[libcamera-devel] [RFC PATCH 2/2] android: jpeg: Add a basic NV12 image thumbnailer
Hirokazu Honda
hiroh at chromium.org
Mon Oct 26 08:29:20 CET 2020
Hi Laurent and Kieran,
On Fri, Oct 23, 2020 at 5:39 PM Kieran Bingham
<kieran.bingham at ideasonboard.com> wrote:
>
> Hi Hiro-san, Laurent,
>
> On 23/10/2020 04:59, Laurent Pinchart wrote:
> > Hi Hiro-san,
> >
> > On Fri, Oct 23, 2020 at 09:37:09AM +0900, Hirokazu Honda wrote:
> >> On Thu, Oct 22, 2020 at 11:00 PM Laurent Pinchart wrote:
> >>> On Thu, Oct 22, 2020 at 08:42:01PM +0900, Hirokazu Honda wrote:
> >>>> Hi Umang,
> >>>> Thanks for the work. I couldn't have time to review these today and
> >>>> will review tomorrow.
> >>>> In my understanding, the thumbnail class is basically to scale frames
> >>>> to the given destination.
> >>>> Can we generalize it as PostProcessor interface so that we will be
> >>>> able to make use of it for down-scaling regardless of whether jpeg
> >>>> encoding is required.
> >>>> It will be necessary when Android HAL adaptation layer has to output
> >>>> multiple different NV12 streams while the native camera can produce a
> >>>> single NV12 stream simultaneously.
> >>>
> >>> One point to consider is that the thumbnailer is a down-scaler using a
> >>> nearest-neighbour algorithm. It's good enough for thumbnails, but will
> >>> lead to awful quality when scaling streams. For that, I think we need a
> >>> better scaler implementation, and we could then replace the thumbnailer
> >>> class with it.
> >>
> >> We can select nearest-neighbor and bilinear algorithms in
> >> libyuv::NV12Scale with the filtering mode argument[1].
> >> We may want to specify the scale algorithm on a class construction.
> >> I would use libyuv here for better performance and less buggy (+ less
> >> lines to be reviewed).
> >> What do you'all think?
> >
> > We have considered libyuv early on, and I think it's an interesting
> > option moving forward. The reason we have decided to go for a minimal
> > scaler implementation to create JPEG thumbnails is to avoid adding an
> > external dependency that isn't packaged by distributions. This allows us
> > to finalize JPEG thumbnail support without delay, and decide on the
> > longer term solution separately.
> >
> > It would be very nice if Google could work with distributions to package
> > libyuv. I'd like to avoid including a copy of libyuv in libcamera.
>
> Lack of packaging is a big factor here. If it's not available we can't
> use it ;-)
>
Thanks. That is fair enough.
> However I looked into this briefly yesterday, - we could use a meson
> wrap. But I don't know how that will affect us distributing packages, as
> then we also have to distribute libyuv.
>
> Or perhaps the meson wrap would link statically.
>
Can we download libyuv somewhere in libcamera, e.g. //third_party,
during meson config if it is necessary?
The code under //third_party is ignored by .gitignore, i.e., is not a
part of libcamera tree.
We may want to maintain a commit hash to be checked out in the libcamera tree.
Best Regards,
-Hiro
>
> > One last comment to provide all the design background. We have also
> > considered creating a format conversion library in libcamera, to gather
> > code from various projects that exist out there. There's format
> > conversion in libv4l2, in raw2rgbpnm, in projects such as ffmpeg, and of
> > course in libyuv. I would be great if that could be consolidated and
> > packaged, but that's a big task in itself.
>
> And that's the bigger picture. This should be consolidated. And that
> consolidated library should support more than yuv ;-)
>
> Given it would be a software format convertor, and scaler, and ideally
> with hardware acceleration support where possible, I think that quickly
> becomes libisp - and the feature creep will ... creep in ;-)
>
>
>
> >
> >> [1] https://chromium.googlesource.com/libyuv/libyuv/+/refs/heads/master/include/libyuv/scale.h#148
> >>
> >>>> On Wed, Oct 21, 2020 at 8:09 PM Kieran Bingham wrote:
> >>>>> On 21/10/2020 12:03, Umang Jain wrote:
> >>>>>> On 10/21/20 4:25 PM, Kieran Bingham wrote:
> >>>>>>> On 21/10/2020 11:51, Umang Jain wrote:
> >>>>>>>> On 10/21/20 2:49 PM, Kieran Bingham wrote:
> >>>>>>>>> On 21/10/2020 09:08, Umang Jain wrote:
> >>>>>>>>>> Add a basic image thumbnailer for NV12 frames being captured.
> >>>>>>>>>> It shall generate a thumbnail image to be embedded as a part of
> >>>>>>>>>> EXIF metadata of the frame. The output of the thumbnail will still
> >>>>>>>>>> be NV12.
> >>>>>>>>>>
> >>>>>>>>>> Signed-off-by: Umang Jain <email at uajain.com>
> >>>>>>>>>> ---
> >>>>>>>>>> src/android/jpeg/post_processor_jpeg.cpp | 39 +++++++++
> >>>>>>>>>> src/android/jpeg/post_processor_jpeg.h | 3 +
> >>>>>>>>>> src/android/jpeg/thumbnailer.cpp | 100 +++++++++++++++++++++++
> >>>>>>>>>> src/android/jpeg/thumbnailer.h | 40 +++++++++
> >>>>>>>>>> src/android/meson.build | 1 +
> >>>>>>>>>> 5 files changed, 183 insertions(+)
> >>>>>>>>>> create mode 100644 src/android/jpeg/thumbnailer.cpp
> >>>>>>>>>> create mode 100644 src/android/jpeg/thumbnailer.h
> >>>>>>>>>>
> >>>>>>>>>> diff --git a/src/android/jpeg/post_processor_jpeg.cpp
> >>>>>>>>>> b/src/android/jpeg/post_processor_jpeg.cpp
> >>>>>>>>>> index 9d452b7..f5f1f78 100644
> >>>>>>>>>> --- a/src/android/jpeg/post_processor_jpeg.cpp
> >>>>>>>>>> +++ b/src/android/jpeg/post_processor_jpeg.cpp
> >>>>>>>>>> @@ -11,6 +11,7 @@
> >>>>>>>>>> #include "../camera_metadata.h"
> >>>>>>>>>> #include "encoder_libjpeg.h"
> >>>>>>>>>> #include "exif.h"
> >>>>>>>>>> +#include "thumbnailer.h"
> >>>>>>>>>> #include <libcamera/formats.h>
> >>>>>>>>>> @@ -44,6 +45,32 @@ int PostProcessorJpeg::configure(const StreamConfiguration &inCfg,
> >>>>>>>>>> return encoder_->configure(inCfg);
> >>>>>>>>>> }
> >>>>>>>>>> +void PostProcessorJpeg::generateThumbnail(const libcamera::Span<uint8_t> &source,
> >>>>>>>>>> + std::vector<unsigned char> &thumbnail)
> >>>>>>>>>> +{
> >>>>>>>>>> + libcamera::Span<uint8_t> destination;
> >>>>>>>>>> + Thumbnailer thumbnailer;
> >>>>>>>>>> +
> >>>>>>>>>> + thumbnailer.configure(streamSize_, formats::NV12);
> >>>>>>>>>> + libcamera::Size targetSize = thumbnailer.computeThumbnailSize();
> >>>>>>>>>> + thumbnailer.scaleBuffer(source, thumbnail);
> >>>>>>>>>> +
> >>>>>>>>>> + if (thumbnail.data()) {
> >>>>>>>>>> + StreamConfiguration thumbnailCfg;
> >>>>>>>>>> +
> >>>>>>>>>> + std::unique_ptr<EncoderLibJpeg> encoder =
> >>>>>>>>>> + std::make_unique<EncoderLibJpeg>();
> >>>>>>>>>> +
> >>>>>>>>>> + thumbnailCfg.pixelFormat = formats::NV12;
> >>>>>>>>>> + thumbnailCfg.size = targetSize;
> >>>>>>>>>> + encoder->configure(thumbnailCfg);
> >>>>>>>>>
> >>>>>>>>> thumbnail.capacity() might be quite low here.
> >>>>>>>>> We need to make sure the vector is big enough at this point, you might
> >>>>>>>>> need to do something like:
> >>>>>>>>>
> >>>>>>>>> thumbnail.resize(targetSize.width * targetSize.height * 4);
> >>>>>>>>
> >>>>>>>> I am not sure I follow. This is compressing the thumbnail part right? So
> >>>>>>>> thumbnail is the "source" for the encoder here (Please refer to
> >>>>>>>> ->encode() below) and why would we resize the source(i.e. 'thumbnail')?
> >>>>>>>
> >>>>>>> Ugh, sorry - tired eyes indeed. Yes, please ignore those comments.
> >>>>>>>
> >>>>>>>>> Really we should obtain that size from the encoder. I thought we had a
> >>>>>>>>> helper to do that already, but it seems we don't.
> >>>>>>>>> We could/should add Encoder::maxOutput(); which would ask the encoder
> >>>>>>>>> "For your given configuration, what is the maximum number of bytes you
> >>>>>>>>> might output".
> >>>>>>>>>
> >>>>>>>>> Then of course we'd do:
> >>>>>>>>> thumbnail.resize(encoder.maxOutput());
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>> + int jpeg_size = encoder->encode({ thumbnail.data(), thumbnail.capacity() },
> >>>>>>>>>> + destination, { });
> >>>>>>>>
> >>>>>>>> As said above, thumbnail is the source here. The compressed thumbnail
> >>>>>>>> output is carried in destination. And, destination is a local span<>
> >>>>>>>> here. Does it makes sense?
> >>>>>>>
> >>>>>>> Yes, I had mixed up source and destination I'm sorry.
> >>>>>>>
> >>>>>>> So who/where is allocating the destination?
> >>>>>>
> >>>>>> Ah, that's a good question and I think I just found a bug. I currently
> >>>>>> pass on a locally allocated Span<> 'destination' to the encoder. But
> >>>>>> there is no explicit allocation here. I think I need to point / create
> >>>>>> the Span with the output size / bytes and then pass to the encode().
> >>>>>
> >>>>> :-) Ok - then please apply all of my earlier comments to that part
> >>>>> instead. ;-)
> >>>>>
> >>>>> We certainly need somewhere to store the compressed image, to be able to
> >>>>> pass it into the exif :-D
> >>>>>
> >>>>>>>>>> + LOG(JPEG, Info) << "Thumbnail compress returned "
> >>>>>>>>>> + << jpeg_size << " bytes";
> >>>>>>>>>
> >>>>>>>>> And I presume we could then do an:
> >>>>>>>>> thumbnail.resize(jpeg_size);
> >>>>>>>>>
> >>>>>>>>> here to update the thumbnail with the correct size. I'd be weary of the
> >>>>>>>>> resize operations doing lots of re-allocations though, so perhaps we
> >>>>>>>>> want to minimize that. But lets get to something that works first
> >>>>>>>>> before worrying about optimising.
> >>>>>>>>>
> >>>>>>>>>> + }
> >>>>>>>>>> +}
> >>>>>>>>>> +
> >>>>>>>>>> int PostProcessorJpeg::process(const libcamera::FrameBuffer *source,
> >>>>>>>>>> const libcamera::Span<uint8_t> &destination,
> >>>>>>>>>> CameraMetadata *metadata)
> >>>>>>>>>> @@ -73,6 +100,18 @@ int PostProcessorJpeg::process(const
> >>>>>>>>>> libcamera::FrameBuffer *source,
> >>>>>>>>>> return jpeg_size;
> >>>>>>>>>> }
> >>>>>>>>>> + std::vector<unsigned char> thumbnail;
> >>>>>>>>>
> >>>>>>>>> You need to resize this somewhere.
> >>>>>>>>> Edit: Now seen a better place above ;-)
> >>>>>>>>
> >>>>>>>> The thumbnail vector just gets resized in Thumbnailer::scaleBuffer(). :)
> >>>>>>>> We just pass in to the thumbnailer, while keeping its ownership in
> >>>>>>>> PostProcessorJpeg.
> >>>>>>>>
> >>>>>>>>>> + generateThumbnail(destination, thumbnail);
> >>>>>>>>>> + /*
> >>>>>>>>>> + * \todo: Write the compressed thumbnail to a file for inspection.
> >>>>>>>>>> + * (I) Check if we can still write the thumbnail to EXIF here.
> >>>>>>>>>> + * If not, we might need to move the thumbnailer logic to encoder.
> >>>>>>>>>> + * And if we do that, first we need to make sure we get can
> >>>>>>>>>> + * compressed data written to destination first before calling
> >>>>>>>>>> + * jpeg_finish_compress operation somehow. Thumbnailing will
> >>>>>>>>>> + * only occur if we have compressed data available first.
> >>>>>>>>>> + */
> >>>>>>>>>> +
> >>>>>>>>>> /*
> >>>>>>>>>> * Fill in the JPEG blob header.
> >>>>>>>>>> *
> >>>>>>>>>> diff --git a/src/android/jpeg/post_processor_jpeg.h
> >>>>>>>>>> b/src/android/jpeg/post_processor_jpeg.h
> >>>>>>>>>> index 62c8650..05601ee 100644
> >>>>>>>>>> --- a/src/android/jpeg/post_processor_jpeg.h
> >>>>>>>>>> +++ b/src/android/jpeg/post_processor_jpeg.h
> >>>>>>>>>> @@ -28,6 +28,9 @@ public:
> >>>>>>>>>> CameraMetadata *metadata) override;
> >>>>>>>>>> private:
> >>>>>>>>>> + void generateThumbnail(const libcamera::Span<uint8_t> &source,
> >>>>>>>>>> + std::vector<unsigned char> &thumbnail);
> >>>>>>>>>> +
> >>>>>>>>>> CameraDevice *cameraDevice_;
> >>>>>>>>>> std::unique_ptr<Encoder> encoder_;
> >>>>>>>>>> libcamera::Size streamSize_;
> >>>>>>>>>> diff --git a/src/android/jpeg/thumbnailer.cpp
> >>>>>>>>>> b/src/android/jpeg/thumbnailer.cpp
> >>>>>>>>>> new file mode 100644
> >>>>>>>>>> index 0000000..3163576
> >>>>>>>>>> --- /dev/null
> >>>>>>>>>> +++ b/src/android/jpeg/thumbnailer.cpp
> >>>>>>>>>> @@ -0,0 +1,100 @@
> >>>>>>>>>> +/* SPDX-License-Identifier: GPL-2.0-or-later */
> >>>>>>>>>> +/*
> >>>>>>>>>> + * Copyright (C) 2020, Google Inc.
> >>>>>>>>>> + *
> >>>>>>>>>> + * thumbnailer.cpp - Basic image thumbnailer from NV12
> >>>>>>>>>> + */
> >>>>>>>>>> +
> >>>>>>>>>> +#include "thumbnailer.h"
> >>>>>>>>>> +
> >>>>>>>>>> +#include <libcamera/formats.h>
> >>>>>>>>>> +
> >>>>>>>>>> +#include "libcamera/internal/file.h"
> >>>>>>>>>> +#include "libcamera/internal/log.h"
> >>>>>>>>>> +
> >>>>>>>>>> +using namespace libcamera;
> >>>>>>>>>> +
> >>>>>>>>>> +LOG_DEFINE_CATEGORY(Thumbnailer)
> >>>>>>>>>> +
> >>>>>>>>>> +Thumbnailer::Thumbnailer()
> >>>>>>>>>> + : validConfiguration_(false)
> >>>>>>>>>> +{
> >>>>>>>>>> +}
> >>>>>>>>>> +
> >>>>>>>>>> +void Thumbnailer::configure(const Size &sourceSize, PixelFormat pixelFormat)
> >>>>>>>>>> +{
> >>>>>>>>>> + sourceSize_ = sourceSize;
> >>>>>>>>>> + pixelFormat_ = pixelFormat;
> >>>>>>>>>> +
> >>>>>>>>>> + if (pixelFormat_ != formats::NV12) {
> >>>>>>>>>> + LOG (Thumbnailer, Error) << "Failed to configure: Pixel Format "
> >>>>>>>>>> + << pixelFormat_.toString() << " unsupported.";
> >>>>>>>>>> + return;
> >>>>>>>>>> + }
> >>>>>>>>>> +
> >>>>>>>>>> + validConfiguration_ = true;
> >>>>>>>>>> +}
> >>>>>>>>>> +
> >>>>>>>>>> +/*
> >>>>>>>>>> + * The Exif specification recommends the width of the thumbnail to be a
> >>>>>>>>>> + * mutiple of 16 (section 4.8.1). Hence, compute the corresponding height
> >>>>>>>>>> + * keeping the aspect ratio same as of the source.
> >>>>>>>>>> + */
> >>>>>>>>>> +Size Thumbnailer::computeThumbnailSize()
> >>>>>>>>>> +{
> >>>>>>>>>> + unsigned int targetHeight;
> >>>>>>>>>> + unsigned int targetWidth = 160;
> >>>>>>>>>> +
> >>>>>>>>>> + targetHeight = targetWidth * sourceSize_.height / sourceSize_.width;
> >>>>>>>>>> +
> >>>>>>>>>> + if (targetHeight & 1)
> >>>>>>>>>> + targetHeight++;
> >>>>>>>>>> +
> >>>>>>>>>> + return Size(targetWidth, targetHeight);
> >>>>>>>>>> +}
> >>>>>>>>>> +
> >>>>>>>>>> +void
> >>>>>>>>>> +Thumbnailer::scaleBuffer(const libcamera::Span<uint8_t> &source,
> >>>>>>>>>> + std::vector<unsigned char> &destination)
> >>>>>>>>>> +{
> >>>>>>>>>> + if (!validConfiguration_) {
> >>>>>>>>>> + LOG(Thumbnailer, Error) << "config is unconfigured or invalid.";
> >>>>>>>>>> + return;
> >>>>>>>>>> + }
> >>>>>>>>>> +
> >>>>>>>>>> + targetSize_ = computeThumbnailSize();
> >>>>>>>>>> +
> >>>>>>>>>> + const unsigned int sw = sourceSize_.width;
> >>>>>>>>>> + const unsigned int sh = sourceSize_.height;
> >>>>>>>>>> + const unsigned int tw = targetSize_.width;
> >>>>>>>>>> + const unsigned int th = targetSize_.height;
> >>>>>>>>>> +
> >>>>>>>>>> + /* Image scaling block implementing nearest-neighbour algorithm. */
> >>>>>>>>>> + unsigned char *src = static_cast<unsigned char *>(source.data());
> >>>>>>>>>> + unsigned char *src_c = src + sh * sw;
> >>>>>>>>>> + unsigned char *src_cb, *src_cr;
> >>>>>>>>>> +
> >>>>>>>>>> + size_t dstSize = (th * tw) + ((th/2) * tw);
> >>>>>>>>>> + destination.reserve(dstSize);
> >>>>>>>>>> + unsigned char *dst = destination.data();
> >>>>>>>>>> + unsigned char *dst_c = dst + th * tw;
> >>>>>>>>>> +
> >>>>>>>>>> + for (unsigned int y = 0; y < th; y+=2) {
> >>>>>>>>>> + unsigned int sourceY = (sh*y + th/2) / th;
> >>>>>>>>>> +
> >>>>>>>>>> + src_cb = src_c + (sourceY/2) * sw + 0;
> >>>>>>>>>> + src_cr = src_c + (sourceY/2) * sw + 1;
> >>>>>>>>>> +
> >>>>>>>>>> + for (unsigned int x = 0; x < tw; x+=2) {
> >>>>>>>>>> + unsigned int sourceX = (sw*x + tw/2) / tw;
> >>>>>>>>>> +
> >>>>>>>>>> + dst[y * tw + x] = src[sw * sourceY + sourceX];
> >>>>>>>>>> + dst[(y+1) * tw + x] = src[sw * (sourceY+1) + sourceX];
> >>>>>>>>>> + dst[y * tw + (x+1)] = src[sw * sourceY + (sourceX+1)];
> >>>>>>>>>> + dst[(y+1) * tw + (x+1)] = src[sw * (sourceY+1) + (sourceX+1)];
> >>>>>>>>>> +
> >>>>>>>>>> + dst_c[(y/2) * tw + x + 0] = src_cb[(sourceX/2) * 2];
> >>>>>>>>>> + dst_c[(y/2) * tw + x + 1] = src_cr[(sourceX/2) * 2];
> >>>>>>>>>> + }
> >>>>>>>>>> + }
> >>>>>>>>>> +}
> >>>>>>>>>> diff --git a/src/android/jpeg/thumbnailer.h
> >>>>>>>>>> b/src/android/jpeg/thumbnailer.h
> >>>>>>>>>> new file mode 100644
> >>>>>>>>>> index 0000000..bab9855
> >>>>>>>>>> --- /dev/null
> >>>>>>>>>> +++ b/src/android/jpeg/thumbnailer.h
> >>>>>>>>>> @@ -0,0 +1,40 @@
> >>>>>>>>>> +/* SPDX-License-Identifier: GPL-2.0-or-later */
> >>>>>>>>>> +/*
> >>>>>>>>>> + * Copyright (C) 2020, Google Inc.
> >>>>>>>>>> + *
> >>>>>>>>>> + * thumbnailer.h - Basic image thumbnailer from NV12
> >>>>>>>>>> + */
> >>>>>>>>>> +#ifndef __ANDROID_JPEG_THUMBNAILER_H__
> >>>>>>>>>> +#define __ANDROID_JPEG_THUMBNAILER_H__
> >>>>>>>>>> +
> >>>>>>>>>> +#include <libcamera/geometry.h>
> >>>>>>>>>> +
> >>>>>>>>>> +#include "libcamera/internal/buffer.h"
> >>>>>>>>>> +#include "libcamera/internal/formats.h"
> >>>>>>>>>> +
> >>>>>>>>>> +class Thumbnailer
> >>>>>>>>>> +{
> >>>>>>>>>> +public:
> >>>>>>>>>> + Thumbnailer();
> >>>>>>>>>> +
> >>>>>>>>>> + void configure(const libcamera::Size &sourceSize,
> >>>>>>>>>> + libcamera::PixelFormat pixelFormat);
> >>>>>>>>>> +
> >>>>>>>>>> + /*
> >>>>>>>>>> + * \todo: Discuss if we can return targetSize_ via configure() or
> >>>>>>>>>> + * scaleBuffer(). We need targetSize_ to re-encode the scaled
> >>>>>>>>>> buffer
> >>>>>>>>>> + * via encoder in PostProcssorJpeg::writeThumbnail().
> >>>>>>>>>> + */
> >>>>>>>>>> + libcamera::Size computeThumbnailSize();
> >>>>>>>>>> + void scaleBuffer(const libcamera::Span<uint8_t> &source,
> >>>>>>>>>> + std::vector<unsigned char> &dest);
> >>>>>>>>>> +
> >>>>>>>>>> +private:
> >>>>>>>>>> + libcamera::PixelFormat pixelFormat_;
> >>>>>>>>>> + libcamera::Size sourceSize_;
> >>>>>>>>>> + libcamera::Size targetSize_;
> >>>>>>>>>> +
> >>>>>>>>>> + bool validConfiguration_;
> >>>>>>>>>> +};
> >>>>>>>>>> +
> >>>>>>>>>> +#endif /* __ANDROID_JPEG_THUMBNAILER_H__ */
> >>>>>>>>>> diff --git a/src/android/meson.build b/src/android/meson.build
> >>>>>>>>>> index 5a01bea..3905e2f 100644
> >>>>>>>>>> --- a/src/android/meson.build
> >>>>>>>>>> +++ b/src/android/meson.build
> >>>>>>>>>> @@ -25,6 +25,7 @@ android_hal_sources = files([
> >>>>>>>>>> 'jpeg/encoder_libjpeg.cpp',
> >>>>>>>>>> 'jpeg/exif.cpp',
> >>>>>>>>>> 'jpeg/post_processor_jpeg.cpp',
> >>>>>>>>>> + 'jpeg/thumbnailer.cpp',
> >>>>>>>>>> ])
> >>>>>>>>>> android_camera_metadata_sources = files([
> >>>>>>>>>>
> >
>
> --
> Regards
> --
> Kieran
More information about the libcamera-devel
mailing list