[libcamera-devel] [RFC PATCH v2 3/4] libcamera: converter: add software converter

Jacopo Mondi jacopo.mondi at ideasonboard.com
Thu Sep 21 12:40:31 CEST 2023


Hi Andrey

Sw converters will be discussed in detail, so I'll only go through the
trivial/cosmetics things and leave discussion on the architecture for
the face-2-face meeting in Paris

On Wed, Sep 20, 2023 at 06:19:20PM +0300, Andrey Konovalov via libcamera-devel wrote:
> Use the Converter interface to implement bilinear Bayer demosaic
> filtering and very simple AWB on CPU.
>
> The only output format currently supported is RGB888.
>
> Signed-off-by: Andrey Konovalov <andrey.konovalov at linaro.org>
> ---
>  .../internal/converter/converter_softw.h      |  90 ++++
>  .../libcamera/internal/converter/meson.build  |   1 +
>  src/libcamera/converter/converter_softw.cpp   | 410 ++++++++++++++++++
>  src/libcamera/converter/meson.build           |   3 +-
>  4 files changed, 503 insertions(+), 1 deletion(-)
>  create mode 100644 include/libcamera/internal/converter/converter_softw.h
>  create mode 100644 src/libcamera/converter/converter_softw.cpp
>
> diff --git a/include/libcamera/internal/converter/converter_softw.h b/include/libcamera/internal/converter/converter_softw.h
> new file mode 100644
> index 00000000..eff9d6b2
> --- /dev/null
> +++ b/include/libcamera/internal/converter/converter_softw.h
> @@ -0,0 +1,90 @@
> +/* SPDX-License-Identifier: LGPL-2.1-or-later */
> +/*
> + * Copyright (C) 2023, Linaro Ltd
> + *
> + * converter_softw.h - interface of software converter (runs 100% on CPU)
> + */
> +

#include <functional>

> +#include <map>

#include <memory>
#include <string>

> +#include <tuple>
> +#include <vector>
> +
> +#include <libcamera/base/log.h>
> +#include <libcamera/base/signal.h>
> +#include <libcamera/base/thread.h>
> +
> +#include <libcamera/pixel_format.h>
> +
> +#include "libcamera/internal/converter.h"
> +
> +namespace libcamera {
> +
> +class FrameBuffer;
> +class MediaDevice;
> +class Size;
> +class SizeRange;
> +struct StreamConfiguration;
> +
> +class SwConverter : public Converter
> +{
> +public:
> +	SwConverter([[maybe_unused]] MediaDevice *media)

This is not super-nice, but I don't have anything better to suggest at
thsi time unless we completely revisit the implementation of [2/4]. I
guess we can live with this even if it might indicate we should be
doing better.

I had a go at using SFINAE to selectively compile in the version of
ConverterFactory::createInstance() that passes in a MediaDevice
selectively, but after 30 minutes I gave up :)


> +		: Converter() {}
> +
> +	int loadConfiguration([[maybe_unused]] const std::string &filename) { return 0; }
> +	bool isValid() const { return true; }
> +
> +	std::vector<PixelFormat> formats(PixelFormat input);
> +	SizeRange sizes(const Size &input);
> +
> +	std::tuple<unsigned int, unsigned int>
> +	strideAndFrameSize(const PixelFormat &pixelFormat, const Size &size);
> +
> +	int configure(const StreamConfiguration &inputCfg,
> +		      const std::vector<std::reference_wrapper<StreamConfiguration>> &outputCfg);
> +	int exportBuffers(unsigned int ouput, unsigned int count,
> +			  std::vector<std::unique_ptr<FrameBuffer>> *buffers);
> +
> +	void process(FrameBuffer *input, FrameBuffer *output);
> +	int start();
> +	void stop();
> +
> +	int queueBuffers(FrameBuffer *input,
> +			 const std::map<unsigned int, FrameBuffer *> &outputs);
> +
> +private:
> +	class Isp : public Object
> +	{
> +	public:
> +		Isp(SwConverter *converter)
> +			: converter_(converter) {}
> +
> +		int configure(const StreamConfiguration &inputCfg,
> +			      const StreamConfiguration &outputCfg);
> +		int exportBuffers(unsigned int count,
> +				  std::vector<std::unique_ptr<FrameBuffer>> *buffers);
> +		void process(FrameBuffer *input, FrameBuffer *output);
> +		int start();
> +		void stop();
> +
> +	private:
> +		void debayer(uint8_t *dst, const uint8_t *src);
> +
> +		SwConverter *converter_;
> +
> +		Thread thread_;
> +
> +		unsigned int width_;
> +		unsigned int height_;
> +		unsigned int stride_;
> +		Point red_shift_;
> +
> +		unsigned long rNumerat_, rDenomin_; /* red gain for AWB */
> +		unsigned long bNumerat_, bDenomin_; /* blue gain for AWB */
> +		unsigned long gNumerat_, gDenomin_; /* green gain for AWB */
> +	};
> +
> +	std::unique_ptr<Isp> isp_;
> +};
> +
> +} /* namespace libcamera */
> diff --git a/include/libcamera/internal/converter/meson.build b/include/libcamera/internal/converter/meson.build
> index 891e79e7..843a0483 100644
> --- a/include/libcamera/internal/converter/meson.build
> +++ b/include/libcamera/internal/converter/meson.build
> @@ -1,5 +1,6 @@
>  # SPDX-License-Identifier: CC0-1.0
>
>  libcamera_internal_headers += files([
> +    'converter_softw.h',
>      'converter_v4l2_m2m.h',
>  ])
> diff --git a/src/libcamera/converter/converter_softw.cpp b/src/libcamera/converter/converter_softw.cpp
> new file mode 100644
> index 00000000..ad7e6f3d
> --- /dev/null
> +++ b/src/libcamera/converter/converter_softw.cpp
> @@ -0,0 +1,410 @@
> +/* SPDX-License-Identifier: LGPL-2.1-or-later */
> +/*
> + * Copyright (C) 2023, Linaro Ltd
> + *
> + * converter_softw.h - interface of software converter (runs 100% on CPU)
> + */
> +
> +#include <sys/mman.h>
> +#include <sys/types.h>
> +#include <tuple>
> +#include <unistd.h>
> +#include <vector>
> +
> +#include <libcamera/formats.h>
> +#include <libcamera/stream.h>
> +
> +#include "libcamera/internal/bayer_format.h"
> +#include "libcamera/internal/framebuffer.h"
> +#include "libcamera/internal/mapped_framebuffer.h"
> +
> +#include "libcamera/internal/converter/converter_softw.h"

Include this first to make sure it's self contained

And you can remove a few duplicated includes like <vector>

> +
> +namespace libcamera {
> +
> +LOG_DECLARE_CATEGORY(Converter)
> +
> +std::vector<PixelFormat> SwConverter::formats(PixelFormat input)
> +{
> +	std::vector<PixelFormat> pixelFormats;
> +	BayerFormat inputFormat = BayerFormat::fromPixelFormat(input);
> +
> +	/* Only RAW10P is currently supported */

Do you care about the bayer ordering or any permutation is fine
(SRGGB10, SBGGR10 etc etc)

> +	if (inputFormat.bitDepth == 10 && inputFormat.packing == BayerFormat::Packing::CSI2)
> +		pixelFormats.push_back(formats::RGB888);
> +
> +	if (pixelFormats.empty())
> +		LOG(Converter, Info)
> +			<< "Unsupported input format " << input.toString();
> +
> +	return pixelFormats;

Isn't
	BayerFormat inputFormat = BayerFormat::fromPixelFormat(input);
	if (inputFormat.bitDepth != 10 ||
	    inputFormat.packing != BayerFormat::Packing::CSI2) {
		LOG(Converter, Info)
			<< "Unsupported input format " << input.toString();
                return {};
	}

        return { formats::RGB888 };

a bit simpler ?


> +}
> +
> +SizeRange SwConverter::sizes(const Size &input)
> +{
> +	if (input.width < 2 || input.height < 2) {
> +		LOG(Converter, Error)
> +			<< "Input format size too small: " << input.toString();
> +		return {};
> +	}
> +
> +	return SizeRange(Size(input.width - 2, input.height - 2));
> +}
> +
> +std::tuple<unsigned int, unsigned int>
> +SwConverter::strideAndFrameSize(const PixelFormat &pixelFormat,
> +				const Size &size)
> +{
> +	/* Only RGB888 output is currently supported */
> +	if (pixelFormat == formats::RGB888) {
> +		int stride = size.width * 3;
> +		return std::make_tuple(stride, stride * (size.height));
> +	}
> +
> +	return std::make_tuple(0, 0);

What about:

	/* Only RGB888 output is currently supported */
	if (pixelFormat != formats::RGB888)
		return {};

	unsigned int stride = size.width * 3;
	return std::make_tuple(stride, stride * size.height);

Up to you, really.

> +}
> +
> +int SwConverter::configure(const StreamConfiguration &inputCfg,
> +			   const std::vector<std::reference_wrapper<StreamConfiguration>> &outputCfgs)
> +{
> +	if (outputCfgs.size() != 1) {
> +		LOG(Converter, Error)
> +			<< "Unsupported number of output streams: "
> +			<< outputCfgs.size();
> +		return -EINVAL;
> +	}
> +
> +	isp_ = std::make_unique<SwConverter::Isp>(this);

Why do you re-create the isp_ at every configuration ?

I think the actual soft ISP implementation could be discussed f2f and
I'll stop here with the nitpicking ;)

Thanks
  j

> +
> +	return isp_->configure(inputCfg, outputCfgs[0]);
> +}
> +
> +int SwConverter::Isp::configure(const StreamConfiguration &inputCfg,
> +		const StreamConfiguration &outputCfg)
> +{
> +	BayerFormat bayerFormat =
> +		BayerFormat::fromPixelFormat(inputCfg.pixelFormat);
> +	width_ = inputCfg.size.width;
> +	height_ = inputCfg.size.height;
> +	stride_ = inputCfg.stride;
> +
> +	if (bayerFormat.bitDepth != 10 ||
> +	    bayerFormat.packing != BayerFormat::Packing::CSI2 ||
> +	    width_ < 2 || height_ < 2) {
> +		LOG(Converter, Error) << "Input format "
> +				      << inputCfg.size << "-"
> +				      << inputCfg.pixelFormat
> +				      << "not supported";
> +		return -EINVAL;
> +	}
> +
> +	switch (bayerFormat.order) {
> +	case BayerFormat::BGGR:
> +		red_shift_ = Point(0, 0);
> +		break;
> +	case BayerFormat::GBRG:
> +		red_shift_ = Point(1, 0);
> +		break;
> +	case BayerFormat::GRBG:
> +		red_shift_ = Point(0, 1);
> +		break;
> +	case BayerFormat::RGGB:
> +	default:
> +		red_shift_ = Point(1, 1);
> +		break;
> +	}
> +
> +	if (outputCfg.size.width != width_ - 2 ||
> +	    outputCfg.size.height != height_ - 2 ||
> +	    outputCfg.stride != (width_ - 2) * 3 ||
> +	    outputCfg.pixelFormat != formats::RGB888) {
> +		LOG(Converter, Error)
> +			<< "Output format not supported";
> +		return -EINVAL;
> +	}
> +
> +	LOG(Converter, Info) << "SwConverter configuration: "
> +			     << inputCfg.size << "-" << inputCfg.pixelFormat
> +			     << " -> "
> +			     << outputCfg.size << "-" << outputCfg.pixelFormat;
> +
> +	/* set r/g/b gains to 1.0 until frame data collected */
> +	rNumerat_ = rDenomin_ = 1;
> +	bNumerat_ = bDenomin_ = 1;
> +	gNumerat_ = gDenomin_ = 1;
> +
> +	return 0;
> +}
> +
> +int SwConverter::exportBuffers(unsigned int output, unsigned int count,
> +			       std::vector<std::unique_ptr<FrameBuffer>> *buffers)
> +{
> +	/* single output for now */
> +	if (output >= 1)
> +		return -EINVAL;
> +
> +	return isp_->exportBuffers(count, buffers);
> +}
> +
> +int SwConverter::Isp::exportBuffers(unsigned int count,
> +				    std::vector<std::unique_ptr<FrameBuffer>> *buffers)
> +{
> +	/* V4L2_PIX_FMT_BGR24 aka 'BGR3' for output: */
> +	unsigned int bufSize = (height_ - 2) * (width_ - 2) * 3;
> +
> +	for (unsigned int i = 0; i < count; i++) {
> +		std::string name = "frame-" + std::to_string(i);
> +
> +		const int ispFd = memfd_create(name.c_str(), 0);
> +		int ret = ftruncate(ispFd, bufSize);
> +		if (ret < 0) {
> +			LOG(Converter, Error) << "ftruncate() for memfd failed "
> +					      << strerror(-ret);
> +			return ret;
> +		}
> +
> +		FrameBuffer::Plane outPlane;
> +		outPlane.fd = SharedFD(std::move(ispFd));
> +		outPlane.offset = 0;
> +		outPlane.length = bufSize;
> +
> +		std::vector<FrameBuffer::Plane> planes{ outPlane };
> +		buffers->emplace_back(std::make_unique<FrameBuffer>(std::move(planes)));
> +	}
> +
> +	return count;
> +}
> +
> +int SwConverter::Isp::start()
> +{
> +	moveToThread(&thread_);
> +	thread_.start();
> +	return 0;
> +}
> +
> +void SwConverter::Isp::stop()
> +{
> +	thread_.exit();
> +	thread_.wait();
> +}
> +
> +int SwConverter::start()
> +{
> +	return isp_->start();
> +}
> +
> +void SwConverter::stop()
> +{
> +	return isp_->stop();
> +}
> +
> +int SwConverter::queueBuffers(FrameBuffer *input,
> +			      const std::map<unsigned int, FrameBuffer *> &outputs)
> +{
> +	unsigned int mask = 0;
> +
> +	/*
> +	 * Validate the outputs as a sanity check: at least one output is
> +	 * required, all outputs must reference a valid stream and no two
> +	 * outputs can reference the same stream.
> +	 */
> +	if (outputs.empty())
> +		return -EINVAL;
> +
> +	for (auto [index, buffer] : outputs) {
> +		if (!buffer)
> +			return -EINVAL;
> +		if (index >= 1) /* only single stream atm */
> +			return -EINVAL;
> +		if (mask & (1 << index))
> +			return -EINVAL;
> +
> +		mask |= 1 << index;
> +	}
> +
> +	process(input, outputs.at(0));
> +
> +	return 0;
> +}
> +
> +void SwConverter::process(FrameBuffer *input, FrameBuffer *output)
> +{
> +	isp_->invokeMethod(&SwConverter::Isp::process,
> +			   ConnectionTypeQueued, input, output);
> +}
> +
> +void SwConverter::Isp::process(FrameBuffer *input, FrameBuffer *output)
> +{
> +	/* Copy metadata from the input buffer */
> +	FrameMetadata &metadata = output->_d()->metadata();
> +	metadata.status = input->metadata().status;
> +	metadata.sequence = input->metadata().sequence;
> +	metadata.timestamp = input->metadata().timestamp;
> +
> +	MappedFrameBuffer in(input, MappedFrameBuffer::MapFlag::Read);
> +	MappedFrameBuffer out(output, MappedFrameBuffer::MapFlag::Write);
> +	if (!in.isValid() || !out.isValid()) {
> +		LOG(Converter, Error) << "mmap-ing buffer(s) failed";
> +		metadata.status = FrameMetadata::FrameError;
> +		converter_->outputBufferReady.emit(output);
> +		converter_->inputBufferReady.emit(input);
> +		return;
> +	}
> +
> +	debayer(out.planes()[0].data(), in.planes()[0].data());
> +	metadata.planes()[0].bytesused = out.planes()[0].size();
> +
> +	converter_->outputBufferReady.emit(output);
> +	converter_->inputBufferReady.emit(input);
> +}
> +
> +void SwConverter::Isp::debayer(uint8_t *dst, const uint8_t *src)
> +{
> +	/* RAW10P input format is assumed */
> +
> +	/* output buffer is in BGR24 format and is of (width-2)*(height-2) */
> +
> +	int w_out = width_ - 2;
> +	int h_out = height_ - 2;
> +
> +	unsigned long sumR = 0;
> +	unsigned long sumB = 0;
> +	unsigned long sumG = 0;
> +
> +	for (int y = 0; y < h_out; y++) {
> +		const uint8_t *pin_base = src + (y + 1) * stride_;
> +		uint8_t *pout = dst + y * w_out * 3;
> +		int phase_y = (y + red_shift_.y) % 2;
> +
> +		for (int x = 0; x < w_out; x++) {
> +			int phase_x = (x + red_shift_.x) % 2;
> +			int phase = 2 * phase_y + phase_x;
> +
> +			/* x part of the offset in the input buffer: */
> +			int x_m1 = x + x / 4;		/* offset for (x-1) */
> +			int x_0 = x + 1 + (x + 1) / 4;	/* offset for x */
> +			int x_p1 = x + 2 + (x + 2) / 4;	/* offset for (x+1) */
> +			/* the colour component value to write to the output */
> +			unsigned val;
> +
> +			switch (phase) {
> +			case 0: /* at R pixel */
> +				/* blue: ((-1,-1)+(1,-1)+(-1,1)+(1,1)) / 4 */
> +				val = ( *(pin_base + x_m1 - stride_)
> +					+ *(pin_base + x_p1 - stride_)
> +					+ *(pin_base + x_m1 + stride_)
> +					+ *(pin_base + x_p1 + stride_) ) >> 2;
> +				val = val * bNumerat_ / bDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* green: ((0,-1)+(-1,0)+(1,0)+(0,1)) / 4 */
> +				val = ( *(pin_base + x_0 - stride_)
> +					+ *(pin_base + x_p1)
> +					+ *(pin_base + x_m1)
> +					+ *(pin_base + x_0 + stride_) ) >> 2;
> +				val = val * gNumerat_ / gDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* red: (0,0) */
> +				val = *(pin_base + x_0);
> +				sumR += val;
> +				val = val * rNumerat_ / rDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				break;
> +			case 1: /* at Gr pixel */
> +				/* blue: ((0,-1) + (0,1)) / 2 */
> +				val = ( *(pin_base + x_0 - stride_)
> +					+ *(pin_base + x_0 + stride_) ) >> 1;
> +				val = val * bNumerat_ / bDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* green: (0,0) */
> +				val = *(pin_base + x_0);
> +				sumG += val;
> +				val = val * gNumerat_ / gDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* red: ((-1,0) + (1,0)) / 2 */
> +				val = ( *(pin_base + x_m1)
> +					+ *(pin_base + x_p1) ) >> 1;
> +				val = val * rNumerat_ / rDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				break;
> +			case 2: /* at Gb pixel */
> +				/* blue: ((-1,0) + (1,0)) / 2 */
> +				val = ( *(pin_base + x_m1)
> +					+ *(pin_base + x_p1) ) >> 1;
> +				val = val * bNumerat_ / bDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* green: (0,0) */
> +				val = *(pin_base + x_0);
> +				sumG += val;
> +				val = val * gNumerat_ / gDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* red: ((0,-1) + (0,1)) / 2 */
> +				val = ( *(pin_base + x_0 - stride_)
> +					+ *(pin_base + x_0 + stride_) ) >> 1;
> +				val = val * rNumerat_ / rDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				break;
> +			default: /* at B pixel */
> +				/* blue: (0,0) */
> +				val = *(pin_base + x_0);
> +				sumB += val;
> +				val = val * bNumerat_ / bDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* green: ((0,-1)+(-1,0)+(1,0)+(0,1)) / 4 */
> +				val = ( *(pin_base + x_0 - stride_)
> +					+ *(pin_base + x_p1)
> +					+ *(pin_base + x_m1)
> +					+ *(pin_base + x_0 + stride_) ) >> 2;
> +				val = val * gNumerat_ / gDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +				/* red: ((-1,-1)+(1,-1)+(-1,1)+(1,1)) / 4 */
> +				val = ( *(pin_base + x_m1 - stride_)
> +					+ *(pin_base + x_p1 - stride_)
> +					+ *(pin_base + x_m1 + stride_)
> +					+ *(pin_base + x_p1 + stride_) ) >> 2;
> +				val = val * rNumerat_ / rDenomin_;
> +				*pout++ = (uint8_t)std::min(val, 0xffU);
> +			}
> +		}
> +	}
> +
> +	/* calculate red and blue gains for simple AWB */
> +	LOG(Converter, Debug) << "sumR = " << sumR
> +			      << ", sumB = " << sumB << ", sumG = " << sumG;
> +
> +	sumG /= 2; /* the number of G pixels is twice as big vs R and B ones */
> +
> +	/* normalize red, blue, and green sums to fit into 22-bit value */
> +	unsigned long fRed = sumR / 0x400000;
> +	unsigned long fBlue = sumB / 0x400000;
> +	unsigned long fGreen = sumG / 0x400000;
> +	unsigned long fNorm = std::max({ 1UL, fRed, fBlue, fGreen });
> +	sumR /= fNorm;
> +	sumB /= fNorm;
> +	sumG /= fNorm;
> +
> +	LOG(Converter, Debug) << "fNorm = " << fNorm;
> +	LOG(Converter, Debug) << "Normalized: sumR = " << sumR
> +			      << ", sumB= " << sumB << ", sumG = " << sumG;
> +
> +	/* make sure red/blue gains never exceed approximately 256 */
> +	unsigned long minDenom;
> +	rNumerat_ = (sumR + sumB + sumG) / 3;
> +	minDenom = rNumerat_ / 0x100;
> +	rDenomin_ = std::max(minDenom, sumR);
> +	bNumerat_ = rNumerat_;
> +	bDenomin_ = std::max(minDenom, sumB);
> +	gNumerat_ = rNumerat_;
> +	gDenomin_ = std::max(minDenom, sumG);
> +
> +	LOG(Converter, Debug) << "rGain = [ "
> +			      << rNumerat_ << " / " << rDenomin_
> +			      << " ], bGain = [ " << bNumerat_ << " / " << bDenomin_
> +			      << " ], gGain = [ " << gNumerat_ << " / " << gDenomin_
> +			      << " (minDenom = " << minDenom << ")";
> +}
> +
> +static std::initializer_list<std::string> compatibles = {};
> +
> +REGISTER_CONVERTER("linaro-sw-converter", SwConverter, compatibles)
> +
> +} /* namespace libcamera */
> diff --git a/src/libcamera/converter/meson.build b/src/libcamera/converter/meson.build
> index 2aa72fe4..1f1e0ea4 100644
> --- a/src/libcamera/converter/meson.build
> +++ b/src/libcamera/converter/meson.build
> @@ -1,5 +1,6 @@
>  # SPDX-License-Identifier: CC0-1.0
>
>  libcamera_sources += files([
> -        'converter_v4l2_m2m.cpp'
> +        'converter_softw.cpp',
> +        'converter_v4l2_m2m.cpp',
>  ])
> --
> 2.34.1
>


More information about the libcamera-devel mailing list