1 #include "mjpeg_encoder.h"
12 #include <libavformat/avformat.h>
16 #include "shared/ffmpeg_raii.h"
18 #include "shared/httpd.h"
19 #include "shared/memcpy_interleaved.h"
20 #include "shared/metrics.h"
21 #include "pbo_frame_allocator.h"
22 #include "shared/timebase.h"
23 #include "shared/va_display.h"
25 #include <movit/colorspace_conversion_effect.h>
28 #include <va/va_drm.h>
29 #include <va/va_x11.h>
31 using namespace Eigen;
32 using namespace bmusb;
33 using namespace movit;
36 static VAImageFormat uyvy_format, nv12_format;
38 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
40 // The inverse of memcpy_interleaved(), with (slow) support for pitch.
41 void interleave_with_pitch(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, size_t src_width, size_t dst_pitch, size_t height)
44 if (dst_pitch == src_width * 2 && (src_width * height) % 16 == 0) {
45 __m128i *dptr = reinterpret_cast<__m128i *>(dst);
46 const __m128i *sptr1 = reinterpret_cast<const __m128i *>(src1);
47 const __m128i *sptr2 = reinterpret_cast<const __m128i *>(src2);
48 for (size_t i = 0; i < src_width * height / 16; ++i) {
49 __m128i data1 = _mm_loadu_si128(sptr1++);
50 __m128i data2 = _mm_loadu_si128(sptr2++);
51 _mm_storeu_si128(dptr++, _mm_unpacklo_epi8(data1, data2));
52 _mm_storeu_si128(dptr++, _mm_unpackhi_epi8(data1, data2));
58 for (size_t y = 0; y < height; ++y) {
59 uint8_t *dptr = dst + y * dst_pitch;
60 const uint8_t *sptr1 = src1 + y * src_width;
61 const uint8_t *sptr2 = src2 + y * src_width;
62 for (size_t x = 0; x < src_width; ++x) {
69 // From libjpeg (although it's of course identical between implementations).
70 static const int jpeg_natural_order[DCTSIZE2] = {
71 0, 1, 8, 16, 9, 2, 3, 10,
72 17, 24, 32, 25, 18, 11, 4, 5,
73 12, 19, 26, 33, 40, 48, 41, 34,
74 27, 20, 13, 6, 7, 14, 21, 28,
75 35, 42, 49, 56, 57, 50, 43, 36,
76 29, 22, 15, 23, 30, 37, 44, 51,
77 58, 59, 52, 45, 38, 31, 39, 46,
78 53, 60, 61, 54, 47, 55, 62, 63,
81 struct VectorDestinationManager {
82 jpeg_destination_mgr pub;
83 std::vector<uint8_t> dest;
85 VectorDestinationManager()
87 pub.init_destination = init_destination_thunk;
88 pub.empty_output_buffer = empty_output_buffer_thunk;
89 pub.term_destination = term_destination_thunk;
92 static void init_destination_thunk(j_compress_ptr ptr)
94 ((VectorDestinationManager *)(ptr->dest))->init_destination();
97 inline void init_destination()
102 static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
104 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
107 inline bool empty_output_buffer()
109 make_room(dest.size()); // Should ignore pub.free_in_buffer!
113 inline void make_room(size_t bytes_used)
115 dest.resize(bytes_used + 4096);
116 dest.resize(dest.capacity());
117 pub.next_output_byte = dest.data() + bytes_used;
118 pub.free_in_buffer = dest.size() - bytes_used;
121 static void term_destination_thunk(j_compress_ptr ptr)
123 ((VectorDestinationManager *)(ptr->dest))->term_destination();
126 inline void term_destination()
128 dest.resize(dest.size() - pub.free_in_buffer);
131 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
133 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
135 WritePacket2Context *ctx = (WritePacket2Context *)opaque;
136 return ctx->mjpeg_encoder->write_packet2(ctx->stream_id, buf, buf_size, type, time);
139 int MJPEGEncoder::write_packet2(HTTPD::StreamID stream_id, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
141 string *mux_header = &streams[stream_id].mux_header;
142 if (type == AVIO_DATA_MARKER_HEADER) {
143 mux_header->append((char *)buf, buf_size);
144 httpd->set_header(stream_id, *mux_header);
146 httpd->add_data(stream_id, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
153 void add_video_stream(AVFormatContext *avctx)
155 AVStream *stream = avformat_new_stream(avctx, nullptr);
156 if (stream == nullptr) {
157 fprintf(stderr, "avformat_new_stream() failed\n");
161 // FFmpeg is very picky about having audio at 1/48000 timebase,
162 // no matter what we write. Even though we'd prefer our usual 1/120000,
163 // put the video on the same one, so that we can have locked audio.
164 stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
165 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
166 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
168 // Used for aspect ratio only. Can change without notice (the mux won't care).
169 stream->codecpar->width = global_flags.width;
170 stream->codecpar->height = global_flags.height;
172 // TODO: We could perhaps use the interpretation for each card here
173 // (or at least the command-line flags) instead of the defaults,
174 // but what would we do when they change?
175 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
176 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
177 stream->codecpar->color_space = AVCOL_SPC_BT709;
178 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
179 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
180 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
183 void add_audio_stream(AVFormatContext *avctx)
185 AVStream *stream = avformat_new_stream(avctx, nullptr);
186 if (stream == nullptr) {
187 fprintf(stderr, "avformat_new_stream() failed\n");
190 stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
191 stream->codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
192 stream->codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
193 stream->codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
194 stream->codecpar->channels = 2;
195 stream->codecpar->sample_rate = OUTPUT_FREQUENCY;
198 void finalize_mux(AVFormatContext *avctx)
200 AVDictionary *options = NULL;
201 vector<pair<string, string>> opts = MUX_OPTS;
202 for (pair<string, string> opt : opts) {
203 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
205 if (avformat_write_header(avctx, &options) < 0) {
206 fprintf(stderr, "avformat_write_header() failed\n");
213 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
216 create_ffmpeg_context(HTTPD::StreamID{ HTTPD::MULTICAM_STREAM, 0 });
217 for (unsigned stream_idx = 0; stream_idx < MAX_VIDEO_CARDS; ++stream_idx) {
218 create_ffmpeg_context(HTTPD::StreamID{ HTTPD::SIPHON_STREAM, stream_idx });
221 add_stream(HTTPD::StreamID{ HTTPD::MULTICAM_STREAM, 0 });
223 // Initialize VA-API.
225 va_dpy = try_open_va(va_display, { VAProfileJPEGBaseline }, VAEntrypointEncPicture,
227 { "4:2:2", VA_RT_FORMAT_YUV422, VA_FOURCC_UYVY, &config_id_422, &uyvy_format },
228 // We'd prefer VA_FOURCC_I420, but it's not supported by Intel's driver.
229 { "4:2:0", VA_RT_FORMAT_YUV420, VA_FOURCC_NV12, &config_id_420, &nv12_format }
231 /*chosen_profile=*/nullptr, &error);
232 if (va_dpy == nullptr) {
233 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
236 encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
237 if (va_dpy != nullptr) {
238 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
241 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
242 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
243 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
244 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
245 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
246 global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
251 MJPEGEncoder::~MJPEGEncoder()
253 for (auto &id_and_stream : streams) {
254 av_free(id_and_stream.second.avctx->pb->buffer);
257 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
258 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
259 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
260 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
261 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
262 global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
265 void MJPEGEncoder::stop()
272 any_frames_to_be_encoded.notify_all();
273 any_frames_encoding.notify_all();
274 encoder_thread.join();
275 if (va_dpy != nullptr) {
276 va_receiver_thread.join();
282 bool is_uyvy(RefCountedFrame frame)
284 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
285 return userdata->pixel_format == PixelFormat_8BitYCbCr && frame->interleaved;
288 bool is_i420(RefCountedFrame frame)
290 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
291 return userdata->pixel_format == PixelFormat_8BitYCbCrPlanar &&
292 userdata->ycbcr_format.chroma_subsampling_x == 2 &&
293 userdata->ycbcr_format.chroma_subsampling_y == 2;
298 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)
300 if (video_format.width == 0 || video_format.height == 0) {
301 ++metric_mjpeg_frames_zero_size_dropped;
304 if (video_format.interlaced) {
305 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
306 ++metric_mjpeg_frames_interlaced_dropped;
309 if (!is_uyvy(frame) && !is_i420(frame)) {
310 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
311 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
314 if (video_format.width > 4096 || video_format.height > 4096) {
315 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
316 ++metric_mjpeg_frames_oversized_dropped;
320 lock_guard<mutex> lock(mu);
321 if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
322 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
323 ++metric_mjpeg_overrun_dropped;
326 ++metric_mjpeg_overrun_submitted;
327 frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance });
328 any_frames_to_be_encoded.notify_all();
331 bool MJPEGEncoder::should_encode_mjpeg_for_card(unsigned card_index)
333 // Only bother doing MJPEG encoding if there are any connected clients
334 // that want the stream.
335 if (httpd->get_num_connected_multicam_clients() == 0 &&
336 httpd->get_num_connected_siphon_clients(card_index) == 0) {
340 auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
341 return (it != global_flags.card_to_mjpeg_stream_export.end());
344 void MJPEGEncoder::encoder_thread_func()
346 pthread_setname_np(pthread_self(), "MJPEG_Encode");
347 posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
348 posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
349 posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
350 posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
355 unique_lock<mutex> lock(mu);
356 any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
357 if (should_quit) break;
358 qf = move(frames_to_be_encoded.front());
359 frames_to_be_encoded.pop();
362 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index)); // Or should_encode_mjpeg_for_card() would have returned false.
363 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
365 if (va_dpy != nullptr) {
366 // Will call back in the receiver thread.
367 encode_jpeg_va(move(qf));
369 update_siphon_streams();
371 HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
372 HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
373 assert(streams.count(multicam_id));
375 // Write audio before video, since Futatabi expects it.
376 if (qf.audio.size() > 0) {
377 write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
378 if (streams.count(siphon_id)) {
379 write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
383 // Encode synchronously, in the same thread.
384 vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
385 write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, jpeg.data(), jpeg.size());
386 if (streams.count(siphon_id)) {
387 write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, jpeg.data(), jpeg.size());
398 void MJPEGEncoder::write_mjpeg_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const uint8_t *jpeg, size_t jpeg_size)
401 memset(&pkt, 0, sizeof(pkt));
403 pkt.data = const_cast<uint8_t *>(jpeg);
404 pkt.size = jpeg_size;
405 pkt.stream_index = stream_index;
406 pkt.flags = AV_PKT_FLAG_KEY;
407 AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
408 pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
411 if (av_write_frame(avctx, &pkt) < 0) {
412 fprintf(stderr, "av_write_frame() failed\n");
417 void MJPEGEncoder::write_audio_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const vector<int32_t> &audio)
420 memset(&pkt, 0, sizeof(pkt));
422 pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
423 pkt.size = audio.size() * sizeof(audio[0]);
424 pkt.stream_index = stream_index;
425 pkt.flags = AV_PKT_FLAG_KEY;
426 AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
427 pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
428 size_t num_stereo_samples = audio.size() / 2;
429 pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
431 if (av_write_frame(avctx, &pkt) < 0) {
432 fprintf(stderr, "av_write_frame() failed\n");
437 class VABufferDestroyer {
439 VABufferDestroyer(VADisplay dpy, VABufferID buf)
440 : dpy(dpy), buf(buf) {}
442 ~VABufferDestroyer() {
443 VAStatus va_status = vaDestroyBuffer(dpy, buf);
444 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
452 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height, uint32_t fourcc)
455 lock_guard<mutex> lock(va_resources_mutex);
456 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
457 if (it->width == width && it->height == height && it->fourcc == fourcc) {
458 VAResources ret = *it;
459 va_resources_freelist.erase(it);
471 VASurfaceAttrib attrib;
472 attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
473 attrib.type = VASurfaceAttribPixelFormat;
474 attrib.value.type = VAGenericValueTypeInteger;
475 attrib.value.value.i = fourcc;
478 VAConfigID config_id;
479 if (fourcc == VA_FOURCC_UYVY) {
480 va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422, width, height, &ret.surface, 1, &attrib, 1);
481 config_id = config_id_422;
483 assert(fourcc == VA_FOURCC_NV12);
484 va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV420, width, height, &ret.surface, 1, &attrib, 1);
485 config_id = config_id_420;
488 va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
489 CHECK_VASTATUS(va_status, "vaCreateContext");
491 va_status = vaCreateBuffer(va_dpy->va_dpy, ret.context, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
492 CHECK_VASTATUS(va_status, "vaCreateBuffer");
494 if (fourcc == VA_FOURCC_UYVY) {
495 va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
496 CHECK_VASTATUS(va_status, "vaCreateImage");
498 assert(fourcc == VA_FOURCC_NV12);
499 va_status = vaCreateImage(va_dpy->va_dpy, &nv12_format, width, height, &ret.image);
500 CHECK_VASTATUS(va_status, "vaCreateImage");
506 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
508 lock_guard<mutex> lock(va_resources_mutex);
509 if (va_resources_freelist.size() > 50) {
510 auto it = va_resources_freelist.end();
513 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
514 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
516 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
517 CHECK_VASTATUS(va_status, "vaDestroyContext");
519 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
520 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
522 va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
523 CHECK_VASTATUS(va_status, "vaDestroyImage");
525 va_resources_freelist.erase(it);
528 va_resources_freelist.push_front(resources);
533 void push16(uint16_t val, string *str)
535 str->push_back(val >> 8);
536 str->push_back(val & 0xff);
539 void push32(uint32_t val, string *str)
541 str->push_back(val >> 24);
542 str->push_back((val >> 16) & 0xff);
543 str->push_back((val >> 8) & 0xff);
544 str->push_back(val & 0xff);
549 void MJPEGEncoder::init_jpeg(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo, int y_h_samp_factor, int y_v_samp_factor)
552 cinfo->err = jpeg_std_error(&jerr);
553 jpeg_create_compress(cinfo);
555 cinfo->dest = (jpeg_destination_mgr *)dest;
557 cinfo->input_components = 3;
558 jpeg_set_defaults(cinfo);
559 jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
561 cinfo->image_width = width;
562 cinfo->image_height = height;
563 cinfo->raw_data_in = true;
564 jpeg_set_colorspace(cinfo, JCS_YCbCr);
565 cinfo->comp_info[0].h_samp_factor = y_h_samp_factor;
566 cinfo->comp_info[0].v_samp_factor = y_v_samp_factor;
567 cinfo->comp_info[1].h_samp_factor = 1;
568 cinfo->comp_info[1].v_samp_factor = 1;
569 cinfo->comp_info[2].h_samp_factor = 1;
570 cinfo->comp_info[2].v_samp_factor = 1;
571 cinfo->CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though.
572 jpeg_start_compress(cinfo, true);
574 if (fabs(white_balance.r - 1.0f) > 1e-3 ||
575 fabs(white_balance.g - 1.0f) > 1e-3 ||
576 fabs(white_balance.b - 1.0f) > 1e-3) {
577 // Convert from (linear) RGB to XYZ.
578 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
579 Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b);
581 // Convert from XYZ to xyz by normalizing.
582 xyz /= (xyz[0] + xyz[1] + xyz[2]);
584 // Create a very rudimentary EXIF header to hold our white point.
587 // Exif header, followed by some padding.
591 // TIFF header first:
592 exif += "MM"; // Big endian.
597 // Offset of first IFD (relative to the MM, immediately after the header).
598 push32(exif.size() - 6 + 4, &exif);
600 // Now the actual IFD.
605 // WhitePoint tag ID.
606 push16(0x13e, &exif);
611 // Two values (x and y; z is implicit due to normalization).
614 // Offset (relative to the MM, immediately after the last IFD).
615 push32(exif.size() - 6 + 8, &exif);
620 // The actual values.
621 push32(lrintf(xyz[0] * 10000.0f), &exif);
622 push32(10000, &exif);
623 push32(lrintf(xyz[1] * 10000.0f), &exif);
624 push32(10000, &exif);
626 jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size());
629 // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
630 // (and nothing else).
631 jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
634 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, int y_h_samp_factor, int y_v_samp_factor, jpeg_compress_struct *cinfo)
636 VectorDestinationManager dest;
637 init_jpeg(width, height, white_balance, &dest, cinfo, y_h_samp_factor, y_v_samp_factor);
639 // Make a dummy black image; there's seemingly no other easy way of
640 // making libjpeg outputting all of its headers.
641 assert(y_v_samp_factor <= 2); // Or we'd need larger JSAMPROW arrays below.
642 size_t block_height_y = 8 * y_v_samp_factor;
643 size_t block_height_cbcr = 8;
645 JSAMPROW yptr[16], cbptr[16], crptr[16];
646 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
647 memset(tmp_y, 0, 4096);
648 memset(tmp_cb, 0, 4096);
649 memset(tmp_cr, 0, 4096);
650 for (unsigned yy = 0; yy < block_height_y; ++yy) {
653 for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
657 for (unsigned y = 0; y < height; y += block_height_y) {
658 jpeg_write_raw_data(cinfo, data, block_height_y);
660 jpeg_finish_compress(cinfo);
662 // We're only interested in the header, not the data after it.
663 dest.term_destination();
664 for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
665 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) { // Start of scan (SOS).
666 unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
667 dest.dest.resize(i + len + 2);
675 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_parameters(unsigned width, unsigned height, unsigned y_h_samp_factor, unsigned y_v_samp_factor, const RGBTriplet &white_balance)
677 VAKey key{width, height, y_h_samp_factor, y_v_samp_factor, white_balance};
678 if (va_data_for_parameters.count(key)) {
679 return va_data_for_parameters[key];
682 // Use libjpeg to generate a header and set sane defaults for e.g.
683 // quantization tables. Then do the actual encode with VA-API.
684 jpeg_compress_struct cinfo;
685 vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, y_h_samp_factor, y_v_samp_factor, &cinfo);
687 // Picture parameters.
688 VAEncPictureParameterBufferJPEG pic_param;
689 memset(&pic_param, 0, sizeof(pic_param));
690 pic_param.reconstructed_picture = VA_INVALID_ID;
691 pic_param.picture_width = cinfo.image_width;
692 pic_param.picture_height = cinfo.image_height;
693 for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
694 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
695 pic_param.component_id[component_idx] = comp->component_id;
696 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
698 pic_param.num_components = cinfo.num_components;
699 pic_param.num_scan = 1;
700 pic_param.sample_bit_depth = 8;
701 pic_param.coded_buf = VA_INVALID_ID; // To be filled out by caller.
702 pic_param.pic_flags.bits.huffman = 1;
703 pic_param.quality = 50; // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
705 // Quantization matrices.
706 VAQMatrixBufferJPEG q;
707 memset(&q, 0, sizeof(q));
709 q.load_lum_quantiser_matrix = true;
710 q.load_chroma_quantiser_matrix = true;
711 for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
712 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
713 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
714 if (qtbl == nullptr) continue;
716 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
717 for (int i = 0; i < 64; ++i) {
718 if (qtbl->quantval[i] > 255) {
719 fprintf(stderr, "Baseline JPEG only!\n");
722 qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
726 // Huffman tables (arithmetic is not supported).
727 VAHuffmanTableBufferJPEGBaseline huff;
728 memset(&huff, 0, sizeof(huff));
730 for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
731 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
732 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
733 if (ac_hufftbl == nullptr) {
734 assert(dc_hufftbl == nullptr);
735 huff.load_huffman_table[huff_tbl_idx] = 0;
737 assert(dc_hufftbl != nullptr);
738 huff.load_huffman_table[huff_tbl_idx] = 1;
740 for (int i = 0; i < 16; ++i) {
741 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
743 for (int i = 0; i < 12; ++i) {
744 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
746 for (int i = 0; i < 16; ++i) {
747 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
749 for (int i = 0; i < 162; ++i) {
750 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
755 // Slice parameters (metadata about the slice).
756 VAEncSliceParameterBufferJPEG parms;
757 memset(&parms, 0, sizeof(parms));
758 for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
759 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
760 parms.components[component_idx].component_selector = comp->component_id;
761 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
762 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
763 if (parms.components[component_idx].dc_table_selector > 1 ||
764 parms.components[component_idx].ac_table_selector > 1) {
765 fprintf(stderr, "Uses too many Huffman tables\n");
769 parms.num_components = cinfo.num_components;
770 parms.restart_interval = cinfo.restart_interval;
772 jpeg_destroy_compress(&cinfo);
775 ret.jpeg_header = move(jpeg_header);
776 ret.pic_param = pic_param;
780 va_data_for_parameters[key] = ret;
784 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
786 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
787 unsigned width = qf.video_format.width;
788 unsigned height = qf.video_format.height;
790 VAResources resources;
791 ReleaseVAResources release;
792 if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
793 assert(is_uyvy(qf.frame));
794 resources = move(userdata->va_resources);
795 release = move(userdata->va_resources_release);
797 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
798 if (is_uyvy(qf.frame)) {
799 resources = get_va_resources(width, height, VA_FOURCC_UYVY);
801 assert(is_i420(qf.frame));
802 resources = get_va_resources(width, height, VA_FOURCC_NV12);
804 release = ReleaseVAResources(this, resources);
807 int y_h_samp_factor, y_v_samp_factor;
808 if (is_uyvy(qf.frame)) {
809 // 4:2:2 (sample Y' twice as often horizontally as Cb or Cr, vertical is left alone).
813 // 4:2:0 (sample Y' twice as often as Cb or Cr, in both directions)
814 assert(is_i420(qf.frame));
819 VAData va_data = get_va_data_for_parameters(width, height, y_h_samp_factor, y_v_samp_factor, qf.white_balance);
820 va_data.pic_param.coded_buf = resources.data_buffer;
822 VABufferID pic_param_buffer;
823 VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
824 CHECK_VASTATUS(va_status, "vaCreateBuffer");
825 VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
828 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
829 CHECK_VASTATUS(va_status, "vaCreateBuffer");
830 VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
832 VABufferID huff_buffer;
833 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
834 CHECK_VASTATUS(va_status, "vaCreateBuffer");
835 VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
837 VABufferID slice_param_buffer;
838 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
839 CHECK_VASTATUS(va_status, "vaCreateBuffer");
840 VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
842 if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
843 // The pixel data is already put into the image by the caller.
844 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
845 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
847 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
849 // Upload the pixel data.
850 uint8_t *surface_p = nullptr;
851 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
853 if (is_uyvy(qf.frame)) {
854 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
855 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
857 const uint8_t *src = qf.frame->data_copy + field_start;
858 uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
859 memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
861 assert(is_i420(qf.frame));
862 assert(!qf.frame->interleaved); // Makes no sense for I420.
864 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
865 const uint8_t *y_src = qf.frame->data + qf.video_format.width * field_start_line;
866 const uint8_t *cb_src = y_src + width * height;
867 const uint8_t *cr_src = cb_src + (width / 2) * (height / 2);
869 uint8_t *y_dst = (unsigned char *)surface_p + resources.image.offsets[0];
870 uint8_t *cbcr_dst = (unsigned char *)surface_p + resources.image.offsets[1];
872 memcpy_with_pitch(y_dst, y_src, qf.video_format.width, resources.image.pitches[0], qf.video_format.height);
873 interleave_with_pitch(cbcr_dst, cb_src, cr_src, qf.video_format.width / 2, resources.image.pitches[1], qf.video_format.height / 2);
876 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
877 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
880 qf.frame->data_copy = nullptr;
882 // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
883 // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
884 // Exactly why is unclear, but it seems to involve L3 cache usage when there
885 // are many high-res (1080p+) images in play.
886 va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
887 CHECK_VASTATUS(va_status, "vaPutImage");
889 // Finally, stick in the JPEG header.
890 VAEncPackedHeaderParameterBuffer header_parm;
891 header_parm.type = VAEncPackedHeaderRawData;
892 header_parm.bit_length = 8 * va_data.jpeg_header.size();
894 VABufferID header_parm_buffer;
895 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
896 CHECK_VASTATUS(va_status, "vaCreateBuffer");
897 VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
899 VABufferID header_data_buffer;
900 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
901 CHECK_VASTATUS(va_status, "vaCreateBuffer");
902 VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
904 va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
905 CHECK_VASTATUS(va_status, "vaBeginPicture");
906 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
907 CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
908 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
909 CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
910 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
911 CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
912 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
913 CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
914 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
915 CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
916 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
917 CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
918 va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
919 CHECK_VASTATUS(va_status, "vaEndPicture");
921 qf.resources = move(resources);
922 qf.resource_releaser = move(release);
924 lock_guard<mutex> lock(mu);
925 frames_encoding.push(move(qf));
926 any_frames_encoding.notify_all();
929 void MJPEGEncoder::va_receiver_thread_func()
931 pthread_setname_np(pthread_self(), "MJPEG_Receive");
935 unique_lock<mutex> lock(mu);
936 any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
937 if (should_quit) return;
938 qf = move(frames_encoding.front());
939 frames_encoding.pop();
942 update_siphon_streams();
944 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index)); // Or should_encode_mjpeg_for_card() would have returned false.
945 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
947 HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
948 HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
949 assert(streams.count(multicam_id));
950 assert(streams[multicam_id].avctx != nullptr);
952 // Write audio before video, since Futatabi expects it.
953 if (qf.audio.size() > 0) {
954 write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
955 if (streams.count(siphon_id)) {
956 write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
960 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
961 CHECK_VASTATUS(va_status, "vaSyncSurface");
963 VACodedBufferSegment *segment;
964 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
965 CHECK_VASTATUS(va_status, "vaMapBuffer");
967 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
968 write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, coded_buf, segment->size);
969 if (streams.count(siphon_id)) {
970 write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, coded_buf, segment->size);
973 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
974 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
978 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
980 unsigned width = qf.video_format.width;
981 unsigned height = qf.video_format.height;
983 VectorDestinationManager dest;
984 jpeg_compress_struct cinfo;
986 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
988 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
989 if (userdata->pixel_format == PixelFormat_8BitYCbCr) {
990 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, /*y_h_samp_factor=*/2, /*y_v_samp_factor=*/1);
992 assert(qf.frame->interleaved);
993 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
995 JSAMPROW yptr[8], cbptr[8], crptr[8];
996 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
997 for (unsigned y = 0; y < qf.video_format.height; y += 8) {
999 src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
1001 memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
1002 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
1003 for (unsigned yy = 0; yy < 8; ++yy) {
1004 yptr[yy] = tmp_y + yy * width;
1005 cbptr[yy] = tmp_cb + yy * width / 2;
1006 crptr[yy] = tmp_cr + yy * width / 2;
1008 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
1011 assert(userdata->pixel_format == PixelFormat_8BitYCbCrPlanar);
1013 const movit::YCbCrFormat &ycbcr = userdata->ycbcr_format;
1014 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, ycbcr.chroma_subsampling_x, ycbcr.chroma_subsampling_y);
1015 assert(ycbcr.chroma_subsampling_y <= 2); // Or we'd need larger JSAMPROW arrays below.
1017 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
1018 const uint8_t *y_start = qf.frame->data + qf.video_format.width * field_start_line;
1019 const uint8_t *cb_start = y_start + width * height;
1020 const uint8_t *cr_start = cb_start + (width / ycbcr.chroma_subsampling_x) * (height / ycbcr.chroma_subsampling_y);
1022 size_t block_height_y = 8 * ycbcr.chroma_subsampling_y;
1023 size_t block_height_cbcr = 8;
1025 JSAMPROW yptr[16], cbptr[16], crptr[16];
1026 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
1027 for (unsigned y = 0; y < qf.video_format.height; y += block_height_y) {
1028 for (unsigned yy = 0; yy < block_height_y; ++yy) {
1029 yptr[yy] = const_cast<JSAMPROW>(y_start) + (y + yy) * width;
1031 unsigned cbcr_y = y / ycbcr.chroma_subsampling_y;
1032 for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
1033 cbptr[yy] = const_cast<JSAMPROW>(cb_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
1034 crptr[yy] = const_cast<JSAMPROW>(cr_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
1036 jpeg_write_raw_data(&cinfo, data, block_height_y);
1039 jpeg_finish_compress(&cinfo);
1044 void MJPEGEncoder::add_stream(HTTPD::StreamID stream_id)
1046 AVFormatContextWithCloser avctx;
1048 // Set up the mux. We don't use the Mux wrapper, because it's geared towards
1049 // a situation with only one video stream (and possibly one audio stream)
1050 // with known width/height, and we don't need the extra functionality it provides.
1051 avctx.reset(avformat_alloc_context());
1052 avctx->oformat = av_guess_format("nut", nullptr, nullptr);
1054 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
1055 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, &ffmpeg_contexts[stream_id], nullptr, nullptr, nullptr);
1056 avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
1057 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
1059 if (stream_id.type == HTTPD::MULTICAM_STREAM) {
1060 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
1061 add_video_stream(avctx.get());
1063 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
1064 add_audio_stream(avctx.get());
1067 assert(stream_id.type == HTTPD::SIPHON_STREAM);
1068 add_video_stream(avctx.get());
1069 add_audio_stream(avctx.get());
1071 finalize_mux(avctx.get());
1074 s.avctx = move(avctx);
1075 streams[stream_id] = move(s);
1078 void MJPEGEncoder::update_siphon_streams()
1080 // Bring the list of streams into sync with what the clients need.
1081 for (auto it = streams.begin(); it != streams.end(); ) {
1082 if (it->first.type != HTTPD::SIPHON_STREAM) {
1086 if (httpd->get_num_connected_siphon_clients(it->first.index) == 0) {
1087 av_free(it->second.avctx->pb->buffer);
1088 streams.erase(it++);
1093 for (unsigned stream_idx = 0; stream_idx < MAX_VIDEO_CARDS; ++stream_idx) {
1094 HTTPD::StreamID stream_id{ HTTPD::SIPHON_STREAM, stream_idx };
1095 if (streams.count(stream_id) == 0 && httpd->get_num_connected_siphon_clients(stream_idx) > 0) {
1096 add_stream(stream_id);
1101 void MJPEGEncoder::create_ffmpeg_context(HTTPD::StreamID stream_id)
1103 ffmpeg_contexts.emplace(stream_id, WritePacket2Context{ this, stream_id });