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 using namespace std;
  15
  16 namespace {
  17
  18 // OpenGL has texel center in (0.5, 0.5), but different formats have
  19 // chroma in various other places. If luma samples are X, the chroma
  20 // sample is *, and subsampling is 3x3, the situation with chroma
  21 // center in (0.5, 0.5) looks approximately like this:
  22 //
  23 //   X   X
  24 //     *
  25 //   X   X
  26 //
  27 // If, on the other hand, chroma center is in (0.0, 0.5) (common
  28 // for e.g. MPEG-4), the figure changes to:
  29 //
  30 //   X   X
  31 //   *
  32 //   X   X
  33 //
  34 // In other words, (0.0, 0.0) means that the chroma sample is exactly
  35 // co-sited on top of the top-left luma sample. Note, however, that
  36 // this is _not_ 0.5 texels to the left, since the OpenGL's texel center
  37 // is in (0.5, 0.5); it is in (0.25, 0.25). In a sense, the four luma samples
  38 // define a square where chroma position (0.0, 0.0) is in texel position
  39 // (0.25, 0.25) and chroma position (1.0, 1.0) is in texel position (0.75, 0.75)
  40 // (the outer border shows the borders of the texel itself, ie. from
  41 // (0, 0) to (1, 1)):
  42 //
  43 //  ---------
  44 // |         |
  45 // |  X---X  |
  46 // |  | * |  |
  47 // |  X---X  |
  48 // |         |
  49 //  ---------
  50 //
  51 // Also note that if we have no subsampling, the square will have zero
  52 // area and the chroma position does not matter at all.
  53 float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
  54 {
  55         float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
  56         return (0.5 - local_chroma_pos) / resolution;
  57 }
  58
  59 }  // namespace
  60
  61 YCbCrInput::YCbCrInput(const ImageFormat &image_format,
  62                        const YCbCrFormat &ycbcr_format,
  63                        unsigned width, unsigned height)
  64         : image_format(image_format),
  65           ycbcr_format(ycbcr_format),
  66           finalized(false),
  67           needs_mipmaps(false),
  68           width(width),
  69           height(height),
  70           resource_pool(NULL)
  71 {
  72         pbos[0] = pbos[1] = pbos[2] = 0;
  73         texture_num[0] = texture_num[1] = texture_num[2] = 0;
  74
  75         assert(width % ycbcr_format.chroma_subsampling_x == 0);
  76         pitch[0] = widths[0] = width;
  77         pitch[1] = widths[1] = width / ycbcr_format.chroma_subsampling_x;
  78         pitch[2] = widths[2] = width / ycbcr_format.chroma_subsampling_x;
  79
  80         assert(height % ycbcr_format.chroma_subsampling_y == 0);
  81         heights[0] = height;
  82         heights[1] = height / ycbcr_format.chroma_subsampling_y;
  83         heights[2] = height / ycbcr_format.chroma_subsampling_y;
  84
  85         pixel_data[0] = pixel_data[1] = pixel_data[2] = NULL;
  86
  87         register_int("needs_mipmaps", &needs_mipmaps);
  88 }
  89
  90 YCbCrInput::~YCbCrInput()
  91 {
  92         for (unsigned channel = 0; channel < 3; ++channel) {
  93                 if (texture_num[channel] != 0) {
  94                         resource_pool->release_2d_texture(texture_num[channel]);
  95                 }
  96         }
  97 }
  98
  99 void YCbCrInput::finalize()
 100 {
 101         finalized = true;
 102 }
 103
 104 void YCbCrInput::set_gl_state(GLuint glsl_program_num, const string& prefix, unsigned *sampler_num)
 105 {
 106         for (unsigned channel = 0; channel < 3; ++channel) {
 107                 glActiveTexture(GL_TEXTURE0 + *sampler_num + channel);
 108                 check_error();
 109
 110                 if (texture_num[channel] == 0) {
 111                         // (Re-)upload the texture.
 112                         texture_num[channel] = resource_pool->create_2d_texture(GL_LUMINANCE8, widths[channel], heights[channel]);
 113                         glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 114                         check_error();
 115                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 116                         check_error();
 117                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, pbos[channel]);
 118                         check_error();
 119                         glPixelStorei(GL_UNPACK_ALIGNMENT, 1);
 120                         check_error();
 121                         glPixelStorei(GL_UNPACK_ROW_LENGTH, pitch[channel]);
 122                         check_error();
 123                         glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, widths[channel], heights[channel], GL_LUMINANCE, GL_UNSIGNED_BYTE, pixel_data[channel]);
 124                         check_error();
 125                         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
 126                         check_error();
 127                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 128                         check_error();
 129                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 130                         check_error();
 131                 } else {
 132                         glBindTexture(GL_TEXTURE_2D, texture_num[channel]);
 133                         check_error();
 134                 }
 135         }
 136
 137         glBindBuffer(GL_PIXEL_UNPACK_BUFFER_ARB, 0);
 138         check_error();
 139
 140         // Bind samplers.
 141         set_uniform_int(glsl_program_num, prefix, "tex_y", *sampler_num + 0);
 142         set_uniform_int(glsl_program_num, prefix, "tex_cb", *sampler_num + 1);
 143         set_uniform_int(glsl_program_num, prefix, "tex_cr", *sampler_num + 2);
 144
 145         *sampler_num += 3;
 146 }
 147
 148 string YCbCrInput::output_fragment_shader()
 149 {
 150         float coeff[3], offset[3], scale[3];
 151
 152         switch (ycbcr_format.luma_coefficients) {
 153         case YCBCR_REC_601:
 154                 // Rec. 601, page 2.
 155                 coeff[0] = 0.299;
 156                 coeff[1] = 0.587;
 157                 coeff[2] = 0.114;
 158                 break;
 159
 160         case YCBCR_REC_709:
 161                 // Rec. 709, page 19.
 162                 coeff[0] = 0.2126;
 163                 coeff[1] = 0.7152;
 164                 coeff[2] = 0.0722;
 165                 break;
 166
 167         case YCBCR_REC_2020:
 168                 // Rec. 2020, page 4.
 169                 coeff[0] = 0.2627;
 170                 coeff[1] = 0.6780;
 171                 coeff[2] = 0.0593;
 172                 break;
 173
 174         default:
 175                 assert(false);
 176         }
 177
 178         if (ycbcr_format.full_range) {
 179                 offset[0] = 0.0 / 255.0;
 180                 offset[1] = 128.0 / 255.0;
 181                 offset[2] = 128.0 / 255.0;
 182
 183                 scale[0] = 1.0;
 184                 scale[1] = 1.0;
 185                 scale[2] = 1.0;
 186         } else {
 187                 // Rec. 601, page 4; Rec. 709, page 19; Rec. 2020, page 4.
 188                 offset[0] = 16.0 / 255.0;
 189                 offset[1] = 128.0 / 255.0;
 190                 offset[2] = 128.0 / 255.0;
 191
 192                 scale[0] = 255.0 / 219.0;
 193                 scale[1] = 255.0 / 224.0;
 194                 scale[2] = 255.0 / 224.0;
 195         }
 196
 197         // Matrix to convert RGB to YCbCr. See e.g. Rec. 601.
 198         Matrix3d rgb_to_ycbcr;
 199         rgb_to_ycbcr(0,0) = coeff[0];
 200         rgb_to_ycbcr(0,1) = coeff[1];
 201         rgb_to_ycbcr(0,2) = coeff[2];
 202
 203         float cb_fac = (224.0 / 219.0) / (coeff[0] + coeff[1] + 1.0f - coeff[2]);
 204         rgb_to_ycbcr(1,0) = -coeff[0] * cb_fac;
 205         rgb_to_ycbcr(1,1) = -coeff[1] * cb_fac;
 206         rgb_to_ycbcr(1,2) = (1.0f - coeff[2]) * cb_fac;
 207
 208         float cr_fac = (224.0 / 219.0) / (1.0f - coeff[0] + coeff[1] + coeff[2]);
 209         rgb_to_ycbcr(2,0) = (1.0f - coeff[0]) * cr_fac;
 210         rgb_to_ycbcr(2,1) = -coeff[1] * cr_fac;
 211         rgb_to_ycbcr(2,2) = -coeff[2] * cr_fac;
 212
 213         // Inverting the matrix gives us what we need to go from YCbCr back to RGB.
 214         Matrix3d ycbcr_to_rgb = rgb_to_ycbcr.inverse();
 215
 216         string frag_shader;
 217
 218         frag_shader = output_glsl_mat3("PREFIX(inv_ycbcr_matrix)", ycbcr_to_rgb);
 219
 220         char buf[256];
 221         sprintf(buf, "const vec3 PREFIX(offset) = vec3(%.8f, %.8f, %.8f);\n",
 222                 offset[0], offset[1], offset[2]);
 223         frag_shader += buf;
 224
 225         sprintf(buf, "const vec3 PREFIX(scale) = vec3(%.8f, %.8f, %.8f);\n",
 226                 scale[0], scale[1], scale[2]);
 227         frag_shader += buf;
 228
 229         float cb_offset_x = compute_chroma_offset(
 230                 ycbcr_format.cb_x_position, ycbcr_format.chroma_subsampling_x, widths[1]);
 231         float cb_offset_y = compute_chroma_offset(
 232                 ycbcr_format.cb_y_position, ycbcr_format.chroma_subsampling_y, heights[1]);
 233         sprintf(buf, "const vec2 PREFIX(cb_offset) = vec2(%.8f, %.8f);\n",
 234                 cb_offset_x, cb_offset_y);
 235         frag_shader += buf;
 236
 237         float cr_offset_x = compute_chroma_offset(
 238                 ycbcr_format.cr_x_position, ycbcr_format.chroma_subsampling_x, widths[2]);
 239         float cr_offset_y = compute_chroma_offset(
 240                 ycbcr_format.cr_y_position, ycbcr_format.chroma_subsampling_y, heights[2]);
 241         sprintf(buf, "const vec2 PREFIX(cr_offset) = vec2(%.8f, %.8f);\n",
 242                 cr_offset_x, cr_offset_y);
 243         frag_shader += buf;
 244
 245         frag_shader += read_file("ycbcr_input.frag");
 246         return frag_shader;
 247 }
 248
 249 void YCbCrInput::invalidate_pixel_data()
 250 {
 251         for (unsigned channel = 0; channel < 3; ++channel) {
 252                 if (texture_num[channel] != 0) {
 253                         resource_pool->release_2d_texture(texture_num[channel]);
 254                         texture_num[channel] = 0;
 255                 }
 256         }
 257 }