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