git.sesse.net Git - narabu/blob - narabu-encoder.cpp

   1 #include <stdio.h>
   2 #include <stdint.h>
   3 #include <stdlib.h>
   4 #include <string.h>
   5 #include <assert.h>
   6 #include <math.h>
   7 #include <SDL2/SDL.h>
   8 #include <SDL2/SDL_error.h>
   9 #include <SDL2/SDL_video.h>
  10 #include <epoxy/gl.h>
  11
  12 #include <algorithm>
  13 #include <chrono>
  14 #include <memory>
  15 #include <numeric>
  16 #include <random>
  17 #include <vector>
  18 #include <unordered_map>
  19
  20 #include <movit/util.h>
  21
  22 #include "util.h"
  23
  24 #define WIDTH 1280
  25 #define HEIGHT 720
  26 #define WIDTH_BLOCKS (WIDTH/8)
  27 #define WIDTH_BLOCKS_CHROMA (WIDTH/16)
  28 #define HEIGHT_BLOCKS (HEIGHT/8)
  29 #define NUM_BLOCKS (WIDTH_BLOCKS * HEIGHT_BLOCKS)
  30 #define NUM_BLOCKS_CHROMA (WIDTH_BLOCKS_CHROMA * HEIGHT_BLOCKS)
  31
  32 #define NUM_SYMS 256
  33 #define ESCAPE_LIMIT (NUM_SYMS - 1)
  34 #define BLOCKS_PER_STREAM 320
  35
  36 static constexpr uint32_t prob_bits = 12;
  37 static constexpr uint32_t prob_scale = 1 << prob_bits;
  38
  39 unsigned char rgb[WIDTH * HEIGHT * 3];
  40 unsigned char pix_y[WIDTH * HEIGHT];
  41 unsigned char pix_cb[(WIDTH/2) * HEIGHT];
  42 unsigned char pix_cr[(WIDTH/2) * HEIGHT];
  43
  44 struct RansCountSSBO {
  45         unsigned dist[4 * 256];
  46         unsigned ransfreq[4 * 256];
  47 };
  48
  49 struct RansDistUBO {
  50         struct {
  51                 uint32_t x_max, rcp_freq, bias, rcp_shift_and_cmpl_freq;
  52         } ransdist[4 * 256];
  53         struct {
  54                 uint32_t val;
  55                 uint32_t padding[3];  // std140 layout.
  56         } sign_biases[4];
  57 };
  58
  59 using namespace std;
  60 using namespace std::chrono;
  61
  62 void write_varint(int x, FILE *fp)
  63 {
  64         while (x >= 128) {
  65                 putc((x & 0x7f) | 0x80, fp);
  66                 x >>= 7;
  67         }
  68         putc(x, fp);
  69 }
  70
  71 void readpix(unsigned char *ptr, const char *filename)
  72 {
  73         FILE *fp = fopen(filename, "rb");
  74         if (fp == nullptr) {
  75                 perror(filename);
  76                 exit(1);
  77         }
  78
  79         fseek(fp, 0, SEEK_END);
  80         long len = ftell(fp);
  81         assert(len >= WIDTH * HEIGHT * 3);
  82         fseek(fp, len - WIDTH * HEIGHT * 3, SEEK_SET);
  83
  84         fread(ptr, 1, WIDTH * HEIGHT * 3, fp);
  85         fclose(fp);
  86 }
  87
  88 // Should be done on the GPU, of course, but irrelevant for the demonstration.
  89 void convert_ycbcr()
  90 {
  91         double coeff[3] = { 0.2126, 0.7152, 0.0722 };  // sum = 1.0
  92         double cb_fac = 1.0 / (coeff[0] + coeff[1] + 1.0f - coeff[2]);  // 0.539
  93         double cr_fac = 1.0 / (1.0f - coeff[0] + coeff[1] + coeff[2]);  // 0.635
  94
  95         unique_ptr<float[]> temp_cb(new float[WIDTH * HEIGHT]);
  96         unique_ptr<float[]> temp_cr(new float[WIDTH * HEIGHT]);
  97         for (unsigned yb = 0; yb < HEIGHT; ++yb) {
  98                 for (unsigned xb = 0; xb < WIDTH; ++xb) {
  99                         int r = rgb[((yb * WIDTH) + xb) * 3 + 0];
 100                         int g = rgb[((yb * WIDTH) + xb) * 3 + 1];
 101                         int b = rgb[((yb * WIDTH) + xb) * 3 + 2];
 102                         double y = std::min(std::max(coeff[0] * r + coeff[1] * g + coeff[2] * b, 0.0), 255.0);
 103                         double cb = (b - y) * cb_fac + 128.0;
 104                         double cr = (r - y) * cr_fac + 128.0;
 105                         pix_y[(yb * WIDTH) + xb] = lrint(y);
 106                         temp_cb[(yb * WIDTH) + xb] = cb;
 107                         temp_cr[(yb * WIDTH) + xb] = cr;
 108                 }
 109         }
 110
 111         // Simple 4:2:2 subsampling with left convention.
 112         for (unsigned yb = 0; yb < HEIGHT; ++yb) {
 113                 for (unsigned xb = 0; xb < WIDTH / 2; ++xb) {
 114                         int c0 = yb * WIDTH + std::max(int(xb) * 2 - 1, 0);
 115                         int c1 = yb * WIDTH + xb * 2;
 116                         int c2 = yb * WIDTH + xb * 2 + 1;
 117
 118                         double cb = 0.25 * temp_cb[c0] + 0.5 * temp_cb[c1] + 0.25 * temp_cb[c2];
 119                         double cr = 0.25 * temp_cr[c0] + 0.5 * temp_cr[c1] + 0.25 * temp_cr[c2];
 120                         cb = std::min(std::max(cb, 0.0), 255.0);
 121                         cr = std::min(std::max(cr, 0.0), 255.0);
 122                         pix_cb[(yb * WIDTH/2) + xb] = lrint(cb);
 123                         pix_cr[(yb * WIDTH/2) + xb] = lrint(cr);
 124                 }
 125         }
 126 }
 127
 128 int main(int argc, char **argv)
 129 {
 130         // Set up an OpenGL context using SDL.
 131         if (SDL_Init(SDL_INIT_VIDEO) == -1) {
 132                 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
 133                 exit(1);
 134         }
 135         SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
 136         SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
 137         SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
 138         SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
 139         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
 140         SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
 141
 142         SDL_Window *window = SDL_CreateWindow("OpenGL window for unit test",
 143                 SDL_WINDOWPOS_UNDEFINED,
 144                 SDL_WINDOWPOS_UNDEFINED,
 145                 32, 32,
 146                 SDL_WINDOW_OPENGL);
 147         SDL_GLContext context = SDL_GL_CreateContext(window);
 148         assert(context != nullptr);
 149
 150         if (argc >= 2)
 151                 readpix(rgb, argv[1]);
 152         else
 153                 readpix(rgb, "color.pnm");
 154         convert_ycbcr();
 155
 156         // Compile the DCT shader.
 157         string shader_src = ::read_file("encoder.shader");
 158         GLuint shader_num = compile_shader(shader_src, GL_COMPUTE_SHADER);
 159         GLuint glsl_program_num = glCreateProgram();
 160         glAttachShader(glsl_program_num, shader_num);
 161         glLinkProgram(glsl_program_num);
 162
 163         GLint success;
 164         glGetProgramiv(glsl_program_num, GL_LINK_STATUS, &success);
 165         if (success == GL_FALSE) {
 166                 GLchar error_log[1024] = {0};
 167                 glGetProgramInfoLog(glsl_program_num, 1024, nullptr, error_log);
 168                 fprintf(stderr, "Error linking program: %s\n", error_log);
 169                 exit(1);
 170         }
 171
 172         // Compile the tally shader.
 173         shader_src = ::read_file("tally.shader");
 174         shader_num = compile_shader(shader_src, GL_COMPUTE_SHADER);
 175         GLuint glsl_tally_program_num = glCreateProgram();
 176         glAttachShader(glsl_tally_program_num, shader_num);
 177         glLinkProgram(glsl_tally_program_num);
 178
 179         glGetProgramiv(glsl_tally_program_num, GL_LINK_STATUS, &success);
 180         if (success == GL_FALSE) {
 181                 GLchar error_log[1024] = {0};
 182                 glGetProgramInfoLog(glsl_tally_program_num, 1024, nullptr, error_log);
 183                 fprintf(stderr, "Error linking program: %s\n", error_log);
 184                 exit(1);
 185         }
 186
 187         // Compile the rANS shader.
 188         shader_src = ::read_file("rans.shader");
 189         shader_num = compile_shader(shader_src, GL_COMPUTE_SHADER);
 190         GLuint glsl_rans_program_num = glCreateProgram();
 191         glAttachShader(glsl_rans_program_num, shader_num);
 192         glLinkProgram(glsl_rans_program_num);
 193
 194         glGetProgramiv(glsl_rans_program_num, GL_LINK_STATUS, &success);
 195         if (success == GL_FALSE) {
 196                 GLchar error_log[1024] = {0};
 197                 glGetProgramInfoLog(glsl_rans_program_num, 1024, nullptr, error_log);
 198                 fprintf(stderr, "Error linking program: %s\n", error_log);
 199                 exit(1);
 200         }
 201         check_error();
 202
 203         // An SSBO for the raw rANS counts.
 204         GLuint ssbo;
 205         glGenBuffers(1, &ssbo);
 206         glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
 207         glNamedBufferStorage(ssbo, sizeof(RansCountSSBO), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 208         check_error();
 209
 210         // UBO for the rANS distributions (copied from an SSBO).
 211         GLuint dist_ssbo;
 212         glGenBuffers(1, &dist_ssbo);
 213         glBindBuffer(GL_SHADER_STORAGE_BUFFER, dist_ssbo);
 214         glNamedBufferStorage(dist_ssbo, sizeof(RansDistUBO), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 215         check_error();
 216
 217         GLuint dist_ubo;
 218         glGenBuffers(1, &dist_ubo);
 219         glBindBuffer(GL_UNIFORM_BUFFER, dist_ubo);
 220         glNamedBufferStorage(dist_ubo, sizeof(RansDistUBO), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 221         check_error();
 222
 223         // SSBOs for the rANS output (data and offsets).
 224         GLuint output_ssbo;
 225         glGenBuffers(1, &output_ssbo);
 226         glBindBuffer(GL_SHADER_STORAGE_BUFFER, output_ssbo);
 227         glNamedBufferStorage(output_ssbo, 45 * 64 * 1024, nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 228         check_error();
 229
 230         GLuint output_offset_ssbo;
 231         glGenBuffers(1, &output_offset_ssbo);
 232         glBindBuffer(GL_SHADER_STORAGE_BUFFER, output_offset_ssbo);
 233         glNamedBufferStorage(output_offset_ssbo, 45 * 64 * sizeof(uint32_t), nullptr, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 234         check_error();
 235
 236         // Bind SSBOs.
 237         glUseProgram(glsl_program_num);
 238         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 9, ssbo);
 239
 240         glUseProgram(glsl_tally_program_num);
 241         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 9, ssbo);
 242         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 12, dist_ssbo);
 243
 244         glUseProgram(glsl_rans_program_num);
 245         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 10, output_ssbo);
 246         glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 11, output_offset_ssbo);
 247         glBindBufferBase(GL_UNIFORM_BUFFER, 13, dist_ubo);
 248
 249         glUseProgram(glsl_program_num);
 250         check_error();
 251
 252         // Upload luma.
 253         GLuint y_tex;
 254         glGenTextures(1, &y_tex);
 255         glBindTexture(GL_TEXTURE_2D, y_tex);
 256         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 257         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 258         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 259         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
 260         glTexImage2D(GL_TEXTURE_2D, 0, GL_R8UI, WIDTH, HEIGHT, 0, GL_RED_INTEGER, GL_UNSIGNED_BYTE, pix_y);
 261         check_error();
 262
 263         // Make destination textures.
 264         GLuint dc_ac7_tex, ac1_ac6_tex, ac2_ac5_tex;
 265         for (GLuint *tex : { &dc_ac7_tex, &ac1_ac6_tex, &ac2_ac5_tex }) {
 266                 glGenTextures(1, tex);
 267                 glBindTexture(GL_TEXTURE_2D, *tex);
 268                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 269                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 270                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 271                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
 272                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R16UI, WIDTH / 8, HEIGHT, 0, GL_RED_INTEGER, GL_UNSIGNED_SHORT, nullptr);
 273                 check_error();
 274         }
 275
 276         GLuint ac3_tex, ac4_tex;
 277         for (GLuint *tex : { &ac3_tex, &ac4_tex }) {
 278                 glGenTextures(1, tex);
 279                 glBindTexture(GL_TEXTURE_2D, *tex);
 280                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
 281                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
 282                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
 283                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
 284                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8I, WIDTH / 8, HEIGHT, 0, GL_RED_INTEGER, GL_BYTE, nullptr);
 285                 check_error();
 286         }
 287
 288         glBindImageTexture(0, dc_ac7_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R16UI);
 289         glBindImageTexture(1, ac1_ac6_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R16UI);
 290         glBindImageTexture(2, ac2_ac5_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R16UI);
 291         glBindImageTexture(3, ac3_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R8I);
 292         glBindImageTexture(4, ac4_tex, 0, GL_FALSE, 0, GL_READ_WRITE, GL_R8I);
 293         glBindImageTexture(5, y_tex, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8UI);
 294         check_error();
 295
 296         // Bind uniforms.
 297         glUseProgram(glsl_program_num);
 298         GLint dc_ac7_tex_uniform = glGetUniformLocation(glsl_program_num, "dc_ac7_tex");
 299         GLint ac1_ac6_tex_uniform = glGetUniformLocation(glsl_program_num, "ac1_ac6_tex");
 300         GLint ac2_ac5_tex_uniform = glGetUniformLocation(glsl_program_num, "ac2_ac5_tex");
 301         GLint ac3_tex_uniform = glGetUniformLocation(glsl_program_num, "ac3_tex");
 302         GLint ac4_tex_uniform = glGetUniformLocation(glsl_program_num, "ac4_tex");
 303         GLint image_tex_uniform = glGetUniformLocation(glsl_program_num, "image_tex");
 304         glUniform1i(dc_ac7_tex_uniform, 0);
 305         glUniform1i(ac1_ac6_tex_uniform, 1);
 306         glUniform1i(ac2_ac5_tex_uniform, 2);
 307         glUniform1i(ac3_tex_uniform, 3);
 308         glUniform1i(ac4_tex_uniform, 4);
 309         glUniform1i(image_tex_uniform, 5);
 310
 311         glUseProgram(glsl_rans_program_num);
 312         dc_ac7_tex_uniform = glGetUniformLocation(glsl_rans_program_num, "dc_ac7_tex");
 313         ac1_ac6_tex_uniform = glGetUniformLocation(glsl_rans_program_num, "ac1_ac6_tex");
 314         ac2_ac5_tex_uniform = glGetUniformLocation(glsl_rans_program_num, "ac2_ac5_tex");
 315         ac3_tex_uniform = glGetUniformLocation(glsl_rans_program_num, "ac3_tex");
 316         ac4_tex_uniform = glGetUniformLocation(glsl_rans_program_num, "ac4_tex");
 317         image_tex_uniform = glGetUniformLocation(glsl_rans_program_num, "image_tex");
 318         glUniform1i(dc_ac7_tex_uniform, 0);
 319         glUniform1i(ac1_ac6_tex_uniform, 1);
 320         glUniform1i(ac2_ac5_tex_uniform, 2);
 321         glUniform1i(ac3_tex_uniform, 3);
 322         glUniform1i(ac4_tex_uniform, 4);
 323
 324         steady_clock::time_point start = steady_clock::now();
 325         unsigned num_iterations = 100;
 326         for (unsigned i = 0; i < num_iterations; ++i) {
 327                 glClearNamedBufferSubData(ssbo, GL_R8, 0, sizeof(RansCountSSBO), GL_RED, GL_UNSIGNED_BYTE, nullptr);
 328                 glUseProgram(glsl_program_num);
 329                 glDispatchCompute(WIDTH_BLOCKS / 16, HEIGHT_BLOCKS, 1);
 330                 glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
 331
 332                 glUseProgram(glsl_tally_program_num);
 333                 glDispatchCompute(4, 1, 1);
 334                 glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
 335
 336                 glCopyNamedBufferSubData(dist_ssbo, dist_ubo, 0, 0, sizeof(RansDistUBO));
 337                 glMemoryBarrier(GL_UNIFORM_BARRIER_BIT);
 338
 339                 glUseProgram(glsl_rans_program_num);
 340                 glDispatchCompute(NUM_BLOCKS / BLOCKS_PER_STREAM, 8, 5);
 341         }
 342         check_error();
 343         glFinish();
 344         check_error();
 345         steady_clock::time_point now = steady_clock::now();
 346
 347 #if 0
 348         printf("%ld bytes + %ld escape bits (%ld) = %ld total bytes\n",
 349                 tot_bytes - extra_bits / 8,
 350                 extra_bits,
 351                 extra_bits / 8,
 352                 tot_bytes);
 353
 354         printf("\n");
 355 #endif
 356
 357         printf("Each iteration took %.3f ms.\n", 1e3 * duration<double>(now - start).count() / num_iterations);
 358
 359         FILE *codedfp = fopen("coded.dat", "wb");
 360         if (codedfp == nullptr) {
 361                 perror("coded.dat");
 362                 exit(1);
 363         }
 364
 365         // Write out the distributions.
 366         const RansCountSSBO *rans_count = (const RansCountSSBO *)glMapNamedBufferRange(ssbo, 0, sizeof(RansCountSSBO), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 367         const RansDistUBO *rans_dist = (const RansDistUBO *)glMapNamedBufferRange(dist_ssbo, 0, sizeof(RansDistUBO), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 368         for (unsigned r = 0; r < 2; ++r) {  // Hack to write fake chroma tables.
 369                 // TODO: rather gamma-k or something
 370                 for (unsigned i = 0; i < 4; ++i) {
 371                         printf("writing table %d\n", i);
 372                         for (unsigned j = 0; j < NUM_SYMS; ++j) {
 373                                 printf("%d,%d: freq=%d  x_max=%d, rcp_freq=%08x, bias=%d, rcp_shift=%d, cmpl_freq=%d\n",
 374                                         i, j, rans_count->ransfreq[i * 256 + j],
 375                                         rans_dist->ransdist[i * 256 + j].x_max,
 376                                         rans_dist->ransdist[i * 256 + j].rcp_freq,
 377                                         rans_dist->ransdist[i * 256 + j].bias,
 378                                         rans_dist->ransdist[i * 256 + j].rcp_shift_and_cmpl_freq & 0xffff,
 379                                         rans_dist->ransdist[i * 256 + j].rcp_shift_and_cmpl_freq >> 16);
 380                                 write_varint(rans_count->ransfreq[i * 256 + j], codedfp);
 381                         }
 382                 }
 383         }
 384
 385         // Write out the actual data.
 386         // TODO: Do the deduplication.
 387
 388         const uint32_t *offsets = (const uint32_t *)glMapNamedBufferRange(output_offset_ssbo, 0, 45 * 64 * sizeof(uint32_t), GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 389 #if 0
 390         for (int i = 0; i < 45*64; ++i) {
 391                 printf("%d,%d,%d: %u\n", i / 64, (i / 8) % 8, i % 8, 1024 * (i + 1) - offsets[i]);
 392         }
 393 #endif
 394
 395         const uint8_t *data = (const uint8_t *)glMapNamedBufferRange(output_ssbo, 0, 45 * 64 * 1024, GL_MAP_READ_BIT | GL_MAP_PERSISTENT_BIT);
 396
 397         string last_block;
 398         for (unsigned y = 0; y < 8; ++y) {
 399                 for (unsigned x = 0; x < 8; ++x) {
 400                         for (unsigned int stream_idx = 0; stream_idx < 45; ++stream_idx) {
 401                                 const uint8_t *out_end = data + (stream_idx * 64 + y * 8 + x + 1) * 1024;
 402                                 const uint8_t *ptr = data + offsets[stream_idx * 64 + y * 8 + x];
 403                                 uint32_t num_rans_bytes = out_end - ptr;
 404                                 assert(num_rans_bytes <= 1024);
 405
 406                                 if (num_rans_bytes == last_block.size() &&
 407                                     memcmp(last_block.data(), ptr, last_block.size()) == 0) {
 408                                         write_varint(0, codedfp);
 409                                 } else {
 410                                         last_block = string((const char *)ptr, num_rans_bytes);
 411                                         write_varint(num_rans_bytes, codedfp);
 412                                         fwrite(ptr, 1, num_rans_bytes, codedfp);
 413                                 }
 414                         }
 415                 }
 416         }
 417         fclose(codedfp);
 418 }