X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=video_stream.cpp;h=69dfbe6f0fea07ff22a09f60fabf09122b0e93e4;hb=4f7fcd79708e2743a5ea16ba83da63c7e962e88c;hp=820a9335846c9115ed2e2c5ac2bba96819bd330d;hpb=00a917a3feec0fdd99db1a58a7d1edf973a0eb57;p=nageru diff --git a/video_stream.cpp b/video_stream.cpp index 820a933..69dfbe6 100644 --- a/video_stream.cpp +++ b/video_stream.cpp @@ -8,6 +8,7 @@ extern "C" { #include #include +#include "chroma_subsampler.h" #include "context.h" #include "flow.h" #include "httpd.h" @@ -16,6 +17,7 @@ extern "C" { #include "mux.h" #include "player.h" #include "util.h" +#include "ycbcr_converter.h" #include @@ -98,7 +100,7 @@ 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; @@ -116,28 +118,27 @@ vector encode_jpeg(const uint8_t *pixel_data, unsigned width, unsigned cinfo.image_width = width; cinfo.image_height = height; - cinfo.input_components = 3; + cinfo.raw_data_in = true; + jpeg_set_colorspace(&cinfo, JCS_YCbCr); cinfo.comp_info[0].h_samp_factor = 2; cinfo.comp_info[0].v_samp_factor = 1; cinfo.comp_info[1].h_samp_factor = 1; cinfo.comp_info[1].v_samp_factor = 1; cinfo.comp_info[2].h_samp_factor = 1; cinfo.comp_info[2].v_samp_factor = 1; - // cinfo.CCIR601_sampling = true; // TODO: Subsample ourselves. + cinfo.CCIR601_sampling = true; // Seems to be mostly ignored by libjpeg, though. jpeg_start_compress(&cinfo, true); - unique_ptr row(new uint8_t[width * 3]); - JSAMPROW row_pointer[1] = { row.get() }; - for (unsigned y = 0; y < height; ++y) { - const uint8_t *sptr = &pixel_data[(height - cinfo.next_scanline - 1) * width * 4]; - uint8_t *dptr = row.get(); - for (unsigned x = 0; x < width; ++x) { - *dptr++ = *sptr++; - *dptr++ = *sptr++; - *dptr++ = *sptr++; - ++sptr; + JSAMPROW yptr[8], cbptr[8], crptr[8]; + JSAMPARRAY data[3] = { yptr, cbptr, crptr }; + for (unsigned y = 0; y < height; y += 8) { + for (unsigned yy = 0; yy < 8; ++yy) { + 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]); } - (void) jpeg_write_scanlines(&cinfo, row_pointer, 1); + + jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8); } jpeg_finish_compress(&cinfo); @@ -148,39 +149,21 @@ 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)); - - ImageFormat inout_format; - inout_format.color_space = COLORSPACE_sRGB; - inout_format.gamma_curve = GAMMA_sRGB; - - check_error(); - ycbcr_convert_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED); - 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]; + 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, 10, input_tex); glCreateTextures(GL_TEXTURE_2D_ARRAY, 10, gray_tex); + glCreateTextures(GL_TEXTURE_2D, 10, fade_y_output_tex); + glCreateTextures(GL_TEXTURE_2D, 10, fade_cbcr_output_tex); + glCreateTextures(GL_TEXTURE_2D, 10, cb_tex); + glCreateTextures(GL_TEXTURE_2D, 10, 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) { @@ -188,22 +171,46 @@ 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]; + 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); check_error(); + glNamedFramebufferTextureLayer(resource.input_fbos[0], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 0); + check_error(); glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT0, input_tex[i], 0, 1); check_error(); + glNamedFramebufferTextureLayer(resource.input_fbos[1], GL_COLOR_ATTACHMENT1, gray_tex[i], 0, 1); + check_error(); + 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 buf = GL_COLOR_ATTACHMENT0; - glNamedFramebufferDrawBuffers(resource.input_fbos[0], 1, &buf); + GLuint bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 }; + glNamedFramebufferDrawBuffers(resource.input_fbos[0], 2, bufs); check_error(); - glNamedFramebufferDrawBuffers(resource.input_fbos[1], 1, &buf); + glNamedFramebufferDrawBuffers(resource.input_fbos[1], 2, bufs); + check_error(); + glNamedFramebufferDrawBuffers(resource.fade_fbo, 2, bufs); check_error(); glCreateBuffers(1, &resource.pbo); @@ -216,9 +223,10 @@ VideoStream::VideoStream() check_error(); - compute_flow.reset(new DISComputeFlow(width, height, operating_point3)); - gray.reset(new GrayscaleConversion); // NOTE: Must come after DISComputeFlow, since it sets up the VBO! - interpolate.reset(new Interpolate(width, height, operating_point3)); + compute_flow.reset(new DISComputeFlow(width, height, operating_point2)); + interpolate.reset(new Interpolate(operating_point2, /*split_ycbcr_output=*/true)); + interpolate_no_split.reset(new Interpolate(operating_point2, /*split_ycbcr_output=*/false)); + chroma_subsampler.reset(new ChromaSubsampler); check_error(); } @@ -268,9 +276,84 @@ void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_id queue_nonempty.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(int64_t output_pts, unsigned stream_idx, int64_t input_pts, int secondary_stream_idx, int64_t secondary_input_pts, 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, input_pts, secondary_input_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.) + InterpolatedFrameResources 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(); + interpolate_resources.pop_front(); + } + + bool did_decode; + + JPEGID jpeg_id1; + jpeg_id1.stream_idx = stream_idx; + jpeg_id1.pts = input_pts; + jpeg_id1.interpolated = false; + shared_ptr frame1 = decode_jpeg_with_cache(jpeg_id1, DECODE_IF_NOT_IN_CACHE, &did_decode); + + JPEGID jpeg_id2; + jpeg_id2.stream_idx = secondary_stream_idx; + jpeg_id2.pts = secondary_input_pts; + jpeg_id2.interpolated = false; + shared_ptr frame2 = decode_jpeg_with_cache(jpeg_id2, 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.type = QueuedFrame::FADED; + qf.output_pts = output_pts; + qf.stream_idx = stream_idx; + qf.resources = resources; + qf.input_first_pts = input_pts; + + qf.secondary_stream_idx = secondary_stream_idx; + qf.secondary_input_pts = secondary_input_pts; + + // 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(); + + unique_lock lock(queue_lock); + frame_queue.push_back(qf); + queue_nonempty.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, int secondary_stream_idx, int64_t secondary_input_pts, float fade_alpha) +{ + if (secondary_stream_idx != -1) { + fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f secondary_pts=%ld fade_alpha=%.2f\n", output_pts, input_first_pts, input_second_pts, alpha, secondary_input_pts, fade_alpha); + } else { + fprintf(stderr, "output_pts=%ld interpolated input_pts1=%ld input_pts2=%ld alpha=%.3f\n", output_pts, input_first_pts, input_second_pts, alpha); + } // Get the temporary OpenGL resources we need for doing the interpolation. InterpolatedFrameResources resources; @@ -278,6 +361,7 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea unique_lock lock(queue_lock); if (interpolate_resources.empty()) { fprintf(stderr, "WARNING: Too many interpolated frames already in transit; dropping one.\n"); + JPEGFrameView::insert_interpolated_frame(stream_idx, output_pts, nullptr); return; } resources = interpolate_resources.front(); @@ -285,7 +369,7 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea } QueuedFrame qf; - qf.type = QueuedFrame::INTERPOLATED; + qf.type = (secondary_stream_idx == -1) ? QueuedFrame::INTERPOLATED : QueuedFrame::FADED_INTERPOLATED; qf.output_pts = output_pts; qf.stream_idx = stream_idx; qf.resources = resources; @@ -293,36 +377,51 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea 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) { - shared_ptr frame = decode_jpeg(filename_for_frame(stream_idx, frame_no == 1 ? input_second_pts : input_first_pts)); - 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); + JPEGID jpeg_id; + jpeg_id.stream_idx = stream_idx; + jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts; + jpeg_id.interpolated = false; + bool did_decode; + shared_ptr frame = decode_jpeg_with_cache(jpeg_id, 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); - - // Compute the interpolated frame. - check_error(); - gray->exec(resources.input_tex, resources.gray_tex, 1280, 720, /*num_layers=*/2); check_error(); 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, qf.flow_tex, 1280, 720, alpha); - check_error(); + if (secondary_stream_idx != -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. + JPEGID jpeg_id; + jpeg_id.stream_idx = secondary_stream_idx; + jpeg_id.pts = secondary_input_pts; + jpeg_id.interpolated = false; + bool did_decode; + shared_ptr frame2 = decode_jpeg_with_cache(jpeg_id, 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); + } 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 @@ -332,7 +431,15 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea glPixelStorei(GL_PACK_ROW_LENGTH, 0); 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_stream_idx != -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); @@ -347,8 +454,40 @@ void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned strea queue_nonempty.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"); QSurface *surface = create_surface(); QOpenGLContext *context = create_context(surface); bool ok = make_current(context, surface); @@ -377,12 +516,38 @@ 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); - } else if (qf.type == QueuedFrame::INTERPOLATED) { + } else if (qf.type == QueuedFrame::FADED) { + glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); + + 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); + 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); + + // Put the frame resources back. + unique_lock lock(queue_lock); + interpolate_resources.push_back(qf.resources); + } else if (qf.type == QueuedFrame::INTERPOLATED || qf.type == QueuedFrame::FADED_INTERPOLATED) { glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED); - vector jpeg = encode_jpeg((const uint8_t *)qf.resources.pbo_contents, 1280, 720); + // Send a copy of the frame on to display. + shared_ptr frame = frame_from_pbo(qf.resources.pbo_contents, 1280, 720); + JPEGFrameView::insert_interpolated_frame(qf.stream_idx, qf.output_pts, frame); // TODO: this is wrong for fades + + // 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); - interpolate->release_texture(qf.output_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); @@ -394,6 +559,8 @@ void VideoStream::encode_thread_func() // Put the frame resources back. unique_lock lock(queue_lock); interpolate_resources.push_back(qf.resources); + } else { + assert(false); } } }