]> git.sesse.net Git - nageru/blob - h264encode.cpp
Fix a memory leak in muxing.
[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         HTTPD *httpd;
190
191         Display *x11_display = nullptr;
192
193         // Encoder parameters
194         VADisplay va_dpy;
195         VAProfile h264_profile = (VAProfile)~0;
196         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
197         int config_attrib_num = 0, enc_packed_header_idx;
198
199         struct GLSurface {
200                 VASurfaceID src_surface, ref_surface;
201                 VABufferID coded_buf;
202
203                 VAImage surface_image;
204                 GLuint y_tex, cbcr_tex;
205
206                 // Only if use_zerocopy == true.
207                 EGLImage y_egl_image, cbcr_egl_image;
208
209                 // Only if use_zerocopy == false.
210                 GLuint pbo;
211                 uint8_t *y_ptr, *cbcr_ptr;
212                 size_t y_offset, cbcr_offset;
213         };
214         GLSurface gl_surfaces[SURFACE_NUM];
215
216         VAConfigID config_id;
217         VAContextID context_id;
218         VAEncSequenceParameterBufferH264 seq_param;
219         VAEncPictureParameterBufferH264 pic_param;
220         VAEncSliceParameterBufferH264 slice_param;
221         VAPictureH264 CurrentCurrPic;
222         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
223
224         // Static quality settings.
225         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
226         static constexpr unsigned int num_ref_frames = 2;
227         static constexpr int initial_qp = 15;
228         static constexpr int minimal_qp = 0;
229         static constexpr int intra_period = 30;
230         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
231
232         // Quality settings that are meant to be static, but might be overridden
233         // by the profile.
234         int constraint_set_flag = 0;
235         int h264_packedheader = 0; /* support pack header? */
236         int h264_maxref = (1<<16|1);
237         int h264_entropy_mode = 1; /* cabac */
238         int ip_period = 3;
239
240         int rc_mode = -1;
241         unsigned int current_frame_num = 0;
242         unsigned int numShortTerm = 0;
243
244         int frame_width;
245         int frame_height;
246         int frame_width_mbaligned;
247         int frame_height_mbaligned;
248 };
249
250 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
251 // but if we don't delete it here, we get leaks. The GStreamer implementation
252 // does the same.
253 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
254 {
255     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
256     CHECK_VASTATUS(va_status, "vaRenderPicture");
257
258     for (int i = 0; i < num_buffers; ++i) {
259         va_status = vaDestroyBuffer(dpy, buffers[i]);
260         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
261     }
262 }
263
264 static unsigned int 
265 va_swap32(unsigned int val)
266 {
267     unsigned char *pval = (unsigned char *)&val;
268
269     return ((pval[0] << 24)     |
270             (pval[1] << 16)     |
271             (pval[2] << 8)      |
272             (pval[3] << 0));
273 }
274
275 static void
276 bitstream_start(bitstream *bs)
277 {
278     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
279     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
280     bs->bit_offset = 0;
281 }
282
283 static void
284 bitstream_end(bitstream *bs)
285 {
286     int pos = (bs->bit_offset >> 5);
287     int bit_offset = (bs->bit_offset & 0x1f);
288     int bit_left = 32 - bit_offset;
289
290     if (bit_offset) {
291         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
292     }
293 }
294  
295 static void
296 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
297 {
298     int pos = (bs->bit_offset >> 5);
299     int bit_offset = (bs->bit_offset & 0x1f);
300     int bit_left = 32 - bit_offset;
301
302     if (!size_in_bits)
303         return;
304
305     bs->bit_offset += size_in_bits;
306
307     if (bit_left > size_in_bits) {
308         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
309     } else {
310         size_in_bits -= bit_left;
311         if (bit_left >= 32) {
312             bs->buffer[pos] = (val >> size_in_bits);
313         } else {
314             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
315         }
316         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
317
318         if (pos + 1 == bs->max_size_in_dword) {
319             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
320             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
321         }
322
323         bs->buffer[pos + 1] = val;
324     }
325 }
326
327 static void
328 bitstream_put_ue(bitstream *bs, unsigned int val)
329 {
330     int size_in_bits = 0;
331     int tmp_val = ++val;
332
333     while (tmp_val) {
334         tmp_val >>= 1;
335         size_in_bits++;
336     }
337
338     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
339     bitstream_put_ui(bs, val, size_in_bits);
340 }
341
342 static void
343 bitstream_put_se(bitstream *bs, int val)
344 {
345     unsigned int new_val;
346
347     if (val <= 0)
348         new_val = -2 * val;
349     else
350         new_val = 2 * val - 1;
351
352     bitstream_put_ue(bs, new_val);
353 }
354
355 static void
356 bitstream_byte_aligning(bitstream *bs, int bit)
357 {
358     int bit_offset = (bs->bit_offset & 0x7);
359     int bit_left = 8 - bit_offset;
360     int new_val;
361
362     if (!bit_offset)
363         return;
364
365     assert(bit == 0 || bit == 1);
366
367     if (bit)
368         new_val = (1 << bit_left) - 1;
369     else
370         new_val = 0;
371
372     bitstream_put_ui(bs, new_val, bit_left);
373 }
374
375 static void 
376 rbsp_trailing_bits(bitstream *bs)
377 {
378     bitstream_put_ui(bs, 1, 1);
379     bitstream_byte_aligning(bs, 0);
380 }
381
382 static void nal_start_code_prefix(bitstream *bs)
383 {
384     bitstream_put_ui(bs, 0x00000001, 32);
385 }
386
387 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
388 {
389     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
390     bitstream_put_ui(bs, nal_ref_idc, 2);
391     bitstream_put_ui(bs, nal_unit_type, 5);
392 }
393
394 void H264EncoderImpl::sps_rbsp(bitstream *bs)
395 {
396     int profile_idc = PROFILE_IDC_BASELINE;
397
398     if (h264_profile  == VAProfileH264High)
399         profile_idc = PROFILE_IDC_HIGH;
400     else if (h264_profile  == VAProfileH264Main)
401         profile_idc = PROFILE_IDC_MAIN;
402
403     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
404     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
405     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
406     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
407     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
408     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
409     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
410     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
411
412     if ( profile_idc == PROFILE_IDC_HIGH) {
413         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */ 
414         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
415         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
416         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
417         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
418     }
419
420     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
421     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
422
423     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
424         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
425     else {
426         assert(0);
427     }
428
429     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
430     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
431
432     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
433     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
434     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
435
436     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
437         assert(0);
438     }
439
440     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
441     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
442
443     if (seq_param.frame_cropping_flag) {
444         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
445         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
446         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
447         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
448     }
449     
450     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
451     if ( false ) {
452         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
453     } else {
454         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
455         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
456         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
457         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
458         {
459             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
460             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
461             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
462             {
463                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
464                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
465                 bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
466             }
467         }
468         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
469         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
470         {
471             bitstream_put_ui(bs, 1, 32);  // FPS
472             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
473             bitstream_put_ui(bs, 1, 1);
474         }
475         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
476         {
477             // hrd_parameters 
478             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
479             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
480             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
481            
482             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
483             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
484             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
485
486             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
487             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
488             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
489             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
490         }
491         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
492         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */ 
493
494         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
495         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
496     }
497
498     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
499 }
500
501
502 void H264EncoderImpl::pps_rbsp(bitstream *bs)
503 {
504     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
505     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
506
507     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
508
509     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
510
511     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
512
513     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
514     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
515
516     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
517     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
518
519     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
520     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
521     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
522
523     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
524     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
525     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
526     
527     /* more_rbsp_data */
528     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
529     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
530     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
531
532     rbsp_trailing_bits(bs);
533 }
534
535 void H264EncoderImpl::slice_header(bitstream *bs)
536 {
537     int first_mb_in_slice = slice_param.macroblock_address;
538
539     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
540     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
541     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
542     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
543
544     /* frame_mbs_only_flag == 1 */
545     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
546         /* FIXME: */
547         assert(0);
548     }
549
550     if (pic_param.pic_fields.bits.idr_pic_flag)
551         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
552
553     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
554         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
555         /* pic_order_present_flag == 0 */
556     } else {
557         /* FIXME: */
558         assert(0);
559     }
560
561     /* redundant_pic_cnt_present_flag == 0 */
562     /* slice type */
563     if (IS_P_SLICE(slice_param.slice_type)) {
564         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
565
566         if (slice_param.num_ref_idx_active_override_flag)
567             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
568
569         /* ref_pic_list_reordering */
570         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
571     } else if (IS_B_SLICE(slice_param.slice_type)) {
572         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
573
574         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
575
576         if (slice_param.num_ref_idx_active_override_flag) {
577             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
578             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
579         }
580
581         /* ref_pic_list_reordering */
582         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
583         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
584     }
585
586     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
587          IS_P_SLICE(slice_param.slice_type)) ||
588         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
589          IS_B_SLICE(slice_param.slice_type))) {
590         /* FIXME: fill weight/offset table */
591         assert(0);
592     }
593
594     /* dec_ref_pic_marking */
595     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
596         unsigned char no_output_of_prior_pics_flag = 0;
597         unsigned char long_term_reference_flag = 0;
598         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
599
600         if (pic_param.pic_fields.bits.idr_pic_flag) {
601             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
602             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
603         } else {
604             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
605         }
606     }
607
608     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
609         !IS_I_SLICE(slice_param.slice_type))
610         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
611
612     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
613
614     /* ignore for SP/SI */
615
616     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
617         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
618
619         if (slice_param.disable_deblocking_filter_idc != 1) {
620             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
621             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
622         }
623     }
624
625     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
626         bitstream_byte_aligning(bs, 1);
627     }
628 }
629
630 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
631 {
632     bitstream bs;
633
634     bitstream_start(&bs);
635     nal_start_code_prefix(&bs);
636     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
637     pps_rbsp(&bs);
638     bitstream_end(&bs);
639
640     *header_buffer = (unsigned char *)bs.buffer;
641     return bs.bit_offset;
642 }
643
644 int
645 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
646 {
647     bitstream bs;
648
649     bitstream_start(&bs);
650     nal_start_code_prefix(&bs);
651     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
652     sps_rbsp(&bs);
653     bitstream_end(&bs);
654
655     *header_buffer = (unsigned char *)bs.buffer;
656     return bs.bit_offset;
657 }
658
659 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
660 {
661     bitstream bs;
662     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
663     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
664
665     bitstream_start(&bs);
666     nal_start_code_prefix(&bs);
667
668     if (IS_I_SLICE(slice_param.slice_type)) {
669         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
670     } else if (IS_P_SLICE(slice_param.slice_type)) {
671         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
672     } else {
673         assert(IS_B_SLICE(slice_param.slice_type));
674         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
675     }
676
677     slice_header(&bs);
678     bitstream_end(&bs);
679
680     *header_buffer = (unsigned char *)bs.buffer;
681     return bs.bit_offset;
682 }
683
684
685 /*
686   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
687   1) period between Frame #X and Frame #N = #X - #N
688   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
689   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
690   4) intra_period and intra_idr_period take precedence over ip_period
691   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods 
692      of I/IDR frames, see bellow examples
693   -------------------------------------------------------------------
694   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
695   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
696   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
697   1            0                ignored    IDRIIIIIII...      (No IDR any more)
698   1            1                ignored    IDR IDR IDR IDR...
699   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
700   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
701   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
702                                               (PBB)(IBB)(PBB)(IBB)... 
703   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
704                                            IDRPPPPP IPPPPP IPPPPP...
705   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
706                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
707                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
708                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
709                                            {IDR(PBB)(PBB)}                     (6/6/3)
710                                            {IDR(PBB)(PBB)}.
711 */
712
713 // General pts/dts strategy:
714 //
715 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
716 // bit tricky. We assume first of all that the frame rate never goes _above_
717 // MAX_FPS, which gives us a frame period N. The decoder can always decode
718 // in at least this speed, as long at dts <= pts (the frame is not attempted
719 // presented before it is decoded). Furthermore, we never have longer chains of
720 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
721 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
722 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
723 // frame, with an order O <= C.)
724 //
725 // Many strategies are possible, but we establish these rules:
726 //
727 //  - Tip frames have dts = pts - (C-O)*N.
728 //  - Non-tip frames have dts = dts_last + N.
729 //
730 // An example, with C=2 and N=10 and the data flow showed with arrows:
731 //
732 //        I  B  P  B  B  P
733 //   pts: 30 40 50 60 70 80
734 //        ↓  ↓     ↓
735 //   dts: 10 30 20 60 50←40
736 //         |  |  ↑        ↑
737 //         `--|--'        |
738 //             `----------'
739 //
740 // To show that this works fine also with irregular spacings, let's say that
741 // the third frame is delayed a bit (something earlier was dropped). Now the
742 // situation looks like this:
743 //
744 //        I  B  P  B  B   P
745 //   pts: 30 40 80 90 100 110
746 //        ↓  ↓     ↓
747 //   dts: 10 30 20 90 50←40
748 //         |  |  ↑        ↑
749 //         `--|--'        |
750 //             `----------'
751 //
752 // The resetting on every tip frame makes sure dts never ends up lagging a lot
753 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
754 //
755 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
756 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
757 // a tip frame and should be given a dts based on the previous one.
758 #define FRAME_P 0
759 #define FRAME_B 1
760 #define FRAME_I 2
761 #define FRAME_IDR 7
762 void encoding2display_order(
763     int encoding_order, int intra_period,
764     int intra_idr_period, int ip_period,
765     int *displaying_order,
766     int *frame_type, int *pts_lag)
767 {
768     int encoding_order_gop = 0;
769
770     *pts_lag = 0;
771
772     if (intra_period == 1) { /* all are I/IDR frames */
773         *displaying_order = encoding_order;
774         if (intra_idr_period == 0)
775             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
776         else
777             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
778         return;
779     }
780
781     if (intra_period == 0)
782         intra_idr_period = 0;
783
784     if (ip_period == 1) {
785         // No B-frames, sequence is like IDR PPPPP IPPPPP.
786         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
787         *displaying_order = encoding_order;
788
789         if (encoding_order_gop == 0) { /* the first frame */
790             *frame_type = FRAME_IDR;
791         } else if (intra_period != 0 && /* have I frames */
792                    encoding_order_gop >= 2 &&
793                    (encoding_order_gop % intra_period == 0)) {
794             *frame_type = FRAME_I;
795         } else {
796             *frame_type = FRAME_P;
797         }
798         return;
799     } 
800
801     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
802     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
803     *pts_lag = -1;  // Most frames are not tip frames.
804          
805     if (encoding_order_gop == 0) { /* the first frame */
806         *frame_type = FRAME_IDR;
807         *displaying_order = encoding_order;
808         // IDR frames are a special case; I honestly can't find the logic behind
809         // why this is the right thing, but it seems to line up nicely in practice :-)
810         *pts_lag = TIMEBASE / MAX_FPS;
811     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
812         *frame_type = FRAME_B;
813         *displaying_order = encoding_order - 1;
814         if ((encoding_order_gop % ip_period) == 0) {
815             *pts_lag = 0;  // Last B-frame.
816         }
817     } else if (intra_period != 0 && /* have I frames */
818                encoding_order_gop >= 2 &&
819                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
820         *frame_type = FRAME_I;
821         *displaying_order = encoding_order + ip_period - 1;
822     } else {
823         *frame_type = FRAME_P;
824         *displaying_order = encoding_order + ip_period - 1;
825     }
826 }
827
828
829 static const char *rc_to_string(int rc_mode)
830 {
831     switch (rc_mode) {
832     case VA_RC_NONE:
833         return "NONE";
834     case VA_RC_CBR:
835         return "CBR";
836     case VA_RC_VBR:
837         return "VBR";
838     case VA_RC_VCM:
839         return "VCM";
840     case VA_RC_CQP:
841         return "CQP";
842     case VA_RC_VBR_CONSTRAINED:
843         return "VBR_CONSTRAINED";
844     default:
845         return "Unknown";
846     }
847 }
848
849 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
850 {
851         if (va_display.empty()) {
852                 x11_display = XOpenDisplay(NULL);
853                 if (!x11_display) {
854                         fprintf(stderr, "error: can't connect to X server!\n");
855                         return NULL;
856                 }
857                 use_zerocopy = true;
858                 return vaGetDisplay(x11_display);
859         } else if (va_display[0] != '/') {
860                 x11_display = XOpenDisplay(va_display.c_str());
861                 if (!x11_display) {
862                         fprintf(stderr, "error: can't connect to X server!\n");
863                         return NULL;
864                 }
865                 use_zerocopy = true;
866                 return vaGetDisplay(x11_display);
867         } else {
868                 drm_fd = open(va_display.c_str(), O_RDWR);
869                 if (drm_fd == -1) {
870                         perror(va_display.c_str());
871                         return NULL;
872                 }
873                 use_zerocopy = false;
874                 return vaGetDisplayDRM(drm_fd);
875         }
876 }
877
878 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
879 {
880         if (x11_display) {
881                 XCloseDisplay(x11_display);
882                 x11_display = nullptr;
883         }
884         if (drm_fd != -1) {
885                 close(drm_fd);
886         }
887 }
888
889 int H264EncoderImpl::init_va(const string &va_display)
890 {
891     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
892     VAEntrypoint *entrypoints;
893     int num_entrypoints, slice_entrypoint;
894     int support_encode = 0;    
895     int major_ver, minor_ver;
896     VAStatus va_status;
897     unsigned int i;
898
899     va_dpy = va_open_display(va_display);
900     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
901     CHECK_VASTATUS(va_status, "vaInitialize");
902
903     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
904     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
905     if (!entrypoints) {
906         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
907         exit(1);
908     }
909
910     /* use the highest profile */
911     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
912         if ((h264_profile != ~0) && h264_profile != profile_list[i])
913             continue;
914         
915         h264_profile = profile_list[i];
916         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
917         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
918             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
919                 support_encode = 1;
920                 break;
921             }
922         }
923         if (support_encode == 1)
924             break;
925     }
926     
927     if (support_encode == 0) {
928         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
929         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
930         printf("to use VA-API against DRM instead of X11.\n");
931         exit(1);
932     } else {
933         switch (h264_profile) {
934             case VAProfileH264Baseline:
935                 ip_period = 1;
936                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
937                 h264_entropy_mode = 0;
938                 break;
939             case VAProfileH264ConstrainedBaseline:
940                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
941                 ip_period = 1;
942                 break;
943
944             case VAProfileH264Main:
945                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
946                 break;
947
948             case VAProfileH264High:
949                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
950                 break;
951             default:
952                 h264_profile = VAProfileH264Baseline;
953                 ip_period = 1;
954                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
955                 break;
956         }
957     }
958
959     VAConfigAttrib attrib[VAConfigAttribTypeMax];
960
961     /* find out the format for the render target, and rate control mode */
962     for (i = 0; i < VAConfigAttribTypeMax; i++)
963         attrib[i].type = (VAConfigAttribType)i;
964
965     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
966                                       &attrib[0], VAConfigAttribTypeMax);
967     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
968     /* check the interested configattrib */
969     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
970         printf("Not find desired YUV420 RT format\n");
971         exit(1);
972     } else {
973         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
974         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
975         config_attrib_num++;
976     }
977     
978     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
979         int tmp = attrib[VAConfigAttribRateControl].value;
980
981         if (rc_mode == -1 || !(rc_mode & tmp))  {
982             if (rc_mode != -1) {
983                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
984             }
985
986             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
987                 if (rc_default_modes[i] & tmp) {
988                     rc_mode = rc_default_modes[i];
989                     break;
990                 }
991             }
992         }
993
994         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
995         config_attrib[config_attrib_num].value = rc_mode;
996         config_attrib_num++;
997     }
998     
999
1000     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1001         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1002
1003         h264_packedheader = 1;
1004         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1005         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1006         
1007         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1008             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1009         }
1010         
1011         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1012             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1013         }
1014         
1015         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1016             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1017         }
1018         
1019         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1020             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1021         }
1022         
1023         enc_packed_header_idx = config_attrib_num;
1024         config_attrib_num++;
1025     }
1026
1027     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1028         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1029         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1030         config_attrib_num++;
1031     }
1032     
1033     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1034         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1035     }
1036
1037     free(entrypoints);
1038     return 0;
1039 }
1040
1041 int H264EncoderImpl::setup_encode()
1042 {
1043     VAStatus va_status;
1044     VASurfaceID *tmp_surfaceid;
1045     int codedbuf_size, i;
1046     static VASurfaceID src_surface[SURFACE_NUM];
1047     static VASurfaceID ref_surface[SURFACE_NUM];
1048     
1049     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1050             &config_attrib[0], config_attrib_num, &config_id);
1051     CHECK_VASTATUS(va_status, "vaCreateConfig");
1052
1053     /* create source surfaces */
1054     va_status = vaCreateSurfaces(va_dpy,
1055                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1056                                  &src_surface[0], SURFACE_NUM,
1057                                  NULL, 0);
1058     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1059
1060     /* create reference surfaces */
1061     va_status = vaCreateSurfaces(va_dpy,
1062                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1063                                  &ref_surface[0], SURFACE_NUM,
1064                                  NULL, 0);
1065     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1066
1067     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1068     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1069     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1070     
1071     /* Create a context for this encode pipe */
1072     va_status = vaCreateContext(va_dpy, config_id,
1073                                 frame_width_mbaligned, frame_height_mbaligned,
1074                                 VA_PROGRESSIVE,
1075                                 tmp_surfaceid, 2 * SURFACE_NUM,
1076                                 &context_id);
1077     CHECK_VASTATUS(va_status, "vaCreateContext");
1078     free(tmp_surfaceid);
1079
1080     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1081
1082     for (i = 0; i < SURFACE_NUM; i++) {
1083         /* create coded buffer once for all
1084          * other VA buffers which won't be used again after vaRenderPicture.
1085          * so APP can always vaCreateBuffer for every frame
1086          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1087          * so VA won't maintain the coded buffer
1088          */
1089         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1090                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1091         CHECK_VASTATUS(va_status, "vaCreateBuffer");
1092     }
1093
1094     /* create OpenGL objects */
1095     //glGenFramebuffers(SURFACE_NUM, fbos);
1096     
1097     for (i = 0; i < SURFACE_NUM; i++) {
1098         glGenTextures(1, &gl_surfaces[i].y_tex);
1099         glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1100
1101         if (!use_zerocopy) {
1102             // Create Y image.
1103             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1104             glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1105
1106             // Create CbCr image.
1107             glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1108             glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1109
1110             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1111             // buffers, due to potentially differing pitch.
1112             glGenBuffers(1, &gl_surfaces[i].pbo);
1113             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1114             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1115             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1116             gl_surfaces[i].y_offset = 0;
1117             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1118             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1119             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1120             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1121         }
1122     }
1123
1124     for (i = 0; i < SURFACE_NUM; i++) {
1125         gl_surfaces[i].src_surface = src_surface[i];
1126         gl_surfaces[i].ref_surface = ref_surface[i];
1127     }
1128     
1129     return 0;
1130 }
1131
1132 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1133 //
1134 //   2 1 0 [3] 4 8 9
1135 template<class T, class C>
1136 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1137 {
1138         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1139         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1140         sort(middle, end, less_than);
1141 }
1142
1143 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1144 {
1145     int i;
1146     
1147     if (frame_type == FRAME_B)
1148         return;
1149
1150     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1151     numShortTerm++;
1152     if (numShortTerm > num_ref_frames)
1153         numShortTerm = num_ref_frames;
1154     for (i=numShortTerm-1; i>0; i--)
1155         ReferenceFrames[i] = ReferenceFrames[i-1];
1156     ReferenceFrames[0] = CurrentCurrPic;
1157     
1158     current_frame_num++;
1159     if (current_frame_num > MaxFrameNum)
1160         current_frame_num = 0;
1161 }
1162
1163
1164 int H264EncoderImpl::update_RefPicList(int frame_type)
1165 {
1166     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1167         return a.frame_idx > b.frame_idx;
1168     };
1169     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1170         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1171     };
1172     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1173         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1174     };
1175     
1176     if (frame_type == FRAME_P) {
1177         memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1178         sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1179     } else if (frame_type == FRAME_B) {
1180         memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1181         sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1182
1183         memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1184         sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1185     }
1186     
1187     return 0;
1188 }
1189
1190
1191 int H264EncoderImpl::render_sequence()
1192 {
1193     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1194     VAStatus va_status;
1195     VAEncMiscParameterBuffer *misc_param;
1196     VAEncMiscParameterRateControl *misc_rate_ctrl;
1197     
1198     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1199     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1200     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1201     seq_param.bits_per_second = frame_bitrate;
1202
1203     seq_param.intra_period = intra_period;
1204     seq_param.intra_idr_period = intra_idr_period;
1205     seq_param.ip_period = ip_period;
1206
1207     seq_param.max_num_ref_frames = num_ref_frames;
1208     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1209     seq_param.time_scale = TIMEBASE * 2;
1210     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1211     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1212     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1213     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1214     seq_param.seq_fields.bits.chroma_format_idc = 1;
1215     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1216     
1217     if (frame_width != frame_width_mbaligned ||
1218         frame_height != frame_height_mbaligned) {
1219         seq_param.frame_cropping_flag = 1;
1220         seq_param.frame_crop_left_offset = 0;
1221         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1222         seq_param.frame_crop_top_offset = 0;
1223         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1224     }
1225     
1226     va_status = vaCreateBuffer(va_dpy, context_id,
1227                                VAEncSequenceParameterBufferType,
1228                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1229     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1230     
1231     va_status = vaCreateBuffer(va_dpy, context_id,
1232                                VAEncMiscParameterBufferType,
1233                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1234                                1, NULL, &rc_param_buf);
1235     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1236     
1237     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1238     misc_param->type = VAEncMiscParameterTypeRateControl;
1239     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1240     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1241     misc_rate_ctrl->bits_per_second = frame_bitrate;
1242     misc_rate_ctrl->target_percentage = 66;
1243     misc_rate_ctrl->window_size = 1000;
1244     misc_rate_ctrl->initial_qp = initial_qp;
1245     misc_rate_ctrl->min_qp = minimal_qp;
1246     misc_rate_ctrl->basic_unit_size = 0;
1247     vaUnmapBuffer(va_dpy, rc_param_buf);
1248
1249     render_id[0] = seq_param_buf;
1250     render_id[1] = rc_param_buf;
1251     
1252     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1253     
1254     return 0;
1255 }
1256
1257 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1258 {
1259     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1260     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1261     int PicOrderCntMsb, TopFieldOrderCnt;
1262     
1263     if (frame_type == FRAME_IDR)
1264         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1265     else {
1266         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1267         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1268     }
1269     
1270     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1271         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1272         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1273     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1274              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1275         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1276     else
1277         PicOrderCntMsb = prevPicOrderCntMsb;
1278     
1279     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1280
1281     if (frame_type != FRAME_B) {
1282         PicOrderCntMsb_ref = PicOrderCntMsb;
1283         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1284     }
1285     
1286     return TopFieldOrderCnt;
1287 }
1288
1289 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1290 {
1291     VABufferID pic_param_buf;
1292     VAStatus va_status;
1293     int i = 0;
1294
1295     pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1296     pic_param.CurrPic.frame_idx = current_frame_num;
1297     pic_param.CurrPic.flags = 0;
1298     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1299     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1300     CurrentCurrPic = pic_param.CurrPic;
1301
1302     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1303     for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1304         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1305         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1306     }
1307     
1308     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1309     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1310     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1311     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1312     pic_param.frame_num = current_frame_num;
1313     pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1314     pic_param.last_picture = false;  // FIXME
1315     pic_param.pic_init_qp = initial_qp;
1316
1317     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1318                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1319     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1320
1321     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1322
1323     return 0;
1324 }
1325
1326 int H264EncoderImpl::render_packedsequence()
1327 {
1328     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1329     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1330     unsigned int length_in_bits;
1331     unsigned char *packedseq_buffer = NULL;
1332     VAStatus va_status;
1333
1334     length_in_bits = build_packed_seq_buffer(&packedseq_buffer); 
1335     
1336     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1337     
1338     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1339     packedheader_param_buffer.has_emulation_bytes = 0;
1340     va_status = vaCreateBuffer(va_dpy,
1341                                context_id,
1342                                VAEncPackedHeaderParameterBufferType,
1343                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1344                                &packedseq_para_bufid);
1345     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1346
1347     va_status = vaCreateBuffer(va_dpy,
1348                                context_id,
1349                                VAEncPackedHeaderDataBufferType,
1350                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1351                                &packedseq_data_bufid);
1352     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1353
1354     render_id[0] = packedseq_para_bufid;
1355     render_id[1] = packedseq_data_bufid;
1356     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1357
1358     free(packedseq_buffer);
1359     
1360     return 0;
1361 }
1362
1363
1364 int H264EncoderImpl::render_packedpicture()
1365 {
1366     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1367     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1368     unsigned int length_in_bits;
1369     unsigned char *packedpic_buffer = NULL;
1370     VAStatus va_status;
1371
1372     length_in_bits = build_packed_pic_buffer(&packedpic_buffer); 
1373     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1374     packedheader_param_buffer.bit_length = length_in_bits;
1375     packedheader_param_buffer.has_emulation_bytes = 0;
1376
1377     va_status = vaCreateBuffer(va_dpy,
1378                                context_id,
1379                                VAEncPackedHeaderParameterBufferType,
1380                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1381                                &packedpic_para_bufid);
1382     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1383
1384     va_status = vaCreateBuffer(va_dpy,
1385                                context_id,
1386                                VAEncPackedHeaderDataBufferType,
1387                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1388                                &packedpic_data_bufid);
1389     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1390
1391     render_id[0] = packedpic_para_bufid;
1392     render_id[1] = packedpic_data_bufid;
1393     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1394
1395     free(packedpic_buffer);
1396     
1397     return 0;
1398 }
1399
1400 void H264EncoderImpl::render_packedslice()
1401 {
1402     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1403     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1404     unsigned int length_in_bits;
1405     unsigned char *packedslice_buffer = NULL;
1406     VAStatus va_status;
1407
1408     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1409     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1410     packedheader_param_buffer.bit_length = length_in_bits;
1411     packedheader_param_buffer.has_emulation_bytes = 0;
1412
1413     va_status = vaCreateBuffer(va_dpy,
1414                                context_id,
1415                                VAEncPackedHeaderParameterBufferType,
1416                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1417                                &packedslice_para_bufid);
1418     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1419
1420     va_status = vaCreateBuffer(va_dpy,
1421                                context_id,
1422                                VAEncPackedHeaderDataBufferType,
1423                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1424                                &packedslice_data_bufid);
1425     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1426
1427     render_id[0] = packedslice_para_bufid;
1428     render_id[1] = packedslice_data_bufid;
1429     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1430
1431     free(packedslice_buffer);
1432 }
1433
1434 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1435 {
1436     VABufferID slice_param_buf;
1437     VAStatus va_status;
1438     int i;
1439
1440     update_RefPicList(frame_type);
1441     
1442     /* one frame, one slice */
1443     slice_param.macroblock_address = 0;
1444     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1445     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1446     if (frame_type == FRAME_IDR) {
1447         if (encoding_frame_num != 0)
1448             ++slice_param.idr_pic_id;
1449     } else if (frame_type == FRAME_P) {
1450         int refpiclist0_max = h264_maxref & 0xffff;
1451         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1452
1453         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1454             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1455             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1456         }
1457     } else if (frame_type == FRAME_B) {
1458         int refpiclist0_max = h264_maxref & 0xffff;
1459         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1460
1461         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1462         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1463             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1464             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1465         }
1466
1467         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1468         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1469             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1470             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1471         }
1472     }
1473
1474     slice_param.slice_alpha_c0_offset_div2 = 0;
1475     slice_param.slice_beta_offset_div2 = 0;
1476     slice_param.direct_spatial_mv_pred_flag = 1;
1477     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1478     
1479
1480     if (h264_packedheader &&
1481         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1482         render_packedslice();
1483
1484     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1485                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1486     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1487
1488     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1489
1490     return 0;
1491 }
1492
1493
1494
1495 void H264EncoderImpl::save_codeddata(storage_task task)
1496 {    
1497     VACodedBufferSegment *buf_list = NULL;
1498     VAStatus va_status;
1499
1500     string data;
1501
1502     const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);  // So we never get negative dts.
1503
1504     va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1505     CHECK_VASTATUS(va_status, "vaMapBuffer");
1506     while (buf_list != NULL) {
1507         data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1508         buf_list = (VACodedBufferSegment *) buf_list->next;
1509     }
1510     vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1511
1512     {
1513         // Add video.
1514         AVPacket pkt;
1515         memset(&pkt, 0, sizeof(pkt));
1516         pkt.buf = nullptr;
1517         pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1518         pkt.size = data.size();
1519         pkt.stream_index = 0;
1520         if (task.frame_type == FRAME_IDR || task.frame_type == FRAME_I) {
1521             pkt.flags = AV_PKT_FLAG_KEY;
1522         } else {
1523             pkt.flags = 0;
1524         }
1525         //pkt.duration = 1;
1526         httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
1527     }
1528     // Encode and add all audio frames up to and including the pts of this video frame.
1529     for ( ;; ) {
1530         int64_t audio_pts;
1531         vector<float> audio;
1532         {
1533              unique_lock<mutex> lock(frame_queue_mutex);
1534              frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1535              if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1536              auto it = pending_audio_frames.begin();
1537              if (it->first > task.pts) break;
1538              audio_pts = it->first;
1539              audio = move(it->second);
1540              pending_audio_frames.erase(it); 
1541         }
1542
1543         AVFrame *frame = av_frame_alloc();
1544         frame->nb_samples = audio.size() / 2;
1545         frame->format = AV_SAMPLE_FMT_S32;
1546         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(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 < 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, 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(frame);
1576         av_free_packet(&pkt);
1577         if (audio_pts == task.pts) break;
1578     }
1579
1580 #if 0
1581     printf("\r      "); /* return back to startpoint */
1582     switch (encode_order % 4) {
1583         case 0:
1584             printf("|");
1585             break;
1586         case 1:
1587             printf("/");
1588             break;
1589         case 2:
1590             printf("-");
1591             break;
1592         case 3:
1593             printf("\\");
1594             break;
1595     }
1596     printf("%08lld", encode_order);
1597 #endif
1598 }
1599
1600
1601 // this is weird. but it seems to put a new frame onto the queue
1602 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1603 {
1604         unique_lock<mutex> lock(storage_task_queue_mutex);
1605         storage_task_queue.push(move(task));
1606         storage_task_queue_changed.notify_all();
1607 }
1608
1609 void H264EncoderImpl::storage_task_thread()
1610 {
1611         for ( ;; ) {
1612                 storage_task current;
1613                 {
1614                         // wait until there's an encoded frame  
1615                         unique_lock<mutex> lock(storage_task_queue_mutex);
1616                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1617                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1618                         current = move(storage_task_queue.front());
1619                         storage_task_queue.pop();
1620                 }
1621
1622                 VAStatus va_status;
1623            
1624                 // waits for data, then saves it to disk.
1625                 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1626                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1627                 save_codeddata(move(current));
1628
1629                 {
1630                         unique_lock<mutex> lock(storage_task_queue_mutex);
1631                         srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1632                         storage_task_queue_changed.notify_all();
1633                 }
1634         }
1635 }
1636
1637 int H264EncoderImpl::release_encode()
1638 {
1639         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1640                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1641                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1642                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1643
1644                 if (!use_zerocopy) {
1645                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1646                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1647                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1648                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1649                 }
1650                 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1651                 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1652         }
1653
1654         vaDestroyContext(va_dpy, context_id);
1655         vaDestroyConfig(va_dpy, config_id);
1656
1657         return 0;
1658 }
1659
1660 int H264EncoderImpl::deinit_va()
1661
1662     vaTerminate(va_dpy);
1663
1664     va_close_display(va_dpy);
1665
1666     return 0;
1667 }
1668
1669
1670 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1671         : current_storage_frame(0), surface(surface), httpd(httpd)
1672 {
1673         AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
1674         context_audio = avcodec_alloc_context3(codec_audio);
1675         context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
1676         context_audio->sample_rate = OUTPUT_FREQUENCY;
1677         context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
1678         context_audio->channels = 2;
1679         context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1680         context_audio->time_base = AVRational{1, TIMEBASE};
1681         if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1682                 fprintf(stderr, "Could not open codec\n");
1683                 exit(1);
1684         }
1685
1686         frame_width = width;
1687         frame_height = height;
1688         frame_width_mbaligned = (frame_width + 15) & (~15);
1689         frame_height_mbaligned = (frame_height + 15) & (~15);
1690
1691         //print_input();
1692
1693         init_va(va_display);
1694         setup_encode();
1695
1696         // No frames are ready yet.
1697         memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1698             
1699         memset(&seq_param, 0, sizeof(seq_param));
1700         memset(&pic_param, 0, sizeof(pic_param));
1701         memset(&slice_param, 0, sizeof(slice_param));
1702
1703         storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1704
1705         encode_thread = thread([this]{
1706                 //SDL_GL_MakeCurrent(window, context);
1707                 QOpenGLContext *context = create_context(this->surface);
1708                 eglBindAPI(EGL_OPENGL_API);
1709                 if (!make_current(context, this->surface)) {
1710                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1711                                 eglGetError());
1712                         exit(1);
1713                 }
1714                 encode_thread_func();
1715         });
1716 }
1717
1718 H264EncoderImpl::~H264EncoderImpl()
1719 {
1720         shutdown();
1721 }
1722
1723 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1724 {
1725         assert(!is_shutdown);
1726         {
1727                 // Wait until this frame slot is done encoding.
1728                 unique_lock<mutex> lock(storage_task_queue_mutex);
1729                 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1730                         fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1731                                 current_storage_frame % SURFACE_NUM, current_storage_frame);
1732                 }
1733                 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1734                 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1735                 if (storage_thread_should_quit) return false;
1736         }
1737
1738         //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1739         GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1740         *y_tex = surf->y_tex;
1741         *cbcr_tex = surf->cbcr_tex;
1742
1743         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1744         CHECK_VASTATUS(va_status, "vaDeriveImage");
1745
1746         if (use_zerocopy) {
1747                 VABufferInfo buf_info;
1748                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1749                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1750                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1751
1752                 // Create Y image.
1753                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1754                 EGLint y_attribs[] = {
1755                         EGL_WIDTH, frame_width,
1756                         EGL_HEIGHT, frame_height,
1757                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1758                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1759                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1760                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1761                         EGL_NONE
1762                 };
1763
1764                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1765                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1766
1767                 // Associate Y image to a texture.
1768                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1769                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1770
1771                 // Create CbCr image.
1772                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1773                 EGLint cbcr_attribs[] = {
1774                         EGL_WIDTH, frame_width,
1775                         EGL_HEIGHT, frame_height,
1776                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1777                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1778                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1779                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1780                         EGL_NONE
1781                 };
1782
1783                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1784                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1785
1786                 // Associate CbCr image to a texture.
1787                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1788                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1789         }
1790
1791         return true;
1792 }
1793
1794 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
1795 {
1796         assert(!is_shutdown);
1797         {
1798                 unique_lock<mutex> lock(frame_queue_mutex);
1799                 pending_audio_frames[pts] = move(audio);
1800         }
1801         frame_queue_nonempty.notify_all();
1802 }
1803
1804 void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
1805 {
1806         assert(!is_shutdown);
1807
1808         if (!use_zerocopy) {
1809                 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1810
1811                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1812                 check_error();
1813
1814                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1815                 check_error();
1816
1817                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1818                 check_error();
1819                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1820                 check_error();
1821
1822                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1823                 check_error();
1824                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1825                 check_error();
1826
1827                 glBindTexture(GL_TEXTURE_2D, 0);
1828                 check_error();
1829                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1830                 check_error();
1831
1832                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1833                 check_error();
1834                 fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1835                 check_error();
1836         }
1837
1838         {
1839                 unique_lock<mutex> lock(frame_queue_mutex);
1840                 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
1841                 ++current_storage_frame;
1842         }
1843         frame_queue_nonempty.notify_all();
1844 }
1845
1846 void H264EncoderImpl::shutdown()
1847 {
1848         if (is_shutdown) {
1849                 return;
1850         }
1851
1852         {
1853                 unique_lock<mutex> lock(frame_queue_mutex);
1854                 encode_thread_should_quit = true;
1855                 frame_queue_nonempty.notify_all();
1856         }
1857         encode_thread.join();
1858         {
1859                 unique_lock<mutex> lock(storage_task_queue_mutex);
1860                 storage_thread_should_quit = true;
1861                 frame_queue_nonempty.notify_all();
1862                 storage_task_queue_changed.notify_all();
1863         }
1864         storage_thread.join();
1865
1866         release_encode();
1867         deinit_va();
1868         is_shutdown = true;
1869 }
1870
1871 void H264EncoderImpl::encode_thread_func()
1872 {
1873         int64_t last_dts = -1;
1874         int gop_start_display_frame_num = 0;
1875         for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1876                 PendingFrame frame;
1877                 int pts_lag;
1878                 int frame_type, display_frame_num;
1879                 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1880                                        &display_frame_num, &frame_type, &pts_lag);
1881                 if (frame_type == FRAME_IDR) {
1882                         numShortTerm = 0;
1883                         current_frame_num = 0;
1884                         gop_start_display_frame_num = display_frame_num;
1885                 }
1886
1887                 {
1888                         unique_lock<mutex> lock(frame_queue_mutex);
1889                         frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1890                                 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1891                         });
1892                         if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1893                                 // We have queued frames that were supposed to be B-frames,
1894                                 // but will be no P-frame to encode them against. Encode them all
1895                                 // as P-frames instead. Note that this happens under the mutex,
1896                                 // but nobody else uses it at this point, since we're shutting down,
1897                                 // so there's no contention.
1898                                 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1899                                 return;
1900                         } else {
1901                                 frame = move(pending_video_frames[display_frame_num]);
1902                                 pending_video_frames.erase(display_frame_num);
1903                         }
1904                 }
1905
1906                 // Determine the dts of this frame.
1907                 int64_t dts;
1908                 if (pts_lag == -1) {
1909                         assert(last_dts != -1);
1910                         dts = last_dts + (TIMEBASE / MAX_FPS);
1911                 } else {
1912                         dts = frame.pts - pts_lag;
1913                 }
1914                 last_dts = dts;
1915
1916                 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
1917         }
1918 }
1919
1920 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1921 {
1922         if (pending_video_frames.empty()) {
1923                 return;
1924         }
1925
1926         for (auto &pending_frame : pending_video_frames) {
1927                 int display_frame_num = pending_frame.first;
1928                 assert(display_frame_num > 0);
1929                 PendingFrame frame = move(pending_frame.second);
1930                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1931                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1932                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
1933                 last_dts = dts;
1934         }
1935 }
1936
1937 namespace {
1938
1939 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1940 {
1941         if (src_width == dst_pitch) {
1942                 memcpy(dst, src, src_width * height);
1943         } else {
1944                 for (size_t y = 0; y < height; ++y) {
1945                         const uint8_t *sptr = src + y * src_width;
1946                         uint8_t *dptr = dst + y * dst_pitch;
1947                         memcpy(dptr, sptr, src_width);
1948                 }
1949         }
1950 }
1951
1952 }  // namespace
1953
1954 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1955                                    int frame_type, int64_t pts, int64_t dts)
1956 {
1957         // Wait for the GPU to be done with the frame.
1958         GLenum sync_status;
1959         do {
1960                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1961                 check_error();
1962         } while (sync_status == GL_TIMEOUT_EXPIRED);
1963         assert(sync_status != GL_WAIT_FAILED);
1964
1965         // Release back any input frames we needed to render this frame.
1966         frame.input_frames.clear();
1967
1968         GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1969         VAStatus va_status;
1970
1971         if (use_zerocopy) {
1972                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1973                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1974                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1975                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1976         } else {
1977                 unsigned char *surface_p = nullptr;
1978                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1979
1980                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1981                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1982
1983                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1984                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1985
1986                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1987                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1988         }
1989
1990         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1991         CHECK_VASTATUS(va_status, "vaDestroyImage");
1992
1993         // Schedule the frame for encoding.
1994         VASurfaceID va_surface = surf->src_surface;
1995         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1996         CHECK_VASTATUS(va_status, "vaBeginPicture");
1997
1998         if (frame_type == FRAME_IDR) {
1999                 render_sequence();
2000                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2001                 if (h264_packedheader) {
2002                         render_packedsequence();
2003                         render_packedpicture();
2004                 }
2005         } else {
2006                 //render_sequence();
2007                 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2008         }
2009         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2010
2011         va_status = vaEndPicture(va_dpy, context_id);
2012         CHECK_VASTATUS(va_status, "vaEndPicture");
2013
2014         // so now the data is done encoding (well, async job kicked off)...
2015         // we send that to the storage thread
2016         storage_task tmp;
2017         tmp.display_order = display_frame_num;
2018         tmp.frame_type = frame_type;
2019         tmp.pts = pts;
2020         tmp.dts = dts;
2021         storage_task_enqueue(move(tmp));
2022
2023         update_ReferenceFrames(frame_type);
2024 }
2025
2026 // Proxy object.
2027 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2028         : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2029
2030 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2031 H264Encoder::~H264Encoder() {}
2032
2033 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2034 {
2035         impl->add_audio(pts, audio);
2036 }
2037
2038 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2039 {
2040         return impl->begin_frame(y_tex, cbcr_tex);
2041 }
2042
2043 void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
2044 {
2045         impl->end_frame(fence, pts, input_frames);
2046 }
2047
2048 void H264Encoder::shutdown()
2049 {
2050         impl->shutdown();
2051 }
2052
2053 // Real class.