]> git.sesse.net Git - nageru/blobdiff - nageru/mjpeg_encoder.cpp
Split out some work from the MJPEG encoder constructor.
[nageru] / nageru / mjpeg_encoder.cpp
index 9ddb9ff15af20598ead26b5b9a6c282f16861e6c..867414d44127d5019cca822102de519ef3f6bcf8 100644 (file)
@@ -16,24 +16,25 @@ extern "C" {
 #include "flags.h"
 #include "shared/httpd.h"
 #include "shared/memcpy_interleaved.h"
+#include "shared/metrics.h"
 #include "pbo_frame_allocator.h"
 #include "shared/timebase.h"
 #include "va_display_with_cleanup.h"
 
+#include <movit/colorspace_conversion_effect.h>
+
 #include <va/va.h>
 #include <va/va_drm.h>
 #include <va/va_x11.h>
 
+using namespace Eigen;
 using namespace bmusb;
+using namespace movit;
 using namespace std;
 
-extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
+static VAImageFormat uyvy_format;
 
-#define CHECK_VASTATUS(va_status, func)                                 \
-    if (va_status != VA_STATUS_SUCCESS) {                               \
-        fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
-        exit(1);                                                        \
-    }
+extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
 
 // From libjpeg (although it's of course identical between implementations).
 static const int jpeg_natural_order[DCTSIZE2] = {
@@ -116,6 +117,68 @@ int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType t
        return buf_size;
 }
 
+namespace {
+
+void add_video_stream(AVFormatContext *avctx)
+{
+       AVStream *stream = avformat_new_stream(avctx, nullptr);
+       if (stream == nullptr) {
+               fprintf(stderr, "avformat_new_stream() failed\n");
+               abort();
+       }
+
+       // FFmpeg is very picky about having audio at 1/48000 timebase,
+       // no matter what we write. Even though we'd prefer our usual 1/120000,
+       // put the video on the same one, so that we can have locked audio.
+       stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
+       stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
+       stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
+
+       // Used for aspect ratio only. Can change without notice (the mux won't care).
+       stream->codecpar->width = global_flags.width;
+       stream->codecpar->height = global_flags.height;
+
+       // TODO: We could perhaps use the interpretation for each card here
+       // (or at least the command-line flags) instead of the defaults,
+       // but what would we do when they change?
+       stream->codecpar->color_primaries = AVCOL_PRI_BT709;
+       stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
+       stream->codecpar->color_space = AVCOL_SPC_BT709;
+       stream->codecpar->color_range = AVCOL_RANGE_MPEG;
+       stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
+       stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
+}
+
+void add_audio_stream(AVFormatContext *avctx)
+{
+       AVStream *stream = avformat_new_stream(avctx, nullptr);
+       if (stream == nullptr) {
+               fprintf(stderr, "avformat_new_stream() failed\n");
+               abort();
+       }
+       stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
+       stream->codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
+       stream->codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
+       stream->codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
+       stream->codecpar->channels = 2;
+       stream->codecpar->sample_rate = OUTPUT_FREQUENCY;
+}
+
+void finalize_mux(AVFormatContext *avctx)
+{
+       AVDictionary *options = NULL;
+       vector<pair<string, string>> opts = MUX_OPTS;
+       for (pair<string, string> opt : opts) {
+               av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
+       }
+       if (avformat_write_header(avctx, &options) < 0) {
+               fprintf(stderr, "avformat_write_header() failed\n");
+               abort();
+       }
+}
+
+}  // namespace
+
 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
        : httpd(httpd)
 {
@@ -123,7 +186,7 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
        // a situation with only one video stream (and possibly one audio stream)
        // with known width/height, and we don't need the extra functionality it provides.
        avctx.reset(avformat_alloc_context());
-       avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
+       avctx->oformat = av_guess_format("nut", nullptr, nullptr);
 
        uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
        avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
@@ -131,39 +194,12 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
        avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 
        for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
-               AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
-               if (stream == nullptr) {
-                       fprintf(stderr, "avformat_new_stream() failed\n");
-                       exit(1);
-               }
-               stream->time_base = AVRational{ 1, TIMEBASE };
-               stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
-               stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
-
-               // Used for aspect ratio only. Can change without notice (the mux won't care).
-               stream->codecpar->width = global_flags.width;
-               stream->codecpar->height = global_flags.height;
-
-               // TODO: We could perhaps use the interpretation for each card here
-               // (or at least the command-line flags) instead of the defaults,
-               // but what would we do when they change?
-               stream->codecpar->color_primaries = AVCOL_PRI_BT709;
-               stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
-               stream->codecpar->color_space = AVCOL_SPC_BT709;
-               stream->codecpar->color_range = AVCOL_RANGE_MPEG;
-               stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
-               stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
+               add_video_stream(avctx.get());
        }
-
-       AVDictionary *options = NULL;
-       vector<pair<string, string>> opts = MUX_OPTS;
-       for (pair<string, string> opt : opts) {
-               av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
-       }
-       if (avformat_write_header(avctx.get(), &options) < 0) {
-               fprintf(stderr, "avformat_write_header() failed\n");
-               exit(1);
+       for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
+               add_audio_stream(avctx.get());
        }
+       finalize_mux(avctx.get());
 
        // Initialize VA-API.
        string error;
@@ -177,12 +213,26 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
                va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
        }
 
+       global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
+       global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
+       global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
+       global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
+       global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
+       global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
+
        running = true;
 }
 
 MJPEGEncoder::~MJPEGEncoder()
 {
        av_free(avctx->pb->buffer);
+
+       global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
+       global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
+       global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
+       global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
+       global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
+       global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
 }
 
 void MJPEGEncoder::stop()
@@ -193,6 +243,7 @@ void MJPEGEncoder::stop()
        running = false;
        should_quit = true;
        any_frames_to_be_encoded.notify_all();
+       any_frames_encoding.notify_all();
        encoder_thread.join();
        if (va_dpy != nullptr) {
                va_receiver_thread.join();
@@ -228,6 +279,7 @@ unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_disp
                return nullptr;
        }
 
+       // TODO: Unify with the code in Futatabi.
        int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
        assert(num_formats > 0);
 
@@ -240,38 +292,72 @@ unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_disp
                return nullptr;
        }
 
+       bool found = false;
+       for (int i = 0; i < num_formats; ++i) {
+               if (formats[i].fourcc == VA_FOURCC_UYVY) {
+                       memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
+                       found = true;
+                       break;
+               }
+       }
+       if (!found) {
+               if (error != nullptr) *error = "UYVY format not found";
+               return nullptr;
+       }
+
        return va_dpy;
 }
 
-void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset)
+void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset, vector<int32_t> audio, const RGBTriplet &white_balance)
 {
        PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
        if (video_format.width == 0 || video_format.height == 0) {
+               ++metric_mjpeg_frames_zero_size_dropped;
                return;
        }
        if (video_format.interlaced) {
                fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
+               ++metric_mjpeg_frames_interlaced_dropped;
                return;
        }
        if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
            !frame->interleaved) {
                fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
+               ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
                return;
        }
        if (video_format.width > 4096 || video_format.height > 4096) {
                fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
+               ++metric_mjpeg_frames_oversized_dropped;
                return;
        }
 
        lock_guard<mutex> lock(mu);
-       if (frames_to_be_encoded.size() + frames_encoding.size() > 10) {
+       if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
                fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
+               ++metric_mjpeg_overrun_dropped;
                return;
        }
-       frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
+       ++metric_mjpeg_overrun_submitted;
+       frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance });
        any_frames_to_be_encoded.notify_all();
 }
 
+int MJPEGEncoder::get_mjpeg_stream_for_card(unsigned card_index)
+{
+       // Only bother doing MJPEG encoding if there are any connected clients
+       // that want the stream.
+       if (httpd->get_num_connected_multicam_clients() == 0) {
+               return -1;
+       }
+
+       auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
+       if (it == global_flags.card_to_mjpeg_stream_export.end()) {
+               return -1;
+       }
+       return it->second;
+}
+
 void MJPEGEncoder::encoder_thread_func()
 {
        pthread_setname_np(pthread_self(), "MJPEG_Encode");
@@ -294,9 +380,14 @@ void MJPEGEncoder::encoder_thread_func()
                        // Will call back in the receiver thread.
                        encode_jpeg_va(move(qf));
                } else {
+                       // Write audio before video, since Futatabi expects it.
+                       if (qf.audio.size() > 0) {
+                               write_audio_packet(qf.pts, qf.card_index, qf.audio);
+                       }
+
                        // Encode synchronously, in the same thread.
                        vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
-                       write_mjpeg_packet(qf.pts, qf.card_index, jpeg);
+                       write_mjpeg_packet(qf.pts, qf.card_index, jpeg.data(), jpeg.size());
                }
        }
 
@@ -306,20 +397,42 @@ void MJPEGEncoder::encoder_thread_func()
        free(tmp_cr);
 }
 
-void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const vector<uint8_t> &jpeg)
+void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const uint8_t *jpeg, size_t jpeg_size)
 {
        AVPacket pkt;
        memset(&pkt, 0, sizeof(pkt));
        pkt.buf = nullptr;
-       pkt.data = const_cast<uint8_t *>(&jpeg[0]);
-       pkt.size = jpeg.size();
+       pkt.data = const_cast<uint8_t *>(jpeg);
+       pkt.size = jpeg_size;
        pkt.stream_index = card_index;
        pkt.flags = AV_PKT_FLAG_KEY;
-       pkt.pts = pkt.dts = pts;
+       AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
+       pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
+       pkt.duration = 0;
+
+       if (av_write_frame(avctx.get(), &pkt) < 0) {
+               fprintf(stderr, "av_write_frame() failed\n");
+               abort();
+       }
+}
+
+void MJPEGEncoder::write_audio_packet(int64_t pts, unsigned card_index, const vector<int32_t> &audio)
+{
+       AVPacket pkt;
+       memset(&pkt, 0, sizeof(pkt));
+       pkt.buf = nullptr;
+       pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
+       pkt.size = audio.size() * sizeof(audio[0]);
+       pkt.stream_index = card_index + global_flags.card_to_mjpeg_stream_export.size();
+       pkt.flags = AV_PKT_FLAG_KEY;
+       AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
+       pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
+       size_t num_stereo_samples = audio.size() / 2;
+       pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
 
        if (av_write_frame(avctx.get(), &pkt) < 0) {
                fprintf(stderr, "av_write_frame() failed\n");
-               exit(1);
+               abort();
        }
 }
 
@@ -373,13 +486,16 @@ MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigne
        va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
 
+       va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
+       CHECK_VASTATUS(va_status, "vaCreateImage");
+
        return ret;
 }
 
 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
 {
        lock_guard<mutex> lock(va_resources_mutex);
-       if (va_resources_freelist.size() > 10) {
+       if (va_resources_freelist.size() > 50) {
                auto it = va_resources_freelist.end();
                --it;
 
@@ -392,13 +508,34 @@ void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
                va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
                CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 
+               va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
+               CHECK_VASTATUS(va_status, "vaDestroyImage");
+
                va_resources_freelist.erase(it);
        }
 
        va_resources_freelist.push_front(resources);
 }
 
-void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
+namespace {
+
+void push16(uint16_t val, string *str)
+{
+       str->push_back(val >> 8);
+       str->push_back(val & 0xff);
+}
+
+void push32(uint32_t val, string *str)
+{
+       str->push_back(val >> 24);
+       str->push_back((val >> 16) & 0xff);
+       str->push_back((val >> 8) & 0xff);
+       str->push_back(val & 0xff);
+}
+
+}  // namespace
+
+void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
 {
        jpeg_error_mgr jerr;
        cinfo->err = jpeg_std_error(&jerr);
@@ -423,15 +560,70 @@ void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinat
        cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
        jpeg_start_compress(cinfo, true);
 
+       if (fabs(white_balance.r - 1.0f) > 1e-3 ||
+           fabs(white_balance.g - 1.0f) > 1e-3 ||
+           fabs(white_balance.b - 1.0f) > 1e-3) {
+               // Convert from (linear) RGB to XYZ.
+               Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
+               Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b);
+
+               // Convert from XYZ to xyz by normalizing.
+               xyz /= (xyz[0] + xyz[1] + xyz[2]);
+
+               // Create a very rudimentary EXIF header to hold our white point.
+               string exif;
+
+               // Exif header, followed by some padding.
+               exif = "Exif";
+               push16(0, &exif);
+
+               // TIFF header first:
+               exif += "MM";  // Big endian.
+
+               // Magic number.
+               push16(42, &exif);
+
+               // Offset of first IFD (relative to the MM, immediately after the header).
+               push32(exif.size() - 6 + 4, &exif);
+
+               // Now the actual IFD.
+
+               // One entry.
+               push16(1, &exif);
+
+               // WhitePoint tag ID.
+               push16(0x13e, &exif);
+
+               // Rational type.
+               push16(5, &exif);
+
+               // Two values (x and y; z is implicit due to normalization).
+               push32(2, &exif);
+
+               // Offset (relative to the MM, immediately after the last IFD).
+               push32(exif.size() - 6 + 8, &exif);
+
+               // No more IFDs.
+               push32(0, &exif);
+
+               // The actual values.
+               push32(lrintf(xyz[0] * 10000.0f), &exif);
+               push32(10000, &exif);
+               push32(lrintf(xyz[1] * 10000.0f), &exif);
+               push32(10000, &exif);
+
+               jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size());
+       }
+
        // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
        // (and nothing else).
        jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 }
 
-vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
+vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, jpeg_compress_struct *cinfo)
 {
        VectorDestinationManager dest;
-       init_jpeg_422(width, height, &dest, cinfo);
+       init_jpeg_422(width, height, white_balance, &dest, cinfo);
 
        // Make a dummy black image; there's seemingly no other easy way of
        // making libjpeg outputting all of its headers.
@@ -463,7 +655,7 @@ vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, j
        return dest.dest;
 }
 
-MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
+MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height, const RGBTriplet &white_balance)
 {
        pair<unsigned, unsigned> key(width, height);
        if (va_data_for_resolution.count(key)) {
@@ -473,7 +665,7 @@ MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, un
        // Use libjpeg to generate a header and set sane defaults for e.g.
        // quantization tables. Then do the actual encode with VA-API.
        jpeg_compress_struct cinfo;
-       vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
+       vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, &cinfo);
 
        // Picture parameters.
        VAEncPictureParameterBufferJPEG pic_param;
@@ -574,13 +766,22 @@ MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, un
 
 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
 {
+       PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
        unsigned width = qf.video_format.width;
        unsigned height = qf.video_format.height;
 
-       VAResources resources = get_va_resources(width, height);
-       ReleaseVAResources release(this, resources);
+       VAResources resources;
+       ReleaseVAResources release;
+       if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
+               resources = move(userdata->va_resources);
+               release = move(userdata->va_resources_release);
+       } else {
+               assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
+               resources = get_va_resources(width, height);
+               release = ReleaseVAResources(this, resources);
+       }
 
-       VAData va_data = get_va_data_for_resolution(width, height);
+       VAData va_data = get_va_data_for_resolution(width, height, qf.white_balance);
        va_data.pic_param.coded_buf = resources.data_buffer;
 
        VABufferID pic_param_buffer;
@@ -603,27 +804,38 @@ void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 
-       VAImage image;
-       va_status = vaDeriveImage(va_dpy->va_dpy, resources.surface, &image);
-       CHECK_VASTATUS(va_status, "vaDeriveImage");
+       if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
+               // The pixel data is already put into the image by the caller.
+               va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
+               CHECK_VASTATUS(va_status, "vaUnmapBuffer");
+       } else {
+               assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
 
-       // Upload the pixel data.
-       uint8_t *surface_p = nullptr;
-       vaMapBuffer(va_dpy->va_dpy, image.buf, (void **)&surface_p);
+               // Upload the pixel data.
+               uint8_t *surface_p = nullptr;
+               vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
 
-       size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
-       size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
+               size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
+               size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 
-       {
-               const uint8_t *src = qf.frame->data_copy + field_start;
-               uint8_t *dst = (unsigned char *)surface_p + image.offsets[0];
-               memcpy_with_pitch(dst, src, qf.video_format.width * 2, image.pitches[0], qf.video_format.height);
+               {
+                       const uint8_t *src = qf.frame->data_copy + field_start;
+                       uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
+                       memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
+               }
+
+               va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
+               CHECK_VASTATUS(va_status, "vaUnmapBuffer");
        }
 
-       va_status = vaUnmapBuffer(va_dpy->va_dpy, image.buf);
-       CHECK_VASTATUS(va_status, "vaUnmapBuffer");
-       va_status = vaDestroyImage(va_dpy->va_dpy, image.image_id);
-       CHECK_VASTATUS(va_status, "vaDestroyImage");
+       qf.frame->data_copy = nullptr;
+
+       // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
+       // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
+       // Exactly why is unclear, but it seems to involve L3 cache usage when there
+       // are many high-res (1080p+) images in play.
+       va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
+       CHECK_VASTATUS(va_status, "vaPutImage");
 
        // Finally, stick in the JPEG header.
        VAEncPackedHeaderParameterBuffer header_parm;
@@ -678,6 +890,11 @@ void MJPEGEncoder::va_receiver_thread_func()
                        frames_encoding.pop();
                }
 
+               // Write audio before video, since Futatabi expects it.
+               if (qf.audio.size() > 0) {
+                       write_audio_packet(qf.pts, qf.card_index, qf.audio);
+               }
+
                VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
                CHECK_VASTATUS(va_status, "vaSyncSurface");
 
@@ -685,13 +902,11 @@ void MJPEGEncoder::va_receiver_thread_func()
                va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
                CHECK_VASTATUS(va_status, "vaMapBuffer");
 
-               const char *coded_buf = reinterpret_cast<char *>(segment->buf);
-               vector<uint8_t> jpeg(coded_buf, coded_buf + segment->size);
+               const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
+               write_mjpeg_packet(qf.pts, qf.card_index, coded_buf, segment->size);
 
                va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
                CHECK_VASTATUS(va_status, "vaUnmapBuffer");
-
-               write_mjpeg_packet(qf.pts, qf.card_index, jpeg);
        }
 }
 
@@ -702,7 +917,7 @@ vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 
        VectorDestinationManager dest;
        jpeg_compress_struct cinfo;
-       init_jpeg_422(width, height, &dest, &cinfo);
+       init_jpeg_422(width, height, qf.white_balance, &dest, &cinfo);
 
        size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
        size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
@@ -712,7 +927,7 @@ vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
        for (unsigned y = 0; y < qf.video_format.height; y += 8) {
                const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
 
-               memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
+               memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
                memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
                for (unsigned yy = 0; yy < 8; ++yy) {
                        yptr[yy] = tmp_y + yy * width;