git.sesse.net Git - movit/blob - ycbcr_input.cpp

   1 #include <string.h>
   2 #include <assert.h>
   3
   4 #include <Eigen/LU>
   5
   6 #include "ycbcr_input.h"
   7 #include "util.h"
   8 #include "opengl.h"
   9
  10 using namespace Eigen;
  11
  12 namespace {
  13
  14 // OpenGL has texel center in (0.5, 0.5), but different formats have
  15 // chroma in various other places. If luma samples are X, the chroma
  16 // sample is *, and subsampling is 3x3, the situation with chroma
  17 // center in (0.5, 0.5) looks approximately like this:
  18 //
  19 //   X   X
  20 //     *
  21 //   X   X
  22 //
  23 // If, on the other hand, chroma center is in (0.0, 0.5) (common
  24 // for e.g. MPEG-4), the figure changes to:
  25 //
  26 //   X   X
  27 //   *
  28 //   X   X
  29 //
  30 // In other words, (0.0, 0.0) means that the chroma sample is exactly
  31 // co-sited on top of the top-left luma sample. Note, however, that
  32 // this is _not_ 0.5 texels to the left, since the OpenGL's texel center
  33 // is in (0.5, 0.5); it is in (0.25, 0.25). In a sense, the four luma samples
  34 // define a square where chroma position (0.0, 0.0) is in texel position
  35 // (0.25, 0.25) and chroma position (1.0, 1.0) is in texel position (0.75, 0.75)
  36 // (the outer border shows the borders of the texel itself, ie. from
  37 // (0, 0) to (1, 1)):
  38 //
  39 //  ---------
  40 // |         |
  41 // |  X---X  |
  42 // |  | * |  |
  43 // |  X---X  |
  44 // |         |
  45 //  ---------
  46 //
  47 // Also note that if we have no subsampling, the square will have zero
  48 // area and the chroma position does not matter at all.
  49 float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
  50 {
  51         float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
  52         return (0.5 - local_chroma_pos) / resolution;
  53 }
  54
  55 }  // namespace
  56
  57 YCbCrInput::YCbCrInput(const ImageFormat &image_format,
  58                        const YCbCrFormat &ycbcr_format,
  59                        unsigned width, unsigned height)
  60         : image_format(image_format),
  61           ycbcr_format(ycbcr_format),
  62           needs_update(false),
  63           needs_pbo_recreate(false),
  64           finalized(false),
  65           needs_mipmaps(false),
  66           width(width),
  67           height(height)
  68 {
  69         pbos[0] = pbos[1] = pbos[2] = 0;
  70         texture_num[0] = texture_num[1] = texture_num[2] = 0;
  71
  72         assert(width % ycbcr_format.chroma_subsampling_x == 0);
  73         pitch[0] = widths[0] = width;
  74         pitch[1] = widths[1] = width / ycbcr_format.chroma_subsampling_x;
  75         pitch[2] = widths[2] = width / ycbcr_format.chroma_subsampling_x;
  76
  77         assert(height % ycbcr_format.chroma_subsampling_y == 0);
  78         heights[0] = height;
  79         heights[1] = height / ycbcr_format.chroma_subsampling_y;
  80         heights[2] = height / ycbcr_format.chroma_subsampling_y;
  81
  82         register_int("needs_mipmaps", &needs_mipmaps);
  83 }
  84
  85 YCbCrInput::~YCbCrInput()
  86 {
  87         if (pbos[0] != 0) {
  88                 glDeleteBuffers(3, pbos);
  89                 check_error();
  90         }
  91         if (texture_num[0] != 0) {
  92                 glDeleteTextures(3, texture_num);
  93                 check_error();
  94         }
  95 }
  96
  97 void YCbCrInput::finalize()
  98 {
  99         glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
 100         check_error();
 101
 102         // Create PBOs to hold the textures holding the input image, and then the texture itself.
 103         glGenBuffers(3, pbos);
 104         check_error();
 105         glGenTextures(3, texture_num);
 106         check_error();
 107
 108         for (unsigned channel = 0; channel < 3; ++channel) {
 109                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
 110                 check_error();
 111                 glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, pitch[channel] * heights[channel], NULL, GL_STREAM_DRAW);
 112                 check_error();
 113                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 114                 check_error();
 115
 116                 glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 117                 check_error();
 118                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 119                 check_error();
 120                 glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
 121                 check_error();
 122                 glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE8, widths[channel], heights[channel], 0, GL_LUMINANCE, GL_UNSIGNED_BYTE, NULL);
 123                 check_error();
 124                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 125                 check_error();
 126         }
 127
 128         needs_update = true;
 129         finalized = true;
 130 }
 131
 132 void YCbCrInput::set_gl_state(GLuint glsl_program_num, const std::string& prefix, unsigned *sampler_num)
 133 {
 134         for (unsigned channel = 0; channel < 3; ++channel) {
 135                 glActiveTexture(GL_TEXTURE0 + *sampler_num + channel);
 136                 check_error();
 137                 glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 138                 check_error();
 139
 140                 if (needs_update || needs_pbo_recreate) {
 141                         // Copy the pixel data into the PBO.
 142                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
 143                         check_error();
 144
 145                         if (needs_pbo_recreate) {
 146                                 // The pitch has changed; we need to reallocate this PBO.
 147                                 glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, pitch[channel] * heights[channel], NULL, GL_STREAM_DRAW);
 148                                 check_error();
 149                         }
 150
 151                         void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
 152                         memcpy(mapped_pbo, pixel_data[channel], pitch[channel] * heights[channel]);
 153
 154                         glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
 155                         check_error();
 156
 157                         // Re-upload the texture from the PBO.
 158                         glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
 159                         check_error();
 160                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], GL_LUMINANCE, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
 161                         check_error();
 162                         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 163                         check_error();
 164                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 165                         check_error();
 166                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 167                         check_error();
 168                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 169                         check_error();
 170                 }
 171         }
 172
 173         // Bind samplers.
 174         set_uniform_int(glsl_program_num, prefix, "tex_y", *sampler_num + 0);
 175         set_uniform_int(glsl_program_num, prefix, "tex_cb", *sampler_num + 1);
 176         set_uniform_int(glsl_program_num, prefix, "tex_cr", *sampler_num + 2);
 177
 178         *sampler_num += 3;
 179         needs_update = false;
 180         needs_pbo_recreate = false;
 181 }
 182
 183 std::string YCbCrInput::output_fragment_shader()
 184 {
 185         float coeff[3], offset[3], scale[3];
 186
 187         switch (ycbcr_format.luma_coefficients) {
 188         case YCBCR_REC_601:
 189                 // Rec. 601, page 2.
 190                 coeff[0] = 0.299;
 191                 coeff[1] = 0.587;
 192                 coeff[2] = 0.114;
 193                 break;
 194
 195         case YCBCR_REC_709:
 196                 // Rec. 709, page 19.
 197                 coeff[0] = 0.2126;
 198                 coeff[1] = 0.7152;
 199                 coeff[2] = 0.0722;
 200                 break;
 201         default:
 202                 assert(false);
 203         }
 204
 205         if (ycbcr_format.full_range) {
 206                 offset[0] = 0.0 / 255.0;
 207                 offset[1] = 128.0 / 255.0;
 208                 offset[2] = 128.0 / 255.0;
 209
 210                 scale[0] = 1.0;
 211                 scale[1] = 1.0;
 212                 scale[2] = 1.0;
 213         } else {
 214                 // Rec. 601, page 4; Rec. 709, page 19.
 215                 offset[0] = 16.0 / 255.0;
 216                 offset[1] = 128.0 / 255.0;
 217                 offset[2] = 128.0 / 255.0;
 218
 219                 scale[0] = 255.0 / 219.0;
 220                 scale[1] = 255.0 / 224.0;
 221                 scale[2] = 255.0 / 224.0;
 222         }
 223
 224         // Matrix to convert RGB to YCbCr. See e.g. Rec. 601.
 225         Matrix3d rgb_to_ycbcr;
 226         rgb_to_ycbcr(0,0) = coeff[0];
 227         rgb_to_ycbcr(0,1) = coeff[1];
 228         rgb_to_ycbcr(0,2) = coeff[2];
 229
 230         float cb_fac = (224.0 / 219.0) / (coeff[0] + coeff[1] + 1.0f - coeff[2]);
 231         rgb_to_ycbcr(1,0) = -coeff[0] * cb_fac;
 232         rgb_to_ycbcr(1,1) = -coeff[1] * cb_fac;
 233         rgb_to_ycbcr(1,2) = (1.0f - coeff[2]) * cb_fac;
 234
 235         float cr_fac = (224.0 / 219.0) / (1.0f - coeff[0] + coeff[1] + coeff[2]);
 236         rgb_to_ycbcr(2,0) = (1.0f - coeff[0]) * cr_fac;
 237         rgb_to_ycbcr(2,1) = -coeff[1] * cr_fac;
 238         rgb_to_ycbcr(2,2) = -coeff[2] * cr_fac;
 239
 240         // Inverting the matrix gives us what we need to go from YCbCr back to RGB.
 241         Matrix3d ycbcr_to_rgb = rgb_to_ycbcr.inverse();
 242
 243         std::string frag_shader;
 244
 245         frag_shader = output_glsl_mat3("PREFIX(inv_ycbcr_matrix)", ycbcr_to_rgb);
 246
 247         char buf[256];
 248         sprintf(buf, "const vec3 PREFIX(offset) = vec3(%.8f, %.8f, %.8f);\n",
 249                 offset[0], offset[1], offset[2]);
 250         frag_shader += buf;
 251
 252         sprintf(buf, "const vec3 PREFIX(scale) = vec3(%.8f, %.8f, %.8f);\n",
 253                 scale[0], scale[1], scale[2]);
 254         frag_shader += buf;
 255
 256         float chroma_offset_x = compute_chroma_offset(
 257                 ycbcr_format.chroma_x_position, ycbcr_format.chroma_subsampling_x, widths[1]);
 258         float chroma_offset_y = compute_chroma_offset(
 259                 ycbcr_format.chroma_y_position, ycbcr_format.chroma_subsampling_y, heights[1]);
 260         sprintf(buf, "const vec2 PREFIX(chroma_offset) = vec2(%.8f, %.8f);\n",
 261                 chroma_offset_x, chroma_offset_y);
 262         frag_shader += buf;
 263
 264         frag_shader += read_file("ycbcr_input.frag");
 265         return frag_shader;
 266 }