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;
28 struct RansDecSymbol {
32 struct RansDecodeTable {
33 int cum2sym[prob_scale];
34 RansDecSymbol dsyms[NUM_SYMS];
36 RansDecodeTable decode_tables[NUM_TABLES];
38 optional<uint32_t> read_varint(const char **ptr, const char *end)
46 x |= (ch & 0x7f) << shift;
47 if ((ch & 0x80) == 0) return x;
50 return nullopt; // Error: Overlong int.
53 return nullopt; // Error: EOF.
57 uint src_offset, src_len;
59 CoeffStream streams[45 * 64]; // HACK
61 int main(int argc, char **argv)
63 // Set up an OpenGL context using SDL.
64 if (SDL_Init(SDL_INIT_VIDEO) == -1) {
65 fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
68 SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
69 SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
70 SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
71 SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
72 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
73 SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
75 SDL_Window *window = SDL_CreateWindow("OpenGL window for unit test",
76 SDL_WINDOWPOS_UNDEFINED,
77 SDL_WINDOWPOS_UNDEFINED,
80 SDL_GLContext context = SDL_GL_CreateContext(window);
81 assert(context != nullptr);
83 //char buf[16] = { 0 };
86 glGetIntegerv(GL_MAX_COMPUTE_SHARED_MEMORY_SIZE, &size);
87 printf("shared_memory_size=%u\n", size);
89 string shader_src = ::read_file("decoder.shader");
90 GLuint shader_num = compile_shader(shader_src, GL_COMPUTE_SHADER);
91 GLuint glsl_program_num = glCreateProgram();
92 glAttachShader(glsl_program_num, shader_num);
93 glLinkProgram(glsl_program_num);
96 glGetProgramiv(glsl_program_num, GL_LINK_STATUS, &success);
97 if (success == GL_FALSE) {
98 GLchar error_log[1024] = {0};
99 glGetProgramInfoLog(glsl_program_num, 1024, nullptr, error_log);
100 fprintf(stderr, "Error linking program: %s\n", error_log);
104 glUseProgram(glsl_program_num);
106 string coded = ::read_file(argc >= 2 ? argv[1] : "coded.dat");
107 const char *ptr = &coded[0];
108 const char *end = ptr + coded.size();
109 GLuint sign_bias[NUM_TABLES];
111 // printf("first few bytes offs=%zu: %d %d %d %d %d %d %d %d\n", ptr - coded.data(),
112 // (uint8_t)ptr[0], (uint8_t)ptr[1], (uint8_t)ptr[2], (uint8_t)ptr[3],
113 // (uint8_t)ptr[4], (uint8_t)ptr[5], (uint8_t)ptr[6], (uint8_t)ptr[7]);
115 // read the rANS tables
116 for (unsigned table = 0; table < NUM_TABLES; ++table) {
117 uint32_t cum_freq = 0;
118 for (unsigned sym = 0; sym < NUM_SYMS; ++sym) {
119 optional<uint32_t> freq = read_varint(&ptr, end);
121 fprintf(stderr, "Error parsing varint for table %d symbol %d\n", table, sym);
125 decode_tables[table].dsyms[(sym + 1) & (NUM_SYMS - 1)].sym_start = cum_freq;
126 decode_tables[table].dsyms[(sym + 1) & (NUM_SYMS - 1)].sym_freq = *freq;
127 for (uint32_t i = 0; i < freq; ++i) {
128 if (cum_freq < prob_scale)
129 decode_tables[table].cum2sym[cum_freq] = (sym + 1) & (NUM_SYMS - 1);
133 sign_bias[table] = cum_freq;
136 // Make cum2sym texture.
137 unique_ptr<uint8_t[]> cum2sym_data(new uint8_t[prob_scale * NUM_TABLES]);
138 for (unsigned table = 0; table < NUM_TABLES; ++table) {
139 for (unsigned i = 0; i < prob_scale; ++i) {
140 cum2sym_data[prob_scale * table + i] = decode_tables[table].cum2sym[i];
144 glGenTextures(1, &cum2sym_tex);
145 glBindTexture(GL_TEXTURE_2D, cum2sym_tex);
146 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
147 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
148 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
149 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
150 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8UI, prob_scale, NUM_TABLES, 0, GL_RED_INTEGER, GL_UNSIGNED_BYTE, cum2sym_data.get());
153 // Make dsyms texture.
154 unique_ptr<pair<uint16_t, uint16_t>[]> dsyms_data(new pair<uint16_t, uint16_t>[NUM_SYMS * NUM_TABLES]);
155 for (unsigned table = 0; table < NUM_TABLES; ++table) {
156 for (unsigned sym = 0; sym < NUM_SYMS; ++sym) {
157 dsyms_data[NUM_SYMS * table + sym].first = decode_tables[table].dsyms[sym].sym_start;
158 dsyms_data[NUM_SYMS * table + sym].second = decode_tables[table].dsyms[sym].sym_freq;
162 glGenTextures(1, &dsyms_tex);
163 glBindTexture(GL_TEXTURE_2D, dsyms_tex);
164 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
165 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
166 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
167 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
168 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16UI, NUM_SYMS, NUM_TABLES, 0, GL_RG_INTEGER, GL_UNSIGNED_SHORT, dsyms_data.get());
172 glGenTextures(1, &coeff_tex);
173 glBindTexture(GL_TEXTURE_2D, coeff_tex);
174 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
176 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
178 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
180 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
182 glTexImage2D(GL_TEXTURE_2D, 0, GL_R16I, WIDTH, HEIGHT, 0, GL_RED_INTEGER, GL_SHORT, nullptr);
186 glGenTextures(1, &out_tex);
187 glBindTexture(GL_TEXTURE_2D, out_tex);
188 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
189 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
190 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
191 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
192 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, WIDTH, HEIGHT, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
193 //glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, WIDTH, HEIGHT, 0, GL_RED, GL_FLOAT, nullptr);
196 GLint cum2sym_tex_pos = glGetUniformLocation(glsl_program_num, "cum2sym_tex");
197 GLint dsyms_tex_pos = glGetUniformLocation(glsl_program_num, "dsyms_tex");
198 GLint out_tex_pos = glGetUniformLocation(glsl_program_num, "out_tex");
199 GLint coeff_tex_pos = glGetUniformLocation(glsl_program_num, "coeff_tex");
200 GLint sign_bias_pos = glGetUniformLocation(glsl_program_num, "sign_bias_per_model");
201 GLint num_blocks_pos = glGetUniformLocation(glsl_program_num, "num_blocks");
202 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);
204 unsigned num_blocks = (HEIGHT / 16);
206 // Bind the textures.
207 glUniform1i(cum2sym_tex_pos, 0);
208 glUniform1i(dsyms_tex_pos, 1);
209 glUniform1i(out_tex_pos, 2);
210 glUniform1i(coeff_tex_pos, 3);
211 glUniform1uiv(sign_bias_pos, 16, sign_bias);
212 glUniform1i(num_blocks_pos, num_blocks);
213 glBindImageTexture(0, cum2sym_tex, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R8UI);
214 glBindImageTexture(1, dsyms_tex, 0, GL_FALSE, 0, GL_READ_ONLY, GL_RG16UI);
215 glBindImageTexture(2, out_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R8);
216 glBindImageTexture(3, coeff_tex, 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_R16I);
217 printf("%d err=0x%x\n", __LINE__, glGetError());
219 // Decode all luma blocks.
220 for (unsigned y = 0; y < 8; ++y) {
221 for (unsigned x = 0; x < 8; ++x) {
222 unsigned coeff_num = y * 8 + x;
224 for (unsigned yb = 0; yb < HEIGHT; yb += 16) {
225 optional<uint32_t> num_rans_bytes = read_varint(&ptr, end);
226 if (!num_rans_bytes) {
227 fprintf(stderr, "Error parsing varint for block %d rANS bytes\n", yb);
231 CoeffStream *stream = &streams[coeff_num * num_blocks + (yb/16)];
232 stream->src_offset = ptr - coded.data();
233 stream->src_len = *num_rans_bytes;
236 ptr += *num_rans_bytes;
238 //printf("read %d rANS bytes, %d sign bytes\n", *num_rans_bytes, *num_sign_bytes);
243 // put the coded data (as a whole) into an SSBO
244 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
246 GLuint ssbo_stream, ssbo, ssbo_out;
248 glGenBuffers(1, &ssbo_stream);
249 glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo_stream);
250 glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(streams), streams, GL_STREAM_DRAW);
251 glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, ssbo_stream);
252 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
254 glGenBuffers(1, &ssbo);
255 glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo);
256 glBufferData(GL_SHADER_STORAGE_BUFFER, coded.size(), coded.data(), GL_STREAM_DRAW);
257 glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 9, ssbo);
258 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
260 glGenBuffers(1, &ssbo_out);
261 glBindBuffer(GL_SHADER_STORAGE_BUFFER, ssbo_out);
262 glBufferData(GL_SHADER_STORAGE_BUFFER, 65536, nullptr, GL_STREAM_DRAW); // ??
263 glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 10, ssbo_out);
266 #define PARALLEL_SLICES 1
267 steady_clock::time_point start = steady_clock::now();
268 for (int i = 0; i < 1000; ++i) {
269 unsigned num_slices = (WIDTH/8)*(HEIGHT/8)/320;
270 glDispatchCompute(1, (num_slices+PARALLEL_SLICES-1)/PARALLEL_SLICES, 1);
274 steady_clock::time_point now = steady_clock::now();
276 unsigned *timing = (unsigned *)glMapBufferRange(GL_SHADER_STORAGE_BUFFER, 0, 65536, GL_MAP_READ_BIT);
277 //setlocale(LC_ALL, "nb_NO.UTF-8");
283 "barrier after rANS decode",
285 "barrier after horizontal IDCT",
288 "barrier after store to texture",
289 "dummy timer for overhead measurement",
292 for (int i = 0; i < 10; ++i) {
293 //printf("%d: %'18.0f [%s]\n", i, double((uint64_t(timing[i * 2 + 1]) << 32) | timing[i * 2]), phases[i].c_str());
294 printf("%d,%s", i, phases[i].c_str());
295 for (int j = 0; j < 512; ++j) {
296 int idx = (j * 10 + i) * 2;
297 uint64_t val = (uint64_t(timing[idx + 1]) << 32) | timing[idx];
298 // printf(" %'18.0f", double(val));
299 // printf(" %'6.0f", double(val) * 1e-6);
300 printf(",%.0f", double(val) * 1e-6);
303 //printf(" [%s]\n", phases[i].c_str());
307 unsigned char *data = new unsigned char[WIDTH * HEIGHT];
308 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED, GL_UNSIGNED_BYTE, data);
310 printf("%d err=0x%x bufsize=%zu\n", __LINE__, glGetError(), coded.size());
313 for (int k = 0; k < 4; ++k) {
314 for (int y = 0; y < 8; ++y) {
315 for (int x = 0; x < 8; ++x) {
316 printf("%3d ", data[y * WIDTH + x + k*8]);
324 for (int k = 0; k < 4; ++k) {
325 for (int y = 0; y < 8; ++y) {
326 for (int x = 0; x < 8; ++x) {
327 //printf("%5.2f ", data[(y+8) * WIDTH + x + (1272-k*8)]);
328 printf("%3d ", data[y * WIDTH + x + k*8]);
337 FILE *fp = fopen("narabu.pgm", "wb");
338 fprintf(fp, "P5\n%d %d\n255\n", WIDTH, HEIGHT);
339 for (int y = 0; y < HEIGHT; ++y) {
340 for (int x = 0; x < WIDTH; ++x) {
341 int k = lrintf(data[y * WIDTH + x]);
343 if (k > 255) k = 255;
349 int16_t *coeff_data = new int16_t[WIDTH * HEIGHT];
350 glBindTexture(GL_TEXTURE_2D, coeff_tex);
352 glGetTexImage(GL_TEXTURE_2D, 0, GL_RED_INTEGER, GL_SHORT, coeff_data);
354 for (int k = 0; k < 4; ++k) {
355 for (int y = 0; y < 8; ++y) {
356 for (int x = 0; x < 8; ++x) {
357 printf("%3d ", coeff_data[y * WIDTH + x + k*8]);
367 glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0); // unbind
369 printf("foo = 0x%x\n", glGetError());
370 printf("Each iteration took %.3f ms.\n", 1e3 * duration<double>(now - start).count() / 1000);