1 #include "quicksync_encoder.h"
7 #include <movit/image_format.h>
8 #include <movit/resource_pool.h> // Must be above the Xlib includes.
9 #include <movit/util.h>
11 #include <EGL/eglplatform.h>
13 #include <epoxy/egl.h>
20 #include <va/va_drmcommon.h>
21 #include <va/va_enc_h264.h>
24 #include <condition_variable>
39 #include <drm_fourcc.h>
40 #include <libavcodec/packet.h>
41 #include <libavformat/avformat.h>
42 #include <libavformat/avio.h>
43 #include <libavutil/error.h>
47 #include "audio_encoder.h"
50 #include "print_latency.h"
51 #include "quicksync_encoder_impl.h"
52 #include "ref_counted_frame.h"
53 #include "shared/context.h"
54 #include "shared/disk_space_estimator.h"
55 #include "shared/ffmpeg_raii.h"
56 #include "shared/metrics.h"
57 #include "shared/mux.h"
58 #include "shared/ref_counted_gl_sync.h"
59 #include "shared/timebase.h"
60 #include "shared/va_display.h"
61 #include "v4l_output.h"
62 #include "x264_encoder.h"
64 using namespace movit;
66 using namespace std::chrono;
67 using namespace std::placeholders;
74 // These need to survive several QuickSyncEncoderImpl instances,
75 // so they are outside.
76 once_flag quick_sync_metrics_inited;
77 LatencyHistogram mixer_latency_histogram, qs_latency_histogram;
78 MuxMetrics current_file_mux_metrics, total_mux_metrics;
79 std::atomic<double> metric_current_file_start_time_seconds{0.0 / 0.0};
80 std::atomic<int64_t> metric_quick_sync_stalled_frames{0};
84 #define CHECK_VASTATUS(va_status, func) \
85 if (va_status != VA_STATUS_SUCCESS) { \
86 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
91 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
93 //#include "loadsurface.h"
95 #define NAL_REF_IDC_NONE 0
96 #define NAL_REF_IDC_LOW 1
97 #define NAL_REF_IDC_MEDIUM 2
98 #define NAL_REF_IDC_HIGH 3
100 #define NAL_NON_IDR 1
106 #define SLICE_TYPE_P 0
107 #define SLICE_TYPE_B 1
108 #define SLICE_TYPE_I 2
109 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
110 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
111 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
114 #define ENTROPY_MODE_CAVLC 0
115 #define ENTROPY_MODE_CABAC 1
117 #define PROFILE_IDC_BASELINE 66
118 #define PROFILE_IDC_MAIN 77
119 #define PROFILE_IDC_HIGH 100
121 #define BITSTREAM_ALLOCATE_STEPPING 4096
123 static constexpr unsigned int MaxFrameNum = (2<<16);
124 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
125 static constexpr unsigned int Log2MaxFrameNum = 16;
126 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
130 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
131 // but if we don't delete it here, we get leaks. The GStreamer implementation
133 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
135 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
136 CHECK_VASTATUS(va_status, "vaRenderPicture");
138 for (int i = 0; i < num_buffers; ++i) {
139 va_status = vaDestroyBuffer(dpy, buffers[i]);
140 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
145 va_swap32(unsigned int val)
147 unsigned char *pval = (unsigned char *)&val;
149 return ((pval[0] << 24) |
156 bitstream_start(bitstream *bs)
158 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
159 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
164 bitstream_end(bitstream *bs)
166 int pos = (bs->bit_offset >> 5);
167 int bit_offset = (bs->bit_offset & 0x1f);
168 int bit_left = 32 - bit_offset;
171 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
176 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
178 int pos = (bs->bit_offset >> 5);
179 int bit_offset = (bs->bit_offset & 0x1f);
180 int bit_left = 32 - bit_offset;
185 bs->bit_offset += size_in_bits;
187 if (bit_left > size_in_bits) {
188 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
190 size_in_bits -= bit_left;
191 if (bit_left >= 32) {
192 bs->buffer[pos] = (val >> size_in_bits);
194 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
196 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
198 if (pos + 1 == bs->max_size_in_dword) {
199 bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
200 bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
203 bs->buffer[pos + 1] = val;
208 bitstream_put_ue(bitstream *bs, unsigned int val)
210 int size_in_bits = 0;
218 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
219 bitstream_put_ui(bs, val, size_in_bits);
223 bitstream_put_se(bitstream *bs, int val)
225 unsigned int new_val;
230 new_val = 2 * val - 1;
232 bitstream_put_ue(bs, new_val);
236 bitstream_byte_aligning(bitstream *bs, int bit)
238 int bit_offset = (bs->bit_offset & 0x7);
239 int bit_left = 8 - bit_offset;
245 assert(bit == 0 || bit == 1);
248 new_val = (1 << bit_left) - 1;
252 bitstream_put_ui(bs, new_val, bit_left);
256 rbsp_trailing_bits(bitstream *bs)
258 bitstream_put_ui(bs, 1, 1);
259 bitstream_byte_aligning(bs, 0);
262 static void nal_start_code_prefix(bitstream *bs)
264 bitstream_put_ui(bs, 0x00000001, 32);
267 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
269 bitstream_put_ui(bs, 0, 1); /* forbidden_zero_bit: 0 */
270 bitstream_put_ui(bs, nal_ref_idc, 2);
271 bitstream_put_ui(bs, nal_unit_type, 5);
274 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
276 int profile_idc = PROFILE_IDC_BASELINE;
278 if (h264_profile == VAProfileH264High)
279 profile_idc = PROFILE_IDC_HIGH;
280 else if (h264_profile == VAProfileH264Main)
281 profile_idc = PROFILE_IDC_MAIN;
283 bitstream_put_ui(bs, profile_idc, 8); /* profile_idc */
284 bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1); /* constraint_set0_flag */
285 bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1); /* constraint_set1_flag */
286 bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1); /* constraint_set2_flag */
287 bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1); /* constraint_set3_flag */
288 bitstream_put_ui(bs, 0, 4); /* reserved_zero_4bits */
289 bitstream_put_ui(bs, seq_param.level_idc, 8); /* level_idc */
290 bitstream_put_ue(bs, seq_param.seq_parameter_set_id); /* seq_parameter_set_id */
292 if ( profile_idc == PROFILE_IDC_HIGH) {
293 bitstream_put_ue(bs, 1); /* chroma_format_idc = 1, 4:2:0 */
294 bitstream_put_ue(bs, 0); /* bit_depth_luma_minus8 */
295 bitstream_put_ue(bs, 0); /* bit_depth_chroma_minus8 */
296 bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
297 bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
300 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
301 bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type); /* pic_order_cnt_type */
303 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
304 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4); /* log2_max_pic_order_cnt_lsb_minus4 */
309 bitstream_put_ue(bs, seq_param.max_num_ref_frames); /* num_ref_frames */
310 bitstream_put_ui(bs, 0, 1); /* gaps_in_frame_num_value_allowed_flag */
312 bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1); /* pic_width_in_mbs_minus1 */
313 bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
314 bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1); /* frame_mbs_only_flag */
316 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
320 bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1); /* direct_8x8_inference_flag */
321 bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1); /* frame_cropping_flag */
323 if (seq_param.frame_cropping_flag) {
324 bitstream_put_ue(bs, seq_param.frame_crop_left_offset); /* frame_crop_left_offset */
325 bitstream_put_ue(bs, seq_param.frame_crop_right_offset); /* frame_crop_right_offset */
326 bitstream_put_ue(bs, seq_param.frame_crop_top_offset); /* frame_crop_top_offset */
327 bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset); /* frame_crop_bottom_offset */
330 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
332 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
334 // See H.264 annex E for the definition of this header.
335 bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
336 bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
337 bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
338 bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
340 bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
341 bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
342 bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
344 bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
345 bitstream_put_ui(bs, 13, 8); /* transfer_characteristics (13 = sRGB) */
346 if (ycbcr_coefficients == YCBCR_REC_709) {
347 bitstream_put_ui(bs, 1, 8); /* matrix_coefficients (1 = BT.709) */
349 assert(ycbcr_coefficients == YCBCR_REC_601);
350 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
354 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
355 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
357 bitstream_put_ui(bs, 1, 32); // FPS
358 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
359 bitstream_put_ui(bs, 1, 1);
361 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
364 bitstream_put_ue(bs, 0); /* cpb_cnt_minus1 */
365 bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
366 bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
368 bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
369 bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
370 bitstream_put_ui(bs, 1, 1); /* cbr_flag[0] */
372 bitstream_put_ui(bs, 23, 5); /* initial_cpb_removal_delay_length_minus1 */
373 bitstream_put_ui(bs, 23, 5); /* cpb_removal_delay_length_minus1 */
374 bitstream_put_ui(bs, 23, 5); /* dpb_output_delay_length_minus1 */
375 bitstream_put_ui(bs, 23, 5); /* time_offset_length */
377 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
378 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
380 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
381 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
384 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
388 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
390 bitstream_put_ue(bs, pic_param.pic_parameter_set_id); /* pic_parameter_set_id */
391 bitstream_put_ue(bs, pic_param.seq_parameter_set_id); /* seq_parameter_set_id */
393 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
395 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
397 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
399 bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1); /* num_ref_idx_l0_active_minus1 */
400 bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1); /* num_ref_idx_l1_active_minus1 1 */
402 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1); /* weighted_pred_flag: 0 */
403 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2); /* weighted_bipred_idc: 0 */
405 bitstream_put_se(bs, pic_param.pic_init_qp - 26); /* pic_init_qp_minus26 */
406 bitstream_put_se(bs, 0); /* pic_init_qs_minus26 */
407 bitstream_put_se(bs, 0); /* chroma_qp_index_offset */
409 bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
410 bitstream_put_ui(bs, 0, 1); /* constrained_intra_pred_flag */
411 bitstream_put_ui(bs, 0, 1); /* redundant_pic_cnt_present_flag */
414 bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1); /*transform_8x8_mode_flag */
415 bitstream_put_ui(bs, 0, 1); /* pic_scaling_matrix_present_flag */
416 bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset ); /*second_chroma_qp_index_offset */
418 rbsp_trailing_bits(bs);
421 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
423 int first_mb_in_slice = slice_param.macroblock_address;
425 bitstream_put_ue(bs, first_mb_in_slice); /* first_mb_in_slice: 0 */
426 bitstream_put_ue(bs, slice_param.slice_type); /* slice_type */
427 bitstream_put_ue(bs, slice_param.pic_parameter_set_id); /* pic_parameter_set_id: 0 */
428 bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
430 /* frame_mbs_only_flag == 1 */
431 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
436 if (pic_param.pic_fields.bits.idr_pic_flag)
437 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
439 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
440 bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
441 /* pic_order_present_flag == 0 */
447 /* redundant_pic_cnt_present_flag == 0 */
449 if (IS_P_SLICE(slice_param.slice_type)) {
450 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
452 if (slice_param.num_ref_idx_active_override_flag)
453 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
455 /* ref_pic_list_reordering */
456 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
457 } else if (IS_B_SLICE(slice_param.slice_type)) {
458 bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1); /* direct_spatial_mv_pred: 1 */
460 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
462 if (slice_param.num_ref_idx_active_override_flag) {
463 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
464 bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
467 /* ref_pic_list_reordering */
468 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
469 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l1: 0 */
472 if ((pic_param.pic_fields.bits.weighted_pred_flag &&
473 IS_P_SLICE(slice_param.slice_type)) ||
474 ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
475 IS_B_SLICE(slice_param.slice_type))) {
476 /* FIXME: fill weight/offset table */
480 /* dec_ref_pic_marking */
481 if (pic_param.pic_fields.bits.reference_pic_flag) { /* nal_ref_idc != 0 */
482 unsigned char no_output_of_prior_pics_flag = 0;
483 unsigned char long_term_reference_flag = 0;
484 unsigned char adaptive_ref_pic_marking_mode_flag = 0;
486 if (pic_param.pic_fields.bits.idr_pic_flag) {
487 bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1); /* no_output_of_prior_pics_flag: 0 */
488 bitstream_put_ui(bs, long_term_reference_flag, 1); /* long_term_reference_flag: 0 */
490 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
494 if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
495 !IS_I_SLICE(slice_param.slice_type))
496 bitstream_put_ue(bs, slice_param.cabac_init_idc); /* cabac_init_idc: 0 */
498 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
500 /* ignore for SP/SI */
502 if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
503 bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc); /* disable_deblocking_filter_idc: 0 */
505 if (slice_param.disable_deblocking_filter_idc != 1) {
506 bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2); /* slice_alpha_c0_offset_div2: 2 */
507 bitstream_put_se(bs, slice_param.slice_beta_offset_div2); /* slice_beta_offset_div2: 2 */
511 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
512 bitstream_byte_aligning(bs, 1);
516 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
520 bitstream_start(&bs);
521 nal_start_code_prefix(&bs);
522 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
526 *header_buffer = (unsigned char *)bs.buffer;
527 return bs.bit_offset;
531 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
535 bitstream_start(&bs);
536 nal_start_code_prefix(&bs);
537 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
538 sps_rbsp(ycbcr_coefficients, &bs);
541 *header_buffer = (unsigned char *)bs.buffer;
542 return bs.bit_offset;
545 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
548 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
549 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
551 bitstream_start(&bs);
552 nal_start_code_prefix(&bs);
554 if (IS_I_SLICE(slice_param.slice_type)) {
555 nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
556 } else if (IS_P_SLICE(slice_param.slice_type)) {
557 nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
559 assert(IS_B_SLICE(slice_param.slice_type));
560 nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
566 *header_buffer = (unsigned char *)bs.buffer;
567 return bs.bit_offset;
572 Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
573 1) period between Frame #X and Frame #N = #X - #N
574 2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
575 3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
576 4) intra_period and intra_idr_period take precedence over ip_period
577 5) if ip_period > 1, intra_period and intra_idr_period are not the strict periods
578 of I/IDR frames, see bellow examples
579 -------------------------------------------------------------------
580 intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
581 0 ignored 1 IDRPPPPPPP ... (No IDR/I any more)
582 0 ignored >=2 IDR(PBB)(PBB)... (No IDR/I any more)
583 1 0 ignored IDRIIIIIII... (No IDR any more)
584 1 1 ignored IDR IDR IDR IDR...
585 1 >=2 ignored IDRII IDRII IDR... (1/3/ignore)
586 >=2 0 1 IDRPPP IPPP I... (3/0/1)
587 >=2 0 >=2 IDR(PBB)(PBB)(IBB) (6/0/3)
588 (PBB)(IBB)(PBB)(IBB)...
589 >=2 >=2 1 IDRPPPPP IPPPPP IPPPPP (6/18/1)
590 IDRPPPPP IPPPPP IPPPPP...
591 >=2 >=2 >=2 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
592 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
593 {IDR(PBB)(PBB)(IBB)(PBB)} (6/12/3)
594 {IDR(PBB)(PBB)(IBB)(PBB)}...
595 {IDR(PBB)(PBB)} (6/6/3)
603 void encoding2display_order(
604 int encoding_order, int intra_period,
605 int intra_idr_period, int ip_period,
606 int *displaying_order,
609 int encoding_order_gop = 0;
611 if (intra_period == 1) { /* all are I/IDR frames */
612 *displaying_order = encoding_order;
613 if (intra_idr_period == 0)
614 *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
616 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
620 if (intra_period == 0)
621 intra_idr_period = 0;
623 if (ip_period == 1) {
624 // No B-frames, sequence is like IDR PPPPP IPPPPP.
625 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
626 *displaying_order = encoding_order;
628 if (encoding_order_gop == 0) { /* the first frame */
629 *frame_type = FRAME_IDR;
630 } else if (intra_period != 0 && /* have I frames */
631 encoding_order_gop >= 2 &&
632 (encoding_order_gop % intra_period == 0)) {
633 *frame_type = FRAME_I;
635 *frame_type = FRAME_P;
640 // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
641 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
643 if (encoding_order_gop == 0) { /* the first frame */
644 *frame_type = FRAME_IDR;
645 *displaying_order = encoding_order;
646 } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
647 *frame_type = FRAME_B;
648 *displaying_order = encoding_order - 1;
649 } else if (intra_period != 0 && /* have I frames */
650 encoding_order_gop >= 2 &&
651 ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
652 *frame_type = FRAME_I;
653 *displaying_order = encoding_order + ip_period - 1;
655 *frame_type = FRAME_P;
656 *displaying_order = encoding_order + ip_period - 1;
660 // General pts/dts strategy:
662 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
663 // bit tricky. This strategy roughly matches what x264 seems to do: We take in
664 // the pts as the frames are encoded, and reuse that as dts in the same order,
667 // If we don't have B-frames (only I and P), this means pts == dts always.
668 // This is the simple case. Now consider the case with a single B-frame:
671 // pts: 30 40 50 60 70
673 // Since we always inherently encode P-frames before B-frames, this means that
674 // we see them in this order, which we can _almost_ use for dts:
676 // dts: 30 50 40 70 60
678 // the only problem here is that for the B-frames, pts < dts. We solve this by
679 // priming the queue at the very start with some made-up dts:
682 // pts: 30 40 50 60 70
683 // dts: xx 30 50 40 70 60
685 // Now we have all the desirable properties: pts >= dts, successive dts delta
686 // is never larger than the decoder can figure out (assuming, of course,
687 // the pts has that property), and there's minimal lag between pts and dts.
688 // For the made-up dts, we assume 1/60 sec per frame, which should generally
689 // be reasonable. dts can go negative, but this is corrected using global_delay()
690 // by delaying both pts and dts (although we probably don't need to).
692 // If there's more than one B-frame possible, we simply insert more of them
693 // (here shown with some irregular spacing, assuming B-frames don't depend
694 // on each other and simply go back-to-front):
697 // pts: 30 40 55 60 65 66 67 68 80
698 // dts: xx yy zz 30 65 60 55 40 80 68 67 66
701 DTSReorderer(int num_b_frames) : num_b_frames(num_b_frames) {}
703 void push_pts(int64_t pts)
705 if (buf.empty() && num_b_frames > 0) { // First frame.
706 int64_t base_dts = pts - num_b_frames * (TIMEBASE / TYPICAL_FPS);
707 for (int i = 0; i < num_b_frames; ++i) {
708 buf.push(base_dts + i * (TIMEBASE / TYPICAL_FPS));
716 assert(!buf.empty());
717 int64_t dts = buf.front();
723 const int num_b_frames;
727 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
729 if (global_flags.x264_video_to_disk) {
730 // Quick Sync is entirely disabled.
731 use_zerocopy = false;
732 } else if (global_flags.x264_video_to_http) {
733 use_zerocopy = false;
734 } else if (global_flags.av1_video_to_http) {
735 use_zerocopy = false;
736 } else if (!global_flags.v4l_output_device.empty()) {
737 use_zerocopy = false;
741 global_flags.use_zerocopy = use_zerocopy;
744 static unique_ptr<VADisplayWithCleanup> try_open_va_h264(const string &va_display, VAProfile *h264_profile, string *error)
746 return try_open_va(va_display, { VAProfileH264High, VAProfileH264Main, VAProfileH264ConstrainedBaseline },
747 VAEntrypointEncSlice, /*desired_configs=*/{}, h264_profile, error);
750 int QuickSyncEncoderImpl::init_va(const string &va_display)
753 va_dpy = try_open_va_h264(va_display, &h264_profile, &error);
754 if (va_dpy == nullptr) {
755 fprintf(stderr, "error: %s\n", error.c_str());
758 if (!va_dpy->can_use_zerocopy) {
759 use_zerocopy = false;
762 switch (h264_profile) {
763 case VAProfileH264ConstrainedBaseline:
764 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
768 case VAProfileH264Main:
769 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
772 case VAProfileH264High:
773 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
776 h264_profile = VAProfileH264ConstrainedBaseline;
778 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
782 VAConfigAttrib attrib[VAConfigAttribTypeMax];
784 /* find out the format for the render target, and rate control mode */
785 for (unsigned i = 0; i < VAConfigAttribTypeMax; i++)
786 attrib[i].type = (VAConfigAttribType)i;
788 VAStatus va_status = vaGetConfigAttributes(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
789 &attrib[0], VAConfigAttribTypeMax);
790 CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
791 /* check the interested configattrib */
792 if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
793 printf("Not find desired YUV420 RT format\n");
796 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
797 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
801 if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
802 if (!(attrib[VAConfigAttribRateControl].value & VA_RC_CQP)) {
803 fprintf(stderr, "ERROR: VA-API encoder does not support CQP mode.\n");
807 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
808 config_attrib[config_attrib_num].value = VA_RC_CQP;
813 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
814 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
816 h264_packedheader = 1;
817 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
818 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
820 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
821 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
824 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
825 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
828 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
829 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
832 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
833 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
836 enc_packed_header_idx = config_attrib_num;
840 if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
841 config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
842 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
846 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
847 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
853 int QuickSyncEncoderImpl::setup_encode()
855 if (!global_flags.x264_video_to_disk) {
857 VASurfaceID *tmp_surfaceid;
859 VASurfaceID src_surface[SURFACE_NUM];
860 VASurfaceID ref_surface[SURFACE_NUM];
862 va_status = vaCreateConfig(va_dpy->va_dpy, h264_profile, VAEntrypointEncSlice,
863 &config_attrib[0], config_attrib_num, &config_id);
864 CHECK_VASTATUS(va_status, "vaCreateConfig");
866 /* create source surfaces */
867 va_status = vaCreateSurfaces(va_dpy->va_dpy,
868 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
869 &src_surface[0], SURFACE_NUM,
871 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
873 /* create reference surfaces */
874 va_status = vaCreateSurfaces(va_dpy->va_dpy,
875 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
876 &ref_surface[0], SURFACE_NUM,
878 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
880 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
881 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
882 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
884 for (int i = 0; i < SURFACE_NUM; i++) {
885 gl_surfaces[i].src_surface = src_surface[i];
886 gl_surfaces[i].ref_surface = ref_surface[i];
889 /* Create a context for this encode pipe */
890 va_status = vaCreateContext(va_dpy->va_dpy, config_id,
891 frame_width_mbaligned, frame_height_mbaligned,
893 tmp_surfaceid, 2 * SURFACE_NUM,
895 CHECK_VASTATUS(va_status, "vaCreateContext");
898 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
900 for (int i = 0; i < SURFACE_NUM; i++) {
901 /* create coded buffer once for all
902 * other VA buffers which won't be used again after vaRenderPicture.
903 * so APP can always vaCreateBuffer for every frame
904 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
905 * so VA won't maintain the coded buffer
907 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncCodedBufferType,
908 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
909 CHECK_VASTATUS(va_status, "vaCreateBuffer");
913 /* create OpenGL objects */
914 for (int i = 0; i < SURFACE_NUM; i++) {
916 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
917 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
919 size_t bytes_per_pixel = (global_flags.bit_depth > 8) ? 2 : 1;
921 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
922 // buffers, due to potentially differing pitch.
923 glGenBuffers(1, &gl_surfaces[i].pbo);
924 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
925 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);
926 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);
927 gl_surfaces[i].y_offset = 0;
928 gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
929 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
930 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
931 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
938 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
941 template<class T, class C>
942 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
944 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
945 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
946 sort(middle, end, less_than);
949 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
951 if (frame_type == FRAME_B)
954 pic_param.CurrPic.frame_idx = current_ref_frame_num;
956 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
957 lock_guard<mutex> lock(storage_task_queue_mutex);
959 // Insert the new frame at the start of the reference queue.
960 reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
962 if (reference_frames.size() > num_ref_frames)
964 // The back frame frame is no longer in use as a reference.
965 int display_frame_num = reference_frames.back().display_number;
966 assert(surface_for_frame.count(display_frame_num));
967 release_gl_surface(display_frame_num);
968 reference_frames.pop_back();
971 // Mark this frame in use as a reference.
972 assert(surface_for_frame.count(current_display_frame));
973 ++surface_for_frame[current_display_frame]->refcount;
975 current_ref_frame_num++;
976 if (current_ref_frame_num > MaxFrameNum)
977 current_ref_frame_num = 0;
981 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
983 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
984 return a.frame_idx > b.frame_idx;
987 for (size_t i = 0; i < reference_frames.size(); ++i) {
988 RefPicList0_P[i] = reference_frames[i].pic;
990 sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
993 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
995 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
996 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
998 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
999 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1002 for (size_t i = 0; i < reference_frames.size(); ++i) {
1003 RefPicList0_B[i] = reference_frames[i].pic;
1004 RefPicList1_B[i] = reference_frames[i].pic;
1006 sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1007 sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1011 int QuickSyncEncoderImpl::render_sequence()
1013 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1015 VAEncMiscParameterBuffer *misc_param;
1016 VAEncMiscParameterRateControl *misc_rate_ctrl;
1018 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1019 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1020 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1021 seq_param.bits_per_second = frame_bitrate;
1023 seq_param.intra_period = intra_period;
1024 seq_param.intra_idr_period = intra_idr_period;
1025 seq_param.ip_period = ip_period;
1027 seq_param.max_num_ref_frames = num_ref_frames;
1028 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1029 seq_param.time_scale = TIMEBASE * 2;
1030 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1031 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1032 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1033 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1034 seq_param.seq_fields.bits.chroma_format_idc = 1;
1035 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1037 if (frame_width != frame_width_mbaligned ||
1038 frame_height != frame_height_mbaligned) {
1039 seq_param.frame_cropping_flag = 1;
1040 seq_param.frame_crop_left_offset = 0;
1041 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1042 seq_param.frame_crop_top_offset = 0;
1043 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1046 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1047 VAEncSequenceParameterBufferType,
1048 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1049 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1051 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id,
1052 VAEncMiscParameterBufferType,
1053 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1054 1, NULL, &rc_param_buf);
1055 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1057 vaMapBuffer(va_dpy->va_dpy, rc_param_buf, (void **)&misc_param);
1058 misc_param->type = VAEncMiscParameterTypeRateControl;
1059 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1060 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1061 misc_rate_ctrl->bits_per_second = frame_bitrate;
1062 misc_rate_ctrl->target_percentage = 66;
1063 misc_rate_ctrl->window_size = 1000;
1064 misc_rate_ctrl->initial_qp = initial_qp;
1065 misc_rate_ctrl->min_qp = minimal_qp;
1066 misc_rate_ctrl->basic_unit_size = 0;
1067 vaUnmapBuffer(va_dpy->va_dpy, rc_param_buf);
1069 render_id[0] = seq_param_buf;
1070 render_id[1] = rc_param_buf;
1072 render_picture_and_delete(va_dpy->va_dpy, context_id, &render_id[0], 2);
1077 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1079 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1080 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1081 int PicOrderCntMsb, TopFieldOrderCnt;
1083 if (frame_type == FRAME_IDR)
1084 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1086 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1087 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1090 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1091 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1092 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1093 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1094 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1095 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1097 PicOrderCntMsb = prevPicOrderCntMsb;
1099 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1101 if (frame_type != FRAME_B) {
1102 PicOrderCntMsb_ref = PicOrderCntMsb;
1103 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1106 return TopFieldOrderCnt;
1109 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1111 VABufferID pic_param_buf;
1115 pic_param.CurrPic.picture_id = surf->ref_surface;
1116 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1117 pic_param.CurrPic.flags = 0;
1118 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1119 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1120 CurrentCurrPic = pic_param.CurrPic;
1122 for (i = 0; i < reference_frames.size(); i++) {
1123 pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1125 for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1126 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1127 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1130 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1131 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1132 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1133 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1134 pic_param.frame_num = current_ref_frame_num; // FIXME: is this correct?
1135 pic_param.coded_buf = surf->coded_buf;
1136 pic_param.last_picture = false; // FIXME
1137 pic_param.pic_init_qp = initial_qp;
1139 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncPictureParameterBufferType,
1140 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1141 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1143 render_picture_and_delete(va_dpy->va_dpy, context_id, &pic_param_buf, 1);
1148 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1150 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1151 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1152 unsigned int length_in_bits;
1153 unsigned char *packedseq_buffer = NULL;
1156 length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1158 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1160 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1161 packedheader_param_buffer.has_emulation_bytes = 0;
1162 va_status = vaCreateBuffer(va_dpy->va_dpy,
1164 VAEncPackedHeaderParameterBufferType,
1165 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1166 &packedseq_para_bufid);
1167 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1169 va_status = vaCreateBuffer(va_dpy->va_dpy,
1171 VAEncPackedHeaderDataBufferType,
1172 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1173 &packedseq_data_bufid);
1174 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1176 render_id[0] = packedseq_para_bufid;
1177 render_id[1] = packedseq_data_bufid;
1178 render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1180 free(packedseq_buffer);
1186 int QuickSyncEncoderImpl::render_packedpicture()
1188 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1189 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1190 unsigned int length_in_bits;
1191 unsigned char *packedpic_buffer = NULL;
1194 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1195 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1196 packedheader_param_buffer.bit_length = length_in_bits;
1197 packedheader_param_buffer.has_emulation_bytes = 0;
1199 va_status = vaCreateBuffer(va_dpy->va_dpy,
1201 VAEncPackedHeaderParameterBufferType,
1202 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1203 &packedpic_para_bufid);
1204 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1206 va_status = vaCreateBuffer(va_dpy->va_dpy,
1208 VAEncPackedHeaderDataBufferType,
1209 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1210 &packedpic_data_bufid);
1211 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1213 render_id[0] = packedpic_para_bufid;
1214 render_id[1] = packedpic_data_bufid;
1215 render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1217 free(packedpic_buffer);
1222 void QuickSyncEncoderImpl::render_packedslice()
1224 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1225 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1226 unsigned int length_in_bits;
1227 unsigned char *packedslice_buffer = NULL;
1230 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1231 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1232 packedheader_param_buffer.bit_length = length_in_bits;
1233 packedheader_param_buffer.has_emulation_bytes = 0;
1235 va_status = vaCreateBuffer(va_dpy->va_dpy,
1237 VAEncPackedHeaderParameterBufferType,
1238 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1239 &packedslice_para_bufid);
1240 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1242 va_status = vaCreateBuffer(va_dpy->va_dpy,
1244 VAEncPackedHeaderDataBufferType,
1245 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1246 &packedslice_data_bufid);
1247 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1249 render_id[0] = packedslice_para_bufid;
1250 render_id[1] = packedslice_data_bufid;
1251 render_picture_and_delete(va_dpy->va_dpy, context_id, render_id, 2);
1253 free(packedslice_buffer);
1256 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1258 VABufferID slice_param_buf;
1262 /* one frame, one slice */
1263 slice_param.macroblock_address = 0;
1264 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1265 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1266 if (frame_type == FRAME_IDR) {
1267 if (encoding_frame_num != 0)
1268 ++slice_param.idr_pic_id;
1269 } else if (frame_type == FRAME_P) {
1270 VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1271 update_RefPicList_P(RefPicList0_P);
1273 int refpiclist0_max = h264_maxref & 0xffff;
1274 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1276 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1277 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1278 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1280 } else if (frame_type == FRAME_B) {
1281 VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1282 update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1284 int refpiclist0_max = h264_maxref & 0xffff;
1285 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1287 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1288 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1289 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1290 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1293 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1294 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1295 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1296 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1300 slice_param.slice_alpha_c0_offset_div2 = 0;
1301 slice_param.slice_beta_offset_div2 = 0;
1302 slice_param.direct_spatial_mv_pred_flag = 1;
1303 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1306 if (h264_packedheader &&
1307 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1308 render_packedslice();
1310 va_status = vaCreateBuffer(va_dpy->va_dpy, context_id, VAEncSliceParameterBufferType,
1311 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1312 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1314 render_picture_and_delete(va_dpy->va_dpy, context_id, &slice_param_buf, 1);
1321 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1323 VACodedBufferSegment *buf_list = NULL;
1328 va_status = vaMapBuffer(va_dpy->va_dpy, surf->coded_buf, (void **)(&buf_list));
1329 CHECK_VASTATUS(va_status, "vaMapBuffer");
1330 while (buf_list != NULL) {
1331 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1332 buf_list = (VACodedBufferSegment *) buf_list->next;
1334 vaUnmapBuffer(va_dpy->va_dpy, surf->coded_buf);
1336 static int frameno = 0;
1337 print_latency("Current Quick Sync latency (video inputs → disk mux):",
1338 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1343 memset(&pkt, 0, sizeof(pkt));
1345 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1346 pkt.size = data.size();
1347 pkt.stream_index = 0;
1348 if (task.frame_type == FRAME_IDR) {
1349 pkt.flags = AV_PKT_FLAG_KEY;
1353 pkt.duration = task.duration;
1355 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1357 if (!global_flags.x264_video_to_http &&
1358 !global_flags.av1_video_to_http) {
1359 http_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1360 if (srt_mux != nullptr) {
1361 srt_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1368 // this is weird. but it seems to put a new frame onto the queue
1369 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1371 assert(task.pts >= task.dts);
1372 lock_guard<mutex> lock(storage_task_queue_mutex);
1373 storage_task_queue.push(move(task));
1374 storage_task_queue_changed.notify_all();
1377 void QuickSyncEncoderImpl::storage_task_thread()
1379 pthread_setname_np(pthread_self(), "QS_Storage");
1381 storage_task current;
1384 // wait until there's an encoded frame
1385 unique_lock<mutex> lock(storage_task_queue_mutex);
1386 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1387 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1388 current = move(storage_task_queue.front());
1389 storage_task_queue.pop();
1390 surf = surface_for_frame[current.display_order];
1391 assert(surf != nullptr);
1396 size_t display_order = current.display_order;
1397 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1399 // waits for data, then saves it to disk.
1400 va_status = vaSyncSurface(va_dpy->va_dpy, surf->src_surface);
1401 CHECK_VASTATUS(va_status, "vaSyncSurface");
1402 save_codeddata(surf, move(current));
1404 // Unlock the frame, and all its references.
1406 lock_guard<mutex> lock(storage_task_queue_mutex);
1407 release_gl_surface(display_order);
1409 for (size_t frame_num : ref_display_frame_numbers) {
1410 release_gl_surface(frame_num);
1416 void QuickSyncEncoderImpl::release_encode()
1418 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1419 VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, gl_surfaces[i].coded_buf);
1420 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
1421 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].src_surface, 1);
1422 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1423 va_status = vaDestroySurfaces(va_dpy->va_dpy, &gl_surfaces[i].ref_surface, 1);
1424 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
1427 VAStatus va_status = vaDestroyContext(va_dpy->va_dpy, context_id);
1428 CHECK_VASTATUS(va_status, "vaDestroyContext");
1429 va_status = vaDestroyConfig(va_dpy->va_dpy, config_id);
1430 CHECK_VASTATUS(va_status, "vaDestroyConfig");
1433 void QuickSyncEncoderImpl::release_gl_resources()
1435 assert(is_shutdown);
1436 if (has_released_gl_resources) {
1440 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1442 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1443 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1445 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1446 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1447 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1448 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1452 has_released_gl_resources = true;
1455 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, const AVOutputFormat *oformat, VideoCodecInterface *http_encoder, VideoCodecInterface *disk_encoder, DiskSpaceEstimator *disk_space_estimator)
1456 : current_storage_frame(0), resource_pool(resource_pool), surface(surface), http_encoder(http_encoder), disk_encoder(disk_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator)
1458 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1459 open_output_file(filename);
1460 file_audio_encoder->add_mux(file_mux.get());
1462 frame_width_mbaligned = (frame_width + 15) & (~15);
1463 frame_height_mbaligned = (frame_height + 15) & (~15);
1467 if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1468 assert(http_encoder != nullptr);
1469 assert(disk_encoder != nullptr);
1470 } else if (global_flags.av1_video_to_http) {
1471 assert(http_encoder != nullptr);
1473 assert(http_encoder == nullptr);
1474 assert(disk_encoder == nullptr);
1477 enable_zerocopy_if_possible();
1478 if (!global_flags.x264_video_to_disk) {
1479 init_va(va_display);
1483 if (!global_flags.x264_video_to_disk) {
1484 memset(&seq_param, 0, sizeof(seq_param));
1485 memset(&pic_param, 0, sizeof(pic_param));
1486 memset(&slice_param, 0, sizeof(slice_param));
1489 if (!global_flags.v4l_output_device.empty()) {
1490 v4l_output.reset(new V4LOutput(global_flags.v4l_output_device.c_str(), width, height));
1493 call_once(quick_sync_metrics_inited, [](){
1494 mixer_latency_histogram.init("mixer");
1495 qs_latency_histogram.init("quick_sync");
1496 current_file_mux_metrics.init({{ "destination", "current_file" }});
1497 total_mux_metrics.init({{ "destination", "files_total" }});
1498 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1499 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1502 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1504 encode_thread = thread([this]{
1505 QOpenGLContext *context = create_context(this->surface);
1506 eglBindAPI(EGL_OPENGL_API);
1507 if (!make_current(context, this->surface)) {
1508 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1512 encode_thread_func();
1513 delete_context(context);
1517 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1520 release_gl_resources();
1523 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1525 for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1526 if (gl_surfaces[i].refcount == 0) {
1527 ++gl_surfaces[i].refcount;
1528 return &gl_surfaces[i];
1534 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1536 assert(surface_for_frame.count(display_frame_num));
1537 QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1538 if (--surf->refcount == 0) {
1539 assert(surface_for_frame.count(display_frame_num));
1540 surface_for_frame.erase(display_frame_num);
1541 storage_task_queue_changed.notify_all();
1545 bool QuickSyncEncoderImpl::is_zerocopy() const
1547 return use_zerocopy;
1550 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1552 assert(!is_shutdown);
1553 GLSurface *surf = nullptr;
1555 // Wait until this frame slot is done encoding.
1556 unique_lock<mutex> lock(storage_task_queue_mutex);
1557 surf = allocate_gl_surface();
1558 if (surf == nullptr) {
1559 fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1560 current_storage_frame);
1561 ++metric_quick_sync_stalled_frames;
1562 storage_task_queue_changed.wait(lock, [this, &surf]{
1563 if (storage_thread_should_quit)
1565 surf = allocate_gl_surface();
1566 return surf != nullptr;
1569 if (storage_thread_should_quit) return false;
1570 assert(surf != nullptr);
1571 surface_for_frame[current_storage_frame] = surf;
1575 *y_tex = surf->y_tex;
1576 *cbcr_tex = surf->cbcr_tex;
1578 surf->y_tex = *y_tex;
1579 surf->cbcr_tex = *cbcr_tex;
1582 if (!global_flags.x264_video_to_disk) {
1583 VAStatus va_status = vaDeriveImage(va_dpy->va_dpy, surf->src_surface, &surf->surface_image);
1584 CHECK_VASTATUS(va_status, "vaDeriveImage");
1587 VABufferInfo buf_info;
1588 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1589 va_status = vaAcquireBufferHandle(va_dpy->va_dpy, surf->surface_image.buf, &buf_info);
1590 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1593 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1594 EGLint y_attribs[] = {
1595 EGL_WIDTH, frame_width,
1596 EGL_HEIGHT, frame_height,
1597 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1598 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1599 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1600 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1604 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1605 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1607 // Associate Y image to a texture.
1608 glBindTexture(GL_TEXTURE_2D, *y_tex);
1609 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1611 // Create CbCr image.
1612 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1613 EGLint cbcr_attribs[] = {
1614 EGL_WIDTH, frame_width / 2,
1615 EGL_HEIGHT, frame_height / 2,
1616 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1617 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1618 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1619 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1623 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1624 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1626 // Associate CbCr image to a texture.
1627 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1628 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1632 current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1637 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1639 lock_guard<mutex> lock(file_audio_encoder_mutex);
1640 assert(!is_shutdown);
1641 file_audio_encoder->encode_audio(audio, pts + global_delay());
1644 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1646 assert(!is_shutdown);
1648 if (!use_zerocopy) {
1649 GLenum type = global_flags.bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1652 lock_guard<mutex> lock(storage_task_queue_mutex);
1653 surf = surface_for_frame[current_storage_frame];
1654 assert(surf != nullptr);
1657 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1660 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1663 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1665 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1668 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1670 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1673 // We don't own these; the caller does.
1674 surf->y_tex = surf->cbcr_tex = 0;
1676 glBindTexture(GL_TEXTURE_2D, 0);
1678 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1681 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1685 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1687 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1691 lock_guard<mutex> lock(frame_queue_mutex);
1692 current_video_frame.fence = fence;
1693 pending_video_frames.push(move(current_video_frame));
1694 ++current_storage_frame;
1696 frame_queue_nonempty.notify_all();
1700 void QuickSyncEncoderImpl::shutdown()
1707 lock_guard<mutex> lock(frame_queue_mutex);
1708 encode_thread_should_quit = true;
1709 frame_queue_nonempty.notify_all();
1711 encode_thread.join();
1713 lock_guard<mutex> lock(storage_task_queue_mutex);
1714 storage_thread_should_quit = true;
1715 frame_queue_nonempty.notify_all();
1716 storage_task_queue_changed.notify_all();
1718 storage_thread.join();
1720 // Encode any leftover audio in the queues, and also any delayed frames.
1722 lock_guard<mutex> lock(file_audio_encoder_mutex);
1723 file_audio_encoder->encode_last_audio();
1726 if (!global_flags.x264_video_to_disk) {
1733 void QuickSyncEncoderImpl::close_file()
1736 metric_current_file_start_time_seconds = 0.0 / 0.0;
1739 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1741 AVFormatContext *avctx = avformat_alloc_context();
1742 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1743 avctx->url = strdup(filename.c_str());
1745 string url = "file:" + filename;
1746 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1748 char tmp[AV_ERROR_MAX_STRING_SIZE];
1749 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1753 string video_extradata; // FIXME: See other comment about global headers.
1754 if (global_flags.x264_video_to_disk) {
1755 video_extradata = disk_encoder->get_global_headers();
1758 current_file_mux_metrics.reset();
1761 lock_guard<mutex> lock(file_audio_encoder_mutex);
1762 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1763 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,
1764 std::bind(&DiskSpaceEstimator::report_append, disk_space_estimator, filename, _1),
1765 Mux::WRITE_BACKGROUND,
1766 { ¤t_file_mux_metrics, &total_mux_metrics }));
1768 metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1770 if (global_flags.x264_video_to_disk) {
1771 disk_encoder->add_mux(file_mux.get());
1775 void QuickSyncEncoderImpl::encode_thread_func()
1777 pthread_setname_np(pthread_self(), "QS_Encode");
1779 DTSReorderer dts_reorder_buf(ip_period - 1);
1781 int64_t last_dts = -1;
1782 int gop_start_display_frame_num = 0;
1783 for (int display_frame_num = 0; ; ++display_frame_num) {
1784 // Wait for the frame to be in the queue. Note that this only means
1785 // we started rendering it.
1788 unique_lock<mutex> lock(frame_queue_mutex);
1789 frame_queue_nonempty.wait(lock, [this]{
1790 return encode_thread_should_quit || !pending_video_frames.empty();
1792 if (encode_thread_should_quit && pending_video_frames.empty()) {
1793 // We may have queued frames left in the reorder buffer
1794 // that were supposed to be B-frames, but have no P-frame
1795 // to be encoded against. If so, encode them all as
1796 // P-frames instead. Note that this happens under the mutex,
1797 // but nobody else uses it at this point, since we're shutting down,
1798 // so there's no contention.
1799 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1802 frame = move(pending_video_frames.front());
1803 pending_video_frames.pop();
1807 dts_reorder_buf.push_pts(frame.pts);
1809 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1810 // Note that this implicitly waits for the frame to be done rendering.
1811 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1813 if (global_flags.x264_video_to_disk) {
1814 lock_guard<mutex> lock(storage_task_queue_mutex);
1815 release_gl_surface(display_frame_num);
1819 reorder_buffer[display_frame_num] = move(frame);
1821 // Now encode as many QuickSync frames as we can using the frames we have available.
1822 // (It could be zero, or it could be multiple.) FIXME: make a function.
1824 int frame_type, quicksync_display_frame_num;
1825 encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1826 &quicksync_display_frame_num, &frame_type);
1827 if (!reorder_buffer.count(quicksync_display_frame_num)) {
1830 frame = move(reorder_buffer[quicksync_display_frame_num]);
1831 reorder_buffer.erase(quicksync_display_frame_num);
1833 if (frame_type == FRAME_IDR) {
1834 // Release any reference frames from the previous GOP.
1836 lock_guard<mutex> lock(storage_task_queue_mutex);
1837 for (const ReferenceFrame &frame : reference_frames) {
1838 release_gl_surface(frame.display_number);
1841 reference_frames.clear();
1842 current_ref_frame_num = 0;
1843 gop_start_display_frame_num = quicksync_display_frame_num;
1846 const int64_t dts = dts_reorder_buf.pop_dts();
1849 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);
1850 ++quicksync_encoding_frame_num;
1855 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1857 if (reorder_buffer.empty()) {
1861 for (auto &pending_frame : reorder_buffer) {
1862 int display_frame_num = pending_frame.first;
1863 assert(display_frame_num > 0);
1864 PendingFrame frame = move(pending_frame.second);
1865 int64_t dts = last_dts + (TIMEBASE / TYPICAL_FPS);
1866 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1871 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1873 if (src_width == dst_pitch) {
1874 memcpy(dst, src, src_width * height);
1876 for (size_t y = 0; y < height; ++y) {
1877 const uint8_t *sptr = src + y * src_width;
1878 uint8_t *dptr = dst + y * dst_pitch;
1879 memcpy(dptr, sptr, src_width);
1884 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1886 // Wait for the GPU to be done with the frame.
1889 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1891 if (sync_status == GL_TIMEOUT_EXPIRED) {
1892 // NVIDIA likes to busy-wait; yield instead.
1893 this_thread::sleep_for(milliseconds(1));
1895 } while (sync_status == GL_TIMEOUT_EXPIRED);
1896 assert(sync_status != GL_WAIT_FAILED);
1898 ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1899 static int frameno = 0;
1900 print_latency("Current mixer latency (video inputs → ready for encode):",
1901 received_ts, false, &frameno, &mixer_latency_histogram);
1903 // Release back any input frames we needed to render this frame.
1904 frame.input_frames.clear();
1908 lock_guard<mutex> lock(storage_task_queue_mutex);
1909 surf = surface_for_frame[display_frame_num];
1910 assert(surf != nullptr);
1912 uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1913 if (http_encoder != nullptr) {
1914 http_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1915 } if (disk_encoder != nullptr && disk_encoder != http_encoder) {
1916 disk_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1919 if (v4l_output != nullptr) {
1920 v4l_output->send_frame(data);
1924 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1925 int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1927 const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1931 lock_guard<mutex> lock(storage_task_queue_mutex);
1932 surf = surface_for_frame[display_frame_num];
1933 assert(surf != nullptr);
1938 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1939 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1940 va_status = vaReleaseBufferHandle(va_dpy->va_dpy, surf->surface_image.buf);
1941 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1943 // Upload the frame to VA-API.
1944 unsigned char *surface_p = nullptr;
1945 vaMapBuffer(va_dpy->va_dpy, surf->surface_image.buf, (void **)&surface_p);
1947 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1948 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1950 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1951 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1953 va_status = vaUnmapBuffer(va_dpy->va_dpy, surf->surface_image.buf);
1954 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1957 va_status = vaDestroyImage(va_dpy->va_dpy, surf->surface_image.image_id);
1958 CHECK_VASTATUS(va_status, "vaDestroyImage");
1960 // Schedule the frame for encoding.
1961 VASurfaceID va_surface = surf->src_surface;
1962 va_status = vaBeginPicture(va_dpy->va_dpy, context_id, va_surface);
1963 CHECK_VASTATUS(va_status, "vaBeginPicture");
1965 if (frame_type == FRAME_IDR) {
1966 // FIXME: If the mux wants global headers, we should not put the
1967 // SPS/PPS before each IDR frame, but rather put it into the
1968 // codec extradata (formatted differently?).
1970 // NOTE: If we change ycbcr_coefficients, it will not take effect
1971 // before the next IDR frame. This is acceptable, as it should only
1972 // happen on a mode change, which is rare.
1974 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1975 if (h264_packedheader) {
1976 render_packedsequence(ycbcr_coefficients);
1977 render_packedpicture();
1980 //render_sequence();
1981 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
1983 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1985 va_status = vaEndPicture(va_dpy->va_dpy, context_id);
1986 CHECK_VASTATUS(va_status, "vaEndPicture");
1988 update_ReferenceFrames(display_frame_num, frame_type);
1990 vector<size_t> ref_display_frame_numbers;
1992 // Lock the references for this frame; otherwise, they could be
1993 // rendered to before this frame is done encoding.
1995 lock_guard<mutex> lock(storage_task_queue_mutex);
1996 for (const ReferenceFrame &frame : reference_frames) {
1997 assert(surface_for_frame.count(frame.display_number));
1998 ++surface_for_frame[frame.display_number]->refcount;
1999 ref_display_frame_numbers.push_back(frame.display_number);
2003 // so now the data is done encoding (well, async job kicked off)...
2004 // we send that to the storage thread
2006 tmp.display_order = display_frame_num;
2007 tmp.frame_type = frame_type;
2010 tmp.duration = duration;
2011 tmp.ycbcr_coefficients = ycbcr_coefficients;
2012 tmp.received_ts = received_ts;
2013 tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2014 storage_task_enqueue(move(tmp));
2018 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, const AVOutputFormat *oformat, VideoCodecInterface *http_encoder, VideoCodecInterface *disk_encoder, DiskSpaceEstimator *disk_space_estimator)
2019 : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, http_encoder, disk_encoder, disk_space_estimator)) {}
2021 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2022 QuickSyncEncoder::~QuickSyncEncoder() {}
2024 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2026 impl->add_audio(pts, audio);
2029 bool QuickSyncEncoder::is_zerocopy() const
2031 return impl->is_zerocopy();
2034 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2036 return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2039 RefCountedGLsync QuickSyncEncoder::end_frame()
2041 return impl->end_frame();
2044 void QuickSyncEncoder::shutdown()
2049 void QuickSyncEncoder::close_file()
2054 void QuickSyncEncoder::set_http_mux(Mux *mux)
2056 impl->set_http_mux(mux);
2059 void QuickSyncEncoder::set_srt_mux(Mux *mux)
2061 impl->set_srt_mux(mux);
2064 int64_t QuickSyncEncoder::global_delay() const {
2065 return impl->global_delay();
2068 string QuickSyncEncoder::get_usable_va_display()
2070 // Reduce the amount of chatter while probing,
2071 // unless the user has specified otherwise.
2072 bool need_env_reset = false;
2073 if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
2074 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
2075 need_env_reset = true;
2078 // First try the default (ie., whatever $DISPLAY is set to).
2079 unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va_h264("", nullptr, nullptr);
2080 if (va_dpy != nullptr) {
2081 if (need_env_reset) {
2082 unsetenv("LIBVA_MESSAGING_LEVEL");
2087 fprintf(stderr, "No --va-display was given, and the X11 display did not expose a VA-API H.264 encoder.\n");
2089 // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
2091 int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
2093 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
2095 for (size_t i = 0; i < g.gl_pathc; ++i) {
2096 string path = g.gl_pathv[i];
2097 va_dpy = try_open_va_h264(path, nullptr, nullptr);
2098 if (va_dpy != nullptr) {
2099 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
2102 if (need_env_reset) {
2103 unsetenv("LIBVA_MESSAGING_LEVEL");
2110 fprintf(stderr, "No suitable VA-API H.264 encoders were found in /dev/dri; giving up.\n");
2111 fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
2112 fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
2113 fprintf(stderr, "to expose Quick Sync. Alternatively, you can use --record-x264-video\n");
2114 fprintf(stderr, "to use software instead of hardware H.264 encoding, at the expense\n");
2115 fprintf(stderr, "of increased CPU usage and possibly bit rate.\n");