Message ID | 20200730111134.641-3-david.plowman@raspberrypi.com |
---|---|
State | Superseded |
Headers | show |
Series |
|
Related | show |
Hi David, Thank you for the patch. On Thu, Jul 30, 2020 at 12:11:33PM +0100, David Plowman wrote: > This fixes a bug where the luminance correction table was not being > resampled according to the camera mode, in the same way as the colour > tables. This could be noticeable if any camera modes crop > aggressively. > > This resampling can be done "up front" in the SwitchMode, as we have > only a single fixed luminance table. In order to protect the > recalculation of the table from the asynchronous thread (which reads > it) I've elected to wait for that thread to go idle (though I doubt it > would have mattered much). As a by-product of stopping the thread, it > no longer needs its own copy of the camera mode (async_camera_mode_). This looks good to me. There are however a couple of potential issues I've noticed when reading the ALSC code and that you may want to check: - In Alsc::restartAsync(), async_start_ and async_started_ are set without holding the mutex_. They could thus potentially be reordered, possibly causing a race condition. - In Alsc::asyncFunc(), sync_signal_.notify_one() is called with the mutex_ held. The function will cause the thread waiting on the sync_signal_ condition variable to wake up, to then go back to sleep immediatelly when it tries to lock the mutex_. The notify_one() call should be moved after the locked scope. - sync_signal_ is actually never used :-) Or rather was never used, as this patch now waits on it in Alsc::waitForAysncThread(). The sync_signal_ in the AWB algorithm is however completely unused. > Signed-off-by: David Plowman <david.plowman@raspberrypi.com> For this patch, Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > --- > src/ipa/raspberrypi/controller/rpi/alsc.cpp | 39 ++++++++++++++------- > src/ipa/raspberrypi/controller/rpi/alsc.hpp | 3 +- > 2 files changed, 28 insertions(+), 14 deletions(-) > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.cpp b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > index 76e2f04..8f7720b 100644 > --- a/src/ipa/raspberrypi/controller/rpi/alsc.cpp > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > @@ -173,19 +173,34 @@ void Alsc::Initialise() > lambda_r_[i] = lambda_b_[i] = 1.0; > } > > +void Alsc::waitForAysncThread() > +{ > + std::unique_lock<std::mutex> lock(mutex_); > + if (async_started_) { > + sync_signal_.wait(lock, [&] { > + return async_finished_; > + }); > + async_started_ = false; > + async_finished_ = false; > + } > +} > + > void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > { > (void)metadata; > > + // Ensure the other thread isn't running while we do this. > + waitForAysncThread(); > + > // There's a bit of a question what we should do if the "crop" of the > - // camera mode has changed. Any calculation currently in flight would > - // not be useful to the new mode, so arguably we should abort it, and > - // generate a new table (like the "first_time" code already here). When > - // the crop doesn't change, we can presumably just leave things > - // alone. For now, I think we'll just wait and see. When the crop does > - // change, any effects should be transient, and if they're not transient > - // enough, we'll revisit the question then. > + // camera mode has changed. Should we effectively reset the algorithm > + // and start over? > camera_mode_ = camera_mode; > + > + // We must resample the luminance table like we do the others, but it's > + // fixed so we can simply do it up front here. > + resample_cal_table(config_.luminance_lut, camera_mode_, luminance_table_); > + > if (first_time_) { > // On the first time, arrange for something sensible in the > // initial tables. Construct the tables for some default colour > @@ -201,7 +216,7 @@ void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > compensate_lambdas_for_cal(cal_table_b, lambda_b_, > async_lambda_b_); > add_luminance_to_tables(sync_results_, async_lambda_r_, 1.0, > - async_lambda_b_, config_.luminance_lut, > + async_lambda_b_, luminance_table_, > config_.luminance_strength); > memcpy(prev_sync_results_, sync_results_, > sizeof(prev_sync_results_)); > @@ -266,8 +281,6 @@ void Alsc::restartAsync(StatisticsPtr &stats, Metadata *image_metadata) > } > copy_stats(statistics_, stats, alsc_status); > frame_phase_ = 0; > - // copy the camera mode so it won't change during the calculations > - async_camera_mode_ = camera_mode_; > async_start_ = true; > async_started_ = true; > async_signal_.notify_one(); > @@ -670,9 +683,9 @@ void Alsc::doAlsc() > // Fetch the new calibrations (if any) for this CT. Resample them in > // case the camera mode is not full-frame. > get_cal_table(ct_, config_.calibrations_Cr, cal_table_tmp); > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_r); > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_r); > get_cal_table(ct_, config_.calibrations_Cb, cal_table_tmp); > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_b); > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_b); > // You could print out the cal tables for this image here, if you're > // tuning the algorithm... > (void)print_cal_table; > @@ -695,7 +708,7 @@ void Alsc::doAlsc() > compensate_lambdas_for_cal(cal_table_b, lambda_b_, async_lambda_b_); > // Fold in the luminance table at the appropriate strength. > add_luminance_to_tables(async_results_, async_lambda_r_, 1.0, > - async_lambda_b_, config_.luminance_lut, > + async_lambda_b_, luminance_table_, > config_.luminance_strength); > } > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.hpp b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > index e895913..95572af 100644 > --- a/src/ipa/raspberrypi/controller/rpi/alsc.hpp > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > @@ -60,10 +60,10 @@ private: > AlscConfig config_; > bool first_time_; > CameraMode camera_mode_; > + double luminance_table_[ALSC_CELLS_X * ALSC_CELLS_Y]; > std::thread async_thread_; > void asyncFunc(); // asynchronous thread function > std::mutex mutex_; > - CameraMode async_camera_mode_; > // condvar for async thread to wait on > std::condition_variable async_signal_; > // condvar for synchronous thread to wait on > @@ -86,6 +86,7 @@ private: > int frame_count2_; > double sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > double prev_sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > + void waitForAysncThread(); > // The following are for the asynchronous thread to use, though the main > // thread can set/reset them if the async thread is known to be idle: > void restartAsync(StatisticsPtr &stats, Metadata *image_metadata);
Hi Laurent Thanks for the review! On Thu, 30 Jul 2020 at 23:57, Laurent Pinchart <laurent.pinchart@ideasonboard.com> wrote: > > Hi David, > > Thank you for the patch. > > On Thu, Jul 30, 2020 at 12:11:33PM +0100, David Plowman wrote: > > This fixes a bug where the luminance correction table was not being > > resampled according to the camera mode, in the same way as the colour > > tables. This could be noticeable if any camera modes crop > > aggressively. > > > > This resampling can be done "up front" in the SwitchMode, as we have > > only a single fixed luminance table. In order to protect the > > recalculation of the table from the asynchronous thread (which reads > > it) I've elected to wait for that thread to go idle (though I doubt it > > would have mattered much). As a by-product of stopping the thread, it > > no longer needs its own copy of the camera mode (async_camera_mode_). > > This looks good to me. There are however a couple of potential issues > I've noticed when reading the ALSC code and that you may want to check: > > - In Alsc::restartAsync(), async_start_ and async_started_ are set > without holding the mutex_. They could thus potentially be reordered, > possibly causing a race condition. async_started_ is used only by the main thread, so all we have to worry about is async_start_. Presumably this is safe assuming updating bools is atomic. But I have some recollection that nothing is guaranteed atomic if you don't mark it explicitly, so it probably wants wrapping in std::atomic, is that right? Same comments probably true about async_abort_. If that's atomic (and marked explicitly), no need for the lock in the destructor. > > - In Alsc::asyncFunc(), sync_signal_.notify_one() is called with the > mutex_ held. The function will cause the thread waiting on the > sync_signal_ condition variable to wake up, to then go back to sleep > immediatelly when it tries to lock the mutex_. The notify_one() call > should be moved after the locked scope. Here too, presumably the lock is in fact redundant if we believe the update to async_finished_ is atomic (though again, the correct thing would be to wrap it in std::atomic) > > - sync_signal_ is actually never used :-) Or rather was never used, as > this patch now waits on it in Alsc::waitForAysncThread(). The > sync_signal_ in the AWB algorithm is however completely unused. Yes indeed. All this code was ported from somewhere where you could force these algorithms to run synchronously (i.e. wait for the calculations to finish and use them immediately), so these other signals were indeed being used. I removed that for the libcamera/VC4 platform and never did anything about this signal. Though I'm quite happy to have left it in now! I'm not sure about the AWB one, I can't think of any case where waiting for that thread to stop would be useful like it is here, so maybe that will be one to remove. Thanks for all the comments! Best regards David > > > Signed-off-by: David Plowman <david.plowman@raspberrypi.com> > > For this patch, > > Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > > > --- > > src/ipa/raspberrypi/controller/rpi/alsc.cpp | 39 ++++++++++++++------- > > src/ipa/raspberrypi/controller/rpi/alsc.hpp | 3 +- > > 2 files changed, 28 insertions(+), 14 deletions(-) > > > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.cpp b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > index 76e2f04..8f7720b 100644 > > --- a/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > @@ -173,19 +173,34 @@ void Alsc::Initialise() > > lambda_r_[i] = lambda_b_[i] = 1.0; > > } > > > > +void Alsc::waitForAysncThread() > > +{ > > + std::unique_lock<std::mutex> lock(mutex_); > > + if (async_started_) { > > + sync_signal_.wait(lock, [&] { > > + return async_finished_; > > + }); > > + async_started_ = false; > > + async_finished_ = false; > > + } > > +} > > + > > void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > > { > > (void)metadata; > > > > + // Ensure the other thread isn't running while we do this. > > + waitForAysncThread(); > > + > > // There's a bit of a question what we should do if the "crop" of the > > - // camera mode has changed. Any calculation currently in flight would > > - // not be useful to the new mode, so arguably we should abort it, and > > - // generate a new table (like the "first_time" code already here). When > > - // the crop doesn't change, we can presumably just leave things > > - // alone. For now, I think we'll just wait and see. When the crop does > > - // change, any effects should be transient, and if they're not transient > > - // enough, we'll revisit the question then. > > + // camera mode has changed. Should we effectively reset the algorithm > > + // and start over? > > camera_mode_ = camera_mode; > > + > > + // We must resample the luminance table like we do the others, but it's > > + // fixed so we can simply do it up front here. > > + resample_cal_table(config_.luminance_lut, camera_mode_, luminance_table_); > > + > > if (first_time_) { > > // On the first time, arrange for something sensible in the > > // initial tables. Construct the tables for some default colour > > @@ -201,7 +216,7 @@ void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > > compensate_lambdas_for_cal(cal_table_b, lambda_b_, > > async_lambda_b_); > > add_luminance_to_tables(sync_results_, async_lambda_r_, 1.0, > > - async_lambda_b_, config_.luminance_lut, > > + async_lambda_b_, luminance_table_, > > config_.luminance_strength); > > memcpy(prev_sync_results_, sync_results_, > > sizeof(prev_sync_results_)); > > @@ -266,8 +281,6 @@ void Alsc::restartAsync(StatisticsPtr &stats, Metadata *image_metadata) > > } > > copy_stats(statistics_, stats, alsc_status); > > frame_phase_ = 0; > > - // copy the camera mode so it won't change during the calculations > > - async_camera_mode_ = camera_mode_; > > async_start_ = true; > > async_started_ = true; > > async_signal_.notify_one(); > > @@ -670,9 +683,9 @@ void Alsc::doAlsc() > > // Fetch the new calibrations (if any) for this CT. Resample them in > > // case the camera mode is not full-frame. > > get_cal_table(ct_, config_.calibrations_Cr, cal_table_tmp); > > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_r); > > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_r); > > get_cal_table(ct_, config_.calibrations_Cb, cal_table_tmp); > > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_b); > > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_b); > > // You could print out the cal tables for this image here, if you're > > // tuning the algorithm... > > (void)print_cal_table; > > @@ -695,7 +708,7 @@ void Alsc::doAlsc() > > compensate_lambdas_for_cal(cal_table_b, lambda_b_, async_lambda_b_); > > // Fold in the luminance table at the appropriate strength. > > add_luminance_to_tables(async_results_, async_lambda_r_, 1.0, > > - async_lambda_b_, config_.luminance_lut, > > + async_lambda_b_, luminance_table_, > > config_.luminance_strength); > > } > > > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.hpp b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > index e895913..95572af 100644 > > --- a/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > @@ -60,10 +60,10 @@ private: > > AlscConfig config_; > > bool first_time_; > > CameraMode camera_mode_; > > + double luminance_table_[ALSC_CELLS_X * ALSC_CELLS_Y]; > > std::thread async_thread_; > > void asyncFunc(); // asynchronous thread function > > std::mutex mutex_; > > - CameraMode async_camera_mode_; > > // condvar for async thread to wait on > > std::condition_variable async_signal_; > > // condvar for synchronous thread to wait on > > @@ -86,6 +86,7 @@ private: > > int frame_count2_; > > double sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > > double prev_sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > > + void waitForAysncThread(); > > // The following are for the asynchronous thread to use, though the main > > // thread can set/reset them if the async thread is known to be idle: > > void restartAsync(StatisticsPtr &stats, Metadata *image_metadata); > > -- > Regards, > > Laurent Pinchart
Hi David, On Fri, Jul 31, 2020 at 09:03:35AM +0100, David Plowman wrote: > On Thu, 30 Jul 2020 at 23:57, Laurent Pinchart wrote: > > On Thu, Jul 30, 2020 at 12:11:33PM +0100, David Plowman wrote: > > > This fixes a bug where the luminance correction table was not being > > > resampled according to the camera mode, in the same way as the colour > > > tables. This could be noticeable if any camera modes crop > > > aggressively. > > > > > > This resampling can be done "up front" in the SwitchMode, as we have > > > only a single fixed luminance table. In order to protect the > > > recalculation of the table from the asynchronous thread (which reads > > > it) I've elected to wait for that thread to go idle (though I doubt it > > > would have mattered much). As a by-product of stopping the thread, it > > > no longer needs its own copy of the camera mode (async_camera_mode_). > > > > This looks good to me. There are however a couple of potential issues > > I've noticed when reading the ALSC code and that you may want to check: > > > > - In Alsc::restartAsync(), async_start_ and async_started_ are set > > without holding the mutex_. They could thus potentially be reordered, > > possibly causing a race condition. > > async_started_ is used only by the main thread, so all we have to > worry about is async_start_. Oops, indeed, my bad. > Presumably this is safe assuming updating bools is atomic. But I have > some recollection that nothing is guaranteed atomic if you don't mark > it explicitly, so it probably wants wrapping in std::atomic, is that > right? If it's used in a single thread there's no issue, atomicity isn't required. > Same comments probably true about async_abort_. If that's atomic (and > marked explicitly), no need for the lock in the destructor. async_abort_ is used in the ALSC thread, so for that we need a lock. It's not just about atomicity, it's also about ensuring that memory accesses are not reordered. If you write async_abort_ = true; async_signal_.notify_one(); without taking a lock, there is no guarantee that the async_abort_ write will be visible to the reader before it gets woken up by async_signal_.notify_one(). The documentation of std::condition_variable::notify_one() states "The effects of notify_one()/notify_all() and each of the three atomic parts of wait()/wait_for()/wait_until() (unlock+wait, wakeup, and lock) take place in a single total order that can be viewed as modification order of an atomic variable: the order is specific to this individual condition_variable. This makes it impossible for notify_one() to, for example, be delayed and unblock a thread that started waiting just after the call to notify_one() was made." This doesn't offer any memory order guarantee (as in https://en.cppreference.com/w/cpp/atomic/memory_order) for the modification of the shared variable. The std::condition_variable documentation state "Even if the shared variable is atomic, it must be modified under the mutex in order to correctly publish the modification to the waiting thread." The lock is thus needed, but the notify_one() call can be moved after unlocking the mutex. > > - In Alsc::asyncFunc(), sync_signal_.notify_one() is called with the > > mutex_ held. The function will cause the thread waiting on the > > sync_signal_ condition variable to wake up, to then go back to sleep > > immediatelly when it tries to lock the mutex_. The notify_one() call > > should be moved after the locked scope. > > Here too, presumably the lock is in fact redundant if we believe the > update to async_finished_ is atomic (though again, the correct thing > would be to wrap it in std::atomic) See above :-) Memory ordering is a tricky topic. > > - sync_signal_ is actually never used :-) Or rather was never used, as > > this patch now waits on it in Alsc::waitForAysncThread(). The > > sync_signal_ in the AWB algorithm is however completely unused. > > Yes indeed. All this code was ported from somewhere where you could > force these algorithms to run synchronously (i.e. wait for the > calculations to finish and use them immediately), so these other > signals were indeed being used. I removed that for the libcamera/VC4 > platform and never did anything about this signal. Though I'm quite > happy to have left it in now! I'm not sure about the AWB one, I can't > think of any case where waiting for that thread to stop would be > useful like it is here, so maybe that will be one to remove. > > Thanks for all the comments! > > > > Signed-off-by: David Plowman <david.plowman@raspberrypi.com> > > > > For this patch, > > > > Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > > > > > --- > > > src/ipa/raspberrypi/controller/rpi/alsc.cpp | 39 ++++++++++++++------- > > > src/ipa/raspberrypi/controller/rpi/alsc.hpp | 3 +- > > > 2 files changed, 28 insertions(+), 14 deletions(-) > > > > > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.cpp b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > > index 76e2f04..8f7720b 100644 > > > --- a/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > > @@ -173,19 +173,34 @@ void Alsc::Initialise() > > > lambda_r_[i] = lambda_b_[i] = 1.0; > > > } > > > > > > +void Alsc::waitForAysncThread() > > > +{ > > > + std::unique_lock<std::mutex> lock(mutex_); > > > + if (async_started_) { > > > + sync_signal_.wait(lock, [&] { > > > + return async_finished_; > > > + }); > > > + async_started_ = false; > > > + async_finished_ = false; > > > + } > > > +} > > > + > > > void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > > > { > > > (void)metadata; > > > > > > + // Ensure the other thread isn't running while we do this. > > > + waitForAysncThread(); > > > + > > > // There's a bit of a question what we should do if the "crop" of the > > > - // camera mode has changed. Any calculation currently in flight would > > > - // not be useful to the new mode, so arguably we should abort it, and > > > - // generate a new table (like the "first_time" code already here). When > > > - // the crop doesn't change, we can presumably just leave things > > > - // alone. For now, I think we'll just wait and see. When the crop does > > > - // change, any effects should be transient, and if they're not transient > > > - // enough, we'll revisit the question then. > > > + // camera mode has changed. Should we effectively reset the algorithm > > > + // and start over? > > > camera_mode_ = camera_mode; > > > + > > > + // We must resample the luminance table like we do the others, but it's > > > + // fixed so we can simply do it up front here. > > > + resample_cal_table(config_.luminance_lut, camera_mode_, luminance_table_); > > > + > > > if (first_time_) { > > > // On the first time, arrange for something sensible in the > > > // initial tables. Construct the tables for some default colour > > > @@ -201,7 +216,7 @@ void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > > > compensate_lambdas_for_cal(cal_table_b, lambda_b_, > > > async_lambda_b_); > > > add_luminance_to_tables(sync_results_, async_lambda_r_, 1.0, > > > - async_lambda_b_, config_.luminance_lut, > > > + async_lambda_b_, luminance_table_, > > > config_.luminance_strength); > > > memcpy(prev_sync_results_, sync_results_, > > > sizeof(prev_sync_results_)); > > > @@ -266,8 +281,6 @@ void Alsc::restartAsync(StatisticsPtr &stats, Metadata *image_metadata) > > > } > > > copy_stats(statistics_, stats, alsc_status); > > > frame_phase_ = 0; > > > - // copy the camera mode so it won't change during the calculations > > > - async_camera_mode_ = camera_mode_; > > > async_start_ = true; > > > async_started_ = true; > > > async_signal_.notify_one(); > > > @@ -670,9 +683,9 @@ void Alsc::doAlsc() > > > // Fetch the new calibrations (if any) for this CT. Resample them in > > > // case the camera mode is not full-frame. > > > get_cal_table(ct_, config_.calibrations_Cr, cal_table_tmp); > > > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_r); > > > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_r); > > > get_cal_table(ct_, config_.calibrations_Cb, cal_table_tmp); > > > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_b); > > > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_b); > > > // You could print out the cal tables for this image here, if you're > > > // tuning the algorithm... > > > (void)print_cal_table; > > > @@ -695,7 +708,7 @@ void Alsc::doAlsc() > > > compensate_lambdas_for_cal(cal_table_b, lambda_b_, async_lambda_b_); > > > // Fold in the luminance table at the appropriate strength. > > > add_luminance_to_tables(async_results_, async_lambda_r_, 1.0, > > > - async_lambda_b_, config_.luminance_lut, > > > + async_lambda_b_, luminance_table_, > > > config_.luminance_strength); > > > } > > > > > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.hpp b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > > index e895913..95572af 100644 > > > --- a/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > > @@ -60,10 +60,10 @@ private: > > > AlscConfig config_; > > > bool first_time_; > > > CameraMode camera_mode_; > > > + double luminance_table_[ALSC_CELLS_X * ALSC_CELLS_Y]; > > > std::thread async_thread_; > > > void asyncFunc(); // asynchronous thread function > > > std::mutex mutex_; > > > - CameraMode async_camera_mode_; > > > // condvar for async thread to wait on > > > std::condition_variable async_signal_; > > > // condvar for synchronous thread to wait on > > > @@ -86,6 +86,7 @@ private: > > > int frame_count2_; > > > double sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > > > double prev_sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > > > + void waitForAysncThread(); > > > // The following are for the asynchronous thread to use, though the main > > > // thread can set/reset them if the async thread is known to be idle: > > > void restartAsync(StatisticsPtr &stats, Metadata *image_metadata);
Hi again Thanks for that. Let me circulate a second version of these patches, and we can also wait for Naush to review next week. David On Fri, 31 Jul 2020 at 13:01, Laurent Pinchart <laurent.pinchart@ideasonboard.com> wrote: > > Hi David, > > On Fri, Jul 31, 2020 at 09:03:35AM +0100, David Plowman wrote: > > On Thu, 30 Jul 2020 at 23:57, Laurent Pinchart wrote: > > > On Thu, Jul 30, 2020 at 12:11:33PM +0100, David Plowman wrote: > > > > This fixes a bug where the luminance correction table was not being > > > > resampled according to the camera mode, in the same way as the colour > > > > tables. This could be noticeable if any camera modes crop > > > > aggressively. > > > > > > > > This resampling can be done "up front" in the SwitchMode, as we have > > > > only a single fixed luminance table. In order to protect the > > > > recalculation of the table from the asynchronous thread (which reads > > > > it) I've elected to wait for that thread to go idle (though I doubt it > > > > would have mattered much). As a by-product of stopping the thread, it > > > > no longer needs its own copy of the camera mode (async_camera_mode_). > > > > > > This looks good to me. There are however a couple of potential issues > > > I've noticed when reading the ALSC code and that you may want to check: > > > > > > - In Alsc::restartAsync(), async_start_ and async_started_ are set > > > without holding the mutex_. They could thus potentially be reordered, > > > possibly causing a race condition. > > > > async_started_ is used only by the main thread, so all we have to > > worry about is async_start_. > > Oops, indeed, my bad. > > > Presumably this is safe assuming updating bools is atomic. But I have > > some recollection that nothing is guaranteed atomic if you don't mark > > it explicitly, so it probably wants wrapping in std::atomic, is that > > right? > > If it's used in a single thread there's no issue, atomicity isn't > required. > > > Same comments probably true about async_abort_. If that's atomic (and > > marked explicitly), no need for the lock in the destructor. > > async_abort_ is used in the ALSC thread, so for that we need a lock. > It's not just about atomicity, it's also about ensuring that memory > accesses are not reordered. If you write > > async_abort_ = true; > async_signal_.notify_one(); > > without taking a lock, there is no guarantee that the async_abort_ write > will be visible to the reader before it gets woken up by > async_signal_.notify_one(). > > The documentation of std::condition_variable::notify_one() states > > "The effects of notify_one()/notify_all() and each of the three atomic > parts of wait()/wait_for()/wait_until() (unlock+wait, wakeup, and lock) > take place in a single total order that can be viewed as modification > order of an atomic variable: the order is specific to this individual > condition_variable. This makes it impossible for notify_one() to, for > example, be delayed and unblock a thread that started waiting just after > the call to notify_one() was made." > > This doesn't offer any memory order guarantee (as in > https://en.cppreference.com/w/cpp/atomic/memory_order) for the > modification of the shared variable. > > The std::condition_variable documentation state > > "Even if the shared variable is atomic, it must be modified under the > mutex in order to correctly publish the modification to the waiting > thread." > > The lock is thus needed, but the notify_one() call can be moved after > unlocking the mutex. > > > > - In Alsc::asyncFunc(), sync_signal_.notify_one() is called with the > > > mutex_ held. The function will cause the thread waiting on the > > > sync_signal_ condition variable to wake up, to then go back to sleep > > > immediatelly when it tries to lock the mutex_. The notify_one() call > > > should be moved after the locked scope. > > > > Here too, presumably the lock is in fact redundant if we believe the > > update to async_finished_ is atomic (though again, the correct thing > > would be to wrap it in std::atomic) > > See above :-) Memory ordering is a tricky topic. > > > > - sync_signal_ is actually never used :-) Or rather was never used, as > > > this patch now waits on it in Alsc::waitForAysncThread(). The > > > sync_signal_ in the AWB algorithm is however completely unused. > > > > Yes indeed. All this code was ported from somewhere where you could > > force these algorithms to run synchronously (i.e. wait for the > > calculations to finish and use them immediately), so these other > > signals were indeed being used. I removed that for the libcamera/VC4 > > platform and never did anything about this signal. Though I'm quite > > happy to have left it in now! I'm not sure about the AWB one, I can't > > think of any case where waiting for that thread to stop would be > > useful like it is here, so maybe that will be one to remove. > > > > Thanks for all the comments! > > > > > > Signed-off-by: David Plowman <david.plowman@raspberrypi.com> > > > > > > For this patch, > > > > > > Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > > > > > > > --- > > > > src/ipa/raspberrypi/controller/rpi/alsc.cpp | 39 ++++++++++++++------- > > > > src/ipa/raspberrypi/controller/rpi/alsc.hpp | 3 +- > > > > 2 files changed, 28 insertions(+), 14 deletions(-) > > > > > > > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.cpp b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > > > index 76e2f04..8f7720b 100644 > > > > --- a/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > > > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.cpp > > > > @@ -173,19 +173,34 @@ void Alsc::Initialise() > > > > lambda_r_[i] = lambda_b_[i] = 1.0; > > > > } > > > > > > > > +void Alsc::waitForAysncThread() > > > > +{ > > > > + std::unique_lock<std::mutex> lock(mutex_); > > > > + if (async_started_) { > > > > + sync_signal_.wait(lock, [&] { > > > > + return async_finished_; > > > > + }); > > > > + async_started_ = false; > > > > + async_finished_ = false; > > > > + } > > > > +} > > > > + > > > > void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > > > > { > > > > (void)metadata; > > > > > > > > + // Ensure the other thread isn't running while we do this. > > > > + waitForAysncThread(); > > > > + > > > > // There's a bit of a question what we should do if the "crop" of the > > > > - // camera mode has changed. Any calculation currently in flight would > > > > - // not be useful to the new mode, so arguably we should abort it, and > > > > - // generate a new table (like the "first_time" code already here). When > > > > - // the crop doesn't change, we can presumably just leave things > > > > - // alone. For now, I think we'll just wait and see. When the crop does > > > > - // change, any effects should be transient, and if they're not transient > > > > - // enough, we'll revisit the question then. > > > > + // camera mode has changed. Should we effectively reset the algorithm > > > > + // and start over? > > > > camera_mode_ = camera_mode; > > > > + > > > > + // We must resample the luminance table like we do the others, but it's > > > > + // fixed so we can simply do it up front here. > > > > + resample_cal_table(config_.luminance_lut, camera_mode_, luminance_table_); > > > > + > > > > if (first_time_) { > > > > // On the first time, arrange for something sensible in the > > > > // initial tables. Construct the tables for some default colour > > > > @@ -201,7 +216,7 @@ void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) > > > > compensate_lambdas_for_cal(cal_table_b, lambda_b_, > > > > async_lambda_b_); > > > > add_luminance_to_tables(sync_results_, async_lambda_r_, 1.0, > > > > - async_lambda_b_, config_.luminance_lut, > > > > + async_lambda_b_, luminance_table_, > > > > config_.luminance_strength); > > > > memcpy(prev_sync_results_, sync_results_, > > > > sizeof(prev_sync_results_)); > > > > @@ -266,8 +281,6 @@ void Alsc::restartAsync(StatisticsPtr &stats, Metadata *image_metadata) > > > > } > > > > copy_stats(statistics_, stats, alsc_status); > > > > frame_phase_ = 0; > > > > - // copy the camera mode so it won't change during the calculations > > > > - async_camera_mode_ = camera_mode_; > > > > async_start_ = true; > > > > async_started_ = true; > > > > async_signal_.notify_one(); > > > > @@ -670,9 +683,9 @@ void Alsc::doAlsc() > > > > // Fetch the new calibrations (if any) for this CT. Resample them in > > > > // case the camera mode is not full-frame. > > > > get_cal_table(ct_, config_.calibrations_Cr, cal_table_tmp); > > > > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_r); > > > > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_r); > > > > get_cal_table(ct_, config_.calibrations_Cb, cal_table_tmp); > > > > - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_b); > > > > + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_b); > > > > // You could print out the cal tables for this image here, if you're > > > > // tuning the algorithm... > > > > (void)print_cal_table; > > > > @@ -695,7 +708,7 @@ void Alsc::doAlsc() > > > > compensate_lambdas_for_cal(cal_table_b, lambda_b_, async_lambda_b_); > > > > // Fold in the luminance table at the appropriate strength. > > > > add_luminance_to_tables(async_results_, async_lambda_r_, 1.0, > > > > - async_lambda_b_, config_.luminance_lut, > > > > + async_lambda_b_, luminance_table_, > > > > config_.luminance_strength); > > > > } > > > > > > > > diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.hpp b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > > > index e895913..95572af 100644 > > > > --- a/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > > > +++ b/src/ipa/raspberrypi/controller/rpi/alsc.hpp > > > > @@ -60,10 +60,10 @@ private: > > > > AlscConfig config_; > > > > bool first_time_; > > > > CameraMode camera_mode_; > > > > + double luminance_table_[ALSC_CELLS_X * ALSC_CELLS_Y]; > > > > std::thread async_thread_; > > > > void asyncFunc(); // asynchronous thread function > > > > std::mutex mutex_; > > > > - CameraMode async_camera_mode_; > > > > // condvar for async thread to wait on > > > > std::condition_variable async_signal_; > > > > // condvar for synchronous thread to wait on > > > > @@ -86,6 +86,7 @@ private: > > > > int frame_count2_; > > > > double sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > > > > double prev_sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; > > > > + void waitForAysncThread(); > > > > // The following are for the asynchronous thread to use, though the main > > > > // thread can set/reset them if the async thread is known to be idle: > > > > void restartAsync(StatisticsPtr &stats, Metadata *image_metadata); > > -- > Regards, > > Laurent Pinchart
diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.cpp b/src/ipa/raspberrypi/controller/rpi/alsc.cpp index 76e2f04..8f7720b 100644 --- a/src/ipa/raspberrypi/controller/rpi/alsc.cpp +++ b/src/ipa/raspberrypi/controller/rpi/alsc.cpp @@ -173,19 +173,34 @@ void Alsc::Initialise() lambda_r_[i] = lambda_b_[i] = 1.0; } +void Alsc::waitForAysncThread() +{ + std::unique_lock<std::mutex> lock(mutex_); + if (async_started_) { + sync_signal_.wait(lock, [&] { + return async_finished_; + }); + async_started_ = false; + async_finished_ = false; + } +} + void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) { (void)metadata; + // Ensure the other thread isn't running while we do this. + waitForAysncThread(); + // There's a bit of a question what we should do if the "crop" of the - // camera mode has changed. Any calculation currently in flight would - // not be useful to the new mode, so arguably we should abort it, and - // generate a new table (like the "first_time" code already here). When - // the crop doesn't change, we can presumably just leave things - // alone. For now, I think we'll just wait and see. When the crop does - // change, any effects should be transient, and if they're not transient - // enough, we'll revisit the question then. + // camera mode has changed. Should we effectively reset the algorithm + // and start over? camera_mode_ = camera_mode; + + // We must resample the luminance table like we do the others, but it's + // fixed so we can simply do it up front here. + resample_cal_table(config_.luminance_lut, camera_mode_, luminance_table_); + if (first_time_) { // On the first time, arrange for something sensible in the // initial tables. Construct the tables for some default colour @@ -201,7 +216,7 @@ void Alsc::SwitchMode(CameraMode const &camera_mode, Metadata *metadata) compensate_lambdas_for_cal(cal_table_b, lambda_b_, async_lambda_b_); add_luminance_to_tables(sync_results_, async_lambda_r_, 1.0, - async_lambda_b_, config_.luminance_lut, + async_lambda_b_, luminance_table_, config_.luminance_strength); memcpy(prev_sync_results_, sync_results_, sizeof(prev_sync_results_)); @@ -266,8 +281,6 @@ void Alsc::restartAsync(StatisticsPtr &stats, Metadata *image_metadata) } copy_stats(statistics_, stats, alsc_status); frame_phase_ = 0; - // copy the camera mode so it won't change during the calculations - async_camera_mode_ = camera_mode_; async_start_ = true; async_started_ = true; async_signal_.notify_one(); @@ -670,9 +683,9 @@ void Alsc::doAlsc() // Fetch the new calibrations (if any) for this CT. Resample them in // case the camera mode is not full-frame. get_cal_table(ct_, config_.calibrations_Cr, cal_table_tmp); - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_r); + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_r); get_cal_table(ct_, config_.calibrations_Cb, cal_table_tmp); - resample_cal_table(cal_table_tmp, async_camera_mode_, cal_table_b); + resample_cal_table(cal_table_tmp, camera_mode_, cal_table_b); // You could print out the cal tables for this image here, if you're // tuning the algorithm... (void)print_cal_table; @@ -695,7 +708,7 @@ void Alsc::doAlsc() compensate_lambdas_for_cal(cal_table_b, lambda_b_, async_lambda_b_); // Fold in the luminance table at the appropriate strength. add_luminance_to_tables(async_results_, async_lambda_r_, 1.0, - async_lambda_b_, config_.luminance_lut, + async_lambda_b_, luminance_table_, config_.luminance_strength); } diff --git a/src/ipa/raspberrypi/controller/rpi/alsc.hpp b/src/ipa/raspberrypi/controller/rpi/alsc.hpp index e895913..95572af 100644 --- a/src/ipa/raspberrypi/controller/rpi/alsc.hpp +++ b/src/ipa/raspberrypi/controller/rpi/alsc.hpp @@ -60,10 +60,10 @@ private: AlscConfig config_; bool first_time_; CameraMode camera_mode_; + double luminance_table_[ALSC_CELLS_X * ALSC_CELLS_Y]; std::thread async_thread_; void asyncFunc(); // asynchronous thread function std::mutex mutex_; - CameraMode async_camera_mode_; // condvar for async thread to wait on std::condition_variable async_signal_; // condvar for synchronous thread to wait on @@ -86,6 +86,7 @@ private: int frame_count2_; double sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; double prev_sync_results_[3][ALSC_CELLS_Y][ALSC_CELLS_X]; + void waitForAysncThread(); // The following are for the asynchronous thread to use, though the main // thread can set/reset them if the async thread is known to be idle: void restartAsync(StatisticsPtr &stats, Metadata *image_metadata);
This fixes a bug where the luminance correction table was not being resampled according to the camera mode, in the same way as the colour tables. This could be noticeable if any camera modes crop aggressively. This resampling can be done "up front" in the SwitchMode, as we have only a single fixed luminance table. In order to protect the recalculation of the table from the asynchronous thread (which reads it) I've elected to wait for that thread to go idle (though I doubt it would have mattered much). As a by-product of stopping the thread, it no longer needs its own copy of the camera mode (async_camera_mode_). Signed-off-by: David Plowman <david.plowman@raspberrypi.com> --- src/ipa/raspberrypi/controller/rpi/alsc.cpp | 39 ++++++++++++++------- src/ipa/raspberrypi/controller/rpi/alsc.hpp | 3 +- 2 files changed, 28 insertions(+), 14 deletions(-)