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,
// Initialize VA-API.
string error;
- va_dpy = try_open_va(va_display, &error, &config_id);
+ va_dpy = try_open_va(va_display, &error, &config_id_422, &config_id_420);
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());
}
}
}
-unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
+unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id_422, VAConfigID *config_id_420)
{
unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
if (va_dpy == nullptr) {
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;
+ {
+ VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
+ va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
+ &attr, 1, config_id_422);
+ if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
+ if (error != nullptr) *error = "No 4:2:2 hardware support";
+ return nullptr;
+ } else if (va_status != VA_STATUS_SUCCESS) {
+ char buf[256];
+ snprintf(buf, sizeof(buf), "vaCreateConfig() for 4:2:2 failed with status %d\n", va_status);
+ if (error != nullptr) *error = buf;
+ return nullptr;
+ }
+ }
+ {
+ VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV420 };
+ va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
+ &attr, 1, config_id_420);
+ if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
+ if (error != nullptr) *error = "No 4:2:0 hardware support";
+ return nullptr;
+ } else if (va_status != VA_STATUS_SUCCESS) {
+ char buf[256];
+ snprintf(buf, sizeof(buf), "vaCreateConfig() for 4:2:0 failed with status %d\n", va_status);
+ if (error != nullptr) *error = buf;
+ return nullptr;
+ }
}
// TODO: Unify with the code in Futatabi.
return nullptr;
}
- bool found = false;
+ bool uyvy_found = false, nv12_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;
+ uyvy_found = true;
+ }
+ if (formats[i].fourcc == VA_FOURCC_NV12) {
+ memcpy(&nv12_format, &formats[i], sizeof(VAImageFormat));
+ nv12_found = true;
}
}
- if (!found) {
+ if (!uyvy_found) {
if (error != nullptr) *error = "UYVY format not found";
return nullptr;
}
+ if (!nv12_found) {
+ if (error != nullptr) *error = "NV12 format not found";
+ return nullptr;
+ }
return va_dpy;
}
-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)
+namespace {
+
+bool is_uyvy(RefCountedFrame frame)
+{
+ PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
+ return userdata->pixel_format == PixelFormat_8BitYCbCr && frame->interleaved;
+}
+
+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)
+{
if (video_format.width == 0 || video_format.height == 0) {
++metric_mjpeg_frames_zero_size_dropped;
return;
++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;
VABufferID buf;
};
-MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
+MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height, uint32_t fourcc)
{
{
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) {
+ if (it->width == width && it->height == height && it->fourcc == fourcc) {
VAResources ret = *it;
va_resources_freelist.erase(it);
return ret;
ret.width = width;
ret.height = height;
+ ret.fourcc = fourcc;
VASurfaceAttrib attrib;
attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
attrib.type = VASurfaceAttribPixelFormat;
attrib.value.type = VAGenericValueTypeInteger;
- attrib.value.value.i = VA_FOURCC_UYVY;
+ attrib.value.value.i = fourcc;
- VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
- width, height,
- &ret.surface, 1, &attrib, 1);
- CHECK_VASTATUS(va_status, "vaCreateSurfaces");
+ VAStatus va_status;
+ VAConfigID config_id;
+ if (fourcc == VA_FOURCC_UYVY) {
+ va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422, width, height, &ret.surface, 1, &attrib, 1);
+ config_id = config_id_422;
+ } else {
+ assert(fourcc == VA_FOURCC_NV12);
+ va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV420, width, height, &ret.surface, 1, &attrib, 1);
+ config_id = config_id_420;
+ }
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, ret.context, 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");
+ if (fourcc == VA_FOURCC_UYVY) {
+ va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
+ CHECK_VASTATUS(va_status, "vaCreateImage");
+ } else {
+ assert(fourcc == VA_FOURCC_NV12);
+ va_status = vaCreateImage(va_dpy->va_dpy, &nv12_format, width, height, &ret.image);
+ CHECK_VASTATUS(va_status, "vaCreateImage");
+ }
return ret;
}
} // 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);
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;
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);
return dest.dest;
}
-MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_parameters(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)
{
- VAKey key{width, height, white_balance};
+ 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;
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);
+ if (is_uyvy(qf.frame)) {
+ resources = get_va_resources(width, height, VA_FOURCC_UYVY);
+ } else {
+ assert(is_i420(qf.frame));
+ // We'd prefer VA_FOURCC_I420, but it's not supported by Intel's driver.
+ resources = get_va_resources(width, height, VA_FOURCC_NV12);
+ }
release = ReleaseVAResources(this, resources);
}
- VAData va_data = get_va_data_for_parameters(width, height, qf.white_balance);
+ int y_h_samp_factor, y_v_samp_factor;
+ if (is_uyvy(qf.frame)) {
+ // 4:2:2 (sample Y' twice as often vertically as Cb or Cr).
+ y_h_samp_factor = 1;
+ y_v_samp_factor = 2;
+ } 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_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;
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);
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);