X-Git-Url: https://git.carlh.net/gitweb/?a=blobdiff_plain;f=src%2Flib%2Fimage.cc;h=228685442dd5016a3a614d63f010cf18e6171947;hb=0f664dab81a835feebd5db92942413f91747481f;hp=48a4e16cd9e8f797cf5d0409359ec569d1d05210;hpb=aa7edc3114fa3b7868088a9f8f22b5ee195a8668;p=dcpomatic.git diff --git a/src/lib/image.cc b/src/lib/image.cc index 48a4e16cd..228685442 100644 --- a/src/lib/image.cc +++ b/src/lib/image.cc @@ -1,19 +1,20 @@ /* - Copyright (C) 2012-2015 Carl Hetherington + Copyright (C) 2012-2016 Carl Hetherington - This program is free software; you can redistribute it and/or modify + This file is part of DCP-o-matic. + + DCP-o-matic is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. - This program is distributed in the hope that it will be useful, + DCP-o-matic is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License - along with this program; if not, write to the Free Software - Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + along with DCP-o-matic. If not, see . */ @@ -26,8 +27,9 @@ #include "timer.h" #include "rect.h" #include "util.h" -#include "md5_digester.h" #include "dcpomatic_socket.h" +#include +#include extern "C" { #include #include @@ -40,6 +42,7 @@ extern "C" { using std::string; using std::min; +using std::max; using std::cout; using std::cerr; using std::list; @@ -48,7 +51,7 @@ using boost::shared_ptr; using dcp::Size; int -Image::line_factor (int n) const +Image::vertical_factor (int n) const { if (n == 0) { return 1; @@ -62,11 +65,8 @@ Image::line_factor (int n) const return pow (2.0f, d->log2_chroma_h); } -/** @param n Component index. - * @return Number of samples (i.e. pixels, unless sub-sampled) in each direction for this component. - */ -dcp::Size -Image::sample_size (int n) const +int +Image::horizontal_factor (int n) const { int horizontal_factor = 1; if (n > 0) { @@ -76,22 +76,19 @@ Image::sample_size (int n) const } horizontal_factor = pow (2.0f, d->log2_chroma_w); } - - return dcp::Size ( - lrint (ceil (static_cast(size().width) / horizontal_factor)), - lrint (ceil (static_cast(size().height) / line_factor (n))) - ); + return horizontal_factor; } -int -Image::components () const +/** @param n Component index. + * @return Number of samples (i.e. pixels, unless sub-sampled) in each direction for this component. + */ +dcp::Size +Image::sample_size (int n) const { - AVPixFmtDescriptor const * d = av_pix_fmt_desc_get(_pixel_format); - if (!d) { - throw PixelFormatError ("components()", _pixel_format); - } - - return d->nb_components; + return dcp::Size ( + lrint (ceil (static_cast(size().width) / horizontal_factor (n))), + lrint (ceil (static_cast(size().height) / vertical_factor (n))) + ); } /** @return Number of planes */ @@ -110,10 +107,19 @@ Image::planes () const return d->nb_components; } -/** Crop this image, scale it to `inter_size' and then place it in a black frame of `out_size' */ +/** 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 out_format Output pixel format. + * @param out_aligned true to make the output image aligned. + * @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::crop_scale_window ( - Crop crop, dcp::Size inter_size, dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_format, bool out_aligned + Crop crop, dcp::Size inter_size, dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_format, bool out_aligned, bool fast ) const { /* Empirical testing suggests that sws_scale() will crash if @@ -124,8 +130,29 @@ Image::crop_scale_window ( 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 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 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 */ @@ -135,7 +162,7 @@ Image::crop_scale_window ( struct SwsContext* scale_context = sws_getContext ( cropped_size.width, cropped_size.height, pixel_format(), inter_size.width, inter_size.height, out_format, - SWS_BICUBIC, 0, 0, 0 + fast ? SWS_FAST_BILINEAR : SWS_BICUBIC, 0, 0, 0 ); if (!scale_context) { @@ -169,7 +196,7 @@ Image::crop_scale_window ( we've cropped all of its Y-channel pixels. */ int const x = lrintf (bytes_per_pixel(c) * crop.left) & ~ ((int) desc->log2_chroma_w); - scale_in_data[c] = data()[c] + x + stride()[c] * (crop.top / line_factor(c)); + scale_in_data[c] = data()[c] + x + stride()[c] * (crop.top / vertical_factor(c)); } /* Corner of the image within out_size */ @@ -192,8 +219,15 @@ Image::crop_scale_window ( return out; } +/** @param out_size Size to scale to. + * @param yuv_to_rgb YUVToRGB transform transform to use, if required. + * @param out_format Output pixel format. + * @param out_aligned true to make an aligned 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::scale (dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_format, bool out_aligned) const +Image::scale (dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_format, bool out_aligned, bool fast) const { /* Empirical testing suggests that sws_scale() will crash if the input image is not aligned. @@ -205,7 +239,7 @@ Image::scale (dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_fo struct SwsContext* scale_context = sws_getContext ( size().width, size().height, pixel_format(), out_size.width, out_size.height, out_format, - SWS_BICUBIC, 0, 0, 0 + fast ? SWS_FAST_BILINEAR : SWS_BICUBIC, 0, 0, 0 ); DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUV_TO_RGB_COUNT); @@ -395,9 +429,40 @@ Image::make_transparent () memset (data()[0], 0, sample_size(0).height * stride()[0]); } +template +void +component ( + int n, + Image* base, + shared_ptr other, + shared_ptr rgba, + int start_base_x, int start_base_y, + int start_other_x, int start_other_y + ) +{ + dcp::Size const base_size = base->sample_size(n); + dcp::Size const other_size = other->sample_size(n); + for (int by = start_base_y, oy = start_other_y; by < base_size.height && oy < other_size.height; ++by, ++oy) { + /* base image */ + T* bp = ((T*) (base->data()[n] + by * base->stride()[n])) + start_base_x; + /* overlay image */ + T* op = ((T*) (other->data()[n] + oy * other->stride()[n])); + /* original RGBA for alpha channel */ + uint8_t* rp = rgba->data()[0] + oy * rgba->stride()[0]; + for (int bx = start_base_x, ox = start_other_x; bx < base_size.width && ox < other_size.width; ++bx, ++ox) { + float const alpha = float (rp[3]) / 255; + *bp = *op * alpha + *bp * (1 - alpha); + ++bp; + ++op; + rp += 4; + } + } +} + void Image::alpha_blend (shared_ptr other, Position position) { + /* We're blending RGBA images; first byte is blue, second byte is green, third byte blue, fourth byte alpha */ DCPOMATIC_ASSERT (other->pixel_format() == AV_PIX_FMT_RGBA); int const other_bpp = 4; @@ -420,15 +485,16 @@ Image::alpha_blend (shared_ptr other, Position position) switch (_pixel_format) { case AV_PIX_FMT_RGB24: { + /* Going onto RGB24. First byte is red, second green, third blue */ int const this_bpp = 3; 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; - tp[0] = op[0] * alpha + tp[0] * (1 - alpha); + tp[0] = op[2] * 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[0] * alpha + tp[2] * (1 - alpha); tp += this_bpp; op += other_bpp; @@ -464,10 +530,10 @@ Image::alpha_blend (shared_ptr other, Position position) 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; @@ -475,8 +541,83 @@ Image::alpha_blend (shared_ptr other, Position position) } break; } + case AV_PIX_FMT_XYZ12LE: + { + dcp::ColourConversion conv = dcp::ColourConversion::srgb_to_xyz(); + double fast_matrix[9]; + dcp::combined_rgb_to_xyz (conv, fast_matrix); + double const * lut_in = conv.in()->lut (8, false); + double const * lut_out = conv.out()->lut (16, true); + int const this_bpp = 6; + for (int ty = start_ty, oy = start_oy; ty < size().height && oy < other->size().height; ++ty, ++oy) { + uint16_t* tp = reinterpret_cast (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. First, input gamma LUT */ + double const r = lut_in[op[2]]; + double const g = lut_in[op[1]]; + double const b = lut_in[op[0]]; + + /* 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])); + double const y = max (0.0, min (65535.0, r * fast_matrix[3] + g * fast_matrix[4] + b * fast_matrix[5])); + double const z = max (0.0, min (65535.0, r * fast_matrix[6] + g * fast_matrix[7] + b * fast_matrix[8])); + + /* Out gamma LUT and blend */ + tp[0] = lrint(lut_out[lrint(x)] * 65535) * alpha + tp[0] * (1 - alpha); + tp[1] = lrint(lut_out[lrint(y)] * 65535) * alpha + tp[1] * (1 - alpha); + tp[2] = lrint(lut_out[lrint(z)] * 65535) * alpha + tp[2] * (1 - alpha); + + tp += this_bpp / 2; + op += other_bpp; + } + } + break; + } + case AV_PIX_FMT_YUV420P: + { + shared_ptr yuv = other->scale (other->size(), dcp::YUV_TO_RGB_REC709, _pixel_format, false, false); + + for (int i = 0; i < 3; ++i) { + dcp::Size const tsize = sample_size(i); + dcp::Size const osize = yuv->sample_size(i); + int const tbpp = ceil (bytes_per_pixel(i) / horizontal_factor(i)); + int const obpp = ceil (yuv->bytes_per_pixel(i) / yuv->horizontal_factor(i)); + int const abpp = other->bytes_per_pixel(0); + start_tx /= horizontal_factor (i); + start_ty /= vertical_factor (i); + start_ox /= yuv->horizontal_factor (i); + start_oy /= yuv->vertical_factor (i); + for (int ty = start_ty, oy = start_oy; ty < tsize.height && oy < osize.height; ++ty, ++oy) { + /* this image */ + uint8_t* tp = data()[i] + ty * stride()[i] + start_tx * tbpp; + /* overlay image */ + uint8_t* op = yuv->data()[i] + oy * yuv->stride()[i]; + /* original RGBA for alpha channel */ + uint8_t* ap = other->data()[0] + oy * other->stride()[0]; + for (int tx = start_tx, ox = start_ox; tx < tsize.width && ox < osize.width; ++tx, ++ox) { + float const alpha = float (ap[3]) / 255; + *tp = *op * alpha + *tp * (1 - alpha); + tp += tbpp; + op += obpp; + ap += abpp; + } + } + } + break; + } + case AV_PIX_FMT_YUV420P10: + { + shared_ptr yuv = other->scale (other->size(), dcp::YUV_TO_RGB_REC709, _pixel_format, false, false); + component (0, this, yuv, other, start_tx, start_ty, start_ox, start_oy); + component (1, this, yuv, other, start_tx, start_ty, start_ox, start_oy); + component (2, this, yuv, other, start_tx, start_ty, start_ox, start_oy); + break; + } default: - DCPOMATIC_ASSERT (false); + throw PixelFormatError ("alpha_blend()", _pixel_format); } } @@ -535,16 +676,29 @@ Image::bytes_per_pixel (int c) const float bpp[4] = { 0, 0, 0, 0 }; - bpp[0] = floor ((d->comp[0].depth_minus1 + 1 + 7) / 8); +#ifdef DCPOMATIC_HAVE_AVCOMPONENTDESCRIPTOR_DEPTH_MINUS1 + bpp[0] = floor ((d->comp[0].depth_minus1 + 8) / 8); + if (d->nb_components > 1) { + bpp[1] = floor ((d->comp[1].depth_minus1 + 8) / 8) / pow (2.0f, d->log2_chroma_w); + } + if (d->nb_components > 2) { + bpp[2] = floor ((d->comp[2].depth_minus1 + 8) / 8) / pow (2.0f, d->log2_chroma_w); + } + if (d->nb_components > 3) { + bpp[3] = floor ((d->comp[3].depth_minus1 + 8) / 8) / pow (2.0f, d->log2_chroma_w); + } +#else + 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); } +#endif if ((d->flags & AV_PIX_FMT_FLAG_PLANAR) == 0) { /* Not planar; sum them up */ @@ -559,11 +713,14 @@ Image::bytes_per_pixel (int c) const * * @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. */ -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 (); } @@ -598,7 +755,7 @@ Image::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); } } @@ -606,6 +763,7 @@ Image::Image (Image const & other) : _size (other._size) , _pixel_format (other._pixel_format) , _aligned (other._aligned) + , _extra_pixels (other._extra_pixels) { allocate (); @@ -625,6 +783,7 @@ Image::Image (AVFrame* frame) : _size (frame->width, frame->height) , _pixel_format (static_cast (frame->format)) , _aligned (true) + , _extra_pixels (0) { allocate (); @@ -645,6 +804,7 @@ Image::Image (shared_ptr other, bool aligned) : _size (other->_size) , _pixel_format (other->_pixel_format) , _aligned (aligned) + , _extra_pixels (other->_extra_pixels) { allocate (); @@ -686,6 +846,7 @@ Image::swap (Image & other) } std::swap (_aligned, other._aligned); + std::swap (_extra_pixels, other._extra_pixels); } /** Destroy a Image */