]> git.sesse.net Git - nageru/blob - nageru/mjpeg_encoder.cpp
Various fixes for non-VA-API MJPEG encoding.
[nageru] / nageru / mjpeg_encoder.cpp
1 #include "mjpeg_encoder.h"
2
3 #include <jpeglib.h>
4 #include <unistd.h>
5 #if __SSE2__
6 #include <immintrin.h>
7 #endif
8 #include <list>
9
10 extern "C" {
11 #include <libavformat/avformat.h>
12 }
13
14 #include "defs.h"
15 #include "shared/ffmpeg_raii.h"
16 #include "flags.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"
23
24 #include <va/va.h>
25 #include <va/va_drm.h>
26 #include <va/va_x11.h>
27
28 using namespace bmusb;
29 using namespace std;
30
31 static VAImageFormat uyvy_format;
32
33 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
34
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,
45 };
46
47 struct VectorDestinationManager {
48         jpeg_destination_mgr pub;
49         std::vector<uint8_t> dest;
50
51         VectorDestinationManager()
52         {
53                 pub.init_destination = init_destination_thunk;
54                 pub.empty_output_buffer = empty_output_buffer_thunk;
55                 pub.term_destination = term_destination_thunk;
56         }
57
58         static void init_destination_thunk(j_compress_ptr ptr)
59         {
60                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
61         }
62
63         inline void init_destination()
64         {
65                 make_room(0);
66         }
67
68         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
69         {
70                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
71         }
72
73         inline bool empty_output_buffer()
74         {
75                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
76                 return true;
77         }
78
79         inline void make_room(size_t bytes_used)
80         {
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;
85         }
86
87         static void term_destination_thunk(j_compress_ptr ptr)
88         {
89                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
90         }
91
92         inline void term_destination()
93         {
94                 dest.resize(dest.size() - pub.free_in_buffer);
95         }
96 };
97 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
98
99 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
100 {
101         MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
102         return engine->write_packet2(buf, buf_size, type, time);
103 }
104
105 int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
106 {
107         if (type == AVIO_DATA_MARKER_HEADER) {
108                 mux_header.append((char *)buf, buf_size);
109                 httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
110         } else {
111                 httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
112         }
113         return buf_size;
114 }
115
116 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
117         : httpd(httpd)
118 {
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("nut", nullptr, nullptr);
124
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;
129
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");
134                         abort();
135                 }
136
137                 // FFmpeg is very picky about having audio at 1/48000 timebase,
138                 // no matter what we write. Even though we'd prefer our usual 1/120000,
139                 // put the video on the same one, so that we can have locked audio.
140                 stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
141                 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
142                 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
143
144                 // Used for aspect ratio only. Can change without notice (the mux won't care).
145                 stream->codecpar->width = global_flags.width;
146                 stream->codecpar->height = global_flags.height;
147
148                 // TODO: We could perhaps use the interpretation for each card here
149                 // (or at least the command-line flags) instead of the defaults,
150                 // but what would we do when they change?
151                 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
152                 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
153                 stream->codecpar->color_space = AVCOL_SPC_BT709;
154                 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
155                 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
156                 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
157         }
158         for (unsigned card_idx = 0; card_idx < global_flags.card_to_mjpeg_stream_export.size(); ++card_idx) {
159                 AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
160                 if (stream == nullptr) {
161                         fprintf(stderr, "avformat_new_stream() failed\n");
162                         abort();
163                 }
164                 stream->time_base = AVRational{ 1, OUTPUT_FREQUENCY };
165                 stream->codecpar->codec_type = AVMEDIA_TYPE_AUDIO;
166                 stream->codecpar->codec_id = AV_CODEC_ID_PCM_S32LE;
167                 stream->codecpar->channel_layout = AV_CH_LAYOUT_STEREO;
168                 stream->codecpar->channels = 2;
169                 stream->codecpar->sample_rate = OUTPUT_FREQUENCY;
170         }
171
172         AVDictionary *options = NULL;
173         vector<pair<string, string>> opts = MUX_OPTS;
174         for (pair<string, string> opt : opts) {
175                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
176         }
177         if (avformat_write_header(avctx.get(), &options) < 0) {
178                 fprintf(stderr, "avformat_write_header() failed\n");
179                 abort();
180         }
181
182         // Initialize VA-API.
183         string error;
184         va_dpy = try_open_va(va_display, &error, &config_id);
185         if (va_dpy == nullptr) {
186                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
187         }
188
189         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
190         if (va_dpy != nullptr) {
191                 va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
192         }
193
194         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }}, &metric_mjpeg_frames_zero_size_dropped);
195         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }}, &metric_mjpeg_frames_interlaced_dropped);
196         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }}, &metric_mjpeg_frames_unsupported_pixel_format_dropped);
197         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }}, &metric_mjpeg_frames_oversized_dropped);
198         global_metrics.add("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }}, &metric_mjpeg_overrun_dropped);
199         global_metrics.add("mjpeg_frames", {{ "status", "submitted" }}, &metric_mjpeg_overrun_submitted);
200
201         running = true;
202 }
203
204 MJPEGEncoder::~MJPEGEncoder()
205 {
206         av_free(avctx->pb->buffer);
207
208         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "zero_size" }});
209         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "interlaced" }});
210         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "unsupported_pixel_format" }});
211         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "oversized" }});
212         global_metrics.remove("mjpeg_frames", {{ "status", "dropped" }, { "reason", "overrun" }});
213         global_metrics.remove("mjpeg_frames", {{ "status", "submitted" }});
214 }
215
216 void MJPEGEncoder::stop()
217 {
218         if (!running) {
219                 return;
220         }
221         running = false;
222         should_quit = true;
223         any_frames_to_be_encoded.notify_all();
224         any_frames_encoding.notify_all();
225         encoder_thread.join();
226         if (va_dpy != nullptr) {
227                 va_receiver_thread.join();
228         }
229 }
230
231 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
232 {
233         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
234         if (va_dpy == nullptr) {
235                 if (error) *error = "Opening VA display failed";
236                 return nullptr;
237         }
238         int major_ver, minor_ver;
239         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
240         if (va_status != VA_STATUS_SUCCESS) {
241                 char buf[256];
242                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
243                 if (error != nullptr) *error = buf;
244                 return nullptr;
245         }
246
247         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
248         va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
249                 &attr, 1, config_id);
250         if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
251                 if (error != nullptr) *error = "No hardware support";
252                 return nullptr;
253         } else if (va_status != VA_STATUS_SUCCESS) {
254                 char buf[256];
255                 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
256                 if (error != nullptr) *error = buf;
257                 return nullptr;
258         }
259
260         // TODO: Unify with the code in Futatabi.
261         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
262         assert(num_formats > 0);
263
264         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
265         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
266         if (va_status != VA_STATUS_SUCCESS) {
267                 char buf[256];
268                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
269                 if (error != nullptr) *error = buf;
270                 return nullptr;
271         }
272
273         bool found = false;
274         for (int i = 0; i < num_formats; ++i) {
275                 if (formats[i].fourcc == VA_FOURCC_UYVY) {
276                         memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
277                         found = true;
278                         break;
279                 }
280         }
281         if (!found) {
282                 if (error != nullptr) *error = "UYVY format not found";
283                 return nullptr;
284         }
285
286         return va_dpy;
287 }
288
289 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)
290 {
291         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
292         if (video_format.width == 0 || video_format.height == 0) {
293                 ++metric_mjpeg_frames_zero_size_dropped;
294                 return;
295         }
296         if (video_format.interlaced) {
297                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
298                 ++metric_mjpeg_frames_interlaced_dropped;
299                 return;
300         }
301         if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
302             !frame->interleaved) {
303                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
304                 ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
305                 return;
306         }
307         if (video_format.width > 4096 || video_format.height > 4096) {
308                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
309                 ++metric_mjpeg_frames_oversized_dropped;
310                 return;
311         }
312
313         lock_guard<mutex> lock(mu);
314         if (frames_to_be_encoded.size() + frames_encoding.size() > 50) {
315                 fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n");
316                 ++metric_mjpeg_overrun_dropped;
317                 return;
318         }
319         ++metric_mjpeg_overrun_submitted;
320         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset, move(audio) });
321         any_frames_to_be_encoded.notify_all();
322 }
323
324 int MJPEGEncoder::get_mjpeg_stream_for_card(unsigned card_index)
325 {
326         // Only bother doing MJPEG encoding if there are any connected clients
327         // that want the stream.
328         if (httpd->get_num_connected_multicam_clients() == 0) {
329                 return -1;
330         }
331
332         auto it = global_flags.card_to_mjpeg_stream_export.find(card_index);
333         if (it == global_flags.card_to_mjpeg_stream_export.end()) {
334                 return -1;
335         }
336         return it->second;
337 }
338
339 void MJPEGEncoder::encoder_thread_func()
340 {
341         pthread_setname_np(pthread_self(), "MJPEG_Encode");
342         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
343         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
344         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
345         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
346
347         for (;;) {
348                 QueuedFrame qf;
349                 {
350                         unique_lock<mutex> lock(mu);
351                         any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
352                         if (should_quit) break;
353                         qf = move(frames_to_be_encoded.front());
354                         frames_to_be_encoded.pop();
355                 }
356
357                 if (va_dpy != nullptr) {
358                         // Will call back in the receiver thread.
359                         encode_jpeg_va(move(qf));
360                 } else {
361                         // Write audio before video, since Futatabi expects it.
362                         if (qf.audio.size() > 0) {
363                                 write_audio_packet(qf.pts, qf.card_index, qf.audio);
364                         }
365
366                         // Encode synchronously, in the same thread.
367                         vector<uint8_t> jpeg = encode_jpeg_libjpeg(qf);
368                         write_mjpeg_packet(qf.pts, qf.card_index, jpeg.data(), jpeg.size());
369                 }
370         }
371
372         free(tmp_y);
373         free(tmp_cbcr);
374         free(tmp_cb);
375         free(tmp_cr);
376 }
377
378 void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const uint8_t *jpeg, size_t jpeg_size)
379 {
380         AVPacket pkt;
381         memset(&pkt, 0, sizeof(pkt));
382         pkt.buf = nullptr;
383         pkt.data = const_cast<uint8_t *>(jpeg);
384         pkt.size = jpeg_size;
385         pkt.stream_index = card_index;
386         pkt.flags = AV_PKT_FLAG_KEY;
387         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
388         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
389         pkt.duration = 0;
390
391         if (av_write_frame(avctx.get(), &pkt) < 0) {
392                 fprintf(stderr, "av_write_frame() failed\n");
393                 abort();
394         }
395 }
396
397 void MJPEGEncoder::write_audio_packet(int64_t pts, unsigned card_index, const vector<int32_t> &audio)
398 {
399         AVPacket pkt;
400         memset(&pkt, 0, sizeof(pkt));
401         pkt.buf = nullptr;
402         pkt.data = reinterpret_cast<uint8_t *>(const_cast<int32_t *>(&audio[0]));
403         pkt.size = audio.size() * sizeof(audio[0]);
404         pkt.stream_index = card_index + global_flags.card_to_mjpeg_stream_export.size();
405         pkt.flags = AV_PKT_FLAG_KEY;
406         AVRational time_base = avctx->streams[pkt.stream_index]->time_base;
407         pkt.pts = pkt.dts = av_rescale_q(pts, AVRational{ 1, TIMEBASE }, time_base);
408         size_t num_stereo_samples = audio.size() / 2;
409         pkt.duration = av_rescale_q(num_stereo_samples, AVRational{ 1, OUTPUT_FREQUENCY }, time_base);
410
411         if (av_write_frame(avctx.get(), &pkt) < 0) {
412                 fprintf(stderr, "av_write_frame() failed\n");
413                 abort();
414         }
415 }
416
417 class VABufferDestroyer {
418 public:
419         VABufferDestroyer(VADisplay dpy, VABufferID buf)
420                 : dpy(dpy), buf(buf) {}
421
422         ~VABufferDestroyer() {
423                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
424                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
425         }
426
427 private:
428         VADisplay dpy;
429         VABufferID buf;
430 };
431
432 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
433 {
434         {
435                 lock_guard<mutex> lock(va_resources_mutex);
436                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
437                         if (it->width == width && it->height == height) {
438                                 VAResources ret = *it;
439                                 va_resources_freelist.erase(it);
440                                 return ret;
441                         }
442                 }
443         }
444
445         VAResources ret;
446
447         ret.width = width;
448         ret.height = height;
449
450         VASurfaceAttrib attrib;
451         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
452         attrib.type = VASurfaceAttribPixelFormat;
453         attrib.value.type = VAGenericValueTypeInteger;
454         attrib.value.value.i = VA_FOURCC_UYVY;
455
456         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
457                 width, height,
458                 &ret.surface, 1, &attrib, 1);
459         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
460
461         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
462         CHECK_VASTATUS(va_status, "vaCreateContext");
463
464         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
465         CHECK_VASTATUS(va_status, "vaCreateBuffer");
466
467         va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
468         CHECK_VASTATUS(va_status, "vaCreateImage");
469
470         return ret;
471 }
472
473 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
474 {
475         lock_guard<mutex> lock(va_resources_mutex);
476         if (va_resources_freelist.size() > 50) {
477                 auto it = va_resources_freelist.end();
478                 --it;
479
480                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
481                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
482
483                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
484                 CHECK_VASTATUS(va_status, "vaDestroyContext");
485
486                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
487                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
488
489                 va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
490                 CHECK_VASTATUS(va_status, "vaDestroyImage");
491
492                 va_resources_freelist.erase(it);
493         }
494
495         va_resources_freelist.push_front(resources);
496 }
497
498 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
499 {
500         jpeg_error_mgr jerr;
501         cinfo->err = jpeg_std_error(&jerr);
502         jpeg_create_compress(cinfo);
503
504         cinfo->dest = (jpeg_destination_mgr *)dest;
505
506         cinfo->input_components = 3;
507         jpeg_set_defaults(cinfo);
508         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
509
510         cinfo->image_width = width;
511         cinfo->image_height = height;
512         cinfo->raw_data_in = true;
513         jpeg_set_colorspace(cinfo, JCS_YCbCr);
514         cinfo->comp_info[0].h_samp_factor = 2;
515         cinfo->comp_info[0].v_samp_factor = 1;
516         cinfo->comp_info[1].h_samp_factor = 1;
517         cinfo->comp_info[1].v_samp_factor = 1;
518         cinfo->comp_info[2].h_samp_factor = 1;
519         cinfo->comp_info[2].v_samp_factor = 1;
520         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
521         jpeg_start_compress(cinfo, true);
522
523         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
524         // (and nothing else).
525         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
526 }
527
528 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
529 {
530         VectorDestinationManager dest;
531         init_jpeg_422(width, height, &dest, cinfo);
532
533         // Make a dummy black image; there's seemingly no other easy way of
534         // making libjpeg outputting all of its headers.
535         JSAMPROW yptr[8], cbptr[8], crptr[8];
536         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
537         memset(tmp_y, 0, 4096);
538         memset(tmp_cb, 0, 4096);
539         memset(tmp_cr, 0, 4096);
540         for (unsigned yy = 0; yy < 8; ++yy) {
541                 yptr[yy] = tmp_y;
542                 cbptr[yy] = tmp_cb;
543                 crptr[yy] = tmp_cr;
544         }
545         for (unsigned y = 0; y < height; y += 8) {
546                 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
547         }
548         jpeg_finish_compress(cinfo);
549
550         // We're only interested in the header, not the data after it.
551         dest.term_destination();
552         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
553                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
554                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
555                         dest.dest.resize(i + len + 2);
556                         break;
557                 }
558         }
559
560         return dest.dest;
561 }
562
563 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
564 {
565         pair<unsigned, unsigned> key(width, height);
566         if (va_data_for_resolution.count(key)) {
567                 return va_data_for_resolution[key];
568         }
569
570         // Use libjpeg to generate a header and set sane defaults for e.g.
571         // quantization tables. Then do the actual encode with VA-API.
572         jpeg_compress_struct cinfo;
573         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
574
575         // Picture parameters.
576         VAEncPictureParameterBufferJPEG pic_param;
577         memset(&pic_param, 0, sizeof(pic_param));
578         pic_param.reconstructed_picture = VA_INVALID_ID;
579         pic_param.picture_width = cinfo.image_width;
580         pic_param.picture_height = cinfo.image_height;
581         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
582                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
583                 pic_param.component_id[component_idx] = comp->component_id;
584                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
585         }
586         pic_param.num_components = cinfo.num_components;
587         pic_param.num_scan = 1;
588         pic_param.sample_bit_depth = 8;
589         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
590         pic_param.pic_flags.bits.huffman = 1;
591         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
592
593         // Quantization matrices.
594         VAQMatrixBufferJPEG q;
595         memset(&q, 0, sizeof(q));
596
597         q.load_lum_quantiser_matrix = true;
598         q.load_chroma_quantiser_matrix = true;
599         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
600                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
601                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
602                 if (qtbl == nullptr) continue;
603
604                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
605                 for (int i = 0; i < 64; ++i) {
606                         if (qtbl->quantval[i] > 255) {
607                                 fprintf(stderr, "Baseline JPEG only!\n");
608                                 abort();
609                         }
610                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
611                 }
612         }
613
614         // Huffman tables (arithmetic is not supported).
615         VAHuffmanTableBufferJPEGBaseline huff;
616         memset(&huff, 0, sizeof(huff));
617
618         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
619                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
620                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
621                 if (ac_hufftbl == nullptr) {
622                         assert(dc_hufftbl == nullptr);
623                         huff.load_huffman_table[huff_tbl_idx] = 0;
624                 } else {
625                         assert(dc_hufftbl != nullptr);
626                         huff.load_huffman_table[huff_tbl_idx] = 1;
627
628                         for (int i = 0; i < 16; ++i) {
629                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
630                         }
631                         for (int i = 0; i < 12; ++i) {
632                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
633                         }
634                         for (int i = 0; i < 16; ++i) {
635                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
636                         }
637                         for (int i = 0; i < 162; ++i) {
638                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
639                         }
640                 }
641         }
642
643         // Slice parameters (metadata about the slice).
644         VAEncSliceParameterBufferJPEG parms;
645         memset(&parms, 0, sizeof(parms));
646         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
647                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
648                 parms.components[component_idx].component_selector = comp->component_id;
649                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
650                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
651                 if (parms.components[component_idx].dc_table_selector > 1 ||
652                     parms.components[component_idx].ac_table_selector > 1) {
653                         fprintf(stderr, "Uses too many Huffman tables\n");
654                         abort();
655                 }
656         }
657         parms.num_components = cinfo.num_components;
658         parms.restart_interval = cinfo.restart_interval;
659
660         jpeg_destroy_compress(&cinfo);
661
662         VAData ret;
663         ret.jpeg_header = move(jpeg_header);
664         ret.pic_param = pic_param;
665         ret.q = q;
666         ret.huff = huff;
667         ret.parms = parms;
668         va_data_for_resolution[key] = ret;
669         return ret;
670 }
671
672 void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
673 {
674         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
675         unsigned width = qf.video_format.width;
676         unsigned height = qf.video_format.height;
677
678         VAResources resources;
679         ReleaseVAResources release;
680         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
681                 resources = move(userdata->va_resources);
682                 release = move(userdata->va_resources_release);
683         } else {
684                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
685                 resources = get_va_resources(width, height);
686                 release = ReleaseVAResources(this, resources);
687         }
688
689         VAData va_data = get_va_data_for_resolution(width, height);
690         va_data.pic_param.coded_buf = resources.data_buffer;
691
692         VABufferID pic_param_buffer;
693         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
694         CHECK_VASTATUS(va_status, "vaCreateBuffer");
695         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
696
697         VABufferID q_buffer;
698         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
699         CHECK_VASTATUS(va_status, "vaCreateBuffer");
700         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
701
702         VABufferID huff_buffer;
703         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
704         CHECK_VASTATUS(va_status, "vaCreateBuffer");
705         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
706
707         VABufferID slice_param_buffer;
708         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
709         CHECK_VASTATUS(va_status, "vaCreateBuffer");
710         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
711
712         if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
713                 // The pixel data is already put into the image by the caller.
714                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
715                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
716         } else {
717                 assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
718
719                 // Upload the pixel data.
720                 uint8_t *surface_p = nullptr;
721                 vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
722
723                 size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
724                 size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
725
726                 {
727                         const uint8_t *src = qf.frame->data_copy + field_start;
728                         uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
729                         memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
730                 }
731
732                 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
733                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
734         }
735
736         qf.frame->data_copy = nullptr;
737
738         // Seemingly vaPutImage() (which triggers a GPU copy) is much nicer to the
739         // CPU than vaDeriveImage() and copying directly into the GPU's buffers.
740         // Exactly why is unclear, but it seems to involve L3 cache usage when there
741         // are many high-res (1080p+) images in play.
742         va_status = vaPutImage(va_dpy->va_dpy, resources.surface, resources.image.image_id, 0, 0, width, height, 0, 0, width, height);
743         CHECK_VASTATUS(va_status, "vaPutImage");
744
745         // Finally, stick in the JPEG header.
746         VAEncPackedHeaderParameterBuffer header_parm;
747         header_parm.type = VAEncPackedHeaderRawData;
748         header_parm.bit_length = 8 * va_data.jpeg_header.size();
749
750         VABufferID header_parm_buffer;
751         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
752         CHECK_VASTATUS(va_status, "vaCreateBuffer");
753         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
754
755         VABufferID header_data_buffer;
756         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
757         CHECK_VASTATUS(va_status, "vaCreateBuffer");
758         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
759
760         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
761         CHECK_VASTATUS(va_status, "vaBeginPicture");
762         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
763         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
764         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
765         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
766         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
767         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
768         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
769         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
770         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
771         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
772         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
773         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
774         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
775         CHECK_VASTATUS(va_status, "vaEndPicture");
776
777         qf.resources = move(resources);
778         qf.resource_releaser = move(release);
779
780         lock_guard<mutex> lock(mu);
781         frames_encoding.push(move(qf));
782         any_frames_encoding.notify_all();
783 }
784
785 void MJPEGEncoder::va_receiver_thread_func()
786 {
787         pthread_setname_np(pthread_self(), "MJPEG_Receive");
788         for (;;) {
789                 QueuedFrame qf;
790                 {
791                         unique_lock<mutex> lock(mu);
792                         any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; });
793                         if (should_quit) return;
794                         qf = move(frames_encoding.front());
795                         frames_encoding.pop();
796                 }
797
798                 // Write audio before video, since Futatabi expects it.
799                 if (qf.audio.size() > 0) {
800                         write_audio_packet(qf.pts, qf.card_index, qf.audio);
801                 }
802
803                 VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface);
804                 CHECK_VASTATUS(va_status, "vaSyncSurface");
805
806                 VACodedBufferSegment *segment;
807                 va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment);
808                 CHECK_VASTATUS(va_status, "vaMapBuffer");
809
810                 const uint8_t *coded_buf = reinterpret_cast<uint8_t *>(segment->buf);
811                 write_mjpeg_packet(qf.pts, qf.card_index, coded_buf, segment->size);
812
813                 va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer);
814                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
815         }
816 }
817
818 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
819 {
820         unsigned width = qf.video_format.width;
821         unsigned height = qf.video_format.height;
822
823         VectorDestinationManager dest;
824         jpeg_compress_struct cinfo;
825         init_jpeg_422(width, height, &dest, &cinfo);
826
827         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
828         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
829
830         JSAMPROW yptr[8], cbptr[8], crptr[8];
831         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
832         for (unsigned y = 0; y < qf.video_format.height; y += 8) {
833                 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
834
835                 memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
836                 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
837                 for (unsigned yy = 0; yy < 8; ++yy) {
838                         yptr[yy] = tmp_y + yy * width;
839                         cbptr[yy] = tmp_cb + yy * width / 2;
840                         crptr[yy] = tmp_cr + yy * width / 2;
841                 }
842                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
843         }
844         jpeg_finish_compress(&cinfo);
845
846         return dest.dest;
847 }