1 #include "mjpeg_encoder.h"
11 #include <libavformat/avformat.h>
15 #include "shared/ffmpeg_raii.h"
17 #include "shared/httpd.h"
18 #include "shared/memcpy_interleaved.h"
19 #include "shared/metrics.h"
20 #include "pbo_frame_allocator.h"
21 #include "shared/timebase.h"
22 #include "va_display_with_cleanup.h"
25 #include <va/va_drm.h>
26 #include <va/va_x11.h>
28 using namespace bmusb;
31 static VAImageFormat uyvy_format;
33 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
35 // From libjpeg (although it's of course identical between implementations).
36 static const int jpeg_natural_order[DCTSIZE2] = {
37 0, 1, 8, 16, 9, 2, 3, 10,
38 17, 24, 32, 25, 18, 11, 4, 5,
39 12, 19, 26, 33, 40, 48, 41, 34,
40 27, 20, 13, 6, 7, 14, 21, 28,
41 35, 42, 49, 56, 57, 50, 43, 36,
42 29, 22, 15, 23, 30, 37, 44, 51,
43 58, 59, 52, 45, 38, 31, 39, 46,
44 53, 60, 61, 54, 47, 55, 62, 63,
47 struct VectorDestinationManager {
48 jpeg_destination_mgr pub;
49 std::vector<uint8_t> dest;
51 VectorDestinationManager()
53 pub.init_destination = init_destination_thunk;
54 pub.empty_output_buffer = empty_output_buffer_thunk;
55 pub.term_destination = term_destination_thunk;
58 static void init_destination_thunk(j_compress_ptr ptr)
60 ((VectorDestinationManager *)(ptr->dest))->init_destination();
63 inline void init_destination()
68 static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
70 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
73 inline bool empty_output_buffer()
75 make_room(dest.size()); // Should ignore pub.free_in_buffer!
79 inline void make_room(size_t bytes_used)
81 dest.resize(bytes_used + 4096);
82 dest.resize(dest.capacity());
83 pub.next_output_byte = dest.data() + bytes_used;
84 pub.free_in_buffer = dest.size() - bytes_used;
87 static void term_destination_thunk(j_compress_ptr ptr)
89 ((VectorDestinationManager *)(ptr->dest))->term_destination();
92 inline void term_destination()
94 dest.resize(dest.size() - pub.free_in_buffer);
97 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
99 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
101 MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
102 return engine->write_packet2(buf, buf_size, type, time);
105 int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
107 if (type == AVIO_DATA_MARKER_HEADER) {
108 mux_header.append((char *)buf, buf_size);
109 httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
111 httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
116 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
119 // Set up the mux. We don't use the Mux wrapper, because it's geared towards
120 // a situation with only one video stream (and possibly one audio stream)
121 // with known width/height, and we don't need the extra functionality it provides.
122 avctx.reset(avformat_alloc_context());
123 avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
125 uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
126 avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
127 avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
128 avctx->flags = AVFMT_FLAG_CUSTOM_IO;
130 for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
131 AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
132 if (stream == nullptr) {
133 fprintf(stderr, "avformat_new_stream() failed\n");
136 stream->time_base = AVRational{ 1, TIMEBASE };
137 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
138 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
140 // Used for aspect ratio only. Can change without notice (the mux won't care).
141 stream->codecpar->width = global_flags.width;
142 stream->codecpar->height = global_flags.height;
144 // TODO: We could perhaps use the interpretation for each card here
145 // (or at least the command-line flags) instead of the defaults,
146 // but what would we do when they change?
147 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
148 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
149 stream->codecpar->color_space = AVCOL_SPC_BT709;
150 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
151 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
152 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
155 AVDictionary *options = NULL;
156 vector<pair<string, string>> opts = MUX_OPTS;
157 for (pair<string, string> opt : opts) {
158 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
160 if (avformat_write_header(avctx.get(), &options) < 0) {
161 fprintf(stderr, "avformat_write_header() failed\n");
165 // Initialize VA-API.
167 va_dpy = try_open_va(va_display, &error, &config_id);
168 if (va_dpy == nullptr) {
169 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
172 encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
173 if (va_dpy != nullptr) {
174 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
177 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
178 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
179 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
180 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
181 global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
182 global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
187 MJPEGEncoder::~MJPEGEncoder()
189 av_free(avctx->pb->buffer);
191 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
192 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
193 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
194 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
195 global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
196 global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
199 void MJPEGEncoder::stop()
206 any_frames_to_be_encoded.notify_all();
207 any_frames_encoding.notify_all();
208 encoder_thread.join();
209 if (va_dpy != nullptr) {
210 va_receiver_thread.join();
214 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
216 unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
217 if (va_dpy == nullptr) {
218 if (error) *error = "Opening VA display failed";
221 int major_ver, minor_ver;
222 VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
223 if (va_status != VA_STATUS_SUCCESS) {
225 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
226 if (error != nullptr) *error = buf;
230 VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
231 va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
232 &attr, 1, config_id);
233 if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
234 if (error != nullptr) *error = "No hardware support";
236 } else if (va_status != VA_STATUS_SUCCESS) {
238 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
239 if (error != nullptr) *error = buf;
243 // TODO: Unify with the code in Futatabi.
244 int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
245 assert(num_formats > 0);
247 unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
248 va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
249 if (va_status != VA_STATUS_SUCCESS) {
251 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
252 if (error != nullptr) *error = buf;
257 for (int i = 0; i < num_formats; ++i) {
258 if (formats[i].fourcc == VA_FOURCC_UYVY) {
259 memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
265 if (error != nullptr) *error = "UYVY format not found";
272 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)
274 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
275 if (video_format.width == 0 || video_format.height == 0) {
276 ++metric_mjpeg_frames_zero_size_dropped;
279 if (video_format.interlaced) {
280 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
281 ++metric_mjpeg_frames_interlaced_dropped;
284 if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
285 !frame->interleaved) {
286 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
287 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
290 if (video_format.width > 4096 || video_format.height > 4096) {
291 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
292 ++metric_mjpeg_frames_oversized_dropped;
296 lock_guard<mutex> lock(mu);
297 if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
298 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
299 ++metric_mjpeg_overrun_dropped;
302 ++metric_mjpeg_overrun_submitted;
303 frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
304 any_frames_to_be_encoded.notify_all();
307 int MJPEGEncoder::get_mjpeg_stream_for_card(unsigned card_index)
309 // Only bother doing MJPEG encoding if there are any connected clients
310 // that want the stream.
311 if (httpd->get_num_connected_multicam_clients() == 0) {
315 auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
316 if (it == global_flags.card_to_mjpeg_stream_export.end()) {
322 void MJPEGEncoder::encoder_thread_func()
324 pthread_setname_np(pthread_self(), "MJPEG_Encode");
325 posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
326 posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
327 posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
328 posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
333 unique_lock<mutex> lock(mu);
334 any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
335 if (should_quit) break;
336 qf = move(frames_to_be_encoded.front());
337 frames_to_be_encoded.pop();
340 if (va_dpy != nullptr) {
341 // Will call back in the receiver thread.
342 encode_jpeg_va(move(qf));
344 // Encode synchronously, in the same thread.
345 vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
346 write_mjpeg_packet(qf.pts, qf.card_index, jpeg.data(), jpeg.size());
356 void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const uint8_t *jpeg, size_t jpeg_size)
359 memset(&pkt, 0, sizeof(pkt));
361 pkt.data = const_cast<uint8_t *>(jpeg);
362 pkt.size = jpeg_size;
363 pkt.stream_index = card_index;
364 pkt.flags = AV_PKT_FLAG_KEY;
365 AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
366 pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
368 if (av_write_frame(avctx.get(), &pkt) < 0) {
369 fprintf(stderr, "av_write_frame() failed\n");
374 class VABufferDestroyer {
376 VABufferDestroyer(VADisplay dpy, VABufferID buf)
377 : dpy(dpy), buf(buf) {}
379 ~VABufferDestroyer() {
380 VAStatus va_status = vaDestroyBuffer(dpy, buf);
381 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
389 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
392 lock_guard<mutex> lock(va_resources_mutex);
393 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
394 if (it->width == width && it->height == height) {
395 VAResources ret = *it;
396 va_resources_freelist.erase(it);
407 VASurfaceAttrib attrib;
408 attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
409 attrib.type = VASurfaceAttribPixelFormat;
410 attrib.value.type = VAGenericValueTypeInteger;
411 attrib.value.value.i = VA_FOURCC_UYVY;
413 VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
415 &ret.surface, 1, &attrib, 1);
416 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
418 va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
419 CHECK_VASTATUS(va_status, "vaCreateContext");
421 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
422 CHECK_VASTATUS(va_status, "vaCreateBuffer");
424 va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
425 CHECK_VASTATUS(va_status, "vaCreateImage");
430 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
432 lock_guard<mutex> lock(va_resources_mutex);
433 if (va_resources_freelist.size() > 50) {
434 auto it = va_resources_freelist.end();
437 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
438 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
440 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
441 CHECK_VASTATUS(va_status, "vaDestroyContext");
443 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
444 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
446 va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
447 CHECK_VASTATUS(va_status, "vaDestroyImage");
449 va_resources_freelist.erase(it);
452 va_resources_freelist.push_front(resources);
455 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
458 cinfo->err = jpeg_std_error(&jerr);
459 jpeg_create_compress(cinfo);
461 cinfo->dest = (jpeg_destination_mgr *)dest;
463 cinfo->input_components = 3;
464 jpeg_set_defaults(cinfo);
465 jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
467 cinfo->image_width = width;
468 cinfo->image_height = height;
469 cinfo->raw_data_in = true;
470 jpeg_set_colorspace(cinfo, JCS_YCbCr);
471 cinfo->comp_info[0].h_samp_factor = 2;
472 cinfo->comp_info[0].v_samp_factor = 1;
473 cinfo->comp_info[1].h_samp_factor = 1;
474 cinfo->comp_info[1].v_samp_factor = 1;
475 cinfo->comp_info[2].h_samp_factor = 1;
476 cinfo->comp_info[2].v_samp_factor = 1;
477 cinfo->CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though.
478 jpeg_start_compress(cinfo, true);
480 // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
481 // (and nothing else).
482 jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
485 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
487 VectorDestinationManager dest;
488 init_jpeg_422(width, height, &dest, cinfo);
490 // Make a dummy black image; there's seemingly no other easy way of
491 // making libjpeg outputting all of its headers.
492 JSAMPROW yptr[8], cbptr[8], crptr[8];
493 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
494 memset(tmp_y, 0, 4096);
495 memset(tmp_cb, 0, 4096);
496 memset(tmp_cr, 0, 4096);
497 for (unsigned yy = 0; yy < 8; ++yy) {
502 for (unsigned y = 0; y < height; y += 8) {
503 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
505 jpeg_finish_compress(cinfo);
507 // We're only interested in the header, not the data after it.
508 dest.term_destination();
509 for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
510 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) { // Start of scan (SOS).
511 unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
512 dest.dest.resize(i + len + 2);
520 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
522 pair<unsigned, unsigned> key(width, height);
523 if (va_data_for_resolution.count(key)) {
524 return va_data_for_resolution[key];
527 // Use libjpeg to generate a header and set sane defaults for e.g.
528 // quantization tables. Then do the actual encode with VA-API.
529 jpeg_compress_struct cinfo;
530 vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
532 // Picture parameters.
533 VAEncPictureParameterBufferJPEG pic_param;
534 memset(&pic_param, 0, sizeof(pic_param));
535 pic_param.reconstructed_picture = VA_INVALID_ID;
536 pic_param.picture_width = cinfo.image_width;
537 pic_param.picture_height = cinfo.image_height;
538 for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
539 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
540 pic_param.component_id[component_idx] = comp->component_id;
541 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
543 pic_param.num_components = cinfo.num_components;
544 pic_param.num_scan = 1;
545 pic_param.sample_bit_depth = 8;
546 pic_param.coded_buf = VA_INVALID_ID; // To be filled out by caller.
547 pic_param.pic_flags.bits.huffman = 1;
548 pic_param.quality = 50; // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
550 // Quantization matrices.
551 VAQMatrixBufferJPEG q;
552 memset(&q, 0, sizeof(q));
554 q.load_lum_quantiser_matrix = true;
555 q.load_chroma_quantiser_matrix = true;
556 for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
557 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
558 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
559 if (qtbl == nullptr) continue;
561 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
562 for (int i = 0; i < 64; ++i) {
563 if (qtbl->quantval[i] > 255) {
564 fprintf(stderr, "Baseline JPEG only!\n");
567 qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
571 // Huffman tables (arithmetic is not supported).
572 VAHuffmanTableBufferJPEGBaseline huff;
573 memset(&huff, 0, sizeof(huff));
575 for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
576 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
577 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
578 if (ac_hufftbl == nullptr) {
579 assert(dc_hufftbl == nullptr);
580 huff.load_huffman_table[huff_tbl_idx] = 0;
582 assert(dc_hufftbl != nullptr);
583 huff.load_huffman_table[huff_tbl_idx] = 1;
585 for (int i = 0; i < 16; ++i) {
586 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
588 for (int i = 0; i < 12; ++i) {
589 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
591 for (int i = 0; i < 16; ++i) {
592 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
594 for (int i = 0; i < 162; ++i) {
595 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
600 // Slice parameters (metadata about the slice).
601 VAEncSliceParameterBufferJPEG parms;
602 memset(&parms, 0, sizeof(parms));
603 for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
604 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
605 parms.components[component_idx].component_selector = comp->component_id;
606 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
607 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
608 if (parms.components[component_idx].dc_table_selector > 1 ||
609 parms.components[component_idx].ac_table_selector > 1) {
610 fprintf(stderr, "Uses too many Huffman tables\n");
614 parms.num_components = cinfo.num_components;
615 parms.restart_interval = cinfo.restart_interval;
617 jpeg_destroy_compress(&cinfo);
620 ret.jpeg_header = move(jpeg_header);
621 ret.pic_param = pic_param;
625 va_data_for_resolution[key] = ret;
629 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
631 PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
632 unsigned width = qf.video_format.width;
633 unsigned height = qf.video_format.height;
635 VAResources resources;
636 ReleaseVAResources release;
637 if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
638 resources = move(userdata->va_resources);
639 release = move(userdata->va_resources_release);
641 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
642 resources = get_va_resources(width, height);
643 release = ReleaseVAResources(this, resources);
646 VAData va_data = get_va_data_for_resolution(width, height);
647 va_data.pic_param.coded_buf = resources.data_buffer;
649 VABufferID pic_param_buffer;
650 VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
651 CHECK_VASTATUS(va_status, "vaCreateBuffer");
652 VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
655 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
656 CHECK_VASTATUS(va_status, "vaCreateBuffer");
657 VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
659 VABufferID huff_buffer;
660 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
661 CHECK_VASTATUS(va_status, "vaCreateBuffer");
662 VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
664 VABufferID slice_param_buffer;
665 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
666 CHECK_VASTATUS(va_status, "vaCreateBuffer");
667 VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
669 if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
670 // The pixel data is already put into the image by the caller.
671 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
672 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
674 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
676 // Upload the pixel data.
677 uint8_t *surface_p = nullptr;
678 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
680 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
681 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
684 const uint8_t *src = qf.frame->data_copy + field_start;
685 uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
686 memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
689 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
690 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
693 qf.frame->data_copy = nullptr;
695 // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
696 // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
697 // Exactly why is unclear, but it seems to involve L3 cache usage when there
698 // are many high-res (1080p+) images in play.
699 va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
700 CHECK_VASTATUS(va_status, "vaPutImage");
702 // Finally, stick in the JPEG header.
703 VAEncPackedHeaderParameterBuffer header_parm;
704 header_parm.type = VAEncPackedHeaderRawData;
705 header_parm.bit_length = 8 * va_data.jpeg_header.size();
707 VABufferID header_parm_buffer;
708 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
709 CHECK_VASTATUS(va_status, "vaCreateBuffer");
710 VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
712 VABufferID header_data_buffer;
713 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
714 CHECK_VASTATUS(va_status, "vaCreateBuffer");
715 VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
717 va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
718 CHECK_VASTATUS(va_status, "vaBeginPicture");
719 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
720 CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
721 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
722 CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
723 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
724 CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
725 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
726 CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
727 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
728 CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
729 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
730 CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
731 va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
732 CHECK_VASTATUS(va_status, "vaEndPicture");
734 qf.resources = move(resources);
735 qf.resource_releaser = move(release);
737 lock_guard<mutex> lock(mu);
738 frames_encoding.push(move(qf));
739 any_frames_encoding.notify_all();
742 void MJPEGEncoder::va_receiver_thread_func()
744 pthread_setname_np(pthread_self(), "MJPEG_Receive");
748 unique_lock<mutex> lock(mu);
749 any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
750 if (should_quit) return;
751 qf = move(frames_encoding.front());
752 frames_encoding.pop();
755 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
756 CHECK_VASTATUS(va_status, "vaSyncSurface");
758 VACodedBufferSegment *segment;
759 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
760 CHECK_VASTATUS(va_status, "vaMapBuffer");
762 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
763 write_mjpeg_packet(qf.pts, qf.card_index, coded_buf, segment->size);
765 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
766 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
770 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
772 unsigned width = qf.video_format.width;
773 unsigned height = qf.video_format.height;
775 VectorDestinationManager dest;
776 jpeg_compress_struct cinfo;
777 init_jpeg_422(width, height, &dest, &cinfo);
779 size_t field_start_line = qf.video_format.extra_lines_top; // No interlacing support.
780 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
782 JSAMPROW yptr[8], cbptr[8], crptr[8];
783 JSAMPARRAY data[3] = { yptr, cbptr, crptr };
784 for (unsigned y = 0; y < qf.video_format.height; y += 8) {
785 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
787 memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
788 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
789 for (unsigned yy = 0; yy < 8; ++yy) {
790 yptr[yy] = tmp_y + yy * width;
791 cbptr[yy] = tmp_cb + yy * width / 2;
792 crptr[yy] = tmp_cr + yy * width / 2;
794 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
796 jpeg_finish_compress(&cinfo);