]> git.sesse.net Git - nageru/blobdiff - nageru/mjpeg_encoder.cpp
Fix compilation with FFmpeg 5.0.
[nageru] / nageru / mjpeg_encoder.cpp
index 19ffd58719aa151c01144d850cc987799cd175ed..193027b38483e42359f9733a3d899dcad0b4f8db 100644 (file)
@@ -10,6 +10,7 @@
 
 extern "C" {
 #include <libavformat/avformat.h>
+#include <libavutil/channel_layout.h>
 }
 
 #include "defs.h"
@@ -20,7 +21,7 @@ extern "C" {
 #include "shared/metrics.h"
 #include "pbo_frame_allocator.h"
 #include "shared/timebase.h"
-#include "va_display_with_cleanup.h"
+#include "shared/va_display.h"
 
 #include <movit/colorspace_conversion_effect.h>
 
@@ -33,10 +34,39 @@ using namespace bmusb;
 using namespace movit;
 using namespace std;
 
-static VAImageFormat uyvy_format;
+static VAImageFormat uyvy_format, nv12_format;
 
 extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
 
+// The inverse of memcpy_interleaved(), with (slow) support for pitch.
+void interleave_with_pitch(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, size_t src_width, size_t dst_pitch, size_t height)
+{
+#if __SSE2__
+       if (dst_pitch == src_width * 2 && (src_width * height) % 16 == 0) {
+               __m128i *dptr = reinterpret_cast<__m128i *>(dst);
+               const __m128i *sptr1 = reinterpret_cast<const __m128i *>(src1);
+               const __m128i *sptr2 = reinterpret_cast<const __m128i *>(src2);
+               for (size_t i = 0; i < src_width * height / 16; ++i) {
+                       __m128i data1 = _mm_loadu_si128(sptr1++);
+                       __m128i data2 = _mm_loadu_si128(sptr2++);
+                       _mm_storeu_si128(dptr++, _mm_unpacklo_epi8(data1, data2));
+                       _mm_storeu_si128(dptr++, _mm_unpackhi_epi8(data1, data2));
+               }
+               return;
+       }
+#endif
+
+       for (size_t y = 0; y < height; ++y) {
+               uint8_t *dptr = dst + y * dst_pitch;
+               const uint8_t *sptr1 = src1 + y * src_width;
+               const uint8_t *sptr2 = src2 + y * src_width;
+               for (size_t x = 0; x < src_width; ++x) {
+                       *dptr++ = *sptr1++;
+                       *dptr++ = *sptr2++;
+               }
+       }
+}
+
 // From libjpeg (although it's of course identical between implementations).
 static const int jpeg_natural_order[DCTSIZE2] = {
         0,  1,  8, 16,  9,  2,  3, 10,
@@ -192,14 +222,22 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
        add_stream(HTTPD::StreamID{ HTTPD::MULTICAM_STREAM, 0 });
 
        // Initialize VA-API.
+       VAConfigID config_id_422, config_id_420;
        string error;
-       va_dpy = try_open_va(va_display, &error, &config_id);
+       va_dpy = try_open_va(va_display, { VAProfileJPEGBaseline }, VAEntrypointEncPicture,
+               {
+                       { "4:2:2", VA_RT_FORMAT_YUV422, VA_FOURCC_UYVY, &config_id_422, &uyvy_format },
+                       // We'd prefer VA_FOURCC_I420, but it's not supported by Intel's driver.
+                       { "4:2:0", VA_RT_FORMAT_YUV420, VA_FOURCC_NV12, &config_id_420, &nv12_format }
+               },
+               /*chosen_profile=*/nullptr, &error);
        if (va_dpy == nullptr) {
                fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
        }
 
        encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
        if (va_dpy != nullptr) {
+               va_pool.reset(new VAResourcePool(va_dpy->va_dpy, uyvy_format, nv12_format, config_id_422, config_id_420, /*with_data_buffer=*/true));
                va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this);
        }
 
@@ -242,67 +280,26 @@ void MJPEGEncoder::stop()
        }
 }
 
-unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
-{
-       unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
-       if (va_dpy == nullptr) {
-               if (error) *error = "Opening VA display failed";
-               return nullptr;
-       }
-       int major_ver, minor_ver;
-       VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
-       if (va_status != VA_STATUS_SUCCESS) {
-               char buf[256];
-               snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
-               if (error != nullptr) *error = buf;
-               return nullptr;
-       }
-
-       VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
-       va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
-               &attr, 1, config_id);
-       if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
-               if (error != nullptr) *error = "No hardware support";
-               return nullptr;
-       } else if (va_status != VA_STATUS_SUCCESS) {
-               char buf[256];
-               snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
-               if (error != nullptr) *error = buf;
-               return nullptr;
-       }
-
-       // TODO: Unify with the code in Futatabi.
-       int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
-       assert(num_formats > 0);
-
-       unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
-       va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
-       if (va_status != VA_STATUS_SUCCESS) {
-               char buf[256];
-               snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
-               if (error != nullptr) *error = buf;
-               return nullptr;
-       }
+namespace {
 
-       bool found = false;
-       for (int i = 0; i < num_formats; ++i) {
-               if (formats[i].fourcc == VA_FOURCC_UYVY) {
-                       memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
-                       found = true;
-                       break;
-               }
-       }
-       if (!found) {
-               if (error != nullptr) *error = "UYVY format not found";
-               return nullptr;
-       }
+bool is_uyvy(RefCountedFrame frame)
+{
+       PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
+       return userdata->pixel_format == PixelFormat_8BitYCbCr && frame->interleaved;
+}
 
-       return va_dpy;
+bool is_i420(RefCountedFrame frame)
+{
+       PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
+       return userdata->pixel_format == PixelFormat_8BitYCbCrPlanar &&
+               userdata->ycbcr_format.chroma_subsampling_x == 2 &&
+               userdata->ycbcr_format.chroma_subsampling_y == 2;
 }
 
+}  // namespace
+
 void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset, vector<int32_t> audio, const RGBTriplet &white_balance)
 {
-       PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
        if (video_format.width == 0 || video_format.height == 0) {
                ++metric_mjpeg_frames_zero_size_dropped;
                return;
@@ -312,8 +309,7 @@ void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFram
                ++metric_mjpeg_frames_interlaced_dropped;
                return;
        }
-       if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
-           !frame->interleaved) {
+       if (!is_uyvy(frame) && !is_i420(frame)) {
                fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
                ++metric_mjpeg_frames_unsupported_pixel_format_dropped;
                return;
@@ -456,72 +452,6 @@ private:
        VABufferID buf;
 };
 
-MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
-{
-       {
-               lock_guard<mutex> lock(va_resources_mutex);
-               for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
-                       if (it->width == width && it->height == height) {
-                               VAResources ret = *it;
-                               va_resources_freelist.erase(it);
-                               return ret;
-                       }
-               }
-       }
-
-       VAResources ret;
-
-       ret.width = width;
-       ret.height = height;
-
-       VASurfaceAttrib attrib;
-       attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
-       attrib.type = VASurfaceAttribPixelFormat;
-       attrib.value.type = VAGenericValueTypeInteger;
-       attrib.value.value.i = VA_FOURCC_UYVY;
-
-       VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
-               width, height,
-               &ret.surface, 1, &attrib, 1);
-       CHECK_VASTATUS(va_status, "vaCreateSurfaces");
-
-       va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
-       CHECK_VASTATUS(va_status, "vaCreateContext");
-
-       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
-       CHECK_VASTATUS(va_status, "vaCreateBuffer");
-
-       va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
-       CHECK_VASTATUS(va_status, "vaCreateImage");
-
-       return ret;
-}
-
-void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
-{
-       lock_guard<mutex> lock(va_resources_mutex);
-       if (va_resources_freelist.size() > 50) {
-               auto it = va_resources_freelist.end();
-               --it;
-
-               VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
-               CHECK_VASTATUS(va_status, "vaDestroyBuffer");
-
-               va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
-               CHECK_VASTATUS(va_status, "vaDestroyContext");
-
-               va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
-               CHECK_VASTATUS(va_status, "vaDestroySurfaces");
-
-               va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
-               CHECK_VASTATUS(va_status, "vaDestroyImage");
-
-               va_resources_freelist.erase(it);
-       }
-
-       va_resources_freelist.push_front(resources);
-}
-
 namespace {
 
 void push16(uint16_t val, string *str)
@@ -540,7 +470,7 @@ void push32(uint32_t val, string *str)
 
 }  // namespace
 
-void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
+void MJPEGEncoder::init_jpeg(unsigned width, unsigned height, const RGBTriplet &white_balance, VectorDestinationManager *dest, jpeg_compress_struct *cinfo, int y_h_samp_factor, int y_v_samp_factor)
 {
        jpeg_error_mgr jerr;
        cinfo->err = jpeg_std_error(&jerr);
@@ -556,8 +486,8 @@ void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, const RGBTripl
        cinfo->image_height = height;
        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[0].h_samp_factor = y_h_samp_factor;
+       cinfo->comp_info[0].v_samp_factor = y_v_samp_factor;
        cinfo->comp_info[1].h_samp_factor = 1;
        cinfo->comp_info[1].v_samp_factor = 1;
        cinfo->comp_info[2].h_samp_factor = 1;
@@ -625,25 +555,31 @@ void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, const RGBTripl
        jpeg_write_marker(cinfo, JPEG_COM, (const JOCTET *)"CS=ITU601", strlen("CS=ITU601"));
 }
 
-vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, jpeg_compress_struct *cinfo)
+vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, const RGBTriplet &white_balance, int y_h_samp_factor, int y_v_samp_factor, jpeg_compress_struct *cinfo)
 {
        VectorDestinationManager dest;
-       init_jpeg_422(width, height, white_balance, &dest, cinfo);
+       init_jpeg(width, height, white_balance, &dest, cinfo, y_h_samp_factor, y_v_samp_factor);
 
        // Make a dummy black image; there's seemingly no other easy way of
        // making libjpeg outputting all of its headers.
-       JSAMPROW yptr[8], cbptr[8], crptr[8];
+       assert(y_v_samp_factor <= 2);  // Or we'd need larger JSAMPROW arrays below.
+       size_t block_height_y = 8 * y_v_samp_factor;
+       size_t block_height_cbcr = 8;
+
+       JSAMPROW yptr[16], cbptr[16], crptr[16];
        JSAMPARRAY data[3] = { yptr, cbptr, crptr };
        memset(tmp_y, 0, 4096);
        memset(tmp_cb, 0, 4096);
        memset(tmp_cr, 0, 4096);
-       for (unsigned yy = 0; yy < 8; ++yy) {
+       for (unsigned yy = 0; yy < block_height_y; ++yy) {
                yptr[yy] = tmp_y;
+       }
+       for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
                cbptr[yy] = tmp_cb;
                crptr[yy] = tmp_cr;
        }
-       for (unsigned y = 0; y < height; y += 8) {
-               jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
+       for (unsigned y = 0; y < height; y += block_height_y) {
+               jpeg_write_raw_data(cinfo, data, block_height_y);
        }
        jpeg_finish_compress(cinfo);
 
@@ -660,17 +596,17 @@ vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, c
        return dest.dest;
 }
 
-MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height, const RGBTriplet &white_balance)
+MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_parameters(unsigned width, unsigned height, unsigned y_h_samp_factor, unsigned y_v_samp_factor, const RGBTriplet &white_balance)
 {
-       pair<unsigned, unsigned> key(width, height);
-       if (va_data_for_resolution.count(key)) {
-               return va_data_for_resolution[key];
+       VAKey key{width, height, y_h_samp_factor, y_v_samp_factor, white_balance};
+       if (va_data_for_parameters.count(key)) {
+               return va_data_for_parameters[key];
        }
 
        // Use libjpeg to generate a header and set sane defaults for e.g.
        // quantization tables. Then do the actual encode with VA-API.
        jpeg_compress_struct cinfo;
-       vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, &cinfo);
+       vector<uint8_t> jpeg_header = get_jpeg_header(width, height, white_balance, y_h_samp_factor, y_v_samp_factor, &cinfo);
 
        // Picture parameters.
        VAEncPictureParameterBufferJPEG pic_param;
@@ -765,7 +701,7 @@ MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, un
        ret.q = q;
        ret.huff = huff;
        ret.parms = parms;
-       va_data_for_resolution[key] = ret;
+       va_data_for_parameters[key] = ret;
        return ret;
 }
 
@@ -775,37 +711,55 @@ void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
        unsigned width = qf.video_format.width;
        unsigned height = qf.video_format.height;
 
-       VAResources resources;
+       VAResourcePool::VAResources resources;
        ReleaseVAResources release;
        if (userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_VA_API) {
+               assert(is_uyvy(qf.frame));
                resources = move(userdata->va_resources);
                release = move(userdata->va_resources_release);
        } else {
                assert(userdata->data_copy_current_src == PBOFrameAllocator::Userdata::FROM_MALLOC);
-               resources = get_va_resources(width, height);
-               release = ReleaseVAResources(this, resources);
+               if (is_uyvy(qf.frame)) {
+                       resources = va_pool->get_va_resources(width, height, VA_FOURCC_UYVY);
+               } else {
+                       assert(is_i420(qf.frame));
+                       resources = va_pool->get_va_resources(width, height, VA_FOURCC_NV12);
+               }
+               release = ReleaseVAResources(va_pool.get(), resources);
+       }
+
+       int y_h_samp_factor, y_v_samp_factor;
+       if (is_uyvy(qf.frame)) {
+               // 4:2:2 (sample Y' twice as often horizontally as Cb or Cr, vertical is left alone).
+               y_h_samp_factor = 2;
+               y_v_samp_factor = 1;
+       } else {
+               // 4:2:0 (sample Y' twice as often as Cb or Cr, in both directions)
+               assert(is_i420(qf.frame));
+               y_h_samp_factor = 2;
+               y_v_samp_factor = 2;
        }
 
-       VAData va_data = get_va_data_for_resolution(width, height, qf.white_balance);
+       VAData va_data = get_va_data_for_parameters(width, height, y_h_samp_factor, y_v_samp_factor, qf.white_balance);
        va_data.pic_param.coded_buf = resources.data_buffer;
 
        VABufferID pic_param_buffer;
-       VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
+       VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 
        VABufferID q_buffer;
-       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
+       va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
 
        VABufferID huff_buffer;
-       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
+       va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 
        VABufferID slice_param_buffer;
-       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
+       va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 
@@ -820,13 +774,27 @@ void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
                uint8_t *surface_p = nullptr;
                vaMapBuffer(va_dpy->va_dpy, resources.image.buf, (void **)&surface_p);
 
-               size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
-               size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
+               if (is_uyvy(qf.frame)) {
+                       size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
+                       size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 
-               {
                        const uint8_t *src = qf.frame->data_copy + field_start;
                        uint8_t *dst = (unsigned char *)surface_p + resources.image.offsets[0];
                        memcpy_with_pitch(dst, src, qf.video_format.width * 2, resources.image.pitches[0], qf.video_format.height);
+               } else {
+                       assert(is_i420(qf.frame));
+                       assert(!qf.frame->interleaved);  // Makes no sense for I420.
+
+                       size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
+                       const uint8_t *y_src = qf.frame->data + qf.video_format.width * field_start_line;
+                       const uint8_t *cb_src = y_src + width * height;
+                       const uint8_t *cr_src = cb_src + (width / 2) * (height / 2);
+
+                       uint8_t *y_dst = (unsigned char *)surface_p + resources.image.offsets[0];
+                       uint8_t *cbcr_dst = (unsigned char *)surface_p + resources.image.offsets[1];
+
+                       memcpy_with_pitch(y_dst, y_src, qf.video_format.width, resources.image.pitches[0], qf.video_format.height);
+                       interleave_with_pitch(cbcr_dst, cb_src, cr_src, qf.video_format.width / 2, resources.image.pitches[1], qf.video_format.height / 2);
                }
 
                va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
@@ -848,12 +816,12 @@ void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf)
        header_parm.bit_length = 8 * va_data.jpeg_header.size();
 
        VABufferID header_parm_buffer;
-       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
+       va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
 
        VABufferID header_data_buffer;
-       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
+       va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
        CHECK_VASTATUS(va_status, "vaCreateBuffer");
        VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
 
@@ -938,24 +906,59 @@ vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
 
        VectorDestinationManager dest;
        jpeg_compress_struct cinfo;
-       init_jpeg_422(width, height, qf.white_balance, &dest, &cinfo);
 
        size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
-       size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
 
-       JSAMPROW yptr[8], cbptr[8], crptr[8];
-       JSAMPARRAY data[3] = { yptr, cbptr, crptr };
-       for (unsigned y = 0; y < qf.video_format.height; y += 8) {
-               const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
-
-               memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
-               memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
-               for (unsigned yy = 0; yy < 8; ++yy) {
-                       yptr[yy] = tmp_y + yy * width;
-                       cbptr[yy] = tmp_cb + yy * width / 2;
-                       crptr[yy] = tmp_cr + yy * width / 2;
+       PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)qf.frame->userdata;
+       if (userdata->pixel_format == PixelFormat_8BitYCbCr) {
+               init_jpeg(width, height, qf.white_balance, &dest, &cinfo, /*y_h_samp_factor=*/2, /*y_v_samp_factor=*/1);
+
+               assert(qf.frame->interleaved);
+               size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
+
+               JSAMPROW yptr[8], cbptr[8], crptr[8];
+               JSAMPARRAY data[3] = { yptr, cbptr, crptr };
+               for (unsigned y = 0; y < qf.video_format.height; y += 8) {
+                       const uint8_t *src;
+                       src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
+
+                       memcpy_interleaved(tmp_cbcr, tmp_y, src, qf.video_format.width * 8 * 2);
+                       memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
+                       for (unsigned yy = 0; yy < 8; ++yy) {
+                               yptr[yy] = tmp_y + yy * width;
+                               cbptr[yy] = tmp_cb + yy * width / 2;
+                               crptr[yy] = tmp_cr + yy * width / 2;
+                       }
+                       jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
+               }
+       } else {
+               assert(userdata->pixel_format == PixelFormat_8BitYCbCrPlanar);
+
+               const movit::YCbCrFormat &ycbcr = userdata->ycbcr_format;
+               init_jpeg(width, height, qf.white_balance, &dest, &cinfo, ycbcr.chroma_subsampling_x, ycbcr.chroma_subsampling_y);
+               assert(ycbcr.chroma_subsampling_y <= 2);  // Or we'd need larger JSAMPROW arrays below.
+
+               size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
+               const uint8_t *y_start = qf.frame->data + qf.video_format.width * field_start_line;
+               const uint8_t *cb_start = y_start + width * height;
+               const uint8_t *cr_start = cb_start + (width / ycbcr.chroma_subsampling_x) * (height / ycbcr.chroma_subsampling_y);
+
+               size_t block_height_y = 8 * ycbcr.chroma_subsampling_y;
+               size_t block_height_cbcr = 8;
+
+               JSAMPROW yptr[16], cbptr[16], crptr[16];
+               JSAMPARRAY data[3] = { yptr, cbptr, crptr };
+               for (unsigned y = 0; y < qf.video_format.height; y += block_height_y) {
+                       for (unsigned yy = 0; yy < block_height_y; ++yy) {
+                               yptr[yy] = const_cast<JSAMPROW>(y_start) + (y + yy) * width;
+                       }
+                       unsigned cbcr_y = y / ycbcr.chroma_subsampling_y;
+                       for (unsigned yy = 0; yy < block_height_cbcr; ++yy) {
+                               cbptr[yy] = const_cast<JSAMPROW>(cb_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
+                               crptr[yy] = const_cast<JSAMPROW>(cr_start) + (cbcr_y + yy) * width / ycbcr.chroma_subsampling_x;
+                       }
+                       jpeg_write_raw_data(&cinfo, data, block_height_y);
                }
-               jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
        }
        jpeg_finish_compress(&cinfo);