git.sesse.net Git - nageru/blob - chroma_subsampler.cpp

   1 #include "chroma_subsampler.h"
   2
   3 #include <vector>
   4
   5 #include <movit/effect_util.h>
   6 #include <movit/resource_pool.h>
   7 #include <movit/util.h>
   8
   9 using namespace movit;
  10 using namespace std;
  11
  12 ChromaSubsampler::ChromaSubsampler(ResourcePool *resource_pool)
  13         : resource_pool(resource_pool)
  14 {
  15         vector<string> frag_shader_outputs;
  16
  17         // Set up stuff for NV12 conversion.
  18         //
  19         // Note: Due to the horizontally co-sited chroma/luma samples in H.264
  20         // (chrome position is left for horizontal and center for vertical),
  21         // we need to be a bit careful in our subsampling. A diagram will make
  22         // this clearer, showing some luma and chroma samples:
  23         //
  24         //     a   b   c   d
  25         //   +---+---+---+---+
  26         //   |   |   |   |   |
  27         //   | Y | Y | Y | Y |
  28         //   |   |   |   |   |
  29         //   +---+---+---+---+
  30         //
  31         // +-------+-------+
  32         // |       |       |
  33         // |   C   |   C   |
  34         // |       |       |
  35         // +-------+-------+
  36         //
  37         // Clearly, the rightmost chroma sample here needs to be equivalent to
  38         // b/4 + c/2 + d/4. (We could also implement more sophisticated filters,
  39         // of course, but as long as the upsampling is not going to be equally
  40         // sophisticated, it's probably not worth it.) If we sample once with
  41         // no mipmapping, we get just c, ie., no actual filtering in the
  42         // horizontal direction. (For the vertical direction, we can just
  43         // sample in the middle to get the right filtering.) One could imagine
  44         // we could use mipmapping (assuming we can create mipmaps cheaply),
  45         // but then, what we'd get is this:
  46         //
  47         //    (a+b)/2 (c+d)/2
  48         //   +-------+-------+
  49         //   |       |       |
  50         //   |   Y   |   Y   |
  51         //   |       |       |
  52         //   +-------+-------+
  53         //
  54         // +-------+-------+
  55         // |       |       |
  56         // |   C   |   C   |
  57         // |       |       |
  58         // +-------+-------+
  59         //
  60         // which ends up sampling equally from a and b, which clearly isn't right. Instead,
  61         // we need to do two (non-mipmapped) chroma samples, both hitting exactly in-between
  62         // source pixels.
  63         //
  64         // Sampling in-between b and c gives us the sample (b+c)/2, and similarly for c and d.
  65         // Taking the average of these gives of (b+c)/4 + (c+d)/4 = b/4 + c/2 + d/4, which is
  66         // exactly what we want.
  67         //
  68         // See also http://www.poynton.com/PDFs/Merging_RGB_and_422.pdf, pages 6–7.
  69
  70         // Cb/Cr shader.
  71         string cbcr_vert_shader =
  72                 "#version 130 \n"
  73                 " \n"
  74                 "in vec2 position; \n"
  75                 "in vec2 texcoord; \n"
  76                 "out vec2 tc0, tc1; \n"
  77                 "uniform vec2 foo_chroma_offset_0; \n"
  78                 "uniform vec2 foo_chroma_offset_1; \n"
  79                 " \n"
  80                 "void main() \n"
  81                 "{ \n"
  82                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
  83                 "    // \n"
  84                 "    //   2.000  0.000  0.000 -1.000 \n"
  85                 "    //   0.000  2.000  0.000 -1.000 \n"
  86                 "    //   0.000  0.000 -2.000 -1.000 \n"
  87                 "    //   0.000  0.000  0.000  1.000 \n"
  88                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
  89                 "    vec2 flipped_tc = texcoord; \n"
  90                 "    tc0 = flipped_tc + foo_chroma_offset_0; \n"
  91                 "    tc1 = flipped_tc + foo_chroma_offset_1; \n"
  92                 "} \n";
  93         string cbcr_frag_shader =
  94                 "#version 130 \n"
  95                 "in vec2 tc0, tc1; \n"
  96                 "uniform sampler2D cbcr_tex; \n"
  97                 "out vec4 FragColor; \n"
  98                 "void main() { \n"
  99                 "    FragColor = 0.5 * (texture(cbcr_tex, tc0) + texture(cbcr_tex, tc1)); \n"
 100                 "} \n";
 101         cbcr_program_num = resource_pool->compile_glsl_program(cbcr_vert_shader, cbcr_frag_shader, frag_shader_outputs);
 102         check_error();
 103
 104         cbcr_texture_sampler_uniform = glGetUniformLocation(cbcr_program_num, "cbcr_tex");
 105         check_error();
 106         cbcr_position_attribute_index = glGetAttribLocation(cbcr_program_num, "position");
 107         check_error();
 108         cbcr_texcoord_attribute_index = glGetAttribLocation(cbcr_program_num, "texcoord");
 109         check_error();
 110
 111         // Same, for UYVY conversion.
 112         string uyvy_vert_shader =
 113                 "#version 130 \n"
 114                 " \n"
 115                 "in vec2 position; \n"
 116                 "in vec2 texcoord; \n"
 117                 "out vec2 y_tc0, y_tc1, cbcr_tc0, cbcr_tc1; \n"
 118                 "uniform vec2 foo_luma_offset_0; \n"
 119                 "uniform vec2 foo_luma_offset_1; \n"
 120                 "uniform vec2 foo_chroma_offset_0; \n"
 121                 "uniform vec2 foo_chroma_offset_1; \n"
 122                 " \n"
 123                 "void main() \n"
 124                 "{ \n"
 125                 "    // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is: \n"
 126                 "    // \n"
 127                 "    //   2.000  0.000  0.000 -1.000 \n"
 128                 "    //   0.000  2.000  0.000 -1.000 \n"
 129                 "    //   0.000  0.000 -2.000 -1.000 \n"
 130                 "    //   0.000  0.000  0.000  1.000 \n"
 131                 "    gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0); \n"
 132                 "    vec2 flipped_tc = texcoord; \n"
 133                 "    y_tc0 = flipped_tc + foo_luma_offset_0; \n"
 134                 "    y_tc1 = flipped_tc + foo_luma_offset_1; \n"
 135                 "    cbcr_tc0 = flipped_tc + foo_chroma_offset_0; \n"
 136                 "    cbcr_tc1 = flipped_tc + foo_chroma_offset_1; \n"
 137                 "} \n";
 138         string uyvy_frag_shader =
 139                 "#version 130 \n"
 140                 "in vec2 y_tc0, y_tc1, cbcr_tc0, cbcr_tc1; \n"
 141                 "uniform sampler2D y_tex, cbcr_tex; \n"
 142                 "out vec4 FragColor; \n"
 143                 "void main() { \n"
 144                 "    float y0 = texture(y_tex, y_tc0).r; \n"
 145                 "    float y1 = texture(y_tex, y_tc1).r; \n"
 146                 "    vec2 cbcr0 = texture(cbcr_tex, cbcr_tc0).rg; \n"
 147                 "    vec2 cbcr1 = texture(cbcr_tex, cbcr_tc1).rg; \n"
 148                 "    vec2 cbcr = 0.5 * (cbcr0 + cbcr1); \n"
 149                 "    FragColor = vec4(cbcr.g, y0, cbcr.r, y1); \n"  // FIXME: swap y0 and y1?
 150                 "} \n";
 151
 152         uyvy_program_num = resource_pool->compile_glsl_program(uyvy_vert_shader, uyvy_frag_shader, frag_shader_outputs);
 153         check_error();
 154
 155         uyvy_y_texture_sampler_uniform = glGetUniformLocation(uyvy_program_num, "y_tex");
 156         check_error();
 157         uyvy_cbcr_texture_sampler_uniform = glGetUniformLocation(uyvy_program_num, "cbcr_tex");
 158         check_error();
 159         uyvy_position_attribute_index = glGetAttribLocation(uyvy_program_num, "position");
 160         check_error();
 161         uyvy_texcoord_attribute_index = glGetAttribLocation(uyvy_program_num, "texcoord");
 162         check_error();
 163
 164         // Shared between the two.
 165         float vertices[] = {
 166                 0.0f, 2.0f,
 167                 0.0f, 0.0f,
 168                 2.0f, 0.0f
 169         };
 170         vbo = generate_vbo(2, GL_FLOAT, sizeof(vertices), vertices);
 171         check_error();
 172 }
 173
 174 ChromaSubsampler::~ChromaSubsampler()
 175 {
 176         resource_pool->release_glsl_program(cbcr_program_num);
 177         check_error();
 178         resource_pool->release_glsl_program(uyvy_program_num);
 179         check_error();
 180         glDeleteBuffers(1, &vbo);
 181         check_error();
 182 }
 183
 184 void ChromaSubsampler::subsample_chroma(GLuint cbcr_tex, unsigned width, unsigned height, GLuint dst_tex)
 185 {
 186         GLuint vao;
 187         glGenVertexArrays(1, &vao);
 188         check_error();
 189
 190         glBindVertexArray(vao);
 191         check_error();
 192
 193         // Extract Cb/Cr.
 194         GLuint fbo = resource_pool->create_fbo(dst_tex);
 195         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 196         glViewport(0, 0, width/2, height/2);
 197         check_error();
 198
 199         glUseProgram(cbcr_program_num);
 200         check_error();
 201
 202         glActiveTexture(GL_TEXTURE0);
 203         check_error();
 204         glBindTexture(GL_TEXTURE_2D, cbcr_tex);
 205         check_error();
 206         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 207         check_error();
 208         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 209         check_error();
 210         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 211         check_error();
 212
 213         float chroma_offset_0[] = { -1.0f / width, 0.0f };
 214         float chroma_offset_1[] = { -0.0f / width, 0.0f };
 215         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_0", chroma_offset_0);
 216         set_uniform_vec2(cbcr_program_num, "foo", "chroma_offset_1", chroma_offset_1);
 217
 218         glUniform1i(cbcr_texture_sampler_uniform, 0);
 219
 220         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 221         check_error();
 222
 223         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 224                 glEnableVertexAttribArray(attr_index);
 225                 check_error();
 226                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 227                 check_error();
 228         }
 229
 230         glDrawArrays(GL_TRIANGLES, 0, 3);
 231         check_error();
 232
 233         for (GLint attr_index : { cbcr_position_attribute_index, cbcr_texcoord_attribute_index }) {
 234                 glDisableVertexAttribArray(attr_index);
 235                 check_error();
 236         }
 237
 238         glUseProgram(0);
 239         check_error();
 240         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 241         check_error();
 242
 243         resource_pool->release_fbo(fbo);
 244         glDeleteVertexArrays(1, &vao);
 245         check_error();
 246 }
 247
 248 void ChromaSubsampler::create_uyvy(GLuint y_tex, GLuint cbcr_tex, unsigned width, unsigned height, GLuint dst_tex)
 249 {
 250         GLuint vao;
 251         glGenVertexArrays(1, &vao);
 252         check_error();
 253
 254         glBindVertexArray(vao);
 255         check_error();
 256
 257         GLuint fbo = resource_pool->create_fbo(dst_tex);
 258         glBindFramebuffer(GL_FRAMEBUFFER, fbo);
 259         glViewport(0, 0, width/2, height);
 260         check_error();
 261
 262         glUseProgram(uyvy_program_num);
 263         check_error();
 264
 265         glUniform1i(uyvy_y_texture_sampler_uniform, 0);
 266         check_error();
 267         glUniform1i(uyvy_cbcr_texture_sampler_uniform, 1);
 268         check_error();
 269
 270         glActiveTexture(GL_TEXTURE0);
 271         check_error();
 272         glBindTexture(GL_TEXTURE_2D, y_tex);
 273         check_error();
 274         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 275         check_error();
 276         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 277         check_error();
 278         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 279         check_error();
 280
 281         glActiveTexture(GL_TEXTURE1);
 282         check_error();
 283         glBindTexture(GL_TEXTURE_2D, cbcr_tex);
 284         check_error();
 285         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 286         check_error();
 287         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 288         check_error();
 289         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 290         check_error();
 291
 292         float y_offset_0[] = { -0.5f / width, 0.0f };
 293         float y_offset_1[] = {  0.5f / width, 0.0f };
 294         float cbcr_offset0[] = { -1.0f / width, 0.0f };
 295         float cbcr_offset1[] = { -0.0f / width, 0.0f };
 296         set_uniform_vec2(uyvy_program_num, "foo", "luma_offset_0", y_offset_0);
 297         set_uniform_vec2(uyvy_program_num, "foo", "luma_offset_1", y_offset_1);
 298         set_uniform_vec2(uyvy_program_num, "foo", "chroma_offset_0", cbcr_offset0);
 299         set_uniform_vec2(uyvy_program_num, "foo", "chroma_offset_1", cbcr_offset1);
 300
 301         glBindBuffer(GL_ARRAY_BUFFER, vbo);
 302         check_error();
 303
 304         for (GLint attr_index : { uyvy_position_attribute_index, uyvy_texcoord_attribute_index }) {
 305                 if (attr_index == -1) continue;
 306                 glEnableVertexAttribArray(attr_index);
 307                 check_error();
 308                 glVertexAttribPointer(attr_index, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
 309                 check_error();
 310         }
 311
 312         glDrawArrays(GL_TRIANGLES, 0, 3);
 313         check_error();
 314
 315         for (GLint attr_index : { uyvy_position_attribute_index, uyvy_texcoord_attribute_index }) {
 316                 if (attr_index == -1) continue;
 317                 glDisableVertexAttribArray(attr_index);
 318                 check_error();
 319         }
 320
 321         glActiveTexture(GL_TEXTURE0);
 322         check_error();
 323         glUseProgram(0);
 324         check_error();
 325         glBindFramebuffer(GL_FRAMEBUFFER, 0);
 326         check_error();
 327
 328         resource_pool->release_fbo(fbo);
 329         glDeleteVertexArrays(1, &vao);
 330 }