cannot + compute exposure and gain separately, the state of the AE component is + unified. There are three states: Idle, Searching, and Converged. + + The state shall be Idle if both ExposureTimeMode and AnalogueGainMode + are set to Manual. If the camera only supports one of the two controls, + then the state shall be Idle if that one control is set to Manual. If + the camera does not support Manual for at least one of the two controls, + then the state will never be Idle, as AE will always be running. + + The state shall be Searching if at least one of exposure or gain calculated + by the AE algorithm is used (that is, at least one of the two modes is Auto), + *and* the value(s) have not converged yet. + + The state shall be Converged if at least one of exposure or gain calculated + by the AE algorithm is used (that is, at least one of the two modes is Auto), + *and* the value(s) have converged.