Remove in-place translations support.
[dcpomatic.git] / src / lib / image.cc
index 4f5b11c90ba63c93dd7e8f842a6c1992f41a000b..2167918f8d1055c1c7fff274021797fb767b84bc 100644 (file)
@@ -27,6 +27,7 @@
 #include "compose.hpp"
 #include "dcpomatic_assert.h"
 #include "dcpomatic_socket.h"
+#include "enum_indexed_vector.h"
 #include "exceptions.h"
 #include "image.h"
 #include "maths_util.h"
@@ -44,7 +45,6 @@ extern "C" {
 #include <libswscale/swscale.h>
 }
 LIBDCP_ENABLE_WARNINGS
-#include <png.h>
 #if HAVE_VALGRIND_MEMCHECK_H
 #include <valgrind/memcheck.h>
 #endif
@@ -237,10 +237,10 @@ Image::crop_scale_window (
        }
 
        DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUVToRGB::COUNT);
-       int const lut[static_cast<int>(dcp::YUVToRGB::COUNT)] = {
-               SWS_CS_ITU601,
-               SWS_CS_ITU709
-       };
+       EnumIndexedVector<int, dcp::YUVToRGB> lut;
+       lut[dcp::YUVToRGB::REC601] = SWS_CS_ITU601;
+       lut[dcp::YUVToRGB::REC709] = SWS_CS_ITU709;
+       lut[dcp::YUVToRGB::REC2020] = SWS_CS_BT2020;
 
        /* The 3rd parameter here is:
           0 -> source range MPEG (i.e. "video", 16-235)
@@ -255,8 +255,8 @@ Image::crop_scale_window (
        */
        sws_setColorspaceDetails (
                scale_context,
-               sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), video_range == VideoRange::VIDEO ? 0 : 1,
-               sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), out_video_range == VideoRange::VIDEO ? 0 : 1,
+               sws_getCoefficients(lut[yuv_to_rgb]), video_range == VideoRange::VIDEO ? 0 : 1,
+               sws_getCoefficients(lut[yuv_to_rgb]), out_video_range == VideoRange::VIDEO ? 0 : 1,
                0, 1 << 16, 1 << 16
                );
 
@@ -293,13 +293,17 @@ Image::crop_scale_window (
 
        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);
-       }
+       /* There are some cases where there will be unwanted image data left in the image at this point:
+        *
+        * 1. When we are cropping without any scaling or pixel format conversion.
+        * 2. When we are scaling to certain sizes and placing the result into a larger
+        *    black frame.
+        *
+        * Clear out the sides of the image to take care of those cases.
+        */
+       auto const pad = (out_size.width - inter_size.width) / 2;
+       out->make_part_black(0, pad);
+       out->make_part_black(corner.x + inter_size.width, pad);
 
        if (
                video_range == VideoRange::VIDEO &&
@@ -324,7 +328,7 @@ Image::convert_pixel_format (dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_format,
 /** @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_aligment Output alignment.
+ *  @param out_alignment Output alignment.
  *  @param fast Try to be fast at the possible expense of quality; at present this means using
  *  fast bilinear rather than bicubic scaling.
  */
@@ -335,6 +339,10 @@ Image::scale (dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_fo
           the input image alignment is not PADDED.
        */
        DCPOMATIC_ASSERT (alignment() == Alignment::PADDED);
+       DCPOMATIC_ASSERT(size().width > 0);
+       DCPOMATIC_ASSERT(size().height > 0);
+       DCPOMATIC_ASSERT(out_size.width > 0);
+       DCPOMATIC_ASSERT(out_size.height > 0);
 
        auto scaled = make_shared<Image>(out_format, out_size, out_alignment);
        auto scale_context = sws_getContext (
@@ -343,11 +351,13 @@ Image::scale (dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_fo
                (fast ? SWS_FAST_BILINEAR : SWS_BICUBIC) | SWS_ACCURATE_RND, 0, 0, 0
                );
 
+       DCPOMATIC_ASSERT(scale_context);
+
        DCPOMATIC_ASSERT (yuv_to_rgb < dcp::YUVToRGB::COUNT);
-       int const lut[static_cast<int>(dcp::YUVToRGB::COUNT)] = {
-               SWS_CS_ITU601,
-               SWS_CS_ITU709
-       };
+       EnumIndexedVector<int, dcp::YUVToRGB> lut;
+       lut[dcp::YUVToRGB::REC601] = SWS_CS_ITU601;
+       lut[dcp::YUVToRGB::REC709] = SWS_CS_ITU709;
+       lut[dcp::YUVToRGB::REC2020] = SWS_CS_BT2020;
 
        /* The 3rd parameter here is:
           0 -> source range MPEG (i.e. "video", 16-235)
@@ -362,8 +372,8 @@ Image::scale (dcp::Size out_size, dcp::YUVToRGB yuv_to_rgb, AVPixelFormat out_fo
        */
        sws_setColorspaceDetails (
                scale_context,
-               sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), 0,
-               sws_getCoefficients (lut[static_cast<int>(yuv_to_rgb)]), 0,
+               sws_getCoefficients(lut[yuv_to_rgb]), 0,
+               sws_getCoefficients(lut[yuv_to_rgb]), 0,
                0, 1 << 16, 1 << 16
                );
 
@@ -475,6 +485,21 @@ Image::make_part_black (int const start, int const width)
                }
                break;
        }
+       case AV_PIX_FMT_YUV444P10LE:
+       {
+               y_part();
+               for (int i = 1; i < 3; ++i) {
+                       auto p = reinterpret_cast<int16_t*>(data()[i]);
+                       int const h = sample_size(i).height;
+                       for (int y = 0; y < h; ++y) {
+                               for (int x = start; x < (start + width); ++x) {
+                                       p[x] = ten_bit_uv;
+                               }
+                               p += stride()[i] / 2;
+                       }
+               }
+               break;
+       }
        default:
                throw PixelFormatError ("make_part_black()", _pixel_format);
        }
@@ -600,7 +625,7 @@ Image::make_black ()
 void
 Image::make_transparent ()
 {
-       if (_pixel_format != AV_PIX_FMT_BGRA && _pixel_format != AV_PIX_FMT_RGBA) {
+       if (_pixel_format != AV_PIX_FMT_BGRA && _pixel_format != AV_PIX_FMT_RGBA && _pixel_format != AV_PIX_FMT_RGBA64BE) {
                throw PixelFormatError ("make_transparent()", _pixel_format);
        }
 
@@ -608,16 +633,333 @@ Image::make_transparent ()
 }
 
 
+struct TargetParams
+{
+       int start_x;
+       int start_y;
+       dcp::Size size;
+       uint8_t* const* data;
+       int const* stride;
+       int bpp;
+
+       uint8_t* line_pointer(int y) const {
+               return data[0] + y * stride[0] + start_x * bpp;
+       }
+};
+
+
+/** Parameters of the other image (the one being blended onto the target) when target and other are RGB */
+struct OtherRGBParams
+{
+       int start_x;
+       int start_y;
+       dcp::Size size;
+       uint8_t* const* data;
+       int const* stride;
+       int bpp;
+
+       uint8_t* line_pointer(int y) const {
+               return data[0] + y * stride[0];
+       }
+
+       float alpha_divisor() const {
+               return pow(2, bpp * 2) - 1;
+       }
+};
+
+
+/** Parameters of the other image (the one being blended onto the target) when target and other are YUV */
+struct OtherYUVParams
+{
+       int start_x;
+       int start_y;
+       dcp::Size size;
+       uint8_t* const* data;
+       int const* stride;
+
+       uint8_t* const* alpha_data;
+       int const* alpha_stride;
+       int alpha_bpp;
+};
+
+
+template <class OtherType>
+void
+alpha_blend_onto_rgb24(TargetParams const& target, OtherRGBParams const& other, int red, int blue, std::function<float (OtherType*)> get, int value_divisor)
+{
+       /* Going onto RGB24.  First byte is red, second green, third blue */
+       auto const alpha_divisor = other.alpha_divisor();
+       for (int ty = target.start_y, oy = other.start_y; ty < target.size.height && oy < other.size.height; ++ty, ++oy) {
+               auto tp = target.line_pointer(ty);
+               auto op = reinterpret_cast<OtherType*>(other.line_pointer(oy));
+               for (int tx = target.start_x, ox = other.start_x; tx < target.size.width && ox < other.size.width; ++tx, ++ox) {
+                       float const alpha = get(op + 3) / alpha_divisor;
+                       tp[0] = (get(op + red) / value_divisor) * alpha + tp[0] * (1 - alpha);
+                       tp[1] = (get(op + 1) / value_divisor) * alpha + tp[1] * (1 - alpha);
+                       tp[2] = (get(op + blue) / value_divisor) * alpha + tp[2] * (1 - alpha);
+
+                       tp += target.bpp;
+                       op += other.bpp / sizeof(OtherType);
+               }
+       }
+}
+
+
+template <class OtherType>
+void
+alpha_blend_onto_bgra(TargetParams const& target, OtherRGBParams const& other, int red, int blue, std::function<float (OtherType*)> get, int value_divisor)
+{
+       auto const alpha_divisor = other.alpha_divisor();
+       for (int ty = target.start_y, oy = other.start_y; ty < target.size.height && oy < other.size.height; ++ty, ++oy) {
+               auto tp = target.line_pointer(ty);
+               auto op = reinterpret_cast<OtherType*>(other.line_pointer(oy));
+               for (int tx = target.start_x, ox = other.start_x; tx < target.size.width && ox < other.size.width; ++tx, ++ox) {
+                       float const alpha = get(op + 3) / alpha_divisor;
+                       tp[0] = (get(op + blue) / value_divisor) * alpha + tp[0] * (1 - alpha);
+                       tp[1] = (get(op + 1) / value_divisor) * alpha + tp[1] * (1 - alpha);
+                       tp[2] = (get(op + red) / value_divisor) * alpha + tp[2] * (1 - alpha);
+                       tp[3] = (get(op + 3) / value_divisor) * alpha + tp[3] * (1 - alpha);
+
+                       tp += target.bpp;
+                       op += other.bpp / sizeof(OtherType);
+               }
+       }
+}
+
+
+template <class OtherType>
+void
+alpha_blend_onto_rgba(TargetParams const& target, OtherRGBParams const& other, int red, int blue, std::function<float (OtherType*)> get, int value_divisor)
+{
+       auto const alpha_divisor = other.alpha_divisor();
+       for (int ty = target.start_y, oy = other.start_y; ty < target.size.height && oy < other.size.height; ++ty, ++oy) {
+               auto tp = target.line_pointer(ty);
+               auto op = reinterpret_cast<OtherType*>(other.line_pointer(oy));
+               for (int tx = target.start_x, ox = other.start_x; tx < target.size.width && ox < other.size.width; ++tx, ++ox) {
+                       float const alpha = get(op + 3) / alpha_divisor;
+                       tp[0] = (get(op + red) / value_divisor) * alpha + tp[0] * (1 - alpha);
+                       tp[1] = (get(op + 1) / value_divisor) * alpha + tp[1] * (1 - alpha);
+                       tp[2] = (get(op + blue) / value_divisor) * alpha + tp[2] * (1 - alpha);
+                       tp[3] = (get(op + 3) / value_divisor) * alpha + tp[3] * (1 - alpha);
+
+                       tp += target.bpp;
+                       op += other.bpp / sizeof(OtherType);
+               }
+       }
+}
+
+
+template <class OtherType>
+void
+alpha_blend_onto_rgb48le(TargetParams const& target, OtherRGBParams const& other, int red, int blue, std::function<float (OtherType*)> get, int value_scale)
+{
+       auto const alpha_divisor = other.alpha_divisor();
+       for (int ty = target.start_y, oy = other.start_y; ty < target.size.height && oy < other.size.height; ++ty, ++oy) {
+               auto tp = reinterpret_cast<uint16_t*>(target.line_pointer(ty));
+               auto op = reinterpret_cast<OtherType*>(other.line_pointer(oy));
+               for (int tx = target.start_x, ox = other.start_x; tx < target.size.width && ox < other.size.width; ++tx, ++ox) {
+                       float const alpha = get(op + 3) / alpha_divisor;
+                       tp[0] = get(op + red) * value_scale * alpha + tp[0] * (1 - alpha);
+                       tp[1] = get(op + 1) * value_scale * alpha + tp[1] * (1 - alpha);
+                       tp[2] = get(op + blue) * value_scale * alpha + tp[2] * (1 - alpha);
+
+                       tp += target.bpp / 2;
+                       op += other.bpp / sizeof(OtherType);
+               }
+       }
+}
+
+
+template <class OtherType>
+void
+alpha_blend_onto_xyz12le(TargetParams const& target, OtherRGBParams const& other, int red, int blue, std::function<float (OtherType*)> get, int value_divisor)
+{
+       auto const alpha_divisor = other.alpha_divisor();
+       auto conv = dcp::ColourConversion::srgb_to_xyz();
+       double fast_matrix[9];
+       dcp::combined_rgb_to_xyz(conv, fast_matrix);
+       auto lut_in = conv.in()->double_lut(0, 1, 8, false);
+       auto lut_out = conv.out()->int_lut(0, 1, 16, true, 65535);
+       for (int ty = target.start_y, oy = other.start_y; ty < target.size.height && oy < other.size.height; ++ty, ++oy) {
+               auto tp = reinterpret_cast<uint16_t*>(target.data[0] + ty * target.stride[0] + target.start_x * target.bpp);
+               auto op = reinterpret_cast<OtherType*>(other.data[0] + oy * other.stride[0]);
+               for (int tx = target.start_x, ox = other.start_x; tx < target.size.width && ox < other.size.width; ++tx, ++ox) {
+                       float const alpha = get(op + 3) / alpha_divisor;
+
+                       /* Convert sRGB to XYZ; op is BGRA.  First, input gamma LUT */
+                       double const r = lut_in[get(op + red) / value_divisor];
+                       double const g = lut_in[get(op + 1) / value_divisor];
+                       double const b = lut_in[get(op + blue) / value_divisor];
+
+                       /* RGB to XYZ, including Bradford transform and DCI companding */
+                       double const x = max(0.0, min(1.0, r * fast_matrix[0] + g * fast_matrix[1] + b * fast_matrix[2]));
+                       double const y = max(0.0, min(1.0, r * fast_matrix[3] + g * fast_matrix[4] + b * fast_matrix[5]));
+                       double const z = max(0.0, min(1.0, r * fast_matrix[6] + g * fast_matrix[7] + b * fast_matrix[8]));
+
+                       /* Out gamma LUT and blend */
+                       tp[0] = lut_out[lrint(x * 65535)] * alpha + tp[0] * (1 - alpha);
+                       tp[1] = lut_out[lrint(y * 65535)] * alpha + tp[1] * (1 - alpha);
+                       tp[2] = lut_out[lrint(z * 65535)] * alpha + tp[2] * (1 - alpha);
+
+                       tp += target.bpp / 2;
+                       op += other.bpp / sizeof(OtherType);
+               }
+       }
+}
+
+
+static
+void
+alpha_blend_onto_yuv420p(TargetParams const& target, OtherYUVParams const& other, std::function<float (uint8_t* data)> get_alpha)
+{
+       auto const ts = target.size;
+       auto const os = other.size;
+       for (int ty = target.start_y, oy = other.start_y; ty < ts.height && oy < os.height; ++ty, ++oy) {
+               int const hty = ty / 2;
+               int const hoy = oy / 2;
+               uint8_t* tY = target.data[0] + (ty * target.stride[0]) + target.start_x;
+               uint8_t* tU = target.data[1] + (hty * target.stride[1]) + target.start_x / 2;
+               uint8_t* tV = target.data[2] + (hty * target.stride[2]) + target.start_x / 2;
+               uint8_t* oY = other.data[0] + (oy * other.stride[0]) + other.start_x;
+               uint8_t* oU = other.data[1] + (hoy * other.stride[1]) + other.start_x / 2;
+               uint8_t* oV = other.data[2] + (hoy * other.stride[2]) + other.start_x / 2;
+               uint8_t* alpha = other.alpha_data[0] + (oy * other.alpha_stride[0]) + other.start_x * other.alpha_bpp;
+               for (int tx = target.start_x, ox = other.start_x; tx < ts.width && ox < os.width; ++tx, ++ox) {
+                       float const a = get_alpha(alpha);
+                       *tY = *oY * a + *tY * (1 - a);
+                       *tU = *oU * a + *tU * (1 - a);
+                       *tV = *oV * a + *tV * (1 - a);
+                       ++tY;
+                       ++oY;
+                       if (tx % 2) {
+                               ++tU;
+                               ++tV;
+                       }
+                       if (ox % 2) {
+                               ++oU;
+                               ++oV;
+                       }
+                       alpha += other.alpha_bpp;
+               }
+       }
+}
+
+
+static
+void
+alpha_blend_onto_yuv420p10(TargetParams const& target, OtherYUVParams const& other, std::function<float (uint8_t* data)> get_alpha)
+{
+       auto const ts = target.size;
+       auto const os = other.size;
+       for (int ty = target.start_y, oy = other.start_y; ty < ts.height && oy < os.height; ++ty, ++oy) {
+               int const hty = ty / 2;
+               int const hoy = oy / 2;
+               uint16_t* tY = reinterpret_cast<uint16_t*>(target.data[0] + (ty * target.stride[0])) + target.start_x;
+               uint16_t* tU = reinterpret_cast<uint16_t*>(target.data[1] + (hty * target.stride[1])) + target.start_x / 2;
+               uint16_t* tV = reinterpret_cast<uint16_t*>(target.data[2] + (hty * target.stride[2])) + target.start_x / 2;
+               uint16_t* oY = reinterpret_cast<uint16_t*>(other.data[0] + (oy * other.stride[0])) + other.start_x;
+               uint16_t* oU = reinterpret_cast<uint16_t*>(other.data[1] + (hoy * other.stride[1])) + other.start_x / 2;
+               uint16_t* oV = reinterpret_cast<uint16_t*>(other.data[2] + (hoy * other.stride[2])) + other.start_x / 2;
+               uint8_t* alpha = other.alpha_data[0] + (oy * other.alpha_stride[0]) + other.start_x * other.alpha_bpp;
+               for (int tx = target.start_x, ox = other.start_x; tx < ts.width && ox < os.width; ++tx, ++ox) {
+                       float const a = get_alpha(alpha);
+                       *tY = *oY * a + *tY * (1 - a);
+                       *tU = *oU * a + *tU * (1 - a);
+                       *tV = *oV * a + *tV * (1 - a);
+                       ++tY;
+                       ++oY;
+                       if (tx % 2) {
+                               ++tU;
+                               ++tV;
+                       }
+                       if (ox % 2) {
+                               ++oU;
+                               ++oV;
+                       }
+                       alpha += other.alpha_bpp;
+               }
+       }
+}
+
+
+static
+void
+alpha_blend_onto_yuv422p9or10le(TargetParams const& target, OtherYUVParams const& other, std::function<float (uint8_t* data)> get_alpha)
+{
+       auto const ts = target.size;
+       auto const os = other.size;
+       for (int ty = target.start_y, oy = other.start_y; ty < ts.height && oy < os.height; ++ty, ++oy) {
+               uint16_t* tY = reinterpret_cast<uint16_t*>(target.data[0] + (ty * target.stride[0])) + target.start_x;
+               uint16_t* tU = reinterpret_cast<uint16_t*>(target.data[1] + (ty * target.stride[1])) + target.start_x / 2;
+               uint16_t* tV = reinterpret_cast<uint16_t*>(target.data[2] + (ty * target.stride[2])) + target.start_x / 2;
+               uint16_t* oY = reinterpret_cast<uint16_t*>(other.data[0] + (oy * other.stride[0])) + other.start_x;
+               uint16_t* oU = reinterpret_cast<uint16_t*>(other.data[1] + (oy * other.stride[1])) + other.start_x / 2;
+               uint16_t* oV = reinterpret_cast<uint16_t*>(other.data[2] + (oy * other.stride[2])) + other.start_x / 2;
+               uint8_t* alpha = other.alpha_data[0] + (oy * other.alpha_stride[0]) + other.start_x * other.alpha_bpp;
+               for (int tx = target.start_x, ox = other.start_x; tx < ts.width && ox < os.width; ++tx, ++ox) {
+                       float const a = get_alpha(alpha);
+                       *tY = *oY * a + *tY * (1 - a);
+                       *tU = *oU * a + *tU * (1 - a);
+                       *tV = *oV * a + *tV * (1 - a);
+                       ++tY;
+                       ++oY;
+                       if (tx % 2) {
+                               ++tU;
+                               ++tV;
+                       }
+                       if (ox % 2) {
+                               ++oU;
+                               ++oV;
+                       }
+                       alpha += other.alpha_bpp;
+               }
+       }
+}
+
+
+static
+void
+alpha_blend_onto_yuv444p9or10le(TargetParams const& target, OtherYUVParams const& other, std::function<float (uint8_t* data)> get_alpha)
+{
+       auto const ts = target.size;
+       auto const os = other.size;
+       for (int ty = target.start_y, oy = other.start_y; ty < ts.height && oy < os.height; ++ty, ++oy) {
+               uint16_t* tY = reinterpret_cast<uint16_t*>(target.data[0] + (ty * target.stride[0])) + target.start_x;
+               uint16_t* tU = reinterpret_cast<uint16_t*>(target.data[1] + (ty * target.stride[1])) + target.start_x;
+               uint16_t* tV = reinterpret_cast<uint16_t*>(target.data[2] + (ty * target.stride[2])) + target.start_x;
+               uint16_t* oY = reinterpret_cast<uint16_t*>(other.data[0] + (oy * other.stride[0])) + other.start_x;
+               uint16_t* oU = reinterpret_cast<uint16_t*>(other.data[1] + (oy * other.stride[1])) + other.start_x;
+               uint16_t* oV = reinterpret_cast<uint16_t*>(other.data[2] + (oy * other.stride[2])) + other.start_x;
+               uint8_t* alpha = other.alpha_data[0] + (oy * other.alpha_stride[0]) + other.start_x * other.alpha_bpp;
+               for (int tx = target.start_x, ox = other.start_x; tx < ts.width && ox < os.width; ++tx, ++ox) {
+                       float const a = get_alpha(alpha);
+                       *tY = *oY * a + *tY * (1 - a);
+                       *tU = *oU * a + *tU * (1 - a);
+                       *tV = *oV * a + *tV * (1 - a);
+                       ++tY;
+                       ++oY;
+                       ++tU;
+                       ++tV;
+                       ++oU;
+                       ++oV;
+                       alpha += other.alpha_bpp;
+               }
+       }
+}
+
+
 void
 Image::alpha_blend (shared_ptr<const Image> other, Position<int> position)
 {
-       /* We're blending RGBA or BGRA images */
-       DCPOMATIC_ASSERT (other->pixel_format() == AV_PIX_FMT_BGRA || other->pixel_format() == AV_PIX_FMT_RGBA);
+       DCPOMATIC_ASSERT(
+               other->pixel_format() == AV_PIX_FMT_BGRA ||
+               other->pixel_format() == AV_PIX_FMT_RGBA ||
+               other->pixel_format() == AV_PIX_FMT_RGBA64BE
+               );
+
        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;
        int start_ox = 0;
 
@@ -634,218 +976,147 @@ Image::alpha_blend (shared_ptr<const Image> other, Position<int> position)
                start_ty = 0;
        }
 
+       TargetParams target_params = {
+               start_tx,
+               start_ty,
+               size(),
+               data(),
+               stride(),
+               0
+       };
+
+       OtherRGBParams other_rgb_params = {
+               start_ox,
+               start_oy,
+               other->size(),
+               other->data(),
+               other->stride(),
+               other->pixel_format() == AV_PIX_FMT_RGBA64BE ? 8 : 4
+       };
+
+       OtherYUVParams other_yuv_params = {
+               start_ox,
+               start_oy,
+               other->size(),
+               other->data(),
+               other->stride(),
+               nullptr,
+               nullptr,
+               other->pixel_format() == AV_PIX_FMT_RGBA64BE ? 8 : 4
+       };
+
+       auto byteswap = [](uint16_t* p) {
+               return (*p >> 8) | ((*p & 0xff) << 8);
+       };
+
+       auto pass = [](uint8_t* p) {
+               return *p;
+       };
+
+       auto get_alpha_64be = [](uint8_t* p) {
+               return ((static_cast<int16_t>(p[6]) << 8) | p[7]) / 65535.0f;
+       };
+
+       auto get_alpha_byte = [](uint8_t* p) {
+               return p[3] / 255.0f;
+       };
+
        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[red] * alpha + tp[0] * (1 - alpha);
-                               tp[1] = op[1] * alpha + tp[1] * (1 - alpha);
-                               tp[2] = op[blue] * alpha + tp[2] * (1 - alpha);
-
-                               tp += this_bpp;
-                               op += other_bpp;
-                       }
+               target_params.bpp = 3;
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_rgb24<uint16_t>(target_params, other_rgb_params, red, blue, byteswap, 256);
+               } else {
+                       alpha_blend_onto_rgb24<uint8_t>(target_params, other_rgb_params, red, blue, pass, 1);
                }
                break;
-       }
        case AV_PIX_FMT_BGRA:
-       {
-               int const this_bpp = 4;
-               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[blue] * alpha + tp[0] * (1 - alpha);
-                               tp[1] = op[1] * alpha + tp[1] * (1 - alpha);
-                               tp[2] = op[red] * alpha + tp[2] * (1 - alpha);
-                               tp[3] = op[3] * alpha + tp[3] * (1 - alpha);
-
-                               tp += this_bpp;
-                               op += other_bpp;
-                       }
+               target_params.bpp = 4;
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_bgra<uint16_t>(target_params, other_rgb_params, red, blue, byteswap, 256);
+               } else {
+                       alpha_blend_onto_bgra<uint8_t>(target_params, other_rgb_params, red, blue, pass, 1);
                }
                break;
-       }
        case AV_PIX_FMT_RGBA:
-       {
-               int const this_bpp = 4;
-               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[red] * alpha + tp[0] * (1 - alpha);
-                               tp[1] = op[1] * alpha + tp[1] * (1 - alpha);
-                               tp[2] = op[blue] * alpha + tp[2] * (1 - alpha);
-                               tp[3] = op[3] * alpha + tp[3] * (1 - alpha);
-
-                               tp += this_bpp;
-                               op += other_bpp;
-                       }
+               target_params.bpp = 4;
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_rgba<uint16_t>(target_params, other_rgb_params, red, blue, byteswap, 256);
+               } else {
+                       alpha_blend_onto_rgba<uint8_t>(target_params, other_rgb_params, red, blue, pass, 1);
                }
                break;
-       }
        case AV_PIX_FMT_RGB48LE:
-       {
-               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;
-                               /* Blend high bytes */
-                               tp[1] = op[red] * alpha + tp[1] * (1 - alpha);
-                               tp[3] = op[1] * alpha + tp[3] * (1 - alpha);
-                               tp[5] = op[blue] * alpha + tp[5] * (1 - alpha);
-
-                               tp += this_bpp;
-                               op += other_bpp;
-                       }
+               target_params.bpp = 6;
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_rgb48le<uint16_t>(target_params, other_rgb_params, red, blue, byteswap, 1);
+               } else {
+                       alpha_blend_onto_rgb48le<uint8_t>(target_params, other_rgb_params, red, blue, pass, 256);
                }
                break;
-       }
        case AV_PIX_FMT_XYZ12LE:
-       {
-               auto 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<uint16_t*> (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[red]];
-                               double const g = lut_in[op[1]];
-                               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]));
-                               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;
-                       }
+               target_params.bpp = 6;
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_xyz12le<uint16_t>(target_params, other_rgb_params, red, blue, byteswap, 256);
+               } else {
+                       alpha_blend_onto_xyz12le<uint8_t>(target_params, other_rgb_params, red, blue, pass, 1);
                }
                break;
-       }
        case AV_PIX_FMT_YUV420P:
        {
                auto yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, Alignment::COMPACT, 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) {
-                       int const hty = ty / 2;
-                       int const hoy = oy / 2;
-                       uint8_t* tY = data()[0] + (ty * stride()[0]) + start_tx;
-                       uint8_t* tU = data()[1] + (hty * stride()[1]) + start_tx / 2;
-                       uint8_t* tV = data()[2] + (hty * stride()[2]) + start_tx / 2;
-                       uint8_t* oY = yuv->data()[0] + (oy * yuv->stride()[0]) + start_ox;
-                       uint8_t* oU = yuv->data()[1] + (hoy * yuv->stride()[1]) + start_ox / 2;
-                       uint8_t* oV = yuv->data()[2] + (hoy * yuv->stride()[2]) + start_ox / 2;
-                       uint8_t* alpha = other->data()[0] + (oy * other->stride()[0]) + start_ox * 4;
-                       for (int tx = start_tx, ox = start_ox; tx < ts.width && ox < os.width; ++tx, ++ox) {
-                               float const a = float(alpha[3]) / 255;
-                               *tY = *oY * a + *tY * (1 - a);
-                               *tU = *oU * a + *tU * (1 - a);
-                               *tV = *oV * a + *tV * (1 - a);
-                               ++tY;
-                               ++oY;
-                               if (tx % 2) {
-                                       ++tU;
-                                       ++tV;
-                               }
-                               if (ox % 2) {
-                                       ++oU;
-                                       ++oV;
-                               }
-                               alpha += 4;
-                       }
+               other_yuv_params.data = yuv->data();
+               other_yuv_params.stride = yuv->stride();
+               other_yuv_params.alpha_data = other->data();
+               other_yuv_params.alpha_stride = other->stride();
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_yuv420p(target_params, other_yuv_params, get_alpha_64be);
+               } else {
+                       alpha_blend_onto_yuv420p(target_params, other_yuv_params, get_alpha_byte);
                }
                break;
        }
        case AV_PIX_FMT_YUV420P10:
        {
                auto yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, Alignment::COMPACT, 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) {
-                       int const hty = ty / 2;
-                       int const hoy = oy / 2;
-                       uint16_t* tY = ((uint16_t *) (data()[0] + (ty * stride()[0]))) + start_tx;
-                       uint16_t* tU = ((uint16_t *) (data()[1] + (hty * stride()[1]))) + start_tx / 2;
-                       uint16_t* tV = ((uint16_t *) (data()[2] + (hty * stride()[2]))) + start_tx / 2;
-                       uint16_t* oY = ((uint16_t *) (yuv->data()[0] + (oy * yuv->stride()[0]))) + start_ox;
-                       uint16_t* oU = ((uint16_t *) (yuv->data()[1] + (hoy * yuv->stride()[1]))) + start_ox / 2;
-                       uint16_t* oV = ((uint16_t *) (yuv->data()[2] + (hoy * yuv->stride()[2]))) + start_ox / 2;
-                       uint8_t* alpha = other->data()[0] + (oy * other->stride()[0]) + start_ox * 4;
-                       for (int tx = start_tx, ox = start_ox; tx < ts.width && ox < os.width; ++tx, ++ox) {
-                               float const a = float(alpha[3]) / 255;
-                               *tY = *oY * a + *tY * (1 - a);
-                               *tU = *oU * a + *tU * (1 - a);
-                               *tV = *oV * a + *tV * (1 - a);
-                               ++tY;
-                               ++oY;
-                               if (tx % 2) {
-                                       ++tU;
-                                       ++tV;
-                               }
-                               if (ox % 2) {
-                                       ++oU;
-                                       ++oV;
-                               }
-                               alpha += 4;
-                       }
+               other_yuv_params.data = yuv->data();
+               other_yuv_params.stride = yuv->stride();
+               other_yuv_params.alpha_data = other->data();
+               other_yuv_params.alpha_stride = other->stride();
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_yuv420p10(target_params, other_yuv_params, get_alpha_64be);
+               } else {
+                       alpha_blend_onto_yuv420p10(target_params, other_yuv_params, get_alpha_byte);
                }
                break;
        }
+       case AV_PIX_FMT_YUV422P9LE:
        case AV_PIX_FMT_YUV422P10LE:
        {
                auto yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, Alignment::COMPACT, 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) {
-                       uint16_t* tY = ((uint16_t *) (data()[0] + (ty * stride()[0]))) + start_tx;
-                       uint16_t* tU = ((uint16_t *) (data()[1] + (ty * stride()[1]))) + start_tx / 2;
-                       uint16_t* tV = ((uint16_t *) (data()[2] + (ty * stride()[2]))) + start_tx / 2;
-                       uint16_t* oY = ((uint16_t *) (yuv->data()[0] + (oy * yuv->stride()[0]))) + start_ox;
-                       uint16_t* oU = ((uint16_t *) (yuv->data()[1] + (oy * yuv->stride()[1]))) + start_ox / 2;
-                       uint16_t* oV = ((uint16_t *) (yuv->data()[2] + (oy * yuv->stride()[2]))) + start_ox / 2;
-                       uint8_t* alpha = other->data()[0] + (oy * other->stride()[0]) + start_ox * 4;
-                       for (int tx = start_tx, ox = start_ox; tx < ts.width && ox < os.width; ++tx, ++ox) {
-                               float const a = float(alpha[3]) / 255;
-                               *tY = *oY * a + *tY * (1 - a);
-                               *tU = *oU * a + *tU * (1 - a);
-                               *tV = *oV * a + *tV * (1 - a);
-                               ++tY;
-                               ++oY;
-                               if (tx % 2) {
-                                       ++tU;
-                                       ++tV;
-                               }
-                               if (ox % 2) {
-                                       ++oU;
-                                       ++oV;
-                               }
-                               alpha += 4;
-                       }
+               other_yuv_params.data = yuv->data();
+               other_yuv_params.stride = yuv->stride();
+               other_yuv_params.alpha_data = other->data();
+               other_yuv_params.alpha_stride = other->stride();
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_yuv422p9or10le(target_params, other_yuv_params, get_alpha_64be);
+               } else {
+                       alpha_blend_onto_yuv422p9or10le(target_params, other_yuv_params, get_alpha_byte);
+               }
+               break;
+       }
+       case AV_PIX_FMT_YUV444P9LE:
+       case AV_PIX_FMT_YUV444P10LE:
+       {
+               auto yuv = other->convert_pixel_format (dcp::YUVToRGB::REC709, _pixel_format, Alignment::COMPACT, false);
+               other_yuv_params.data = yuv->data();
+               other_yuv_params.stride = yuv->stride();
+               other_yuv_params.alpha_data = other->data();
+               other_yuv_params.alpha_stride = other->stride();
+               if (other->pixel_format() == AV_PIX_FMT_RGBA64BE) {
+                       alpha_blend_onto_yuv444p9or10le(target_params, other_yuv_params, get_alpha_64be);
+               } else {
+                       alpha_blend_onto_yuv444p9or10le(target_params, other_yuv_params, get_alpha_byte);
                }
                break;
        }
@@ -1371,109 +1642,6 @@ Image::memory_used () const
 }
 
 
-class Memory
-{
-public:
-       Memory ()
-               : data(0)
-               , size(0)
-       {}
-
-       ~Memory ()
-       {
-               free (data);
-       }
-
-       uint8_t* data;
-       size_t size;
-};
-
-
-static void
-png_write_data (png_structp png_ptr, png_bytep data, png_size_t length)
-{
-       auto mem = reinterpret_cast<Memory*>(png_get_io_ptr(png_ptr));
-       size_t size = mem->size + length;
-
-       if (mem->data) {
-               mem->data = reinterpret_cast<uint8_t*>(realloc(mem->data, size));
-       } else {
-               mem->data = reinterpret_cast<uint8_t*>(malloc(size));
-       }
-
-       if (!mem->data) {
-               throw EncodeError (N_("could not allocate memory for PNG"));
-       }
-
-       memcpy (mem->data + mem->size, data, length);
-       mem->size += length;
-}
-
-
-static void
-png_flush (png_structp)
-{
-
-}
-
-
-static void
-png_error_fn (png_structp png_ptr, char const * message)
-{
-       reinterpret_cast<Image*>(png_get_error_ptr(png_ptr))->png_error (message);
-}
-
-
-void
-Image::png_error (char const * message)
-{
-       throw EncodeError (String::compose ("Error during PNG write: %1", message));
-}
-
-
-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::YUVToRGB::REC709, AV_PIX_FMT_RGBA, Image::Alignment::PADDED, 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);
-       if (!png_ptr) {
-               throw EncodeError (N_("could not create PNG write struct"));
-       }
-
-       Memory state;
-
-       png_set_write_fn (png_ptr, &state, png_write_data, png_flush);
-
-       png_infop info_ptr = png_create_info_struct(png_ptr);
-       if (!info_ptr) {
-               png_destroy_write_struct (&png_ptr, &info_ptr);
-               throw EncodeError (N_("could not create PNG info struct"));
-       }
-
-       png_set_IHDR (png_ptr, info_ptr, size().width, size().height, 8, PNG_COLOR_TYPE_RGBA, PNG_INTERLACE_NONE, PNG_COMPRESSION_TYPE_DEFAULT, PNG_FILTER_TYPE_DEFAULT);
-
-       png_byte ** row_pointers = reinterpret_cast<png_byte **>(png_malloc(png_ptr, size().height * sizeof(png_byte *)));
-       for (int i = 0; i < size().height; ++i) {
-               row_pointers[i] = (png_byte *) (data()[0] + i * stride()[0]);
-       }
-
-       png_write_info (png_ptr, info_ptr);
-       png_write_image (png_ptr, row_pointers);
-       png_write_end (png_ptr, info_ptr);
-
-       png_destroy_write_struct (&png_ptr, &info_ptr);
-       png_free (png_ptr, row_pointers);
-
-       return dcp::ArrayData (state.data, state.size);
-}
-
-
 void
 Image::video_range_to_full_range ()
 {