using boost::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.
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)..|
-
- 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
+ shared_ptr<Image> out (new Image(out_format, out_size, out_aligned));
+ out->make_black ();
- To get around this, we ask Image to overallocate its buffers by the overrun.
- */
+ AVPixFmtDescriptor const * in_desc = av_pix_fmt_desc_get (_pixel_format);
+ if (!in_desc) {
+ throw PixelFormatError ("crop_scale_window()", _pixel_format);
+ }
- shared_ptr<Image> out (new Image (out_format, out_size, out_aligned, (out_size.width - inter_size.width) / 2));
- out->make_black ();
+ /* Round down so that we crop only the number of pixels that is straightforward
+ * considering any subsampling.
+ */
+ Crop rounded_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)
+ );
/* Size of the image after any crop */
- dcp::Size const cropped_size = crop.apply (size ());
+ dcp::Size const cropped_size = rounded_crop.apply (size());
+
+ /* Hack: if we're not doing quite the crop that we were asked for, and we carry on scaling
+ * to the inter_size we were asked for, there is a small but noticeable wobble in the image
+ * luminance (#1872). This hack means we will jump in steps of the subsampling distance
+ * in both crop and scale.
+ */
+ inter_size.width = round_width_for_subsampling(inter_size.width, in_desc);
+ inter_size.height = round_width_for_subsampling(inter_size.height, in_desc);
/* Scale context for a scale from cropped_size to inter_size */
struct SwsContext* scale_context = sws_getContext (
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) * rounded_crop.left);
+ scale_in_data[c] = data()[c] + x + stride()[c] * (rounded_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 (rounded_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;
}
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 ()
{
void
Image::make_transparent ()
{
- if (_pixel_format != AV_PIX_FMT_BGRA) {
+ if (_pixel_format != AV_PIX_FMT_BGRA && _pixel_format != AV_PIX_FMT_RGBA) {
throw PixelFormatError ("make_transparent()", _pixel_format);
}
void
Image::alpha_blend (shared_ptr<const Image> other, Position<int> position)
{
- /* We're blending BGRA images; first byte is blue, second byte is green, third byte red, fourth byte alpha */
- DCPOMATIC_ASSERT (other->pixel_format() == AV_PIX_FMT_BGRA);
+ /* We're blending RGBA or BGRA images */
+ DCPOMATIC_ASSERT (other->pixel_format() == AV_PIX_FMT_BGRA || other->pixel_format() == AV_PIX_FMT_RGBA);
+ int const blue = other->pixel_format() == AV_PIX_FMT_BGRA ? 0 : 2;
+ int const red = other->pixel_format() == AV_PIX_FMT_BGRA ? 2 : 0;
+
int const other_bpp = 4;
int start_tx = position.x;
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;
- tp[0] = op[2] * alpha + tp[0] * (1 - alpha);
+ tp[0] = op[red] * alpha + tp[0] * (1 - alpha);
tp[1] = op[1] * alpha + tp[1] * (1 - alpha);
- tp[2] = op[0] * alpha + tp[2] * (1 - alpha);
+ tp[2] = op[blue] * alpha + tp[2] * (1 - alpha);
tp += this_bpp;
op += other_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;
- tp[0] = op[0] * alpha + tp[0] * (1 - alpha);
+ tp[0] = op[blue] * alpha + tp[0] * (1 - alpha);
tp[1] = op[1] * alpha + tp[1] * (1 - alpha);
- tp[2] = op[2] * alpha + tp[2] * (1 - alpha);
+ tp[2] = op[red] * alpha + tp[2] * (1 - alpha);
tp[3] = op[3] * alpha + tp[3] * (1 - alpha);
tp += 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;
- tp[0] = op[2] * alpha + tp[0] * (1 - alpha);
+ tp[0] = op[red] * alpha + tp[0] * (1 - alpha);
tp[1] = op[1] * alpha + tp[1] * (1 - alpha);
- tp[2] = op[0] * alpha + tp[2] * (1 - alpha);
+ tp[2] = op[blue] * alpha + tp[2] * (1 - alpha);
tp[3] = op[3] * alpha + tp[3] * (1 - alpha);
tp += this_bpp;
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[2] * alpha + tp[1] * (1 - alpha);
+ tp[1] = op[red] * alpha + tp[1] * (1 - alpha);
tp[3] = op[1] * alpha + tp[3] * (1 - alpha);
- tp[5] = op[0] * alpha + tp[5] * (1 - alpha);
+ tp[5] = op[blue] * alpha + tp[5] * (1 - alpha);
tp += this_bpp;
op += other_bpp;
float const alpha = float (op[3]) / 255;
/* Convert sRGB to XYZ; op is BGRA. First, input gamma LUT */
- double const r = lut_in[op[2]];
+ double const r = lut_in[op[red]];
double const g = lut_in[op[1]];
- double const b = lut_in[op[0]];
+ double const b = lut_in[op[blue]];
/* RGB to XYZ, including Bradford transform and DCI companding */
double const x = max (0.0, min (65535.0, r * fast_matrix[0] + g * fast_matrix[1] + b * fast_matrix[2]));
*
* @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
}
}
, _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 */
void
Image::fade (float f)
{
+ /* U/V black value for 8-bit colour */
+ static int const eight_bit_uv = (1 << 7) - 1;
+ /* U/V black value for 10-bit colour */
+ static uint16_t const ten_bit_uv = (1 << 9) - 1;
+
switch (_pixel_format) {
case AV_PIX_FMT_YUV420P:
- case AV_PIX_FMT_YUV422P:
- case AV_PIX_FMT_YUV444P:
- case AV_PIX_FMT_YUV411P:
- case AV_PIX_FMT_YUVJ420P:
- case AV_PIX_FMT_YUVJ422P:
- case AV_PIX_FMT_YUVJ444P:
- 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:
- /* 8-bit */
- for (int c = 0; c < 3; ++c) {
+ {
+ /* Y */
+ 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(float(*q) * f);
+ ++q;
+ }
+ p += stride()[0];
+ }
+
+ /* U, V */
+ for (int c = 1; c < 3; ++c) {
uint8_t* p = data()[c];
int const lines = sample_size(c).height;
for (int y = 0; y < lines; ++y) {
uint8_t* q = p;
for (int x = 0; x < line_size()[c]; ++x) {
- *q = int (float (*q) * f);
+ *q = eight_bit_uv + int((int(*q) - eight_bit_uv) * f);
++q;
}
p += stride()[c];
}
}
+
break;
+ }
+
+ case AV_PIX_FMT_RGB24:
+ {
+ /* 8-bit */
+ 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 (float (*q) * f);
+ ++q;
+ }
+ p += stride()[0];
+ }
+ break;
+ }
- case AV_PIX_FMT_YUV422P9LE:
- case AV_PIX_FMT_YUV444P9LE:
- case AV_PIX_FMT_YUV422P10LE:
- case AV_PIX_FMT_YUV444P10LE:
- case AV_PIX_FMT_YUV422P16LE:
- case AV_PIX_FMT_YUV444P16LE:
- case AV_PIX_FMT_YUVA420P9LE:
- case AV_PIX_FMT_YUVA422P9LE:
- case AV_PIX_FMT_YUVA444P9LE:
- case AV_PIX_FMT_YUVA420P10LE:
- case AV_PIX_FMT_YUVA422P10LE:
- case AV_PIX_FMT_YUVA444P10LE:
- case AV_PIX_FMT_RGB48LE:
case AV_PIX_FMT_XYZ12LE:
+ case AV_PIX_FMT_RGB48LE:
/* 16-bit little-endian */
for (int c = 0; c < 3; ++c) {
int const stride_pixels = stride()[c] / 2;
}
break;
- case AV_PIX_FMT_YUV422P9BE:
- case AV_PIX_FMT_YUV444P9BE:
- case AV_PIX_FMT_YUV444P10BE:
- case AV_PIX_FMT_YUV422P10BE:
- case AV_PIX_FMT_YUVA420P9BE:
- case AV_PIX_FMT_YUVA422P9BE:
- case AV_PIX_FMT_YUVA444P9BE:
- case AV_PIX_FMT_YUVA420P10BE:
- case AV_PIX_FMT_YUVA422P10BE:
- case AV_PIX_FMT_YUVA444P10BE:
- case AV_PIX_FMT_YUVA420P16BE:
- case AV_PIX_FMT_YUVA422P16BE:
- case AV_PIX_FMT_YUVA444P16BE:
- case AV_PIX_FMT_RGB48BE:
- /* 16-bit big-endian */
- for (int c = 0; c < 3; ++c) {
+ case AV_PIX_FMT_YUV422P10LE:
+ {
+ /* Y */
+ {
+ int const stride_pixels = stride()[0] / 2;
+ int const line_size_pixels = line_size()[0] / 2;
+ uint16_t* p = reinterpret_cast<uint16_t*> (data()[0]);
+ int const lines = sample_size(0).height;
+ for (int y = 0; y < lines; ++y) {
+ uint16_t* q = p;
+ for (int x = 0; x < line_size_pixels; ++x) {
+ *q = int(float(*q) * f);
+ ++q;
+ }
+ p += stride_pixels;
+ }
+ }
+
+ /* U, V */
+ for (int c = 1; c < 3; ++c) {
int const stride_pixels = stride()[c] / 2;
int const line_size_pixels = line_size()[c] / 2;
uint16_t* p = reinterpret_cast<uint16_t*> (data()[c]);
for (int y = 0; y < lines; ++y) {
uint16_t* q = p;
for (int x = 0; x < line_size_pixels; ++x) {
- *q = swap_16 (int (float (swap_16 (*q)) * f));
+ *q = ten_bit_uv + int((int(*q) - ten_bit_uv) * f);
++q;
}
p += stride_pixels;
}
break;
- case AV_PIX_FMT_UYVY422:
- {
- int const Y = sample_size(0).height;
- int const X = line_size()[0];
- uint8_t* p = data()[0];
- for (int y = 0; y < Y; ++y) {
- for (int x = 0; x < X; ++x) {
- *p = int (float (*p) * f);
- ++p;
- }
- }
- break;
}
default:
{
DCPOMATIC_ASSERT (bytes_per_pixel(0) == 4);
DCPOMATIC_ASSERT (planes() == 1);
- DCPOMATIC_ASSERT (pixel_format() == AV_PIX_FMT_BGRA);
+ if (pixel_format() != AV_PIX_FMT_RGBA) {
+ return convert_pixel_format(dcp::YUV_TO_RGB_REC709, AV_PIX_FMT_RGBA, true, false)->as_png();
+ }
/* error handling? */
png_structp png_ptr = png_create_write_struct(PNG_LIBPNG_VER_STRING, reinterpret_cast<void*>(const_cast<Image*>(this)), png_error_fn, 0);
projects / dcpomatic.git / blobdiff