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>
16 #include <va/va_drm.h>
17 #include <va/va_drmcommon.h>
18 #include <va/va_enc_h264.h>
19 #include <va/va_x11.h>
21 #include <condition_variable>
35 #include <libavcodec/avcodec.h>
36 #include <libavformat/avio.h>
37 #include <libavutil/error.h>
38 #include <libdrm/drm_fourcc.h>
42 #include "audio_encoder.h"
45 #include "ffmpeg_raii.h"
48 #include "ref_counted_frame.h"
50 #include "x264_encoder.h"
57 #define CHECK_VASTATUS(va_status, func) \
58 if (va_status != VA_STATUS_SUCCESS) { \
59 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
63 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
65 //#include "loadsurface.h"
67 #define NAL_REF_IDC_NONE 0
68 #define NAL_REF_IDC_LOW 1
69 #define NAL_REF_IDC_MEDIUM 2
70 #define NAL_REF_IDC_HIGH 3
78 #define SLICE_TYPE_P 0
79 #define SLICE_TYPE_B 1
80 #define SLICE_TYPE_I 2
81 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
82 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
83 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
86 #define ENTROPY_MODE_CAVLC 0
87 #define ENTROPY_MODE_CABAC 1
89 #define PROFILE_IDC_BASELINE 66
90 #define PROFILE_IDC_MAIN 77
91 #define PROFILE_IDC_HIGH 100
93 #define BITSTREAM_ALLOCATE_STEPPING 4096
94 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
95 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
96 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
98 static constexpr unsigned int MaxFrameNum = (2<<16);
99 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
100 static constexpr unsigned int Log2MaxFrameNum = 16;
101 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
102 static constexpr int rc_default_modes[] = { // Priority list of modes.
105 VA_RC_VBR_CONSTRAINED,
111 /* thread to save coded data */
112 #define SRC_SURFACE_FREE 0
113 #define SRC_SURFACE_IN_ENCODING 1
116 unsigned int *buffer;
118 int max_size_in_dword;
120 typedef struct __bitstream bitstream;
124 // H.264 video comes out in encoding order (e.g. with two B-frames:
125 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
126 // come in the right order. Since we do everything, including waiting
127 // for the frames to come out of OpenGL, in encoding order, we need
128 // a reordering buffer for uncompressed frames so that they come out
129 // correctly. We go the super-lazy way of not making it understand
130 // anything about the true order (which introduces some extra latency,
131 // though); we know that for N B-frames we need at most (N-1) frames
132 // in the reorder buffer, and can just sort on that.
134 // The class also deals with keeping a freelist as needed.
135 class FrameReorderer {
137 FrameReorderer(unsigned queue_length, int width, int height);
140 int64_t pts, duration;
143 // Invert to get the smallest pts first.
144 bool operator< (const Frame &other) const { return pts > other.pts; }
147 // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
148 // Does _not_ take ownership of data; a copy is taken if needed.
149 // The returned pointer is valid until the next call to reorder_frame, or destruction.
150 // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
151 Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data);
153 // The same as reorder_frame, but without inserting anything. Used to empty the queue.
154 Frame get_first_frame();
156 bool empty() const { return frames.empty(); }
159 unsigned queue_length;
162 priority_queue<Frame> frames;
163 stack<uint8_t *> freelist; // Includes the last value returned from reorder_frame.
165 // Owns all the pointers. Normally, freelist and frames could do this themselves,
166 // except priority_queue doesn't work well with movable-only types.
167 vector<unique_ptr<uint8_t[]>> owner;
170 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
171 : queue_length(queue_length), width(width), height(height)
173 for (unsigned i = 0; i < queue_length; ++i) {
174 owner.emplace_back(new uint8_t[width * height * 2]);
175 freelist.push(owner.back().get());
179 FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data)
181 if (queue_length == 0) {
182 return Frame{pts, duration, data};
185 assert(!freelist.empty());
186 uint8_t *storage = freelist.top();
188 memcpy(storage, data, width * height * 2);
189 frames.push(Frame{pts, duration, storage});
191 if (frames.size() >= queue_length) {
192 return get_first_frame();
194 return Frame{-1, -1, nullptr};
198 FrameReorderer::Frame FrameReorderer::get_first_frame()
200 assert(!frames.empty());
201 Frame storage = frames.top();
203 freelist.push(storage.data);
207 class QuickSyncEncoderImpl {
209 QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder);
210 ~QuickSyncEncoderImpl();
211 void add_audio(int64_t pts, vector<float> audio);
212 bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
213 RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames);
215 void release_gl_resources();
216 void set_stream_mux(Mux *mux)
221 // So we never get negative dts.
222 int64_t global_delay() const {
223 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
227 struct storage_task {
228 unsigned long long display_order;
231 int64_t pts, dts, duration;
233 struct PendingFrame {
234 RefCountedGLsync fence;
235 vector<RefCountedFrame> input_frames;
236 int64_t pts, duration;
239 void open_output_file(const std::string &filename);
240 void encode_thread_func();
241 void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
242 void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
243 void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
244 int frame_type, int64_t pts, int64_t dts, int64_t duration);
245 void storage_task_thread();
246 void storage_task_enqueue(storage_task task);
247 void save_codeddata(storage_task task);
248 int render_packedsequence();
249 int render_packedpicture();
250 void render_packedslice();
251 int render_sequence();
252 int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
253 void sps_rbsp(bitstream *bs);
254 void pps_rbsp(bitstream *bs);
255 int build_packed_pic_buffer(unsigned char **header_buffer);
256 int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
257 void slice_header(bitstream *bs);
258 int build_packed_seq_buffer(unsigned char **header_buffer);
259 int build_packed_slice_buffer(unsigned char **header_buffer);
260 int init_va(const string &va_display);
262 void enable_zerocopy_if_possible();
263 VADisplay va_open_display(const string &va_display);
264 void va_close_display(VADisplay va_dpy);
266 void release_encode();
267 void update_ReferenceFrames(int frame_type);
268 int update_RefPicList(int frame_type);
270 bool is_shutdown = false;
271 bool has_released_gl_resources = false;
275 thread encode_thread, storage_thread;
277 mutex storage_task_queue_mutex;
278 condition_variable storage_task_queue_changed;
279 int srcsurface_status[SURFACE_NUM]; // protected by storage_task_queue_mutex
280 queue<storage_task> storage_task_queue; // protected by storage_task_queue_mutex
281 bool storage_thread_should_quit = false; // protected by storage_task_queue_mutex
283 mutex frame_queue_mutex;
284 condition_variable frame_queue_nonempty;
285 bool encode_thread_should_quit = false; // under frame_queue_mutex
287 int current_storage_frame;
289 map<int, PendingFrame> pending_video_frames; // under frame_queue_mutex
290 movit::ResourcePool *resource_pool;
293 unique_ptr<AudioEncoder> file_audio_encoder;
295 unique_ptr<FrameReorderer> reorderer;
296 X264Encoder *x264_encoder; // nullptr if not using x264.
298 Mux* stream_mux = nullptr; // To HTTP.
299 unique_ptr<Mux> file_mux; // To local disk.
301 Display *x11_display = nullptr;
303 // Encoder parameters
305 VAProfile h264_profile = (VAProfile)~0;
306 VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
307 int config_attrib_num = 0, enc_packed_header_idx;
310 VASurfaceID src_surface, ref_surface;
311 VABufferID coded_buf;
313 VAImage surface_image;
314 GLuint y_tex, cbcr_tex;
316 // Only if use_zerocopy == true.
317 EGLImage y_egl_image, cbcr_egl_image;
319 // Only if use_zerocopy == false.
321 uint8_t *y_ptr, *cbcr_ptr;
322 size_t y_offset, cbcr_offset;
324 GLSurface gl_surfaces[SURFACE_NUM];
326 VAConfigID config_id;
327 VAContextID context_id;
328 VAEncSequenceParameterBufferH264 seq_param;
329 VAEncPictureParameterBufferH264 pic_param;
330 VAEncSliceParameterBufferH264 slice_param;
331 VAPictureH264 CurrentCurrPic;
332 VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
334 // Static quality settings.
335 static constexpr unsigned int frame_bitrate = 15000000 / 60; // Doesn't really matter; only initial_qp does.
336 static constexpr unsigned int num_ref_frames = 2;
337 static constexpr int initial_qp = 15;
338 static constexpr int minimal_qp = 0;
339 static constexpr int intra_period = 30;
340 static constexpr int intra_idr_period = MAX_FPS; // About a second; more at lower frame rates. Not ideal.
342 // Quality settings that are meant to be static, but might be overridden
344 int constraint_set_flag = 0;
345 int h264_packedheader = 0; /* support pack header? */
346 int h264_maxref = (1<<16|1);
347 int h264_entropy_mode = 1; /* cabac */
351 unsigned int current_frame_num = 0;
352 unsigned int numShortTerm = 0;
356 int frame_width_mbaligned;
357 int frame_height_mbaligned;
360 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
361 // but if we don't delete it here, we get leaks. The GStreamer implementation
363 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
365 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
366 CHECK_VASTATUS(va_status, "vaRenderPicture");
368 for (int i = 0; i < num_buffers; ++i) {
369 va_status = vaDestroyBuffer(dpy, buffers[i]);
370 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
375 va_swap32(unsigned int val)
377 unsigned char *pval = (unsigned char *)&val;
379 return ((pval[0] << 24) |
386 bitstream_start(bitstream *bs)
388 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
389 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
394 bitstream_end(bitstream *bs)
396 int pos = (bs->bit_offset >> 5);
397 int bit_offset = (bs->bit_offset & 0x1f);
398 int bit_left = 32 - bit_offset;
401 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
406 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
408 int pos = (bs->bit_offset >> 5);
409 int bit_offset = (bs->bit_offset & 0x1f);
410 int bit_left = 32 - bit_offset;
415 bs->bit_offset += size_in_bits;
417 if (bit_left > size_in_bits) {
418 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
420 size_in_bits -= bit_left;
421 if (bit_left >= 32) {
422 bs->buffer[pos] = (val >> size_in_bits);
424 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
426 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
428 if (pos + 1 == bs->max_size_in_dword) {
429 bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
430 bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
433 bs->buffer[pos + 1] = val;
438 bitstream_put_ue(bitstream *bs, unsigned int val)
440 int size_in_bits = 0;
448 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
449 bitstream_put_ui(bs, val, size_in_bits);
453 bitstream_put_se(bitstream *bs, int val)
455 unsigned int new_val;
460 new_val = 2 * val - 1;
462 bitstream_put_ue(bs, new_val);
466 bitstream_byte_aligning(bitstream *bs, int bit)
468 int bit_offset = (bs->bit_offset & 0x7);
469 int bit_left = 8 - bit_offset;
475 assert(bit == 0 || bit == 1);
478 new_val = (1 << bit_left) - 1;
482 bitstream_put_ui(bs, new_val, bit_left);
486 rbsp_trailing_bits(bitstream *bs)
488 bitstream_put_ui(bs, 1, 1);
489 bitstream_byte_aligning(bs, 0);
492 static void nal_start_code_prefix(bitstream *bs)
494 bitstream_put_ui(bs, 0x00000001, 32);
497 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
499 bitstream_put_ui(bs, 0, 1); /* forbidden_zero_bit: 0 */
500 bitstream_put_ui(bs, nal_ref_idc, 2);
501 bitstream_put_ui(bs, nal_unit_type, 5);
504 void QuickSyncEncoderImpl::sps_rbsp(bitstream *bs)
506 int profile_idc = PROFILE_IDC_BASELINE;
508 if (h264_profile == VAProfileH264High)
509 profile_idc = PROFILE_IDC_HIGH;
510 else if (h264_profile == VAProfileH264Main)
511 profile_idc = PROFILE_IDC_MAIN;
513 bitstream_put_ui(bs, profile_idc, 8); /* profile_idc */
514 bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1); /* constraint_set0_flag */
515 bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1); /* constraint_set1_flag */
516 bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1); /* constraint_set2_flag */
517 bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1); /* constraint_set3_flag */
518 bitstream_put_ui(bs, 0, 4); /* reserved_zero_4bits */
519 bitstream_put_ui(bs, seq_param.level_idc, 8); /* level_idc */
520 bitstream_put_ue(bs, seq_param.seq_parameter_set_id); /* seq_parameter_set_id */
522 if ( profile_idc == PROFILE_IDC_HIGH) {
523 bitstream_put_ue(bs, 1); /* chroma_format_idc = 1, 4:2:0 */
524 bitstream_put_ue(bs, 0); /* bit_depth_luma_minus8 */
525 bitstream_put_ue(bs, 0); /* bit_depth_chroma_minus8 */
526 bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
527 bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
530 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
531 bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type); /* pic_order_cnt_type */
533 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
534 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4); /* log2_max_pic_order_cnt_lsb_minus4 */
539 bitstream_put_ue(bs, seq_param.max_num_ref_frames); /* num_ref_frames */
540 bitstream_put_ui(bs, 0, 1); /* gaps_in_frame_num_value_allowed_flag */
542 bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1); /* pic_width_in_mbs_minus1 */
543 bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
544 bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1); /* frame_mbs_only_flag */
546 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
550 bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1); /* direct_8x8_inference_flag */
551 bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1); /* frame_cropping_flag */
553 if (seq_param.frame_cropping_flag) {
554 bitstream_put_ue(bs, seq_param.frame_crop_left_offset); /* frame_crop_left_offset */
555 bitstream_put_ue(bs, seq_param.frame_crop_right_offset); /* frame_crop_right_offset */
556 bitstream_put_ue(bs, seq_param.frame_crop_top_offset); /* frame_crop_top_offset */
557 bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset); /* frame_crop_bottom_offset */
560 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
562 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
564 bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
565 bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
566 bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
567 bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
569 bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
570 bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
571 bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
573 bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
574 bitstream_put_ui(bs, 2, 8); /* transfer_characteristics (2 = unspecified, since we use sRGB) */
575 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
578 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
579 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
581 bitstream_put_ui(bs, 1, 32); // FPS
582 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
583 bitstream_put_ui(bs, 1, 1);
585 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
588 bitstream_put_ue(bs, 0); /* cpb_cnt_minus1 */
589 bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
590 bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
592 bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
593 bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
594 bitstream_put_ui(bs, 1, 1); /* cbr_flag[0] */
596 bitstream_put_ui(bs, 23, 5); /* initial_cpb_removal_delay_length_minus1 */
597 bitstream_put_ui(bs, 23, 5); /* cpb_removal_delay_length_minus1 */
598 bitstream_put_ui(bs, 23, 5); /* dpb_output_delay_length_minus1 */
599 bitstream_put_ui(bs, 23, 5); /* time_offset_length */
601 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
602 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
604 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
605 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
608 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
612 void QuickSyncEncoderImpl::pps_rbsp(bitstream *bs)
614 bitstream_put_ue(bs, pic_param.pic_parameter_set_id); /* pic_parameter_set_id */
615 bitstream_put_ue(bs, pic_param.seq_parameter_set_id); /* seq_parameter_set_id */
617 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
619 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
621 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
623 bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1); /* num_ref_idx_l0_active_minus1 */
624 bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1); /* num_ref_idx_l1_active_minus1 1 */
626 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1); /* weighted_pred_flag: 0 */
627 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2); /* weighted_bipred_idc: 0 */
629 bitstream_put_se(bs, pic_param.pic_init_qp - 26); /* pic_init_qp_minus26 */
630 bitstream_put_se(bs, 0); /* pic_init_qs_minus26 */
631 bitstream_put_se(bs, 0); /* chroma_qp_index_offset */
633 bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
634 bitstream_put_ui(bs, 0, 1); /* constrained_intra_pred_flag */
635 bitstream_put_ui(bs, 0, 1); /* redundant_pic_cnt_present_flag */
638 bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1); /*transform_8x8_mode_flag */
639 bitstream_put_ui(bs, 0, 1); /* pic_scaling_matrix_present_flag */
640 bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset ); /*second_chroma_qp_index_offset */
642 rbsp_trailing_bits(bs);
645 void QuickSyncEncoderImpl::slice_header(bitstream *bs)
647 int first_mb_in_slice = slice_param.macroblock_address;
649 bitstream_put_ue(bs, first_mb_in_slice); /* first_mb_in_slice: 0 */
650 bitstream_put_ue(bs, slice_param.slice_type); /* slice_type */
651 bitstream_put_ue(bs, slice_param.pic_parameter_set_id); /* pic_parameter_set_id: 0 */
652 bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
654 /* frame_mbs_only_flag == 1 */
655 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
660 if (pic_param.pic_fields.bits.idr_pic_flag)
661 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
663 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
664 bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
665 /* pic_order_present_flag == 0 */
671 /* redundant_pic_cnt_present_flag == 0 */
673 if (IS_P_SLICE(slice_param.slice_type)) {
674 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
676 if (slice_param.num_ref_idx_active_override_flag)
677 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
679 /* ref_pic_list_reordering */
680 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
681 } else if (IS_B_SLICE(slice_param.slice_type)) {
682 bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1); /* direct_spatial_mv_pred: 1 */
684 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
686 if (slice_param.num_ref_idx_active_override_flag) {
687 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
688 bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
691 /* ref_pic_list_reordering */
692 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
693 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l1: 0 */
696 if ((pic_param.pic_fields.bits.weighted_pred_flag &&
697 IS_P_SLICE(slice_param.slice_type)) ||
698 ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
699 IS_B_SLICE(slice_param.slice_type))) {
700 /* FIXME: fill weight/offset table */
704 /* dec_ref_pic_marking */
705 if (pic_param.pic_fields.bits.reference_pic_flag) { /* nal_ref_idc != 0 */
706 unsigned char no_output_of_prior_pics_flag = 0;
707 unsigned char long_term_reference_flag = 0;
708 unsigned char adaptive_ref_pic_marking_mode_flag = 0;
710 if (pic_param.pic_fields.bits.idr_pic_flag) {
711 bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1); /* no_output_of_prior_pics_flag: 0 */
712 bitstream_put_ui(bs, long_term_reference_flag, 1); /* long_term_reference_flag: 0 */
714 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
718 if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
719 !IS_I_SLICE(slice_param.slice_type))
720 bitstream_put_ue(bs, slice_param.cabac_init_idc); /* cabac_init_idc: 0 */
722 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
724 /* ignore for SP/SI */
726 if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
727 bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc); /* disable_deblocking_filter_idc: 0 */
729 if (slice_param.disable_deblocking_filter_idc != 1) {
730 bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2); /* slice_alpha_c0_offset_div2: 2 */
731 bitstream_put_se(bs, slice_param.slice_beta_offset_div2); /* slice_beta_offset_div2: 2 */
735 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
736 bitstream_byte_aligning(bs, 1);
740 int QuickSyncEncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
744 bitstream_start(&bs);
745 nal_start_code_prefix(&bs);
746 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
750 *header_buffer = (unsigned char *)bs.buffer;
751 return bs.bit_offset;
755 QuickSyncEncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
759 bitstream_start(&bs);
760 nal_start_code_prefix(&bs);
761 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
765 *header_buffer = (unsigned char *)bs.buffer;
766 return bs.bit_offset;
769 int QuickSyncEncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
772 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
773 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
775 bitstream_start(&bs);
776 nal_start_code_prefix(&bs);
778 if (IS_I_SLICE(slice_param.slice_type)) {
779 nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
780 } else if (IS_P_SLICE(slice_param.slice_type)) {
781 nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
783 assert(IS_B_SLICE(slice_param.slice_type));
784 nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
790 *header_buffer = (unsigned char *)bs.buffer;
791 return bs.bit_offset;
796 Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
797 1) period between Frame #X and Frame #N = #X - #N
798 2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
799 3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
800 4) intra_period and intra_idr_period take precedence over ip_period
801 5) if ip_period > 1, intra_period and intra_idr_period are not the strict periods
802 of I/IDR frames, see bellow examples
803 -------------------------------------------------------------------
804 intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
805 0 ignored 1 IDRPPPPPPP ... (No IDR/I any more)
806 0 ignored >=2 IDR(PBB)(PBB)... (No IDR/I any more)
807 1 0 ignored IDRIIIIIII... (No IDR any more)
808 1 1 ignored IDR IDR IDR IDR...
809 1 >=2 ignored IDRII IDRII IDR... (1/3/ignore)
810 >=2 0 1 IDRPPP IPPP I... (3/0/1)
811 >=2 0 >=2 IDR(PBB)(PBB)(IBB) (6/0/3)
812 (PBB)(IBB)(PBB)(IBB)...
813 >=2 >=2 1 IDRPPPPP IPPPPP IPPPPP (6/18/1)
814 IDRPPPPP IPPPPP IPPPPP...
815 >=2 >=2 >=2 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
816 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
817 {IDR(PBB)(PBB)(IBB)(PBB)} (6/12/3)
818 {IDR(PBB)(PBB)(IBB)(PBB)}...
819 {IDR(PBB)(PBB)} (6/6/3)
823 // General pts/dts strategy:
825 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
826 // bit tricky. We assume first of all that the frame rate never goes _above_
827 // MAX_FPS, which gives us a frame period N. The decoder can always decode
828 // in at least this speed, as long at dts <= pts (the frame is not attempted
829 // presented before it is decoded). Furthermore, we never have longer chains of
830 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
831 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
832 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
833 // frame, with an order O <= C.)
835 // Many strategies are possible, but we establish these rules:
837 // - Tip frames have dts = pts - (C-O)*N.
838 // - Non-tip frames have dts = dts_last + N.
840 // An example, with C=2 and N=10 and the data flow showed with arrows:
843 // pts: 30 40 50 60 70 80
845 // dts: 10 30 20 60 50←40
850 // To show that this works fine also with irregular spacings, let's say that
851 // the third frame is delayed a bit (something earlier was dropped). Now the
852 // situation looks like this:
855 // pts: 30 40 80 90 100 110
857 // dts: 10 30 20 90 50←40
862 // The resetting on every tip frame makes sure dts never ends up lagging a lot
863 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
865 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
866 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
867 // a tip frame and should be given a dts based on the previous one.
872 void encoding2display_order(
873 int encoding_order, int intra_period,
874 int intra_idr_period, int ip_period,
875 int *displaying_order,
876 int *frame_type, int *pts_lag)
878 int encoding_order_gop = 0;
882 if (intra_period == 1) { /* all are I/IDR frames */
883 *displaying_order = encoding_order;
884 if (intra_idr_period == 0)
885 *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
887 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
891 if (intra_period == 0)
892 intra_idr_period = 0;
894 if (ip_period == 1) {
895 // No B-frames, sequence is like IDR PPPPP IPPPPP.
896 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
897 *displaying_order = encoding_order;
899 if (encoding_order_gop == 0) { /* the first frame */
900 *frame_type = FRAME_IDR;
901 } else if (intra_period != 0 && /* have I frames */
902 encoding_order_gop >= 2 &&
903 (encoding_order_gop % intra_period == 0)) {
904 *frame_type = FRAME_I;
906 *frame_type = FRAME_P;
911 // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
912 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
913 *pts_lag = -1; // Most frames are not tip frames.
915 if (encoding_order_gop == 0) { /* the first frame */
916 *frame_type = FRAME_IDR;
917 *displaying_order = encoding_order;
918 // IDR frames are a special case; I honestly can't find the logic behind
919 // why this is the right thing, but it seems to line up nicely in practice :-)
920 *pts_lag = TIMEBASE / MAX_FPS;
921 } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
922 *frame_type = FRAME_B;
923 *displaying_order = encoding_order - 1;
924 if ((encoding_order_gop % ip_period) == 0) {
925 *pts_lag = 0; // Last B-frame.
927 } else if (intra_period != 0 && /* have I frames */
928 encoding_order_gop >= 2 &&
929 ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
930 *frame_type = FRAME_I;
931 *displaying_order = encoding_order + ip_period - 1;
933 *frame_type = FRAME_P;
934 *displaying_order = encoding_order + ip_period - 1;
939 static const char *rc_to_string(int rc_mode)
952 case VA_RC_VBR_CONSTRAINED:
953 return "VBR_CONSTRAINED";
959 void QuickSyncEncoderImpl::enable_zerocopy_if_possible()
961 if (global_flags.uncompressed_video_to_http) {
962 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-uncompressed-video.\n");
963 use_zerocopy = false;
964 } else if (global_flags.x264_video_to_http) {
965 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --http-x264-video.\n");
966 use_zerocopy = false;
972 VADisplay QuickSyncEncoderImpl::va_open_display(const string &va_display)
974 if (va_display.empty()) {
975 x11_display = XOpenDisplay(NULL);
977 fprintf(stderr, "error: can't connect to X server!\n");
980 enable_zerocopy_if_possible();
981 return vaGetDisplay(x11_display);
982 } else if (va_display[0] != '/') {
983 x11_display = XOpenDisplay(va_display.c_str());
985 fprintf(stderr, "error: can't connect to X server!\n");
988 enable_zerocopy_if_possible();
989 return vaGetDisplay(x11_display);
991 drm_fd = open(va_display.c_str(), O_RDWR);
993 perror(va_display.c_str());
996 use_zerocopy = false;
997 return vaGetDisplayDRM(drm_fd);
1001 void QuickSyncEncoderImpl::va_close_display(VADisplay va_dpy)
1004 XCloseDisplay(x11_display);
1005 x11_display = nullptr;
1012 int QuickSyncEncoderImpl::init_va(const string &va_display)
1014 VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1015 VAEntrypoint *entrypoints;
1016 int num_entrypoints, slice_entrypoint;
1017 int support_encode = 0;
1018 int major_ver, minor_ver;
1022 va_dpy = va_open_display(va_display);
1023 va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1024 CHECK_VASTATUS(va_status, "vaInitialize");
1026 num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1027 entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1029 fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1033 /* use the highest profile */
1034 for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1035 if ((h264_profile != ~0) && h264_profile != profile_list[i])
1038 h264_profile = profile_list[i];
1039 vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1040 for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1041 if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1046 if (support_encode == 1)
1050 if (support_encode == 0) {
1051 printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1052 printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1053 printf("to use VA-API against DRM instead of X11.\n");
1056 switch (h264_profile) {
1057 case VAProfileH264Baseline:
1059 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1060 h264_entropy_mode = 0;
1062 case VAProfileH264ConstrainedBaseline:
1063 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1067 case VAProfileH264Main:
1068 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1071 case VAProfileH264High:
1072 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1075 h264_profile = VAProfileH264Baseline;
1077 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1082 VAConfigAttrib attrib[VAConfigAttribTypeMax];
1084 /* find out the format for the render target, and rate control mode */
1085 for (i = 0; i < VAConfigAttribTypeMax; i++)
1086 attrib[i].type = (VAConfigAttribType)i;
1088 va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1089 &attrib[0], VAConfigAttribTypeMax);
1090 CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1091 /* check the interested configattrib */
1092 if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1093 printf("Not find desired YUV420 RT format\n");
1096 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1097 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1098 config_attrib_num++;
1101 if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1102 int tmp = attrib[VAConfigAttribRateControl].value;
1104 if (rc_mode == -1 || !(rc_mode & tmp)) {
1105 if (rc_mode != -1) {
1106 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1109 for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1110 if (rc_default_modes[i] & tmp) {
1111 rc_mode = rc_default_modes[i];
1117 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1118 config_attrib[config_attrib_num].value = rc_mode;
1119 config_attrib_num++;
1123 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1124 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1126 h264_packedheader = 1;
1127 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1128 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1130 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1131 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1134 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1135 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1138 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1139 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1142 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1143 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1146 enc_packed_header_idx = config_attrib_num;
1147 config_attrib_num++;
1150 if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1151 config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1152 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1153 config_attrib_num++;
1156 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1157 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1164 int QuickSyncEncoderImpl::setup_encode()
1167 VASurfaceID *tmp_surfaceid;
1168 int codedbuf_size, i;
1169 VASurfaceID src_surface[SURFACE_NUM];
1170 VASurfaceID ref_surface[SURFACE_NUM];
1172 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1173 &config_attrib[0], config_attrib_num, &config_id);
1174 CHECK_VASTATUS(va_status, "vaCreateConfig");
1176 /* create source surfaces */
1177 va_status = vaCreateSurfaces(va_dpy,
1178 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1179 &src_surface[0], SURFACE_NUM,
1181 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1183 /* create reference surfaces */
1184 va_status = vaCreateSurfaces(va_dpy,
1185 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1186 &ref_surface[0], SURFACE_NUM,
1188 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1190 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1191 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1192 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1194 /* Create a context for this encode pipe */
1195 va_status = vaCreateContext(va_dpy, config_id,
1196 frame_width_mbaligned, frame_height_mbaligned,
1198 tmp_surfaceid, 2 * SURFACE_NUM,
1200 CHECK_VASTATUS(va_status, "vaCreateContext");
1201 free(tmp_surfaceid);
1203 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1205 for (i = 0; i < SURFACE_NUM; i++) {
1206 /* create coded buffer once for all
1207 * other VA buffers which won't be used again after vaRenderPicture.
1208 * so APP can always vaCreateBuffer for every frame
1209 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1210 * so VA won't maintain the coded buffer
1212 va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1213 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1214 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1217 /* create OpenGL objects */
1218 //glGenFramebuffers(SURFACE_NUM, fbos);
1220 for (i = 0; i < SURFACE_NUM; i++) {
1222 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, 1, 1);
1223 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, 1, 1);
1225 gl_surfaces[i].y_tex = resource_pool->create_2d_texture(GL_R8, frame_width, frame_height);
1226 gl_surfaces[i].cbcr_tex = resource_pool->create_2d_texture(GL_RG8, frame_width / 2, frame_height / 2);
1228 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1229 // buffers, due to potentially differing pitch.
1230 glGenBuffers(1, &gl_surfaces[i].pbo);
1231 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1232 glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1233 uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1234 gl_surfaces[i].y_offset = 0;
1235 gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1236 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1237 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1238 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1242 for (i = 0; i < SURFACE_NUM; i++) {
1243 gl_surfaces[i].src_surface = src_surface[i];
1244 gl_surfaces[i].ref_surface = ref_surface[i];
1250 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1253 template<class T, class C>
1254 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1256 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1257 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1258 sort(middle, end, less_than);
1261 void QuickSyncEncoderImpl::update_ReferenceFrames(int frame_type)
1265 if (frame_type == FRAME_B)
1268 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1270 if (numShortTerm > num_ref_frames)
1271 numShortTerm = num_ref_frames;
1272 for (i=numShortTerm-1; i>0; i--)
1273 ReferenceFrames[i] = ReferenceFrames[i-1];
1274 ReferenceFrames[0] = CurrentCurrPic;
1276 current_frame_num++;
1277 if (current_frame_num > MaxFrameNum)
1278 current_frame_num = 0;
1282 int QuickSyncEncoderImpl::update_RefPicList(int frame_type)
1284 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1285 return a.frame_idx > b.frame_idx;
1287 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1288 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1290 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1291 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1294 if (frame_type == FRAME_P) {
1295 memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1296 sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1297 } else if (frame_type == FRAME_B) {
1298 memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1299 sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1301 memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1302 sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1309 int QuickSyncEncoderImpl::render_sequence()
1311 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1313 VAEncMiscParameterBuffer *misc_param;
1314 VAEncMiscParameterRateControl *misc_rate_ctrl;
1316 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1317 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1318 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1319 seq_param.bits_per_second = frame_bitrate;
1321 seq_param.intra_period = intra_period;
1322 seq_param.intra_idr_period = intra_idr_period;
1323 seq_param.ip_period = ip_period;
1325 seq_param.max_num_ref_frames = num_ref_frames;
1326 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1327 seq_param.time_scale = TIMEBASE * 2;
1328 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1329 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1330 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1331 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1332 seq_param.seq_fields.bits.chroma_format_idc = 1;
1333 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1335 if (frame_width != frame_width_mbaligned ||
1336 frame_height != frame_height_mbaligned) {
1337 seq_param.frame_cropping_flag = 1;
1338 seq_param.frame_crop_left_offset = 0;
1339 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1340 seq_param.frame_crop_top_offset = 0;
1341 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1344 va_status = vaCreateBuffer(va_dpy, context_id,
1345 VAEncSequenceParameterBufferType,
1346 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1347 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1349 va_status = vaCreateBuffer(va_dpy, context_id,
1350 VAEncMiscParameterBufferType,
1351 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1352 1, NULL, &rc_param_buf);
1353 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1355 vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1356 misc_param->type = VAEncMiscParameterTypeRateControl;
1357 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1358 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1359 misc_rate_ctrl->bits_per_second = frame_bitrate;
1360 misc_rate_ctrl->target_percentage = 66;
1361 misc_rate_ctrl->window_size = 1000;
1362 misc_rate_ctrl->initial_qp = initial_qp;
1363 misc_rate_ctrl->min_qp = minimal_qp;
1364 misc_rate_ctrl->basic_unit_size = 0;
1365 vaUnmapBuffer(va_dpy, rc_param_buf);
1367 render_id[0] = seq_param_buf;
1368 render_id[1] = rc_param_buf;
1370 render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1375 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1377 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1378 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1379 int PicOrderCntMsb, TopFieldOrderCnt;
1381 if (frame_type == FRAME_IDR)
1382 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1384 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1385 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1388 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1389 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1390 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1391 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1392 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1393 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1395 PicOrderCntMsb = prevPicOrderCntMsb;
1397 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1399 if (frame_type != FRAME_B) {
1400 PicOrderCntMsb_ref = PicOrderCntMsb;
1401 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1404 return TopFieldOrderCnt;
1407 int QuickSyncEncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1409 VABufferID pic_param_buf;
1413 pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1414 pic_param.CurrPic.frame_idx = current_frame_num;
1415 pic_param.CurrPic.flags = 0;
1416 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1417 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1418 CurrentCurrPic = pic_param.CurrPic;
1420 memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1421 for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1422 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1423 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1426 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1427 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1428 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1429 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1430 pic_param.frame_num = current_frame_num;
1431 pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1432 pic_param.last_picture = false; // FIXME
1433 pic_param.pic_init_qp = initial_qp;
1435 va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1436 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1437 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1439 render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1444 int QuickSyncEncoderImpl::render_packedsequence()
1446 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1447 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1448 unsigned int length_in_bits;
1449 unsigned char *packedseq_buffer = NULL;
1452 length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1454 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1456 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1457 packedheader_param_buffer.has_emulation_bytes = 0;
1458 va_status = vaCreateBuffer(va_dpy,
1460 VAEncPackedHeaderParameterBufferType,
1461 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1462 &packedseq_para_bufid);
1463 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1465 va_status = vaCreateBuffer(va_dpy,
1467 VAEncPackedHeaderDataBufferType,
1468 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1469 &packedseq_data_bufid);
1470 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1472 render_id[0] = packedseq_para_bufid;
1473 render_id[1] = packedseq_data_bufid;
1474 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1476 free(packedseq_buffer);
1482 int QuickSyncEncoderImpl::render_packedpicture()
1484 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1485 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1486 unsigned int length_in_bits;
1487 unsigned char *packedpic_buffer = NULL;
1490 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1491 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1492 packedheader_param_buffer.bit_length = length_in_bits;
1493 packedheader_param_buffer.has_emulation_bytes = 0;
1495 va_status = vaCreateBuffer(va_dpy,
1497 VAEncPackedHeaderParameterBufferType,
1498 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1499 &packedpic_para_bufid);
1500 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1502 va_status = vaCreateBuffer(va_dpy,
1504 VAEncPackedHeaderDataBufferType,
1505 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1506 &packedpic_data_bufid);
1507 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1509 render_id[0] = packedpic_para_bufid;
1510 render_id[1] = packedpic_data_bufid;
1511 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1513 free(packedpic_buffer);
1518 void QuickSyncEncoderImpl::render_packedslice()
1520 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1521 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1522 unsigned int length_in_bits;
1523 unsigned char *packedslice_buffer = NULL;
1526 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1527 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1528 packedheader_param_buffer.bit_length = length_in_bits;
1529 packedheader_param_buffer.has_emulation_bytes = 0;
1531 va_status = vaCreateBuffer(va_dpy,
1533 VAEncPackedHeaderParameterBufferType,
1534 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1535 &packedslice_para_bufid);
1536 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1538 va_status = vaCreateBuffer(va_dpy,
1540 VAEncPackedHeaderDataBufferType,
1541 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1542 &packedslice_data_bufid);
1543 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1545 render_id[0] = packedslice_para_bufid;
1546 render_id[1] = packedslice_data_bufid;
1547 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1549 free(packedslice_buffer);
1552 int QuickSyncEncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1554 VABufferID slice_param_buf;
1558 update_RefPicList(frame_type);
1560 /* one frame, one slice */
1561 slice_param.macroblock_address = 0;
1562 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1563 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1564 if (frame_type == FRAME_IDR) {
1565 if (encoding_frame_num != 0)
1566 ++slice_param.idr_pic_id;
1567 } else if (frame_type == FRAME_P) {
1568 int refpiclist0_max = h264_maxref & 0xffff;
1569 memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
1571 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1572 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1573 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1575 } else if (frame_type == FRAME_B) {
1576 int refpiclist0_max = h264_maxref & 0xffff;
1577 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1579 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1580 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1581 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1582 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1585 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1586 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1587 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1588 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1592 slice_param.slice_alpha_c0_offset_div2 = 0;
1593 slice_param.slice_beta_offset_div2 = 0;
1594 slice_param.direct_spatial_mv_pred_flag = 1;
1595 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1598 if (h264_packedheader &&
1599 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1600 render_packedslice();
1602 va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1603 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1604 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1606 render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1613 void QuickSyncEncoderImpl::save_codeddata(storage_task task)
1615 VACodedBufferSegment *buf_list = NULL;
1620 va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1621 CHECK_VASTATUS(va_status, "vaMapBuffer");
1622 while (buf_list != NULL) {
1623 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1624 buf_list = (VACodedBufferSegment *) buf_list->next;
1626 vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1631 memset(&pkt, 0, sizeof(pkt));
1633 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1634 pkt.size = data.size();
1635 pkt.stream_index = 0;
1636 if (task.frame_type == FRAME_IDR) {
1637 pkt.flags = AV_PKT_FLAG_KEY;
1641 pkt.duration = task.duration;
1643 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1645 if (!global_flags.uncompressed_video_to_http &&
1646 !global_flags.x264_video_to_http) {
1647 stream_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1653 // this is weird. but it seems to put a new frame onto the queue
1654 void QuickSyncEncoderImpl::storage_task_enqueue(storage_task task)
1656 unique_lock<mutex> lock(storage_task_queue_mutex);
1657 storage_task_queue.push(move(task));
1658 storage_task_queue_changed.notify_all();
1661 void QuickSyncEncoderImpl::storage_task_thread()
1664 storage_task current;
1666 // wait until there's an encoded frame
1667 unique_lock<mutex> lock(storage_task_queue_mutex);
1668 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1669 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1670 current = move(storage_task_queue.front());
1671 storage_task_queue.pop();
1676 // waits for data, then saves it to disk.
1677 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1678 CHECK_VASTATUS(va_status, "vaSyncSurface");
1679 save_codeddata(move(current));
1682 unique_lock<mutex> lock(storage_task_queue_mutex);
1683 srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1684 storage_task_queue_changed.notify_all();
1689 void QuickSyncEncoderImpl::release_encode()
1691 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1692 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1693 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1694 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1697 vaDestroyContext(va_dpy, context_id);
1698 vaDestroyConfig(va_dpy, config_id);
1701 void QuickSyncEncoderImpl::release_gl_resources()
1703 assert(is_shutdown);
1704 if (has_released_gl_resources) {
1708 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1709 if (!use_zerocopy) {
1710 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1711 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1712 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1713 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1715 resource_pool->release_2d_texture(gl_surfaces[i].y_tex);
1716 resource_pool->release_2d_texture(gl_surfaces[i].cbcr_tex);
1719 has_released_gl_resources = true;
1722 int QuickSyncEncoderImpl::deinit_va()
1724 vaTerminate(va_dpy);
1726 va_close_display(va_dpy);
1731 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)
1732 : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height)
1734 file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat));
1735 open_output_file(filename);
1736 file_audio_encoder->add_mux(file_mux.get());
1738 frame_width_mbaligned = (frame_width + 15) & (~15);
1739 frame_height_mbaligned = (frame_height + 15) & (~15);
1743 if (global_flags.uncompressed_video_to_http ||
1744 global_flags.x264_video_to_http) {
1745 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1747 if (global_flags.x264_video_to_http) {
1748 assert(x264_encoder != nullptr);
1750 assert(x264_encoder == nullptr);
1753 init_va(va_display);
1756 // No frames are ready yet.
1757 memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1759 memset(&seq_param, 0, sizeof(seq_param));
1760 memset(&pic_param, 0, sizeof(pic_param));
1761 memset(&slice_param, 0, sizeof(slice_param));
1763 storage_thread = thread(&QuickSyncEncoderImpl::storage_task_thread, this);
1765 encode_thread = thread([this]{
1766 //SDL_GL_MakeCurrent(window, context);
1767 QOpenGLContext *context = create_context(this->surface);
1768 eglBindAPI(EGL_OPENGL_API);
1769 if (!make_current(context, this->surface)) {
1770 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1774 encode_thread_func();
1775 delete_context(context);
1779 QuickSyncEncoderImpl::~QuickSyncEncoderImpl()
1782 release_gl_resources();
1785 bool QuickSyncEncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1787 assert(!is_shutdown);
1789 // Wait until this frame slot is done encoding.
1790 unique_lock<mutex> lock(storage_task_queue_mutex);
1791 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1792 fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1793 current_storage_frame % SURFACE_NUM, current_storage_frame);
1795 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1796 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1797 if (storage_thread_should_quit) return false;
1800 //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1801 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1802 *y_tex = surf->y_tex;
1803 *cbcr_tex = surf->cbcr_tex;
1805 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1806 CHECK_VASTATUS(va_status, "vaDeriveImage");
1809 VABufferInfo buf_info;
1810 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1811 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1812 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1815 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1816 EGLint y_attribs[] = {
1817 EGL_WIDTH, frame_width,
1818 EGL_HEIGHT, frame_height,
1819 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1820 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1821 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1822 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1826 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1827 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1829 // Associate Y image to a texture.
1830 glBindTexture(GL_TEXTURE_2D, *y_tex);
1831 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1833 // Create CbCr image.
1834 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1835 EGLint cbcr_attribs[] = {
1836 EGL_WIDTH, frame_width,
1837 EGL_HEIGHT, frame_height,
1838 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1839 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1840 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1841 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1845 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1846 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1848 // Associate CbCr image to a texture.
1849 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
1850 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
1856 void QuickSyncEncoderImpl::add_audio(int64_t pts, vector<float> audio)
1858 assert(!is_shutdown);
1859 file_audio_encoder->encode_audio(audio, pts + global_delay());
1862 RefCountedGLsync QuickSyncEncoderImpl::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
1864 assert(!is_shutdown);
1866 if (!use_zerocopy) {
1867 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1869 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
1872 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
1875 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
1877 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
1880 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
1882 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
1885 glBindTexture(GL_TEXTURE_2D, 0);
1887 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1890 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
1894 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
1896 glFlush(); // Make the H.264 thread see the fence as soon as possible.
1900 unique_lock<mutex> lock(frame_queue_mutex);
1901 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts, duration };
1902 ++current_storage_frame;
1904 frame_queue_nonempty.notify_all();
1908 void QuickSyncEncoderImpl::shutdown()
1915 unique_lock<mutex> lock(frame_queue_mutex);
1916 encode_thread_should_quit = true;
1917 frame_queue_nonempty.notify_all();
1919 encode_thread.join();
1921 unique_lock<mutex> lock(storage_task_queue_mutex);
1922 storage_thread_should_quit = true;
1923 frame_queue_nonempty.notify_all();
1924 storage_task_queue_changed.notify_all();
1926 storage_thread.join();
1928 // Encode any leftover audio in the queues, and also any delayed frames.
1929 file_audio_encoder->encode_last_audio();
1937 void QuickSyncEncoderImpl::open_output_file(const std::string &filename)
1939 AVFormatContext *avctx = avformat_alloc_context();
1940 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
1941 assert(filename.size() < sizeof(avctx->filename) - 1);
1942 strcpy(avctx->filename, filename.c_str());
1944 string url = "file:" + filename;
1945 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
1947 char tmp[AV_ERROR_MAX_STRING_SIZE];
1948 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
1952 string video_extradata = ""; // FIXME: See other comment about global headers.
1953 AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters();
1954 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE));
1957 void QuickSyncEncoderImpl::encode_thread_func()
1959 int64_t last_dts = -1;
1960 int gop_start_display_frame_num = 0;
1961 for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
1964 int frame_type, display_frame_num;
1965 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
1966 &display_frame_num, &frame_type, &pts_lag);
1967 if (frame_type == FRAME_IDR) {
1969 current_frame_num = 0;
1970 gop_start_display_frame_num = display_frame_num;
1974 unique_lock<mutex> lock(frame_queue_mutex);
1975 frame_queue_nonempty.wait(lock, [this, display_frame_num]{
1976 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
1978 if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
1979 // We have queued frames that were supposed to be B-frames,
1980 // but will be no P-frame to encode them against. Encode them all
1981 // as P-frames instead. Note that this happens under the mutex,
1982 // but nobody else uses it at this point, since we're shutting down,
1983 // so there's no contention.
1984 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
1987 frame = move(pending_video_frames[display_frame_num]);
1988 pending_video_frames.erase(display_frame_num);
1992 // Determine the dts of this frame.
1994 if (pts_lag == -1) {
1995 assert(last_dts != -1);
1996 dts = last_dts + (TIMEBASE / MAX_FPS);
1998 dts = frame.pts - pts_lag;
2002 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts, frame.duration);
2006 void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2008 if (pending_video_frames.empty()) {
2012 for (auto &pending_frame : pending_video_frames) {
2013 int display_frame_num = pending_frame.first;
2014 assert(display_frame_num > 0);
2015 PendingFrame frame = move(pending_frame.second);
2016 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2017 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2018 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts, frame.duration);
2022 if (global_flags.uncompressed_video_to_http ||
2023 global_flags.x264_video_to_http) {
2024 // Add frames left in reorderer.
2025 while (!reorderer->empty()) {
2026 FrameReorderer::Frame output_frame = reorderer->get_first_frame();
2027 if (global_flags.uncompressed_video_to_http) {
2028 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2030 assert(global_flags.x264_video_to_http);
2031 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2037 void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data)
2040 memset(&pkt, 0, sizeof(pkt));
2042 pkt.data = const_cast<uint8_t *>(data);
2043 pkt.size = frame_width * frame_height * 2;
2044 pkt.stream_index = 0;
2045 pkt.flags = AV_PKT_FLAG_KEY;
2046 pkt.duration = duration;
2047 stream_mux->add_packet(pkt, pts, pts);
2052 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2054 if (src_width == dst_pitch) {
2055 memcpy(dst, src, src_width * height);
2057 for (size_t y = 0; y < height; ++y) {
2058 const uint8_t *sptr = src + y * src_width;
2059 uint8_t *dptr = dst + y * dst_pitch;
2060 memcpy(dptr, sptr, src_width);
2067 void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2068 int frame_type, int64_t pts, int64_t dts, int64_t duration)
2070 // Wait for the GPU to be done with the frame.
2073 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2075 } while (sync_status == GL_TIMEOUT_EXPIRED);
2076 assert(sync_status != GL_WAIT_FAILED);
2078 // Release back any input frames we needed to render this frame.
2079 frame.input_frames.clear();
2081 GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2085 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2086 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2087 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2088 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2090 unsigned char *surface_p = nullptr;
2091 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2093 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2094 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2096 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2097 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2099 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2100 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2102 if (global_flags.uncompressed_video_to_http ||
2103 global_flags.x264_video_to_http) {
2104 // Add uncompressed video. (Note that pts == dts here.)
2105 // Delay needs to match audio.
2106 FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast<uint8_t *>(surf->y_ptr));
2107 if (output_frame.data != nullptr) {
2108 if (global_flags.uncompressed_video_to_http) {
2109 add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data);
2111 assert(global_flags.x264_video_to_http);
2112 x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data);
2118 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2119 CHECK_VASTATUS(va_status, "vaDestroyImage");
2121 // Schedule the frame for encoding.
2122 VASurfaceID va_surface = surf->src_surface;
2123 va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2124 CHECK_VASTATUS(va_status, "vaBeginPicture");
2126 if (frame_type == FRAME_IDR) {
2127 // FIXME: If the mux wants global headers, we should not put the
2128 // SPS/PPS before each IDR frame, but rather put it into the
2129 // codec extradata (formatted differently?).
2131 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2132 if (h264_packedheader) {
2133 render_packedsequence();
2134 render_packedpicture();
2137 //render_sequence();
2138 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2140 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2142 va_status = vaEndPicture(va_dpy, context_id);
2143 CHECK_VASTATUS(va_status, "vaEndPicture");
2145 // so now the data is done encoding (well, async job kicked off)...
2146 // we send that to the storage thread
2148 tmp.display_order = display_frame_num;
2149 tmp.frame_type = frame_type;
2152 tmp.duration = duration;
2153 storage_task_enqueue(move(tmp));
2155 update_ReferenceFrames(frame_type);
2159 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)
2160 : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder)) {}
2162 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2163 QuickSyncEncoder::~QuickSyncEncoder() {}
2165 void QuickSyncEncoder::add_audio(int64_t pts, vector<float> audio)
2167 impl->add_audio(pts, audio);
2170 bool QuickSyncEncoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2172 return impl->begin_frame(y_tex, cbcr_tex);
2175 RefCountedGLsync QuickSyncEncoder::end_frame(int64_t pts, int64_t duration, const vector<RefCountedFrame> &input_frames)
2177 return impl->end_frame(pts, duration, input_frames);
2180 void QuickSyncEncoder::shutdown()
2185 void QuickSyncEncoder::set_stream_mux(Mux *mux)
2187 impl->set_stream_mux(mux);
2190 int64_t QuickSyncEncoder::global_delay() const {
2191 return impl->global_delay();