2 #include "h264encode.h"
4 #include <movit/util.h>
5 #include <EGL/eglplatform.h>
11 #include <libavcodec/avcodec.h>
12 #include <libavformat/avformat.h>
13 #include <libavresample/avresample.h>
14 #include <libavutil/channel_layout.h>
15 #include <libavutil/frame.h>
16 #include <libavutil/rational.h>
17 #include <libavutil/samplefmt.h>
18 #include <libavutil/opt.h>
20 #include <libdrm/drm_fourcc.h>
26 #include <va/va_drm.h>
27 #include <va/va_drmcommon.h>
28 #include <va/va_enc_h264.h>
29 #include <va/va_x11.h>
31 #include <condition_variable>
46 #include "x264encode.h"
53 #define CHECK_VASTATUS(va_status, func) \
54 if (va_status != VA_STATUS_SUCCESS) { \
55 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
59 #define BUFFER_OFFSET(i) ((char *)NULL + (i))
61 //#include "loadsurface.h"
63 #define NAL_REF_IDC_NONE 0
64 #define NAL_REF_IDC_LOW 1
65 #define NAL_REF_IDC_MEDIUM 2
66 #define NAL_REF_IDC_HIGH 3
74 #define SLICE_TYPE_P 0
75 #define SLICE_TYPE_B 1
76 #define SLICE_TYPE_I 2
77 #define IS_P_SLICE(type) (SLICE_TYPE_P == (type))
78 #define IS_B_SLICE(type) (SLICE_TYPE_B == (type))
79 #define IS_I_SLICE(type) (SLICE_TYPE_I == (type))
82 #define ENTROPY_MODE_CAVLC 0
83 #define ENTROPY_MODE_CABAC 1
85 #define PROFILE_IDC_BASELINE 66
86 #define PROFILE_IDC_MAIN 77
87 #define PROFILE_IDC_HIGH 100
89 #define BITSTREAM_ALLOCATE_STEPPING 4096
90 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
91 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
92 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
94 static constexpr unsigned int MaxFrameNum = (2<<16);
95 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
96 static constexpr unsigned int Log2MaxFrameNum = 16;
97 static constexpr unsigned int Log2MaxPicOrderCntLsb = 8;
98 static constexpr int rc_default_modes[] = { // Priority list of modes.
101 VA_RC_VBR_CONSTRAINED,
107 /* thread to save coded data */
108 #define SRC_SURFACE_FREE 0
109 #define SRC_SURFACE_IN_ENCODING 1
112 unsigned int *buffer;
114 int max_size_in_dword;
116 typedef struct __bitstream bitstream;
120 // H.264 video comes out in encoding order (e.g. with two B-frames:
121 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
122 // come in the right order. Since we do everything, including waiting
123 // for the frames to come out of OpenGL, in encoding order, we need
124 // a reordering buffer for uncompressed frames so that they come out
125 // correctly. We go the super-lazy way of not making it understand
126 // anything about the true order (which introduces some extra latency,
127 // though); we know that for N B-frames we need at most (N-1) frames
128 // in the reorder buffer, and can just sort on that.
130 // The class also deals with keeping a freelist as needed.
131 class FrameReorderer {
133 FrameReorderer(unsigned queue_length, int width, int height);
135 // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
136 // Does _not_ take ownership of data; a copy is taken if needed.
137 // The returned pointer is valid until the next call to reorder_frame, or destruction.
138 // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
139 pair<int64_t, const uint8_t *> reorder_frame(int64_t pts, const uint8_t *data);
141 // The same as reorder_frame, but without inserting anything. Used to empty the queue.
142 pair<int64_t, const uint8_t *> get_first_frame();
144 bool empty() const { return frames.empty(); }
147 unsigned queue_length;
150 priority_queue<pair<int64_t, uint8_t *>> frames;
151 stack<uint8_t *> freelist; // Includes the last value returned from reorder_frame.
153 // Owns all the pointers. Normally, freelist and frames could do this themselves,
154 // except priority_queue doesn't work well with movable-only types.
155 vector<unique_ptr<uint8_t[]>> owner;
158 FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
159 : queue_length(queue_length), width(width), height(height)
161 for (unsigned i = 0; i < queue_length; ++i) {
162 owner.emplace_back(new uint8_t[width * height * 2]);
163 freelist.push(owner.back().get());
167 pair<int64_t, const uint8_t *> FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data)
169 if (queue_length == 0) {
170 return make_pair(pts, data);
173 assert(!freelist.empty());
174 uint8_t *storage = freelist.top();
176 memcpy(storage, data, width * height * 2);
177 frames.emplace(-pts, storage); // Invert pts to get smallest first.
179 if (frames.size() >= queue_length) {
180 return get_first_frame();
182 return make_pair(-1, nullptr);
186 pair<int64_t, const uint8_t *> FrameReorderer::get_first_frame()
188 assert(!frames.empty());
189 pair<int64_t, uint8_t *> storage = frames.top();
191 int64_t pts = storage.first;
192 freelist.push(storage.second);
193 return make_pair(-pts, storage.second); // Re-invert pts (see reorder_frame()).
196 class H264EncoderImpl {
198 H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
200 void add_audio(int64_t pts, vector<float> audio);
201 bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
202 RefCountedGLsync end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames);
204 void open_output_file(const std::string &filename);
205 void close_output_file();
208 struct storage_task {
209 unsigned long long display_order;
214 struct PendingFrame {
215 RefCountedGLsync fence;
216 vector<RefCountedFrame> input_frames;
220 // So we never get negative dts.
221 int64_t global_delay() const {
222 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
225 void encode_thread_func();
226 void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
227 void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data);
228 void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
229 int frame_type, int64_t pts, int64_t dts);
230 void storage_task_thread();
231 void encode_audio(const vector<float> &audio,
232 vector<float> *audio_queue,
235 AVAudioResampleContext *resampler,
236 const vector<PacketDestination *> &destinations);
237 void encode_audio_one_frame(const float *audio,
238 size_t num_samples, // In each channel.
241 AVAudioResampleContext *resampler,
242 const vector<PacketDestination *> &destinations);
243 void storage_task_enqueue(storage_task task);
244 void save_codeddata(storage_task task);
245 int render_packedsequence();
246 int render_packedpicture();
247 void render_packedslice();
248 int render_sequence();
249 int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
250 void sps_rbsp(bitstream *bs);
251 void pps_rbsp(bitstream *bs);
252 int build_packed_pic_buffer(unsigned char **header_buffer);
253 int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
254 void slice_header(bitstream *bs);
255 int build_packed_seq_buffer(unsigned char **header_buffer);
256 int build_packed_slice_buffer(unsigned char **header_buffer);
257 int init_va(const string &va_display);
259 void enable_zerocopy_if_possible();
260 VADisplay va_open_display(const string &va_display);
261 void va_close_display(VADisplay va_dpy);
263 int release_encode();
264 void update_ReferenceFrames(int frame_type);
265 int update_RefPicList(int frame_type);
267 bool is_shutdown = false;
271 thread encode_thread, storage_thread;
273 mutex storage_task_queue_mutex;
274 condition_variable storage_task_queue_changed;
275 int srcsurface_status[SURFACE_NUM]; // protected by storage_task_queue_mutex
276 queue<storage_task> storage_task_queue; // protected by storage_task_queue_mutex
277 bool storage_thread_should_quit = false; // protected by storage_task_queue_mutex
279 mutex frame_queue_mutex;
280 condition_variable frame_queue_nonempty;
281 bool encode_thread_should_quit = false; // under frame_queue_mutex
283 int current_storage_frame;
285 map<int, PendingFrame> pending_video_frames; // under frame_queue_mutex
286 map<int64_t, vector<float>> pending_audio_frames; // under frame_queue_mutex
289 AVCodecContext *context_audio_file;
290 AVCodecContext *context_audio_stream = nullptr; // nullptr = don't code separate audio for stream.
292 AVAudioResampleContext *resampler_audio_file;
293 AVAudioResampleContext *resampler_audio_stream;
295 vector<float> audio_queue_file;
296 vector<float> audio_queue_stream;
298 AVFrame *audio_frame = nullptr;
300 unique_ptr<FrameReorderer> reorderer;
301 unique_ptr<X264Encoder> x264_encoder; // nullptr if not using x264.
303 Display *x11_display = nullptr;
305 // Encoder parameters
307 VAProfile h264_profile = (VAProfile)~0;
308 VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
309 int config_attrib_num = 0, enc_packed_header_idx;
312 VASurfaceID src_surface, ref_surface;
313 VABufferID coded_buf;
315 VAImage surface_image;
316 GLuint y_tex, cbcr_tex;
318 // Only if use_zerocopy == true.
319 EGLImage y_egl_image, cbcr_egl_image;
321 // Only if use_zerocopy == false.
323 uint8_t *y_ptr, *cbcr_ptr;
324 size_t y_offset, cbcr_offset;
326 GLSurface gl_surfaces[SURFACE_NUM];
328 VAConfigID config_id;
329 VAContextID context_id;
330 VAEncSequenceParameterBufferH264 seq_param;
331 VAEncPictureParameterBufferH264 pic_param;
332 VAEncSliceParameterBufferH264 slice_param;
333 VAPictureH264 CurrentCurrPic;
334 VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
336 // Static quality settings.
337 static constexpr unsigned int frame_bitrate = 15000000 / 60; // Doesn't really matter; only initial_qp does.
338 static constexpr unsigned int num_ref_frames = 2;
339 static constexpr int initial_qp = 15;
340 static constexpr int minimal_qp = 0;
341 static constexpr int intra_period = 30;
342 static constexpr int intra_idr_period = MAX_FPS; // About a second; more at lower frame rates. Not ideal.
344 // Quality settings that are meant to be static, but might be overridden
346 int constraint_set_flag = 0;
347 int h264_packedheader = 0; /* support pack header? */
348 int h264_maxref = (1<<16|1);
349 int h264_entropy_mode = 1; /* cabac */
353 unsigned int current_frame_num = 0;
354 unsigned int numShortTerm = 0;
358 int frame_width_mbaligned;
359 int frame_height_mbaligned;
361 unique_ptr<Mux> file_mux; // To local disk.
364 // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
365 // but if we don't delete it here, we get leaks. The GStreamer implementation
367 static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
369 VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
370 CHECK_VASTATUS(va_status, "vaRenderPicture");
372 for (int i = 0; i < num_buffers; ++i) {
373 va_status = vaDestroyBuffer(dpy, buffers[i]);
374 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
379 va_swap32(unsigned int val)
381 unsigned char *pval = (unsigned char *)&val;
383 return ((pval[0] << 24) |
390 bitstream_start(bitstream *bs)
392 bs->max_size_in_dword = BITSTREAM_ALLOCATE_STEPPING;
393 bs->buffer = (unsigned int *)calloc(bs->max_size_in_dword * sizeof(int), 1);
398 bitstream_end(bitstream *bs)
400 int pos = (bs->bit_offset >> 5);
401 int bit_offset = (bs->bit_offset & 0x1f);
402 int bit_left = 32 - bit_offset;
405 bs->buffer[pos] = va_swap32((bs->buffer[pos] << bit_left));
410 bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits)
412 int pos = (bs->bit_offset >> 5);
413 int bit_offset = (bs->bit_offset & 0x1f);
414 int bit_left = 32 - bit_offset;
419 bs->bit_offset += size_in_bits;
421 if (bit_left > size_in_bits) {
422 bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val);
424 size_in_bits -= bit_left;
425 if (bit_left >= 32) {
426 bs->buffer[pos] = (val >> size_in_bits);
428 bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits);
430 bs->buffer[pos] = va_swap32(bs->buffer[pos]);
432 if (pos + 1 == bs->max_size_in_dword) {
433 bs->max_size_in_dword += BITSTREAM_ALLOCATE_STEPPING;
434 bs->buffer = (unsigned int *)realloc(bs->buffer, bs->max_size_in_dword * sizeof(unsigned int));
437 bs->buffer[pos + 1] = val;
442 bitstream_put_ue(bitstream *bs, unsigned int val)
444 int size_in_bits = 0;
452 bitstream_put_ui(bs, 0, size_in_bits - 1); // leading zero
453 bitstream_put_ui(bs, val, size_in_bits);
457 bitstream_put_se(bitstream *bs, int val)
459 unsigned int new_val;
464 new_val = 2 * val - 1;
466 bitstream_put_ue(bs, new_val);
470 bitstream_byte_aligning(bitstream *bs, int bit)
472 int bit_offset = (bs->bit_offset & 0x7);
473 int bit_left = 8 - bit_offset;
479 assert(bit == 0 || bit == 1);
482 new_val = (1 << bit_left) - 1;
486 bitstream_put_ui(bs, new_val, bit_left);
490 rbsp_trailing_bits(bitstream *bs)
492 bitstream_put_ui(bs, 1, 1);
493 bitstream_byte_aligning(bs, 0);
496 static void nal_start_code_prefix(bitstream *bs)
498 bitstream_put_ui(bs, 0x00000001, 32);
501 static void nal_header(bitstream *bs, int nal_ref_idc, int nal_unit_type)
503 bitstream_put_ui(bs, 0, 1); /* forbidden_zero_bit: 0 */
504 bitstream_put_ui(bs, nal_ref_idc, 2);
505 bitstream_put_ui(bs, nal_unit_type, 5);
508 void H264EncoderImpl::sps_rbsp(bitstream *bs)
510 int profile_idc = PROFILE_IDC_BASELINE;
512 if (h264_profile == VAProfileH264High)
513 profile_idc = PROFILE_IDC_HIGH;
514 else if (h264_profile == VAProfileH264Main)
515 profile_idc = PROFILE_IDC_MAIN;
517 bitstream_put_ui(bs, profile_idc, 8); /* profile_idc */
518 bitstream_put_ui(bs, !!(constraint_set_flag & 1), 1); /* constraint_set0_flag */
519 bitstream_put_ui(bs, !!(constraint_set_flag & 2), 1); /* constraint_set1_flag */
520 bitstream_put_ui(bs, !!(constraint_set_flag & 4), 1); /* constraint_set2_flag */
521 bitstream_put_ui(bs, !!(constraint_set_flag & 8), 1); /* constraint_set3_flag */
522 bitstream_put_ui(bs, 0, 4); /* reserved_zero_4bits */
523 bitstream_put_ui(bs, seq_param.level_idc, 8); /* level_idc */
524 bitstream_put_ue(bs, seq_param.seq_parameter_set_id); /* seq_parameter_set_id */
526 if ( profile_idc == PROFILE_IDC_HIGH) {
527 bitstream_put_ue(bs, 1); /* chroma_format_idc = 1, 4:2:0 */
528 bitstream_put_ue(bs, 0); /* bit_depth_luma_minus8 */
529 bitstream_put_ue(bs, 0); /* bit_depth_chroma_minus8 */
530 bitstream_put_ui(bs, 0, 1); /* qpprime_y_zero_transform_bypass_flag */
531 bitstream_put_ui(bs, 0, 1); /* seq_scaling_matrix_present_flag */
534 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_frame_num_minus4); /* log2_max_frame_num_minus4 */
535 bitstream_put_ue(bs, seq_param.seq_fields.bits.pic_order_cnt_type); /* pic_order_cnt_type */
537 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0)
538 bitstream_put_ue(bs, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4); /* log2_max_pic_order_cnt_lsb_minus4 */
543 bitstream_put_ue(bs, seq_param.max_num_ref_frames); /* num_ref_frames */
544 bitstream_put_ui(bs, 0, 1); /* gaps_in_frame_num_value_allowed_flag */
546 bitstream_put_ue(bs, seq_param.picture_width_in_mbs - 1); /* pic_width_in_mbs_minus1 */
547 bitstream_put_ue(bs, seq_param.picture_height_in_mbs - 1); /* pic_height_in_map_units_minus1 */
548 bitstream_put_ui(bs, seq_param.seq_fields.bits.frame_mbs_only_flag, 1); /* frame_mbs_only_flag */
550 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
554 bitstream_put_ui(bs, seq_param.seq_fields.bits.direct_8x8_inference_flag, 1); /* direct_8x8_inference_flag */
555 bitstream_put_ui(bs, seq_param.frame_cropping_flag, 1); /* frame_cropping_flag */
557 if (seq_param.frame_cropping_flag) {
558 bitstream_put_ue(bs, seq_param.frame_crop_left_offset); /* frame_crop_left_offset */
559 bitstream_put_ue(bs, seq_param.frame_crop_right_offset); /* frame_crop_right_offset */
560 bitstream_put_ue(bs, seq_param.frame_crop_top_offset); /* frame_crop_top_offset */
561 bitstream_put_ue(bs, seq_param.frame_crop_bottom_offset); /* frame_crop_bottom_offset */
564 //if ( frame_bit_rate < 0 ) { //TODO EW: the vui header isn't correct
566 bitstream_put_ui(bs, 0, 1); /* vui_parameters_present_flag */
568 bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
569 bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
570 bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
571 bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
573 bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
574 bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
575 bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
577 bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
578 bitstream_put_ui(bs, 2, 8); /* transfer_characteristics (2 = unspecified, since we use sRGB) */
579 bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
582 bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
583 bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
585 bitstream_put_ui(bs, 1, 32); // FPS
586 bitstream_put_ui(bs, TIMEBASE * 2, 32); // FPS
587 bitstream_put_ui(bs, 1, 1);
589 bitstream_put_ui(bs, 1, 1); /* nal_hrd_parameters_present_flag */
592 bitstream_put_ue(bs, 0); /* cpb_cnt_minus1 */
593 bitstream_put_ui(bs, 4, 4); /* bit_rate_scale */
594 bitstream_put_ui(bs, 6, 4); /* cpb_size_scale */
596 bitstream_put_ue(bs, frame_bitrate - 1); /* bit_rate_value_minus1[0] */
597 bitstream_put_ue(bs, frame_bitrate*8 - 1); /* cpb_size_value_minus1[0] */
598 bitstream_put_ui(bs, 1, 1); /* cbr_flag[0] */
600 bitstream_put_ui(bs, 23, 5); /* initial_cpb_removal_delay_length_minus1 */
601 bitstream_put_ui(bs, 23, 5); /* cpb_removal_delay_length_minus1 */
602 bitstream_put_ui(bs, 23, 5); /* dpb_output_delay_length_minus1 */
603 bitstream_put_ui(bs, 23, 5); /* time_offset_length */
605 bitstream_put_ui(bs, 0, 1); /* vcl_hrd_parameters_present_flag */
606 bitstream_put_ui(bs, 0, 1); /* low_delay_hrd_flag */
608 bitstream_put_ui(bs, 0, 1); /* pic_struct_present_flag */
609 bitstream_put_ui(bs, 0, 1); /* bitstream_restriction_flag */
612 rbsp_trailing_bits(bs); /* rbsp_trailing_bits */
616 void H264EncoderImpl::pps_rbsp(bitstream *bs)
618 bitstream_put_ue(bs, pic_param.pic_parameter_set_id); /* pic_parameter_set_id */
619 bitstream_put_ue(bs, pic_param.seq_parameter_set_id); /* seq_parameter_set_id */
621 bitstream_put_ui(bs, pic_param.pic_fields.bits.entropy_coding_mode_flag, 1); /* entropy_coding_mode_flag */
623 bitstream_put_ui(bs, 0, 1); /* pic_order_present_flag: 0 */
625 bitstream_put_ue(bs, 0); /* num_slice_groups_minus1 */
627 bitstream_put_ue(bs, pic_param.num_ref_idx_l0_active_minus1); /* num_ref_idx_l0_active_minus1 */
628 bitstream_put_ue(bs, pic_param.num_ref_idx_l1_active_minus1); /* num_ref_idx_l1_active_minus1 1 */
630 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_pred_flag, 1); /* weighted_pred_flag: 0 */
631 bitstream_put_ui(bs, pic_param.pic_fields.bits.weighted_bipred_idc, 2); /* weighted_bipred_idc: 0 */
633 bitstream_put_se(bs, pic_param.pic_init_qp - 26); /* pic_init_qp_minus26 */
634 bitstream_put_se(bs, 0); /* pic_init_qs_minus26 */
635 bitstream_put_se(bs, 0); /* chroma_qp_index_offset */
637 bitstream_put_ui(bs, pic_param.pic_fields.bits.deblocking_filter_control_present_flag, 1); /* deblocking_filter_control_present_flag */
638 bitstream_put_ui(bs, 0, 1); /* constrained_intra_pred_flag */
639 bitstream_put_ui(bs, 0, 1); /* redundant_pic_cnt_present_flag */
642 bitstream_put_ui(bs, pic_param.pic_fields.bits.transform_8x8_mode_flag, 1); /*transform_8x8_mode_flag */
643 bitstream_put_ui(bs, 0, 1); /* pic_scaling_matrix_present_flag */
644 bitstream_put_se(bs, pic_param.second_chroma_qp_index_offset ); /*second_chroma_qp_index_offset */
646 rbsp_trailing_bits(bs);
649 void H264EncoderImpl::slice_header(bitstream *bs)
651 int first_mb_in_slice = slice_param.macroblock_address;
653 bitstream_put_ue(bs, first_mb_in_slice); /* first_mb_in_slice: 0 */
654 bitstream_put_ue(bs, slice_param.slice_type); /* slice_type */
655 bitstream_put_ue(bs, slice_param.pic_parameter_set_id); /* pic_parameter_set_id: 0 */
656 bitstream_put_ui(bs, pic_param.frame_num, seq_param.seq_fields.bits.log2_max_frame_num_minus4 + 4); /* frame_num */
658 /* frame_mbs_only_flag == 1 */
659 if (!seq_param.seq_fields.bits.frame_mbs_only_flag) {
664 if (pic_param.pic_fields.bits.idr_pic_flag)
665 bitstream_put_ue(bs, slice_param.idr_pic_id); /* idr_pic_id: 0 */
667 if (seq_param.seq_fields.bits.pic_order_cnt_type == 0) {
668 bitstream_put_ui(bs, pic_param.CurrPic.TopFieldOrderCnt, seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 + 4);
669 /* pic_order_present_flag == 0 */
675 /* redundant_pic_cnt_present_flag == 0 */
677 if (IS_P_SLICE(slice_param.slice_type)) {
678 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
680 if (slice_param.num_ref_idx_active_override_flag)
681 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
683 /* ref_pic_list_reordering */
684 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
685 } else if (IS_B_SLICE(slice_param.slice_type)) {
686 bitstream_put_ui(bs, slice_param.direct_spatial_mv_pred_flag, 1); /* direct_spatial_mv_pred: 1 */
688 bitstream_put_ui(bs, slice_param.num_ref_idx_active_override_flag, 1); /* num_ref_idx_active_override_flag: */
690 if (slice_param.num_ref_idx_active_override_flag) {
691 bitstream_put_ue(bs, slice_param.num_ref_idx_l0_active_minus1);
692 bitstream_put_ue(bs, slice_param.num_ref_idx_l1_active_minus1);
695 /* ref_pic_list_reordering */
696 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l0: 0 */
697 bitstream_put_ui(bs, 0, 1); /* ref_pic_list_reordering_flag_l1: 0 */
700 if ((pic_param.pic_fields.bits.weighted_pred_flag &&
701 IS_P_SLICE(slice_param.slice_type)) ||
702 ((pic_param.pic_fields.bits.weighted_bipred_idc == 1) &&
703 IS_B_SLICE(slice_param.slice_type))) {
704 /* FIXME: fill weight/offset table */
708 /* dec_ref_pic_marking */
709 if (pic_param.pic_fields.bits.reference_pic_flag) { /* nal_ref_idc != 0 */
710 unsigned char no_output_of_prior_pics_flag = 0;
711 unsigned char long_term_reference_flag = 0;
712 unsigned char adaptive_ref_pic_marking_mode_flag = 0;
714 if (pic_param.pic_fields.bits.idr_pic_flag) {
715 bitstream_put_ui(bs, no_output_of_prior_pics_flag, 1); /* no_output_of_prior_pics_flag: 0 */
716 bitstream_put_ui(bs, long_term_reference_flag, 1); /* long_term_reference_flag: 0 */
718 bitstream_put_ui(bs, adaptive_ref_pic_marking_mode_flag, 1); /* adaptive_ref_pic_marking_mode_flag: 0 */
722 if (pic_param.pic_fields.bits.entropy_coding_mode_flag &&
723 !IS_I_SLICE(slice_param.slice_type))
724 bitstream_put_ue(bs, slice_param.cabac_init_idc); /* cabac_init_idc: 0 */
726 bitstream_put_se(bs, slice_param.slice_qp_delta); /* slice_qp_delta: 0 */
728 /* ignore for SP/SI */
730 if (pic_param.pic_fields.bits.deblocking_filter_control_present_flag) {
731 bitstream_put_ue(bs, slice_param.disable_deblocking_filter_idc); /* disable_deblocking_filter_idc: 0 */
733 if (slice_param.disable_deblocking_filter_idc != 1) {
734 bitstream_put_se(bs, slice_param.slice_alpha_c0_offset_div2); /* slice_alpha_c0_offset_div2: 2 */
735 bitstream_put_se(bs, slice_param.slice_beta_offset_div2); /* slice_beta_offset_div2: 2 */
739 if (pic_param.pic_fields.bits.entropy_coding_mode_flag) {
740 bitstream_byte_aligning(bs, 1);
744 int H264EncoderImpl::build_packed_pic_buffer(unsigned char **header_buffer)
748 bitstream_start(&bs);
749 nal_start_code_prefix(&bs);
750 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_PPS);
754 *header_buffer = (unsigned char *)bs.buffer;
755 return bs.bit_offset;
759 H264EncoderImpl::build_packed_seq_buffer(unsigned char **header_buffer)
763 bitstream_start(&bs);
764 nal_start_code_prefix(&bs);
765 nal_header(&bs, NAL_REF_IDC_HIGH, NAL_SPS);
769 *header_buffer = (unsigned char *)bs.buffer;
770 return bs.bit_offset;
773 int H264EncoderImpl::build_packed_slice_buffer(unsigned char **header_buffer)
776 int is_idr = !!pic_param.pic_fields.bits.idr_pic_flag;
777 int is_ref = !!pic_param.pic_fields.bits.reference_pic_flag;
779 bitstream_start(&bs);
780 nal_start_code_prefix(&bs);
782 if (IS_I_SLICE(slice_param.slice_type)) {
783 nal_header(&bs, NAL_REF_IDC_HIGH, is_idr ? NAL_IDR : NAL_NON_IDR);
784 } else if (IS_P_SLICE(slice_param.slice_type)) {
785 nal_header(&bs, NAL_REF_IDC_MEDIUM, NAL_NON_IDR);
787 assert(IS_B_SLICE(slice_param.slice_type));
788 nal_header(&bs, is_ref ? NAL_REF_IDC_LOW : NAL_REF_IDC_NONE, NAL_NON_IDR);
794 *header_buffer = (unsigned char *)bs.buffer;
795 return bs.bit_offset;
800 Assume frame sequence is: Frame#0, #1, #2, ..., #M, ..., #X, ... (encoding order)
801 1) period between Frame #X and Frame #N = #X - #N
802 2) 0 means infinite for intra_period/intra_idr_period, and 0 is invalid for ip_period
803 3) intra_idr_period % intra_period (intra_period > 0) and intra_period % ip_period must be 0
804 4) intra_period and intra_idr_period take precedence over ip_period
805 5) if ip_period > 1, intra_period and intra_idr_period are not the strict periods
806 of I/IDR frames, see bellow examples
807 -------------------------------------------------------------------
808 intra_period intra_idr_period ip_period frame sequence (intra_period/intra_idr_period/ip_period)
809 0 ignored 1 IDRPPPPPPP ... (No IDR/I any more)
810 0 ignored >=2 IDR(PBB)(PBB)... (No IDR/I any more)
811 1 0 ignored IDRIIIIIII... (No IDR any more)
812 1 1 ignored IDR IDR IDR IDR...
813 1 >=2 ignored IDRII IDRII IDR... (1/3/ignore)
814 >=2 0 1 IDRPPP IPPP I... (3/0/1)
815 >=2 0 >=2 IDR(PBB)(PBB)(IBB) (6/0/3)
816 (PBB)(IBB)(PBB)(IBB)...
817 >=2 >=2 1 IDRPPPPP IPPPPP IPPPPP (6/18/1)
818 IDRPPPPP IPPPPP IPPPPP...
819 >=2 >=2 >=2 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)} (6/18/3)
820 {IDR(PBB)(PBB)(IBB)(PBB)(IBB)(PBB)}...
821 {IDR(PBB)(PBB)(IBB)(PBB)} (6/12/3)
822 {IDR(PBB)(PBB)(IBB)(PBB)}...
823 {IDR(PBB)(PBB)} (6/6/3)
827 // General pts/dts strategy:
829 // Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
830 // bit tricky. We assume first of all that the frame rate never goes _above_
831 // MAX_FPS, which gives us a frame period N. The decoder can always decode
832 // in at least this speed, as long at dts <= pts (the frame is not attempted
833 // presented before it is decoded). Furthermore, we never have longer chains of
834 // B-frames than a fixed constant C. (In a B-frame chain, we say that the base
835 // I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
836 // etc. The last frame in the chain, which no B-frames depend on, is the “tip”
837 // frame, with an order O <= C.)
839 // Many strategies are possible, but we establish these rules:
841 // - Tip frames have dts = pts - (C-O)*N.
842 // - Non-tip frames have dts = dts_last + N.
844 // An example, with C=2 and N=10 and the data flow showed with arrows:
847 // pts: 30 40 50 60 70 80
849 // dts: 10 30 20 60 50←40
854 // To show that this works fine also with irregular spacings, let's say that
855 // the third frame is delayed a bit (something earlier was dropped). Now the
856 // situation looks like this:
859 // pts: 30 40 80 90 100 110
861 // dts: 10 30 20 90 50←40
866 // The resetting on every tip frame makes sure dts never ends up lagging a lot
867 // behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
869 // In the output of this function, if <dts_lag> is >= 0, it means to reset the
870 // dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
871 // a tip frame and should be given a dts based on the previous one.
876 void encoding2display_order(
877 int encoding_order, int intra_period,
878 int intra_idr_period, int ip_period,
879 int *displaying_order,
880 int *frame_type, int *pts_lag)
882 int encoding_order_gop = 0;
886 if (intra_period == 1) { /* all are I/IDR frames */
887 *displaying_order = encoding_order;
888 if (intra_idr_period == 0)
889 *frame_type = (encoding_order == 0)?FRAME_IDR:FRAME_I;
891 *frame_type = (encoding_order % intra_idr_period == 0)?FRAME_IDR:FRAME_I;
895 if (intra_period == 0)
896 intra_idr_period = 0;
898 if (ip_period == 1) {
899 // No B-frames, sequence is like IDR PPPPP IPPPPP.
900 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
901 *displaying_order = encoding_order;
903 if (encoding_order_gop == 0) { /* the first frame */
904 *frame_type = FRAME_IDR;
905 } else if (intra_period != 0 && /* have I frames */
906 encoding_order_gop >= 2 &&
907 (encoding_order_gop % intra_period == 0)) {
908 *frame_type = FRAME_I;
910 *frame_type = FRAME_P;
915 // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
916 encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
917 *pts_lag = -1; // Most frames are not tip frames.
919 if (encoding_order_gop == 0) { /* the first frame */
920 *frame_type = FRAME_IDR;
921 *displaying_order = encoding_order;
922 // IDR frames are a special case; I honestly can't find the logic behind
923 // why this is the right thing, but it seems to line up nicely in practice :-)
924 *pts_lag = TIMEBASE / MAX_FPS;
925 } else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
926 *frame_type = FRAME_B;
927 *displaying_order = encoding_order - 1;
928 if ((encoding_order_gop % ip_period) == 0) {
929 *pts_lag = 0; // Last B-frame.
931 } else if (intra_period != 0 && /* have I frames */
932 encoding_order_gop >= 2 &&
933 ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
934 *frame_type = FRAME_I;
935 *displaying_order = encoding_order + ip_period - 1;
937 *frame_type = FRAME_P;
938 *displaying_order = encoding_order + ip_period - 1;
943 static const char *rc_to_string(int rc_mode)
956 case VA_RC_VBR_CONSTRAINED:
957 return "VBR_CONSTRAINED";
963 void H264EncoderImpl::enable_zerocopy_if_possible()
965 if (global_flags.uncompressed_video_to_http) {
966 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
967 use_zerocopy = false;
968 } else if (global_flags.x264_video_to_http) {
969 fprintf(stderr, "Disabling zerocopy H.264 encoding due to --x264_video_to_http.\n");
970 use_zerocopy = false;
976 VADisplay H264EncoderImpl::va_open_display(const string &va_display)
978 if (va_display.empty()) {
979 x11_display = XOpenDisplay(NULL);
981 fprintf(stderr, "error: can't connect to X server!\n");
984 enable_zerocopy_if_possible();
985 return vaGetDisplay(x11_display);
986 } else if (va_display[0] != '/') {
987 x11_display = XOpenDisplay(va_display.c_str());
989 fprintf(stderr, "error: can't connect to X server!\n");
992 enable_zerocopy_if_possible();
993 return vaGetDisplay(x11_display);
995 drm_fd = open(va_display.c_str(), O_RDWR);
997 perror(va_display.c_str());
1000 use_zerocopy = false;
1001 return vaGetDisplayDRM(drm_fd);
1005 void H264EncoderImpl::va_close_display(VADisplay va_dpy)
1008 XCloseDisplay(x11_display);
1009 x11_display = nullptr;
1016 int H264EncoderImpl::init_va(const string &va_display)
1018 VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline};
1019 VAEntrypoint *entrypoints;
1020 int num_entrypoints, slice_entrypoint;
1021 int support_encode = 0;
1022 int major_ver, minor_ver;
1026 va_dpy = va_open_display(va_display);
1027 va_status = vaInitialize(va_dpy, &major_ver, &minor_ver);
1028 CHECK_VASTATUS(va_status, "vaInitialize");
1030 num_entrypoints = vaMaxNumEntrypoints(va_dpy);
1031 entrypoints = (VAEntrypoint *)malloc(num_entrypoints * sizeof(*entrypoints));
1033 fprintf(stderr, "error: failed to initialize VA entrypoints array\n");
1037 /* use the highest profile */
1038 for (i = 0; i < sizeof(profile_list)/sizeof(profile_list[0]); i++) {
1039 if ((h264_profile != ~0) && h264_profile != profile_list[i])
1042 h264_profile = profile_list[i];
1043 vaQueryConfigEntrypoints(va_dpy, h264_profile, entrypoints, &num_entrypoints);
1044 for (slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
1045 if (entrypoints[slice_entrypoint] == VAEntrypointEncSlice) {
1050 if (support_encode == 1)
1054 if (support_encode == 0) {
1055 printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
1056 printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
1057 printf("to use VA-API against DRM instead of X11.\n");
1060 switch (h264_profile) {
1061 case VAProfileH264Baseline:
1063 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1064 h264_entropy_mode = 0;
1066 case VAProfileH264ConstrainedBaseline:
1067 constraint_set_flag |= (1 << 0 | 1 << 1); /* Annex A.2.2 */
1071 case VAProfileH264Main:
1072 constraint_set_flag |= (1 << 1); /* Annex A.2.2 */
1075 case VAProfileH264High:
1076 constraint_set_flag |= (1 << 3); /* Annex A.2.4 */
1079 h264_profile = VAProfileH264Baseline;
1081 constraint_set_flag |= (1 << 0); /* Annex A.2.1 */
1086 VAConfigAttrib attrib[VAConfigAttribTypeMax];
1088 /* find out the format for the render target, and rate control mode */
1089 for (i = 0; i < VAConfigAttribTypeMax; i++)
1090 attrib[i].type = (VAConfigAttribType)i;
1092 va_status = vaGetConfigAttributes(va_dpy, h264_profile, VAEntrypointEncSlice,
1093 &attrib[0], VAConfigAttribTypeMax);
1094 CHECK_VASTATUS(va_status, "vaGetConfigAttributes");
1095 /* check the interested configattrib */
1096 if ((attrib[VAConfigAttribRTFormat].value & VA_RT_FORMAT_YUV420) == 0) {
1097 printf("Not find desired YUV420 RT format\n");
1100 config_attrib[config_attrib_num].type = VAConfigAttribRTFormat;
1101 config_attrib[config_attrib_num].value = VA_RT_FORMAT_YUV420;
1102 config_attrib_num++;
1105 if (attrib[VAConfigAttribRateControl].value != VA_ATTRIB_NOT_SUPPORTED) {
1106 int tmp = attrib[VAConfigAttribRateControl].value;
1108 if (rc_mode == -1 || !(rc_mode & tmp)) {
1109 if (rc_mode != -1) {
1110 printf("Warning: Don't support the specified RateControl mode: %s!!!, switch to ", rc_to_string(rc_mode));
1113 for (i = 0; i < sizeof(rc_default_modes) / sizeof(rc_default_modes[0]); i++) {
1114 if (rc_default_modes[i] & tmp) {
1115 rc_mode = rc_default_modes[i];
1121 config_attrib[config_attrib_num].type = VAConfigAttribRateControl;
1122 config_attrib[config_attrib_num].value = rc_mode;
1123 config_attrib_num++;
1127 if (attrib[VAConfigAttribEncPackedHeaders].value != VA_ATTRIB_NOT_SUPPORTED) {
1128 int tmp = attrib[VAConfigAttribEncPackedHeaders].value;
1130 h264_packedheader = 1;
1131 config_attrib[config_attrib_num].type = VAConfigAttribEncPackedHeaders;
1132 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1134 if (tmp & VA_ENC_PACKED_HEADER_SEQUENCE) {
1135 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SEQUENCE;
1138 if (tmp & VA_ENC_PACKED_HEADER_PICTURE) {
1139 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_PICTURE;
1142 if (tmp & VA_ENC_PACKED_HEADER_SLICE) {
1143 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_SLICE;
1146 if (tmp & VA_ENC_PACKED_HEADER_MISC) {
1147 config_attrib[config_attrib_num].value |= VA_ENC_PACKED_HEADER_MISC;
1150 enc_packed_header_idx = config_attrib_num;
1151 config_attrib_num++;
1154 if (attrib[VAConfigAttribEncInterlaced].value != VA_ATTRIB_NOT_SUPPORTED) {
1155 config_attrib[config_attrib_num].type = VAConfigAttribEncInterlaced;
1156 config_attrib[config_attrib_num].value = VA_ENC_PACKED_HEADER_NONE;
1157 config_attrib_num++;
1160 if (attrib[VAConfigAttribEncMaxRefFrames].value != VA_ATTRIB_NOT_SUPPORTED) {
1161 h264_maxref = attrib[VAConfigAttribEncMaxRefFrames].value;
1168 int H264EncoderImpl::setup_encode()
1171 VASurfaceID *tmp_surfaceid;
1172 int codedbuf_size, i;
1173 static VASurfaceID src_surface[SURFACE_NUM];
1174 static VASurfaceID ref_surface[SURFACE_NUM];
1176 va_status = vaCreateConfig(va_dpy, h264_profile, VAEntrypointEncSlice,
1177 &config_attrib[0], config_attrib_num, &config_id);
1178 CHECK_VASTATUS(va_status, "vaCreateConfig");
1180 /* create source surfaces */
1181 va_status = vaCreateSurfaces(va_dpy,
1182 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1183 &src_surface[0], SURFACE_NUM,
1185 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1187 /* create reference surfaces */
1188 va_status = vaCreateSurfaces(va_dpy,
1189 VA_RT_FORMAT_YUV420, frame_width_mbaligned, frame_height_mbaligned,
1190 &ref_surface[0], SURFACE_NUM,
1192 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
1194 tmp_surfaceid = (VASurfaceID *)calloc(2 * SURFACE_NUM, sizeof(VASurfaceID));
1195 memcpy(tmp_surfaceid, src_surface, SURFACE_NUM * sizeof(VASurfaceID));
1196 memcpy(tmp_surfaceid + SURFACE_NUM, ref_surface, SURFACE_NUM * sizeof(VASurfaceID));
1198 /* Create a context for this encode pipe */
1199 va_status = vaCreateContext(va_dpy, config_id,
1200 frame_width_mbaligned, frame_height_mbaligned,
1202 tmp_surfaceid, 2 * SURFACE_NUM,
1204 CHECK_VASTATUS(va_status, "vaCreateContext");
1205 free(tmp_surfaceid);
1207 codedbuf_size = (frame_width_mbaligned * frame_height_mbaligned * 400) / (16*16);
1209 for (i = 0; i < SURFACE_NUM; i++) {
1210 /* create coded buffer once for all
1211 * other VA buffers which won't be used again after vaRenderPicture.
1212 * so APP can always vaCreateBuffer for every frame
1213 * but coded buffer need to be mapped and accessed after vaRenderPicture/vaEndPicture
1214 * so VA won't maintain the coded buffer
1216 va_status = vaCreateBuffer(va_dpy, context_id, VAEncCodedBufferType,
1217 codedbuf_size, 1, NULL, &gl_surfaces[i].coded_buf);
1218 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1221 /* create OpenGL objects */
1222 //glGenFramebuffers(SURFACE_NUM, fbos);
1224 for (i = 0; i < SURFACE_NUM; i++) {
1225 glGenTextures(1, &gl_surfaces[i].y_tex);
1226 glGenTextures(1, &gl_surfaces[i].cbcr_tex);
1228 if (!use_zerocopy) {
1230 glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex);
1231 glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height);
1233 // Create CbCr image.
1234 glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex);
1235 glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2);
1237 // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API
1238 // buffers, due to potentially differing pitch.
1239 glGenBuffers(1, &gl_surfaces[i].pbo);
1240 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1241 glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1242 uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
1243 gl_surfaces[i].y_offset = 0;
1244 gl_surfaces[i].cbcr_offset = frame_width * frame_height;
1245 gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset;
1246 gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset;
1247 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1251 for (i = 0; i < SURFACE_NUM; i++) {
1252 gl_surfaces[i].src_surface = src_surface[i];
1253 gl_surfaces[i].ref_surface = ref_surface[i];
1259 // Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce
1262 template<class T, class C>
1263 static void sort_two(T *begin, T *end, const T &pivot, const C &less_than)
1265 T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); });
1266 sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); });
1267 sort(middle, end, less_than);
1270 void H264EncoderImpl::update_ReferenceFrames(int frame_type)
1274 if (frame_type == FRAME_B)
1277 CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE;
1279 if (numShortTerm > num_ref_frames)
1280 numShortTerm = num_ref_frames;
1281 for (i=numShortTerm-1; i>0; i--)
1282 ReferenceFrames[i] = ReferenceFrames[i-1];
1283 ReferenceFrames[0] = CurrentCurrPic;
1285 current_frame_num++;
1286 if (current_frame_num > MaxFrameNum)
1287 current_frame_num = 0;
1291 int H264EncoderImpl::update_RefPicList(int frame_type)
1293 const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1294 return a.frame_idx > b.frame_idx;
1296 const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1297 return a.TopFieldOrderCnt < b.TopFieldOrderCnt;
1299 const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) {
1300 return a.TopFieldOrderCnt > b.TopFieldOrderCnt;
1303 if (frame_type == FRAME_P) {
1304 memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1305 sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx);
1306 } else if (frame_type == FRAME_B) {
1307 memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1308 sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt);
1310 memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264));
1311 sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt);
1318 int H264EncoderImpl::render_sequence()
1320 VABufferID seq_param_buf, rc_param_buf, render_id[2];
1322 VAEncMiscParameterBuffer *misc_param;
1323 VAEncMiscParameterRateControl *misc_rate_ctrl;
1325 seq_param.level_idc = 41 /*SH_LEVEL_3*/;
1326 seq_param.picture_width_in_mbs = frame_width_mbaligned / 16;
1327 seq_param.picture_height_in_mbs = frame_height_mbaligned / 16;
1328 seq_param.bits_per_second = frame_bitrate;
1330 seq_param.intra_period = intra_period;
1331 seq_param.intra_idr_period = intra_idr_period;
1332 seq_param.ip_period = ip_period;
1334 seq_param.max_num_ref_frames = num_ref_frames;
1335 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1336 seq_param.time_scale = TIMEBASE * 2;
1337 seq_param.num_units_in_tick = 1; /* Tc = num_units_in_tick / scale */
1338 seq_param.seq_fields.bits.log2_max_pic_order_cnt_lsb_minus4 = Log2MaxPicOrderCntLsb - 4;
1339 seq_param.seq_fields.bits.log2_max_frame_num_minus4 = Log2MaxFrameNum - 4;;
1340 seq_param.seq_fields.bits.frame_mbs_only_flag = 1;
1341 seq_param.seq_fields.bits.chroma_format_idc = 1;
1342 seq_param.seq_fields.bits.direct_8x8_inference_flag = 1;
1344 if (frame_width != frame_width_mbaligned ||
1345 frame_height != frame_height_mbaligned) {
1346 seq_param.frame_cropping_flag = 1;
1347 seq_param.frame_crop_left_offset = 0;
1348 seq_param.frame_crop_right_offset = (frame_width_mbaligned - frame_width)/2;
1349 seq_param.frame_crop_top_offset = 0;
1350 seq_param.frame_crop_bottom_offset = (frame_height_mbaligned - frame_height)/2;
1353 va_status = vaCreateBuffer(va_dpy, context_id,
1354 VAEncSequenceParameterBufferType,
1355 sizeof(seq_param), 1, &seq_param, &seq_param_buf);
1356 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1358 va_status = vaCreateBuffer(va_dpy, context_id,
1359 VAEncMiscParameterBufferType,
1360 sizeof(VAEncMiscParameterBuffer) + sizeof(VAEncMiscParameterRateControl),
1361 1, NULL, &rc_param_buf);
1362 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1364 vaMapBuffer(va_dpy, rc_param_buf, (void **)&misc_param);
1365 misc_param->type = VAEncMiscParameterTypeRateControl;
1366 misc_rate_ctrl = (VAEncMiscParameterRateControl *)misc_param->data;
1367 memset(misc_rate_ctrl, 0, sizeof(*misc_rate_ctrl));
1368 misc_rate_ctrl->bits_per_second = frame_bitrate;
1369 misc_rate_ctrl->target_percentage = 66;
1370 misc_rate_ctrl->window_size = 1000;
1371 misc_rate_ctrl->initial_qp = initial_qp;
1372 misc_rate_ctrl->min_qp = minimal_qp;
1373 misc_rate_ctrl->basic_unit_size = 0;
1374 vaUnmapBuffer(va_dpy, rc_param_buf);
1376 render_id[0] = seq_param_buf;
1377 render_id[1] = rc_param_buf;
1379 render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
1384 static int calc_poc(int pic_order_cnt_lsb, int frame_type)
1386 static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0;
1387 int prevPicOrderCntMsb, prevPicOrderCntLsb;
1388 int PicOrderCntMsb, TopFieldOrderCnt;
1390 if (frame_type == FRAME_IDR)
1391 prevPicOrderCntMsb = prevPicOrderCntLsb = 0;
1393 prevPicOrderCntMsb = PicOrderCntMsb_ref;
1394 prevPicOrderCntLsb = pic_order_cnt_lsb_ref;
1397 if ((pic_order_cnt_lsb < prevPicOrderCntLsb) &&
1398 ((prevPicOrderCntLsb - pic_order_cnt_lsb) >= (int)(MaxPicOrderCntLsb / 2)))
1399 PicOrderCntMsb = prevPicOrderCntMsb + MaxPicOrderCntLsb;
1400 else if ((pic_order_cnt_lsb > prevPicOrderCntLsb) &&
1401 ((pic_order_cnt_lsb - prevPicOrderCntLsb) > (int)(MaxPicOrderCntLsb / 2)))
1402 PicOrderCntMsb = prevPicOrderCntMsb - MaxPicOrderCntLsb;
1404 PicOrderCntMsb = prevPicOrderCntMsb;
1406 TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb;
1408 if (frame_type != FRAME_B) {
1409 PicOrderCntMsb_ref = PicOrderCntMsb;
1410 pic_order_cnt_lsb_ref = pic_order_cnt_lsb;
1413 return TopFieldOrderCnt;
1416 int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num)
1418 VABufferID pic_param_buf;
1422 pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface;
1423 pic_param.CurrPic.frame_idx = current_frame_num;
1424 pic_param.CurrPic.flags = 0;
1425 pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type);
1426 pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt;
1427 CurrentCurrPic = pic_param.CurrPic;
1429 memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
1430 for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
1431 pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
1432 pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
1435 pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR);
1436 pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B);
1437 pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode;
1438 pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1;
1439 pic_param.frame_num = current_frame_num;
1440 pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf;
1441 pic_param.last_picture = false; // FIXME
1442 pic_param.pic_init_qp = initial_qp;
1444 va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
1445 sizeof(pic_param), 1, &pic_param, &pic_param_buf);
1446 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1448 render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
1453 int H264EncoderImpl::render_packedsequence()
1455 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1456 VABufferID packedseq_para_bufid, packedseq_data_bufid, render_id[2];
1457 unsigned int length_in_bits;
1458 unsigned char *packedseq_buffer = NULL;
1461 length_in_bits = build_packed_seq_buffer(&packedseq_buffer);
1463 packedheader_param_buffer.type = VAEncPackedHeaderSequence;
1465 packedheader_param_buffer.bit_length = length_in_bits; /*length_in_bits*/
1466 packedheader_param_buffer.has_emulation_bytes = 0;
1467 va_status = vaCreateBuffer(va_dpy,
1469 VAEncPackedHeaderParameterBufferType,
1470 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1471 &packedseq_para_bufid);
1472 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1474 va_status = vaCreateBuffer(va_dpy,
1476 VAEncPackedHeaderDataBufferType,
1477 (length_in_bits + 7) / 8, 1, packedseq_buffer,
1478 &packedseq_data_bufid);
1479 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1481 render_id[0] = packedseq_para_bufid;
1482 render_id[1] = packedseq_data_bufid;
1483 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1485 free(packedseq_buffer);
1491 int H264EncoderImpl::render_packedpicture()
1493 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1494 VABufferID packedpic_para_bufid, packedpic_data_bufid, render_id[2];
1495 unsigned int length_in_bits;
1496 unsigned char *packedpic_buffer = NULL;
1499 length_in_bits = build_packed_pic_buffer(&packedpic_buffer);
1500 packedheader_param_buffer.type = VAEncPackedHeaderPicture;
1501 packedheader_param_buffer.bit_length = length_in_bits;
1502 packedheader_param_buffer.has_emulation_bytes = 0;
1504 va_status = vaCreateBuffer(va_dpy,
1506 VAEncPackedHeaderParameterBufferType,
1507 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1508 &packedpic_para_bufid);
1509 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1511 va_status = vaCreateBuffer(va_dpy,
1513 VAEncPackedHeaderDataBufferType,
1514 (length_in_bits + 7) / 8, 1, packedpic_buffer,
1515 &packedpic_data_bufid);
1516 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1518 render_id[0] = packedpic_para_bufid;
1519 render_id[1] = packedpic_data_bufid;
1520 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1522 free(packedpic_buffer);
1527 void H264EncoderImpl::render_packedslice()
1529 VAEncPackedHeaderParameterBuffer packedheader_param_buffer;
1530 VABufferID packedslice_para_bufid, packedslice_data_bufid, render_id[2];
1531 unsigned int length_in_bits;
1532 unsigned char *packedslice_buffer = NULL;
1535 length_in_bits = build_packed_slice_buffer(&packedslice_buffer);
1536 packedheader_param_buffer.type = VAEncPackedHeaderSlice;
1537 packedheader_param_buffer.bit_length = length_in_bits;
1538 packedheader_param_buffer.has_emulation_bytes = 0;
1540 va_status = vaCreateBuffer(va_dpy,
1542 VAEncPackedHeaderParameterBufferType,
1543 sizeof(packedheader_param_buffer), 1, &packedheader_param_buffer,
1544 &packedslice_para_bufid);
1545 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1547 va_status = vaCreateBuffer(va_dpy,
1549 VAEncPackedHeaderDataBufferType,
1550 (length_in_bits + 7) / 8, 1, packedslice_buffer,
1551 &packedslice_data_bufid);
1552 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1554 render_id[0] = packedslice_para_bufid;
1555 render_id[1] = packedslice_data_bufid;
1556 render_picture_and_delete(va_dpy, context_id, render_id, 2);
1558 free(packedslice_buffer);
1561 int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type)
1563 VABufferID slice_param_buf;
1567 update_RefPicList(frame_type);
1569 /* one frame, one slice */
1570 slice_param.macroblock_address = 0;
1571 slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */
1572 slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type;
1573 if (frame_type == FRAME_IDR) {
1574 if (encoding_frame_num != 0)
1575 ++slice_param.idr_pic_id;
1576 } else if (frame_type == FRAME_P) {
1577 int refpiclist0_max = h264_maxref & 0xffff;
1578 memcpy(slice_param.RefPicList0, RefPicList0_P, 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;
1584 } else if (frame_type == FRAME_B) {
1585 int refpiclist0_max = h264_maxref & 0xffff;
1586 int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
1588 memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
1589 for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
1590 slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
1591 slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
1594 memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
1595 for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
1596 slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
1597 slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
1601 slice_param.slice_alpha_c0_offset_div2 = 0;
1602 slice_param.slice_beta_offset_div2 = 0;
1603 slice_param.direct_spatial_mv_pred_flag = 1;
1604 slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb;
1607 if (h264_packedheader &&
1608 config_attrib[enc_packed_header_idx].value & VA_ENC_PACKED_HEADER_SLICE)
1609 render_packedslice();
1611 va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
1612 sizeof(slice_param), 1, &slice_param, &slice_param_buf);
1613 CHECK_VASTATUS(va_status, "vaCreateBuffer");
1615 render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
1622 void H264EncoderImpl::save_codeddata(storage_task task)
1624 VACodedBufferSegment *buf_list = NULL;
1629 va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
1630 CHECK_VASTATUS(va_status, "vaMapBuffer");
1631 while (buf_list != NULL) {
1632 data.append(reinterpret_cast<const char *>(buf_list->buf), buf_list->size);
1633 buf_list = (VACodedBufferSegment *) buf_list->next;
1635 vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
1640 memset(&pkt, 0, sizeof(pkt));
1642 pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
1643 pkt.size = data.size();
1644 pkt.stream_index = 0;
1645 if (task.frame_type == FRAME_IDR) {
1646 pkt.flags = AV_PKT_FLAG_KEY;
1652 file_mux->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1654 if (!global_flags.uncompressed_video_to_http &&
1655 !global_flags.x264_video_to_http) {
1656 httpd->add_packet(pkt, task.pts + global_delay(), task.dts + global_delay());
1659 // Encode and add all audio frames up to and including the pts of this video frame.
1662 vector<float> audio;
1664 unique_lock<mutex> lock(frame_queue_mutex);
1665 frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); });
1666 if (storage_thread_should_quit && pending_audio_frames.empty()) return;
1667 auto it = pending_audio_frames.begin();
1668 if (it->first > task.pts) break;
1669 audio_pts = it->first;
1670 audio = move(it->second);
1671 pending_audio_frames.erase(it);
1674 if (context_audio_stream) {
1675 encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { file_mux.get() });
1676 encode_audio(audio, &audio_queue_stream, audio_pts, context_audio_stream, resampler_audio_stream, { httpd });
1678 encode_audio(audio, &audio_queue_file, audio_pts, context_audio_file, resampler_audio_file, { httpd, file_mux.get() });
1681 if (audio_pts == task.pts) break;
1685 void H264EncoderImpl::encode_audio(
1686 const vector<float> &audio,
1687 vector<float> *audio_queue,
1689 AVCodecContext *ctx,
1690 AVAudioResampleContext *resampler,
1691 const vector<PacketDestination *> &destinations)
1693 if (ctx->frame_size == 0) {
1694 // No queueing needed.
1695 assert(audio_queue->empty());
1696 assert(audio.size() % 2 == 0);
1697 encode_audio_one_frame(&audio[0], audio.size() / 2, audio_pts, ctx, resampler, destinations);
1701 audio_queue->insert(audio_queue->end(), audio.begin(), audio.end());
1703 for (sample_num = 0;
1704 sample_num + ctx->frame_size * 2 <= audio_queue->size();
1705 sample_num += ctx->frame_size * 2) {
1706 encode_audio_one_frame(&(*audio_queue)[sample_num],
1708 audio_pts, // FIXME: Must be increased or decreased as needed.
1713 audio_queue->erase(audio_queue->begin(), audio_queue->begin() + sample_num);
1716 void H264EncoderImpl::encode_audio_one_frame(
1720 AVCodecContext *ctx,
1721 AVAudioResampleContext *resampler,
1722 const vector<PacketDestination *> &destinations)
1724 audio_frame->nb_samples = num_samples;
1725 audio_frame->channel_layout = AV_CH_LAYOUT_STEREO;
1726 audio_frame->format = ctx->sample_fmt;
1727 audio_frame->sample_rate = OUTPUT_FREQUENCY;
1729 if (av_samples_alloc(audio_frame->data, nullptr, 2, num_samples, ctx->sample_fmt, 0) < 0) {
1730 fprintf(stderr, "Could not allocate %ld samples.\n", num_samples);
1734 if (avresample_convert(resampler, audio_frame->data, 0, num_samples,
1735 (uint8_t **)&audio, 0, num_samples) < 0) {
1736 fprintf(stderr, "Audio conversion failed.\n");
1741 av_init_packet(&pkt);
1745 avcodec_encode_audio2(ctx, &pkt, audio_frame, &got_output);
1747 pkt.stream_index = 1;
1748 pkt.flags = AV_PKT_FLAG_KEY;
1749 for (PacketDestination *dest : destinations) {
1750 dest->add_packet(pkt, audio_pts + global_delay(), audio_pts + global_delay());
1753 // TODO: Delayed frames.
1754 av_frame_unref(audio_frame);
1755 av_free_packet(&pkt);
1757 av_freep(&audio_frame->data[0]);
1758 av_freep(&audio_frame->linesize[0]);
1761 // this is weird. but it seems to put a new frame onto the queue
1762 void H264EncoderImpl::storage_task_enqueue(storage_task task)
1764 unique_lock<mutex> lock(storage_task_queue_mutex);
1765 storage_task_queue.push(move(task));
1766 storage_task_queue_changed.notify_all();
1769 void H264EncoderImpl::storage_task_thread()
1772 storage_task current;
1774 // wait until there's an encoded frame
1775 unique_lock<mutex> lock(storage_task_queue_mutex);
1776 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); });
1777 if (storage_thread_should_quit && storage_task_queue.empty()) return;
1778 current = move(storage_task_queue.front());
1779 storage_task_queue.pop();
1784 // waits for data, then saves it to disk.
1785 va_status = vaSyncSurface(va_dpy, gl_surfaces[current.display_order % SURFACE_NUM].src_surface);
1786 CHECK_VASTATUS(va_status, "vaSyncSurface");
1787 save_codeddata(move(current));
1790 unique_lock<mutex> lock(storage_task_queue_mutex);
1791 srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
1792 storage_task_queue_changed.notify_all();
1797 int H264EncoderImpl::release_encode()
1799 for (unsigned i = 0; i < SURFACE_NUM; i++) {
1800 vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf);
1801 vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1);
1802 vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1);
1804 if (!use_zerocopy) {
1805 glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo);
1806 glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
1807 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
1808 glDeleteBuffers(1, &gl_surfaces[i].pbo);
1810 glDeleteTextures(1, &gl_surfaces[i].y_tex);
1811 glDeleteTextures(1, &gl_surfaces[i].cbcr_tex);
1814 vaDestroyContext(va_dpy, context_id);
1815 vaDestroyConfig(va_dpy, config_id);
1820 int H264EncoderImpl::deinit_va()
1822 vaTerminate(va_dpy);
1824 va_close_display(va_dpy);
1831 void init_audio_encoder(const string &codec_name, int bit_rate, AVCodecContext **ctx, AVAudioResampleContext **resampler)
1833 AVCodec *codec_audio = avcodec_find_encoder_by_name(codec_name.c_str());
1834 if (codec_audio == nullptr) {
1835 fprintf(stderr, "ERROR: Could not find codec '%s'\n", codec_name.c_str());
1839 AVCodecContext *context_audio = avcodec_alloc_context3(codec_audio);
1840 context_audio->bit_rate = bit_rate;
1841 context_audio->sample_rate = OUTPUT_FREQUENCY;
1842 context_audio->sample_fmt = codec_audio->sample_fmts[0];
1843 context_audio->channels = 2;
1844 context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
1845 context_audio->time_base = AVRational{1, TIMEBASE};
1846 if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
1847 fprintf(stderr, "Could not open codec '%s'\n", codec_name.c_str());
1851 *ctx = context_audio;
1853 // FIXME: These leak on close.
1854 *resampler = avresample_alloc_context();
1855 if (*resampler == nullptr) {
1856 fprintf(stderr, "Allocating resampler failed.\n");
1860 av_opt_set_int(*resampler, "in_channel_layout", AV_CH_LAYOUT_STEREO, 0);
1861 av_opt_set_int(*resampler, "out_channel_layout", AV_CH_LAYOUT_STEREO, 0);
1862 av_opt_set_int(*resampler, "in_sample_rate", OUTPUT_FREQUENCY, 0);
1863 av_opt_set_int(*resampler, "out_sample_rate", OUTPUT_FREQUENCY, 0);
1864 av_opt_set_int(*resampler, "in_sample_fmt", AV_SAMPLE_FMT_FLT, 0);
1865 av_opt_set_int(*resampler, "out_sample_fmt", context_audio->sample_fmt, 0);
1867 if (avresample_open(*resampler) < 0) {
1868 fprintf(stderr, "Could not open resample context.\n");
1875 H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
1876 : current_storage_frame(0), surface(surface), httpd(httpd)
1878 init_audio_encoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, &context_audio_file, &resampler_audio_file);
1880 if (!global_flags.stream_audio_codec_name.empty()) {
1881 init_audio_encoder(global_flags.stream_audio_codec_name,
1882 global_flags.stream_audio_codec_bitrate, &context_audio_stream, &resampler_audio_stream);
1885 audio_frame = av_frame_alloc();
1887 frame_width = width;
1888 frame_height = height;
1889 frame_width_mbaligned = (frame_width + 15) & (~15);
1890 frame_height_mbaligned = (frame_height + 15) & (~15);
1894 if (global_flags.uncompressed_video_to_http ||
1895 global_flags.x264_video_to_http) {
1896 reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
1898 if (global_flags.x264_video_to_http) {
1899 x264_encoder.reset(new X264Encoder(httpd));
1902 init_va(va_display);
1905 // No frames are ready yet.
1906 memset(srcsurface_status, SRC_SURFACE_FREE, sizeof(srcsurface_status));
1908 memset(&seq_param, 0, sizeof(seq_param));
1909 memset(&pic_param, 0, sizeof(pic_param));
1910 memset(&slice_param, 0, sizeof(slice_param));
1912 storage_thread = thread(&H264EncoderImpl::storage_task_thread, this);
1914 encode_thread = thread([this]{
1915 //SDL_GL_MakeCurrent(window, context);
1916 QOpenGLContext *context = create_context(this->surface);
1917 eglBindAPI(EGL_OPENGL_API);
1918 if (!make_current(context, this->surface)) {
1919 printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
1923 encode_thread_func();
1927 H264EncoderImpl::~H264EncoderImpl()
1930 av_frame_free(&audio_frame);
1932 // TODO: Destroy context.
1935 bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
1937 assert(!is_shutdown);
1939 // Wait until this frame slot is done encoding.
1940 unique_lock<mutex> lock(storage_task_queue_mutex);
1941 if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
1942 fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
1943 current_storage_frame % SURFACE_NUM, current_storage_frame);
1945 storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
1946 srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
1947 if (storage_thread_should_quit) return false;
1950 //*fbo = fbos[current_storage_frame % SURFACE_NUM];
1951 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
1952 *y_tex = surf->y_tex;
1953 *cbcr_tex = surf->cbcr_tex;
1955 VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image);
1956 CHECK_VASTATUS(va_status, "vaDeriveImage");
1959 VABufferInfo buf_info;
1960 buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM?
1961 va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info);
1962 CHECK_VASTATUS(va_status, "vaAcquireBufferHandle");
1965 surf->y_egl_image = EGL_NO_IMAGE_KHR;
1966 EGLint y_attribs[] = {
1967 EGL_WIDTH, frame_width,
1968 EGL_HEIGHT, frame_height,
1969 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '),
1970 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1971 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]),
1972 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]),
1976 surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs);
1977 assert(surf->y_egl_image != EGL_NO_IMAGE_KHR);
1979 // Associate Y image to a texture.
1980 glBindTexture(GL_TEXTURE_2D, *y_tex);
1981 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image);
1983 // Create CbCr image.
1984 surf->cbcr_egl_image = EGL_NO_IMAGE_KHR;
1985 EGLint cbcr_attribs[] = {
1986 EGL_WIDTH, frame_width,
1987 EGL_HEIGHT, frame_height,
1988 EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'),
1989 EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle),
1990 EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]),
1991 EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]),
1995 surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs);
1996 assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR);
1998 // Associate CbCr image to a texture.
1999 glBindTexture(GL_TEXTURE_2D, *cbcr_tex);
2000 glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image);
2006 void H264EncoderImpl::add_audio(int64_t pts, vector<float> audio)
2008 assert(!is_shutdown);
2010 unique_lock<mutex> lock(frame_queue_mutex);
2011 pending_audio_frames[pts] = move(audio);
2013 frame_queue_nonempty.notify_all();
2016 RefCountedGLsync H264EncoderImpl::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2018 assert(!is_shutdown);
2020 if (!use_zerocopy) {
2021 GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM];
2023 glPixelStorei(GL_PACK_ROW_LENGTH, 0);
2026 glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo);
2029 glBindTexture(GL_TEXTURE_2D, surf->y_tex);
2031 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset));
2034 glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex);
2036 glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset));
2039 glBindTexture(GL_TEXTURE_2D, 0);
2041 glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
2044 glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT);
2048 RefCountedGLsync fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
2050 glFlush(); // Make the H.264 thread see the fence as soon as possible.
2054 unique_lock<mutex> lock(frame_queue_mutex);
2055 pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
2056 ++current_storage_frame;
2058 frame_queue_nonempty.notify_all();
2062 void H264EncoderImpl::shutdown()
2069 unique_lock<mutex> lock(frame_queue_mutex);
2070 encode_thread_should_quit = true;
2071 frame_queue_nonempty.notify_all();
2073 encode_thread.join();
2075 unique_lock<mutex> lock(storage_task_queue_mutex);
2076 storage_thread_should_quit = true;
2077 frame_queue_nonempty.notify_all();
2078 storage_task_queue_changed.notify_all();
2080 storage_thread.join();
2087 void H264EncoderImpl::open_output_file(const std::string &filename)
2089 AVFormatContext *avctx = avformat_alloc_context();
2090 avctx->oformat = av_guess_format(NULL, filename.c_str(), NULL);
2091 assert(filename.size() < sizeof(avctx->filename) - 1);
2092 strcpy(avctx->filename, filename.c_str());
2094 string url = "file:" + filename;
2095 int ret = avio_open2(&avctx->pb, url.c_str(), AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL);
2097 char tmp[AV_ERROR_MAX_STRING_SIZE];
2098 fprintf(stderr, "%s: avio_open2() failed: %s\n", filename.c_str(), av_make_error_string(tmp, sizeof(tmp), ret));
2102 file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, context_audio_file->codec, TIMEBASE, DEFAULT_AUDIO_OUTPUT_BIT_RATE));
2105 void H264EncoderImpl::close_output_file()
2110 void H264EncoderImpl::encode_thread_func()
2112 int64_t last_dts = -1;
2113 int gop_start_display_frame_num = 0;
2114 for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
2117 int frame_type, display_frame_num;
2118 encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
2119 &display_frame_num, &frame_type, &pts_lag);
2120 if (frame_type == FRAME_IDR) {
2122 current_frame_num = 0;
2123 gop_start_display_frame_num = display_frame_num;
2127 unique_lock<mutex> lock(frame_queue_mutex);
2128 frame_queue_nonempty.wait(lock, [this, display_frame_num]{
2129 return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0;
2131 if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) {
2132 // We have queued frames that were supposed to be B-frames,
2133 // but will be no P-frame to encode them against. Encode them all
2134 // as P-frames instead. Note that this happens under the mutex,
2135 // but nobody else uses it at this point, since we're shutting down,
2136 // so there's no contention.
2137 encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts);
2140 frame = move(pending_video_frames[display_frame_num]);
2141 pending_video_frames.erase(display_frame_num);
2145 // Determine the dts of this frame.
2147 if (pts_lag == -1) {
2148 assert(last_dts != -1);
2149 dts = last_dts + (TIMEBASE / MAX_FPS);
2151 dts = frame.pts - pts_lag;
2155 encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
2159 void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts)
2161 if (pending_video_frames.empty()) {
2165 for (auto &pending_frame : pending_video_frames) {
2166 int display_frame_num = pending_frame.first;
2167 assert(display_frame_num > 0);
2168 PendingFrame frame = move(pending_frame.second);
2169 int64_t dts = last_dts + (TIMEBASE / MAX_FPS);
2170 printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num);
2171 encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
2175 if (global_flags.uncompressed_video_to_http ||
2176 global_flags.x264_video_to_http) {
2177 // Add frames left in reorderer.
2178 while (!reorderer->empty()) {
2179 pair<int64_t, const uint8_t *> output_frame = reorderer->get_first_frame();
2180 if (global_flags.uncompressed_video_to_http) {
2181 add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2183 assert(global_flags.x264_video_to_http);
2184 x264_encoder->add_frame(output_frame.first, output_frame.second);
2190 void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data)
2193 memset(&pkt, 0, sizeof(pkt));
2195 pkt.data = const_cast<uint8_t *>(data);
2196 pkt.size = frame_width * frame_height * 2;
2197 pkt.stream_index = 0;
2198 pkt.flags = AV_PKT_FLAG_KEY;
2199 httpd->add_packet(pkt, pts, pts);
2204 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
2206 if (src_width == dst_pitch) {
2207 memcpy(dst, src, src_width * height);
2209 for (size_t y = 0; y < height; ++y) {
2210 const uint8_t *sptr = src + y * src_width;
2211 uint8_t *dptr = dst + y * dst_pitch;
2212 memcpy(dptr, sptr, src_width);
2219 void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
2220 int frame_type, int64_t pts, int64_t dts)
2222 // Wait for the GPU to be done with the frame.
2225 sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
2227 } while (sync_status == GL_TIMEOUT_EXPIRED);
2228 assert(sync_status != GL_WAIT_FAILED);
2230 // Release back any input frames we needed to render this frame.
2231 frame.input_frames.clear();
2233 GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM];
2237 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
2238 eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
2239 va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
2240 CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
2242 unsigned char *surface_p = nullptr;
2243 vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p);
2245 unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0];
2246 memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height);
2248 unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1];
2249 memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2);
2251 va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
2252 CHECK_VASTATUS(va_status, "vaUnmapBuffer");
2254 if (global_flags.uncompressed_video_to_http ||
2255 global_flags.x264_video_to_http) {
2256 // Add uncompressed video. (Note that pts == dts here.)
2257 // Delay needs to match audio.
2258 pair<int64_t, const uint8_t *> output_frame = reorderer->reorder_frame(pts + global_delay(), reinterpret_cast<uint8_t *>(surf->y_ptr));
2259 if (output_frame.second != nullptr) {
2260 if (global_flags.uncompressed_video_to_http) {
2261 add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
2263 assert(global_flags.x264_video_to_http);
2264 x264_encoder->add_frame(output_frame.first, output_frame.second);
2270 va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
2271 CHECK_VASTATUS(va_status, "vaDestroyImage");
2273 // Schedule the frame for encoding.
2274 VASurfaceID va_surface = surf->src_surface;
2275 va_status = vaBeginPicture(va_dpy, context_id, va_surface);
2276 CHECK_VASTATUS(va_status, "vaBeginPicture");
2278 if (frame_type == FRAME_IDR) {
2280 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2281 if (h264_packedheader) {
2282 render_packedsequence();
2283 render_packedpicture();
2286 //render_sequence();
2287 render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
2289 render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type);
2291 va_status = vaEndPicture(va_dpy, context_id);
2292 CHECK_VASTATUS(va_status, "vaEndPicture");
2294 // so now the data is done encoding (well, async job kicked off)...
2295 // we send that to the storage thread
2297 tmp.display_order = display_frame_num;
2298 tmp.frame_type = frame_type;
2301 storage_task_enqueue(move(tmp));
2303 update_ReferenceFrames(frame_type);
2307 H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd)
2308 : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {}
2310 // Must be defined here because unique_ptr<> destructor needs to know the impl.
2311 H264Encoder::~H264Encoder() {}
2313 void H264Encoder::add_audio(int64_t pts, vector<float> audio)
2315 impl->add_audio(pts, audio);
2318 bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
2320 return impl->begin_frame(y_tex, cbcr_tex);
2323 RefCountedGLsync H264Encoder::end_frame(int64_t pts, const vector<RefCountedFrame> &input_frames)
2325 return impl->end_frame(pts, input_frames);
2328 void H264Encoder::shutdown()
2333 void H264Encoder::open_output_file(const std::string &filename)
2335 impl->open_output_file(filename);
2338 void H264Encoder::close_output_file()
2340 impl->close_output_file();