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.
224 VAConfigID config_id_422, config_id_420;
226 va_dpy = try_open_va(va_display, { VAProfileJPEGBaseline }, VAEntrypointEncPicture,
228 { "4:2:2", VA_RT_FORMAT_YUV422, VA_FOURCC_UYVY, &config_id_422, &uyvy_format },
229 // We'd prefer VA_FOURCC_I420, but it's not supported by Intel's driver.
230 { "4:2:0", VA_RT_FORMAT_YUV420, VA_FOURCC_NV12, &config_id_420, &nv12_format }
232 /*chosen_profile=*/nullptr, &error);
233 if (va_dpy == nullptr) {
234 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
237 encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
238 if (va_dpy != nullptr) {
239 va_pool.reset(new VAResourcePool(va_dpy->va_dpy, uyvy_format, nv12_format, config_id_422, config_id_420, /*with_data_buffer=*/true));
240 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
243 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
244 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
245 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
246 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
247 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
248 global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
253 MJPEGEncoder::~MJPEGEncoder()
255 for (auto &id_and_stream : streams) {
256 av_free(id_and_stream.second.avctx->pb->buffer);
259 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
260 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
261 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
262 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
263 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
264 global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
267 void MJPEGEncoder::stop()
274 any_frames_to_be_encoded.notify_all();
275 any_frames_encoding.notify_all();
276 encoder_thread.join();
277 if (va_dpy != nullptr) {
278 va_receiver_thread.join();
284 bool is_uyvy(RefCountedFrame frame)
286 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
287 return userdata->pixel_format == PixelFormat_8BitYCbCr && frame->interleaved;
290 bool is_i420(RefCountedFrame frame)
292 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
293 return userdata->pixel_format == PixelFormat_8BitYCbCrPlanar &&
294 userdata->ycbcr_format.chroma_subsampling_x == 2 &&
295 userdata->ycbcr_format.chroma_subsampling_y == 2;
300 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)
302 if (video_format.width == 0 || video_format.height == 0) {
303 ++metric_mjpeg_frames_zero_size_dropped;
306 if (video_format.interlaced) {
307 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
308 ++metric_mjpeg_frames_interlaced_dropped;
311 if (!is_uyvy(frame) && !is_i420(frame)) {
312 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
313 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
316 if (video_format.width > 4096 || video_format.height > 4096) {
317 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
318 ++metric_mjpeg_frames_oversized_dropped;
322 lock_guard<mutex> lock(mu);
323 if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
324 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
325 ++metric_mjpeg_overrun_dropped;
328 ++metric_mjpeg_overrun_submitted;
329 frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio), white_balance });
330 any_frames_to_be_encoded.notify_all();
333 bool MJPEGEncoder::should_encode_mjpeg_for_card(unsigned card_index)
335 // Only bother doing MJPEG encoding if there are any connected clients
336 // that want the stream.
337 if (httpd->get_num_connected_multicam_clients() == 0 &&
338 httpd->get_num_connected_siphon_clients(card_index) == 0) {
342 auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
343 return (it != global_flags.card_to_mjpeg_stream_export.end());
346 void MJPEGEncoder::encoder_thread_func()
348 pthread_setname_np(pthread_self(), "MJPEG_Encode");
349 posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
350 posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
351 posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
352 posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
357 unique_lock<mutex> lock(mu);
358 any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
359 if (should_quit) break;
360 qf = move(frames_to_be_encoded.front());
361 frames_to_be_encoded.pop();
364 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index)); // Or should_encode_mjpeg_for_card() would have returned false.
365 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
367 if (va_dpy != nullptr) {
368 // Will call back in the receiver thread.
369 encode_jpeg_va(move(qf));
371 update_siphon_streams();
373 HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
374 HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
375 assert(streams.count(multicam_id));
377 // Write audio before video, since Futatabi expects it.
378 if (qf.audio.size() > 0) {
379 write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
380 if (streams.count(siphon_id)) {
381 write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
385 // Encode synchronously, in the same thread.
386 vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
387 write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, jpeg.data(), jpeg.size());
388 if (streams.count(siphon_id)) {
389 write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, jpeg.data(), jpeg.size());
400 void MJPEGEncoder::write_mjpeg_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const uint8_t *jpeg, size_t jpeg_size)
403 memset(&pkt, 0, sizeof(pkt));
405 pkt.data = const_cast<uint8_t *>(jpeg);
406 pkt.size = jpeg_size;
407 pkt.stream_index = stream_index;
408 pkt.flags = AV_PKT_FLAG_KEY;
409 AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
410 pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
413 if (av_write_frame(avctx, &pkt) < 0) {
414 fprintf(stderr, "av_write_frame() failed\n");
419 void MJPEGEncoder::write_audio_packet(AVFormatContext *avctx, int64_t pts, unsigned stream_index, const vector<int32_t> &audio)
422 memset(&pkt, 0, sizeof(pkt));
424 pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
425 pkt.size = audio.size() * sizeof(audio[0]);
426 pkt.stream_index = stream_index;
427 pkt.flags = AV_PKT_FLAG_KEY;
428 AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
429 pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
430 size_t num_stereo_samples = audio.size() / 2;
431 pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
433 if (av_write_frame(avctx, &pkt) < 0) {
434 fprintf(stderr, "av_write_frame() failed\n");
439 class VABufferDestroyer {
441 VABufferDestroyer(VADisplay dpy, VABufferID buf)
442 : dpy(dpy), buf(buf) {}
444 ~VABufferDestroyer() {
445 VAStatus va_status = vaDestroyBuffer(dpy, buf);
446 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
456 void push16(uint16_t val, string *str)
458 str->push_back(val >> 8);
459 str->push_back(val & 0xff);
462 void push32(uint32_t val, string *str)
464 str->push_back(val >> 24);
465 str->push_back((val >> 16) & 0xff);
466 str->push_back((val >> 8) & 0xff);
467 str->push_back(val & 0xff);
472 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)
475 cinfo->err = jpeg_std_error(&jerr);
476 jpeg_create_compress(cinfo);
478 cinfo->dest = (jpeg_destination_mgr *)dest;
480 cinfo->input_components = 3;
481 jpeg_set_defaults(cinfo);
482 jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
484 cinfo->image_width = width;
485 cinfo->image_height = height;
486 cinfo->raw_data_in = true;
487 jpeg_set_colorspace(cinfo, JCS_YCbCr);
488 cinfo->comp_info[0].h_samp_factor = y_h_samp_factor;
489 cinfo->comp_info[0].v_samp_factor = y_v_samp_factor;
490 cinfo->comp_info[1].h_samp_factor = 1;
491 cinfo->comp_info[1].v_samp_factor = 1;
492 cinfo->comp_info[2].h_samp_factor = 1;
493 cinfo->comp_info[2].v_samp_factor = 1;
494 cinfo->CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though.
495 jpeg_start_compress(cinfo, true);
497 if (fabs(white_balance.r - 1.0f) > 1e-3 ||
498 fabs(white_balance.g - 1.0f) > 1e-3 ||
499 fabs(white_balance.b - 1.0f) > 1e-3) {
500 // Convert from (linear) RGB to XYZ.
501 Matrix3d rgb_to_xyz_matrix = movit::ColorspaceConversionEffect::get_xyz_matrix(COLORSPACE_sRGB);
502 Vector3d xyz = rgb_to_xyz_matrix * Vector3d(white_balance.r, white_balance.g, white_balance.b);
504 // Convert from XYZ to xyz by normalizing.
505 xyz /= (xyz[0] + xyz[1] + xyz[2]);
507 // Create a very rudimentary EXIF header to hold our white point.
510 // Exif header, followed by some padding.
514 // TIFF header first:
515 exif += "MM"; // Big endian.
520 // Offset of first IFD (relative to the MM, immediately after the header).
521 push32(exif.size() - 6 + 4, &exif);
523 // Now the actual IFD.
528 // WhitePoint tag ID.
529 push16(0x13e, &exif);
534 // Two values (x and y; z is implicit due to normalization).
537 // Offset (relative to the MM, immediately after the last IFD).
538 push32(exif.size() - 6 + 8, &exif);
543 // The actual values.
544 push32(lrintf(xyz[0] * 10000.0f), &exif);
545 push32(10000, &exif);
546 push32(lrintf(xyz[1] * 10000.0f), &exif);
547 push32(10000, &exif);
549 jpeg_write_marker(cinfo, JPEG_APP0 + 1, (const JOCTET *)exif.data(), exif.size());
552 // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
553 // (and nothing else).
554 jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
557 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)
559 VectorDestinationManager dest;
560 init_jpeg(width, height, white_balance, &dest, cinfo, y_h_samp_factor, y_v_samp_factor);
562 // Make a dummy black image; there's seemingly no other easy way of
563 // making libjpeg outputting all of its headers.
564 assert(y_v_samp_factor <= 2); // Or we'd need larger JSAMPROW arrays below.
565 size_t block_height_y = 8 * y_v_samp_factor;
566 size_t block_height_cbcr = 8;
568 JSAMPROW yptr[16], cbptr[16], crptr[16];
569 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
570 memset(tmp_y, 0, 4096);
571 memset(tmp_cb, 0, 4096);
572 memset(tmp_cr, 0, 4096);
573 for (unsigned yy = 0; yy < block_height_y; ++yy) {
576 for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
580 for (unsigned y = 0; y < height; y += block_height_y) {
581 jpeg_write_raw_data(cinfo, data, block_height_y);
583 jpeg_finish_compress(cinfo);
585 // We're only interested in the header, not the data after it.
586 dest.term_destination();
587 for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
588 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) { // Start of scan (SOS).
589 unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
590 dest.dest.resize(i + len + 2);
598 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)
600 VAKey key{width, height, y_h_samp_factor, y_v_samp_factor, white_balance};
601 if (va_data_for_parameters.count(key)) {
602 return va_data_for_parameters[key];
605 // Use libjpeg to generate a header and set sane defaults for e.g.
606 // quantization tables. Then do the actual encode with VA-API.
607 jpeg_compress_struct cinfo;
608 vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, y_h_samp_factor, y_v_samp_factor, &cinfo);
610 // Picture parameters.
611 VAEncPictureParameterBufferJPEG pic_param;
612 memset(&pic_param, 0, sizeof(pic_param));
613 pic_param.reconstructed_picture = VA_INVALID_ID;
614 pic_param.picture_width = cinfo.image_width;
615 pic_param.picture_height = cinfo.image_height;
616 for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
617 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
618 pic_param.component_id[component_idx] = comp->component_id;
619 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
621 pic_param.num_components = cinfo.num_components;
622 pic_param.num_scan = 1;
623 pic_param.sample_bit_depth = 8;
624 pic_param.coded_buf = VA_INVALID_ID; // To be filled out by caller.
625 pic_param.pic_flags.bits.huffman = 1;
626 pic_param.quality = 50; // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
628 // Quantization matrices.
629 VAQMatrixBufferJPEG q;
630 memset(&q, 0, sizeof(q));
632 q.load_lum_quantiser_matrix = true;
633 q.load_chroma_quantiser_matrix = true;
634 for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
635 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
636 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
637 if (qtbl == nullptr) continue;
639 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
640 for (int i = 0; i < 64; ++i) {
641 if (qtbl->quantval[i] > 255) {
642 fprintf(stderr, "Baseline JPEG only!\n");
645 qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
649 // Huffman tables (arithmetic is not supported).
650 VAHuffmanTableBufferJPEGBaseline huff;
651 memset(&huff, 0, sizeof(huff));
653 for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
654 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
655 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
656 if (ac_hufftbl == nullptr) {
657 assert(dc_hufftbl == nullptr);
658 huff.load_huffman_table[huff_tbl_idx] = 0;
660 assert(dc_hufftbl != nullptr);
661 huff.load_huffman_table[huff_tbl_idx] = 1;
663 for (int i = 0; i < 16; ++i) {
664 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
666 for (int i = 0; i < 12; ++i) {
667 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
669 for (int i = 0; i < 16; ++i) {
670 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
672 for (int i = 0; i < 162; ++i) {
673 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
678 // Slice parameters (metadata about the slice).
679 VAEncSliceParameterBufferJPEG parms;
680 memset(&parms, 0, sizeof(parms));
681 for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
682 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
683 parms.components[component_idx].component_selector = comp->component_id;
684 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
685 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
686 if (parms.components[component_idx].dc_table_selector > 1 ||
687 parms.components[component_idx].ac_table_selector > 1) {
688 fprintf(stderr, "Uses too many Huffman tables\n");
692 parms.num_components = cinfo.num_components;
693 parms.restart_interval = cinfo.restart_interval;
695 jpeg_destroy_compress(&cinfo);
698 ret.jpeg_header = move(jpeg_header);
699 ret.pic_param = pic_param;
703 va_data_for_parameters[key] = ret;
707 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
709 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
710 unsigned width = qf.video_format.width;
711 unsigned height = qf.video_format.height;
713 VAResourcePool::VAResources resources;
714 ReleaseVAResources release;
715 if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
716 assert(is_uyvy(qf.frame));
717 resources = move(userdata->va_resources);
718 release = move(userdata->va_resources_release);
720 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
721 if (is_uyvy(qf.frame)) {
722 resources = va_pool->get_va_resources(width, height, VA_FOURCC_UYVY);
724 assert(is_i420(qf.frame));
725 resources = va_pool->get_va_resources(width, height, VA_FOURCC_NV12);
727 release = ReleaseVAResources(va_pool.get(), resources);
730 int y_h_samp_factor, y_v_samp_factor;
731 if (is_uyvy(qf.frame)) {
732 // 4:2:2 (sample Y' twice as often horizontally as Cb or Cr, vertical is left alone).
736 // 4:2:0 (sample Y' twice as often as Cb or Cr, in both directions)
737 assert(is_i420(qf.frame));
742 VAData va_data = get_va_data_for_parameters(width, height, y_h_samp_factor, y_v_samp_factor, qf.white_balance);
743 va_data.pic_param.coded_buf = resources.data_buffer;
745 VABufferID pic_param_buffer;
746 VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
747 CHECK_VASTATUS(va_status, "vaCreateBuffer");
748 VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
751 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
752 CHECK_VASTATUS(va_status, "vaCreateBuffer");
753 VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
755 VABufferID huff_buffer;
756 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
757 CHECK_VASTATUS(va_status, "vaCreateBuffer");
758 VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
760 VABufferID slice_param_buffer;
761 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
762 CHECK_VASTATUS(va_status, "vaCreateBuffer");
763 VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
765 if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
766 // The pixel data is already put into the image by the caller.
767 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
768 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
770 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
772 // Upload the pixel data.
773 uint8_t *surface_p = nullptr;
774 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
776 if (is_uyvy(qf.frame)) {
777 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
778 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
780 const uint8_t *src = qf.frame->data_copy + field_start;
781 uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
782 memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
784 assert(is_i420(qf.frame));
785 assert(!qf.frame->interleaved); // Makes no sense for I420.
787 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
788 const uint8_t *y_src = qf.frame->data + qf.video_format.width * field_start_line;
789 const uint8_t *cb_src = y_src + width * height;
790 const uint8_t *cr_src = cb_src + (width / 2) * (height / 2);
792 uint8_t *y_dst = (unsigned char *)surface_p + resources.image.offsets[0];
793 uint8_t *cbcr_dst = (unsigned char *)surface_p + resources.image.offsets[1];
795 memcpy_with_pitch(y_dst, y_src, qf.video_format.width, resources.image.pitches[0], qf.video_format.height);
796 interleave_with_pitch(cbcr_dst, cb_src, cr_src, qf.video_format.width / 2, resources.image.pitches[1], qf.video_format.height / 2);
799 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
800 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
803 qf.frame->data_copy = nullptr;
805 // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
806 // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
807 // Exactly why is unclear, but it seems to involve L3 cache usage when there
808 // are many high-res (1080p+) images in play.
809 va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
810 CHECK_VASTATUS(va_status, "vaPutImage");
812 // Finally, stick in the JPEG header.
813 VAEncPackedHeaderParameterBuffer header_parm;
814 header_parm.type = VAEncPackedHeaderRawData;
815 header_parm.bit_length = 8 * va_data.jpeg_header.size();
817 VABufferID header_parm_buffer;
818 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
819 CHECK_VASTATUS(va_status, "vaCreateBuffer");
820 VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
822 VABufferID header_data_buffer;
823 va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
824 CHECK_VASTATUS(va_status, "vaCreateBuffer");
825 VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
827 va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
828 CHECK_VASTATUS(va_status, "vaBeginPicture");
829 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
830 CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
831 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
832 CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
833 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
834 CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
835 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
836 CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
837 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
838 CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
839 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
840 CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
841 va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
842 CHECK_VASTATUS(va_status, "vaEndPicture");
844 qf.resources = move(resources);
845 qf.resource_releaser = move(release);
847 lock_guard<mutex> lock(mu);
848 frames_encoding.push(move(qf));
849 any_frames_encoding.notify_all();
852 void MJPEGEncoder::va_receiver_thread_func()
854 pthread_setname_np(pthread_self(), "MJPEG_Receive");
858 unique_lock<mutex> lock(mu);
859 any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
860 if (should_quit) return;
861 qf = move(frames_encoding.front());
862 frames_encoding.pop();
865 update_siphon_streams();
867 assert(global_flags.card_to_mjpeg_stream_export.count(qf.card_index)); // Or should_encode_mjpeg_for_card() would have returned false.
868 int stream_index = global_flags.card_to_mjpeg_stream_export[qf.card_index];
870 HTTPD::StreamID multicam_id{ HTTPD::MULTICAM_STREAM, 0 };
871 HTTPD::StreamID siphon_id{ HTTPD::SIPHON_STREAM, qf.card_index };
872 assert(streams.count(multicam_id));
873 assert(streams[multicam_id].avctx != nullptr);
875 // Write audio before video, since Futatabi expects it.
876 if (qf.audio.size() > 0) {
877 write_audio_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index + global_flags.card_to_mjpeg_stream_export.size(), qf.audio);
878 if (streams.count(siphon_id)) {
879 write_audio_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/1, qf.audio);
883 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
884 CHECK_VASTATUS(va_status, "vaSyncSurface");
886 VACodedBufferSegment *segment;
887 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
888 CHECK_VASTATUS(va_status, "vaMapBuffer");
890 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
891 write_mjpeg_packet(streams[multicam_id].avctx.get(), qf.pts, stream_index, coded_buf, segment->size);
892 if (streams.count(siphon_id)) {
893 write_mjpeg_packet(streams[siphon_id].avctx.get(), qf.pts, /*stream_index=*/0, coded_buf, segment->size);
896 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
897 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
901 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
903 unsigned width = qf.video_format.width;
904 unsigned height = qf.video_format.height;
906 VectorDestinationManager dest;
907 jpeg_compress_struct cinfo;
909 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
911 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
912 if (userdata->pixel_format == PixelFormat_8BitYCbCr) {
913 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, /*y_h_samp_factor=*/2, /*y_v_samp_factor=*/1);
915 assert(qf.frame->interleaved);
916 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
918 JSAMPROW yptr[8], cbptr[8], crptr[8];
919 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
920 for (unsigned y = 0; y < qf.video_format.height; y += 8) {
922 src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
924 memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
925 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
926 for (unsigned yy = 0; yy < 8; ++yy) {
927 yptr[yy] = tmp_y + yy * width;
928 cbptr[yy] = tmp_cb + yy * width / 2;
929 crptr[yy] = tmp_cr + yy * width / 2;
931 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
934 assert(userdata->pixel_format == PixelFormat_8BitYCbCrPlanar);
936 const movit::YCbCrFormat &ycbcr = userdata->ycbcr_format;
937 init_jpeg(width, height, qf.white_balance, &dest, &cinfo, ycbcr.chroma_subsampling_x, ycbcr.chroma_subsampling_y);
938 assert(ycbcr.chroma_subsampling_y <= 2); // Or we'd need larger JSAMPROW arrays below.
940 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
941 const uint8_t *y_start = qf.frame->data + qf.video_format.width * field_start_line;
942 const uint8_t *cb_start = y_start + width * height;
943 const uint8_t *cr_start = cb_start + (width / ycbcr.chroma_subsampling_x) * (height / ycbcr.chroma_subsampling_y);
945 size_t block_height_y = 8 * ycbcr.chroma_subsampling_y;
946 size_t block_height_cbcr = 8;
948 JSAMPROW yptr[16], cbptr[16], crptr[16];
949 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
950 for (unsigned y = 0; y < qf.video_format.height; y += block_height_y) {
951 for (unsigned yy = 0; yy < block_height_y; ++yy) {
952 yptr[yy] = const_cast<JSAMPROW>(y_start) + (y + yy) * width;
954 unsigned cbcr_y = y / ycbcr.chroma_subsampling_y;
955 for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
956 cbptr[yy] = const_cast<JSAMPROW>(cb_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
957 crptr[yy] = const_cast<JSAMPROW>(cr_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
959 jpeg_write_raw_data(&cinfo, data, block_height_y);
962 jpeg_finish_compress(&cinfo);
967 void MJPEGEncoder::add_stream(HTTPD::StreamID stream_id)
969 AVFormatContextWithCloser avctx;
971 // Set up the mux. We don't use the Mux wrapper, because it's geared towards
972 // a situation with only one video stream (and possibly one audio stream)
973 // with known width/height, and we don't need the extra functionality it provides.
974 avctx.reset(avformat_alloc_context());
975 avctx->oformat = av_guess_format("nut", nullptr, nullptr);
977 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
978 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, &ffmpeg_contexts[stream_id], nullptr, nullptr, nullptr);
979 avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
980 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
982 if (stream_id.type == HTTPD::MULTICAM_STREAM) {
983 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
984 add_video_stream(avctx.get());
986 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
987 add_audio_stream(avctx.get());
990 assert(stream_id.type == HTTPD::SIPHON_STREAM);
991 add_video_stream(avctx.get());
992 add_audio_stream(avctx.get());
994 finalize_mux(avctx.get());
997 s.avctx = move(avctx);
998 streams[stream_id] = move(s);
1001 void MJPEGEncoder::update_siphon_streams()
1003 // Bring the list of streams into sync with what the clients need.
1004 for (auto it = streams.begin(); it != streams.end(); ) {
1005 if (it->first.type != HTTPD::SIPHON_STREAM) {
1009 if (httpd->get_num_connected_siphon_clients(it->first.index) == 0) {
1010 av_free(it->second.avctx->pb->buffer);
1011 streams.erase(it++);
1016 for (unsigned stream_idx = 0; stream_idx < MAX_VIDEO_CARDS; ++stream_idx) {
1017 HTTPD::StreamID stream_id{ HTTPD::SIPHON_STREAM, stream_idx };
1018 if (streams.count(stream_id) == 0 && httpd->get_num_connected_siphon_clients(stream_idx) > 0) {
1019 add_stream(stream_id);
1024 void MJPEGEncoder::create_ffmpeg_context(HTTPD::StreamID stream_id)
1026 ffmpeg_contexts.emplace(stream_id, WritePacket2Context{ this, stream_id });