DCPOMATIC_ASSERT (out_size.width >= inter_size.width);
DCPOMATIC_ASSERT (out_size.height >= inter_size.height);
- /* Here's an image of out_size */
- shared_ptr<Image> out (new Image (out_format, out_size, out_aligned));
+ /* 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)..|
+
+ 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
+
+ To get around this, we ask Image to overallocate its buffers by the overrun.
+ */
+
+ shared_ptr<Image> out (new Image (out_format, out_size, out_aligned, (out_size.width - inter_size.width) / 2));
out->make_black ();
/* Size of the image after any crop */
case AV_PIX_FMT_RGB555LE:
case AV_PIX_FMT_RGB48LE:
case AV_PIX_FMT_RGB48BE:
+ case AV_PIX_FMT_XYZ12LE:
memset (data()[0], 0, sample_size(0).height * stride()[0]);
break;
uint8_t* op = other->data()[0] + oy * other->stride()[0];
for (int tx = start_tx, ox = start_ox; tx < size().width && ox < other->size().width; ++tx, ++ox) {
float const alpha = float (op[3]) / 255;
- /* Blend high bytes */
- tp[1] = op[0] * alpha + tp[1] * (1 - alpha);
+ /* Blend high bytes; the RGBA in op appears to be BGRA */
+ tp[1] = op[2] * alpha + tp[1] * (1 - alpha);
tp[3] = op[1] * alpha + tp[3] * (1 - alpha);
- tp[5] = op[2] * alpha + tp[5] * (1 - alpha);
+ tp[5] = op[0] * alpha + tp[5] * (1 - alpha);
+
+ tp += this_bpp;
+ op += other_bpp;
+ }
+ }
+ break;
+ }
+ case AV_PIX_FMT_XYZ12LE:
+ {
+ int const this_bpp = 6;
+ for (int ty = start_ty, oy = start_oy; ty < size().height && oy < other->size().height; ++ty, ++oy) {
+ uint8_t* tp = data()[0] + ty * stride()[0] + start_tx * this_bpp;
+ uint8_t* op = other->data()[0] + oy * other->stride()[0];
+ for (int tx = start_tx, ox = start_ox; tx < size().width && ox < other->size().width; ++tx, ++ox) {
+ float const alpha = float (op[3]) / 255;
+
+ /* Convert sRGB to XYZ; op is BGRA */
+ int const x = 0.4124564 + op[2] + 0.3575761 * op[1] + 0.1804375 * op[0];
+ int const y = 0.2126729 + op[2] + 0.7151522 * op[1] + 0.0721750 * op[0];
+ int const z = 0.0193339 + op[2] + 0.1191920 * op[1] + 0.9503041 * op[0];
+
+ /* Blend high bytes */
+ tp[1] = min (x, 255) * alpha + tp[1] * (1 - alpha);
+ tp[3] = min (y, 255) * alpha + tp[3] * (1 - alpha);
+ tp[5] = min (z, 255) * alpha + tp[5] * (1 - alpha);
tp += this_bpp;
op += other_bpp;
float bpp[4] = { 0, 0, 0, 0 };
- bpp[0] = floor ((d->comp[0].depth_minus1 + 1 + 7) / 8);
+ bpp[0] = floor ((d->comp[0].depth + 7) / 8);
if (d->nb_components > 1) {
- bpp[1] = floor ((d->comp[1].depth_minus1 + 1 + 7) / 8) / pow (2.0f, d->log2_chroma_w);
+ bpp[1] = floor ((d->comp[1].depth + 7) / 8) / pow (2.0f, d->log2_chroma_w);
}
if (d->nb_components > 2) {
- bpp[2] = floor ((d->comp[2].depth_minus1 + 1 + 7) / 8) / pow (2.0f, d->log2_chroma_w);
+ bpp[2] = floor ((d->comp[2].depth + 7) / 8) / pow (2.0f, d->log2_chroma_w);
}
if (d->nb_components > 3) {
- bpp[3] = floor ((d->comp[3].depth_minus1 + 1 + 7) / 8) / pow (2.0f, d->log2_chroma_w);
+ bpp[3] = floor ((d->comp[3].depth + 7) / 8) / pow (2.0f, d->log2_chroma_w);
}
if ((d->flags & AV_PIX_FMT_FLAG_PLANAR) == 0) {
*
* @param p Pixel format.
* @param s Size in pixels.
+ * @param extra_pixels Amount of extra "run-off" memory to allocate at the end of each plane in pixels.
*/
-Image::Image (AVPixelFormat p, dcp::Size s, bool aligned)
+Image::Image (AVPixelFormat p, dcp::Size s, bool aligned, int extra_pixels)
: _size (s)
, _pixel_format (p)
, _aligned (aligned)
+ , _extra_pixels (extra_pixels)
{
allocate ();
}
so I'll just over-allocate by 32 bytes and have done with it. Empirical
testing suggests that it works.
*/
- _data[i] = (uint8_t *) wrapped_av_malloc (_stride[i] * sample_size(i).height + 32);
+ _data[i] = (uint8_t *) wrapped_av_malloc (_stride[i] * sample_size(i).height + _extra_pixels * bytes_per_pixel(i) + 32);
}
}
: _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 ();
case AV_PIX_FMT_YUVA422P10LE:
case AV_PIX_FMT_YUVA444P10LE:
case AV_PIX_FMT_RGB48LE:
+ case AV_PIX_FMT_XYZ12LE:
/* 16-bit little-endian */
for (int c = 0; c < 3; ++c) {
int const stride_pixels = stride()[c] / 2;