[libcamera-devel] [PATCH v1 10/10] ipa: raspberrypi: Generalise the autofocus algorithm
Nick Hollinghurst
nick.hollinghurst at raspberrypi.com
Fri Mar 24 12:35:15 CET 2023
Hi Jacopo,
Thanks - comments inline.
Nick
On Fri, 24 Mar 2023 at 10:21, Jacopo Mondi
<jacopo.mondi at ideasonboard.com> wrote:
>
> Hi Nick
>
> On Wed, Mar 22, 2023 at 01:06:12PM +0000, Naushir Patuck via libcamera-devel wrote:
> > From: Nick Hollinghurst <nick.hollinghurst at raspberrypi.com>
> >
> > Remove any hard-coded assumptions about the target hardware platform
> > from the autofocus algorithm. Instead, use the "target" string provided
> > by the camera tuning config and generalised statistics structures to
> > determing parameters such as grid and region sizes.
> >
> > Additionally, PDAF statistics are represented by a generalised region
> > statistics structure to be device agnostic.
> >
> > These changes also require the autofocus algorithm to initialise
> > region weights on the first frame's prepare()/process() call rather
> > than during initialisation.
> >
> > Signed-off-by: Nick Hollinghurst <nick.hollinghurst at raspberrypi.com>
> > Signed-off-by: Naushir Patuck <naush at raspberrypi.com>
> > Tested-by: Naushir Patuck <naush at raspberrypi.com>
> > ---
> > src/ipa/raspberrypi/cam_helper_imx708.cpp | 23 ++-
> > src/ipa/raspberrypi/controller/pdaf_data.h | 21 +--
> > src/ipa/raspberrypi/controller/rpi/af.cpp | 161 ++++++++++-----------
> > src/ipa/raspberrypi/controller/rpi/af.h | 28 ++--
> > 4 files changed, 119 insertions(+), 114 deletions(-)
> >
> > diff --git a/src/ipa/raspberrypi/cam_helper_imx708.cpp b/src/ipa/raspberrypi/cam_helper_imx708.cpp
> > index 1f213d3c0833..641ba18f4b9d 100644
> > --- a/src/ipa/raspberrypi/cam_helper_imx708.cpp
> > +++ b/src/ipa/raspberrypi/cam_helper_imx708.cpp
> > @@ -69,11 +69,14 @@ private:
> > /* Largest long exposure scale factor given as a left shift on the frame length. */
> > static constexpr int longExposureShiftMax = 7;
> >
> > + static constexpr int pdafStatsRows = 12;
> > + static constexpr int pdafStatsCols = 16;
> > +
> > void populateMetadata(const MdParser::RegisterMap ®isters,
> > Metadata &metadata) const override;
> >
> > static bool parsePdafData(const uint8_t *ptr, size_t len, unsigned bpp,
> > - PdafData &pdaf);
> > + PdafRegions &pdaf);
> >
> > bool parseAEHist(const uint8_t *ptr, size_t len, unsigned bpp);
> > void putAGCStatistics(StatisticsPtr stats);
> > @@ -120,11 +123,11 @@ void CamHelperImx708::prepare(libcamera::Span<const uint8_t> buffer, Metadata &m
> > size_t bytesPerLine = (mode_.width * mode_.bitdepth) >> 3;
> >
> > if (buffer.size() > 2 * bytesPerLine) {
> > - PdafData pdaf;
> > + PdafRegions pdaf;
> > if (parsePdafData(&buffer[2 * bytesPerLine],
> > buffer.size() - 2 * bytesPerLine,
> > mode_.bitdepth, pdaf))
> > - metadata.set("pdaf.data", pdaf);
> > + metadata.set("pdaf.regions", pdaf);
> > }
> >
> > /* Parse AE-HIST data where present */
> > @@ -239,7 +242,7 @@ void CamHelperImx708::populateMetadata(const MdParser::RegisterMap ®isters,
> > }
> >
> > bool CamHelperImx708::parsePdafData(const uint8_t *ptr, size_t len,
> > - unsigned bpp, PdafData &pdaf)
> > + unsigned bpp, PdafRegions &pdaf)
> > {
> > size_t step = bpp >> 1; /* bytes per PDAF grid entry */
> >
> > @@ -248,13 +251,17 @@ bool CamHelperImx708::parsePdafData(const uint8_t *ptr, size_t len,
> > return false;
> > }
> >
> > + pdaf.init({ pdafStatsCols, pdafStatsRows });
> > +
> > ptr += 2 * step;
> > - for (unsigned i = 0; i < PDAF_DATA_ROWS; ++i) {
> > - for (unsigned j = 0; j < PDAF_DATA_COLS; ++j) {
> > + for (unsigned i = 0; i < pdafStatsRows; ++i) {
> > + for (unsigned j = 0; j < pdafStatsCols; ++j) {
> > unsigned c = (ptr[0] << 3) | (ptr[1] >> 5);
> > int p = (((ptr[1] & 0x0F) - (ptr[1] & 0x10)) << 6) | (ptr[2] >> 2);
> > - pdaf.conf[i][j] = c;
> > - pdaf.phase[i][j] = c ? p : 0;
> > + PdafData pdafData;
> > + pdafData.conf = c;
> > + pdafData.phase = c ? p : 0;
> > + pdaf.set(libcamera::Point(j, i), { pdafData, 1, 0 });
> > ptr += step;
> > }
> > }
> > diff --git a/src/ipa/raspberrypi/controller/pdaf_data.h b/src/ipa/raspberrypi/controller/pdaf_data.h
> > index 03c00d72c0e8..ae6ab996ded0 100644
> > --- a/src/ipa/raspberrypi/controller/pdaf_data.h
> > +++ b/src/ipa/raspberrypi/controller/pdaf_data.h
> > @@ -2,20 +2,23 @@
> > /*
> > * Copyright (C) 2022, Raspberry Pi Ltd
> > *
> > - * pdaf_data.h - PDAF Metadata; for now this is
> > - * largely based on IMX708's PDAF "Type 1" output.
> > + * pdaf_data.h - PDAF Metadata
> > */
> > #pragma once
> >
> > #include <stdint.h>
> >
> > -#define PDAF_DATA_ROWS 12
> > -#define PDAF_DATA_COLS 16
> > +#include "region_stats.h"
> >
> > -struct PdafData {
> > - /* Confidence values, in raster order, in arbitrary units */
> > - uint16_t conf[PDAF_DATA_ROWS][PDAF_DATA_COLS];
> > +namespace RPiController {
> >
> > - /* Phase error, in raster order, in s11 Q4 format (S.6.4) */
> > - int16_t phase[PDAF_DATA_ROWS][PDAF_DATA_COLS];
> > +struct PdafData {
> > + /* Confidence, in arbitrary units */
> > + uint16_t conf;
> > + /* Phase error, in s16 Q4 format (S.11.4) */
> > + int16_t phase;
> > };
> > +
> > +using PdafRegions = RegionStats<PdafData>;
> > +
> > +}; // namespace RPiController
> > diff --git a/src/ipa/raspberrypi/controller/rpi/af.cpp b/src/ipa/raspberrypi/controller/rpi/af.cpp
> > index a623651875f2..3d2dad974d8b 100644
> > --- a/src/ipa/raspberrypi/controller/rpi/af.cpp
> > +++ b/src/ipa/raspberrypi/controller/rpi/af.cpp
> > @@ -174,9 +174,8 @@ Af::Af(Controller *controller)
> > statsRegion_(0, 0, 0, 0),
> > windows_(),
> > useWindows_(false),
> > - phaseWeights_{},
> > - contrastWeights_{},
> > - sumWeights_(0),
> > + phaseWeights_(),
> > + contrastWeights_(),
> > scanState_(ScanState::Idle),
> > initted_(false),
> > ftarget_(-1.0),
> > @@ -190,6 +189,8 @@ Af::Af(Controller *controller)
> > scanData_(),
> > reportState_(AfState::Idle)
> > {
> > + phaseWeights_.w.reserve(16 * 12);
>
> Isn't the point of the patch to get these values from the camHelper ?
Ideally it would.. But right now, the helper doesn't tell us the
dimensions of the grids in advance, so I had to "guess".
All these reserve() calls are perhaps an unnecessary affectation, an
attempt to reduce memory fragmentation -- should I remove them?
> > + contrastWeights_.w.reserve(4 * 3);
> > scanData_.reserve(24);
> > }
> >
> > @@ -226,7 +227,7 @@ void Af::switchMode(CameraMode const &cameraMode, [[maybe_unused]] Metadata *met
> > << statsRegion_.y << ','
> > << statsRegion_.width << ','
> > << statsRegion_.height;
> > - computeWeights();
> > + invalidateWeights();
> >
> > if (scanState_ >= ScanState::Coarse && scanState_ < ScanState::Settle) {
> > /*
> > @@ -239,111 +240,94 @@ void Af::switchMode(CameraMode const &cameraMode, [[maybe_unused]] Metadata *met
> > skipCount_ = cfg_.skipFrames;
> > }
> >
> > -void Af::computeWeights()
> > +void Af::computeWeights(RegionWeights *wgts, unsigned rows, unsigned cols)
> > {
> > - constexpr int MaxCellWeight = 240 / (int)MaxWindows;
> > + wgts->rows = rows;
> > + wgts->cols = cols;
> > + wgts->sum = 0;
> > + wgts->w.resize(rows * cols);
> > + std::fill(wgts->w.begin(), wgts->w.end(), 0);
> >
> > - sumWeights_ = 0;
> > - for (int i = 0; i < PDAF_DATA_ROWS; ++i)
> > - std::fill(phaseWeights_[i], phaseWeights_[i] + PDAF_DATA_COLS, 0);
> > -
> > - if (useWindows_ &&
> > - statsRegion_.width >= PDAF_DATA_COLS && statsRegion_.height >= PDAF_DATA_ROWS) {
> > + if (rows > 0 && cols > 0 && useWindows_ &&
> > + statsRegion_.height >= rows && statsRegion_.width >= cols) {
> > /*
> > * Here we just merge all of the given windows, weighted by area.
> > * \todo Perhaps a better approach might be to find the phase in each
> > * window and choose either the closest or the highest-confidence one?
> > - *
> > - * Using mostly "int" arithmetic, because Rectangle has signed x, y
> > */
> > - int cellH = (int)(statsRegion_.height / PDAF_DATA_ROWS);
> > - int cellW = (int)(statsRegion_.width / PDAF_DATA_COLS);
> > - int cellA = cellH * cellW;
> > + const unsigned maxCellWeight = 46080u / (MaxWindows * rows * cols);
>
> Where does 46080u comes from ? I can't find it in the diff
It's a "highly divisible" number, close to but not exceeding 65536,
chosen to anticipate the grid dimensions having only factors of 2 and
3, and MaxWindows==10. I'm not sure if there's a better way to express
that in the code? For the IMX708 PDAF grid it yields maxCellWeight =
24, as before.
Actually, a numerator of 65535 or 65536 (for denominator > 1) would
work almost as well, but it would have different rounding behaviour in
cases where a focus window intersects a fraction of a grid cell.
> > + const unsigned cellH = statsRegion_.height / rows;
> > + const unsigned cellW = statsRegion_.width / cols;
> > + const unsigned cellA = cellH * cellW;
> >
> > for (auto &w : windows_) {
> > - for (int i = 0; i < PDAF_DATA_ROWS; ++i) {
> > - int y0 = std::max(statsRegion_.y + cellH * i, w.y);
> > - int y1 = std::min(statsRegion_.y + cellH * (i + 1), w.y + (int)(w.height));
> > + for (unsigned r = 0; r < rows; ++r) {
> > + int y0 = std::max(statsRegion_.y + (int)(cellH * r), w.y);
> > + int y1 = std::min(statsRegion_.y + (int)(cellH * (r + 1)), w.y + (int)(w.height));
> > if (y0 >= y1)
> > continue;
> > y1 -= y0;
> > - for (int j = 0; j < PDAF_DATA_COLS; ++j) {
> > - int x0 = std::max(statsRegion_.x + cellW * j, w.x);
> > - int x1 = std::min(statsRegion_.x + cellW * (j + 1), w.x + (int)(w.width));
> > + for (unsigned c = 0; c < cols; ++c) {
> > + int x0 = std::max(statsRegion_.x + (int)(cellW * c), w.x);
> > + int x1 = std::min(statsRegion_.x + (int)(cellW * (c + 1)), w.x + (int)(w.width));
> > if (x0 >= x1)
> > continue;
> > - int a = y1 * (x1 - x0);
> > - a = (MaxCellWeight * a + cellA - 1) / cellA;
> > - phaseWeights_[i][j] += a;
> > - sumWeights_ += a;
> > + unsigned a = y1 * (x1 - x0);
> > + a = (maxCellWeight * a + cellA - 1) / cellA;
> > + wgts->w[r * cols + c] += a;
> > + wgts->sum += a;
> > }
> > }
> > }
> > }
> >
> > - if (sumWeights_ == 0) {
> > - /*
> > - * Default AF window is the middle 1/2 width of the middle 1/3 height
> > - * since this maps nicely to both PDAF (16x12) and Focus (4x3) grids.
> > - */
> > - for (int i = PDAF_DATA_ROWS / 3; i < 2 * PDAF_DATA_ROWS / 3; ++i) {
> > - for (int j = PDAF_DATA_COLS / 4; j < 3 * PDAF_DATA_COLS / 4; ++j) {
> > - phaseWeights_[i][j] = MaxCellWeight;
> > - sumWeights_ += MaxCellWeight;
> > + if (wgts->sum == 0) {
> > + /* Default AF window is the middle 1/2 width of the middle 1/3 height */
> > + for (unsigned r = rows / 3; r < rows - rows / 3; ++r) {
> > + for (unsigned c = cols / 4; c < cols - cols / 4; ++c) {
> > + wgts->w[r * cols + c] = 1;
> > + wgts->sum += 1;
>
> This seems functionally different, or was MaxCellWeight == 1 ?
When used, we divide by the total weight (rounding down). I realised
that the original MaxCellWeight factor was unnecessary.
> > }
> > }
> > }
> > +}
> >
> > - /* Scale from PDAF to Focus Statistics grid (which has fixed size 4x3) */
> > - constexpr int FocusStatsRows = 3;
> > - constexpr int FocusStatsCols = 4;
> > - static_assert(FOCUS_REGIONS == FocusStatsRows * FocusStatsCols);
> > - static_assert(PDAF_DATA_ROWS % FocusStatsRows == 0);
> > - static_assert(PDAF_DATA_COLS % FocusStatsCols == 0);
> > - constexpr int YFactor = PDAF_DATA_ROWS / FocusStatsRows;
> > - constexpr int XFactor = PDAF_DATA_COLS / FocusStatsCols;
> > -
> > - LOG(RPiAf, Debug) << "Recomputed weights:";
> > - for (int i = 0; i < FocusStatsRows; ++i) {
> > - for (int j = 0; j < FocusStatsCols; ++j) {
> > - unsigned w = 0;
> > - for (int y = 0; y < YFactor; ++y)
> > - for (int x = 0; x < XFactor; ++x)
> > - w += phaseWeights_[YFactor * i + y][XFactor * j + x];
> > - contrastWeights_[FocusStatsCols * i + j] = w;
> > - }
> > - LOG(RPiAf, Debug) << " "
> > - << contrastWeights_[FocusStatsCols * i + 0] << " "
> > - << contrastWeights_[FocusStatsCols * i + 1] << " "
> > - << contrastWeights_[FocusStatsCols * i + 2] << " "
> > - << contrastWeights_[FocusStatsCols * i + 3];
> > - }
>
> I might have missed where re-scaling is gone.. as you've tested this I
> presum it's just me not fully getting what's happening here
It's missing because we now call the function twice: to construct
weights for each of the PDAF (from the sensor) and CDAF (from the ISP)
grids in turn.
The code no longer assumes or asserts that one grid's size is an exact
integer fraction of the other.
>
> > +void Af::invalidateWeights()
> > +{
> > + phaseWeights_.sum = 0;
> > + contrastWeights_.sum = 0;
> > }
> >
> > -bool Af::getPhase(PdafData const &data, double &phase, double &conf) const
> > +bool Af::getPhase(PdafRegions const ®ions, double &phase, double &conf)
> > {
> > + libcamera::Size size = regions.size();
> > + if (size.height != phaseWeights_.rows || size.width != phaseWeights_.cols ||
> > + phaseWeights_.sum == 0) {
> > + LOG(RPiAf, Debug) << "Recompute Phase weights " << size.width << 'x' << size.height;
> > + computeWeights(&phaseWeights_, size.height, size.width);
> > + }
> > +
> > uint32_t sumWc = 0;
> > int64_t sumWcp = 0;
> > -
> > - for (unsigned i = 0; i < PDAF_DATA_ROWS; ++i) {
> > - for (unsigned j = 0; j < PDAF_DATA_COLS; ++j) {
> > - if (phaseWeights_[i][j]) {
> > - uint32_t c = data.conf[i][j];
> > - if (c >= cfg_.confThresh) {
> > - if (c > cfg_.confClip)
> > - c = cfg_.confClip;
> > - c -= (cfg_.confThresh >> 2);
> > - sumWc += phaseWeights_[i][j] * c;
> > - c -= (cfg_.confThresh >> 2);
> > - sumWcp += phaseWeights_[i][j] * data.phase[i][j] * (int64_t)c;
> > - }
> > + for (unsigned i = 0; i < regions.numRegions(); ++i) {
> > + unsigned w = phaseWeights_.w[i];
> > + if (w) {
> > + const PdafData &data = regions.get(i).val;
> > + unsigned c = data.conf;
> > + if (c >= cfg_.confThresh) {
> > + if (c > cfg_.confClip)
> > + c = cfg_.confClip;
> > + c -= (cfg_.confThresh >> 2);
> > + sumWc += w * c;
> > + c -= (cfg_.confThresh >> 2);
> > + sumWcp += (int64_t)(w * c) * (int64_t)data.phase;
> > }
> > }
> > }
> >
> > - if (0 < sumWeights_ && sumWeights_ <= sumWc) {
> > + if (0 < phaseWeights_.sum && phaseWeights_.sum <= sumWc) {
> > phase = (double)sumWcp / (double)sumWc;
> > - conf = (double)sumWc / (double)sumWeights_;
> > + conf = (double)sumWc / (double)phaseWeights_.sum;
> > return true;
> > } else {
> > phase = 0.0;
> > @@ -352,14 +336,19 @@ bool Af::getPhase(PdafData const &data, double &phase, double &conf) const
> > }
> > }
> >
> > -double Af::getContrast(const FocusRegions &focusStats) const
> > +double Af::getContrast(const FocusRegions &focusStats)
> > {
> > - uint32_t sumWc = 0;
> > + libcamera::Size size = focusStats.size();
> > + if (size.height != contrastWeights_.rows || size.width != contrastWeights_.cols || contrastWeights_.sum == 0) {
>
> Please break this to multiple lines
>
> > + LOG(RPiAf, Debug) << "Recompute Contrast weights " << size.width << 'x' << size.height;
>
> it's easy to do the same here
OK
> > + computeWeights(&contrastWeights_, size.height, size.width);
> > + }
> >
> > + uint64_t sumWc = 0;
> > for (unsigned i = 0; i < focusStats.numRegions(); ++i)
> > - sumWc += contrastWeights_[i] * focusStats.get(i).val;
> > + sumWc += contrastWeights_.w[i] * focusStats.get(i).val;
> >
> > - return (sumWeights_ == 0) ? 0.0 : (double)sumWc / (double)sumWeights_;
> > + return (contrastWeights_.sum > 0) ? ((double)sumWc / (double)contrastWeights_.sum) : 0.0;
> > }
> >
> > void Af::doPDAF(double phase, double conf)
> > @@ -623,14 +612,14 @@ void Af::prepare(Metadata *imageMetadata)
> >
> > if (initted_) {
> > /* Get PDAF from the embedded metadata, and run AF algorithm core */
> > - PdafData data;
> > + PdafRegions regions;
> > double phase = 0.0, conf = 0.0;
> > double oldFt = ftarget_;
> > double oldFs = fsmooth_;
> > ScanState oldSs = scanState_;
> > uint32_t oldSt = stepCount_;
> > - if (imageMetadata->get("pdaf.data", data) == 0)
> > - getPhase(data, phase, conf);
> > + if (imageMetadata->get("pdaf.regions", regions) == 0)
> > + getPhase(regions, phase, conf);
> > doAF(prevContrast_, phase, conf);
> > updateLensPosition();
> > LOG(RPiAf, Debug) << std::fixed << std::setprecision(2)
> > @@ -691,7 +680,7 @@ void Af::setMetering(bool mode)
> > {
> > if (useWindows_ != mode) {
> > useWindows_ = mode;
> > - computeWeights();
> > + invalidateWeights();
> > }
> > }
> >
> > @@ -708,7 +697,9 @@ void Af::setWindows(libcamera::Span<libcamera::Rectangle const> const &wins)
> > if (windows_.size() >= MaxWindows)
> > break;
> > }
> > - computeWeights();
> > +
> > + if (useWindows_)
> > + invalidateWeights();
> > }
> >
> > bool Af::setLensPosition(double dioptres, int *hwpos)
> > diff --git a/src/ipa/raspberrypi/controller/rpi/af.h b/src/ipa/raspberrypi/controller/rpi/af.h
> > index 7959371baf64..b479feb88c39 100644
> > --- a/src/ipa/raspberrypi/controller/rpi/af.h
> > +++ b/src/ipa/raspberrypi/controller/rpi/af.h
> > @@ -11,12 +11,6 @@
> > #include "../pdaf_data.h"
> > #include "../pwl.h"
> >
> > -/*
> > - * \todo FOCUS_REGIONS is taken from bcm2835-isp.h, but should be made as a
> > - * generic RegionStats structure.
> > - */
> > -#define FOCUS_REGIONS 12
> > -
> > /*
> > * This algorithm implements a hybrid of CDAF and PDAF, favouring PDAF.
> > *
> > @@ -121,9 +115,20 @@ private:
> > double conf;
> > };
> >
> > - void computeWeights();
> > - bool getPhase(PdafData const &data, double &phase, double &conf) const;
> > - double getContrast(const FocusRegions &focusStats) const;
> > + struct RegionWeights {
> > + unsigned rows;
> > + unsigned cols;
> > + uint32_t sum;
> > + std::vector<uint16_t> w;
> > +
> > + RegionWeights()
> > + : rows(0), cols(0), sum(0), w() {}
> > + };
> > +
> > + void computeWeights(RegionWeights *wgts, unsigned rows, unsigned cols);
> > + void invalidateWeights();
> > + bool getPhase(PdafRegions const ®ions, double &phase, double &conf);
> > + double getContrast(const FocusRegions &focusStats);
> > void doPDAF(double phase, double conf);
> > bool earlyTerminationByPhase(double phase);
> > double findPeak(unsigned index) const;
> > @@ -143,9 +148,8 @@ private:
> > libcamera::Rectangle statsRegion_;
> > std::vector<libcamera::Rectangle> windows_;
> > bool useWindows_;
> > - uint8_t phaseWeights_[PDAF_DATA_ROWS][PDAF_DATA_COLS];
> > - uint16_t contrastWeights_[FOCUS_REGIONS];
> > - uint32_t sumWeights_;
> > + RegionWeights phaseWeights_;
> > + RegionWeights contrastWeights_;
> >
> > /* Working state. */
> > ScanState scanState_;
> > --
> > 2.34.1
> >
More information about the libcamera-devel
mailing list