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

   1 #include <Eigen/Core>
   2 #include <Eigen/LU>
   3 #include <GL/glew.h>
   4 #include <assert.h>
   5 #include <stdio.h>
   6 #include <string.h>
   7
   8 #include "effect_util.h"
   9 #include "resource_pool.h"
  10 #include "util.h"
  11 #include "ycbcr_input.h"
  12
  13 using namespace Eigen;
  14
  15 namespace {
  16
  17 // OpenGL has texel center in (0.5, 0.5), but different formats have
  18 // chroma in various other places. If luma samples are X, the chroma
  19 // sample is *, and subsampling is 3x3, the situation with chroma
  20 // center in (0.5, 0.5) looks approximately like this:
  21 //
  22 //   X   X
  23 //     *
  24 //   X   X
  25 //
  26 // If, on the other hand, chroma center is in (0.0, 0.5) (common
  27 // for e.g. MPEG-4), the figure changes to:
  28 //
  29 //   X   X
  30 //   *
  31 //   X   X
  32 //
  33 // In other words, (0.0, 0.0) means that the chroma sample is exactly
  34 // co-sited on top of the top-left luma sample. Note, however, that
  35 // this is _not_ 0.5 texels to the left, since the OpenGL's texel center
  36 // is in (0.5, 0.5); it is in (0.25, 0.25). In a sense, the four luma samples
  37 // define a square where chroma position (0.0, 0.0) is in texel position
  38 // (0.25, 0.25) and chroma position (1.0, 1.0) is in texel position (0.75, 0.75)
  39 // (the outer border shows the borders of the texel itself, ie. from
  40 // (0, 0) to (1, 1)):
  41 //
  42 //  ---------
  43 // |         |
  44 // |  X---X  |
  45 // |  | * |  |
  46 // |  X---X  |
  47 // |         |
  48 //  ---------
  49 //
  50 // Also note that if we have no subsampling, the square will have zero
  51 // area and the chroma position does not matter at all.
  52 float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
  53 {
  54         float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
  55         return (0.5 - local_chroma_pos) / resolution;
  56 }
  57
  58 }  // namespace
  59
  60 YCbCrInput::YCbCrInput(const ImageFormat &image_format,
  61                        const YCbCrFormat &ycbcr_format,
  62                        unsigned width, unsigned height)
  63         : image_format(image_format),
  64           ycbcr_format(ycbcr_format),
  65           needs_update(false),
  66           finalized(false),
  67           needs_mipmaps(false),
  68           width(width),
  69           height(height)
  70 {
  71         pbos[0] = pbos[1] = pbos[2] = 0;
  72         texture_num[0] = texture_num[1] = texture_num[2] = 0;
  73
  74         assert(width % ycbcr_format.chroma_subsampling_x == 0);
  75         pitch[0] = widths[0] = width;
  76         pitch[1] = widths[1] = width / ycbcr_format.chroma_subsampling_x;
  77         pitch[2] = widths[2] = width / ycbcr_format.chroma_subsampling_x;
  78
  79         assert(height % ycbcr_format.chroma_subsampling_y == 0);
  80         heights[0] = height;
  81         heights[1] = height / ycbcr_format.chroma_subsampling_y;
  82         heights[2] = height / ycbcr_format.chroma_subsampling_y;
  83
  84         pixel_data[0] = pixel_data[1] = pixel_data[2] = NULL;
  85
  86         register_int("needs_mipmaps", &needs_mipmaps);
  87 }
  88
  89 YCbCrInput::~YCbCrInput()
  90 {
  91         for (unsigned channel = 0; channel < 3; ++channel) {
  92                 if (texture_num[channel] != 0) {
  93                         resource_pool->release_2d_texture(texture_num[channel]);
  94                 }
  95         }
  96 }
  97
  98 void YCbCrInput::finalize()
  99 {
 100         // Create the textures themselves.
 101         for (unsigned channel = 0; channel < 3; ++channel) {
 102                 texture_num[channel] = resource_pool->create_2d_texture(GL_LUMINANCE8, widths[channel], heights[channel]);
 103                 glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 104                 check_error();
 105                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 106                 check_error();
 107         }
 108
 109         needs_update = true;
 110         finalized = true;
 111 }
 112
 113 void YCbCrInput::set_gl_state(GLuint glsl_program_num, const std::string& prefix, unsigned *sampler_num)
 114 {
 115         for (unsigned channel = 0; channel < 3; ++channel) {
 116                 glActiveTexture(GL_TEXTURE0 + *sampler_num + channel);
 117                 check_error();
 118                 glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 119                 check_error();
 120
 121                 if (needs_update) {
 122                         // Re-upload the texture.
 123                         // Copy the pixel data into the PBO.
 124                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
 125                         check_error();
 126                         glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
 127                         check_error();
 128                         glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
 129                         check_error();
 130                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], GL_LUMINANCE, GL_UNSIGNED_BYTE, pixel_data[channel]);
 131                         check_error();
 132                         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 133                         check_error();
 134                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 135                         check_error();
 136                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 137                         check_error();
 138                 }
 139         }
 140
 141         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 142         check_error();
 143
 144         // Bind samplers.
 145         set_uniform_int(glsl_program_num, prefix, "tex_y", *sampler_num + 0);
 146         set_uniform_int(glsl_program_num, prefix, "tex_cb", *sampler_num + 1);
 147         set_uniform_int(glsl_program_num, prefix, "tex_cr", *sampler_num + 2);
 148
 149         *sampler_num += 3;
 150         needs_update = false;
 151 }
 152
 153 std::string YCbCrInput::output_fragment_shader()
 154 {
 155         float coeff[3], offset[3], scale[3];
 156
 157         switch (ycbcr_format.luma_coefficients) {
 158         case YCBCR_REC_601:
 159                 // Rec. 601, page 2.
 160                 coeff[0] = 0.299;
 161                 coeff[1] = 0.587;
 162                 coeff[2] = 0.114;
 163                 break;
 164
 165         case YCBCR_REC_709:
 166                 // Rec. 709, page 19.
 167                 coeff[0] = 0.2126;
 168                 coeff[1] = 0.7152;
 169                 coeff[2] = 0.0722;
 170                 break;
 171
 172         case YCBCR_REC_2020:
 173                 // Rec. 2020, page 4.
 174                 coeff[0] = 0.2627;
 175                 coeff[1] = 0.6780;
 176                 coeff[2] = 0.0593;
 177                 break;
 178
 179         default:
 180                 assert(false);
 181         }
 182
 183         if (ycbcr_format.full_range) {
 184                 offset[0] = 0.0 / 255.0;
 185                 offset[1] = 128.0 / 255.0;
 186                 offset[2] = 128.0 / 255.0;
 187
 188                 scale[0] = 1.0;
 189                 scale[1] = 1.0;
 190                 scale[2] = 1.0;
 191         } else {
 192                 // Rec. 601, page 4; Rec. 709, page 19; Rec. 2020, page 4.
 193                 offset[0] = 16.0 / 255.0;
 194                 offset[1] = 128.0 / 255.0;
 195                 offset[2] = 128.0 / 255.0;
 196
 197                 scale[0] = 255.0 / 219.0;
 198                 scale[1] = 255.0 / 224.0;
 199                 scale[2] = 255.0 / 224.0;
 200         }
 201
 202         // Matrix to convert RGB to YCbCr. See e.g. Rec. 601.
 203         Matrix3d rgb_to_ycbcr;
 204         rgb_to_ycbcr(0,0) = coeff[0];
 205         rgb_to_ycbcr(0,1) = coeff[1];
 206         rgb_to_ycbcr(0,2) = coeff[2];
 207
 208         float cb_fac = (224.0 / 219.0) / (coeff[0] + coeff[1] + 1.0f - coeff[2]);
 209         rgb_to_ycbcr(1,0) = -coeff[0] * cb_fac;
 210         rgb_to_ycbcr(1,1) = -coeff[1] * cb_fac;
 211         rgb_to_ycbcr(1,2) = (1.0f - coeff[2]) * cb_fac;
 212
 213         float cr_fac = (224.0 / 219.0) / (1.0f - coeff[0] + coeff[1] + coeff[2]);
 214         rgb_to_ycbcr(2,0) = (1.0f - coeff[0]) * cr_fac;
 215         rgb_to_ycbcr(2,1) = -coeff[1] * cr_fac;
 216         rgb_to_ycbcr(2,2) = -coeff[2] * cr_fac;
 217
 218         // Inverting the matrix gives us what we need to go from YCbCr back to RGB.
 219         Matrix3d ycbcr_to_rgb = rgb_to_ycbcr.inverse();
 220
 221         std::string frag_shader;
 222
 223         frag_shader = output_glsl_mat3("PREFIX(inv_ycbcr_matrix)", ycbcr_to_rgb);
 224
 225         char buf[256];
 226         sprintf(buf, "const vec3 PREFIX(offset) = vec3(%.8f, %.8f, %.8f);\n",
 227                 offset[0], offset[1], offset[2]);
 228         frag_shader += buf;
 229
 230         sprintf(buf, "const vec3 PREFIX(scale) = vec3(%.8f, %.8f, %.8f);\n",
 231                 scale[0], scale[1], scale[2]);
 232         frag_shader += buf;
 233
 234         float cb_offset_x = compute_chroma_offset(
 235                 ycbcr_format.cb_x_position, ycbcr_format.chroma_subsampling_x, widths[1]);
 236         float cb_offset_y = compute_chroma_offset(
 237                 ycbcr_format.cb_y_position, ycbcr_format.chroma_subsampling_y, heights[1]);
 238         sprintf(buf, "const vec2 PREFIX(cb_offset) = vec2(%.8f, %.8f);\n",
 239                 cb_offset_x, cb_offset_y);
 240         frag_shader += buf;
 241
 242         float cr_offset_x = compute_chroma_offset(
 243                 ycbcr_format.cr_x_position, ycbcr_format.chroma_subsampling_x, widths[2]);
 244         float cr_offset_y = compute_chroma_offset(
 245                 ycbcr_format.cr_y_position, ycbcr_format.chroma_subsampling_y, heights[2]);
 246         sprintf(buf, "const vec2 PREFIX(cr_offset) = vec2(%.8f, %.8f);\n",
 247                 cr_offset_x, cr_offset_y);
 248         frag_shader += buf;
 249
 250         frag_shader += read_file("ycbcr_input.frag");
 251         return frag_shader;
 252 }