};
static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
-vector<uint8_t> encode_jpeg(const uint8_t *pixel_data, unsigned width, unsigned height)
+vector<uint8_t> encode_jpeg(const uint8_t *y_data, const uint8_t *cbcr_data, unsigned width, unsigned height)
{
VectorDestinationManager dest;
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<uint8_t[]> 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;
+ // TODO: Subsample on the GPU.
+ unique_ptr<uint8_t[]> cbdata(new uint8_t[(width/2) * 8]);
+ unique_ptr<uint8_t[]> 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) {
+ cbptr[yy] = cbdata.get() + yy * (width / 2);
+ crptr[yy] = crdata.get() + yy * (width / 2);
+ }
+ for (unsigned y = 0; y < height; y += 8) {
+ uint8_t *cbptr = cbdata.get();
+ uint8_t *crptr = crdata.get();
+ for (unsigned yy = 0; yy < 8; ++yy) {
+ yptr[yy] = const_cast<JSAMPROW>(&y_data[(height - y - yy - 1) * width]);
+ const uint8_t *sptr = &cbcr_data[(height - y - yy - 1) * width * 2];
+ for (unsigned x = 0; x < width; x += 2) {
+ *cbptr++ = (sptr[0] + sptr[2]) / 2;
+ *crptr++ = (sptr[1] + sptr[3]) / 2;
+ sptr += 4;
+ }
}
- (void) jpeg_write_scanlines(&cinfo, row_pointer, 1);
+
+ jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
}
jpeg_finish_compress(&cinfo);
image_format.color_space = COLORSPACE_sRGB;
image_format.gamma_curve = GAMMA_sRGB;
ycbcr_format.luma_coefficients = YCBCR_REC_709;
- ycbcr_format.full_range = false;
+ 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.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();
- ycbcr_convert_chain->add_output(inout_format, OUTPUT_ALPHA_FORMAT_POSTMULTIPLIED);
+
+ // 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();
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();
- 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();
glCreateBuffers(1, &resource.pbo);
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(width, height, operating_point2, /*split_ycbcr_output=*/true));
check_error();
}
void VideoStream::schedule_original_frame(int64_t output_pts, unsigned stream_idx, int64_t input_pts)
{
+ fprintf(stderr, "output_pts=%ld original input_pts=%ld\n", output_pts, input_pts);
+
QueuedFrame qf;
qf.type = QueuedFrame::ORIGINAL;
qf.output_pts = output_pts;
void VideoStream::schedule_interpolated_frame(int64_t output_pts, unsigned stream_idx, int64_t input_first_pts, int64_t input_second_pts, float 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);
+
// Get the temporary OpenGL resources we need for doing the interpolation.
InterpolatedFrameResources resources;
{
// 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> frame = decode_jpeg(filename_for_frame(stream_idx, frame_no == 1 ? input_second_pts : input_first_pts));
+ JPEGID jpeg_id;
+ jpeg_id.stream_idx = stream_idx;
+ jpeg_id.pts = frame_no == 1 ? input_second_pts : input_first_pts;
+ bool did_decode;
+ shared_ptr<Frame> 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);
}
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();
- GLuint 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, flow_tex, 1280, 720, alpha);
+ // Compute the interpolated frame.
+ qf.flow_tex = compute_flow->exec(resources.gray_tex, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+ check_error();
+ tie(qf.output_tex, qf.output2_tex) = interpolate->exec(resources.input_tex, resources.gray_tex, qf.flow_tex, 1280, 720, alpha);
check_error();
+ // 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);
check_error();
- glGetTextureImage(qf.output_tex, 0, GL_RGBA, GL_UNSIGNED_BYTE, 1280 * 720 * 4, nullptr);
+ glGetTextureImage(qf.output_tex, 0, GL_RED, GL_UNSIGNED_BYTE, 1280 * 720 * 4, BUFFER_OFFSET(0));
+ check_error();
+ glGetTextureImage(qf.output2_tex, 0, GL_RG, GL_UNSIGNED_BYTE, 1280 * 720 * 3, BUFFER_OFFSET(1280 * 720));
check_error();
glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
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);
} else if (qf.type == QueuedFrame::INTERPOLATED) {
glClientWaitSync(qf.fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
- vector<uint8_t> jpeg = encode_jpeg((const uint8_t *)qf.resources.pbo_contents, 1280, 720);
+ vector<uint8_t> jpeg = encode_jpeg(
+ (const uint8_t *)qf.resources.pbo_contents,
+ (const uint8_t *)qf.resources.pbo_contents + 1280 * 720,
+ 1280, 720);
+ compute_flow->release_texture(qf.flow_tex);
+ interpolate->release_texture(qf.output_tex);
+ interpolate->release_texture(qf.output2_tex);
AVPacket pkt;
av_init_packet(&pkt);
projects / nageru / blobdiff