using std::cerr;
using std::list;
using std::runtime_error;
-using boost::shared_ptr;
+using std::shared_ptr;
using dcp::Size;
+
+/** The memory alignment, in bytes, used for each row of an image if aligment is requested */
+#define ALIGNMENT 64
+
+
int
Image::vertical_factor (int n) const
{
throw PixelFormatError ("line_factor()", _pixel_format);
}
- return pow (2.0f, d->log2_chroma_h);
+ return lrintf(powf(2.0f, d->log2_chroma_h));
}
int
throw PixelFormatError ("sample_size()", _pixel_format);
}
- return pow (2.0f, d->log2_chroma_w);
+ return lrintf(powf(2.0f, d->log2_chroma_w));
}
/** @param n Component index.
return d->nb_components;
}
+
+static
+int
+round_width_for_subsampling (int p, AVPixFmtDescriptor const * desc)
+{
+ return p & ~ ((1 << desc->log2_chroma_w) - 1);
+}
+
+
+static
+int
+round_height_for_subsampling (int p, AVPixFmtDescriptor const * desc)
+{
+ return p & ~ ((1 << desc->log2_chroma_h) - 1);
+}
+
+
/** Crop this image, scale it to `inter_size' and then place it in a black frame of `out_size'.
* @param crop Amount to crop by.
* @param inter_size Size to scale the cropped image to.
* @param out_size Size of output frame; if this is larger than inter_size there will be black padding.
* @param yuv_to_rgb YUV to RGB transformation to use, if required.
+ * @param video_range Video range of the image.
* @param out_format Output pixel format.
* @param out_aligned true to make the output image aligned.
+ * @param out_video_range Video range to use for the output image.
* @param fast Try to be fast at the possible expense of quality; at present this means using
* fast bilinear rather than bicubic scaling.
*/
shared_ptr<Image>
Image::crop_scale_window (
- Crop crop, dcp::Size inter_size, dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_format, bool out_aligned, bool fast
+ Crop crop,
+ dcp::Size inter_size,
+ dcp::Size out_size,
+ dcp::YUVToRGB yuv_to_rgb,
+ VideoRange video_range,
+ AVPixelFormat out_format,
+ VideoRange out_video_range,
+ bool out_aligned,
+ bool fast
) const
{
/* Empirical testing suggests that sws_scale() will crash if
DCPOMATIC_ASSERT (out_size.width >= inter_size.width);
DCPOMATIC_ASSERT (out_size.height >= inter_size.height);
- /* Here's an image of out_size. Below we may write to it starting at an offset so we get some padding.
- Hence we want to write in the following pattern:
-
- block start write start line end
- |..(padding)..|<------line-size------------->|..(padding)..|
- |..(padding)..|<------line-size------------->|..(padding)..|
- |..(padding)..|<------line-size------------->|..(padding)..|
+ shared_ptr<Image> out (new Image(out_format, out_size, out_aligned));
+ out->make_black ();
- where line-size is of the smaller (inter_size) image and the full padded line length is that of
- out_size. To get things to work we have to tell FFmpeg that the stride is that of out_size.
- However some parts of FFmpeg (notably rgb48Toxyz12 in swscale.c) process data for the full
- specified *stride*. This does not matter until we get to the last line:
+ AVPixFmtDescriptor const * in_desc = av_pix_fmt_desc_get (_pixel_format);
+ if (!in_desc) {
+ throw PixelFormatError ("crop_scale_window()", _pixel_format);
+ }
- block start write start line end
- |..(padding)..|<------line-size------------->|XXXwrittenXXX|
- |XXXwrittenXXX|<------line-size------------->|XXXwrittenXXX|
- |XXXwrittenXXX|<------line-size------------->|XXXwrittenXXXXXXwrittenXXX
- ^^^^ out of bounds
+ /* Round down so that we crop only the number of pixels that is straightforward
+ * considering any subsampling.
+ */
+ Crop corrected_crop(
+ round_width_for_subsampling(crop.left, in_desc),
+ round_width_for_subsampling(crop.right, in_desc),
+ round_height_for_subsampling(crop.top, in_desc),
+ round_height_for_subsampling(crop.bottom, in_desc)
+ );
- To get around this, we ask Image to overallocate its buffers by the overrun.
- */
+ /* Also check that we aren't cropping more image than there actually is */
+ if ((corrected_crop.left + corrected_crop.right) >= (size().width - 4)) {
+ corrected_crop.left = 0;
+ corrected_crop.right = size().width - 4;
+ }
- shared_ptr<Image> out (new Image (out_format, out_size, out_aligned, (out_size.width - inter_size.width) / 2));
- out->make_black ();
+ if ((corrected_crop.top + corrected_crop.bottom) >= (size().height - 4)) {
+ corrected_crop.top = 0;
+ corrected_crop.bottom = size().height - 4;
+ }
/* Size of the image after any crop */
- dcp::Size const cropped_size = crop.apply (size ());
+ dcp::Size const cropped_size = corrected_crop.apply (size());
/* Scale context for a scale from cropped_size to inter_size */
struct SwsContext* scale_context = sws_getContext (
throw runtime_error (N_("Could not allocate SwsContext"));
}
- DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUV_TO_RGB_COUNT);
- int const lut[dcp::YUV_TO_RGB_COUNT] = {
+ DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUVToRGB::COUNT);
+ int const lut[static_cast<int>(dcp::YUVToRGB::COUNT)] = {
SWS_CS_ITU601,
SWS_CS_ITU709
};
0 -> destination range MPEG (i.e. "video", 16-235)
1 -> destination range JPEG (i.e. "full", 0-255)
- But remember: sws_setColorspaceDetails ignores
- these parameters unless the image isYUV or isGray
- (if it's neither, it uses video range for source
- and destination).
+ But remember: sws_setColorspaceDetails ignores these
+ parameters unless the both source and destination images
+ are isYUV or isGray. (If either is not, it uses video range).
*/
sws_setColorspaceDetails (
scale_context,
- sws_getCoefficients (lut[yuv_to_rgb]), 0,
- sws_getCoefficients (lut[yuv_to_rgb]), 0,
+ sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), video_range == VIDEO_RANGE_VIDEO ? 0 : 1,
+ sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), out_video_range == VIDEO_RANGE_VIDEO ? 0 : 1,
0, 1 << 16, 1 << 16
);
- AVPixFmtDescriptor const * in_desc = av_pix_fmt_desc_get (_pixel_format);
- if (!in_desc) {
- throw PixelFormatError ("crop_scale_window()", _pixel_format);
- }
-
/* Prepare input data pointers with crop */
uint8_t* scale_in_data[planes()];
for (int c = 0; c < planes(); ++c) {
- /* To work out the crop in bytes, start by multiplying
- the crop by the (average) bytes per pixel. Then
- round down so that we don't crop a subsampled pixel until
- we've cropped all of its Y-channel pixels.
- */
- int const x = lrintf (bytes_per_pixel(c) * crop.left) & ~ ((int) in_desc->log2_chroma_w);
- scale_in_data[c] = data()[c] + x + stride()[c] * (crop.top / vertical_factor(c));
+ int const x = lrintf(bytes_per_pixel(c) * corrected_crop.left);
+ scale_in_data[c] = data()[c] + x + stride()[c] * (corrected_crop.top / vertical_factor(c));
}
- /* Corner of the image within out_size */
- Position<int> const corner ((out_size.width - inter_size.width) / 2, (out_size.height - inter_size.height) / 2);
-
AVPixFmtDescriptor const * out_desc = av_pix_fmt_desc_get (out_format);
if (!out_desc) {
throw PixelFormatError ("crop_scale_window()", out_format);
}
+ /* Corner of the image within out_size */
+ Position<int> const corner (
+ round_width_for_subsampling((out_size.width - inter_size.width) / 2, out_desc),
+ round_height_for_subsampling((out_size.height - inter_size.height) / 2, out_desc)
+ );
+
uint8_t* scale_out_data[out->planes()];
for (int c = 0; c < out->planes(); ++c) {
- /* See the note in the crop loop above */
- int const x = lrintf (out->bytes_per_pixel(c) * corner.x) & ~ ((int) out_desc->log2_chroma_w);
+ int const x = lrintf(out->bytes_per_pixel(c) * corner.x);
scale_out_data[c] = out->data()[c] + x + out->stride()[c] * (corner.y / out->vertical_factor(c));
}
sws_freeContext (scale_context);
+ if (corrected_crop != Crop() && cropped_size == inter_size) {
+ /* We are cropping without any scaling or pixel format conversion, so FFmpeg may have left some
+ data behind in our image. Clear it out. It may get to the point where we should just stop
+ trying to be clever with cropping.
+ */
+ out->make_part_black (corner.x + cropped_size.width, out_size.width - cropped_size.width);
+ }
+
return out;
}
(fast ? SWS_FAST_BILINEAR : SWS_BICUBIC) | SWS_ACCURATE_RND, 0, 0, 0
);
- DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUV_TO_RGB_COUNT);
- int const lut[dcp::YUV_TO_RGB_COUNT] = {
+ DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUVToRGB::COUNT);
+ int const lut[static_cast<int>(dcp::YUVToRGB::COUNT)] = {
SWS_CS_ITU601,
SWS_CS_ITU709
};
0 -> destination range MPEG (i.e. "video", 16-235)
1 -> destination range JPEG (i.e. "full", 0-255)
- But remember: sws_setColorspaceDetails ignores
- these parameters unless the image isYUV or isGray
- (if it's neither, it uses video range for source
- and destination).
+ But remember: sws_setColorspaceDetails ignores these
+ parameters unless the corresponding image isYUV or isGray.
+ (If it's neither, it uses video range).
*/
sws_setColorspaceDetails (
scale_context,
- sws_getCoefficients (lut[yuv_to_rgb]), 0,
- sws_getCoefficients (lut[yuv_to_rgb]), 0,
+ sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), 0,
+ sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), 0,
0, 1 << 16, 1 << 16
);
return ((v >> 8) & 0xff) | ((v & 0xff) << 8);
}
+void
+Image::make_part_black (int x, int w)
+{
+ switch (_pixel_format) {
+ case AV_PIX_FMT_RGB24:
+ case AV_PIX_FMT_ARGB:
+ case AV_PIX_FMT_RGBA:
+ case AV_PIX_FMT_ABGR:
+ case AV_PIX_FMT_BGRA:
+ case AV_PIX_FMT_RGB555LE:
+ case AV_PIX_FMT_RGB48LE:
+ case AV_PIX_FMT_RGB48BE:
+ case AV_PIX_FMT_XYZ12LE:
+ {
+ int const h = sample_size(0).height;
+ int const bpp = bytes_per_pixel(0);
+ int const s = stride()[0];
+ uint8_t* p = data()[0];
+ for (int y = 0; y < h; y++) {
+ memset (p + x * bpp, 0, w * bpp);
+ p += s;
+ }
+ break;
+ }
+
+ default:
+ throw PixelFormatError ("make_part_black()", _pixel_format);
+ }
+}
+
void
Image::make_black ()
{
}
case AV_PIX_FMT_YUV420P:
{
- shared_ptr<Image> yuv = other->convert_pixel_format (dcp::YUV_TO_RGB_REC709, _pixel_format, false, false);
+ shared_ptr<Image> yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, false, false);
dcp::Size const ts = size();
dcp::Size const os = yuv->size();
for (int ty = start_ty, oy = start_oy; ty < ts.height && oy < os.height; ++ty, ++oy) {
}
case AV_PIX_FMT_YUV420P10:
{
- shared_ptr<Image> yuv = other->convert_pixel_format (dcp::YUV_TO_RGB_REC709, _pixel_format, false, false);
+ shared_ptr<Image> yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, false, false);
dcp::Size const ts = size();
dcp::Size const os = yuv->size();
for (int ty = start_ty, oy = start_oy; ty < ts.height && oy < os.height; ++ty, ++oy) {
}
case AV_PIX_FMT_YUV422P10LE:
{
- shared_ptr<Image> yuv = other->convert_pixel_format (dcp::YUV_TO_RGB_REC709, _pixel_format, false, false);
+ shared_ptr<Image> yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, false, false);
dcp::Size const ts = size();
dcp::Size const os = yuv->size();
for (int ty = start_ty, oy = start_oy; ty < ts.height && oy < os.height; ++ty, ++oy) {
*
* @param p Pixel format.
* @param s Size in pixels.
- * @param aligned true to make each row of this image aligned to a 32-byte boundary.
- * @param extra_pixels Amount of extra "run-off" memory to allocate at the end of each plane in pixels.
+ * @param aligned true to make each row of this image aligned to a ALIGNMENT-byte boundary.
*/
-Image::Image (AVPixelFormat p, dcp::Size s, bool aligned, int extra_pixels)
+Image::Image (AVPixelFormat p, dcp::Size s, bool aligned)
: _size (s)
, _pixel_format (p)
, _aligned (aligned)
- , _extra_pixels (extra_pixels)
{
allocate ();
}
for (int i = 0; i < planes(); ++i) {
_line_size[i] = ceil (_size.width * bytes_per_pixel(i));
- _stride[i] = stride_round_up (i, _line_size, _aligned ? 32 : 1);
+ _stride[i] = stride_round_up (i, _line_size, _aligned ? ALIGNMENT : 1);
/* The assembler function ff_rgb24ToY_avx (in libswscale/x86/input.asm)
uses a 16-byte fetch to read three bytes (R/G/B) of image data.
Further to the above, valgrind is now telling me that ff_rgb24ToY_ssse3
over-reads by more then _avx. I can't follow the code to work out how much,
- so I'll just over-allocate by 32 bytes and have done with it. Empirical
+ so I'll just over-allocate by ALIGNMENT bytes and have done with it. Empirical
testing suggests that it works.
+
+ In addition to these concerns, we may read/write as much as a whole extra line
+ at the end of each plane in cases where we are messing with offsets in order to
+ do pad or crop. To solve this we over-allocate by an extra _stride[i] bytes.
+
+ As an example: we may write to images starting at an offset so we get some padding.
+ Hence we want to write in the following pattern:
+
+ block start write start line end
+ |..(padding)..|<------line-size------------->|..(padding)..|
+ |..(padding)..|<------line-size------------->|..(padding)..|
+ |..(padding)..|<------line-size------------->|..(padding)..|
+
+ where line-size is of the smaller (inter_size) image and the full padded line length is that of
+ out_size. To get things to work we have to tell FFmpeg that the stride is that of out_size.
+ However some parts of FFmpeg (notably rgb48Toxyz12 in swscale.c) process data for the full
+ specified *stride*. This does not matter until we get to the last line:
+
+ block start write start line end
+ |..(padding)..|<------line-size------------->|XXXwrittenXXX|
+ |XXXwrittenXXX|<------line-size------------->|XXXwrittenXXX|
+ |XXXwrittenXXX|<------line-size------------->|XXXwrittenXXXXXXwrittenXXX
+ ^^^^ out of bounds
*/
- _data[i] = (uint8_t *) wrapped_av_malloc (_stride[i] * sample_size(i).height + _extra_pixels * bytes_per_pixel(i) + 32);
+ _data[i] = (uint8_t *) wrapped_av_malloc (_stride[i] * (sample_size(i).height + 1) + ALIGNMENT);
#if HAVE_VALGRIND_MEMCHECK_H
/* The data between the end of the line size and the stride is undefined but processed by
libswscale, causing lots of valgrind errors. Mark it all defined to quell these errors.
*/
- VALGRIND_MAKE_MEM_DEFINED (_data[i], _stride[i] * sample_size(i).height + _extra_pixels * bytes_per_pixel(i) + 32);
+ VALGRIND_MAKE_MEM_DEFINED (_data[i], _stride[i] * (sample_size(i).height + 1) + ALIGNMENT);
#endif
}
}
Image::Image (Image const & other)
- : boost::enable_shared_from_this<Image>(other)
+ : std::enable_shared_from_this<Image>(other)
, _size (other._size)
, _pixel_format (other._pixel_format)
, _aligned (other._aligned)
- , _extra_pixels (other._extra_pixels)
{
allocate ();
: _size (frame->width, frame->height)
, _pixel_format (static_cast<AVPixelFormat> (frame->format))
, _aligned (true)
- , _extra_pixels (0)
{
allocate ();
: _size (other->_size)
, _pixel_format (other->_pixel_format)
, _aligned (aligned)
- , _extra_pixels (other->_extra_pixels)
{
allocate ();
}
std::swap (_aligned, other._aligned);
- std::swap (_extra_pixels, other._extra_pixels);
}
-/** Destroy a Image */
Image::~Image ()
{
for (int i = 0; i < planes(); ++i) {
throw EncodeError (String::compose ("Error during PNG write: %1", message));
}
-dcp::Data
+dcp::ArrayData
Image::as_png () const
{
DCPOMATIC_ASSERT (bytes_per_pixel(0) == 4);
DCPOMATIC_ASSERT (planes() == 1);
if (pixel_format() != AV_PIX_FMT_RGBA) {
- return convert_pixel_format(dcp::YUV_TO_RGB_REC709, AV_PIX_FMT_RGBA, true, false)->as_png();
+ return convert_pixel_format(dcp::YUVToRGB::REC709, AV_PIX_FMT_RGBA, true, false)->as_png();
}
/* error handling? */
png_destroy_write_struct (&png_ptr, &info_ptr);
png_free (png_ptr, row_pointers);
- return dcp::Data (state.data, state.size);
+ return dcp::ArrayData (state.data, state.size);
+}
+
+
+void
+Image::video_range_to_full_range ()
+{
+ switch (_pixel_format) {
+ case AV_PIX_FMT_RGB24:
+ {
+ float const factor = 256.0 / 219.0;
+ uint8_t* p = data()[0];
+ int const lines = sample_size(0).height;
+ for (int y = 0; y < lines; ++y) {
+ uint8_t* q = p;
+ for (int x = 0; x < line_size()[0]; ++x) {
+ *q = int((*q - 16) * factor);
+ ++q;
+ }
+ p += stride()[0];
+ }
+ break;
+ }
+ case AV_PIX_FMT_GBRP12LE:
+ {
+ float const factor = 4096.0 / 3504.0;
+ for (int c = 0; c < 3; ++c) {
+ uint16_t* p = reinterpret_cast<uint16_t*>(data()[c]);
+ int const lines = sample_size(c).height;
+ for (int y = 0; y < lines; ++y) {
+ uint16_t* q = p;
+ int const line_size_pixels = line_size()[c] / 2;
+ for (int x = 0; x < line_size_pixels; ++x) {
+ *q = int((*q - 256) * factor);
+ ++q;
+ }
+ }
+ }
+ break;
+ }
+ default:
+ throw PixelFormatError ("video_range_to_full_range()", _pixel_format);
+ }
}
+
projects / dcpomatic.git / blobdiff