]> git.sesse.net Git - nageru/blob - h264encode.cpp
876e956ec36cfcf7f0abb38c59ad3dc9772933b2
[nageru] / h264encode.cpp
1 //#include "sysdeps.h"
2 #include "h264encode.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 extern "C" {
11 #include <libavcodec/avcodec.h>
12 #include <libavformat/avformat.h>
13 #include <libavutil/channel_layout.h>
14 #include <libavutil/frame.h>
15 #include <libavutil/rational.h>
16 #include <libavutil/samplefmt.h>
17 }
18 #include <libdrm/drm_fourcc.h>
19 #include <stdio.h>
20 #include <stdlib.h>
21 #include <string.h>
22 #include <fcntl.h>
23 #include <va/va.h>
24 #include <va/va_drm.h>
25 #include <va/va_drmcommon.h>
26 #include <va/va_enc_h264.h>
27 #include <va/va_x11.h>
28 #include <algorithm>
29 #include <condition_variable>
30 #include <cstdint>
31 #include <map>
32 #include <memory>
33 #include <mutex>
34 #include <queue>
35 #include <string>
36 #include <thread>
37 #include <utility>
38
39 #include "context.h"
40 #include "defs.h"
41 #include "httpd.h"
42 #include "timebase.h"
43
44 using namespace std;
45
46 class QOpenGLContext;
47 class QSurface;
48
49 #define CHECK_VASTATUS(va_status, func)                                 \
50     if (va_status != VA_STATUS_SUCCESS) {                               \
51         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
52         exit(1);                                                        \
53     }
54
55 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
56
57 //#include "loadsurface.h"
58
59 #define NAL_REF_IDC_NONE        0
60 #define NAL_REF_IDC_LOW         1
61 #define NAL_REF_IDC_MEDIUM      2
62 #define NAL_REF_IDC_HIGH        3
63
64 #define NAL_NON_IDR             1
65 #define NAL_IDR                 5
66 #define NAL_SPS                 7
67 #define NAL_PPS                 8
68 #define NAL_SEI                 6
69
70 #define SLICE_TYPE_P            0
71 #define SLICE_TYPE_B            1
72 #define SLICE_TYPE_I            2
73 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
74 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
75 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
76
77
78 #define ENTROPY_MODE_CAVLC      0
79 #define ENTROPY_MODE_CABAC      1
80
81 #define PROFILE_IDC_BASELINE    66
82 #define PROFILE_IDC_MAIN        77
83 #define PROFILE_IDC_HIGH        100
84    
85 #define BITSTREAM_ALLOCATE_STEPPING     4096
86 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
87 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
88 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
89
90 static constexpr unsigned int MaxFrameNum = (2<<16);
91 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
92 static constexpr unsigned int Log2MaxFrameNum = 16;
93 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
94 static constexpr int rc_default_modes[] = {  // Priority list of modes.
95     VA_RC_VBR,
96     VA_RC_CQP,
97     VA_RC_VBR_CONSTRAINED,
98     VA_RC_CBR,
99     VA_RC_VCM,
100     VA_RC_NONE,
101 };
102
103 /* thread to save coded data */
104 #define SRC_SURFACE_FREE        0
105 #define SRC_SURFACE_IN_ENCODING 1
106     
107 struct __bitstream {
108     unsigned int *buffer;
109     int bit_offset;
110     int max_size_in_dword;
111 };
112 typedef struct __bitstream bitstream;
113
114 using namespace std;
115
116 class H264EncoderImpl {
117 public:
118         H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
119         ~H264EncoderImpl();
120         void add_audio(int64_t pts, vector<float> audio);
121         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
122         void end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames);
123         void shutdown();
124
125 private:
126         struct storage_task {
127                 unsigned long long display_order;
128                 int frame_type;
129                 vector<float> audio;
130                 int64_t pts, dts;
131         };
132         struct PendingFrame {
133                 RefCountedGLsync fence;
134                 vector<RefCountedFrame> input_frames;
135                 int64_t pts;
136         };
137
138         void encode_thread_func();
139         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
140         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
141                           int frame_type, int64_t pts, int64_t dts);
142         void storage_task_thread();
143         void storage_task_enqueue(storage_task task);
144         void save_codeddata(storage_task task);
145         int render_packedsequence();
146         int render_packedpicture();
147         void render_packedslice();
148         int render_sequence();
149         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
150         void sps_rbsp(bitstream *bs);
151         void pps_rbsp(bitstream *bs);
152         int build_packed_pic_buffer(unsigned char **header_buffer);
153         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
154         void slice_header(bitstream *bs);
155         int build_packed_seq_buffer(unsigned char **header_buffer);
156         int build_packed_slice_buffer(unsigned char **header_buffer);
157         int init_va(const string &va_display);
158         int deinit_va();
159         VADisplay va_open_display(const string &va_display);
160         void va_close_display(VADisplay va_dpy);
161         int setup_encode();
162         int release_encode();
163         void update_ReferenceFrames(int frame_type);
164         int update_RefPicList(int frame_type);
165
166         bool is_shutdown = false;
167         bool use_zerocopy;
168         int drm_fd = -1;
169
170         thread encode_thread, storage_thread;
171
172         mutex storage_task_queue_mutex;
173         condition_variable storage_task_queue_changed;
174         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
175         queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
176         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
177
178         mutex frame_queue_mutex;
179         condition_variable frame_queue_nonempty;
180         bool encode_thread_should_quit = false;  // under frame_queue_mutex
181
182         int current_storage_frame;
183
184         map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
185         map<int64_t, vector<float>> pending_audio_frames;  // under frame_queue_mutex
186         QSurface *surface;
187
188         AVCodecContext *context_audio;
189         AVFrame *audio_frame = nullptr;
190         HTTPD *httpd;
191
192         Display *x11_display = nullptr;
193
194         // Encoder parameters
195         VADisplay va_dpy;
196         VAProfile h264_profile = (VAProfile)~0;
197         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
198         int config_attrib_num = 0, enc_packed_header_idx;
199
200         struct GLSurface {
201                 VASurfaceID src_surface, ref_surface;
202                 VABufferID coded_buf;
203
204                 VAImage surface_image;
205                 GLuint y_tex, cbcr_tex;
206
207                 // Only if use_zerocopy == true.
208                 EGLImage y_egl_image, cbcr_egl_image;
209
210                 // Only if use_zerocopy == false.
211                 GLuint pbo;
212                 uint8_t *y_ptr, *cbcr_ptr;
213                 size_t y_offset, cbcr_offset;
214         };
215         GLSurface gl_surfaces[SURFACE_NUM];
216
217         VAConfigID config_id;
218         VAContextID context_id;
219         VAEncSequenceParameterBufferH264 seq_param;
220         VAEncPictureParameterBufferH264 pic_param;
221         VAEncSliceParameterBufferH264 slice_param;
222         VAPictureH264 CurrentCurrPic;
223         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
224
225         // Static quality settings.
226         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
227         static constexpr unsigned int num_ref_frames = 2;
228         static constexpr int initial_qp = 15;
229         static constexpr int minimal_qp = 0;
230         static constexpr int intra_period = 30;
231         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
232
233         // Quality settings that are meant to be static, but might be overridden
234         // by the profile.
235         int constraint_set_flag = 0;
236         int h264_packedheader = 0; /* support pack header? */
237         int h264_maxref = (1<<16|1);
238         int h264_entropy_mode = 1; /* cabac */
239         int ip_period = 3;
240
241         int rc_mode = -1;
242         unsigned int current_frame_num = 0;
243         unsigned int numShortTerm = 0;
244
245         int frame_width;
246         int frame_height;
247         int frame_width_mbaligned;
248         int frame_height_mbaligned;
249 };
250
251 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
252 // but if we don't delete it here, we get leaks. The GStreamer implementation
253 // does the same.
254 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
255 {
256     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
257     CHECK_VASTATUS(va_status, "vaRenderPicture");
258
259     for (int i = 0; i < num_buffers; ++i) {
260         va_status = vaDestroyBuffer(dpy, buffers[i]);
261         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
262     }
263 }
264
265 static unsigned int 
266 va_swap32(unsigned int val)
267 {
268     unsigned char *pval = (unsigned char *)&val;
269
270     return ((pval[0] << 24)     |
271             (pval[1] << 16)     |
272             (pval[2] << 8)      |
273             (pval[3] << 0));
274 }
275
276 static void
277 bitstream_start(bitstream *bs)
278 {
279     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
280     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
281     bs->bit_offset = 0;
282 }
283
284 static void
285 bitstream_end(bitstream *bs)
286 {
287     int pos = (bs->bit_offset >> 5);
288     int bit_offset = (bs->bit_offset & 0x1f);
289     int bit_left = 32 - bit_offset;
290
291     if (bit_offset) {
292         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
293     }
294 }
295  
296 static void
297 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
298 {
299     int pos = (bs->bit_offset >> 5);
300     int bit_offset = (bs->bit_offset & 0x1f);
301     int bit_left = 32 - bit_offset;
302
303     if (!size_in_bits)
304         return;
305
306     bs->bit_offset += size_in_bits;
307
308     if (bit_left > size_in_bits) {
309         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
310     } else {
311         size_in_bits -= bit_left;
312         if (bit_left >= 32) {
313             bs->buffer[pos] = (val >> size_in_bits);
314         } else {
315             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
316         }
317         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
318
319         if (pos + 1 == bs->max_size_in_dword) {
320             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
321             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
322         }
323
324         bs->buffer[pos + 1] = val;
325     }
326 }
327
328 static void
329 bitstream_put_ue(bitstream *bs, unsigned int val)
330 {
331     int size_in_bits = 0;
332     int tmp_val = ++val;
333
334     while (tmp_val) {
335         tmp_val >>= 1;
336         size_in_bits++;
337     }
338
339     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
340     bitstream_put_ui(bs, val, size_in_bits);
341 }
342
343 static void
344 bitstream_put_se(bitstream *bs, int val)
345 {
346     unsigned int new_val;
347
348     if (val <= 0)
349         new_val = -2 * val;
350     else
351         new_val = 2 * val - 1;
352
353     bitstream_put_ue(bs, new_val);
354 }
355
356 static void
357 bitstream_byte_aligning(bitstream *bs, int bit)
358 {
359     int bit_offset = (bs->bit_offset & 0x7);
360     int bit_left = 8 - bit_offset;
361     int new_val;
362
363     if (!bit_offset)
364         return;
365
366     assert(bit == 0 || bit == 1);
367
368     if (bit)
369         new_val = (1 << bit_left) - 1;
370     else
371         new_val = 0;
372
373     bitstream_put_ui(bs, new_val, bit_left);
374 }
375
376 static void 
377 rbsp_trailing_bits(bitstream *bs)
378 {
379     bitstream_put_ui(bs, 1, 1);
380     bitstream_byte_aligning(bs, 0);
381 }
382
383 static void nal_start_code_prefix(bitstream *bs)
384 {
385     bitstream_put_ui(bs, 0x00000001, 32);
386 }
387
388 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
389 {
390     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
391     bitstream_put_ui(bs, nal_ref_idc, 2);
392     bitstream_put_ui(bs, nal_unit_type, 5);
393 }
394
395 void H264EncoderImpl::sps_rbsp(bitstream *bs)
396 {
397     int profile_idc = PROFILE_IDC_BASELINE;
398
399     if (h264_profile  == VAProfileH264High)
400         profile_idc = PROFILE_IDC_HIGH;
401     else if (h264_profile  == VAProfileH264Main)
402         profile_idc = PROFILE_IDC_MAIN;
403
404     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
405     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
406     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
407     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
408     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
409     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
410     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
411     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
412
413     if ( profile_idc == PROFILE_IDC_HIGH) {
414         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */ 
415         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
416         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
417         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
418         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
419     }
420
421     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
422     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
423
424     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
425         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
426     else {
427         assert(0);
428     }
429
430     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
431     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
432
433     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
434     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
435     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
436
437     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
438         assert(0);
439     }
440
441     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
442     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
443
444     if (seq_param.frame_cropping_flag) {
445         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
446         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
447         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
448         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
449     }
450     
451     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
452     if ( false ) {
453         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
454     } else {
455         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
456         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
457         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
458         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
459         {
460             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
461             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
462             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
463             {
464                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
465                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
466                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
467             }
468         }
469         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
470         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
471         {
472             bitstream_put_ui(bs, 1, 32);  // FPS
473             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
474             bitstream_put_ui(bs, 1, 1);
475         }
476         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
477         {
478             // hrd_parameters 
479             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
480             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
481             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
482            
483             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
484             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
485             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
486
487             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
488             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
489             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
490             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
491         }
492         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
493         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */ 
494
495         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
496         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
497     }
498
499     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
500 }
501
502
503 void H264EncoderImpl::pps_rbsp(bitstream *bs)
504 {
505     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
506     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
507
508     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
509
510     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
511
512     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
513
514     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
515     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
516
517     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
518     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
519
520     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
521     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
522     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
523
524     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
525     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
526     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
527     
528     /* more_rbsp_data */
529     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
530     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
531     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
532
533     rbsp_trailing_bits(bs);
534 }
535
536 void H264EncoderImpl::slice_header(bitstream *bs)
537 {
538     int first_mb_in_slice = slice_param.macroblock_address;
539
540     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
541     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
542     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
543     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
544
545     /* frame_mbs_only_flag == 1 */
546     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
547         /* FIXME: */
548         assert(0);
549     }
550
551     if (pic_param.pic_fields.bits.idr_pic_flag)
552         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
553
554     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
555         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
556         /* pic_order_present_flag == 0 */
557     } else {
558         /* FIXME: */
559         assert(0);
560     }
561
562     /* redundant_pic_cnt_present_flag == 0 */
563     /* slice type */
564     if (IS_P_SLICE(slice_param.slice_type)) {
565         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
566
567         if (slice_param.num_ref_idx_active_override_flag)
568             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
569
570         /* ref_pic_list_reordering */
571         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
572     } else if (IS_B_SLICE(slice_param.slice_type)) {
573         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
574
575         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
576
577         if (slice_param.num_ref_idx_active_override_flag) {
578             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
579             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
580         }
581
582         /* ref_pic_list_reordering */
583         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
584         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
585     }
586
587     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
588          IS_P_SLICE(slice_param.slice_type)) ||
589         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
590          IS_B_SLICE(slice_param.slice_type))) {
591         /* FIXME: fill weight/offset table */
592         assert(0);
593     }
594
595     /* dec_ref_pic_marking */
596     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
597         unsigned char no_output_of_prior_pics_flag = 0;
598         unsigned char long_term_reference_flag = 0;
599         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
600
601         if (pic_param.pic_fields.bits.idr_pic_flag) {
602             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
603             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
604         } else {
605             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
606         }
607     }
608
609     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
610         !IS_I_SLICE(slice_param.slice_type))
611         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
612
613     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
614
615     /* ignore for SP/SI */
616
617     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
618         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
619
620         if (slice_param.disable_deblocking_filter_idc != 1) {
621             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
622             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
623         }
624     }
625
626     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
627         bitstream_byte_aligning(bs, 1);
628     }
629 }
630
631 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
632 {
633     bitstream bs;
634
635     bitstream_start(&bs);
636     nal_start_code_prefix(&bs);
637     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
638     pps_rbsp(&bs);
639     bitstream_end(&bs);
640
641     *header_buffer = (unsigned char *)bs.buffer;
642     return bs.bit_offset;
643 }
644
645 int
646 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
647 {
648     bitstream bs;
649
650     bitstream_start(&bs);
651     nal_start_code_prefix(&bs);
652     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
653     sps_rbsp(&bs);
654     bitstream_end(&bs);
655
656     *header_buffer = (unsigned char *)bs.buffer;
657     return bs.bit_offset;
658 }
659
660 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
661 {
662     bitstream bs;
663     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
664     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
665
666     bitstream_start(&bs);
667     nal_start_code_prefix(&bs);
668
669     if (IS_I_SLICE(slice_param.slice_type)) {
670         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
671     } else if (IS_P_SLICE(slice_param.slice_type)) {
672         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
673     } else {
674         assert(IS_B_SLICE(slice_param.slice_type));
675         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
676     }
677
678     slice_header(&bs);
679     bitstream_end(&bs);
680
681     *header_buffer = (unsigned char *)bs.buffer;
682     return bs.bit_offset;
683 }
684
685
686 /*
687   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
688   1) period between Frame #X and Frame #N = #X - #N
689   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
690   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
691   4) intra_period and intra_idr_period take precedence over ip_period
692   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods 
693      of I/IDR frames, see bellow examples
694   -------------------------------------------------------------------
695   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
696   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
697   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
698   1            0                ignored    IDRIIIIIII...      (No IDR any more)
699   1            1                ignored    IDR IDR IDR IDR...
700   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
701   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
702   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
703                                               (PBB)(IBB)(PBB)(IBB)... 
704   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
705                                            IDRPPPPP IPPPPP IPPPPP...
706   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
707                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
708                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
709                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
710                                            {IDR(PBB)(PBB)}                     (6/6/3)
711                                            {IDR(PBB)(PBB)}.
712 */
713
714 // General pts/dts strategy:
715 //
716 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
717 // bit tricky. We assume first of all that the frame rate never goes _above_
718 // MAX_FPS, which gives us a frame period N. The decoder can always decode
719 // in at least this speed, as long at dts <= pts (the frame is not attempted
720 // presented before it is decoded). Furthermore, we never have longer chains of
721 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
722 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
723 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
724 // frame, with an order O <= C.)
725 //
726 // Many strategies are possible, but we establish these rules:
727 //
728 //  - Tip frames have dts = pts - (C-O)*N.
729 //  - Non-tip frames have dts = dts_last + N.
730 //
731 // An example, with C=2 and N=10 and the data flow showed with arrows:
732 //
733 //        I  B  P  B  B  P
734 //   pts: 30 40 50 60 70 80
735 //        ↓  ↓     ↓
736 //   dts: 10 30 20 60 50←40
737 //         |  |  ↑        ↑
738 //         `--|--'        |
739 //             `----------'
740 //
741 // To show that this works fine also with irregular spacings, let's say that
742 // the third frame is delayed a bit (something earlier was dropped). Now the
743 // situation looks like this:
744 //
745 //        I  B  P  B  B   P
746 //   pts: 30 40 80 90 100 110
747 //        ↓  ↓     ↓
748 //   dts: 10 30 20 90 50←40
749 //         |  |  ↑        ↑
750 //         `--|--'        |
751 //             `----------'
752 //
753 // The resetting on every tip frame makes sure dts never ends up lagging a lot
754 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
755 //
756 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
757 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
758 // a tip frame and should be given a dts based on the previous one.
759 #define FRAME_P 0
760 #define FRAME_B 1
761 #define FRAME_I 2
762 #define FRAME_IDR 7
763 void encoding2display_order(
764     int encoding_order, int intra_period,
765     int intra_idr_period, int ip_period,
766     int *displaying_order,
767     int *frame_type, int *pts_lag)
768 {
769     int encoding_order_gop = 0;
770
771     *pts_lag = 0;
772
773     if (intra_period == 1) { /* all are I/IDR frames */
774         *displaying_order = encoding_order;
775         if (intra_idr_period == 0)
776             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
777         else
778             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
779         return;
780     }
781
782     if (intra_period == 0)
783         intra_idr_period = 0;
784
785     if (ip_period == 1) {
786         // No B-frames, sequence is like IDR PPPPP IPPPPP.
787         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
788         *displaying_order = encoding_order;
789
790         if (encoding_order_gop == 0) { /* the first frame */
791             *frame_type = FRAME_IDR;
792         } else if (intra_period != 0 && /* have I frames */
793                    encoding_order_gop >= 2 &&
794                    (encoding_order_gop % intra_period == 0)) {
795             *frame_type = FRAME_I;
796         } else {
797             *frame_type = FRAME_P;
798         }
799         return;
800     } 
801
802     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
803     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
804     *pts_lag = -1;  // Most frames are not tip frames.
805          
806     if (encoding_order_gop == 0) { /* the first frame */
807         *frame_type = FRAME_IDR;
808         *displaying_order = encoding_order;
809         // IDR frames are a special case; I honestly can't find the logic behind
810         // why this is the right thing, but it seems to line up nicely in practice :-)
811         *pts_lag = TIMEBASE / MAX_FPS;
812     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
813         *frame_type = FRAME_B;
814         *displaying_order = encoding_order - 1;
815         if ((encoding_order_gop % ip_period) == 0) {
816             *pts_lag = 0;  // Last B-frame.
817         }
818     } else if (intra_period != 0 && /* have I frames */
819                encoding_order_gop >= 2 &&
820                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
821         *frame_type = FRAME_I;
822         *displaying_order = encoding_order + ip_period - 1;
823     } else {
824         *frame_type = FRAME_P;
825         *displaying_order = encoding_order + ip_period - 1;
826     }
827 }
828
829
830 static const char *rc_to_string(int rc_mode)
831 {
832     switch (rc_mode) {
833     case VA_RC_NONE:
834         return "NONE";
835     case VA_RC_CBR:
836         return "CBR";
837     case VA_RC_VBR:
838         return "VBR";
839     case VA_RC_VCM:
840         return "VCM";
841     case VA_RC_CQP:
842         return "CQP";
843     case VA_RC_VBR_CONSTRAINED:
844         return "VBR_CONSTRAINED";
845     default:
846         return "Unknown";
847     }
848 }
849
850 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
851 {
852         if (va_display.empty()) {
853                 x11_display = XOpenDisplay(NULL);
854                 if (!x11_display) {
855                         fprintf(stderr, "error: can't connect to X server!\n");
856                         return NULL;
857                 }
858                 use_zerocopy = true;
859                 return vaGetDisplay(x11_display);
860         } else if (va_display[0] != '/') {
861                 x11_display = XOpenDisplay(va_display.c_str());
862                 if (!x11_display) {
863                         fprintf(stderr, "error: can't connect to X server!\n");
864                         return NULL;
865                 }
866                 use_zerocopy = true;
867                 return vaGetDisplay(x11_display);
868         } else {
869                 drm_fd = open(va_display.c_str(), O_RDWR);
870                 if (drm_fd == -1) {
871                         perror(va_display.c_str());
872                         return NULL;
873                 }
874                 use_zerocopy = false;
875                 return vaGetDisplayDRM(drm_fd);
876         }
877 }
878
879 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
880 {
881         if (x11_display) {
882                 XCloseDisplay(x11_display);
883                 x11_display = nullptr;
884         }
885         if (drm_fd != -1) {
886                 close(drm_fd);
887         }
888 }
889
890 int H264EncoderImpl::init_va(const string &va_display)
891 {
892     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
893     VAEntrypoint *entrypoints;
894     int num_entrypoints, slice_entrypoint;
895     int support_encode = 0;    
896     int major_ver, minor_ver;
897     VAStatus va_status;
898     unsigned int i;
899
900     va_dpy = va_open_display(va_display);
901     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
902     CHECK_VASTATUS(va_status, "vaInitialize");
903
904     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
905     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
906     if (!entrypoints) {
907         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
908         exit(1);
909     }
910
911     /* use the highest profile */
912     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
913         if ((h264_profile != ~0) && h264_profile != profile_list[i])
914             continue;
915         
916         h264_profile = profile_list[i];
917         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
918         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
919             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
920                 support_encode = 1;
921                 break;
922             }
923         }
924         if (support_encode == 1)
925             break;
926     }
927     
928     if (support_encode == 0) {
929         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
930         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
931         printf("to use VA-API against DRM instead of X11.\n");
932         exit(1);
933     } else {
934         switch (h264_profile) {
935             case VAProfileH264Baseline:
936                 ip_period = 1;
937                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
938                 h264_entropy_mode = 0;
939                 break;
940             case VAProfileH264ConstrainedBaseline:
941                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
942                 ip_period = 1;
943                 break;
944
945             case VAProfileH264Main:
946                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
947                 break;
948
949             case VAProfileH264High:
950                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
951                 break;
952             default:
953                 h264_profile = VAProfileH264Baseline;
954                 ip_period = 1;
955                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
956                 break;
957         }
958     }
959
960     VAConfigAttrib attrib[VAConfigAttribTypeMax];
961
962     /* find out the format for the render target, and rate control mode */
963     for (i = 0; i < VAConfigAttribTypeMax; i++)
964         attrib[i].type = (VAConfigAttribType)i;
965
966     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
967                                       &attrib[0], VAConfigAttribTypeMax);
968     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
969     /* check the interested configattrib */
970     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
971         printf("Not find desired YUV420 RT format\n");
972         exit(1);
973     } else {
974         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
975         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
976         config_attrib_num++;
977     }
978     
979     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
980         int tmp = attrib[VAConfigAttribRateControl].value;
981
982         if (rc_mode == -1 || !(rc_mode & tmp))  {
983             if (rc_mode != -1) {
984                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
985             }
986
987             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
988                 if (rc_default_modes[i] & tmp) {
989                     rc_mode = rc_default_modes[i];
990                     break;
991                 }
992             }
993         }
994
995         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
996         config_attrib[config_attrib_num].value = rc_mode;
997         config_attrib_num++;
998     }
999     
1000
1001     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1002         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1003
1004         h264_packedheader = 1;
1005         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1006         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1007         
1008         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1009             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1010         }
1011         
1012         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1013             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1014         }
1015         
1016         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1017             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1018         }
1019         
1020         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1021             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1022         }
1023         
1024         enc_packed_header_idx = config_attrib_num;
1025         config_attrib_num++;
1026     }
1027
1028     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1029         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1030         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1031         config_attrib_num++;
1032     }
1033     
1034     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1035         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1036     }
1037
1038     free(entrypoints);
1039     return 0;
1040 }
1041
1042 int H264EncoderImpl::setup_encode()
1043 {
1044     VAStatus va_status;
1045     VASurfaceID *tmp_surfaceid;
1046     int codedbuf_size, i;
1047     static VASurfaceID src_surface[SURFACE_NUM];
1048     static VASurfaceID ref_surface[SURFACE_NUM];
1049     
1050     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1051             &config_attrib[0], config_attrib_num, &config_id);
1052     CHECK_VASTATUS(va_status, "vaCreateConfig");
1053
1054     /* create source surfaces */
1055     va_status = vaCreateSurfaces(va_dpy,
1056                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1057                                  &src_surface[0], SURFACE_NUM,
1058                                  NULL, 0);
1059     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1060
1061     /* create reference surfaces */
1062     va_status = vaCreateSurfaces(va_dpy,
1063                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1064                                  &ref_surface[0], SURFACE_NUM,
1065                                  NULL, 0);
1066     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1067
1068     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1069     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1070     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1071     
1072     /* Create a context for this encode pipe */
1073     va_status = vaCreateContext(va_dpy, config_id,
1074                                 frame_width_mbaligned, frame_height_mbaligned,
1075                                 VA_PROGRESSIVE,
1076                                 tmp_surfaceid, 2 * SURFACE_NUM,
1077                                 &context_id);
1078     CHECK_VASTATUS(va_status, "vaCreateContext");
1079     free(tmp_surfaceid);
1080
1081     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1082
1083     for (i = 0; i < SURFACE_NUM; i++) {
1084         /* create coded buffer once for all
1085          * other VA buffers which won't be used again after vaRenderPicture.
1086          * so APP can always vaCreateBuffer for every frame
1087          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1088          * so VA won't maintain the coded buffer
1089          */
1090         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1091                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1092         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1093     }
1094
1095     /* create OpenGL objects */
1096     //glGenFramebuffers(SURFACE_NUM, fbos);
1097     
1098     for (i = 0; i < SURFACE_NUM; i++) {
1099         glGenTextures(1, &gl_surfaces[i].y_tex);
1100         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1101
1102         if (!use_zerocopy) {
1103             // Create Y image.
1104             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1105             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1106
1107             // Create CbCr image.
1108             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1109             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1110
1111             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1112             // buffers, due to potentially differing pitch.
1113             glGenBuffers(1, &gl_surfaces[i].pbo);
1114             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1115             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1116             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1117             gl_surfaces[i].y_offset = 0;
1118             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1119             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1120             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1121             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1122         }
1123     }
1124
1125     for (i = 0; i < SURFACE_NUM; i++) {
1126         gl_surfaces[i].src_surface = src_surface[i];
1127         gl_surfaces[i].ref_surface = ref_surface[i];
1128     }
1129     
1130     return 0;
1131 }
1132
1133 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1134 //
1135 //   2 1 0 [3] 4 8 9
1136 template<class T, class C>
1137 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1138 {
1139         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1140         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1141         sort(middle, end, less_than);
1142 }
1143
1144 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1145 {
1146     int i;
1147     
1148     if (frame_type == FRAME_B)
1149         return;
1150
1151     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1152     numShortTerm++;
1153     if (numShortTerm > num_ref_frames)
1154         numShortTerm = num_ref_frames;
1155     for (i=numShortTerm-1; i>0; i--)
1156         ReferenceFrames[i] = ReferenceFrames[i-1];
1157     ReferenceFrames[0] = CurrentCurrPic;
1158     
1159     current_frame_num++;
1160     if (current_frame_num > MaxFrameNum)
1161         current_frame_num = 0;
1162 }
1163
1164
1165 int H264EncoderImpl::update_RefPicList(int frame_type)
1166 {
1167     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1168         return a.frame_idx > b.frame_idx;
1169     };
1170     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1171         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1172     };
1173     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1174         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1175     };
1176     
1177     if (frame_type == FRAME_P) {
1178         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1179         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1180     } else if (frame_type == FRAME_B) {
1181         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1182         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1183
1184         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1185         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1186     }
1187     
1188     return 0;
1189 }
1190
1191
1192 int H264EncoderImpl::render_sequence()
1193 {
1194     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1195     VAStatus va_status;
1196     VAEncMiscParameterBuffer *misc_param;
1197     VAEncMiscParameterRateControl *misc_rate_ctrl;
1198     
1199     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1200     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1201     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1202     seq_param.bits_per_second = frame_bitrate;
1203
1204     seq_param.intra_period = intra_period;
1205     seq_param.intra_idr_period = intra_idr_period;
1206     seq_param.ip_period = ip_period;
1207
1208     seq_param.max_num_ref_frames = num_ref_frames;
1209     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1210     seq_param.time_scale = TIMEBASE * 2;
1211     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1212     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1213     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1214     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1215     seq_param.seq_fields.bits.chroma_format_idc = 1;
1216     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1217     
1218     if (frame_width != frame_width_mbaligned ||
1219         frame_height != frame_height_mbaligned) {
1220         seq_param.frame_cropping_flag = 1;
1221         seq_param.frame_crop_left_offset = 0;
1222         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1223         seq_param.frame_crop_top_offset = 0;
1224         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1225     }
1226     
1227     va_status = vaCreateBuffer(va_dpy, context_id,
1228                                VAEncSequenceParameterBufferType,
1229                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1230     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1231     
1232     va_status = vaCreateBuffer(va_dpy, context_id,
1233                                VAEncMiscParameterBufferType,
1234                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1235                                1, NULL, &rc_param_buf);
1236     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1237     
1238     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1239     misc_param->type = VAEncMiscParameterTypeRateControl;
1240     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1241     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1242     misc_rate_ctrl->bits_per_second = frame_bitrate;
1243     misc_rate_ctrl->target_percentage = 66;
1244     misc_rate_ctrl->window_size = 1000;
1245     misc_rate_ctrl->initial_qp = initial_qp;
1246     misc_rate_ctrl->min_qp = minimal_qp;
1247     misc_rate_ctrl->basic_unit_size = 0;
1248     vaUnmapBuffer(va_dpy, rc_param_buf);
1249
1250     render_id[0] = seq_param_buf;
1251     render_id[1] = rc_param_buf;
1252     
1253     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1254     
1255     return 0;
1256 }
1257
1258 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1259 {
1260     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1261     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1262     int PicOrderCntMsb, TopFieldOrderCnt;
1263     
1264     if (frame_type == FRAME_IDR)
1265         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1266     else {
1267         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1268         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1269     }
1270     
1271     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1272         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1273         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1274     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1275              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1276         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1277     else
1278         PicOrderCntMsb = prevPicOrderCntMsb;
1279     
1280     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1281
1282     if (frame_type != FRAME_B) {
1283         PicOrderCntMsb_ref = PicOrderCntMsb;
1284         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1285     }
1286     
1287     return TopFieldOrderCnt;
1288 }
1289
1290 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1291 {
1292     VABufferID pic_param_buf;
1293     VAStatus va_status;
1294     int i = 0;
1295
1296     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1297     pic_param.CurrPic.frame_idx = current_frame_num;
1298     pic_param.CurrPic.flags = 0;
1299     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1300     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1301     CurrentCurrPic = pic_param.CurrPic;
1302
1303     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1304     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1305         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1306         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1307     }
1308     
1309     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1310     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1311     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1312     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1313     pic_param.frame_num = current_frame_num;
1314     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1315     pic_param.last_picture = false;  // FIXME
1316     pic_param.pic_init_qp = initial_qp;
1317
1318     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1319                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1320     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1321
1322     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1323
1324     return 0;
1325 }
1326
1327 int H264EncoderImpl::render_packedsequence()
1328 {
1329     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1330     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1331     unsigned int length_in_bits;
1332     unsigned char *packedseq_buffer = NULL;
1333     VAStatus va_status;
1334
1335     length_in_bits = build_packed_seq_buffer(&packedseq_buffer); 
1336     
1337     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1338     
1339     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1340     packedheader_param_buffer.has_emulation_bytes = 0;
1341     va_status = vaCreateBuffer(va_dpy,
1342                                context_id,
1343                                VAEncPackedHeaderParameterBufferType,
1344                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1345                                &packedseq_para_bufid);
1346     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1347
1348     va_status = vaCreateBuffer(va_dpy,
1349                                context_id,
1350                                VAEncPackedHeaderDataBufferType,
1351                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1352                                &packedseq_data_bufid);
1353     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1354
1355     render_id[0] = packedseq_para_bufid;
1356     render_id[1] = packedseq_data_bufid;
1357     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1358
1359     free(packedseq_buffer);
1360     
1361     return 0;
1362 }
1363
1364
1365 int H264EncoderImpl::render_packedpicture()
1366 {
1367     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1368     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1369     unsigned int length_in_bits;
1370     unsigned char *packedpic_buffer = NULL;
1371     VAStatus va_status;
1372
1373     length_in_bits = build_packed_pic_buffer(&packedpic_buffer); 
1374     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1375     packedheader_param_buffer.bit_length = length_in_bits;
1376     packedheader_param_buffer.has_emulation_bytes = 0;
1377
1378     va_status = vaCreateBuffer(va_dpy,
1379                                context_id,
1380                                VAEncPackedHeaderParameterBufferType,
1381                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1382                                &packedpic_para_bufid);
1383     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1384
1385     va_status = vaCreateBuffer(va_dpy,
1386                                context_id,
1387                                VAEncPackedHeaderDataBufferType,
1388                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1389                                &packedpic_data_bufid);
1390     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1391
1392     render_id[0] = packedpic_para_bufid;
1393     render_id[1] = packedpic_data_bufid;
1394     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1395
1396     free(packedpic_buffer);
1397     
1398     return 0;
1399 }
1400
1401 void H264EncoderImpl::render_packedslice()
1402 {
1403     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1404     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1405     unsigned int length_in_bits;
1406     unsigned char *packedslice_buffer = NULL;
1407     VAStatus va_status;
1408
1409     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1410     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1411     packedheader_param_buffer.bit_length = length_in_bits;
1412     packedheader_param_buffer.has_emulation_bytes = 0;
1413
1414     va_status = vaCreateBuffer(va_dpy,
1415                                context_id,
1416                                VAEncPackedHeaderParameterBufferType,
1417                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1418                                &packedslice_para_bufid);
1419     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1420
1421     va_status = vaCreateBuffer(va_dpy,
1422                                context_id,
1423                                VAEncPackedHeaderDataBufferType,
1424                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1425                                &packedslice_data_bufid);
1426     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1427
1428     render_id[0] = packedslice_para_bufid;
1429     render_id[1] = packedslice_data_bufid;
1430     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1431
1432     free(packedslice_buffer);
1433 }
1434
1435 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1436 {
1437     VABufferID slice_param_buf;
1438     VAStatus va_status;
1439     int i;
1440
1441     update_RefPicList(frame_type);
1442     
1443     /* one frame, one slice */
1444     slice_param.macroblock_address = 0;
1445     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1446     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1447     if (frame_type == FRAME_IDR) {
1448         if (encoding_frame_num != 0)
1449             ++slice_param.idr_pic_id;
1450     } else if (frame_type == FRAME_P) {
1451         int refpiclist0_max = h264_maxref & 0xffff;
1452         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1453
1454         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1455             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1456             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1457         }
1458     } else if (frame_type == FRAME_B) {
1459         int refpiclist0_max = h264_maxref & 0xffff;
1460         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1461
1462         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1463         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1464             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1465             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1466         }
1467
1468         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1469         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1470             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1471             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1472         }
1473     }
1474
1475     slice_param.slice_alpha_c0_offset_div2 = 0;
1476     slice_param.slice_beta_offset_div2 = 0;
1477     slice_param.direct_spatial_mv_pred_flag = 1;
1478     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1479     
1480
1481     if (h264_packedheader &&
1482         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1483         render_packedslice();
1484
1485     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1486                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1487     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1488
1489     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1490
1491     return 0;
1492 }
1493
1494
1495
1496 void H264EncoderImpl::save_codeddata(storage_task task)
1497 {    
1498     VACodedBufferSegment *buf_list = NULL;
1499     VAStatus va_status;
1500
1501     string data;
1502
1503     const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1504
1505     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1506     CHECK_VASTATUS(va_status, "vaMapBuffer");
1507     while (buf_list != NULL) {
1508         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1509         buf_list = (VACodedBufferSegment *) buf_list->next;
1510     }
1511     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1512
1513     {
1514         // Add video.
1515         AVPacket pkt;
1516         memset(&pkt, 0, sizeof(pkt));
1517         pkt.buf = nullptr;
1518         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1519         pkt.size = data.size();
1520         pkt.stream_index = 0;
1521         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1522             pkt.flags = AV_PKT_FLAG_KEY;
1523         } else {
1524             pkt.flags = 0;
1525         }
1526         //pkt.duration = 1;
1527         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1528     }
1529     // Encode and add all audio frames up to and including the pts of this video frame.
1530     for ( ;; ) {
1531         int64_t audio_pts;
1532         vector<float> audio;
1533         {
1534              unique_lock<mutex> lock(frame_queue_mutex);
1535              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1536              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1537              auto it = pending_audio_frames.begin();
1538              if (it->first > task.pts) break;
1539              audio_pts = it->first;
1540              audio = move(it->second);
1541              pending_audio_frames.erase(it); 
1542         }
1543
1544         audio_frame->nb_samples = audio.size() / 2;
1545         audio_frame->format = AV_SAMPLE_FMT_S32;
1546         audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
1547
1548         unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
1549         int ret = avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
1550         if (ret < 0) {
1551             fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
1552             exit(1);
1553         }
1554         for (int i = 0; i < audio_frame->nb_samples * 2; ++i) {
1555             if (audio[i] >= 1.0f) {
1556                 int_samples[i] = 2147483647;
1557             } else if (audio[i] <= -1.0f) {
1558                 int_samples[i] = -2147483647;
1559             } else {
1560                 int_samples[i] = lrintf(audio[i] * 2147483647.0f);
1561             }
1562         }
1563
1564         AVPacket pkt;
1565         av_init_packet(&pkt);
1566         pkt.data = nullptr;
1567         pkt.size = 0;
1568         int got_output;
1569         avcodec_encode_audio2(context_audio, &pkt, audio_frame, &got_output);
1570         if (got_output) {
1571             pkt.stream_index = 1;
1572             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
1573         }
1574         // TODO: Delayed frames.
1575         av_frame_unref(audio_frame);
1576         av_free_packet(&pkt);
1577         if (audio_pts == task.pts) break;
1578     }
1579 }
1580
1581
1582 // this is weird. but it seems to put a new frame onto the queue
1583 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1584 {
1585         unique_lock<mutex> lock(storage_task_queue_mutex);
1586         storage_task_queue.push(move(task));
1587         storage_task_queue_changed.notify_all();
1588 }
1589
1590 void H264EncoderImpl::storage_task_thread()
1591 {
1592         for ( ;; ) {
1593                 storage_task current;
1594                 {
1595                         // wait until there's an encoded frame  
1596                         unique_lock<mutex> lock(storage_task_queue_mutex);
1597                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1598                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1599                         current = move(storage_task_queue.front());
1600                         storage_task_queue.pop();
1601                 }
1602
1603                 VAStatus va_status;
1604            
1605                 // waits for data, then saves it to disk.
1606                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1607                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1608                 save_codeddata(move(current));
1609
1610                 {
1611                         unique_lock<mutex> lock(storage_task_queue_mutex);
1612                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1613                         storage_task_queue_changed.notify_all();
1614                 }
1615         }
1616 }
1617
1618 int H264EncoderImpl::release_encode()
1619 {
1620         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1621                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1622                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1623                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1624
1625                 if (!use_zerocopy) {
1626                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1627                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1628                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1629                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1630                 }
1631                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1632                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1633         }
1634
1635         vaDestroyContext(va_dpy, context_id);
1636         vaDestroyConfig(va_dpy, config_id);
1637
1638         return 0;
1639 }
1640
1641 int H264EncoderImpl::deinit_va()
1642
1643     vaTerminate(va_dpy);
1644
1645     va_close_display(va_dpy);
1646
1647     return 0;
1648 }
1649
1650
1651 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1652         : current_storage_frame(0), surface(surface), httpd(httpd)
1653 {
1654         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1655         context_audio = avcodec_alloc_context3(codec_audio);
1656         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1657         context_audio->sample_rate = OUTPUT_FREQUENCY;
1658         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1659         context_audio->channels = 2;
1660         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1661         context_audio->time_base = AVRational{1, TIMEBASE};
1662         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1663                 fprintf(stderr, "Could not open codec\n");
1664                 exit(1);
1665         }
1666         audio_frame = av_frame_alloc();
1667
1668         frame_width = width;
1669         frame_height = height;
1670         frame_width_mbaligned = (frame_width + 15) & (~15);
1671         frame_height_mbaligned = (frame_height + 15) & (~15);
1672
1673         //print_input();
1674
1675         init_va(va_display);
1676         setup_encode();
1677
1678         // No frames are ready yet.
1679         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1680             
1681         memset(&seq_param, 0, sizeof(seq_param));
1682         memset(&pic_param, 0, sizeof(pic_param));
1683         memset(&slice_param, 0, sizeof(slice_param));
1684
1685         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1686
1687         encode_thread = thread([this]{
1688                 //SDL_GL_MakeCurrent(window, context);
1689                 QOpenGLContext *context = create_context(this->surface);
1690                 eglBindAPI(EGL_OPENGL_API);
1691                 if (!make_current(context, this->surface)) {
1692                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1693                                 eglGetError());
1694                         exit(1);
1695                 }
1696                 encode_thread_func();
1697         });
1698 }
1699
1700 H264EncoderImpl::~H264EncoderImpl()
1701 {
1702         shutdown();
1703         av_frame_free(&audio_frame);
1704
1705         // TODO: Destroy context.
1706 }
1707
1708 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1709 {
1710         assert(!is_shutdown);
1711         {
1712                 // Wait until this frame slot is done encoding.
1713                 unique_lock<mutex> lock(storage_task_queue_mutex);
1714                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1715                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1716                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1717                 }
1718                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1719                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1720                 if (storage_thread_should_quit) return false;
1721         }
1722
1723         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1724         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1725         *y_tex = surf->y_tex;
1726         *cbcr_tex = surf->cbcr_tex;
1727
1728         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1729         CHECK_VASTATUS(va_status, "vaDeriveImage");
1730
1731         if (use_zerocopy) {
1732                 VABufferInfo buf_info;
1733                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1734                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1735                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1736
1737                 // Create Y image.
1738                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1739                 EGLint y_attribs[] = {
1740                         EGL_WIDTH, frame_width,
1741                         EGL_HEIGHT, frame_height,
1742                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1743                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1744                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1745                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1746                         EGL_NONE
1747                 };
1748
1749                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1750                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1751
1752                 // Associate Y image to a texture.
1753                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1754                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1755
1756                 // Create CbCr image.
1757                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1758                 EGLint cbcr_attribs[] = {
1759                         EGL_WIDTH, frame_width,
1760                         EGL_HEIGHT, frame_height,
1761                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1762                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1763                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1764                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1765                         EGL_NONE
1766                 };
1767
1768                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1769                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1770
1771                 // Associate CbCr image to a texture.
1772                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1773                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1774         }
1775
1776         return true;
1777 }
1778
1779 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1780 {
1781         assert(!is_shutdown);
1782         {
1783                 unique_lock<mutex> lock(frame_queue_mutex);
1784                 pending_audio_frames[pts] = move(audio);
1785         }
1786         frame_queue_nonempty.notify_all();
1787 }
1788
1789 void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1790 {
1791         assert(!is_shutdown);
1792
1793         if (!use_zerocopy) {
1794                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1795
1796                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1797                 check_error();
1798
1799                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1800                 check_error();
1801
1802                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1803                 check_error();
1804                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1805                 check_error();
1806
1807                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1808                 check_error();
1809                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1810                 check_error();
1811
1812                 glBindTexture(GL_TEXTURE_2D, 0);
1813                 check_error();
1814                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1815                 check_error();
1816
1817                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1818                 check_error();
1819                 fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1820                 check_error();
1821         }
1822
1823         {
1824                 unique_lock<mutex> lock(frame_queue_mutex);
1825                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1826                 ++current_storage_frame;
1827         }
1828         frame_queue_nonempty.notify_all();
1829 }
1830
1831 void H264EncoderImpl::shutdown()
1832 {
1833         if (is_shutdown) {
1834                 return;
1835         }
1836
1837         {
1838                 unique_lock<mutex> lock(frame_queue_mutex);
1839                 encode_thread_should_quit = true;
1840                 frame_queue_nonempty.notify_all();
1841         }
1842         encode_thread.join();
1843         {
1844                 unique_lock<mutex> lock(storage_task_queue_mutex);
1845                 storage_thread_should_quit = true;
1846                 frame_queue_nonempty.notify_all();
1847                 storage_task_queue_changed.notify_all();
1848         }
1849         storage_thread.join();
1850
1851         release_encode();
1852         deinit_va();
1853         is_shutdown = true;
1854 }
1855
1856 void H264EncoderImpl::encode_thread_func()
1857 {
1858         int64_t last_dts = -1;
1859         int gop_start_display_frame_num = 0;
1860         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1861                 PendingFrame frame;
1862                 int pts_lag;
1863                 int frame_type, display_frame_num;
1864                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1865                                        &display_frame_num, &frame_type, &pts_lag);
1866                 if (frame_type == FRAME_IDR) {
1867                         numShortTerm = 0;
1868                         current_frame_num = 0;
1869                         gop_start_display_frame_num = display_frame_num;
1870                 }
1871
1872                 {
1873                         unique_lock<mutex> lock(frame_queue_mutex);
1874                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1875                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1876                         });
1877                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1878                                 // We have queued frames that were supposed to be B-frames,
1879                                 // but will be no P-frame to encode them against. Encode them all
1880                                 // as P-frames instead. Note that this happens under the mutex,
1881                                 // but nobody else uses it at this point, since we're shutting down,
1882                                 // so there's no contention.
1883                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1884                                 return;
1885                         } else {
1886                                 frame = move(pending_video_frames[display_frame_num]);
1887                                 pending_video_frames.erase(display_frame_num);
1888                         }
1889                 }
1890
1891                 // Determine the dts of this frame.
1892                 int64_t dts;
1893                 if (pts_lag == -1) {
1894                         assert(last_dts != -1);
1895                         dts = last_dts + (TIMEBASE / MAX_FPS);
1896                 } else {
1897                         dts = frame.pts - pts_lag;
1898                 }
1899                 last_dts = dts;
1900
1901                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
1902         }
1903 }
1904
1905 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1906 {
1907         if (pending_video_frames.empty()) {
1908                 return;
1909         }
1910
1911         for (auto &pending_frame : pending_video_frames) {
1912                 int display_frame_num = pending_frame.first;
1913                 assert(display_frame_num > 0);
1914                 PendingFrame frame = move(pending_frame.second);
1915                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1916                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1917                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
1918                 last_dts = dts;
1919         }
1920 }
1921
1922 namespace {
1923
1924 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1925 {
1926         if (src_width == dst_pitch) {
1927                 memcpy(dst, src, src_width * height);
1928         } else {
1929                 for (size_t y = 0; y < height; ++y) {
1930                         const uint8_t *sptr = src + y * src_width;
1931                         uint8_t *dptr = dst + y * dst_pitch;
1932                         memcpy(dptr, sptr, src_width);
1933                 }
1934         }
1935 }
1936
1937 }  // namespace
1938
1939 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1940                                    int frame_type, int64_t pts, int64_t dts)
1941 {
1942         // Wait for the GPU to be done with the frame.
1943         GLenum sync_status;
1944         do {
1945                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1946                 check_error();
1947         } while (sync_status == GL_TIMEOUT_EXPIRED);
1948         assert(sync_status != GL_WAIT_FAILED);
1949
1950         // Release back any input frames we needed to render this frame.
1951         frame.input_frames.clear();
1952
1953         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1954         VAStatus va_status;
1955
1956         if (use_zerocopy) {
1957                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1958                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1959                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1960                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1961         } else {
1962                 unsigned char *surface_p = nullptr;
1963                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1964
1965                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1966                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1967
1968                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1969                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1970
1971                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1972                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1973         }
1974
1975         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1976         CHECK_VASTATUS(va_status, "vaDestroyImage");
1977
1978         // Schedule the frame for encoding.
1979         VASurfaceID va_surface = surf->src_surface;
1980         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1981         CHECK_VASTATUS(va_status, "vaBeginPicture");
1982
1983         if (frame_type == FRAME_IDR) {
1984                 render_sequence();
1985                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1986                 if (h264_packedheader) {
1987                         render_packedsequence();
1988                         render_packedpicture();
1989                 }
1990         } else {
1991                 //render_sequence();
1992                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1993         }
1994         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1995
1996         va_status = vaEndPicture(va_dpy, context_id);
1997         CHECK_VASTATUS(va_status, "vaEndPicture");
1998
1999         // so now the data is done encoding (well, async job kicked off)...
2000         // we send that to the storage thread
2001         storage_task tmp;
2002         tmp.display_order = display_frame_num;
2003         tmp.frame_type = frame_type;
2004         tmp.pts = pts;
2005         tmp.dts = dts;
2006         storage_task_enqueue(move(tmp));
2007
2008         update_ReferenceFrames(frame_type);
2009 }
2010
2011 // Proxy object.
2012 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2013         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2014
2015 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2016 H264Encoder::~H264Encoder() {}
2017
2018 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2019 {
2020         impl->add_audio(pts, audio);
2021 }
2022
2023 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2024 {
2025         return impl->begin_frame(y_tex, cbcr_tex);
2026 }
2027
2028 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
2029 {
2030         impl->end_frame(fence, pts, input_frames);
2031 }
2032
2033 void H264Encoder::shutdown()
2034 {
2035         impl->shutdown();
2036 }
2037
2038 // Real class.