1 #include "quicksync_encoder.h"
3 #include <movit/resource_pool.h> // Must be above the Xlib includes.
4 #include <movit/util.h>
6 #include <EGL/eglplatform.h>
17 #include <va/va_drm.h>
18 #include <va/va_drmcommon.h>
19 #include <va/va_enc_h264.h>
20 #include <va/va_x11.h>
23 #include <condition_variable>
38 #include <libavcodec/avcodec.h>
39 #include <libavformat/avio.h>
40 #include <libavutil/error.h>
41 #include <libdrm/drm_fourcc.h>
45 #include "audio_encoder.h"
48 #include "disk_space_estimator.h"
49 #include "ffmpeg_raii.h"
52 #include "print_latency.h"
53 #include "quicksync_encoder_impl.h"
54 #include "ref_counted_frame.h"
56 #include "x264_encoder.h"
59 using namespace std::chrono;
60 using namespace std::placeholders;
65 #define CHECK_VASTATUS(va_status, func) \
66 if (va_status != VA_STATUS_SUCCESS) { \
67 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
71 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
73 //#include "loadsurface.h"
75 #define NAL_REF_IDC_NONE 0
76 #define NAL_REF_IDC_LOW 1
77 #define NAL_REF_IDC_MEDIUM 2
78 #define NAL_REF_IDC_HIGH 3
86 #define SLICE_TYPE_P 0
87 #define SLICE_TYPE_B 1
88 #define SLICE_TYPE_I 2
89 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
90 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
91 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
94 #define ENTROPY_MODE_CAVLC 0
95 #define ENTROPY_MODE_CABAC 1
97 #define PROFILE_IDC_BASELINE 66
98 #define PROFILE_IDC_MAIN 77
99 #define PROFILE_IDC_HIGH 100
101 #define BITSTREAM_ALLOCATE_STEPPING 4096
103 static constexpr unsigned int MaxFrameNum = (2<<16);
104 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
105 static constexpr unsigned int Log2MaxFrameNum = 16;
106 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
107 static constexpr int rc_default_modes[] = { // Priority list of modes.
110 VA_RC_VBR_CONSTRAINED,
116 /* thread to save coded data */
117 #define SRC_SURFACE_FREE 0
118 #define SRC_SURFACE_IN_ENCODING 1
122 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
123 // but if we don't delete it here, we get leaks. The GStreamer implementation
125 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
127 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
128 CHECK_VASTATUS(va_status, "vaRenderPicture");
130 for (int i = 0; i < num_buffers; ++i) {
131 va_status = vaDestroyBuffer(dpy, buffers[i]);
132 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
137 va_swap32(unsigned int val)
139 unsigned char *pval = (unsigned char *)&val;
141 return ((pval[0] << 24) |
148 bitstream_start(bitstream *bs)
150 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
151 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
156 bitstream_end(bitstream *bs)
158 int pos = (bs->bit_offset >> 5);
159 int bit_offset = (bs->bit_offset & 0x1f);
160 int bit_left = 32 - bit_offset;
163 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
168 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
170 int pos = (bs->bit_offset >> 5);
171 int bit_offset = (bs->bit_offset & 0x1f);
172 int bit_left = 32 - bit_offset;
177 bs->bit_offset += size_in_bits;
179 if (bit_left > size_in_bits) {
180 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
182 size_in_bits -= bit_left;
183 if (bit_left >= 32) {
184 bs->buffer[pos] = (val >> size_in_bits);
186 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
188 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
190 if (pos + 1 == bs->max_size_in_dword) {
191 bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
192 bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
195 bs->buffer[pos + 1] = val;
200 bitstream_put_ue(bitstream *bs, unsigned int val)
202 int size_in_bits = 0;
210 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
211 bitstream_put_ui(bs, val, size_in_bits);
215 bitstream_put_se(bitstream *bs, int val)
217 unsigned int new_val;
222 new_val = 2 * val - 1;
224 bitstream_put_ue(bs, new_val);
228 bitstream_byte_aligning(bitstream *bs, int bit)
230 int bit_offset = (bs->bit_offset & 0x7);
231 int bit_left = 8 - bit_offset;
237 assert(bit == 0 || bit == 1);
240 new_val = (1 << bit_left) - 1;
244 bitstream_put_ui(bs, new_val, bit_left);
248 rbsp_trailing_bits(bitstream *bs)
250 bitstream_put_ui(bs, 1, 1);
251 bitstream_byte_aligning(bs, 0);
254 static void nal_start_code_prefix(bitstream *bs)
256 bitstream_put_ui(bs, 0x00000001, 32);
259 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
261 bitstream_put_ui(bs, 0, 1); /* forbidden_zero_bit: 0 */
262 bitstream_put_ui(bs, nal_ref_idc, 2);
263 bitstream_put_ui(bs, nal_unit_type, 5);
266 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
268 int profile_idc = PROFILE_IDC_BASELINE;
270 if (h264_profile == VAProfileH264High)
271 profile_idc = PROFILE_IDC_HIGH;
272 else if (h264_profile == VAProfileH264Main)
273 profile_idc = PROFILE_IDC_MAIN;
275 bitstream_put_ui(bs, profile_idc, 8); /* profile_idc */
276 bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1); /* constraint_set0_flag */
277 bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1); /* constraint_set1_flag */
278 bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1); /* constraint_set2_flag */
279 bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1); /* constraint_set3_flag */
280 bitstream_put_ui(bs, 0, 4); /* reserved_zero_4bits */
281 bitstream_put_ui(bs, seq_param.level_idc, 8); /* level_idc */
282 bitstream_put_ue(bs, seq_param.seq_parameter_set_id); /* seq_parameter_set_id */
284 if ( profile_idc == PROFILE_IDC_HIGH) {
285 bitstream_put_ue(bs, 1); /* chroma_format_idc = 1, 4:2:0 */
286 bitstream_put_ue(bs, 0); /* bit_depth_luma_minus8 */
287 bitstream_put_ue(bs, 0); /* bit_depth_chroma_minus8 */
288 bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
289 bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
292 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
293 bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type); /* pic_order_cnt_type */
295 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
296 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4); /* log2_max_pic_order_cnt_lsb_minus4 */
301 bitstream_put_ue(bs, seq_param.max_num_ref_frames); /* num_ref_frames */
302 bitstream_put_ui(bs, 0, 1); /* gaps_in_frame_num_value_allowed_flag */
304 bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1); /* pic_width_in_mbs_minus1 */
305 bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
306 bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1); /* frame_mbs_only_flag */
308 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
312 bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1); /* direct_8x8_inference_flag */
313 bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1); /* frame_cropping_flag */
315 if (seq_param.frame_cropping_flag) {
316 bitstream_put_ue(bs, seq_param.frame_crop_left_offset); /* frame_crop_left_offset */
317 bitstream_put_ue(bs, seq_param.frame_crop_right_offset); /* frame_crop_right_offset */
318 bitstream_put_ue(bs, seq_param.frame_crop_top_offset); /* frame_crop_top_offset */
319 bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset); /* frame_crop_bottom_offset */
322 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
324 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
326 // See H.264 annex E for the definition of this header.
327 bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
328 bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
329 bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
330 bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
332 bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
333 bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
334 bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
336 bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
337 bitstream_put_ui(bs, 2, 8); /* transfer_characteristics (2 = unspecified, since we use sRGB) */
338 if (global_flags.ycbcr_rec709_coefficients) {
339 bitstream_put_ui(bs, 1, 8); /* matrix_coefficients (1 = BT.709) */
341 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
345 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
346 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
348 bitstream_put_ui(bs, 1, 32); // FPS
349 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
350 bitstream_put_ui(bs, 1, 1);
352 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
355 bitstream_put_ue(bs, 0); /* cpb_cnt_minus1 */
356 bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
357 bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
359 bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
360 bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
361 bitstream_put_ui(bs, 1, 1); /* cbr_flag[0] */
363 bitstream_put_ui(bs, 23, 5); /* initial_cpb_removal_delay_length_minus1 */
364 bitstream_put_ui(bs, 23, 5); /* cpb_removal_delay_length_minus1 */
365 bitstream_put_ui(bs, 23, 5); /* dpb_output_delay_length_minus1 */
366 bitstream_put_ui(bs, 23, 5); /* time_offset_length */
368 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
369 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
371 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
372 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
375 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
379 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
381 bitstream_put_ue(bs, pic_param.pic_parameter_set_id); /* pic_parameter_set_id */
382 bitstream_put_ue(bs, pic_param.seq_parameter_set_id); /* seq_parameter_set_id */
384 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
386 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
388 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
390 bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1); /* num_ref_idx_l0_active_minus1 */
391 bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1); /* num_ref_idx_l1_active_minus1 1 */
393 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1); /* weighted_pred_flag: 0 */
394 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2); /* weighted_bipred_idc: 0 */
396 bitstream_put_se(bs, pic_param.pic_init_qp - 26); /* pic_init_qp_minus26 */
397 bitstream_put_se(bs, 0); /* pic_init_qs_minus26 */
398 bitstream_put_se(bs, 0); /* chroma_qp_index_offset */
400 bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
401 bitstream_put_ui(bs, 0, 1); /* constrained_intra_pred_flag */
402 bitstream_put_ui(bs, 0, 1); /* redundant_pic_cnt_present_flag */
405 bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1); /*transform_8x8_mode_flag */
406 bitstream_put_ui(bs, 0, 1); /* pic_scaling_matrix_present_flag */
407 bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset ); /*second_chroma_qp_index_offset */
409 rbsp_trailing_bits(bs);
412 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
414 int first_mb_in_slice = slice_param.macroblock_address;
416 bitstream_put_ue(bs, first_mb_in_slice); /* first_mb_in_slice: 0 */
417 bitstream_put_ue(bs, slice_param.slice_type); /* slice_type */
418 bitstream_put_ue(bs, slice_param.pic_parameter_set_id); /* pic_parameter_set_id: 0 */
419 bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
421 /* frame_mbs_only_flag == 1 */
422 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
427 if (pic_param.pic_fields.bits.idr_pic_flag)
428 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
430 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
431 bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
432 /* pic_order_present_flag == 0 */
438 /* redundant_pic_cnt_present_flag == 0 */
440 if (IS_P_SLICE(slice_param.slice_type)) {
441 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
443 if (slice_param.num_ref_idx_active_override_flag)
444 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
446 /* ref_pic_list_reordering */
447 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
448 } else if (IS_B_SLICE(slice_param.slice_type)) {
449 bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1); /* direct_spatial_mv_pred: 1 */
451 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
453 if (slice_param.num_ref_idx_active_override_flag) {
454 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
455 bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
458 /* ref_pic_list_reordering */
459 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
460 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l1: 0 */
463 if ((pic_param.pic_fields.bits.weighted_pred_flag &&
464 IS_P_SLICE(slice_param.slice_type)) ||
465 ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
466 IS_B_SLICE(slice_param.slice_type))) {
467 /* FIXME: fill weight/offset table */
471 /* dec_ref_pic_marking */
472 if (pic_param.pic_fields.bits.reference_pic_flag) { /* nal_ref_idc != 0 */
473 unsigned char no_output_of_prior_pics_flag = 0;
474 unsigned char long_term_reference_flag = 0;
475 unsigned char adaptive_ref_pic_marking_mode_flag = 0;
477 if (pic_param.pic_fields.bits.idr_pic_flag) {
478 bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1); /* no_output_of_prior_pics_flag: 0 */
479 bitstream_put_ui(bs, long_term_reference_flag, 1); /* long_term_reference_flag: 0 */
481 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
485 if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
486 !IS_I_SLICE(slice_param.slice_type))
487 bitstream_put_ue(bs, slice_param.cabac_init_idc); /* cabac_init_idc: 0 */
489 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
491 /* ignore for SP/SI */
493 if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
494 bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc); /* disable_deblocking_filter_idc: 0 */
496 if (slice_param.disable_deblocking_filter_idc != 1) {
497 bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2); /* slice_alpha_c0_offset_div2: 2 */
498 bitstream_put_se(bs, slice_param.slice_beta_offset_div2); /* slice_beta_offset_div2: 2 */
502 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
503 bitstream_byte_aligning(bs, 1);
507 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
511 bitstream_start(&bs);
512 nal_start_code_prefix(&bs);
513 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
517 *header_buffer = (unsigned char *)bs.buffer;
518 return bs.bit_offset;
522 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
526 bitstream_start(&bs);
527 nal_start_code_prefix(&bs);
528 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
532 *header_buffer = (unsigned char *)bs.buffer;
533 return bs.bit_offset;
536 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
539 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
540 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
542 bitstream_start(&bs);
543 nal_start_code_prefix(&bs);
545 if (IS_I_SLICE(slice_param.slice_type)) {
546 nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
547 } else if (IS_P_SLICE(slice_param.slice_type)) {
548 nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
550 assert(IS_B_SLICE(slice_param.slice_type));
551 nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
557 *header_buffer = (unsigned char *)bs.buffer;
558 return bs.bit_offset;
563 Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
564 1) period between Frame #X and Frame #N = #X - #N
565 2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
566 3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
567 4) intra_period and intra_idr_period take precedence over ip_period
568 5) if ip_period > 1, intra_period and intra_idr_period are not the strict periods
569 of I/IDR frames, see bellow examples
570 -------------------------------------------------------------------
571 intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
572 0 ignored 1 IDRPPPPPPP ... (No IDR/I any more)
573 0 ignored >=2 IDR(PBB)(PBB)... (No IDR/I any more)
574 1 0 ignored IDRIIIIIII... (No IDR any more)
575 1 1 ignored IDR IDR IDR IDR...
576 1 >=2 ignored IDRII IDRII IDR... (1/3/ignore)
577 >=2 0 1 IDRPPP IPPP I... (3/0/1)
578 >=2 0 >=2 IDR(PBB)(PBB)(IBB) (6/0/3)
579 (PBB)(IBB)(PBB)(IBB)...
580 >=2 >=2 1 IDRPPPPP IPPPPP IPPPPP (6/18/1)
581 IDRPPPPP IPPPPP IPPPPP...
582 >=2 >=2 >=2 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
583 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
584 {IDR(PBB)(PBB)(IBB)(PBB)} (6/12/3)
585 {IDR(PBB)(PBB)(IBB)(PBB)}...
586 {IDR(PBB)(PBB)} (6/6/3)
590 // General pts/dts strategy:
592 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
593 // bit tricky. We assume first of all that the frame rate never goes _above_
594 // MAX_FPS, which gives us a frame period N. The decoder can always decode
595 // in at least this speed, as long at dts <= pts (the frame is not attempted
596 // presented before it is decoded). Furthermore, we never have longer chains of
597 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
598 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
599 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
600 // frame, with an order O <= C.)
602 // Many strategies are possible, but we establish these rules:
604 // - Tip frames have dts = pts - (C-O)*N.
605 // - Non-tip frames have dts = dts_last + N.
607 // An example, with C=2 and N=10 and the data flow showed with arrows:
610 // pts: 30 40 50 60 70 80
612 // dts: 10 30 20 60 50←40
617 // To show that this works fine also with irregular spacings, let's say that
618 // the third frame is delayed a bit (something earlier was dropped). Now the
619 // situation looks like this:
622 // pts: 30 40 80 90 100 110
624 // dts: 10 30 20 90 50←40
629 // The resetting on every tip frame makes sure dts never ends up lagging a lot
630 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
632 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
633 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
634 // a tip frame and should be given a dts based on the previous one.
639 void encoding2display_order(
640 int encoding_order, int intra_period,
641 int intra_idr_period, int ip_period,
642 int *displaying_order,
643 int *frame_type, int *pts_lag)
645 int encoding_order_gop = 0;
649 if (intra_period == 1) { /* all are I/IDR frames */
650 *displaying_order = encoding_order;
651 if (intra_idr_period == 0)
652 *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
654 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
658 if (intra_period == 0)
659 intra_idr_period = 0;
661 if (ip_period == 1) {
662 // No B-frames, sequence is like IDR PPPPP IPPPPP.
663 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
664 *displaying_order = encoding_order;
666 if (encoding_order_gop == 0) { /* the first frame */
667 *frame_type = FRAME_IDR;
668 } else if (intra_period != 0 && /* have I frames */
669 encoding_order_gop >= 2 &&
670 (encoding_order_gop % intra_period == 0)) {
671 *frame_type = FRAME_I;
673 *frame_type = FRAME_P;
678 // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
679 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
680 *pts_lag = -1; // Most frames are not tip frames.
682 if (encoding_order_gop == 0) { /* the first frame */
683 *frame_type = FRAME_IDR;
684 *displaying_order = encoding_order;
685 // IDR frames are a special case; I honestly can't find the logic behind
686 // why this is the right thing, but it seems to line up nicely in practice :-)
687 *pts_lag = TIMEBASE / MAX_FPS;
688 } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
689 *frame_type = FRAME_B;
690 *displaying_order = encoding_order - 1;
691 if ((encoding_order_gop % ip_period) == 0) {
692 *pts_lag = 0; // Last B-frame.
694 } else if (intra_period != 0 && /* have I frames */
695 encoding_order_gop >= 2 &&
696 ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
697 *frame_type = FRAME_I;
698 *displaying_order = encoding_order + ip_period - 1;
700 *frame_type = FRAME_P;
701 *displaying_order = encoding_order + ip_period - 1;
706 static const char *rc_to_string(int rc_mode)
719 case VA_RC_VBR_CONSTRAINED:
720 return "VBR_CONSTRAINED";
726 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
728 if (global_flags.uncompressed_video_to_http) {
729 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
730 use_zerocopy = false;
731 } else if (global_flags.x264_video_to_http) {
732 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
733 use_zerocopy = false;
739 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
741 if (va_display.empty()) {
742 x11_display = XOpenDisplay(NULL);
744 fprintf(stderr, "error: can't connect to X server!\n");
747 enable_zerocopy_if_possible();
748 return vaGetDisplay(x11_display);
749 } else if (va_display[0] != '/') {
750 x11_display = XOpenDisplay(va_display.c_str());
752 fprintf(stderr, "error: can't connect to X server!\n");
755 enable_zerocopy_if_possible();
756 return vaGetDisplay(x11_display);
758 drm_fd = open(va_display.c_str(), O_RDWR);
760 perror(va_display.c_str());
763 use_zerocopy = false;
764 return vaGetDisplayDRM(drm_fd);
768 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
771 XCloseDisplay(x11_display);
772 x11_display = nullptr;
779 int QuickSyncEncoderImpl::init_va(const string &va_display)
781 VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
782 VAEntrypoint *entrypoints;
783 int num_entrypoints, slice_entrypoint;
784 int support_encode = 0;
785 int major_ver, minor_ver;
789 va_dpy = va_open_display(va_display);
790 va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
791 CHECK_VASTATUS(va_status, "vaInitialize");
793 num_entrypoints = vaMaxNumEntrypoints(va_dpy);
794 entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
796 fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
800 /* use the highest profile */
801 for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
802 if ((h264_profile != ~0) && h264_profile != profile_list[i])
805 h264_profile = profile_list[i];
806 vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
807 for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
808 if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
813 if (support_encode == 1)
817 if (support_encode == 0) {
818 printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
819 printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
820 printf("to use VA-API against DRM instead of X11.\n");
823 switch (h264_profile) {
824 case VAProfileH264Baseline:
826 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
827 h264_entropy_mode = 0;
829 case VAProfileH264ConstrainedBaseline:
830 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
834 case VAProfileH264Main:
835 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
838 case VAProfileH264High:
839 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
842 h264_profile = VAProfileH264Baseline;
844 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
849 VAConfigAttrib attrib[VAConfigAttribTypeMax];
851 /* find out the format for the render target, and rate control mode */
852 for (i = 0; i < VAConfigAttribTypeMax; i++)
853 attrib[i].type = (VAConfigAttribType)i;
855 va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
856 &attrib[0], VAConfigAttribTypeMax);
857 CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
858 /* check the interested configattrib */
859 if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
860 printf("Not find desired YUV420 RT format\n");
863 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
864 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
868 if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
869 int tmp = attrib[VAConfigAttribRateControl].value;
871 if (rc_mode == -1 || !(rc_mode & tmp)) {
873 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
876 for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
877 if (rc_default_modes[i] & tmp) {
878 rc_mode = rc_default_modes[i];
884 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
885 config_attrib[config_attrib_num].value = rc_mode;
890 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
891 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
893 h264_packedheader = 1;
894 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
895 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
897 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
898 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
901 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
902 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
905 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
906 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
909 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
910 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
913 enc_packed_header_idx = config_attrib_num;
917 if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
918 config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
919 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
923 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
924 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
931 int QuickSyncEncoderImpl::setup_encode()
934 VASurfaceID *tmp_surfaceid;
935 int codedbuf_size, i;
936 VASurfaceID src_surface[SURFACE_NUM];
937 VASurfaceID ref_surface[SURFACE_NUM];
939 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
940 &config_attrib[0], config_attrib_num, &config_id);
941 CHECK_VASTATUS(va_status, "vaCreateConfig");
943 /* create source surfaces */
944 va_status = vaCreateSurfaces(va_dpy,
945 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
946 &src_surface[0], SURFACE_NUM,
948 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
950 /* create reference surfaces */
951 va_status = vaCreateSurfaces(va_dpy,
952 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
953 &ref_surface[0], SURFACE_NUM,
955 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
957 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
958 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
959 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
961 /* Create a context for this encode pipe */
962 va_status = vaCreateContext(va_dpy, config_id,
963 frame_width_mbaligned, frame_height_mbaligned,
965 tmp_surfaceid, 2 * SURFACE_NUM,
967 CHECK_VASTATUS(va_status, "vaCreateContext");
970 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
972 for (i = 0; i < SURFACE_NUM; i++) {
973 /* create coded buffer once for all
974 * other VA buffers which won't be used again after vaRenderPicture.
975 * so APP can always vaCreateBuffer for every frame
976 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
977 * so VA won't maintain the coded buffer
979 va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
980 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
981 CHECK_VASTATUS(va_status, "vaCreateBuffer");
984 /* create OpenGL objects */
985 //glGenFramebuffers(SURFACE_NUM, fbos);
987 for (i = 0; i < SURFACE_NUM; i++) {
989 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
990 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
992 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
993 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
995 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
996 // buffers, due to potentially differing pitch.
997 glGenBuffers(1, &gl_surfaces[i].pbo);
998 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
999 glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1000 uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1001 gl_surfaces[i].y_offset = 0;
1002 gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1003 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1004 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1005 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1009 for (i = 0; i < SURFACE_NUM; i++) {
1010 gl_surfaces[i].src_surface = src_surface[i];
1011 gl_surfaces[i].ref_surface = ref_surface[i];
1017 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1020 template<class T, class C>
1021 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1023 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1024 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1025 sort(middle, end, less_than);
1028 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1032 if (frame_type == FRAME_B)
1035 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1037 if (numShortTerm > num_ref_frames)
1038 numShortTerm = num_ref_frames;
1039 for (i=numShortTerm-1; i>0; i--)
1040 ReferenceFrames[i] = ReferenceFrames[i-1];
1041 ReferenceFrames[0] = CurrentCurrPic;
1043 current_frame_num++;
1044 if (current_frame_num > MaxFrameNum)
1045 current_frame_num = 0;
1049 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1051 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1052 return a.frame_idx > b.frame_idx;
1054 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1055 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1057 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1058 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1061 if (frame_type == FRAME_P) {
1062 memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1063 sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1064 } else if (frame_type == FRAME_B) {
1065 memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1066 sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1068 memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1069 sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1076 int QuickSyncEncoderImpl::render_sequence()
1078 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1080 VAEncMiscParameterBuffer *misc_param;
1081 VAEncMiscParameterRateControl *misc_rate_ctrl;
1083 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1084 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1085 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1086 seq_param.bits_per_second = frame_bitrate;
1088 seq_param.intra_period = intra_period;
1089 seq_param.intra_idr_period = intra_idr_period;
1090 seq_param.ip_period = ip_period;
1092 seq_param.max_num_ref_frames = num_ref_frames;
1093 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1094 seq_param.time_scale = TIMEBASE * 2;
1095 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1096 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1097 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1098 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1099 seq_param.seq_fields.bits.chroma_format_idc = 1;
1100 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1102 if (frame_width != frame_width_mbaligned ||
1103 frame_height != frame_height_mbaligned) {
1104 seq_param.frame_cropping_flag = 1;
1105 seq_param.frame_crop_left_offset = 0;
1106 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1107 seq_param.frame_crop_top_offset = 0;
1108 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1111 va_status = vaCreateBuffer(va_dpy, context_id,
1112 VAEncSequenceParameterBufferType,
1113 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1114 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1116 va_status = vaCreateBuffer(va_dpy, context_id,
1117 VAEncMiscParameterBufferType,
1118 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1119 1, NULL, &rc_param_buf);
1120 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1122 vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1123 misc_param->type = VAEncMiscParameterTypeRateControl;
1124 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1125 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1126 misc_rate_ctrl->bits_per_second = frame_bitrate;
1127 misc_rate_ctrl->target_percentage = 66;
1128 misc_rate_ctrl->window_size = 1000;
1129 misc_rate_ctrl->initial_qp = initial_qp;
1130 misc_rate_ctrl->min_qp = minimal_qp;
1131 misc_rate_ctrl->basic_unit_size = 0;
1132 vaUnmapBuffer(va_dpy, rc_param_buf);
1134 render_id[0] = seq_param_buf;
1135 render_id[1] = rc_param_buf;
1137 render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1142 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1144 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1145 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1146 int PicOrderCntMsb, TopFieldOrderCnt;
1148 if (frame_type == FRAME_IDR)
1149 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1151 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1152 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1155 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1156 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1157 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1158 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1159 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1160 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1162 PicOrderCntMsb = prevPicOrderCntMsb;
1164 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1166 if (frame_type != FRAME_B) {
1167 PicOrderCntMsb_ref = PicOrderCntMsb;
1168 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1171 return TopFieldOrderCnt;
1174 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1176 VABufferID pic_param_buf;
1180 pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1181 pic_param.CurrPic.frame_idx = current_frame_num;
1182 pic_param.CurrPic.flags = 0;
1183 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1184 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1185 CurrentCurrPic = pic_param.CurrPic;
1187 memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1188 for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1189 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1190 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1193 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1194 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1195 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1196 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1197 pic_param.frame_num = current_frame_num;
1198 pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1199 pic_param.last_picture = false; // FIXME
1200 pic_param.pic_init_qp = initial_qp;
1202 va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1203 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1204 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1206 render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1211 int QuickSyncEncoderImpl::render_packedsequence()
1213 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1214 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1215 unsigned int length_in_bits;
1216 unsigned char *packedseq_buffer = NULL;
1219 length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1221 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1223 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1224 packedheader_param_buffer.has_emulation_bytes = 0;
1225 va_status = vaCreateBuffer(va_dpy,
1227 VAEncPackedHeaderParameterBufferType,
1228 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1229 &packedseq_para_bufid);
1230 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1232 va_status = vaCreateBuffer(va_dpy,
1234 VAEncPackedHeaderDataBufferType,
1235 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1236 &packedseq_data_bufid);
1237 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1239 render_id[0] = packedseq_para_bufid;
1240 render_id[1] = packedseq_data_bufid;
1241 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1243 free(packedseq_buffer);
1249 int QuickSyncEncoderImpl::render_packedpicture()
1251 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1252 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1253 unsigned int length_in_bits;
1254 unsigned char *packedpic_buffer = NULL;
1257 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1258 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1259 packedheader_param_buffer.bit_length = length_in_bits;
1260 packedheader_param_buffer.has_emulation_bytes = 0;
1262 va_status = vaCreateBuffer(va_dpy,
1264 VAEncPackedHeaderParameterBufferType,
1265 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1266 &packedpic_para_bufid);
1267 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1269 va_status = vaCreateBuffer(va_dpy,
1271 VAEncPackedHeaderDataBufferType,
1272 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1273 &packedpic_data_bufid);
1274 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1276 render_id[0] = packedpic_para_bufid;
1277 render_id[1] = packedpic_data_bufid;
1278 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1280 free(packedpic_buffer);
1285 void QuickSyncEncoderImpl::render_packedslice()
1287 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1288 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1289 unsigned int length_in_bits;
1290 unsigned char *packedslice_buffer = NULL;
1293 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1294 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1295 packedheader_param_buffer.bit_length = length_in_bits;
1296 packedheader_param_buffer.has_emulation_bytes = 0;
1298 va_status = vaCreateBuffer(va_dpy,
1300 VAEncPackedHeaderParameterBufferType,
1301 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1302 &packedslice_para_bufid);
1303 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1305 va_status = vaCreateBuffer(va_dpy,
1307 VAEncPackedHeaderDataBufferType,
1308 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1309 &packedslice_data_bufid);
1310 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1312 render_id[0] = packedslice_para_bufid;
1313 render_id[1] = packedslice_data_bufid;
1314 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1316 free(packedslice_buffer);
1319 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1321 VABufferID slice_param_buf;
1325 update_RefPicList(frame_type);
1327 /* one frame, one slice */
1328 slice_param.macroblock_address = 0;
1329 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1330 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1331 if (frame_type == FRAME_IDR) {
1332 if (encoding_frame_num != 0)
1333 ++slice_param.idr_pic_id;
1334 } else if (frame_type == FRAME_P) {
1335 int refpiclist0_max = h264_maxref & 0xffff;
1336 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1338 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1339 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1340 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1342 } else if (frame_type == FRAME_B) {
1343 int refpiclist0_max = h264_maxref & 0xffff;
1344 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1346 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1347 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1348 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1349 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1352 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1353 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1354 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1355 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1359 slice_param.slice_alpha_c0_offset_div2 = 0;
1360 slice_param.slice_beta_offset_div2 = 0;
1361 slice_param.direct_spatial_mv_pred_flag = 1;
1362 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1365 if (h264_packedheader &&
1366 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1367 render_packedslice();
1369 va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1370 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1371 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1373 render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1380 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1382 VACodedBufferSegment *buf_list = NULL;
1387 va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1388 CHECK_VASTATUS(va_status, "vaMapBuffer");
1389 while (buf_list != NULL) {
1390 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1391 buf_list = (VACodedBufferSegment *) buf_list->next;
1393 vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1395 static int frameno = 0;
1396 print_latency("Current QuickSync latency (video inputs → disk mux):",
1397 task.received_ts, (task.frame_type == FRAME_B), &frameno);
1402 memset(&pkt, 0, sizeof(pkt));
1404 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1405 pkt.size = data.size();
1406 pkt.stream_index = 0;
1407 if (task.frame_type == FRAME_IDR) {
1408 pkt.flags = AV_PKT_FLAG_KEY;
1412 pkt.duration = task.duration;
1414 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1416 if (!global_flags.uncompressed_video_to_http &&
1417 !global_flags.x264_video_to_http) {
1418 stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1424 // this is weird. but it seems to put a new frame onto the queue
1425 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1427 unique_lock<mutex> lock(storage_task_queue_mutex);
1428 storage_task_queue.push(move(task));
1429 storage_task_queue_changed.notify_all();
1432 void QuickSyncEncoderImpl::storage_task_thread()
1434 pthread_setname_np(pthread_self(), "QS_Storage");
1436 storage_task current;
1438 // wait until there's an encoded frame
1439 unique_lock<mutex> lock(storage_task_queue_mutex);
1440 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1441 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1442 current = move(storage_task_queue.front());
1443 storage_task_queue.pop();
1448 // waits for data, then saves it to disk.
1449 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1450 CHECK_VASTATUS(va_status, "vaSyncSurface");
1451 save_codeddata(move(current));
1454 unique_lock<mutex> lock(storage_task_queue_mutex);
1455 srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1456 storage_task_queue_changed.notify_all();
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++) {
1481 if (!use_zerocopy) {
1482 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1483 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1484 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1485 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1487 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1488 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1491 has_released_gl_resources = true;
1494 int QuickSyncEncoderImpl::deinit_va()
1496 vaTerminate(va_dpy);
1498 va_close_display(va_dpy);
1503 QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1504 : 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)
1506 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1507 open_output_file(filename);
1508 file_audio_encoder->add_mux(file_mux.get());
1510 frame_width_mbaligned = (frame_width + 15) & (~15);
1511 frame_height_mbaligned = (frame_height + 15) & (~15);
1515 if (global_flags.x264_video_to_http) {
1516 assert(x264_encoder != nullptr);
1518 assert(x264_encoder == nullptr);
1521 init_va(va_display);
1524 // No frames are ready yet.
1525 memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1527 memset(&seq_param, 0, sizeof(seq_param));
1528 memset(&pic_param, 0, sizeof(pic_param));
1529 memset(&slice_param, 0, sizeof(slice_param));
1531 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1533 encode_thread = thread([this]{
1534 //SDL_GL_MakeCurrent(window, context);
1535 QOpenGLContext *context = create_context(this->surface);
1536 eglBindAPI(EGL_OPENGL_API);
1537 if (!make_current(context, this->surface)) {
1538 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1542 encode_thread_func();
1543 delete_context(context);
1547 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1550 release_gl_resources();
1553 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1555 assert(!is_shutdown);
1557 // Wait until this frame slot is done encoding.
1558 unique_lock<mutex> lock(storage_task_queue_mutex);
1559 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1560 fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1561 current_storage_frame % SURFACE_NUM, current_storage_frame);
1563 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1564 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1565 if (storage_thread_should_quit) return false;
1568 //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1569 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1570 *y_tex = surf->y_tex;
1571 *cbcr_tex = surf->cbcr_tex;
1573 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1574 CHECK_VASTATUS(va_status, "vaDeriveImage");
1577 VABufferInfo buf_info;
1578 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1579 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1580 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1583 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1584 EGLint y_attribs[] = {
1585 EGL_WIDTH, frame_width,
1586 EGL_HEIGHT, frame_height,
1587 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1588 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1589 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1590 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1594 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1595 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1597 // Associate Y image to a texture.
1598 glBindTexture(GL_TEXTURE_2D, *y_tex);
1599 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1601 // Create CbCr image.
1602 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1603 EGLint cbcr_attribs[] = {
1604 EGL_WIDTH, frame_width,
1605 EGL_HEIGHT, frame_height,
1606 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1607 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1608 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1609 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1613 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1614 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1616 // Associate CbCr image to a texture.
1617 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1618 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1624 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1626 assert(!is_shutdown);
1627 file_audio_encoder->encode_audio(audio, pts + global_delay());
1630 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1632 assert(!is_shutdown);
1634 if (!use_zerocopy) {
1635 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1637 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1640 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1643 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1645 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1648 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1650 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1653 glBindTexture(GL_TEXTURE_2D, 0);
1655 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1658 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1662 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1664 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1668 unique_lock<mutex> lock(frame_queue_mutex);
1669 pending_video_frames.push(PendingFrame{ fence, input_frames, pts, duration });
1670 ++current_storage_frame;
1672 frame_queue_nonempty.notify_all();
1676 void QuickSyncEncoderImpl::shutdown()
1683 unique_lock<mutex> lock(frame_queue_mutex);
1684 encode_thread_should_quit = true;
1685 frame_queue_nonempty.notify_all();
1687 encode_thread.join();
1689 unique_lock<mutex> lock(storage_task_queue_mutex);
1690 storage_thread_should_quit = true;
1691 frame_queue_nonempty.notify_all();
1692 storage_task_queue_changed.notify_all();
1694 storage_thread.join();
1696 // Encode any leftover audio in the queues, and also any delayed frames.
1697 file_audio_encoder->encode_last_audio();
1705 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1707 AVFormatContext *avctx = avformat_alloc_context();
1708 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1709 assert(filename.size() < sizeof(avctx->filename) - 1);
1710 strcpy(avctx->filename, filename.c_str());
1712 string url = "file:" + filename;
1713 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1715 char tmp[AV_ERROR_MAX_STRING_SIZE];
1716 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1720 string video_extradata = ""; // FIXME: See other comment about global headers.
1721 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1722 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1723 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1726 void QuickSyncEncoderImpl::encode_thread_func()
1728 pthread_setname_np(pthread_self(), "QS_Encode");
1730 int64_t last_dts = -1;
1731 int gop_start_display_frame_num = 0;
1732 for (int display_frame_num = 0; ; ++display_frame_num) {
1733 // Wait for the frame to be in the queue. Note that this only means
1734 // we started rendering it.
1737 unique_lock<mutex> lock(frame_queue_mutex);
1738 frame_queue_nonempty.wait(lock, [this]{
1739 return encode_thread_should_quit || !pending_video_frames.empty();
1741 if (encode_thread_should_quit && pending_video_frames.empty()) {
1742 // We may have queued frames left in the reorder buffer
1743 // that were supposed to be B-frames, but have no P-frame
1744 // to be encoded against. If so, encode them all as
1745 // P-frames instead. Note that this happens under the mutex,
1746 // but nobody else uses it at this point, since we're shutting down,
1747 // so there's no contention.
1748 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1751 frame = move(pending_video_frames.front());
1752 pending_video_frames.pop();
1756 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1757 // Note that this implicitly waits for the frame to be done rendering.
1758 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1759 reorder_buffer[display_frame_num] = move(frame);
1761 // Now encode as many QuickSync frames as we can using the frames we have available.
1762 // (It could be zero, or it could be multiple.) FIXME: make a function.
1765 int frame_type, quicksync_display_frame_num;
1766 encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1767 &quicksync_display_frame_num, &frame_type, &pts_lag);
1768 if (!reorder_buffer.count(quicksync_display_frame_num)) {
1771 frame = move(reorder_buffer[quicksync_display_frame_num]);
1772 reorder_buffer.erase(quicksync_display_frame_num);
1774 if (frame_type == FRAME_IDR) {
1776 current_frame_num = 0;
1777 gop_start_display_frame_num = quicksync_display_frame_num;
1780 // Determine the dts of this frame.
1782 if (pts_lag == -1) {
1783 assert(last_dts != -1);
1784 dts = last_dts + (TIMEBASE / MAX_FPS);
1786 dts = frame.pts - pts_lag;
1790 encode_frame(frame, quicksync_encoding_frame_num, quicksync_display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
1791 ++quicksync_encoding_frame_num;
1796 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1798 if (reorder_buffer.empty()) {
1802 for (auto &pending_frame : reorder_buffer) {
1803 int display_frame_num = pending_frame.first;
1804 assert(display_frame_num > 0);
1805 PendingFrame frame = move(pending_frame.second);
1806 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1807 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1808 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
1813 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1816 memset(&pkt, 0, sizeof(pkt));
1818 pkt.data = const_cast<uint8_t *>(data);
1819 pkt.size = frame_width * frame_height * 2;
1820 pkt.stream_index = 0;
1821 pkt.flags = AV_PKT_FLAG_KEY;
1822 pkt.duration = duration;
1823 stream_mux->add_packet(pkt, pts, pts);
1828 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1830 if (src_width == dst_pitch) {
1831 memcpy(dst, src, src_width * height);
1833 for (size_t y = 0; y < height; ++y) {
1834 const uint8_t *sptr = src + y * src_width;
1835 uint8_t *dptr = dst + y * dst_pitch;
1836 memcpy(dptr, sptr, src_width);
1843 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1845 // Wait for the GPU to be done with the frame.
1848 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1850 } while (sync_status == GL_TIMEOUT_EXPIRED);
1851 assert(sync_status != GL_WAIT_FAILED);
1853 ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1854 static int frameno = 0;
1855 print_latency("Current mixer latency (video inputs → ready for encode):",
1856 received_ts, false, &frameno);
1858 // Release back any input frames we needed to render this frame.
1859 frame.input_frames.clear();
1861 GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1862 uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1863 if (global_flags.uncompressed_video_to_http) {
1864 add_packet_for_uncompressed_frame(pts, duration, data);
1865 } else if (global_flags.x264_video_to_http) {
1866 x264_encoder->add_frame(pts, duration, data, received_ts);
1870 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1871 int frame_type, int64_t pts, int64_t dts, int64_t duration)
1873 const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1875 GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
1879 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1880 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1881 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1882 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1884 // Upload the frame to VA-API.
1885 unsigned char *surface_p = nullptr;
1886 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
1888 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
1889 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
1891 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
1892 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
1894 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
1895 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
1898 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
1899 CHECK_VASTATUS(va_status, "vaDestroyImage");
1901 // Schedule the frame for encoding.
1902 VASurfaceID va_surface = surf->src_surface;
1903 va_status = vaBeginPicture(va_dpy, context_id, va_surface);
1904 CHECK_VASTATUS(va_status, "vaBeginPicture");
1906 if (frame_type == FRAME_IDR) {
1907 // FIXME: If the mux wants global headers, we should not put the
1908 // SPS/PPS before each IDR frame, but rather put it into the
1909 // codec extradata (formatted differently?).
1911 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1912 if (h264_packedheader) {
1913 render_packedsequence();
1914 render_packedpicture();
1917 //render_sequence();
1918 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
1920 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
1922 va_status = vaEndPicture(va_dpy, context_id);
1923 CHECK_VASTATUS(va_status, "vaEndPicture");
1925 // so now the data is done encoding (well, async job kicked off)...
1926 // we send that to the storage thread
1928 tmp.display_order = display_frame_num;
1929 tmp.frame_type = frame_type;
1932 tmp.duration = duration;
1933 tmp.received_ts = received_ts;
1934 storage_task_enqueue(move(tmp));
1936 update_ReferenceFrames(frame_type);
1940 QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator)
1941 : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
1943 // Must be defined here because unique_ptr<> destructor needs to know the impl.
1944 QuickSyncEncoder::~QuickSyncEncoder() {}
1946 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
1948 impl->add_audio(pts, audio);
1951 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1953 return impl->begin_frame(y_tex, cbcr_tex);
1956 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1958 return impl->end_frame(pts, duration, input_frames);
1961 void QuickSyncEncoder::shutdown()
1966 void QuickSyncEncoder::set_stream_mux(Mux *mux)
1968 impl->set_stream_mux(mux);
1971 int64_t QuickSyncEncoder::global_delay() const {
1972 return impl->global_delay();