/* SPDX-License-Identifier: LGPL-2.1-or-later */ /* * Copyright (C) 2019, Google Inc. * * rkisp1.cpp - RkISP1 Image Processing Algorithms */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "libcamera/internal/log.h" namespace libcamera { LOG_DEFINE_CATEGORY(IPARkISP1) namespace ipa::rkisp1 { class IPARkISP1 : public IPARkISP1Interface { public: int init(unsigned int hwRevision) override; int start() override; void stop() override {} int configure(const CameraSensorInfo &info, const std::map &streamConfig, const std::map &entityControls) override; void mapBuffers(const std::vector &buffers) override; void unmapBuffers(const std::vector &ids) override; void processEvent(const RkISP1Event &event) override; private: void queueRequest(unsigned int frame, rkisp1_params_cfg *params, const ControlList &controls); void updateStatistics(unsigned int frame, const rkisp1_stat_buffer *stats); void setControls(unsigned int frame); void metadataReady(unsigned int frame, unsigned int aeState); std::map buffers_; std::map buffersMemory_; ControlInfoMap ctrls_; /* Camera sensor controls. */ bool autoExposure_; uint32_t exposure_; uint32_t minExposure_; uint32_t maxExposure_; uint32_t gain_; uint32_t minGain_; uint32_t maxGain_; }; int IPARkISP1::init(unsigned int hwRevision) { /* \todo Add support for other revisions */ if (hwRevision != RKISP1_V10) { LOG(IPARkISP1, Error) << "Hardware revision " << hwRevision << " is currently not supported"; return -ENODEV; } LOG(IPARkISP1, Debug) << "Hardware revision is " << hwRevision; return 0; } int IPARkISP1::start() { setControls(0); return 0; } /** * \todo The RkISP1 pipeline currently provides an empty CameraSensorInfo * if the connected sensor does not provide enough information to properly * assemble one. Make sure the reported sensor information are relevant * before accessing them. */ int IPARkISP1::configure([[maybe_unused]] const CameraSensorInfo &info, [[maybe_unused]] const std::map &streamConfig, const std::map &entityControls) { if (entityControls.empty()) return -EINVAL; ctrls_ = entityControls.at(0); const auto itExp = ctrls_.find(V4L2_CID_EXPOSURE); if (itExp == ctrls_.end()) { LOG(IPARkISP1, Error) << "Can't find exposure control"; return -EINVAL; } const auto itGain = ctrls_.find(V4L2_CID_ANALOGUE_GAIN); if (itGain == ctrls_.end()) { LOG(IPARkISP1, Error) << "Can't find gain control"; return -EINVAL; } autoExposure_ = true; minExposure_ = std::max(itExp->second.min().get(), 1); maxExposure_ = itExp->second.max().get(); exposure_ = minExposure_; minGain_ = std::max(itGain->second.min().get(), 1); maxGain_ = itGain->second.max().get(); gain_ = minGain_; LOG(IPARkISP1, Info) << "Exposure: " << minExposure_ << "-" << maxExposure_ << " Gain: " << minGain_ << "-" << maxGain_; return 0; } void IPARkISP1::mapBuffers(const std::vector &buffers) { for (const IPABuffer &buffer : buffers) { auto elem = buffers_.emplace(std::piecewise_construct, std::forward_as_tuple(buffer.id), std::forward_as_tuple(buffer.planes)); const FrameBuffer &fb = elem.first->second; /* * \todo Provide a helper to mmap() buffers (possibly exposed * to applications). */ buffersMemory_[buffer.id] = mmap(NULL, fb.planes()[0].length, PROT_READ | PROT_WRITE, MAP_SHARED, fb.planes()[0].fd.fd(), 0); if (buffersMemory_[buffer.id] == MAP_FAILED) { int ret = -errno; LOG(IPARkISP1, Fatal) << "Failed to mmap buffer: " << strerror(-ret); } } } void IPARkISP1::unmapBuffers(const std::vector &ids) { for (unsigned int id : ids) { const auto fb = buffers_.find(id); if (fb == buffers_.end()) continue; munmap(buffersMemory_[id], fb->second.planes()[0].length); buffersMemory_.erase(id); buffers_.erase(id); } } void IPARkISP1::processEvent(const RkISP1Event &event) { switch (event.op) { case EventSignalStatBuffer: { unsigned int frame = event.frame; unsigned int bufferId = event.bufferId; const rkisp1_stat_buffer *stats = static_cast(buffersMemory_[bufferId]); updateStatistics(frame, stats); break; } case EventQueueRequest: { unsigned int frame = event.frame; unsigned int bufferId = event.bufferId; rkisp1_params_cfg *params = static_cast(buffersMemory_[bufferId]); queueRequest(frame, params, event.controls); break; } default: LOG(IPARkISP1, Error) << "Unknown event " << event.op; break; } } void IPARkISP1::queueRequest(unsigned int frame, rkisp1_params_cfg *params, const ControlList &controls) { /* Prepare parameters buffer. */ memset(params, 0, sizeof(*params)); /* Auto Exposure on/off. */ if (controls.contains(controls::AeEnable)) { autoExposure_ = controls.get(controls::AeEnable); if (autoExposure_) params->module_ens = RKISP1_CIF_ISP_MODULE_AEC; params->module_en_update = RKISP1_CIF_ISP_MODULE_AEC; } RkISP1Action op; op.op = ActionParamFilled; queueFrameAction.emit(frame, op); } void IPARkISP1::updateStatistics(unsigned int frame, const rkisp1_stat_buffer *stats) { const rkisp1_cif_isp_stat *params = &stats->params; unsigned int aeState = 0; if (stats->meas_type & RKISP1_CIF_ISP_STAT_AUTOEXP) { const rkisp1_cif_isp_ae_stat *ae = ¶ms->ae; const unsigned int target = 60; unsigned int value = 0; unsigned int num = 0; for (int i = 0; i < RKISP1_CIF_ISP_AE_MEAN_MAX_V10; i++) { if (ae->exp_mean[i] <= 15) continue; value += ae->exp_mean[i]; num++; } value /= num; double factor = (double)target / value; if (frame % 3 == 0) { double exposure; exposure = factor * exposure_ * gain_ / minGain_; exposure_ = std::clamp((uint64_t)exposure, minExposure_, maxExposure_); exposure = exposure / exposure_ * minGain_; gain_ = std::clamp((uint64_t)exposure, minGain_, maxGain_); setControls(frame + 1); } aeState = fabs(factor - 1.0f) < 0.05f ? 2 : 1; } metadataReady(frame, aeState); } void IPARkISP1::setControls(unsigned int frame) { RkISP1Action op; op.op = ActionV4L2Set; ControlList ctrls(ctrls_); ctrls.set(V4L2_CID_EXPOSURE, static_cast(exposure_)); ctrls.set(V4L2_CID_ANALOGUE_GAIN, static_cast(gain_)); op.controls = ctrls; queueFrameAction.emit(frame, op); } void IPARkISP1::metadataReady(unsigned int frame, unsigned int aeState) { ControlList ctrls(controls::controls); if (aeState) ctrls.set(controls::AeLocked, aeState == 2); RkISP1Action op; op.op = ActionMetadata; op.controls = ctrls; queueFrameAction.emit(frame, op); } } /* namespace ipa::rkisp1 */ /* * External IPA module interface */ extern "C" { const struct IPAModuleInfo ipaModuleInfo = { IPA_MODULE_API_VERSION, 1, "PipelineHandlerRkISP1", "rkisp1", }; IPAInterface *ipaCreate() { return new ipa::rkisp1::IPARkISP1(); } } } /* namespace libcamera */ 'n61' href='#n61'>61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647

/* SPDX-License-Identifier: BSD-2-Clause */
/*
 * Copyright (C) 2019, Raspberry Pi (Trading) Limited
 *
 * awb.cpp - AWB control algorithm
 */

#include "libcamera/internal/log.h"

#include "../lux_status.h"

#include "awb.hpp"

using namespace RPiController;
using namespace libcamera;

LOG_DEFINE_CATEGORY(RPiAwb)

#define NAME "rpi.awb"

#define AWB_STATS_SIZE_X DEFAULT_AWB_REGIONS_X
#define AWB_STATS_SIZE_Y DEFAULT_AWB_REGIONS_Y

// todo - the locking in this algorithm needs some tidying up as has been done
// elsewhere (ALSC and AGC).

void AwbMode::Read(boost::property_tree::ptree const &params)
{
	ct_lo = params.get<double>("lo");
	ct_hi = params.get<double>("hi");
}

void AwbPrior::Read(boost::property_tree::ptree const &params)
{
	lux = params.get<double>("lux");
	prior.Read(params.get_child("prior"));
}

static void read_ct_curve(Pwl &ct_r, Pwl &ct_b,
			  boost::property_tree::ptree const &params)
{
	int num = 0;
	for (auto it = params.begin(); it != params.end(); it++) {
		double ct = it->second.get_value<double>();
		assert(it == params.begin() || ct != ct_r.Domain().end);
		if (++it == params.end())
			throw std::runtime_error(
				"AwbConfig: incomplete CT curve entry");
		ct_r.Append(ct, it->second.get_value<double>());
		if (++it == params.end())
			throw std::runtime_error(
				"AwbConfig: incomplete CT curve entry");
		ct_b.Append(ct, it->second.get_value<double>());
		num++;
	}
	if (num < 2)
		throw std::runtime_error(
			"AwbConfig: insufficient points in CT curve");
}

void AwbConfig::Read(boost::property_tree::ptree const &params)
{
	bayes = params.get<int>("bayes", 1);
	frame_period = params.get<uint16_t>("frame_period", 10);
	startup_frames = params.get<uint16_t>("startup_frames", 10);
	convergence_frames = params.get<unsigned int>("convergence_frames", 3);
	speed = params.get<double>("speed", 0.05);
	if (params.get_child_optional("ct_curve"))
		read_ct_curve(ct_r, ct_b, params.get_child("ct_curve"));
	if (params.get_child_optional("priors")) {
		for (auto &p : params.get_child("priors")) {
			AwbPrior prior;
			prior.Read(p.second);
			if (!priors.empty() && prior.lux <= priors.back().lux)
				throw std::runtime_error(
					"AwbConfig: Prior must be ordered in increasing lux value");
			priors.push_back(prior);
		}
		if (priors.empty())
			throw std::runtime_error(
				"AwbConfig: no AWB priors configured");
	}
	if (params.get_child_optional("modes")) {
		for (auto &p : params.get_child("modes")) {
			modes[p.first].Read(p.second);
			if (default_mode == nullptr)
				default_mode = &modes[p.first];
		}
		if (default_mode == nullptr)
			throw std::runtime_error(
				"AwbConfig: no AWB modes configured");
	}
	min_pixels = params.get<double>("min_pixels", 16.0);
	min_G = params.get<uint16_t>("min_G", 32);
	min_regions = params.get<uint32_t>("min_regions", 10);
	delta_limit = params.get<double>("delta_limit", 0.2);
	coarse_step = params.get<double>("coarse_step", 0.2);
	transverse_pos = params.get<double>("transverse_pos", 0.01);
	transverse_neg = params.get<double>("transverse_neg", 0.01);
	if (transverse_pos <= 0 || transverse_neg <= 0)
		throw std::runtime_error(
			"AwbConfig: transverse_pos/neg must be > 0");
	sensitivity_r = params.get<double>("sensitivity_r", 1.0);
	sensitivity_b = params.get<double>("sensitivity_b", 1.0);
	if (bayes) {
		if (ct_r.Empty() || ct_b.Empty() || priors.empty() ||
		    default_mode == nullptr) {
			LOG(RPiAwb, Warning)
				<< "Bayesian AWB mis-configured - switch to Grey method";
			bayes = false;
		}
	}
	fast = params.get<int>(
		"fast", bayes); // default to fast for Bayesian, otherwise slow
	whitepoint_r = params.get<double>("whitepoint_r", 0.0);
	whitepoint_b = params.get<double>("whitepoint_b", 0.0);
	if (bayes == false)
		sensitivity_r = sensitivity_b =
			1.0; // nor do sensitivities make any sense
}

Awb::Awb(Controller *controller)
	: AwbAlgorithm(controller)
{
	async_abort_ = async_start_ = async_started_ = async_finished_ = false;
	mode_ = nullptr;
	manual_r_ = manual_b_ = 0.0;
	first_switch_mode_ = true;
	async_thread_ = std::thread(std::bind(&Awb::asyncFunc, this));
}

Awb::~Awb()
{
	{
		std::lock_guard<std::mutex> lock(mutex_);
		async_abort_ = true;
	}
	async_signal_.notify_one();
	async_thread_.join();
}

char const *Awb::Name() const
{
	return NAME;
}

void Awb::Read(boost::property_tree::ptree const &params)
{
	config_.Read(params);
}

void Awb::Initialise()
{
	frame_count_ = frame_phase_ = 0;
	// Put something sane into the status that we are filtering towards,
	// just in case the first few frames don't have anything meaningful in
	// them.
	if (!config_.ct_r.Empty() && !config_.ct_b.Empty()) {
		sync_results_.temperature_K = config_.ct_r.Domain().Clip(4000);
		sync_results_.gain_r =
			1.0 / config_.ct_r.Eval(sync_results_.temperature_K);
		sync_results_.gain_g = 1.0;
		sync_results_.gain_b =
			1.0 / config_.ct_b.Eval(sync_results_.temperature_K);
	} else {
		// random values just to stop the world blowing up
		sync_results_.temperature_K = 4500;
		sync_results_.gain_r = sync_results_.gain_g =
			sync_results_.gain_b = 1.0;
	}
	prev_sync_results_ = sync_results_;
	async_results_ = sync_results_;
}

unsigned int Awb::GetConvergenceFrames() const
{
	// If not in auto mode, there is no convergence
	// to happen, so no need to drop any frames - return zero.
	if (!isAutoEnabled())
		return 0;
	else
		return config_.convergence_frames;
}

void Awb::SetMode(std::string const &mode_name)
{
	mode_name_ = mode_name;
}

void Awb::SetManualGains(double manual_r, double manual_b)
{
	// If any of these are 0.0, we swich back to auto.
	manual_r_ = manual_r;
	manual_b_ = manual_b;
	// If not in auto mode, set these values into the sync_results which
	// means that Prepare() will adopt them immediately.
	if (!isAutoEnabled()) {
		sync_results_.gain_r = prev_sync_results_.gain_r = manual_r_;
		sync_results_.gain_g = prev_sync_results_.gain_g = 1.0;
		sync_results_.gain_b = prev_sync_results_.gain_b = manual_b_;
	}
}

void Awb::SwitchMode([[maybe_unused]] CameraMode const &camera_mode,
		     Metadata *metadata)
{
	// On the first mode switch we'll have no meaningful colour
	// temperature, so try to dead reckon one if in manual mode.
	if (!isAutoEnabled() && first_switch_mode_ && config_.bayes) {
		Pwl ct_r_inverse = config_.ct_r.Inverse();
		Pwl ct_b_inverse = config_.ct_b.Inverse();
		double ct_r = ct_r_inverse.Eval(ct_r_inverse.Domain().Clip(1 / manual_r_));
		double ct_b = ct_b_inverse.Eval(ct_b_inverse.Domain().Clip(1 / manual_b_));
		prev_sync_results_.temperature_K = (ct_r + ct_b) / 2;
		sync_results_.temperature_K = prev_sync_results_.temperature_K;
	}
	// Let other algorithms know the current white balance values.
	metadata->Set("awb.status", prev_sync_results_);
	first_switch_mode_ = false;
}

bool Awb::isAutoEnabled() const
{
	return manual_r_ == 0.0 || manual_b_ == 0.0;
}

void Awb::fetchAsyncResults()
{
	LOG(RPiAwb, Debug) << "Fetch AWB results";
	async_finished_ = false;
	async_started_ = false;
	// It's possible manual gains could be set even while the async
	// thread was running, so only copy the results if still in auto mode.
	if (isAutoEnabled())
		sync_results_ = async_results_;
}

void Awb::restartAsync(StatisticsPtr &stats, double lux)
{
	LOG(RPiAwb, Debug) << "Starting AWB calculation";
	// this makes a new reference which belongs to the asynchronous thread
	statistics_ = stats;
	// store the mode as it could technically change
	auto m = config_.modes.find(mode_name_);
	mode_ = m != config_.modes.end()
			? &m->second
			: (mode_ == nullptr ? config_.default_mode : mode_);
	lux_ = lux;
	frame_phase_ = 0;
	async_started_ = true;
	size_t len = mode_name_.copy(async_results_.mode,
				     sizeof(async_results_.mode) - 1);
	async_results_.mode[len] = '\0';
	{
		std::lock_guard<std::mutex> lock(mutex_);
		async_start_ = true;
	}
	async_signal_.notify_one();
}

void Awb::Prepare(Metadata *image_metadata)
{
	if (frame_count_ < (int)config_.startup_frames)
		frame_count_++;
	double speed = frame_count_ < (int)config_.startup_frames
			       ? 1.0
			       : config_.speed;
	LOG(RPiAwb, Debug)
		<< "frame_count " << frame_count_ << " speed " << speed;
	{
		std::unique_lock<std::mutex> lock(mutex_);
		if (async_started_ && async_finished_)
			fetchAsyncResults();
	}
	// Finally apply IIR filter to results and put into metadata.
	memcpy(prev_sync_results_.mode, sync_results_.mode,
	       sizeof(prev_sync_results_.mode));
	prev_sync_results_.temperature_K =
		speed * sync_results_.temperature_K +
		(1.0 - speed) * prev_sync_results_.temperature_K;
	prev_sync_results_.gain_r = speed * sync_results_.gain_r +
				    (1.0 - speed) * prev_sync_results_.gain_r;
	prev_sync_results_.gain_g = speed * sync_results_.gain_g +
				    (1.0 - speed) * prev_sync_results_.gain_g;
	prev_sync_results_.gain_b = speed * sync_results_.gain_b +
				    (1.0 - speed) * prev_sync_results_.gain_b;
	image_metadata->Set("awb.status", prev_sync_results_);
	LOG(RPiAwb, Debug)
		<< "Using AWB gains r " << prev_sync_results_.gain_r << " g "
		<< prev_sync_results_.gain_g << " b "
		<< prev_sync_results_.gain_b;
}

void Awb::Process(StatisticsPtr &stats, Metadata *image_metadata)
{
	// Count frames since we last poked the async thread.
	if (frame_phase_ < (int)config_.frame_period)
		frame_phase_++;
	LOG(RPiAwb, Debug) << "frame_phase " << frame_phase_;
	// We do not restart the async thread if we're not in auto mode.
	if (isAutoEnabled() &&
	    (frame_phase_ >= (int)config_.frame_period ||
	     frame_count_ < (int)config_.startup_frames)) {
		// Update any settings and any image metadata that we need.
		struct LuxStatus lux_status = {};
		lux_status.lux = 400; // in case no metadata
		if (image_metadata->Get("lux.status", lux_status) != 0)
			LOG(RPiAwb, Debug) << "No lux metadata found";
		LOG(RPiAwb, Debug) << "Awb lux value is " << lux_status.lux;

		if (async_started_ == false)
			restartAsync(stats, lux_status.lux);
	}
}

void Awb::asyncFunc()
{
	while (true) {
		{
			std::unique_lock<std::mutex> lock(mutex_);
			async_signal_.wait(lock, [&] {
				return async_start_ || async_abort_;
			});
			async_start_ = false;
			if (async_abort_)
				break;
		}
		doAwb();
		{
			std::lock_guard<std::mutex> lock(mutex_);
			async_finished_ = true;
		}
		sync_signal_.notify_one();
	}
}

static void generate_stats(std::vector<Awb::RGB> &zones,
			   bcm2835_isp_stats_region *stats, double min_pixels,
			   double min_G)
{
	for (int i = 0; i < AWB_STATS_SIZE_X * AWB_STATS_SIZE_Y; i++) {
		Awb::RGB zone;
		double counted = stats[i].counted;
		if (counted >= min_pixels) {
			zone.G = stats[i].g_sum / counted;
			if (zone.G >= min_G) {
				zone.R = stats[i].r_sum / counted;
				zone.B = stats[i].b_sum / counted;
				zones.push_back(zone);
			}
		}
	}
}

void Awb::prepareStats()
{
	zones_.clear();
	// LSC has already been applied to the stats in this pipeline, so stop
	// any LSC compensation.  We also ignore config_.fast in this version.
	generate_stats(zones_, statistics_->awb_stats, config_.min_pixels,
		       config_.min_G);
	// we're done with these; we may as well relinquish our hold on the
	// pointer.
	statistics_.reset();
	// apply sensitivities, so values appear to come from our "canonical"
	// sensor.
	for (auto &zone : zones_)
		zone.R *= config_.sensitivity_r,
			zone.B *= config_.sensitivity_b;
}

double Awb::computeDelta2Sum(double gain_r, double gain_b)
{
	// Compute the sum of the squared colour error (non-greyness) as it
	// appears in the log likelihood equation.
	double delta2_sum = 0;
	for (auto &z : zones_) {
		double delta_r = gain_r * z.R - 1 - config_.whitepoint_r;
		double delta_b = gain_b * z.B - 1 - config_.whitepoint_b;
		double delta2 = delta_r * delta_r + delta_b * delta_b;
		//LOG(RPiAwb, Debug) << "delta_r " << delta_r << " delta_b " << delta_b << " delta2 " << delta2;
		delta2 = std::min(delta2, config_.delta_limit);
		delta2_sum += delta2;
	}
	return delta2_sum;
}

Pwl Awb::interpolatePrior()
{
	// Interpolate the prior log likelihood function for our current lux
	// value.
	if (lux_ <= config_.priors.front().lux)
		return config_.priors.front().prior;
	else if (lux_ >= config_.priors.back().lux)
		return config_.priors.back().prior;
	else {
		int idx = 0;
		// find which two we lie between
		while (config_.priors[idx + 1].lux < lux_)
			idx++;
		double lux0 = config_.priors[idx].lux,
		       lux1 = config_.priors[idx + 1].lux;
		return Pwl::Combine(config_.priors[idx].prior,
				    config_.priors[idx + 1].prior,
				    [&](double /*x*/, double y0, double y1) {
					    return y0 + (y1 - y0) *
							(lux_ - lux0) / (lux1 - lux0);
				    });
	}
}

static double interpolate_quadatric(Pwl::Point const &A, Pwl::Point const &B,
				    Pwl::Point const &C)
{
	// Given 3 points on a curve, find the extremum of the function in that
	// interval by fitting a quadratic.
	const double eps = 1e-3;
	Pwl::Point CA = C - A, BA = B - A;
	double denominator = 2 * (BA.y * CA.x - CA.y * BA.x);
	if (abs(denominator) > eps) {
		double numerator = BA.y * CA.x * CA.x - CA.y * BA.x * BA.x;
		double result = numerator / denominator + A.x;
		return std::max(A.x, std::min(C.x, result));
	}
	// has degenerated to straight line segment
	return A.y < C.y - eps ? A.x : (C.y < A.y - eps ? C.x : B.x);
}

double Awb::coarseSearch(Pwl const &prior)
{
	points_.clear(); // assume doesn't deallocate memory
	size_t best_point = 0;
	double t = mode_->ct_lo;
	int span_r = 0, span_b = 0;
	// Step down the CT curve evaluating log likelihood.
	while (true) {
		double r = config_.ct_r.Eval(t, &span_r);
		double b = config_.ct_b.Eval(t, &span_b);
		double gain_r = 1 / r, gain_b = 1 / b;
		double delta2_sum = computeDelta2Sum(gain_r, gain_b);
		double prior_log_likelihood =
			prior.Eval(prior.Domain().Clip(t));
		double final_log_likelihood = delta2_sum - prior_log_likelihood;
		LOG(RPiAwb, Debug)
			<< "t: " << t << " gain_r " << gain_r << " gain_b "
			<< gain_b << " delta2_sum " << delta2_sum
			<< " prior " << prior_log_likelihood << " final "
			<< final_log_likelihood;
		points_.push_back(Pwl::Point(t, final_log_likelihood));
		if (points_.back().y < points_[best_point].y)
			best_point = points_.size() - 1;
		if (t == mode_->ct_hi)
			break;
		// for even steps along the r/b curve scale them by the current t
		t = std::min(t + t / 10 * config_.coarse_step,
			     mode_->ct_hi);
	}
	t = points_[best_point].x;
	LOG(RPiAwb, Debug) << "Coarse search found CT " << t;
	// We have the best point of the search, but refine it with a quadratic
	// interpolation around its neighbours.
	if (points_.size() > 2) {
		unsigned long bp = std::min(best_point, points_.size() - 2);
		best_point = std::max(1UL, bp);
		t = interpolate_quadatric(points_[best_point - 1],
					  points_[best_point],
					  points_[best_point + 1]);
		LOG(RPiAwb, Debug)
			<< "After quadratic refinement, coarse search has CT "
			<< t;
	}
	return t;
}

void Awb::fineSearch(double &t, double &r, double &b, Pwl const &prior)
{
	int span_r = -1, span_b = -1;
	config_.ct_r.Eval(t, &span_r);
	config_.ct_b.Eval(t, &span_b);
	double step = t / 10 * config_.coarse_step * 0.1;
	int nsteps = 5;
	double r_diff = config_.ct_r.Eval(t + nsteps * step, &span_r) -
			config_.ct_r.Eval(t - nsteps * step, &span_r);
	double b_diff = config_.ct_b.Eval(t + nsteps * step, &span_b) -
			config_.ct_b.Eval(t - nsteps * step, &span_b);
	Pwl::Point transverse(b_diff, -r_diff);
	if (transverse.Len2() < 1e-6)
		return;
	// unit vector orthogonal to the b vs. r function (pointing outwards
	// with r and b increasing)
	transverse = transverse / transverse.Len();
	double best_log_likelihood = 0, best_t = 0, best_r = 0, best_b = 0;
	double transverse_range =
		config_.transverse_neg + config_.transverse_pos;
	const int MAX_NUM_DELTAS = 12;
	// a transverse step approximately every 0.01 r/b units
	int num_deltas = floor(transverse_range * 100 + 0.5) + 1;
	num_deltas = num_deltas < 3 ? 3 :
		     (num_deltas > MAX_NUM_DELTAS ? MAX_NUM_DELTAS : num_deltas);
	// Step down CT curve. March a bit further if the transverse range is
	// large.
	nsteps += num_deltas;
	for (int i = -nsteps; i <= nsteps; i++) {
		double t_test = t + i * step;
		double prior_log_likelihood =
			prior.Eval(prior.Domain().Clip(t_test));
		double r_curve = config_.ct_r.Eval(t_test, &span_r);
		double b_curve = config_.ct_b.Eval(t_test, &span_b);
		// x will be distance off the curve, y the log likelihood there
		Pwl::Point points[MAX_NUM_DELTAS];
		int best_point = 0;
		// Take some measurements transversely *off* the CT curve.
		for (int j = 0; j < num_deltas; j++) {
			points[j].x = -config_.transverse_neg +
				      (transverse_range * j) / (num_deltas - 1);
			Pwl::Point rb_test = Pwl::Point(r_curve, b_curve) +
					     transverse * points[j].x;
			double r_test = rb_test.x, b_test = rb_test.y;
			double gain_r = 1 / r_test, gain_b = 1 / b_test;
			double delta2_sum = computeDelta2Sum(gain_r, gain_b);
			points[j].y = delta2_sum - prior_log_likelihood;
			LOG(RPiAwb, Debug)
				<< "At t " << t_test << " r " << r_test << " b "
				<< b_test << ": " << points[j].y;
			if (points[j].y < points[best_point].y)
				best_point = j;
		}
		// We have NUM_DELTAS points transversely across the CT curve,
		// now let's do a quadratic interpolation for the best result.
		best_point = std::max(1, std::min(best_point, num_deltas - 2));
		Pwl::Point rb_test =
			Pwl::Point(r_curve, b_curve) +
			transverse *
				interpolate_quadatric(points[best_point - 1],
						      points[best_point],
						      points[best_point + 1]);
		double r_test = rb_test.x, b_test = rb_test.y;
		double gain_r = 1 / r_test, gain_b = 1 / b_test;
		double delta2_sum = computeDelta2Sum(gain_r, gain_b);
		double final_log_likelihood = delta2_sum - prior_log_likelihood;
		LOG(RPiAwb, Debug)
			<< "Finally "
			<< t_test << " r " << r_test << " b " << b_test << ": "
			<< final_log_likelihood
			<< (final_log_likelihood < best_log_likelihood ? " BEST" : "");
		if (best_t == 0 || final_log_likelihood < best_log_likelihood)
			best_log_likelihood = final_log_likelihood,
			best_t = t_test, best_r = r_test, best_b = b_test;
	}
	t = best_t, r = best_r, b = best_b;
	LOG(RPiAwb, Debug)
		<< "Fine search found t " << t << " r " << r << " b " << b;
}

void Awb::awbBayes()
{
	// May as well divide out G to save computeDelta2Sum from doing it over
	// and over.
	for (auto &z : zones_)
		z.R = z.R / (z.G + 1), z.B = z.B / (z.G + 1);
	// Get the current prior, and scale according to how many zones are
	// valid... not entirely sure about this.
	Pwl prior = interpolatePrior();
	prior *= zones_.size() / (double)(AWB_STATS_SIZE_X * AWB_STATS_SIZE_Y);
	prior.Map([](double x, double y) {
		LOG(RPiAwb, Debug) << "(" << x << "," << y << ")";
	});
	double t = coarseSearch(prior);
	double r = config_.ct_r.Eval(t);
	double b = config_.ct_b.Eval(t);
	LOG(RPiAwb, Debug)
		<< "After coarse search: r " << r << " b " << b << " (gains r "
		<< 1 / r << " b " << 1 / b << ")";
	// Not entirely sure how to handle the fine search yet. Mostly the
	// estimated CT is already good enough, but the fine search allows us to
	// wander transverely off the CT curve. Under some illuminants, where
	// there may be more or less green light, this may prove beneficial,
	// though I probably need more real datasets before deciding exactly how
	// this should be controlled and tuned.
	fineSearch(t, r, b, prior);
	LOG(RPiAwb, Debug)
		<< "After fine search: r " << r << " b " << b << " (gains r "
		<< 1 / r << " b " << 1 / b << ")";
	// Write results out for the main thread to pick up. Remember to adjust
	// the gains from the ones that the "canonical sensor" would require to
	// the ones needed by *this* sensor.
	async_results_.temperature_K = t;
	async_results_.gain_r = 1.0 / r * config_.sensitivity_r;
	async_results_.gain_g = 1.0;
	async_results_.gain_b = 1.0 / b * config_.sensitivity_b;
}

void Awb::awbGrey()
{
	LOG(RPiAwb, Debug) << "Grey world AWB";
	// Make a separate list of the derivatives for each of red and blue, so
	// that we can sort them to exclude the extreme gains.  We could
	// consider some variations, such as normalising all the zones first, or
	// doing an L2 average etc.
	std::vector<RGB> &derivs_R(zones_);
	std::vector<RGB> derivs_B(derivs_R);
	std::sort(derivs_R.begin(), derivs_R.end(),
		  [](RGB const &a, RGB const &b) {
			  return a.G * b.R < b.G * a.R;
		  });
	std::sort(derivs_B.begin(), derivs_B.end(),
		  [](RGB const &a, RGB const &b) {
			  return a.G * b.B < b.G * a.B;
		  });
	// Average the middle half of the values.
	int discard = derivs_R.size() / 4;
	RGB sum_R(0, 0, 0), sum_B(0, 0, 0);
	for (auto ri = derivs_R.begin() + discard,
		  bi = derivs_B.begin() + discard;
	     ri != derivs_R.end() - discard; ri++, bi++)
		sum_R += *ri, sum_B += *bi;
	double gain_r = sum_R.G / (sum_R.R + 1),
	       gain_b = sum_B.G / (sum_B.B + 1);
	async_results_.temperature_K = 4500; // don't know what it is
	async_results_.gain_r = gain_r;
	async_results_.gain_g = 1.0;
	async_results_.gain_b = gain_b;
}

void Awb::doAwb()
{
	prepareStats();
	LOG(RPiAwb, Debug) << "Valid zones: " << zones_.size();
	if (zones_.size() > config_.min_regions) {
		if (config_.bayes)
			awbBayes();
		else
			awbGrey();
		LOG(RPiAwb, Debug)
			<< "CT found is "
			<< async_results_.temperature_K
			<< " with gains r " << async_results_.gain_r
			<< " and b " << async_results_.gain_b;
	}
}

// Register algorithm with the system.
static Algorithm *Create(Controller *controller)
{
	return (Algorithm *)new Awb(controller);
}
static RegisterAlgorithm reg(NAME, &Create);