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;
722 } else if (!global_flags.v4l_output_device.empty()) {
723 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --v4l-output.\n");
724 use_zerocopy = false;
728 global_flags.use_zerocopy = use_zerocopy;
731 static unique_ptr<VADisplayWithCleanup> try_open_va_h264(const string &va_display, VAProfile *h264_profile, string *error)
733 return try_open_va(va_display, { VAProfileH264High, VAProfileH264Main, VAProfileH264ConstrainedBaseline },
734 VAEntrypointEncSlice, /*desired_configs=*/{}, h264_profile, error);
737 int QuickSyncEncoderImpl::init_va(const string &va_display)
740 va_dpy = try_open_va_h264(va_display, &h264_profile, &error);
741 if (va_dpy == nullptr) {
742 fprintf(stderr, "error: %s\n", error.c_str());
745 if (!va_dpy->can_use_zerocopy) {
746 use_zerocopy = false;
749 switch (h264_profile) {
750 case VAProfileH264ConstrainedBaseline:
751 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
755 case VAProfileH264Main:
756 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
759 case VAProfileH264High:
760 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
763 h264_profile = VAProfileH264ConstrainedBaseline;
765 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
769 VAConfigAttrib attrib[VAConfigAttribTypeMax];
771 /* find out the format for the render target, and rate control mode */
772 for (unsigned i = 0; i < VAConfigAttribTypeMax; i++)
773 attrib[i].type = (VAConfigAttribType)i;
775 VAStatus va_status = vaGetConfigAttributes(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
776 &attrib[0], VAConfigAttribTypeMax);
777 CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
778 /* check the interested configattrib */
779 if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
780 printf("Not find desired YUV420 RT format\n");
783 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
784 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
788 if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
789 if (!(attrib[VAConfigAttribRateControl].value & VA_RC_CQP)) {
790 fprintf(stderr, "ERROR: VA-API encoder does not support CQP mode.\n");
794 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
795 config_attrib[config_attrib_num].value = VA_RC_CQP;
800 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
801 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
803 h264_packedheader = 1;
804 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
805 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
807 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
808 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
811 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
812 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
815 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
816 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
819 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
820 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
823 enc_packed_header_idx = config_attrib_num;
827 if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
828 config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
829 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
833 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
834 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
840 int QuickSyncEncoderImpl::setup_encode()
842 if (!global_flags.x264_video_to_disk) {
844 VASurfaceID *tmp_surfaceid;
846 VASurfaceID src_surface[SURFACE_NUM];
847 VASurfaceID ref_surface[SURFACE_NUM];
849 va_status = vaCreateConfig(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
850 &config_attrib[0], config_attrib_num, &config_id);
851 CHECK_VASTATUS(va_status, "vaCreateConfig");
853 /* create source surfaces */
854 va_status = vaCreateSurfaces(va_dpy->va_dpy,
855 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
856 &src_surface[0], SURFACE_NUM,
858 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
860 /* create reference surfaces */
861 va_status = vaCreateSurfaces(va_dpy->va_dpy,
862 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
863 &ref_surface[0], SURFACE_NUM,
865 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
867 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
868 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
869 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
871 for (int i = 0; i < SURFACE_NUM; i++) {
872 gl_surfaces[i].src_surface = src_surface[i];
873 gl_surfaces[i].ref_surface = ref_surface[i];
876 /* Create a context for this encode pipe */
877 va_status = vaCreateContext(va_dpy->va_dpy, config_id,
878 frame_width_mbaligned, frame_height_mbaligned,
880 tmp_surfaceid, 2 * SURFACE_NUM,
882 CHECK_VASTATUS(va_status, "vaCreateContext");
885 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
887 for (int i = 0; i < SURFACE_NUM; i++) {
888 /* create coded buffer once for all
889 * other VA buffers which won't be used again after vaRenderPicture.
890 * so APP can always vaCreateBuffer for every frame
891 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
892 * so VA won't maintain the coded buffer
894 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncCodedBufferType,
895 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
896 CHECK_VASTATUS(va_status, "vaCreateBuffer");
900 /* create OpenGL objects */
901 for (int i = 0; i < SURFACE_NUM; i++) {
903 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
904 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
906 size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
908 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
909 // buffers, due to potentially differing pitch.
910 glGenBuffers(1, &gl_surfaces[i].pbo);
911 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
912 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);
913 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);
914 gl_surfaces[i].y_offset = 0;
915 gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
916 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
917 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
918 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
925 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
928 template<class T, class C>
929 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
931 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
932 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
933 sort(middle, end, less_than);
936 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
938 if (frame_type == FRAME_B)
941 pic_param.CurrPic.frame_idx = current_ref_frame_num;
943 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
944 lock_guard<mutex> lock(storage_task_queue_mutex);
946 // Insert the new frame at the start of the reference queue.
947 reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
949 if (reference_frames.size() > num_ref_frames)
951 // The back frame frame is no longer in use as a reference.
952 int display_frame_num = reference_frames.back().display_number;
953 assert(surface_for_frame.count(display_frame_num));
954 release_gl_surface(display_frame_num);
955 reference_frames.pop_back();
958 // Mark this frame in use as a reference.
959 assert(surface_for_frame.count(current_display_frame));
960 ++surface_for_frame[current_display_frame]->refcount;
962 current_ref_frame_num++;
963 if (current_ref_frame_num > MaxFrameNum)
964 current_ref_frame_num = 0;
968 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
970 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
971 return a.frame_idx > b.frame_idx;
974 for (size_t i = 0; i < reference_frames.size(); ++i) {
975 RefPicList0_P[i] = reference_frames[i].pic;
977 sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
980 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
982 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
983 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
985 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
986 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
989 for (size_t i = 0; i < reference_frames.size(); ++i) {
990 RefPicList0_B[i] = reference_frames[i].pic;
991 RefPicList1_B[i] = reference_frames[i].pic;
993 sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
994 sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
998 int QuickSyncEncoderImpl::render_sequence()
1000 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1002 VAEncMiscParameterBuffer *misc_param;
1003 VAEncMiscParameterRateControl *misc_rate_ctrl;
1005 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1006 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1007 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1008 seq_param.bits_per_second = frame_bitrate;
1010 seq_param.intra_period = intra_period;
1011 seq_param.intra_idr_period = intra_idr_period;
1012 seq_param.ip_period = ip_period;
1014 seq_param.max_num_ref_frames = num_ref_frames;
1015 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1016 seq_param.time_scale = TIMEBASE * 2;
1017 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1018 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1019 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1020 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1021 seq_param.seq_fields.bits.chroma_format_idc = 1;
1022 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1024 if (frame_width != frame_width_mbaligned ||
1025 frame_height != frame_height_mbaligned) {
1026 seq_param.frame_cropping_flag = 1;
1027 seq_param.frame_crop_left_offset = 0;
1028 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1029 seq_param.frame_crop_top_offset = 0;
1030 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1033 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1034 VAEncSequenceParameterBufferType,
1035 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1036 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1038 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1039 VAEncMiscParameterBufferType,
1040 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1041 1, NULL, &rc_param_buf);
1042 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1044 vaMapBuffer(va_dpy->va_dpy, rc_param_buf, (void **)&misc_param);
1045 misc_param->type = VAEncMiscParameterTypeRateControl;
1046 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1047 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1048 misc_rate_ctrl->bits_per_second = frame_bitrate;
1049 misc_rate_ctrl->target_percentage = 66;
1050 misc_rate_ctrl->window_size = 1000;
1051 misc_rate_ctrl->initial_qp = initial_qp;
1052 misc_rate_ctrl->min_qp = minimal_qp;
1053 misc_rate_ctrl->basic_unit_size = 0;
1054 vaUnmapBuffer(va_dpy->va_dpy, rc_param_buf);
1056 render_id[0] = seq_param_buf;
1057 render_id[1] = rc_param_buf;
1059 render_picture_and_delete(va_dpy->va_dpy, context_id, &render_id[0], 2);
1064 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1066 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1067 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1068 int PicOrderCntMsb, TopFieldOrderCnt;
1070 if (frame_type == FRAME_IDR)
1071 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1073 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1074 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1077 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1078 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1079 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1080 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1081 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1082 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1084 PicOrderCntMsb = prevPicOrderCntMsb;
1086 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1088 if (frame_type != FRAME_B) {
1089 PicOrderCntMsb_ref = PicOrderCntMsb;
1090 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1093 return TopFieldOrderCnt;
1096 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1098 VABufferID pic_param_buf;
1102 pic_param.CurrPic.picture_id = surf->ref_surface;
1103 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1104 pic_param.CurrPic.flags = 0;
1105 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1106 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1107 CurrentCurrPic = pic_param.CurrPic;
1109 for (i = 0; i < reference_frames.size(); i++) {
1110 pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1112 for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1113 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1114 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1117 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1118 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1119 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1120 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1121 pic_param.frame_num = current_ref_frame_num; // FIXME: is this correct?
1122 pic_param.coded_buf = surf->coded_buf;
1123 pic_param.last_picture = false; // FIXME
1124 pic_param.pic_init_qp = initial_qp;
1126 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncPictureParameterBufferType,
1127 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1128 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1130 render_picture_and_delete(va_dpy->va_dpy, context_id, &pic_param_buf, 1);
1135 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1137 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1138 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1139 unsigned int length_in_bits;
1140 unsigned char *packedseq_buffer = NULL;
1143 length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1145 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1147 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1148 packedheader_param_buffer.has_emulation_bytes = 0;
1149 va_status = vaCreateBuffer(va_dpy->va_dpy,
1151 VAEncPackedHeaderParameterBufferType,
1152 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1153 &packedseq_para_bufid);
1154 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1156 va_status = vaCreateBuffer(va_dpy->va_dpy,
1158 VAEncPackedHeaderDataBufferType,
1159 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1160 &packedseq_data_bufid);
1161 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1163 render_id[0] = packedseq_para_bufid;
1164 render_id[1] = packedseq_data_bufid;
1165 render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1167 free(packedseq_buffer);
1173 int QuickSyncEncoderImpl::render_packedpicture()
1175 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1176 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1177 unsigned int length_in_bits;
1178 unsigned char *packedpic_buffer = NULL;
1181 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1182 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1183 packedheader_param_buffer.bit_length = length_in_bits;
1184 packedheader_param_buffer.has_emulation_bytes = 0;
1186 va_status = vaCreateBuffer(va_dpy->va_dpy,
1188 VAEncPackedHeaderParameterBufferType,
1189 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1190 &packedpic_para_bufid);
1191 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1193 va_status = vaCreateBuffer(va_dpy->va_dpy,
1195 VAEncPackedHeaderDataBufferType,
1196 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1197 &packedpic_data_bufid);
1198 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1200 render_id[0] = packedpic_para_bufid;
1201 render_id[1] = packedpic_data_bufid;
1202 render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1204 free(packedpic_buffer);
1209 void QuickSyncEncoderImpl::render_packedslice()
1211 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1212 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1213 unsigned int length_in_bits;
1214 unsigned char *packedslice_buffer = NULL;
1217 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1218 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1219 packedheader_param_buffer.bit_length = length_in_bits;
1220 packedheader_param_buffer.has_emulation_bytes = 0;
1222 va_status = vaCreateBuffer(va_dpy->va_dpy,
1224 VAEncPackedHeaderParameterBufferType,
1225 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1226 &packedslice_para_bufid);
1227 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1229 va_status = vaCreateBuffer(va_dpy->va_dpy,
1231 VAEncPackedHeaderDataBufferType,
1232 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1233 &packedslice_data_bufid);
1234 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1236 render_id[0] = packedslice_para_bufid;
1237 render_id[1] = packedslice_data_bufid;
1238 render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1240 free(packedslice_buffer);
1243 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1245 VABufferID slice_param_buf;
1249 /* one frame, one slice */
1250 slice_param.macroblock_address = 0;
1251 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1252 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1253 if (frame_type == FRAME_IDR) {
1254 if (encoding_frame_num != 0)
1255 ++slice_param.idr_pic_id;
1256 } else if (frame_type == FRAME_P) {
1257 VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1258 update_RefPicList_P(RefPicList0_P);
1260 int refpiclist0_max = h264_maxref & 0xffff;
1261 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1263 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1264 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1265 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1267 } else if (frame_type == FRAME_B) {
1268 VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1269 update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1271 int refpiclist0_max = h264_maxref & 0xffff;
1272 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1274 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1275 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1276 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1277 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1280 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1281 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1282 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1283 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1287 slice_param.slice_alpha_c0_offset_div2 = 0;
1288 slice_param.slice_beta_offset_div2 = 0;
1289 slice_param.direct_spatial_mv_pred_flag = 1;
1290 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1293 if (h264_packedheader &&
1294 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1295 render_packedslice();
1297 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncSliceParameterBufferType,
1298 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1299 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1301 render_picture_and_delete(va_dpy->va_dpy, context_id, &slice_param_buf, 1);
1308 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1310 VACodedBufferSegment *buf_list = NULL;
1315 va_status = vaMapBuffer(va_dpy->va_dpy, surf->coded_buf, (void **)(&buf_list));
1316 CHECK_VASTATUS(va_status, "vaMapBuffer");
1317 while (buf_list != NULL) {
1318 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1319 buf_list = (VACodedBufferSegment *) buf_list->next;
1321 vaUnmapBuffer(va_dpy->va_dpy, surf->coded_buf);
1323 static int frameno = 0;
1324 print_latency("Current Quick Sync latency (video inputs → disk mux):",
1325 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1330 memset(&pkt, 0, sizeof(pkt));
1332 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1333 pkt.size = data.size();
1334 pkt.stream_index = 0;
1335 if (task.frame_type == FRAME_IDR) {
1336 pkt.flags = AV_PKT_FLAG_KEY;
1340 pkt.duration = task.duration;
1342 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1344 if (!global_flags.uncompressed_video_to_http &&
1345 !global_flags.x264_video_to_http) {
1346 stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1352 // this is weird. but it seems to put a new frame onto the queue
1353 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1355 lock_guard<mutex> lock(storage_task_queue_mutex);
1356 storage_task_queue.push(move(task));
1357 storage_task_queue_changed.notify_all();
1360 void QuickSyncEncoderImpl::storage_task_thread()
1362 pthread_setname_np(pthread_self(), "QS_Storage");
1364 storage_task current;
1367 // wait until there's an encoded frame
1368 unique_lock<mutex> lock(storage_task_queue_mutex);
1369 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1370 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1371 current = move(storage_task_queue.front());
1372 storage_task_queue.pop();
1373 surf = surface_for_frame[current.display_order];
1374 assert(surf != nullptr);
1379 size_t display_order = current.display_order;
1380 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1382 // waits for data, then saves it to disk.
1383 va_status = vaSyncSurface(va_dpy->va_dpy, surf->src_surface);
1384 CHECK_VASTATUS(va_status, "vaSyncSurface");
1385 save_codeddata(surf, move(current));
1387 // Unlock the frame, and all its references.
1389 lock_guard<mutex> lock(storage_task_queue_mutex);
1390 release_gl_surface(display_order);
1392 for (size_t frame_num : ref_display_frame_numbers) {
1393 release_gl_surface(frame_num);
1399 void QuickSyncEncoderImpl::release_encode()
1401 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1402 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, gl_surfaces[i].coded_buf);
1403 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
1404 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].src_surface, 1);
1405 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1406 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].ref_surface, 1);
1407 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1410 VAStatus va_status = vaDestroyContext(va_dpy->va_dpy, context_id);
1411 CHECK_VASTATUS(va_status, "vaDestroyContext");
1412 va_status = vaDestroyConfig(va_dpy->va_dpy, config_id);
1413 CHECK_VASTATUS(va_status, "vaDestroyConfig");
1416 void QuickSyncEncoderImpl::release_gl_resources()
1418 assert(is_shutdown);
1419 if (has_released_gl_resources) {
1423 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1425 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1426 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1428 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1429 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1430 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1431 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1435 has_released_gl_resources = true;
1438 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)
1439 : 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)
1441 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1442 open_output_file(filename);
1443 file_audio_encoder->add_mux(file_mux.get());
1445 frame_width_mbaligned = (frame_width + 15) & (~15);
1446 frame_height_mbaligned = (frame_height + 15) & (~15);
1450 if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1451 assert(x264_encoder != nullptr);
1453 assert(x264_encoder == nullptr);
1456 enable_zerocopy_if_possible();
1457 if (!global_flags.x264_video_to_disk) {
1458 init_va(va_display);
1462 if (!global_flags.x264_video_to_disk) {
1463 memset(&seq_param, 0, sizeof(seq_param));
1464 memset(&pic_param, 0, sizeof(pic_param));
1465 memset(&slice_param, 0, sizeof(slice_param));
1468 if (!global_flags.v4l_output_device.empty()) {
1469 v4l_output.reset(new V4LOutput(global_flags.v4l_output_device.c_str(), width, height));
1472 call_once(quick_sync_metrics_inited, [](){
1473 mixer_latency_histogram.init("mixer");
1474 qs_latency_histogram.init("quick_sync");
1475 current_file_mux_metrics.init({{ "destination", "current_file" }});
1476 total_mux_metrics.init({{ "destination", "files_total" }});
1477 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1478 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1481 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1483 encode_thread = thread([this]{
1484 QOpenGLContext *context = create_context(this->surface);
1485 eglBindAPI(EGL_OPENGL_API);
1486 if (!make_current(context, this->surface)) {
1487 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1491 encode_thread_func();
1492 delete_context(context);
1496 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1499 release_gl_resources();
1502 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1504 for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1505 if (gl_surfaces[i].refcount == 0) {
1506 ++gl_surfaces[i].refcount;
1507 return &gl_surfaces[i];
1513 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1515 assert(surface_for_frame.count(display_frame_num));
1516 QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1517 if (--surf->refcount == 0) {
1518 assert(surface_for_frame.count(display_frame_num));
1519 surface_for_frame.erase(display_frame_num);
1520 storage_task_queue_changed.notify_all();
1524 bool QuickSyncEncoderImpl::is_zerocopy() const
1526 return use_zerocopy;
1529 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1531 assert(!is_shutdown);
1532 GLSurface *surf = nullptr;
1534 // Wait until this frame slot is done encoding.
1535 unique_lock<mutex> lock(storage_task_queue_mutex);
1536 surf = allocate_gl_surface();
1537 if (surf == nullptr) {
1538 fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1539 current_storage_frame);
1540 ++metric_quick_sync_stalled_frames;
1541 storage_task_queue_changed.wait(lock, [this, &surf]{
1542 if (storage_thread_should_quit)
1544 surf = allocate_gl_surface();
1545 return surf != nullptr;
1548 if (storage_thread_should_quit) return false;
1549 assert(surf != nullptr);
1550 surface_for_frame[current_storage_frame] = surf;
1554 *y_tex = surf->y_tex;
1555 *cbcr_tex = surf->cbcr_tex;
1557 surf->y_tex = *y_tex;
1558 surf->cbcr_tex = *cbcr_tex;
1561 if (!global_flags.x264_video_to_disk) {
1562 VAStatus va_status = vaDeriveImage(va_dpy->va_dpy, surf->src_surface, &surf->surface_image);
1563 CHECK_VASTATUS(va_status, "vaDeriveImage");
1566 VABufferInfo buf_info;
1567 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1568 va_status = vaAcquireBufferHandle(va_dpy->va_dpy, surf->surface_image.buf, &buf_info);
1569 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1572 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1573 EGLint y_attribs[] = {
1574 EGL_WIDTH, frame_width,
1575 EGL_HEIGHT, frame_height,
1576 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1577 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1578 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1579 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1583 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1584 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1586 // Associate Y image to a texture.
1587 glBindTexture(GL_TEXTURE_2D, *y_tex);
1588 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1590 // Create CbCr image.
1591 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1592 EGLint cbcr_attribs[] = {
1593 EGL_WIDTH, frame_width / 2,
1594 EGL_HEIGHT, frame_height / 2,
1595 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1596 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1597 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1598 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1602 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1603 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1605 // Associate CbCr image to a texture.
1606 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1607 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1611 current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1616 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1618 lock_guard<mutex> lock(file_audio_encoder_mutex);
1619 assert(!is_shutdown);
1620 file_audio_encoder->encode_audio(audio, pts + global_delay());
1623 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1625 assert(!is_shutdown);
1627 if (!use_zerocopy) {
1628 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1631 lock_guard<mutex> lock(storage_task_queue_mutex);
1632 surf = surface_for_frame[current_storage_frame];
1633 assert(surf != nullptr);
1636 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1639 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1642 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1644 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1647 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1649 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1652 // We don't own these; the caller does.
1653 surf->y_tex = surf->cbcr_tex = 0;
1655 glBindTexture(GL_TEXTURE_2D, 0);
1657 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1660 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1664 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1666 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1670 lock_guard<mutex> lock(frame_queue_mutex);
1671 current_video_frame.fence = fence;
1672 pending_video_frames.push(move(current_video_frame));
1673 ++current_storage_frame;
1675 frame_queue_nonempty.notify_all();
1679 void QuickSyncEncoderImpl::shutdown()
1686 lock_guard<mutex> lock(frame_queue_mutex);
1687 encode_thread_should_quit = true;
1688 frame_queue_nonempty.notify_all();
1690 encode_thread.join();
1692 lock_guard<mutex> lock(storage_task_queue_mutex);
1693 storage_thread_should_quit = true;
1694 frame_queue_nonempty.notify_all();
1695 storage_task_queue_changed.notify_all();
1697 storage_thread.join();
1699 // Encode any leftover audio in the queues, and also any delayed frames.
1701 lock_guard<mutex> lock(file_audio_encoder_mutex);
1702 file_audio_encoder->encode_last_audio();
1705 if (!global_flags.x264_video_to_disk) {
1712 void QuickSyncEncoderImpl::close_file()
1715 metric_current_file_start_time_seconds = 0.0 / 0.0;
1718 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1720 AVFormatContext *avctx = avformat_alloc_context();
1721 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1722 avctx->url = strdup(filename.c_str());
1724 string url = "file:" + filename;
1725 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1727 char tmp[AV_ERROR_MAX_STRING_SIZE];
1728 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1732 string video_extradata; // FIXME: See other comment about global headers.
1733 if (global_flags.x264_video_to_disk) {
1734 video_extradata = x264_encoder->get_global_headers();
1737 current_file_mux_metrics.reset();
1740 lock_guard<mutex> lock(file_audio_encoder_mutex);
1741 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1742 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), TIMEBASE,
1743 std::bind(&DiskSpaceEstimator::report_append, disk_space_estimator, filename, _1),
1744 Mux::WRITE_BACKGROUND,
1745 { ¤t_file_mux_metrics, &total_mux_metrics }));
1747 metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1749 if (global_flags.x264_video_to_disk) {
1750 x264_encoder->add_mux(file_mux.get());
1754 void QuickSyncEncoderImpl::encode_thread_func()
1756 pthread_setname_np(pthread_self(), "QS_Encode");
1758 int64_t last_dts = -1;
1759 int gop_start_display_frame_num = 0;
1760 for (int display_frame_num = 0; ; ++display_frame_num) {
1761 // Wait for the frame to be in the queue. Note that this only means
1762 // we started rendering it.
1765 unique_lock<mutex> lock(frame_queue_mutex);
1766 frame_queue_nonempty.wait(lock, [this]{
1767 return encode_thread_should_quit || !pending_video_frames.empty();
1769 if (encode_thread_should_quit && pending_video_frames.empty()) {
1770 // We may have queued frames left in the reorder buffer
1771 // that were supposed to be B-frames, but have no P-frame
1772 // to be encoded against. If so, encode them all as
1773 // P-frames instead. Note that this happens under the mutex,
1774 // but nobody else uses it at this point, since we're shutting down,
1775 // so there's no contention.
1776 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1779 frame = move(pending_video_frames.front());
1780 pending_video_frames.pop();
1784 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1785 // Note that this implicitly waits for the frame to be done rendering.
1786 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1788 if (global_flags.x264_video_to_disk) {
1789 lock_guard<mutex> lock(storage_task_queue_mutex);
1790 release_gl_surface(display_frame_num);
1794 reorder_buffer[display_frame_num] = move(frame);
1796 // Now encode as many QuickSync frames as we can using the frames we have available.
1797 // (It could be zero, or it could be multiple.) FIXME: make a function.
1800 int frame_type, quicksync_display_frame_num;
1801 encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1802 &quicksync_display_frame_num, &frame_type, &pts_lag);
1803 if (!reorder_buffer.count(quicksync_display_frame_num)) {
1806 frame = move(reorder_buffer[quicksync_display_frame_num]);
1807 reorder_buffer.erase(quicksync_display_frame_num);
1809 if (frame_type == FRAME_IDR) {
1810 // Release any reference frames from the previous GOP.
1812 lock_guard<mutex> lock(storage_task_queue_mutex);
1813 for (const ReferenceFrame &frame : reference_frames) {
1814 release_gl_surface(frame.display_number);
1817 reference_frames.clear();
1818 current_ref_frame_num = 0;
1819 gop_start_display_frame_num = quicksync_display_frame_num;
1822 // Determine the dts of this frame.
1824 if (pts_lag == -1) {
1825 assert(last_dts != -1);
1826 dts = last_dts + (TIMEBASE / MAX_FPS);
1828 dts = frame.pts - pts_lag;
1832 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);
1833 ++quicksync_encoding_frame_num;
1838 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1840 if (reorder_buffer.empty()) {
1844 for (auto &pending_frame : reorder_buffer) {
1845 int display_frame_num = pending_frame.first;
1846 assert(display_frame_num > 0);
1847 PendingFrame frame = move(pending_frame.second);
1848 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1849 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1850 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1855 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1858 memset(&pkt, 0, sizeof(pkt));
1860 pkt.data = const_cast<uint8_t *>(data);
1861 pkt.size = frame_width * frame_height * 2;
1862 pkt.stream_index = 0;
1863 pkt.flags = AV_PKT_FLAG_KEY;
1864 pkt.duration = duration;
1865 stream_mux->add_packet(pkt, pts, pts);
1868 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1870 if (src_width == dst_pitch) {
1871 memcpy(dst, src, src_width * height);
1873 for (size_t y = 0; y < height; ++y) {
1874 const uint8_t *sptr = src + y * src_width;
1875 uint8_t *dptr = dst + y * dst_pitch;
1876 memcpy(dptr, sptr, src_width);
1881 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1883 // Wait for the GPU to be done with the frame.
1886 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1888 if (sync_status == GL_TIMEOUT_EXPIRED) {
1889 // NVIDIA likes to busy-wait; yield instead.
1890 this_thread::sleep_for(milliseconds(1));
1892 } while (sync_status == GL_TIMEOUT_EXPIRED);
1893 assert(sync_status != GL_WAIT_FAILED);
1895 ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1896 static int frameno = 0;
1897 print_latency("Current mixer latency (video inputs → ready for encode):",
1898 received_ts, false, &frameno, &mixer_latency_histogram);
1900 // Release back any input frames we needed to render this frame.
1901 frame.input_frames.clear();
1905 lock_guard<mutex> lock(storage_task_queue_mutex);
1906 surf = surface_for_frame[display_frame_num];
1907 assert(surf != nullptr);
1909 uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1910 if (global_flags.uncompressed_video_to_http) {
1911 add_packet_for_uncompressed_frame(pts, duration, data);
1912 } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1913 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1916 if (v4l_output != nullptr) {
1917 v4l_output->send_frame(data);
1921 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1922 int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1924 const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1928 lock_guard<mutex> lock(storage_task_queue_mutex);
1929 surf = surface_for_frame[display_frame_num];
1930 assert(surf != nullptr);
1935 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1936 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1937 va_status = vaReleaseBufferHandle(va_dpy->va_dpy, surf->surface_image.buf);
1938 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1940 // Upload the frame to VA-API.
1941 unsigned char *surface_p = nullptr;
1942 vaMapBuffer(va_dpy->va_dpy, surf->surface_image.buf, (void **)&surface_p);
1944 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1945 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1947 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1948 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1950 va_status = vaUnmapBuffer(va_dpy->va_dpy, surf->surface_image.buf);
1951 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1954 va_status = vaDestroyImage(va_dpy->va_dpy, surf->surface_image.image_id);
1955 CHECK_VASTATUS(va_status, "vaDestroyImage");
1957 // Schedule the frame for encoding.
1958 VASurfaceID va_surface = surf->src_surface;
1959 va_status = vaBeginPicture(va_dpy->va_dpy, context_id, va_surface);
1960 CHECK_VASTATUS(va_status, "vaBeginPicture");
1962 if (frame_type == FRAME_IDR) {
1963 // FIXME: If the mux wants global headers, we should not put the
1964 // SPS/PPS before each IDR frame, but rather put it into the
1965 // codec extradata (formatted differently?).
1967 // NOTE: If we change ycbcr_coefficients, it will not take effect
1968 // before the next IDR frame. This is acceptable, as it should only
1969 // happen on a mode change, which is rare.
1971 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1972 if (h264_packedheader) {
1973 render_packedsequence(ycbcr_coefficients);
1974 render_packedpicture();
1977 //render_sequence();
1978 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1980 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1982 va_status = vaEndPicture(va_dpy->va_dpy, context_id);
1983 CHECK_VASTATUS(va_status, "vaEndPicture");
1985 update_ReferenceFrames(display_frame_num, frame_type);
1987 vector<size_t> ref_display_frame_numbers;
1989 // Lock the references for this frame; otherwise, they could be
1990 // rendered to before this frame is done encoding.
1992 lock_guard<mutex> lock(storage_task_queue_mutex);
1993 for (const ReferenceFrame &frame : reference_frames) {
1994 assert(surface_for_frame.count(frame.display_number));
1995 ++surface_for_frame[frame.display_number]->refcount;
1996 ref_display_frame_numbers.push_back(frame.display_number);
2000 // so now the data is done encoding (well, async job kicked off)...
2001 // we send that to the storage thread
2003 tmp.display_order = display_frame_num;
2004 tmp.frame_type = frame_type;
2007 tmp.duration = duration;
2008 tmp.ycbcr_coefficients = ycbcr_coefficients;
2009 tmp.received_ts = received_ts;
2010 tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2011 storage_task_enqueue(move(tmp));
2015 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)
2016 : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2018 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2019 QuickSyncEncoder::~QuickSyncEncoder() {}
2021 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2023 impl->add_audio(pts, audio);
2026 bool QuickSyncEncoder::is_zerocopy() const
2028 return impl->is_zerocopy();
2031 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2033 return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2036 RefCountedGLsync QuickSyncEncoder::end_frame()
2038 return impl->end_frame();
2041 void QuickSyncEncoder::shutdown()
2046 void QuickSyncEncoder::close_file()
2051 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2053 impl->set_stream_mux(mux);
2056 int64_t QuickSyncEncoder::global_delay() const {
2057 return impl->global_delay();
2060 string QuickSyncEncoder::get_usable_va_display()
2062 // Reduce the amount of chatter while probing,
2063 // unless the user has specified otherwise.
2064 bool need_env_reset = false;
2065 if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
2066 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
2067 need_env_reset = true;
2070 // First try the default (ie., whatever $DISPLAY is set to).
2071 unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va_h264("", nullptr, nullptr);
2072 if (va_dpy != nullptr) {
2073 if (need_env_reset) {
2074 unsetenv("LIBVA_MESSAGING_LEVEL");
2079 fprintf(stderr, "No --va-display was given, and the X11 display did not expose a VA-API H.264 encoder.\n");
2081 // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
2083 int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
2085 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
2087 for (size_t i = 0; i < g.gl_pathc; ++i) {
2088 string path = g.gl_pathv[i];
2089 va_dpy = try_open_va_h264(path, nullptr, nullptr);
2090 if (va_dpy != nullptr) {
2091 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
2094 if (need_env_reset) {
2095 unsetenv("LIBVA_MESSAGING_LEVEL");
2102 fprintf(stderr, "No suitable VA-API H.264 encoders were found in /dev/dri; giving up.\n");
2103 fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
2104 fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
2105 fprintf(stderr, "to expose Quick Sync. Alternatively, you can use --record-x264-video\n");
2106 fprintf(stderr, "to use software instead of hardware H.264 encoding, at the expense\n");
2107 fprintf(stderr, "of increased CPU usage and possibly bit rate.\n");