X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=h264encode.cpp;h=fbeeba707949284bb0ba982cdd86d46a4b0c01b6;hb=bb8cf44c90b78921e6c6b2f62e232addff78b5ed;hp=96e0165746ef71928ca1ca3b72dfd3bccf877fb4;hpb=ded4cffec7a53c01389a306ebc5caf1a4e965c05;p=nageru diff --git a/h264encode.cpp b/h264encode.cpp index 96e0165..fbeeba7 100644 --- a/h264encode.cpp +++ b/h264encode.cpp @@ -38,6 +38,7 @@ extern "C" { #include "context.h" #include "defs.h" +#include "flags.h" #include "httpd.h" #include "timebase.h" @@ -84,6 +85,8 @@ class QSurface; #define BITSTREAM_ALLOCATE_STEPPING 4096 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */ +#define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM +#define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM static constexpr unsigned int MaxFrameNum = (2<<16); static constexpr unsigned int MaxPicOrderCntLsb = (2<<8); @@ -111,6 +114,82 @@ typedef struct __bitstream bitstream; using namespace std; +// H.264 video comes out in encoding order (e.g. with two B-frames: +// 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to +// come in the right order. Since we do everything, including waiting +// for the frames to come out of OpenGL, in encoding order, we need +// a reordering buffer for uncompressed frames so that they come out +// correctly. We go the super-lazy way of not making it understand +// anything about the true order (which introduces some extra latency, +// though); we know that for N B-frames we need at most (N-1) frames +// in the reorder buffer, and can just sort on that. +// +// The class also deals with keeping a freelist as needed. +class FrameReorderer { +public: + FrameReorderer(unsigned queue_length, int width, int height); + + // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr. + // Does _not_ take ownership of data; a copy is taken if needed. + // The returned pointer is valid until the next call to reorder_frame, or destruction. + // As a special case, if queue_length == 0, will just return pts and data (no reordering needed). + pair reorder_frame(int64_t pts, const uint8_t *data); + + // The same as reorder_frame, but without inserting anything. Used to empty the queue. + pair get_first_frame(); + + bool empty() const { return frames.empty(); } + +private: + unsigned queue_length; + int width, height; + + priority_queue> frames; + stack freelist; // Includes the last value returned from reorder_frame. + + // Owns all the pointers. Normally, freelist and frames could do this themselves, + // except priority_queue doesn't work well with movable-only types. + vector> owner; +}; + +FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height) + : queue_length(queue_length), width(width), height(height) +{ + for (unsigned i = 0; i < queue_length; ++i) { + owner.emplace_back(new uint8_t[width * height * 2]); + freelist.push(owner.back().get()); + } +} + +pair FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data) +{ + if (queue_length == 0) { + return make_pair(pts, data); + } + + assert(!freelist.empty()); + uint8_t *storage = freelist.top(); + freelist.pop(); + memcpy(storage, data, width * height * 2); + frames.emplace(-pts, storage); // Invert pts to get smallest first. + + if (frames.size() >= queue_length) { + return get_first_frame(); + } else { + return make_pair(-1, nullptr); + } +} + +pair FrameReorderer::get_first_frame() +{ + assert(!frames.empty()); + pair storage = frames.top(); + frames.pop(); + int64_t pts = storage.first; + freelist.push(storage.second); + return make_pair(-pts, storage.second); // Re-invert pts (see reorder_frame()). +} + class H264EncoderImpl { public: H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd); @@ -135,6 +214,7 @@ private: void encode_thread_func(); void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts); + void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data); void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type, int64_t pts, int64_t dts); void storage_task_thread(); @@ -184,7 +264,9 @@ private: QSurface *surface; AVCodecContext *context_audio; + AVFrame *audio_frame = nullptr; HTTPD *httpd; + unique_ptr reorderer; Display *x11_display = nullptr; @@ -205,7 +287,6 @@ private: EGLImage y_egl_image, cbcr_egl_image; // Only if use_zerocopy == false. - RefCountedGLsync readback_done_fence; GLuint pbo; uint8_t *y_ptr, *cbcr_ptr; size_t y_offset, cbcr_offset; @@ -218,7 +299,7 @@ private: VAEncPictureParameterBufferH264 pic_param; VAEncSliceParameterBufferH264 slice_param; VAPictureH264 CurrentCurrPic; - VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32]; + VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2]; // Static quality settings. static constexpr unsigned int frame_bitrate = 15000000 / 60; // Doesn't really matter; only initial_qp does. @@ -854,6 +935,10 @@ VADisplay H264EncoderImpl::va_open_display(const string &va_display) return NULL; } use_zerocopy = true; + if (global_flags.uncompressed_video_to_http) { + fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n"); + use_zerocopy = false; + } return vaGetDisplay(x11_display); } else if (va_display[0] != '/') { x11_display = XOpenDisplay(va_display.c_str()); @@ -862,6 +947,10 @@ VADisplay H264EncoderImpl::va_open_display(const string &va_display) return NULL; } use_zerocopy = true; + if (global_flags.uncompressed_video_to_http) { + fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n"); + use_zerocopy = false; + } return vaGetDisplay(x11_display); } else { drm_fd = open(va_display.c_str(), O_RDWR); @@ -925,7 +1014,7 @@ int H264EncoderImpl::init_va(const string &va_display) if (support_encode == 0) { printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n"); - printf("but have one in your system, try launching Nageru with --va-display /dev/dri/card0\n"); + printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n"); printf("to use VA-API against DRM instead of X11.\n"); exit(1); } else { @@ -1112,10 +1201,10 @@ int H264EncoderImpl::setup_encode() glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo); glBufferStorage(GL_PIXEL_PACK_BUFFER, frame_width * frame_height * 2, nullptr, GL_MAP_READ_BIT | GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT); uint8_t *ptr = (uint8_t *)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, frame_width * frame_height * 2, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); - gl_surfaces[i].y_ptr = ptr; - gl_surfaces[i].cbcr_ptr = ptr + frame_width * frame_height; gl_surfaces[i].y_offset = 0; gl_surfaces[i].cbcr_offset = frame_width * frame_height; + gl_surfaces[i].y_ptr = ptr + gl_surfaces[i].y_offset; + gl_surfaces[i].cbcr_ptr = ptr + gl_surfaces[i].cbcr_offset; glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); } } @@ -1299,7 +1388,7 @@ int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int g CurrentCurrPic = pic_param.CurrPic; memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264)); - for (i = numShortTerm; i < SURFACE_NUM; i++) { + for (i = numShortTerm; i < MAX_NUM_REF1; i++) { pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE; pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID; } @@ -1449,7 +1538,7 @@ int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int refpiclist0_max = h264_maxref & 0xffff; memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264)); - for (i = refpiclist0_max; i < 32; i++) { + for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) { slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE; slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID; } @@ -1458,13 +1547,13 @@ int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num, int refpiclist1_max = (h264_maxref >> 16) & 0xffff; memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264)); - for (i = refpiclist0_max; i < 32; i++) { + for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) { slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE; slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID; } memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264)); - for (i = refpiclist1_max; i < 32; i++) { + for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) { slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE; slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID; } @@ -1522,7 +1611,8 @@ void H264EncoderImpl::save_codeddata(storage_task task) pkt.flags = 0; } //pkt.duration = 1; - httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay); + httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay, + global_flags.uncompressed_video_to_http ? HTTPD::DESTINATION_FILE_ONLY : HTTPD::DESTINATION_FILE_AND_HTTP); } // Encode and add all audio frames up to and including the pts of this video frame. for ( ;; ) { @@ -1539,18 +1629,17 @@ void H264EncoderImpl::save_codeddata(storage_task task) pending_audio_frames.erase(it); } - AVFrame *frame = avcodec_alloc_frame(); - frame->nb_samples = audio.size() / 2; - frame->format = AV_SAMPLE_FMT_S32; - frame->channel_layout = AV_CH_LAYOUT_STEREO; + audio_frame->nb_samples = audio.size() / 2; + audio_frame->format = AV_SAMPLE_FMT_S32; + audio_frame->channel_layout = AV_CH_LAYOUT_STEREO; unique_ptr int_samples(new int32_t[audio.size()]); - int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1); + int ret = avcodec_fill_audio_frame(audio_frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1); if (ret < 0) { fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret); exit(1); } - for (int i = 0; i < frame->nb_samples * 2; ++i) { + for (int i = 0; i < audio_frame->nb_samples * 2; ++i) { if (audio[i] >= 1.0f) { int_samples[i] = 2147483647; } else if (audio[i] <= -1.0f) { @@ -1565,35 +1654,16 @@ void H264EncoderImpl::save_codeddata(storage_task task) pkt.data = nullptr; pkt.size = 0; int got_output; - avcodec_encode_audio2(context_audio, &pkt, frame, &got_output); + avcodec_encode_audio2(context_audio, &pkt, audio_frame, &got_output); if (got_output) { pkt.stream_index = 1; - httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay); + httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay, HTTPD::DESTINATION_FILE_AND_HTTP); } // TODO: Delayed frames. - avcodec_free_frame(&frame); + av_frame_unref(audio_frame); av_free_packet(&pkt); if (audio_pts == task.pts) break; } - -#if 0 - printf("\r "); /* return back to startpoint */ - switch (encode_order % 4) { - case 0: - printf("|"); - break; - case 1: - printf("/"); - break; - case 2: - printf("-"); - break; - case 3: - printf("\\"); - break; - } - printf("%08lld", encode_order); -#endif } @@ -1602,7 +1672,6 @@ void H264EncoderImpl::storage_task_enqueue(storage_task task) { unique_lock lock(storage_task_queue_mutex); storage_task_queue.push(move(task)); - srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING; storage_task_queue_changed.notify_all(); } @@ -1682,6 +1751,7 @@ H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, in fprintf(stderr, "Could not open codec\n"); exit(1); } + audio_frame = av_frame_alloc(); frame_width = width; frame_height = height; @@ -1690,6 +1760,10 @@ H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, in //print_input(); + if (global_flags.uncompressed_video_to_http) { + reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height)); + } + init_va(va_display); setup_encode(); @@ -1718,6 +1792,9 @@ H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, in H264EncoderImpl::~H264EncoderImpl() { shutdown(); + av_frame_free(&audio_frame); + + // TODO: Destroy context. } bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) @@ -1726,7 +1803,12 @@ bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) { // Wait until this frame slot is done encoding. unique_lock lock(storage_task_queue_mutex); + if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) { + fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n", + current_storage_frame % SURFACE_NUM, current_storage_frame); + } storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); }); + srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING; if (storage_thread_should_quit) return false; } @@ -1927,6 +2009,26 @@ void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int g encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts); last_dts = dts; } + + if (global_flags.uncompressed_video_to_http) { + // Add frames left in reorderer. + while (!reorderer->empty()) { + pair output_frame = reorderer->get_first_frame(); + add_packet_for_uncompressed_frame(output_frame.first, output_frame.second); + } + } +} + +void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data) +{ + AVPacket pkt; + memset(&pkt, 0, sizeof(pkt)); + pkt.buf = nullptr; + pkt.data = const_cast(data); + pkt.size = frame_width * frame_height * 2; + pkt.stream_index = 0; + pkt.flags = AV_PKT_FLAG_KEY; + httpd->add_packet(pkt, pts, pts, HTTPD::DESTINATION_HTTP_ONLY); } namespace { @@ -1950,7 +2052,12 @@ void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int enco int frame_type, int64_t pts, int64_t dts) { // Wait for the GPU to be done with the frame. - glClientWaitSync(frame.fence.get(), 0, 0); + GLenum sync_status; + do { + sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000); + check_error(); + } while (sync_status == GL_TIMEOUT_EXPIRED); + assert(sync_status != GL_WAIT_FAILED); // Release back any input frames we needed to render this frame. frame.input_frames.clear(); @@ -1967,14 +2074,23 @@ void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int enco unsigned char *surface_p = nullptr; vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p); - unsigned char *y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0]; - memcpy_with_pitch(y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height); + unsigned char *va_y_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[0]; + memcpy_with_pitch(va_y_ptr, surf->y_ptr, frame_width, surf->surface_image.pitches[0], frame_height); - unsigned char *cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1]; - memcpy_with_pitch(cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2); + unsigned char *va_cbcr_ptr = (unsigned char *)surface_p + surf->surface_image.offsets[1]; + memcpy_with_pitch(va_cbcr_ptr, surf->cbcr_ptr, (frame_width / 2) * sizeof(uint16_t), surf->surface_image.pitches[1], frame_height / 2); va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf); CHECK_VASTATUS(va_status, "vaUnmapBuffer"); + + if (global_flags.uncompressed_video_to_http) { + // Add uncompressed video. (Note that pts == dts here.) + const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS); // Needs to match audio. + pair output_frame = reorderer->reorder_frame(pts + global_delay, reinterpret_cast(surf->y_ptr)); + if (output_frame.second != nullptr) { + add_packet_for_uncompressed_frame(output_frame.first, output_frame.second); + } + } } va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);