X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=h264encode.cpp;h=a478c2a1a2459a2fa24d3ed06c34a598ea1bf8d1;hb=71fc6732eae50bb053e3a3618ab53280f72b4fff;hp=7fc62b3a920cde84d44ea68913138aa51acd9764;hpb=8c2ebd3fdae0bd663ed02098a693078e87d49451;p=nageru diff --git a/h264encode.cpp b/h264encode.cpp index 7fc62b3..a478c2a 100644 --- a/h264encode.cpp +++ b/h264encode.cpp @@ -1,6 +1,7 @@ //#include "sysdeps.h" #include "h264encode.h" +#include #include #include #include @@ -18,10 +19,13 @@ extern "C" { #include #include #include +#include #include +#include #include #include #include +#include #include #include #include @@ -48,6 +52,8 @@ class QSurface; exit(1); \ } +#define BUFFER_OFFSET(i) ((char *)NULL + (i)) + //#include "loadsurface.h" #define NAL_REF_IDC_NONE 0 @@ -107,22 +113,23 @@ using namespace std; class H264EncoderImpl { public: - H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd); + H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd); ~H264EncoderImpl(); - void add_audio(int64_t pts, std::vector audio); // Needs to come before end_frame() of same pts. + void add_audio(int64_t pts, vector audio); bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex); - void end_frame(RefCountedGLsync fence, int64_t pts, const std::vector &input_frames); + void end_frame(RefCountedGLsync fence, int64_t pts, const vector &input_frames); + void shutdown(); private: struct storage_task { unsigned long long display_order; int frame_type; - std::vector audio; + vector audio; int64_t pts, dts; }; struct PendingFrame { RefCountedGLsync fence; - std::vector input_frames; + vector input_frames; int64_t pts; }; @@ -145,38 +152,41 @@ private: void slice_header(bitstream *bs); int build_packed_seq_buffer(unsigned char **header_buffer); int build_packed_slice_buffer(unsigned char **header_buffer); - int init_va(); + int init_va(const string &va_display); int deinit_va(); - VADisplay va_open_display(void); + VADisplay va_open_display(const string &va_display); void va_close_display(VADisplay va_dpy); int setup_encode(); int release_encode(); void update_ReferenceFrames(int frame_type); int update_RefPicList(int frame_type); - std::thread encode_thread, storage_thread; + bool is_shutdown = false; + bool use_zerocopy; + int drm_fd = -1; + + thread encode_thread, storage_thread; - std::mutex storage_task_queue_mutex; - std::condition_variable storage_task_queue_changed; + mutex storage_task_queue_mutex; + condition_variable storage_task_queue_changed; int srcsurface_status[SURFACE_NUM]; // protected by storage_task_queue_mutex - std::queue storage_task_queue; // protected by storage_task_queue_mutex + queue storage_task_queue; // protected by storage_task_queue_mutex bool storage_thread_should_quit = false; // protected by storage_task_queue_mutex - std::mutex frame_queue_mutex; - std::condition_variable frame_queue_nonempty; + mutex frame_queue_mutex; + condition_variable frame_queue_nonempty; bool encode_thread_should_quit = false; // under frame_queue_mutex int current_storage_frame; - std::map pending_video_frames; // under frame_queue_mutex - std::map> pending_audio_frames; // under frame_queue_mutex + map pending_video_frames; // under frame_queue_mutex + map> pending_audio_frames; // under frame_queue_mutex QSurface *surface; AVCodecContext *context_audio; HTTPD *httpd; - Display *x11_display; - Window x11_window; + Display *x11_display = nullptr; // Encoder parameters VADisplay va_dpy; @@ -190,7 +200,14 @@ private: VAImage surface_image; GLuint y_tex, cbcr_tex; + + // Only if use_zerocopy == true. EGLImage y_egl_image, cbcr_egl_image; + + // Only if use_zerocopy == false. + GLuint pbo; + uint8_t *y_ptr, *cbcr_ptr; + size_t y_offset, cbcr_offset; }; GLSurface gl_surfaces[SURFACE_NUM]; @@ -228,7 +245,6 @@ private: int frame_height_mbaligned; }; - // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly, // but if we don't delete it here, we get leaks. The GStreamer implementation // does the same. @@ -290,7 +306,11 @@ bitstream_put_ui(bitstream *bs, unsigned int val, int size_in_bits) bs->buffer[pos] = (bs->buffer[pos] << size_in_bits | val); } else { size_in_bits -= bit_left; - bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits); + if (bit_left >= 32) { + bs->buffer[pos] = (val >> size_in_bits); + } else { + bs->buffer[pos] = (bs->buffer[pos] << bit_left) | (val >> size_in_bits); + } bs->buffer[pos] = va_swap32(bs->buffer[pos]); if (pos + 1 == bs->max_size_in_dword) { @@ -824,31 +844,47 @@ static const char *rc_to_string(int rc_mode) } } -VADisplay H264EncoderImpl::va_open_display(void) +VADisplay H264EncoderImpl::va_open_display(const string &va_display) { - x11_display = XOpenDisplay(NULL); - if (!x11_display) { - fprintf(stderr, "error: can't connect to X server!\n"); - return NULL; - } - return vaGetDisplay(x11_display); + if (va_display.empty()) { + x11_display = XOpenDisplay(NULL); + if (!x11_display) { + fprintf(stderr, "error: can't connect to X server!\n"); + return NULL; + } + use_zerocopy = true; + return vaGetDisplay(x11_display); + } else if (va_display[0] != '/') { + x11_display = XOpenDisplay(va_display.c_str()); + if (!x11_display) { + fprintf(stderr, "error: can't connect to X server!\n"); + return NULL; + } + use_zerocopy = true; + return vaGetDisplay(x11_display); + } else { + drm_fd = open(va_display.c_str(), O_RDWR); + if (drm_fd == -1) { + perror(va_display.c_str()); + return NULL; + } + use_zerocopy = false; + return vaGetDisplayDRM(drm_fd); + } } void H264EncoderImpl::va_close_display(VADisplay va_dpy) { - if (!x11_display) - return; - - if (x11_window) { - XUnmapWindow(x11_display, x11_window); - XDestroyWindow(x11_display, x11_window); - x11_window = None; - } - XCloseDisplay(x11_display); - x11_display = NULL; + if (x11_display) { + XCloseDisplay(x11_display); + x11_display = nullptr; + } + if (drm_fd != -1) { + close(drm_fd); + } } -int H264EncoderImpl::init_va() +int H264EncoderImpl::init_va(const string &va_display) { VAProfile profile_list[]={VAProfileH264High, VAProfileH264Main, VAProfileH264Baseline, VAProfileH264ConstrainedBaseline}; VAEntrypoint *entrypoints; @@ -858,7 +894,7 @@ int H264EncoderImpl::init_va() VAStatus va_status; unsigned int i; - va_dpy = va_open_display(); + va_dpy = va_open_display(va_display); va_status = vaInitialize(va_dpy, &major_ver, &minor_ver); CHECK_VASTATUS(va_status, "vaInitialize"); @@ -887,7 +923,9 @@ int H264EncoderImpl::init_va() } if (support_encode == 0) { - printf("Can't find VAEntrypointEncSlice for H264 profiles\n"); + 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/renderD128\n"); + printf("to use VA-API against DRM instead of X11.\n"); exit(1); } else { switch (h264_profile) { @@ -1057,6 +1095,28 @@ int H264EncoderImpl::setup_encode() for (i = 0; i < SURFACE_NUM; i++) { glGenTextures(1, &gl_surfaces[i].y_tex); glGenTextures(1, &gl_surfaces[i].cbcr_tex); + + if (!use_zerocopy) { + // Create Y image. + glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].y_tex); + glTexStorage2D(GL_TEXTURE_2D, 1, GL_R8, frame_width, frame_height); + + // Create CbCr image. + glBindTexture(GL_TEXTURE_2D, gl_surfaces[i].cbcr_tex); + glTexStorage2D(GL_TEXTURE_2D, 1, GL_RG8, frame_width / 2, frame_height / 2); + + // Generate a PBO to read into. It doesn't necessarily fit 1:1 with the VA-API + // buffers, due to potentially differing pitch. + glGenBuffers(1, &gl_surfaces[i].pbo); + 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_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); + } } for (i = 0; i < SURFACE_NUM; i++) { @@ -1067,68 +1127,15 @@ int H264EncoderImpl::setup_encode() return 0; } - - -#define partition(ref, field, key, ascending) \ - while (i <= j) { \ - if (ascending) { \ - while (ref[i].field < key) \ - i++; \ - while (ref[j].field > key) \ - j--; \ - } else { \ - while (ref[i].field > key) \ - i++; \ - while (ref[j].field < key) \ - j--; \ - } \ - if (i <= j) { \ - tmp = ref[i]; \ - ref[i] = ref[j]; \ - ref[j] = tmp; \ - i++; \ - j--; \ - } \ - } \ - -static void sort_one(VAPictureH264 ref[], int left, int right, - int ascending, int frame_idx) -{ - int i = left, j = right; - unsigned int key; - VAPictureH264 tmp; - - if (frame_idx) { - key = ref[(left + right) / 2].frame_idx; - partition(ref, frame_idx, key, ascending); - } else { - key = ref[(left + right) / 2].TopFieldOrderCnt; - partition(ref, TopFieldOrderCnt, (signed int)key, ascending); - } - - /* recursion */ - if (left < j) - sort_one(ref, left, j, ascending, frame_idx); - - if (i < right) - sort_one(ref, i, right, ascending, frame_idx); -} - -static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, unsigned int frame_idx, - int partition_ascending, int list0_ascending, int list1_ascending) +// Given a list like 1 9 3 0 2 8 4 and a pivot element 3, will produce +// +// 2 1 0 [3] 4 8 9 +template +static void sort_two(T *begin, T *end, const T &pivot, const C &less_than) { - int i = left, j = right; - VAPictureH264 tmp; - - if (frame_idx) { - partition(ref, frame_idx, key, partition_ascending); - } else { - partition(ref, TopFieldOrderCnt, (signed int)key, partition_ascending); - } - - - sort_one(ref, left, i-1, list0_ascending, frame_idx); - sort_one(ref, j+1, right, list1_ascending, frame_idx); + T *middle = partition(begin, end, [&](const T &elem) { return less_than(elem, pivot); }); + sort(begin, middle, [&](const T &a, const T &b) { return less_than(b, a); }); + sort(middle, end, less_than); } void H264EncoderImpl::update_ReferenceFrames(int frame_type) @@ -1154,21 +1161,25 @@ void H264EncoderImpl::update_ReferenceFrames(int frame_type) int H264EncoderImpl::update_RefPicList(int frame_type) { - unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt; + const auto descending_by_frame_idx = [](const VAPictureH264 &a, const VAPictureH264 &b) { + return a.frame_idx > b.frame_idx; + }; + const auto ascending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) { + return a.TopFieldOrderCnt < b.TopFieldOrderCnt; + }; + const auto descending_by_top_field_order_cnt = [](const VAPictureH264 &a, const VAPictureH264 &b) { + return a.TopFieldOrderCnt > b.TopFieldOrderCnt; + }; if (frame_type == FRAME_P) { memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264)); - sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1); - } - - if (frame_type == FRAME_B) { + sort(&RefPicList0_P[0], &RefPicList0_P[numShortTerm], descending_by_frame_idx); + } else if (frame_type == FRAME_B) { memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264)); - sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0, - 1, 0, 1); + sort_two(&RefPicList0_B[0], &RefPicList0_B[numShortTerm], CurrentCurrPic, ascending_by_top_field_order_cnt); memcpy(RefPicList1_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264)); - sort_two(RefPicList1_B, 0, numShortTerm-1, current_poc, 0, - 0, 1, 0); + sort_two(&RefPicList1_B[0], &RefPicList1_B[numShortTerm], CurrentCurrPic, descending_by_top_field_order_cnt); } return 0; @@ -1486,7 +1497,7 @@ void H264EncoderImpl::save_codeddata(storage_task task) string data; - const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS); // So we never get negative dts. + const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS); // So we never get negative dts. va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list)); CHECK_VASTATUS(va_status, "vaMapBuffer"); @@ -1624,18 +1635,25 @@ void H264EncoderImpl::storage_task_thread() int H264EncoderImpl::release_encode() { - int i; - - for (i = 0; i < SURFACE_NUM; i++) { - vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf); - vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1); - vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1); - } - - vaDestroyContext(va_dpy, context_id); - vaDestroyConfig(va_dpy, config_id); + for (unsigned i = 0; i < SURFACE_NUM; i++) { + vaDestroyBuffer(va_dpy, gl_surfaces[i].coded_buf); + vaDestroySurfaces(va_dpy, &gl_surfaces[i].src_surface, 1); + vaDestroySurfaces(va_dpy, &gl_surfaces[i].ref_surface, 1); + + if (!use_zerocopy) { + glBindBuffer(GL_PIXEL_PACK_BUFFER, gl_surfaces[i].pbo); + glUnmapBuffer(GL_PIXEL_PACK_BUFFER); + glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); + glDeleteBuffers(1, &gl_surfaces[i].pbo); + } + glDeleteTextures(1, &gl_surfaces[i].y_tex); + glDeleteTextures(1, &gl_surfaces[i].cbcr_tex); + } - return 0; + vaDestroyContext(va_dpy, context_id); + vaDestroyConfig(va_dpy, config_id); + + return 0; } int H264EncoderImpl::deinit_va() @@ -1648,7 +1666,7 @@ int H264EncoderImpl::deinit_va() } -H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd) +H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd) : current_storage_frame(0), surface(surface), httpd(httpd) { AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC); @@ -1671,7 +1689,7 @@ H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD //print_input(); - init_va(); + init_va(va_display); setup_encode(); // No frames are ready yet. @@ -1698,26 +1716,12 @@ H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD H264EncoderImpl::~H264EncoderImpl() { - { - unique_lock lock(frame_queue_mutex); - encode_thread_should_quit = true; - frame_queue_nonempty.notify_all(); - } - encode_thread.join(); - { - unique_lock lock(storage_task_queue_mutex); - storage_thread_should_quit = true; - frame_queue_nonempty.notify_all(); - storage_task_queue_changed.notify_all(); - } - storage_thread.join(); - - release_encode(); - deinit_va(); + shutdown(); } bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) { + assert(!is_shutdown); { // Wait until this frame slot is done encoding. unique_lock lock(storage_task_queue_mutex); @@ -1730,58 +1734,60 @@ bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) *y_tex = surf->y_tex; *cbcr_tex = surf->cbcr_tex; - VASurfaceID surface = surf->src_surface; - VAStatus va_status = vaDeriveImage(va_dpy, surface, &surf->surface_image); - CHECK_VASTATUS(va_status, "vaDeriveImage"); - - VABufferInfo buf_info; - buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM? - va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info); - CHECK_VASTATUS(va_status, "vaAcquireBufferHandle"); - - // Create Y image. - surf->y_egl_image = EGL_NO_IMAGE_KHR; - EGLint y_attribs[] = { - EGL_WIDTH, frame_width, - EGL_HEIGHT, frame_height, - EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '), - EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle), - EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]), - EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]), - EGL_NONE - }; - - surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs); - assert(surf->y_egl_image != EGL_NO_IMAGE_KHR); - - // Associate Y image to a texture. - glBindTexture(GL_TEXTURE_2D, *y_tex); - glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image); - - // Create CbCr image. - surf->cbcr_egl_image = EGL_NO_IMAGE_KHR; - EGLint cbcr_attribs[] = { - EGL_WIDTH, frame_width, - EGL_HEIGHT, frame_height, - EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'), - EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle), - EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]), - EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]), - EGL_NONE - }; - - surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs); - assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR); - - // Associate CbCr image to a texture. - glBindTexture(GL_TEXTURE_2D, *cbcr_tex); - glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image); + VAStatus va_status = vaDeriveImage(va_dpy, surf->src_surface, &surf->surface_image); + CHECK_VASTATUS(va_status, "vaDeriveImage"); + + if (use_zerocopy) { + VABufferInfo buf_info; + buf_info.mem_type = VA_SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME; // or VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM? + va_status = vaAcquireBufferHandle(va_dpy, surf->surface_image.buf, &buf_info); + CHECK_VASTATUS(va_status, "vaAcquireBufferHandle"); + + // Create Y image. + surf->y_egl_image = EGL_NO_IMAGE_KHR; + EGLint y_attribs[] = { + EGL_WIDTH, frame_width, + EGL_HEIGHT, frame_height, + EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('R', '8', ' ', ' '), + EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle), + EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[0]), + EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[0]), + EGL_NONE + }; + + surf->y_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, y_attribs); + assert(surf->y_egl_image != EGL_NO_IMAGE_KHR); + + // Associate Y image to a texture. + glBindTexture(GL_TEXTURE_2D, *y_tex); + glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->y_egl_image); + + // Create CbCr image. + surf->cbcr_egl_image = EGL_NO_IMAGE_KHR; + EGLint cbcr_attribs[] = { + EGL_WIDTH, frame_width, + EGL_HEIGHT, frame_height, + EGL_LINUX_DRM_FOURCC_EXT, fourcc_code('G', 'R', '8', '8'), + EGL_DMA_BUF_PLANE0_FD_EXT, EGLint(buf_info.handle), + EGL_DMA_BUF_PLANE0_OFFSET_EXT, EGLint(surf->surface_image.offsets[1]), + EGL_DMA_BUF_PLANE0_PITCH_EXT, EGLint(surf->surface_image.pitches[1]), + EGL_NONE + }; + + surf->cbcr_egl_image = eglCreateImageKHR(eglGetCurrentDisplay(), EGL_NO_CONTEXT, EGL_LINUX_DMA_BUF_EXT, NULL, cbcr_attribs); + assert(surf->cbcr_egl_image != EGL_NO_IMAGE_KHR); + + // Associate CbCr image to a texture. + glBindTexture(GL_TEXTURE_2D, *cbcr_tex); + glEGLImageTargetTexture2DOES(GL_TEXTURE_2D, surf->cbcr_egl_image); + } return true; } void H264EncoderImpl::add_audio(int64_t pts, vector audio) { + assert(!is_shutdown); { unique_lock lock(frame_queue_mutex); pending_audio_frames[pts] = move(audio); @@ -1791,6 +1797,38 @@ void H264EncoderImpl::add_audio(int64_t pts, vector audio) void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector &input_frames) { + assert(!is_shutdown); + + if (!use_zerocopy) { + GLSurface *surf = &gl_surfaces[current_storage_frame % SURFACE_NUM]; + + glPixelStorei(GL_PACK_ROW_LENGTH, 0); + check_error(); + + glBindBuffer(GL_PIXEL_PACK_BUFFER, surf->pbo); + check_error(); + + glBindTexture(GL_TEXTURE_2D, surf->y_tex); + check_error(); + glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->y_offset)); + check_error(); + + glBindTexture(GL_TEXTURE_2D, surf->cbcr_tex); + check_error(); + glGetTexImage(GL_TEXTURE_2D, 0, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(surf->cbcr_offset)); + check_error(); + + glBindTexture(GL_TEXTURE_2D, 0); + check_error(); + glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); + check_error(); + + glMemoryBarrier(GL_TEXTURE_UPDATE_BARRIER_BIT | GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT); + check_error(); + fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); + check_error(); + } + { unique_lock lock(frame_queue_mutex); pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts }; @@ -1799,6 +1837,31 @@ void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vecto frame_queue_nonempty.notify_all(); } +void H264EncoderImpl::shutdown() +{ + if (is_shutdown) { + return; + } + + { + unique_lock lock(frame_queue_mutex); + encode_thread_should_quit = true; + frame_queue_nonempty.notify_all(); + } + encode_thread.join(); + { + unique_lock lock(storage_task_queue_mutex); + storage_thread_should_quit = true; + frame_queue_nonempty.notify_all(); + storage_task_queue_changed.notify_all(); + } + storage_thread.join(); + + release_encode(); + deinit_va(); + is_shutdown = true; +} + void H264EncoderImpl::encode_thread_func() { int64_t last_dts = -1; @@ -1865,28 +1928,65 @@ void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int g } } +namespace { + +void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height) +{ + if (src_width == dst_pitch) { + memcpy(dst, src, src_width * height); + } else { + for (size_t y = 0; y < height; ++y) { + const uint8_t *sptr = src + y * src_width; + uint8_t *dptr = dst + y * dst_pitch; + memcpy(dptr, sptr, src_width); + } + } +} + +} // namespace + void H264EncoderImpl::encode_frame(H264EncoderImpl::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) { // 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(); - // Unmap the image. GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM]; - eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image); - eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image); - VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf); - CHECK_VASTATUS(va_status, "vaReleaseBufferHandle"); + VAStatus va_status; + + if (use_zerocopy) { + eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image); + eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image); + va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf); + CHECK_VASTATUS(va_status, "vaReleaseBufferHandle"); + } else { + unsigned char *surface_p = nullptr; + vaMapBuffer(va_dpy, surf->surface_image.buf, (void **)&surface_p); + + 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 *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"); + } + va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id); CHECK_VASTATUS(va_status, "vaDestroyImage"); - VASurfaceID surface = surf->src_surface; - // Schedule the frame for encoding. - va_status = vaBeginPicture(va_dpy, context_id, surface); + VASurfaceID va_surface = surf->src_surface; + va_status = vaBeginPicture(va_dpy, context_id, va_surface); CHECK_VASTATUS(va_status, "vaBeginPicture"); if (frame_type == FRAME_IDR) { @@ -1918,13 +2018,13 @@ void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int enco } // Proxy object. -H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd) - : impl(new H264EncoderImpl(surface, width, height, httpd)) {} +H264Encoder::H264Encoder(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd) + : impl(new H264EncoderImpl(surface, va_display, width, height, httpd)) {} // Must be defined here because unique_ptr<> destructor needs to know the impl. H264Encoder::~H264Encoder() {} -void H264Encoder::add_audio(int64_t pts, std::vector audio) +void H264Encoder::add_audio(int64_t pts, vector audio) { impl->add_audio(pts, audio); } @@ -1934,9 +2034,14 @@ bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) return impl->begin_frame(y_tex, cbcr_tex); } -void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector &input_frames) +void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector &input_frames) { impl->end_frame(fence, pts, input_frames); } +void H264Encoder::shutdown() +{ + impl->shutdown(); +} + // Real class.