4 #include <SDL2/SDL_error.h>
5 #include <SDL2/SDL_video.h>
7 #include <movit/util.h>
18 using namespace std::chrono;
23 const unsigned prob_bits = 12;
24 const unsigned prob_scale = 1 << prob_bits;
25 const unsigned NUM_SYMS = 256;
26 const unsigned NUM_TABLES = 8;
27 const unsigned BLOCKS_PER_STREAM = 320;
29 struct RansDecSymbol {
33 struct RansDecodeTable {
34 int cum2sym[prob_scale];
35 RansDecSymbol dsyms[NUM_SYMS];
37 RansDecodeTable decode_tables[NUM_TABLES];
39 optional<uint32_t> read_varint(const char **ptr, const char *end)
47 x |= (ch & 0x7f) << shift;
48 if ((ch & 0x80) == 0) return x;
51 return nullopt; // Error: Overlong int.
54 return nullopt; // Error: EOF.
58 uint src_offset, src_len;
60 CoeffStream streams[45 * 64]; // HACK
62 int main(int argc, char **argv)
64 // Set up an OpenGL context using SDL.
65 if (SDL_Init(SDL_INIT_VIDEO) == -1) {
66 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
69 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
70 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
71 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
72 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
73 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
74 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
76 SDL_Window *window = SDL_CreateWindow("OpenGL window for unit test",
77 SDL_WINDOWPOS_UNDEFINED,
78 SDL_WINDOWPOS_UNDEFINED,
81 SDL_GLContext context = SDL_GL_CreateContext(window);
82 assert(context != nullptr);
84 //char buf[16] = { 0 };
87 glGetIntegerv(GL_MAX_COMPUTE_SHARED_MEMORY_SIZE, &size);
88 printf("shared_memory_size=%u\n", size);
90 string shader_src = ::read_file("decoder.shader");
91 GLuint shader_num = compile_shader(shader_src, GL_COMPUTE_SHADER);
92 GLuint glsl_program_num = glCreateProgram();
93 glAttachShader(glsl_program_num, shader_num);
94 glLinkProgram(glsl_program_num);
97 glGetProgramiv(glsl_program_num, GL_LINK_STATUS, &success);
98 if (success == GL_FALSE) {
99 GLchar error_log[1024] = {0};
100 glGetProgramInfoLog(glsl_program_num, 1024, nullptr, error_log);
101 fprintf(stderr, "Error linking program: %s\n", error_log);
105 glUseProgram(glsl_program_num);
107 string coded = ::read_file(argc >= 2 ? argv[1] : "coded.dat");
108 const char *ptr = &coded[0];
109 const char *end = ptr + coded.size();
110 GLuint sign_bias[NUM_TABLES];
112 // printf("first few bytes offs=%zu: %d %d %d %d %d %d %d %d\n", ptr - coded.data(),
113 // (uint8_t)ptr[0], (uint8_t)ptr[1], (uint8_t)ptr[2], (uint8_t)ptr[3],
114 // (uint8_t)ptr[4], (uint8_t)ptr[5], (uint8_t)ptr[6], (uint8_t)ptr[7]);
116 // read the rANS tables
117 for (unsigned table = 0; table < NUM_TABLES; ++table) {
118 uint32_t cum_freq = 0;
119 for (unsigned sym = 0; sym < NUM_SYMS; ++sym) {
120 optional<uint32_t> freq = read_varint(&ptr, end);
122 fprintf(stderr, "Error parsing varint for table %d symbol %d\n", table, sym);
126 decode_tables[table].dsyms[(sym + 1) & (NUM_SYMS - 1)].sym_start = cum_freq;
127 decode_tables[table].dsyms[(sym + 1) & (NUM_SYMS - 1)].sym_freq = *freq;
128 for (uint32_t i = 0; i < freq; ++i) {
129 if (cum_freq < prob_scale)
130 decode_tables[table].cum2sym[cum_freq] = (sym + 1) & (NUM_SYMS - 1);
134 sign_bias[table] = cum_freq;
137 // Make cum2sym texture.
138 unique_ptr<uint8_t[]> cum2sym_data(new uint8_t[prob_scale * NUM_TABLES]);
139 for (unsigned table = 0; table < NUM_TABLES; ++table) {
140 for (unsigned i = 0; i < prob_scale; ++i) {
141 cum2sym_data[prob_scale * table + i] = decode_tables[table].cum2sym[i];
145 glGenTextures(1, &cum2sym_tex);
146 glBindTexture(GL_TEXTURE_2D, cum2sym_tex);
147 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
148 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
149 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
150 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
151 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8UI, prob_scale, NUM_TABLES, 0, GL_RED_INTEGER, GL_UNSIGNED_BYTE, cum2sym_data.get());
154 // Make dsyms texture.
155 unique_ptr<pair<uint16_t, uint16_t>[]> dsyms_data(new pair<uint16_t, uint16_t>[NUM_SYMS * NUM_TABLES]);
156 for (unsigned table = 0; table < NUM_TABLES; ++table) {
157 for (unsigned sym = 0; sym < NUM_SYMS; ++sym) {
158 dsyms_data[NUM_SYMS * table + sym].first = decode_tables[table].dsyms[sym].sym_start;
159 dsyms_data[NUM_SYMS * table + sym].second = decode_tables[table].dsyms[sym].sym_freq;
163 glGenTextures(1, &dsyms_tex);
164 glBindTexture(GL_TEXTURE_2D, dsyms_tex);
165 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
166 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
167 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
168 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
169 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16UI, NUM_SYMS, NUM_TABLES, 0, GL_RG_INTEGER, GL_UNSIGNED_SHORT, dsyms_data.get());
173 glGenTextures(1, &coeff_tex);
174 glBindTexture(GL_TEXTURE_2D, coeff_tex);
175 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
177 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
179 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
181 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
183 glTexImage2D(GL_TEXTURE_2D, 0, GL_R16I, WIDTH, HEIGHT, 0, GL_RED_INTEGER, GL_SHORT, nullptr);
187 glGenTextures(1, &out_tex);
188 glBindTexture(GL_TEXTURE_2D, out_tex);
189 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
190 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
191 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
192 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
193 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, WIDTH, HEIGHT, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
194 //glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, WIDTH, HEIGHT, 0, GL_RED, GL_FLOAT, nullptr);
197 GLint cum2sym_tex_pos = glGetUniformLocation(glsl_program_num, "cum2sym_tex");
198 GLint dsyms_tex_pos = glGetUniformLocation(glsl_program_num, "dsyms_tex");
199 GLint out_tex_pos = glGetUniformLocation(glsl_program_num, "out_tex");
200 GLint coeff_tex_pos = glGetUniformLocation(glsl_program_num, "coeff_tex");
201 GLint sign_bias_pos = glGetUniformLocation(glsl_program_num, "sign_bias_per_model");
202 GLint num_blocks_pos = glGetUniformLocation(glsl_program_num, "num_blocks");
203 printf("%d err=0x%x pos=%d,%d,%d,%d\n", __LINE__, glGetError(), cum2sym_tex_pos, dsyms_tex_pos, out_tex_pos, sign_bias_pos);
205 unsigned num_blocks = (HEIGHT / 16);
207 // Bind the textures.
208 glUniform1i(cum2sym_tex_pos, 0);
209 glUniform1i(dsyms_tex_pos, 1);
210 glUniform1i(out_tex_pos, 2);
211 glUniform1i(coeff_tex_pos, 3);
212 glUniform1uiv(sign_bias_pos, 16, sign_bias);
213 glUniform1i(num_blocks_pos, num_blocks);
214 glBindImageTexture(0, cum2sym_tex, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8UI);
215 glBindImageTexture(1, dsyms_tex, 0, GL_FALSE, 0, GL_READ_ONLY, GL_RG16UI);
216 glBindImageTexture(2, out_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R8);
217 glBindImageTexture(3, coeff_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R16I);
218 printf("%d err=0x%x\n", __LINE__, glGetError());
220 // Decode all luma blocks.
221 for (unsigned y = 0; y < 8; ++y) {
222 for (unsigned x = 0; x < 8; ++x) {
223 unsigned coeff_num = y * 8 + x;
225 for (unsigned yb = 0; yb < HEIGHT; yb += 16) {
226 optional<uint32_t> num_rans_bytes = read_varint(&ptr, end);
227 if (!num_rans_bytes) {
228 fprintf(stderr, "Error parsing varint for block %d rANS bytes\n", yb);
232 CoeffStream *stream = &streams[coeff_num * num_blocks + (yb/16)];
233 stream->src_offset = ptr - coded.data();
234 stream->src_len = *num_rans_bytes;
237 ptr += *num_rans_bytes;
239 //printf("read %d rANS bytes, %d sign bytes\n", *num_rans_bytes, *num_sign_bytes);
244 // put the coded data (as a whole) into an SSBO
245 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
247 GLuint ssbo_stream, ssbo, ssbo_out;
249 glGenBuffers(1, &ssbo_stream);
250 glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo_stream);
251 glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(streams), streams, GL_STREAM_DRAW);
252 glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, ssbo_stream);
253 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
255 glGenBuffers(1, &ssbo);
256 glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
257 glBufferData(GL_SHADER_STORAGE_BUFFER, coded.size(), coded.data(), GL_STREAM_DRAW);
258 glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 9, ssbo);
259 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
261 glGenBuffers(1, &ssbo_out);
262 glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo_out);
263 glBufferData(GL_SHADER_STORAGE_BUFFER, 65536, nullptr, GL_STREAM_DRAW); // ??
264 glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 10, ssbo_out);
267 #define PARALLEL_SLICES 1
268 steady_clock::time_point start = steady_clock::now();
269 for (int i = 0; i < 1000; ++i) {
270 unsigned num_slices = (WIDTH/8)*(HEIGHT/8)/BLOCKS_PER_STREAM;
271 glDispatchCompute(1, (num_slices+PARALLEL_SLICES-1)/PARALLEL_SLICES, 1);
275 steady_clock::time_point now = steady_clock::now();
277 unsigned *timing = (unsigned *)glMapBufferRange(GL_SHADER_STORAGE_BUFFER, 0, 65536, GL_MAP_READ_BIT);
278 //setlocale(LC_ALL, "nb_NO.UTF-8");
284 "barrier after rANS decode",
286 "barrier after horizontal IDCT",
289 "barrier after store to texture",
290 "dummy timer for overhead measurement",
293 for (int i = 0; i < 10; ++i) {
294 //printf("%d: %'18.0f [%s]\n", i, double((uint64_t(timing[i * 2 + 1]) << 32) | timing[i * 2]), phases[i].c_str());
295 printf("%d,%s", i, phases[i].c_str());
296 for (int j = 0; j < 512; ++j) {
297 int idx = (j * 10 + i) * 2;
298 uint64_t val = (uint64_t(timing[idx + 1]) << 32) | timing[idx];
299 // printf(" %'18.0f", double(val));
300 // printf(" %'6.0f", double(val) * 1e-6);
301 printf(",%.0f", double(val) * 1e-6);
304 //printf(" [%s]\n", phases[i].c_str());
308 unsigned char *data = new unsigned char[WIDTH * HEIGHT];
309 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, data);
311 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
314 for (int k = 0; k < 4; ++k) {
315 for (int y = 0; y < 8; ++y) {
316 for (int x = 0; x < 8; ++x) {
317 printf("%3d ", data[y * WIDTH + x + k*8]);
325 for (int k = 0; k < 4; ++k) {
326 for (int y = 0; y < 8; ++y) {
327 for (int x = 0; x < 8; ++x) {
328 //printf("%5.2f ", data[(y+8) * WIDTH + x + (1272-k*8)]);
329 printf("%3d ", data[y * WIDTH + x + k*8]);
338 FILE *fp = fopen("narabu.pgm", "wb");
339 fprintf(fp, "P5\n%d %d\n255\n", WIDTH, HEIGHT);
340 for (int y = 0; y < HEIGHT; ++y) {
341 for (int x = 0; x < WIDTH; ++x) {
342 int k = lrintf(data[y * WIDTH + x]);
344 if (k > 255) k = 255;
350 int16_t *coeff_data = new int16_t[WIDTH * HEIGHT];
351 glBindTexture(GL_TEXTURE_2D, coeff_tex);
353 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED_INTEGER, GL_SHORT, coeff_data);
355 for (int k = 0; k < 4; ++k) {
356 for (int y = 0; y < 8; ++y) {
357 for (int x = 0; x < 8; ++x) {
358 printf("%3d ", coeff_data[y * WIDTH + x + k*8]);
368 glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0); // unbind
370 printf("foo = 0x%x\n", glGetError());
371 printf("Each iteration took %.3f ms.\n", 1e3 * duration<double>(now - start).count() / 1000);