X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=video_stream.cpp;h=a9ed0610e4518729e9aaa9614bb98ab0083d33ce;hb=bdef311c334b674ba39a931805fb7d32ce8698da;hp=8d485ad02b20e79ab3c2a2ef578c7395835d2e9b;hpb=8deea0eee4f9b18540463ba253747c232873527b;p=nageru diff --git a/video_stream.cpp b/video_stream.cpp index 8d485ad..a9ed061 100644 --- a/video_stream.cpp +++ b/video_stream.cpp @@ -5,10 +5,9 @@ extern "C" { #include } -#include -#include - +#include "chroma_subsampler.h" #include "context.h" +#include "flags.h" #include "flow.h" #include "httpd.h" #include "jpeg_frame_view.h" @@ -16,36 +15,17 @@ extern "C" { #include "mux.h" #include "player.h" #include "util.h" +#include "ycbcr_converter.h" #include +#include +#include using namespace std; +using namespace std::chrono; extern HTTPD *global_httpd; -namespace { - -string read_file(const string &filename) -{ - FILE *fp = fopen(filename.c_str(), "rb"); - if (fp == nullptr) { - perror(filename.c_str()); - return ""; - } - - fseek(fp, 0, SEEK_END); - long len = ftell(fp); - rewind(fp); - - string ret; - ret.resize(len); - fread(&ret[0], len, 1, fp); - fclose(fp); - return ret; -} - -} // namespace - struct VectorDestinationManager { jpeg_destination_mgr pub; std::vector dest; @@ -98,7 +78,7 @@ struct VectorDestinationManager { }; static_assert(std::is_standard_layout::value, ""); -vector encode_jpeg(const uint8_t *pixel_data, unsigned width, unsigned height) +vector encode_jpeg(const uint8_t *y_data, const uint8_t *cb_data, const uint8_t *cr_data, unsigned width, unsigned height) { VectorDestinationManager dest; @@ -127,31 +107,13 @@ vector encode_jpeg(const uint8_t *pixel_data, unsigned width, unsigned cinfo.CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though. jpeg_start_compress(&cinfo, true); - // TODO: Subsample and deinterleave on the GPU. - - unique_ptr ydata(new uint8_t[width * 8]); - unique_ptr cbdata(new uint8_t[(width/2) * 8]); - unique_ptr crdata(new uint8_t[(width/2) * 8]); JSAMPROW yptr[8], cbptr[8], crptr[8]; JSAMPARRAY data[3] = { yptr, cbptr, crptr }; - for (unsigned yy = 0; yy < 8; ++yy) { - yptr[yy] = ydata.get() + yy * width; - cbptr[yy] = cbdata.get() + yy * (width / 2); - crptr[yy] = crdata.get() + yy * (width / 2); - } for (unsigned y = 0; y < height; y += 8) { - uint8_t *yptr = ydata.get(); - uint8_t *cbptr = cbdata.get(); - uint8_t *crptr = crdata.get(); for (unsigned yy = 0; yy < 8; ++yy) { - const uint8_t *sptr = &pixel_data[(height - y - yy - 1) * width * 4]; - for (unsigned x = 0; x < width; x += 2) { - *yptr++ = sptr[0]; - *yptr++ = sptr[4]; - *cbptr++ = (sptr[1] + sptr[5]) / 2; - *crptr++ = (sptr[2] + sptr[6]) / 2; - sptr += 8; - } + yptr[yy] = const_cast(&y_data[(y + yy) * width]); + cbptr[yy] = const_cast(&cb_data[(y + yy) * width / 2]); + crptr[yy] = const_cast(&cr_data[(y + yy) * width / 2]); } jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8); @@ -165,48 +127,21 @@ vector encode_jpeg(const uint8_t *pixel_data, unsigned width, unsigned VideoStream::VideoStream() { - using namespace movit; - // TODO: deduplicate code against JPEGFrameView? - ycbcr_convert_chain.reset(new EffectChain(1280, 720)); - ImageFormat image_format; - image_format.color_space = COLORSPACE_sRGB; - image_format.gamma_curve = GAMMA_sRGB; - ycbcr_format.luma_coefficients = YCBCR_REC_709; - ycbcr_format.full_range = true; // JPEG. - ycbcr_format.num_levels = 256; - ycbcr_format.chroma_subsampling_x = 2; - ycbcr_format.chroma_subsampling_y = 1; - ycbcr_format.cb_x_position = 0.0f; // H.264 -- _not_ JPEG, even though our input is MJPEG-encoded - ycbcr_format.cb_y_position = 0.5f; // Irrelevant. - ycbcr_format.cr_x_position = 0.0f; - ycbcr_format.cr_y_position = 0.5f; - ycbcr_input = (movit::YCbCrInput *)ycbcr_convert_chain->add_input(new YCbCrInput(image_format, ycbcr_format, 1280, 720)); - - YCbCrFormat ycbcr_output_format = ycbcr_format; - ycbcr_output_format.chroma_subsampling_x = 1; - - ImageFormat inout_format; - inout_format.color_space = COLORSPACE_sRGB; - inout_format.gamma_curve = GAMMA_sRGB; - - check_error(); - - // One full Y'CbCr texture (for interpolation), one that's just Y (throwing away the - // Cb and Cr channels). The second copy is sort of redundant, but it's the easiest way - // of getting the gray data into a layered texture. - ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format); - check_error(); - ycbcr_convert_chain->add_ycbcr_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED, ycbcr_output_format); - check_error(); - ycbcr_convert_chain->set_dither_bits(8); - check_error(); - ycbcr_convert_chain->finalize(); - check_error(); + ycbcr_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_DUAL_YCBCR, /*resource_pool=*/nullptr)); + ycbcr_semiplanar_converter.reset(new YCbCrConverter(YCbCrConverter::OUTPUT_TO_SEMIPLANAR, /*resource_pool=*/nullptr)); GLuint input_tex[num_interpolate_slots], gray_tex[num_interpolate_slots]; - glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, input_tex); - glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex); + GLuint fade_y_output_tex[num_interpolate_slots], fade_cbcr_output_tex[num_interpolate_slots]; + GLuint cb_tex[num_interpolate_slots], cr_tex[num_interpolate_slots]; + + glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, input_tex); + glCreateTextures(GL_TEXTURE_2D_ARRAY, num_interpolate_slots, gray_tex); + glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_y_output_tex); + glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, fade_cbcr_output_tex); + glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cb_tex); + glCreateTextures(GL_TEXTURE_2D, num_interpolate_slots, cr_tex); check_error(); + constexpr size_t width = 1280, height = 720; // FIXME: adjustable width, height int levels = find_num_levels(width, height); for (size_t i = 0; i < num_interpolate_slots; ++i) { @@ -214,41 +149,84 @@ VideoStream::VideoStream() check_error(); glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2); check_error(); + glTextureStorage2D(fade_y_output_tex[i], 1, GL_R8, width, height); + check_error(); + glTextureStorage2D(fade_cbcr_output_tex[i], 1, GL_RG8, width, height); + check_error(); + glTextureStorage2D(cb_tex[i], 1, GL_R8, width / 2, height); + check_error(); + glTextureStorage2D(cr_tex[i], 1, GL_R8, width / 2, height); + check_error(); - InterpolatedFrameResources resource; - resource.input_tex = input_tex[i]; - resource.gray_tex = gray_tex[i]; - glCreateFramebuffers(2, resource.input_fbos); + unique_ptr resource(new InterpolatedFrameResources); + resource->owner = this; + resource->input_tex = input_tex[i]; + resource->gray_tex = gray_tex[i]; + resource->fade_y_output_tex = fade_y_output_tex[i]; + resource->fade_cbcr_output_tex = fade_cbcr_output_tex[i]; + resource->cb_tex = cb_tex[i]; + resource->cr_tex = cr_tex[i]; + glCreateFramebuffers(2, resource->input_fbos); + check_error(); + glCreateFramebuffers(1, &resource->fade_fbo); check_error(); - glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0); + glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 0); + check_error(); + glNamedFramebufferTextureLayer(resource->input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0); check_error(); - glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0); + glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1); check_error(); - glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1); + glNamedFramebufferTextureLayer(resource->input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1); check_error(); - glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1); + glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT0, fade_y_output_tex[i], 0); + check_error(); + glNamedFramebufferTexture(resource->fade_fbo, GL_COLOR_ATTACHMENT1, fade_cbcr_output_tex[i], 0); check_error(); GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 }; - glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs); + glNamedFramebufferDrawBuffers(resource->input_fbos[0], 2, bufs); + check_error(); + glNamedFramebufferDrawBuffers(resource->input_fbos[1], 2, bufs); check_error(); - glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs); + glNamedFramebufferDrawBuffers(resource->fade_fbo, 2, bufs); check_error(); - glCreateBuffers(1, &resource.pbo); + glCreateBuffers(1, &resource->pbo); check_error(); - glNamedBufferStorage(resource.pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); + glNamedBufferStorage(resource->pbo, width * height * 4, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); check_error(); - resource.pbo_contents = glMapNamedBufferRange(resource.pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); - interpolate_resources.push_back(resource); + resource->pbo_contents = glMapNamedBufferRange(resource->pbo, 0, width * height * 4, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT); + interpolate_resources.push_back(move(resource)); } check_error(); - compute_flow.reset(new DISComputeFlow(width, height, operating_point3)); - interpolate.reset(new Interpolate(width, height, operating_point3)); + OperatingPoint op; + if (global_flags.interpolation_quality == 1) { + op = operating_point1; + } else if (global_flags.interpolation_quality == 2) { + op = operating_point2; + } else if (global_flags.interpolation_quality == 3) { + op = operating_point3; + } else if (global_flags.interpolation_quality == 4) { + op = operating_point4; + } else { + assert(false); + } + + compute_flow.reset(new DISComputeFlow(width, height, op)); + interpolate.reset(new Interpolate(op, /*split_ycbcr_output=*/true)); + interpolate_no_split.reset(new Interpolate(op, /*split_ycbcr_output=*/false)); + chroma_subsampler.reset(new ChromaSubsampler); check_error(); + + // The “last frame” is initially black. + unique_ptr y(new uint8_t[1280 * 720]); + unique_ptr cb_or_cr(new uint8_t[640 * 720]); + memset(y.get(), 16, 1280 * 720); + memset(cb_or_cr.get(), 128, 640 * 720); + last_frame = encode_jpeg(y.get(), cb_or_cr.get(), cb_or_cr.get(), 1280, 720); } VideoStream::~VideoStream() {} @@ -282,87 +260,214 @@ void VideoStream::stop() encode_thread.join(); } -void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_idx, int64_t input_pts) +void VideoStream::clear_queue() { - fprintf(stderr, "output_pts=%ld original input_pts=%ld\n", output_pts, input_pts); + deque q; + + { + unique_lock lock(queue_lock); + q = move(frame_queue); + } + + // These are not RAII-ed, unfortunately, so we'll need to clean them ourselves. + // Note that release_texture() is thread-safe. + for (const QueuedFrame &qf : q) { + if (qf.type == QueuedFrame::INTERPOLATED || + qf.type == QueuedFrame::FADED_INTERPOLATED) { + compute_flow->release_texture(qf.flow_tex); + } + if (qf.type == QueuedFrame::INTERPOLATED) { + interpolate->release_texture(qf.output_tex); + interpolate->release_texture(qf.cbcr_tex); + } + } + + // Destroy q outside the mutex, as that would be a double-lock. +} + +void VideoStream::schedule_original_frame(steady_clock::time_point local_pts, + int64_t output_pts, function &&display_func, + QueueSpotHolder &&queue_spot_holder, + FrameOnDisk frame) +{ + fprintf(stderr, "output_pts=%ld original input_pts=%ld\n", output_pts, frame.pts); + + // Preload the file from disk, so that the encoder thread does not get stalled. + // TODO: Consider sending it through the queue instead. + (void)read_frame(frame); QueuedFrame qf; + qf.local_pts = local_pts; qf.type = QueuedFrame::ORIGINAL; qf.output_pts = output_pts; - qf.stream_idx = stream_idx; - qf.input_first_pts = input_pts; + qf.frame1 = frame; + qf.display_func = move(display_func); + qf.queue_spot_holder = move(queue_spot_holder); unique_lock lock(queue_lock); - frame_queue.push_back(qf); - queue_nonempty.notify_all(); + frame_queue.push_back(move(qf)); + queue_changed.notify_all(); } -void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float alpha) +void VideoStream::schedule_faded_frame(steady_clock::time_point local_pts, int64_t output_pts, + function &&display_func, + QueueSpotHolder &&queue_spot_holder, + FrameOnDisk frame1_spec, FrameOnDisk frame2_spec, + float fade_alpha) { - fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha); + fprintf(stderr, "output_pts=%ld faded input_pts=%ld,%ld fade_alpha=%.2f\n", output_pts, frame1_spec.pts, frame2_spec.pts, fade_alpha); + + // Get the temporary OpenGL resources we need for doing the fade. + // (We share these with interpolated frames, which is slightly + // overkill, but there's no need to waste resources on keeping + // separate pools around.) + BorrowedInterpolatedFrameResources resources; + { + unique_lock lock(queue_lock); + if (interpolate_resources.empty()) { + fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n"); + return; + } + resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release()); + interpolate_resources.pop_front(); + } + + bool did_decode; + + shared_ptr frame1 = decode_jpeg_with_cache(frame1_spec, DECODE_IF_NOT_IN_CACHE, &did_decode); + shared_ptr frame2 = decode_jpeg_with_cache(frame2_spec, DECODE_IF_NOT_IN_CACHE, &did_decode); + + ycbcr_semiplanar_converter->prepare_chain_for_fade(frame1, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720); + + QueuedFrame qf; + qf.local_pts = local_pts; + qf.type = QueuedFrame::FADED; + qf.output_pts = output_pts; + qf.frame1 = frame1_spec; + qf.display_func = move(display_func); + qf.queue_spot_holder = move(queue_spot_holder); + + qf.secondary_frame = frame2_spec; + + // Subsample and split Cb/Cr. + chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex); + + // Read it down (asynchronously) to the CPU. + glPixelStorei(GL_PACK_ROW_LENGTH, 0); + glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo); + check_error(); + glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0)); + check_error(); + glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720)); + check_error(); + glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720)); + check_error(); + glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); + + // Set a fence we can wait for to make sure the CPU sees the read. + glMemoryBarrier(GL_CLIENT_MAPPED_BUFFER_BARRIER_BIT); + check_error(); + qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); + check_error(); + qf.resources = move(resources); + qf.local_pts = local_pts; + + unique_lock lock(queue_lock); + frame_queue.push_back(move(qf)); + queue_changed.notify_all(); +} + +void VideoStream::schedule_interpolated_frame(steady_clock::time_point local_pts, + int64_t output_pts, function)> &&display_func, + QueueSpotHolder &&queue_spot_holder, + FrameOnDisk frame1, FrameOnDisk frame2, + float alpha, FrameOnDisk secondary_frame, float fade_alpha) +{ + if (secondary_frame.pts != -1) { + fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f secondary_pts=%ld fade_alpha=%.2f\n", output_pts, frame1.pts, frame2.pts, alpha, secondary_frame.pts, fade_alpha); + } else { + fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, frame1.pts, frame2.pts, alpha); + } // Get the temporary OpenGL resources we need for doing the interpolation. - InterpolatedFrameResources resources; + BorrowedInterpolatedFrameResources resources; { unique_lock lock(queue_lock); if (interpolate_resources.empty()) { fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n"); return; } - resources = interpolate_resources.front(); + resources = BorrowedInterpolatedFrameResources(interpolate_resources.front().release()); interpolate_resources.pop_front(); } QueuedFrame qf; - qf.type = QueuedFrame::INTERPOLATED; + qf.type = (secondary_frame.pts == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED; qf.output_pts = output_pts; - qf.stream_idx = stream_idx; - qf.resources = resources; + qf.display_decoded_func = move(display_func); + qf.queue_spot_holder = move(queue_spot_holder); + qf.local_pts = local_pts; check_error(); // Convert frame0 and frame1 to OpenGL textures. - // TODO: Deduplicate against JPEGFrameView::setDecodedFrame? for (size_t frame_no = 0; frame_no < 2; ++frame_no) { - JPEGID jpeg_id; - jpeg_id.stream_idx = stream_idx; - jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts; + FrameOnDisk frame_spec = frame_no == 1 ? frame2 : frame1; bool did_decode; - shared_ptr frame = decode_jpeg_with_cache(jpeg_id, DECODE_IF_NOT_IN_CACHE, &did_decode); - ycbcr_format.chroma_subsampling_x = frame->chroma_subsampling_x; - ycbcr_format.chroma_subsampling_y = frame->chroma_subsampling_y; - ycbcr_input->change_ycbcr_format(ycbcr_format); - ycbcr_input->set_width(frame->width); - ycbcr_input->set_height(frame->height); - ycbcr_input->set_pixel_data(0, frame->y.get()); - ycbcr_input->set_pixel_data(1, frame->cb.get()); - ycbcr_input->set_pixel_data(2, frame->cr.get()); - ycbcr_input->set_pitch(0, frame->pitch_y); - ycbcr_input->set_pitch(1, frame->pitch_chroma); - ycbcr_input->set_pitch(2, frame->pitch_chroma); - ycbcr_convert_chain->render_to_fbo(resources.input_fbos[frame_no], 1280, 720); + shared_ptr frame = decode_jpeg_with_cache(frame_spec, DECODE_IF_NOT_IN_CACHE, &did_decode); + ycbcr_converter->prepare_chain_for_conversion(frame)->render_to_fbo(resources->input_fbos[frame_no], 1280, 720); } - glGenerateTextureMipmap(resources.input_tex); + glGenerateTextureMipmap(resources->input_tex); check_error(); - glGenerateTextureMipmap(resources.gray_tex); + glGenerateTextureMipmap(resources->gray_tex); check_error(); // Compute the interpolated frame. - qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW); - check_error(); - qf.output_tex = interpolate->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha); + qf.flow_tex = compute_flow->exec(resources->gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW); check_error(); + if (secondary_frame.pts != -1) { + // Fade. First kick off the interpolation. + tie(qf.output_tex, ignore) = interpolate_no_split->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha); + check_error(); + + // Now decode the image we are fading against. + bool did_decode; + shared_ptr frame2 = decode_jpeg_with_cache(secondary_frame, DECODE_IF_NOT_IN_CACHE, &did_decode); + + // Then fade against it, putting it into the fade Y' and CbCr textures. + ycbcr_semiplanar_converter->prepare_chain_for_fade_from_texture(qf.output_tex, frame2, fade_alpha)->render_to_fbo(resources->fade_fbo, 1280, 720); + + // Subsample and split Cb/Cr. + chroma_subsampler->subsample_chroma(resources->fade_cbcr_output_tex, 1280, 720, resources->cb_tex, resources->cr_tex); + + interpolate_no_split->release_texture(qf.output_tex); + } else { + tie(qf.output_tex, qf.cbcr_tex) = interpolate->exec(resources->input_tex, resources->gray_tex, qf.flow_tex, 1280, 720, alpha); + check_error(); + + // Subsample and split Cb/Cr. + chroma_subsampler->subsample_chroma(qf.cbcr_tex, 1280, 720, resources->cb_tex, resources->cr_tex); + } + // We could have released qf.flow_tex here, but to make sure we don't cause a stall // when trying to reuse it for the next frame, we can just as well hold on to it // and release it only when the readback is done. // Read it down (asynchronously) to the CPU. glPixelStorei(GL_PACK_ROW_LENGTH, 0); - glBindBuffer(GL_PIXEL_PACK_BUFFER, resources.pbo); + glBindBuffer(GL_PIXEL_PACK_BUFFER, resources->pbo); check_error(); - glGetTextureImage(qf.output_tex, 0, GL_RGBA, GL_UNSIGNED_BYTE, 1280 * 720 * 4, nullptr); + if (secondary_frame.pts != -1) { + glGetTextureImage(resources->fade_y_output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0)); + } else { + glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0)); + } + check_error(); + glGetTextureImage(resources->cb_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720)); + check_error(); + glGetTextureImage(resources->cr_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 3 - 640 * 720, BUFFER_OFFSET(1280 * 720 + 640 * 720)); check_error(); glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); @@ -371,12 +476,59 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea check_error(); qf.fence = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0); check_error(); + qf.resources = move(resources); + + unique_lock lock(queue_lock); + frame_queue.push_back(move(qf)); + queue_changed.notify_all(); +} + +void VideoStream::schedule_refresh_frame(steady_clock::time_point local_pts, + int64_t output_pts, function &&display_func, + QueueSpotHolder &&queue_spot_holder) +{ + QueuedFrame qf; + qf.type = QueuedFrame::REFRESH; + qf.output_pts = output_pts; + qf.display_func = move(display_func); + qf.queue_spot_holder = move(queue_spot_holder); unique_lock lock(queue_lock); - frame_queue.push_back(qf); - queue_nonempty.notify_all(); + frame_queue.push_back(move(qf)); + queue_changed.notify_all(); +} + +namespace { + +shared_ptr frame_from_pbo(void *contents, size_t width, size_t height) +{ + size_t chroma_width = width / 2; + + const uint8_t *y = (const uint8_t *)contents; + const uint8_t *cb = (const uint8_t *)contents + width * height; + const uint8_t *cr = (const uint8_t *)contents + width * height + chroma_width * height; + + shared_ptr frame(new Frame); + frame->y.reset(new uint8_t[width * height]); + frame->cb.reset(new uint8_t[chroma_width * height]); + frame->cr.reset(new uint8_t[chroma_width * height]); + for (unsigned yy = 0; yy < height; ++yy) { + memcpy(frame->y.get() + width * yy, y + width * yy, width); + memcpy(frame->cb.get() + chroma_width * yy, cb + chroma_width * yy, chroma_width); + memcpy(frame->cr.get() + chroma_width * yy, cr + chroma_width * yy, chroma_width); + } + frame->is_semiplanar = false; + frame->width = width; + frame->height = height; + frame->chroma_subsampling_x = 2; + frame->chroma_subsampling_y = 1; + frame->pitch_y = width; + frame->pitch_chroma = chroma_width; + return frame; } +} // namespace + void VideoStream::encode_thread_func() { pthread_setname_np(pthread_self(), "VideoStream"); @@ -392,28 +544,44 @@ void VideoStream::encode_thread_func() QueuedFrame qf; { unique_lock lock(queue_lock); - queue_nonempty.wait(lock, [this]{ + + // Wait until we have a frame to play. + queue_changed.wait(lock, [this]{ return !frame_queue.empty(); }); - qf = frame_queue.front(); + steady_clock::time_point frame_start = frame_queue.front().local_pts; + + // Now sleep until the frame is supposed to start (the usual case), + // _or_ clear_queue() happened. + bool aborted = queue_changed.wait_until(lock, frame_start, [this, frame_start]{ + return frame_queue.empty() || frame_queue.front().local_pts != frame_start; + }); + if (aborted) { + // clear_queue() happened, so don't play this frame after all. + continue; + } + qf = move(frame_queue.front()); frame_queue.pop_front(); } if (qf.type == QueuedFrame::ORIGINAL) { // Send the JPEG frame on, unchanged. - string jpeg = read_file(filename_for_frame(qf.stream_idx, qf.input_first_pts)); + string jpeg = read_frame(qf.frame1); AVPacket pkt; av_init_packet(&pkt); pkt.stream_index = 0; pkt.data = (uint8_t *)jpeg.data(); pkt.size = jpeg.size(); stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts); - } else if (qf.type == QueuedFrame::INTERPOLATED) { + + last_frame.assign(&jpeg[0], &jpeg[0] + jpeg.size()); + } else if (qf.type == QueuedFrame::FADED) { glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); - vector jpeg = encode_jpeg((const uint8_t *)qf.resources.pbo_contents, 1280, 720); - compute_flow->release_texture(qf.flow_tex); - interpolate->release_texture(qf.output_tex); + shared_ptr frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720); + + // Now JPEG encode it, and send it on to the stream. + vector jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720); AVPacket pkt; av_init_packet(&pkt); @@ -421,10 +589,43 @@ void VideoStream::encode_thread_func() pkt.data = (uint8_t *)jpeg.data(); pkt.size = jpeg.size(); stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts); + last_frame = move(jpeg); + } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) { + glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); - // Put the frame resources back. - unique_lock lock(queue_lock); - interpolate_resources.push_back(qf.resources); + // Send it on to display. + shared_ptr frame = frame_from_pbo(qf.resources->pbo_contents, 1280, 720); + if (qf.display_decoded_func != nullptr) { + qf.display_decoded_func(frame); + } + + // Now JPEG encode it, and send it on to the stream. + vector jpeg = encode_jpeg(frame->y.get(), frame->cb.get(), frame->cr.get(), 1280, 720); + compute_flow->release_texture(qf.flow_tex); + if (qf.type != QueuedFrame::FADED_INTERPOLATED) { + interpolate->release_texture(qf.output_tex); + interpolate->release_texture(qf.cbcr_tex); + } + + AVPacket pkt; + av_init_packet(&pkt); + pkt.stream_index = 0; + pkt.data = (uint8_t *)jpeg.data(); + pkt.size = jpeg.size(); + stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts); + last_frame = move(jpeg); + } else if (qf.type == QueuedFrame::REFRESH) { + AVPacket pkt; + av_init_packet(&pkt); + pkt.stream_index = 0; + pkt.data = (uint8_t *)last_frame.data(); + pkt.size = last_frame.size(); + stream_mux->add_packet(pkt, qf.output_pts, qf.output_pts); + } else { + assert(false); + } + if (qf.display_func != nullptr) { + qf.display_func(); } } } @@ -453,4 +654,3 @@ int VideoStream::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType ty } return buf_size; } -