1 #include "quicksync_encoder.h"
3 #include <movit/image_format.h>
4 #include <movit/resource_pool.h> // Must be above the Xlib includes.
5 #include <movit/util.h>
7 #include <EGL/eglplatform.h>
10 #include <epoxy/egl.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>
25 #include <condition_variable>
40 #include <libavcodec/avcodec.h>
41 #include <libavformat/avio.h>
42 #include <libavutil/error.h>
43 #include <libdrm/drm_fourcc.h>
47 #include "audio_encoder.h"
48 #include "shared/context.h"
50 #include "shared/disk_space_estimator.h"
51 #include "shared/ffmpeg_raii.h"
53 #include "shared/mux.h"
54 #include "print_latency.h"
55 #include "quicksync_encoder_impl.h"
56 #include "ref_counted_frame.h"
57 #include "shared/timebase.h"
58 #include "x264_encoder.h"
60 using namespace movit;
62 using namespace std::chrono;
63 using namespace std::placeholders;
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};
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); \
87 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
89 //#include "loadsurface.h"
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
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))
110 #define ENTROPY_MODE_CAVLC 0
111 #define ENTROPY_MODE_CABAC 1
113 #define PROFILE_IDC_BASELINE 66
114 #define PROFILE_IDC_MAIN 77
115 #define PROFILE_IDC_HIGH 100
117 #define BITSTREAM_ALLOCATE_STEPPING 4096
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;
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
129 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
131 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
132 CHECK_VASTATUS(va_status, "vaRenderPicture");
134 for (int i = 0; i < num_buffers; ++i) {
135 va_status = vaDestroyBuffer(dpy, buffers[i]);
136 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
141 va_swap32(unsigned int val)
143 unsigned char *pval = (unsigned char *)&val;
145 return ((pval[0] << 24) |
152 bitstream_start(bitstream *bs)
154 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
155 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
160 bitstream_end(bitstream *bs)
162 int pos = (bs->bit_offset >> 5);
163 int bit_offset = (bs->bit_offset & 0x1f);
164 int bit_left = 32 - bit_offset;
167 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
172 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
174 int pos = (bs->bit_offset >> 5);
175 int bit_offset = (bs->bit_offset & 0x1f);
176 int bit_left = 32 - bit_offset;
181 bs->bit_offset += size_in_bits;
183 if (bit_left > size_in_bits) {
184 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
186 size_in_bits -= bit_left;
187 if (bit_left >= 32) {
188 bs->buffer[pos] = (val >> size_in_bits);
190 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
192 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
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));
199 bs->buffer[pos + 1] = val;
204 bitstream_put_ue(bitstream *bs, unsigned int val)
206 int size_in_bits = 0;
214 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
215 bitstream_put_ui(bs, val, size_in_bits);
219 bitstream_put_se(bitstream *bs, int val)
221 unsigned int new_val;
226 new_val = 2 * val - 1;
228 bitstream_put_ue(bs, new_val);
232 bitstream_byte_aligning(bitstream *bs, int bit)
234 int bit_offset = (bs->bit_offset & 0x7);
235 int bit_left = 8 - bit_offset;
241 assert(bit == 0 || bit == 1);
244 new_val = (1 << bit_left) - 1;
248 bitstream_put_ui(bs, new_val, bit_left);
252 rbsp_trailing_bits(bitstream *bs)
254 bitstream_put_ui(bs, 1, 1);
255 bitstream_byte_aligning(bs, 0);
258 static void nal_start_code_prefix(bitstream *bs)
260 bitstream_put_ui(bs, 0x00000001, 32);
263 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
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);
270 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
272 int profile_idc = PROFILE_IDC_BASELINE;
274 if (h264_profile == VAProfileH264High)
275 profile_idc = PROFILE_IDC_HIGH;
276 else if (h264_profile == VAProfileH264Main)
277 profile_idc = PROFILE_IDC_MAIN;
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 */
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 */
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 */
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 */
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 */
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 */
312 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
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 */
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 */
326 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
328 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
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 */
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 */
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) */
345 assert(ycbcr_coefficients == YCBCR_REC_601);
346 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
350 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
351 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
353 bitstream_put_ui(bs, 1, 32); // FPS
354 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
355 bitstream_put_ui(bs, 1, 1);
357 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
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 */
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] */
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 */
373 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
374 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
376 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
377 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
380 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
384 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
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 */
389 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
391 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
393 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
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 */
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 */
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 */
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 */
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 */
414 rbsp_trailing_bits(bs);
417 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
419 int first_mb_in_slice = slice_param.macroblock_address;
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 */
426 /* frame_mbs_only_flag == 1 */
427 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
432 if (pic_param.pic_fields.bits.idr_pic_flag)
433 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
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 */
443 /* redundant_pic_cnt_present_flag == 0 */
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: */
448 if (slice_param.num_ref_idx_active_override_flag)
449 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
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 */
456 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
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);
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 */
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 */
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;
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 */
486 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
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 */
494 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
496 /* ignore for SP/SI */
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 */
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 */
507 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
508 bitstream_byte_aligning(bs, 1);
512 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
516 bitstream_start(&bs);
517 nal_start_code_prefix(&bs);
518 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
522 *header_buffer = (unsigned char *)bs.buffer;
523 return bs.bit_offset;
527 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
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);
537 *header_buffer = (unsigned char *)bs.buffer;
538 return bs.bit_offset;
541 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
544 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
545 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
547 bitstream_start(&bs);
548 nal_start_code_prefix(&bs);
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);
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);
562 *header_buffer = (unsigned char *)bs.buffer;
563 return bs.bit_offset;
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)
595 // General pts/dts strategy:
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.)
607 // Many strategies are possible, but we establish these rules:
609 // - Tip frames have dts = pts - (C-O)*N.
610 // - Non-tip frames have dts = dts_last + N.
612 // An example, with C=2 and N=10 and the data flow showed with arrows:
615 // pts: 30 40 50 60 70 80
617 // dts: 10 30 20 60 50←40
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:
627 // pts: 30 40 80 90 100 110
629 // dts: 10 30 20 90 50←40
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.
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.
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)
650 int encoding_order_gop = 0;
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;
659 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
663 if (intra_period == 0)
664 intra_idr_period = 0;
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;
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;
678 *frame_type = FRAME_P;
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.
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.
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;
705 *frame_type = FRAME_P;
706 *displaying_order = encoding_order + ip_period - 1;
711 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
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;
725 global_flags.use_zerocopy = use_zerocopy;
728 VADisplayWithCleanup::~VADisplayWithCleanup()
730 if (va_dpy != nullptr) {
733 if (x11_display != nullptr) {
734 XCloseDisplay(x11_display);
741 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
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");
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) {
758 } else { // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
759 int drm_fd = open(va_display.c_str(), O_RDWR);
761 perror(va_display.c_str());
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) {
775 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, VAProfile *h264_profile, string *error)
777 unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
778 if (va_dpy == nullptr) {
779 if (error) *error = "Opening VA display failed";
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) {
786 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
787 if (error != nullptr) *error = buf;
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";
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) {
807 // We found a usable encoder, so return it.
808 if (h264_profile != nullptr) {
809 *h264_profile = profile_list[i];
815 if (error != nullptr) *error = "Can't find VAEntrypointEncSlice for H264 profiles";
819 int QuickSyncEncoderImpl::init_va(const string &va_display)
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());
827 if (!va_dpy->can_use_zerocopy) {
828 use_zerocopy = false;
831 switch (h264_profile) {
832 case VAProfileH264ConstrainedBaseline:
833 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
837 case VAProfileH264Main:
838 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
841 case VAProfileH264High:
842 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
845 h264_profile = VAProfileH264ConstrainedBaseline;
847 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
851 VAConfigAttrib attrib[VAConfigAttribTypeMax];
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;
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");
865 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
866 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
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");
876 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
877 config_attrib[config_attrib_num].value = VA_RC_CQP;
882 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
883 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
885 h264_packedheader = 1;
886 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
887 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
889 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
890 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
893 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
894 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
897 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
898 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
901 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
902 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
905 enc_packed_header_idx = config_attrib_num;
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;
915 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
916 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
922 int QuickSyncEncoderImpl::setup_encode()
924 if (!global_flags.x264_video_to_disk) {
926 VASurfaceID *tmp_surfaceid;
928 VASurfaceID src_surface[SURFACE_NUM];
929 VASurfaceID ref_surface[SURFACE_NUM];
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");
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,
940 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
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,
947 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
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));
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];
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,
962 tmp_surfaceid, 2 * SURFACE_NUM,
964 CHECK_VASTATUS(va_status, "vaCreateContext");
967 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
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
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");
982 /* create OpenGL objects */
983 for (int i = 0; i < SURFACE_NUM; i++) {
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);
988 size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
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);
1007 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1010 template<class T, class C>
1011 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
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);
1018 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1020 if (frame_type == FRAME_B)
1023 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1025 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1026 unique_lock<mutex> lock(storage_task_queue_mutex);
1028 // Insert the new frame at the start of the reference queue.
1029 reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1031 if (reference_frames.size() > num_ref_frames)
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();
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;
1044 current_ref_frame_num++;
1045 if (current_ref_frame_num > MaxFrameNum)
1046 current_ref_frame_num = 0;
1050 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1052 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1053 return a.frame_idx > b.frame_idx;
1056 for (size_t i = 0; i < reference_frames.size(); ++i) {
1057 RefPicList0_P[i] = reference_frames[i].pic;
1059 sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1062 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1064 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1065 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1067 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1068 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
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;
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);
1080 int QuickSyncEncoderImpl::render_sequence()
1082 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1084 VAEncMiscParameterBuffer *misc_param;
1085 VAEncMiscParameterRateControl *misc_rate_ctrl;
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;
1092 seq_param.intra_period = intra_period;
1093 seq_param.intra_idr_period = intra_idr_period;
1094 seq_param.ip_period = ip_period;
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;
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;
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");
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");
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);
1138 render_id[0] = seq_param_buf;
1139 render_id[1] = rc_param_buf;
1141 render_picture_and_delete(va_dpy->va_dpy, context_id, &render_id[0], 2);
1146 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1148 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1149 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1150 int PicOrderCntMsb, TopFieldOrderCnt;
1152 if (frame_type == FRAME_IDR)
1153 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1155 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1156 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
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;
1166 PicOrderCntMsb = prevPicOrderCntMsb;
1168 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1170 if (frame_type != FRAME_B) {
1171 PicOrderCntMsb_ref = PicOrderCntMsb;
1172 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1175 return TopFieldOrderCnt;
1178 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1180 VABufferID pic_param_buf;
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;
1191 for (i = 0; i < reference_frames.size(); i++) {
1192 pic_param.ReferenceFrames[i] = reference_frames[i].pic;
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;
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;
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");
1212 render_picture_and_delete(va_dpy->va_dpy, context_id, &pic_param_buf, 1);
1217 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
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;
1225 length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1227 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
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,
1233 VAEncPackedHeaderParameterBufferType,
1234 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1235 &packedseq_para_bufid);
1236 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1238 va_status = vaCreateBuffer(va_dpy->va_dpy,
1240 VAEncPackedHeaderDataBufferType,
1241 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1242 &packedseq_data_bufid);
1243 CHECK_VASTATUS(va_status, "vaCreateBuffer");
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);
1249 free(packedseq_buffer);
1255 int QuickSyncEncoderImpl::render_packedpicture()
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;
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;
1268 va_status = vaCreateBuffer(va_dpy->va_dpy,
1270 VAEncPackedHeaderParameterBufferType,
1271 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1272 &packedpic_para_bufid);
1273 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1275 va_status = vaCreateBuffer(va_dpy->va_dpy,
1277 VAEncPackedHeaderDataBufferType,
1278 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1279 &packedpic_data_bufid);
1280 CHECK_VASTATUS(va_status, "vaCreateBuffer");
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);
1286 free(packedpic_buffer);
1291 void QuickSyncEncoderImpl::render_packedslice()
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;
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;
1304 va_status = vaCreateBuffer(va_dpy->va_dpy,
1306 VAEncPackedHeaderParameterBufferType,
1307 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1308 &packedslice_para_bufid);
1309 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1311 va_status = vaCreateBuffer(va_dpy->va_dpy,
1313 VAEncPackedHeaderDataBufferType,
1314 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1315 &packedslice_data_bufid);
1316 CHECK_VASTATUS(va_status, "vaCreateBuffer");
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);
1322 free(packedslice_buffer);
1325 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1327 VABufferID slice_param_buf;
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);
1342 int refpiclist0_max = h264_maxref & 0xffff;
1343 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
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;
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);
1353 int refpiclist0_max = h264_maxref & 0xffff;
1354 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
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;
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;
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;
1375 if (h264_packedheader &&
1376 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1377 render_packedslice();
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");
1383 render_picture_and_delete(va_dpy->va_dpy, context_id, &slice_param_buf, 1);
1390 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1392 VACodedBufferSegment *buf_list = NULL;
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;
1403 vaUnmapBuffer(va_dpy->va_dpy, surf->coded_buf);
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);
1412 memset(&pkt, 0, sizeof(pkt));
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;
1422 pkt.duration = task.duration;
1424 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
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());
1434 // this is weird. but it seems to put a new frame onto the queue
1435 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1437 unique_lock<mutex> lock(storage_task_queue_mutex);
1438 storage_task_queue.push(move(task));
1439 storage_task_queue_changed.notify_all();
1442 void QuickSyncEncoderImpl::storage_task_thread()
1444 pthread_setname_np(pthread_self(), "QS_Storage");
1446 storage_task current;
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);
1461 size_t display_order = current.display_order;
1462 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
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));
1469 // Unlock the frame, and all its references.
1471 unique_lock<mutex> lock(storage_task_queue_mutex);
1472 release_gl_surface(display_order);
1474 for (size_t frame_num : ref_display_frame_numbers) {
1475 release_gl_surface(frame_num);
1481 void QuickSyncEncoderImpl::release_encode()
1483 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1484 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, gl_surfaces[i].coded_buf);
1485 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
1486 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].src_surface, 1);
1487 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1488 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].ref_surface, 1);
1489 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1492 VAStatus va_status = vaDestroyContext(va_dpy->va_dpy, context_id);
1493 CHECK_VASTATUS(va_status, "vaDestroyContext");
1494 va_status = vaDestroyConfig(va_dpy->va_dpy, config_id);
1495 CHECK_VASTATUS(va_status, "vaDestroyConfig");
1498 void QuickSyncEncoderImpl::release_gl_resources()
1500 assert(is_shutdown);
1501 if (has_released_gl_resources) {
1505 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1507 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1508 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1510 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1511 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1512 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1513 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1517 has_released_gl_resources = true;
1520 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)
1521 : 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)
1523 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1524 open_output_file(filename);
1525 file_audio_encoder->add_mux(file_mux.get());
1527 frame_width_mbaligned = (frame_width + 15) & (~15);
1528 frame_height_mbaligned = (frame_height + 15) & (~15);
1532 if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1533 assert(x264_encoder != nullptr);
1535 assert(x264_encoder == nullptr);
1538 enable_zerocopy_if_possible();
1539 if (!global_flags.x264_video_to_disk) {
1540 init_va(va_display);
1544 if (!global_flags.x264_video_to_disk) {
1545 memset(&seq_param, 0, sizeof(seq_param));
1546 memset(&pic_param, 0, sizeof(pic_param));
1547 memset(&slice_param, 0, sizeof(slice_param));
1550 call_once(quick_sync_metrics_inited, [](){
1551 mixer_latency_histogram.init("mixer");
1552 qs_latency_histogram.init("quick_sync");
1553 current_file_mux_metrics.init({{ "destination", "current_file" }});
1554 total_mux_metrics.init({{ "destination", "files_total" }});
1555 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1556 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1559 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1561 encode_thread = thread([this]{
1562 QOpenGLContext *context = create_context(this->surface);
1563 eglBindAPI(EGL_OPENGL_API);
1564 if (!make_current(context, this->surface)) {
1565 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1569 encode_thread_func();
1570 delete_context(context);
1574 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1577 release_gl_resources();
1580 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1582 for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1583 if (gl_surfaces[i].refcount == 0) {
1584 ++gl_surfaces[i].refcount;
1585 return &gl_surfaces[i];
1591 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1593 assert(surface_for_frame.count(display_frame_num));
1594 QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1595 if (--surf->refcount == 0) {
1596 assert(surface_for_frame.count(display_frame_num));
1597 surface_for_frame.erase(display_frame_num);
1598 storage_task_queue_changed.notify_all();
1602 bool QuickSyncEncoderImpl::is_zerocopy() const
1604 return use_zerocopy;
1607 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1609 assert(!is_shutdown);
1610 GLSurface *surf = nullptr;
1612 // Wait until this frame slot is done encoding.
1613 unique_lock<mutex> lock(storage_task_queue_mutex);
1614 surf = allocate_gl_surface();
1615 if (surf == nullptr) {
1616 fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1617 current_storage_frame);
1618 ++metric_quick_sync_stalled_frames;
1619 storage_task_queue_changed.wait(lock, [this, &surf]{
1620 if (storage_thread_should_quit)
1622 surf = allocate_gl_surface();
1623 return surf != nullptr;
1626 if (storage_thread_should_quit) return false;
1627 assert(surf != nullptr);
1628 surface_for_frame[current_storage_frame] = surf;
1632 *y_tex = surf->y_tex;
1633 *cbcr_tex = surf->cbcr_tex;
1635 surf->y_tex = *y_tex;
1636 surf->cbcr_tex = *cbcr_tex;
1639 if (!global_flags.x264_video_to_disk) {
1640 VAStatus va_status = vaDeriveImage(va_dpy->va_dpy, surf->src_surface, &surf->surface_image);
1641 CHECK_VASTATUS(va_status, "vaDeriveImage");
1644 VABufferInfo buf_info;
1645 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1646 va_status = vaAcquireBufferHandle(va_dpy->va_dpy, surf->surface_image.buf, &buf_info);
1647 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1650 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1651 EGLint y_attribs[] = {
1652 EGL_WIDTH, frame_width,
1653 EGL_HEIGHT, frame_height,
1654 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1655 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1656 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1657 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1661 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1662 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1664 // Associate Y image to a texture.
1665 glBindTexture(GL_TEXTURE_2D, *y_tex);
1666 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1668 // Create CbCr image.
1669 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1670 EGLint cbcr_attribs[] = {
1671 EGL_WIDTH, frame_width / 2,
1672 EGL_HEIGHT, frame_height / 2,
1673 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1674 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1675 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1676 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1680 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1681 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1683 // Associate CbCr image to a texture.
1684 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1685 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1689 current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1694 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1696 lock_guard<mutex> lock(file_audio_encoder_mutex);
1697 assert(!is_shutdown);
1698 file_audio_encoder->encode_audio(audio, pts + global_delay());
1701 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1703 assert(!is_shutdown);
1705 if (!use_zerocopy) {
1706 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1709 unique_lock<mutex> lock(storage_task_queue_mutex);
1710 surf = surface_for_frame[current_storage_frame];
1711 assert(surf != nullptr);
1714 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1717 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1720 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1722 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1725 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1727 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1730 // We don't own these; the caller does.
1731 surf->y_tex = surf->cbcr_tex = 0;
1733 glBindTexture(GL_TEXTURE_2D, 0);
1735 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1738 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1742 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1744 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1748 unique_lock<mutex> lock(frame_queue_mutex);
1749 current_video_frame.fence = fence;
1750 pending_video_frames.push(move(current_video_frame));
1751 ++current_storage_frame;
1753 frame_queue_nonempty.notify_all();
1757 void QuickSyncEncoderImpl::shutdown()
1764 unique_lock<mutex> lock(frame_queue_mutex);
1765 encode_thread_should_quit = true;
1766 frame_queue_nonempty.notify_all();
1768 encode_thread.join();
1770 unique_lock<mutex> lock(storage_task_queue_mutex);
1771 storage_thread_should_quit = true;
1772 frame_queue_nonempty.notify_all();
1773 storage_task_queue_changed.notify_all();
1775 storage_thread.join();
1777 // Encode any leftover audio in the queues, and also any delayed frames.
1779 lock_guard<mutex> lock(file_audio_encoder_mutex);
1780 file_audio_encoder->encode_last_audio();
1783 if (!global_flags.x264_video_to_disk) {
1790 void QuickSyncEncoderImpl::close_file()
1793 metric_current_file_start_time_seconds = 0.0 / 0.0;
1796 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1798 AVFormatContext *avctx = avformat_alloc_context();
1799 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1800 avctx->url = strdup(filename.c_str());
1802 string url = "file:" + filename;
1803 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1805 char tmp[AV_ERROR_MAX_STRING_SIZE];
1806 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1810 string video_extradata; // FIXME: See other comment about global headers.
1811 if (global_flags.x264_video_to_disk) {
1812 video_extradata = x264_encoder->get_global_headers();
1815 current_file_mux_metrics.reset();
1818 lock_guard<mutex> lock(file_audio_encoder_mutex);
1819 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1820 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), get_color_space(global_flags.ycbcr_rec709_coefficients), Mux::WITH_AUDIO, TIMEBASE,
1821 std::bind(&DiskSpaceEstimator::report_append, disk_space_estimator, filename, _1),
1822 Mux::WRITE_BACKGROUND,
1823 { ¤t_file_mux_metrics, &total_mux_metrics }));
1825 metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1827 if (global_flags.x264_video_to_disk) {
1828 x264_encoder->add_mux(file_mux.get());
1832 void QuickSyncEncoderImpl::encode_thread_func()
1834 pthread_setname_np(pthread_self(), "QS_Encode");
1836 int64_t last_dts = -1;
1837 int gop_start_display_frame_num = 0;
1838 for (int display_frame_num = 0; ; ++display_frame_num) {
1839 // Wait for the frame to be in the queue. Note that this only means
1840 // we started rendering it.
1843 unique_lock<mutex> lock(frame_queue_mutex);
1844 frame_queue_nonempty.wait(lock, [this]{
1845 return encode_thread_should_quit || !pending_video_frames.empty();
1847 if (encode_thread_should_quit && pending_video_frames.empty()) {
1848 // We may have queued frames left in the reorder buffer
1849 // that were supposed to be B-frames, but have no P-frame
1850 // to be encoded against. If so, encode them all as
1851 // P-frames instead. Note that this happens under the mutex,
1852 // but nobody else uses it at this point, since we're shutting down,
1853 // so there's no contention.
1854 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1857 frame = move(pending_video_frames.front());
1858 pending_video_frames.pop();
1862 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1863 // Note that this implicitly waits for the frame to be done rendering.
1864 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1866 if (global_flags.x264_video_to_disk) {
1867 unique_lock<mutex> lock(storage_task_queue_mutex);
1868 release_gl_surface(display_frame_num);
1872 reorder_buffer[display_frame_num] = move(frame);
1874 // Now encode as many QuickSync frames as we can using the frames we have available.
1875 // (It could be zero, or it could be multiple.) FIXME: make a function.
1878 int frame_type, quicksync_display_frame_num;
1879 encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1880 &quicksync_display_frame_num, &frame_type, &pts_lag);
1881 if (!reorder_buffer.count(quicksync_display_frame_num)) {
1884 frame = move(reorder_buffer[quicksync_display_frame_num]);
1885 reorder_buffer.erase(quicksync_display_frame_num);
1887 if (frame_type == FRAME_IDR) {
1888 // Release any reference frames from the previous GOP.
1890 unique_lock<mutex> lock(storage_task_queue_mutex);
1891 for (const ReferenceFrame &frame : reference_frames) {
1892 release_gl_surface(frame.display_number);
1895 reference_frames.clear();
1896 current_ref_frame_num = 0;
1897 gop_start_display_frame_num = quicksync_display_frame_num;
1900 // Determine the dts of this frame.
1902 if (pts_lag == -1) {
1903 assert(last_dts != -1);
1904 dts = last_dts + (TIMEBASE / MAX_FPS);
1906 dts = frame.pts - pts_lag;
1910 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);
1911 ++quicksync_encoding_frame_num;
1916 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1918 if (reorder_buffer.empty()) {
1922 for (auto &pending_frame : reorder_buffer) {
1923 int display_frame_num = pending_frame.first;
1924 assert(display_frame_num > 0);
1925 PendingFrame frame = move(pending_frame.second);
1926 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1927 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1928 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1933 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1936 memset(&pkt, 0, sizeof(pkt));
1938 pkt.data = const_cast<uint8_t *>(data);
1939 pkt.size = frame_width * frame_height * 2;
1940 pkt.stream_index = 0;
1941 pkt.flags = AV_PKT_FLAG_KEY;
1942 pkt.duration = duration;
1943 stream_mux->add_packet(pkt, pts, pts);
1946 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1948 if (src_width == dst_pitch) {
1949 memcpy(dst, src, src_width * height);
1951 for (size_t y = 0; y < height; ++y) {
1952 const uint8_t *sptr = src + y * src_width;
1953 uint8_t *dptr = dst + y * dst_pitch;
1954 memcpy(dptr, sptr, src_width);
1959 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1961 // Wait for the GPU to be done with the frame.
1964 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1966 if (sync_status == GL_TIMEOUT_EXPIRED) {
1967 // NVIDIA likes to busy-wait; yield instead.
1968 this_thread::sleep_for(milliseconds(1));
1970 } while (sync_status == GL_TIMEOUT_EXPIRED);
1971 assert(sync_status != GL_WAIT_FAILED);
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);
1978 // Release back any input frames we needed to render this frame.
1979 frame.input_frames.clear();
1983 unique_lock<mutex> lock(storage_task_queue_mutex);
1984 surf = surface_for_frame[display_frame_num];
1985 assert(surf != nullptr);
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);
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)
1998 const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
2002 unique_lock<mutex> lock(storage_task_queue_mutex);
2003 surf = surface_for_frame[display_frame_num];
2004 assert(surf != nullptr);
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");
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);
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);
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);
2024 va_status = vaUnmapBuffer(va_dpy->va_dpy, surf->surface_image.buf);
2025 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2028 va_status = vaDestroyImage(va_dpy->va_dpy, surf->surface_image.image_id);
2029 CHECK_VASTATUS(va_status, "vaDestroyImage");
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");
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?).
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.
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();
2051 //render_sequence();
2052 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2054 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2056 va_status = vaEndPicture(va_dpy->va_dpy, context_id);
2057 CHECK_VASTATUS(va_status, "vaEndPicture");
2059 update_ReferenceFrames(display_frame_num, frame_type);
2061 vector<size_t> ref_display_frame_numbers;
2063 // Lock the references for this frame; otherwise, they could be
2064 // rendered to before this frame is done encoding.
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);
2074 // so now the data is done encoding (well, async job kicked off)...
2075 // we send that to the storage thread
2077 tmp.display_order = display_frame_num;
2078 tmp.frame_type = frame_type;
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));
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)) {}
2092 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2093 QuickSyncEncoder::~QuickSyncEncoder() {}
2095 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2097 impl->add_audio(pts, audio);
2100 bool QuickSyncEncoder::is_zerocopy() const
2102 return impl->is_zerocopy();
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)
2107 return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2110 RefCountedGLsync QuickSyncEncoder::end_frame()
2112 return impl->end_frame();
2115 void QuickSyncEncoder::shutdown()
2120 void QuickSyncEncoder::close_file()
2125 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2127 impl->set_stream_mux(mux);
2130 int64_t QuickSyncEncoder::global_delay() const {
2131 return impl->global_delay();
2134 string QuickSyncEncoder::get_usable_va_display()
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;
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");
2153 fprintf(stderr, "No --va-display was given, and the X11 display did not expose a VA-API H.264 encoder.\n");
2155 // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
2157 int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
2159 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
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",
2168 if (need_env_reset) {
2169 unsetenv("LIBVA_MESSAGING_LEVEL");
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");