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>
18 #include <va/va_drm.h>
19 #include <va/va_drmcommon.h>
20 #include <va/va_enc_h264.h>
21 #include <va/va_x11.h>
24 #include <condition_variable>
39 #include <libavcodec/avcodec.h>
40 #include <libavformat/avio.h>
41 #include <libavutil/error.h>
42 #include <libdrm/drm_fourcc.h>
46 #include "audio_encoder.h"
49 #include "disk_space_estimator.h"
50 #include "ffmpeg_raii.h"
53 #include "print_latency.h"
54 #include "quicksync_encoder_impl.h"
55 #include "ref_counted_frame.h"
57 #include "x264_encoder.h"
59 using namespace movit;
61 using namespace std::chrono;
62 using namespace std::placeholders;
69 // These need to survive several QuickSyncEncoderImpl instances,
70 // so they are outside.
71 once_flag quick_sync_metrics_inited;
72 LatencyHistogram mixer_latency_histogram, qs_latency_histogram;
73 MuxMetrics current_file_mux_metrics, total_mux_metrics;
74 std::atomic<double> metric_current_file_start_time_seconds{0.0 / 0.0};
75 std::atomic<int64_t> metric_quick_sync_stalled_frames{0};
79 #define CHECK_VASTATUS(va_status, func) \
80 if (va_status != VA_STATUS_SUCCESS) { \
81 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
86 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
88 //#include "loadsurface.h"
90 #define NAL_REF_IDC_NONE 0
91 #define NAL_REF_IDC_LOW 1
92 #define NAL_REF_IDC_MEDIUM 2
93 #define NAL_REF_IDC_HIGH 3
101 #define SLICE_TYPE_P 0
102 #define SLICE_TYPE_B 1
103 #define SLICE_TYPE_I 2
104 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
105 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
106 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
109 #define ENTROPY_MODE_CAVLC 0
110 #define ENTROPY_MODE_CABAC 1
112 #define PROFILE_IDC_BASELINE 66
113 #define PROFILE_IDC_MAIN 77
114 #define PROFILE_IDC_HIGH 100
116 #define BITSTREAM_ALLOCATE_STEPPING 4096
118 static constexpr unsigned int MaxFrameNum = (2<<16);
119 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
120 static constexpr unsigned int Log2MaxFrameNum = 16;
121 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
125 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
126 // but if we don't delete it here, we get leaks. The GStreamer implementation
128 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
130 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
131 CHECK_VASTATUS(va_status, "vaRenderPicture");
133 for (int i = 0; i < num_buffers; ++i) {
134 va_status = vaDestroyBuffer(dpy, buffers[i]);
135 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
140 va_swap32(unsigned int val)
142 unsigned char *pval = (unsigned char *)&val;
144 return ((pval[0] << 24) |
151 bitstream_start(bitstream *bs)
153 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
154 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
159 bitstream_end(bitstream *bs)
161 int pos = (bs->bit_offset >> 5);
162 int bit_offset = (bs->bit_offset & 0x1f);
163 int bit_left = 32 - bit_offset;
166 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
171 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
173 int pos = (bs->bit_offset >> 5);
174 int bit_offset = (bs->bit_offset & 0x1f);
175 int bit_left = 32 - bit_offset;
180 bs->bit_offset += size_in_bits;
182 if (bit_left > size_in_bits) {
183 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
185 size_in_bits -= bit_left;
186 if (bit_left >= 32) {
187 bs->buffer[pos] = (val >> size_in_bits);
189 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
191 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
193 if (pos + 1 == bs->max_size_in_dword) {
194 bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
195 bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
198 bs->buffer[pos + 1] = val;
203 bitstream_put_ue(bitstream *bs, unsigned int val)
205 int size_in_bits = 0;
213 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
214 bitstream_put_ui(bs, val, size_in_bits);
218 bitstream_put_se(bitstream *bs, int val)
220 unsigned int new_val;
225 new_val = 2 * val - 1;
227 bitstream_put_ue(bs, new_val);
231 bitstream_byte_aligning(bitstream *bs, int bit)
233 int bit_offset = (bs->bit_offset & 0x7);
234 int bit_left = 8 - bit_offset;
240 assert(bit == 0 || bit == 1);
243 new_val = (1 << bit_left) - 1;
247 bitstream_put_ui(bs, new_val, bit_left);
251 rbsp_trailing_bits(bitstream *bs)
253 bitstream_put_ui(bs, 1, 1);
254 bitstream_byte_aligning(bs, 0);
257 static void nal_start_code_prefix(bitstream *bs)
259 bitstream_put_ui(bs, 0x00000001, 32);
262 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
264 bitstream_put_ui(bs, 0, 1); /* forbidden_zero_bit: 0 */
265 bitstream_put_ui(bs, nal_ref_idc, 2);
266 bitstream_put_ui(bs, nal_unit_type, 5);
269 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
271 int profile_idc = PROFILE_IDC_BASELINE;
273 if (h264_profile == VAProfileH264High)
274 profile_idc = PROFILE_IDC_HIGH;
275 else if (h264_profile == VAProfileH264Main)
276 profile_idc = PROFILE_IDC_MAIN;
278 bitstream_put_ui(bs, profile_idc, 8); /* profile_idc */
279 bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1); /* constraint_set0_flag */
280 bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1); /* constraint_set1_flag */
281 bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1); /* constraint_set2_flag */
282 bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1); /* constraint_set3_flag */
283 bitstream_put_ui(bs, 0, 4); /* reserved_zero_4bits */
284 bitstream_put_ui(bs, seq_param.level_idc, 8); /* level_idc */
285 bitstream_put_ue(bs, seq_param.seq_parameter_set_id); /* seq_parameter_set_id */
287 if ( profile_idc == PROFILE_IDC_HIGH) {
288 bitstream_put_ue(bs, 1); /* chroma_format_idc = 1, 4:2:0 */
289 bitstream_put_ue(bs, 0); /* bit_depth_luma_minus8 */
290 bitstream_put_ue(bs, 0); /* bit_depth_chroma_minus8 */
291 bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
292 bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
295 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
296 bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type); /* pic_order_cnt_type */
298 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
299 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4); /* log2_max_pic_order_cnt_lsb_minus4 */
304 bitstream_put_ue(bs, seq_param.max_num_ref_frames); /* num_ref_frames */
305 bitstream_put_ui(bs, 0, 1); /* gaps_in_frame_num_value_allowed_flag */
307 bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1); /* pic_width_in_mbs_minus1 */
308 bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
309 bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1); /* frame_mbs_only_flag */
311 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
315 bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1); /* direct_8x8_inference_flag */
316 bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1); /* frame_cropping_flag */
318 if (seq_param.frame_cropping_flag) {
319 bitstream_put_ue(bs, seq_param.frame_crop_left_offset); /* frame_crop_left_offset */
320 bitstream_put_ue(bs, seq_param.frame_crop_right_offset); /* frame_crop_right_offset */
321 bitstream_put_ue(bs, seq_param.frame_crop_top_offset); /* frame_crop_top_offset */
322 bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset); /* frame_crop_bottom_offset */
325 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
327 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
329 // See H.264 annex E for the definition of this header.
330 bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
331 bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
332 bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
333 bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
335 bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
336 bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
337 bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
339 bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
340 bitstream_put_ui(bs, 13, 8); /* transfer_characteristics (13 = sRGB) */
341 if (ycbcr_coefficients == YCBCR_REC_709) {
342 bitstream_put_ui(bs, 1, 8); /* matrix_coefficients (1 = BT.709) */
344 assert(ycbcr_coefficients == YCBCR_REC_601);
345 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
349 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
350 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
352 bitstream_put_ui(bs, 1, 32); // FPS
353 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
354 bitstream_put_ui(bs, 1, 1);
356 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
359 bitstream_put_ue(bs, 0); /* cpb_cnt_minus1 */
360 bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
361 bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
363 bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
364 bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
365 bitstream_put_ui(bs, 1, 1); /* cbr_flag[0] */
367 bitstream_put_ui(bs, 23, 5); /* initial_cpb_removal_delay_length_minus1 */
368 bitstream_put_ui(bs, 23, 5); /* cpb_removal_delay_length_minus1 */
369 bitstream_put_ui(bs, 23, 5); /* dpb_output_delay_length_minus1 */
370 bitstream_put_ui(bs, 23, 5); /* time_offset_length */
372 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
373 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
375 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
376 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
379 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
383 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
385 bitstream_put_ue(bs, pic_param.pic_parameter_set_id); /* pic_parameter_set_id */
386 bitstream_put_ue(bs, pic_param.seq_parameter_set_id); /* seq_parameter_set_id */
388 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
390 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
392 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
394 bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1); /* num_ref_idx_l0_active_minus1 */
395 bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1); /* num_ref_idx_l1_active_minus1 1 */
397 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1); /* weighted_pred_flag: 0 */
398 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2); /* weighted_bipred_idc: 0 */
400 bitstream_put_se(bs, pic_param.pic_init_qp - 26); /* pic_init_qp_minus26 */
401 bitstream_put_se(bs, 0); /* pic_init_qs_minus26 */
402 bitstream_put_se(bs, 0); /* chroma_qp_index_offset */
404 bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
405 bitstream_put_ui(bs, 0, 1); /* constrained_intra_pred_flag */
406 bitstream_put_ui(bs, 0, 1); /* redundant_pic_cnt_present_flag */
409 bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1); /*transform_8x8_mode_flag */
410 bitstream_put_ui(bs, 0, 1); /* pic_scaling_matrix_present_flag */
411 bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset ); /*second_chroma_qp_index_offset */
413 rbsp_trailing_bits(bs);
416 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
418 int first_mb_in_slice = slice_param.macroblock_address;
420 bitstream_put_ue(bs, first_mb_in_slice); /* first_mb_in_slice: 0 */
421 bitstream_put_ue(bs, slice_param.slice_type); /* slice_type */
422 bitstream_put_ue(bs, slice_param.pic_parameter_set_id); /* pic_parameter_set_id: 0 */
423 bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
425 /* frame_mbs_only_flag == 1 */
426 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
431 if (pic_param.pic_fields.bits.idr_pic_flag)
432 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
434 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
435 bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
436 /* pic_order_present_flag == 0 */
442 /* redundant_pic_cnt_present_flag == 0 */
444 if (IS_P_SLICE(slice_param.slice_type)) {
445 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
447 if (slice_param.num_ref_idx_active_override_flag)
448 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
450 /* ref_pic_list_reordering */
451 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
452 } else if (IS_B_SLICE(slice_param.slice_type)) {
453 bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1); /* direct_spatial_mv_pred: 1 */
455 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
457 if (slice_param.num_ref_idx_active_override_flag) {
458 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
459 bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
462 /* ref_pic_list_reordering */
463 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
464 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l1: 0 */
467 if ((pic_param.pic_fields.bits.weighted_pred_flag &&
468 IS_P_SLICE(slice_param.slice_type)) ||
469 ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
470 IS_B_SLICE(slice_param.slice_type))) {
471 /* FIXME: fill weight/offset table */
475 /* dec_ref_pic_marking */
476 if (pic_param.pic_fields.bits.reference_pic_flag) { /* nal_ref_idc != 0 */
477 unsigned char no_output_of_prior_pics_flag = 0;
478 unsigned char long_term_reference_flag = 0;
479 unsigned char adaptive_ref_pic_marking_mode_flag = 0;
481 if (pic_param.pic_fields.bits.idr_pic_flag) {
482 bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1); /* no_output_of_prior_pics_flag: 0 */
483 bitstream_put_ui(bs, long_term_reference_flag, 1); /* long_term_reference_flag: 0 */
485 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
489 if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
490 !IS_I_SLICE(slice_param.slice_type))
491 bitstream_put_ue(bs, slice_param.cabac_init_idc); /* cabac_init_idc: 0 */
493 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
495 /* ignore for SP/SI */
497 if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
498 bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc); /* disable_deblocking_filter_idc: 0 */
500 if (slice_param.disable_deblocking_filter_idc != 1) {
501 bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2); /* slice_alpha_c0_offset_div2: 2 */
502 bitstream_put_se(bs, slice_param.slice_beta_offset_div2); /* slice_beta_offset_div2: 2 */
506 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
507 bitstream_byte_aligning(bs, 1);
511 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
515 bitstream_start(&bs);
516 nal_start_code_prefix(&bs);
517 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
521 *header_buffer = (unsigned char *)bs.buffer;
522 return bs.bit_offset;
526 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
530 bitstream_start(&bs);
531 nal_start_code_prefix(&bs);
532 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
533 sps_rbsp(ycbcr_coefficients, &bs);
536 *header_buffer = (unsigned char *)bs.buffer;
537 return bs.bit_offset;
540 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
543 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
544 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
546 bitstream_start(&bs);
547 nal_start_code_prefix(&bs);
549 if (IS_I_SLICE(slice_param.slice_type)) {
550 nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
551 } else if (IS_P_SLICE(slice_param.slice_type)) {
552 nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
554 assert(IS_B_SLICE(slice_param.slice_type));
555 nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
561 *header_buffer = (unsigned char *)bs.buffer;
562 return bs.bit_offset;
567 Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
568 1) period between Frame #X and Frame #N = #X - #N
569 2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
570 3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
571 4) intra_period and intra_idr_period take precedence over ip_period
572 5) if ip_period > 1, intra_period and intra_idr_period are not the strict periods
573 of I/IDR frames, see bellow examples
574 -------------------------------------------------------------------
575 intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
576 0 ignored 1 IDRPPPPPPP ... (No IDR/I any more)
577 0 ignored >=2 IDR(PBB)(PBB)... (No IDR/I any more)
578 1 0 ignored IDRIIIIIII... (No IDR any more)
579 1 1 ignored IDR IDR IDR IDR...
580 1 >=2 ignored IDRII IDRII IDR... (1/3/ignore)
581 >=2 0 1 IDRPPP IPPP I... (3/0/1)
582 >=2 0 >=2 IDR(PBB)(PBB)(IBB) (6/0/3)
583 (PBB)(IBB)(PBB)(IBB)...
584 >=2 >=2 1 IDRPPPPP IPPPPP IPPPPP (6/18/1)
585 IDRPPPPP IPPPPP IPPPPP...
586 >=2 >=2 >=2 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
587 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
588 {IDR(PBB)(PBB)(IBB)(PBB)} (6/12/3)
589 {IDR(PBB)(PBB)(IBB)(PBB)}...
590 {IDR(PBB)(PBB)} (6/6/3)
594 // General pts/dts strategy:
596 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
597 // bit tricky. We assume first of all that the frame rate never goes _above_
598 // MAX_FPS, which gives us a frame period N. The decoder can always decode
599 // in at least this speed, as long at dts <= pts (the frame is not attempted
600 // presented before it is decoded). Furthermore, we never have longer chains of
601 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
602 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
603 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
604 // frame, with an order O <= C.)
606 // Many strategies are possible, but we establish these rules:
608 // - Tip frames have dts = pts - (C-O)*N.
609 // - Non-tip frames have dts = dts_last + N.
611 // An example, with C=2 and N=10 and the data flow showed with arrows:
614 // pts: 30 40 50 60 70 80
616 // dts: 10 30 20 60 50←40
621 // To show that this works fine also with irregular spacings, let's say that
622 // the third frame is delayed a bit (something earlier was dropped). Now the
623 // situation looks like this:
626 // pts: 30 40 80 90 100 110
628 // dts: 10 30 20 90 50←40
633 // The resetting on every tip frame makes sure dts never ends up lagging a lot
634 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
636 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
637 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
638 // a tip frame and should be given a dts based on the previous one.
643 void encoding2display_order(
644 int encoding_order, int intra_period,
645 int intra_idr_period, int ip_period,
646 int *displaying_order,
647 int *frame_type, int *pts_lag)
649 int encoding_order_gop = 0;
653 if (intra_period == 1) { /* all are I/IDR frames */
654 *displaying_order = encoding_order;
655 if (intra_idr_period == 0)
656 *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
658 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
662 if (intra_period == 0)
663 intra_idr_period = 0;
665 if (ip_period == 1) {
666 // No B-frames, sequence is like IDR PPPPP IPPPPP.
667 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
668 *displaying_order = encoding_order;
670 if (encoding_order_gop == 0) { /* the first frame */
671 *frame_type = FRAME_IDR;
672 } else if (intra_period != 0 && /* have I frames */
673 encoding_order_gop >= 2 &&
674 (encoding_order_gop % intra_period == 0)) {
675 *frame_type = FRAME_I;
677 *frame_type = FRAME_P;
682 // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
683 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
684 *pts_lag = -1; // Most frames are not tip frames.
686 if (encoding_order_gop == 0) { /* the first frame */
687 *frame_type = FRAME_IDR;
688 *displaying_order = encoding_order;
689 // IDR frames are a special case; I honestly can't find the logic behind
690 // why this is the right thing, but it seems to line up nicely in practice :-)
691 *pts_lag = TIMEBASE / MAX_FPS;
692 } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
693 *frame_type = FRAME_B;
694 *displaying_order = encoding_order - 1;
695 if ((encoding_order_gop % ip_period) == 0) {
696 *pts_lag = 0; // Last B-frame.
698 } else if (intra_period != 0 && /* have I frames */
699 encoding_order_gop >= 2 &&
700 ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
701 *frame_type = FRAME_I;
702 *displaying_order = encoding_order + ip_period - 1;
704 *frame_type = FRAME_P;
705 *displaying_order = encoding_order + ip_period - 1;
710 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
712 if (global_flags.x264_video_to_disk) {
713 // Quick Sync is entirely disabled.
714 use_zerocopy = false;
715 } else if (global_flags.uncompressed_video_to_http) {
716 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
717 use_zerocopy = false;
718 } else if (global_flags.x264_video_to_http) {
719 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
720 use_zerocopy = false;
724 global_flags.use_zerocopy = use_zerocopy;
727 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
729 if (va_display.empty()) {
730 x11_display = XOpenDisplay(NULL);
732 fprintf(stderr, "error: can't connect to X server!\n");
735 return vaGetDisplay(x11_display);
736 } else if (va_display[0] != '/') {
737 x11_display = XOpenDisplay(va_display.c_str());
739 fprintf(stderr, "error: can't connect to X server!\n");
742 return vaGetDisplay(x11_display);
744 drm_fd = open(va_display.c_str(), O_RDWR);
746 perror(va_display.c_str());
749 use_zerocopy = false;
750 return vaGetDisplayDRM(drm_fd);
754 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
757 XCloseDisplay(x11_display);
758 x11_display = nullptr;
765 int QuickSyncEncoderImpl::init_va(const string &va_display)
767 VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264ConstrainedBaseline};
768 VAEntrypoint *entrypoints;
769 int num_entrypoints, slice_entrypoint;
770 int support_encode = 0;
771 int major_ver, minor_ver;
775 va_dpy = va_open_display(va_display);
776 va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
777 CHECK_VASTATUS(va_status, "vaInitialize");
779 num_entrypoints = vaMaxNumEntrypoints(va_dpy);
780 entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
782 fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
786 /* use the highest profile */
787 for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
788 if ((h264_profile != ~0) && h264_profile != profile_list[i])
791 h264_profile = profile_list[i];
792 vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
793 for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
794 if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
799 if (support_encode == 1)
803 if (support_encode == 0) {
804 printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
805 printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
806 printf("to use VA-API against DRM instead of X11.\n");
809 switch (h264_profile) {
810 case VAProfileH264ConstrainedBaseline:
811 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
815 case VAProfileH264Main:
816 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
819 case VAProfileH264High:
820 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
823 h264_profile = VAProfileH264ConstrainedBaseline;
825 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
830 VAConfigAttrib attrib[VAConfigAttribTypeMax];
832 /* find out the format for the render target, and rate control mode */
833 for (i = 0; i < VAConfigAttribTypeMax; i++)
834 attrib[i].type = (VAConfigAttribType)i;
836 va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
837 &attrib[0], VAConfigAttribTypeMax);
838 CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
839 /* check the interested configattrib */
840 if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
841 printf("Not find desired YUV420 RT format\n");
844 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
845 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
849 if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
850 if (!(attrib[VAConfigAttribRateControl].value & VA_RC_CQP)) {
851 fprintf(stderr, "ERROR: VA-API encoder does not support CQP mode.\n");
855 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
856 config_attrib[config_attrib_num].value = VA_RC_CQP;
861 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
862 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
864 h264_packedheader = 1;
865 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
866 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
868 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
869 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
872 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
873 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
876 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
877 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
880 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
881 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
884 enc_packed_header_idx = config_attrib_num;
888 if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
889 config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
890 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
894 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
895 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
902 int QuickSyncEncoderImpl::setup_encode()
904 if (!global_flags.x264_video_to_disk) {
906 VASurfaceID *tmp_surfaceid;
908 VASurfaceID src_surface[SURFACE_NUM];
909 VASurfaceID ref_surface[SURFACE_NUM];
911 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
912 &config_attrib[0], config_attrib_num, &config_id);
913 CHECK_VASTATUS(va_status, "vaCreateConfig");
915 /* create source surfaces */
916 va_status = vaCreateSurfaces(va_dpy,
917 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
918 &src_surface[0], SURFACE_NUM,
920 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
922 /* create reference surfaces */
923 va_status = vaCreateSurfaces(va_dpy,
924 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
925 &ref_surface[0], SURFACE_NUM,
927 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
929 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
930 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
931 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
933 for (int i = 0; i < SURFACE_NUM; i++) {
934 gl_surfaces[i].src_surface = src_surface[i];
935 gl_surfaces[i].ref_surface = ref_surface[i];
938 /* Create a context for this encode pipe */
939 va_status = vaCreateContext(va_dpy, config_id,
940 frame_width_mbaligned, frame_height_mbaligned,
942 tmp_surfaceid, 2 * SURFACE_NUM,
944 CHECK_VASTATUS(va_status, "vaCreateContext");
947 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
949 for (int i = 0; i < SURFACE_NUM; i++) {
950 /* create coded buffer once for all
951 * other VA buffers which won't be used again after vaRenderPicture.
952 * so APP can always vaCreateBuffer for every frame
953 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
954 * so VA won't maintain the coded buffer
956 va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
957 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
958 CHECK_VASTATUS(va_status, "vaCreateBuffer");
962 /* create OpenGL objects */
963 for (int i = 0; i < SURFACE_NUM; i++) {
965 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
966 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
968 size_t bytes_per_pixel = (global_flags.x264_bit_depth > 8) ? 2 : 1;
970 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
971 // buffers, due to potentially differing pitch.
972 glGenBuffers(1, &gl_surfaces[i].pbo);
973 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
974 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);
975 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);
976 gl_surfaces[i].y_offset = 0;
977 gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
978 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
979 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
980 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
987 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
990 template<class T, class C>
991 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
993 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
994 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
995 sort(middle, end, less_than);
998 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1000 if (frame_type == FRAME_B)
1003 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1005 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1006 unique_lock<mutex> lock(storage_task_queue_mutex);
1008 // Insert the new frame at the start of the reference queue.
1009 reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1011 if (reference_frames.size() > num_ref_frames)
1013 // The back frame frame is no longer in use as a reference.
1014 int display_frame_num = reference_frames.back().display_number;
1015 assert(surface_for_frame.count(display_frame_num));
1016 release_gl_surface(display_frame_num);
1017 reference_frames.pop_back();
1020 // Mark this frame in use as a reference.
1021 assert(surface_for_frame.count(current_display_frame));
1022 ++surface_for_frame[current_display_frame]->refcount;
1024 current_ref_frame_num++;
1025 if (current_ref_frame_num > MaxFrameNum)
1026 current_ref_frame_num = 0;
1030 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1032 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1033 return a.frame_idx > b.frame_idx;
1036 for (size_t i = 0; i < reference_frames.size(); ++i) {
1037 RefPicList0_P[i] = reference_frames[i].pic;
1039 sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1042 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1044 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1045 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1047 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1048 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1051 for (size_t i = 0; i < reference_frames.size(); ++i) {
1052 RefPicList0_B[i] = reference_frames[i].pic;
1053 RefPicList1_B[i] = reference_frames[i].pic;
1055 sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1056 sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1060 int QuickSyncEncoderImpl::render_sequence()
1062 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1064 VAEncMiscParameterBuffer *misc_param;
1065 VAEncMiscParameterRateControl *misc_rate_ctrl;
1067 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1068 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1069 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1070 seq_param.bits_per_second = frame_bitrate;
1072 seq_param.intra_period = intra_period;
1073 seq_param.intra_idr_period = intra_idr_period;
1074 seq_param.ip_period = ip_period;
1076 seq_param.max_num_ref_frames = num_ref_frames;
1077 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1078 seq_param.time_scale = TIMEBASE * 2;
1079 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1080 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1081 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1082 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1083 seq_param.seq_fields.bits.chroma_format_idc = 1;
1084 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1086 if (frame_width != frame_width_mbaligned ||
1087 frame_height != frame_height_mbaligned) {
1088 seq_param.frame_cropping_flag = 1;
1089 seq_param.frame_crop_left_offset = 0;
1090 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1091 seq_param.frame_crop_top_offset = 0;
1092 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1095 va_status = vaCreateBuffer(va_dpy, context_id,
1096 VAEncSequenceParameterBufferType,
1097 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1098 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1100 va_status = vaCreateBuffer(va_dpy, context_id,
1101 VAEncMiscParameterBufferType,
1102 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1103 1, NULL, &rc_param_buf);
1104 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1106 vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1107 misc_param->type = VAEncMiscParameterTypeRateControl;
1108 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1109 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1110 misc_rate_ctrl->bits_per_second = frame_bitrate;
1111 misc_rate_ctrl->target_percentage = 66;
1112 misc_rate_ctrl->window_size = 1000;
1113 misc_rate_ctrl->initial_qp = initial_qp;
1114 misc_rate_ctrl->min_qp = minimal_qp;
1115 misc_rate_ctrl->basic_unit_size = 0;
1116 vaUnmapBuffer(va_dpy, rc_param_buf);
1118 render_id[0] = seq_param_buf;
1119 render_id[1] = rc_param_buf;
1121 render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1126 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1128 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1129 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1130 int PicOrderCntMsb, TopFieldOrderCnt;
1132 if (frame_type == FRAME_IDR)
1133 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1135 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1136 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1139 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1140 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1141 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1142 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1143 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1144 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1146 PicOrderCntMsb = prevPicOrderCntMsb;
1148 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1150 if (frame_type != FRAME_B) {
1151 PicOrderCntMsb_ref = PicOrderCntMsb;
1152 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1155 return TopFieldOrderCnt;
1158 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1160 VABufferID pic_param_buf;
1164 pic_param.CurrPic.picture_id = surf->ref_surface;
1165 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1166 pic_param.CurrPic.flags = 0;
1167 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1168 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1169 CurrentCurrPic = pic_param.CurrPic;
1171 for (i = 0; i < reference_frames.size(); i++) {
1172 pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1174 for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1175 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1176 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1179 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1180 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1181 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1182 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1183 pic_param.frame_num = current_ref_frame_num; // FIXME: is this correct?
1184 pic_param.coded_buf = surf->coded_buf;
1185 pic_param.last_picture = false; // FIXME
1186 pic_param.pic_init_qp = initial_qp;
1188 va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1189 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1190 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1192 render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1197 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1199 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1200 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1201 unsigned int length_in_bits;
1202 unsigned char *packedseq_buffer = NULL;
1205 length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1207 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1209 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1210 packedheader_param_buffer.has_emulation_bytes = 0;
1211 va_status = vaCreateBuffer(va_dpy,
1213 VAEncPackedHeaderParameterBufferType,
1214 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1215 &packedseq_para_bufid);
1216 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1218 va_status = vaCreateBuffer(va_dpy,
1220 VAEncPackedHeaderDataBufferType,
1221 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1222 &packedseq_data_bufid);
1223 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1225 render_id[0] = packedseq_para_bufid;
1226 render_id[1] = packedseq_data_bufid;
1227 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1229 free(packedseq_buffer);
1235 int QuickSyncEncoderImpl::render_packedpicture()
1237 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1238 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1239 unsigned int length_in_bits;
1240 unsigned char *packedpic_buffer = NULL;
1243 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1244 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1245 packedheader_param_buffer.bit_length = length_in_bits;
1246 packedheader_param_buffer.has_emulation_bytes = 0;
1248 va_status = vaCreateBuffer(va_dpy,
1250 VAEncPackedHeaderParameterBufferType,
1251 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1252 &packedpic_para_bufid);
1253 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1255 va_status = vaCreateBuffer(va_dpy,
1257 VAEncPackedHeaderDataBufferType,
1258 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1259 &packedpic_data_bufid);
1260 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1262 render_id[0] = packedpic_para_bufid;
1263 render_id[1] = packedpic_data_bufid;
1264 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1266 free(packedpic_buffer);
1271 void QuickSyncEncoderImpl::render_packedslice()
1273 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1274 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1275 unsigned int length_in_bits;
1276 unsigned char *packedslice_buffer = NULL;
1279 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1280 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1281 packedheader_param_buffer.bit_length = length_in_bits;
1282 packedheader_param_buffer.has_emulation_bytes = 0;
1284 va_status = vaCreateBuffer(va_dpy,
1286 VAEncPackedHeaderParameterBufferType,
1287 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1288 &packedslice_para_bufid);
1289 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1291 va_status = vaCreateBuffer(va_dpy,
1293 VAEncPackedHeaderDataBufferType,
1294 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1295 &packedslice_data_bufid);
1296 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1298 render_id[0] = packedslice_para_bufid;
1299 render_id[1] = packedslice_data_bufid;
1300 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1302 free(packedslice_buffer);
1305 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1307 VABufferID slice_param_buf;
1311 /* one frame, one slice */
1312 slice_param.macroblock_address = 0;
1313 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1314 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1315 if (frame_type == FRAME_IDR) {
1316 if (encoding_frame_num != 0)
1317 ++slice_param.idr_pic_id;
1318 } else if (frame_type == FRAME_P) {
1319 VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1320 update_RefPicList_P(RefPicList0_P);
1322 int refpiclist0_max = h264_maxref & 0xffff;
1323 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1325 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1326 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1327 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1329 } else if (frame_type == FRAME_B) {
1330 VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1331 update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1333 int refpiclist0_max = h264_maxref & 0xffff;
1334 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1336 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1337 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1338 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1339 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1342 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1343 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1344 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1345 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1349 slice_param.slice_alpha_c0_offset_div2 = 0;
1350 slice_param.slice_beta_offset_div2 = 0;
1351 slice_param.direct_spatial_mv_pred_flag = 1;
1352 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1355 if (h264_packedheader &&
1356 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1357 render_packedslice();
1359 va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1360 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1361 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1363 render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1370 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1372 VACodedBufferSegment *buf_list = NULL;
1377 va_status = vaMapBuffer(va_dpy, surf->coded_buf, (void **)(&buf_list));
1378 CHECK_VASTATUS(va_status, "vaMapBuffer");
1379 while (buf_list != NULL) {
1380 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1381 buf_list = (VACodedBufferSegment *) buf_list->next;
1383 vaUnmapBuffer(va_dpy, surf->coded_buf);
1385 static int frameno = 0;
1386 print_latency("Current Quick Sync latency (video inputs → disk mux):",
1387 task.received_ts, (task.frame_type == FRAME_B), &frameno, &qs_latency_histogram);
1392 memset(&pkt, 0, sizeof(pkt));
1394 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1395 pkt.size = data.size();
1396 pkt.stream_index = 0;
1397 if (task.frame_type == FRAME_IDR) {
1398 pkt.flags = AV_PKT_FLAG_KEY;
1402 pkt.duration = task.duration;
1404 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1406 if (!global_flags.uncompressed_video_to_http &&
1407 !global_flags.x264_video_to_http) {
1408 stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1414 // this is weird. but it seems to put a new frame onto the queue
1415 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1417 unique_lock<mutex> lock(storage_task_queue_mutex);
1418 storage_task_queue.push(move(task));
1419 storage_task_queue_changed.notify_all();
1422 void QuickSyncEncoderImpl::storage_task_thread()
1424 pthread_setname_np(pthread_self(), "QS_Storage");
1426 storage_task current;
1429 // wait until there's an encoded frame
1430 unique_lock<mutex> lock(storage_task_queue_mutex);
1431 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1432 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1433 current = move(storage_task_queue.front());
1434 storage_task_queue.pop();
1435 surf = surface_for_frame[current.display_order];
1436 assert(surf != nullptr);
1441 size_t display_order = current.display_order;
1442 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1444 // waits for data, then saves it to disk.
1445 va_status = vaSyncSurface(va_dpy, surf->src_surface);
1446 CHECK_VASTATUS(va_status, "vaSyncSurface");
1447 save_codeddata(surf, move(current));
1449 // Unlock the frame, and all its references.
1451 unique_lock<mutex> lock(storage_task_queue_mutex);
1452 release_gl_surface(display_order);
1454 for (size_t frame_num : ref_display_frame_numbers) {
1455 release_gl_surface(frame_num);
1461 void QuickSyncEncoderImpl::release_encode()
1463 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1464 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1465 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1466 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1469 vaDestroyContext(va_dpy, context_id);
1470 vaDestroyConfig(va_dpy, config_id);
1473 void QuickSyncEncoderImpl::release_gl_resources()
1475 assert(is_shutdown);
1476 if (has_released_gl_resources) {
1480 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1482 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1483 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1485 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1486 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1487 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1488 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1492 has_released_gl_resources = true;
1495 int QuickSyncEncoderImpl::deinit_va()
1497 vaTerminate(va_dpy);
1499 va_close_display(va_dpy);
1504 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)
1505 : 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)
1507 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1508 open_output_file(filename);
1509 file_audio_encoder->add_mux(file_mux.get());
1511 frame_width_mbaligned = (frame_width + 15) & (~15);
1512 frame_height_mbaligned = (frame_height + 15) & (~15);
1516 if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1517 assert(x264_encoder != nullptr);
1519 assert(x264_encoder == nullptr);
1522 enable_zerocopy_if_possible();
1523 if (!global_flags.x264_video_to_disk) {
1524 init_va(va_display);
1528 if (!global_flags.x264_video_to_disk) {
1529 memset(&seq_param, 0, sizeof(seq_param));
1530 memset(&pic_param, 0, sizeof(pic_param));
1531 memset(&slice_param, 0, sizeof(slice_param));
1534 call_once(quick_sync_metrics_inited, [](){
1535 mixer_latency_histogram.init("mixer");
1536 qs_latency_histogram.init("quick_sync");
1537 current_file_mux_metrics.init({{ "destination", "current_file" }});
1538 total_mux_metrics.init({{ "destination", "files_total" }});
1539 global_metrics.add("current_file_start_time_seconds", &metric_current_file_start_time_seconds, Metrics::TYPE_GAUGE);
1540 global_metrics.add("quick_sync_stalled_frames", &metric_quick_sync_stalled_frames);
1543 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1545 encode_thread = thread([this]{
1546 QOpenGLContext *context = create_context(this->surface);
1547 eglBindAPI(EGL_OPENGL_API);
1548 if (!make_current(context, this->surface)) {
1549 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1553 encode_thread_func();
1554 delete_context(context);
1558 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1561 release_gl_resources();
1564 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1566 for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1567 if (gl_surfaces[i].refcount == 0) {
1568 ++gl_surfaces[i].refcount;
1569 return &gl_surfaces[i];
1575 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1577 assert(surface_for_frame.count(display_frame_num));
1578 QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1579 if (--surf->refcount == 0) {
1580 assert(surface_for_frame.count(display_frame_num));
1581 surface_for_frame.erase(display_frame_num);
1582 storage_task_queue_changed.notify_all();
1586 bool QuickSyncEncoderImpl::is_zerocopy() const
1588 return use_zerocopy;
1591 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1593 assert(!is_shutdown);
1594 GLSurface *surf = nullptr;
1596 // Wait until this frame slot is done encoding.
1597 unique_lock<mutex> lock(storage_task_queue_mutex);
1598 surf = allocate_gl_surface();
1599 if (surf == nullptr) {
1600 fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1601 current_storage_frame);
1602 ++metric_quick_sync_stalled_frames;
1603 storage_task_queue_changed.wait(lock, [this, &surf]{
1604 if (storage_thread_should_quit)
1606 surf = allocate_gl_surface();
1607 return surf != nullptr;
1610 if (storage_thread_should_quit) return false;
1611 assert(surf != nullptr);
1612 surface_for_frame[current_storage_frame] = surf;
1616 *y_tex = surf->y_tex;
1617 *cbcr_tex = surf->cbcr_tex;
1619 surf->y_tex = *y_tex;
1620 surf->cbcr_tex = *cbcr_tex;
1623 if (!global_flags.x264_video_to_disk) {
1624 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1625 CHECK_VASTATUS(va_status, "vaDeriveImage");
1628 VABufferInfo buf_info;
1629 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1630 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1631 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1634 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1635 EGLint y_attribs[] = {
1636 EGL_WIDTH, frame_width,
1637 EGL_HEIGHT, frame_height,
1638 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1639 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1640 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1641 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1645 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1646 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1648 // Associate Y image to a texture.
1649 glBindTexture(GL_TEXTURE_2D, *y_tex);
1650 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1652 // Create CbCr image.
1653 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1654 EGLint cbcr_attribs[] = {
1655 EGL_WIDTH, frame_width / 2,
1656 EGL_HEIGHT, frame_height / 2,
1657 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1658 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1659 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1660 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1664 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1665 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1667 // Associate CbCr image to a texture.
1668 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1669 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1673 current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1678 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1680 lock_guard<mutex> lock(file_audio_encoder_mutex);
1681 assert(!is_shutdown);
1682 file_audio_encoder->encode_audio(audio, pts + global_delay());
1685 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1687 assert(!is_shutdown);
1689 if (!use_zerocopy) {
1690 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1693 unique_lock<mutex> lock(storage_task_queue_mutex);
1694 surf = surface_for_frame[current_storage_frame];
1695 assert(surf != nullptr);
1698 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1701 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1704 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1706 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1709 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1711 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1714 // We don't own these; the caller does.
1715 surf->y_tex = surf->cbcr_tex = 0;
1717 glBindTexture(GL_TEXTURE_2D, 0);
1719 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1722 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1726 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1728 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1732 unique_lock<mutex> lock(frame_queue_mutex);
1733 current_video_frame.fence = fence;
1734 pending_video_frames.push(move(current_video_frame));
1735 ++current_storage_frame;
1737 frame_queue_nonempty.notify_all();
1741 void QuickSyncEncoderImpl::shutdown()
1748 unique_lock<mutex> lock(frame_queue_mutex);
1749 encode_thread_should_quit = true;
1750 frame_queue_nonempty.notify_all();
1752 encode_thread.join();
1754 unique_lock<mutex> lock(storage_task_queue_mutex);
1755 storage_thread_should_quit = true;
1756 frame_queue_nonempty.notify_all();
1757 storage_task_queue_changed.notify_all();
1759 storage_thread.join();
1761 // Encode any leftover audio in the queues, and also any delayed frames.
1763 lock_guard<mutex> lock(file_audio_encoder_mutex);
1764 file_audio_encoder->encode_last_audio();
1767 if (!global_flags.x264_video_to_disk) {
1774 void QuickSyncEncoderImpl::close_file()
1777 metric_current_file_start_time_seconds = 0.0 / 0.0;
1780 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1782 AVFormatContext *avctx = avformat_alloc_context();
1783 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1784 assert(filename.size() < sizeof(avctx->filename) - 1);
1785 strcpy(avctx->filename, filename.c_str());
1787 string url = "file:" + filename;
1788 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1790 char tmp[AV_ERROR_MAX_STRING_SIZE];
1791 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1795 string video_extradata; // FIXME: See other comment about global headers.
1796 if (global_flags.x264_video_to_disk) {
1797 video_extradata = x264_encoder->get_global_headers();
1800 current_file_mux_metrics.reset();
1803 lock_guard<mutex> lock(file_audio_encoder_mutex);
1804 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1805 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1806 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1),
1807 Mux::WRITE_BACKGROUND,
1808 { ¤t_file_mux_metrics, &total_mux_metrics }));
1810 metric_current_file_start_time_seconds = get_timestamp_for_metrics();
1812 if (global_flags.x264_video_to_disk) {
1813 x264_encoder->add_mux(file_mux.get());
1817 void QuickSyncEncoderImpl::encode_thread_func()
1819 pthread_setname_np(pthread_self(), "QS_Encode");
1821 int64_t last_dts = -1;
1822 int gop_start_display_frame_num = 0;
1823 for (int display_frame_num = 0; ; ++display_frame_num) {
1824 // Wait for the frame to be in the queue. Note that this only means
1825 // we started rendering it.
1828 unique_lock<mutex> lock(frame_queue_mutex);
1829 frame_queue_nonempty.wait(lock, [this]{
1830 return encode_thread_should_quit || !pending_video_frames.empty();
1832 if (encode_thread_should_quit && pending_video_frames.empty()) {
1833 // We may have queued frames left in the reorder buffer
1834 // that were supposed to be B-frames, but have no P-frame
1835 // to be encoded against. If so, encode them all as
1836 // P-frames instead. Note that this happens under the mutex,
1837 // but nobody else uses it at this point, since we're shutting down,
1838 // so there's no contention.
1839 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1842 frame = move(pending_video_frames.front());
1843 pending_video_frames.pop();
1847 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1848 // Note that this implicitly waits for the frame to be done rendering.
1849 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1851 if (global_flags.x264_video_to_disk) {
1852 unique_lock<mutex> lock(storage_task_queue_mutex);
1853 release_gl_surface(display_frame_num);
1857 reorder_buffer[display_frame_num] = move(frame);
1859 // Now encode as many QuickSync frames as we can using the frames we have available.
1860 // (It could be zero, or it could be multiple.) FIXME: make a function.
1863 int frame_type, quicksync_display_frame_num;
1864 encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1865 &quicksync_display_frame_num, &frame_type, &pts_lag);
1866 if (!reorder_buffer.count(quicksync_display_frame_num)) {
1869 frame = move(reorder_buffer[quicksync_display_frame_num]);
1870 reorder_buffer.erase(quicksync_display_frame_num);
1872 if (frame_type == FRAME_IDR) {
1873 // Release any reference frames from the previous GOP.
1875 unique_lock<mutex> lock(storage_task_queue_mutex);
1876 for (const ReferenceFrame &frame : reference_frames) {
1877 release_gl_surface(frame.display_number);
1880 reference_frames.clear();
1881 current_ref_frame_num = 0;
1882 gop_start_display_frame_num = quicksync_display_frame_num;
1885 // Determine the dts of this frame.
1887 if (pts_lag == -1) {
1888 assert(last_dts != -1);
1889 dts = last_dts + (TIMEBASE / MAX_FPS);
1891 dts = frame.pts - pts_lag;
1895 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);
1896 ++quicksync_encoding_frame_num;
1901 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1903 if (reorder_buffer.empty()) {
1907 for (auto &pending_frame : reorder_buffer) {
1908 int display_frame_num = pending_frame.first;
1909 assert(display_frame_num > 0);
1910 PendingFrame frame = move(pending_frame.second);
1911 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1912 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1913 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1918 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1921 memset(&pkt, 0, sizeof(pkt));
1923 pkt.data = const_cast<uint8_t *>(data);
1924 pkt.size = frame_width * frame_height * 2;
1925 pkt.stream_index = 0;
1926 pkt.flags = AV_PKT_FLAG_KEY;
1927 pkt.duration = duration;
1928 stream_mux->add_packet(pkt, pts, pts);
1933 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1935 if (src_width == dst_pitch) {
1936 memcpy(dst, src, src_width * height);
1938 for (size_t y = 0; y < height; ++y) {
1939 const uint8_t *sptr = src + y * src_width;
1940 uint8_t *dptr = dst + y * dst_pitch;
1941 memcpy(dptr, sptr, src_width);
1948 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1950 // Wait for the GPU to be done with the frame.
1953 sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
1955 if (sync_status == GL_TIMEOUT_EXPIRED) {
1956 // NVIDIA likes to busy-wait; yield instead.
1957 this_thread::sleep_for(milliseconds(1));
1959 } while (sync_status == GL_TIMEOUT_EXPIRED);
1960 assert(sync_status != GL_WAIT_FAILED);
1962 ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1963 static int frameno = 0;
1964 print_latency("Current mixer latency (video inputs → ready for encode):",
1965 received_ts, false, &frameno, &mixer_latency_histogram);
1967 // Release back any input frames we needed to render this frame.
1968 frame.input_frames.clear();
1972 unique_lock<mutex> lock(storage_task_queue_mutex);
1973 surf = surface_for_frame[display_frame_num];
1974 assert(surf != nullptr);
1976 uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1977 if (global_flags.uncompressed_video_to_http) {
1978 add_packet_for_uncompressed_frame(pts, duration, data);
1979 } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1980 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1984 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1985 int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1987 const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1991 unique_lock<mutex> lock(storage_task_queue_mutex);
1992 surf = surface_for_frame[display_frame_num];
1993 assert(surf != nullptr);
1998 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1999 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2000 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2001 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2003 // Upload the frame to VA-API.
2004 unsigned char *surface_p = nullptr;
2005 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2007 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2008 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2010 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2011 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2013 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2014 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2017 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2018 CHECK_VASTATUS(va_status, "vaDestroyImage");
2020 // Schedule the frame for encoding.
2021 VASurfaceID va_surface = surf->src_surface;
2022 va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2023 CHECK_VASTATUS(va_status, "vaBeginPicture");
2025 if (frame_type == FRAME_IDR) {
2026 // FIXME: If the mux wants global headers, we should not put the
2027 // SPS/PPS before each IDR frame, but rather put it into the
2028 // codec extradata (formatted differently?).
2030 // NOTE: If we change ycbcr_coefficients, it will not take effect
2031 // before the next IDR frame. This is acceptable, as it should only
2032 // happen on a mode change, which is rare.
2034 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2035 if (h264_packedheader) {
2036 render_packedsequence(ycbcr_coefficients);
2037 render_packedpicture();
2040 //render_sequence();
2041 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2043 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2045 va_status = vaEndPicture(va_dpy, context_id);
2046 CHECK_VASTATUS(va_status, "vaEndPicture");
2048 update_ReferenceFrames(display_frame_num, frame_type);
2050 vector<size_t> ref_display_frame_numbers;
2052 // Lock the references for this frame; otherwise, they could be
2053 // rendered to before this frame is done encoding.
2055 unique_lock<mutex> lock(storage_task_queue_mutex);
2056 for (const ReferenceFrame &frame : reference_frames) {
2057 assert(surface_for_frame.count(frame.display_number));
2058 ++surface_for_frame[frame.display_number]->refcount;
2059 ref_display_frame_numbers.push_back(frame.display_number);
2063 // so now the data is done encoding (well, async job kicked off)...
2064 // we send that to the storage thread
2066 tmp.display_order = display_frame_num;
2067 tmp.frame_type = frame_type;
2070 tmp.duration = duration;
2071 tmp.ycbcr_coefficients = ycbcr_coefficients;
2072 tmp.received_ts = received_ts;
2073 tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2074 storage_task_enqueue(move(tmp));
2078 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)
2079 : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2081 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2082 QuickSyncEncoder::~QuickSyncEncoder() {}
2084 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2086 impl->add_audio(pts, audio);
2089 bool QuickSyncEncoder::is_zerocopy() const
2091 return impl->is_zerocopy();
2094 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2096 return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2099 RefCountedGLsync QuickSyncEncoder::end_frame()
2101 return impl->end_frame();
2104 void QuickSyncEncoder::shutdown()
2109 void QuickSyncEncoder::close_file()
2114 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2116 impl->set_stream_mux(mux);
2119 int64_t QuickSyncEncoder::global_delay() const {
2120 return impl->global_delay();