]> git.sesse.net Git - nageru/blob - nageru/mjpeg_encoder.cpp
Set a FFmpeg-private JPEG marker to signal some Y'CbCr information.
[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 "pbo_frame_allocator.h"
20 #include "shared/timebase.h"
21 #include "va_display_with_cleanup.h"
22
23 #include <va/va.h>
24 #include <va/va_drm.h>
25 #include <va/va_x11.h>
26
27 using namespace bmusb;
28 using namespace std;
29
30 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
31
32 #define CHECK_VASTATUS(va_status, func)                                 \
33     if (va_status != VA_STATUS_SUCCESS) {                               \
34         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
35         exit(1);                                                        \
36     }
37
38 // From libjpeg (although it's of course identical between implementations).
39 static const int jpeg_natural_order[DCTSIZE2] = {
40          0,  1,  8, 16,  9,  2,  3, 10,
41         17, 24, 32, 25, 18, 11,  4,  5,
42         12, 19, 26, 33, 40, 48, 41, 34,
43         27, 20, 13,  6,  7, 14, 21, 28,
44         35, 42, 49, 56, 57, 50, 43, 36,
45         29, 22, 15, 23, 30, 37, 44, 51,
46         58, 59, 52, 45, 38, 31, 39, 46,
47         53, 60, 61, 54, 47, 55, 62, 63,
48 };
49
50 struct VectorDestinationManager {
51         jpeg_destination_mgr pub;
52         std::vector<uint8_t> dest;
53
54         VectorDestinationManager()
55         {
56                 pub.init_destination = init_destination_thunk;
57                 pub.empty_output_buffer = empty_output_buffer_thunk;
58                 pub.term_destination = term_destination_thunk;
59         }
60
61         static void init_destination_thunk(j_compress_ptr ptr)
62         {
63                 ((VectorDestinationManager *)(ptr->dest))->init_destination();
64         }
65
66         inline void init_destination()
67         {
68                 make_room(0);
69         }
70
71         static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
72         {
73                 return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
74         }
75
76         inline bool empty_output_buffer()
77         {
78                 make_room(dest.size());  // Should ignore pub.free_in_buffer!
79                 return true;
80         }
81
82         inline void make_room(size_t bytes_used)
83         {
84                 dest.resize(bytes_used + 4096);
85                 dest.resize(dest.capacity());
86                 pub.next_output_byte = dest.data() + bytes_used;
87                 pub.free_in_buffer = dest.size() - bytes_used;
88         }
89
90         static void term_destination_thunk(j_compress_ptr ptr)
91         {
92                 ((VectorDestinationManager *)(ptr->dest))->term_destination();
93         }
94
95         inline void term_destination()
96         {
97                 dest.resize(dest.size() - pub.free_in_buffer);
98         }
99 };
100 static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
101
102 int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
103 {
104         MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
105         return engine->write_packet2(buf, buf_size, type, time);
106 }
107
108 int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
109 {
110         if (type == AVIO_DATA_MARKER_HEADER) {
111                 mux_header.append((char *)buf, buf_size);
112                 httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
113         } else {
114                 httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
115         }
116         return buf_size;
117 }
118
119 MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
120         : httpd(httpd)
121 {
122         encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
123
124         // Set up the mux. We don't use the Mux wrapper, because it's geared towards
125         // a situation with only one video stream (and possibly one audio stream)
126         // with known width/height, and we don't need the extra functionality it provides.
127         avctx.reset(avformat_alloc_context());
128         avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
129
130         uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
131         avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
132         avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
133         avctx->flags = AVFMT_FLAG_CUSTOM_IO;
134
135         for (int card_idx = 0; card_idx < global_flags.num_cards; ++card_idx) {
136                 AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
137                 if (stream == nullptr) {
138                         fprintf(stderr, "avformat_new_stream() failed\n");
139                         exit(1);
140                 }
141                 stream->time_base = AVRational{ 1, TIMEBASE };
142                 stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
143                 stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
144
145                 // Used for aspect ratio only. Can change without notice (the mux won't care).
146                 stream->codecpar->width = global_flags.width;
147                 stream->codecpar->height = global_flags.height;
148
149                 // TODO: We could perhaps use the interpretation for each card here
150                 // (or at least the command-line flags) instead of the defaults,
151                 // but what would we do when they change?
152                 stream->codecpar->color_primaries = AVCOL_PRI_BT709;
153                 stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
154                 stream->codecpar->color_space = AVCOL_SPC_BT709;
155                 stream->codecpar->color_range = AVCOL_RANGE_MPEG;
156                 stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
157                 stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
158         }
159
160         AVDictionary *options = NULL;
161         vector<pair<string, string>> opts = MUX_OPTS;
162         for (pair<string, string> opt : opts) {
163                 av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
164         }
165         if (avformat_write_header(avctx.get(), &options) < 0) {
166                 fprintf(stderr, "avformat_write_header() failed\n");
167                 exit(1);
168         }
169
170         // Initialize VA-API.
171         string error;
172         va_dpy = try_open_va(va_display, &error, &config_id);
173         if (va_dpy == nullptr) {
174                 fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
175         }
176
177         running = true;
178 }
179
180 void MJPEGEncoder::stop()
181 {
182         if (!running) {
183                 return;
184         }
185         running = false;
186         should_quit = true;
187         any_frames_to_be_encoded.notify_all();
188         encoder_thread.join();
189 }
190
191 unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
192 {
193         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
194         if (va_dpy == nullptr) {
195                 if (error) *error = "Opening VA display failed";
196                 return nullptr;
197         }
198         int major_ver, minor_ver;
199         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
200         if (va_status != VA_STATUS_SUCCESS) {
201                 char buf[256];
202                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
203                 if (error != nullptr) *error = buf;
204                 return nullptr;
205         }
206
207         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
208         va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
209                 &attr, 1, config_id);
210         if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
211                 if (error != nullptr) *error = "No hardware support";
212                 return nullptr;
213         } else if (va_status != VA_STATUS_SUCCESS) {
214                 char buf[256];
215                 snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
216                 if (error != nullptr) *error = buf;
217                 return nullptr;
218         }
219
220         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
221         assert(num_formats > 0);
222
223         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
224         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
225         if (va_status != VA_STATUS_SUCCESS) {
226                 char buf[256];
227                 snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
228                 if (error != nullptr) *error = buf;
229                 return nullptr;
230         }
231
232         return va_dpy;
233 }
234
235 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)
236 {
237         PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
238         if (video_format.width == 0 || video_format.height == 0) {
239                 return;
240         }
241         if (video_format.interlaced) {
242                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
243                 return;
244         }
245         if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
246             !frame->interleaved) {
247                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
248                 return;
249         }
250         if (video_format.width > 4096 || video_format.height > 4096) {
251                 fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
252                 return;
253         }
254
255         lock_guard<mutex> lock(mu);
256         frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
257         any_frames_to_be_encoded.notify_all();
258 }
259
260 void MJPEGEncoder::encoder_thread_func()
261 {
262         pthread_setname_np(pthread_self(), "MJPEG_Encode");
263         posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
264         posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
265         posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
266         posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
267
268         unique_lock<mutex> lock(mu);
269         for (;;) {
270                 any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
271                 if (should_quit) return;
272                 QueuedFrame qf = move(frames_to_be_encoded.front());
273                 frames_to_be_encoded.pop();
274
275                 vector<uint8_t> jpeg = encode_jpeg(qf);
276
277                 AVPacket pkt;
278                 memset(&pkt, 0, sizeof(pkt));
279                 pkt.buf = nullptr;
280                 pkt.data = &jpeg[0];
281                 pkt.size = jpeg.size();
282                 pkt.stream_index = qf.card_index;
283                 pkt.flags = AV_PKT_FLAG_KEY;
284                 pkt.pts = pkt.dts = qf.pts;
285
286                 if (av_write_frame(avctx.get(), &pkt) < 0) {
287                         fprintf(stderr, "av_write_frame() failed\n");
288                         exit(1);
289                 }
290         }
291 }
292
293 class VABufferDestroyer {
294 public:
295         VABufferDestroyer(VADisplay dpy, VABufferID buf)
296                 : dpy(dpy), buf(buf) {}
297
298         ~VABufferDestroyer() {
299                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
300                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
301         }
302
303 private:
304         VADisplay dpy;
305         VABufferID buf;
306 };
307
308 MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
309 {
310         {
311                 lock_guard<mutex> lock(va_resources_mutex);
312                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
313                         if (it->width == width && it->height == height) {
314                                 VAResources ret = *it;
315                                 va_resources_freelist.erase(it);
316                                 return ret;
317                         }
318                 }
319         }
320
321         VAResources ret;
322
323         ret.width = width;
324         ret.height = height;
325
326         VASurfaceAttrib attrib;
327         attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
328         attrib.type = VASurfaceAttribPixelFormat;
329         attrib.value.type = VAGenericValueTypeInteger;
330         attrib.value.value.i = VA_FOURCC_UYVY;
331
332         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
333                 width, height,
334                 &ret.surface, 1, &attrib, 1);
335         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
336
337         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
338         CHECK_VASTATUS(va_status, "vaCreateContext");
339
340         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
341         CHECK_VASTATUS(va_status, "vaCreateBuffer");
342
343         return ret;
344 }
345
346 void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
347 {
348         lock_guard<mutex> lock(va_resources_mutex);
349         if (va_resources_freelist.size() > 10) {
350                 auto it = va_resources_freelist.end();
351                 --it;
352
353                 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
354                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
355
356                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
357                 CHECK_VASTATUS(va_status, "vaDestroyContext");
358
359                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
360                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
361
362                 va_resources_freelist.erase(it);
363         }
364
365         va_resources_freelist.push_front(resources);
366 }
367
368 void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
369 {
370         jpeg_error_mgr jerr;
371         cinfo->err = jpeg_std_error(&jerr);
372         jpeg_create_compress(cinfo);
373
374         cinfo->dest = (jpeg_destination_mgr *)dest;
375
376         cinfo->input_components = 3;
377         jpeg_set_defaults(cinfo);
378         jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
379
380         cinfo->image_width = width;
381         cinfo->image_height = height;
382         cinfo->raw_data_in = true;
383         jpeg_set_colorspace(cinfo, JCS_YCbCr);
384         cinfo->comp_info[0].h_samp_factor = 2;
385         cinfo->comp_info[0].v_samp_factor = 1;
386         cinfo->comp_info[1].h_samp_factor = 1;
387         cinfo->comp_info[1].v_samp_factor = 1;
388         cinfo->comp_info[2].h_samp_factor = 1;
389         cinfo->comp_info[2].v_samp_factor = 1;
390         cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
391         jpeg_start_compress(cinfo, true);
392
393         // This comment marker is private to FFmpeg. It signals limited Y'CbCr range
394         // (and nothing else).
395         jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
396 }
397
398 vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
399 {
400         VectorDestinationManager dest;
401         init_jpeg_422(width, height, &dest, cinfo);
402
403         // Make a dummy black image; there's seemingly no other easy way of
404         // making libjpeg outputting all of its headers.
405         JSAMPROW yptr[8], cbptr[8], crptr[8];
406         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
407         memset(tmp_y, 0, 4096);
408         memset(tmp_cb, 0, 4096);
409         memset(tmp_cr, 0, 4096);
410         for (unsigned yy = 0; yy < 8; ++yy) {
411                 yptr[yy] = tmp_y;
412                 cbptr[yy] = tmp_cb;
413                 crptr[yy] = tmp_cr;
414         }
415         for (unsigned y = 0; y < height; y += 8) {
416                 jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
417         }
418         jpeg_finish_compress(cinfo);
419
420         // We're only interested in the header, not the data after it.
421         dest.term_destination();
422         for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
423                 if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
424                         unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
425                         dest.dest.resize(i + len + 2);
426                         break;
427                 }
428         }
429
430         return dest.dest;
431 }
432
433 MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
434 {
435         pair<unsigned, unsigned> key(width, height);
436         if (va_data_for_resolution.count(key)) {
437                 return va_data_for_resolution[key];
438         }
439
440         // Use libjpeg to generate a header and set sane defaults for e.g.
441         // quantization tables. Then do the actual encode with VA-API.
442         jpeg_compress_struct cinfo;
443         vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
444
445         // Picture parameters.
446         VAEncPictureParameterBufferJPEG pic_param;
447         memset(&pic_param, 0, sizeof(pic_param));
448         pic_param.reconstructed_picture = VA_INVALID_ID;
449         pic_param.picture_width = cinfo.image_width;
450         pic_param.picture_height = cinfo.image_height;
451         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
452                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
453                 pic_param.component_id[component_idx] = comp->component_id;
454                 pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
455         }
456         pic_param.num_components = cinfo.num_components;
457         pic_param.num_scan = 1;
458         pic_param.sample_bit_depth = 8;
459         pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
460         pic_param.pic_flags.bits.huffman = 1;
461         pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
462
463         // Quantization matrices.
464         VAQMatrixBufferJPEG q;
465         memset(&q, 0, sizeof(q));
466
467         q.load_lum_quantiser_matrix = true;
468         q.load_chroma_quantiser_matrix = true;
469         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
470                 const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
471                 assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
472                 if (qtbl == nullptr) continue;
473
474                 uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
475                 for (int i = 0; i < 64; ++i) {
476                         if (qtbl->quantval[i] > 255) {
477                                 fprintf(stderr, "Baseline JPEG only!\n");
478                                 abort();
479                         }
480                         qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
481                 }
482         }
483
484         // Huffman tables (arithmetic is not supported).
485         VAHuffmanTableBufferJPEGBaseline huff;
486         memset(&huff, 0, sizeof(huff));
487
488         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
489                 const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
490                 const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
491                 if (ac_hufftbl == nullptr) {
492                         assert(dc_hufftbl == nullptr);
493                         huff.load_huffman_table[huff_tbl_idx] = 0;
494                 } else {
495                         assert(dc_hufftbl != nullptr);
496                         huff.load_huffman_table[huff_tbl_idx] = 1;
497
498                         for (int i = 0; i < 16; ++i) {
499                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
500                         }
501                         for (int i = 0; i < 12; ++i) {
502                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
503                         }
504                         for (int i = 0; i < 16; ++i) {
505                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
506                         }
507                         for (int i = 0; i < 162; ++i) {
508                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
509                         }
510                 }
511         }
512
513         // Slice parameters (metadata about the slice).
514         VAEncSliceParameterBufferJPEG parms;
515         memset(&parms, 0, sizeof(parms));
516         for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
517                 const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
518                 parms.components[component_idx].component_selector = comp->component_id;
519                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
520                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
521                 if (parms.components[component_idx].dc_table_selector > 1 ||
522                     parms.components[component_idx].ac_table_selector > 1) {
523                         fprintf(stderr, "Uses too many Huffman tables\n");
524                         abort();
525                 }
526         }
527         parms.num_components = cinfo.num_components;
528         parms.restart_interval = cinfo.restart_interval;
529
530         jpeg_destroy_compress(&cinfo);
531
532         VAData ret;
533         ret.jpeg_header = move(jpeg_header);
534         ret.pic_param = pic_param;
535         ret.q = q;
536         ret.huff = huff;
537         ret.parms = parms;
538         va_data_for_resolution[key] = ret;
539         return ret;
540 }
541
542 vector<uint8_t> MJPEGEncoder::encode_jpeg(const QueuedFrame &qf)
543 {
544         if (va_dpy != nullptr) {
545                 return encode_jpeg_va(qf);
546         } else {
547                 return encode_jpeg_libjpeg(qf);
548         }
549 }
550
551 vector<uint8_t> MJPEGEncoder::encode_jpeg_va(const QueuedFrame &qf)
552 {
553         unsigned width = qf.video_format.width;
554         unsigned height = qf.video_format.height;
555
556         VAResources resources = get_va_resources(width, height);
557         ReleaseVAResources release(this, resources);
558
559         VAData va_data = get_va_data_for_resolution(width, height);
560         va_data.pic_param.coded_buf = resources.data_buffer;
561
562         VABufferID pic_param_buffer;
563         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
564         CHECK_VASTATUS(va_status, "vaCreateBuffer");
565         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
566
567         VABufferID q_buffer;
568         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
569         CHECK_VASTATUS(va_status, "vaCreateBuffer");
570         VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
571
572         VABufferID huff_buffer;
573         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
574         CHECK_VASTATUS(va_status, "vaCreateBuffer");
575         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
576
577         VABufferID slice_param_buffer;
578         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
579         CHECK_VASTATUS(va_status, "vaCreateBuffer");
580         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
581
582         VAImage image;
583         va_status = vaDeriveImage(va_dpy->va_dpy, resources.surface, &image);
584         CHECK_VASTATUS(va_status, "vaDeriveImage");
585
586         // Upload the pixel data.
587         uint8_t *surface_p = nullptr;
588         vaMapBuffer(va_dpy->va_dpy, image.buf, (void **)&surface_p);
589
590         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
591         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
592
593         {
594                 const uint8_t *src = qf.frame->data_copy + field_start;
595                 uint8_t *dst = (unsigned char *)surface_p + image.offsets[0];
596                 memcpy_with_pitch(dst, src, qf.video_format.width * 2, image.pitches[0], qf.video_format.height);
597         }
598
599         va_status = vaUnmapBuffer(va_dpy->va_dpy, image.buf);
600         CHECK_VASTATUS(va_status, "vaUnmapBuffer");
601         va_status = vaDestroyImage(va_dpy->va_dpy, image.image_id);
602         CHECK_VASTATUS(va_status, "vaDestroyImage");
603
604         // Finally, stick in the JPEG header.
605         VAEncPackedHeaderParameterBuffer header_parm;
606         header_parm.type = VAEncPackedHeaderRawData;
607         header_parm.bit_length = 8 * va_data.jpeg_header.size();
608
609         VABufferID header_parm_buffer;
610         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
611         CHECK_VASTATUS(va_status, "vaCreateBuffer");
612         VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
613
614         VABufferID header_data_buffer;
615         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
616         CHECK_VASTATUS(va_status, "vaCreateBuffer");
617         VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
618
619         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
620         CHECK_VASTATUS(va_status, "vaBeginPicture");
621         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
622         CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
623         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
624         CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
625         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
626         CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
627         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
628         CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
629         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
630         CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
631         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
632         CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
633         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
634         CHECK_VASTATUS(va_status, "vaEndPicture");
635
636         va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
637         CHECK_VASTATUS(va_status, "vaSyncSurface");
638
639         VACodedBufferSegment *segment;
640         va_status = vaMapBuffer(va_dpy->va_dpy, resources.data_buffer, (void **)&segment);
641         CHECK_VASTATUS(va_status, "vaMapBuffer");
642
643         const char *coded_buf = reinterpret_cast<char *>(segment->buf);
644         vector<uint8_t> jpeg(coded_buf, coded_buf + segment->size); 
645
646         va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.data_buffer);
647         CHECK_VASTATUS(va_status, "vaUnmapBuffer");
648
649         return jpeg;
650 }
651
652 vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
653 {
654         unsigned width = qf.video_format.width;
655         unsigned height = qf.video_format.height;
656
657         VectorDestinationManager dest;
658         jpeg_compress_struct cinfo;
659         init_jpeg_422(width, height, &dest, &cinfo);
660
661         size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
662         size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
663
664         JSAMPROW yptr[8], cbptr[8], crptr[8];
665         JSAMPARRAY data[3] = { yptr, cbptr, crptr };
666         for (unsigned y = 0; y < qf.video_format.height; y += 8) {
667                 const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
668
669                 memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
670                 memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
671                 for (unsigned yy = 0; yy < 8; ++yy) {
672                         yptr[yy] = tmp_y + yy * width;
673                         cbptr[yy] = tmp_cb + yy * width / 2;
674                         crptr[yy] = tmp_cr + yy * width / 2;
675                 }
676                 jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
677         }
678         jpeg_finish_compress(&cinfo);
679
680         return dest.dest;
681 }