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;
67 #define CHECK_VASTATUS(va_status, func) \
68 if (va_status != VA_STATUS_SUCCESS) { \
69 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
73 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
75 //#include "loadsurface.h"
77 #define NAL_REF_IDC_NONE 0
78 #define NAL_REF_IDC_LOW 1
79 #define NAL_REF_IDC_MEDIUM 2
80 #define NAL_REF_IDC_HIGH 3
88 #define SLICE_TYPE_P 0
89 #define SLICE_TYPE_B 1
90 #define SLICE_TYPE_I 2
91 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
92 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
93 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
96 #define ENTROPY_MODE_CAVLC 0
97 #define ENTROPY_MODE_CABAC 1
99 #define PROFILE_IDC_BASELINE 66
100 #define PROFILE_IDC_MAIN 77
101 #define PROFILE_IDC_HIGH 100
103 #define BITSTREAM_ALLOCATE_STEPPING 4096
105 static constexpr unsigned int MaxFrameNum = (2<<16);
106 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
107 static constexpr unsigned int Log2MaxFrameNum = 16;
108 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
109 static constexpr int rc_default_modes[] = { // Priority list of modes.
112 VA_RC_VBR_CONSTRAINED,
120 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
121 // but if we don't delete it here, we get leaks. The GStreamer implementation
123 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
125 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
126 CHECK_VASTATUS(va_status, "vaRenderPicture");
128 for (int i = 0; i < num_buffers; ++i) {
129 va_status = vaDestroyBuffer(dpy, buffers[i]);
130 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
135 va_swap32(unsigned int val)
137 unsigned char *pval = (unsigned char *)&val;
139 return ((pval[0] << 24) |
146 bitstream_start(bitstream *bs)
148 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
149 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
154 bitstream_end(bitstream *bs)
156 int pos = (bs->bit_offset >> 5);
157 int bit_offset = (bs->bit_offset & 0x1f);
158 int bit_left = 32 - bit_offset;
161 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
166 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
168 int pos = (bs->bit_offset >> 5);
169 int bit_offset = (bs->bit_offset & 0x1f);
170 int bit_left = 32 - bit_offset;
175 bs->bit_offset += size_in_bits;
177 if (bit_left > size_in_bits) {
178 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
180 size_in_bits -= bit_left;
181 if (bit_left >= 32) {
182 bs->buffer[pos] = (val >> size_in_bits);
184 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
186 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
188 if (pos + 1 == bs->max_size_in_dword) {
189 bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
190 bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
193 bs->buffer[pos + 1] = val;
198 bitstream_put_ue(bitstream *bs, unsigned int val)
200 int size_in_bits = 0;
208 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
209 bitstream_put_ui(bs, val, size_in_bits);
213 bitstream_put_se(bitstream *bs, int val)
215 unsigned int new_val;
220 new_val = 2 * val - 1;
222 bitstream_put_ue(bs, new_val);
226 bitstream_byte_aligning(bitstream *bs, int bit)
228 int bit_offset = (bs->bit_offset & 0x7);
229 int bit_left = 8 - bit_offset;
235 assert(bit == 0 || bit == 1);
238 new_val = (1 << bit_left) - 1;
242 bitstream_put_ui(bs, new_val, bit_left);
246 rbsp_trailing_bits(bitstream *bs)
248 bitstream_put_ui(bs, 1, 1);
249 bitstream_byte_aligning(bs, 0);
252 static void nal_start_code_prefix(bitstream *bs)
254 bitstream_put_ui(bs, 0x00000001, 32);
257 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
259 bitstream_put_ui(bs, 0, 1); /* forbidden_zero_bit: 0 */
260 bitstream_put_ui(bs, nal_ref_idc, 2);
261 bitstream_put_ui(bs, nal_unit_type, 5);
264 void QuickSyncEncoderImpl::sps_rbsp(YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs)
266 int profile_idc = PROFILE_IDC_BASELINE;
268 if (h264_profile == VAProfileH264High)
269 profile_idc = PROFILE_IDC_HIGH;
270 else if (h264_profile == VAProfileH264Main)
271 profile_idc = PROFILE_IDC_MAIN;
273 bitstream_put_ui(bs, profile_idc, 8); /* profile_idc */
274 bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1); /* constraint_set0_flag */
275 bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1); /* constraint_set1_flag */
276 bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1); /* constraint_set2_flag */
277 bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1); /* constraint_set3_flag */
278 bitstream_put_ui(bs, 0, 4); /* reserved_zero_4bits */
279 bitstream_put_ui(bs, seq_param.level_idc, 8); /* level_idc */
280 bitstream_put_ue(bs, seq_param.seq_parameter_set_id); /* seq_parameter_set_id */
282 if ( profile_idc == PROFILE_IDC_HIGH) {
283 bitstream_put_ue(bs, 1); /* chroma_format_idc = 1, 4:2:0 */
284 bitstream_put_ue(bs, 0); /* bit_depth_luma_minus8 */
285 bitstream_put_ue(bs, 0); /* bit_depth_chroma_minus8 */
286 bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
287 bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
290 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
291 bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type); /* pic_order_cnt_type */
293 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
294 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4); /* log2_max_pic_order_cnt_lsb_minus4 */
299 bitstream_put_ue(bs, seq_param.max_num_ref_frames); /* num_ref_frames */
300 bitstream_put_ui(bs, 0, 1); /* gaps_in_frame_num_value_allowed_flag */
302 bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1); /* pic_width_in_mbs_minus1 */
303 bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
304 bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1); /* frame_mbs_only_flag */
306 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
310 bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1); /* direct_8x8_inference_flag */
311 bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1); /* frame_cropping_flag */
313 if (seq_param.frame_cropping_flag) {
314 bitstream_put_ue(bs, seq_param.frame_crop_left_offset); /* frame_crop_left_offset */
315 bitstream_put_ue(bs, seq_param.frame_crop_right_offset); /* frame_crop_right_offset */
316 bitstream_put_ue(bs, seq_param.frame_crop_top_offset); /* frame_crop_top_offset */
317 bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset); /* frame_crop_bottom_offset */
320 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
322 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
324 // See H.264 annex E for the definition of this header.
325 bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
326 bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
327 bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
328 bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
330 bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
331 bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
332 bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
334 bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
335 bitstream_put_ui(bs, 2, 8); /* transfer_characteristics (2 = unspecified, since we use sRGB) */
336 if (ycbcr_coefficients == YCBCR_REC_709) {
337 bitstream_put_ui(bs, 1, 8); /* matrix_coefficients (1 = BT.709) */
339 assert(ycbcr_coefficients == YCBCR_REC_601);
340 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
344 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
345 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
347 bitstream_put_ui(bs, 1, 32); // FPS
348 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
349 bitstream_put_ui(bs, 1, 1);
351 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
354 bitstream_put_ue(bs, 0); /* cpb_cnt_minus1 */
355 bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
356 bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
358 bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
359 bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
360 bitstream_put_ui(bs, 1, 1); /* cbr_flag[0] */
362 bitstream_put_ui(bs, 23, 5); /* initial_cpb_removal_delay_length_minus1 */
363 bitstream_put_ui(bs, 23, 5); /* cpb_removal_delay_length_minus1 */
364 bitstream_put_ui(bs, 23, 5); /* dpb_output_delay_length_minus1 */
365 bitstream_put_ui(bs, 23, 5); /* time_offset_length */
367 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
368 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
370 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
371 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
374 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
378 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
380 bitstream_put_ue(bs, pic_param.pic_parameter_set_id); /* pic_parameter_set_id */
381 bitstream_put_ue(bs, pic_param.seq_parameter_set_id); /* seq_parameter_set_id */
383 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
385 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
387 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
389 bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1); /* num_ref_idx_l0_active_minus1 */
390 bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1); /* num_ref_idx_l1_active_minus1 1 */
392 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1); /* weighted_pred_flag: 0 */
393 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2); /* weighted_bipred_idc: 0 */
395 bitstream_put_se(bs, pic_param.pic_init_qp - 26); /* pic_init_qp_minus26 */
396 bitstream_put_se(bs, 0); /* pic_init_qs_minus26 */
397 bitstream_put_se(bs, 0); /* chroma_qp_index_offset */
399 bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
400 bitstream_put_ui(bs, 0, 1); /* constrained_intra_pred_flag */
401 bitstream_put_ui(bs, 0, 1); /* redundant_pic_cnt_present_flag */
404 bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1); /*transform_8x8_mode_flag */
405 bitstream_put_ui(bs, 0, 1); /* pic_scaling_matrix_present_flag */
406 bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset ); /*second_chroma_qp_index_offset */
408 rbsp_trailing_bits(bs);
411 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
413 int first_mb_in_slice = slice_param.macroblock_address;
415 bitstream_put_ue(bs, first_mb_in_slice); /* first_mb_in_slice: 0 */
416 bitstream_put_ue(bs, slice_param.slice_type); /* slice_type */
417 bitstream_put_ue(bs, slice_param.pic_parameter_set_id); /* pic_parameter_set_id: 0 */
418 bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
420 /* frame_mbs_only_flag == 1 */
421 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
426 if (pic_param.pic_fields.bits.idr_pic_flag)
427 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
429 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
430 bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
431 /* pic_order_present_flag == 0 */
437 /* redundant_pic_cnt_present_flag == 0 */
439 if (IS_P_SLICE(slice_param.slice_type)) {
440 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
442 if (slice_param.num_ref_idx_active_override_flag)
443 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
445 /* ref_pic_list_reordering */
446 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
447 } else if (IS_B_SLICE(slice_param.slice_type)) {
448 bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1); /* direct_spatial_mv_pred: 1 */
450 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
452 if (slice_param.num_ref_idx_active_override_flag) {
453 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
454 bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
457 /* ref_pic_list_reordering */
458 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
459 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l1: 0 */
462 if ((pic_param.pic_fields.bits.weighted_pred_flag &&
463 IS_P_SLICE(slice_param.slice_type)) ||
464 ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
465 IS_B_SLICE(slice_param.slice_type))) {
466 /* FIXME: fill weight/offset table */
470 /* dec_ref_pic_marking */
471 if (pic_param.pic_fields.bits.reference_pic_flag) { /* nal_ref_idc != 0 */
472 unsigned char no_output_of_prior_pics_flag = 0;
473 unsigned char long_term_reference_flag = 0;
474 unsigned char adaptive_ref_pic_marking_mode_flag = 0;
476 if (pic_param.pic_fields.bits.idr_pic_flag) {
477 bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1); /* no_output_of_prior_pics_flag: 0 */
478 bitstream_put_ui(bs, long_term_reference_flag, 1); /* long_term_reference_flag: 0 */
480 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
484 if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
485 !IS_I_SLICE(slice_param.slice_type))
486 bitstream_put_ue(bs, slice_param.cabac_init_idc); /* cabac_init_idc: 0 */
488 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
490 /* ignore for SP/SI */
492 if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
493 bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc); /* disable_deblocking_filter_idc: 0 */
495 if (slice_param.disable_deblocking_filter_idc != 1) {
496 bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2); /* slice_alpha_c0_offset_div2: 2 */
497 bitstream_put_se(bs, slice_param.slice_beta_offset_div2); /* slice_beta_offset_div2: 2 */
501 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
502 bitstream_byte_aligning(bs, 1);
506 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
510 bitstream_start(&bs);
511 nal_start_code_prefix(&bs);
512 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
516 *header_buffer = (unsigned char *)bs.buffer;
517 return bs.bit_offset;
521 QuickSyncEncoderImpl::build_packed_seq_buffer(YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer)
525 bitstream_start(&bs);
526 nal_start_code_prefix(&bs);
527 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
528 sps_rbsp(ycbcr_coefficients, &bs);
531 *header_buffer = (unsigned char *)bs.buffer;
532 return bs.bit_offset;
535 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
538 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
539 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
541 bitstream_start(&bs);
542 nal_start_code_prefix(&bs);
544 if (IS_I_SLICE(slice_param.slice_type)) {
545 nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
546 } else if (IS_P_SLICE(slice_param.slice_type)) {
547 nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
549 assert(IS_B_SLICE(slice_param.slice_type));
550 nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
556 *header_buffer = (unsigned char *)bs.buffer;
557 return bs.bit_offset;
562 Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
563 1) period between Frame #X and Frame #N = #X - #N
564 2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
565 3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
566 4) intra_period and intra_idr_period take precedence over ip_period
567 5) if ip_period > 1, intra_period and intra_idr_period are not the strict periods
568 of I/IDR frames, see bellow examples
569 -------------------------------------------------------------------
570 intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
571 0 ignored 1 IDRPPPPPPP ... (No IDR/I any more)
572 0 ignored >=2 IDR(PBB)(PBB)... (No IDR/I any more)
573 1 0 ignored IDRIIIIIII... (No IDR any more)
574 1 1 ignored IDR IDR IDR IDR...
575 1 >=2 ignored IDRII IDRII IDR... (1/3/ignore)
576 >=2 0 1 IDRPPP IPPP I... (3/0/1)
577 >=2 0 >=2 IDR(PBB)(PBB)(IBB) (6/0/3)
578 (PBB)(IBB)(PBB)(IBB)...
579 >=2 >=2 1 IDRPPPPP IPPPPP IPPPPP (6/18/1)
580 IDRPPPPP IPPPPP IPPPPP...
581 >=2 >=2 >=2 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
582 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
583 {IDR(PBB)(PBB)(IBB)(PBB)} (6/12/3)
584 {IDR(PBB)(PBB)(IBB)(PBB)}...
585 {IDR(PBB)(PBB)} (6/6/3)
589 // General pts/dts strategy:
591 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
592 // bit tricky. We assume first of all that the frame rate never goes _above_
593 // MAX_FPS, which gives us a frame period N. The decoder can always decode
594 // in at least this speed, as long at dts <= pts (the frame is not attempted
595 // presented before it is decoded). Furthermore, we never have longer chains of
596 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
597 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
598 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
599 // frame, with an order O <= C.)
601 // Many strategies are possible, but we establish these rules:
603 // - Tip frames have dts = pts - (C-O)*N.
604 // - Non-tip frames have dts = dts_last + N.
606 // An example, with C=2 and N=10 and the data flow showed with arrows:
609 // pts: 30 40 50 60 70 80
611 // dts: 10 30 20 60 50←40
616 // To show that this works fine also with irregular spacings, let's say that
617 // the third frame is delayed a bit (something earlier was dropped). Now the
618 // situation looks like this:
621 // pts: 30 40 80 90 100 110
623 // dts: 10 30 20 90 50←40
628 // The resetting on every tip frame makes sure dts never ends up lagging a lot
629 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
631 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
632 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
633 // a tip frame and should be given a dts based on the previous one.
638 void encoding2display_order(
639 int encoding_order, int intra_period,
640 int intra_idr_period, int ip_period,
641 int *displaying_order,
642 int *frame_type, int *pts_lag)
644 int encoding_order_gop = 0;
648 if (intra_period == 1) { /* all are I/IDR frames */
649 *displaying_order = encoding_order;
650 if (intra_idr_period == 0)
651 *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
653 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
657 if (intra_period == 0)
658 intra_idr_period = 0;
660 if (ip_period == 1) {
661 // No B-frames, sequence is like IDR PPPPP IPPPPP.
662 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
663 *displaying_order = encoding_order;
665 if (encoding_order_gop == 0) { /* the first frame */
666 *frame_type = FRAME_IDR;
667 } else if (intra_period != 0 && /* have I frames */
668 encoding_order_gop >= 2 &&
669 (encoding_order_gop % intra_period == 0)) {
670 *frame_type = FRAME_I;
672 *frame_type = FRAME_P;
677 // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
678 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
679 *pts_lag = -1; // Most frames are not tip frames.
681 if (encoding_order_gop == 0) { /* the first frame */
682 *frame_type = FRAME_IDR;
683 *displaying_order = encoding_order;
684 // IDR frames are a special case; I honestly can't find the logic behind
685 // why this is the right thing, but it seems to line up nicely in practice :-)
686 *pts_lag = TIMEBASE / MAX_FPS;
687 } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
688 *frame_type = FRAME_B;
689 *displaying_order = encoding_order - 1;
690 if ((encoding_order_gop % ip_period) == 0) {
691 *pts_lag = 0; // Last B-frame.
693 } else if (intra_period != 0 && /* have I frames */
694 encoding_order_gop >= 2 &&
695 ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
696 *frame_type = FRAME_I;
697 *displaying_order = encoding_order + ip_period - 1;
699 *frame_type = FRAME_P;
700 *displaying_order = encoding_order + ip_period - 1;
705 static const char *rc_to_string(int rc_mode)
718 case VA_RC_VBR_CONSTRAINED:
719 return "VBR_CONSTRAINED";
725 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
727 if (global_flags.x264_video_to_disk) {
728 // Quick Sync is entirely disabled.
729 use_zerocopy = false;
730 } else if (global_flags.uncompressed_video_to_http) {
731 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
732 use_zerocopy = false;
733 } else if (global_flags.x264_video_to_http) {
734 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
735 use_zerocopy = false;
741 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
743 if (va_display.empty()) {
744 x11_display = XOpenDisplay(NULL);
746 fprintf(stderr, "error: can't connect to X server!\n");
749 return vaGetDisplay(x11_display);
750 } else if (va_display[0] != '/') {
751 x11_display = XOpenDisplay(va_display.c_str());
753 fprintf(stderr, "error: can't connect to X server!\n");
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()
933 if (!global_flags.x264_video_to_disk) {
935 VASurfaceID *tmp_surfaceid;
937 VASurfaceID src_surface[SURFACE_NUM];
938 VASurfaceID ref_surface[SURFACE_NUM];
940 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
941 &config_attrib[0], config_attrib_num, &config_id);
942 CHECK_VASTATUS(va_status, "vaCreateConfig");
944 /* create source surfaces */
945 va_status = vaCreateSurfaces(va_dpy,
946 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
947 &src_surface[0], SURFACE_NUM,
949 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
951 /* create reference surfaces */
952 va_status = vaCreateSurfaces(va_dpy,
953 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
954 &ref_surface[0], SURFACE_NUM,
956 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
958 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
959 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
960 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
962 for (int i = 0; i < SURFACE_NUM; i++) {
963 gl_surfaces[i].src_surface = src_surface[i];
964 gl_surfaces[i].ref_surface = ref_surface[i];
967 /* Create a context for this encode pipe */
968 va_status = vaCreateContext(va_dpy, config_id,
969 frame_width_mbaligned, frame_height_mbaligned,
971 tmp_surfaceid, 2 * SURFACE_NUM,
973 CHECK_VASTATUS(va_status, "vaCreateContext");
976 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
978 for (int i = 0; i < SURFACE_NUM; i++) {
979 /* create coded buffer once for all
980 * other VA buffers which won't be used again after vaRenderPicture.
981 * so APP can always vaCreateBuffer for every frame
982 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
983 * so VA won't maintain the coded buffer
985 va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
986 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
987 CHECK_VASTATUS(va_status, "vaCreateBuffer");
991 /* create OpenGL objects */
992 for (int i = 0; i < SURFACE_NUM; i++) {
994 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
995 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
997 size_t bytes_per_pixel;
998 if (global_flags.x264_bit_depth > 8) {
1000 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R16, frame_width, frame_height);
1001 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG16, frame_width / 2, frame_height / 2);
1003 bytes_per_pixel = 1;
1004 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1005 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1008 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1009 // buffers, due to potentially differing pitch.
1010 glGenBuffers(1, &gl_surfaces[i].pbo);
1011 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1012 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);
1013 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);
1014 gl_surfaces[i].y_offset = 0;
1015 gl_surfaces[i].cbcr_offset = frame_width * frame_height * bytes_per_pixel;
1016 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1017 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1018 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1025 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1028 template<class T, class C>
1029 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1031 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1032 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1033 sort(middle, end, less_than);
1036 void QuickSyncEncoderImpl::update_ReferenceFrames(int current_display_frame, int frame_type)
1038 if (frame_type == FRAME_B)
1041 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1043 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1044 unique_lock<mutex> lock(storage_task_queue_mutex);
1046 // Insert the new frame at the start of the reference queue.
1047 reference_frames.push_front(ReferenceFrame{ CurrentCurrPic, current_display_frame });
1049 if (reference_frames.size() > num_ref_frames)
1051 // The back frame frame is no longer in use as a reference.
1052 int display_frame_num = reference_frames.back().display_number;
1053 assert(surface_for_frame.count(display_frame_num));
1054 release_gl_surface(display_frame_num);
1055 reference_frames.pop_back();
1058 // Mark this frame in use as a reference.
1059 assert(surface_for_frame.count(current_display_frame));
1060 ++surface_for_frame[current_display_frame]->refcount;
1062 current_ref_frame_num++;
1063 if (current_ref_frame_num > MaxFrameNum)
1064 current_ref_frame_num = 0;
1068 void QuickSyncEncoderImpl::update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2])
1070 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1071 return a.frame_idx > b.frame_idx;
1074 for (size_t i = 0; i < reference_frames.size(); ++i) {
1075 RefPicList0_P[i] = reference_frames[i].pic;
1077 sort(&RefPicList0_P[0], &RefPicList0_P[reference_frames.size()], descending_by_frame_idx);
1080 void QuickSyncEncoderImpl::update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2])
1082 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1083 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1085 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1086 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1089 for (size_t i = 0; i < reference_frames.size(); ++i) {
1090 RefPicList0_B[i] = reference_frames[i].pic;
1091 RefPicList1_B[i] = reference_frames[i].pic;
1093 sort_two(&RefPicList0_B[0], &RefPicList0_B[reference_frames.size()], CurrentCurrPic, ascending_by_top_field_order_cnt);
1094 sort_two(&RefPicList1_B[0], &RefPicList1_B[reference_frames.size()], CurrentCurrPic, descending_by_top_field_order_cnt);
1098 int QuickSyncEncoderImpl::render_sequence()
1100 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1102 VAEncMiscParameterBuffer *misc_param;
1103 VAEncMiscParameterRateControl *misc_rate_ctrl;
1105 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1106 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1107 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1108 seq_param.bits_per_second = frame_bitrate;
1110 seq_param.intra_period = intra_period;
1111 seq_param.intra_idr_period = intra_idr_period;
1112 seq_param.ip_period = ip_period;
1114 seq_param.max_num_ref_frames = num_ref_frames;
1115 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1116 seq_param.time_scale = TIMEBASE * 2;
1117 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1118 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1119 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1120 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1121 seq_param.seq_fields.bits.chroma_format_idc = 1;
1122 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1124 if (frame_width != frame_width_mbaligned ||
1125 frame_height != frame_height_mbaligned) {
1126 seq_param.frame_cropping_flag = 1;
1127 seq_param.frame_crop_left_offset = 0;
1128 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1129 seq_param.frame_crop_top_offset = 0;
1130 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1133 va_status = vaCreateBuffer(va_dpy, context_id,
1134 VAEncSequenceParameterBufferType,
1135 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1136 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1138 va_status = vaCreateBuffer(va_dpy, context_id,
1139 VAEncMiscParameterBufferType,
1140 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1141 1, NULL, &rc_param_buf);
1142 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1144 vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1145 misc_param->type = VAEncMiscParameterTypeRateControl;
1146 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1147 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1148 misc_rate_ctrl->bits_per_second = frame_bitrate;
1149 misc_rate_ctrl->target_percentage = 66;
1150 misc_rate_ctrl->window_size = 1000;
1151 misc_rate_ctrl->initial_qp = initial_qp;
1152 misc_rate_ctrl->min_qp = minimal_qp;
1153 misc_rate_ctrl->basic_unit_size = 0;
1154 vaUnmapBuffer(va_dpy, rc_param_buf);
1156 render_id[0] = seq_param_buf;
1157 render_id[1] = rc_param_buf;
1159 render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1164 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1166 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1167 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1168 int PicOrderCntMsb, TopFieldOrderCnt;
1170 if (frame_type == FRAME_IDR)
1171 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1173 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1174 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1177 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1178 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1179 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1180 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1181 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1182 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1184 PicOrderCntMsb = prevPicOrderCntMsb;
1186 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1188 if (frame_type != FRAME_B) {
1189 PicOrderCntMsb_ref = PicOrderCntMsb;
1190 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1193 return TopFieldOrderCnt;
1196 int QuickSyncEncoderImpl::render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num)
1198 VABufferID pic_param_buf;
1202 pic_param.CurrPic.picture_id = surf->ref_surface;
1203 pic_param.CurrPic.frame_idx = current_ref_frame_num;
1204 pic_param.CurrPic.flags = 0;
1205 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1206 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1207 CurrentCurrPic = pic_param.CurrPic;
1209 for (i = 0; i < reference_frames.size(); i++) {
1210 pic_param.ReferenceFrames[i] = reference_frames[i].pic;
1212 for (i = reference_frames.size(); i < MAX_NUM_REF1; i++) {
1213 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1214 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1217 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1218 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1219 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1220 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1221 pic_param.frame_num = current_ref_frame_num; // FIXME: is this correct?
1222 pic_param.coded_buf = surf->coded_buf;
1223 pic_param.last_picture = false; // FIXME
1224 pic_param.pic_init_qp = initial_qp;
1226 va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1227 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1228 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1230 render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1235 int QuickSyncEncoderImpl::render_packedsequence(YCbCrLumaCoefficients ycbcr_coefficients)
1237 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1238 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1239 unsigned int length_in_bits;
1240 unsigned char *packedseq_buffer = NULL;
1243 length_in_bits = build_packed_seq_buffer(ycbcr_coefficients, &packedseq_buffer);
1245 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1247 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1248 packedheader_param_buffer.has_emulation_bytes = 0;
1249 va_status = vaCreateBuffer(va_dpy,
1251 VAEncPackedHeaderParameterBufferType,
1252 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1253 &packedseq_para_bufid);
1254 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1256 va_status = vaCreateBuffer(va_dpy,
1258 VAEncPackedHeaderDataBufferType,
1259 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1260 &packedseq_data_bufid);
1261 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1263 render_id[0] = packedseq_para_bufid;
1264 render_id[1] = packedseq_data_bufid;
1265 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1267 free(packedseq_buffer);
1273 int QuickSyncEncoderImpl::render_packedpicture()
1275 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1276 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1277 unsigned int length_in_bits;
1278 unsigned char *packedpic_buffer = NULL;
1281 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1282 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1283 packedheader_param_buffer.bit_length = length_in_bits;
1284 packedheader_param_buffer.has_emulation_bytes = 0;
1286 va_status = vaCreateBuffer(va_dpy,
1288 VAEncPackedHeaderParameterBufferType,
1289 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1290 &packedpic_para_bufid);
1291 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1293 va_status = vaCreateBuffer(va_dpy,
1295 VAEncPackedHeaderDataBufferType,
1296 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1297 &packedpic_data_bufid);
1298 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1300 render_id[0] = packedpic_para_bufid;
1301 render_id[1] = packedpic_data_bufid;
1302 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1304 free(packedpic_buffer);
1309 void QuickSyncEncoderImpl::render_packedslice()
1311 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1312 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1313 unsigned int length_in_bits;
1314 unsigned char *packedslice_buffer = NULL;
1317 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1318 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1319 packedheader_param_buffer.bit_length = length_in_bits;
1320 packedheader_param_buffer.has_emulation_bytes = 0;
1322 va_status = vaCreateBuffer(va_dpy,
1324 VAEncPackedHeaderParameterBufferType,
1325 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1326 &packedslice_para_bufid);
1327 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1329 va_status = vaCreateBuffer(va_dpy,
1331 VAEncPackedHeaderDataBufferType,
1332 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1333 &packedslice_data_bufid);
1334 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1336 render_id[0] = packedslice_para_bufid;
1337 render_id[1] = packedslice_data_bufid;
1338 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1340 free(packedslice_buffer);
1343 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1345 VABufferID slice_param_buf;
1349 /* one frame, one slice */
1350 slice_param.macroblock_address = 0;
1351 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1352 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1353 if (frame_type == FRAME_IDR) {
1354 if (encoding_frame_num != 0)
1355 ++slice_param.idr_pic_id;
1356 } else if (frame_type == FRAME_P) {
1357 VAPictureH264 RefPicList0_P[MAX_NUM_REF2];
1358 update_RefPicList_P(RefPicList0_P);
1360 int refpiclist0_max = h264_maxref & 0xffff;
1361 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1363 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1364 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1365 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1367 } else if (frame_type == FRAME_B) {
1368 VAPictureH264 RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
1369 update_RefPicList_B(RefPicList0_B, RefPicList1_B);
1371 int refpiclist0_max = h264_maxref & 0xffff;
1372 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1374 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1375 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1376 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1377 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1380 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1381 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1382 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1383 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1387 slice_param.slice_alpha_c0_offset_div2 = 0;
1388 slice_param.slice_beta_offset_div2 = 0;
1389 slice_param.direct_spatial_mv_pred_flag = 1;
1390 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1393 if (h264_packedheader &&
1394 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1395 render_packedslice();
1397 va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1398 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1399 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1401 render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1408 void QuickSyncEncoderImpl::save_codeddata(GLSurface *surf, storage_task task)
1410 VACodedBufferSegment *buf_list = NULL;
1415 va_status = vaMapBuffer(va_dpy, surf->coded_buf, (void **)(&buf_list));
1416 CHECK_VASTATUS(va_status, "vaMapBuffer");
1417 while (buf_list != NULL) {
1418 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1419 buf_list = (VACodedBufferSegment *) buf_list->next;
1421 vaUnmapBuffer(va_dpy, surf->coded_buf);
1423 static int frameno = 0;
1424 print_latency("Current QuickSync latency (video inputs → disk mux):",
1425 task.received_ts, (task.frame_type == FRAME_B), &frameno);
1430 memset(&pkt, 0, sizeof(pkt));
1432 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1433 pkt.size = data.size();
1434 pkt.stream_index = 0;
1435 if (task.frame_type == FRAME_IDR) {
1436 pkt.flags = AV_PKT_FLAG_KEY;
1440 pkt.duration = task.duration;
1442 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1444 if (!global_flags.uncompressed_video_to_http &&
1445 !global_flags.x264_video_to_http) {
1446 stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1452 // this is weird. but it seems to put a new frame onto the queue
1453 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1455 unique_lock<mutex> lock(storage_task_queue_mutex);
1456 storage_task_queue.push(move(task));
1457 storage_task_queue_changed.notify_all();
1460 void QuickSyncEncoderImpl::storage_task_thread()
1462 pthread_setname_np(pthread_self(), "QS_Storage");
1464 storage_task current;
1467 // wait until there's an encoded frame
1468 unique_lock<mutex> lock(storage_task_queue_mutex);
1469 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1470 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1471 current = move(storage_task_queue.front());
1472 storage_task_queue.pop();
1473 surf = surface_for_frame[current.display_order];
1474 assert(surf != nullptr);
1479 size_t display_order = current.display_order;
1480 vector<size_t> ref_display_frame_numbers = move(current.ref_display_frame_numbers);
1482 // waits for data, then saves it to disk.
1483 va_status = vaSyncSurface(va_dpy, surf->src_surface);
1484 CHECK_VASTATUS(va_status, "vaSyncSurface");
1485 save_codeddata(surf, move(current));
1487 // Unlock the frame, and all its references.
1489 unique_lock<mutex> lock(storage_task_queue_mutex);
1490 release_gl_surface(display_order);
1492 for (size_t frame_num : ref_display_frame_numbers) {
1493 release_gl_surface(frame_num);
1499 void QuickSyncEncoderImpl::release_encode()
1501 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1502 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1503 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1504 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1507 vaDestroyContext(va_dpy, context_id);
1508 vaDestroyConfig(va_dpy, config_id);
1511 void QuickSyncEncoderImpl::release_gl_resources()
1513 assert(is_shutdown);
1514 if (has_released_gl_resources) {
1518 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1519 if (!use_zerocopy) {
1520 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1521 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1522 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1523 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1525 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1526 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1529 has_released_gl_resources = true;
1532 int QuickSyncEncoderImpl::deinit_va()
1534 vaTerminate(va_dpy);
1536 va_close_display(va_dpy);
1541 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)
1542 : 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)
1544 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1545 open_output_file(filename);
1546 file_audio_encoder->add_mux(file_mux.get());
1548 frame_width_mbaligned = (frame_width + 15) & (~15);
1549 frame_height_mbaligned = (frame_height + 15) & (~15);
1553 if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1554 assert(x264_encoder != nullptr);
1556 assert(x264_encoder == nullptr);
1559 enable_zerocopy_if_possible();
1560 if (!global_flags.x264_video_to_disk) {
1561 init_va(va_display);
1565 if (!global_flags.x264_video_to_disk) {
1566 memset(&seq_param, 0, sizeof(seq_param));
1567 memset(&pic_param, 0, sizeof(pic_param));
1568 memset(&slice_param, 0, sizeof(slice_param));
1571 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1573 encode_thread = thread([this]{
1574 QOpenGLContext *context = create_context(this->surface);
1575 eglBindAPI(EGL_OPENGL_API);
1576 if (!make_current(context, this->surface)) {
1577 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1581 encode_thread_func();
1582 delete_context(context);
1586 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1589 release_gl_resources();
1592 QuickSyncEncoderImpl::GLSurface *QuickSyncEncoderImpl::allocate_gl_surface()
1594 for (unsigned i = 0; i < SURFACE_NUM; ++i) {
1595 if (gl_surfaces[i].refcount == 0) {
1596 ++gl_surfaces[i].refcount;
1597 return &gl_surfaces[i];
1603 void QuickSyncEncoderImpl::release_gl_surface(size_t display_frame_num)
1605 assert(surface_for_frame.count(display_frame_num));
1606 QuickSyncEncoderImpl::GLSurface *surf = surface_for_frame[display_frame_num];
1607 if (--surf->refcount == 0) {
1608 assert(surface_for_frame.count(display_frame_num));
1609 surface_for_frame.erase(display_frame_num);
1610 storage_task_queue_changed.notify_all();
1614 bool QuickSyncEncoderImpl::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
1616 assert(!is_shutdown);
1617 GLSurface *surf = nullptr;
1619 // Wait until this frame slot is done encoding.
1620 unique_lock<mutex> lock(storage_task_queue_mutex);
1621 surf = allocate_gl_surface();
1622 if (surf == nullptr) {
1623 fprintf(stderr, "Warning: No free slots for frame %d, rendering has to wait for H.264 encoder\n",
1624 current_storage_frame);
1625 storage_task_queue_changed.wait(lock, [this, &surf]{
1626 if (storage_thread_should_quit)
1628 surf = allocate_gl_surface();
1629 return surf != nullptr;
1632 if (storage_thread_should_quit) return false;
1633 assert(surf != nullptr);
1634 surface_for_frame[current_storage_frame] = surf;
1637 *y_tex = surf->y_tex;
1638 *cbcr_tex = surf->cbcr_tex;
1640 if (!global_flags.x264_video_to_disk) {
1641 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1642 CHECK_VASTATUS(va_status, "vaDeriveImage");
1645 VABufferInfo buf_info;
1646 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1647 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1648 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1651 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1652 EGLint y_attribs[] = {
1653 EGL_WIDTH, frame_width,
1654 EGL_HEIGHT, frame_height,
1655 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1656 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1657 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1658 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1662 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1663 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1665 // Associate Y image to a texture.
1666 glBindTexture(GL_TEXTURE_2D, *y_tex);
1667 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1669 // Create CbCr image.
1670 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1671 EGLint cbcr_attribs[] = {
1672 EGL_WIDTH, frame_width,
1673 EGL_HEIGHT, frame_height,
1674 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1675 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1676 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1677 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1681 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1682 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1684 // Associate CbCr image to a texture.
1685 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1686 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1690 current_video_frame = PendingFrame{ {}, input_frames, pts, duration, ycbcr_coefficients };
1695 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1697 assert(!is_shutdown);
1698 file_audio_encoder->encode_audio(audio, pts + global_delay());
1701 RefCountedGLsync QuickSyncEncoderImpl::end_frame()
1703 assert(!is_shutdown);
1705 if (!use_zerocopy) {
1706 GLenum type = global_flags.x264_bit_depth > 8 ? GL_UNSIGNED_SHORT : GL_UNSIGNED_BYTE;
1709 unique_lock<mutex> lock(storage_task_queue_mutex);
1710 surf = surface_for_frame[current_storage_frame];
1711 assert(surf != nullptr);
1714 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1717 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1720 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1722 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, type, BUFFER_OFFSET(surf->y_offset));
1725 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1727 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, type, BUFFER_OFFSET(surf->cbcr_offset));
1730 glBindTexture(GL_TEXTURE_2D, 0);
1732 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1735 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1739 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1741 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1745 unique_lock<mutex> lock(frame_queue_mutex);
1746 current_video_frame.fence = fence;
1747 pending_video_frames.push(move(current_video_frame));
1748 ++current_storage_frame;
1750 frame_queue_nonempty.notify_all();
1754 void QuickSyncEncoderImpl::shutdown()
1761 unique_lock<mutex> lock(frame_queue_mutex);
1762 encode_thread_should_quit = true;
1763 frame_queue_nonempty.notify_all();
1765 encode_thread.join();
1767 unique_lock<mutex> lock(storage_task_queue_mutex);
1768 storage_thread_should_quit = true;
1769 frame_queue_nonempty.notify_all();
1770 storage_task_queue_changed.notify_all();
1772 storage_thread.join();
1774 // Encode any leftover audio in the queues, and also any delayed frames.
1775 file_audio_encoder->encode_last_audio();
1777 if (!global_flags.x264_video_to_disk) {
1784 void QuickSyncEncoderImpl::close_file()
1789 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1791 AVFormatContext *avctx = avformat_alloc_context();
1792 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1793 assert(filename.size() < sizeof(avctx->filename) - 1);
1794 strcpy(avctx->filename, filename.c_str());
1796 string url = "file:" + filename;
1797 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1799 char tmp[AV_ERROR_MAX_STRING_SIZE];
1800 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1804 string video_extradata; // FIXME: See other comment about global headers.
1805 if (global_flags.x264_video_to_disk) {
1806 video_extradata = x264_encoder->get_global_headers();
1809 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1810 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE,
1811 std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1)));
1813 if (global_flags.x264_video_to_disk) {
1814 x264_encoder->add_mux(file_mux.get());
1818 void QuickSyncEncoderImpl::encode_thread_func()
1820 pthread_setname_np(pthread_self(), "QS_Encode");
1822 int64_t last_dts = -1;
1823 int gop_start_display_frame_num = 0;
1824 for (int display_frame_num = 0; ; ++display_frame_num) {
1825 // Wait for the frame to be in the queue. Note that this only means
1826 // we started rendering it.
1829 unique_lock<mutex> lock(frame_queue_mutex);
1830 frame_queue_nonempty.wait(lock, [this]{
1831 return encode_thread_should_quit || !pending_video_frames.empty();
1833 if (encode_thread_should_quit && pending_video_frames.empty()) {
1834 // We may have queued frames left in the reorder buffer
1835 // that were supposed to be B-frames, but have no P-frame
1836 // to be encoded against. If so, encode them all as
1837 // P-frames instead. Note that this happens under the mutex,
1838 // but nobody else uses it at this point, since we're shutting down,
1839 // so there's no contention.
1840 encode_remaining_frames_as_p(quicksync_encoding_frame_num, gop_start_display_frame_num, last_dts);
1843 frame = move(pending_video_frames.front());
1844 pending_video_frames.pop();
1848 // Pass the frame on to x264 (or uncompressed to HTTP) as needed.
1849 // Note that this implicitly waits for the frame to be done rendering.
1850 pass_frame(frame, display_frame_num, frame.pts, frame.duration);
1852 if (global_flags.x264_video_to_disk) {
1853 unique_lock<mutex> lock(storage_task_queue_mutex);
1854 release_gl_surface(display_frame_num);
1858 reorder_buffer[display_frame_num] = move(frame);
1860 // Now encode as many QuickSync frames as we can using the frames we have available.
1861 // (It could be zero, or it could be multiple.) FIXME: make a function.
1864 int frame_type, quicksync_display_frame_num;
1865 encoding2display_order(quicksync_encoding_frame_num, intra_period, intra_idr_period, ip_period,
1866 &quicksync_display_frame_num, &frame_type, &pts_lag);
1867 if (!reorder_buffer.count(quicksync_display_frame_num)) {
1870 frame = move(reorder_buffer[quicksync_display_frame_num]);
1871 reorder_buffer.erase(quicksync_display_frame_num);
1873 if (frame_type == FRAME_IDR) {
1874 // Release any reference frames from the previous GOP.
1875 for (const ReferenceFrame &frame : reference_frames) {
1876 release_gl_surface(frame.display_number);
1878 reference_frames.clear();
1879 current_ref_frame_num = 0;
1880 gop_start_display_frame_num = quicksync_display_frame_num;
1883 // Determine the dts of this frame.
1885 if (pts_lag == -1) {
1886 assert(last_dts != -1);
1887 dts = last_dts + (TIMEBASE / MAX_FPS);
1889 dts = frame.pts - pts_lag;
1893 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);
1894 ++quicksync_encoding_frame_num;
1899 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
1901 if (reorder_buffer.empty()) {
1905 for (auto &pending_frame : reorder_buffer) {
1906 int display_frame_num = pending_frame.first;
1907 assert(display_frame_num > 0);
1908 PendingFrame frame = move(pending_frame.second);
1909 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
1910 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
1911 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration, frame.ycbcr_coefficients);
1916 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
1919 memset(&pkt, 0, sizeof(pkt));
1921 pkt.data = const_cast<uint8_t *>(data);
1922 pkt.size = frame_width * frame_height * 2;
1923 pkt.stream_index = 0;
1924 pkt.flags = AV_PKT_FLAG_KEY;
1925 pkt.duration = duration;
1926 stream_mux->add_packet(pkt, pts, pts);
1931 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
1933 if (src_width == dst_pitch) {
1934 memcpy(dst, src, src_width * height);
1936 for (size_t y = 0; y < height; ++y) {
1937 const uint8_t *sptr = src + y * src_width;
1938 uint8_t *dptr = dst + y * dst_pitch;
1939 memcpy(dptr, sptr, src_width);
1946 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
1948 // Wait for the GPU to be done with the frame.
1951 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
1953 } while (sync_status == GL_TIMEOUT_EXPIRED);
1954 assert(sync_status != GL_WAIT_FAILED);
1956 ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1957 static int frameno = 0;
1958 print_latency("Current mixer latency (video inputs → ready for encode):",
1959 received_ts, false, &frameno);
1961 // Release back any input frames we needed to render this frame.
1962 frame.input_frames.clear();
1966 unique_lock<mutex> lock(storage_task_queue_mutex);
1967 surf = surface_for_frame[display_frame_num];
1968 assert(surf != nullptr);
1970 uint8_t *data = reinterpret_cast<uint8_t *>(surf->y_ptr);
1971 if (global_flags.uncompressed_video_to_http) {
1972 add_packet_for_uncompressed_frame(pts, duration, data);
1973 } else if (global_flags.x264_video_to_http || global_flags.x264_video_to_disk) {
1974 x264_encoder->add_frame(pts, duration, frame.ycbcr_coefficients, data, received_ts);
1978 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
1979 int frame_type, int64_t pts, int64_t dts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients)
1981 const ReceivedTimestamps received_ts = find_received_timestamp(frame.input_frames);
1985 unique_lock<mutex> lock(storage_task_queue_mutex);
1986 surf = surface_for_frame[display_frame_num];
1987 assert(surf != nullptr);
1992 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
1993 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
1994 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
1995 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
1997 // Upload the frame to VA-API.
1998 unsigned char *surface_p = nullptr;
1999 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2001 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2002 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2004 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2005 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2007 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2008 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2011 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2012 CHECK_VASTATUS(va_status, "vaDestroyImage");
2014 // Schedule the frame for encoding.
2015 VASurfaceID va_surface = surf->src_surface;
2016 va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2017 CHECK_VASTATUS(va_status, "vaBeginPicture");
2019 if (frame_type == FRAME_IDR) {
2020 // FIXME: If the mux wants global headers, we should not put the
2021 // SPS/PPS before each IDR frame, but rather put it into the
2022 // codec extradata (formatted differently?).
2024 // NOTE: If we change ycbcr_coefficients, it will not take effect
2025 // before the next IDR frame. This is acceptable, as it should only
2026 // happen on a mode change, which is rare.
2028 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2029 if (h264_packedheader) {
2030 render_packedsequence(ycbcr_coefficients);
2031 render_packedpicture();
2034 //render_sequence();
2035 render_picture(surf, frame_type, display_frame_num, gop_start_display_frame_num);
2037 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2039 va_status = vaEndPicture(va_dpy, context_id);
2040 CHECK_VASTATUS(va_status, "vaEndPicture");
2042 update_ReferenceFrames(display_frame_num, frame_type);
2044 vector<size_t> ref_display_frame_numbers;
2046 // Lock the references for this frame; otherwise, they could be
2047 // rendered to before this frame is done encoding.
2049 unique_lock<mutex> lock(storage_task_queue_mutex);
2050 for (const ReferenceFrame &frame : reference_frames) {
2051 assert(surface_for_frame.count(frame.display_number));
2052 ++surface_for_frame[frame.display_number]->refcount;
2053 ref_display_frame_numbers.push_back(frame.display_number);
2057 // so now the data is done encoding (well, async job kicked off)...
2058 // we send that to the storage thread
2060 tmp.display_order = display_frame_num;
2061 tmp.frame_type = frame_type;
2064 tmp.duration = duration;
2065 tmp.ycbcr_coefficients = ycbcr_coefficients;
2066 tmp.received_ts = received_ts;
2067 tmp.ref_display_frame_numbers = move(ref_display_frame_numbers);
2068 storage_task_enqueue(move(tmp));
2072 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)
2073 : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {}
2075 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2076 QuickSyncEncoder::~QuickSyncEncoder() {}
2078 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2080 impl->add_audio(pts, audio);
2083 bool QuickSyncEncoder::begin_frame(int64_t pts, int64_t duration, YCbCrLumaCoefficients ycbcr_coefficients, const vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex)
2085 return impl->begin_frame(pts, duration, ycbcr_coefficients, input_frames, y_tex, cbcr_tex);
2088 RefCountedGLsync QuickSyncEncoder::end_frame()
2090 return impl->end_frame();
2093 void QuickSyncEncoder::shutdown()
2098 void QuickSyncEncoder::close_file()
2103 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2105 impl->set_stream_mux(mux);
2108 int64_t QuickSyncEncoder::global_delay() const {
2109 return impl->global_delay();