]> git.sesse.net Git - nageru/blob - quicksync_encoder.cpp
Make QuickSync surfaces be allocated dynamically.
[nageru] / quicksync_encoder.cpp
1 #include "quicksync_encoder.h"
2
3 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
4 #include <movit/util.h>
5
6 #include <EGL/eglplatform.h>
7 #include <X11/Xlib.h>
8 #include <assert.h>
9 #include <epoxy/egl.h>
10 #include <fcntl.h>
11 #include <pthread.h>
12 #include <stdio.h>
13 #include <stdlib.h>
14 #include <string.h>
15 #include <unistd.h>
16 #include <va/va.h>
17 #include <va/va_drm.h>
18 #include <va/va_drmcommon.h>
19 #include <va/va_enc_h264.h>
20 #include <va/va_x11.h>
21 #include <algorithm>
22 #include <chrono>
23 #include <condition_variable>
24 #include <cstddef>
25 #include <cstdint>
26 #include <functional>
27 #include <map>
28 #include <memory>
29 #include <mutex>
30 #include <queue>
31 #include <stack>
32 #include <string>
33 #include <thread>
34 #include <utility>
35
36 extern "C" {
37
38 #include <libavcodec/avcodec.h>
39 #include <libavformat/avio.h>
40 #include <libavutil/error.h>
41 #include <libdrm/drm_fourcc.h>
42
43 }  // namespace
44
45 #include "audio_encoder.h"
46 #include "context.h"
47 #include "defs.h"
48 #include "disk_space_estimator.h"
49 #include "ffmpeg_raii.h"
50 #include "flags.h"
51 #include "mux.h"
52 #include "print_latency.h"
53 #include "quicksync_encoder_impl.h"
54 #include "ref_counted_frame.h"
55 #include "timebase.h"
56 #include "x264_encoder.h"
57
58 using namespace std;
59 using namespace std::chrono;
60 using namespace std::placeholders;
61
62 class QOpenGLContext;
63 class QSurface;
64
65 #define CHECK_VASTATUS(va_status, func)                                 \
66     if (va_status != VA_STATUS_SUCCESS) {                               \
67         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
68         exit(1);                                                        \
69     }
70
71 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
72
73 //#include "loadsurface.h"
74
75 #define NAL_REF_IDC_NONE        0
76 #define NAL_REF_IDC_LOW         1
77 #define NAL_REF_IDC_MEDIUM      2
78 #define NAL_REF_IDC_HIGH        3
79
80 #define NAL_NON_IDR             1
81 #define NAL_IDR                 5
82 #define NAL_SPS                 7
83 #define NAL_PPS                 8
84 #define NAL_SEI                 6
85
86 #define SLICE_TYPE_P            0
87 #define SLICE_TYPE_B            1
88 #define SLICE_TYPE_I            2
89 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
90 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
91 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
92
93
94 #define ENTROPY_MODE_CAVLC      0
95 #define ENTROPY_MODE_CABAC      1
96
97 #define PROFILE_IDC_BASELINE    66
98 #define PROFILE_IDC_MAIN        77
99 #define PROFILE_IDC_HIGH        100
100    
101 #define BITSTREAM_ALLOCATE_STEPPING     4096
102
103 static constexpr unsigned int MaxFrameNum = (2<<16);
104 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
105 static constexpr unsigned int Log2MaxFrameNum = 16;
106 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
107 static constexpr int rc_default_modes[] = {  // Priority list of modes.
108     VA_RC_VBR,
109     VA_RC_CQP,
110     VA_RC_VBR_CONSTRAINED,
111     VA_RC_CBR,
112     VA_RC_VCM,
113     VA_RC_NONE,
114 };
115
116 using namespace std;
117
118 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
119 // but if we don't delete it here, we get leaks. The GStreamer implementation
120 // does the same.
121 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
122 {
123     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
124     CHECK_VASTATUS(va_status, "vaRenderPicture");
125
126     for (int i = 0; i < num_buffers; ++i) {
127         va_status = vaDestroyBuffer(dpy, buffers[i]);
128         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
129     }
130 }
131
132 static unsigned int 
133 va_swap32(unsigned int val)
134 {
135     unsigned char *pval = (unsigned char *)&val;
136
137     return ((pval[0] << 24)     |
138             (pval[1] << 16)     |
139             (pval[2] << 8)      |
140             (pval[3] << 0));
141 }
142
143 static void
144 bitstream_start(bitstream *bs)
145 {
146     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
147     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
148     bs->bit_offset = 0;
149 }
150
151 static void
152 bitstream_end(bitstream *bs)
153 {
154     int pos = (bs->bit_offset >> 5);
155     int bit_offset = (bs->bit_offset & 0x1f);
156     int bit_left = 32 - bit_offset;
157
158     if (bit_offset) {
159         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
160     }
161 }
162  
163 static void
164 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
165 {
166     int pos = (bs->bit_offset >> 5);
167     int bit_offset = (bs->bit_offset & 0x1f);
168     int bit_left = 32 - bit_offset;
169
170     if (!size_in_bits)
171         return;
172
173     bs->bit_offset += size_in_bits;
174
175     if (bit_left > size_in_bits) {
176         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
177     } else {
178         size_in_bits -= bit_left;
179         if (bit_left >= 32) {
180             bs->buffer[pos] = (val >> size_in_bits);
181         } else {
182             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
183         }
184         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
185
186         if (pos + 1 == bs->max_size_in_dword) {
187             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
188             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
189         }
190
191         bs->buffer[pos + 1] = val;
192     }
193 }
194
195 static void
196 bitstream_put_ue(bitstream *bs, unsigned int val)
197 {
198     int size_in_bits = 0;
199     int tmp_val = ++val;
200
201     while (tmp_val) {
202         tmp_val >>= 1;
203         size_in_bits++;
204     }
205
206     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
207     bitstream_put_ui(bs, val, size_in_bits);
208 }
209
210 static void
211 bitstream_put_se(bitstream *bs, int val)
212 {
213     unsigned int new_val;
214
215     if (val <= 0)
216         new_val = -2 * val;
217     else
218         new_val = 2 * val - 1;
219
220     bitstream_put_ue(bs, new_val);
221 }
222
223 static void
224 bitstream_byte_aligning(bitstream *bs, int bit)
225 {
226     int bit_offset = (bs->bit_offset & 0x7);
227     int bit_left = 8 - bit_offset;
228     int new_val;
229
230     if (!bit_offset)
231         return;
232
233     assert(bit == 0 || bit == 1);
234
235     if (bit)
236         new_val = (1 << bit_left) - 1;
237     else
238         new_val = 0;
239
240     bitstream_put_ui(bs, new_val, bit_left);
241 }
242
243 static void 
244 rbsp_trailing_bits(bitstream *bs)
245 {
246     bitstream_put_ui(bs, 1, 1);
247     bitstream_byte_aligning(bs, 0);
248 }
249
250 static void nal_start_code_prefix(bitstream *bs)
251 {
252     bitstream_put_ui(bs, 0x00000001, 32);
253 }
254
255 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
256 {
257     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
258     bitstream_put_ui(bs, nal_ref_idc, 2);
259     bitstream_put_ui(bs, nal_unit_type, 5);
260 }
261
262 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
263 {
264     int profile_idc = PROFILE_IDC_BASELINE;
265
266     if (h264_profile  == VAProfileH264High)
267         profile_idc = PROFILE_IDC_HIGH;
268     else if (h264_profile  == VAProfileH264Main)
269         profile_idc = PROFILE_IDC_MAIN;
270
271     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
272     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
273     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
274     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
275     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
276     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
277     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
278     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
279
280     if ( profile_idc == PROFILE_IDC_HIGH) {
281         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */ 
282         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
283         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
284         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
285         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
286     }
287
288     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
289     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
290
291     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
292         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
293     else {
294         assert(0);
295     }
296
297     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
298     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
299
300     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
301     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
302     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
303
304     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
305         assert(0);
306     }
307
308     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
309     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
310
311     if (seq_param.frame_cropping_flag) {
312         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
313         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
314         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
315         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
316     }
317     
318     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
319     if ( false ) {
320         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
321     } else {
322         // See H.264 annex E for the definition of this header.
323         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
324         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
325         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
326         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
327         {
328             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
329             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
330             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
331             {
332                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
333                 bitstream_put_ui(bs, 2, 8);  /* transfer_characteristics (2 = unspecified, since we use sRGB) */
334                 if (global_flags.ycbcr_rec709_coefficients) {
335                     bitstream_put_ui(bs, 1, 8);  /* matrix_coefficients (1 = BT.709) */
336                 } else {
337                     bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
338                 }
339             }
340         }
341         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
342         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
343         {
344             bitstream_put_ui(bs, 1, 32);  // FPS
345             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
346             bitstream_put_ui(bs, 1, 1);
347         }
348         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
349         {
350             // hrd_parameters 
351             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
352             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
353             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
354            
355             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
356             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
357             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
358
359             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
360             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
361             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
362             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
363         }
364         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
365         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */ 
366
367         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
368         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
369     }
370
371     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
372 }
373
374
375 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
376 {
377     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
378     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
379
380     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
381
382     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
383
384     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
385
386     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
387     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
388
389     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
390     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
391
392     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
393     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
394     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
395
396     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
397     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
398     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
399     
400     /* more_rbsp_data */
401     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
402     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
403     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
404
405     rbsp_trailing_bits(bs);
406 }
407
408 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
409 {
410     int first_mb_in_slice = slice_param.macroblock_address;
411
412     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
413     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
414     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
415     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
416
417     /* frame_mbs_only_flag == 1 */
418     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
419         /* FIXME: */
420         assert(0);
421     }
422
423     if (pic_param.pic_fields.bits.idr_pic_flag)
424         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
425
426     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
427         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
428         /* pic_order_present_flag == 0 */
429     } else {
430         /* FIXME: */
431         assert(0);
432     }
433
434     /* redundant_pic_cnt_present_flag == 0 */
435     /* slice type */
436     if (IS_P_SLICE(slice_param.slice_type)) {
437         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
438
439         if (slice_param.num_ref_idx_active_override_flag)
440             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
441
442         /* ref_pic_list_reordering */
443         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
444     } else if (IS_B_SLICE(slice_param.slice_type)) {
445         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
446
447         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
448
449         if (slice_param.num_ref_idx_active_override_flag) {
450             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
451             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
452         }
453
454         /* ref_pic_list_reordering */
455         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
456         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
457     }
458
459     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
460          IS_P_SLICE(slice_param.slice_type)) ||
461         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
462          IS_B_SLICE(slice_param.slice_type))) {
463         /* FIXME: fill weight/offset table */
464         assert(0);
465     }
466
467     /* dec_ref_pic_marking */
468     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
469         unsigned char no_output_of_prior_pics_flag = 0;
470         unsigned char long_term_reference_flag = 0;
471         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
472
473         if (pic_param.pic_fields.bits.idr_pic_flag) {
474             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
475             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
476         } else {
477             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
478         }
479     }
480
481     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
482         !IS_I_SLICE(slice_param.slice_type))
483         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
484
485     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
486
487     /* ignore for SP/SI */
488
489     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
490         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
491
492         if (slice_param.disable_deblocking_filter_idc != 1) {
493             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
494             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
495         }
496     }
497
498     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
499         bitstream_byte_aligning(bs, 1);
500     }
501 }
502
503 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
504 {
505     bitstream bs;
506
507     bitstream_start(&bs);
508     nal_start_code_prefix(&bs);
509     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
510     pps_rbsp(&bs);
511     bitstream_end(&bs);
512
513     *header_buffer = (unsigned char *)bs.buffer;
514     return bs.bit_offset;
515 }
516
517 int
518 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
519 {
520     bitstream bs;
521
522     bitstream_start(&bs);
523     nal_start_code_prefix(&bs);
524     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
525     sps_rbsp(&bs);
526     bitstream_end(&bs);
527
528     *header_buffer = (unsigned char *)bs.buffer;
529     return bs.bit_offset;
530 }
531
532 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
533 {
534     bitstream bs;
535     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
536     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
537
538     bitstream_start(&bs);
539     nal_start_code_prefix(&bs);
540
541     if (IS_I_SLICE(slice_param.slice_type)) {
542         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
543     } else if (IS_P_SLICE(slice_param.slice_type)) {
544         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
545     } else {
546         assert(IS_B_SLICE(slice_param.slice_type));
547         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
548     }
549
550     slice_header(&bs);
551     bitstream_end(&bs);
552
553     *header_buffer = (unsigned char *)bs.buffer;
554     return bs.bit_offset;
555 }
556
557
558 /*
559   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
560   1) period between Frame #X and Frame #N = #X - #N
561   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
562   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
563   4) intra_period and intra_idr_period take precedence over ip_period
564   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods 
565      of I/IDR frames, see bellow examples
566   -------------------------------------------------------------------
567   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
568   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
569   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
570   1            0                ignored    IDRIIIIIII...      (No IDR any more)
571   1            1                ignored    IDR IDR IDR IDR...
572   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
573   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
574   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
575                                               (PBB)(IBB)(PBB)(IBB)... 
576   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
577                                            IDRPPPPP IPPPPP IPPPPP...
578   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
579                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
580                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
581                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
582                                            {IDR(PBB)(PBB)}                     (6/6/3)
583                                            {IDR(PBB)(PBB)}.
584 */
585
586 // General pts/dts strategy:
587 //
588 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
589 // bit tricky. We assume first of all that the frame rate never goes _above_
590 // MAX_FPS, which gives us a frame period N. The decoder can always decode
591 // in at least this speed, as long at dts <= pts (the frame is not attempted
592 // presented before it is decoded). Furthermore, we never have longer chains of
593 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
594 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
595 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
596 // frame, with an order O <= C.)
597 //
598 // Many strategies are possible, but we establish these rules:
599 //
600 //  - Tip frames have dts = pts - (C-O)*N.
601 //  - Non-tip frames have dts = dts_last + N.
602 //
603 // An example, with C=2 and N=10 and the data flow showed with arrows:
604 //
605 //        I  B  P  B  B  P
606 //   pts: 30 40 50 60 70 80
607 //        ↓  ↓     ↓
608 //   dts: 10 30 20 60 50←40
609 //         |  |  ↑        ↑
610 //         `--|--'        |
611 //             `----------'
612 //
613 // To show that this works fine also with irregular spacings, let's say that
614 // the third frame is delayed a bit (something earlier was dropped). Now the
615 // situation looks like this:
616 //
617 //        I  B  P  B  B   P
618 //   pts: 30 40 80 90 100 110
619 //        ↓  ↓     ↓
620 //   dts: 10 30 20 90 50←40
621 //         |  |  ↑        ↑
622 //         `--|--'        |
623 //             `----------'
624 //
625 // The resetting on every tip frame makes sure dts never ends up lagging a lot
626 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
627 //
628 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
629 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
630 // a tip frame and should be given a dts based on the previous one.
631 #define FRAME_P 0
632 #define FRAME_B 1
633 #define FRAME_I 2
634 #define FRAME_IDR 7
635 void encoding2display_order(
636     int encoding_order, int intra_period,
637     int intra_idr_period, int ip_period,
638     int *displaying_order,
639     int *frame_type, int *pts_lag)
640 {
641     int encoding_order_gop = 0;
642
643     *pts_lag = 0;
644
645     if (intra_period == 1) { /* all are I/IDR frames */
646         *displaying_order = encoding_order;
647         if (intra_idr_period == 0)
648             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
649         else
650             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
651         return;
652     }
653
654     if (intra_period == 0)
655         intra_idr_period = 0;
656
657     if (ip_period == 1) {
658         // No B-frames, sequence is like IDR PPPPP IPPPPP.
659         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
660         *displaying_order = encoding_order;
661
662         if (encoding_order_gop == 0) { /* the first frame */
663             *frame_type = FRAME_IDR;
664         } else if (intra_period != 0 && /* have I frames */
665                    encoding_order_gop >= 2 &&
666                    (encoding_order_gop % intra_period == 0)) {
667             *frame_type = FRAME_I;
668         } else {
669             *frame_type = FRAME_P;
670         }
671         return;
672     } 
673
674     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
675     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
676     *pts_lag = -1;  // Most frames are not tip frames.
677          
678     if (encoding_order_gop == 0) { /* the first frame */
679         *frame_type = FRAME_IDR;
680         *displaying_order = encoding_order;
681         // IDR frames are a special case; I honestly can't find the logic behind
682         // why this is the right thing, but it seems to line up nicely in practice :-)
683         *pts_lag = TIMEBASE / MAX_FPS;
684     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
685         *frame_type = FRAME_B;
686         *displaying_order = encoding_order - 1;
687         if ((encoding_order_gop % ip_period) == 0) {
688             *pts_lag = 0;  // Last B-frame.
689         }
690     } else if (intra_period != 0 && /* have I frames */
691                encoding_order_gop >= 2 &&
692                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
693         *frame_type = FRAME_I;
694         *displaying_order = encoding_order + ip_period - 1;
695     } else {
696         *frame_type = FRAME_P;
697         *displaying_order = encoding_order + ip_period - 1;
698     }
699 }
700
701
702 static const char *rc_to_string(int rc_mode)
703 {
704     switch (rc_mode) {
705     case VA_RC_NONE:
706         return "NONE";
707     case VA_RC_CBR:
708         return "CBR";
709     case VA_RC_VBR:
710         return "VBR";
711     case VA_RC_VCM:
712         return "VCM";
713     case VA_RC_CQP:
714         return "CQP";
715     case VA_RC_VBR_CONSTRAINED:
716         return "VBR_CONSTRAINED";
717     default:
718         return "Unknown";
719     }
720 }
721
722 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
723 {
724         if (global_flags.uncompressed_video_to_http) {
725                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
726                 use_zerocopy = false;
727         } else if (global_flags.x264_video_to_http) {
728                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
729                 use_zerocopy = false;
730         } else {
731                 use_zerocopy = true;
732         }
733 }
734
735 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
736 {
737         if (va_display.empty()) {
738                 x11_display = XOpenDisplay(NULL);
739                 if (!x11_display) {
740                         fprintf(stderr, "error: can't connect to X server!\n");
741                         return NULL;
742                 }
743                 enable_zerocopy_if_possible();
744                 return vaGetDisplay(x11_display);
745         } else if (va_display[0] != '/') {
746                 x11_display = XOpenDisplay(va_display.c_str());
747                 if (!x11_display) {
748                         fprintf(stderr, "error: can't connect to X server!\n");
749                         return NULL;
750                 }
751                 enable_zerocopy_if_possible();
752                 return vaGetDisplay(x11_display);
753         } else {
754                 drm_fd = open(va_display.c_str(), O_RDWR);
755                 if (drm_fd == -1) {
756                         perror(va_display.c_str());
757                         return NULL;
758                 }
759                 use_zerocopy = false;
760                 return vaGetDisplayDRM(drm_fd);
761         }
762 }
763
764 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
765 {
766         if (x11_display) {
767                 XCloseDisplay(x11_display);
768                 x11_display = nullptr;
769         }
770         if (drm_fd != -1) {
771                 close(drm_fd);
772         }
773 }
774
775 int QuickSyncEncoderImpl::init_va(const string &va_display)
776 {
777     VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
778     VAEntrypoint *entrypoints;
779     int num_entrypoints, slice_entrypoint;
780     int support_encode = 0;    
781     int major_ver, minor_ver;
782     VAStatus va_status;
783     unsigned int i;
784
785     va_dpy = va_open_display(va_display);
786     va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
787     CHECK_VASTATUS(va_status, "vaInitialize");
788
789     num_entrypoints = vaMaxNumEntrypoints(va_dpy);
790     entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
791     if (!entrypoints) {
792         fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
793         exit(1);
794     }
795
796     /* use the highest profile */
797     for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
798         if ((h264_profile != ~0) && h264_profile != profile_list[i])
799             continue;
800         
801         h264_profile = profile_list[i];
802         vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
803         for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
804             if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
805                 support_encode = 1;
806                 break;
807             }
808         }
809         if (support_encode == 1)
810             break;
811     }
812     
813     if (support_encode == 0) {
814         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
815         printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
816         printf("to use VA-API against DRM instead of X11.\n");
817         exit(1);
818     } else {
819         switch (h264_profile) {
820             case VAProfileH264Baseline:
821                 ip_period = 1;
822                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
823                 h264_entropy_mode = 0;
824                 break;
825             case VAProfileH264ConstrainedBaseline:
826                 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
827                 ip_period = 1;
828                 break;
829
830             case VAProfileH264Main:
831                 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
832                 break;
833
834             case VAProfileH264High:
835                 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
836                 break;
837             default:
838                 h264_profile = VAProfileH264Baseline;
839                 ip_period = 1;
840                 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
841                 break;
842         }
843     }
844
845     VAConfigAttrib attrib[VAConfigAttribTypeMax];
846
847     /* find out the format for the render target, and rate control mode */
848     for (i = 0; i < VAConfigAttribTypeMax; i++)
849         attrib[i].type = (VAConfigAttribType)i;
850
851     va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
852                                       &attrib[0], VAConfigAttribTypeMax);
853     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
854     /* check the interested configattrib */
855     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
856         printf("Not find desired YUV420 RT format\n");
857         exit(1);
858     } else {
859         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
860         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
861         config_attrib_num++;
862     }
863     
864     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
865         int tmp = attrib[VAConfigAttribRateControl].value;
866
867         if (rc_mode == -1 || !(rc_mode & tmp))  {
868             if (rc_mode != -1) {
869                 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
870             }
871
872             for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
873                 if (rc_default_modes[i] & tmp) {
874                     rc_mode = rc_default_modes[i];
875                     break;
876                 }
877             }
878         }
879
880         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
881         config_attrib[config_attrib_num].value = rc_mode;
882         config_attrib_num++;
883     }
884     
885
886     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
887         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
888
889         h264_packedheader = 1;
890         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
891         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
892         
893         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
894             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
895         }
896         
897         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
898             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
899         }
900         
901         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
902             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
903         }
904         
905         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
906             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
907         }
908         
909         enc_packed_header_idx = config_attrib_num;
910         config_attrib_num++;
911     }
912
913     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
914         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
915         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
916         config_attrib_num++;
917     }
918     
919     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
920         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
921     }
922
923     free(entrypoints);
924     return 0;
925 }
926
927 int QuickSyncEncoderImpl::setup_encode()
928 {
929     VAStatus va_status;
930     VASurfaceID *tmp_surfaceid;
931     int codedbuf_size, i;
932     VASurfaceID src_surface[SURFACE_NUM];
933     VASurfaceID ref_surface[SURFACE_NUM];
934     
935     va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
936             &config_attrib[0], config_attrib_num, &config_id);
937     CHECK_VASTATUS(va_status, "vaCreateConfig");
938
939     /* create source surfaces */
940     va_status = vaCreateSurfaces(va_dpy,
941                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
942                                  &src_surface[0], SURFACE_NUM,
943                                  NULL, 0);
944     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
945
946     /* create reference surfaces */
947     va_status = vaCreateSurfaces(va_dpy,
948                                  VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
949                                  &ref_surface[0], SURFACE_NUM,
950                                  NULL, 0);
951     CHECK_VASTATUS(va_status, "vaCreateSurfaces");
952
953     tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
954     memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
955     memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
956     
957     /* Create a context for this encode pipe */
958     va_status = vaCreateContext(va_dpy, config_id,
959                                 frame_width_mbaligned, frame_height_mbaligned,
960                                 VA_PROGRESSIVE,
961                                 tmp_surfaceid, 2 * SURFACE_NUM,
962                                 &context_id);
963     CHECK_VASTATUS(va_status, "vaCreateContext");
964     free(tmp_surfaceid);
965
966     codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
967
968     for (i = 0; i < SURFACE_NUM; i++) {
969         /* create coded buffer once for all
970          * other VA buffers which won't be used again after vaRenderPicture.
971          * so APP can always vaCreateBuffer for every frame
972          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
973          * so VA won't maintain the coded buffer
974          */
975         va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
976                 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
977         CHECK_VASTATUS(va_status, "vaCreateBuffer");
978     }
979
980     /* create OpenGL objects */
981     //glGenFramebuffers(SURFACE_NUM, fbos);
982     
983     for (i = 0; i < SURFACE_NUM; i++) {
984         if (use_zerocopy) {
985             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
986             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
987         } else {
988             gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
989             gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
990
991             // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
992             // buffers, due to potentially differing pitch.
993             glGenBuffers(1, &gl_surfaces[i].pbo);
994             glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
995             glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
996             uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
997             gl_surfaces[i].y_offset = 0;
998             gl_surfaces[i].cbcr_offset = frame_width * frame_height;
999             gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1000             gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1001             glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1002         }
1003     }
1004
1005     for (i = 0; i < SURFACE_NUM; i++) {
1006         gl_surfaces[i].src_surface = src_surface[i];
1007         gl_surfaces[i].ref_surface = ref_surface[i];
1008     }
1009     
1010     return 0;
1011 }
1012
1013 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1014 //
1015 //   2 1 0 [3] 4 8 9
1016 template<class T, class C>
1017 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1018 {
1019         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1020         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1021         sort(middle, end, less_than);
1022 }
1023
1024 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1025 {
1026     if (frame_type == FRAME_B)
1027         return;
1028
1029     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1030
1031     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1032     unique_lock<mutex> lock(storage_task_queue_mutex);
1033
1034     // Insert the new frame at the start of the reference queue.
1035     reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1036
1037     if (reference_frames.size() > num_ref_frames)
1038     {
1039         // The back frame frame is no longer in use as a reference.
1040         int display_frame_num = reference_frames.back().display_number;
1041         assert(surface_for_frame.count(display_frame_num));
1042         release_gl_surface(display_frame_num);
1043         reference_frames.pop_back();
1044     }
1045
1046     // Mark this frame in use as a reference.
1047     assert(surface_for_frame.count(current_display_frame));
1048     ++surface_for_frame[current_display_frame]->refcount;
1049     
1050     current_ref_frame_num++;
1051     if (current_ref_frame_num > MaxFrameNum)
1052         current_ref_frame_num = 0;
1053 }
1054
1055
1056 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1057 {
1058     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1059         return a.frame_idx > b.frame_idx;
1060     };
1061
1062     for (size_t i = 0; i < reference_frames.size(); ++i) {
1063         RefPicList0_P[i] = reference_frames[i].pic;
1064     }
1065     sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1066 }
1067
1068 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1069 {
1070     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1071         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1072     };
1073     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1074         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1075     };
1076
1077     for (size_t i = 0; i < reference_frames.size(); ++i) {
1078         RefPicList0_B[i] = reference_frames[i].pic;
1079         RefPicList1_B[i] = reference_frames[i].pic;
1080     }
1081     sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1082     sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1083 }
1084
1085
1086 int QuickSyncEncoderImpl::render_sequence()
1087 {
1088     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1089     VAStatus va_status;
1090     VAEncMiscParameterBuffer *misc_param;
1091     VAEncMiscParameterRateControl *misc_rate_ctrl;
1092     
1093     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1094     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1095     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1096     seq_param.bits_per_second = frame_bitrate;
1097
1098     seq_param.intra_period = intra_period;
1099     seq_param.intra_idr_period = intra_idr_period;
1100     seq_param.ip_period = ip_period;
1101
1102     seq_param.max_num_ref_frames = num_ref_frames;
1103     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1104     seq_param.time_scale = TIMEBASE * 2;
1105     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1106     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1107     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1108     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1109     seq_param.seq_fields.bits.chroma_format_idc = 1;
1110     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1111     
1112     if (frame_width != frame_width_mbaligned ||
1113         frame_height != frame_height_mbaligned) {
1114         seq_param.frame_cropping_flag = 1;
1115         seq_param.frame_crop_left_offset = 0;
1116         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1117         seq_param.frame_crop_top_offset = 0;
1118         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1119     }
1120     
1121     va_status = vaCreateBuffer(va_dpy, context_id,
1122                                VAEncSequenceParameterBufferType,
1123                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1124     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1125     
1126     va_status = vaCreateBuffer(va_dpy, context_id,
1127                                VAEncMiscParameterBufferType,
1128                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1129                                1, NULL, &rc_param_buf);
1130     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1131     
1132     vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1133     misc_param->type = VAEncMiscParameterTypeRateControl;
1134     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1135     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1136     misc_rate_ctrl->bits_per_second = frame_bitrate;
1137     misc_rate_ctrl->target_percentage = 66;
1138     misc_rate_ctrl->window_size = 1000;
1139     misc_rate_ctrl->initial_qp = initial_qp;
1140     misc_rate_ctrl->min_qp = minimal_qp;
1141     misc_rate_ctrl->basic_unit_size = 0;
1142     vaUnmapBuffer(va_dpy, rc_param_buf);
1143
1144     render_id[0] = seq_param_buf;
1145     render_id[1] = rc_param_buf;
1146     
1147     render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1148     
1149     return 0;
1150 }
1151
1152 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1153 {
1154     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1155     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1156     int PicOrderCntMsb, TopFieldOrderCnt;
1157     
1158     if (frame_type == FRAME_IDR)
1159         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1160     else {
1161         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1162         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1163     }
1164     
1165     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1166         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1167         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1168     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1169              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1170         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1171     else
1172         PicOrderCntMsb = prevPicOrderCntMsb;
1173     
1174     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1175
1176     if (frame_type != FRAME_B) {
1177         PicOrderCntMsb_ref = PicOrderCntMsb;
1178         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1179     }
1180     
1181     return TopFieldOrderCnt;
1182 }
1183
1184 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1185 {
1186     VABufferID pic_param_buf;
1187     VAStatus va_status;
1188     size_t i = 0;
1189
1190     pic_param.CurrPic.picture_id = surf->ref_surface;
1191     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1192     pic_param.CurrPic.flags = 0;
1193     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1194     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1195     CurrentCurrPic = pic_param.CurrPic;
1196
1197     for (i = 0; i < reference_frames.size(); i++) {
1198         pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1199     }
1200     for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1201         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1202         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1203     }
1204     
1205     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1206     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1207     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1208     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1209     pic_param.frame_num = current_ref_frame_num;  // FIXME: is this correct?
1210     pic_param.coded_buf = surf->coded_buf;
1211     pic_param.last_picture = false;  // FIXME
1212     pic_param.pic_init_qp = initial_qp;
1213
1214     va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1215                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1216     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1217
1218     render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1219
1220     return 0;
1221 }
1222
1223 int QuickSyncEncoderImpl::render_packedsequence()
1224 {
1225     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1226     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1227     unsigned int length_in_bits;
1228     unsigned char *packedseq_buffer = NULL;
1229     VAStatus va_status;
1230
1231     length_in_bits = build_packed_seq_buffer(&packedseq_buffer); 
1232     
1233     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1234     
1235     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1236     packedheader_param_buffer.has_emulation_bytes = 0;
1237     va_status = vaCreateBuffer(va_dpy,
1238                                context_id,
1239                                VAEncPackedHeaderParameterBufferType,
1240                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1241                                &packedseq_para_bufid);
1242     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1243
1244     va_status = vaCreateBuffer(va_dpy,
1245                                context_id,
1246                                VAEncPackedHeaderDataBufferType,
1247                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1248                                &packedseq_data_bufid);
1249     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1250
1251     render_id[0] = packedseq_para_bufid;
1252     render_id[1] = packedseq_data_bufid;
1253     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1254
1255     free(packedseq_buffer);
1256     
1257     return 0;
1258 }
1259
1260
1261 int QuickSyncEncoderImpl::render_packedpicture()
1262 {
1263     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1264     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1265     unsigned int length_in_bits;
1266     unsigned char *packedpic_buffer = NULL;
1267     VAStatus va_status;
1268
1269     length_in_bits = build_packed_pic_buffer(&packedpic_buffer); 
1270     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1271     packedheader_param_buffer.bit_length = length_in_bits;
1272     packedheader_param_buffer.has_emulation_bytes = 0;
1273
1274     va_status = vaCreateBuffer(va_dpy,
1275                                context_id,
1276                                VAEncPackedHeaderParameterBufferType,
1277                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1278                                &packedpic_para_bufid);
1279     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1280
1281     va_status = vaCreateBuffer(va_dpy,
1282                                context_id,
1283                                VAEncPackedHeaderDataBufferType,
1284                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1285                                &packedpic_data_bufid);
1286     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1287
1288     render_id[0] = packedpic_para_bufid;
1289     render_id[1] = packedpic_data_bufid;
1290     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1291
1292     free(packedpic_buffer);
1293     
1294     return 0;
1295 }
1296
1297 void QuickSyncEncoderImpl::render_packedslice()
1298 {
1299     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1300     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1301     unsigned int length_in_bits;
1302     unsigned char *packedslice_buffer = NULL;
1303     VAStatus va_status;
1304
1305     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1306     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1307     packedheader_param_buffer.bit_length = length_in_bits;
1308     packedheader_param_buffer.has_emulation_bytes = 0;
1309
1310     va_status = vaCreateBuffer(va_dpy,
1311                                context_id,
1312                                VAEncPackedHeaderParameterBufferType,
1313                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1314                                &packedslice_para_bufid);
1315     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1316
1317     va_status = vaCreateBuffer(va_dpy,
1318                                context_id,
1319                                VAEncPackedHeaderDataBufferType,
1320                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1321                                &packedslice_data_bufid);
1322     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1323
1324     render_id[0] = packedslice_para_bufid;
1325     render_id[1] = packedslice_data_bufid;
1326     render_picture_and_delete(va_dpy, context_id, render_id, 2);
1327
1328     free(packedslice_buffer);
1329 }
1330
1331 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1332 {
1333     VABufferID slice_param_buf;
1334     VAStatus va_status;
1335     int i;
1336
1337     /* one frame, one slice */
1338     slice_param.macroblock_address = 0;
1339     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1340     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1341     if (frame_type == FRAME_IDR) {
1342         if (encoding_frame_num != 0)
1343             ++slice_param.idr_pic_id;
1344     } else if (frame_type == FRAME_P) {
1345         VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1346         update_RefPicList_P(RefPicList0_P);
1347
1348         int refpiclist0_max = h264_maxref & 0xffff;
1349         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1350
1351         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1352             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1353             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1354         }
1355     } else if (frame_type == FRAME_B) {
1356         VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1357         update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1358
1359         int refpiclist0_max = h264_maxref & 0xffff;
1360         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1361
1362         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1363         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1364             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1365             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1366         }
1367
1368         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1369         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1370             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1371             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1372         }
1373     }
1374
1375     slice_param.slice_alpha_c0_offset_div2 = 0;
1376     slice_param.slice_beta_offset_div2 = 0;
1377     slice_param.direct_spatial_mv_pred_flag = 1;
1378     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1379     
1380
1381     if (h264_packedheader &&
1382         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1383         render_packedslice();
1384
1385     va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1386                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1387     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1388
1389     render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1390
1391     return 0;
1392 }
1393
1394
1395
1396 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1397 {    
1398         VACodedBufferSegment *buf_list = NULL;
1399         VAStatus va_status;
1400
1401         string data;
1402
1403         va_status = vaMapBuffer(va_dpy, surf->coded_buf, (void **)(&buf_list));
1404         CHECK_VASTATUS(va_status, "vaMapBuffer");
1405         while (buf_list != NULL) {
1406                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1407                 buf_list = (VACodedBufferSegment *) buf_list->next;
1408         }
1409         vaUnmapBuffer(va_dpy, surf->coded_buf);
1410
1411         static int frameno = 0;
1412         print_latency("Current QuickSync latency (video inputs → disk mux):",
1413                 task.received_ts, (task.frame_type == FRAME_B), &frameno);
1414
1415         {
1416                 // Add video.
1417                 AVPacket pkt;
1418                 memset(&pkt, 0, sizeof(pkt));
1419                 pkt.buf = nullptr;
1420                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1421                 pkt.size = data.size();
1422                 pkt.stream_index = 0;
1423                 if (task.frame_type == FRAME_IDR) {
1424                         pkt.flags = AV_PKT_FLAG_KEY;
1425                 } else {
1426                         pkt.flags = 0;
1427                 }
1428                 pkt.duration = task.duration;
1429                 if (file_mux) {
1430                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1431                 }
1432                 if (!global_flags.uncompressed_video_to_http &&
1433                     !global_flags.x264_video_to_http) {
1434                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1435                 }
1436         }
1437 }
1438
1439
1440 // this is weird. but it seems to put a new frame onto the queue
1441 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1442 {
1443         unique_lock<mutex> lock(storage_task_queue_mutex);
1444         storage_task_queue.push(move(task));
1445         storage_task_queue_changed.notify_all();
1446 }
1447
1448 void QuickSyncEncoderImpl::storage_task_thread()
1449 {
1450         pthread_setname_np(pthread_self(), "QS_Storage");
1451         for ( ;; ) {
1452                 storage_task current;
1453                 GLSurface *surf;
1454                 {
1455                         // wait until there's an encoded frame  
1456                         unique_lock<mutex> lock(storage_task_queue_mutex);
1457                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1458                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1459                         current = move(storage_task_queue.front());
1460                         storage_task_queue.pop();
1461                         surf = surface_for_frame[current.display_order];
1462                         assert(surf != nullptr);
1463                 }
1464
1465                 VAStatus va_status;
1466
1467                 size_t display_order = current.display_order;
1468                 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1469            
1470                 // waits for data, then saves it to disk.
1471                 va_status = vaSyncSurface(va_dpy, surf->src_surface);
1472                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1473                 save_codeddata(surf, move(current));
1474
1475                 // Unlock the frame, and all its references.
1476                 {
1477                         unique_lock<mutex> lock(storage_task_queue_mutex);
1478                         release_gl_surface(display_order);
1479
1480                         for (size_t frame_num : ref_display_frame_numbers) {
1481                                 release_gl_surface(frame_num);
1482                         }
1483                 }
1484         }
1485 }
1486
1487 void QuickSyncEncoderImpl::release_encode()
1488 {
1489         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1490                 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1491                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1492                 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1493         }
1494
1495         vaDestroyContext(va_dpy, context_id);
1496         vaDestroyConfig(va_dpy, config_id);
1497 }
1498
1499 void QuickSyncEncoderImpl::release_gl_resources()
1500 {
1501         assert(is_shutdown);
1502         if (has_released_gl_resources) {
1503                 return;
1504         }
1505
1506         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1507                 if (!use_zerocopy) {
1508                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1509                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1510                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1511                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1512                 }
1513                 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1514                 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1515         }
1516
1517         has_released_gl_resources = true;
1518 }
1519
1520 int QuickSyncEncoderImpl::deinit_va()
1521
1522     vaTerminate(va_dpy);
1523
1524     va_close_display(va_dpy);
1525
1526     return 0;
1527 }
1528
1529 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1530         : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator)
1531 {
1532         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1533         open_output_file(filename);
1534         file_audio_encoder->add_mux(file_mux.get());
1535
1536         frame_width_mbaligned = (frame_width + 15) & (~15);
1537         frame_height_mbaligned = (frame_height + 15) & (~15);
1538
1539         //print_input();
1540
1541         if (global_flags.x264_video_to_http) {
1542                 assert(x264_encoder != nullptr);
1543         } else {
1544                 assert(x264_encoder == nullptr);
1545         }
1546
1547         init_va(va_display);
1548         setup_encode();
1549
1550         memset(&seq_param, 0, sizeof(seq_param));
1551         memset(&pic_param, 0, sizeof(pic_param));
1552         memset(&slice_param, 0, sizeof(slice_param));
1553
1554         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1555
1556         encode_thread = thread([this]{
1557                 //SDL_GL_MakeCurrent(window, context);
1558                 QOpenGLContext *context = create_context(this->surface);
1559                 eglBindAPI(EGL_OPENGL_API);
1560                 if (!make_current(context, this->surface)) {
1561                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1562                                 eglGetError());
1563                         exit(1);
1564                 }
1565                 encode_thread_func();
1566                 delete_context(context);
1567         });
1568 }
1569
1570 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1571 {
1572         shutdown();
1573         release_gl_resources();
1574 }
1575
1576 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1577 {
1578         for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1579                 if (gl_surfaces[i].refcount == 0) {
1580                         ++gl_surfaces[i].refcount;
1581                         return &gl_surfaces[i];
1582                 }
1583         }
1584         return nullptr;
1585 }
1586
1587 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1588 {
1589         assert(surface_for_frame.count(display_frame_num));
1590         QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1591         if (--surf->refcount == 0) {
1592                 assert(surface_for_frame.count(display_frame_num));
1593                 surface_for_frame.erase(display_frame_num);
1594                 storage_task_queue_changed.notify_all();
1595         }
1596 }
1597
1598 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1599 {
1600         assert(!is_shutdown);
1601         GLSurface *surf = nullptr;
1602         {
1603                 // Wait until this frame slot is done encoding.
1604                 unique_lock<mutex> lock(storage_task_queue_mutex);
1605                 surf = allocate_gl_surface();
1606                 if (surf == nullptr) {
1607                         fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1608                                 current_storage_frame);
1609                         storage_task_queue_changed.wait(lock, [this, &surf]{
1610                                 if (storage_thread_should_quit)
1611                                         return true;
1612                                 surf = allocate_gl_surface();
1613                                 return surf != nullptr;
1614                         });
1615                 }
1616                 if (storage_thread_should_quit) return false;
1617                 assert(surf != nullptr);
1618                 surface_for_frame[current_storage_frame] = surf;
1619         }
1620
1621         *y_tex = surf->y_tex;
1622         *cbcr_tex = surf->cbcr_tex;
1623
1624         VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1625         CHECK_VASTATUS(va_status, "vaDeriveImage");
1626
1627         if (use_zerocopy) {
1628                 VABufferInfo buf_info;
1629                 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1630                 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1631                 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1632
1633                 // Create Y image.
1634                 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1635                 EGLint y_attribs[] = {
1636                         EGL_WIDTH, frame_width,
1637                         EGL_HEIGHT, frame_height,
1638                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1639                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1640                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1641                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1642                         EGL_NONE
1643                 };
1644
1645                 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1646                 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1647
1648                 // Associate Y image to a texture.
1649                 glBindTexture(GL_TEXTURE_2D, *y_tex);
1650                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1651
1652                 // Create CbCr image.
1653                 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1654                 EGLint cbcr_attribs[] = {
1655                         EGL_WIDTH, frame_width,
1656                         EGL_HEIGHT, frame_height,
1657                         EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1658                         EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1659                         EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1660                         EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1661                         EGL_NONE
1662                 };
1663
1664                 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1665                 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1666
1667                 // Associate CbCr image to a texture.
1668                 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1669                 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1670         }
1671
1672         return true;
1673 }
1674
1675 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1676 {
1677         assert(!is_shutdown);
1678         file_audio_encoder->encode_audio(audio, pts + global_delay());
1679 }
1680
1681 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1682 {
1683         assert(!is_shutdown);
1684
1685         if (!use_zerocopy) {
1686                 GLSurface *surf;
1687                 {
1688                         unique_lock<mutex> lock(storage_task_queue_mutex);
1689                         surf = surface_for_frame[current_storage_frame];
1690                         assert(surf != nullptr);
1691                 }
1692
1693                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1694                 check_error();
1695
1696                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1697                 check_error();
1698
1699                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1700                 check_error();
1701                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1702                 check_error();
1703
1704                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1705                 check_error();
1706                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1707                 check_error();
1708
1709                 glBindTexture(GL_TEXTURE_2D, 0);
1710                 check_error();
1711                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1712                 check_error();
1713
1714                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1715                 check_error();
1716         }
1717
1718         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1719         check_error();
1720         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1721         check_error();
1722
1723         {
1724                 unique_lock<mutex> lock(frame_queue_mutex);
1725                 pending_video_frames.push(PendingFrame{ fence, input_frames, pts, duration });
1726                 ++current_storage_frame;
1727         }
1728         frame_queue_nonempty.notify_all();
1729         return fence;
1730 }
1731
1732 void QuickSyncEncoderImpl::shutdown()
1733 {
1734         if (is_shutdown) {
1735                 return;
1736         }
1737
1738         {
1739                 unique_lock<mutex> lock(frame_queue_mutex);
1740                 encode_thread_should_quit = true;
1741                 frame_queue_nonempty.notify_all();
1742         }
1743         encode_thread.join();
1744         {
1745                 unique_lock<mutex> lock(storage_task_queue_mutex);
1746                 storage_thread_should_quit = true;
1747                 frame_queue_nonempty.notify_all();
1748                 storage_task_queue_changed.notify_all();
1749         }
1750         storage_thread.join();
1751
1752         // Encode any leftover audio in the queues, and also any delayed frames.
1753         file_audio_encoder->encode_last_audio();
1754
1755         release_encode();
1756         deinit_va();
1757         file_mux.reset();
1758         is_shutdown = true;
1759 }
1760
1761 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1762 {
1763         AVFormatContext *avctx = avformat_alloc_context();
1764         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1765         assert(filename.size() < sizeof(avctx->filename) - 1);
1766         strcpy(avctx->filename, filename.c_str());
1767
1768         string url = "file:" + filename;
1769         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1770         if (ret < 0) {
1771                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1772                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1773                 exit(1);
1774         }
1775
1776         string video_extradata = "";  // FIXME: See other comment about global headers.
1777         AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1778         file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1779                 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1780 }
1781
1782 void QuickSyncEncoderImpl::encode_thread_func()
1783 {
1784         pthread_setname_np(pthread_self(), "QS_Encode");
1785
1786         int64_t last_dts = -1;
1787         int gop_start_display_frame_num = 0;
1788         for (int display_frame_num = 0; ; ++display_frame_num) {
1789                 // Wait for the frame to be in the queue. Note that this only means
1790                 // we started rendering it.
1791                 PendingFrame frame;
1792                 {
1793                         unique_lock<mutex> lock(frame_queue_mutex);
1794                         frame_queue_nonempty.wait(lock, [this]{
1795                                 return encode_thread_should_quit || !pending_video_frames.empty();
1796                         });
1797                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1798                                 // We may have queued frames left in the reorder buffer
1799                                 // that were supposed to be B-frames, but have no P-frame
1800                                 // to be encoded against. If so, encode them all as
1801                                 // P-frames instead. Note that this happens under the mutex,
1802                                 // but nobody else uses it at this point, since we're shutting down,
1803                                 // so there's no contention.
1804                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1805                                 return;
1806                         } else {
1807                                 frame = move(pending_video_frames.front());
1808                                 pending_video_frames.pop();
1809                         }
1810                 }
1811
1812                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1813                 // Note that this implicitly waits for the frame to be done rendering.
1814                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1815                 reorder_buffer[display_frame_num] = move(frame);
1816
1817                 // Now encode as many QuickSync frames as we can using the frames we have available.
1818                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1819                 for ( ;; ) {
1820                         int pts_lag;
1821                         int frame_type, quicksync_display_frame_num;
1822                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1823                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1824                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1825                                 break;
1826                         }
1827                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1828                         reorder_buffer.erase(quicksync_display_frame_num);
1829
1830                         if (frame_type == FRAME_IDR) {
1831                                 // Release any reference frames from the previous GOP.
1832                                 for (const ReferenceFrame &frame : reference_frames) {
1833                                         release_gl_surface(frame.display_number);
1834                                 }
1835                                 reference_frames.clear();
1836                                 current_ref_frame_num = 0;
1837                                 gop_start_display_frame_num = quicksync_display_frame_num;
1838                         }
1839
1840                         // Determine the dts of this frame.
1841                         int64_t dts;
1842                         if (pts_lag == -1) {
1843                                 assert(last_dts != -1);
1844                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1845                         } else {
1846                                 dts = frame.pts - pts_lag;
1847                         }
1848                         last_dts = dts;
1849
1850                         encode_frame(frame, quicksync_encoding_frame_num, quicksync_display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
1851                         ++quicksync_encoding_frame_num;
1852                 }
1853         }
1854 }
1855
1856 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1857 {
1858         if (reorder_buffer.empty()) {
1859                 return;
1860         }
1861
1862         for (auto &pending_frame : reorder_buffer) {
1863                 int display_frame_num = pending_frame.first;
1864                 assert(display_frame_num > 0);
1865                 PendingFrame frame = move(pending_frame.second);
1866                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1867                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1868                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
1869                 last_dts = dts;
1870         }
1871 }
1872
1873 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1874 {
1875         AVPacket pkt;
1876         memset(&pkt, 0, sizeof(pkt));
1877         pkt.buf = nullptr;
1878         pkt.data = const_cast<uint8_t *>(data);
1879         pkt.size = frame_width * frame_height * 2;
1880         pkt.stream_index = 0;
1881         pkt.flags = AV_PKT_FLAG_KEY;
1882         pkt.duration = duration;
1883         stream_mux->add_packet(pkt, pts, pts);
1884 }
1885
1886 namespace {
1887
1888 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1889 {
1890         if (src_width == dst_pitch) {
1891                 memcpy(dst, src, src_width * height);
1892         } else {
1893                 for (size_t y = 0; y < height; ++y) {
1894                         const uint8_t *sptr = src + y * src_width;
1895                         uint8_t *dptr = dst + y * dst_pitch;
1896                         memcpy(dptr, sptr, src_width);
1897                 }
1898         }
1899 }
1900
1901 }  // namespace
1902
1903 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1904 {
1905         // Wait for the GPU to be done with the frame.
1906         GLenum sync_status;
1907         do {
1908                 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1909                 check_error();
1910         } while (sync_status == GL_TIMEOUT_EXPIRED);
1911         assert(sync_status != GL_WAIT_FAILED);
1912
1913         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1914         static int frameno = 0;
1915         print_latency("Current mixer latency (video inputs → ready for encode):",
1916                 received_ts, false, &frameno);
1917
1918         // Release back any input frames we needed to render this frame.
1919         frame.input_frames.clear();
1920
1921         GLSurface *surf;
1922         {
1923                 unique_lock<mutex> lock(storage_task_queue_mutex);
1924                 surf = surface_for_frame[display_frame_num];
1925                 assert(surf != nullptr);
1926         }
1927         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1928         if (global_flags.uncompressed_video_to_http) {
1929                 add_packet_for_uncompressed_frame(pts, duration, data);
1930         } else if (global_flags.x264_video_to_http) {
1931                 x264_encoder->add_frame(pts, duration, data, received_ts);
1932         }
1933 }
1934
1935 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1936                                         int frame_type, int64_t pts, int64_t dts, int64_t duration)
1937 {
1938         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1939
1940         GLSurface *surf;
1941         {
1942                 unique_lock<mutex> lock(storage_task_queue_mutex);
1943                 surf = surface_for_frame[display_frame_num];
1944                 assert(surf != nullptr);
1945         }
1946         VAStatus va_status;
1947
1948         if (use_zerocopy) {
1949                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1950                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1951                 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1952                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1953         } else {
1954                 // Upload the frame to VA-API.
1955                 unsigned char *surface_p = nullptr;
1956                 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1957
1958                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1959                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1960
1961                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1962                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1963
1964                 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1965                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1966         }
1967
1968         va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1969         CHECK_VASTATUS(va_status, "vaDestroyImage");
1970
1971         // Schedule the frame for encoding.
1972         VASurfaceID va_surface = surf->src_surface;
1973         va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1974         CHECK_VASTATUS(va_status, "vaBeginPicture");
1975
1976         if (frame_type == FRAME_IDR) {
1977                 // FIXME: If the mux wants global headers, we should not put the
1978                 // SPS/PPS before each IDR frame, but rather put it into the
1979                 // codec extradata (formatted differently?).
1980                 render_sequence();
1981                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1982                 if (h264_packedheader) {
1983                         render_packedsequence();
1984                         render_packedpicture();
1985                 }
1986         } else {
1987                 //render_sequence();
1988                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1989         }
1990         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1991
1992         va_status = vaEndPicture(va_dpy, context_id);
1993         CHECK_VASTATUS(va_status, "vaEndPicture");
1994
1995         update_ReferenceFrames(display_frame_num, frame_type);
1996
1997         vector<size_t> ref_display_frame_numbers;
1998
1999         // Lock the references for this frame; otherwise, they could be
2000         // rendered to before this frame is done encoding.
2001         {
2002                 unique_lock<mutex> lock(storage_task_queue_mutex);
2003                 for (const ReferenceFrame &frame : reference_frames) {
2004                         assert(surface_for_frame.count(frame.display_number));
2005                         ++surface_for_frame[frame.display_number]->refcount;
2006                         ref_display_frame_numbers.push_back(frame.display_number);
2007                 }
2008         }
2009
2010         // so now the data is done encoding (well, async job kicked off)...
2011         // we send that to the storage thread
2012         storage_task tmp;
2013         tmp.display_order = display_frame_num;
2014         tmp.frame_type = frame_type;
2015         tmp.pts = pts;
2016         tmp.dts = dts;
2017         tmp.duration = duration;
2018         tmp.received_ts = received_ts;
2019         tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2020         storage_task_enqueue(move(tmp));
2021 }
2022
2023 // Proxy object.
2024 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
2025         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2026
2027 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2028 QuickSyncEncoder::~QuickSyncEncoder() {}
2029
2030 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2031 {
2032         impl->add_audio(pts, audio);
2033 }
2034
2035 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2036 {
2037         return impl->begin_frame(y_tex, cbcr_tex);
2038 }
2039
2040 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2041 {
2042         return impl->end_frame(pts, duration, input_frames);
2043 }
2044
2045 void QuickSyncEncoder::shutdown()
2046 {
2047         impl->shutdown();
2048 }
2049
2050 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2051 {
2052         impl->set_stream_mux(mux);
2053 }
2054
2055 int64_t QuickSyncEncoder::global_delay() const {
2056         return impl->global_delay();
2057 }