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

   1 #include <string.h>
   2 #include <assert.h>
   3 #include <GL/glew.h>
   4
   5 #include <Eigen/LU>
   6
   7 #include "ycbcr_input.h"
   8 #include "util.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         pixel_data[0] = pixel_data[1] = pixel_data[2] = NULL;
  83
  84         register_int("needs_mipmaps", &needs_mipmaps);
  85 }
  86
  87 YCbCrInput::~YCbCrInput()
  88 {
  89         if (pbos[0] != 0) {
  90                 glDeleteBuffers(3, pbos);
  91                 check_error();
  92         }
  93         if (texture_num[0] != 0) {
  94                 glDeleteTextures(3, texture_num);
  95                 check_error();
  96         }
  97 }
  98
  99 void YCbCrInput::finalize()
 100 {
 101         glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
 102         check_error();
 103
 104         // Create PBOs to hold the textures holding the input image, and then the texture itself.
 105         glGenBuffers(3, pbos);
 106         check_error();
 107         glGenTextures(3, texture_num);
 108         check_error();
 109
 110         for (unsigned channel = 0; channel < 3; ++channel) {
 111                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
 112                 check_error();
 113                 glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, pitch[channel] * heights[channel], NULL, GL_STREAM_DRAW);
 114                 check_error();
 115                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 116                 check_error();
 117
 118                 glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 119                 check_error();
 120                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 121                 check_error();
 122                 glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
 123                 check_error();
 124                 glTexImage2D(GL_TEXTURE_2D, 0, GL_LUMINANCE8, widths[channel], heights[channel], 0, GL_LUMINANCE, GL_UNSIGNED_BYTE, NULL);
 125                 check_error();
 126                 glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 127                 check_error();
 128         }
 129
 130         needs_update = true;
 131         finalized = true;
 132 }
 133
 134 void YCbCrInput::set_gl_state(GLuint glsl_program_num, const std::string& prefix, unsigned *sampler_num)
 135 {
 136         for (unsigned channel = 0; channel < 3; ++channel) {
 137                 glActiveTexture(GL_TEXTURE0 + *sampler_num + channel);
 138                 check_error();
 139                 glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 140                 check_error();
 141
 142                 if (needs_update || needs_pbo_recreate) {
 143                         // Copy the pixel data into the PBO.
 144                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
 145                         check_error();
 146
 147                         if (needs_pbo_recreate) {
 148                                 // The pitch has changed; we need to reallocate this PBO.
 149                                 glBufferData(GL_PIXEL_UNPACK_BUFFER_ARB, pitch[channel] * heights[channel], NULL, GL_STREAM_DRAW);
 150                                 check_error();
 151                         }
 152
 153                         void *mapped_pbo = glMapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB, GL_WRITE_ONLY);
 154                         memcpy(mapped_pbo, pixel_data[channel], pitch[channel] * heights[channel]);
 155
 156                         glUnmapBufferARB(GL_PIXEL_UNPACK_BUFFER_ARB);
 157                         check_error();
 158
 159                         // Re-upload the texture from the PBO.
 160                         glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
 161                         check_error();
 162                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], GL_LUMINANCE, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
 163                         check_error();
 164                         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 165                         check_error();
 166                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 167                         check_error();
 168                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 169                         check_error();
 170                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 171                         check_error();
 172                 }
 173         }
 174
 175         // Bind samplers.
 176         set_uniform_int(glsl_program_num, prefix, "tex_y", *sampler_num + 0);
 177         set_uniform_int(glsl_program_num, prefix, "tex_cb", *sampler_num + 1);
 178         set_uniform_int(glsl_program_num, prefix, "tex_cr", *sampler_num + 2);
 179
 180         *sampler_num += 3;
 181         needs_update = false;
 182         needs_pbo_recreate = false;
 183 }
 184
 185 std::string YCbCrInput::output_fragment_shader()
 186 {
 187         float coeff[3], offset[3], scale[3];
 188
 189         switch (ycbcr_format.luma_coefficients) {
 190         case YCBCR_REC_601:
 191                 // Rec. 601, page 2.
 192                 coeff[0] = 0.299;
 193                 coeff[1] = 0.587;
 194                 coeff[2] = 0.114;
 195                 break;
 196
 197         case YCBCR_REC_709:
 198                 // Rec. 709, page 19.
 199                 coeff[0] = 0.2126;
 200                 coeff[1] = 0.7152;
 201                 coeff[2] = 0.0722;
 202                 break;
 203         default:
 204                 assert(false);
 205         }
 206
 207         if (ycbcr_format.full_range) {
 208                 offset[0] = 0.0 / 255.0;
 209                 offset[1] = 128.0 / 255.0;
 210                 offset[2] = 128.0 / 255.0;
 211
 212                 scale[0] = 1.0;
 213                 scale[1] = 1.0;
 214                 scale[2] = 1.0;
 215         } else {
 216                 // Rec. 601, page 4; Rec. 709, page 19.
 217                 offset[0] = 16.0 / 255.0;
 218                 offset[1] = 128.0 / 255.0;
 219                 offset[2] = 128.0 / 255.0;
 220
 221                 scale[0] = 255.0 / 219.0;
 222                 scale[1] = 255.0 / 224.0;
 223                 scale[2] = 255.0 / 224.0;
 224         }
 225
 226         // Matrix to convert RGB to YCbCr. See e.g. Rec. 601.
 227         Matrix3d rgb_to_ycbcr;
 228         rgb_to_ycbcr(0,0) = coeff[0];
 229         rgb_to_ycbcr(0,1) = coeff[1];
 230         rgb_to_ycbcr(0,2) = coeff[2];
 231
 232         float cb_fac = (224.0 / 219.0) / (coeff[0] + coeff[1] + 1.0f - coeff[2]);
 233         rgb_to_ycbcr(1,0) = -coeff[0] * cb_fac;
 234         rgb_to_ycbcr(1,1) = -coeff[1] * cb_fac;
 235         rgb_to_ycbcr(1,2) = (1.0f - coeff[2]) * cb_fac;
 236
 237         float cr_fac = (224.0 / 219.0) / (1.0f - coeff[0] + coeff[1] + coeff[2]);
 238         rgb_to_ycbcr(2,0) = (1.0f - coeff[0]) * cr_fac;
 239         rgb_to_ycbcr(2,1) = -coeff[1] * cr_fac;
 240         rgb_to_ycbcr(2,2) = -coeff[2] * cr_fac;
 241
 242         // Inverting the matrix gives us what we need to go from YCbCr back to RGB.
 243         Matrix3d ycbcr_to_rgb = rgb_to_ycbcr.inverse();
 244
 245         std::string frag_shader;
 246
 247         frag_shader = output_glsl_mat3("PREFIX(inv_ycbcr_matrix)", ycbcr_to_rgb);
 248
 249         char buf[256];
 250         sprintf(buf, "const vec3 PREFIX(offset) = vec3(%.8f, %.8f, %.8f);\n",
 251                 offset[0], offset[1], offset[2]);
 252         frag_shader += buf;
 253
 254         sprintf(buf, "const vec3 PREFIX(scale) = vec3(%.8f, %.8f, %.8f);\n",
 255                 scale[0], scale[1], scale[2]);
 256         frag_shader += buf;
 257
 258         float cb_offset_x = compute_chroma_offset(
 259                 ycbcr_format.cb_x_position, ycbcr_format.chroma_subsampling_x, widths[1]);
 260         float cb_offset_y = compute_chroma_offset(
 261                 ycbcr_format.cb_y_position, ycbcr_format.chroma_subsampling_y, heights[1]);
 262         sprintf(buf, "const vec2 PREFIX(cb_offset) = vec2(%.8f, %.8f);\n",
 263                 cb_offset_x, cb_offset_y);
 264         frag_shader += buf;
 265
 266         float cr_offset_x = compute_chroma_offset(
 267                 ycbcr_format.cr_x_position, ycbcr_format.chroma_subsampling_x, widths[2]);
 268         float cr_offset_y = compute_chroma_offset(
 269                 ycbcr_format.cr_y_position, ycbcr_format.chroma_subsampling_y, heights[2]);
 270         sprintf(buf, "const vec2 PREFIX(cr_offset) = vec2(%.8f, %.8f);\n",
 271                 cr_offset_x, cr_offset_y);
 272         frag_shader += buf;
 273
 274         frag_shader += read_file("ycbcr_input.frag");
 275         return frag_shader;
 276 }