#include <jpeglib.h>
#include <unistd.h>
+#include "chroma_subsampler.h"
#include "context.h"
#include "flow.h"
#include "httpd.h"
#include "mux.h"
#include "player.h"
#include "util.h"
+#include "ycbcr_converter.h"
#include <epoxy/glx.h>
};
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 *cb_data, const uint8_t *cr_data, unsigned width, unsigned height)
{
VectorDestinationManager dest;
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<uint8_t[]> ydata(new uint8_t[width * 8]);
- 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) {
- 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<JSAMPROW>(&y_data[(y + yy) * width]);
+ cbptr[yy] = const_cast<JSAMPROW>(&cb_data[(y + yy) * width/2]);
+ crptr[yy] = const_cast<JSAMPROW>(&cr_data[(y + yy) * width/2]);
}
jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
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));
GLuint input_tex[num_interpolate_slots], gray_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, 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);
check_error();
glTextureStorage3D(gray_tex[i], levels, GL_R8, width, height, 2);
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.cb_tex = cb_tex[i];
+ resource.cr_tex = cr_tex[i];
glCreateFramebuffers(2, resource.input_fbos);
check_error();
check_error();
- compute_flow.reset(new DISComputeFlow(width, height, operating_point3));
- 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));
+ chroma_subsampler.reset(new ChromaSubsampler);
check_error();
}
unique_lock<mutex> 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();
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;
+ jpeg_id.interpolated = false;
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);
- 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);
+ 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.
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);
+ 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.
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(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);
queue_nonempty.notify_all();
}
+namespace {
+
+shared_ptr<Frame> 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> 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");
} 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);
+
+ // Send a copy of the frame on to display.
+ shared_ptr<Frame> frame = frame_from_pbo(qf.resources.pbo_contents, 1280, 720);
+ JPEGFrameView::insert_interpolated_frame(qf.stream_idx, qf.output_pts, frame);
+
+ // Now JPEG encode it, and send it on to the stream.
+ vector<uint8_t> 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);
+ interpolate->release_texture(qf.cbcr_tex);
AVPacket pkt;
av_init_packet(&pkt);
projects / nageru / blobdiff