]> git.sesse.net Git - nageru/blob - quicksync_encoder.cpp
fa00a3409f336bfd383c5afde67850207853614b
[nageru] / quicksync_encoder.cpp
1 //#include "sysdeps.h"
2 #include "quicksync_encoder.h"
3
4 #include <movit/util.h>
5 #include <EGL/eglplatform.h>
6 #include <X11/X.h>
7 #include <X11/Xlib.h>
8 #include <assert.h>
9 #include <epoxy/egl.h>
10 #include <libdrm/drm_fourcc.h>
11 #include <stdio.h>
12 #include <stdlib.h>
13 #include <string.h>
14 #include <fcntl.h>
15 #include <va/va.h>
16 #include <va/va_drm.h>
17 #include <va/va_drmcommon.h>
18 #include <va/va_enc_h264.h>
19 #include <va/va_x11.h>
20 #include <algorithm>
21 #include <condition_variable>
22 #include <cstdint>
23 #include <map>
24 #include <memory>
25 #include <mutex>
26 #include <queue>
27 #include <string>
28 #include <thread>
29 #include <utility>
30
31 #include "audio_encoder.h"
32 #include "context.h"
33 #include "defs.h"
34 #include "flags.h"
35 #include "mux.h"
36 #include "timebase.h"
37 #include "x264_encoder.h"
38
39 using namespace std;
40
41 class QOpenGLContext;
42 class QSurface;
43
44 #define CHECK_VASTATUS(va_status, func)                                 \
45     if (va_status != VA_STATUS_SUCCESS) {                               \
46         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
47         exit(1);                                                        \
48     }
49
50 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
51
52 //#include "loadsurface.h"
53
54 #define NAL_REF_IDC_NONE        0
55 #define NAL_REF_IDC_LOW         1
56 #define NAL_REF_IDC_MEDIUM      2
57 #define NAL_REF_IDC_HIGH        3
58
59 #define NAL_NON_IDR             1
60 #define NAL_IDR                 5
61 #define NAL_SPS                 7
62 #define NAL_PPS                 8
63 #define NAL_SEI                 6
64
65 #define SLICE_TYPE_P            0
66 #define SLICE_TYPE_B            1
67 #define SLICE_TYPE_I            2
68 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
69 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
70 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
71
72
73 #define ENTROPY_MODE_CAVLC      0
74 #define ENTROPY_MODE_CABAC      1
75
76 #define PROFILE_IDC_BASELINE    66
77 #define PROFILE_IDC_MAIN        77
78 #define PROFILE_IDC_HIGH        100
79    
80 #define BITSTREAM_ALLOCATE_STEPPING     4096
81 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
82 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
83 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
84
85 static constexpr unsigned int MaxFrameNum = (2<<16);
86 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
87 static constexpr unsigned int Log2MaxFrameNum = 16;
88 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
89 static constexpr int rc_default_modes[] = {  // Priority list of modes.
90     VA_RC_VBR,
91     VA_RC_CQP,
92     VA_RC_VBR_CONSTRAINED,
93     VA_RC_CBR,
94     VA_RC_VCM,
95     VA_RC_NONE,
96 };
97
98 /* thread to save coded data */
99 #define SRC_SURFACE_FREE        0
100 #define SRC_SURFACE_IN_ENCODING 1
101     
102 struct __bitstream {
103     unsigned int *buffer;
104     int bit_offset;
105     int max_size_in_dword;
106 };
107 typedef struct __bitstream bitstream;
108
109 using namespace std;
110
111 // H.264 video comes out in encoding order (e.g. with two B-frames:
112 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
113 // come in the right order. Since we do everything, including waiting
114 // for the frames to come out of OpenGL, in encoding order, we need
115 // a reordering buffer for uncompressed frames so that they come out
116 // correctly. We go the super-lazy way of not making it understand
117 // anything about the true order (which introduces some extra latency,
118 // though); we know that for N B-frames we need at most (N-1) frames
119 // in the reorder buffer, and can just sort on that.
120 //
121 // The class also deals with keeping a freelist as needed.
122 class FrameReorderer {
123 public:
124         FrameReorderer(unsigned queue_length, int width, int height);
125
126         struct Frame {
127                 int64_t pts, duration;
128                 uint8_t *data;
129
130                 // Invert to get the smallest pts first.
131                 bool operator< (const Frame &other) const { return pts > other.pts; }
132         };
133
134         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
135         // Does _not_ take ownership of data; a copy is taken if needed.
136         // The returned pointer is valid until the next call to reorder_frame, or destruction.
137         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
138         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
139
140         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
141         Frame get_first_frame();
142
143         bool empty() const { return frames.empty(); }
144
145 private:
146         unsigned queue_length;
147         int width, height;
148
149         priority_queue<Frame> frames;
150         stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
151
152         // Owns all the pointers. Normally, freelist and frames could do this themselves,
153         // except priority_queue doesn't work well with movable-only types.
154         vector<unique_ptr<uint8_t[]>> owner;
155 };
156
157 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
158     : queue_length(queue_length), width(width), height(height)
159 {
160         for (unsigned i = 0; i < queue_length; ++i) {
161                 owner.emplace_back(new uint8_t[width * height * 2]);
162                 freelist.push(owner.back().get());
163         }
164 }
165
166 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
167 {
168         if (queue_length == 0) {
169                 return Frame{pts, duration, data};
170         }
171
172         assert(!freelist.empty());
173         uint8_t *storage = freelist.top();
174         freelist.pop();
175         memcpy(storage, data, width * height * 2);
176         frames.push(Frame{pts, duration, storage});
177
178         if (frames.size() >= queue_length) {
179                 return get_first_frame();
180         } else {
181                 return Frame{-1, -1, nullptr};
182         }
183 }
184
185 FrameReorderer::Frame FrameReorderer::get_first_frame()
186 {
187         assert(!frames.empty());
188         Frame storage = frames.top();
189         frames.pop();
190         freelist.push(storage.data);
191         return storage;
192 }
193
194 class QuickSyncEncoderImpl {
195 public:
196         QuickSyncEncoderImpl(const std::string &filename, QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder);
197         ~QuickSyncEncoderImpl();
198         void add_audio(int64_t pts, vector<float> audio);
199         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
200         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
201         void shutdown();
202
203 private:
204         struct storage_task {
205                 unsigned long long display_order;
206                 int frame_type;
207                 vector<float> audio;
208                 int64_t pts, dts, duration;
209         };
210         struct PendingFrame {
211                 RefCountedGLsync fence;
212                 vector<RefCountedFrame> input_frames;
213                 int64_t pts, duration;
214         };
215
216         // So we never get negative dts.
217         int64_t global_delay() const {
218                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
219         }
220
221         void open_output_file(const std::string &filename);
222         void encode_thread_func();
223         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
224         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
225         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
226                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
227         void storage_task_thread();
228         void encode_remaining_audio();
229         void storage_task_enqueue(storage_task task);
230         void save_codeddata(storage_task task);
231         int render_packedsequence();
232         int render_packedpicture();
233         void render_packedslice();
234         int render_sequence();
235         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
236         void sps_rbsp(bitstream *bs);
237         void pps_rbsp(bitstream *bs);
238         int build_packed_pic_buffer(unsigned char **header_buffer);
239         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
240         void slice_header(bitstream *bs);
241         int build_packed_seq_buffer(unsigned char **header_buffer);
242         int build_packed_slice_buffer(unsigned char **header_buffer);
243         int init_va(const string &va_display);
244         int deinit_va();
245         void enable_zerocopy_if_possible();
246         VADisplay va_open_display(const string &va_display);
247         void va_close_display(VADisplay va_dpy);
248         int setup_encode();
249         int release_encode();
250         void update_ReferenceFrames(int frame_type);
251         int update_RefPicList(int frame_type);
252
253         bool is_shutdown = false;
254         bool use_zerocopy;
255         int drm_fd = -1;
256
257         thread encode_thread, storage_thread;
258
259         mutex storage_task_queue_mutex;
260         condition_variable storage_task_queue_changed;
261         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
262         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
263         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
264
265         mutex frame_queue_mutex;
266         condition_variable frame_queue_nonempty;
267         bool encode_thread_should_quit = false;  // under frame_queue_mutex
268
269         int current_storage_frame;
270
271         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
272         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
273         QSurface *surface;
274
275         unique_ptr<AudioEncoder> file_audio_encoder;
276         AudioEncoder *stream_audio_encoder;
277
278         unique_ptr<FrameReorderer> reorderer;
279         X264Encoder *x264_encoder;  // nullptr if not using x264.
280
281         Mux* stream_mux;  // To HTTP.
282         unique_ptr<Mux> file_mux;  // To local disk.
283
284         Display *x11_display = nullptr;
285
286         // Encoder parameters
287         VADisplay va_dpy;
288         VAProfile h264_profile = (VAProfile)~0;
289         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
290         int config_attrib_num = 0, enc_packed_header_idx;
291
292         struct GLSurface {
293                 VASurfaceID src_surface, ref_surface;
294                 VABufferID coded_buf;
295
296                 VAImage surface_image;
297                 GLuint y_tex, cbcr_tex;
298
299                 // Only if use_zerocopy == true.
300                 EGLImage y_egl_image, cbcr_egl_image;
301
302                 // Only if use_zerocopy == false.
303                 GLuint pbo;
304                 uint8_t *y_ptr, *cbcr_ptr;
305                 size_t y_offset, cbcr_offset;
306         };
307         GLSurface gl_surfaces[SURFACE_NUM];
308
309         VAConfigID config_id;
310         VAContextID context_id;
311         VAEncSequenceParameterBufferH264 seq_param;
312         VAEncPictureParameterBufferH264 pic_param;
313         VAEncSliceParameterBufferH264 slice_param;
314         VAPictureH264 CurrentCurrPic;
315         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
316
317         // Static quality settings.
318         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
319         static constexpr unsigned int num_ref_frames = 2;
320         static constexpr int initial_qp = 15;
321         static constexpr int minimal_qp = 0;
322         static constexpr int intra_period = 30;
323         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
324
325         // Quality settings that are meant to be static, but might be overridden
326         // by the profile.
327         int constraint_set_flag = 0;
328         int h264_packedheader = 0; /* support pack header? */
329         int h264_maxref = (1<<16|1);
330         int h264_entropy_mode = 1; /* cabac */
331         int ip_period = 3;
332
333         int rc_mode = -1;
334         unsigned int current_frame_num = 0;
335         unsigned int numShortTerm = 0;
336
337         int frame_width;
338         int frame_height;
339         int frame_width_mbaligned;
340         int frame_height_mbaligned;
341 };
342
343 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
344 // but if we don't delete it here, we get leaks. The GStreamer implementation
345 // does the same.
346 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
347 {
348     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
349     CHECK_VASTATUS(va_status, "vaRenderPicture");
350
351     for (int i = 0; i < num_buffers; ++i) {
352         va_status = vaDestroyBuffer(dpy, buffers[i]);
353         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
354     }
355 }
356
357 static unsigned int 
358 va_swap32(unsigned int val)
359 {
360     unsigned char *pval = (unsigned char *)&val;
361
362     return ((pval[0] << 24)     |
363             (pval[1] << 16)     |
364             (pval[2] << 8)      |
365             (pval[3] << 0));
366 }
367
368 static void
369 bitstream_start(bitstream *bs)
370 {
371     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
372     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
373     bs->bit_offset = 0;
374 }
375
376 static void
377 bitstream_end(bitstream *bs)
378 {
379     int pos = (bs->bit_offset >> 5);
380     int bit_offset = (bs->bit_offset & 0x1f);
381     int bit_left = 32 - bit_offset;
382
383     if (bit_offset) {
384         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
385     }
386 }
387  
388 static void
389 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
390 {
391     int pos = (bs->bit_offset >> 5);
392     int bit_offset = (bs->bit_offset & 0x1f);
393     int bit_left = 32 - bit_offset;
394
395     if (!size_in_bits)
396         return;
397
398     bs->bit_offset += size_in_bits;
399
400     if (bit_left > size_in_bits) {
401         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
402     } else {
403         size_in_bits -= bit_left;
404         if (bit_left >= 32) {
405             bs->buffer[pos] = (val >> size_in_bits);
406         } else {
407             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
408         }
409         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
410
411         if (pos + 1 == bs->max_size_in_dword) {
412             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
413             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
414         }
415
416         bs->buffer[pos + 1] = val;
417     }
418 }
419
420 static void
421 bitstream_put_ue(bitstream *bs, unsigned int val)
422 {
423     int size_in_bits = 0;
424     int tmp_val = ++val;
425
426     while (tmp_val) {
427         tmp_val >>= 1;
428         size_in_bits++;
429     }
430
431     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
432     bitstream_put_ui(bs, val, size_in_bits);
433 }
434
435 static void
436 bitstream_put_se(bitstream *bs, int val)
437 {
438     unsigned int new_val;
439
440     if (val <= 0)
441         new_val = -2 * val;
442     else
443         new_val = 2 * val - 1;
444
445     bitstream_put_ue(bs, new_val);
446 }
447
448 static void
449 bitstream_byte_aligning(bitstream *bs, int bit)
450 {
451     int bit_offset = (bs->bit_offset & 0x7);
452     int bit_left = 8 - bit_offset;
453     int new_val;
454
455     if (!bit_offset)
456         return;
457
458     assert(bit == 0 || bit == 1);
459
460     if (bit)
461         new_val = (1 << bit_left) - 1;
462     else
463         new_val = 0;
464
465     bitstream_put_ui(bs, new_val, bit_left);
466 }
467
468 static void 
469 rbsp_trailing_bits(bitstream *bs)
470 {
471     bitstream_put_ui(bs, 1, 1);
472     bitstream_byte_aligning(bs, 0);
473 }
474
475 static void nal_start_code_prefix(bitstream *bs)
476 {
477     bitstream_put_ui(bs, 0x00000001, 32);
478 }
479
480 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
481 {
482     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
483     bitstream_put_ui(bs, nal_ref_idc, 2);
484     bitstream_put_ui(bs, nal_unit_type, 5);
485 }
486
487 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
488 {
489     int profile_idc = PROFILE_IDC_BASELINE;
490
491     if (h264_profile  == VAProfileH264High)
492         profile_idc = PROFILE_IDC_HIGH;
493     else if (h264_profile  == VAProfileH264Main)
494         profile_idc = PROFILE_IDC_MAIN;
495
496     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
497     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
498     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
499     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
500     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
501     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
502     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
503     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
504
505     if ( profile_idc == PROFILE_IDC_HIGH) {
506         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */ 
507         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
508         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
509         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
510         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
511     }
512
513     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
514     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
515
516     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
517         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
518     else {
519         assert(0);
520     }
521
522     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
523     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
524
525     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
526     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
527     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
528
529     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
530         assert(0);
531     }
532
533     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
534     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
535
536     if (seq_param.frame_cropping_flag) {
537         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
538         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
539         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
540         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
541     }
542     
543     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
544     if ( false ) {
545         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
546     } else {
547         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
548         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
549         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
550         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
551         {
552             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
553             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
554             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
555             {
556                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
557                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
558                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
559             }
560         }
561         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
562         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
563         {
564             bitstream_put_ui(bs, 1, 32);  // FPS
565             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
566             bitstream_put_ui(bs, 1, 1);
567         }
568         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
569         {
570             // hrd_parameters 
571             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
572             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
573             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
574            
575             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
576             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
577             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
578
579             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
580             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
581             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
582             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
583         }
584         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
585         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */ 
586
587         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
588         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
589     }
590
591     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
592 }
593
594
595 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
596 {
597     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
598     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
599
600     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
601
602     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
603
604     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
605
606     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
607     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
608
609     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
610     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
611
612     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
613     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
614     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
615
616     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
617     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
618     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
619     
620     /* more_rbsp_data */
621     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
622     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
623     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
624
625     rbsp_trailing_bits(bs);
626 }
627
628 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
629 {
630     int first_mb_in_slice = slice_param.macroblock_address;
631
632     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
633     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
634     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
635     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
636
637     /* frame_mbs_only_flag == 1 */
638     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
639         /* FIXME: */
640         assert(0);
641     }
642
643     if (pic_param.pic_fields.bits.idr_pic_flag)
644         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
645
646     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
647         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
648         /* pic_order_present_flag == 0 */
649     } else {
650         /* FIXME: */
651         assert(0);
652     }
653
654     /* redundant_pic_cnt_present_flag == 0 */
655     /* slice type */
656     if (IS_P_SLICE(slice_param.slice_type)) {
657         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
658
659         if (slice_param.num_ref_idx_active_override_flag)
660             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
661
662         /* ref_pic_list_reordering */
663         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
664     } else if (IS_B_SLICE(slice_param.slice_type)) {
665         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
666
667         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
668
669         if (slice_param.num_ref_idx_active_override_flag) {
670             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
671             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
672         }
673
674         /* ref_pic_list_reordering */
675         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
676         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
677     }
678
679     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
680          IS_P_SLICE(slice_param.slice_type)) ||
681         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
682          IS_B_SLICE(slice_param.slice_type))) {
683         /* FIXME: fill weight/offset table */
684         assert(0);
685     }
686
687     /* dec_ref_pic_marking */
688     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
689         unsigned char no_output_of_prior_pics_flag = 0;
690         unsigned char long_term_reference_flag = 0;
691         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
692
693         if (pic_param.pic_fields.bits.idr_pic_flag) {
694             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
695             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
696         } else {
697             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
698         }
699     }
700
701     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
702         !IS_I_SLICE(slice_param.slice_type))
703         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
704
705     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
706
707     /* ignore for SP/SI */
708
709     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
710         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
711
712         if (slice_param.disable_deblocking_filter_idc != 1) {
713             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
714             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
715         }
716     }
717
718     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
719         bitstream_byte_aligning(bs, 1);
720     }
721 }
722
723 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
724 {
725     bitstream bs;
726
727     bitstream_start(&bs);
728     nal_start_code_prefix(&bs);
729     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
730     pps_rbsp(&bs);
731     bitstream_end(&bs);
732
733     *header_buffer = (unsigned char *)bs.buffer;
734     return bs.bit_offset;
735 }
736
737 int
738 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
739 {
740     bitstream bs;
741
742     bitstream_start(&bs);
743     nal_start_code_prefix(&bs);
744     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
745     sps_rbsp(&bs);
746     bitstream_end(&bs);
747
748     *header_buffer = (unsigned char *)bs.buffer;
749     return bs.bit_offset;
750 }
751
752 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
753 {
754     bitstream bs;
755     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
756     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
757
758     bitstream_start(&bs);
759     nal_start_code_prefix(&bs);
760
761     if (IS_I_SLICE(slice_param.slice_type)) {
762         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
763     } else if (IS_P_SLICE(slice_param.slice_type)) {
764         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
765     } else {
766         assert(IS_B_SLICE(slice_param.slice_type));
767         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
768     }
769
770     slice_header(&bs);
771     bitstream_end(&bs);
772
773     *header_buffer = (unsigned char *)bs.buffer;
774     return bs.bit_offset;
775 }
776
777
778 /*
779   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
780   1) period between Frame #X and Frame #N = #X - #N
781   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
782   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
783   4) intra_period and intra_idr_period take precedence over ip_period
784   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods 
785      of I/IDR frames, see bellow examples
786   -------------------------------------------------------------------
787   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
788   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
789   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
790   1            0                ignored    IDRIIIIIII...      (No IDR any more)
791   1            1                ignored    IDR IDR IDR IDR...
792   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
793   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
794   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
795                                               (PBB)(IBB)(PBB)(IBB)... 
796   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
797                                            IDRPPPPP IPPPPP IPPPPP...
798   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
799                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
800                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
801                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
802                                            {IDR(PBB)(PBB)}                     (6/6/3)
803                                            {IDR(PBB)(PBB)}.
804 */
805
806 // General pts/dts strategy:
807 //
808 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
809 // bit tricky. We assume first of all that the frame rate never goes _above_
810 // MAX_FPS, which gives us a frame period N. The decoder can always decode
811 // in at least this speed, as long at dts <= pts (the frame is not attempted
812 // presented before it is decoded). Furthermore, we never have longer chains of
813 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
814 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
815 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
816 // frame, with an order O <= C.)
817 //
818 // Many strategies are possible, but we establish these rules:
819 //
820 //  - Tip frames have dts = pts - (C-O)*N.
821 //  - Non-tip frames have dts = dts_last + N.
822 //
823 // An example, with C=2 and N=10 and the data flow showed with arrows:
824 //
825 //        I  B  P  B  B  P
826 //   pts: 30 40 50 60 70 80
827 //        ↓  ↓     ↓
828 //   dts: 10 30 20 60 50←40
829 //         |  |  ↑        ↑
830 //         `--|--'        |
831 //             `----------'
832 //
833 // To show that this works fine also with irregular spacings, let's say that
834 // the third frame is delayed a bit (something earlier was dropped). Now the
835 // situation looks like this:
836 //
837 //        I  B  P  B  B   P
838 //   pts: 30 40 80 90 100 110
839 //        ↓  ↓     ↓
840 //   dts: 10 30 20 90 50←40
841 //         |  |  ↑        ↑
842 //         `--|--'        |
843 //             `----------'
844 //
845 // The resetting on every tip frame makes sure dts never ends up lagging a lot
846 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
847 //
848 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
849 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
850 // a tip frame and should be given a dts based on the previous one.
851 #define FRAME_P 0
852 #define FRAME_B 1
853 #define FRAME_I 2
854 #define FRAME_IDR 7
855 void encoding2display_order(
856     int encoding_order, int intra_period,
857     int intra_idr_period, int ip_period,
858     int *displaying_order,
859     int *frame_type, int *pts_lag)
860 {
861     int encoding_order_gop = 0;
862
863     *pts_lag = 0;
864
865     if (intra_period == 1) { /* all are I/IDR frames */
866         *displaying_order = encoding_order;
867         if (intra_idr_period == 0)
868             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
869         else
870             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
871         return;
872     }
873
874     if (intra_period == 0)
875         intra_idr_period = 0;
876
877     if (ip_period == 1) {
878         // No B-frames, sequence is like IDR PPPPP IPPPPP.
879         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
880         *displaying_order = encoding_order;
881
882         if (encoding_order_gop == 0) { /* the first frame */
883             *frame_type = FRAME_IDR;
884         } else if (intra_period != 0 && /* have I frames */
885                    encoding_order_gop >= 2 &&
886                    (encoding_order_gop % intra_period == 0)) {
887             *frame_type = FRAME_I;
888         } else {
889             *frame_type = FRAME_P;
890         }
891         return;
892     } 
893
894     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
895     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
896     *pts_lag = -1;  // Most frames are not tip frames.
897          
898     if (encoding_order_gop == 0) { /* the first frame */
899         *frame_type = FRAME_IDR;
900         *displaying_order = encoding_order;
901         // IDR frames are a special case; I honestly can't find the logic behind
902         // why this is the right thing, but it seems to line up nicely in practice :-)
903         *pts_lag = TIMEBASE / MAX_FPS;
904     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
905         *frame_type = FRAME_B;
906         *displaying_order = encoding_order - 1;
907         if ((encoding_order_gop % ip_period) == 0) {
908             *pts_lag = 0;  // Last B-frame.
909         }
910     } else if (intra_period != 0 && /* have I frames */
911                encoding_order_gop >= 2 &&
912                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
913         *frame_type = FRAME_I;
914         *displaying_order = encoding_order + ip_period - 1;
915     } else {
916         *frame_type = FRAME_P;
917         *displaying_order = encoding_order + ip_period - 1;
918     }
919 }
920
921
922 static const char *rc_to_string(int rc_mode)
923 {
924     switch (rc_mode) {
925     case VA_RC_NONE:
926         return "NONE";
927     case VA_RC_CBR:
928         return "CBR";
929     case VA_RC_VBR:
930         return "VBR";
931     case VA_RC_VCM:
932         return "VCM";
933     case VA_RC_CQP:
934         return "CQP";
935     case VA_RC_VBR_CONSTRAINED:
936         return "VBR_CONSTRAINED";
937     default:
938         return "Unknown";
939     }
940 }
941
942 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
943 {
944         if (global_flags.uncompressed_video_to_http) {
945                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
946                 use_zerocopy = false;
947         } else if (global_flags.x264_video_to_http) {
948                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
949                 use_zerocopy = false;
950         } else {
951                 use_zerocopy = true;
952         }
953 }
954
955 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
956 {
957         if (va_display.empty()) {
958                 x11_display = XOpenDisplay(NULL);
959                 if (!x11_display) {
960                         fprintf(stderr, "error: can't connect to X server!\n");
961                         return NULL;
962                 }
963                 enable_zerocopy_if_possible();
964                 return vaGetDisplay(x11_display);
965         } else if (va_display[0] != '/') {
966                 x11_display = XOpenDisplay(va_display.c_str());
967                 if (!x11_display) {
968                         fprintf(stderr, "error: can't connect to X server!\n");
969                         return NULL;
970                 }
971                 enable_zerocopy_if_possible();
972                 return vaGetDisplay(x11_display);
973         } else {
974                 drm_fd = open(va_display.c_str(), O_RDWR);
975                 if (drm_fd == -1) {
976                         perror(va_display.c_str());
977                         return NULL;
978                 }
979                 use_zerocopy = false;
980                 return vaGetDisplayDRM(drm_fd);
981         }
982 }
983
984 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
985 {
986         if (x11_display) {
987                 XCloseDisplay(x11_display);
988                 x11_display = nullptr;
989         }
990         if (drm_fd != -1) {
991                 close(drm_fd);
992         }
993 }
994
995 int QuickSyncEncoderImpl::init_va(const string &va_display)
996 {
997     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
998     VAEntrypoint *entrypoints;
999     int num_entrypoints, slice_entrypoint;
1000     int support_encode = 0;    
1001     int major_ver, minor_ver;
1002     VAStatus va_status;
1003     unsigned int i;
1004
1005     va_dpy = va_open_display(va_display);
1006     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1007     CHECK_VASTATUS(va_status, "vaInitialize");
1008
1009     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1010     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1011     if (!entrypoints) {
1012         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1013         exit(1);
1014     }
1015
1016     /* use the highest profile */
1017     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1018         if ((h264_profile != ~0) && h264_profile != profile_list[i])
1019             continue;
1020         
1021         h264_profile = profile_list[i];
1022         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1023         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1024             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1025                 support_encode = 1;
1026                 break;
1027             }
1028         }
1029         if (support_encode == 1)
1030             break;
1031     }
1032     
1033     if (support_encode == 0) {
1034         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1035         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1036         printf("to use VA-API against DRM instead of X11.\n");
1037         exit(1);
1038     } else {
1039         switch (h264_profile) {
1040             case VAProfileH264Baseline:
1041                 ip_period = 1;
1042                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1043                 h264_entropy_mode = 0;
1044                 break;
1045             case VAProfileH264ConstrainedBaseline:
1046                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1047                 ip_period = 1;
1048                 break;
1049
1050             case VAProfileH264Main:
1051                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1052                 break;
1053
1054             case VAProfileH264High:
1055                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1056                 break;
1057             default:
1058                 h264_profile = VAProfileH264Baseline;
1059                 ip_period = 1;
1060                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1061                 break;
1062         }
1063     }
1064
1065     VAConfigAttrib attrib[VAConfigAttribTypeMax];
1066
1067     /* find out the format for the render target, and rate control mode */
1068     for (i = 0; i < VAConfigAttribTypeMax; i++)
1069         attrib[i].type = (VAConfigAttribType)i;
1070
1071     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1072                                       &attrib[0], VAConfigAttribTypeMax);
1073     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1074     /* check the interested configattrib */
1075     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1076         printf("Not find desired YUV420 RT format\n");
1077         exit(1);
1078     } else {
1079         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1080         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1081         config_attrib_num++;
1082     }
1083     
1084     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1085         int tmp = attrib[VAConfigAttribRateControl].value;
1086
1087         if (rc_mode == -1 || !(rc_mode & tmp))  {
1088             if (rc_mode != -1) {
1089                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1090             }
1091
1092             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1093                 if (rc_default_modes[i] & tmp) {
1094                     rc_mode = rc_default_modes[i];
1095                     break;
1096                 }
1097             }
1098         }
1099
1100         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1101         config_attrib[config_attrib_num].value = rc_mode;
1102         config_attrib_num++;
1103     }
1104     
1105
1106     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1107         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1108
1109         h264_packedheader = 1;
1110         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1111         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1112         
1113         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1114             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1115         }
1116         
1117         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1118             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1119         }
1120         
1121         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1122             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1123         }
1124         
1125         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1126             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1127         }
1128         
1129         enc_packed_header_idx = config_attrib_num;
1130         config_attrib_num++;
1131     }
1132
1133     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1134         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1135         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1136         config_attrib_num++;
1137     }
1138     
1139     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1140         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1141     }
1142
1143     free(entrypoints);
1144     return 0;
1145 }
1146
1147 int QuickSyncEncoderImpl::setup_encode()
1148 {
1149     VAStatus va_status;
1150     VASurfaceID *tmp_surfaceid;
1151     int codedbuf_size, i;
1152     static VASurfaceID src_surface[SURFACE_NUM];
1153     static VASurfaceID ref_surface[SURFACE_NUM];
1154     
1155     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1156             &config_attrib[0], config_attrib_num, &config_id);
1157     CHECK_VASTATUS(va_status, "vaCreateConfig");
1158
1159     /* create source surfaces */
1160     va_status = vaCreateSurfaces(va_dpy,
1161                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1162                                  &src_surface[0], SURFACE_NUM,
1163                                  NULL, 0);
1164     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1165
1166     /* create reference surfaces */
1167     va_status = vaCreateSurfaces(va_dpy,
1168                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1169                                  &ref_surface[0], SURFACE_NUM,
1170                                  NULL, 0);
1171     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1172
1173     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1174     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1175     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1176     
1177     /* Create a context for this encode pipe */
1178     va_status = vaCreateContext(va_dpy, config_id,
1179                                 frame_width_mbaligned, frame_height_mbaligned,
1180                                 VA_PROGRESSIVE,
1181                                 tmp_surfaceid, 2 * SURFACE_NUM,
1182                                 &context_id);
1183     CHECK_VASTATUS(va_status, "vaCreateContext");
1184     free(tmp_surfaceid);
1185
1186     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1187
1188     for (i = 0; i < SURFACE_NUM; i++) {
1189         /* create coded buffer once for all
1190          * other VA buffers which won't be used again after vaRenderPicture.
1191          * so APP can always vaCreateBuffer for every frame
1192          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1193          * so VA won't maintain the coded buffer
1194          */
1195         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1196                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1197         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1198     }
1199
1200     /* create OpenGL objects */
1201     //glGenFramebuffers(SURFACE_NUM, fbos);
1202     
1203     for (i = 0; i < SURFACE_NUM; i++) {
1204         glGenTextures(1, &gl_surfaces[i].y_tex);
1205         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1206
1207         if (!use_zerocopy) {
1208             // Create Y image.
1209             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1210             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1211
1212             // Create CbCr image.
1213             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1214             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1215
1216             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1217             // buffers, due to potentially differing pitch.
1218             glGenBuffers(1, &gl_surfaces[i].pbo);
1219             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1220             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1221             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1222             gl_surfaces[i].y_offset = 0;
1223             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1224             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1225             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1226             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1227         }
1228     }
1229
1230     for (i = 0; i < SURFACE_NUM; i++) {
1231         gl_surfaces[i].src_surface = src_surface[i];
1232         gl_surfaces[i].ref_surface = ref_surface[i];
1233     }
1234     
1235     return 0;
1236 }
1237
1238 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1239 //
1240 //   2 1 0 [3] 4 8 9
1241 template<class T, class C>
1242 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1243 {
1244         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1245         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1246         sort(middle, end, less_than);
1247 }
1248
1249 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1250 {
1251     int i;
1252     
1253     if (frame_type == FRAME_B)
1254         return;
1255
1256     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1257     numShortTerm++;
1258     if (numShortTerm > num_ref_frames)
1259         numShortTerm = num_ref_frames;
1260     for (i=numShortTerm-1; i>0; i--)
1261         ReferenceFrames[i] = ReferenceFrames[i-1];
1262     ReferenceFrames[0] = CurrentCurrPic;
1263     
1264     current_frame_num++;
1265     if (current_frame_num > MaxFrameNum)
1266         current_frame_num = 0;
1267 }
1268
1269
1270 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1271 {
1272     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1273         return a.frame_idx > b.frame_idx;
1274     };
1275     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1276         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1277     };
1278     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1279         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1280     };
1281     
1282     if (frame_type == FRAME_P) {
1283         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1284         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1285     } else if (frame_type == FRAME_B) {
1286         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1287         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1288
1289         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1290         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1291     }
1292     
1293     return 0;
1294 }
1295
1296
1297 int QuickSyncEncoderImpl::render_sequence()
1298 {
1299     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1300     VAStatus va_status;
1301     VAEncMiscParameterBuffer *misc_param;
1302     VAEncMiscParameterRateControl *misc_rate_ctrl;
1303     
1304     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1305     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1306     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1307     seq_param.bits_per_second = frame_bitrate;
1308
1309     seq_param.intra_period = intra_period;
1310     seq_param.intra_idr_period = intra_idr_period;
1311     seq_param.ip_period = ip_period;
1312
1313     seq_param.max_num_ref_frames = num_ref_frames;
1314     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1315     seq_param.time_scale = TIMEBASE * 2;
1316     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1317     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1318     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1319     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1320     seq_param.seq_fields.bits.chroma_format_idc = 1;
1321     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1322     
1323     if (frame_width != frame_width_mbaligned ||
1324         frame_height != frame_height_mbaligned) {
1325         seq_param.frame_cropping_flag = 1;
1326         seq_param.frame_crop_left_offset = 0;
1327         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1328         seq_param.frame_crop_top_offset = 0;
1329         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1330     }
1331     
1332     va_status = vaCreateBuffer(va_dpy, context_id,
1333                                VAEncSequenceParameterBufferType,
1334                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1335     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1336     
1337     va_status = vaCreateBuffer(va_dpy, context_id,
1338                                VAEncMiscParameterBufferType,
1339                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1340                                1, NULL, &rc_param_buf);
1341     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1342     
1343     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1344     misc_param->type = VAEncMiscParameterTypeRateControl;
1345     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1346     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1347     misc_rate_ctrl->bits_per_second = frame_bitrate;
1348     misc_rate_ctrl->target_percentage = 66;
1349     misc_rate_ctrl->window_size = 1000;
1350     misc_rate_ctrl->initial_qp = initial_qp;
1351     misc_rate_ctrl->min_qp = minimal_qp;
1352     misc_rate_ctrl->basic_unit_size = 0;
1353     vaUnmapBuffer(va_dpy, rc_param_buf);
1354
1355     render_id[0] = seq_param_buf;
1356     render_id[1] = rc_param_buf;
1357     
1358     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1359     
1360     return 0;
1361 }
1362
1363 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1364 {
1365     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1366     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1367     int PicOrderCntMsb, TopFieldOrderCnt;
1368     
1369     if (frame_type == FRAME_IDR)
1370         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1371     else {
1372         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1373         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1374     }
1375     
1376     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1377         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1378         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1379     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1380              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1381         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1382     else
1383         PicOrderCntMsb = prevPicOrderCntMsb;
1384     
1385     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1386
1387     if (frame_type != FRAME_B) {
1388         PicOrderCntMsb_ref = PicOrderCntMsb;
1389         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1390     }
1391     
1392     return TopFieldOrderCnt;
1393 }
1394
1395 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1396 {
1397     VABufferID pic_param_buf;
1398     VAStatus va_status;
1399     int i = 0;
1400
1401     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1402     pic_param.CurrPic.frame_idx = current_frame_num;
1403     pic_param.CurrPic.flags = 0;
1404     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1405     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1406     CurrentCurrPic = pic_param.CurrPic;
1407
1408     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1409     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1410         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1411         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1412     }
1413     
1414     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1415     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1416     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1417     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1418     pic_param.frame_num = current_frame_num;
1419     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1420     pic_param.last_picture = false;  // FIXME
1421     pic_param.pic_init_qp = initial_qp;
1422
1423     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1424                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1425     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1426
1427     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1428
1429     return 0;
1430 }
1431
1432 int QuickSyncEncoderImpl::render_packedsequence()
1433 {
1434     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1435     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1436     unsigned int length_in_bits;
1437     unsigned char *packedseq_buffer = NULL;
1438     VAStatus va_status;
1439
1440     length_in_bits = build_packed_seq_buffer(&packedseq_buffer); 
1441     
1442     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1443     
1444     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1445     packedheader_param_buffer.has_emulation_bytes = 0;
1446     va_status = vaCreateBuffer(va_dpy,
1447                                context_id,
1448                                VAEncPackedHeaderParameterBufferType,
1449                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1450                                &packedseq_para_bufid);
1451     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1452
1453     va_status = vaCreateBuffer(va_dpy,
1454                                context_id,
1455                                VAEncPackedHeaderDataBufferType,
1456                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1457                                &packedseq_data_bufid);
1458     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1459
1460     render_id[0] = packedseq_para_bufid;
1461     render_id[1] = packedseq_data_bufid;
1462     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1463
1464     free(packedseq_buffer);
1465     
1466     return 0;
1467 }
1468
1469
1470 int QuickSyncEncoderImpl::render_packedpicture()
1471 {
1472     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1473     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1474     unsigned int length_in_bits;
1475     unsigned char *packedpic_buffer = NULL;
1476     VAStatus va_status;
1477
1478     length_in_bits = build_packed_pic_buffer(&packedpic_buffer); 
1479     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1480     packedheader_param_buffer.bit_length = length_in_bits;
1481     packedheader_param_buffer.has_emulation_bytes = 0;
1482
1483     va_status = vaCreateBuffer(va_dpy,
1484                                context_id,
1485                                VAEncPackedHeaderParameterBufferType,
1486                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1487                                &packedpic_para_bufid);
1488     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1489
1490     va_status = vaCreateBuffer(va_dpy,
1491                                context_id,
1492                                VAEncPackedHeaderDataBufferType,
1493                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1494                                &packedpic_data_bufid);
1495     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1496
1497     render_id[0] = packedpic_para_bufid;
1498     render_id[1] = packedpic_data_bufid;
1499     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1500
1501     free(packedpic_buffer);
1502     
1503     return 0;
1504 }
1505
1506 void QuickSyncEncoderImpl::render_packedslice()
1507 {
1508     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1509     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1510     unsigned int length_in_bits;
1511     unsigned char *packedslice_buffer = NULL;
1512     VAStatus va_status;
1513
1514     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1515     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1516     packedheader_param_buffer.bit_length = length_in_bits;
1517     packedheader_param_buffer.has_emulation_bytes = 0;
1518
1519     va_status = vaCreateBuffer(va_dpy,
1520                                context_id,
1521                                VAEncPackedHeaderParameterBufferType,
1522                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1523                                &packedslice_para_bufid);
1524     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1525
1526     va_status = vaCreateBuffer(va_dpy,
1527                                context_id,
1528                                VAEncPackedHeaderDataBufferType,
1529                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1530                                &packedslice_data_bufid);
1531     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1532
1533     render_id[0] = packedslice_para_bufid;
1534     render_id[1] = packedslice_data_bufid;
1535     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1536
1537     free(packedslice_buffer);
1538 }
1539
1540 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1541 {
1542     VABufferID slice_param_buf;
1543     VAStatus va_status;
1544     int i;
1545
1546     update_RefPicList(frame_type);
1547     
1548     /* one frame, one slice */
1549     slice_param.macroblock_address = 0;
1550     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1551     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1552     if (frame_type == FRAME_IDR) {
1553         if (encoding_frame_num != 0)
1554             ++slice_param.idr_pic_id;
1555     } else if (frame_type == FRAME_P) {
1556         int refpiclist0_max = h264_maxref & 0xffff;
1557         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1558
1559         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1560             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1561             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1562         }
1563     } else if (frame_type == FRAME_B) {
1564         int refpiclist0_max = h264_maxref & 0xffff;
1565         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1566
1567         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1568         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1569             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1570             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1571         }
1572
1573         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1574         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1575             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1576             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1577         }
1578     }
1579
1580     slice_param.slice_alpha_c0_offset_div2 = 0;
1581     slice_param.slice_beta_offset_div2 = 0;
1582     slice_param.direct_spatial_mv_pred_flag = 1;
1583     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1584     
1585
1586     if (h264_packedheader &&
1587         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1588         render_packedslice();
1589
1590     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1591                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1592     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1593
1594     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1595
1596     return 0;
1597 }
1598
1599
1600
1601 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1602 {    
1603         VACodedBufferSegment *buf_list = NULL;
1604         VAStatus va_status;
1605
1606         string data;
1607
1608         va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1609         CHECK_VASTATUS(va_status, "vaMapBuffer");
1610         while (buf_list != NULL) {
1611                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1612                 buf_list = (VACodedBufferSegment *) buf_list->next;
1613         }
1614         vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1615
1616         {
1617                 // Add video.
1618                 AVPacket pkt;
1619                 memset(&pkt, 0, sizeof(pkt));
1620                 pkt.buf = nullptr;
1621                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1622                 pkt.size = data.size();
1623                 pkt.stream_index = 0;
1624                 if (task.frame_type == FRAME_IDR) {
1625                         pkt.flags = AV_PKT_FLAG_KEY;
1626                 } else {
1627                         pkt.flags = 0;
1628                 }
1629                 pkt.duration = task.duration;
1630                 if (file_mux) {
1631                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1632                 }
1633                 if (!global_flags.uncompressed_video_to_http &&
1634                     !global_flags.x264_video_to_http) {
1635                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1636                 }
1637         }
1638         // Encode and add all audio frames up to and including the pts of this video frame.
1639         for ( ;; ) {
1640                 int64_t audio_pts;
1641                 vector<float> audio;
1642                 {
1643                         unique_lock<mutex> lock(frame_queue_mutex);
1644                         frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1645                         if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1646                         auto it = pending_audio_frames.begin();
1647                         if (it->first > task.pts) break;
1648                         audio_pts = it->first;
1649                         audio = move(it->second);
1650                         pending_audio_frames.erase(it); 
1651                 }
1652
1653                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1654                 stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
1655
1656                 if (audio_pts == task.pts) break;
1657         }
1658 }
1659
1660
1661 // this is weird. but it seems to put a new frame onto the queue
1662 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1663 {
1664         unique_lock<mutex> lock(storage_task_queue_mutex);
1665         storage_task_queue.push(move(task));
1666         storage_task_queue_changed.notify_all();
1667 }
1668
1669 void QuickSyncEncoderImpl::storage_task_thread()
1670 {
1671         for ( ;; ) {
1672                 storage_task current;
1673                 {
1674                         // wait until there's an encoded frame  
1675                         unique_lock<mutex> lock(storage_task_queue_mutex);
1676                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1677                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1678                         current = move(storage_task_queue.front());
1679                         storage_task_queue.pop();
1680                 }
1681
1682                 VAStatus va_status;
1683            
1684                 // waits for data, then saves it to disk.
1685                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1686                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1687                 save_codeddata(move(current));
1688
1689                 {
1690                         unique_lock<mutex> lock(storage_task_queue_mutex);
1691                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1692                         storage_task_queue_changed.notify_all();
1693                 }
1694         }
1695 }
1696
1697 int QuickSyncEncoderImpl::release_encode()
1698 {
1699         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1700                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1701                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1702                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1703
1704                 if (!use_zerocopy) {
1705                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1706                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1707                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1708                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1709                 }
1710                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1711                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1712         }
1713
1714         vaDestroyContext(va_dpy, context_id);
1715         vaDestroyConfig(va_dpy, config_id);
1716
1717         return 0;
1718 }
1719
1720 int QuickSyncEncoderImpl::deinit_va()
1721
1722     vaTerminate(va_dpy);
1723
1724     va_close_display(va_dpy);
1725
1726     return 0;
1727 }
1728
1729 namespace {
1730
1731 }  // namespace
1732
1733 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder)
1734         : current_storage_frame(0), surface(surface), stream_audio_encoder(stream_audio_encoder), x264_encoder(x264_encoder), stream_mux(stream_mux), frame_width(width), frame_height(height)
1735 {
1736         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE));
1737         open_output_file(filename);
1738         file_audio_encoder->add_mux(file_mux.get());
1739
1740         frame_width_mbaligned = (frame_width + 15) & (~15);
1741         frame_height_mbaligned = (frame_height + 15) & (~15);
1742
1743         //print_input();
1744
1745         if (global_flags.uncompressed_video_to_http ||
1746             global_flags.x264_video_to_http) {
1747                 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1748         }
1749         if (global_flags.x264_video_to_http) {
1750                 assert(x264_encoder != nullptr);
1751         } else {
1752                 assert(x264_encoder == nullptr);
1753         }
1754
1755         init_va(va_display);
1756         setup_encode();
1757
1758         // No frames are ready yet.
1759         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1760             
1761         memset(&seq_param, 0, sizeof(seq_param));
1762         memset(&pic_param, 0, sizeof(pic_param));
1763         memset(&slice_param, 0, sizeof(slice_param));
1764
1765         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1766
1767         encode_thread = thread([this]{
1768                 //SDL_GL_MakeCurrent(window, context);
1769                 QOpenGLContext *context = create_context(this->surface);
1770                 eglBindAPI(EGL_OPENGL_API);
1771                 if (!make_current(context, this->surface)) {
1772                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1773                                 eglGetError());
1774                         exit(1);
1775                 }
1776                 encode_thread_func();
1777                 delete_context(context);
1778         });
1779 }
1780
1781 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1782 {
1783         shutdown();
1784 }
1785
1786 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1787 {
1788         assert(!is_shutdown);
1789         {
1790                 // Wait until this frame slot is done encoding.
1791                 unique_lock<mutex> lock(storage_task_queue_mutex);
1792                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1793                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1794                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1795                 }
1796                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1797                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1798                 if (storage_thread_should_quit) return false;
1799         }
1800
1801         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1802         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1803         *y_tex = surf->y_tex;
1804         *cbcr_tex = surf->cbcr_tex;
1805
1806         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1807         CHECK_VASTATUS(va_status, "vaDeriveImage");
1808
1809         if (use_zerocopy) {
1810                 VABufferInfo buf_info;
1811                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1812                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1813                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1814
1815                 // Create Y image.
1816                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1817                 EGLint y_attribs[] = {
1818                         EGL_WIDTH, frame_width,
1819                         EGL_HEIGHT, frame_height,
1820                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1821                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1822                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1823                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1824                         EGL_NONE
1825                 };
1826
1827                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1828                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1829
1830                 // Associate Y image to a texture.
1831                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1832                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1833
1834                 // Create CbCr image.
1835                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1836                 EGLint cbcr_attribs[] = {
1837                         EGL_WIDTH, frame_width,
1838                         EGL_HEIGHT, frame_height,
1839                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1840                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1841                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1842                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1843                         EGL_NONE
1844                 };
1845
1846                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1847                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1848
1849                 // Associate CbCr image to a texture.
1850                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1851                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1852         }
1853
1854         return true;
1855 }
1856
1857 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1858 {
1859         assert(!is_shutdown);
1860         {
1861                 unique_lock<mutex> lock(frame_queue_mutex);
1862                 pending_audio_frames[pts] = move(audio);
1863         }
1864         frame_queue_nonempty.notify_all();
1865 }
1866
1867 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1868 {
1869         assert(!is_shutdown);
1870
1871         if (!use_zerocopy) {
1872                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1873
1874                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1875                 check_error();
1876
1877                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1878                 check_error();
1879
1880                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1881                 check_error();
1882                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1883                 check_error();
1884
1885                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1886                 check_error();
1887                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1888                 check_error();
1889
1890                 glBindTexture(GL_TEXTURE_2D, 0);
1891                 check_error();
1892                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1893                 check_error();
1894
1895                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1896                 check_error();
1897         }
1898
1899         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1900         check_error();
1901         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1902         check_error();
1903
1904         {
1905                 unique_lock<mutex> lock(frame_queue_mutex);
1906                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1907                 ++current_storage_frame;
1908         }
1909         frame_queue_nonempty.notify_all();
1910         return fence;
1911 }
1912
1913 void QuickSyncEncoderImpl::shutdown()
1914 {
1915         if (is_shutdown) {
1916                 return;
1917         }
1918
1919         {
1920                 unique_lock<mutex> lock(frame_queue_mutex);
1921                 encode_thread_should_quit = true;
1922                 frame_queue_nonempty.notify_all();
1923         }
1924         encode_thread.join();
1925         {
1926                 unique_lock<mutex> lock(storage_task_queue_mutex);
1927                 storage_thread_should_quit = true;
1928                 frame_queue_nonempty.notify_all();
1929                 storage_task_queue_changed.notify_all();
1930         }
1931         storage_thread.join();
1932         encode_remaining_audio();
1933
1934         release_encode();
1935         deinit_va();
1936         file_mux.reset();
1937         is_shutdown = true;
1938 }
1939
1940 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1941 {
1942         AVFormatContext *avctx = avformat_alloc_context();
1943         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1944         assert(filename.size() < sizeof(avctx->filename) - 1);
1945         strcpy(avctx->filename, filename.c_str());
1946
1947         string url = "file:" + filename;
1948         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1949         if (ret < 0) {
1950                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1951                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1952                 exit(1);
1953         }
1954
1955         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, file_audio_encoder->get_codec(), TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE, nullptr));
1956 }
1957
1958 void QuickSyncEncoderImpl::encode_thread_func()
1959 {
1960         int64_t last_dts = -1;
1961         int gop_start_display_frame_num = 0;
1962         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1963                 PendingFrame frame;
1964                 int pts_lag;
1965                 int frame_type, display_frame_num;
1966                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1967                                        &display_frame_num, &frame_type, &pts_lag);
1968                 if (frame_type == FRAME_IDR) {
1969                         numShortTerm = 0;
1970                         current_frame_num = 0;
1971                         gop_start_display_frame_num = display_frame_num;
1972                 }
1973
1974                 {
1975                         unique_lock<mutex> lock(frame_queue_mutex);
1976                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1977                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1978                         });
1979                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1980                                 // We have queued frames that were supposed to be B-frames,
1981                                 // but will be no P-frame to encode them against. Encode them all
1982                                 // as P-frames instead. Note that this happens under the mutex,
1983                                 // but nobody else uses it at this point, since we're shutting down,
1984                                 // so there's no contention.
1985                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1986                                 return;
1987                         } else {
1988                                 frame = move(pending_video_frames[display_frame_num]);
1989                                 pending_video_frames.erase(display_frame_num);
1990                         }
1991                 }
1992
1993                 // Determine the dts of this frame.
1994                 int64_t dts;
1995                 if (pts_lag == -1) {
1996                         assert(last_dts != -1);
1997                         dts = last_dts + (TIMEBASE / MAX_FPS);
1998                 } else {
1999                         dts = frame.pts - pts_lag;
2000                 }
2001                 last_dts = dts;
2002
2003                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2004         }
2005 }
2006
2007 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2008 {
2009         if (pending_video_frames.empty()) {
2010                 return;
2011         }
2012
2013         for (auto &pending_frame : pending_video_frames) {
2014                 int display_frame_num = pending_frame.first;
2015                 assert(display_frame_num > 0);
2016                 PendingFrame frame = move(pending_frame.second);
2017                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2018                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2019                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2020                 last_dts = dts;
2021         }
2022
2023         if (global_flags.uncompressed_video_to_http ||
2024             global_flags.x264_video_to_http) {
2025                 // Add frames left in reorderer.
2026                 while (!reorderer->empty()) {
2027                         FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2028                         if (global_flags.uncompressed_video_to_http) {
2029                                 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2030                         } else {
2031                                 assert(global_flags.x264_video_to_http);
2032                                 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2033                         }
2034                 }
2035         }
2036 }
2037
2038 void QuickSyncEncoderImpl::encode_remaining_audio()
2039 {
2040         // This really ought to be empty by now, but just to be sure...
2041         for (auto &pending_frame : pending_audio_frames) {
2042                 int64_t audio_pts = pending_frame.first;
2043                 vector<float> audio = move(pending_frame.second);
2044
2045                 file_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2046                 if (stream_audio_encoder) {
2047                         stream_audio_encoder->encode_audio(audio, audio_pts + global_delay());
2048                 }
2049         }
2050         pending_audio_frames.clear();
2051
2052         // Encode any leftover audio in the queues, and also any delayed frames.
2053         // Note: stream_audio_encoder is not owned by us, so don't call encode_last_audio().
2054         file_audio_encoder->encode_last_audio();
2055 }
2056
2057 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2058 {
2059         AVPacket pkt;
2060         memset(&pkt, 0, sizeof(pkt));
2061         pkt.buf = nullptr;
2062         pkt.data = const_cast<uint8_t *>(data);
2063         pkt.size = frame_width * frame_height * 2;
2064         pkt.stream_index = 0;
2065         pkt.flags = AV_PKT_FLAG_KEY;
2066         pkt.duration = duration;
2067         stream_mux->add_packet(pkt, pts, pts);
2068 }
2069
2070 namespace {
2071
2072 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2073 {
2074         if (src_width == dst_pitch) {
2075                 memcpy(dst, src, src_width * height);
2076         } else {
2077                 for (size_t y = 0; y < height; ++y) {
2078                         const uint8_t *sptr = src + y * src_width;
2079                         uint8_t *dptr = dst + y * dst_pitch;
2080                         memcpy(dptr, sptr, src_width);
2081                 }
2082         }
2083 }
2084
2085 }  // namespace
2086
2087 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2088                                    int frame_type, int64_t pts, int64_t dts, int64_t duration)
2089 {
2090         // Wait for the GPU to be done with the frame.
2091         GLenum sync_status;
2092         do {
2093                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2094                 check_error();
2095         } while (sync_status == GL_TIMEOUT_EXPIRED);
2096         assert(sync_status != GL_WAIT_FAILED);
2097
2098         // Release back any input frames we needed to render this frame.
2099         frame.input_frames.clear();
2100
2101         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2102         VAStatus va_status;
2103
2104         if (use_zerocopy) {
2105                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2106                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2107                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2108                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2109         } else {
2110                 unsigned char *surface_p = nullptr;
2111                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2112
2113                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2114                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2115
2116                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2117                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2118
2119                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2120                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2121
2122                 if (global_flags.uncompressed_video_to_http ||
2123                     global_flags.x264_video_to_http) {
2124                         // Add uncompressed video. (Note that pts == dts here.)
2125                         // Delay needs to match audio.
2126                         FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2127                         if (output_frame.data != nullptr) {
2128                                 if (global_flags.uncompressed_video_to_http) {
2129                                         add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2130                                 } else {
2131                                         assert(global_flags.x264_video_to_http);
2132                                         x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2133                                 }
2134                         }
2135                 }
2136         }
2137
2138         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2139         CHECK_VASTATUS(va_status, "vaDestroyImage");
2140
2141         // Schedule the frame for encoding.
2142         VASurfaceID va_surface = surf->src_surface;
2143         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2144         CHECK_VASTATUS(va_status, "vaBeginPicture");
2145
2146         if (frame_type == FRAME_IDR) {
2147                 render_sequence();
2148                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2149                 if (h264_packedheader) {
2150                         render_packedsequence();
2151                         render_packedpicture();
2152                 }
2153         } else {
2154                 //render_sequence();
2155                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2156         }
2157         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2158
2159         va_status = vaEndPicture(va_dpy, context_id);
2160         CHECK_VASTATUS(va_status, "vaEndPicture");
2161
2162         // so now the data is done encoding (well, async job kicked off)...
2163         // we send that to the storage thread
2164         storage_task tmp;
2165         tmp.display_order = display_frame_num;
2166         tmp.frame_type = frame_type;
2167         tmp.pts = pts;
2168         tmp.dts = dts;
2169         tmp.duration = duration;
2170         storage_task_enqueue(move(tmp));
2171
2172         update_ReferenceFrames(frame_type);
2173 }
2174
2175 // Proxy object.
2176 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, QSurface *surface, const string &va_display, int width, int height, Mux *stream_mux, AudioEncoder *stream_audio_encoder, X264Encoder *x264_encoder)
2177         : impl(new QuickSyncEncoderImpl(filename, surface, va_display, width, height, stream_mux, stream_audio_encoder, x264_encoder)) {}
2178
2179 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2180 QuickSyncEncoder::~QuickSyncEncoder() {}
2181
2182 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2183 {
2184         impl->add_audio(pts, audio);
2185 }
2186
2187 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2188 {
2189         return impl->begin_frame(y_tex, cbcr_tex);
2190 }
2191
2192 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2193 {
2194         return impl->end_frame(pts, duration, input_frames);
2195 }
2196
2197 void QuickSyncEncoder::shutdown()
2198 {
2199         impl->shutdown();
2200 }