]> git.sesse.net Git - nageru/blob - quicksync_encoder.cpp
Support audio-only FFmpeg inputs. Somewhat wonky, though.
[nageru] / quicksync_encoder.cpp
1 #include "quicksync_encoder.h"
2
3 #include <movit/image_format.h>
4 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
5 #include <movit/util.h>
6
7 #include <EGL/eglplatform.h>
8 #include <X11/Xlib.h>
9 #include <assert.h>
10 #include <epoxy/egl.h>
11 #include <fcntl.h>
12 #include <glob.h>
13 #include <pthread.h>
14 #include <stdio.h>
15 #include <stdlib.h>
16 #include <string.h>
17 #include <unistd.h>
18 #include <va/va.h>
19 #include <va/va_drm.h>
20 #include <va/va_drmcommon.h>
21 #include <va/va_enc_h264.h>
22 #include <va/va_x11.h>
23 #include <algorithm>
24 #include <chrono>
25 #include <condition_variable>
26 #include <cstddef>
27 #include <cstdint>
28 #include <functional>
29 #include <map>
30 #include <memory>
31 #include <mutex>
32 #include <queue>
33 #include <stack>
34 #include <string>
35 #include <thread>
36 #include <utility>
37
38 extern "C" {
39
40 #include <libavcodec/avcodec.h>
41 #include <libavformat/avio.h>
42 #include <libavutil/error.h>
43 #include <libdrm/drm_fourcc.h>
44
45 }  // namespace
46
47 #include "audio_encoder.h"
48 #include "context.h"
49 #include "defs.h"
50 #include "disk_space_estimator.h"
51 #include "ffmpeg_raii.h"
52 #include "flags.h"
53 #include "mux.h"
54 #include "print_latency.h"
55 #include "quicksync_encoder_impl.h"
56 #include "ref_counted_frame.h"
57 #include "timebase.h"
58 #include "x264_encoder.h"
59
60 using namespace movit;
61 using namespace std;
62 using namespace std::chrono;
63 using namespace std::placeholders;
64
65 class QOpenGLContext;
66 class QSurface;
67
68 namespace {
69
70 // These need to survive several QuickSyncEncoderImpl instances,
71 // so they are outside.
72 once_flag quick_sync_metrics_inited;
73 LatencyHistogram mixer_latency_histogram, qs_latency_histogram;
74 MuxMetrics current_file_mux_metrics, total_mux_metrics;
75 std::atomic<double> metric_current_file_start_time_seconds{0.0 / 0.0};
76 std::atomic<int64_t> metric_quick_sync_stalled_frames{0};
77
78 }  // namespace
79
80 #define CHECK_VASTATUS(va_status, func)                                 \
81     if (va_status != VA_STATUS_SUCCESS) {                               \
82         fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
83         exit(1);                                                        \
84     }
85
86 #undef BUFFER_OFFSET
87 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
88
89 //#include "loadsurface.h"
90
91 #define NAL_REF_IDC_NONE        0
92 #define NAL_REF_IDC_LOW         1
93 #define NAL_REF_IDC_MEDIUM      2
94 #define NAL_REF_IDC_HIGH        3
95
96 #define NAL_NON_IDR             1
97 #define NAL_IDR                 5
98 #define NAL_SPS                 7
99 #define NAL_PPS                 8
100 #define NAL_SEI                 6
101
102 #define SLICE_TYPE_P            0
103 #define SLICE_TYPE_B            1
104 #define SLICE_TYPE_I            2
105 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
106 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
107 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
108
109
110 #define ENTROPY_MODE_CAVLC      0
111 #define ENTROPY_MODE_CABAC      1
112
113 #define PROFILE_IDC_BASELINE    66
114 #define PROFILE_IDC_MAIN        77
115 #define PROFILE_IDC_HIGH        100
116    
117 #define BITSTREAM_ALLOCATE_STEPPING     4096
118
119 static constexpr unsigned int MaxFrameNum = (2<<16);
120 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
121 static constexpr unsigned int Log2MaxFrameNum = 16;
122 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
123
124 using namespace std;
125
126 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
127 // but if we don't delete it here, we get leaks. The GStreamer implementation
128 // does the same.
129 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
130 {
131     VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
132     CHECK_VASTATUS(va_status, "vaRenderPicture");
133
134     for (int i = 0; i < num_buffers; ++i) {
135         va_status = vaDestroyBuffer(dpy, buffers[i]);
136         CHECK_VASTATUS(va_status, "vaDestroyBuffer");
137     }
138 }
139
140 static unsigned int 
141 va_swap32(unsigned int val)
142 {
143     unsigned char *pval = (unsigned char *)&val;
144
145     return ((pval[0] << 24)     |
146             (pval[1] << 16)     |
147             (pval[2] << 8)      |
148             (pval[3] << 0));
149 }
150
151 static void
152 bitstream_start(bitstream *bs)
153 {
154     bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
155     bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
156     bs->bit_offset = 0;
157 }
158
159 static void
160 bitstream_end(bitstream *bs)
161 {
162     int pos = (bs->bit_offset >> 5);
163     int bit_offset = (bs->bit_offset & 0x1f);
164     int bit_left = 32 - bit_offset;
165
166     if (bit_offset) {
167         bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
168     }
169 }
170  
171 static void
172 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
173 {
174     int pos = (bs->bit_offset >> 5);
175     int bit_offset = (bs->bit_offset & 0x1f);
176     int bit_left = 32 - bit_offset;
177
178     if (!size_in_bits)
179         return;
180
181     bs->bit_offset += size_in_bits;
182
183     if (bit_left > size_in_bits) {
184         bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
185     } else {
186         size_in_bits -= bit_left;
187         if (bit_left >= 32) {
188             bs->buffer[pos] = (val >> size_in_bits);
189         } else {
190             bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
191         }
192         bs->buffer[pos] = va_swap32(bs->buffer[pos]);
193
194         if (pos + 1 == bs->max_size_in_dword) {
195             bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
196             bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
197         }
198
199         bs->buffer[pos + 1] = val;
200     }
201 }
202
203 static void
204 bitstream_put_ue(bitstream *bs, unsigned int val)
205 {
206     int size_in_bits = 0;
207     int tmp_val = ++val;
208
209     while (tmp_val) {
210         tmp_val >>= 1;
211         size_in_bits++;
212     }
213
214     bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
215     bitstream_put_ui(bs, val, size_in_bits);
216 }
217
218 static void
219 bitstream_put_se(bitstream *bs, int val)
220 {
221     unsigned int new_val;
222
223     if (val <= 0)
224         new_val = -2 * val;
225     else
226         new_val = 2 * val - 1;
227
228     bitstream_put_ue(bs, new_val);
229 }
230
231 static void
232 bitstream_byte_aligning(bitstream *bs, int bit)
233 {
234     int bit_offset = (bs->bit_offset & 0x7);
235     int bit_left = 8 - bit_offset;
236     int new_val;
237
238     if (!bit_offset)
239         return;
240
241     assert(bit == 0 || bit == 1);
242
243     if (bit)
244         new_val = (1 << bit_left) - 1;
245     else
246         new_val = 0;
247
248     bitstream_put_ui(bs, new_val, bit_left);
249 }
250
251 static void 
252 rbsp_trailing_bits(bitstream *bs)
253 {
254     bitstream_put_ui(bs, 1, 1);
255     bitstream_byte_aligning(bs, 0);
256 }
257
258 static void nal_start_code_prefix(bitstream *bs)
259 {
260     bitstream_put_ui(bs, 0x00000001, 32);
261 }
262
263 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
264 {
265     bitstream_put_ui(bs, 0, 1);                /* forbidden_zero_bit: 0 */
266     bitstream_put_ui(bs, nal_ref_idc, 2);
267     bitstream_put_ui(bs, nal_unit_type, 5);
268 }
269
270 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
271 {
272     int profile_idc = PROFILE_IDC_BASELINE;
273
274     if (h264_profile  == VAProfileH264High)
275         profile_idc = PROFILE_IDC_HIGH;
276     else if (h264_profile  == VAProfileH264Main)
277         profile_idc = PROFILE_IDC_MAIN;
278
279     bitstream_put_ui(bs, profile_idc, 8);               /* profile_idc */
280     bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1);                         /* constraint_set0_flag */
281     bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1);                         /* constraint_set1_flag */
282     bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1);                         /* constraint_set2_flag */
283     bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1);                         /* constraint_set3_flag */
284     bitstream_put_ui(bs, 0, 4);                         /* reserved_zero_4bits */
285     bitstream_put_ui(bs, seq_param.level_idc, 8);      /* level_idc */
286     bitstream_put_ue(bs, seq_param.seq_parameter_set_id);      /* seq_parameter_set_id */
287
288     if ( profile_idc == PROFILE_IDC_HIGH) {
289         bitstream_put_ue(bs, 1);        /* chroma_format_idc = 1, 4:2:0 */ 
290         bitstream_put_ue(bs, 0);        /* bit_depth_luma_minus8 */
291         bitstream_put_ue(bs, 0);        /* bit_depth_chroma_minus8 */
292         bitstream_put_ui(bs, 0, 1);     /* qpprime_y_zero_transform_bypass_flag */
293         bitstream_put_ui(bs, 0, 1);     /* seq_scaling_matrix_present_flag */
294     }
295
296     bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
297     bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type);        /* pic_order_cnt_type */
298
299     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
300         bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4);     /* log2_max_pic_order_cnt_lsb_minus4 */
301     else {
302         assert(0);
303     }
304
305     bitstream_put_ue(bs, seq_param.max_num_ref_frames);        /* num_ref_frames */
306     bitstream_put_ui(bs, 0, 1);                                 /* gaps_in_frame_num_value_allowed_flag */
307
308     bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1);  /* pic_width_in_mbs_minus1 */
309     bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
310     bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1);    /* frame_mbs_only_flag */
311
312     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
313         assert(0);
314     }
315
316     bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1);      /* direct_8x8_inference_flag */
317     bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1);            /* frame_cropping_flag */
318
319     if (seq_param.frame_cropping_flag) {
320         bitstream_put_ue(bs, seq_param.frame_crop_left_offset);        /* frame_crop_left_offset */
321         bitstream_put_ue(bs, seq_param.frame_crop_right_offset);       /* frame_crop_right_offset */
322         bitstream_put_ue(bs, seq_param.frame_crop_top_offset);         /* frame_crop_top_offset */
323         bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset);      /* frame_crop_bottom_offset */
324     }
325     
326     //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
327     if ( false ) {
328         bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
329     } else {
330         // See H.264 annex E for the definition of this header.
331         bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
332         bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
333         bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
334         bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
335         {
336             bitstream_put_ui(bs, 5, 3);  /* video_format (5 = Unspecified) */
337             bitstream_put_ui(bs, 0, 1);  /* video_full_range_flag */
338             bitstream_put_ui(bs, 1, 1);  /* colour_description_present_flag */
339             {
340                 bitstream_put_ui(bs, 1, 8);  /* colour_primaries (1 = BT.709) */
341                 bitstream_put_ui(bs, 13, 8);  /* transfer_characteristics (13 = sRGB) */
342                 if (ycbcr_coefficients == YCBCR_REC_709) {
343                     bitstream_put_ui(bs, 1, 8);  /* matrix_coefficients (1 = BT.709) */
344                 } else {
345                     assert(ycbcr_coefficients == YCBCR_REC_601);
346                     bitstream_put_ui(bs, 6, 8);  /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
347                 }
348             }
349         }
350         bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
351         bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
352         {
353             bitstream_put_ui(bs, 1, 32);  // FPS
354             bitstream_put_ui(bs, TIMEBASE * 2, 32);  // FPS
355             bitstream_put_ui(bs, 1, 1);
356         }
357         bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
358         {
359             // hrd_parameters 
360             bitstream_put_ue(bs, 0);    /* cpb_cnt_minus1 */
361             bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
362             bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
363            
364             bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
365             bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
366             bitstream_put_ui(bs, 1, 1);  /* cbr_flag[0] */
367
368             bitstream_put_ui(bs, 23, 5);   /* initial_cpb_removal_delay_length_minus1 */
369             bitstream_put_ui(bs, 23, 5);   /* cpb_removal_delay_length_minus1 */
370             bitstream_put_ui(bs, 23, 5);   /* dpb_output_delay_length_minus1 */
371             bitstream_put_ui(bs, 23, 5);   /* time_offset_length  */
372         }
373         bitstream_put_ui(bs, 0, 1);   /* vcl_hrd_parameters_present_flag */
374         bitstream_put_ui(bs, 0, 1);   /* low_delay_hrd_flag */ 
375
376         bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
377         bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
378     }
379
380     rbsp_trailing_bits(bs);     /* rbsp_trailing_bits */
381 }
382
383
384 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
385 {
386     bitstream_put_ue(bs, pic_param.pic_parameter_set_id);      /* pic_parameter_set_id */
387     bitstream_put_ue(bs, pic_param.seq_parameter_set_id);      /* seq_parameter_set_id */
388
389     bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1);  /* entropy_coding_mode_flag */
390
391     bitstream_put_ui(bs, 0, 1);                         /* pic_order_present_flag: 0 */
392
393     bitstream_put_ue(bs, 0);                            /* num_slice_groups_minus1 */
394
395     bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1);      /* num_ref_idx_l0_active_minus1 */
396     bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1);      /* num_ref_idx_l1_active_minus1 1 */
397
398     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1);     /* weighted_pred_flag: 0 */
399     bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2);     /* weighted_bipred_idc: 0 */
400
401     bitstream_put_se(bs, pic_param.pic_init_qp - 26);  /* pic_init_qp_minus26 */
402     bitstream_put_se(bs, 0);                            /* pic_init_qs_minus26 */
403     bitstream_put_se(bs, 0);                            /* chroma_qp_index_offset */
404
405     bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
406     bitstream_put_ui(bs, 0, 1);                         /* constrained_intra_pred_flag */
407     bitstream_put_ui(bs, 0, 1);                         /* redundant_pic_cnt_present_flag */
408     
409     /* more_rbsp_data */
410     bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1);    /*transform_8x8_mode_flag */
411     bitstream_put_ui(bs, 0, 1);                         /* pic_scaling_matrix_present_flag */
412     bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset );    /*second_chroma_qp_index_offset */
413
414     rbsp_trailing_bits(bs);
415 }
416
417 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
418 {
419     int first_mb_in_slice = slice_param.macroblock_address;
420
421     bitstream_put_ue(bs, first_mb_in_slice);        /* first_mb_in_slice: 0 */
422     bitstream_put_ue(bs, slice_param.slice_type);   /* slice_type */
423     bitstream_put_ue(bs, slice_param.pic_parameter_set_id);        /* pic_parameter_set_id: 0 */
424     bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
425
426     /* frame_mbs_only_flag == 1 */
427     if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
428         /* FIXME: */
429         assert(0);
430     }
431
432     if (pic_param.pic_fields.bits.idr_pic_flag)
433         bitstream_put_ue(bs, slice_param.idr_pic_id);           /* idr_pic_id: 0 */
434
435     if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
436         bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
437         /* pic_order_present_flag == 0 */
438     } else {
439         /* FIXME: */
440         assert(0);
441     }
442
443     /* redundant_pic_cnt_present_flag == 0 */
444     /* slice type */
445     if (IS_P_SLICE(slice_param.slice_type)) {
446         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);            /* num_ref_idx_active_override_flag: */
447
448         if (slice_param.num_ref_idx_active_override_flag)
449             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
450
451         /* ref_pic_list_reordering */
452         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
453     } else if (IS_B_SLICE(slice_param.slice_type)) {
454         bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1);            /* direct_spatial_mv_pred: 1 */
455
456         bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1);       /* num_ref_idx_active_override_flag: */
457
458         if (slice_param.num_ref_idx_active_override_flag) {
459             bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
460             bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
461         }
462
463         /* ref_pic_list_reordering */
464         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l0: 0 */
465         bitstream_put_ui(bs, 0, 1);            /* ref_pic_list_reordering_flag_l1: 0 */
466     }
467
468     if ((pic_param.pic_fields.bits.weighted_pred_flag &&
469          IS_P_SLICE(slice_param.slice_type)) ||
470         ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
471          IS_B_SLICE(slice_param.slice_type))) {
472         /* FIXME: fill weight/offset table */
473         assert(0);
474     }
475
476     /* dec_ref_pic_marking */
477     if (pic_param.pic_fields.bits.reference_pic_flag) {     /* nal_ref_idc != 0 */
478         unsigned char no_output_of_prior_pics_flag = 0;
479         unsigned char long_term_reference_flag = 0;
480         unsigned char adaptive_ref_pic_marking_mode_flag = 0;
481
482         if (pic_param.pic_fields.bits.idr_pic_flag) {
483             bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1);            /* no_output_of_prior_pics_flag: 0 */
484             bitstream_put_ui(bs, long_term_reference_flag, 1);            /* long_term_reference_flag: 0 */
485         } else {
486             bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1);            /* adaptive_ref_pic_marking_mode_flag: 0 */
487         }
488     }
489
490     if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
491         !IS_I_SLICE(slice_param.slice_type))
492         bitstream_put_ue(bs, slice_param.cabac_init_idc);               /* cabac_init_idc: 0 */
493
494     bitstream_put_se(bs, slice_param.slice_qp_delta);                   /* slice_qp_delta: 0 */
495
496     /* ignore for SP/SI */
497
498     if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
499         bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc);           /* disable_deblocking_filter_idc: 0 */
500
501         if (slice_param.disable_deblocking_filter_idc != 1) {
502             bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2);          /* slice_alpha_c0_offset_div2: 2 */
503             bitstream_put_se(bs, slice_param.slice_beta_offset_div2);              /* slice_beta_offset_div2: 2 */
504         }
505     }
506
507     if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
508         bitstream_byte_aligning(bs, 1);
509     }
510 }
511
512 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
513 {
514     bitstream bs;
515
516     bitstream_start(&bs);
517     nal_start_code_prefix(&bs);
518     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
519     pps_rbsp(&bs);
520     bitstream_end(&bs);
521
522     *header_buffer = (unsigned char *)bs.buffer;
523     return bs.bit_offset;
524 }
525
526 int
527 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
528 {
529     bitstream bs;
530
531     bitstream_start(&bs);
532     nal_start_code_prefix(&bs);
533     nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
534     sps_rbsp(ycbcr_coefficients, &bs);
535     bitstream_end(&bs);
536
537     *header_buffer = (unsigned char *)bs.buffer;
538     return bs.bit_offset;
539 }
540
541 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
542 {
543     bitstream bs;
544     int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
545     int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
546
547     bitstream_start(&bs);
548     nal_start_code_prefix(&bs);
549
550     if (IS_I_SLICE(slice_param.slice_type)) {
551         nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
552     } else if (IS_P_SLICE(slice_param.slice_type)) {
553         nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
554     } else {
555         assert(IS_B_SLICE(slice_param.slice_type));
556         nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
557     }
558
559     slice_header(&bs);
560     bitstream_end(&bs);
561
562     *header_buffer = (unsigned char *)bs.buffer;
563     return bs.bit_offset;
564 }
565
566
567 /*
568   Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
569   1) period between Frame #X and Frame #N = #X - #N
570   2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
571   3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
572   4) intra_period and intra_idr_period take precedence over ip_period
573   5) if ip_period > 1, intra_period and intra_idr_period are not  the strict periods 
574      of I/IDR frames, see bellow examples
575   -------------------------------------------------------------------
576   intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
577   0            ignored          1          IDRPPPPPPP ...     (No IDR/I any more)
578   0            ignored        >=2          IDR(PBB)(PBB)...   (No IDR/I any more)
579   1            0                ignored    IDRIIIIIII...      (No IDR any more)
580   1            1                ignored    IDR IDR IDR IDR...
581   1            >=2              ignored    IDRII IDRII IDR... (1/3/ignore)
582   >=2          0                1          IDRPPP IPPP I...   (3/0/1)
583   >=2          0              >=2          IDR(PBB)(PBB)(IBB) (6/0/3)
584                                               (PBB)(IBB)(PBB)(IBB)... 
585   >=2          >=2              1          IDRPPPPP IPPPPP IPPPPP (6/18/1)
586                                            IDRPPPPP IPPPPP IPPPPP...
587   >=2          >=2              >=2        {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
588                                            {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
589                                            {IDR(PBB)(PBB)(IBB)(PBB)}           (6/12/3)
590                                            {IDR(PBB)(PBB)(IBB)(PBB)}...
591                                            {IDR(PBB)(PBB)}                     (6/6/3)
592                                            {IDR(PBB)(PBB)}.
593 */
594
595 // General pts/dts strategy:
596 //
597 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
598 // bit tricky. We assume first of all that the frame rate never goes _above_
599 // MAX_FPS, which gives us a frame period N. The decoder can always decode
600 // in at least this speed, as long at dts <= pts (the frame is not attempted
601 // presented before it is decoded). Furthermore, we never have longer chains of
602 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
603 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
604 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
605 // frame, with an order O <= C.)
606 //
607 // Many strategies are possible, but we establish these rules:
608 //
609 //  - Tip frames have dts = pts - (C-O)*N.
610 //  - Non-tip frames have dts = dts_last + N.
611 //
612 // An example, with C=2 and N=10 and the data flow showed with arrows:
613 //
614 //        I  B  P  B  B  P
615 //   pts: 30 40 50 60 70 80
616 //        ↓  ↓     ↓
617 //   dts: 10 30 20 60 50←40
618 //         |  |  ↑        ↑
619 //         `--|--'        |
620 //             `----------'
621 //
622 // To show that this works fine also with irregular spacings, let's say that
623 // the third frame is delayed a bit (something earlier was dropped). Now the
624 // situation looks like this:
625 //
626 //        I  B  P  B  B   P
627 //   pts: 30 40 80 90 100 110
628 //        ↓  ↓     ↓
629 //   dts: 10 30 20 90 50←40
630 //         |  |  ↑        ↑
631 //         `--|--'        |
632 //             `----------'
633 //
634 // The resetting on every tip frame makes sure dts never ends up lagging a lot
635 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
636 //
637 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
638 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
639 // a tip frame and should be given a dts based on the previous one.
640 #define FRAME_P 0
641 #define FRAME_B 1
642 #define FRAME_I 2
643 #define FRAME_IDR 7
644 void encoding2display_order(
645     int encoding_order, int intra_period,
646     int intra_idr_period, int ip_period,
647     int *displaying_order,
648     int *frame_type, int *pts_lag)
649 {
650     int encoding_order_gop = 0;
651
652     *pts_lag = 0;
653
654     if (intra_period == 1) { /* all are I/IDR frames */
655         *displaying_order = encoding_order;
656         if (intra_idr_period == 0)
657             *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
658         else
659             *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
660         return;
661     }
662
663     if (intra_period == 0)
664         intra_idr_period = 0;
665
666     if (ip_period == 1) {
667         // No B-frames, sequence is like IDR PPPPP IPPPPP.
668         encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
669         *displaying_order = encoding_order;
670
671         if (encoding_order_gop == 0) { /* the first frame */
672             *frame_type = FRAME_IDR;
673         } else if (intra_period != 0 && /* have I frames */
674                    encoding_order_gop >= 2 &&
675                    (encoding_order_gop % intra_period == 0)) {
676             *frame_type = FRAME_I;
677         } else {
678             *frame_type = FRAME_P;
679         }
680         return;
681     } 
682
683     // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
684     encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
685     *pts_lag = -1;  // Most frames are not tip frames.
686          
687     if (encoding_order_gop == 0) { /* the first frame */
688         *frame_type = FRAME_IDR;
689         *displaying_order = encoding_order;
690         // IDR frames are a special case; I honestly can't find the logic behind
691         // why this is the right thing, but it seems to line up nicely in practice :-)
692         *pts_lag = TIMEBASE / MAX_FPS;
693     } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
694         *frame_type = FRAME_B;
695         *displaying_order = encoding_order - 1;
696         if ((encoding_order_gop % ip_period) == 0) {
697             *pts_lag = 0;  // Last B-frame.
698         }
699     } else if (intra_period != 0 && /* have I frames */
700                encoding_order_gop >= 2 &&
701                ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
702         *frame_type = FRAME_I;
703         *displaying_order = encoding_order + ip_period - 1;
704     } else {
705         *frame_type = FRAME_P;
706         *displaying_order = encoding_order + ip_period - 1;
707     }
708 }
709
710
711 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
712 {
713         if (global_flags.x264_video_to_disk) {
714                 // Quick Sync is entirely disabled.
715                 use_zerocopy = false;
716         } else if (global_flags.uncompressed_video_to_http) {
717                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
718                 use_zerocopy = false;
719         } else if (global_flags.x264_video_to_http) {
720                 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
721                 use_zerocopy = false;
722         } else {
723                 use_zerocopy = true;
724         }
725         global_flags.use_zerocopy = use_zerocopy;
726 }
727
728 VADisplayWithCleanup::~VADisplayWithCleanup()
729 {
730         if (va_dpy != nullptr) {
731                 vaTerminate(va_dpy);
732         }
733         if (x11_display != nullptr) {
734                 XCloseDisplay(x11_display);
735         }
736         if (drm_fd != -1) {
737                 close(drm_fd);
738         }
739 }
740
741 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
742 {
743         if (va_display.empty() || va_display[0] != '/') {  // An X display.
744                 Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str());
745                 if (x11_display == nullptr) {
746                         fprintf(stderr, "error: can't connect to X server!\n");
747                         return nullptr;
748                 }
749
750                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
751                 ret->x11_display = x11_display;
752                 ret->can_use_zerocopy = true;
753                 ret->va_dpy = vaGetDisplay(x11_display);
754                 if (ret->va_dpy == nullptr) {
755                         return nullptr;
756                 }
757                 return ret;
758         } else {  // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
759                 int drm_fd = open(va_display.c_str(), O_RDWR);
760                 if (drm_fd == -1) {
761                         perror(va_display.c_str());
762                         return NULL;
763                 }
764                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
765                 ret->drm_fd = drm_fd;
766                 ret->can_use_zerocopy = false;
767                 ret->va_dpy = vaGetDisplayDRM(drm_fd);
768                 if (ret->va_dpy == nullptr) {
769                         return nullptr;
770                 }
771                 return ret;
772         }
773 }
774
775 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, VAProfile *h264_profile, string *error)
776 {
777         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
778         if (va_dpy == nullptr) {
779                 if (error) *error = "Opening VA display failed";
780                 return nullptr;
781         }
782         int major_ver, minor_ver;
783         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
784         if (va_status != VA_STATUS_SUCCESS) {
785                 char buf[256];
786                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
787                 if (error != nullptr) *error = buf;
788                 return nullptr;
789         }
790
791         int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy);
792         unique_ptr<VAEntrypoint[]> entrypoints(new VAEntrypoint[num_entrypoints]);
793         if (entrypoints == nullptr) {
794                 if (error != nullptr) *error = "Failed to allocate memory for VA entry points";
795                 return nullptr;
796         }
797
798         // Try the profiles from highest to lowest until we find one that can be encoded.
799         constexpr VAProfile profile_list[] = { VAProfileH264High, VAProfileH264Main, VAProfileH264ConstrainedBaseline };
800         for (unsigned i = 0; i < sizeof(profile_list) / sizeof(profile_list[0]); ++i) {
801                 vaQueryConfigEntrypoints(va_dpy->va_dpy, profile_list[i], entrypoints.get(), &num_entrypoints);
802                 for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
803                         if (entrypoints[slice_entrypoint] != VAEntrypointEncSlice) {
804                                 continue;
805                         }
806
807                         // We found a usable encoder, so return it.
808                         if (h264_profile != nullptr) {
809                                 *h264_profile = profile_list[i];
810                         }
811                         return va_dpy;
812                 }
813         }
814
815         if (error != nullptr) *error = "Can't find VAEntrypointEncSlice for H264 profiles";
816         return nullptr;
817 }
818
819 int QuickSyncEncoderImpl::init_va(const string &va_display)
820 {
821     string error;
822     va_dpy = try_open_va(va_display, &h264_profile, &error);
823     if (va_dpy == nullptr) {
824         fprintf(stderr, "error: %s\n", error.c_str());
825         exit(1);
826     }
827     if (!va_dpy->can_use_zerocopy) {
828         use_zerocopy = false;
829     }
830     
831     switch (h264_profile) {
832         case VAProfileH264ConstrainedBaseline:
833             constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
834             ip_period = 1;
835             break;
836
837         case VAProfileH264Main:
838             constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
839             break;
840
841         case VAProfileH264High:
842             constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
843             break;
844         default:
845             h264_profile = VAProfileH264ConstrainedBaseline;
846             ip_period = 1;
847             constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
848             break;
849     }
850
851     VAConfigAttrib attrib[VAConfigAttribTypeMax];
852
853     /* find out the format for the render target, and rate control mode */
854     for (unsigned i = 0; i < VAConfigAttribTypeMax; i++)
855         attrib[i].type = (VAConfigAttribType)i;
856
857     VAStatus va_status = vaGetConfigAttributes(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
858                                       &attrib[0], VAConfigAttribTypeMax);
859     CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
860     /* check the interested configattrib */
861     if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
862         printf("Not find desired YUV420 RT format\n");
863         exit(1);
864     } else {
865         config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
866         config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
867         config_attrib_num++;
868     }
869     
870     if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
871         if (!(attrib[VAConfigAttribRateControl].value & VA_RC_CQP)) {
872             fprintf(stderr, "ERROR: VA-API encoder does not support CQP mode.\n");
873             exit(1);
874         }
875
876         config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
877         config_attrib[config_attrib_num].value = VA_RC_CQP;
878         config_attrib_num++;
879     }
880     
881
882     if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
883         int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
884
885         h264_packedheader = 1;
886         config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
887         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
888         
889         if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
890             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
891         }
892         
893         if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
894             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
895         }
896         
897         if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
898             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
899         }
900         
901         if (tmp & VA_ENC_PACKED_HEADER_MISC) {
902             config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
903         }
904         
905         enc_packed_header_idx = config_attrib_num;
906         config_attrib_num++;
907     }
908
909     if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
910         config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
911         config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
912         config_attrib_num++;
913     }
914     
915     if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
916         h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
917     }
918
919     return 0;
920 }
921
922 int QuickSyncEncoderImpl::setup_encode()
923 {
924         if (!global_flags.x264_video_to_disk) {
925                 VAStatus va_status;
926                 VASurfaceID *tmp_surfaceid;
927                 int codedbuf_size;
928                 VASurfaceID src_surface[SURFACE_NUM];
929                 VASurfaceID ref_surface[SURFACE_NUM];
930
931                 va_status = vaCreateConfig(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
932                                 &config_attrib[0], config_attrib_num, &config_id);
933                 CHECK_VASTATUS(va_status, "vaCreateConfig");
934
935                 /* create source surfaces */
936                 va_status = vaCreateSurfaces(va_dpy->va_dpy,
937                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
938                                 &src_surface[0], SURFACE_NUM,
939                                 NULL, 0);
940                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
941
942                 /* create reference surfaces */
943                 va_status = vaCreateSurfaces(va_dpy->va_dpy,
944                                 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
945                                 &ref_surface[0], SURFACE_NUM,
946                                 NULL, 0);
947                 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
948
949                 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
950                 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
951                 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
952
953                 for (int i = 0; i < SURFACE_NUM; i++) {
954                         gl_surfaces[i].src_surface = src_surface[i];
955                         gl_surfaces[i].ref_surface = ref_surface[i];
956                 }
957
958                 /* Create a context for this encode pipe */
959                 va_status = vaCreateContext(va_dpy->va_dpy, config_id,
960                                 frame_width_mbaligned, frame_height_mbaligned,
961                                 VA_PROGRESSIVE,
962                                 tmp_surfaceid, 2 * SURFACE_NUM,
963                                 &context_id);
964                 CHECK_VASTATUS(va_status, "vaCreateContext");
965                 free(tmp_surfaceid);
966
967                 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
968
969                 for (int i = 0; i < SURFACE_NUM; i++) {
970                         /* create coded buffer once for all
971                          * other VA buffers which won't be used again after vaRenderPicture.
972                          * so APP can always vaCreateBuffer for every frame
973                          * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
974                          * so VA won't maintain the coded buffer
975                          */
976                         va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncCodedBufferType,
977                                         codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
978                         CHECK_VASTATUS(va_status, "vaCreateBuffer");
979                 }
980         }
981
982         /* create OpenGL objects */
983         for (int 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                         size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
989
990                         // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
991                         // buffers, due to potentially differing pitch.
992                         glGenBuffers(1, &gl_surfaces[i].pbo);
993                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
994                         glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2 * bytes_per_pixel, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
995                         uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2 * bytes_per_pixel, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
996                         gl_surfaces[i].y_offset = 0;
997                         gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
998                         gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
999                         gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1000                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1001                 }
1002         }
1003
1004         return 0;
1005 }
1006
1007 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1008 //
1009 //   2 1 0 [3] 4 8 9
1010 template<class T, class C>
1011 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1012 {
1013         T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1014         sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1015         sort(middle, end, less_than);
1016 }
1017
1018 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1019 {
1020     if (frame_type == FRAME_B)
1021         return;
1022
1023     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1024
1025     CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1026     unique_lock<mutex> lock(storage_task_queue_mutex);
1027
1028     // Insert the new frame at the start of the reference queue.
1029     reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1030
1031     if (reference_frames.size() > num_ref_frames)
1032     {
1033         // The back frame frame is no longer in use as a reference.
1034         int display_frame_num = reference_frames.back().display_number;
1035         assert(surface_for_frame.count(display_frame_num));
1036         release_gl_surface(display_frame_num);
1037         reference_frames.pop_back();
1038     }
1039
1040     // Mark this frame in use as a reference.
1041     assert(surface_for_frame.count(current_display_frame));
1042     ++surface_for_frame[current_display_frame]->refcount;
1043     
1044     current_ref_frame_num++;
1045     if (current_ref_frame_num > MaxFrameNum)
1046         current_ref_frame_num = 0;
1047 }
1048
1049
1050 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1051 {
1052     const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1053         return a.frame_idx > b.frame_idx;
1054     };
1055
1056     for (size_t i = 0; i < reference_frames.size(); ++i) {
1057         RefPicList0_P[i] = reference_frames[i].pic;
1058     }
1059     sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1060 }
1061
1062 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1063 {
1064     const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1065         return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1066     };
1067     const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1068         return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1069     };
1070
1071     for (size_t i = 0; i < reference_frames.size(); ++i) {
1072         RefPicList0_B[i] = reference_frames[i].pic;
1073         RefPicList1_B[i] = reference_frames[i].pic;
1074     }
1075     sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1076     sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1077 }
1078
1079
1080 int QuickSyncEncoderImpl::render_sequence()
1081 {
1082     VABufferID seq_param_buf, rc_param_buf, render_id[2];
1083     VAStatus va_status;
1084     VAEncMiscParameterBuffer *misc_param;
1085     VAEncMiscParameterRateControl *misc_rate_ctrl;
1086     
1087     seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1088     seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1089     seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1090     seq_param.bits_per_second = frame_bitrate;
1091
1092     seq_param.intra_period = intra_period;
1093     seq_param.intra_idr_period = intra_idr_period;
1094     seq_param.ip_period = ip_period;
1095
1096     seq_param.max_num_ref_frames = num_ref_frames;
1097     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1098     seq_param.time_scale = TIMEBASE * 2;
1099     seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1100     seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1101     seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1102     seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1103     seq_param.seq_fields.bits.chroma_format_idc = 1;
1104     seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1105     
1106     if (frame_width != frame_width_mbaligned ||
1107         frame_height != frame_height_mbaligned) {
1108         seq_param.frame_cropping_flag = 1;
1109         seq_param.frame_crop_left_offset = 0;
1110         seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1111         seq_param.frame_crop_top_offset = 0;
1112         seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1113     }
1114     
1115     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1116                                VAEncSequenceParameterBufferType,
1117                                sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1118     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1119     
1120     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1121                                VAEncMiscParameterBufferType,
1122                                sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1123                                1, NULL, &rc_param_buf);
1124     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1125     
1126     vaMapBuffer(va_dpy->va_dpy, rc_param_buf, (void **)&misc_param);
1127     misc_param->type = VAEncMiscParameterTypeRateControl;
1128     misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1129     memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1130     misc_rate_ctrl->bits_per_second = frame_bitrate;
1131     misc_rate_ctrl->target_percentage = 66;
1132     misc_rate_ctrl->window_size = 1000;
1133     misc_rate_ctrl->initial_qp = initial_qp;
1134     misc_rate_ctrl->min_qp = minimal_qp;
1135     misc_rate_ctrl->basic_unit_size = 0;
1136     vaUnmapBuffer(va_dpy->va_dpy, rc_param_buf);
1137
1138     render_id[0] = seq_param_buf;
1139     render_id[1] = rc_param_buf;
1140     
1141     render_picture_and_delete(va_dpy->va_dpy, context_id, &render_id[0], 2);
1142     
1143     return 0;
1144 }
1145
1146 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1147 {
1148     static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1149     int prevPicOrderCntMsb, prevPicOrderCntLsb;
1150     int PicOrderCntMsb, TopFieldOrderCnt;
1151     
1152     if (frame_type == FRAME_IDR)
1153         prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1154     else {
1155         prevPicOrderCntMsb = PicOrderCntMsb_ref;
1156         prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1157     }
1158     
1159     if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1160         ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1161         PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1162     else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1163              ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1164         PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1165     else
1166         PicOrderCntMsb = prevPicOrderCntMsb;
1167     
1168     TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1169
1170     if (frame_type != FRAME_B) {
1171         PicOrderCntMsb_ref = PicOrderCntMsb;
1172         pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1173     }
1174     
1175     return TopFieldOrderCnt;
1176 }
1177
1178 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1179 {
1180     VABufferID pic_param_buf;
1181     VAStatus va_status;
1182     size_t i = 0;
1183
1184     pic_param.CurrPic.picture_id = surf->ref_surface;
1185     pic_param.CurrPic.frame_idx = current_ref_frame_num;
1186     pic_param.CurrPic.flags = 0;
1187     pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1188     pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1189     CurrentCurrPic = pic_param.CurrPic;
1190
1191     for (i = 0; i < reference_frames.size(); i++) {
1192         pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1193     }
1194     for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1195         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1196         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1197     }
1198     
1199     pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1200     pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1201     pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1202     pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1203     pic_param.frame_num = current_ref_frame_num;  // FIXME: is this correct?
1204     pic_param.coded_buf = surf->coded_buf;
1205     pic_param.last_picture = false;  // FIXME
1206     pic_param.pic_init_qp = initial_qp;
1207
1208     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncPictureParameterBufferType,
1209                                sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1210     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1211
1212     render_picture_and_delete(va_dpy->va_dpy, context_id, &pic_param_buf, 1);
1213
1214     return 0;
1215 }
1216
1217 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1218 {
1219     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1220     VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1221     unsigned int length_in_bits;
1222     unsigned char *packedseq_buffer = NULL;
1223     VAStatus va_status;
1224
1225     length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer); 
1226     
1227     packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1228     
1229     packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1230     packedheader_param_buffer.has_emulation_bytes = 0;
1231     va_status = vaCreateBuffer(va_dpy->va_dpy,
1232                                context_id,
1233                                VAEncPackedHeaderParameterBufferType,
1234                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1235                                &packedseq_para_bufid);
1236     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1237
1238     va_status = vaCreateBuffer(va_dpy->va_dpy,
1239                                context_id,
1240                                VAEncPackedHeaderDataBufferType,
1241                                (length_in_bits + 7) / 8, 1, packedseq_buffer,
1242                                &packedseq_data_bufid);
1243     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1244
1245     render_id[0] = packedseq_para_bufid;
1246     render_id[1] = packedseq_data_bufid;
1247     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1248
1249     free(packedseq_buffer);
1250     
1251     return 0;
1252 }
1253
1254
1255 int QuickSyncEncoderImpl::render_packedpicture()
1256 {
1257     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1258     VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1259     unsigned int length_in_bits;
1260     unsigned char *packedpic_buffer = NULL;
1261     VAStatus va_status;
1262
1263     length_in_bits = build_packed_pic_buffer(&packedpic_buffer); 
1264     packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1265     packedheader_param_buffer.bit_length = length_in_bits;
1266     packedheader_param_buffer.has_emulation_bytes = 0;
1267
1268     va_status = vaCreateBuffer(va_dpy->va_dpy,
1269                                context_id,
1270                                VAEncPackedHeaderParameterBufferType,
1271                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1272                                &packedpic_para_bufid);
1273     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1274
1275     va_status = vaCreateBuffer(va_dpy->va_dpy,
1276                                context_id,
1277                                VAEncPackedHeaderDataBufferType,
1278                                (length_in_bits + 7) / 8, 1, packedpic_buffer,
1279                                &packedpic_data_bufid);
1280     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1281
1282     render_id[0] = packedpic_para_bufid;
1283     render_id[1] = packedpic_data_bufid;
1284     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1285
1286     free(packedpic_buffer);
1287     
1288     return 0;
1289 }
1290
1291 void QuickSyncEncoderImpl::render_packedslice()
1292 {
1293     VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1294     VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1295     unsigned int length_in_bits;
1296     unsigned char *packedslice_buffer = NULL;
1297     VAStatus va_status;
1298
1299     length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1300     packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1301     packedheader_param_buffer.bit_length = length_in_bits;
1302     packedheader_param_buffer.has_emulation_bytes = 0;
1303
1304     va_status = vaCreateBuffer(va_dpy->va_dpy,
1305                                context_id,
1306                                VAEncPackedHeaderParameterBufferType,
1307                                sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1308                                &packedslice_para_bufid);
1309     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1310
1311     va_status = vaCreateBuffer(va_dpy->va_dpy,
1312                                context_id,
1313                                VAEncPackedHeaderDataBufferType,
1314                                (length_in_bits + 7) / 8, 1, packedslice_buffer,
1315                                &packedslice_data_bufid);
1316     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1317
1318     render_id[0] = packedslice_para_bufid;
1319     render_id[1] = packedslice_data_bufid;
1320     render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1321
1322     free(packedslice_buffer);
1323 }
1324
1325 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1326 {
1327     VABufferID slice_param_buf;
1328     VAStatus va_status;
1329     int i;
1330
1331     /* one frame, one slice */
1332     slice_param.macroblock_address = 0;
1333     slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1334     slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1335     if (frame_type == FRAME_IDR) {
1336         if (encoding_frame_num != 0)
1337             ++slice_param.idr_pic_id;
1338     } else if (frame_type == FRAME_P) {
1339         VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1340         update_RefPicList_P(RefPicList0_P);
1341
1342         int refpiclist0_max = h264_maxref & 0xffff;
1343         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1344
1345         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1346             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1347             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1348         }
1349     } else if (frame_type == FRAME_B) {
1350         VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1351         update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1352
1353         int refpiclist0_max = h264_maxref & 0xffff;
1354         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1355
1356         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1357         for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1358             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1359             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1360         }
1361
1362         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1363         for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1364             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1365             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1366         }
1367     }
1368
1369     slice_param.slice_alpha_c0_offset_div2 = 0;
1370     slice_param.slice_beta_offset_div2 = 0;
1371     slice_param.direct_spatial_mv_pred_flag = 1;
1372     slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1373     
1374
1375     if (h264_packedheader &&
1376         config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1377         render_packedslice();
1378
1379     va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncSliceParameterBufferType,
1380                                sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1381     CHECK_VASTATUS(va_status, "vaCreateBuffer");
1382
1383     render_picture_and_delete(va_dpy->va_dpy, context_id, &slice_param_buf, 1);
1384
1385     return 0;
1386 }
1387
1388
1389
1390 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1391 {    
1392         VACodedBufferSegment *buf_list = NULL;
1393         VAStatus va_status;
1394
1395         string data;
1396
1397         va_status = vaMapBuffer(va_dpy->va_dpy, surf->coded_buf, (void **)(&buf_list));
1398         CHECK_VASTATUS(va_status, "vaMapBuffer");
1399         while (buf_list != NULL) {
1400                 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1401                 buf_list = (VACodedBufferSegment *) buf_list->next;
1402         }
1403         vaUnmapBuffer(va_dpy->va_dpy, surf->coded_buf);
1404
1405         static int frameno = 0;
1406         print_latency("Current Quick Sync latency (video inputs → disk mux):",
1407                 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1408
1409         {
1410                 // Add video.
1411                 AVPacket pkt;
1412                 memset(&pkt, 0, sizeof(pkt));
1413                 pkt.buf = nullptr;
1414                 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1415                 pkt.size = data.size();
1416                 pkt.stream_index = 0;
1417                 if (task.frame_type == FRAME_IDR) {
1418                         pkt.flags = AV_PKT_FLAG_KEY;
1419                 } else {
1420                         pkt.flags = 0;
1421                 }
1422                 pkt.duration = task.duration;
1423                 if (file_mux) {
1424                         file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1425                 }
1426                 if (!global_flags.uncompressed_video_to_http &&
1427                     !global_flags.x264_video_to_http) {
1428                         stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1429                 }
1430         }
1431 }
1432
1433
1434 // this is weird. but it seems to put a new frame onto the queue
1435 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1436 {
1437         unique_lock<mutex> lock(storage_task_queue_mutex);
1438         storage_task_queue.push(move(task));
1439         storage_task_queue_changed.notify_all();
1440 }
1441
1442 void QuickSyncEncoderImpl::storage_task_thread()
1443 {
1444         pthread_setname_np(pthread_self(), "QS_Storage");
1445         for ( ;; ) {
1446                 storage_task current;
1447                 GLSurface *surf;
1448                 {
1449                         // wait until there's an encoded frame  
1450                         unique_lock<mutex> lock(storage_task_queue_mutex);
1451                         storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1452                         if (storage_thread_should_quit && storage_task_queue.empty()) return;
1453                         current = move(storage_task_queue.front());
1454                         storage_task_queue.pop();
1455                         surf = surface_for_frame[current.display_order];
1456                         assert(surf != nullptr);
1457                 }
1458
1459                 VAStatus va_status;
1460
1461                 size_t display_order = current.display_order;
1462                 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1463            
1464                 // waits for data, then saves it to disk.
1465                 va_status = vaSyncSurface(va_dpy->va_dpy, surf->src_surface);
1466                 CHECK_VASTATUS(va_status, "vaSyncSurface");
1467                 save_codeddata(surf, move(current));
1468
1469                 // Unlock the frame, and all its references.
1470                 {
1471                         unique_lock<mutex> lock(storage_task_queue_mutex);
1472                         release_gl_surface(display_order);
1473
1474                         for (size_t frame_num : ref_display_frame_numbers) {
1475                                 release_gl_surface(frame_num);
1476                         }
1477                 }
1478         }
1479 }
1480
1481 void QuickSyncEncoderImpl::release_encode()
1482 {
1483         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1484                 vaDestroyBuffer(va_dpy->va_dpy, gl_surfaces[i].coded_buf);
1485                 vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].src_surface, 1);
1486                 vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].ref_surface, 1);
1487         }
1488
1489         vaDestroyContext(va_dpy->va_dpy, context_id);
1490         vaDestroyConfig(va_dpy->va_dpy, config_id);
1491 }
1492
1493 void QuickSyncEncoderImpl::release_gl_resources()
1494 {
1495         assert(is_shutdown);
1496         if (has_released_gl_resources) {
1497                 return;
1498         }
1499
1500         for (unsigned i = 0; i < SURFACE_NUM; i++) {
1501                 if (use_zerocopy) {
1502                         resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1503                         resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1504                 } else {
1505                         glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1506                         glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1507                         glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1508                         glDeleteBuffers(1, &gl_surfaces[i].pbo);
1509                 }
1510         }
1511
1512         has_released_gl_resources = true;
1513 }
1514
1515 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1516         : 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)
1517 {
1518         file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1519         open_output_file(filename);
1520         file_audio_encoder->add_mux(file_mux.get());
1521
1522         frame_width_mbaligned = (frame_width + 15) & (~15);
1523         frame_height_mbaligned = (frame_height + 15) & (~15);
1524
1525         //print_input();
1526
1527         if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1528                 assert(x264_encoder != nullptr);
1529         } else {
1530                 assert(x264_encoder == nullptr);
1531         }
1532
1533         enable_zerocopy_if_possible();
1534         if (!global_flags.x264_video_to_disk) {
1535                 init_va(va_display);
1536         }
1537         setup_encode();
1538
1539         if (!global_flags.x264_video_to_disk) {
1540                 memset(&seq_param, 0, sizeof(seq_param));
1541                 memset(&pic_param, 0, sizeof(pic_param));
1542                 memset(&slice_param, 0, sizeof(slice_param));
1543         }
1544
1545         call_once(quick_sync_metrics_inited, [](){
1546                 mixer_latency_histogram.init("mixer");
1547                 qs_latency_histogram.init("quick_sync");
1548                 current_file_mux_metrics.init({{ "destination", "current_file" }});
1549                 total_mux_metrics.init({{ "destination", "files_total" }});
1550                 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1551                 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1552         });
1553
1554         storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1555
1556         encode_thread = thread([this]{
1557                 QOpenGLContext *context = create_context(this->surface);
1558                 eglBindAPI(EGL_OPENGL_API);
1559                 if (!make_current(context, this->surface)) {
1560                         printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1561                                 eglGetError());
1562                         exit(1);
1563                 }
1564                 encode_thread_func();
1565                 delete_context(context);
1566         });
1567 }
1568
1569 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1570 {
1571         shutdown();
1572         release_gl_resources();
1573 }
1574
1575 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1576 {
1577         for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1578                 if (gl_surfaces[i].refcount == 0) {
1579                         ++gl_surfaces[i].refcount;
1580                         return &gl_surfaces[i];
1581                 }
1582         }
1583         return nullptr;
1584 }
1585
1586 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1587 {
1588         assert(surface_for_frame.count(display_frame_num));
1589         QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1590         if (--surf->refcount == 0) {
1591                 assert(surface_for_frame.count(display_frame_num));
1592                 surface_for_frame.erase(display_frame_num);
1593                 storage_task_queue_changed.notify_all();
1594         }
1595 }
1596
1597 bool QuickSyncEncoderImpl::is_zerocopy() const
1598 {
1599         return use_zerocopy;
1600 }
1601
1602 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1603 {
1604         assert(!is_shutdown);
1605         GLSurface *surf = nullptr;
1606         {
1607                 // Wait until this frame slot is done encoding.
1608                 unique_lock<mutex> lock(storage_task_queue_mutex);
1609                 surf = allocate_gl_surface();
1610                 if (surf == nullptr) {
1611                         fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1612                                 current_storage_frame);
1613                         ++metric_quick_sync_stalled_frames;
1614                         storage_task_queue_changed.wait(lock, [this, &surf]{
1615                                 if (storage_thread_should_quit)
1616                                         return true;
1617                                 surf = allocate_gl_surface();
1618                                 return surf != nullptr;
1619                         });
1620                 }
1621                 if (storage_thread_should_quit) return false;
1622                 assert(surf != nullptr);
1623                 surface_for_frame[current_storage_frame] = surf;
1624         }
1625
1626         if (use_zerocopy) {
1627                 *y_tex = surf->y_tex;
1628                 *cbcr_tex = surf->cbcr_tex;
1629         } else {
1630                 surf->y_tex = *y_tex;
1631                 surf->cbcr_tex = *cbcr_tex;
1632         }
1633
1634         if (!global_flags.x264_video_to_disk) {
1635                 VAStatus va_status = vaDeriveImage(va_dpy->va_dpy, surf->src_surface, &surf->surface_image);
1636                 CHECK_VASTATUS(va_status, "vaDeriveImage");
1637
1638                 if (use_zerocopy) {
1639                         VABufferInfo buf_info;
1640                         buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME;  // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1641                         va_status = vaAcquireBufferHandle(va_dpy->va_dpy, surf->surface_image.buf, &buf_info);
1642                         CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1643
1644                         // Create Y image.
1645                         surf->y_egl_image = EGL_NO_IMAGE_KHR;
1646                         EGLint y_attribs[] = {
1647                                 EGL_WIDTH, frame_width,
1648                                 EGL_HEIGHT, frame_height,
1649                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1650                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1651                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1652                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1653                                 EGL_NONE
1654                         };
1655
1656                         surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1657                         assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1658
1659                         // Associate Y image to a texture.
1660                         glBindTexture(GL_TEXTURE_2D, *y_tex);
1661                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1662
1663                         // Create CbCr image.
1664                         surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1665                         EGLint cbcr_attribs[] = {
1666                                 EGL_WIDTH, frame_width / 2,
1667                                 EGL_HEIGHT, frame_height / 2,
1668                                 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1669                                 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1670                                 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1671                                 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1672                                 EGL_NONE
1673                         };
1674
1675                         surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1676                         assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1677
1678                         // Associate CbCr image to a texture.
1679                         glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1680                         glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1681                 }
1682         }
1683
1684         current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1685
1686         return true;
1687 }
1688
1689 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1690 {
1691         lock_guard<mutex> lock(file_audio_encoder_mutex);
1692         assert(!is_shutdown);
1693         file_audio_encoder->encode_audio(audio, pts + global_delay());
1694 }
1695
1696 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1697 {
1698         assert(!is_shutdown);
1699
1700         if (!use_zerocopy) {
1701                 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1702                 GLSurface *surf;
1703                 {
1704                         unique_lock<mutex> lock(storage_task_queue_mutex);
1705                         surf = surface_for_frame[current_storage_frame];
1706                         assert(surf != nullptr);
1707                 }
1708
1709                 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1710                 check_error();
1711
1712                 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1713                 check_error();
1714
1715                 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1716                 check_error();
1717                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1718                 check_error();
1719
1720                 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1721                 check_error();
1722                 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1723                 check_error();
1724
1725                 // We don't own these; the caller does.
1726                 surf->y_tex = surf->cbcr_tex = 0;
1727
1728                 glBindTexture(GL_TEXTURE_2D, 0);
1729                 check_error();
1730                 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1731                 check_error();
1732
1733                 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1734                 check_error();
1735         }
1736
1737         RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1738         check_error();
1739         glFlush();  // Make the H.264 thread see the fence as soon as possible.
1740         check_error();
1741
1742         {
1743                 unique_lock<mutex> lock(frame_queue_mutex);
1744                 current_video_frame.fence = fence;
1745                 pending_video_frames.push(move(current_video_frame));
1746                 ++current_storage_frame;
1747         }
1748         frame_queue_nonempty.notify_all();
1749         return fence;
1750 }
1751
1752 void QuickSyncEncoderImpl::shutdown()
1753 {
1754         if (is_shutdown) {
1755                 return;
1756         }
1757
1758         {
1759                 unique_lock<mutex> lock(frame_queue_mutex);
1760                 encode_thread_should_quit = true;
1761                 frame_queue_nonempty.notify_all();
1762         }
1763         encode_thread.join();
1764         {
1765                 unique_lock<mutex> lock(storage_task_queue_mutex);
1766                 storage_thread_should_quit = true;
1767                 frame_queue_nonempty.notify_all();
1768                 storage_task_queue_changed.notify_all();
1769         }
1770         storage_thread.join();
1771
1772         // Encode any leftover audio in the queues, and also any delayed frames.
1773         {
1774                 lock_guard<mutex> lock(file_audio_encoder_mutex);
1775                 file_audio_encoder->encode_last_audio();
1776         }
1777
1778         if (!global_flags.x264_video_to_disk) {
1779                 release_encode();
1780                 va_dpy.reset();
1781         }
1782         is_shutdown = true;
1783 }
1784
1785 void QuickSyncEncoderImpl::close_file()
1786 {
1787         file_mux.reset();
1788         metric_current_file_start_time_seconds = 0.0 / 0.0;
1789 }
1790
1791 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1792 {
1793         AVFormatContext *avctx = avformat_alloc_context();
1794         avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1795         assert(filename.size() < sizeof(avctx->filename) - 1);
1796         strcpy(avctx->filename, filename.c_str());
1797
1798         string url = "file:" + filename;
1799         int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1800         if (ret < 0) {
1801                 char tmp[AV_ERROR_MAX_STRING_SIZE];
1802                 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1803                 exit(1);
1804         }
1805
1806         string video_extradata;  // FIXME: See other comment about global headers.
1807         if (global_flags.x264_video_to_disk) {
1808                 video_extradata = x264_encoder->get_global_headers();
1809         }
1810
1811         current_file_mux_metrics.reset();
1812
1813         {
1814                 lock_guard<mutex> lock(file_audio_encoder_mutex);
1815                 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1816                 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1817                         std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1),
1818                         Mux::WRITE_BACKGROUND,
1819                         { &current_file_mux_metrics, &total_mux_metrics }));
1820         }
1821         metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1822
1823         if (global_flags.x264_video_to_disk) {
1824                 x264_encoder->add_mux(file_mux.get());
1825         }
1826 }
1827
1828 void QuickSyncEncoderImpl::encode_thread_func()
1829 {
1830         pthread_setname_np(pthread_self(), "QS_Encode");
1831
1832         int64_t last_dts = -1;
1833         int gop_start_display_frame_num = 0;
1834         for (int display_frame_num = 0; ; ++display_frame_num) {
1835                 // Wait for the frame to be in the queue. Note that this only means
1836                 // we started rendering it.
1837                 PendingFrame frame;
1838                 {
1839                         unique_lock<mutex> lock(frame_queue_mutex);
1840                         frame_queue_nonempty.wait(lock, [this]{
1841                                 return encode_thread_should_quit || !pending_video_frames.empty();
1842                         });
1843                         if (encode_thread_should_quit && pending_video_frames.empty()) {
1844                                 // We may have queued frames left in the reorder buffer
1845                                 // that were supposed to be B-frames, but have no P-frame
1846                                 // to be encoded against. If so, encode them all as
1847                                 // P-frames instead. Note that this happens under the mutex,
1848                                 // but nobody else uses it at this point, since we're shutting down,
1849                                 // so there's no contention.
1850                                 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1851                                 return;
1852                         } else {
1853                                 frame = move(pending_video_frames.front());
1854                                 pending_video_frames.pop();
1855                         }
1856                 }
1857
1858                 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1859                 // Note that this implicitly waits for the frame to be done rendering.
1860                 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1861
1862                 if (global_flags.x264_video_to_disk) {
1863                         unique_lock<mutex> lock(storage_task_queue_mutex);
1864                         release_gl_surface(display_frame_num);
1865                         continue;
1866                 }
1867
1868                 reorder_buffer[display_frame_num] = move(frame);
1869
1870                 // Now encode as many QuickSync frames as we can using the frames we have available.
1871                 // (It could be zero, or it could be multiple.) FIXME: make a function.
1872                 for ( ;; ) {
1873                         int pts_lag;
1874                         int frame_type, quicksync_display_frame_num;
1875                         encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1876                                                &quicksync_display_frame_num, &frame_type, &pts_lag);
1877                         if (!reorder_buffer.count(quicksync_display_frame_num)) {
1878                                 break;
1879                         }
1880                         frame = move(reorder_buffer[quicksync_display_frame_num]);
1881                         reorder_buffer.erase(quicksync_display_frame_num);
1882
1883                         if (frame_type == FRAME_IDR) {
1884                                 // Release any reference frames from the previous GOP.
1885                                 {
1886                                         unique_lock<mutex> lock(storage_task_queue_mutex);
1887                                         for (const ReferenceFrame &frame : reference_frames) {
1888                                                 release_gl_surface(frame.display_number);
1889                                         }
1890                                 }
1891                                 reference_frames.clear();
1892                                 current_ref_frame_num = 0;
1893                                 gop_start_display_frame_num = quicksync_display_frame_num;
1894                         }
1895
1896                         // Determine the dts of this frame.
1897                         int64_t dts;
1898                         if (pts_lag == -1) {
1899                                 assert(last_dts != -1);
1900                                 dts = last_dts + (TIMEBASE / MAX_FPS);
1901                         } else {
1902                                 dts = frame.pts - pts_lag;
1903                         }
1904                         last_dts = dts;
1905
1906                         encode_frame(frame, quicksync_encoding_frame_num, quicksync_display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1907                         ++quicksync_encoding_frame_num;
1908                 }
1909         }
1910 }
1911
1912 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1913 {
1914         if (reorder_buffer.empty()) {
1915                 return;
1916         }
1917
1918         for (auto &pending_frame : reorder_buffer) {
1919                 int display_frame_num = pending_frame.first;
1920                 assert(display_frame_num > 0);
1921                 PendingFrame frame = move(pending_frame.second);
1922                 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1923                 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1924                 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1925                 last_dts = dts;
1926         }
1927 }
1928
1929 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1930 {
1931         AVPacket pkt;
1932         memset(&pkt, 0, sizeof(pkt));
1933         pkt.buf = nullptr;
1934         pkt.data = const_cast<uint8_t *>(data);
1935         pkt.size = frame_width * frame_height * 2;
1936         pkt.stream_index = 0;
1937         pkt.flags = AV_PKT_FLAG_KEY;
1938         pkt.duration = duration;
1939         stream_mux->add_packet(pkt, pts, pts);
1940 }
1941
1942 namespace {
1943
1944 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1945 {
1946         if (src_width == dst_pitch) {
1947                 memcpy(dst, src, src_width * height);
1948         } else {
1949                 for (size_t y = 0; y < height; ++y) {
1950                         const uint8_t *sptr = src + y * src_width;
1951                         uint8_t *dptr = dst + y * dst_pitch;
1952                         memcpy(dptr, sptr, src_width);
1953                 }
1954         }
1955 }
1956
1957 }  // namespace
1958
1959 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1960 {
1961         // Wait for the GPU to be done with the frame.
1962         GLenum sync_status;
1963         do {
1964                 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1965                 check_error();
1966                 if (sync_status == GL_TIMEOUT_EXPIRED) {
1967                         // NVIDIA likes to busy-wait; yield instead.
1968                         this_thread::sleep_for(milliseconds(1));
1969                 }
1970         } while (sync_status == GL_TIMEOUT_EXPIRED);
1971         assert(sync_status != GL_WAIT_FAILED);
1972
1973         ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1974         static int frameno = 0;
1975         print_latency("Current mixer latency (video inputs → ready for encode):",
1976                 received_ts, false, &frameno, &mixer_latency_histogram);
1977
1978         // Release back any input frames we needed to render this frame.
1979         frame.input_frames.clear();
1980
1981         GLSurface *surf;
1982         {
1983                 unique_lock<mutex> lock(storage_task_queue_mutex);
1984                 surf = surface_for_frame[display_frame_num];
1985                 assert(surf != nullptr);
1986         }
1987         uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1988         if (global_flags.uncompressed_video_to_http) {
1989                 add_packet_for_uncompressed_frame(pts, duration, data);
1990         } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1991                 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1992         }
1993 }
1994
1995 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1996                                         int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1997 {
1998         const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1999
2000         GLSurface *surf;
2001         {
2002                 unique_lock<mutex> lock(storage_task_queue_mutex);
2003                 surf = surface_for_frame[display_frame_num];
2004                 assert(surf != nullptr);
2005         }
2006         VAStatus va_status;
2007
2008         if (use_zerocopy) {
2009                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2010                 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2011                 va_status = vaReleaseBufferHandle(va_dpy->va_dpy, surf->surface_image.buf);
2012                 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2013         } else {
2014                 // Upload the frame to VA-API.
2015                 unsigned char *surface_p = nullptr;
2016                 vaMapBuffer(va_dpy->va_dpy, surf->surface_image.buf, (void **)&surface_p);
2017
2018                 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2019                 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2020
2021                 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2022                 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2023
2024                 va_status = vaUnmapBuffer(va_dpy->va_dpy, surf->surface_image.buf);
2025                 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2026         }
2027
2028         va_status = vaDestroyImage(va_dpy->va_dpy, surf->surface_image.image_id);
2029         CHECK_VASTATUS(va_status, "vaDestroyImage");
2030
2031         // Schedule the frame for encoding.
2032         VASurfaceID va_surface = surf->src_surface;
2033         va_status = vaBeginPicture(va_dpy->va_dpy, context_id, va_surface);
2034         CHECK_VASTATUS(va_status, "vaBeginPicture");
2035
2036         if (frame_type == FRAME_IDR) {
2037                 // FIXME: If the mux wants global headers, we should not put the
2038                 // SPS/PPS before each IDR frame, but rather put it into the
2039                 // codec extradata (formatted differently?).
2040                 //
2041                 // NOTE: If we change ycbcr_coefficients, it will not take effect
2042                 // before the next IDR frame. This is acceptable, as it should only
2043                 // happen on a mode change, which is rare.
2044                 render_sequence();
2045                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2046                 if (h264_packedheader) {
2047                         render_packedsequence(ycbcr_coefficients);
2048                         render_packedpicture();
2049                 }
2050         } else {
2051                 //render_sequence();
2052                 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2053         }
2054         render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2055
2056         va_status = vaEndPicture(va_dpy->va_dpy, context_id);
2057         CHECK_VASTATUS(va_status, "vaEndPicture");
2058
2059         update_ReferenceFrames(display_frame_num, frame_type);
2060
2061         vector<size_t> ref_display_frame_numbers;
2062
2063         // Lock the references for this frame; otherwise, they could be
2064         // rendered to before this frame is done encoding.
2065         {
2066                 unique_lock<mutex> lock(storage_task_queue_mutex);
2067                 for (const ReferenceFrame &frame : reference_frames) {
2068                         assert(surface_for_frame.count(frame.display_number));
2069                         ++surface_for_frame[frame.display_number]->refcount;
2070                         ref_display_frame_numbers.push_back(frame.display_number);
2071                 }
2072         }
2073
2074         // so now the data is done encoding (well, async job kicked off)...
2075         // we send that to the storage thread
2076         storage_task tmp;
2077         tmp.display_order = display_frame_num;
2078         tmp.frame_type = frame_type;
2079         tmp.pts = pts;
2080         tmp.dts = dts;
2081         tmp.duration = duration;
2082         tmp.ycbcr_coefficients = ycbcr_coefficients;
2083         tmp.received_ts = received_ts;
2084         tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2085         storage_task_enqueue(move(tmp));
2086 }
2087
2088 // Proxy object.
2089 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
2090         : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2091
2092 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2093 QuickSyncEncoder::~QuickSyncEncoder() {}
2094
2095 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2096 {
2097         impl->add_audio(pts, audio);
2098 }
2099
2100 bool QuickSyncEncoder::is_zerocopy() const
2101 {
2102         return impl->is_zerocopy();
2103 }
2104
2105 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2106 {
2107         return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2108 }
2109
2110 RefCountedGLsync QuickSyncEncoder::end_frame()
2111 {
2112         return impl->end_frame();
2113 }
2114
2115 void QuickSyncEncoder::shutdown()
2116 {
2117         impl->shutdown();
2118 }
2119
2120 void QuickSyncEncoder::close_file()
2121 {
2122         impl->shutdown();
2123 }
2124
2125 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2126 {
2127         impl->set_stream_mux(mux);
2128 }
2129
2130 int64_t QuickSyncEncoder::global_delay() const {
2131         return impl->global_delay();
2132 }
2133
2134 string QuickSyncEncoder::get_usable_va_display()
2135 {
2136         // Reduce the amount of chatter while probing,
2137         // unless the user has specified otherwise.
2138         bool need_env_reset = false;
2139         if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
2140                 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
2141                 need_env_reset = true;
2142         }
2143
2144         // First try the default (ie., whatever $DISPLAY is set to).
2145         unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va("", nullptr, nullptr);
2146         if (va_dpy != nullptr) {
2147                 if (need_env_reset) {
2148                         unsetenv("LIBVA_MESSAGING_LEVEL");
2149                 }
2150                 return "";
2151         }
2152
2153         fprintf(stderr, "No --va-display was given, and the X11 display did not expose a VA-API H.264 encoder.\n");
2154
2155         // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
2156         glob_t g;
2157         int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
2158         if (err != 0) {
2159                 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
2160         } else {
2161                 for (size_t i = 0; i < g.gl_pathc; ++i) {
2162                         string path = g.gl_pathv[i];
2163                         va_dpy = try_open_va(path, nullptr, nullptr);
2164                         if (va_dpy != nullptr) {
2165                                 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
2166                                         path.c_str());
2167                                 globfree(&g);
2168                                 if (need_env_reset) {
2169                                         unsetenv("LIBVA_MESSAGING_LEVEL");
2170                                 }
2171                                 return path;
2172                         }
2173                 }
2174         }
2175
2176         fprintf(stderr, "No suitable VA-API H.264 encoders were found in /dev/dri; giving up.\n");
2177         fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
2178         fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
2179         fprintf(stderr, "to expose Quick Sync. Alternatively, you can use --record-x264-video\n");
2180         fprintf(stderr, "to use software instead of hardware H.264 encoding, at the expense\n");
2181         fprintf(stderr, "of increased CPU usage and possibly bit rate.\n");
2182         exit(1);
2183 }