| 1 | // basisu_frontend.cpp |
|---|---|
| 2 | // Copyright (C) 2019-2021 Binomial LLC. All Rights Reserved. |
| 3 | // |
| 4 | // Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | // you may not use this file except in compliance with the License. |
| 6 | // You may obtain a copy of the License at |
| 7 | // |
| 8 | // http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | // |
| 10 | // Unless required by applicable law or agreed to in writing, software |
| 11 | // distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | // See the License for the specific language governing permissions and |
| 14 | // limitations under the License. |
| 15 | // |
| 16 | // TODO: |
| 17 | // This code originally supported full ETC1 and ETC1S, so there's some legacy stuff to be cleaned up in here. |
| 18 | // Add endpoint tiling support (where we force adjacent blocks to use the same endpoints during quantization), for a ~10% or more increase in bitrate at same SSIM. The backend already supports this. |
| 19 | // |
| 20 | #include "../transcoder/basisu.h" |
| 21 | #include "basisu_frontend.h" |
| 22 | #include "basisu_opencl.h" |
| 23 | #include <unordered_set> |
| 24 | #include <unordered_map> |
| 25 | |
| 26 | #if BASISU_SUPPORT_SSE |
| 27 | #define CPPSPMD_NAME(a) a##_sse41 |
| 28 | #include "basisu_kernels_declares.h" |
| 29 | #endif |
| 30 | |
| 31 | #define BASISU_FRONTEND_VERIFY(c) do { if (!(c)) handle_verify_failure(__LINE__); } while(0) |
| 32 | |
| 33 | namespace basisu |
| 34 | { |
| 35 | const uint32_t cMaxCodebookCreationThreads = 8; |
| 36 | |
| 37 | const uint32_t BASISU_MAX_ENDPOINT_REFINEMENT_STEPS = 3; |
| 38 | //const uint32_t BASISU_MAX_SELECTOR_REFINEMENT_STEPS = 3; |
| 39 | |
| 40 | const uint32_t BASISU_ENDPOINT_PARENT_CODEBOOK_SIZE = 16; |
| 41 | const uint32_t BASISU_SELECTOR_PARENT_CODEBOOK_SIZE_COMP_LEVEL_01 = 32; |
| 42 | const uint32_t BASISU_SELECTOR_PARENT_CODEBOOK_SIZE_COMP_LEVEL_DEFAULT = 16; |
| 43 | |
| 44 | // TODO - How to handle internal verifies in the basisu lib |
| 45 | static inline void handle_verify_failure(int line) |
| 46 | { |
| 47 | error_printf("basisu_frontend: verify check failed at line %i!\n", line); |
| 48 | abort(); |
| 49 | } |
| 50 | |
| 51 | bool basisu_frontend::init(const params &p) |
| 52 | { |
| 53 | debug_printf("basisu_frontend::init: Multithreaded: %u, Job pool total threads: %u, NumEndpointClusters: %u, NumSelectorClusters: %u, Perceptual: %u, CompressionLevel: %u\n", |
| 54 | p.m_multithreaded, p.m_pJob_pool ? p.m_pJob_pool->get_total_threads() : 0, |
| 55 | p.m_max_endpoint_clusters, p.m_max_selector_clusters, p.m_perceptual, p.m_compression_level); |
| 56 | |
| 57 | if ((p.m_max_endpoint_clusters < 1) || (p.m_max_endpoint_clusters > cMaxEndpointClusters)) |
| 58 | return false; |
| 59 | if ((p.m_max_selector_clusters < 1) || (p.m_max_selector_clusters > cMaxSelectorClusters)) |
| 60 | return false; |
| 61 | |
| 62 | m_source_blocks.resize(0); |
| 63 | append_vector(m_source_blocks, p.m_pSource_blocks, p.m_num_source_blocks); |
| 64 | |
| 65 | m_params = p; |
| 66 | |
| 67 | if (m_params.m_pOpenCL_context) |
| 68 | { |
| 69 | BASISU_ASSUME(sizeof(cl_pixel_block) == sizeof(pixel_block)); |
| 70 | |
| 71 | // Upload the RGBA pixel blocks a single time. |
| 72 | if (!opencl_set_pixel_blocks(m_params.m_pOpenCL_context, m_source_blocks.size(), (cl_pixel_block*)m_source_blocks.data())) |
| 73 | { |
| 74 | // This is not fatal, we just won't use OpenCL. |
| 75 | error_printf("basisu_frontend::init: opencl_set_pixel_blocks() failed\n"); |
| 76 | m_params.m_pOpenCL_context = nullptr; |
| 77 | m_opencl_failed = true; |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | m_encoded_blocks.resize(m_params.m_num_source_blocks); |
| 82 | memset(&m_encoded_blocks[0], 0, m_encoded_blocks.size() * sizeof(m_encoded_blocks[0])); |
| 83 | |
| 84 | m_num_endpoint_codebook_iterations = 1; |
| 85 | m_num_selector_codebook_iterations = 1; |
| 86 | |
| 87 | switch (p.m_compression_level) |
| 88 | { |
| 89 | case 0: |
| 90 | { |
| 91 | m_endpoint_refinement = false; |
| 92 | m_use_hierarchical_endpoint_codebooks = true; |
| 93 | m_use_hierarchical_selector_codebooks = true; |
| 94 | break; |
| 95 | } |
| 96 | case 1: |
| 97 | { |
| 98 | m_endpoint_refinement = true; |
| 99 | m_use_hierarchical_endpoint_codebooks = true; |
| 100 | m_use_hierarchical_selector_codebooks = true; |
| 101 | |
| 102 | break; |
| 103 | } |
| 104 | case 2: |
| 105 | { |
| 106 | m_endpoint_refinement = true; |
| 107 | m_use_hierarchical_endpoint_codebooks = true; |
| 108 | m_use_hierarchical_selector_codebooks = true; |
| 109 | |
| 110 | break; |
| 111 | } |
| 112 | case 3: |
| 113 | { |
| 114 | m_endpoint_refinement = true; |
| 115 | m_use_hierarchical_endpoint_codebooks = false; |
| 116 | m_use_hierarchical_selector_codebooks = false; |
| 117 | break; |
| 118 | } |
| 119 | case 4: |
| 120 | { |
| 121 | m_endpoint_refinement = true; |
| 122 | m_use_hierarchical_endpoint_codebooks = true; |
| 123 | m_use_hierarchical_selector_codebooks = true; |
| 124 | m_num_endpoint_codebook_iterations = BASISU_MAX_ENDPOINT_REFINEMENT_STEPS; |
| 125 | m_num_selector_codebook_iterations = BASISU_MAX_ENDPOINT_REFINEMENT_STEPS; |
| 126 | break; |
| 127 | } |
| 128 | case 5: |
| 129 | { |
| 130 | m_endpoint_refinement = true; |
| 131 | m_use_hierarchical_endpoint_codebooks = false; |
| 132 | m_use_hierarchical_selector_codebooks = false; |
| 133 | m_num_endpoint_codebook_iterations = BASISU_MAX_ENDPOINT_REFINEMENT_STEPS; |
| 134 | m_num_selector_codebook_iterations = BASISU_MAX_ENDPOINT_REFINEMENT_STEPS; |
| 135 | break; |
| 136 | } |
| 137 | case 6: |
| 138 | default: |
| 139 | { |
| 140 | m_endpoint_refinement = true; |
| 141 | m_use_hierarchical_endpoint_codebooks = false; |
| 142 | m_use_hierarchical_selector_codebooks = false; |
| 143 | m_num_endpoint_codebook_iterations = BASISU_MAX_ENDPOINT_REFINEMENT_STEPS*2; |
| 144 | m_num_selector_codebook_iterations = BASISU_MAX_ENDPOINT_REFINEMENT_STEPS*2; |
| 145 | break; |
| 146 | } |
| 147 | |
| 148 | } |
| 149 | |
| 150 | if (m_params.m_disable_hierarchical_endpoint_codebooks) |
| 151 | m_use_hierarchical_endpoint_codebooks = false; |
| 152 | |
| 153 | debug_printf("Endpoint refinement: %u, Hierarchical endpoint codebooks: %u, Hierarchical selector codebooks: %u, Endpoint codebook iters: %u, Selector codebook iters: %u\n", |
| 154 | m_endpoint_refinement, m_use_hierarchical_endpoint_codebooks, m_use_hierarchical_selector_codebooks, m_num_endpoint_codebook_iterations, m_num_selector_codebook_iterations); |
| 155 | |
| 156 | return true; |
| 157 | } |
| 158 | |
| 159 | bool basisu_frontend::compress() |
| 160 | { |
| 161 | debug_printf("basisu_frontend::compress\n"); |
| 162 | |
| 163 | m_total_blocks = m_params.m_num_source_blocks; |
| 164 | m_total_pixels = m_total_blocks * cPixelBlockTotalPixels; |
| 165 | |
| 166 | // Encode the initial high quality ETC1S texture |
| 167 | |
| 168 | init_etc1_images(); |
| 169 | |
| 170 | // First quantize the ETC1S endpoints |
| 171 | |
| 172 | if (m_params.m_pGlobal_codebooks) |
| 173 | { |
| 174 | init_global_codebooks(); |
| 175 | } |
| 176 | else |
| 177 | { |
| 178 | init_endpoint_training_vectors(); |
| 179 | |
| 180 | generate_endpoint_clusters(); |
| 181 | |
| 182 | for (uint32_t refine_endpoint_step = 0; refine_endpoint_step < m_num_endpoint_codebook_iterations; refine_endpoint_step++) |
| 183 | { |
| 184 | if (m_params.m_validate) |
| 185 | { |
| 186 | BASISU_FRONTEND_VERIFY(check_etc1s_constraints()); |
| 187 | |
| 188 | BASISU_FRONTEND_VERIFY(validate_endpoint_cluster_hierarchy(false)); |
| 189 | } |
| 190 | |
| 191 | if (refine_endpoint_step) |
| 192 | { |
| 193 | introduce_new_endpoint_clusters(); |
| 194 | } |
| 195 | |
| 196 | if (m_params.m_validate) |
| 197 | { |
| 198 | BASISU_FRONTEND_VERIFY(validate_endpoint_cluster_hierarchy(false)); |
| 199 | } |
| 200 | |
| 201 | generate_endpoint_codebook(refine_endpoint_step); |
| 202 | |
| 203 | if ((m_params.m_debug_images) && (m_params.m_dump_endpoint_clusterization)) |
| 204 | { |
| 205 | char buf[256]; |
| 206 | snprintf(buf, sizeof(buf), "endpoint_cluster_vis_pre_%u.png", refine_endpoint_step); |
| 207 | dump_endpoint_clusterization_visualization(buf, false); |
| 208 | } |
| 209 | |
| 210 | bool early_out = false; |
| 211 | |
| 212 | if (m_endpoint_refinement) |
| 213 | { |
| 214 | //dump_endpoint_clusterization_visualization("endpoint_clusters_before_refinement.png"); |
| 215 | |
| 216 | if (!refine_endpoint_clusterization()) |
| 217 | early_out = true; |
| 218 | |
| 219 | if ((m_params.m_tex_type == basist::cBASISTexTypeVideoFrames) && (!refine_endpoint_step) && (m_num_endpoint_codebook_iterations == 1)) |
| 220 | { |
| 221 | eliminate_redundant_or_empty_endpoint_clusters(); |
| 222 | generate_endpoint_codebook(basisu::maximum(1U, refine_endpoint_step)); |
| 223 | } |
| 224 | |
| 225 | if ((m_params.m_debug_images) && (m_params.m_dump_endpoint_clusterization)) |
| 226 | { |
| 227 | char buf[256]; |
| 228 | snprintf(buf, sizeof(buf), "endpoint_cluster_vis_post_%u.png", refine_endpoint_step); |
| 229 | |
| 230 | dump_endpoint_clusterization_visualization(buf, false); |
| 231 | snprintf(buf, sizeof(buf), "endpoint_cluster_colors_vis_post_%u.png", refine_endpoint_step); |
| 232 | |
| 233 | dump_endpoint_clusterization_visualization(buf, true); |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | if (m_params.m_validate) |
| 238 | { |
| 239 | BASISU_FRONTEND_VERIFY(validate_endpoint_cluster_hierarchy(false)); |
| 240 | } |
| 241 | |
| 242 | eliminate_redundant_or_empty_endpoint_clusters(); |
| 243 | |
| 244 | if (m_params.m_validate) |
| 245 | { |
| 246 | BASISU_FRONTEND_VERIFY(validate_endpoint_cluster_hierarchy(false)); |
| 247 | } |
| 248 | |
| 249 | if (m_params.m_debug_stats) |
| 250 | debug_printf("Total endpoint clusters: %u\n", (uint32_t)m_endpoint_clusters.size()); |
| 251 | |
| 252 | if (early_out) |
| 253 | break; |
| 254 | } |
| 255 | |
| 256 | if (m_params.m_validate) |
| 257 | { |
| 258 | BASISU_FRONTEND_VERIFY(check_etc1s_constraints()); |
| 259 | } |
| 260 | |
| 261 | generate_block_endpoint_clusters(); |
| 262 | |
| 263 | create_initial_packed_texture(); |
| 264 | |
| 265 | // Now quantize the ETC1S selectors |
| 266 | |
| 267 | generate_selector_clusters(); |
| 268 | |
| 269 | if (m_use_hierarchical_selector_codebooks) |
| 270 | compute_selector_clusters_within_each_parent_cluster(); |
| 271 | |
| 272 | if (m_params.m_compression_level == 0) |
| 273 | { |
| 274 | create_optimized_selector_codebook(0); |
| 275 | |
| 276 | find_optimal_selector_clusters_for_each_block(); |
| 277 | |
| 278 | introduce_special_selector_clusters(); |
| 279 | } |
| 280 | else |
| 281 | { |
| 282 | const uint32_t num_refine_selector_steps = m_num_selector_codebook_iterations; |
| 283 | for (uint32_t refine_selector_steps = 0; refine_selector_steps < num_refine_selector_steps; refine_selector_steps++) |
| 284 | { |
| 285 | create_optimized_selector_codebook(refine_selector_steps); |
| 286 | |
| 287 | find_optimal_selector_clusters_for_each_block(); |
| 288 | |
| 289 | introduce_special_selector_clusters(); |
| 290 | |
| 291 | if ((m_params.m_compression_level >= 4) || (m_params.m_tex_type == basist::cBASISTexTypeVideoFrames)) |
| 292 | { |
| 293 | if (!refine_block_endpoints_given_selectors()) |
| 294 | break; |
| 295 | } |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | optimize_selector_codebook(); |
| 300 | |
| 301 | if (m_params.m_debug_stats) |
| 302 | debug_printf("Total selector clusters: %u\n", (uint32_t)m_selector_cluster_block_indices.size()); |
| 303 | } |
| 304 | |
| 305 | finalize(); |
| 306 | |
| 307 | if (m_params.m_validate) |
| 308 | { |
| 309 | if (!validate_output()) |
| 310 | return false; |
| 311 | } |
| 312 | |
| 313 | debug_printf("basisu_frontend::compress: Done\n"); |
| 314 | |
| 315 | return true; |
| 316 | } |
| 317 | |
| 318 | bool basisu_frontend::init_global_codebooks() |
| 319 | { |
| 320 | const basist::basisu_lowlevel_etc1s_transcoder* pTranscoder = m_params.m_pGlobal_codebooks; |
| 321 | |
| 322 | const basist::basisu_lowlevel_etc1s_transcoder::endpoint_vec& endpoints = pTranscoder->get_endpoints(); |
| 323 | const basist::basisu_lowlevel_etc1s_transcoder::selector_vec& selectors = pTranscoder->get_selectors(); |
| 324 | |
| 325 | m_endpoint_cluster_etc_params.resize(endpoints.size()); |
| 326 | for (uint32_t i = 0; i < endpoints.size(); i++) |
| 327 | { |
| 328 | m_endpoint_cluster_etc_params[i].m_inten_table[0] = endpoints[i].m_inten5; |
| 329 | m_endpoint_cluster_etc_params[i].m_inten_table[1] = endpoints[i].m_inten5; |
| 330 | |
| 331 | m_endpoint_cluster_etc_params[i].m_color_unscaled[0].set(endpoints[i].m_color5.r, endpoints[i].m_color5.g, endpoints[i].m_color5.b, 255); |
| 332 | m_endpoint_cluster_etc_params[i].m_color_used[0] = true; |
| 333 | m_endpoint_cluster_etc_params[i].m_valid = true; |
| 334 | } |
| 335 | |
| 336 | m_optimized_cluster_selectors.resize(selectors.size()); |
| 337 | for (uint32_t i = 0; i < m_optimized_cluster_selectors.size(); i++) |
| 338 | { |
| 339 | for (uint32_t y = 0; y < 4; y++) |
| 340 | for (uint32_t x = 0; x < 4; x++) |
| 341 | m_optimized_cluster_selectors[i].set_selector(x, y, selectors[i].get_selector(x, y)); |
| 342 | } |
| 343 | |
| 344 | m_block_endpoint_clusters_indices.resize(m_total_blocks); |
| 345 | |
| 346 | m_orig_encoded_blocks.resize(m_total_blocks); |
| 347 | |
| 348 | m_block_selector_cluster_index.resize(m_total_blocks); |
| 349 | |
| 350 | #if 0 |
| 351 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 352 | { |
| 353 | const uint32_t first_index = block_index_iter; |
| 354 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 355 | |
| 356 | #ifndef __EMSCRIPTEN__ |
| 357 | m_params.m_pJob_pool->add_job([this, first_index, last_index] { |
| 358 | #endif |
| 359 | |
| 360 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 361 | { |
| 362 | const etc_block& blk = m_etc1_blocks_etc1s[block_index]; |
| 363 | |
| 364 | const uint32_t block_endpoint_index = m_block_endpoint_clusters_indices[block_index][0]; |
| 365 | |
| 366 | etc_block trial_blk; |
| 367 | trial_blk.set_block_color5_etc1s(blk.m_color_unscaled[0]); |
| 368 | trial_blk.set_flip_bit(true); |
| 369 | |
| 370 | uint64_t best_err = UINT64_MAX; |
| 371 | uint32_t best_index = 0; |
| 372 | |
| 373 | for (uint32_t i = 0; i < m_optimized_cluster_selectors.size(); i++) |
| 374 | { |
| 375 | trial_blk.set_raw_selector_bits(m_optimized_cluster_selectors[i].get_raw_selector_bits()); |
| 376 | |
| 377 | const uint64_t cur_err = trial_blk.evaluate_etc1_error(get_source_pixel_block(block_index).get_ptr(), m_params.m_perceptual); |
| 378 | if (cur_err < best_err) |
| 379 | { |
| 380 | best_err = cur_err; |
| 381 | best_index = i; |
| 382 | if (!cur_err) |
| 383 | break; |
| 384 | } |
| 385 | |
| 386 | } // block_index |
| 387 | |
| 388 | m_block_selector_cluster_index[block_index] = best_index; |
| 389 | } |
| 390 | |
| 391 | #ifndef __EMSCRIPTEN__ |
| 392 | }); |
| 393 | #endif |
| 394 | |
| 395 | } |
| 396 | |
| 397 | #ifndef __EMSCRIPTEN__ |
| 398 | m_params.m_pJob_pool->wait_for_all(); |
| 399 | #endif |
| 400 | |
| 401 | m_encoded_blocks.resize(m_total_blocks); |
| 402 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 403 | { |
| 404 | const uint32_t endpoint_index = m_block_endpoint_clusters_indices[block_index][0]; |
| 405 | const uint32_t selector_index = m_block_selector_cluster_index[block_index]; |
| 406 | |
| 407 | etc_block& blk = m_encoded_blocks[block_index]; |
| 408 | |
| 409 | blk.set_block_color5_etc1s(m_endpoint_cluster_etc_params[endpoint_index].m_color_unscaled[0]); |
| 410 | blk.set_inten_tables_etc1s(m_endpoint_cluster_etc_params[endpoint_index].m_inten_table[0]); |
| 411 | blk.set_flip_bit(true); |
| 412 | blk.set_raw_selector_bits(m_optimized_cluster_selectors[selector_index].get_raw_selector_bits()); |
| 413 | } |
| 414 | #endif |
| 415 | |
| 416 | // HACK HACK |
| 417 | const uint32_t NUM_PASSES = 3; |
| 418 | for (uint32_t pass = 0; pass < NUM_PASSES; pass++) |
| 419 | { |
| 420 | debug_printf("init_global_codebooks: pass %u\n", pass); |
| 421 | |
| 422 | const uint32_t N = 128; |
| 423 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 424 | { |
| 425 | const uint32_t first_index = block_index_iter; |
| 426 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 427 | |
| 428 | #ifndef __EMSCRIPTEN__ |
| 429 | m_params.m_pJob_pool->add_job([this, first_index, last_index, pass] { |
| 430 | #endif |
| 431 | |
| 432 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 433 | { |
| 434 | const etc_block& blk = pass ? m_encoded_blocks[block_index] : m_etc1_blocks_etc1s[block_index]; |
| 435 | const uint32_t blk_raw_selector_bits = blk.get_raw_selector_bits(); |
| 436 | |
| 437 | etc_block trial_blk(blk); |
| 438 | trial_blk.set_raw_selector_bits(blk_raw_selector_bits); |
| 439 | trial_blk.set_flip_bit(true); |
| 440 | |
| 441 | uint64_t best_err = UINT64_MAX; |
| 442 | uint32_t best_index = 0; |
| 443 | etc_block best_block(trial_blk); |
| 444 | |
| 445 | for (uint32_t i = 0; i < m_endpoint_cluster_etc_params.size(); i++) |
| 446 | { |
| 447 | if (m_endpoint_cluster_etc_params[i].m_inten_table[0] > blk.get_inten_table(0)) |
| 448 | continue; |
| 449 | |
| 450 | trial_blk.set_block_color5_etc1s(m_endpoint_cluster_etc_params[i].m_color_unscaled[0]); |
| 451 | trial_blk.set_inten_tables_etc1s(m_endpoint_cluster_etc_params[i].m_inten_table[0]); |
| 452 | |
| 453 | const color_rgba* pSource_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 454 | uint64_t cur_err; |
| 455 | if (!pass) |
| 456 | cur_err = trial_blk.determine_selectors(pSource_pixels, m_params.m_perceptual); |
| 457 | else |
| 458 | cur_err = trial_blk.evaluate_etc1_error(pSource_pixels, m_params.m_perceptual); |
| 459 | |
| 460 | if (cur_err < best_err) |
| 461 | { |
| 462 | best_err = cur_err; |
| 463 | best_index = i; |
| 464 | best_block = trial_blk; |
| 465 | |
| 466 | if (!cur_err) |
| 467 | break; |
| 468 | } |
| 469 | } |
| 470 | |
| 471 | m_block_endpoint_clusters_indices[block_index][0] = best_index; |
| 472 | m_block_endpoint_clusters_indices[block_index][1] = best_index; |
| 473 | |
| 474 | m_orig_encoded_blocks[block_index] = best_block; |
| 475 | |
| 476 | } // block_index |
| 477 | |
| 478 | #ifndef __EMSCRIPTEN__ |
| 479 | }); |
| 480 | #endif |
| 481 | |
| 482 | } |
| 483 | |
| 484 | #ifndef __EMSCRIPTEN__ |
| 485 | m_params.m_pJob_pool->wait_for_all(); |
| 486 | #endif |
| 487 | |
| 488 | m_endpoint_clusters.resize(0); |
| 489 | m_endpoint_clusters.resize(endpoints.size()); |
| 490 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 491 | { |
| 492 | const uint32_t endpoint_cluster_index = m_block_endpoint_clusters_indices[block_index][0]; |
| 493 | m_endpoint_clusters[endpoint_cluster_index].push_back(block_index * 2); |
| 494 | m_endpoint_clusters[endpoint_cluster_index].push_back(block_index * 2 + 1); |
| 495 | } |
| 496 | |
| 497 | m_block_selector_cluster_index.resize(m_total_blocks); |
| 498 | |
| 499 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 500 | { |
| 501 | const uint32_t first_index = block_index_iter; |
| 502 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 503 | |
| 504 | #ifndef __EMSCRIPTEN__ |
| 505 | m_params.m_pJob_pool->add_job([this, first_index, last_index] { |
| 506 | #endif |
| 507 | |
| 508 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 509 | { |
| 510 | const uint32_t block_endpoint_index = m_block_endpoint_clusters_indices[block_index][0]; |
| 511 | |
| 512 | etc_block trial_blk; |
| 513 | trial_blk.set_block_color5_etc1s(m_endpoint_cluster_etc_params[block_endpoint_index].m_color_unscaled[0]); |
| 514 | trial_blk.set_inten_tables_etc1s(m_endpoint_cluster_etc_params[block_endpoint_index].m_inten_table[0]); |
| 515 | trial_blk.set_flip_bit(true); |
| 516 | |
| 517 | uint64_t best_err = UINT64_MAX; |
| 518 | uint32_t best_index = 0; |
| 519 | |
| 520 | for (uint32_t i = 0; i < m_optimized_cluster_selectors.size(); i++) |
| 521 | { |
| 522 | trial_blk.set_raw_selector_bits(m_optimized_cluster_selectors[i].get_raw_selector_bits()); |
| 523 | |
| 524 | const uint64_t cur_err = trial_blk.evaluate_etc1_error(get_source_pixel_block(block_index).get_ptr(), m_params.m_perceptual); |
| 525 | if (cur_err < best_err) |
| 526 | { |
| 527 | best_err = cur_err; |
| 528 | best_index = i; |
| 529 | if (!cur_err) |
| 530 | break; |
| 531 | } |
| 532 | |
| 533 | } // block_index |
| 534 | |
| 535 | m_block_selector_cluster_index[block_index] = best_index; |
| 536 | } |
| 537 | |
| 538 | #ifndef __EMSCRIPTEN__ |
| 539 | }); |
| 540 | #endif |
| 541 | |
| 542 | } |
| 543 | |
| 544 | #ifndef __EMSCRIPTEN__ |
| 545 | m_params.m_pJob_pool->wait_for_all(); |
| 546 | #endif |
| 547 | |
| 548 | m_encoded_blocks.resize(m_total_blocks); |
| 549 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 550 | { |
| 551 | const uint32_t endpoint_index = m_block_endpoint_clusters_indices[block_index][0]; |
| 552 | const uint32_t selector_index = m_block_selector_cluster_index[block_index]; |
| 553 | |
| 554 | etc_block& blk = m_encoded_blocks[block_index]; |
| 555 | |
| 556 | blk.set_block_color5_etc1s(m_endpoint_cluster_etc_params[endpoint_index].m_color_unscaled[0]); |
| 557 | blk.set_inten_tables_etc1s(m_endpoint_cluster_etc_params[endpoint_index].m_inten_table[0]); |
| 558 | blk.set_flip_bit(true); |
| 559 | blk.set_raw_selector_bits(m_optimized_cluster_selectors[selector_index].get_raw_selector_bits()); |
| 560 | } |
| 561 | |
| 562 | } // pass |
| 563 | |
| 564 | m_selector_cluster_block_indices.resize(selectors.size()); |
| 565 | for (uint32_t block_index = 0; block_index < m_etc1_blocks_etc1s.size(); block_index++) |
| 566 | m_selector_cluster_block_indices[m_block_selector_cluster_index[block_index]].push_back(block_index); |
| 567 | |
| 568 | return true; |
| 569 | } |
| 570 | |
| 571 | void basisu_frontend::introduce_special_selector_clusters() |
| 572 | { |
| 573 | debug_printf("introduce_special_selector_clusters\n"); |
| 574 | |
| 575 | uint32_t total_blocks_relocated = 0; |
| 576 | const uint32_t initial_selector_clusters = (uint32_t)m_selector_cluster_block_indices.size(); |
| 577 | |
| 578 | bool_vec block_relocated_flags(m_total_blocks); |
| 579 | |
| 580 | // Make sure the selector codebook always has pure flat blocks for each possible selector, to avoid obvious artifacts. |
| 581 | // optimize_selector_codebook() will clean up any redundant clusters we create here. |
| 582 | for (uint32_t sel = 0; sel < 4; sel++) |
| 583 | { |
| 584 | etc_block blk; |
| 585 | clear_obj(blk); |
| 586 | for (uint32_t j = 0; j < 16; j++) |
| 587 | blk.set_selector(j & 3, j >> 2, sel); |
| 588 | |
| 589 | int k; |
| 590 | for (k = 0; k < (int)m_optimized_cluster_selectors.size(); k++) |
| 591 | if (m_optimized_cluster_selectors[k].get_raw_selector_bits() == blk.get_raw_selector_bits()) |
| 592 | break; |
| 593 | if (k < (int)m_optimized_cluster_selectors.size()) |
| 594 | continue; |
| 595 | |
| 596 | debug_printf("Introducing sel %u\n", sel); |
| 597 | |
| 598 | const uint32_t new_selector_cluster_index = (uint32_t)m_optimized_cluster_selectors.size(); |
| 599 | |
| 600 | m_optimized_cluster_selectors.push_back(blk); |
| 601 | |
| 602 | vector_ensure_element_is_valid(m_selector_cluster_block_indices, new_selector_cluster_index); |
| 603 | |
| 604 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 605 | { |
| 606 | if (m_orig_encoded_blocks[block_index].get_raw_selector_bits() != blk.get_raw_selector_bits()) |
| 607 | continue; |
| 608 | |
| 609 | // See if using flat selectors actually decreases the block's error. |
| 610 | const uint32_t old_selector_cluster_index = m_block_selector_cluster_index[block_index]; |
| 611 | |
| 612 | etc_block cur_blk; |
| 613 | const uint32_t endpoint_cluster_index = get_subblock_endpoint_cluster_index(block_index, 0); |
| 614 | cur_blk.set_block_color5_etc1s(get_endpoint_cluster_unscaled_color(endpoint_cluster_index, false)); |
| 615 | cur_blk.set_inten_tables_etc1s(get_endpoint_cluster_inten_table(endpoint_cluster_index, false)); |
| 616 | cur_blk.set_raw_selector_bits(get_selector_cluster_selector_bits(old_selector_cluster_index).get_raw_selector_bits()); |
| 617 | cur_blk.set_flip_bit(true); |
| 618 | |
| 619 | const uint64_t cur_err = cur_blk.evaluate_etc1_error(get_source_pixel_block(block_index).get_ptr(), m_params.m_perceptual); |
| 620 | |
| 621 | cur_blk.set_raw_selector_bits(blk.get_raw_selector_bits()); |
| 622 | |
| 623 | const uint64_t new_err = cur_blk.evaluate_etc1_error(get_source_pixel_block(block_index).get_ptr(), m_params.m_perceptual); |
| 624 | |
| 625 | if (new_err >= cur_err) |
| 626 | continue; |
| 627 | |
| 628 | // Change the block to use the new cluster |
| 629 | m_block_selector_cluster_index[block_index] = new_selector_cluster_index; |
| 630 | |
| 631 | m_selector_cluster_block_indices[new_selector_cluster_index].push_back(block_index); |
| 632 | |
| 633 | block_relocated_flags[block_index] = true; |
| 634 | |
| 635 | #if 0 |
| 636 | int j = vector_find(m_selector_cluster_block_indices[old_selector_cluster_index], block_index); |
| 637 | if (j >= 0) |
| 638 | m_selector_cluster_block_indices[old_selector_cluster_index].erase(m_selector_cluster_block_indices[old_selector_cluster_index].begin() + j); |
| 639 | #endif |
| 640 | |
| 641 | total_blocks_relocated++; |
| 642 | |
| 643 | m_encoded_blocks[block_index].set_raw_selector_bits(blk.get_raw_selector_bits()); |
| 644 | |
| 645 | } // block_index |
| 646 | |
| 647 | } // sel |
| 648 | |
| 649 | if (total_blocks_relocated) |
| 650 | { |
| 651 | debug_printf("Fixing selector codebook\n"); |
| 652 | |
| 653 | for (int selector_cluster_index = 0; selector_cluster_index < (int)initial_selector_clusters; selector_cluster_index++) |
| 654 | { |
| 655 | uint_vec& block_indices = m_selector_cluster_block_indices[selector_cluster_index]; |
| 656 | |
| 657 | uint32_t dst_ofs = 0; |
| 658 | |
| 659 | for (uint32_t i = 0; i < block_indices.size(); i++) |
| 660 | { |
| 661 | const uint32_t block_index = block_indices[i]; |
| 662 | if (!block_relocated_flags[block_index]) |
| 663 | block_indices[dst_ofs++] = block_index; |
| 664 | } |
| 665 | |
| 666 | block_indices.resize(dst_ofs); |
| 667 | } |
| 668 | } |
| 669 | |
| 670 | debug_printf("Total blocks relocated to new flat selector clusters: %u\n", total_blocks_relocated); |
| 671 | } |
| 672 | |
| 673 | // This method will change the number and ordering of the selector codebook clusters. |
| 674 | void basisu_frontend::optimize_selector_codebook() |
| 675 | { |
| 676 | debug_printf("optimize_selector_codebook\n"); |
| 677 | |
| 678 | const uint32_t orig_total_selector_clusters = (uint32_t)m_optimized_cluster_selectors.size(); |
| 679 | |
| 680 | bool_vec selector_cluster_was_used(m_optimized_cluster_selectors.size()); |
| 681 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 682 | selector_cluster_was_used[m_block_selector_cluster_index[i]] = true; |
| 683 | |
| 684 | int_vec old_to_new(m_optimized_cluster_selectors.size()); |
| 685 | int_vec new_to_old; |
| 686 | uint32_t total_new_entries = 0; |
| 687 | |
| 688 | std::unordered_map<uint32_t, uint32_t> selector_hashmap; |
| 689 | |
| 690 | for (int i = 0; i < static_cast<int>(m_optimized_cluster_selectors.size()); i++) |
| 691 | { |
| 692 | if (!selector_cluster_was_used[i]) |
| 693 | { |
| 694 | old_to_new[i] = -1; |
| 695 | continue; |
| 696 | } |
| 697 | |
| 698 | const uint32_t raw_selector_bits = m_optimized_cluster_selectors[i].get_raw_selector_bits(); |
| 699 | |
| 700 | auto find_res = selector_hashmap.insert(std::make_pair(raw_selector_bits, total_new_entries)); |
| 701 | if (!find_res.second) |
| 702 | { |
| 703 | old_to_new[i] = (find_res.first)->second; |
| 704 | continue; |
| 705 | } |
| 706 | |
| 707 | old_to_new[i] = total_new_entries++; |
| 708 | new_to_old.push_back(i); |
| 709 | } |
| 710 | |
| 711 | debug_printf("Original selector clusters: %u, new cluster selectors: %u\n", orig_total_selector_clusters, total_new_entries); |
| 712 | |
| 713 | for (uint32_t i = 0; i < m_block_selector_cluster_index.size(); i++) |
| 714 | { |
| 715 | BASISU_FRONTEND_VERIFY((old_to_new[m_block_selector_cluster_index[i]] >= 0) && (old_to_new[m_block_selector_cluster_index[i]] < (int)total_new_entries)); |
| 716 | m_block_selector_cluster_index[i] = old_to_new[m_block_selector_cluster_index[i]]; |
| 717 | } |
| 718 | |
| 719 | basisu::vector<etc_block> new_optimized_cluster_selectors(m_optimized_cluster_selectors.size() ? total_new_entries : 0); |
| 720 | basisu::vector<uint_vec> new_selector_cluster_indices(m_selector_cluster_block_indices.size() ? total_new_entries : 0); |
| 721 | |
| 722 | for (uint32_t i = 0; i < total_new_entries; i++) |
| 723 | { |
| 724 | if (m_optimized_cluster_selectors.size()) |
| 725 | new_optimized_cluster_selectors[i] = m_optimized_cluster_selectors[new_to_old[i]]; |
| 726 | |
| 727 | //if (m_selector_cluster_block_indices.size()) |
| 728 | // new_selector_cluster_indices[i] = m_selector_cluster_block_indices[new_to_old[i]]; |
| 729 | } |
| 730 | |
| 731 | for (uint32_t i = 0; i < m_block_selector_cluster_index.size(); i++) |
| 732 | { |
| 733 | new_selector_cluster_indices[m_block_selector_cluster_index[i]].push_back(i); |
| 734 | } |
| 735 | |
| 736 | m_optimized_cluster_selectors.swap(new_optimized_cluster_selectors); |
| 737 | m_selector_cluster_block_indices.swap(new_selector_cluster_indices); |
| 738 | |
| 739 | // This isn't strictly necessary - doing it for completeness/future sanity. |
| 740 | if (m_selector_clusters_within_each_parent_cluster.size()) |
| 741 | { |
| 742 | for (uint32_t i = 0; i < m_selector_clusters_within_each_parent_cluster.size(); i++) |
| 743 | for (uint32_t j = 0; j < m_selector_clusters_within_each_parent_cluster[i].size(); j++) |
| 744 | m_selector_clusters_within_each_parent_cluster[i][j] = old_to_new[m_selector_clusters_within_each_parent_cluster[i][j]]; |
| 745 | } |
| 746 | |
| 747 | debug_printf("optimize_selector_codebook: Before: %u After: %u\n", orig_total_selector_clusters, total_new_entries); |
| 748 | } |
| 749 | |
| 750 | void basisu_frontend::init_etc1_images() |
| 751 | { |
| 752 | debug_printf("basisu_frontend::init_etc1_images\n"); |
| 753 | |
| 754 | interval_timer tm; |
| 755 | tm.start(); |
| 756 | |
| 757 | m_etc1_blocks_etc1s.resize(m_total_blocks); |
| 758 | |
| 759 | bool use_cpu = true; |
| 760 | |
| 761 | if (m_params.m_pOpenCL_context) |
| 762 | { |
| 763 | uint32_t total_perms = 64; |
| 764 | if (m_params.m_compression_level == 0) |
| 765 | total_perms = 4; |
| 766 | else if (m_params.m_compression_level == 1) |
| 767 | total_perms = 16; |
| 768 | else if (m_params.m_compression_level == BASISU_MAX_COMPRESSION_LEVEL) |
| 769 | total_perms = OPENCL_ENCODE_ETC1S_MAX_PERMS; |
| 770 | |
| 771 | bool status = opencl_encode_etc1s_blocks(m_params.m_pOpenCL_context, m_etc1_blocks_etc1s.data(), m_params.m_perceptual, total_perms); |
| 772 | if (status) |
| 773 | use_cpu = false; |
| 774 | else |
| 775 | { |
| 776 | error_printf("basisu_frontend::init_etc1_images: opencl_encode_etc1s_blocks() failed! Using CPU.\n"); |
| 777 | m_params.m_pOpenCL_context = nullptr; |
| 778 | m_opencl_failed = true; |
| 779 | } |
| 780 | } |
| 781 | |
| 782 | if (use_cpu) |
| 783 | { |
| 784 | const uint32_t N = 4096; |
| 785 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 786 | { |
| 787 | const uint32_t first_index = block_index_iter; |
| 788 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 789 | |
| 790 | #ifndef __EMSCRIPTEN__ |
| 791 | m_params.m_pJob_pool->add_job([this, first_index, last_index] { |
| 792 | #endif |
| 793 | |
| 794 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 795 | { |
| 796 | const pixel_block& source_blk = get_source_pixel_block(block_index); |
| 797 | |
| 798 | etc1_optimizer optimizer; |
| 799 | etc1_optimizer::params optimizer_params; |
| 800 | etc1_optimizer::results optimizer_results; |
| 801 | |
| 802 | if (m_params.m_compression_level == 0) |
| 803 | optimizer_params.m_quality = cETCQualityFast; |
| 804 | else if (m_params.m_compression_level == 1) |
| 805 | optimizer_params.m_quality = cETCQualityMedium; |
| 806 | else if (m_params.m_compression_level == BASISU_MAX_COMPRESSION_LEVEL) |
| 807 | optimizer_params.m_quality = cETCQualityUber; |
| 808 | |
| 809 | optimizer_params.m_num_src_pixels = 16; |
| 810 | optimizer_params.m_pSrc_pixels = source_blk.get_ptr(); |
| 811 | optimizer_params.m_perceptual = m_params.m_perceptual; |
| 812 | |
| 813 | uint8_t selectors[16]; |
| 814 | optimizer_results.m_pSelectors = selectors; |
| 815 | optimizer_results.m_n = 16; |
| 816 | |
| 817 | optimizer.init(optimizer_params, optimizer_results); |
| 818 | if (!optimizer.compute()) |
| 819 | BASISU_FRONTEND_VERIFY(false); |
| 820 | |
| 821 | etc_block& blk = m_etc1_blocks_etc1s[block_index]; |
| 822 | |
| 823 | memset(&blk, 0, sizeof(blk)); |
| 824 | blk.set_block_color5_etc1s(optimizer_results.m_block_color_unscaled); |
| 825 | blk.set_inten_tables_etc1s(optimizer_results.m_block_inten_table); |
| 826 | blk.set_flip_bit(true); |
| 827 | |
| 828 | for (uint32_t y = 0; y < 4; y++) |
| 829 | for (uint32_t x = 0; x < 4; x++) |
| 830 | blk.set_selector(x, y, selectors[x + y * 4]); |
| 831 | } |
| 832 | |
| 833 | #ifndef __EMSCRIPTEN__ |
| 834 | }); |
| 835 | #endif |
| 836 | |
| 837 | } |
| 838 | |
| 839 | #ifndef __EMSCRIPTEN__ |
| 840 | m_params.m_pJob_pool->wait_for_all(); |
| 841 | #endif |
| 842 | |
| 843 | } // use_cpu |
| 844 | |
| 845 | debug_printf("init_etc1_images: Elapsed time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 846 | } |
| 847 | |
| 848 | void basisu_frontend::init_endpoint_training_vectors() |
| 849 | { |
| 850 | debug_printf("init_endpoint_training_vectors\n"); |
| 851 | |
| 852 | vec6F_quantizer::array_of_weighted_training_vecs &training_vecs = m_endpoint_clusterizer.get_training_vecs(); |
| 853 | |
| 854 | training_vecs.resize(m_total_blocks * 2); |
| 855 | |
| 856 | const uint32_t N = 16384; |
| 857 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 858 | { |
| 859 | const uint32_t first_index = block_index_iter; |
| 860 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 861 | |
| 862 | #ifndef __EMSCRIPTEN__ |
| 863 | m_params.m_pJob_pool->add_job( [this, first_index, last_index, &training_vecs] { |
| 864 | #endif |
| 865 | |
| 866 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 867 | { |
| 868 | const etc_block &blk = m_etc1_blocks_etc1s[block_index]; |
| 869 | |
| 870 | color_rgba block_colors[2]; |
| 871 | blk.get_block_low_high_colors(block_colors, 0); |
| 872 | |
| 873 | vec6F v; |
| 874 | v[0] = block_colors[0].r * (1.0f / 255.0f); |
| 875 | v[1] = block_colors[0].g * (1.0f / 255.0f); |
| 876 | v[2] = block_colors[0].b * (1.0f / 255.0f); |
| 877 | v[3] = block_colors[1].r * (1.0f / 255.0f); |
| 878 | v[4] = block_colors[1].g * (1.0f / 255.0f); |
| 879 | v[5] = block_colors[1].b * (1.0f / 255.0f); |
| 880 | |
| 881 | training_vecs[block_index * 2 + 0] = std::make_pair(v, 1); |
| 882 | training_vecs[block_index * 2 + 1] = std::make_pair(v, 1); |
| 883 | |
| 884 | } // block_index; |
| 885 | |
| 886 | #ifndef __EMSCRIPTEN__ |
| 887 | } ); |
| 888 | #endif |
| 889 | |
| 890 | } // block_index_iter |
| 891 | |
| 892 | #ifndef __EMSCRIPTEN__ |
| 893 | m_params.m_pJob_pool->wait_for_all(); |
| 894 | #endif |
| 895 | } |
| 896 | |
| 897 | void basisu_frontend::generate_endpoint_clusters() |
| 898 | { |
| 899 | debug_printf("Begin endpoint quantization\n"); |
| 900 | |
| 901 | const uint32_t parent_codebook_size = (m_params.m_max_endpoint_clusters >= 256) ? BASISU_ENDPOINT_PARENT_CODEBOOK_SIZE : 0; |
| 902 | uint32_t max_threads = 0; |
| 903 | max_threads = m_params.m_multithreaded ? minimum<int>(std::thread::hardware_concurrency(), cMaxCodebookCreationThreads) : 0; |
| 904 | if (m_params.m_pJob_pool) |
| 905 | max_threads = minimum<int>((int)m_params.m_pJob_pool->get_total_threads(), max_threads); |
| 906 | |
| 907 | debug_printf("max_threads: %u\n", max_threads); |
| 908 | bool status = generate_hierarchical_codebook_threaded(m_endpoint_clusterizer, |
| 909 | m_params.m_max_endpoint_clusters, m_use_hierarchical_endpoint_codebooks ? parent_codebook_size : 0, |
| 910 | m_endpoint_clusters, |
| 911 | m_endpoint_parent_clusters, |
| 912 | max_threads, m_params.m_pJob_pool, true); |
| 913 | BASISU_FRONTEND_VERIFY(status); |
| 914 | |
| 915 | if (m_use_hierarchical_endpoint_codebooks) |
| 916 | { |
| 917 | if (!m_endpoint_parent_clusters.size()) |
| 918 | { |
| 919 | m_endpoint_parent_clusters.resize(0); |
| 920 | m_endpoint_parent_clusters.resize(1); |
| 921 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 922 | { |
| 923 | m_endpoint_parent_clusters[0].push_back(i*2); |
| 924 | m_endpoint_parent_clusters[0].push_back(i*2+1); |
| 925 | } |
| 926 | } |
| 927 | |
| 928 | BASISU_ASSUME(BASISU_ENDPOINT_PARENT_CODEBOOK_SIZE <= UINT8_MAX); |
| 929 | |
| 930 | m_block_parent_endpoint_cluster.resize(0); |
| 931 | m_block_parent_endpoint_cluster.resize(m_total_blocks); |
| 932 | vector_set_all(m_block_parent_endpoint_cluster, 0xFF); |
| 933 | for (uint32_t parent_cluster_index = 0; parent_cluster_index < m_endpoint_parent_clusters.size(); parent_cluster_index++) |
| 934 | { |
| 935 | const uint_vec &cluster = m_endpoint_parent_clusters[parent_cluster_index]; |
| 936 | for (uint32_t j = 0; j < cluster.size(); j++) |
| 937 | { |
| 938 | const uint32_t block_index = cluster[j] >> 1; |
| 939 | m_block_parent_endpoint_cluster[block_index] = static_cast<uint8_t>(parent_cluster_index); |
| 940 | } |
| 941 | } |
| 942 | |
| 943 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 944 | { |
| 945 | BASISU_FRONTEND_VERIFY(m_block_parent_endpoint_cluster[i] != 0xFF); |
| 946 | } |
| 947 | |
| 948 | // Ensure that all the blocks within each cluster are all in the same parent cluster, or something is very wrong. |
| 949 | for (uint32_t cluster_index = 0; cluster_index < m_endpoint_clusters.size(); cluster_index++) |
| 950 | { |
| 951 | const uint_vec &cluster = m_endpoint_clusters[cluster_index]; |
| 952 | |
| 953 | uint32_t parent_cluster_index = 0; |
| 954 | for (uint32_t j = 0; j < cluster.size(); j++) |
| 955 | { |
| 956 | const uint32_t block_index = cluster[j] >> 1; |
| 957 | |
| 958 | BASISU_FRONTEND_VERIFY(block_index < m_block_parent_endpoint_cluster.size()); |
| 959 | |
| 960 | if (!j) |
| 961 | { |
| 962 | parent_cluster_index = m_block_parent_endpoint_cluster[block_index]; |
| 963 | } |
| 964 | else |
| 965 | { |
| 966 | BASISU_FRONTEND_VERIFY(m_block_parent_endpoint_cluster[block_index] == parent_cluster_index); |
| 967 | } |
| 968 | } |
| 969 | } |
| 970 | } |
| 971 | |
| 972 | if (m_params.m_debug_stats) |
| 973 | debug_printf("Total endpoint clusters: %u, parent clusters: %u\n", (uint32_t)m_endpoint_clusters.size(), (uint32_t)m_endpoint_parent_clusters.size()); |
| 974 | } |
| 975 | |
| 976 | // Iterate through each array of endpoint cluster block indices and set the m_block_endpoint_clusters_indices[][] array to indicaste which cluster index each block uses. |
| 977 | void basisu_frontend::generate_block_endpoint_clusters() |
| 978 | { |
| 979 | m_block_endpoint_clusters_indices.resize(m_total_blocks); |
| 980 | |
| 981 | for (int cluster_index = 0; cluster_index < static_cast<int>(m_endpoint_clusters.size()); cluster_index++) |
| 982 | { |
| 983 | const basisu::vector<uint32_t>& cluster_indices = m_endpoint_clusters[cluster_index]; |
| 984 | |
| 985 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 986 | { |
| 987 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 988 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 989 | |
| 990 | m_block_endpoint_clusters_indices[block_index][subblock_index] = cluster_index; |
| 991 | |
| 992 | } // cluster_indices_iter |
| 993 | } |
| 994 | |
| 995 | if (m_params.m_validate) |
| 996 | { |
| 997 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 998 | { |
| 999 | uint32_t cluster_0 = m_block_endpoint_clusters_indices[block_index][0]; |
| 1000 | uint32_t cluster_1 = m_block_endpoint_clusters_indices[block_index][1]; |
| 1001 | BASISU_FRONTEND_VERIFY(cluster_0 == cluster_1); |
| 1002 | } |
| 1003 | } |
| 1004 | } |
| 1005 | |
| 1006 | void basisu_frontend::compute_endpoint_clusters_within_each_parent_cluster() |
| 1007 | { |
| 1008 | generate_block_endpoint_clusters(); |
| 1009 | |
| 1010 | m_endpoint_clusters_within_each_parent_cluster.resize(0); |
| 1011 | m_endpoint_clusters_within_each_parent_cluster.resize(m_endpoint_parent_clusters.size()); |
| 1012 | |
| 1013 | // Note: It's possible that some blocks got moved into the same cluster, but live in different parent clusters. |
| 1014 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 1015 | { |
| 1016 | const uint32_t cluster_index = m_block_endpoint_clusters_indices[block_index][0]; |
| 1017 | const uint32_t parent_cluster_index = m_block_parent_endpoint_cluster[block_index]; |
| 1018 | |
| 1019 | m_endpoint_clusters_within_each_parent_cluster[parent_cluster_index].push_back(cluster_index); |
| 1020 | } |
| 1021 | |
| 1022 | for (uint32_t i = 0; i < m_endpoint_clusters_within_each_parent_cluster.size(); i++) |
| 1023 | { |
| 1024 | uint_vec &cluster_indices = m_endpoint_clusters_within_each_parent_cluster[i]; |
| 1025 | |
| 1026 | BASISU_FRONTEND_VERIFY(cluster_indices.size()); |
| 1027 | |
| 1028 | vector_sort(cluster_indices); |
| 1029 | |
| 1030 | auto last = std::unique(cluster_indices.begin(), cluster_indices.end()); |
| 1031 | cluster_indices.erase(last, cluster_indices.end()); |
| 1032 | } |
| 1033 | } |
| 1034 | |
| 1035 | void basisu_frontend::compute_endpoint_subblock_error_vec() |
| 1036 | { |
| 1037 | m_subblock_endpoint_quant_err_vec.resize(0); |
| 1038 | |
| 1039 | const uint32_t N = 512; |
| 1040 | for (uint32_t cluster_index_iter = 0; cluster_index_iter < m_endpoint_clusters.size(); cluster_index_iter += N) |
| 1041 | { |
| 1042 | const uint32_t first_index = cluster_index_iter; |
| 1043 | const uint32_t last_index = minimum<uint32_t>((uint32_t)m_endpoint_clusters.size(), cluster_index_iter + N); |
| 1044 | |
| 1045 | #ifndef __EMSCRIPTEN__ |
| 1046 | m_params.m_pJob_pool->add_job( [this, first_index, last_index] { |
| 1047 | #endif |
| 1048 | |
| 1049 | for (uint32_t cluster_index = first_index; cluster_index < last_index; cluster_index++) |
| 1050 | { |
| 1051 | const basisu::vector<uint32_t>& cluster_indices = m_endpoint_clusters[cluster_index]; |
| 1052 | |
| 1053 | assert(cluster_indices.size()); |
| 1054 | |
| 1055 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1056 | { |
| 1057 | basisu::vector<color_rgba> cluster_pixels(8); |
| 1058 | |
| 1059 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1060 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1061 | |
| 1062 | const bool flipped = true; |
| 1063 | |
| 1064 | const color_rgba *pSource_block_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 1065 | |
| 1066 | for (uint32_t pixel_index = 0; pixel_index < 8; pixel_index++) |
| 1067 | { |
| 1068 | cluster_pixels[pixel_index] = pSource_block_pixels[g_etc1_pixel_indices[flipped][subblock_index][pixel_index]]; |
| 1069 | } |
| 1070 | |
| 1071 | const endpoint_cluster_etc_params &etc_params = m_endpoint_cluster_etc_params[cluster_index]; |
| 1072 | |
| 1073 | assert(etc_params.m_valid); |
| 1074 | |
| 1075 | color_rgba block_colors[4]; |
| 1076 | etc_block::get_block_colors5(block_colors, etc_params.m_color_unscaled[0], etc_params.m_inten_table[0], true); |
| 1077 | |
| 1078 | uint64_t total_err = 0; |
| 1079 | |
| 1080 | for (uint32_t i = 0; i < 8; i++) |
| 1081 | { |
| 1082 | const color_rgba &c = cluster_pixels[i]; |
| 1083 | |
| 1084 | uint64_t best_err = UINT64_MAX; |
| 1085 | //uint32_t best_index = 0; |
| 1086 | |
| 1087 | for (uint32_t s = 0; s < 4; s++) |
| 1088 | { |
| 1089 | uint64_t err = color_distance(m_params.m_perceptual, c, block_colors[s], false); |
| 1090 | if (err < best_err) |
| 1091 | { |
| 1092 | best_err = err; |
| 1093 | //best_index = s; |
| 1094 | } |
| 1095 | } |
| 1096 | |
| 1097 | total_err += best_err; |
| 1098 | } |
| 1099 | |
| 1100 | subblock_endpoint_quant_err quant_err; |
| 1101 | quant_err.m_total_err = total_err; |
| 1102 | quant_err.m_cluster_index = cluster_index; |
| 1103 | quant_err.m_cluster_subblock_index = cluster_indices_iter; |
| 1104 | quant_err.m_block_index = block_index; |
| 1105 | quant_err.m_subblock_index = subblock_index; |
| 1106 | |
| 1107 | { |
| 1108 | std::lock_guard<std::mutex> lock(m_lock); |
| 1109 | |
| 1110 | m_subblock_endpoint_quant_err_vec.push_back(quant_err); |
| 1111 | } |
| 1112 | } |
| 1113 | } // cluster_index |
| 1114 | |
| 1115 | #ifndef __EMSCRIPTEN__ |
| 1116 | } ); |
| 1117 | #endif |
| 1118 | |
| 1119 | } // cluster_index_iter |
| 1120 | |
| 1121 | #ifndef __EMSCRIPTEN__ |
| 1122 | m_params.m_pJob_pool->wait_for_all(); |
| 1123 | #endif |
| 1124 | |
| 1125 | vector_sort(m_subblock_endpoint_quant_err_vec); |
| 1126 | } |
| 1127 | |
| 1128 | void basisu_frontend::introduce_new_endpoint_clusters() |
| 1129 | { |
| 1130 | debug_printf("introduce_new_endpoint_clusters\n"); |
| 1131 | |
| 1132 | generate_block_endpoint_clusters(); |
| 1133 | |
| 1134 | int num_new_endpoint_clusters = m_params.m_max_endpoint_clusters - (uint32_t)m_endpoint_clusters.size(); |
| 1135 | if (num_new_endpoint_clusters <= 0) |
| 1136 | return; |
| 1137 | |
| 1138 | compute_endpoint_subblock_error_vec(); |
| 1139 | |
| 1140 | const uint32_t num_orig_endpoint_clusters = (uint32_t)m_endpoint_clusters.size(); |
| 1141 | |
| 1142 | std::unordered_set<uint32_t> training_vector_was_relocated; |
| 1143 | |
| 1144 | uint_vec cluster_sizes(num_orig_endpoint_clusters); |
| 1145 | for (uint32_t i = 0; i < num_orig_endpoint_clusters; i++) |
| 1146 | cluster_sizes[i] = (uint32_t)m_endpoint_clusters[i].size(); |
| 1147 | |
| 1148 | std::unordered_set<uint32_t> ignore_cluster; |
| 1149 | |
| 1150 | uint32_t total_new_clusters = 0; |
| 1151 | |
| 1152 | while (num_new_endpoint_clusters) |
| 1153 | { |
| 1154 | if (m_subblock_endpoint_quant_err_vec.size() == 0) |
| 1155 | break; |
| 1156 | |
| 1157 | subblock_endpoint_quant_err subblock_to_move(m_subblock_endpoint_quant_err_vec.back()); |
| 1158 | |
| 1159 | m_subblock_endpoint_quant_err_vec.pop_back(); |
| 1160 | |
| 1161 | if (unordered_set_contains(ignore_cluster, subblock_to_move.m_cluster_index)) |
| 1162 | continue; |
| 1163 | |
| 1164 | uint32_t training_vector_index = subblock_to_move.m_block_index * 2 + subblock_to_move.m_subblock_index; |
| 1165 | |
| 1166 | if (cluster_sizes[subblock_to_move.m_cluster_index] <= 2) |
| 1167 | continue; |
| 1168 | |
| 1169 | if (unordered_set_contains(training_vector_was_relocated, training_vector_index)) |
| 1170 | continue; |
| 1171 | |
| 1172 | if (unordered_set_contains(training_vector_was_relocated, training_vector_index ^ 1)) |
| 1173 | continue; |
| 1174 | |
| 1175 | #if 0 |
| 1176 | const uint32_t block_index = subblock_to_move.m_block_index; |
| 1177 | const etc_block& blk = m_etc1_blocks_etc1s[block_index]; |
| 1178 | uint32_t ls, hs; |
| 1179 | blk.get_selector_range(ls, hs); |
| 1180 | if (ls != hs) |
| 1181 | continue; |
| 1182 | #endif |
| 1183 | |
| 1184 | //const uint32_t new_endpoint_cluster_index = (uint32_t)m_endpoint_clusters.size(); |
| 1185 | |
| 1186 | enlarge_vector(m_endpoint_clusters, 1)->push_back(training_vector_index); |
| 1187 | enlarge_vector(m_endpoint_cluster_etc_params, 1); |
| 1188 | |
| 1189 | assert(m_endpoint_clusters.size() == m_endpoint_cluster_etc_params.size()); |
| 1190 | |
| 1191 | training_vector_was_relocated.insert(training_vector_index); |
| 1192 | |
| 1193 | m_endpoint_clusters.back().push_back(training_vector_index ^ 1); |
| 1194 | training_vector_was_relocated.insert(training_vector_index ^ 1); |
| 1195 | |
| 1196 | BASISU_FRONTEND_VERIFY(cluster_sizes[subblock_to_move.m_cluster_index] >= 2); |
| 1197 | cluster_sizes[subblock_to_move.m_cluster_index] -= 2; |
| 1198 | |
| 1199 | ignore_cluster.insert(subblock_to_move.m_cluster_index); |
| 1200 | |
| 1201 | total_new_clusters++; |
| 1202 | |
| 1203 | num_new_endpoint_clusters--; |
| 1204 | } |
| 1205 | |
| 1206 | debug_printf("Introduced %i new endpoint clusters\n", total_new_clusters); |
| 1207 | |
| 1208 | for (uint32_t i = 0; i < num_orig_endpoint_clusters; i++) |
| 1209 | { |
| 1210 | uint_vec &cluster_indices = m_endpoint_clusters[i]; |
| 1211 | |
| 1212 | uint_vec new_cluster_indices; |
| 1213 | for (uint32_t j = 0; j < cluster_indices.size(); j++) |
| 1214 | { |
| 1215 | uint32_t training_vector_index = cluster_indices[j]; |
| 1216 | |
| 1217 | if (!unordered_set_contains(training_vector_was_relocated, training_vector_index)) |
| 1218 | new_cluster_indices.push_back(training_vector_index); |
| 1219 | } |
| 1220 | |
| 1221 | if (cluster_indices.size() != new_cluster_indices.size()) |
| 1222 | { |
| 1223 | BASISU_FRONTEND_VERIFY(new_cluster_indices.size() > 0); |
| 1224 | cluster_indices.swap(new_cluster_indices); |
| 1225 | } |
| 1226 | } |
| 1227 | |
| 1228 | generate_block_endpoint_clusters(); |
| 1229 | } |
| 1230 | |
| 1231 | struct color_rgba_hasher |
| 1232 | { |
| 1233 | inline std::size_t operator()(const color_rgba& k) const |
| 1234 | { |
| 1235 | uint32_t v = *(const uint32_t*)&k; |
| 1236 | |
| 1237 | //return bitmix32(v); |
| 1238 | |
| 1239 | //v ^= (v << 10); |
| 1240 | //v ^= (v >> 12); |
| 1241 | |
| 1242 | return v; |
| 1243 | } |
| 1244 | }; |
| 1245 | |
| 1246 | // Given each endpoint cluster, gather all the block pixels which are in that cluster and compute optimized ETC1S endpoints for them. |
| 1247 | // TODO: Don't optimize endpoint clusters which haven't changed. |
| 1248 | // If step>=1, we check to ensure the new endpoint values actually decrease quantization error. |
| 1249 | void basisu_frontend::generate_endpoint_codebook(uint32_t step) |
| 1250 | { |
| 1251 | debug_printf("generate_endpoint_codebook\n"); |
| 1252 | |
| 1253 | interval_timer tm; |
| 1254 | tm.start(); |
| 1255 | |
| 1256 | m_endpoint_cluster_etc_params.resize(m_endpoint_clusters.size()); |
| 1257 | |
| 1258 | bool use_cpu = true; |
| 1259 | // TODO: Get this working when step>0 |
| 1260 | if (m_params.m_pOpenCL_context && !step) |
| 1261 | { |
| 1262 | const uint32_t total_clusters = m_endpoint_clusters.size(); |
| 1263 | |
| 1264 | basisu::vector<cl_pixel_cluster> pixel_clusters(total_clusters); |
| 1265 | |
| 1266 | std::vector<color_rgba> input_pixels; |
| 1267 | input_pixels.reserve(m_total_blocks * 16); |
| 1268 | |
| 1269 | std::vector<uint32_t> pixel_weights; |
| 1270 | pixel_weights.reserve(m_total_blocks * 16); |
| 1271 | |
| 1272 | uint_vec cluster_sizes(total_clusters); |
| 1273 | |
| 1274 | //typedef basisu::hash_map<color_rgba, uint32_t, color_rgba_hasher> color_hasher_type; |
| 1275 | //color_hasher_type color_hasher; |
| 1276 | //color_hasher.reserve(2048); |
| 1277 | |
| 1278 | interval_timer hash_tm; |
| 1279 | hash_tm.start(); |
| 1280 | |
| 1281 | basisu::vector<uint32_t> colors, colors2; |
| 1282 | colors.reserve(65536); |
| 1283 | colors2.reserve(65536); |
| 1284 | |
| 1285 | for (uint32_t cluster_index = 0; cluster_index < m_endpoint_clusters.size(); cluster_index++) |
| 1286 | { |
| 1287 | const basisu::vector<uint32_t>& cluster_indices = m_endpoint_clusters[cluster_index]; |
| 1288 | assert((cluster_indices.size() & 1) == 0); |
| 1289 | |
| 1290 | #if 0 |
| 1291 | uint64_t first_pixel_index = input_pixels.size(); |
| 1292 | const uint32_t total_pixels = 16 * (cluster_indices.size() / 2); |
| 1293 | |
| 1294 | input_pixels.resize(input_pixels.size() + total_pixels); |
| 1295 | pixel_weights.resize(pixel_weights.size() + total_pixels); |
| 1296 | |
| 1297 | uint64_t dst_ofs = first_pixel_index; |
| 1298 | |
| 1299 | uint64_t total_r = 0, total_g = 0, total_b = 0; |
| 1300 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1301 | { |
| 1302 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1303 | if (subblock_index) |
| 1304 | continue; |
| 1305 | |
| 1306 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1307 | const color_rgba* pBlock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 1308 | |
| 1309 | for (uint32_t i = 0; i < 16; i++) |
| 1310 | { |
| 1311 | input_pixels[dst_ofs] = pBlock_pixels[i]; |
| 1312 | pixel_weights[dst_ofs] = 1; |
| 1313 | dst_ofs++; |
| 1314 | |
| 1315 | total_r += pBlock_pixels[i].r; |
| 1316 | total_g += pBlock_pixels[i].g; |
| 1317 | total_b += pBlock_pixels[i].b; |
| 1318 | } |
| 1319 | } |
| 1320 | |
| 1321 | //printf("%i %f %f %f\n", cluster_index, total_r / (float)total_pixels, total_g / (float)total_pixels, total_b / (float)total_pixels); |
| 1322 | |
| 1323 | pixel_clusters[cluster_index].m_first_pixel_index = first_pixel_index; |
| 1324 | pixel_clusters[cluster_index].m_total_pixels = total_pixels; |
| 1325 | cluster_sizes[cluster_index] = total_pixels; |
| 1326 | #elif 1 |
| 1327 | colors.resize(cluster_indices.size() * 8); |
| 1328 | colors2.resize(cluster_indices.size() * 8); |
| 1329 | uint32_t dst_ofs = 0; |
| 1330 | |
| 1331 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1332 | { |
| 1333 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1334 | if (subblock_index) |
| 1335 | continue; |
| 1336 | |
| 1337 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1338 | const color_rgba* pBlock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 1339 | |
| 1340 | memcpy(colors.data() + dst_ofs, pBlock_pixels, sizeof(color_rgba) * 16); |
| 1341 | dst_ofs += 16; |
| 1342 | |
| 1343 | } // cluster_indices_iter |
| 1344 | |
| 1345 | uint32_t* pSorted = radix_sort(colors.size(), colors.data(), colors2.data(), 0, 3); |
| 1346 | |
| 1347 | const uint64_t first_pixel_index = input_pixels.size(); |
| 1348 | |
| 1349 | uint32_t prev_color = 0, cur_weight = 0; |
| 1350 | |
| 1351 | for (uint32_t i = 0; i < colors.size(); i++) |
| 1352 | { |
| 1353 | uint32_t cur_color = pSorted[i]; |
| 1354 | if (cur_color == prev_color) |
| 1355 | { |
| 1356 | if (++cur_weight == 0) |
| 1357 | cur_weight--; |
| 1358 | } |
| 1359 | else |
| 1360 | { |
| 1361 | if (cur_weight) |
| 1362 | { |
| 1363 | input_pixels.push_back(*(const color_rgba*)&prev_color); |
| 1364 | pixel_weights.push_back(cur_weight); |
| 1365 | } |
| 1366 | |
| 1367 | prev_color = cur_color; |
| 1368 | cur_weight = 1; |
| 1369 | } |
| 1370 | } |
| 1371 | |
| 1372 | if (cur_weight) |
| 1373 | { |
| 1374 | input_pixels.push_back(*(const color_rgba*)&prev_color); |
| 1375 | pixel_weights.push_back(cur_weight); |
| 1376 | } |
| 1377 | |
| 1378 | uint32_t total_unique_pixels = (uint32_t)(input_pixels.size() - first_pixel_index); |
| 1379 | |
| 1380 | pixel_clusters[cluster_index].m_first_pixel_index = first_pixel_index; |
| 1381 | pixel_clusters[cluster_index].m_total_pixels = total_unique_pixels; |
| 1382 | |
| 1383 | cluster_sizes[cluster_index] = total_unique_pixels; |
| 1384 | #else |
| 1385 | color_hasher.reset(); |
| 1386 | |
| 1387 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1388 | { |
| 1389 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1390 | if (subblock_index) |
| 1391 | continue; |
| 1392 | |
| 1393 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1394 | const color_rgba* pBlock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 1395 | |
| 1396 | uint32_t *pPrev_weight = nullptr; |
| 1397 | color_rgba prev_color; |
| 1398 | |
| 1399 | { |
| 1400 | color_rgba cur_color = pBlock_pixels[0]; |
| 1401 | auto res = color_hasher.insert(cur_color, 0); |
| 1402 | |
| 1403 | uint32_t& weight = (res.first)->second; |
| 1404 | if (weight != UINT32_MAX) |
| 1405 | weight++; |
| 1406 | |
| 1407 | prev_color = cur_color; |
| 1408 | pPrev_weight = &(res.first)->second; |
| 1409 | } |
| 1410 | |
| 1411 | for (uint32_t i = 1; i < 16; i++) |
| 1412 | { |
| 1413 | color_rgba cur_color = pBlock_pixels[i]; |
| 1414 | |
| 1415 | if (cur_color == prev_color) |
| 1416 | { |
| 1417 | if (*pPrev_weight != UINT32_MAX) |
| 1418 | *pPrev_weight = *pPrev_weight + 1; |
| 1419 | } |
| 1420 | else |
| 1421 | { |
| 1422 | auto res = color_hasher.insert(cur_color, 0); |
| 1423 | |
| 1424 | uint32_t& weight = (res.first)->second; |
| 1425 | if (weight != UINT32_MAX) |
| 1426 | weight++; |
| 1427 | |
| 1428 | prev_color = cur_color; |
| 1429 | pPrev_weight = &(res.first)->second; |
| 1430 | } |
| 1431 | } |
| 1432 | |
| 1433 | } // cluster_indices_iter |
| 1434 | |
| 1435 | const uint64_t first_pixel_index = input_pixels.size(); |
| 1436 | uint32_t total_unique_pixels = color_hasher.size(); |
| 1437 | |
| 1438 | pixel_clusters[cluster_index].m_first_pixel_index = first_pixel_index; |
| 1439 | pixel_clusters[cluster_index].m_total_pixels = total_unique_pixels; |
| 1440 | |
| 1441 | input_pixels.resize(first_pixel_index + total_unique_pixels); |
| 1442 | pixel_weights.resize(first_pixel_index + total_unique_pixels); |
| 1443 | |
| 1444 | uint32_t j = 0; |
| 1445 | |
| 1446 | for (auto it = color_hasher.begin(); it != color_hasher.end(); ++it, ++j) |
| 1447 | { |
| 1448 | input_pixels[first_pixel_index + j] = it->first; |
| 1449 | pixel_weights[first_pixel_index + j] = it->second; |
| 1450 | } |
| 1451 | |
| 1452 | cluster_sizes[cluster_index] = total_unique_pixels; |
| 1453 | #endif |
| 1454 | |
| 1455 | } // cluster_index |
| 1456 | |
| 1457 | debug_printf("Total hash time: %3.3f secs\n", hash_tm.get_elapsed_secs()); |
| 1458 | |
| 1459 | debug_printf("Total unique colors: %llu\n", input_pixels.size()); |
| 1460 | |
| 1461 | uint_vec sorted_cluster_indices_new_to_old(total_clusters); |
| 1462 | indirect_sort(total_clusters, sorted_cluster_indices_new_to_old.data(), cluster_sizes.data()); |
| 1463 | //for (uint32_t i = 0; i < total_clusters; i++) |
| 1464 | // sorted_cluster_indices_new_to_old[i] = i; |
| 1465 | |
| 1466 | uint_vec sorted_cluster_indices_old_to_new(total_clusters); |
| 1467 | for (uint32_t i = 0; i < total_clusters; i++) |
| 1468 | sorted_cluster_indices_old_to_new[sorted_cluster_indices_new_to_old[i]] = i; |
| 1469 | |
| 1470 | basisu::vector<cl_pixel_cluster> sorted_pixel_clusters(total_clusters); |
| 1471 | for (uint32_t i = 0; i < total_clusters; i++) |
| 1472 | sorted_pixel_clusters[i] = pixel_clusters[sorted_cluster_indices_new_to_old[i]]; |
| 1473 | |
| 1474 | uint32_t total_perms = 64; |
| 1475 | if (m_params.m_compression_level <= 1) |
| 1476 | total_perms = 16; |
| 1477 | else if (m_params.m_compression_level == BASISU_MAX_COMPRESSION_LEVEL) |
| 1478 | total_perms = OPENCL_ENCODE_ETC1S_MAX_PERMS; |
| 1479 | |
| 1480 | basisu::vector<etc_block> output_blocks(total_clusters); |
| 1481 | |
| 1482 | if (opencl_encode_etc1s_pixel_clusters( |
| 1483 | m_params.m_pOpenCL_context, |
| 1484 | output_blocks.data(), |
| 1485 | total_clusters, |
| 1486 | sorted_pixel_clusters.data(), |
| 1487 | input_pixels.size(), |
| 1488 | input_pixels.data(), |
| 1489 | pixel_weights.data(), |
| 1490 | m_params.m_perceptual, total_perms)) |
| 1491 | { |
| 1492 | for (uint32_t old_cluster_index = 0; old_cluster_index < m_endpoint_clusters.size(); old_cluster_index++) |
| 1493 | { |
| 1494 | const uint32_t new_cluster_index = sorted_cluster_indices_old_to_new[old_cluster_index]; |
| 1495 | |
| 1496 | const etc_block& blk = output_blocks[new_cluster_index]; |
| 1497 | |
| 1498 | endpoint_cluster_etc_params& prev_etc_params = m_endpoint_cluster_etc_params[old_cluster_index]; |
| 1499 | |
| 1500 | prev_etc_params.m_valid = true; |
| 1501 | etc_block::unpack_color5(prev_etc_params.m_color_unscaled[0], blk.get_base5_color(), false); |
| 1502 | prev_etc_params.m_inten_table[0] = blk.get_inten_table(0); |
| 1503 | prev_etc_params.m_color_error[0] = 0; // dummy value - we don't actually use this |
| 1504 | } |
| 1505 | |
| 1506 | use_cpu = false; |
| 1507 | } |
| 1508 | else |
| 1509 | { |
| 1510 | error_printf("basisu_frontend::generate_endpoint_codebook: opencl_encode_etc1s_pixel_clusters() failed! Using CPU.\n"); |
| 1511 | m_params.m_pOpenCL_context = nullptr; |
| 1512 | m_opencl_failed = true; |
| 1513 | } |
| 1514 | |
| 1515 | } // if (opencl_is_available() && m_params.m_use_opencl) |
| 1516 | |
| 1517 | if (use_cpu) |
| 1518 | { |
| 1519 | const uint32_t N = 128; |
| 1520 | for (uint32_t cluster_index_iter = 0; cluster_index_iter < m_endpoint_clusters.size(); cluster_index_iter += N) |
| 1521 | { |
| 1522 | const uint32_t first_index = cluster_index_iter; |
| 1523 | const uint32_t last_index = minimum<uint32_t>((uint32_t)m_endpoint_clusters.size(), cluster_index_iter + N); |
| 1524 | |
| 1525 | #ifndef __EMSCRIPTEN__ |
| 1526 | m_params.m_pJob_pool->add_job([this, first_index, last_index, step] { |
| 1527 | #endif |
| 1528 | |
| 1529 | for (uint32_t cluster_index = first_index; cluster_index < last_index; cluster_index++) |
| 1530 | { |
| 1531 | const basisu::vector<uint32_t>& cluster_indices = m_endpoint_clusters[cluster_index]; |
| 1532 | |
| 1533 | BASISU_FRONTEND_VERIFY(cluster_indices.size()); |
| 1534 | |
| 1535 | const uint32_t total_pixels = (uint32_t)cluster_indices.size() * 8; |
| 1536 | |
| 1537 | basisu::vector<color_rgba> cluster_pixels(total_pixels); |
| 1538 | |
| 1539 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1540 | { |
| 1541 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1542 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1543 | |
| 1544 | const bool flipped = true; |
| 1545 | |
| 1546 | const color_rgba* pBlock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 1547 | |
| 1548 | for (uint32_t pixel_index = 0; pixel_index < 8; pixel_index++) |
| 1549 | { |
| 1550 | const color_rgba& c = pBlock_pixels[g_etc1_pixel_indices[flipped][subblock_index][pixel_index]]; |
| 1551 | cluster_pixels[cluster_indices_iter * 8 + pixel_index] = c; |
| 1552 | } |
| 1553 | } |
| 1554 | |
| 1555 | endpoint_cluster_etc_params new_subblock_params; |
| 1556 | |
| 1557 | { |
| 1558 | etc1_optimizer optimizer; |
| 1559 | etc1_solution_coordinates solutions[2]; |
| 1560 | |
| 1561 | etc1_optimizer::params cluster_optimizer_params; |
| 1562 | cluster_optimizer_params.m_num_src_pixels = total_pixels; |
| 1563 | cluster_optimizer_params.m_pSrc_pixels = &cluster_pixels[0]; |
| 1564 | |
| 1565 | cluster_optimizer_params.m_use_color4 = false; |
| 1566 | cluster_optimizer_params.m_perceptual = m_params.m_perceptual; |
| 1567 | |
| 1568 | if (m_params.m_compression_level <= 1) |
| 1569 | cluster_optimizer_params.m_quality = cETCQualityMedium; |
| 1570 | else if (m_params.m_compression_level == BASISU_MAX_COMPRESSION_LEVEL) |
| 1571 | cluster_optimizer_params.m_quality = cETCQualityUber; |
| 1572 | |
| 1573 | etc1_optimizer::results cluster_optimizer_results; |
| 1574 | |
| 1575 | basisu::vector<uint8_t> cluster_selectors(total_pixels); |
| 1576 | cluster_optimizer_results.m_n = total_pixels; |
| 1577 | cluster_optimizer_results.m_pSelectors = &cluster_selectors[0]; |
| 1578 | |
| 1579 | optimizer.init(cluster_optimizer_params, cluster_optimizer_results); |
| 1580 | |
| 1581 | if (!optimizer.compute()) |
| 1582 | BASISU_FRONTEND_VERIFY(false); |
| 1583 | |
| 1584 | new_subblock_params.m_color_unscaled[0] = cluster_optimizer_results.m_block_color_unscaled; |
| 1585 | new_subblock_params.m_inten_table[0] = cluster_optimizer_results.m_block_inten_table; |
| 1586 | new_subblock_params.m_color_error[0] = cluster_optimizer_results.m_error; |
| 1587 | } |
| 1588 | |
| 1589 | endpoint_cluster_etc_params& prev_etc_params = m_endpoint_cluster_etc_params[cluster_index]; |
| 1590 | |
| 1591 | bool use_new_subblock_params = false; |
| 1592 | if ((!step) || (!prev_etc_params.m_valid)) |
| 1593 | use_new_subblock_params = true; |
| 1594 | else |
| 1595 | { |
| 1596 | assert(prev_etc_params.m_valid); |
| 1597 | |
| 1598 | uint64_t total_prev_err = 0; |
| 1599 | |
| 1600 | { |
| 1601 | color_rgba block_colors[4]; |
| 1602 | |
| 1603 | etc_block::get_block_colors5(block_colors, prev_etc_params.m_color_unscaled[0], prev_etc_params.m_inten_table[0], false); |
| 1604 | |
| 1605 | uint64_t total_err = 0; |
| 1606 | |
| 1607 | for (uint32_t i = 0; i < total_pixels; i++) |
| 1608 | { |
| 1609 | const color_rgba& c = cluster_pixels[i]; |
| 1610 | |
| 1611 | uint64_t best_err = UINT64_MAX; |
| 1612 | //uint32_t best_index = 0; |
| 1613 | |
| 1614 | for (uint32_t s = 0; s < 4; s++) |
| 1615 | { |
| 1616 | uint64_t err = color_distance(m_params.m_perceptual, c, block_colors[s], false); |
| 1617 | if (err < best_err) |
| 1618 | { |
| 1619 | best_err = err; |
| 1620 | //best_index = s; |
| 1621 | } |
| 1622 | } |
| 1623 | |
| 1624 | total_err += best_err; |
| 1625 | } |
| 1626 | |
| 1627 | total_prev_err += total_err; |
| 1628 | } |
| 1629 | |
| 1630 | // See if we should update this cluster's endpoints (if the error has actually fallen) |
| 1631 | if (total_prev_err > new_subblock_params.m_color_error[0]) |
| 1632 | { |
| 1633 | use_new_subblock_params = true; |
| 1634 | } |
| 1635 | } |
| 1636 | |
| 1637 | if (use_new_subblock_params) |
| 1638 | { |
| 1639 | new_subblock_params.m_valid = true; |
| 1640 | |
| 1641 | prev_etc_params = new_subblock_params; |
| 1642 | } |
| 1643 | |
| 1644 | } // cluster_index |
| 1645 | |
| 1646 | #ifndef __EMSCRIPTEN__ |
| 1647 | }); |
| 1648 | #endif |
| 1649 | |
| 1650 | } // cluster_index_iter |
| 1651 | |
| 1652 | #ifndef __EMSCRIPTEN__ |
| 1653 | m_params.m_pJob_pool->wait_for_all(); |
| 1654 | #endif |
| 1655 | } |
| 1656 | |
| 1657 | debug_printf("Elapsed time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 1658 | } |
| 1659 | |
| 1660 | bool basisu_frontend::check_etc1s_constraints() const |
| 1661 | { |
| 1662 | basisu::vector<vec2U> block_clusters(m_total_blocks); |
| 1663 | |
| 1664 | for (int cluster_index = 0; cluster_index < static_cast<int>(m_endpoint_clusters.size()); cluster_index++) |
| 1665 | { |
| 1666 | const basisu::vector<uint32_t>& cluster_indices = m_endpoint_clusters[cluster_index]; |
| 1667 | |
| 1668 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1669 | { |
| 1670 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1671 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1672 | |
| 1673 | block_clusters[block_index][subblock_index] = cluster_index; |
| 1674 | |
| 1675 | } // cluster_indices_iter |
| 1676 | } |
| 1677 | |
| 1678 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 1679 | { |
| 1680 | if (block_clusters[i][0] != block_clusters[i][1]) |
| 1681 | return false; |
| 1682 | } |
| 1683 | |
| 1684 | return true; |
| 1685 | } |
| 1686 | |
| 1687 | // For each block, determine which ETC1S endpoint cluster can encode that block with lowest error. |
| 1688 | // This reassigns blocks to different endpoint clusters. |
| 1689 | uint32_t basisu_frontend::refine_endpoint_clusterization() |
| 1690 | { |
| 1691 | debug_printf("refine_endpoint_clusterization\n"); |
| 1692 | |
| 1693 | if (m_use_hierarchical_endpoint_codebooks) |
| 1694 | compute_endpoint_clusters_within_each_parent_cluster(); |
| 1695 | |
| 1696 | // Note: It's possible that an endpoint cluster may live in more than one parent cluster after the first refinement step. |
| 1697 | |
| 1698 | basisu::vector<vec2U> block_clusters(m_total_blocks); |
| 1699 | |
| 1700 | for (int cluster_index = 0; cluster_index < static_cast<int>(m_endpoint_clusters.size()); cluster_index++) |
| 1701 | { |
| 1702 | const basisu::vector<uint32_t>& cluster_indices = m_endpoint_clusters[cluster_index]; |
| 1703 | |
| 1704 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 1705 | { |
| 1706 | const uint32_t block_index = cluster_indices[cluster_indices_iter] >> 1; |
| 1707 | const uint32_t subblock_index = cluster_indices[cluster_indices_iter] & 1; |
| 1708 | |
| 1709 | block_clusters[block_index][subblock_index] = cluster_index; |
| 1710 | |
| 1711 | } // cluster_indices_iter |
| 1712 | } |
| 1713 | |
| 1714 | //---------------------------------------------------------- |
| 1715 | |
| 1716 | // Create a new endpoint clusterization |
| 1717 | |
| 1718 | interval_timer tm; |
| 1719 | tm.start(); |
| 1720 | |
| 1721 | uint_vec best_cluster_indices(m_total_blocks); |
| 1722 | |
| 1723 | bool use_cpu = true; |
| 1724 | // TODO: Support non-hierarchical endpoint codebooks here |
| 1725 | if (m_params.m_pOpenCL_context && m_use_hierarchical_endpoint_codebooks) |
| 1726 | { |
| 1727 | // For the OpenCL kernel, we order the parent endpoint clusters by smallest to largest for efficiency. |
| 1728 | // We also prepare an array of block info structs that point into this new parent endpoint cluster array. |
| 1729 | const uint32_t total_parent_clusters = m_endpoint_clusters_within_each_parent_cluster.size(); |
| 1730 | |
| 1731 | basisu::vector<cl_block_info_struct> cl_block_info_structs(m_total_blocks); |
| 1732 | |
| 1733 | // the size of each parent cluster, in total clusters |
| 1734 | uint_vec parent_cluster_sizes(total_parent_clusters); |
| 1735 | for (uint32_t i = 0; i < total_parent_clusters; i++) |
| 1736 | parent_cluster_sizes[i] = m_endpoint_clusters_within_each_parent_cluster[i].size(); |
| 1737 | |
| 1738 | uint_vec first_parent_cluster_ofs(total_parent_clusters); |
| 1739 | uint32_t cur_ofs = 0; |
| 1740 | for (uint32_t i = 0; i < total_parent_clusters; i++) |
| 1741 | { |
| 1742 | first_parent_cluster_ofs[i] = cur_ofs; |
| 1743 | |
| 1744 | cur_ofs += parent_cluster_sizes[i]; |
| 1745 | } |
| 1746 | |
| 1747 | // Note: total_actual_endpoint_clusters is not necessarly equal to m_endpoint_clusters.size(), because clusters may live in multiple parent clusters after the first refinement step. |
| 1748 | BASISU_FRONTEND_VERIFY(cur_ofs >= m_endpoint_clusters.size()); |
| 1749 | const uint32_t total_actual_endpoint_clusters = cur_ofs; |
| 1750 | basisu::vector<cl_endpoint_cluster_struct> cl_endpoint_cluster_structs(total_actual_endpoint_clusters); |
| 1751 | |
| 1752 | for (uint32_t i = 0; i < total_parent_clusters; i++) |
| 1753 | { |
| 1754 | const uint32_t dst_ofs = first_parent_cluster_ofs[i]; |
| 1755 | |
| 1756 | const uint32_t parent_cluster_size = parent_cluster_sizes[i]; |
| 1757 | |
| 1758 | assert(m_endpoint_clusters_within_each_parent_cluster[i].size() == parent_cluster_size); |
| 1759 | |
| 1760 | for (uint32_t j = 0; j < parent_cluster_size; j++) |
| 1761 | { |
| 1762 | const uint32_t endpoint_cluster_index = m_endpoint_clusters_within_each_parent_cluster[i][j]; |
| 1763 | |
| 1764 | color_rgba cluster_etc_base_color(m_endpoint_cluster_etc_params[endpoint_cluster_index].m_color_unscaled[0]); |
| 1765 | uint32_t cluster_etc_inten = m_endpoint_cluster_etc_params[endpoint_cluster_index].m_inten_table[0]; |
| 1766 | |
| 1767 | cl_endpoint_cluster_structs[dst_ofs + j].m_unscaled_color = cluster_etc_base_color; |
| 1768 | cl_endpoint_cluster_structs[dst_ofs + j].m_etc_inten = (uint8_t)cluster_etc_inten; |
| 1769 | cl_endpoint_cluster_structs[dst_ofs + j].m_cluster_index = (uint16_t)endpoint_cluster_index; |
| 1770 | } |
| 1771 | } |
| 1772 | |
| 1773 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 1774 | { |
| 1775 | const uint32_t block_parent_endpoint_cluster_index = m_block_parent_endpoint_cluster[block_index]; |
| 1776 | |
| 1777 | cl_block_info_structs[block_index].m_num_clusters = (uint16_t)(parent_cluster_sizes[block_parent_endpoint_cluster_index]); |
| 1778 | cl_block_info_structs[block_index].m_first_cluster_ofs = (uint16_t)(first_parent_cluster_ofs[block_parent_endpoint_cluster_index]); |
| 1779 | |
| 1780 | const uint32_t block_cluster_index = block_clusters[block_index][0]; |
| 1781 | cl_block_info_structs[block_index].m_cur_cluster_index = (uint16_t)block_cluster_index; |
| 1782 | cl_block_info_structs[block_index].m_cur_cluster_etc_inten = (uint8_t)m_endpoint_cluster_etc_params[block_cluster_index].m_inten_table[0]; |
| 1783 | } |
| 1784 | |
| 1785 | uint_vec block_cluster_indices(m_total_blocks); |
| 1786 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 1787 | block_cluster_indices[i] = block_clusters[i][0]; |
| 1788 | |
| 1789 | uint_vec sorted_block_indices(m_total_blocks); |
| 1790 | indirect_sort(m_total_blocks, sorted_block_indices.data(), block_cluster_indices.data()); |
| 1791 | |
| 1792 | bool status = opencl_refine_endpoint_clusterization( |
| 1793 | m_params.m_pOpenCL_context, |
| 1794 | cl_block_info_structs.data(), |
| 1795 | total_actual_endpoint_clusters, |
| 1796 | cl_endpoint_cluster_structs.data(), |
| 1797 | sorted_block_indices.data(), |
| 1798 | best_cluster_indices.data(), |
| 1799 | m_params.m_perceptual); |
| 1800 | |
| 1801 | if (status) |
| 1802 | { |
| 1803 | use_cpu = false; |
| 1804 | } |
| 1805 | else |
| 1806 | { |
| 1807 | error_printf("basisu_frontend::refine_endpoint_clusterization: opencl_refine_endpoint_clusterization() failed! Using CPU.\n"); |
| 1808 | m_params.m_pOpenCL_context = nullptr; |
| 1809 | m_opencl_failed = true; |
| 1810 | } |
| 1811 | } |
| 1812 | |
| 1813 | if (use_cpu) |
| 1814 | { |
| 1815 | const uint32_t N = 1024; |
| 1816 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 1817 | { |
| 1818 | const uint32_t first_index = block_index_iter; |
| 1819 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 1820 | |
| 1821 | #ifndef __EMSCRIPTEN__ |
| 1822 | m_params.m_pJob_pool->add_job([this, first_index, last_index, &best_cluster_indices, &block_clusters] { |
| 1823 | #endif |
| 1824 | |
| 1825 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 1826 | { |
| 1827 | const uint32_t cluster_index = block_clusters[block_index][0]; |
| 1828 | BASISU_FRONTEND_VERIFY(cluster_index == block_clusters[block_index][1]); |
| 1829 | |
| 1830 | const color_rgba* pSubblock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 1831 | const uint32_t num_subblock_pixels = 16; |
| 1832 | |
| 1833 | uint64_t best_cluster_err = INT64_MAX; |
| 1834 | uint32_t best_cluster_index = 0; |
| 1835 | |
| 1836 | const uint32_t block_parent_endpoint_cluster_index = m_block_parent_endpoint_cluster.size() ? m_block_parent_endpoint_cluster[block_index] : 0; |
| 1837 | const uint_vec* pCluster_indices = m_endpoint_clusters_within_each_parent_cluster.size() ? &m_endpoint_clusters_within_each_parent_cluster[block_parent_endpoint_cluster_index] : nullptr; |
| 1838 | |
| 1839 | const uint32_t total_clusters = m_use_hierarchical_endpoint_codebooks ? (uint32_t)pCluster_indices->size() : (uint32_t)m_endpoint_clusters.size(); |
| 1840 | |
| 1841 | for (uint32_t i = 0; i < total_clusters; i++) |
| 1842 | { |
| 1843 | const uint32_t cluster_iter = m_use_hierarchical_endpoint_codebooks ? (*pCluster_indices)[i] : i; |
| 1844 | |
| 1845 | color_rgba cluster_etc_base_color(m_endpoint_cluster_etc_params[cluster_iter].m_color_unscaled[0]); |
| 1846 | uint32_t cluster_etc_inten = m_endpoint_cluster_etc_params[cluster_iter].m_inten_table[0]; |
| 1847 | |
| 1848 | uint64_t total_err = 0; |
| 1849 | |
| 1850 | const uint32_t low_selector = 0;//subblock_etc_params_vec[j].m_low_selectors[0]; |
| 1851 | const uint32_t high_selector = 3;//subblock_etc_params_vec[j].m_high_selectors[0]; |
| 1852 | color_rgba subblock_colors[4]; |
| 1853 | // Can't assign it here - may result in too much error when selector quant occurs |
| 1854 | if (cluster_etc_inten > m_endpoint_cluster_etc_params[cluster_index].m_inten_table[0]) |
| 1855 | { |
| 1856 | total_err = INT64_MAX; |
| 1857 | goto skip_cluster; |
| 1858 | } |
| 1859 | |
| 1860 | etc_block::get_block_colors5(subblock_colors, cluster_etc_base_color, cluster_etc_inten); |
| 1861 | |
| 1862 | #if 0 |
| 1863 | for (uint32_t p = 0; p < num_subblock_pixels; p++) |
| 1864 | { |
| 1865 | uint64_t best_err = UINT64_MAX; |
| 1866 | |
| 1867 | for (uint32_t r = low_selector; r <= high_selector; r++) |
| 1868 | { |
| 1869 | uint64_t err = color_distance(m_params.m_perceptual, pSubblock_pixels[p], subblock_colors[r], false); |
| 1870 | best_err = minimum(best_err, err); |
| 1871 | if (!best_err) |
| 1872 | break; |
| 1873 | } |
| 1874 | |
| 1875 | total_err += best_err; |
| 1876 | if (total_err > best_cluster_err) |
| 1877 | break; |
| 1878 | } // p |
| 1879 | #else |
| 1880 | if (m_params.m_perceptual) |
| 1881 | { |
| 1882 | if (!g_cpu_supports_sse41) |
| 1883 | { |
| 1884 | for (uint32_t p = 0; p < num_subblock_pixels; p++) |
| 1885 | { |
| 1886 | uint64_t best_err = UINT64_MAX; |
| 1887 | |
| 1888 | for (uint32_t r = low_selector; r <= high_selector; r++) |
| 1889 | { |
| 1890 | uint64_t err = color_distance(true, pSubblock_pixels[p], subblock_colors[r], false); |
| 1891 | best_err = minimum(best_err, err); |
| 1892 | if (!best_err) |
| 1893 | break; |
| 1894 | } |
| 1895 | |
| 1896 | total_err += best_err; |
| 1897 | if (total_err > best_cluster_err) |
| 1898 | break; |
| 1899 | } // p |
| 1900 | } |
| 1901 | else |
| 1902 | { |
| 1903 | #if BASISU_SUPPORT_SSE |
| 1904 | find_lowest_error_perceptual_rgb_4_N_sse41((int64_t*)&total_err, subblock_colors, pSubblock_pixels, num_subblock_pixels, best_cluster_err); |
| 1905 | #endif |
| 1906 | } |
| 1907 | } |
| 1908 | else |
| 1909 | { |
| 1910 | if (!g_cpu_supports_sse41) |
| 1911 | { |
| 1912 | for (uint32_t p = 0; p < num_subblock_pixels; p++) |
| 1913 | { |
| 1914 | uint64_t best_err = UINT64_MAX; |
| 1915 | |
| 1916 | for (uint32_t r = low_selector; r <= high_selector; r++) |
| 1917 | { |
| 1918 | uint64_t err = color_distance(false, pSubblock_pixels[p], subblock_colors[r], false); |
| 1919 | best_err = minimum(best_err, err); |
| 1920 | if (!best_err) |
| 1921 | break; |
| 1922 | } |
| 1923 | |
| 1924 | total_err += best_err; |
| 1925 | if (total_err > best_cluster_err) |
| 1926 | break; |
| 1927 | } // p |
| 1928 | } |
| 1929 | else |
| 1930 | { |
| 1931 | #if BASISU_SUPPORT_SSE |
| 1932 | find_lowest_error_linear_rgb_4_N_sse41((int64_t*)&total_err, subblock_colors, pSubblock_pixels, num_subblock_pixels, best_cluster_err); |
| 1933 | #endif |
| 1934 | } |
| 1935 | } |
| 1936 | #endif |
| 1937 | |
| 1938 | skip_cluster: |
| 1939 | if ((total_err < best_cluster_err) || |
| 1940 | ((cluster_iter == cluster_index) && (total_err == best_cluster_err))) |
| 1941 | { |
| 1942 | best_cluster_err = total_err; |
| 1943 | best_cluster_index = cluster_iter; |
| 1944 | |
| 1945 | if (!best_cluster_err) |
| 1946 | break; |
| 1947 | } |
| 1948 | } // j |
| 1949 | |
| 1950 | best_cluster_indices[block_index] = best_cluster_index; |
| 1951 | |
| 1952 | } // block_index |
| 1953 | |
| 1954 | #ifndef __EMSCRIPTEN__ |
| 1955 | }); |
| 1956 | #endif |
| 1957 | |
| 1958 | } // block_index_iter |
| 1959 | |
| 1960 | #ifndef __EMSCRIPTEN__ |
| 1961 | m_params.m_pJob_pool->wait_for_all(); |
| 1962 | #endif |
| 1963 | |
| 1964 | } // use_cpu |
| 1965 | |
| 1966 | debug_printf("refine_endpoint_clusterization time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 1967 | |
| 1968 | basisu::vector<typename basisu::vector<uint32_t> > optimized_endpoint_clusters(m_endpoint_clusters.size()); |
| 1969 | uint32_t total_subblocks_reassigned = 0; |
| 1970 | |
| 1971 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 1972 | { |
| 1973 | const uint32_t training_vector_index = block_index * 2 + 0; |
| 1974 | |
| 1975 | const uint32_t orig_cluster_index = block_clusters[block_index][0]; |
| 1976 | const uint32_t best_cluster_index = best_cluster_indices[block_index]; |
| 1977 | |
| 1978 | optimized_endpoint_clusters[best_cluster_index].push_back(training_vector_index); |
| 1979 | optimized_endpoint_clusters[best_cluster_index].push_back(training_vector_index + 1); |
| 1980 | |
| 1981 | if (best_cluster_index != orig_cluster_index) |
| 1982 | { |
| 1983 | total_subblocks_reassigned++; |
| 1984 | } |
| 1985 | } |
| 1986 | |
| 1987 | debug_printf("total_subblocks_reassigned: %u\n", total_subblocks_reassigned); |
| 1988 | |
| 1989 | m_endpoint_clusters = optimized_endpoint_clusters; |
| 1990 | |
| 1991 | return total_subblocks_reassigned; |
| 1992 | } |
| 1993 | |
| 1994 | void basisu_frontend::eliminate_redundant_or_empty_endpoint_clusters() |
| 1995 | { |
| 1996 | debug_printf("eliminate_redundant_or_empty_endpoint_clusters\n"); |
| 1997 | |
| 1998 | // Step 1: Sort endpoint clusters by the base colors/intens |
| 1999 | |
| 2000 | uint_vec sorted_endpoint_cluster_indices(m_endpoint_clusters.size()); |
| 2001 | for (uint32_t i = 0; i < m_endpoint_clusters.size(); i++) |
| 2002 | sorted_endpoint_cluster_indices[i] = i; |
| 2003 | |
| 2004 | indirect_sort((uint32_t)m_endpoint_clusters.size(), &sorted_endpoint_cluster_indices[0], &m_endpoint_cluster_etc_params[0]); |
| 2005 | |
| 2006 | basisu::vector<basisu::vector<uint32_t> > new_endpoint_clusters(m_endpoint_clusters.size()); |
| 2007 | basisu::vector<endpoint_cluster_etc_params> new_subblock_etc_params(m_endpoint_clusters.size()); |
| 2008 | |
| 2009 | for (uint32_t i = 0; i < m_endpoint_clusters.size(); i++) |
| 2010 | { |
| 2011 | uint32_t j = sorted_endpoint_cluster_indices[i]; |
| 2012 | new_endpoint_clusters[i] = m_endpoint_clusters[j]; |
| 2013 | new_subblock_etc_params[i] = m_endpoint_cluster_etc_params[j]; |
| 2014 | } |
| 2015 | |
| 2016 | new_endpoint_clusters.swap(m_endpoint_clusters); |
| 2017 | new_subblock_etc_params.swap(m_endpoint_cluster_etc_params); |
| 2018 | |
| 2019 | // Step 2: Eliminate redundant endpoint clusters, or empty endpoint clusters |
| 2020 | |
| 2021 | new_endpoint_clusters.resize(0); |
| 2022 | new_subblock_etc_params.resize(0); |
| 2023 | |
| 2024 | for (int i = 0; i < (int)m_endpoint_clusters.size(); ) |
| 2025 | { |
| 2026 | if (!m_endpoint_clusters[i].size()) |
| 2027 | { |
| 2028 | i++; |
| 2029 | continue; |
| 2030 | } |
| 2031 | |
| 2032 | int j; |
| 2033 | for (j = i + 1; j < (int)m_endpoint_clusters.size(); j++) |
| 2034 | { |
| 2035 | if (!(m_endpoint_cluster_etc_params[i] == m_endpoint_cluster_etc_params[j])) |
| 2036 | break; |
| 2037 | } |
| 2038 | |
| 2039 | new_endpoint_clusters.push_back(m_endpoint_clusters[i]); |
| 2040 | new_subblock_etc_params.push_back(m_endpoint_cluster_etc_params[i]); |
| 2041 | |
| 2042 | for (int k = i + 1; k < j; k++) |
| 2043 | { |
| 2044 | append_vector(new_endpoint_clusters.back(), m_endpoint_clusters[k]); |
| 2045 | } |
| 2046 | |
| 2047 | i = j; |
| 2048 | } |
| 2049 | |
| 2050 | if (m_endpoint_clusters.size() != new_endpoint_clusters.size()) |
| 2051 | { |
| 2052 | if (m_params.m_debug_stats) |
| 2053 | debug_printf("Eliminated %u redundant or empty clusters\n", (uint32_t)(m_endpoint_clusters.size() - new_endpoint_clusters.size())); |
| 2054 | |
| 2055 | m_endpoint_clusters.swap(new_endpoint_clusters); |
| 2056 | |
| 2057 | m_endpoint_cluster_etc_params.swap(new_subblock_etc_params); |
| 2058 | } |
| 2059 | } |
| 2060 | |
| 2061 | void basisu_frontend::create_initial_packed_texture() |
| 2062 | { |
| 2063 | debug_printf("create_initial_packed_texture\n"); |
| 2064 | |
| 2065 | interval_timer tm; |
| 2066 | tm.start(); |
| 2067 | |
| 2068 | bool use_cpu = true; |
| 2069 | |
| 2070 | if ((m_params.m_pOpenCL_context) && (opencl_is_available())) |
| 2071 | { |
| 2072 | basisu::vector<color_rgba> block_etc5_color_intens(m_total_blocks); |
| 2073 | |
| 2074 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 2075 | { |
| 2076 | uint32_t cluster0 = m_block_endpoint_clusters_indices[block_index][0]; |
| 2077 | |
| 2078 | const color_rgba& color_unscaled = m_endpoint_cluster_etc_params[cluster0].m_color_unscaled[0]; |
| 2079 | uint32_t inten = m_endpoint_cluster_etc_params[cluster0].m_inten_table[0]; |
| 2080 | |
| 2081 | block_etc5_color_intens[block_index].set(color_unscaled.r, color_unscaled.g, color_unscaled.b, inten); |
| 2082 | } |
| 2083 | |
| 2084 | bool status = opencl_determine_selectors(m_params.m_pOpenCL_context, block_etc5_color_intens.data(), |
| 2085 | m_encoded_blocks.data(), |
| 2086 | m_params.m_perceptual); |
| 2087 | if (!status) |
| 2088 | { |
| 2089 | error_printf("basisu_frontend::create_initial_packed_texture: opencl_determine_selectors() failed! Using CPU.\n"); |
| 2090 | m_params.m_pOpenCL_context = nullptr; |
| 2091 | m_opencl_failed = true; |
| 2092 | } |
| 2093 | else |
| 2094 | { |
| 2095 | use_cpu = false; |
| 2096 | } |
| 2097 | } |
| 2098 | |
| 2099 | if (use_cpu) |
| 2100 | { |
| 2101 | const uint32_t N = 4096; |
| 2102 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 2103 | { |
| 2104 | const uint32_t first_index = block_index_iter; |
| 2105 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 2106 | |
| 2107 | #ifndef __EMSCRIPTEN__ |
| 2108 | m_params.m_pJob_pool->add_job([this, first_index, last_index] { |
| 2109 | #endif |
| 2110 | |
| 2111 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 2112 | { |
| 2113 | uint32_t cluster0 = m_block_endpoint_clusters_indices[block_index][0]; |
| 2114 | uint32_t cluster1 = m_block_endpoint_clusters_indices[block_index][1]; |
| 2115 | BASISU_FRONTEND_VERIFY(cluster0 == cluster1); |
| 2116 | |
| 2117 | const color_rgba* pSource_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 2118 | |
| 2119 | etc_block& blk = m_encoded_blocks[block_index]; |
| 2120 | |
| 2121 | color_rgba unscaled[2] = { m_endpoint_cluster_etc_params[cluster0].m_color_unscaled[0], m_endpoint_cluster_etc_params[cluster1].m_color_unscaled[0] }; |
| 2122 | uint32_t inten[2] = { m_endpoint_cluster_etc_params[cluster0].m_inten_table[0], m_endpoint_cluster_etc_params[cluster1].m_inten_table[0] }; |
| 2123 | |
| 2124 | blk.set_block_color5(unscaled[0], unscaled[1]); |
| 2125 | blk.set_flip_bit(true); |
| 2126 | |
| 2127 | blk.set_inten_table(0, inten[0]); |
| 2128 | blk.set_inten_table(1, inten[1]); |
| 2129 | |
| 2130 | blk.determine_selectors(pSource_pixels, m_params.m_perceptual); |
| 2131 | |
| 2132 | } // block_index |
| 2133 | |
| 2134 | #ifndef __EMSCRIPTEN__ |
| 2135 | }); |
| 2136 | #endif |
| 2137 | |
| 2138 | } // block_index_iter |
| 2139 | |
| 2140 | #ifndef __EMSCRIPTEN__ |
| 2141 | m_params.m_pJob_pool->wait_for_all(); |
| 2142 | #endif |
| 2143 | |
| 2144 | } // use_cpu |
| 2145 | |
| 2146 | m_orig_encoded_blocks = m_encoded_blocks; |
| 2147 | |
| 2148 | debug_printf("Elapsed time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 2149 | } |
| 2150 | |
| 2151 | void basisu_frontend::compute_selector_clusters_within_each_parent_cluster() |
| 2152 | { |
| 2153 | uint_vec block_selector_cluster_indices(m_total_blocks); |
| 2154 | |
| 2155 | for (int cluster_index = 0; cluster_index < static_cast<int>(m_selector_cluster_block_indices.size()); cluster_index++) |
| 2156 | { |
| 2157 | const basisu::vector<uint32_t>& cluster_indices = m_selector_cluster_block_indices[cluster_index]; |
| 2158 | |
| 2159 | for (uint32_t cluster_indices_iter = 0; cluster_indices_iter < cluster_indices.size(); cluster_indices_iter++) |
| 2160 | { |
| 2161 | const uint32_t block_index = cluster_indices[cluster_indices_iter]; |
| 2162 | |
| 2163 | block_selector_cluster_indices[block_index] = cluster_index; |
| 2164 | |
| 2165 | } // cluster_indices_iter |
| 2166 | |
| 2167 | } // cluster_index |
| 2168 | |
| 2169 | m_selector_clusters_within_each_parent_cluster.resize(0); |
| 2170 | m_selector_clusters_within_each_parent_cluster.resize(m_selector_parent_cluster_block_indices.size()); |
| 2171 | |
| 2172 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 2173 | { |
| 2174 | const uint32_t cluster_index = block_selector_cluster_indices[block_index]; |
| 2175 | const uint32_t parent_cluster_index = m_block_parent_selector_cluster[block_index]; |
| 2176 | |
| 2177 | m_selector_clusters_within_each_parent_cluster[parent_cluster_index].push_back(cluster_index); |
| 2178 | } |
| 2179 | |
| 2180 | for (uint32_t i = 0; i < m_selector_clusters_within_each_parent_cluster.size(); i++) |
| 2181 | { |
| 2182 | uint_vec &cluster_indices = m_selector_clusters_within_each_parent_cluster[i]; |
| 2183 | |
| 2184 | BASISU_FRONTEND_VERIFY(cluster_indices.size()); |
| 2185 | |
| 2186 | vector_sort(cluster_indices); |
| 2187 | |
| 2188 | auto last = std::unique(cluster_indices.begin(), cluster_indices.end()); |
| 2189 | cluster_indices.erase(last, cluster_indices.end()); |
| 2190 | } |
| 2191 | } |
| 2192 | |
| 2193 | void basisu_frontend::generate_selector_clusters() |
| 2194 | { |
| 2195 | debug_printf("generate_selector_clusters\n"); |
| 2196 | |
| 2197 | typedef tree_vector_quant<vec16F> vec16F_clusterizer; |
| 2198 | |
| 2199 | vec16F_clusterizer::array_of_weighted_training_vecs training_vecs(m_total_blocks); |
| 2200 | |
| 2201 | const uint32_t N = 4096; |
| 2202 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 2203 | { |
| 2204 | const uint32_t first_index = block_index_iter; |
| 2205 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 2206 | |
| 2207 | #ifndef __EMSCRIPTEN__ |
| 2208 | m_params.m_pJob_pool->add_job( [this, first_index, last_index, &training_vecs] { |
| 2209 | #endif |
| 2210 | |
| 2211 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 2212 | { |
| 2213 | const etc_block &blk = m_encoded_blocks[block_index]; |
| 2214 | |
| 2215 | vec16F v; |
| 2216 | for (uint32_t y = 0; y < 4; y++) |
| 2217 | for (uint32_t x = 0; x < 4; x++) |
| 2218 | v[x + y * 4] = static_cast<float>(blk.get_selector(x, y)); |
| 2219 | |
| 2220 | const uint32_t subblock_index = (blk.get_inten_table(0) > blk.get_inten_table(1)) ? 0 : 1; |
| 2221 | |
| 2222 | color_rgba block_colors[2]; |
| 2223 | blk.get_block_low_high_colors(block_colors, subblock_index); |
| 2224 | |
| 2225 | const uint32_t dist = color_distance(m_params.m_perceptual, block_colors[0], block_colors[1], false); |
| 2226 | |
| 2227 | const uint32_t cColorDistToWeight = 300; |
| 2228 | const uint32_t cMaxWeight = 4096; |
| 2229 | uint32_t weight = clamp<uint32_t>(dist / cColorDistToWeight, 1, cMaxWeight); |
| 2230 | |
| 2231 | training_vecs[block_index].first = v; |
| 2232 | training_vecs[block_index].second = weight; |
| 2233 | |
| 2234 | } // block_index |
| 2235 | |
| 2236 | #ifndef __EMSCRIPTEN__ |
| 2237 | } ); |
| 2238 | #endif |
| 2239 | |
| 2240 | } // block_index_iter |
| 2241 | |
| 2242 | #ifndef __EMSCRIPTEN__ |
| 2243 | m_params.m_pJob_pool->wait_for_all(); |
| 2244 | #endif |
| 2245 | |
| 2246 | vec16F_clusterizer selector_clusterizer; |
| 2247 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 2248 | selector_clusterizer.add_training_vec(training_vecs[i].first, training_vecs[i].second); |
| 2249 | |
| 2250 | const int selector_parent_codebook_size = (m_params.m_compression_level <= 1) ? BASISU_SELECTOR_PARENT_CODEBOOK_SIZE_COMP_LEVEL_01 : BASISU_SELECTOR_PARENT_CODEBOOK_SIZE_COMP_LEVEL_DEFAULT; |
| 2251 | const uint32_t parent_codebook_size = (m_params.m_max_selector_clusters >= 256) ? selector_parent_codebook_size : 0; |
| 2252 | debug_printf("Using selector parent codebook size %u\n", parent_codebook_size); |
| 2253 | |
| 2254 | uint32_t max_threads = 0; |
| 2255 | max_threads = m_params.m_multithreaded ? minimum<int>(std::thread::hardware_concurrency(), cMaxCodebookCreationThreads) : 0; |
| 2256 | if (m_params.m_pJob_pool) |
| 2257 | max_threads = minimum<int>((int)m_params.m_pJob_pool->get_total_threads(), max_threads); |
| 2258 | |
| 2259 | bool status = generate_hierarchical_codebook_threaded(selector_clusterizer, |
| 2260 | m_params.m_max_selector_clusters, m_use_hierarchical_selector_codebooks ? parent_codebook_size : 0, |
| 2261 | m_selector_cluster_block_indices, |
| 2262 | m_selector_parent_cluster_block_indices, |
| 2263 | max_threads, m_params.m_pJob_pool, false); |
| 2264 | BASISU_FRONTEND_VERIFY(status); |
| 2265 | |
| 2266 | if (m_use_hierarchical_selector_codebooks) |
| 2267 | { |
| 2268 | if (!m_selector_parent_cluster_block_indices.size()) |
| 2269 | { |
| 2270 | m_selector_parent_cluster_block_indices.resize(0); |
| 2271 | m_selector_parent_cluster_block_indices.resize(1); |
| 2272 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 2273 | m_selector_parent_cluster_block_indices[0].push_back(i); |
| 2274 | } |
| 2275 | |
| 2276 | BASISU_ASSUME(BASISU_SELECTOR_PARENT_CODEBOOK_SIZE_COMP_LEVEL_01 <= UINT8_MAX); |
| 2277 | BASISU_ASSUME(BASISU_SELECTOR_PARENT_CODEBOOK_SIZE_COMP_LEVEL_DEFAULT <= UINT8_MAX); |
| 2278 | |
| 2279 | m_block_parent_selector_cluster.resize(0); |
| 2280 | m_block_parent_selector_cluster.resize(m_total_blocks); |
| 2281 | vector_set_all(m_block_parent_selector_cluster, 0xFF); |
| 2282 | |
| 2283 | for (uint32_t parent_cluster_index = 0; parent_cluster_index < m_selector_parent_cluster_block_indices.size(); parent_cluster_index++) |
| 2284 | { |
| 2285 | const uint_vec &cluster = m_selector_parent_cluster_block_indices[parent_cluster_index]; |
| 2286 | for (uint32_t j = 0; j < cluster.size(); j++) |
| 2287 | m_block_parent_selector_cluster[cluster[j]] = static_cast<uint8_t>(parent_cluster_index); |
| 2288 | } |
| 2289 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 2290 | { |
| 2291 | BASISU_FRONTEND_VERIFY(m_block_parent_selector_cluster[i] != 0xFF); |
| 2292 | } |
| 2293 | |
| 2294 | // Ensure that all the blocks within each cluster are all in the same parent cluster, or something is very wrong. |
| 2295 | for (uint32_t cluster_index = 0; cluster_index < m_selector_cluster_block_indices.size(); cluster_index++) |
| 2296 | { |
| 2297 | const uint_vec &cluster = m_selector_cluster_block_indices[cluster_index]; |
| 2298 | |
| 2299 | uint32_t parent_cluster_index = 0; |
| 2300 | for (uint32_t j = 0; j < cluster.size(); j++) |
| 2301 | { |
| 2302 | const uint32_t block_index = cluster[j]; |
| 2303 | if (!j) |
| 2304 | { |
| 2305 | parent_cluster_index = m_block_parent_selector_cluster[block_index]; |
| 2306 | } |
| 2307 | else |
| 2308 | { |
| 2309 | BASISU_FRONTEND_VERIFY(m_block_parent_selector_cluster[block_index] == parent_cluster_index); |
| 2310 | } |
| 2311 | } |
| 2312 | } |
| 2313 | } |
| 2314 | |
| 2315 | debug_printf("Total selector clusters: %u, total parent selector clusters: %u\n", (uint32_t)m_selector_cluster_block_indices.size(), (uint32_t)m_selector_parent_cluster_block_indices.size()); |
| 2316 | } |
| 2317 | |
| 2318 | void basisu_frontend::create_optimized_selector_codebook(uint32_t iter) |
| 2319 | { |
| 2320 | debug_printf("create_optimized_selector_codebook\n"); |
| 2321 | |
| 2322 | interval_timer tm; |
| 2323 | tm.start(); |
| 2324 | |
| 2325 | const uint32_t total_selector_clusters = (uint32_t)m_selector_cluster_block_indices.size(); |
| 2326 | |
| 2327 | debug_printf("Total selector clusters (from m_selector_cluster_block_indices.size()): %u\n", (uint32_t)m_selector_cluster_block_indices.size()); |
| 2328 | |
| 2329 | m_optimized_cluster_selectors.resize(total_selector_clusters); |
| 2330 | |
| 2331 | // For each selector codebook entry, and for each of the 4x4 selectors, determine which selector minimizes the error across all the blocks that use that quantized selector. |
| 2332 | const uint32_t N = 256; |
| 2333 | for (uint32_t cluster_index_iter = 0; cluster_index_iter < total_selector_clusters; cluster_index_iter += N) |
| 2334 | { |
| 2335 | const uint32_t first_index = cluster_index_iter; |
| 2336 | const uint32_t last_index = minimum<uint32_t>((uint32_t)total_selector_clusters, cluster_index_iter + N); |
| 2337 | |
| 2338 | #ifndef __EMSCRIPTEN__ |
| 2339 | m_params.m_pJob_pool->add_job([this, first_index, last_index] { |
| 2340 | #endif |
| 2341 | |
| 2342 | for (uint32_t cluster_index = first_index; cluster_index < last_index; cluster_index++) |
| 2343 | { |
| 2344 | const basisu::vector<uint32_t>& cluster_block_indices = m_selector_cluster_block_indices[cluster_index]; |
| 2345 | |
| 2346 | if (!cluster_block_indices.size()) |
| 2347 | continue; |
| 2348 | |
| 2349 | uint64_t overall_best_err = 0; |
| 2350 | |
| 2351 | uint64_t total_err[4][4][4]; |
| 2352 | clear_obj(total_err); |
| 2353 | |
| 2354 | for (uint32_t cluster_block_index = 0; cluster_block_index < cluster_block_indices.size(); cluster_block_index++) |
| 2355 | { |
| 2356 | const uint32_t block_index = cluster_block_indices[cluster_block_index]; |
| 2357 | |
| 2358 | const etc_block& blk = m_encoded_blocks[block_index]; |
| 2359 | |
| 2360 | color_rgba blk_colors[4]; |
| 2361 | blk.get_block_colors(blk_colors, 0); |
| 2362 | |
| 2363 | for (uint32_t y = 0; y < 4; y++) |
| 2364 | { |
| 2365 | for (uint32_t x = 0; x < 4; x++) |
| 2366 | { |
| 2367 | const color_rgba& orig_color = get_source_pixel_block(block_index)(x, y); |
| 2368 | |
| 2369 | if (m_params.m_perceptual) |
| 2370 | { |
| 2371 | for (uint32_t s = 0; s < 4; s++) |
| 2372 | total_err[y][x][s] += color_distance(true, blk_colors[s], orig_color, false); |
| 2373 | } |
| 2374 | else |
| 2375 | { |
| 2376 | for (uint32_t s = 0; s < 4; s++) |
| 2377 | total_err[y][x][s] += color_distance(false, blk_colors[s], orig_color, false); |
| 2378 | } |
| 2379 | } // x |
| 2380 | } // y |
| 2381 | |
| 2382 | } // cluster_block_index |
| 2383 | |
| 2384 | for (uint32_t y = 0; y < 4; y++) |
| 2385 | { |
| 2386 | for (uint32_t x = 0; x < 4; x++) |
| 2387 | { |
| 2388 | uint64_t best_err = total_err[y][x][0]; |
| 2389 | uint8_t best_sel = 0; |
| 2390 | |
| 2391 | for (uint32_t s = 1; s < 4; s++) |
| 2392 | { |
| 2393 | if (total_err[y][x][s] < best_err) |
| 2394 | { |
| 2395 | best_err = total_err[y][x][s]; |
| 2396 | best_sel = (uint8_t)s; |
| 2397 | } |
| 2398 | } |
| 2399 | |
| 2400 | m_optimized_cluster_selectors[cluster_index].set_selector(x, y, best_sel); |
| 2401 | |
| 2402 | overall_best_err += best_err; |
| 2403 | } // x |
| 2404 | } // y |
| 2405 | |
| 2406 | } // cluster_index |
| 2407 | |
| 2408 | #ifndef __EMSCRIPTEN__ |
| 2409 | }); |
| 2410 | #endif |
| 2411 | |
| 2412 | } // cluster_index_iter |
| 2413 | |
| 2414 | #ifndef __EMSCRIPTEN__ |
| 2415 | m_params.m_pJob_pool->wait_for_all(); |
| 2416 | #endif |
| 2417 | |
| 2418 | debug_printf("Elapsed time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 2419 | |
| 2420 | if (m_params.m_debug_images) |
| 2421 | { |
| 2422 | uint32_t max_selector_cluster_size = 0; |
| 2423 | |
| 2424 | for (uint32_t i = 0; i < m_selector_cluster_block_indices.size(); i++) |
| 2425 | max_selector_cluster_size = maximum<uint32_t>(max_selector_cluster_size, (uint32_t)m_selector_cluster_block_indices[i].size()); |
| 2426 | |
| 2427 | if ((max_selector_cluster_size * 5) < 32768) |
| 2428 | { |
| 2429 | const uint32_t x_spacer_len = 16; |
| 2430 | image selector_cluster_vis(x_spacer_len + max_selector_cluster_size * 5, (uint32_t)m_selector_cluster_block_indices.size() * 5); |
| 2431 | |
| 2432 | for (uint32_t selector_cluster_index = 0; selector_cluster_index < m_selector_cluster_block_indices.size(); selector_cluster_index++) |
| 2433 | { |
| 2434 | const basisu::vector<uint32_t> &cluster_block_indices = m_selector_cluster_block_indices[selector_cluster_index]; |
| 2435 | |
| 2436 | for (uint32_t y = 0; y < 4; y++) |
| 2437 | for (uint32_t x = 0; x < 4; x++) |
| 2438 | selector_cluster_vis.set_clipped(x_spacer_len + x - 12, selector_cluster_index * 5 + y, color_rgba((m_optimized_cluster_selectors[selector_cluster_index].get_selector(x, y) * 255) / 3)); |
| 2439 | |
| 2440 | for (uint32_t i = 0; i < cluster_block_indices.size(); i++) |
| 2441 | { |
| 2442 | uint32_t block_index = cluster_block_indices[i]; |
| 2443 | |
| 2444 | const etc_block &blk = m_orig_encoded_blocks[block_index]; |
| 2445 | |
| 2446 | for (uint32_t y = 0; y < 4; y++) |
| 2447 | for (uint32_t x = 0; x < 4; x++) |
| 2448 | selector_cluster_vis.set_clipped(x_spacer_len + x + 5 * i, selector_cluster_index * 5 + y, color_rgba((blk.get_selector(x, y) * 255) / 3)); |
| 2449 | } |
| 2450 | } |
| 2451 | |
| 2452 | char buf[256]; |
| 2453 | snprintf(buf, sizeof(buf), "selector_cluster_vis_%u.png", iter); |
| 2454 | save_png(buf, selector_cluster_vis); |
| 2455 | } |
| 2456 | } |
| 2457 | } |
| 2458 | |
| 2459 | // For each block: Determine which quantized selectors best encode that block, given its quantized endpoints. |
| 2460 | // Note that this method may leave some empty clusters (i.e. arrays with no block indices), including at the end. |
| 2461 | void basisu_frontend::find_optimal_selector_clusters_for_each_block() |
| 2462 | { |
| 2463 | debug_printf("find_optimal_selector_clusters_for_each_block\n"); |
| 2464 | |
| 2465 | interval_timer tm; |
| 2466 | tm.start(); |
| 2467 | |
| 2468 | if (m_params.m_validate) |
| 2469 | { |
| 2470 | // Sanity checks |
| 2471 | BASISU_FRONTEND_VERIFY(m_selector_cluster_block_indices.size() == m_optimized_cluster_selectors.size()); |
| 2472 | for (uint32_t i = 0; i < m_selector_clusters_within_each_parent_cluster.size(); i++) |
| 2473 | { |
| 2474 | for (uint32_t j = 0; j < m_selector_clusters_within_each_parent_cluster[i].size(); j++) |
| 2475 | { |
| 2476 | BASISU_FRONTEND_VERIFY(m_selector_clusters_within_each_parent_cluster[i][j] < m_optimized_cluster_selectors.size()); |
| 2477 | } |
| 2478 | } |
| 2479 | } |
| 2480 | |
| 2481 | m_block_selector_cluster_index.resize(m_total_blocks); |
| 2482 | |
| 2483 | if (m_params.m_compression_level == 0) |
| 2484 | { |
| 2485 | // Just leave the blocks in their original selector clusters. |
| 2486 | for (uint32_t selector_cluster_index = 0; selector_cluster_index < m_selector_cluster_block_indices.size(); selector_cluster_index++) |
| 2487 | { |
| 2488 | for (uint32_t j = 0; j < m_selector_cluster_block_indices[selector_cluster_index].size(); j++) |
| 2489 | { |
| 2490 | const uint32_t block_index = m_selector_cluster_block_indices[selector_cluster_index][j]; |
| 2491 | |
| 2492 | m_block_selector_cluster_index[block_index] = selector_cluster_index; |
| 2493 | |
| 2494 | etc_block& blk = m_encoded_blocks[block_index]; |
| 2495 | blk.set_raw_selector_bits(m_optimized_cluster_selectors[selector_cluster_index].get_raw_selector_bits()); |
| 2496 | } |
| 2497 | } |
| 2498 | |
| 2499 | debug_printf("Elapsed time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 2500 | |
| 2501 | return; |
| 2502 | } |
| 2503 | |
| 2504 | bool use_cpu = true; |
| 2505 | |
| 2506 | if ((m_params.m_pOpenCL_context) && m_use_hierarchical_selector_codebooks) |
| 2507 | { |
| 2508 | const uint32_t num_parent_clusters = m_selector_clusters_within_each_parent_cluster.size(); |
| 2509 | |
| 2510 | basisu::vector<fosc_selector_struct> selector_structs; |
| 2511 | selector_structs.reserve(m_optimized_cluster_selectors.size()); |
| 2512 | |
| 2513 | uint_vec parent_selector_cluster_offsets(num_parent_clusters); |
| 2514 | |
| 2515 | uint_vec selector_cluster_indices; |
| 2516 | selector_cluster_indices.reserve(m_optimized_cluster_selectors.size()); |
| 2517 | |
| 2518 | uint32_t cur_ofs = 0; |
| 2519 | for (uint32_t parent_index = 0; parent_index < num_parent_clusters; parent_index++) |
| 2520 | { |
| 2521 | parent_selector_cluster_offsets[parent_index] = cur_ofs; |
| 2522 | |
| 2523 | for (uint32_t j = 0; j < m_selector_clusters_within_each_parent_cluster[parent_index].size(); j++) |
| 2524 | { |
| 2525 | const uint32_t selector_cluster_index = m_selector_clusters_within_each_parent_cluster[parent_index][j]; |
| 2526 | |
| 2527 | uint32_t sel_bits = 0; |
| 2528 | for (uint32_t p = 0; p < 16; p++) |
| 2529 | sel_bits |= (m_optimized_cluster_selectors[selector_cluster_index].get_selector(p & 3, p >> 2) << (p * 2)); |
| 2530 | |
| 2531 | selector_structs.enlarge(1)->m_packed_selectors = sel_bits; |
| 2532 | |
| 2533 | selector_cluster_indices.push_back(selector_cluster_index); |
| 2534 | } |
| 2535 | |
| 2536 | cur_ofs += m_selector_clusters_within_each_parent_cluster[parent_index].size(); |
| 2537 | } |
| 2538 | |
| 2539 | const uint32_t total_input_selectors = cur_ofs; |
| 2540 | |
| 2541 | basisu::vector<fosc_block_struct> block_structs(m_total_blocks); |
| 2542 | for (uint32_t i = 0; i < m_total_blocks; i++) |
| 2543 | { |
| 2544 | const uint32_t parent_selector_cluster = m_block_parent_selector_cluster[i]; |
| 2545 | |
| 2546 | const etc_block& blk = m_encoded_blocks[i]; |
| 2547 | blk.unpack_color5(block_structs[i].m_etc_color5_inten, blk.get_base5_color(), false); |
| 2548 | |
| 2549 | block_structs[i].m_etc_color5_inten.a = (uint8_t)blk.get_inten_table(0); |
| 2550 | block_structs[i].m_first_selector = parent_selector_cluster_offsets[parent_selector_cluster]; |
| 2551 | block_structs[i].m_num_selectors = m_selector_clusters_within_each_parent_cluster[parent_selector_cluster].size(); |
| 2552 | } |
| 2553 | |
| 2554 | uint_vec output_selector_cluster_indices(m_total_blocks); |
| 2555 | |
| 2556 | bool status = opencl_find_optimal_selector_clusters_for_each_block( |
| 2557 | m_params.m_pOpenCL_context, |
| 2558 | block_structs.data(), |
| 2559 | total_input_selectors, |
| 2560 | selector_structs.data(), |
| 2561 | selector_cluster_indices.data(), |
| 2562 | output_selector_cluster_indices.data(), |
| 2563 | m_params.m_perceptual); |
| 2564 | |
| 2565 | if (!status) |
| 2566 | { |
| 2567 | error_printf("basisu_frontend::find_optimal_selector_clusters_for_each_block: opencl_find_optimal_selector_clusters_for_each_block() failed! Using CPU.\n"); |
| 2568 | m_params.m_pOpenCL_context = nullptr; |
| 2569 | m_opencl_failed = true; |
| 2570 | } |
| 2571 | else |
| 2572 | { |
| 2573 | for (uint32_t i = 0; i < m_selector_cluster_block_indices.size(); i++) |
| 2574 | { |
| 2575 | m_selector_cluster_block_indices[i].resize(0); |
| 2576 | m_selector_cluster_block_indices[i].reserve(128); |
| 2577 | } |
| 2578 | |
| 2579 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 2580 | { |
| 2581 | etc_block& blk = m_encoded_blocks[block_index]; |
| 2582 | |
| 2583 | uint32_t best_cluster_index = output_selector_cluster_indices[block_index]; |
| 2584 | |
| 2585 | blk.set_raw_selector_bits(m_optimized_cluster_selectors[best_cluster_index].get_raw_selector_bits()); |
| 2586 | |
| 2587 | m_block_selector_cluster_index[block_index] = best_cluster_index; |
| 2588 | |
| 2589 | vector_ensure_element_is_valid(m_selector_cluster_block_indices, best_cluster_index); |
| 2590 | m_selector_cluster_block_indices[best_cluster_index].push_back(block_index); |
| 2591 | } |
| 2592 | |
| 2593 | use_cpu = false; |
| 2594 | } |
| 2595 | } |
| 2596 | |
| 2597 | if (use_cpu) |
| 2598 | { |
| 2599 | basisu::vector<uint8_t> unpacked_optimized_cluster_selectors(16 * m_optimized_cluster_selectors.size()); |
| 2600 | for (uint32_t cluster_index = 0; cluster_index < m_optimized_cluster_selectors.size(); cluster_index++) |
| 2601 | { |
| 2602 | for (uint32_t y = 0; y < 4; y++) |
| 2603 | { |
| 2604 | for (uint32_t x = 0; x < 4; x++) |
| 2605 | { |
| 2606 | unpacked_optimized_cluster_selectors[cluster_index * 16 + y * 4 + x] = (uint8_t)m_optimized_cluster_selectors[cluster_index].get_selector(x, y); |
| 2607 | } |
| 2608 | } |
| 2609 | } |
| 2610 | |
| 2611 | const uint32_t N = 2048; |
| 2612 | for (uint32_t block_index_iter = 0; block_index_iter < m_total_blocks; block_index_iter += N) |
| 2613 | { |
| 2614 | const uint32_t first_index = block_index_iter; |
| 2615 | const uint32_t last_index = minimum<uint32_t>(m_total_blocks, first_index + N); |
| 2616 | |
| 2617 | #ifndef __EMSCRIPTEN__ |
| 2618 | m_params.m_pJob_pool->add_job( [this, first_index, last_index, &unpacked_optimized_cluster_selectors] { |
| 2619 | #endif |
| 2620 | |
| 2621 | int prev_best_cluster_index = 0; |
| 2622 | |
| 2623 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 2624 | { |
| 2625 | const pixel_block& block = get_source_pixel_block(block_index); |
| 2626 | |
| 2627 | etc_block& blk = m_encoded_blocks[block_index]; |
| 2628 | |
| 2629 | if ((block_index > first_index) && (block == get_source_pixel_block(block_index - 1))) |
| 2630 | { |
| 2631 | blk.set_raw_selector_bits(m_optimized_cluster_selectors[prev_best_cluster_index].get_raw_selector_bits()); |
| 2632 | |
| 2633 | m_block_selector_cluster_index[block_index] = prev_best_cluster_index; |
| 2634 | |
| 2635 | continue; |
| 2636 | } |
| 2637 | |
| 2638 | const color_rgba* pBlock_pixels = block.get_ptr(); |
| 2639 | |
| 2640 | color_rgba trial_block_colors[4]; |
| 2641 | blk.get_block_colors_etc1s(trial_block_colors); |
| 2642 | |
| 2643 | // precompute errors for the i-th block pixel and selector sel: [sel][i] |
| 2644 | uint32_t trial_errors[4][16]; |
| 2645 | |
| 2646 | if (m_params.m_perceptual) |
| 2647 | { |
| 2648 | for (uint32_t sel = 0; sel < 4; ++sel) |
| 2649 | for (uint32_t i = 0; i < 16; ++i) |
| 2650 | trial_errors[sel][i] = color_distance(true, pBlock_pixels[i], trial_block_colors[sel], false); |
| 2651 | } |
| 2652 | else |
| 2653 | { |
| 2654 | for (uint32_t sel = 0; sel < 4; ++sel) |
| 2655 | for (uint32_t i = 0; i < 16; ++i) |
| 2656 | trial_errors[sel][i] = color_distance(false, pBlock_pixels[i], trial_block_colors[sel], false); |
| 2657 | } |
| 2658 | |
| 2659 | // Compute the minimum possible errors (given any selectors) for pixels 0-15 |
| 2660 | uint64_t min_possible_error_0_15 = 0; |
| 2661 | for (uint32_t i = 0; i < 16; i++) |
| 2662 | min_possible_error_0_15 += basisu::minimum(trial_errors[0][i], trial_errors[1][i], trial_errors[2][i], trial_errors[3][i]); |
| 2663 | |
| 2664 | // Compute the minimum possible errors (given any selectors) for pixels 4-15 |
| 2665 | uint64_t min_possible_error_4_15 = 0; |
| 2666 | for (uint32_t i = 4; i < 16; i++) |
| 2667 | min_possible_error_4_15 += basisu::minimum(trial_errors[0][i], trial_errors[1][i], trial_errors[2][i], trial_errors[3][i]); |
| 2668 | |
| 2669 | // Compute the minimum possible errors (given any selectors) for pixels 8-15 |
| 2670 | uint64_t min_possible_error_8_15 = 0; |
| 2671 | for (uint32_t i = 8; i < 16; i++) |
| 2672 | min_possible_error_8_15 += basisu::minimum(trial_errors[0][i], trial_errors[1][i], trial_errors[2][i], trial_errors[3][i]); |
| 2673 | |
| 2674 | // Compute the minimum possible errors (given any selectors) for pixels 12-15 |
| 2675 | uint64_t min_possible_error_12_15 = 0; |
| 2676 | for (uint32_t i = 12; i < 16; i++) |
| 2677 | min_possible_error_12_15 += basisu::minimum(trial_errors[0][i], trial_errors[1][i], trial_errors[2][i], trial_errors[3][i]); |
| 2678 | |
| 2679 | uint64_t best_cluster_err = INT64_MAX; |
| 2680 | uint32_t best_cluster_index = 0; |
| 2681 | |
| 2682 | const uint32_t parent_selector_cluster = m_block_parent_selector_cluster.size() ? m_block_parent_selector_cluster[block_index] : 0; |
| 2683 | const uint_vec *pCluster_indices = m_selector_clusters_within_each_parent_cluster.size() ? &m_selector_clusters_within_each_parent_cluster[parent_selector_cluster] : nullptr; |
| 2684 | |
| 2685 | const uint32_t total_clusters = m_use_hierarchical_selector_codebooks ? (uint32_t)pCluster_indices->size() : (uint32_t)m_selector_cluster_block_indices.size(); |
| 2686 | |
| 2687 | #if 0 |
| 2688 | for (uint32_t cluster_iter = 0; cluster_iter < total_clusters; cluster_iter++) |
| 2689 | { |
| 2690 | const uint32_t cluster_index = m_use_hierarchical_selector_codebooks ? (*pCluster_indices)[cluster_iter] : cluster_iter; |
| 2691 | |
| 2692 | const etc_block& cluster_blk = m_optimized_cluster_selectors[cluster_index]; |
| 2693 | |
| 2694 | uint64_t trial_err = 0; |
| 2695 | for (int y = 0; y < 4; y++) |
| 2696 | { |
| 2697 | for (int x = 0; x < 4; x++) |
| 2698 | { |
| 2699 | const uint32_t sel = cluster_blk.get_selector(x, y); |
| 2700 | |
| 2701 | trial_err += color_distance(m_params.m_perceptual, trial_block_colors[sel], pBlock_pixels[x + y * 4], false); |
| 2702 | if (trial_err > best_cluster_err) |
| 2703 | goto early_out; |
| 2704 | } |
| 2705 | } |
| 2706 | |
| 2707 | if (trial_err < best_cluster_err) |
| 2708 | { |
| 2709 | best_cluster_err = trial_err; |
| 2710 | best_cluster_index = cluster_index; |
| 2711 | if (!best_cluster_err) |
| 2712 | break; |
| 2713 | } |
| 2714 | |
| 2715 | early_out: |
| 2716 | ; |
| 2717 | } |
| 2718 | #else |
| 2719 | for (uint32_t cluster_iter = 0; cluster_iter < total_clusters; cluster_iter++) |
| 2720 | { |
| 2721 | const uint32_t cluster_index = m_use_hierarchical_selector_codebooks ? (*pCluster_indices)[cluster_iter] : cluster_iter; |
| 2722 | |
| 2723 | const uint8_t* pSels = &unpacked_optimized_cluster_selectors[cluster_index * 16]; |
| 2724 | |
| 2725 | uint64_t trial_err = (uint64_t)trial_errors[pSels[0]][0] + trial_errors[pSels[1]][1] + trial_errors[pSels[2]][2] + trial_errors[pSels[3]][3]; |
| 2726 | if ((trial_err + min_possible_error_4_15) >= best_cluster_err) |
| 2727 | continue; |
| 2728 | |
| 2729 | trial_err += (uint64_t)trial_errors[pSels[4]][4] + trial_errors[pSels[5]][5] + trial_errors[pSels[6]][6] + trial_errors[pSels[7]][7]; |
| 2730 | if ((trial_err + min_possible_error_8_15) >= best_cluster_err) |
| 2731 | continue; |
| 2732 | |
| 2733 | trial_err += (uint64_t)trial_errors[pSels[8]][8] + trial_errors[pSels[9]][9] + trial_errors[pSels[10]][10] + trial_errors[pSels[11]][11]; |
| 2734 | if ((trial_err + min_possible_error_12_15) >= best_cluster_err) |
| 2735 | continue; |
| 2736 | |
| 2737 | trial_err += (uint64_t)trial_errors[pSels[12]][12] + trial_errors[pSels[13]][13] + trial_errors[pSels[14]][14] + trial_errors[pSels[15]][15]; |
| 2738 | |
| 2739 | if (trial_err < best_cluster_err) |
| 2740 | { |
| 2741 | best_cluster_err = trial_err; |
| 2742 | best_cluster_index = cluster_index; |
| 2743 | if (best_cluster_err == min_possible_error_0_15) |
| 2744 | break; |
| 2745 | } |
| 2746 | |
| 2747 | } // cluster_iter |
| 2748 | #endif |
| 2749 | |
| 2750 | blk.set_raw_selector_bits(m_optimized_cluster_selectors[best_cluster_index].get_raw_selector_bits()); |
| 2751 | |
| 2752 | m_block_selector_cluster_index[block_index] = best_cluster_index; |
| 2753 | |
| 2754 | prev_best_cluster_index = best_cluster_index; |
| 2755 | |
| 2756 | } // block_index |
| 2757 | |
| 2758 | #ifndef __EMSCRIPTEN__ |
| 2759 | } ); |
| 2760 | #endif |
| 2761 | |
| 2762 | } // block_index_iter |
| 2763 | |
| 2764 | #ifndef __EMSCRIPTEN__ |
| 2765 | m_params.m_pJob_pool->wait_for_all(); |
| 2766 | #endif |
| 2767 | |
| 2768 | for (uint32_t i = 0; i < m_selector_cluster_block_indices.size(); i++) |
| 2769 | { |
| 2770 | m_selector_cluster_block_indices[i].resize(0); |
| 2771 | m_selector_cluster_block_indices[i].reserve(128); |
| 2772 | } |
| 2773 | |
| 2774 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 2775 | { |
| 2776 | const uint32_t best_cluster_index = m_block_selector_cluster_index[block_index]; |
| 2777 | |
| 2778 | vector_ensure_element_is_valid(m_selector_cluster_block_indices, best_cluster_index); |
| 2779 | m_selector_cluster_block_indices[best_cluster_index].push_back(block_index); |
| 2780 | } |
| 2781 | |
| 2782 | } // if (use_cpu) |
| 2783 | |
| 2784 | debug_printf("Elapsed time: %3.3f secs\n", tm.get_elapsed_secs()); |
| 2785 | } |
| 2786 | |
| 2787 | // TODO: Remove old ETC1 specific stuff, and thread this. |
| 2788 | uint32_t basisu_frontend::refine_block_endpoints_given_selectors() |
| 2789 | { |
| 2790 | debug_printf("refine_block_endpoints_given_selectors\n"); |
| 2791 | |
| 2792 | for (int block_index = 0; block_index < static_cast<int>(m_total_blocks); block_index++) |
| 2793 | { |
| 2794 | //uint32_t selector_cluster = m_block_selector_cluster_index(block_x, block_y); |
| 2795 | vec2U &endpoint_clusters = m_block_endpoint_clusters_indices[block_index]; |
| 2796 | |
| 2797 | m_endpoint_cluster_etc_params[endpoint_clusters[0]].m_subblocks.push_back(block_index * 2); |
| 2798 | |
| 2799 | m_endpoint_cluster_etc_params[endpoint_clusters[1]].m_subblocks.push_back(block_index * 2 + 1); |
| 2800 | } |
| 2801 | |
| 2802 | uint32_t total_subblocks_refined = 0; |
| 2803 | uint32_t total_subblocks_examined = 0; |
| 2804 | |
| 2805 | for (uint32_t endpoint_cluster_index = 0; endpoint_cluster_index < m_endpoint_cluster_etc_params.size(); endpoint_cluster_index++) |
| 2806 | { |
| 2807 | endpoint_cluster_etc_params &subblock_params = m_endpoint_cluster_etc_params[endpoint_cluster_index]; |
| 2808 | |
| 2809 | const uint_vec &subblocks = subblock_params.m_subblocks; |
| 2810 | //uint32_t total_pixels = subblock.m_subblocks.size() * 8; |
| 2811 | |
| 2812 | basisu::vector<color_rgba> subblock_colors[2]; // [use_individual_mode] |
| 2813 | uint8_vec subblock_selectors[2]; |
| 2814 | |
| 2815 | uint64_t cur_subblock_err[2] = { 0, 0 }; |
| 2816 | |
| 2817 | for (uint32_t subblock_iter = 0; subblock_iter < subblocks.size(); subblock_iter++) |
| 2818 | { |
| 2819 | uint32_t training_vector_index = subblocks[subblock_iter]; |
| 2820 | |
| 2821 | uint32_t block_index = training_vector_index >> 1; |
| 2822 | uint32_t subblock_index = training_vector_index & 1; |
| 2823 | const bool is_flipped = true; |
| 2824 | |
| 2825 | const etc_block &blk = m_encoded_blocks[block_index]; |
| 2826 | |
| 2827 | const bool use_individual_mode = !blk.get_diff_bit(); |
| 2828 | |
| 2829 | const color_rgba *pSource_block_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 2830 | |
| 2831 | color_rgba unpacked_block_pixels[16]; |
| 2832 | unpack_etc1(blk, unpacked_block_pixels); |
| 2833 | |
| 2834 | for (uint32_t i = 0; i < 8; i++) |
| 2835 | { |
| 2836 | const uint32_t pixel_index = g_etc1_pixel_indices[is_flipped][subblock_index][i]; |
| 2837 | const etc_coord2 &coords = g_etc1_pixel_coords[is_flipped][subblock_index][i]; |
| 2838 | |
| 2839 | subblock_colors[use_individual_mode].push_back(pSource_block_pixels[pixel_index]); |
| 2840 | |
| 2841 | cur_subblock_err[use_individual_mode] += color_distance(m_params.m_perceptual, pSource_block_pixels[pixel_index], unpacked_block_pixels[pixel_index], false); |
| 2842 | |
| 2843 | subblock_selectors[use_individual_mode].push_back(static_cast<uint8_t>(blk.get_selector(coords.m_x, coords.m_y))); |
| 2844 | } |
| 2845 | } // subblock_iter |
| 2846 | |
| 2847 | etc1_optimizer::results cluster_optimizer_results[2]; |
| 2848 | bool results_valid[2] = { false, false }; |
| 2849 | |
| 2850 | clear_obj(cluster_optimizer_results); |
| 2851 | |
| 2852 | basisu::vector<uint8_t> cluster_selectors[2]; |
| 2853 | |
| 2854 | for (uint32_t use_individual_mode = 0; use_individual_mode < 2; use_individual_mode++) |
| 2855 | { |
| 2856 | const uint32_t total_pixels = (uint32_t)subblock_colors[use_individual_mode].size(); |
| 2857 | |
| 2858 | if (!total_pixels) |
| 2859 | continue; |
| 2860 | |
| 2861 | total_subblocks_examined += total_pixels / 8; |
| 2862 | |
| 2863 | etc1_optimizer optimizer; |
| 2864 | etc1_solution_coordinates solutions[2]; |
| 2865 | |
| 2866 | etc1_optimizer::params cluster_optimizer_params; |
| 2867 | cluster_optimizer_params.m_num_src_pixels = total_pixels; |
| 2868 | cluster_optimizer_params.m_pSrc_pixels = &subblock_colors[use_individual_mode][0]; |
| 2869 | |
| 2870 | cluster_optimizer_params.m_use_color4 = use_individual_mode != 0; |
| 2871 | cluster_optimizer_params.m_perceptual = m_params.m_perceptual; |
| 2872 | |
| 2873 | cluster_optimizer_params.m_pForce_selectors = &subblock_selectors[use_individual_mode][0]; |
| 2874 | cluster_optimizer_params.m_quality = cETCQualityUber; |
| 2875 | |
| 2876 | cluster_selectors[use_individual_mode].resize(total_pixels); |
| 2877 | |
| 2878 | cluster_optimizer_results[use_individual_mode].m_n = total_pixels; |
| 2879 | cluster_optimizer_results[use_individual_mode].m_pSelectors = &cluster_selectors[use_individual_mode][0]; |
| 2880 | |
| 2881 | optimizer.init(cluster_optimizer_params, cluster_optimizer_results[use_individual_mode]); |
| 2882 | |
| 2883 | if (!optimizer.compute()) |
| 2884 | continue; |
| 2885 | |
| 2886 | if (cluster_optimizer_results[use_individual_mode].m_error < cur_subblock_err[use_individual_mode]) |
| 2887 | results_valid[use_individual_mode] = true; |
| 2888 | |
| 2889 | } // use_individual_mode |
| 2890 | |
| 2891 | for (uint32_t use_individual_mode = 0; use_individual_mode < 2; use_individual_mode++) |
| 2892 | { |
| 2893 | if (!results_valid[use_individual_mode]) |
| 2894 | continue; |
| 2895 | |
| 2896 | uint32_t num_passes = use_individual_mode ? 1 : 2; |
| 2897 | |
| 2898 | bool all_passed5 = true; |
| 2899 | |
| 2900 | for (uint32_t pass = 0; pass < num_passes; pass++) |
| 2901 | { |
| 2902 | for (uint32_t subblock_iter = 0; subblock_iter < subblocks.size(); subblock_iter++) |
| 2903 | { |
| 2904 | const uint32_t training_vector_index = subblocks[subblock_iter]; |
| 2905 | |
| 2906 | const uint32_t block_index = training_vector_index >> 1; |
| 2907 | const uint32_t subblock_index = training_vector_index & 1; |
| 2908 | //const bool is_flipped = true; |
| 2909 | |
| 2910 | etc_block &blk = m_encoded_blocks[block_index]; |
| 2911 | |
| 2912 | if (!blk.get_diff_bit() != static_cast<bool>(use_individual_mode != 0)) |
| 2913 | continue; |
| 2914 | |
| 2915 | if (use_individual_mode) |
| 2916 | { |
| 2917 | blk.set_base4_color(subblock_index, etc_block::pack_color4(cluster_optimizer_results[1].m_block_color_unscaled, false)); |
| 2918 | blk.set_inten_table(subblock_index, cluster_optimizer_results[1].m_block_inten_table); |
| 2919 | |
| 2920 | subblock_params.m_color_error[1] = cluster_optimizer_results[1].m_error; |
| 2921 | subblock_params.m_inten_table[1] = cluster_optimizer_results[1].m_block_inten_table; |
| 2922 | subblock_params.m_color_unscaled[1] = cluster_optimizer_results[1].m_block_color_unscaled; |
| 2923 | |
| 2924 | total_subblocks_refined++; |
| 2925 | } |
| 2926 | else |
| 2927 | { |
| 2928 | const uint16_t base_color5 = blk.get_base5_color(); |
| 2929 | const uint16_t delta_color3 = blk.get_delta3_color(); |
| 2930 | |
| 2931 | uint32_t r[2], g[2], b[2]; |
| 2932 | etc_block::unpack_color5(r[0], g[0], b[0], base_color5, false); |
| 2933 | bool success = etc_block::unpack_color5(r[1], g[1], b[1], base_color5, delta_color3, false); |
| 2934 | assert(success); |
| 2935 | BASISU_NOTE_UNUSED(success); |
| 2936 | |
| 2937 | r[subblock_index] = cluster_optimizer_results[0].m_block_color_unscaled.r; |
| 2938 | g[subblock_index] = cluster_optimizer_results[0].m_block_color_unscaled.g; |
| 2939 | b[subblock_index] = cluster_optimizer_results[0].m_block_color_unscaled.b; |
| 2940 | |
| 2941 | color_rgba colors[2] = { color_rgba(r[0], g[0], b[0], 255), color_rgba(r[1], g[1], b[1], 255) }; |
| 2942 | |
| 2943 | if (!etc_block::try_pack_color5_delta3(colors)) |
| 2944 | { |
| 2945 | all_passed5 = false; |
| 2946 | break; |
| 2947 | } |
| 2948 | |
| 2949 | if ((pass == 1) && (all_passed5)) |
| 2950 | { |
| 2951 | blk.set_block_color5(colors[0], colors[1]); |
| 2952 | blk.set_inten_table(subblock_index, cluster_optimizer_results[0].m_block_inten_table); |
| 2953 | |
| 2954 | subblock_params.m_color_error[0] = cluster_optimizer_results[0].m_error; |
| 2955 | subblock_params.m_inten_table[0] = cluster_optimizer_results[0].m_block_inten_table; |
| 2956 | subblock_params.m_color_unscaled[0] = cluster_optimizer_results[0].m_block_color_unscaled; |
| 2957 | |
| 2958 | total_subblocks_refined++; |
| 2959 | } |
| 2960 | } |
| 2961 | |
| 2962 | } // subblock_iter |
| 2963 | |
| 2964 | } // pass |
| 2965 | |
| 2966 | } // use_individual_mode |
| 2967 | |
| 2968 | } // endpoint_cluster_index |
| 2969 | |
| 2970 | if (m_params.m_debug_stats) |
| 2971 | debug_printf("Total subblock endpoints refined: %u (%3.1f%%)\n", total_subblocks_refined, total_subblocks_refined * 100.0f / total_subblocks_examined); |
| 2972 | |
| 2973 | return total_subblocks_refined; |
| 2974 | } |
| 2975 | |
| 2976 | void basisu_frontend::dump_endpoint_clusterization_visualization(const char *pFilename, bool vis_endpoint_colors) |
| 2977 | { |
| 2978 | debug_printf("dump_endpoint_clusterization_visualization\n"); |
| 2979 | |
| 2980 | uint32_t max_endpoint_cluster_size = 0; |
| 2981 | |
| 2982 | basisu::vector<uint32_t> cluster_sizes(m_endpoint_clusters.size()); |
| 2983 | basisu::vector<uint32_t> sorted_cluster_indices(m_endpoint_clusters.size()); |
| 2984 | for (uint32_t i = 0; i < m_endpoint_clusters.size(); i++) |
| 2985 | { |
| 2986 | max_endpoint_cluster_size = maximum<uint32_t>(max_endpoint_cluster_size, (uint32_t)m_endpoint_clusters[i].size()); |
| 2987 | cluster_sizes[i] = (uint32_t)m_endpoint_clusters[i].size(); |
| 2988 | } |
| 2989 | |
| 2990 | if (!max_endpoint_cluster_size) |
| 2991 | return; |
| 2992 | |
| 2993 | for (uint32_t i = 0; i < m_endpoint_clusters.size(); i++) |
| 2994 | sorted_cluster_indices[i] = i; |
| 2995 | |
| 2996 | //indexed_heap_sort(endpoint_clusters.size(), cluster_sizes.get_ptr(), sorted_cluster_indices.get_ptr()); |
| 2997 | |
| 2998 | image endpoint_cluster_vis(12 + minimum<uint32_t>(max_endpoint_cluster_size, 2048) * 5, (uint32_t)m_endpoint_clusters.size() * 3); |
| 2999 | |
| 3000 | for (uint32_t unsorted_cluster_iter = 0; unsorted_cluster_iter < m_endpoint_clusters.size(); unsorted_cluster_iter++) |
| 3001 | { |
| 3002 | const uint32_t cluster_iter = sorted_cluster_indices[unsorted_cluster_iter]; |
| 3003 | |
| 3004 | etc_block blk; |
| 3005 | blk.clear(); |
| 3006 | blk.set_flip_bit(false); |
| 3007 | blk.set_diff_bit(true); |
| 3008 | blk.set_inten_tables_etc1s(m_endpoint_cluster_etc_params[cluster_iter].m_inten_table[0]); |
| 3009 | blk.set_base5_color(etc_block::pack_color5(m_endpoint_cluster_etc_params[cluster_iter].m_color_unscaled[0], false)); |
| 3010 | |
| 3011 | color_rgba blk_colors[4]; |
| 3012 | blk.get_block_colors(blk_colors, 0); |
| 3013 | for (uint32_t i = 0; i < 4; i++) |
| 3014 | endpoint_cluster_vis.fill_box(i * 2, 3 * unsorted_cluster_iter, 2, 2, blk_colors[i]); |
| 3015 | |
| 3016 | for (uint32_t subblock_iter = 0; subblock_iter < m_endpoint_clusters[cluster_iter].size(); subblock_iter++) |
| 3017 | { |
| 3018 | uint32_t training_vector_index = m_endpoint_clusters[cluster_iter][subblock_iter]; |
| 3019 | |
| 3020 | const uint32_t block_index = training_vector_index >> 1; |
| 3021 | const uint32_t subblock_index = training_vector_index & 1; |
| 3022 | |
| 3023 | const etc_block& blk2 = m_etc1_blocks_etc1s[block_index]; |
| 3024 | |
| 3025 | const color_rgba *pBlock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 3026 | |
| 3027 | color_rgba subblock_pixels[8]; |
| 3028 | |
| 3029 | if (vis_endpoint_colors) |
| 3030 | { |
| 3031 | color_rgba colors[2]; |
| 3032 | blk2.get_block_low_high_colors(colors, subblock_index); |
| 3033 | for (uint32_t i = 0; i < 8; i++) |
| 3034 | subblock_pixels[i] = colors[subblock_index]; |
| 3035 | } |
| 3036 | else |
| 3037 | { |
| 3038 | for (uint32_t i = 0; i < 8; i++) |
| 3039 | subblock_pixels[i] = pBlock_pixels[g_etc1_pixel_indices[blk2.get_flip_bit()][subblock_index][i]]; |
| 3040 | } |
| 3041 | |
| 3042 | endpoint_cluster_vis.set_block_clipped(subblock_pixels, 12 + 5 * subblock_iter, 3 * unsorted_cluster_iter, 4, 2); |
| 3043 | } |
| 3044 | } |
| 3045 | |
| 3046 | save_png(pFilename, endpoint_cluster_vis); |
| 3047 | debug_printf("Wrote debug visualization file %s\n", pFilename); |
| 3048 | } |
| 3049 | |
| 3050 | void basisu_frontend::finalize() |
| 3051 | { |
| 3052 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 3053 | { |
| 3054 | for (uint32_t subblock_index = 0; subblock_index < 2; subblock_index++) |
| 3055 | { |
| 3056 | const uint32_t endpoint_cluster_index = get_subblock_endpoint_cluster_index(block_index, subblock_index); |
| 3057 | |
| 3058 | m_endpoint_cluster_etc_params[endpoint_cluster_index].m_color_used[0] = true; |
| 3059 | } |
| 3060 | } |
| 3061 | } |
| 3062 | |
| 3063 | // The backend has remapped the block endpoints while optimizing the output symbols for better rate distortion performance, so let's go and reoptimize the endpoint codebook. |
| 3064 | // This is currently the only place where the backend actually goes and changes the quantization and calls the frontend to fix things up. |
| 3065 | // This is basically a bottom up clusterization stage, where some leaves can be combined. |
| 3066 | void basisu_frontend::reoptimize_remapped_endpoints(const uint_vec &new_block_endpoints, int_vec &old_to_new_endpoint_cluster_indices, bool optimize_final_codebook, uint_vec *pBlock_selector_indices) |
| 3067 | { |
| 3068 | debug_printf("reoptimize_remapped_endpoints\n"); |
| 3069 | |
| 3070 | basisu::vector<uint_vec> new_endpoint_cluster_block_indices(m_endpoint_clusters.size()); |
| 3071 | for (uint32_t i = 0; i < new_block_endpoints.size(); i++) |
| 3072 | new_endpoint_cluster_block_indices[new_block_endpoints[i]].push_back(i); |
| 3073 | |
| 3074 | basisu::vector<uint8_t> cluster_valid(new_endpoint_cluster_block_indices.size()); |
| 3075 | basisu::vector<uint8_t> cluster_improved(new_endpoint_cluster_block_indices.size()); |
| 3076 | |
| 3077 | const uint32_t N = 256; |
| 3078 | for (uint32_t cluster_index_iter = 0; cluster_index_iter < new_endpoint_cluster_block_indices.size(); cluster_index_iter += N) |
| 3079 | { |
| 3080 | const uint32_t first_index = cluster_index_iter; |
| 3081 | const uint32_t last_index = minimum<uint32_t>((uint32_t)new_endpoint_cluster_block_indices.size(), cluster_index_iter + N); |
| 3082 | |
| 3083 | #ifndef __EMSCRIPTEN__ |
| 3084 | m_params.m_pJob_pool->add_job( [this, first_index, last_index, &cluster_improved, &cluster_valid, &new_endpoint_cluster_block_indices, &pBlock_selector_indices ] { |
| 3085 | #endif |
| 3086 | |
| 3087 | for (uint32_t cluster_index = first_index; cluster_index < last_index; cluster_index++) |
| 3088 | { |
| 3089 | const basisu::vector<uint32_t>& cluster_block_indices = new_endpoint_cluster_block_indices[cluster_index]; |
| 3090 | |
| 3091 | if (!cluster_block_indices.size()) |
| 3092 | continue; |
| 3093 | |
| 3094 | const uint32_t total_pixels = (uint32_t)cluster_block_indices.size() * 16; |
| 3095 | |
| 3096 | basisu::vector<color_rgba> cluster_pixels(total_pixels); |
| 3097 | uint8_vec force_selectors(total_pixels); |
| 3098 | |
| 3099 | etc_block blk; |
| 3100 | blk.set_block_color5_etc1s(get_endpoint_cluster_unscaled_color(cluster_index, false)); |
| 3101 | blk.set_inten_tables_etc1s(get_endpoint_cluster_inten_table(cluster_index, false)); |
| 3102 | blk.set_flip_bit(true); |
| 3103 | |
| 3104 | uint64_t cur_err = 0; |
| 3105 | |
| 3106 | for (uint32_t cluster_block_indices_iter = 0; cluster_block_indices_iter < cluster_block_indices.size(); cluster_block_indices_iter++) |
| 3107 | { |
| 3108 | const uint32_t block_index = cluster_block_indices[cluster_block_indices_iter]; |
| 3109 | |
| 3110 | const color_rgba *pBlock_pixels = get_source_pixel_block(block_index).get_ptr(); |
| 3111 | |
| 3112 | memcpy(&cluster_pixels[cluster_block_indices_iter * 16], pBlock_pixels, 16 * sizeof(color_rgba)); |
| 3113 | |
| 3114 | const uint32_t selector_cluster_index = pBlock_selector_indices ? (*pBlock_selector_indices)[block_index] : get_block_selector_cluster_index(block_index); |
| 3115 | |
| 3116 | const etc_block &blk_selectors = get_selector_cluster_selector_bits(selector_cluster_index); |
| 3117 | |
| 3118 | blk.set_raw_selector_bits(blk_selectors.get_raw_selector_bits()); |
| 3119 | |
| 3120 | cur_err += blk.evaluate_etc1_error(pBlock_pixels, m_params.m_perceptual); |
| 3121 | |
| 3122 | for (uint32_t y = 0; y < 4; y++) |
| 3123 | for (uint32_t x = 0; x < 4; x++) |
| 3124 | force_selectors[cluster_block_indices_iter * 16 + x + y * 4] = static_cast<uint8_t>(blk_selectors.get_selector(x, y)); |
| 3125 | } |
| 3126 | |
| 3127 | endpoint_cluster_etc_params new_endpoint_cluster_etc_params; |
| 3128 | |
| 3129 | { |
| 3130 | etc1_optimizer optimizer; |
| 3131 | etc1_solution_coordinates solutions[2]; |
| 3132 | |
| 3133 | etc1_optimizer::params cluster_optimizer_params; |
| 3134 | cluster_optimizer_params.m_num_src_pixels = total_pixels; |
| 3135 | cluster_optimizer_params.m_pSrc_pixels = &cluster_pixels[0]; |
| 3136 | |
| 3137 | cluster_optimizer_params.m_use_color4 = false; |
| 3138 | cluster_optimizer_params.m_perceptual = m_params.m_perceptual; |
| 3139 | cluster_optimizer_params.m_pForce_selectors = &force_selectors[0]; |
| 3140 | |
| 3141 | if (m_params.m_compression_level == BASISU_MAX_COMPRESSION_LEVEL) |
| 3142 | cluster_optimizer_params.m_quality = cETCQualityUber; |
| 3143 | else |
| 3144 | cluster_optimizer_params.m_quality = cETCQualitySlow; |
| 3145 | |
| 3146 | etc1_optimizer::results cluster_optimizer_results; |
| 3147 | |
| 3148 | basisu::vector<uint8_t> cluster_selectors(total_pixels); |
| 3149 | cluster_optimizer_results.m_n = total_pixels; |
| 3150 | cluster_optimizer_results.m_pSelectors = &cluster_selectors[0]; |
| 3151 | |
| 3152 | optimizer.init(cluster_optimizer_params, cluster_optimizer_results); |
| 3153 | |
| 3154 | if (!optimizer.compute()) |
| 3155 | BASISU_FRONTEND_VERIFY(false); |
| 3156 | |
| 3157 | new_endpoint_cluster_etc_params.m_color_unscaled[0] = cluster_optimizer_results.m_block_color_unscaled; |
| 3158 | new_endpoint_cluster_etc_params.m_inten_table[0] = cluster_optimizer_results.m_block_inten_table; |
| 3159 | new_endpoint_cluster_etc_params.m_color_error[0] = cluster_optimizer_results.m_error; |
| 3160 | new_endpoint_cluster_etc_params.m_color_used[0] = true; |
| 3161 | new_endpoint_cluster_etc_params.m_valid = true; |
| 3162 | } |
| 3163 | |
| 3164 | if (new_endpoint_cluster_etc_params.m_color_error[0] < cur_err) |
| 3165 | { |
| 3166 | m_endpoint_cluster_etc_params[cluster_index] = new_endpoint_cluster_etc_params; |
| 3167 | |
| 3168 | cluster_improved[cluster_index] = true; |
| 3169 | } |
| 3170 | |
| 3171 | cluster_valid[cluster_index] = true; |
| 3172 | |
| 3173 | } // cluster_index |
| 3174 | |
| 3175 | #ifndef __EMSCRIPTEN__ |
| 3176 | } ); |
| 3177 | #endif |
| 3178 | |
| 3179 | } // cluster_index_iter |
| 3180 | |
| 3181 | #ifndef __EMSCRIPTEN__ |
| 3182 | m_params.m_pJob_pool->wait_for_all(); |
| 3183 | #endif |
| 3184 | |
| 3185 | uint32_t total_unused_clusters = 0; |
| 3186 | uint32_t total_improved_clusters = 0; |
| 3187 | |
| 3188 | old_to_new_endpoint_cluster_indices.resize(m_endpoint_clusters.size()); |
| 3189 | vector_set_all(old_to_new_endpoint_cluster_indices, -1); |
| 3190 | |
| 3191 | int total_new_endpoint_clusters = 0; |
| 3192 | |
| 3193 | for (uint32_t old_cluster_index = 0; old_cluster_index < m_endpoint_clusters.size(); old_cluster_index++) |
| 3194 | { |
| 3195 | if (!cluster_valid[old_cluster_index]) |
| 3196 | total_unused_clusters++; |
| 3197 | else |
| 3198 | old_to_new_endpoint_cluster_indices[old_cluster_index] = total_new_endpoint_clusters++; |
| 3199 | |
| 3200 | if (cluster_improved[old_cluster_index]) |
| 3201 | total_improved_clusters++; |
| 3202 | } |
| 3203 | |
| 3204 | debug_printf("Total unused clusters: %u\n", total_unused_clusters); |
| 3205 | debug_printf("Total improved_clusters: %u\n", total_improved_clusters); |
| 3206 | debug_printf("Total endpoint clusters: %u\n", total_new_endpoint_clusters); |
| 3207 | |
| 3208 | if (optimize_final_codebook) |
| 3209 | { |
| 3210 | cluster_subblock_etc_params_vec new_endpoint_cluster_etc_params(total_new_endpoint_clusters); |
| 3211 | |
| 3212 | for (uint32_t old_cluster_index = 0; old_cluster_index < m_endpoint_clusters.size(); old_cluster_index++) |
| 3213 | { |
| 3214 | if (old_to_new_endpoint_cluster_indices[old_cluster_index] >= 0) |
| 3215 | new_endpoint_cluster_etc_params[old_to_new_endpoint_cluster_indices[old_cluster_index]] = m_endpoint_cluster_etc_params[old_cluster_index]; |
| 3216 | } |
| 3217 | |
| 3218 | debug_printf("basisu_frontend::reoptimize_remapped_endpoints: stage 1\n"); |
| 3219 | |
| 3220 | basisu::vector<uint_vec> new_endpoint_clusters(total_new_endpoint_clusters); |
| 3221 | |
| 3222 | for (uint32_t block_index = 0; block_index < new_block_endpoints.size(); block_index++) |
| 3223 | { |
| 3224 | const uint32_t old_endpoint_cluster_index = new_block_endpoints[block_index]; |
| 3225 | |
| 3226 | const int new_endpoint_cluster_index = old_to_new_endpoint_cluster_indices[old_endpoint_cluster_index]; |
| 3227 | BASISU_FRONTEND_VERIFY(new_endpoint_cluster_index >= 0); |
| 3228 | |
| 3229 | BASISU_FRONTEND_VERIFY(new_endpoint_cluster_index < (int)new_endpoint_clusters.size()); |
| 3230 | |
| 3231 | new_endpoint_clusters[new_endpoint_cluster_index].push_back(block_index * 2 + 0); |
| 3232 | new_endpoint_clusters[new_endpoint_cluster_index].push_back(block_index * 2 + 1); |
| 3233 | |
| 3234 | BASISU_FRONTEND_VERIFY(new_endpoint_cluster_index < (int)new_endpoint_cluster_etc_params.size()); |
| 3235 | |
| 3236 | new_endpoint_cluster_etc_params[new_endpoint_cluster_index].m_subblocks.push_back(block_index * 2 + 0); |
| 3237 | new_endpoint_cluster_etc_params[new_endpoint_cluster_index].m_subblocks.push_back(block_index * 2 + 1); |
| 3238 | |
| 3239 | m_block_endpoint_clusters_indices[block_index][0] = new_endpoint_cluster_index; |
| 3240 | m_block_endpoint_clusters_indices[block_index][1] = new_endpoint_cluster_index; |
| 3241 | } |
| 3242 | |
| 3243 | debug_printf("basisu_frontend::reoptimize_remapped_endpoints: stage 2\n"); |
| 3244 | |
| 3245 | m_endpoint_clusters = new_endpoint_clusters; |
| 3246 | m_endpoint_cluster_etc_params = new_endpoint_cluster_etc_params; |
| 3247 | |
| 3248 | eliminate_redundant_or_empty_endpoint_clusters(); |
| 3249 | |
| 3250 | debug_printf("basisu_frontend::reoptimize_remapped_endpoints: stage 3\n"); |
| 3251 | |
| 3252 | for (uint32_t new_cluster_index = 0; new_cluster_index < m_endpoint_clusters.size(); new_cluster_index++) |
| 3253 | { |
| 3254 | for (uint32_t cluster_block_iter = 0; cluster_block_iter < m_endpoint_clusters[new_cluster_index].size(); cluster_block_iter++) |
| 3255 | { |
| 3256 | const uint32_t subblock_index = m_endpoint_clusters[new_cluster_index][cluster_block_iter]; |
| 3257 | const uint32_t block_index = subblock_index >> 1; |
| 3258 | |
| 3259 | m_block_endpoint_clusters_indices[block_index][0] = new_cluster_index; |
| 3260 | m_block_endpoint_clusters_indices[block_index][1] = new_cluster_index; |
| 3261 | |
| 3262 | const uint32_t old_cluster_index = new_block_endpoints[block_index]; |
| 3263 | |
| 3264 | old_to_new_endpoint_cluster_indices[old_cluster_index] = new_cluster_index; |
| 3265 | } |
| 3266 | } |
| 3267 | |
| 3268 | debug_printf("basisu_frontend::reoptimize_remapped_endpoints: stage 4\n"); |
| 3269 | |
| 3270 | for (uint32_t block_index = 0; block_index < m_encoded_blocks.size(); block_index++) |
| 3271 | { |
| 3272 | const uint32_t endpoint_cluster_index = get_subblock_endpoint_cluster_index(block_index, 0); |
| 3273 | |
| 3274 | m_encoded_blocks[block_index].set_block_color5_etc1s(get_endpoint_cluster_unscaled_color(endpoint_cluster_index, false)); |
| 3275 | m_encoded_blocks[block_index].set_inten_tables_etc1s(get_endpoint_cluster_inten_table(endpoint_cluster_index, false)); |
| 3276 | } |
| 3277 | |
| 3278 | debug_printf("Final (post-RDO) endpoint clusters: %u\n", m_endpoint_clusters.size()); |
| 3279 | } |
| 3280 | |
| 3281 | //debug_printf("validate_output: %u\n", validate_output()); |
| 3282 | } |
| 3283 | |
| 3284 | // Endpoint clusterization hierarchy integrity checker. |
| 3285 | // Note this doesn't check for empty clusters. |
| 3286 | bool basisu_frontend::validate_endpoint_cluster_hierarchy(bool ensure_clusters_have_same_parents) const |
| 3287 | { |
| 3288 | if (!m_endpoint_parent_clusters.size()) |
| 3289 | return true; |
| 3290 | |
| 3291 | int_vec subblock_parent_indices(m_total_blocks * 2); |
| 3292 | subblock_parent_indices.set_all(-1); |
| 3293 | |
| 3294 | int_vec subblock_cluster_indices(m_total_blocks * 2); |
| 3295 | subblock_cluster_indices.set_all(-1); |
| 3296 | |
| 3297 | for (uint32_t parent_index = 0; parent_index < m_endpoint_parent_clusters.size(); parent_index++) |
| 3298 | { |
| 3299 | for (uint32_t i = 0; i < m_endpoint_parent_clusters[parent_index].size(); i++) |
| 3300 | { |
| 3301 | uint32_t subblock_index = m_endpoint_parent_clusters[parent_index][i]; |
| 3302 | if (subblock_index >= m_total_blocks * 2) |
| 3303 | return false; |
| 3304 | |
| 3305 | // If the endpoint cluster lives in more than one parent node, that's wrong. |
| 3306 | if (subblock_parent_indices[subblock_index] != -1) |
| 3307 | return false; |
| 3308 | |
| 3309 | subblock_parent_indices[subblock_index] = parent_index; |
| 3310 | } |
| 3311 | } |
| 3312 | |
| 3313 | // Make sure all endpoint clusters are present in the parent cluster. |
| 3314 | for (uint32_t i = 0; i < subblock_parent_indices.size(); i++) |
| 3315 | { |
| 3316 | if (subblock_parent_indices[i] == -1) |
| 3317 | return false; |
| 3318 | } |
| 3319 | |
| 3320 | for (uint32_t cluster_index = 0; cluster_index < m_endpoint_clusters.size(); cluster_index++) |
| 3321 | { |
| 3322 | int parent_index = 0; |
| 3323 | |
| 3324 | for (uint32_t i = 0; i < m_endpoint_clusters[cluster_index].size(); i++) |
| 3325 | { |
| 3326 | uint32_t subblock_index = m_endpoint_clusters[cluster_index][i]; |
| 3327 | if (subblock_index >= m_total_blocks * 2) |
| 3328 | return false; |
| 3329 | |
| 3330 | if (subblock_cluster_indices[subblock_index] != -1) |
| 3331 | return false; |
| 3332 | |
| 3333 | subblock_cluster_indices[subblock_index] = cluster_index; |
| 3334 | |
| 3335 | // There are transformations on the endpoint clusters that can break the strict tree requirement |
| 3336 | if (ensure_clusters_have_same_parents) |
| 3337 | { |
| 3338 | // Make sure all the subblocks are in the same parent cluster |
| 3339 | if (!i) |
| 3340 | parent_index = subblock_parent_indices[subblock_index]; |
| 3341 | else if (subblock_parent_indices[subblock_index] != parent_index) |
| 3342 | return false; |
| 3343 | } |
| 3344 | } |
| 3345 | } |
| 3346 | |
| 3347 | // Make sure all endpoint clusters are present in the parent cluster. |
| 3348 | for (uint32_t i = 0; i < subblock_cluster_indices.size(); i++) |
| 3349 | { |
| 3350 | if (subblock_cluster_indices[i] == -1) |
| 3351 | return false; |
| 3352 | } |
| 3353 | |
| 3354 | return true; |
| 3355 | } |
| 3356 | |
| 3357 | // This is very slow and only intended for debugging/development. It's enabled using the "-validate_etc1s" command line option. |
| 3358 | bool basisu_frontend::validate_output() const |
| 3359 | { |
| 3360 | debug_printf("validate_output\n"); |
| 3361 | |
| 3362 | if (!check_etc1s_constraints()) |
| 3363 | return false; |
| 3364 | |
| 3365 | for (uint32_t block_index = 0; block_index < m_total_blocks; block_index++) |
| 3366 | { |
| 3367 | //#define CHECK(x) do { if (!(x)) { DebugBreak(); return false; } } while(0) |
| 3368 | #define CHECK(x) BASISU_FRONTEND_VERIFY(x); |
| 3369 | |
| 3370 | CHECK(get_output_block(block_index).get_flip_bit() == true); |
| 3371 | |
| 3372 | const bool diff_flag = get_diff_flag(block_index); |
| 3373 | CHECK(diff_flag == true); |
| 3374 | |
| 3375 | etc_block blk; |
| 3376 | memset(&blk, 0, sizeof(blk)); |
| 3377 | blk.set_flip_bit(true); |
| 3378 | blk.set_diff_bit(true); |
| 3379 | |
| 3380 | const uint32_t endpoint_cluster0_index = get_subblock_endpoint_cluster_index(block_index, 0); |
| 3381 | const uint32_t endpoint_cluster1_index = get_subblock_endpoint_cluster_index(block_index, 1); |
| 3382 | |
| 3383 | // basisu only supports ETC1S, so these must be equal. |
| 3384 | CHECK(endpoint_cluster0_index == endpoint_cluster1_index); |
| 3385 | |
| 3386 | CHECK(blk.set_block_color5_check(get_endpoint_cluster_unscaled_color(endpoint_cluster0_index, false), get_endpoint_cluster_unscaled_color(endpoint_cluster1_index, false))); |
| 3387 | |
| 3388 | CHECK(get_endpoint_cluster_color_is_used(endpoint_cluster0_index, false)); |
| 3389 | |
| 3390 | blk.set_inten_table(0, get_endpoint_cluster_inten_table(endpoint_cluster0_index, false)); |
| 3391 | blk.set_inten_table(1, get_endpoint_cluster_inten_table(endpoint_cluster1_index, false)); |
| 3392 | |
| 3393 | const uint32_t selector_cluster_index = get_block_selector_cluster_index(block_index); |
| 3394 | CHECK(selector_cluster_index < get_total_selector_clusters()); |
| 3395 | |
| 3396 | CHECK(vector_find(get_selector_cluster_block_indices(selector_cluster_index), block_index) != -1); |
| 3397 | |
| 3398 | blk.set_raw_selector_bits(get_selector_cluster_selector_bits(selector_cluster_index).get_raw_selector_bits()); |
| 3399 | |
| 3400 | const etc_block &rdo_output_block = get_output_block(block_index); |
| 3401 | |
| 3402 | CHECK(rdo_output_block.get_flip_bit() == blk.get_flip_bit()); |
| 3403 | CHECK(rdo_output_block.get_diff_bit() == blk.get_diff_bit()); |
| 3404 | CHECK(rdo_output_block.get_inten_table(0) == blk.get_inten_table(0)); |
| 3405 | CHECK(rdo_output_block.get_inten_table(1) == blk.get_inten_table(1)); |
| 3406 | CHECK(rdo_output_block.get_base5_color() == blk.get_base5_color()); |
| 3407 | CHECK(rdo_output_block.get_delta3_color() == blk.get_delta3_color()); |
| 3408 | CHECK(rdo_output_block.get_raw_selector_bits() == blk.get_raw_selector_bits()); |
| 3409 | |
| 3410 | #undef CHECK |
| 3411 | } |
| 3412 | |
| 3413 | return true; |
| 3414 | } |
| 3415 | |
| 3416 | void basisu_frontend::dump_debug_image(const char *pFilename, uint32_t first_block, uint32_t num_blocks_x, uint32_t num_blocks_y, bool output_blocks) |
| 3417 | { |
| 3418 | gpu_image g; |
| 3419 | g.init(texture_format::cETC1, num_blocks_x * 4, num_blocks_y * 4); |
| 3420 | |
| 3421 | for (uint32_t y = 0; y < num_blocks_y; y++) |
| 3422 | { |
| 3423 | for (uint32_t x = 0; x < num_blocks_x; x++) |
| 3424 | { |
| 3425 | const uint32_t block_index = first_block + x + y * num_blocks_x; |
| 3426 | |
| 3427 | etc_block &blk = *(etc_block *)g.get_block_ptr(x, y); |
| 3428 | |
| 3429 | if (output_blocks) |
| 3430 | blk = get_output_block(block_index); |
| 3431 | else |
| 3432 | { |
| 3433 | const bool diff_flag = get_diff_flag(block_index); |
| 3434 | |
| 3435 | blk.set_diff_bit(diff_flag); |
| 3436 | blk.set_flip_bit(true); |
| 3437 | |
| 3438 | const uint32_t endpoint_cluster0_index = get_subblock_endpoint_cluster_index(block_index, 0); |
| 3439 | const uint32_t endpoint_cluster1_index = get_subblock_endpoint_cluster_index(block_index, 1); |
| 3440 | |
| 3441 | if (diff_flag) |
| 3442 | blk.set_block_color5(get_endpoint_cluster_unscaled_color(endpoint_cluster0_index, false), get_endpoint_cluster_unscaled_color(endpoint_cluster1_index, false)); |
| 3443 | else |
| 3444 | blk.set_block_color4(get_endpoint_cluster_unscaled_color(endpoint_cluster0_index, true), get_endpoint_cluster_unscaled_color(endpoint_cluster1_index, true)); |
| 3445 | |
| 3446 | blk.set_inten_table(0, get_endpoint_cluster_inten_table(endpoint_cluster0_index, !diff_flag)); |
| 3447 | blk.set_inten_table(1, get_endpoint_cluster_inten_table(endpoint_cluster1_index, !diff_flag)); |
| 3448 | |
| 3449 | const uint32_t selector_cluster_index = get_block_selector_cluster_index(block_index); |
| 3450 | blk.set_raw_selector_bits(get_selector_cluster_selector_bits(selector_cluster_index).get_raw_selector_bits()); |
| 3451 | } |
| 3452 | } |
| 3453 | } |
| 3454 | |
| 3455 | image img; |
| 3456 | g.unpack(img); |
| 3457 | |
| 3458 | save_png(pFilename, img); |
| 3459 | } |
| 3460 | |
| 3461 | } // namespace basisu |
| 3462 | |
| 3463 |
Generated on 2023-Sep-17 from project Godot revision 4.2
Powered by 
Code Browser 2.1
Generator usage only permitted with license.