| 1 | // basisu_uastc_enc.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 | #include "basisu_uastc_enc.h" |
| 16 | |
| 17 | #if BASISU_USE_ASTC_DECOMPRESS |
| 18 | #include "basisu_astc_decomp.h" |
| 19 | #endif |
| 20 | |
| 21 | #include "basisu_gpu_texture.h" |
| 22 | #include "basisu_bc7enc.h" |
| 23 | |
| 24 | #ifdef _DEBUG |
| 25 | // When BASISU_VALIDATE_UASTC_ENC is 1, we pack and unpack to/from UASTC and ASTC, then validate that each codec returns the exact same results. This is slower. |
| 26 | #define BASISU_VALIDATE_UASTC_ENC 1 |
| 27 | #endif |
| 28 | |
| 29 | #define BASISU_SUPPORT_FORCE_MODE 0 |
| 30 | |
| 31 | using namespace basist; |
| 32 | |
| 33 | namespace basisu |
| 34 | { |
| 35 | const uint32_t MAX_ENCODE_RESULTS = 512; |
| 36 | |
| 37 | #if BASISU_VALIDATE_UASTC_ENC |
| 38 | static void validate_func(bool condition, int line) |
| 39 | { |
| 40 | if (!condition) |
| 41 | { |
| 42 | fprintf(stderr, "basisu_uastc_enc: Internal validation failed on line %u!\n", line); |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | #define VALIDATE(c) validate_func(c, __LINE__); |
| 47 | #else |
| 48 | #define VALIDATE(c) |
| 49 | #endif |
| 50 | |
| 51 | enum dxt_constants |
| 52 | { |
| 53 | cDXT1SelectorBits = 2U, cDXT1SelectorValues = 1U << cDXT1SelectorBits, cDXT1SelectorMask = cDXT1SelectorValues - 1U, |
| 54 | cDXT5SelectorBits = 3U, cDXT5SelectorValues = 1U << cDXT5SelectorBits, cDXT5SelectorMask = cDXT5SelectorValues - 1U, |
| 55 | }; |
| 56 | |
| 57 | struct dxt1_block |
| 58 | { |
| 59 | enum { cTotalEndpointBytes = 2, cTotalSelectorBytes = 4 }; |
| 60 | |
| 61 | uint8_t m_low_color[cTotalEndpointBytes]; |
| 62 | uint8_t m_high_color[cTotalEndpointBytes]; |
| 63 | uint8_t m_selectors[cTotalSelectorBytes]; |
| 64 | |
| 65 | inline void clear() { basisu::clear_obj(*this); } |
| 66 | |
| 67 | inline uint32_t get_high_color() const { return m_high_color[0] | (m_high_color[1] << 8U); } |
| 68 | inline uint32_t get_low_color() const { return m_low_color[0] | (m_low_color[1] << 8U); } |
| 69 | inline void set_low_color(uint16_t c) { m_low_color[0] = static_cast<uint8_t>(c & 0xFF); m_low_color[1] = static_cast<uint8_t>((c >> 8) & 0xFF); } |
| 70 | inline void set_high_color(uint16_t c) { m_high_color[0] = static_cast<uint8_t>(c & 0xFF); m_high_color[1] = static_cast<uint8_t>((c >> 8) & 0xFF); } |
| 71 | inline uint32_t get_selector(uint32_t x, uint32_t y) const { assert((x < 4U) && (y < 4U)); return (m_selectors[y] >> (x * cDXT1SelectorBits))& cDXT1SelectorMask; } |
| 72 | inline void set_selector(uint32_t x, uint32_t y, uint32_t val) { assert((x < 4U) && (y < 4U) && (val < 4U)); m_selectors[y] &= (~(cDXT1SelectorMask << (x * cDXT1SelectorBits))); m_selectors[y] |= (val << (x * cDXT1SelectorBits)); } |
| 73 | |
| 74 | static uint16_t pack_color(const color_rgba& color, bool scaled, uint32_t bias = 127U) |
| 75 | { |
| 76 | uint32_t r = color.r, g = color.g, b = color.b; |
| 77 | if (scaled) |
| 78 | { |
| 79 | r = (r * 31U + bias) / 255U; |
| 80 | g = (g * 63U + bias) / 255U; |
| 81 | b = (b * 31U + bias) / 255U; |
| 82 | } |
| 83 | return static_cast<uint16_t>(basisu::minimum(b, 31U) | (basisu::minimum(g, 63U) << 5U) | (basisu::minimum(r, 31U) << 11U)); |
| 84 | } |
| 85 | |
| 86 | static uint16_t pack_unscaled_color(uint32_t r, uint32_t g, uint32_t b) { return static_cast<uint16_t>(b | (g << 5U) | (r << 11U)); } |
| 87 | }; |
| 88 | |
| 89 | #define UASTC_WRITE_MODE_DESCS 0 |
| 90 | |
| 91 | static inline void uastc_write_bits(uint8_t* pBuf, uint32_t& bit_offset, uint64_t code, uint32_t codesize, const char* pDesc) |
| 92 | { |
| 93 | (void)pDesc; |
| 94 | |
| 95 | #if UASTC_WRITE_MODE_DESCS |
| 96 | if (pDesc) |
| 97 | printf("%s: %u %u\n", pDesc, bit_offset, codesize); |
| 98 | #endif |
| 99 | |
| 100 | assert((codesize == 64) || (code < (1ULL << codesize))); |
| 101 | |
| 102 | while (codesize) |
| 103 | { |
| 104 | uint32_t byte_bit_offset = bit_offset & 7; |
| 105 | uint32_t bits_to_write = basisu::minimum<int>(codesize, 8 - byte_bit_offset); |
| 106 | |
| 107 | pBuf[bit_offset >> 3] |= (code << byte_bit_offset); |
| 108 | |
| 109 | code >>= bits_to_write; |
| 110 | codesize -= bits_to_write; |
| 111 | bit_offset += bits_to_write; |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | void pack_uastc(basist::uastc_block& blk, const uastc_encode_results& result, const etc_block& etc1_blk, uint32_t etc1_bias, const eac_a8_block& etc_eac_a8_blk, bool bc1_hint0, bool bc1_hint1) |
| 116 | { |
| 117 | if ((g_uastc_mode_has_alpha[result.m_uastc_mode]) && (result.m_uastc_mode != UASTC_MODE_INDEX_SOLID_COLOR)) |
| 118 | { |
| 119 | assert(etc_eac_a8_blk.m_multiplier >= 1); |
| 120 | } |
| 121 | |
| 122 | uint8_t buf[32]; |
| 123 | memset(buf, 0, sizeof(buf)); |
| 124 | |
| 125 | uint32_t block_bit_offset = 0; |
| 126 | |
| 127 | #if UASTC_WRITE_MODE_DESCS |
| 128 | printf("**** Mode: %u\n", result.m_uastc_mode); |
| 129 | #endif |
| 130 | |
| 131 | uastc_write_bits(buf, block_bit_offset, g_uastc_mode_huff_codes[result.m_uastc_mode][0], g_uastc_mode_huff_codes[result.m_uastc_mode][1], "mode"); |
| 132 | |
| 133 | if (result.m_uastc_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 134 | { |
| 135 | uastc_write_bits(buf, block_bit_offset, result.m_solid_color.r, 8, "R"); |
| 136 | uastc_write_bits(buf, block_bit_offset, result.m_solid_color.g, 8, "G"); |
| 137 | uastc_write_bits(buf, block_bit_offset, result.m_solid_color.b, 8, "B"); |
| 138 | uastc_write_bits(buf, block_bit_offset, result.m_solid_color.a, 8, "A"); |
| 139 | |
| 140 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_diff_bit(), 1, "ETC1D"); |
| 141 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_inten_table(0), 3, "ETC1I"); |
| 142 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_selector(0, 0), 2, "ETC1S"); |
| 143 | |
| 144 | uint32_t r, g, b; |
| 145 | if (etc1_blk.get_diff_bit()) |
| 146 | etc_block::unpack_color5(r, g, b, etc1_blk.get_base5_color(), false); |
| 147 | else |
| 148 | etc_block::unpack_color4(r, g, b, etc1_blk.get_base4_color(0), false); |
| 149 | |
| 150 | uastc_write_bits(buf, block_bit_offset, r, 5, "ETC1R"); |
| 151 | uastc_write_bits(buf, block_bit_offset, g, 5, "ETC1G"); |
| 152 | uastc_write_bits(buf, block_bit_offset, b, 5, "ETC1B"); |
| 153 | |
| 154 | memcpy(&blk, buf, sizeof(blk)); |
| 155 | return; |
| 156 | } |
| 157 | |
| 158 | if (g_uastc_mode_has_bc1_hint0[result.m_uastc_mode]) |
| 159 | uastc_write_bits(buf, block_bit_offset, bc1_hint0, 1, "BC1H0"); |
| 160 | else |
| 161 | { |
| 162 | assert(bc1_hint0 == false); |
| 163 | } |
| 164 | |
| 165 | if (g_uastc_mode_has_bc1_hint1[result.m_uastc_mode]) |
| 166 | uastc_write_bits(buf, block_bit_offset, bc1_hint1, 1, "BC1H1"); |
| 167 | else |
| 168 | { |
| 169 | assert(bc1_hint1 == false); |
| 170 | } |
| 171 | |
| 172 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_flip_bit(), 1, "ETC1F"); |
| 173 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_diff_bit(), 1, "ETC1D"); |
| 174 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_inten_table(0), 3, "ETC1I0"); |
| 175 | uastc_write_bits(buf, block_bit_offset, etc1_blk.get_inten_table(1), 3, "ETC1I1"); |
| 176 | |
| 177 | if (g_uastc_mode_has_etc1_bias[result.m_uastc_mode]) |
| 178 | uastc_write_bits(buf, block_bit_offset, etc1_bias, 5, "ETC1BIAS"); |
| 179 | else |
| 180 | { |
| 181 | assert(etc1_bias == 0); |
| 182 | } |
| 183 | |
| 184 | if (g_uastc_mode_has_alpha[result.m_uastc_mode]) |
| 185 | { |
| 186 | const uint32_t etc2_hints = etc_eac_a8_blk.m_table | (etc_eac_a8_blk.m_multiplier << 4); |
| 187 | |
| 188 | assert(etc2_hints > 0 && etc2_hints <= 0xFF); |
| 189 | uastc_write_bits(buf, block_bit_offset, etc2_hints, 8, "ETC2TM"); |
| 190 | } |
| 191 | |
| 192 | uint32_t subsets = 1; |
| 193 | switch (result.m_uastc_mode) |
| 194 | { |
| 195 | case 2: |
| 196 | case 4: |
| 197 | case 7: |
| 198 | case 9: |
| 199 | case 16: |
| 200 | uastc_write_bits(buf, block_bit_offset, result.m_common_pattern, 5, "PAT"); |
| 201 | subsets = 2; |
| 202 | break; |
| 203 | case 3: |
| 204 | uastc_write_bits(buf, block_bit_offset, result.m_common_pattern, 4, "PAT"); |
| 205 | subsets = 3; |
| 206 | break; |
| 207 | default: |
| 208 | break; |
| 209 | } |
| 210 | |
| 211 | #ifdef _DEBUG |
| 212 | uint32_t part_seed = 0; |
| 213 | switch (result.m_uastc_mode) |
| 214 | { |
| 215 | case 2: |
| 216 | case 4: |
| 217 | case 9: |
| 218 | case 16: |
| 219 | part_seed = g_astc_bc7_common_partitions2[result.m_common_pattern].m_astc; |
| 220 | break; |
| 221 | case 3: |
| 222 | part_seed = g_astc_bc7_common_partitions3[result.m_common_pattern].m_astc; |
| 223 | break; |
| 224 | case 7: |
| 225 | part_seed = g_bc7_3_astc2_common_partitions[result.m_common_pattern].m_astc2; |
| 226 | break; |
| 227 | default: |
| 228 | break; |
| 229 | } |
| 230 | #endif |
| 231 | |
| 232 | uint32_t total_planes = 1; |
| 233 | switch (result.m_uastc_mode) |
| 234 | { |
| 235 | case 6: |
| 236 | case 11: |
| 237 | case 13: |
| 238 | uastc_write_bits(buf, block_bit_offset, result.m_astc.m_ccs, 2, "COMPSEL"); |
| 239 | total_planes = 2; |
| 240 | break; |
| 241 | case 17: |
| 242 | // CCS field is always 3 for dual plane LA. |
| 243 | assert(result.m_astc.m_ccs == 3); |
| 244 | total_planes = 2; |
| 245 | break; |
| 246 | default: |
| 247 | break; |
| 248 | } |
| 249 | |
| 250 | uint8_t weights[32]; |
| 251 | memcpy(weights, result.m_astc.m_weights, 16 * total_planes); |
| 252 | |
| 253 | uint8_t endpoints[18]; |
| 254 | memcpy(endpoints, result.m_astc.m_endpoints, sizeof(endpoints)); |
| 255 | |
| 256 | const uint32_t total_comps = g_uastc_mode_comps[result.m_uastc_mode]; |
| 257 | |
| 258 | // LLAA |
| 259 | // LLAA LLAA |
| 260 | // LLAA LLAA LLAA |
| 261 | // RRGGBB |
| 262 | // RRGGBB RRGGBB |
| 263 | // RRGGBB RRGGBB RRGGBB |
| 264 | // RRGGBBAA |
| 265 | // RRGGBBAA RRGGBBAA |
| 266 | |
| 267 | const uint32_t weight_bits = g_uastc_mode_weight_bits[result.m_uastc_mode]; |
| 268 | |
| 269 | const uint8_t* pPartition_pattern; |
| 270 | const uint8_t* pSubset_anchor_indices = basist::get_anchor_indices(subsets, result.m_uastc_mode, result.m_common_pattern, pPartition_pattern); |
| 271 | |
| 272 | for (uint32_t plane_index = 0; plane_index < total_planes; plane_index++) |
| 273 | { |
| 274 | for (uint32_t subset_index = 0; subset_index < subsets; subset_index++) |
| 275 | { |
| 276 | const uint32_t anchor_index = pSubset_anchor_indices[subset_index]; |
| 277 | |
| 278 | #ifdef _DEBUG |
| 279 | if (subsets >= 2) |
| 280 | { |
| 281 | for (uint32_t i = 0; i < 16; i++) |
| 282 | { |
| 283 | const uint32_t part_index = astc_compute_texel_partition(part_seed, i & 3, i >> 2, 0, subsets, true); |
| 284 | if (part_index == subset_index) |
| 285 | { |
| 286 | assert(anchor_index == i); |
| 287 | break; |
| 288 | } |
| 289 | } |
| 290 | } |
| 291 | else |
| 292 | { |
| 293 | assert(!anchor_index); |
| 294 | } |
| 295 | #endif |
| 296 | |
| 297 | // Check anchor weight's MSB - if it's set then invert this subset's weights and swap the endpoints |
| 298 | if (weights[anchor_index * total_planes + plane_index] & (1 << (weight_bits - 1))) |
| 299 | { |
| 300 | for (uint32_t i = 0; i < 16; i++) |
| 301 | { |
| 302 | const uint32_t part_index = pPartition_pattern[i]; |
| 303 | |
| 304 | #ifdef _DEBUG |
| 305 | if (subsets >= 2) |
| 306 | { |
| 307 | assert(part_index == (uint32_t)astc_compute_texel_partition(part_seed, i & 3, i >> 2, 0, subsets, true)); |
| 308 | } |
| 309 | else |
| 310 | { |
| 311 | assert(!part_index); |
| 312 | } |
| 313 | #endif |
| 314 | |
| 315 | if (part_index == subset_index) |
| 316 | weights[i * total_planes + plane_index] = ((1 << weight_bits) - 1) - weights[i * total_planes + plane_index]; |
| 317 | } |
| 318 | |
| 319 | if (total_planes == 2) |
| 320 | { |
| 321 | for (int c = 0; c < (int)total_comps; c++) |
| 322 | { |
| 323 | const uint32_t comp_plane = (total_comps == 2) ? c : ((c == result.m_astc.m_ccs) ? 1 : 0); |
| 324 | |
| 325 | if (comp_plane == plane_index) |
| 326 | std::swap(endpoints[c * 2 + 0], endpoints[c * 2 + 1]); |
| 327 | } |
| 328 | } |
| 329 | else |
| 330 | { |
| 331 | for (uint32_t c = 0; c < total_comps; c++) |
| 332 | std::swap(endpoints[subset_index * total_comps * 2 + c * 2 + 0], endpoints[subset_index * total_comps * 2 + c * 2 + 1]); |
| 333 | } |
| 334 | } |
| 335 | } // subset_index |
| 336 | } // plane_index |
| 337 | |
| 338 | const uint32_t total_values = total_comps * 2 * subsets; |
| 339 | const uint32_t endpoint_range = g_uastc_mode_endpoint_ranges[result.m_uastc_mode]; |
| 340 | |
| 341 | uint32_t bit_values[18]; |
| 342 | uint32_t tq_values[8]; |
| 343 | uint32_t total_tq_values = 0; |
| 344 | uint32_t tq_accum = 0; |
| 345 | uint32_t tq_mul = 1; |
| 346 | |
| 347 | const uint32_t ep_bits = g_astc_bise_range_table[endpoint_range][0]; |
| 348 | const uint32_t ep_trits = g_astc_bise_range_table[endpoint_range][1]; |
| 349 | const uint32_t ep_quints = g_astc_bise_range_table[endpoint_range][2]; |
| 350 | |
| 351 | for (uint32_t i = 0; i < total_values; i++) |
| 352 | { |
| 353 | uint32_t val = endpoints[i]; |
| 354 | |
| 355 | uint32_t bits = val & ((1 << ep_bits) - 1); |
| 356 | uint32_t tq = val >> ep_bits; |
| 357 | |
| 358 | bit_values[i] = bits; |
| 359 | |
| 360 | if (ep_trits) |
| 361 | { |
| 362 | assert(tq < 3); |
| 363 | tq_accum += tq * tq_mul; |
| 364 | tq_mul *= 3; |
| 365 | if (tq_mul == 243) |
| 366 | { |
| 367 | tq_values[total_tq_values++] = tq_accum; |
| 368 | tq_accum = 0; |
| 369 | tq_mul = 1; |
| 370 | } |
| 371 | } |
| 372 | else if (ep_quints) |
| 373 | { |
| 374 | assert(tq < 5); |
| 375 | tq_accum += tq * tq_mul; |
| 376 | tq_mul *= 5; |
| 377 | if (tq_mul == 125) |
| 378 | { |
| 379 | tq_values[total_tq_values++] = tq_accum; |
| 380 | tq_accum = 0; |
| 381 | tq_mul = 1; |
| 382 | } |
| 383 | } |
| 384 | } |
| 385 | |
| 386 | uint32_t total_endpoint_bits = 0; |
| 387 | |
| 388 | for (uint32_t i = 0; i < total_tq_values; i++) |
| 389 | { |
| 390 | const uint32_t num_bits = ep_trits ? 8 : 7; |
| 391 | uastc_write_bits(buf, block_bit_offset, tq_values[i], num_bits, "ETQ"); |
| 392 | total_endpoint_bits += num_bits; |
| 393 | } |
| 394 | |
| 395 | if (tq_mul > 1) |
| 396 | { |
| 397 | uint32_t num_bits; |
| 398 | if (ep_trits) |
| 399 | { |
| 400 | if (tq_mul == 3) |
| 401 | num_bits = 2; |
| 402 | else if (tq_mul == 9) |
| 403 | num_bits = 4; |
| 404 | else if (tq_mul == 27) |
| 405 | num_bits = 5; |
| 406 | else //if (tq_mul == 81) |
| 407 | num_bits = 7; |
| 408 | } |
| 409 | else |
| 410 | { |
| 411 | if (tq_mul == 5) |
| 412 | num_bits = 3; |
| 413 | else //if (tq_mul == 25) |
| 414 | num_bits = 5; |
| 415 | } |
| 416 | uastc_write_bits(buf, block_bit_offset, tq_accum, num_bits, "ETQ"); |
| 417 | total_endpoint_bits += num_bits; |
| 418 | } |
| 419 | |
| 420 | for (uint32_t i = 0; i < total_values; i++) |
| 421 | { |
| 422 | uastc_write_bits(buf, block_bit_offset, bit_values[i], ep_bits, "EBITS"); |
| 423 | total_endpoint_bits += ep_bits; |
| 424 | } |
| 425 | |
| 426 | #if UASTC_WRITE_MODE_DESCS |
| 427 | uint32_t weight_start = block_bit_offset; |
| 428 | #endif |
| 429 | |
| 430 | uint32_t total_weight_bits = 0; |
| 431 | const uint32_t plane_shift = (total_planes == 2) ? 1 : 0; |
| 432 | for (uint32_t i = 0; i < 16 * total_planes; i++) |
| 433 | { |
| 434 | uint32_t numbits = weight_bits; |
| 435 | for (uint32_t s = 0; s < subsets; s++) |
| 436 | { |
| 437 | if (pSubset_anchor_indices[s] == (i >> plane_shift)) |
| 438 | { |
| 439 | numbits--; |
| 440 | break; |
| 441 | } |
| 442 | } |
| 443 | |
| 444 | uastc_write_bits(buf, block_bit_offset, weights[i], numbits, nullptr); |
| 445 | |
| 446 | total_weight_bits += numbits; |
| 447 | } |
| 448 | |
| 449 | #if UASTC_WRITE_MODE_DESCS |
| 450 | printf("WEIGHTS: %u %u\n", weight_start, total_weight_bits); |
| 451 | #endif |
| 452 | |
| 453 | assert(block_bit_offset <= 128); |
| 454 | memcpy(&blk, buf, sizeof(blk)); |
| 455 | |
| 456 | #if UASTC_WRITE_MODE_DESCS |
| 457 | printf("Total bits: %u, endpoint bits: %u, weight bits: %u\n", block_bit_offset, total_endpoint_bits, total_weight_bits); |
| 458 | #endif |
| 459 | } |
| 460 | |
| 461 | // MODE 0 |
| 462 | // 0. DualPlane: 0, WeightRange: 8 (16), Subsets: 1, CEM: 8 (RGB Direct ), EndpointRange: 19 (192) MODE6 RGB |
| 463 | // 18. DualPlane: 0, WeightRange: 11 (32), Subsets: 1, CEM: 8 (RGB Direct ), EndpointRange: 11 (32) MODE6 RGB |
| 464 | static void astc_mode0_or_18(uint32_t mode, const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params, const uint8_t *pForce_selectors = nullptr) |
| 465 | { |
| 466 | const uint32_t endpoint_range = (mode == 18) ? 11 : 19; |
| 467 | const uint32_t weight_range = (mode == 18) ? 11 : 8; |
| 468 | |
| 469 | color_cell_compressor_params ccell_params; |
| 470 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 471 | |
| 472 | ccell_params.m_num_pixels = 16; |
| 473 | ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 474 | ccell_params.m_num_selector_weights = (mode == 18) ? 32 : 16; |
| 475 | ccell_params.m_pSelector_weights = (mode == 18) ? g_astc_weights5 : g_astc_weights4; |
| 476 | ccell_params.m_pSelector_weightsx = (mode == 18) ? (const bc7enc_vec4F*)g_astc_weights5x : (const bc7enc_vec4F*)g_astc_weights4x; |
| 477 | ccell_params.m_astc_endpoint_range = endpoint_range; |
| 478 | ccell_params.m_weights[0] = 1; |
| 479 | ccell_params.m_weights[1] = 1; |
| 480 | ccell_params.m_weights[2] = 1; |
| 481 | ccell_params.m_weights[3] = 1; |
| 482 | ccell_params.m_pForce_selectors = pForce_selectors; |
| 483 | |
| 484 | color_cell_compressor_results ccell_results; |
| 485 | uint8_t ccell_result_selectors[16]; |
| 486 | uint8_t ccell_result_selectors_temp[16]; |
| 487 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 488 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 489 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 490 | |
| 491 | uint64_t part_err = color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 492 | |
| 493 | // ASTC |
| 494 | astc_block_desc astc_results; |
| 495 | memset(&astc_results, 0, sizeof(astc_results)); |
| 496 | |
| 497 | astc_results.m_dual_plane = false; |
| 498 | astc_results.m_weight_range = weight_range;// (mode == 18) ? 11 : 8; |
| 499 | |
| 500 | astc_results.m_ccs = 0; |
| 501 | astc_results.m_subsets = 1; |
| 502 | astc_results.m_partition_seed = 0; |
| 503 | astc_results.m_cem = 8; |
| 504 | |
| 505 | astc_results.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 506 | astc_results.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 507 | astc_results.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[1]; |
| 508 | astc_results.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[1]; |
| 509 | astc_results.m_endpoints[4] = ccell_results.m_astc_low_endpoint.m_c[2]; |
| 510 | astc_results.m_endpoints[5] = ccell_results.m_astc_high_endpoint.m_c[2]; |
| 511 | |
| 512 | bool invert = false; |
| 513 | |
| 514 | if (pForce_selectors == nullptr) |
| 515 | { |
| 516 | int s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[4]].m_unquant; |
| 517 | int s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[5]].m_unquant; |
| 518 | if (s1 < s0) |
| 519 | { |
| 520 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 521 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 522 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 523 | invert = true; |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | for (uint32_t y = 0; y < 4; y++) |
| 528 | { |
| 529 | for (uint32_t x = 0; x < 4; x++) |
| 530 | { |
| 531 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 532 | |
| 533 | if (invert) |
| 534 | astc_results.m_weights[x + y * 4] = ((mode == 18) ? 31 : 15) - astc_results.m_weights[x + y * 4]; |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | assert(total_results < MAX_ENCODE_RESULTS); |
| 539 | if (total_results < MAX_ENCODE_RESULTS) |
| 540 | { |
| 541 | pResults[total_results].m_uastc_mode = mode; |
| 542 | pResults[total_results].m_common_pattern = 0; |
| 543 | pResults[total_results].m_astc = astc_results; |
| 544 | pResults[total_results].m_astc_err = part_err; |
| 545 | total_results++; |
| 546 | } |
| 547 | } |
| 548 | |
| 549 | // MODE 1 |
| 550 | // 1-subset, 2-bit indices, 8-bit endpoints, BC7 mode 3 |
| 551 | // DualPlane: 0, WeightRange: 2 (4), Subsets: 1, CEM: 8 (RGB Direct ), EndpointRange: 20 (256) MODE3 or MODE5 RGB |
| 552 | static void astc_mode1(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 553 | { |
| 554 | color_cell_compressor_params ccell_params; |
| 555 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 556 | |
| 557 | ccell_params.m_num_pixels = 16; |
| 558 | ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 559 | ccell_params.m_num_selector_weights = 4; |
| 560 | ccell_params.m_pSelector_weights = g_bc7_weights2; |
| 561 | ccell_params.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 562 | ccell_params.m_astc_endpoint_range = 20; |
| 563 | ccell_params.m_weights[0] = 1; |
| 564 | ccell_params.m_weights[1] = 1; |
| 565 | ccell_params.m_weights[2] = 1; |
| 566 | ccell_params.m_weights[3] = 1; |
| 567 | |
| 568 | color_cell_compressor_results ccell_results; |
| 569 | uint8_t ccell_result_selectors[16]; |
| 570 | uint8_t ccell_result_selectors_temp[16]; |
| 571 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 572 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 573 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 574 | |
| 575 | uint64_t part_err = color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 576 | |
| 577 | // ASTC |
| 578 | astc_block_desc astc_results; |
| 579 | memset(&astc_results, 0, sizeof(astc_results)); |
| 580 | |
| 581 | astc_results.m_dual_plane = false; |
| 582 | astc_results.m_weight_range = 2; |
| 583 | |
| 584 | astc_results.m_ccs = 0; |
| 585 | astc_results.m_subsets = 1; |
| 586 | astc_results.m_partition_seed = 0; |
| 587 | astc_results.m_cem = 8; |
| 588 | |
| 589 | astc_results.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 590 | astc_results.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 591 | astc_results.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[1]; |
| 592 | astc_results.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[1]; |
| 593 | astc_results.m_endpoints[4] = ccell_results.m_astc_low_endpoint.m_c[2]; |
| 594 | astc_results.m_endpoints[5] = ccell_results.m_astc_high_endpoint.m_c[2]; |
| 595 | |
| 596 | const uint32_t range = 20; |
| 597 | |
| 598 | bool invert = false; |
| 599 | |
| 600 | int s0 = g_astc_unquant[range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[4]].m_unquant; |
| 601 | int s1 = g_astc_unquant[range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[5]].m_unquant; |
| 602 | if (s1 < s0) |
| 603 | { |
| 604 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 605 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 606 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 607 | invert = true; |
| 608 | } |
| 609 | |
| 610 | for (uint32_t y = 0; y < 4; y++) |
| 611 | { |
| 612 | for (uint32_t x = 0; x < 4; x++) |
| 613 | { |
| 614 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 615 | |
| 616 | if (invert) |
| 617 | astc_results.m_weights[x + y * 4] = 3 - astc_results.m_weights[x + y * 4]; |
| 618 | } |
| 619 | } |
| 620 | |
| 621 | assert(total_results < MAX_ENCODE_RESULTS); |
| 622 | if (total_results < MAX_ENCODE_RESULTS) |
| 623 | { |
| 624 | pResults[total_results].m_uastc_mode = 1; |
| 625 | pResults[total_results].m_common_pattern = 0; |
| 626 | pResults[total_results].m_astc = astc_results; |
| 627 | pResults[total_results].m_astc_err = part_err; |
| 628 | total_results++; |
| 629 | } |
| 630 | } |
| 631 | |
| 632 | static uint32_t estimate_partition2(uint32_t num_weights, uint32_t num_comps, const uint32_t* pWeights, const color_rgba block[4][4], const uint32_t weights[4]) |
| 633 | { |
| 634 | assert(pWeights[0] == 0 && pWeights[num_weights - 1] == 64); |
| 635 | |
| 636 | uint64_t best_err = UINT64_MAX; |
| 637 | uint32_t best_common_pattern = 0; |
| 638 | |
| 639 | for (uint32_t common_pattern = 0; common_pattern < TOTAL_ASTC_BC7_COMMON_PARTITIONS2; common_pattern++) |
| 640 | { |
| 641 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions2[common_pattern].m_bc7; |
| 642 | |
| 643 | const uint8_t* pPartition = &g_bc7_partition2[bc7_pattern * 16]; |
| 644 | |
| 645 | color_quad_u8 subset_colors[2][16]; |
| 646 | uint32_t subset_total_colors[2] = { 0, 0 }; |
| 647 | for (uint32_t index = 0; index < 16; index++) |
| 648 | subset_colors[pPartition[index]][subset_total_colors[pPartition[index]]++] = ((const color_quad_u8*)block)[index]; |
| 649 | |
| 650 | uint64_t total_subset_err = 0; |
| 651 | for (uint32_t subset = 0; (subset < 2) && (total_subset_err < best_err); subset++) |
| 652 | total_subset_err += color_cell_compression_est_astc(num_weights, num_comps, pWeights, subset_total_colors[subset], &subset_colors[subset][0], best_err, weights); |
| 653 | |
| 654 | if (total_subset_err < best_err) |
| 655 | { |
| 656 | best_err = total_subset_err; |
| 657 | best_common_pattern = common_pattern; |
| 658 | } |
| 659 | } |
| 660 | |
| 661 | return best_common_pattern; |
| 662 | } |
| 663 | |
| 664 | // MODE 2 |
| 665 | // 2-subset, 3-bit indices, 4-bit endpoints, BC7 mode 1 |
| 666 | // DualPlane: 0, WeightRange: 5 (8), Subsets: 2, CEM: 8 (RGB Direct ), EndpointRange: 8 (16) MODE1 |
| 667 | static void astc_mode2(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params, bool estimate_partition) |
| 668 | { |
| 669 | uint32_t first_common_pattern = 0; |
| 670 | uint32_t last_common_pattern = TOTAL_ASTC_BC7_COMMON_PARTITIONS2; |
| 671 | |
| 672 | if (estimate_partition) |
| 673 | { |
| 674 | const uint32_t weights[4] = { 1, 1, 1, 1 }; |
| 675 | first_common_pattern = estimate_partition2(8, 3, g_bc7_weights3, block, weights); |
| 676 | last_common_pattern = first_common_pattern + 1; |
| 677 | } |
| 678 | |
| 679 | for (uint32_t common_pattern = first_common_pattern; common_pattern < last_common_pattern; common_pattern++) |
| 680 | { |
| 681 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions2[common_pattern].m_bc7; |
| 682 | |
| 683 | color_rgba part_pixels[2][16]; |
| 684 | uint32_t part_pixel_index[4][4]; |
| 685 | uint32_t num_part_pixels[2] = { 0, 0 }; |
| 686 | |
| 687 | for (uint32_t y = 0; y < 4; y++) |
| 688 | { |
| 689 | for (uint32_t x = 0; x < 4; x++) |
| 690 | { |
| 691 | const uint32_t part = g_bc7_partition2[16 * bc7_pattern + x + y * 4]; |
| 692 | part_pixel_index[y][x] = num_part_pixels[part]; |
| 693 | part_pixels[part][num_part_pixels[part]++] = block[y][x]; |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | color_cell_compressor_params ccell_params[2]; |
| 698 | color_cell_compressor_results ccell_results[2]; |
| 699 | uint8_t ccell_result_selectors[2][16]; |
| 700 | uint8_t ccell_result_selectors_temp[2][16]; |
| 701 | |
| 702 | uint64_t total_part_err = 0; |
| 703 | for (uint32_t part = 0; part < 2; part++) |
| 704 | { |
| 705 | memset(&ccell_params[part], 0, sizeof(ccell_params[part])); |
| 706 | |
| 707 | ccell_params[part].m_num_pixels = num_part_pixels[part]; |
| 708 | ccell_params[part].m_pPixels = (color_quad_u8*)&part_pixels[part][0]; |
| 709 | ccell_params[part].m_num_selector_weights = 8; |
| 710 | ccell_params[part].m_pSelector_weights = g_bc7_weights3; |
| 711 | ccell_params[part].m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights3x; |
| 712 | ccell_params[part].m_astc_endpoint_range = 8; |
| 713 | ccell_params[part].m_weights[0] = 1; |
| 714 | ccell_params[part].m_weights[1] = 1; |
| 715 | ccell_params[part].m_weights[2] = 1; |
| 716 | ccell_params[part].m_weights[3] = 1; |
| 717 | |
| 718 | memset(&ccell_results[part], 0, sizeof(ccell_results[part])); |
| 719 | ccell_results[part].m_pSelectors = &ccell_result_selectors[part][0]; |
| 720 | ccell_results[part].m_pSelectors_temp = &ccell_result_selectors_temp[part][0]; |
| 721 | |
| 722 | uint64_t part_err = color_cell_compression(255, &ccell_params[part], &ccell_results[part], &comp_params); |
| 723 | total_part_err += part_err; |
| 724 | } // part |
| 725 | |
| 726 | { |
| 727 | // ASTC |
| 728 | astc_block_desc astc_results; |
| 729 | memset(&astc_results, 0, sizeof(astc_results)); |
| 730 | |
| 731 | astc_results.m_dual_plane = false; |
| 732 | astc_results.m_weight_range = 5; |
| 733 | |
| 734 | astc_results.m_ccs = 0; |
| 735 | astc_results.m_subsets = 2; |
| 736 | astc_results.m_partition_seed = g_astc_bc7_common_partitions2[common_pattern].m_astc; |
| 737 | astc_results.m_cem = 8; |
| 738 | |
| 739 | uint32_t p0 = 0; |
| 740 | uint32_t p1 = 1; |
| 741 | if (g_astc_bc7_common_partitions2[common_pattern].m_invert) |
| 742 | std::swap(p0, p1); |
| 743 | |
| 744 | astc_results.m_endpoints[0] = ccell_results[p0].m_astc_low_endpoint.m_c[0]; |
| 745 | astc_results.m_endpoints[1] = ccell_results[p0].m_astc_high_endpoint.m_c[0]; |
| 746 | astc_results.m_endpoints[2] = ccell_results[p0].m_astc_low_endpoint.m_c[1]; |
| 747 | astc_results.m_endpoints[3] = ccell_results[p0].m_astc_high_endpoint.m_c[1]; |
| 748 | astc_results.m_endpoints[4] = ccell_results[p0].m_astc_low_endpoint.m_c[2]; |
| 749 | astc_results.m_endpoints[5] = ccell_results[p0].m_astc_high_endpoint.m_c[2]; |
| 750 | |
| 751 | const uint32_t range = 8; |
| 752 | |
| 753 | bool invert[2] = { false, false }; |
| 754 | |
| 755 | int s0 = g_astc_unquant[range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[4]].m_unquant; |
| 756 | int s1 = g_astc_unquant[range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[5]].m_unquant; |
| 757 | if (s1 < s0) |
| 758 | { |
| 759 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 760 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 761 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 762 | invert[0] = true; |
| 763 | } |
| 764 | |
| 765 | astc_results.m_endpoints[6] = ccell_results[p1].m_astc_low_endpoint.m_c[0]; |
| 766 | astc_results.m_endpoints[7] = ccell_results[p1].m_astc_high_endpoint.m_c[0]; |
| 767 | astc_results.m_endpoints[8] = ccell_results[p1].m_astc_low_endpoint.m_c[1]; |
| 768 | astc_results.m_endpoints[9] = ccell_results[p1].m_astc_high_endpoint.m_c[1]; |
| 769 | astc_results.m_endpoints[10] = ccell_results[p1].m_astc_low_endpoint.m_c[2]; |
| 770 | astc_results.m_endpoints[11] = ccell_results[p1].m_astc_high_endpoint.m_c[2]; |
| 771 | |
| 772 | s0 = g_astc_unquant[range][astc_results.m_endpoints[0 + 6]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[2 + 6]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[4 + 6]].m_unquant; |
| 773 | s1 = g_astc_unquant[range][astc_results.m_endpoints[1 + 6]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[3 + 6]].m_unquant + g_astc_unquant[range][astc_results.m_endpoints[5 + 6]].m_unquant; |
| 774 | |
| 775 | if (s1 < s0) |
| 776 | { |
| 777 | std::swap(astc_results.m_endpoints[0 + 6], astc_results.m_endpoints[1 + 6]); |
| 778 | std::swap(astc_results.m_endpoints[2 + 6], astc_results.m_endpoints[3 + 6]); |
| 779 | std::swap(astc_results.m_endpoints[4 + 6], astc_results.m_endpoints[5 + 6]); |
| 780 | invert[1] = true; |
| 781 | } |
| 782 | |
| 783 | for (uint32_t y = 0; y < 4; y++) |
| 784 | { |
| 785 | for (uint32_t x = 0; x < 4; x++) |
| 786 | { |
| 787 | const uint32_t bc7_part = g_bc7_partition2[16 * bc7_pattern + x + y * 4]; |
| 788 | |
| 789 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[bc7_part][part_pixel_index[y][x]]; |
| 790 | |
| 791 | uint32_t astc_part = bc7_part; |
| 792 | if (g_astc_bc7_common_partitions2[common_pattern].m_invert) |
| 793 | astc_part = 1 - astc_part; |
| 794 | |
| 795 | if (invert[astc_part]) |
| 796 | astc_results.m_weights[x + y * 4] = 7 - astc_results.m_weights[x + y * 4]; |
| 797 | } |
| 798 | } |
| 799 | |
| 800 | assert(total_results < MAX_ENCODE_RESULTS); |
| 801 | if (total_results < MAX_ENCODE_RESULTS) |
| 802 | { |
| 803 | pResults[total_results].m_uastc_mode = 2; |
| 804 | pResults[total_results].m_common_pattern = common_pattern; |
| 805 | pResults[total_results].m_astc = astc_results; |
| 806 | pResults[total_results].m_astc_err = total_part_err; |
| 807 | total_results++; |
| 808 | } |
| 809 | } |
| 810 | |
| 811 | } // common_pattern |
| 812 | } |
| 813 | |
| 814 | // MODE 3 |
| 815 | // 3-subsets, 2-bit indices, [0,11] endpoints, BC7 mode 2 |
| 816 | // DualPlane: 0, WeightRange: 2 (4), Subsets: 3, CEM: 8 (RGB Direct ), EndpointRange: 7 (12) MODE2 |
| 817 | static void astc_mode3(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params, bool estimate_partition) |
| 818 | { |
| 819 | uint32_t first_common_pattern = 0; |
| 820 | uint32_t last_common_pattern = TOTAL_ASTC_BC7_COMMON_PARTITIONS3; |
| 821 | |
| 822 | if (estimate_partition) |
| 823 | { |
| 824 | uint64_t best_err = UINT64_MAX; |
| 825 | uint32_t best_common_pattern = 0; |
| 826 | const uint32_t weights[4] = { 1, 1, 1, 1 }; |
| 827 | |
| 828 | for (uint32_t common_pattern = 0; common_pattern < TOTAL_ASTC_BC7_COMMON_PARTITIONS3; common_pattern++) |
| 829 | { |
| 830 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions3[common_pattern].m_bc7; |
| 831 | |
| 832 | const uint8_t* pPartition = &g_bc7_partition3[bc7_pattern * 16]; |
| 833 | |
| 834 | color_quad_u8 subset_colors[3][16]; |
| 835 | uint32_t subset_total_colors[3] = { 0, 0 }; |
| 836 | for (uint32_t index = 0; index < 16; index++) |
| 837 | subset_colors[pPartition[index]][subset_total_colors[pPartition[index]]++] = ((const color_quad_u8*)block)[index]; |
| 838 | |
| 839 | uint64_t total_subset_err = 0; |
| 840 | for (uint32_t subset = 0; (subset < 3) && (total_subset_err < best_err); subset++) |
| 841 | total_subset_err += color_cell_compression_est_astc(4, 3, g_bc7_weights2, subset_total_colors[subset], &subset_colors[subset][0], best_err, weights); |
| 842 | |
| 843 | if (total_subset_err < best_err) |
| 844 | { |
| 845 | best_err = total_subset_err; |
| 846 | best_common_pattern = common_pattern; |
| 847 | } |
| 848 | } |
| 849 | |
| 850 | first_common_pattern = best_common_pattern; |
| 851 | last_common_pattern = best_common_pattern + 1; |
| 852 | } |
| 853 | |
| 854 | for (uint32_t common_pattern = first_common_pattern; common_pattern < last_common_pattern; common_pattern++) |
| 855 | { |
| 856 | const uint32_t endpoint_range = 7; |
| 857 | |
| 858 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions3[common_pattern].m_bc7; |
| 859 | |
| 860 | color_rgba part_pixels[3][16]; |
| 861 | uint32_t part_pixel_index[4][4]; |
| 862 | uint32_t num_part_pixels[3] = { 0, 0, 0 }; |
| 863 | |
| 864 | for (uint32_t y = 0; y < 4; y++) |
| 865 | { |
| 866 | for (uint32_t x = 0; x < 4; x++) |
| 867 | { |
| 868 | const uint32_t bc7_part = g_bc7_partition3[16 * bc7_pattern + x + y * 4]; |
| 869 | part_pixel_index[y][x] = num_part_pixels[bc7_part]; |
| 870 | part_pixels[bc7_part][num_part_pixels[bc7_part]++] = block[y][x]; |
| 871 | } |
| 872 | } |
| 873 | |
| 874 | color_cell_compressor_params ccell_params[3]; |
| 875 | color_cell_compressor_results ccell_results[3]; |
| 876 | uint8_t ccell_result_selectors[3][16]; |
| 877 | uint8_t ccell_result_selectors_temp[3][16]; |
| 878 | |
| 879 | uint64_t total_part_err = 0; |
| 880 | for (uint32_t bc7_part = 0; bc7_part < 3; bc7_part++) |
| 881 | { |
| 882 | memset(&ccell_params[bc7_part], 0, sizeof(ccell_params[bc7_part])); |
| 883 | |
| 884 | ccell_params[bc7_part].m_num_pixels = num_part_pixels[bc7_part]; |
| 885 | ccell_params[bc7_part].m_pPixels = (color_quad_u8*)&part_pixels[bc7_part][0]; |
| 886 | ccell_params[bc7_part].m_num_selector_weights = 4; |
| 887 | ccell_params[bc7_part].m_pSelector_weights = g_bc7_weights2; |
| 888 | ccell_params[bc7_part].m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 889 | ccell_params[bc7_part].m_astc_endpoint_range = endpoint_range; |
| 890 | ccell_params[bc7_part].m_weights[0] = 1; |
| 891 | ccell_params[bc7_part].m_weights[1] = 1; |
| 892 | ccell_params[bc7_part].m_weights[2] = 1; |
| 893 | ccell_params[bc7_part].m_weights[3] = 1; |
| 894 | |
| 895 | memset(&ccell_results[bc7_part], 0, sizeof(ccell_results[bc7_part])); |
| 896 | ccell_results[bc7_part].m_pSelectors = &ccell_result_selectors[bc7_part][0]; |
| 897 | ccell_results[bc7_part].m_pSelectors_temp = &ccell_result_selectors_temp[bc7_part][0]; |
| 898 | |
| 899 | uint64_t part_err = color_cell_compression(255, &ccell_params[bc7_part], &ccell_results[bc7_part], &comp_params); |
| 900 | total_part_err += part_err; |
| 901 | } // part |
| 902 | |
| 903 | { |
| 904 | // ASTC |
| 905 | astc_block_desc astc_results; |
| 906 | memset(&astc_results, 0, sizeof(astc_results)); |
| 907 | |
| 908 | astc_results.m_dual_plane = false; |
| 909 | astc_results.m_weight_range = 2; |
| 910 | |
| 911 | astc_results.m_ccs = 0; |
| 912 | astc_results.m_subsets = 3; |
| 913 | astc_results.m_partition_seed = g_astc_bc7_common_partitions3[common_pattern].m_astc; |
| 914 | astc_results.m_cem = 8; |
| 915 | |
| 916 | uint32_t astc_to_bc7_part[3]; // converts ASTC to BC7 partition index |
| 917 | const uint32_t perm = g_astc_bc7_common_partitions3[common_pattern].m_astc_to_bc7_perm; |
| 918 | astc_to_bc7_part[0] = g_astc_to_bc7_partition_index_perm_tables[perm][0]; |
| 919 | astc_to_bc7_part[1] = g_astc_to_bc7_partition_index_perm_tables[perm][1]; |
| 920 | astc_to_bc7_part[2] = g_astc_to_bc7_partition_index_perm_tables[perm][2]; |
| 921 | |
| 922 | bool invert_astc_part[3] = { false, false, false }; |
| 923 | |
| 924 | for (uint32_t astc_part = 0; astc_part < 3; astc_part++) |
| 925 | { |
| 926 | uint8_t* pEndpoints = &astc_results.m_endpoints[6 * astc_part]; |
| 927 | |
| 928 | pEndpoints[0] = ccell_results[astc_to_bc7_part[astc_part]].m_astc_low_endpoint.m_c[0]; |
| 929 | pEndpoints[1] = ccell_results[astc_to_bc7_part[astc_part]].m_astc_high_endpoint.m_c[0]; |
| 930 | pEndpoints[2] = ccell_results[astc_to_bc7_part[astc_part]].m_astc_low_endpoint.m_c[1]; |
| 931 | pEndpoints[3] = ccell_results[astc_to_bc7_part[astc_part]].m_astc_high_endpoint.m_c[1]; |
| 932 | pEndpoints[4] = ccell_results[astc_to_bc7_part[astc_part]].m_astc_low_endpoint.m_c[2]; |
| 933 | pEndpoints[5] = ccell_results[astc_to_bc7_part[astc_part]].m_astc_high_endpoint.m_c[2]; |
| 934 | |
| 935 | int s0 = g_astc_unquant[endpoint_range][pEndpoints[0]].m_unquant + g_astc_unquant[endpoint_range][pEndpoints[2]].m_unquant + g_astc_unquant[endpoint_range][pEndpoints[4]].m_unquant; |
| 936 | int s1 = g_astc_unquant[endpoint_range][pEndpoints[1]].m_unquant + g_astc_unquant[endpoint_range][pEndpoints[3]].m_unquant + g_astc_unquant[endpoint_range][pEndpoints[5]].m_unquant; |
| 937 | if (s1 < s0) |
| 938 | { |
| 939 | std::swap(pEndpoints[0], pEndpoints[1]); |
| 940 | std::swap(pEndpoints[2], pEndpoints[3]); |
| 941 | std::swap(pEndpoints[4], pEndpoints[5]); |
| 942 | invert_astc_part[astc_part] = true; |
| 943 | } |
| 944 | } |
| 945 | |
| 946 | for (uint32_t y = 0; y < 4; y++) |
| 947 | { |
| 948 | for (uint32_t x = 0; x < 4; x++) |
| 949 | { |
| 950 | const uint32_t bc7_part = g_bc7_partition3[16 * bc7_pattern + x + y * 4]; |
| 951 | |
| 952 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[bc7_part][part_pixel_index[y][x]]; |
| 953 | |
| 954 | uint32_t astc_part = 0; |
| 955 | for (uint32_t i = 0; i < 3; i++) |
| 956 | { |
| 957 | if (astc_to_bc7_part[i] == bc7_part) |
| 958 | { |
| 959 | astc_part = i; |
| 960 | break; |
| 961 | } |
| 962 | } |
| 963 | |
| 964 | if (invert_astc_part[astc_part]) |
| 965 | astc_results.m_weights[x + y * 4] = 3 - astc_results.m_weights[x + y * 4]; |
| 966 | } |
| 967 | } |
| 968 | |
| 969 | assert(total_results < MAX_ENCODE_RESULTS); |
| 970 | if (total_results < MAX_ENCODE_RESULTS) |
| 971 | { |
| 972 | pResults[total_results].m_uastc_mode = 3; |
| 973 | pResults[total_results].m_common_pattern = common_pattern; |
| 974 | pResults[total_results].m_astc = astc_results; |
| 975 | pResults[total_results].m_astc_err = total_part_err; |
| 976 | total_results++; |
| 977 | } |
| 978 | |
| 979 | } |
| 980 | |
| 981 | } // common_pattern |
| 982 | } |
| 983 | |
| 984 | // MODE 4 |
| 985 | // DualPlane: 0, WeightRange: 2 (4), Subsets: 2, CEM: 8 (RGB Direct ), EndpointRange: 12 (40) MODE3 |
| 986 | static void astc_mode4(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params, bool estimate_partition) |
| 987 | { |
| 988 | //const uint32_t weight_range = 2; |
| 989 | const uint32_t endpoint_range = 12; |
| 990 | |
| 991 | uint32_t first_common_pattern = 0; |
| 992 | uint32_t last_common_pattern = TOTAL_ASTC_BC7_COMMON_PARTITIONS2; |
| 993 | |
| 994 | if (estimate_partition) |
| 995 | { |
| 996 | const uint32_t weights[4] = { 1, 1, 1, 1 }; |
| 997 | first_common_pattern = estimate_partition2(4, 3, g_bc7_weights2, block, weights); |
| 998 | last_common_pattern = first_common_pattern + 1; |
| 999 | } |
| 1000 | |
| 1001 | for (uint32_t common_pattern = first_common_pattern; common_pattern < last_common_pattern; common_pattern++) |
| 1002 | { |
| 1003 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions2[common_pattern].m_bc7; |
| 1004 | |
| 1005 | color_rgba part_pixels[2][16]; |
| 1006 | uint32_t part_pixel_index[4][4]; |
| 1007 | uint32_t num_part_pixels[2] = { 0, 0 }; |
| 1008 | |
| 1009 | for (uint32_t y = 0; y < 4; y++) |
| 1010 | { |
| 1011 | for (uint32_t x = 0; x < 4; x++) |
| 1012 | { |
| 1013 | const uint32_t part = g_bc7_partition2[16 * bc7_pattern + x + y * 4]; |
| 1014 | part_pixel_index[y][x] = num_part_pixels[part]; |
| 1015 | part_pixels[part][num_part_pixels[part]++] = block[y][x]; |
| 1016 | } |
| 1017 | } |
| 1018 | |
| 1019 | color_cell_compressor_params ccell_params[2]; |
| 1020 | color_cell_compressor_results ccell_results[2]; |
| 1021 | uint8_t ccell_result_selectors[2][16]; |
| 1022 | uint8_t ccell_result_selectors_temp[2][16]; |
| 1023 | |
| 1024 | uint64_t total_part_err = 0; |
| 1025 | for (uint32_t part = 0; part < 2; part++) |
| 1026 | { |
| 1027 | memset(&ccell_params[part], 0, sizeof(ccell_params[part])); |
| 1028 | |
| 1029 | ccell_params[part].m_num_pixels = num_part_pixels[part]; |
| 1030 | ccell_params[part].m_pPixels = (color_quad_u8*)&part_pixels[part][0]; |
| 1031 | ccell_params[part].m_num_selector_weights = 4; |
| 1032 | ccell_params[part].m_pSelector_weights = g_bc7_weights2; |
| 1033 | ccell_params[part].m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1034 | ccell_params[part].m_astc_endpoint_range = endpoint_range; |
| 1035 | ccell_params[part].m_weights[0] = 1; |
| 1036 | ccell_params[part].m_weights[1] = 1; |
| 1037 | ccell_params[part].m_weights[2] = 1; |
| 1038 | ccell_params[part].m_weights[3] = 1; |
| 1039 | |
| 1040 | memset(&ccell_results[part], 0, sizeof(ccell_results[part])); |
| 1041 | ccell_results[part].m_pSelectors = &ccell_result_selectors[part][0]; |
| 1042 | ccell_results[part].m_pSelectors_temp = &ccell_result_selectors_temp[part][0]; |
| 1043 | |
| 1044 | uint64_t part_err = color_cell_compression(255, &ccell_params[part], &ccell_results[part], &comp_params); |
| 1045 | total_part_err += part_err; |
| 1046 | } // part |
| 1047 | |
| 1048 | // ASTC |
| 1049 | astc_block_desc astc_results; |
| 1050 | memset(&astc_results, 0, sizeof(astc_results)); |
| 1051 | |
| 1052 | astc_results.m_dual_plane = false; |
| 1053 | astc_results.m_weight_range = 2; |
| 1054 | |
| 1055 | astc_results.m_ccs = 0; |
| 1056 | astc_results.m_subsets = 2; |
| 1057 | astc_results.m_partition_seed = g_astc_bc7_common_partitions2[common_pattern].m_astc; |
| 1058 | astc_results.m_cem = 8; |
| 1059 | |
| 1060 | uint32_t p0 = 0; |
| 1061 | uint32_t p1 = 1; |
| 1062 | if (g_astc_bc7_common_partitions2[common_pattern].m_invert) |
| 1063 | std::swap(p0, p1); |
| 1064 | |
| 1065 | astc_results.m_endpoints[0] = ccell_results[p0].m_astc_low_endpoint.m_c[0]; |
| 1066 | astc_results.m_endpoints[1] = ccell_results[p0].m_astc_high_endpoint.m_c[0]; |
| 1067 | astc_results.m_endpoints[2] = ccell_results[p0].m_astc_low_endpoint.m_c[1]; |
| 1068 | astc_results.m_endpoints[3] = ccell_results[p0].m_astc_high_endpoint.m_c[1]; |
| 1069 | astc_results.m_endpoints[4] = ccell_results[p0].m_astc_low_endpoint.m_c[2]; |
| 1070 | astc_results.m_endpoints[5] = ccell_results[p0].m_astc_high_endpoint.m_c[2]; |
| 1071 | |
| 1072 | bool invert[2] = { false, false }; |
| 1073 | |
| 1074 | int s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[4]].m_unquant; |
| 1075 | int s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[5]].m_unquant; |
| 1076 | if (s1 < s0) |
| 1077 | { |
| 1078 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 1079 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 1080 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 1081 | invert[0] = true; |
| 1082 | } |
| 1083 | |
| 1084 | astc_results.m_endpoints[6] = ccell_results[p1].m_astc_low_endpoint.m_c[0]; |
| 1085 | astc_results.m_endpoints[7] = ccell_results[p1].m_astc_high_endpoint.m_c[0]; |
| 1086 | astc_results.m_endpoints[8] = ccell_results[p1].m_astc_low_endpoint.m_c[1]; |
| 1087 | astc_results.m_endpoints[9] = ccell_results[p1].m_astc_high_endpoint.m_c[1]; |
| 1088 | astc_results.m_endpoints[10] = ccell_results[p1].m_astc_low_endpoint.m_c[2]; |
| 1089 | astc_results.m_endpoints[11] = ccell_results[p1].m_astc_high_endpoint.m_c[2]; |
| 1090 | |
| 1091 | s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[0 + 6]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[2 + 6]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[4 + 6]].m_unquant; |
| 1092 | s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[1 + 6]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[3 + 6]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[5 + 6]].m_unquant; |
| 1093 | |
| 1094 | if (s1 < s0) |
| 1095 | { |
| 1096 | std::swap(astc_results.m_endpoints[0 + 6], astc_results.m_endpoints[1 + 6]); |
| 1097 | std::swap(astc_results.m_endpoints[2 + 6], astc_results.m_endpoints[3 + 6]); |
| 1098 | std::swap(astc_results.m_endpoints[4 + 6], astc_results.m_endpoints[5 + 6]); |
| 1099 | invert[1] = true; |
| 1100 | } |
| 1101 | |
| 1102 | for (uint32_t y = 0; y < 4; y++) |
| 1103 | { |
| 1104 | for (uint32_t x = 0; x < 4; x++) |
| 1105 | { |
| 1106 | const uint32_t bc7_part = g_bc7_partition2[16 * bc7_pattern + x + y * 4]; |
| 1107 | |
| 1108 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[bc7_part][part_pixel_index[y][x]]; |
| 1109 | |
| 1110 | uint32_t astc_part = bc7_part; |
| 1111 | if (g_astc_bc7_common_partitions2[common_pattern].m_invert) |
| 1112 | astc_part = 1 - astc_part; |
| 1113 | |
| 1114 | if (invert[astc_part]) |
| 1115 | astc_results.m_weights[x + y * 4] = 3 - astc_results.m_weights[x + y * 4]; |
| 1116 | } |
| 1117 | } |
| 1118 | |
| 1119 | assert(total_results < MAX_ENCODE_RESULTS); |
| 1120 | if (total_results < MAX_ENCODE_RESULTS) |
| 1121 | { |
| 1122 | pResults[total_results].m_uastc_mode = 4; |
| 1123 | pResults[total_results].m_common_pattern = common_pattern; |
| 1124 | pResults[total_results].m_astc = astc_results; |
| 1125 | pResults[total_results].m_astc_err = total_part_err; |
| 1126 | total_results++; |
| 1127 | } |
| 1128 | |
| 1129 | } // common_pattern |
| 1130 | } |
| 1131 | |
| 1132 | // MODE 5 |
| 1133 | // DualPlane: 0, WeightRange: 5 (8), Subsets: 1, CEM: 8 (RGB Direct ), EndpointRange: 20 (256) BC7 MODE 6 (or MODE 1 1-subset) |
| 1134 | static void astc_mode5(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 1135 | { |
| 1136 | const uint32_t weight_range = 5; |
| 1137 | const uint32_t endpoint_range = 20; |
| 1138 | |
| 1139 | color_cell_compressor_params ccell_params; |
| 1140 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 1141 | |
| 1142 | ccell_params.m_num_pixels = 16; |
| 1143 | ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 1144 | ccell_params.m_num_selector_weights = 8; |
| 1145 | ccell_params.m_pSelector_weights = g_bc7_weights3; |
| 1146 | ccell_params.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights3x; |
| 1147 | ccell_params.m_astc_endpoint_range = endpoint_range; |
| 1148 | ccell_params.m_weights[0] = 1; |
| 1149 | ccell_params.m_weights[1] = 1; |
| 1150 | ccell_params.m_weights[2] = 1; |
| 1151 | ccell_params.m_weights[3] = 1; |
| 1152 | |
| 1153 | color_cell_compressor_results ccell_results; |
| 1154 | uint8_t ccell_result_selectors[16]; |
| 1155 | uint8_t ccell_result_selectors_temp[16]; |
| 1156 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 1157 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 1158 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 1159 | |
| 1160 | uint64_t part_err = color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 1161 | |
| 1162 | // ASTC |
| 1163 | astc_block_desc blk; |
| 1164 | memset(&blk, 0, sizeof(blk)); |
| 1165 | |
| 1166 | blk.m_dual_plane = false; |
| 1167 | blk.m_weight_range = weight_range; |
| 1168 | |
| 1169 | blk.m_ccs = 0; |
| 1170 | blk.m_subsets = 1; |
| 1171 | blk.m_partition_seed = 0; |
| 1172 | blk.m_cem = 8; |
| 1173 | |
| 1174 | blk.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 1175 | blk.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 1176 | blk.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[1]; |
| 1177 | blk.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[1]; |
| 1178 | blk.m_endpoints[4] = ccell_results.m_astc_low_endpoint.m_c[2]; |
| 1179 | blk.m_endpoints[5] = ccell_results.m_astc_high_endpoint.m_c[2]; |
| 1180 | |
| 1181 | bool invert = false; |
| 1182 | |
| 1183 | int s0 = g_astc_unquant[endpoint_range][blk.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[4]].m_unquant; |
| 1184 | int s1 = g_astc_unquant[endpoint_range][blk.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[5]].m_unquant; |
| 1185 | if (s1 < s0) |
| 1186 | { |
| 1187 | std::swap(blk.m_endpoints[0], blk.m_endpoints[1]); |
| 1188 | std::swap(blk.m_endpoints[2], blk.m_endpoints[3]); |
| 1189 | std::swap(blk.m_endpoints[4], blk.m_endpoints[5]); |
| 1190 | invert = true; |
| 1191 | } |
| 1192 | |
| 1193 | for (uint32_t y = 0; y < 4; y++) |
| 1194 | { |
| 1195 | for (uint32_t x = 0; x < 4; x++) |
| 1196 | { |
| 1197 | blk.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 1198 | |
| 1199 | if (invert) |
| 1200 | blk.m_weights[x + y * 4] = 7 - blk.m_weights[x + y * 4]; |
| 1201 | } |
| 1202 | } |
| 1203 | |
| 1204 | assert(total_results < MAX_ENCODE_RESULTS); |
| 1205 | if (total_results < MAX_ENCODE_RESULTS) |
| 1206 | { |
| 1207 | pResults[total_results].m_uastc_mode = 5; |
| 1208 | pResults[total_results].m_common_pattern = 0; |
| 1209 | pResults[total_results].m_astc = blk; |
| 1210 | pResults[total_results].m_astc_err = part_err; |
| 1211 | total_results++; |
| 1212 | } |
| 1213 | } |
| 1214 | |
| 1215 | // MODE 6 |
| 1216 | // DualPlane: 1, WeightRange: 2 (4), Subsets: 1, CEM: 8 (RGB Direct ), EndpointRange: 18 (160) BC7 MODE5 |
| 1217 | static void astc_mode6(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 1218 | { |
| 1219 | for (uint32_t rot_comp = 0; rot_comp < 3; rot_comp++) |
| 1220 | { |
| 1221 | const uint32_t weight_range = 2; |
| 1222 | const uint32_t endpoint_range = 18; |
| 1223 | |
| 1224 | color_quad_u8 block_rgb[16]; |
| 1225 | color_quad_u8 block_a[16]; |
| 1226 | for (uint32_t i = 0; i < 16; i++) |
| 1227 | { |
| 1228 | block_rgb[i] = ((color_quad_u8*)&block[0][0])[i]; |
| 1229 | block_a[i] = block_rgb[i]; |
| 1230 | |
| 1231 | uint8_t c = block_a[i].m_c[rot_comp]; |
| 1232 | block_a[i].m_c[0] = c; |
| 1233 | block_a[i].m_c[1] = c; |
| 1234 | block_a[i].m_c[2] = c; |
| 1235 | block_a[i].m_c[3] = 255; |
| 1236 | |
| 1237 | block_rgb[i].m_c[rot_comp] = 255; |
| 1238 | } |
| 1239 | |
| 1240 | uint8_t ccell_result_selectors_temp[16]; |
| 1241 | |
| 1242 | color_cell_compressor_params ccell_params_rgb; |
| 1243 | memset(&ccell_params_rgb, 0, sizeof(ccell_params_rgb)); |
| 1244 | |
| 1245 | ccell_params_rgb.m_num_pixels = 16; |
| 1246 | ccell_params_rgb.m_pPixels = block_rgb; |
| 1247 | ccell_params_rgb.m_num_selector_weights = 4; |
| 1248 | ccell_params_rgb.m_pSelector_weights = g_bc7_weights2; |
| 1249 | ccell_params_rgb.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1250 | ccell_params_rgb.m_astc_endpoint_range = endpoint_range; |
| 1251 | ccell_params_rgb.m_weights[0] = 1; |
| 1252 | ccell_params_rgb.m_weights[1] = 1; |
| 1253 | ccell_params_rgb.m_weights[2] = 1; |
| 1254 | ccell_params_rgb.m_weights[3] = 1; |
| 1255 | |
| 1256 | color_cell_compressor_results ccell_results_rgb; |
| 1257 | uint8_t ccell_result_selectors_rgb[16]; |
| 1258 | memset(&ccell_results_rgb, 0, sizeof(ccell_results_rgb)); |
| 1259 | ccell_results_rgb.m_pSelectors = &ccell_result_selectors_rgb[0]; |
| 1260 | ccell_results_rgb.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 1261 | |
| 1262 | uint64_t part_err_rgb = color_cell_compression(255, &ccell_params_rgb, &ccell_results_rgb, &comp_params); |
| 1263 | |
| 1264 | color_cell_compressor_params ccell_params_a; |
| 1265 | memset(&ccell_params_a, 0, sizeof(ccell_params_a)); |
| 1266 | |
| 1267 | ccell_params_a.m_num_pixels = 16; |
| 1268 | ccell_params_a.m_pPixels = block_a; |
| 1269 | ccell_params_a.m_num_selector_weights = 4; |
| 1270 | ccell_params_a.m_pSelector_weights = g_bc7_weights2; |
| 1271 | ccell_params_a.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1272 | ccell_params_a.m_astc_endpoint_range = endpoint_range; |
| 1273 | ccell_params_a.m_weights[0] = 1; |
| 1274 | ccell_params_a.m_weights[1] = 1; |
| 1275 | ccell_params_a.m_weights[2] = 1; |
| 1276 | ccell_params_a.m_weights[3] = 1; |
| 1277 | |
| 1278 | color_cell_compressor_results ccell_results_a; |
| 1279 | uint8_t ccell_result_selectors_a[16]; |
| 1280 | memset(&ccell_results_a, 0, sizeof(ccell_results_a)); |
| 1281 | ccell_results_a.m_pSelectors = &ccell_result_selectors_a[0]; |
| 1282 | ccell_results_a.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 1283 | |
| 1284 | uint64_t part_err_a = color_cell_compression(255, &ccell_params_a, &ccell_results_a, &comp_params) / 3; |
| 1285 | |
| 1286 | uint64_t total_err = part_err_rgb + part_err_a; |
| 1287 | |
| 1288 | // ASTC |
| 1289 | astc_block_desc blk; |
| 1290 | memset(&blk, 0, sizeof(blk)); |
| 1291 | |
| 1292 | blk.m_dual_plane = true; |
| 1293 | blk.m_weight_range = weight_range; |
| 1294 | |
| 1295 | blk.m_ccs = rot_comp; |
| 1296 | blk.m_subsets = 1; |
| 1297 | blk.m_partition_seed = 0; |
| 1298 | blk.m_cem = 8; |
| 1299 | |
| 1300 | blk.m_endpoints[0] = (rot_comp == 0 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[0]; |
| 1301 | blk.m_endpoints[1] = (rot_comp == 0 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[0]; |
| 1302 | blk.m_endpoints[2] = (rot_comp == 1 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[1]; |
| 1303 | blk.m_endpoints[3] = (rot_comp == 1 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[1]; |
| 1304 | blk.m_endpoints[4] = (rot_comp == 2 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[2]; |
| 1305 | blk.m_endpoints[5] = (rot_comp == 2 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[2]; |
| 1306 | |
| 1307 | bool invert = false; |
| 1308 | |
| 1309 | int s0 = g_astc_unquant[endpoint_range][blk.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[4]].m_unquant; |
| 1310 | int s1 = g_astc_unquant[endpoint_range][blk.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[5]].m_unquant; |
| 1311 | if (s1 < s0) |
| 1312 | { |
| 1313 | std::swap(blk.m_endpoints[0], blk.m_endpoints[1]); |
| 1314 | std::swap(blk.m_endpoints[2], blk.m_endpoints[3]); |
| 1315 | std::swap(blk.m_endpoints[4], blk.m_endpoints[5]); |
| 1316 | invert = true; |
| 1317 | } |
| 1318 | |
| 1319 | for (uint32_t y = 0; y < 4; y++) |
| 1320 | { |
| 1321 | for (uint32_t x = 0; x < 4; x++) |
| 1322 | { |
| 1323 | uint32_t rgb_index = ccell_result_selectors_rgb[x + y * 4]; |
| 1324 | uint32_t a_index = ccell_result_selectors_a[x + y * 4]; |
| 1325 | |
| 1326 | if (invert) |
| 1327 | { |
| 1328 | rgb_index = 3 - rgb_index; |
| 1329 | a_index = 3 - a_index; |
| 1330 | } |
| 1331 | |
| 1332 | blk.m_weights[(x + y * 4) * 2 + 0] = (uint8_t)rgb_index; |
| 1333 | blk.m_weights[(x + y * 4) * 2 + 1] = (uint8_t)a_index; |
| 1334 | } |
| 1335 | } |
| 1336 | |
| 1337 | assert(total_results < MAX_ENCODE_RESULTS); |
| 1338 | if (total_results < MAX_ENCODE_RESULTS) |
| 1339 | { |
| 1340 | pResults[total_results].m_uastc_mode = 6; |
| 1341 | pResults[total_results].m_common_pattern = 0; |
| 1342 | pResults[total_results].m_astc = blk; |
| 1343 | pResults[total_results].m_astc_err = total_err; |
| 1344 | total_results++; |
| 1345 | } |
| 1346 | } // rot_comp |
| 1347 | } |
| 1348 | |
| 1349 | // MODE 7 - 2 subset ASTC, 3 subset BC7 |
| 1350 | // DualPlane: 0, WeightRange: 2 (4), Subsets: 2, CEM: 8 (RGB Direct ), EndpointRange: 12 (40) MODE2 |
| 1351 | static void astc_mode7(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params, bool estimate_partition) |
| 1352 | { |
| 1353 | uint32_t first_common_pattern = 0; |
| 1354 | uint32_t last_common_pattern = TOTAL_BC7_3_ASTC2_COMMON_PARTITIONS; |
| 1355 | |
| 1356 | if (estimate_partition) |
| 1357 | { |
| 1358 | uint64_t best_err = UINT64_MAX; |
| 1359 | uint32_t best_common_pattern = 0; |
| 1360 | const uint32_t weights[4] = { 1, 1, 1, 1 }; |
| 1361 | |
| 1362 | for (uint32_t common_pattern = 0; common_pattern < TOTAL_BC7_3_ASTC2_COMMON_PARTITIONS; common_pattern++) |
| 1363 | { |
| 1364 | const uint8_t* pPartition = &g_bc7_3_astc2_patterns2[common_pattern][0]; |
| 1365 | |
| 1366 | #ifdef _DEBUG |
| 1367 | const uint32_t astc_pattern = g_bc7_3_astc2_common_partitions[common_pattern].m_astc2; |
| 1368 | const uint32_t bc7_pattern = g_bc7_3_astc2_common_partitions[common_pattern].m_bc73; |
| 1369 | const uint32_t common_pattern_k = g_bc7_3_astc2_common_partitions[common_pattern].k; |
| 1370 | |
| 1371 | for (uint32_t y = 0; y < 4; y++) |
| 1372 | { |
| 1373 | for (uint32_t x = 0; x < 4; x++) |
| 1374 | { |
| 1375 | const uint32_t astc_part = bc7_convert_partition_index_3_to_2(g_bc7_partition3[16 * bc7_pattern + x + y * 4], common_pattern_k); |
| 1376 | assert((int)astc_part == astc_compute_texel_partition(astc_pattern, x, y, 0, 2, true)); |
| 1377 | assert(astc_part == pPartition[x + y * 4]); |
| 1378 | } |
| 1379 | } |
| 1380 | #endif |
| 1381 | |
| 1382 | color_quad_u8 subset_colors[2][16]; |
| 1383 | uint32_t subset_total_colors[2] = { 0, 0 }; |
| 1384 | for (uint32_t index = 0; index < 16; index++) |
| 1385 | subset_colors[pPartition[index]][subset_total_colors[pPartition[index]]++] = ((const color_quad_u8*)block)[index]; |
| 1386 | |
| 1387 | uint64_t total_subset_err = 0; |
| 1388 | for (uint32_t subset = 0; (subset < 2) && (total_subset_err < best_err); subset++) |
| 1389 | total_subset_err += color_cell_compression_est_astc(4, 3, g_bc7_weights2, subset_total_colors[subset], &subset_colors[subset][0], best_err, weights); |
| 1390 | |
| 1391 | if (total_subset_err < best_err) |
| 1392 | { |
| 1393 | best_err = total_subset_err; |
| 1394 | best_common_pattern = common_pattern; |
| 1395 | } |
| 1396 | } |
| 1397 | |
| 1398 | first_common_pattern = best_common_pattern; |
| 1399 | last_common_pattern = best_common_pattern + 1; |
| 1400 | } |
| 1401 | |
| 1402 | //const uint32_t weight_range = 2; |
| 1403 | const uint32_t endpoint_range = 12; |
| 1404 | |
| 1405 | for (uint32_t common_pattern = first_common_pattern; common_pattern < last_common_pattern; common_pattern++) |
| 1406 | { |
| 1407 | const uint32_t astc_pattern = g_bc7_3_astc2_common_partitions[common_pattern].m_astc2; |
| 1408 | const uint32_t bc7_pattern = g_bc7_3_astc2_common_partitions[common_pattern].m_bc73; |
| 1409 | const uint32_t common_pattern_k = g_bc7_3_astc2_common_partitions[common_pattern].k; |
| 1410 | |
| 1411 | color_rgba part_pixels[2][16]; |
| 1412 | uint32_t part_pixel_index[4][4]; |
| 1413 | uint32_t num_part_pixels[2] = { 0, 0 }; |
| 1414 | |
| 1415 | for (uint32_t y = 0; y < 4; y++) |
| 1416 | { |
| 1417 | for (uint32_t x = 0; x < 4; x++) |
| 1418 | { |
| 1419 | const uint32_t astc_part = bc7_convert_partition_index_3_to_2(g_bc7_partition3[16 * bc7_pattern + x + y * 4], common_pattern_k); |
| 1420 | #ifdef _DEBUG |
| 1421 | assert((int)astc_part == astc_compute_texel_partition(astc_pattern, x, y, 0, 2, true)); |
| 1422 | #endif |
| 1423 | |
| 1424 | part_pixel_index[y][x] = num_part_pixels[astc_part]; |
| 1425 | part_pixels[astc_part][num_part_pixels[astc_part]++] = block[y][x]; |
| 1426 | } |
| 1427 | } |
| 1428 | |
| 1429 | color_cell_compressor_params ccell_params[2]; |
| 1430 | color_cell_compressor_results ccell_results[2]; |
| 1431 | uint8_t ccell_result_selectors[2][16]; |
| 1432 | uint8_t ccell_result_selectors_temp[2][16]; |
| 1433 | |
| 1434 | uint64_t total_part_err = 0; |
| 1435 | for (uint32_t part = 0; part < 2; part++) |
| 1436 | { |
| 1437 | memset(&ccell_params[part], 0, sizeof(ccell_params[part])); |
| 1438 | |
| 1439 | ccell_params[part].m_num_pixels = num_part_pixels[part]; |
| 1440 | ccell_params[part].m_pPixels = (color_quad_u8*)&part_pixels[part][0]; |
| 1441 | ccell_params[part].m_num_selector_weights = 4; |
| 1442 | ccell_params[part].m_pSelector_weights = g_bc7_weights2; |
| 1443 | ccell_params[part].m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1444 | ccell_params[part].m_astc_endpoint_range = endpoint_range; |
| 1445 | ccell_params[part].m_weights[0] = 1; |
| 1446 | ccell_params[part].m_weights[1] = 1; |
| 1447 | ccell_params[part].m_weights[2] = 1; |
| 1448 | ccell_params[part].m_weights[3] = 1; |
| 1449 | |
| 1450 | memset(&ccell_results[part], 0, sizeof(ccell_results[part])); |
| 1451 | ccell_results[part].m_pSelectors = &ccell_result_selectors[part][0]; |
| 1452 | ccell_results[part].m_pSelectors_temp = &ccell_result_selectors_temp[part][0]; |
| 1453 | |
| 1454 | uint64_t part_err = color_cell_compression(255, &ccell_params[part], &ccell_results[part], &comp_params); |
| 1455 | total_part_err += part_err; |
| 1456 | } // part |
| 1457 | |
| 1458 | // ASTC |
| 1459 | astc_block_desc blk; |
| 1460 | memset(&blk, 0, sizeof(blk)); |
| 1461 | |
| 1462 | blk.m_dual_plane = false; |
| 1463 | blk.m_weight_range = 2; |
| 1464 | |
| 1465 | blk.m_ccs = 0; |
| 1466 | blk.m_subsets = 2; |
| 1467 | blk.m_partition_seed = astc_pattern; |
| 1468 | blk.m_cem = 8; |
| 1469 | |
| 1470 | const uint32_t p0 = 0; |
| 1471 | const uint32_t p1 = 1; |
| 1472 | |
| 1473 | blk.m_endpoints[0] = ccell_results[p0].m_astc_low_endpoint.m_c[0]; |
| 1474 | blk.m_endpoints[1] = ccell_results[p0].m_astc_high_endpoint.m_c[0]; |
| 1475 | blk.m_endpoints[2] = ccell_results[p0].m_astc_low_endpoint.m_c[1]; |
| 1476 | blk.m_endpoints[3] = ccell_results[p0].m_astc_high_endpoint.m_c[1]; |
| 1477 | blk.m_endpoints[4] = ccell_results[p0].m_astc_low_endpoint.m_c[2]; |
| 1478 | blk.m_endpoints[5] = ccell_results[p0].m_astc_high_endpoint.m_c[2]; |
| 1479 | |
| 1480 | bool invert[2] = { false, false }; |
| 1481 | |
| 1482 | int s0 = g_astc_unquant[endpoint_range][blk.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[4]].m_unquant; |
| 1483 | int s1 = g_astc_unquant[endpoint_range][blk.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[5]].m_unquant; |
| 1484 | if (s1 < s0) |
| 1485 | { |
| 1486 | std::swap(blk.m_endpoints[0], blk.m_endpoints[1]); |
| 1487 | std::swap(blk.m_endpoints[2], blk.m_endpoints[3]); |
| 1488 | std::swap(blk.m_endpoints[4], blk.m_endpoints[5]); |
| 1489 | invert[0] = true; |
| 1490 | } |
| 1491 | |
| 1492 | blk.m_endpoints[6] = ccell_results[p1].m_astc_low_endpoint.m_c[0]; |
| 1493 | blk.m_endpoints[7] = ccell_results[p1].m_astc_high_endpoint.m_c[0]; |
| 1494 | blk.m_endpoints[8] = ccell_results[p1].m_astc_low_endpoint.m_c[1]; |
| 1495 | blk.m_endpoints[9] = ccell_results[p1].m_astc_high_endpoint.m_c[1]; |
| 1496 | blk.m_endpoints[10] = ccell_results[p1].m_astc_low_endpoint.m_c[2]; |
| 1497 | blk.m_endpoints[11] = ccell_results[p1].m_astc_high_endpoint.m_c[2]; |
| 1498 | |
| 1499 | s0 = g_astc_unquant[endpoint_range][blk.m_endpoints[0 + 6]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[2 + 6]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[4 + 6]].m_unquant; |
| 1500 | s1 = g_astc_unquant[endpoint_range][blk.m_endpoints[1 + 6]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[3 + 6]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[5 + 6]].m_unquant; |
| 1501 | |
| 1502 | if (s1 < s0) |
| 1503 | { |
| 1504 | std::swap(blk.m_endpoints[0 + 6], blk.m_endpoints[1 + 6]); |
| 1505 | std::swap(blk.m_endpoints[2 + 6], blk.m_endpoints[3 + 6]); |
| 1506 | std::swap(blk.m_endpoints[4 + 6], blk.m_endpoints[5 + 6]); |
| 1507 | invert[1] = true; |
| 1508 | } |
| 1509 | |
| 1510 | for (uint32_t y = 0; y < 4; y++) |
| 1511 | { |
| 1512 | for (uint32_t x = 0; x < 4; x++) |
| 1513 | { |
| 1514 | const uint32_t astc_part = bc7_convert_partition_index_3_to_2(g_bc7_partition3[16 * bc7_pattern + x + y * 4], common_pattern_k); |
| 1515 | |
| 1516 | blk.m_weights[x + y * 4] = ccell_result_selectors[astc_part][part_pixel_index[y][x]]; |
| 1517 | |
| 1518 | if (invert[astc_part]) |
| 1519 | blk.m_weights[x + y * 4] = 3 - blk.m_weights[x + y * 4]; |
| 1520 | } |
| 1521 | } |
| 1522 | |
| 1523 | assert(total_results < MAX_ENCODE_RESULTS); |
| 1524 | if (total_results < MAX_ENCODE_RESULTS) |
| 1525 | { |
| 1526 | pResults[total_results].m_uastc_mode = 7; |
| 1527 | pResults[total_results].m_common_pattern = common_pattern; |
| 1528 | pResults[total_results].m_astc = blk; |
| 1529 | pResults[total_results].m_astc_err = total_part_err; |
| 1530 | total_results++; |
| 1531 | } |
| 1532 | |
| 1533 | } // common_pattern |
| 1534 | } |
| 1535 | |
| 1536 | static void estimate_partition2_list(uint32_t num_weights, uint32_t num_comps, const uint32_t* pWeights, const color_rgba block[4][4], uint32_t* pParts, uint32_t max_parts, const uint32_t weights[4]) |
| 1537 | { |
| 1538 | assert(pWeights[0] == 0 && pWeights[num_weights - 1] == 64); |
| 1539 | |
| 1540 | const uint32_t MAX_PARTS = 8; |
| 1541 | assert(max_parts <= MAX_PARTS); |
| 1542 | |
| 1543 | uint64_t part_error[MAX_PARTS]; |
| 1544 | memset(part_error, 0xFF, sizeof(part_error)); |
| 1545 | memset(pParts, 0, sizeof(pParts[0]) * max_parts); |
| 1546 | |
| 1547 | for (uint32_t common_pattern = 0; common_pattern < TOTAL_ASTC_BC7_COMMON_PARTITIONS2; common_pattern++) |
| 1548 | { |
| 1549 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions2[common_pattern].m_bc7; |
| 1550 | |
| 1551 | const uint8_t* pPartition = &g_bc7_partition2[bc7_pattern * 16]; |
| 1552 | |
| 1553 | color_quad_u8 subset_colors[2][16]; |
| 1554 | uint32_t subset_total_colors[2] = { 0, 0 }; |
| 1555 | for (uint32_t index = 0; index < 16; index++) |
| 1556 | subset_colors[pPartition[index]][subset_total_colors[pPartition[index]]++] = ((const color_quad_u8*)block)[index]; |
| 1557 | |
| 1558 | uint64_t total_subset_err = 0; |
| 1559 | for (uint32_t subset = 0; subset < 2; subset++) |
| 1560 | total_subset_err += color_cell_compression_est_astc(num_weights, num_comps, pWeights, subset_total_colors[subset], &subset_colors[subset][0], UINT64_MAX, weights); |
| 1561 | |
| 1562 | for (int i = 0; i < (int)max_parts; i++) |
| 1563 | { |
| 1564 | if (total_subset_err < part_error[i]) |
| 1565 | { |
| 1566 | for (int j = max_parts - 1; j > i; --j) |
| 1567 | { |
| 1568 | pParts[j] = pParts[j - 1]; |
| 1569 | part_error[j] = part_error[j - 1]; |
| 1570 | } |
| 1571 | |
| 1572 | pParts[i] = common_pattern; |
| 1573 | part_error[i] = total_subset_err; |
| 1574 | |
| 1575 | break; |
| 1576 | } |
| 1577 | } |
| 1578 | } |
| 1579 | |
| 1580 | #ifdef _DEBUG |
| 1581 | for (uint32_t i = 0; i < max_parts - 1; i++) |
| 1582 | { |
| 1583 | assert(part_error[i] <= part_error[i + 1]); |
| 1584 | } |
| 1585 | #endif |
| 1586 | } |
| 1587 | |
| 1588 | // 9. DualPlane: 0, WeightRange: 2 (4), Subsets: 2, CEM: 12 (RGBA Direct), EndpointRange: 8 (16) - BC7 MODE 7 |
| 1589 | // 16. DualPlane: 0, WeightRange : 2 (4), Subsets : 2, CEM: 4 (LA Direct), EndpointRange : 20 (256) - BC7 MODE 7 |
| 1590 | static void astc_mode9_or_16(uint32_t mode, const color_rgba source_block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params, uint32_t estimate_partition_list_size) |
| 1591 | { |
| 1592 | assert(mode == 9 || mode == 16); |
| 1593 | |
| 1594 | const color_rgba* pBlock = &source_block[0][0]; |
| 1595 | |
| 1596 | color_rgba temp_block[16]; |
| 1597 | if (mode == 16) |
| 1598 | { |
| 1599 | for (uint32_t i = 0; i < 16; i++) |
| 1600 | { |
| 1601 | if (mode == 16) |
| 1602 | { |
| 1603 | assert(pBlock[i].r == pBlock[i].g); |
| 1604 | assert(pBlock[i].r == pBlock[i].b); |
| 1605 | } |
| 1606 | |
| 1607 | const uint32_t l = pBlock[i].r; |
| 1608 | const uint32_t a = pBlock[i].a; |
| 1609 | |
| 1610 | // Use (l,0,0,a) not (l,l,l,a) so both components are treated equally. |
| 1611 | temp_block[i].set_noclamp_rgba(l, 0, 0, a); |
| 1612 | } |
| 1613 | |
| 1614 | pBlock = temp_block; |
| 1615 | } |
| 1616 | |
| 1617 | const uint32_t weights[4] = { 1, 1, 1, 1 }; |
| 1618 | |
| 1619 | //const uint32_t weight_range = 2; |
| 1620 | const uint32_t endpoint_range = (mode == 16) ? 20 : 8; |
| 1621 | |
| 1622 | uint32_t first_common_pattern = 0; |
| 1623 | uint32_t last_common_pattern = TOTAL_ASTC_BC7_COMMON_PARTITIONS2; |
| 1624 | bool use_part_list = false; |
| 1625 | |
| 1626 | const uint32_t MAX_PARTS = 8; |
| 1627 | uint32_t parts[MAX_PARTS]; |
| 1628 | |
| 1629 | if (estimate_partition_list_size == 1) |
| 1630 | { |
| 1631 | first_common_pattern = estimate_partition2(4, 4, g_bc7_weights2, (const color_rgba(*)[4])pBlock, weights); |
| 1632 | last_common_pattern = first_common_pattern + 1; |
| 1633 | } |
| 1634 | else if (estimate_partition_list_size > 0) |
| 1635 | { |
| 1636 | assert(estimate_partition_list_size <= MAX_PARTS); |
| 1637 | estimate_partition_list_size = basisu::minimum(estimate_partition_list_size, MAX_PARTS); |
| 1638 | |
| 1639 | estimate_partition2_list(4, 4, g_bc7_weights2, (const color_rgba(*)[4])pBlock, parts, estimate_partition_list_size, weights); |
| 1640 | |
| 1641 | first_common_pattern = 0; |
| 1642 | last_common_pattern = estimate_partition_list_size; |
| 1643 | use_part_list = true; |
| 1644 | |
| 1645 | #ifdef _DEBUG |
| 1646 | assert(parts[0] == estimate_partition2(4, 4, g_bc7_weights2, (const color_rgba(*)[4])pBlock, weights)); |
| 1647 | #endif |
| 1648 | } |
| 1649 | |
| 1650 | for (uint32_t common_pattern_iter = first_common_pattern; common_pattern_iter < last_common_pattern; common_pattern_iter++) |
| 1651 | { |
| 1652 | const uint32_t common_pattern = use_part_list ? parts[common_pattern_iter] : common_pattern_iter; |
| 1653 | |
| 1654 | const uint32_t bc7_pattern = g_astc_bc7_common_partitions2[common_pattern].m_bc7; |
| 1655 | |
| 1656 | color_rgba part_pixels[2][16]; |
| 1657 | uint32_t part_pixel_index[4][4]; |
| 1658 | uint32_t num_part_pixels[2] = { 0, 0 }; |
| 1659 | |
| 1660 | for (uint32_t y = 0; y < 4; y++) |
| 1661 | { |
| 1662 | for (uint32_t x = 0; x < 4; x++) |
| 1663 | { |
| 1664 | const uint32_t part = g_bc7_partition2[16 * bc7_pattern + x + y * 4]; |
| 1665 | part_pixel_index[y][x] = num_part_pixels[part]; |
| 1666 | part_pixels[part][num_part_pixels[part]++] = pBlock[y * 4 + x]; |
| 1667 | } |
| 1668 | } |
| 1669 | |
| 1670 | color_cell_compressor_params ccell_params[2]; |
| 1671 | color_cell_compressor_results ccell_results[2]; |
| 1672 | uint8_t ccell_result_selectors[2][16]; |
| 1673 | uint8_t ccell_result_selectors_temp[2][16]; |
| 1674 | |
| 1675 | uint64_t total_err = 0; |
| 1676 | for (uint32_t subset = 0; subset < 2; subset++) |
| 1677 | { |
| 1678 | memset(&ccell_params[subset], 0, sizeof(ccell_params[subset])); |
| 1679 | |
| 1680 | ccell_params[subset].m_num_pixels = num_part_pixels[subset]; |
| 1681 | ccell_params[subset].m_pPixels = (color_quad_u8*)&part_pixels[subset][0]; |
| 1682 | ccell_params[subset].m_num_selector_weights = 4; |
| 1683 | ccell_params[subset].m_pSelector_weights = g_bc7_weights2; |
| 1684 | ccell_params[subset].m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1685 | ccell_params[subset].m_astc_endpoint_range = endpoint_range; |
| 1686 | ccell_params[subset].m_weights[0] = weights[0]; |
| 1687 | ccell_params[subset].m_weights[1] = weights[1]; |
| 1688 | ccell_params[subset].m_weights[2] = weights[2]; |
| 1689 | ccell_params[subset].m_weights[3] = weights[3]; |
| 1690 | ccell_params[subset].m_has_alpha = true; |
| 1691 | |
| 1692 | memset(&ccell_results[subset], 0, sizeof(ccell_results[subset])); |
| 1693 | ccell_results[subset].m_pSelectors = &ccell_result_selectors[subset][0]; |
| 1694 | ccell_results[subset].m_pSelectors_temp = &ccell_result_selectors_temp[subset][0]; |
| 1695 | |
| 1696 | uint64_t subset_err = color_cell_compression(255, &ccell_params[subset], &ccell_results[subset], &comp_params); |
| 1697 | |
| 1698 | if (mode == 16) |
| 1699 | { |
| 1700 | color_rgba colors[4]; |
| 1701 | for (uint32_t c = 0; c < 4; c++) |
| 1702 | { |
| 1703 | colors[0].m_comps[c] = g_astc_unquant[endpoint_range][ccell_results[subset].m_astc_low_endpoint.m_c[(c < 3) ? 0 : 3]].m_unquant; |
| 1704 | colors[3].m_comps[c] = g_astc_unquant[endpoint_range][ccell_results[subset].m_astc_high_endpoint.m_c[(c < 3) ? 0 : 3]].m_unquant; |
| 1705 | } |
| 1706 | |
| 1707 | for (uint32_t i = 1; i < 4 - 1; i++) |
| 1708 | for (uint32_t c = 0; c < 4; c++) |
| 1709 | colors[i].m_comps[c] = (uint8_t)astc_interpolate(colors[0].m_comps[c], colors[3].m_comps[c], g_bc7_weights2[i], false); |
| 1710 | |
| 1711 | for (uint32_t p = 0; p < ccell_params[subset].m_num_pixels; p++) |
| 1712 | { |
| 1713 | color_rgba orig_pix(part_pixels[subset][p]); |
| 1714 | orig_pix.g = orig_pix.r; |
| 1715 | orig_pix.b = orig_pix.r; |
| 1716 | total_err += color_distance_la(orig_pix, colors[ccell_result_selectors[subset][p]]); |
| 1717 | } |
| 1718 | } |
| 1719 | else |
| 1720 | { |
| 1721 | total_err += subset_err; |
| 1722 | } |
| 1723 | } // subset |
| 1724 | |
| 1725 | // ASTC |
| 1726 | astc_block_desc astc_results; |
| 1727 | memset(&astc_results, 0, sizeof(astc_results)); |
| 1728 | |
| 1729 | astc_results.m_dual_plane = false; |
| 1730 | astc_results.m_weight_range = 2; |
| 1731 | |
| 1732 | astc_results.m_ccs = 0; |
| 1733 | astc_results.m_subsets = 2; |
| 1734 | astc_results.m_partition_seed = g_astc_bc7_common_partitions2[common_pattern].m_astc; |
| 1735 | astc_results.m_cem = (mode == 16) ? 4 : 12; |
| 1736 | |
| 1737 | uint32_t part[2] = { 0, 1 }; |
| 1738 | if (g_astc_bc7_common_partitions2[common_pattern].m_invert) |
| 1739 | std::swap(part[0], part[1]); |
| 1740 | |
| 1741 | bool invert[2] = { false, false }; |
| 1742 | |
| 1743 | for (uint32_t p = 0; p < 2; p++) |
| 1744 | { |
| 1745 | if (mode == 16) |
| 1746 | { |
| 1747 | astc_results.m_endpoints[p * 4 + 0] = ccell_results[part[p]].m_astc_low_endpoint.m_c[0]; |
| 1748 | astc_results.m_endpoints[p * 4 + 1] = ccell_results[part[p]].m_astc_high_endpoint.m_c[0]; |
| 1749 | |
| 1750 | astc_results.m_endpoints[p * 4 + 2] = ccell_results[part[p]].m_astc_low_endpoint.m_c[3]; |
| 1751 | astc_results.m_endpoints[p * 4 + 3] = ccell_results[part[p]].m_astc_high_endpoint.m_c[3]; |
| 1752 | } |
| 1753 | else |
| 1754 | { |
| 1755 | for (uint32_t c = 0; c < 4; c++) |
| 1756 | { |
| 1757 | astc_results.m_endpoints[p * 8 + c * 2] = ccell_results[part[p]].m_astc_low_endpoint.m_c[c]; |
| 1758 | astc_results.m_endpoints[p * 8 + c * 2 + 1] = ccell_results[part[p]].m_astc_high_endpoint.m_c[c]; |
| 1759 | } |
| 1760 | |
| 1761 | int s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[p * 8 + 0]].m_unquant + |
| 1762 | g_astc_unquant[endpoint_range][astc_results.m_endpoints[p * 8 + 2]].m_unquant + |
| 1763 | g_astc_unquant[endpoint_range][astc_results.m_endpoints[p * 8 + 4]].m_unquant; |
| 1764 | |
| 1765 | int s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[p * 8 + 1]].m_unquant + |
| 1766 | g_astc_unquant[endpoint_range][astc_results.m_endpoints[p * 8 + 3]].m_unquant + |
| 1767 | g_astc_unquant[endpoint_range][astc_results.m_endpoints[p * 8 + 5]].m_unquant; |
| 1768 | |
| 1769 | if (s1 < s0) |
| 1770 | { |
| 1771 | std::swap(astc_results.m_endpoints[p * 8 + 0], astc_results.m_endpoints[p * 8 + 1]); |
| 1772 | std::swap(astc_results.m_endpoints[p * 8 + 2], astc_results.m_endpoints[p * 8 + 3]); |
| 1773 | std::swap(astc_results.m_endpoints[p * 8 + 4], astc_results.m_endpoints[p * 8 + 5]); |
| 1774 | std::swap(astc_results.m_endpoints[p * 8 + 6], astc_results.m_endpoints[p * 8 + 7]); |
| 1775 | invert[p] = true; |
| 1776 | } |
| 1777 | } |
| 1778 | } |
| 1779 | |
| 1780 | for (uint32_t y = 0; y < 4; y++) |
| 1781 | { |
| 1782 | for (uint32_t x = 0; x < 4; x++) |
| 1783 | { |
| 1784 | const uint32_t bc7_part = g_bc7_partition2[16 * bc7_pattern + x + y * 4]; |
| 1785 | |
| 1786 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[bc7_part][part_pixel_index[y][x]]; |
| 1787 | |
| 1788 | uint32_t astc_part = bc7_part; |
| 1789 | if (g_astc_bc7_common_partitions2[common_pattern].m_invert) |
| 1790 | astc_part = 1 - astc_part; |
| 1791 | |
| 1792 | if (invert[astc_part]) |
| 1793 | astc_results.m_weights[x + y * 4] = 3 - astc_results.m_weights[x + y * 4]; |
| 1794 | } |
| 1795 | } |
| 1796 | |
| 1797 | assert(total_results < MAX_ENCODE_RESULTS); |
| 1798 | if (total_results < MAX_ENCODE_RESULTS) |
| 1799 | { |
| 1800 | pResults[total_results].m_uastc_mode = mode; |
| 1801 | pResults[total_results].m_common_pattern = common_pattern; |
| 1802 | pResults[total_results].m_astc = astc_results; |
| 1803 | pResults[total_results].m_astc_err = total_err; |
| 1804 | total_results++; |
| 1805 | } |
| 1806 | |
| 1807 | } // common_pattern |
| 1808 | } |
| 1809 | |
| 1810 | // MODE 10 |
| 1811 | // DualPlane: 0, WeightRange: 8 (16), Subsets: 1, CEM: 12 (RGBA Direct ), EndpointRange: 13 (48) MODE6 |
| 1812 | static void astc_mode10(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 1813 | { |
| 1814 | const uint32_t weight_range = 8; |
| 1815 | const uint32_t endpoint_range = 13; |
| 1816 | |
| 1817 | color_cell_compressor_params ccell_params; |
| 1818 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 1819 | |
| 1820 | ccell_params.m_num_pixels = 16; |
| 1821 | ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 1822 | ccell_params.m_num_selector_weights = 16; |
| 1823 | ccell_params.m_pSelector_weights = g_astc_weights4; |
| 1824 | ccell_params.m_pSelector_weightsx = (const bc7enc_vec4F*)g_astc_weights4x; |
| 1825 | ccell_params.m_astc_endpoint_range = endpoint_range; |
| 1826 | ccell_params.m_weights[0] = 1; |
| 1827 | ccell_params.m_weights[1] = 1; |
| 1828 | ccell_params.m_weights[2] = 1; |
| 1829 | ccell_params.m_weights[3] = 1; |
| 1830 | ccell_params.m_has_alpha = true; |
| 1831 | |
| 1832 | color_cell_compressor_results ccell_results; |
| 1833 | uint8_t ccell_result_selectors[16]; |
| 1834 | uint8_t ccell_result_selectors_temp[16]; |
| 1835 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 1836 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 1837 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 1838 | |
| 1839 | uint64_t part_err = color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 1840 | |
| 1841 | // ASTC |
| 1842 | astc_block_desc astc_results; |
| 1843 | memset(&astc_results, 0, sizeof(astc_results)); |
| 1844 | |
| 1845 | astc_results.m_dual_plane = false; |
| 1846 | astc_results.m_weight_range = weight_range; |
| 1847 | |
| 1848 | astc_results.m_ccs = 0; |
| 1849 | astc_results.m_subsets = 1; |
| 1850 | astc_results.m_partition_seed = 0; |
| 1851 | astc_results.m_cem = 12; |
| 1852 | |
| 1853 | astc_results.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 1854 | astc_results.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 1855 | astc_results.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[1]; |
| 1856 | astc_results.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[1]; |
| 1857 | astc_results.m_endpoints[4] = ccell_results.m_astc_low_endpoint.m_c[2]; |
| 1858 | astc_results.m_endpoints[5] = ccell_results.m_astc_high_endpoint.m_c[2]; |
| 1859 | astc_results.m_endpoints[6] = ccell_results.m_astc_low_endpoint.m_c[3]; |
| 1860 | astc_results.m_endpoints[7] = ccell_results.m_astc_high_endpoint.m_c[3]; |
| 1861 | |
| 1862 | bool invert = false; |
| 1863 | |
| 1864 | int s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[4]].m_unquant; |
| 1865 | int s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[5]].m_unquant; |
| 1866 | if (s1 < s0) |
| 1867 | { |
| 1868 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 1869 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 1870 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 1871 | std::swap(astc_results.m_endpoints[6], astc_results.m_endpoints[7]); |
| 1872 | invert = true; |
| 1873 | } |
| 1874 | |
| 1875 | for (uint32_t y = 0; y < 4; y++) |
| 1876 | { |
| 1877 | for (uint32_t x = 0; x < 4; x++) |
| 1878 | { |
| 1879 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 1880 | |
| 1881 | if (invert) |
| 1882 | astc_results.m_weights[x + y * 4] = 15 - astc_results.m_weights[x + y * 4]; |
| 1883 | } |
| 1884 | } |
| 1885 | |
| 1886 | assert(total_results < MAX_ENCODE_RESULTS); |
| 1887 | if (total_results < MAX_ENCODE_RESULTS) |
| 1888 | { |
| 1889 | pResults[total_results].m_uastc_mode = 10; |
| 1890 | pResults[total_results].m_common_pattern = 0; |
| 1891 | pResults[total_results].m_astc = astc_results; |
| 1892 | pResults[total_results].m_astc_err = part_err; |
| 1893 | total_results++; |
| 1894 | } |
| 1895 | } |
| 1896 | |
| 1897 | // 11. DualPlane: 1, WeightRange: 2 (4), Subsets: 1, CEM: 12 (RGBA Direct), EndpointRange: 13 (48) MODE5 |
| 1898 | // 17. DualPlane: 1, WeightRange : 2 (4), Subsets : 1, CEM : 4 (LA Direct), EndpointRange : 20 (256) BC7 MODE5 |
| 1899 | static void astc_mode11_or_17(uint32_t mode, const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 1900 | { |
| 1901 | assert((mode == 11) || (mode == 17)); |
| 1902 | |
| 1903 | const uint32_t weight_range = 2; |
| 1904 | const uint32_t endpoint_range = (mode == 17) ? 20 : 13; |
| 1905 | |
| 1906 | bc7enc_compress_block_params local_comp_params(comp_params); |
| 1907 | local_comp_params.m_perceptual = false; |
| 1908 | local_comp_params.m_weights[0] = 1; |
| 1909 | local_comp_params.m_weights[1] = 1; |
| 1910 | local_comp_params.m_weights[2] = 1; |
| 1911 | local_comp_params.m_weights[3] = 1; |
| 1912 | |
| 1913 | const uint32_t last_rot_comp = (mode == 17) ? 1 : 4; |
| 1914 | |
| 1915 | for (uint32_t rot_comp = 0; rot_comp < last_rot_comp; rot_comp++) |
| 1916 | { |
| 1917 | color_quad_u8 block_rgb[16]; |
| 1918 | color_quad_u8 block_a[16]; |
| 1919 | for (uint32_t i = 0; i < 16; i++) |
| 1920 | { |
| 1921 | block_rgb[i] = ((color_quad_u8*)&block[0][0])[i]; |
| 1922 | block_a[i] = block_rgb[i]; |
| 1923 | |
| 1924 | if (mode == 17) |
| 1925 | { |
| 1926 | assert(block_rgb[i].m_c[0] == block_rgb[i].m_c[1]); |
| 1927 | assert(block_rgb[i].m_c[0] == block_rgb[i].m_c[2]); |
| 1928 | |
| 1929 | block_a[i].m_c[0] = block_rgb[i].m_c[3]; |
| 1930 | block_a[i].m_c[1] = block_rgb[i].m_c[3]; |
| 1931 | block_a[i].m_c[2] = block_rgb[i].m_c[3]; |
| 1932 | block_a[i].m_c[3] = 255; |
| 1933 | |
| 1934 | block_rgb[i].m_c[1] = block_rgb[i].m_c[0]; |
| 1935 | block_rgb[i].m_c[2] = block_rgb[i].m_c[0]; |
| 1936 | block_rgb[i].m_c[3] = 255; |
| 1937 | } |
| 1938 | else |
| 1939 | { |
| 1940 | uint8_t c = block_a[i].m_c[rot_comp]; |
| 1941 | block_a[i].m_c[0] = c; |
| 1942 | block_a[i].m_c[1] = c; |
| 1943 | block_a[i].m_c[2] = c; |
| 1944 | block_a[i].m_c[3] = 255; |
| 1945 | |
| 1946 | block_rgb[i].m_c[rot_comp] = block_rgb[i].m_c[3]; |
| 1947 | block_rgb[i].m_c[3] = 255; |
| 1948 | } |
| 1949 | } |
| 1950 | |
| 1951 | uint8_t ccell_result_selectors_temp[16]; |
| 1952 | |
| 1953 | color_cell_compressor_params ccell_params_rgb; |
| 1954 | memset(&ccell_params_rgb, 0, sizeof(ccell_params_rgb)); |
| 1955 | |
| 1956 | ccell_params_rgb.m_num_pixels = 16; |
| 1957 | ccell_params_rgb.m_pPixels = block_rgb; |
| 1958 | ccell_params_rgb.m_num_selector_weights = 4; |
| 1959 | ccell_params_rgb.m_pSelector_weights = g_bc7_weights2; |
| 1960 | ccell_params_rgb.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1961 | ccell_params_rgb.m_astc_endpoint_range = endpoint_range; |
| 1962 | ccell_params_rgb.m_weights[0] = 1; |
| 1963 | ccell_params_rgb.m_weights[1] = 1; |
| 1964 | ccell_params_rgb.m_weights[2] = 1; |
| 1965 | ccell_params_rgb.m_weights[3] = 1; |
| 1966 | |
| 1967 | color_cell_compressor_results ccell_results_rgb; |
| 1968 | uint8_t ccell_result_selectors_rgb[16]; |
| 1969 | memset(&ccell_results_rgb, 0, sizeof(ccell_results_rgb)); |
| 1970 | ccell_results_rgb.m_pSelectors = &ccell_result_selectors_rgb[0]; |
| 1971 | ccell_results_rgb.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 1972 | |
| 1973 | uint64_t part_err_rgb = color_cell_compression(255, &ccell_params_rgb, &ccell_results_rgb, &local_comp_params); |
| 1974 | |
| 1975 | color_cell_compressor_params ccell_params_a; |
| 1976 | memset(&ccell_params_a, 0, sizeof(ccell_params_a)); |
| 1977 | |
| 1978 | ccell_params_a.m_num_pixels = 16; |
| 1979 | ccell_params_a.m_pPixels = block_a; |
| 1980 | ccell_params_a.m_num_selector_weights = 4; |
| 1981 | ccell_params_a.m_pSelector_weights = g_bc7_weights2; |
| 1982 | ccell_params_a.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 1983 | ccell_params_a.m_astc_endpoint_range = endpoint_range; |
| 1984 | ccell_params_a.m_weights[0] = 1; |
| 1985 | ccell_params_a.m_weights[1] = 1; |
| 1986 | ccell_params_a.m_weights[2] = 1; |
| 1987 | ccell_params_a.m_weights[3] = 1; |
| 1988 | |
| 1989 | color_cell_compressor_results ccell_results_a; |
| 1990 | uint8_t ccell_result_selectors_a[16]; |
| 1991 | memset(&ccell_results_a, 0, sizeof(ccell_results_a)); |
| 1992 | ccell_results_a.m_pSelectors = &ccell_result_selectors_a[0]; |
| 1993 | ccell_results_a.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 1994 | |
| 1995 | uint64_t part_err_a = color_cell_compression(255, &ccell_params_a, &ccell_results_a, &local_comp_params) / 3; |
| 1996 | |
| 1997 | uint64_t total_err = (mode == 17) ? ((part_err_rgb / 3) + part_err_a) : (part_err_rgb + part_err_a); |
| 1998 | |
| 1999 | // ASTC |
| 2000 | astc_block_desc blk; |
| 2001 | memset(&blk, 0, sizeof(blk)); |
| 2002 | |
| 2003 | blk.m_dual_plane = true; |
| 2004 | blk.m_weight_range = weight_range; |
| 2005 | |
| 2006 | blk.m_ccs = (mode == 17) ? 3 : rot_comp; |
| 2007 | blk.m_subsets = 1; |
| 2008 | blk.m_partition_seed = 0; |
| 2009 | blk.m_cem = (mode == 17) ? 4 : 12; |
| 2010 | |
| 2011 | bool invert = false; |
| 2012 | |
| 2013 | if (mode == 17) |
| 2014 | { |
| 2015 | assert(ccell_results_rgb.m_astc_low_endpoint.m_c[0] == ccell_results_rgb.m_astc_low_endpoint.m_c[1]); |
| 2016 | assert(ccell_results_rgb.m_astc_low_endpoint.m_c[0] == ccell_results_rgb.m_astc_low_endpoint.m_c[2]); |
| 2017 | |
| 2018 | assert(ccell_results_rgb.m_astc_high_endpoint.m_c[0] == ccell_results_rgb.m_astc_high_endpoint.m_c[1]); |
| 2019 | assert(ccell_results_rgb.m_astc_high_endpoint.m_c[0] == ccell_results_rgb.m_astc_high_endpoint.m_c[2]); |
| 2020 | |
| 2021 | blk.m_endpoints[0] = ccell_results_rgb.m_astc_low_endpoint.m_c[0]; |
| 2022 | blk.m_endpoints[1] = ccell_results_rgb.m_astc_high_endpoint.m_c[0]; |
| 2023 | |
| 2024 | blk.m_endpoints[2] = ccell_results_a.m_astc_low_endpoint.m_c[0]; |
| 2025 | blk.m_endpoints[3] = ccell_results_a.m_astc_high_endpoint.m_c[0]; |
| 2026 | } |
| 2027 | else |
| 2028 | { |
| 2029 | blk.m_endpoints[0] = (rot_comp == 0 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[0]; |
| 2030 | blk.m_endpoints[1] = (rot_comp == 0 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[0]; |
| 2031 | blk.m_endpoints[2] = (rot_comp == 1 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[1]; |
| 2032 | blk.m_endpoints[3] = (rot_comp == 1 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[1]; |
| 2033 | blk.m_endpoints[4] = (rot_comp == 2 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[2]; |
| 2034 | blk.m_endpoints[5] = (rot_comp == 2 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[2]; |
| 2035 | if (rot_comp == 3) |
| 2036 | { |
| 2037 | blk.m_endpoints[6] = ccell_results_a.m_astc_low_endpoint.m_c[0]; |
| 2038 | blk.m_endpoints[7] = ccell_results_a.m_astc_high_endpoint.m_c[0]; |
| 2039 | } |
| 2040 | else |
| 2041 | { |
| 2042 | blk.m_endpoints[6] = ccell_results_rgb.m_astc_low_endpoint.m_c[rot_comp]; |
| 2043 | blk.m_endpoints[7] = ccell_results_rgb.m_astc_high_endpoint.m_c[rot_comp]; |
| 2044 | } |
| 2045 | |
| 2046 | int s0 = g_astc_unquant[endpoint_range][blk.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[4]].m_unquant; |
| 2047 | int s1 = g_astc_unquant[endpoint_range][blk.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[5]].m_unquant; |
| 2048 | if (s1 < s0) |
| 2049 | { |
| 2050 | std::swap(blk.m_endpoints[0], blk.m_endpoints[1]); |
| 2051 | std::swap(blk.m_endpoints[2], blk.m_endpoints[3]); |
| 2052 | std::swap(blk.m_endpoints[4], blk.m_endpoints[5]); |
| 2053 | std::swap(blk.m_endpoints[6], blk.m_endpoints[7]); |
| 2054 | invert = true; |
| 2055 | } |
| 2056 | } |
| 2057 | |
| 2058 | for (uint32_t y = 0; y < 4; y++) |
| 2059 | { |
| 2060 | for (uint32_t x = 0; x < 4; x++) |
| 2061 | { |
| 2062 | uint32_t rgb_index = ccell_result_selectors_rgb[x + y * 4]; |
| 2063 | uint32_t a_index = ccell_result_selectors_a[x + y * 4]; |
| 2064 | |
| 2065 | if (invert) |
| 2066 | { |
| 2067 | rgb_index = 3 - rgb_index; |
| 2068 | a_index = 3 - a_index; |
| 2069 | } |
| 2070 | |
| 2071 | blk.m_weights[(x + y * 4) * 2 + 0] = (uint8_t)rgb_index; |
| 2072 | blk.m_weights[(x + y * 4) * 2 + 1] = (uint8_t)a_index; |
| 2073 | } |
| 2074 | } |
| 2075 | |
| 2076 | assert(total_results < MAX_ENCODE_RESULTS); |
| 2077 | if (total_results < MAX_ENCODE_RESULTS) |
| 2078 | { |
| 2079 | pResults[total_results].m_uastc_mode = mode; |
| 2080 | pResults[total_results].m_common_pattern = 0; |
| 2081 | pResults[total_results].m_astc = blk; |
| 2082 | pResults[total_results].m_astc_err = total_err; |
| 2083 | total_results++; |
| 2084 | } |
| 2085 | } // rot_comp |
| 2086 | } |
| 2087 | |
| 2088 | // MODE 12 |
| 2089 | // DualPlane: 0, WeightRange: 5 (8), Subsets: 1, CEM: 12 (RGBA Direct ), EndpointRange: 19 (192) MODE6 |
| 2090 | static void astc_mode12(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 2091 | { |
| 2092 | const uint32_t weight_range = 5; |
| 2093 | const uint32_t endpoint_range = 19; |
| 2094 | |
| 2095 | color_cell_compressor_params ccell_params; |
| 2096 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 2097 | |
| 2098 | ccell_params.m_num_pixels = 16; |
| 2099 | ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 2100 | ccell_params.m_num_selector_weights = 8; |
| 2101 | ccell_params.m_pSelector_weights = g_bc7_weights3; |
| 2102 | ccell_params.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights3x; |
| 2103 | ccell_params.m_astc_endpoint_range = endpoint_range; |
| 2104 | ccell_params.m_weights[0] = 1; |
| 2105 | ccell_params.m_weights[1] = 1; |
| 2106 | ccell_params.m_weights[2] = 1; |
| 2107 | ccell_params.m_weights[3] = 1; |
| 2108 | ccell_params.m_has_alpha = true; |
| 2109 | |
| 2110 | color_cell_compressor_results ccell_results; |
| 2111 | uint8_t ccell_result_selectors[16]; |
| 2112 | uint8_t ccell_result_selectors_temp[16]; |
| 2113 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 2114 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 2115 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 2116 | |
| 2117 | uint64_t part_err = color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 2118 | |
| 2119 | // ASTC |
| 2120 | astc_block_desc astc_results; |
| 2121 | memset(&astc_results, 0, sizeof(astc_results)); |
| 2122 | |
| 2123 | astc_results.m_dual_plane = false; |
| 2124 | astc_results.m_weight_range = weight_range; |
| 2125 | |
| 2126 | astc_results.m_ccs = 0; |
| 2127 | astc_results.m_subsets = 1; |
| 2128 | astc_results.m_partition_seed = 0; |
| 2129 | astc_results.m_cem = 12; |
| 2130 | |
| 2131 | astc_results.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 2132 | astc_results.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 2133 | astc_results.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[1]; |
| 2134 | astc_results.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[1]; |
| 2135 | astc_results.m_endpoints[4] = ccell_results.m_astc_low_endpoint.m_c[2]; |
| 2136 | astc_results.m_endpoints[5] = ccell_results.m_astc_high_endpoint.m_c[2]; |
| 2137 | astc_results.m_endpoints[6] = ccell_results.m_astc_low_endpoint.m_c[3]; |
| 2138 | astc_results.m_endpoints[7] = ccell_results.m_astc_high_endpoint.m_c[3]; |
| 2139 | |
| 2140 | bool invert = false; |
| 2141 | |
| 2142 | int s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[4]].m_unquant; |
| 2143 | int s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[5]].m_unquant; |
| 2144 | if (s1 < s0) |
| 2145 | { |
| 2146 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 2147 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 2148 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 2149 | std::swap(astc_results.m_endpoints[6], astc_results.m_endpoints[7]); |
| 2150 | invert = true; |
| 2151 | } |
| 2152 | |
| 2153 | for (uint32_t y = 0; y < 4; y++) |
| 2154 | { |
| 2155 | for (uint32_t x = 0; x < 4; x++) |
| 2156 | { |
| 2157 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 2158 | |
| 2159 | if (invert) |
| 2160 | astc_results.m_weights[x + y * 4] = 7 - astc_results.m_weights[x + y * 4]; |
| 2161 | } |
| 2162 | } |
| 2163 | |
| 2164 | assert(total_results < MAX_ENCODE_RESULTS); |
| 2165 | if (total_results < MAX_ENCODE_RESULTS) |
| 2166 | { |
| 2167 | pResults[total_results].m_uastc_mode = 12; |
| 2168 | pResults[total_results].m_common_pattern = 0; |
| 2169 | pResults[total_results].m_astc = astc_results; |
| 2170 | pResults[total_results].m_astc_err = part_err; |
| 2171 | total_results++; |
| 2172 | } |
| 2173 | } |
| 2174 | |
| 2175 | // 13. DualPlane: 1, WeightRange: 0 (2), Subsets: 1, CEM: 12 (RGBA Direct ), EndpointRange: 20 (256) MODE5 |
| 2176 | static void astc_mode13(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 2177 | { |
| 2178 | bc7enc_compress_block_params local_comp_params(comp_params); |
| 2179 | local_comp_params.m_perceptual = false; |
| 2180 | local_comp_params.m_weights[0] = 1; |
| 2181 | local_comp_params.m_weights[1] = 1; |
| 2182 | local_comp_params.m_weights[2] = 1; |
| 2183 | local_comp_params.m_weights[3] = 1; |
| 2184 | |
| 2185 | for (uint32_t rot_comp = 0; rot_comp < 4; rot_comp++) |
| 2186 | { |
| 2187 | const uint32_t weight_range = 0; |
| 2188 | const uint32_t endpoint_range = 20; |
| 2189 | |
| 2190 | color_quad_u8 block_rgb[16]; |
| 2191 | color_quad_u8 block_a[16]; |
| 2192 | for (uint32_t i = 0; i < 16; i++) |
| 2193 | { |
| 2194 | block_rgb[i] = ((color_quad_u8*)&block[0][0])[i]; |
| 2195 | block_a[i] = block_rgb[i]; |
| 2196 | |
| 2197 | uint8_t c = block_a[i].m_c[rot_comp]; |
| 2198 | block_a[i].m_c[0] = c; |
| 2199 | block_a[i].m_c[1] = c; |
| 2200 | block_a[i].m_c[2] = c; |
| 2201 | block_a[i].m_c[3] = 255; |
| 2202 | |
| 2203 | block_rgb[i].m_c[rot_comp] = block_rgb[i].m_c[3]; |
| 2204 | block_rgb[i].m_c[3] = 255; |
| 2205 | } |
| 2206 | |
| 2207 | uint8_t ccell_result_selectors_temp[16]; |
| 2208 | |
| 2209 | color_cell_compressor_params ccell_params_rgb; |
| 2210 | memset(&ccell_params_rgb, 0, sizeof(ccell_params_rgb)); |
| 2211 | |
| 2212 | ccell_params_rgb.m_num_pixels = 16; |
| 2213 | ccell_params_rgb.m_pPixels = block_rgb; |
| 2214 | ccell_params_rgb.m_num_selector_weights = 2; |
| 2215 | ccell_params_rgb.m_pSelector_weights = g_bc7_weights1; |
| 2216 | ccell_params_rgb.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights1x; |
| 2217 | ccell_params_rgb.m_astc_endpoint_range = endpoint_range; |
| 2218 | ccell_params_rgb.m_weights[0] = 1; |
| 2219 | ccell_params_rgb.m_weights[1] = 1; |
| 2220 | ccell_params_rgb.m_weights[2] = 1; |
| 2221 | ccell_params_rgb.m_weights[3] = 1; |
| 2222 | |
| 2223 | color_cell_compressor_results ccell_results_rgb; |
| 2224 | uint8_t ccell_result_selectors_rgb[16]; |
| 2225 | memset(&ccell_results_rgb, 0, sizeof(ccell_results_rgb)); |
| 2226 | ccell_results_rgb.m_pSelectors = &ccell_result_selectors_rgb[0]; |
| 2227 | ccell_results_rgb.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 2228 | |
| 2229 | uint64_t part_err_rgb = color_cell_compression(255, &ccell_params_rgb, &ccell_results_rgb, &local_comp_params); |
| 2230 | |
| 2231 | color_cell_compressor_params ccell_params_a; |
| 2232 | memset(&ccell_params_a, 0, sizeof(ccell_params_a)); |
| 2233 | |
| 2234 | ccell_params_a.m_num_pixels = 16; |
| 2235 | ccell_params_a.m_pPixels = block_a; |
| 2236 | ccell_params_a.m_num_selector_weights = 2; |
| 2237 | ccell_params_a.m_pSelector_weights = g_bc7_weights1; |
| 2238 | ccell_params_a.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights1x; |
| 2239 | ccell_params_a.m_astc_endpoint_range = endpoint_range; |
| 2240 | ccell_params_a.m_weights[0] = 1; |
| 2241 | ccell_params_a.m_weights[1] = 1; |
| 2242 | ccell_params_a.m_weights[2] = 1; |
| 2243 | ccell_params_a.m_weights[3] = 1; |
| 2244 | |
| 2245 | color_cell_compressor_results ccell_results_a; |
| 2246 | uint8_t ccell_result_selectors_a[16]; |
| 2247 | memset(&ccell_results_a, 0, sizeof(ccell_results_a)); |
| 2248 | ccell_results_a.m_pSelectors = &ccell_result_selectors_a[0]; |
| 2249 | ccell_results_a.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 2250 | |
| 2251 | uint64_t part_err_a = color_cell_compression(255, &ccell_params_a, &ccell_results_a, &local_comp_params) / 3; |
| 2252 | |
| 2253 | uint64_t total_err = part_err_rgb + part_err_a; |
| 2254 | |
| 2255 | // ASTC |
| 2256 | astc_block_desc blk; |
| 2257 | memset(&blk, 0, sizeof(blk)); |
| 2258 | |
| 2259 | blk.m_dual_plane = true; |
| 2260 | blk.m_weight_range = weight_range; |
| 2261 | |
| 2262 | blk.m_ccs = rot_comp; |
| 2263 | blk.m_subsets = 1; |
| 2264 | blk.m_partition_seed = 0; |
| 2265 | blk.m_cem = 12; |
| 2266 | |
| 2267 | blk.m_endpoints[0] = (rot_comp == 0 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[0]; |
| 2268 | blk.m_endpoints[1] = (rot_comp == 0 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[0]; |
| 2269 | blk.m_endpoints[2] = (rot_comp == 1 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[1]; |
| 2270 | blk.m_endpoints[3] = (rot_comp == 1 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[1]; |
| 2271 | blk.m_endpoints[4] = (rot_comp == 2 ? ccell_results_a : ccell_results_rgb).m_astc_low_endpoint.m_c[2]; |
| 2272 | blk.m_endpoints[5] = (rot_comp == 2 ? ccell_results_a : ccell_results_rgb).m_astc_high_endpoint.m_c[2]; |
| 2273 | if (rot_comp == 3) |
| 2274 | { |
| 2275 | blk.m_endpoints[6] = ccell_results_a.m_astc_low_endpoint.m_c[0]; |
| 2276 | blk.m_endpoints[7] = ccell_results_a.m_astc_high_endpoint.m_c[0]; |
| 2277 | } |
| 2278 | else |
| 2279 | { |
| 2280 | blk.m_endpoints[6] = ccell_results_rgb.m_astc_low_endpoint.m_c[rot_comp]; |
| 2281 | blk.m_endpoints[7] = ccell_results_rgb.m_astc_high_endpoint.m_c[rot_comp]; |
| 2282 | } |
| 2283 | |
| 2284 | bool invert = false; |
| 2285 | |
| 2286 | int s0 = g_astc_unquant[endpoint_range][blk.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[4]].m_unquant; |
| 2287 | int s1 = g_astc_unquant[endpoint_range][blk.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][blk.m_endpoints[5]].m_unquant; |
| 2288 | if (s1 < s0) |
| 2289 | { |
| 2290 | std::swap(blk.m_endpoints[0], blk.m_endpoints[1]); |
| 2291 | std::swap(blk.m_endpoints[2], blk.m_endpoints[3]); |
| 2292 | std::swap(blk.m_endpoints[4], blk.m_endpoints[5]); |
| 2293 | std::swap(blk.m_endpoints[6], blk.m_endpoints[7]); |
| 2294 | invert = true; |
| 2295 | } |
| 2296 | |
| 2297 | for (uint32_t y = 0; y < 4; y++) |
| 2298 | { |
| 2299 | for (uint32_t x = 0; x < 4; x++) |
| 2300 | { |
| 2301 | uint32_t rgb_index = ccell_result_selectors_rgb[x + y * 4]; |
| 2302 | uint32_t a_index = ccell_result_selectors_a[x + y * 4]; |
| 2303 | |
| 2304 | if (invert) |
| 2305 | { |
| 2306 | rgb_index = 1 - rgb_index; |
| 2307 | a_index = 1 - a_index; |
| 2308 | } |
| 2309 | |
| 2310 | blk.m_weights[(x + y * 4) * 2 + 0] = (uint8_t)rgb_index; |
| 2311 | blk.m_weights[(x + y * 4) * 2 + 1] = (uint8_t)a_index; |
| 2312 | } |
| 2313 | } |
| 2314 | |
| 2315 | assert(total_results < MAX_ENCODE_RESULTS); |
| 2316 | if (total_results < MAX_ENCODE_RESULTS) |
| 2317 | { |
| 2318 | pResults[total_results].m_uastc_mode = 13; |
| 2319 | pResults[total_results].m_common_pattern = 0; |
| 2320 | pResults[total_results].m_astc = blk; |
| 2321 | pResults[total_results].m_astc_err = total_err; |
| 2322 | total_results++; |
| 2323 | } |
| 2324 | } // rot_comp |
| 2325 | } |
| 2326 | |
| 2327 | // MODE14 |
| 2328 | // DualPlane: 0, WeightRange: 2 (4), Subsets: 1, CEM: 12 (RGBA Direct ), EndpointRange: 20 (256) MODE6 |
| 2329 | static void astc_mode14(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 2330 | { |
| 2331 | const uint32_t weight_range = 2; |
| 2332 | const uint32_t endpoint_range = 20; |
| 2333 | |
| 2334 | color_cell_compressor_params ccell_params; |
| 2335 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 2336 | |
| 2337 | ccell_params.m_num_pixels = 16; |
| 2338 | ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 2339 | ccell_params.m_num_selector_weights = 4; |
| 2340 | ccell_params.m_pSelector_weights = g_bc7_weights2; |
| 2341 | ccell_params.m_pSelector_weightsx = (const bc7enc_vec4F*)g_bc7_weights2x; |
| 2342 | ccell_params.m_astc_endpoint_range = endpoint_range; |
| 2343 | ccell_params.m_weights[0] = 1; |
| 2344 | ccell_params.m_weights[1] = 1; |
| 2345 | ccell_params.m_weights[2] = 1; |
| 2346 | ccell_params.m_weights[3] = 1; |
| 2347 | ccell_params.m_has_alpha = true; |
| 2348 | |
| 2349 | color_cell_compressor_results ccell_results; |
| 2350 | uint8_t ccell_result_selectors[16]; |
| 2351 | uint8_t ccell_result_selectors_temp[16]; |
| 2352 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 2353 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 2354 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 2355 | |
| 2356 | uint64_t part_err = color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 2357 | |
| 2358 | // ASTC |
| 2359 | astc_block_desc astc_results; |
| 2360 | memset(&astc_results, 0, sizeof(astc_results)); |
| 2361 | |
| 2362 | astc_results.m_dual_plane = false; |
| 2363 | astc_results.m_weight_range = weight_range; |
| 2364 | |
| 2365 | astc_results.m_ccs = 0; |
| 2366 | astc_results.m_subsets = 1; |
| 2367 | astc_results.m_partition_seed = 0; |
| 2368 | astc_results.m_cem = 12; |
| 2369 | |
| 2370 | astc_results.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 2371 | astc_results.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 2372 | astc_results.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[1]; |
| 2373 | astc_results.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[1]; |
| 2374 | astc_results.m_endpoints[4] = ccell_results.m_astc_low_endpoint.m_c[2]; |
| 2375 | astc_results.m_endpoints[5] = ccell_results.m_astc_high_endpoint.m_c[2]; |
| 2376 | astc_results.m_endpoints[6] = ccell_results.m_astc_low_endpoint.m_c[3]; |
| 2377 | astc_results.m_endpoints[7] = ccell_results.m_astc_high_endpoint.m_c[3]; |
| 2378 | |
| 2379 | bool invert = false; |
| 2380 | |
| 2381 | int s0 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[0]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[2]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[4]].m_unquant; |
| 2382 | int s1 = g_astc_unquant[endpoint_range][astc_results.m_endpoints[1]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[3]].m_unquant + g_astc_unquant[endpoint_range][astc_results.m_endpoints[5]].m_unquant; |
| 2383 | if (s1 < s0) |
| 2384 | { |
| 2385 | std::swap(astc_results.m_endpoints[0], astc_results.m_endpoints[1]); |
| 2386 | std::swap(astc_results.m_endpoints[2], astc_results.m_endpoints[3]); |
| 2387 | std::swap(astc_results.m_endpoints[4], astc_results.m_endpoints[5]); |
| 2388 | std::swap(astc_results.m_endpoints[6], astc_results.m_endpoints[7]); |
| 2389 | invert = true; |
| 2390 | } |
| 2391 | |
| 2392 | for (uint32_t y = 0; y < 4; y++) |
| 2393 | { |
| 2394 | for (uint32_t x = 0; x < 4; x++) |
| 2395 | { |
| 2396 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 2397 | |
| 2398 | if (invert) |
| 2399 | astc_results.m_weights[x + y * 4] = 3 - astc_results.m_weights[x + y * 4]; |
| 2400 | } |
| 2401 | } |
| 2402 | |
| 2403 | assert(total_results < MAX_ENCODE_RESULTS); |
| 2404 | if (total_results < MAX_ENCODE_RESULTS) |
| 2405 | { |
| 2406 | pResults[total_results].m_uastc_mode = 14; |
| 2407 | pResults[total_results].m_common_pattern = 0; |
| 2408 | pResults[total_results].m_astc = astc_results; |
| 2409 | pResults[total_results].m_astc_err = part_err; |
| 2410 | total_results++; |
| 2411 | } |
| 2412 | } |
| 2413 | |
| 2414 | // MODE 15 |
| 2415 | // DualPlane: 0, WeightRange : 8 (16), Subsets : 1, CEM : 4 (LA Direct), EndpointRange : 20 (256) BC7 MODE6 |
| 2416 | static void astc_mode15(const color_rgba block[4][4], uastc_encode_results* pResults, uint32_t& total_results, bc7enc_compress_block_params& comp_params) |
| 2417 | { |
| 2418 | const uint32_t weight_range = 8; |
| 2419 | const uint32_t endpoint_range = 20; |
| 2420 | |
| 2421 | color_cell_compressor_params ccell_params; |
| 2422 | memset(&ccell_params, 0, sizeof(ccell_params)); |
| 2423 | |
| 2424 | color_rgba temp_block[16]; |
| 2425 | for (uint32_t i = 0; i < 16; i++) |
| 2426 | { |
| 2427 | const uint32_t l = ((const color_rgba*)block)[i].r; |
| 2428 | const uint32_t a = ((const color_rgba*)block)[i].a; |
| 2429 | |
| 2430 | // Use (l,0,0,a) not (l,l,l,a) so both components are treated equally. |
| 2431 | temp_block[i].set_noclamp_rgba(l, 0, 0, a); |
| 2432 | } |
| 2433 | |
| 2434 | ccell_params.m_num_pixels = 16; |
| 2435 | //ccell_params.m_pPixels = (color_quad_u8*)&block[0][0]; |
| 2436 | ccell_params.m_pPixels = (color_quad_u8*)temp_block; |
| 2437 | ccell_params.m_num_selector_weights = 16; |
| 2438 | ccell_params.m_pSelector_weights = g_astc_weights4; |
| 2439 | ccell_params.m_pSelector_weightsx = (const bc7enc_vec4F*)g_astc_weights4x; |
| 2440 | ccell_params.m_astc_endpoint_range = endpoint_range; |
| 2441 | ccell_params.m_weights[0] = 1; |
| 2442 | ccell_params.m_weights[1] = 1; |
| 2443 | ccell_params.m_weights[2] = 1; |
| 2444 | ccell_params.m_weights[3] = 1; |
| 2445 | ccell_params.m_has_alpha = true; |
| 2446 | |
| 2447 | color_cell_compressor_results ccell_results; |
| 2448 | uint8_t ccell_result_selectors[16]; |
| 2449 | uint8_t ccell_result_selectors_temp[16]; |
| 2450 | memset(&ccell_results, 0, sizeof(ccell_results)); |
| 2451 | ccell_results.m_pSelectors = &ccell_result_selectors[0]; |
| 2452 | ccell_results.m_pSelectors_temp = &ccell_result_selectors_temp[0]; |
| 2453 | |
| 2454 | color_cell_compression(255, &ccell_params, &ccell_results, &comp_params); |
| 2455 | |
| 2456 | // ASTC |
| 2457 | astc_block_desc astc_results; |
| 2458 | memset(&astc_results, 0, sizeof(astc_results)); |
| 2459 | |
| 2460 | astc_results.m_dual_plane = false; |
| 2461 | astc_results.m_weight_range = weight_range; |
| 2462 | |
| 2463 | astc_results.m_ccs = 0; |
| 2464 | astc_results.m_subsets = 1; |
| 2465 | astc_results.m_partition_seed = 0; |
| 2466 | astc_results.m_cem = 4; |
| 2467 | |
| 2468 | astc_results.m_endpoints[0] = ccell_results.m_astc_low_endpoint.m_c[0]; |
| 2469 | astc_results.m_endpoints[1] = ccell_results.m_astc_high_endpoint.m_c[0]; |
| 2470 | |
| 2471 | astc_results.m_endpoints[2] = ccell_results.m_astc_low_endpoint.m_c[3]; |
| 2472 | astc_results.m_endpoints[3] = ccell_results.m_astc_high_endpoint.m_c[3]; |
| 2473 | |
| 2474 | for (uint32_t y = 0; y < 4; y++) |
| 2475 | for (uint32_t x = 0; x < 4; x++) |
| 2476 | astc_results.m_weights[x + y * 4] = ccell_result_selectors[x + y * 4]; |
| 2477 | |
| 2478 | color_rgba colors[16]; |
| 2479 | for (uint32_t c = 0; c < 4; c++) |
| 2480 | { |
| 2481 | colors[0].m_comps[c] = g_astc_unquant[endpoint_range][ccell_results.m_astc_low_endpoint.m_c[(c < 3) ? 0 : 3]].m_unquant; |
| 2482 | colors[15].m_comps[c] = g_astc_unquant[endpoint_range][ccell_results.m_astc_high_endpoint.m_c[(c < 3) ? 0 : 3]].m_unquant; |
| 2483 | } |
| 2484 | |
| 2485 | for (uint32_t i = 1; i < 16 - 1; i++) |
| 2486 | for (uint32_t c = 0; c < 4; c++) |
| 2487 | colors[i].m_comps[c] = (uint8_t)astc_interpolate(colors[0].m_comps[c], colors[15].m_comps[c], g_astc_weights4[i], false); |
| 2488 | |
| 2489 | uint64_t total_err = 0; |
| 2490 | for (uint32_t p = 0; p < 16; p++) |
| 2491 | total_err += color_distance_la(((const color_rgba*)block)[p], colors[ccell_result_selectors[p]]); |
| 2492 | |
| 2493 | assert(total_results < MAX_ENCODE_RESULTS); |
| 2494 | if (total_results < MAX_ENCODE_RESULTS) |
| 2495 | { |
| 2496 | pResults[total_results].m_uastc_mode = 15; |
| 2497 | pResults[total_results].m_common_pattern = 0; |
| 2498 | pResults[total_results].m_astc = astc_results; |
| 2499 | pResults[total_results].m_astc_err = total_err; |
| 2500 | total_results++; |
| 2501 | } |
| 2502 | } |
| 2503 | |
| 2504 | static void compute_block_error(const color_rgba block[4][4], const color_rgba decoded_block[4][4], uint64_t &total_rgb_err, uint64_t &total_rgba_err, uint64_t &total_la_err) |
| 2505 | { |
| 2506 | uint64_t total_err_r = 0, total_err_g = 0, total_err_b = 0, total_err_a = 0; |
| 2507 | |
| 2508 | for (uint32_t y = 0; y < 4; y++) |
| 2509 | { |
| 2510 | for (uint32_t x = 0; x < 4; x++) |
| 2511 | { |
| 2512 | const int dr = (int)block[y][x].m_comps[0] - (int)decoded_block[y][x].m_comps[0]; |
| 2513 | const int dg = (int)block[y][x].m_comps[1] - (int)decoded_block[y][x].m_comps[1]; |
| 2514 | const int db = (int)block[y][x].m_comps[2] - (int)decoded_block[y][x].m_comps[2]; |
| 2515 | const int da = (int)block[y][x].m_comps[3] - (int)decoded_block[y][x].m_comps[3]; |
| 2516 | |
| 2517 | total_err_r += dr * dr; |
| 2518 | total_err_g += dg * dg; |
| 2519 | total_err_b += db * db; |
| 2520 | total_err_a += da * da; |
| 2521 | } |
| 2522 | } |
| 2523 | |
| 2524 | total_la_err = total_err_r + total_err_a; |
| 2525 | total_rgb_err = total_err_r + total_err_g + total_err_b; |
| 2526 | total_rgba_err = total_rgb_err + total_err_a; |
| 2527 | } |
| 2528 | |
| 2529 | static void compute_bc1_hints(bool &bc1_hint0, bool &bc1_hint1, const uastc_encode_results &best_results, const color_rgba block[4][4], const color_rgba decoded_uastc_block[4][4]) |
| 2530 | { |
| 2531 | const uint32_t best_mode = best_results.m_uastc_mode; |
| 2532 | const bool perceptual = false; |
| 2533 | |
| 2534 | bc1_hint0 = false; |
| 2535 | bc1_hint1 = false; |
| 2536 | |
| 2537 | if (best_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 2538 | return; |
| 2539 | |
| 2540 | if (!g_uastc_mode_has_bc1_hint0[best_mode] && !g_uastc_mode_has_bc1_hint1[best_mode]) |
| 2541 | return; |
| 2542 | |
| 2543 | color_rgba tblock_bc1[4][4]; |
| 2544 | dxt1_block tbc1_block[8]; |
| 2545 | basist::encode_bc1(tbc1_block, (const uint8_t*)&decoded_uastc_block[0][0], 0); |
| 2546 | unpack_block(texture_format::cBC1, tbc1_block, &tblock_bc1[0][0]); |
| 2547 | |
| 2548 | color_rgba tblock_hint0_bc1[4][4]; |
| 2549 | color_rgba tblock_hint1_bc1[4][4]; |
| 2550 | |
| 2551 | etc_block etc1_blk; |
| 2552 | memset(&etc1_blk, 0, sizeof(etc1_blk)); |
| 2553 | |
| 2554 | eac_a8_block etc2_blk; |
| 2555 | memset(&etc2_blk, 0, sizeof(etc2_blk)); |
| 2556 | etc2_blk.m_multiplier = 1; |
| 2557 | |
| 2558 | // Pack to UASTC, then unpack, because the endpoints may be swapped. |
| 2559 | |
| 2560 | uastc_block temp_ublock; |
| 2561 | pack_uastc(temp_ublock, best_results, etc1_blk, 0, etc2_blk, false, false); |
| 2562 | |
| 2563 | unpacked_uastc_block temp_ublock_unpacked; |
| 2564 | unpack_uastc(temp_ublock, temp_ublock_unpacked, false); |
| 2565 | |
| 2566 | unpacked_uastc_block ublock; |
| 2567 | memset(&ublock, 0, sizeof(ublock)); |
| 2568 | ublock.m_mode = best_results.m_uastc_mode; |
| 2569 | ublock.m_common_pattern = best_results.m_common_pattern; |
| 2570 | ublock.m_astc = temp_ublock_unpacked.m_astc; |
| 2571 | |
| 2572 | dxt1_block b; |
| 2573 | |
| 2574 | // HINT1 |
| 2575 | if (!g_uastc_mode_has_bc1_hint1[best_mode]) |
| 2576 | { |
| 2577 | memset(tblock_hint1_bc1, 0, sizeof(tblock_hint1_bc1)); |
| 2578 | } |
| 2579 | else |
| 2580 | { |
| 2581 | transcode_uastc_to_bc1_hint1(ublock, (color32 (*)[4]) decoded_uastc_block, &b, false); |
| 2582 | |
| 2583 | unpack_block(texture_format::cBC1, &b, &tblock_hint1_bc1[0][0]); |
| 2584 | } |
| 2585 | |
| 2586 | // HINT0 |
| 2587 | if (!g_uastc_mode_has_bc1_hint0[best_mode]) |
| 2588 | { |
| 2589 | memset(tblock_hint0_bc1, 0, sizeof(tblock_hint0_bc1)); |
| 2590 | } |
| 2591 | else |
| 2592 | { |
| 2593 | transcode_uastc_to_bc1_hint0(ublock, &b); |
| 2594 | |
| 2595 | unpack_block(texture_format::cBC1, &b, &tblock_hint0_bc1[0][0]); |
| 2596 | } |
| 2597 | |
| 2598 | // Compute block errors |
| 2599 | uint64_t total_t_err = 0, total_hint0_err = 0, total_hint1_err = 0; |
| 2600 | for (uint32_t y = 0; y < 4; y++) |
| 2601 | { |
| 2602 | for (uint32_t x = 0; x < 4; x++) |
| 2603 | { |
| 2604 | total_t_err += color_distance(perceptual, block[y][x], tblock_bc1[y][x], false); |
| 2605 | total_hint0_err += color_distance(perceptual, block[y][x], tblock_hint0_bc1[y][x], false); |
| 2606 | total_hint1_err += color_distance(perceptual, block[y][x], tblock_hint1_bc1[y][x], false); |
| 2607 | } |
| 2608 | } |
| 2609 | |
| 2610 | const float t_err = sqrtf((float)total_t_err); |
| 2611 | const float t_err_hint0 = sqrtf((float)total_hint0_err); |
| 2612 | const float t_err_hint1 = sqrtf((float)total_hint1_err); |
| 2613 | |
| 2614 | const float err_thresh0 = 1.075f; |
| 2615 | const float err_thresh1 = 1.075f; |
| 2616 | |
| 2617 | if ((g_uastc_mode_has_bc1_hint0[best_mode]) && (t_err_hint0 <= t_err * err_thresh0)) |
| 2618 | bc1_hint0 = true; |
| 2619 | |
| 2620 | if ((g_uastc_mode_has_bc1_hint1[best_mode]) && (t_err_hint1 <= t_err * err_thresh1)) |
| 2621 | bc1_hint1 = true; |
| 2622 | } |
| 2623 | |
| 2624 | struct ycbcr |
| 2625 | { |
| 2626 | int32_t m_y; |
| 2627 | int32_t m_cb; |
| 2628 | int32_t m_cr; |
| 2629 | }; |
| 2630 | |
| 2631 | static inline void rgb_to_y_cb_cr(const color_rgba& c, ycbcr& dst) |
| 2632 | { |
| 2633 | const int y = c.r * 54 + c.g * 183 + c.b * 19; |
| 2634 | dst.m_y = y; |
| 2635 | dst.m_cb = (c.b << 8) - y; |
| 2636 | dst.m_cr = (c.r << 8) - y; |
| 2637 | } |
| 2638 | |
| 2639 | static inline uint64_t color_diff(const ycbcr& a, const ycbcr& b) |
| 2640 | { |
| 2641 | const int y_delta = a.m_y - b.m_y; |
| 2642 | const int cb_delta = a.m_cb - b.m_cb; |
| 2643 | const int cr_delta = a.m_cr - b.m_cr; |
| 2644 | return ((int64_t)y_delta * y_delta * 4) + ((int64_t)cr_delta * cr_delta) + ((int64_t)cb_delta * cb_delta); |
| 2645 | } |
| 2646 | |
| 2647 | static inline int gray_distance2(const color_rgba& c, int r, int g, int b) |
| 2648 | { |
| 2649 | int gray_dist = (((int)c[0] - r) + ((int)c[1] - g) + ((int)c[2] - b) + 1) / 3; |
| 2650 | |
| 2651 | int gray_point_r = clamp255(r + gray_dist); |
| 2652 | int gray_point_g = clamp255(g + gray_dist); |
| 2653 | int gray_point_b = clamp255(b + gray_dist); |
| 2654 | |
| 2655 | int dist_to_gray_point_r = c[0] - gray_point_r; |
| 2656 | int dist_to_gray_point_g = c[1] - gray_point_g; |
| 2657 | int dist_to_gray_point_b = c[2] - gray_point_b; |
| 2658 | |
| 2659 | return (dist_to_gray_point_r * dist_to_gray_point_r) + (dist_to_gray_point_g * dist_to_gray_point_g) + (dist_to_gray_point_b * dist_to_gray_point_b); |
| 2660 | } |
| 2661 | |
| 2662 | static bool pack_etc1_estimate_flipped(const color_rgba* pSrc_pixels) |
| 2663 | { |
| 2664 | int sums[3][2][2]; |
| 2665 | |
| 2666 | #define GET_XY(x, y, c) pSrc_pixels[(x) + ((y) * 4)][c] |
| 2667 | |
| 2668 | for (uint32_t c = 0; c < 3; c++) |
| 2669 | { |
| 2670 | sums[c][0][0] = GET_XY(0, 0, c) + GET_XY(0, 1, c) + GET_XY(1, 0, c) + GET_XY(1, 1, c); |
| 2671 | sums[c][1][0] = GET_XY(2, 0, c) + GET_XY(2, 1, c) + GET_XY(3, 0, c) + GET_XY(3, 1, c); |
| 2672 | sums[c][0][1] = GET_XY(0, 2, c) + GET_XY(0, 3, c) + GET_XY(1, 2, c) + GET_XY(1, 3, c); |
| 2673 | sums[c][1][1] = GET_XY(2, 2, c) + GET_XY(2, 3, c) + GET_XY(3, 2, c) + GET_XY(3, 3, c); |
| 2674 | } |
| 2675 | |
| 2676 | int upper_avg[3], lower_avg[3], left_avg[3], right_avg[3]; |
| 2677 | for (uint32_t c = 0; c < 3; c++) |
| 2678 | { |
| 2679 | upper_avg[c] = (sums[c][0][0] + sums[c][1][0] + 4) / 8; |
| 2680 | lower_avg[c] = (sums[c][0][1] + sums[c][1][1] + 4) / 8; |
| 2681 | left_avg[c] = (sums[c][0][0] + sums[c][0][1] + 4) / 8; |
| 2682 | right_avg[c] = (sums[c][1][0] + sums[c][1][1] + 4) / 8; |
| 2683 | } |
| 2684 | |
| 2685 | #undef GET_XY |
| 2686 | #define GET_XY(x, y, a) gray_distance2(pSrc_pixels[(x) + ((y) * 4)], a[0], a[1], a[2]) |
| 2687 | |
| 2688 | int upper_gray_dist = 0, lower_gray_dist = 0, left_gray_dist = 0, right_gray_dist = 0; |
| 2689 | for (uint32_t i = 0; i < 4; i++) |
| 2690 | { |
| 2691 | for (uint32_t j = 0; j < 2; j++) |
| 2692 | { |
| 2693 | upper_gray_dist += GET_XY(i, j, upper_avg); |
| 2694 | lower_gray_dist += GET_XY(i, 2 + j, lower_avg); |
| 2695 | left_gray_dist += GET_XY(j, i, left_avg); |
| 2696 | right_gray_dist += GET_XY(2 + j, i, right_avg); |
| 2697 | } |
| 2698 | } |
| 2699 | |
| 2700 | #undef GET_XY |
| 2701 | |
| 2702 | int upper_lower_sum = upper_gray_dist + lower_gray_dist; |
| 2703 | int left_right_sum = left_gray_dist + right_gray_dist; |
| 2704 | |
| 2705 | return upper_lower_sum < left_right_sum; |
| 2706 | } |
| 2707 | |
| 2708 | static void compute_etc1_hints(etc_block& best_etc1_blk, uint32_t& best_etc1_bias, const uastc_encode_results& best_results, const color_rgba block[4][4], const color_rgba decoded_uastc_block[4][4], int level, uint32_t flags) |
| 2709 | { |
| 2710 | best_etc1_bias = 0; |
| 2711 | |
| 2712 | if (best_results.m_uastc_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 2713 | { |
| 2714 | pack_etc1_block_solid_color(best_etc1_blk, &best_results.m_solid_color.m_comps[0]); |
| 2715 | return; |
| 2716 | } |
| 2717 | |
| 2718 | const bool faster_etc1 = (flags & cPackUASTCETC1FasterHints) != 0; |
| 2719 | const bool fastest_etc1 = (flags & cPackUASTCETC1FastestHints) != 0; |
| 2720 | |
| 2721 | const bool has_bias = g_uastc_mode_has_etc1_bias[best_results.m_uastc_mode]; |
| 2722 | |
| 2723 | // 0 should be at the top, but we need 13 first because it represents bias (0,0,0). |
| 2724 | const uint8_t s_sorted_bias_modes[32] = { 13, 0, 22, 29, 27, 12, 26, 9, 30, 31, 8, 10, 25, 2, 23, 5, 15, 7, 3, 11, 6, 17, 28, 18, 1, 19, 20, 21, 24, 4, 14, 16 }; |
| 2725 | |
| 2726 | uint32_t last_bias = 1; |
| 2727 | bool use_faster_bias_mode_table = false; |
| 2728 | const bool flip_estimate = (level <= cPackUASTCLevelFaster) || (faster_etc1) || (fastest_etc1); |
| 2729 | if (has_bias) |
| 2730 | { |
| 2731 | switch (level) |
| 2732 | { |
| 2733 | case cPackUASTCLevelFastest: |
| 2734 | { |
| 2735 | last_bias = fastest_etc1 ? 1 : (faster_etc1 ? 1 : 2); |
| 2736 | use_faster_bias_mode_table = true; |
| 2737 | break; |
| 2738 | } |
| 2739 | case cPackUASTCLevelFaster: |
| 2740 | { |
| 2741 | last_bias = fastest_etc1 ? 1 : (faster_etc1 ? 3 : 5); |
| 2742 | use_faster_bias_mode_table = true; |
| 2743 | break; |
| 2744 | } |
| 2745 | case cPackUASTCLevelDefault: |
| 2746 | { |
| 2747 | last_bias = fastest_etc1 ? 1 : (faster_etc1 ? 10 : 20); |
| 2748 | use_faster_bias_mode_table = true; |
| 2749 | break; |
| 2750 | } |
| 2751 | case cPackUASTCLevelSlower: |
| 2752 | { |
| 2753 | last_bias = fastest_etc1 ? 1 : (faster_etc1 ? 16 : 32); |
| 2754 | use_faster_bias_mode_table = true; |
| 2755 | break; |
| 2756 | } |
| 2757 | default: |
| 2758 | { |
| 2759 | last_bias = 32; |
| 2760 | break; |
| 2761 | } |
| 2762 | } |
| 2763 | } |
| 2764 | |
| 2765 | memset(&best_etc1_blk, 0, sizeof(best_etc1_blk)); |
| 2766 | uint64_t best_err = UINT64_MAX; |
| 2767 | |
| 2768 | etc_block trial_block; |
| 2769 | memset(&trial_block, 0, sizeof(trial_block)); |
| 2770 | |
| 2771 | ycbcr block_ycbcr[4][4], decoded_uastc_block_ycbcr[4][4]; |
| 2772 | for (uint32_t y = 0; y < 4; y++) |
| 2773 | { |
| 2774 | for (uint32_t x = 0; x < 4; x++) |
| 2775 | { |
| 2776 | rgb_to_y_cb_cr(block[y][x], block_ycbcr[y][x]); |
| 2777 | rgb_to_y_cb_cr(decoded_uastc_block[y][x], decoded_uastc_block_ycbcr[y][x]); |
| 2778 | } |
| 2779 | } |
| 2780 | |
| 2781 | uint32_t first_flip = 0, last_flip = 2; |
| 2782 | uint32_t first_individ = 0, last_individ = 2; |
| 2783 | |
| 2784 | if (flags & cPackUASTCETC1DisableFlipAndIndividual) |
| 2785 | { |
| 2786 | last_flip = 1; |
| 2787 | last_individ = 1; |
| 2788 | } |
| 2789 | else if (flip_estimate) |
| 2790 | { |
| 2791 | if (pack_etc1_estimate_flipped(&decoded_uastc_block[0][0])) |
| 2792 | first_flip = 1; |
| 2793 | last_flip = first_flip + 1; |
| 2794 | } |
| 2795 | |
| 2796 | for (uint32_t flip = first_flip; flip < last_flip; flip++) |
| 2797 | { |
| 2798 | trial_block.set_flip_bit(flip != 0); |
| 2799 | |
| 2800 | for (uint32_t individ = first_individ; individ < last_individ; individ++) |
| 2801 | { |
| 2802 | const uint32_t mul = individ ? 15 : 31; |
| 2803 | |
| 2804 | trial_block.set_diff_bit(individ == 0); |
| 2805 | |
| 2806 | color_rgba unbiased_block_colors[2]; |
| 2807 | |
| 2808 | int min_r[2] = { 255, 255 }, min_g[2] = { 255, 255 }, min_b[2] = { 255, 255 }, max_r[2] = { 0, 0 }, max_g[2] = { 0, 0 }, max_b[2] = { 0, 0 }; |
| 2809 | |
| 2810 | for (uint32_t subset = 0; subset < 2; subset++) |
| 2811 | { |
| 2812 | uint32_t avg_color[3]; |
| 2813 | memset(avg_color, 0, sizeof(avg_color)); |
| 2814 | |
| 2815 | for (uint32_t j = 0; j < 8; j++) |
| 2816 | { |
| 2817 | const etc_coord2 &c = g_etc1_pixel_coords[flip][subset][j]; |
| 2818 | const color_rgba& p = decoded_uastc_block[c.m_y][c.m_x]; |
| 2819 | |
| 2820 | avg_color[0] += p.r; |
| 2821 | avg_color[1] += p.g; |
| 2822 | avg_color[2] += p.b; |
| 2823 | |
| 2824 | min_r[subset] = basisu::minimum<uint32_t>(min_r[subset], p.r); |
| 2825 | min_g[subset] = basisu::minimum<uint32_t>(min_g[subset], p.g); |
| 2826 | min_b[subset] = basisu::minimum<uint32_t>(min_b[subset], p.b); |
| 2827 | |
| 2828 | max_r[subset] = basisu::maximum<uint32_t>(max_r[subset], p.r); |
| 2829 | max_g[subset] = basisu::maximum<uint32_t>(max_g[subset], p.g); |
| 2830 | max_b[subset] = basisu::maximum<uint32_t>(max_b[subset], p.b); |
| 2831 | } // j |
| 2832 | |
| 2833 | unbiased_block_colors[subset][0] = (uint8_t)((avg_color[0] * mul + 1020) / (8 * 255)); |
| 2834 | unbiased_block_colors[subset][1] = (uint8_t)((avg_color[1] * mul + 1020) / (8 * 255)); |
| 2835 | unbiased_block_colors[subset][2] = (uint8_t)((avg_color[2] * mul + 1020) / (8 * 255)); |
| 2836 | unbiased_block_colors[subset][3] = 0; |
| 2837 | |
| 2838 | } // subset |
| 2839 | |
| 2840 | for (uint32_t bias_iter = 0; bias_iter < last_bias; bias_iter++) |
| 2841 | { |
| 2842 | const uint32_t bias = use_faster_bias_mode_table ? s_sorted_bias_modes[bias_iter] : bias_iter; |
| 2843 | |
| 2844 | color_rgba block_colors[2]; |
| 2845 | for (uint32_t subset = 0; subset < 2; subset++) |
| 2846 | block_colors[subset] = has_bias ? apply_etc1_bias((color32&)unbiased_block_colors[subset], bias, mul, subset) : unbiased_block_colors[subset]; |
| 2847 | |
| 2848 | if (individ) |
| 2849 | trial_block.set_block_color4(block_colors[0], block_colors[1]); |
| 2850 | else |
| 2851 | trial_block.set_block_color5_clamp(block_colors[0], block_colors[1]); |
| 2852 | |
| 2853 | uint32_t range[2]; |
| 2854 | for (uint32_t subset = 0; subset < 2; subset++) |
| 2855 | { |
| 2856 | const color_rgba base_c(trial_block.get_block_color(subset, true)); |
| 2857 | |
| 2858 | const int pos_r = iabs(max_r[subset] - base_c.r); |
| 2859 | const int neg_r = iabs(base_c.r - min_r[subset]); |
| 2860 | |
| 2861 | const int pos_g = iabs(max_g[subset] - base_c.g); |
| 2862 | const int neg_g = iabs(base_c.g - min_g[subset]); |
| 2863 | |
| 2864 | const int pos_b = iabs(max_b[subset] - base_c.b); |
| 2865 | const int neg_b = iabs(base_c.b - min_b[subset]); |
| 2866 | |
| 2867 | range[subset] = maximum(maximum(pos_r, neg_r, pos_g, neg_g), pos_b, neg_b); |
| 2868 | } |
| 2869 | |
| 2870 | uint32_t best_inten_table[2] = { 0, 0 }; |
| 2871 | |
| 2872 | for (uint32_t subset = 0; subset < 2; subset++) |
| 2873 | { |
| 2874 | uint64_t best_subset_err = UINT64_MAX; |
| 2875 | |
| 2876 | const uint32_t inten_table_limit = (level == cPackUASTCLevelVerySlow) ? 8 : ((range[subset] > 51) ? 8 : (range[subset] >= 7 ? 4 : 2)); |
| 2877 | |
| 2878 | for (uint32_t inten_table = 0; inten_table < inten_table_limit; inten_table++) |
| 2879 | { |
| 2880 | trial_block.set_inten_table(subset, inten_table); |
| 2881 | |
| 2882 | color_rgba color_table[4]; |
| 2883 | trial_block.get_block_colors(color_table, subset); |
| 2884 | |
| 2885 | ycbcr color_table_ycbcr[4]; |
| 2886 | for (uint32_t i = 0; i < 4; i++) |
| 2887 | rgb_to_y_cb_cr(color_table[i], color_table_ycbcr[i]); |
| 2888 | |
| 2889 | uint64_t total_error = 0; |
| 2890 | if (flip) |
| 2891 | { |
| 2892 | for (uint32_t y = 0; y < 2; y++) |
| 2893 | { |
| 2894 | { |
| 2895 | const ycbcr& c = decoded_uastc_block_ycbcr[subset * 2 + y][0]; |
| 2896 | total_error += minimum(color_diff(color_table_ycbcr[0], c), color_diff(color_table_ycbcr[1], c), color_diff(color_table_ycbcr[2], c), color_diff(color_table_ycbcr[3], c)); |
| 2897 | } |
| 2898 | { |
| 2899 | const ycbcr& c = decoded_uastc_block_ycbcr[subset * 2 + y][1]; |
| 2900 | total_error += minimum(color_diff(color_table_ycbcr[0], c), color_diff(color_table_ycbcr[1], c), color_diff(color_table_ycbcr[2], c), color_diff(color_table_ycbcr[3], c)); |
| 2901 | } |
| 2902 | { |
| 2903 | const ycbcr& c = decoded_uastc_block_ycbcr[subset * 2 + y][2]; |
| 2904 | total_error += minimum(color_diff(color_table_ycbcr[0], c), color_diff(color_table_ycbcr[1], c), color_diff(color_table_ycbcr[2], c), color_diff(color_table_ycbcr[3], c)); |
| 2905 | } |
| 2906 | { |
| 2907 | const ycbcr& c = decoded_uastc_block_ycbcr[subset * 2 + y][3]; |
| 2908 | total_error += minimum(color_diff(color_table_ycbcr[0], c), color_diff(color_table_ycbcr[1], c), color_diff(color_table_ycbcr[2], c), color_diff(color_table_ycbcr[3], c)); |
| 2909 | } |
| 2910 | if (total_error >= best_subset_err) |
| 2911 | break; |
| 2912 | } |
| 2913 | } |
| 2914 | else |
| 2915 | { |
| 2916 | for (uint32_t y = 0; y < 4; y++) |
| 2917 | { |
| 2918 | { |
| 2919 | const ycbcr& c = decoded_uastc_block_ycbcr[y][subset * 2 + 0]; |
| 2920 | total_error += minimum(color_diff(color_table_ycbcr[0], c), color_diff(color_table_ycbcr[1], c), color_diff(color_table_ycbcr[2], c), color_diff(color_table_ycbcr[3], c)); |
| 2921 | } |
| 2922 | { |
| 2923 | const ycbcr& c = decoded_uastc_block_ycbcr[y][subset * 2 + 1]; |
| 2924 | total_error += minimum(color_diff(color_table_ycbcr[0], c), color_diff(color_table_ycbcr[1], c), color_diff(color_table_ycbcr[2], c), color_diff(color_table_ycbcr[3], c)); |
| 2925 | } |
| 2926 | } |
| 2927 | if (total_error >= best_subset_err) |
| 2928 | break; |
| 2929 | } |
| 2930 | |
| 2931 | if (total_error < best_subset_err) |
| 2932 | { |
| 2933 | best_subset_err = total_error; |
| 2934 | best_inten_table[subset] = inten_table; |
| 2935 | } |
| 2936 | |
| 2937 | } // inten_table |
| 2938 | |
| 2939 | } // subset |
| 2940 | |
| 2941 | trial_block.set_inten_table(0, best_inten_table[0]); |
| 2942 | trial_block.set_inten_table(1, best_inten_table[1]); |
| 2943 | |
| 2944 | // Compute error against the ORIGINAL block. |
| 2945 | uint64_t err = 0; |
| 2946 | |
| 2947 | for (uint32_t subset = 0; subset < 2; subset++) |
| 2948 | { |
| 2949 | color_rgba color_table[4]; |
| 2950 | trial_block.get_block_colors(color_table, subset); |
| 2951 | |
| 2952 | ycbcr color_table_ycbcr[4]; |
| 2953 | for (uint32_t i = 0; i < 4; i++) |
| 2954 | rgb_to_y_cb_cr(color_table[i], color_table_ycbcr[i]); |
| 2955 | |
| 2956 | if (flip) |
| 2957 | { |
| 2958 | for (uint32_t y = 0; y < 2; y++) |
| 2959 | { |
| 2960 | for (uint32_t x = 0; x < 4; x++) |
| 2961 | { |
| 2962 | const ycbcr& c = decoded_uastc_block_ycbcr[subset * 2 + y][x]; |
| 2963 | const uint64_t best_index_err = minimum(color_diff(color_table_ycbcr[0], c) << 2, (color_diff(color_table_ycbcr[1], c) << 2) + 1, (color_diff(color_table_ycbcr[2], c) << 2) + 2, (color_diff(color_table_ycbcr[3], c) << 2) + 3); |
| 2964 | |
| 2965 | const uint32_t best_index = (uint32_t)best_index_err & 3; |
| 2966 | err += color_diff(block_ycbcr[subset * 2 + y][x], color_table_ycbcr[best_index]); |
| 2967 | } |
| 2968 | if (err >= best_err) |
| 2969 | break; |
| 2970 | } |
| 2971 | } |
| 2972 | else |
| 2973 | { |
| 2974 | for (uint32_t y = 0; y < 4; y++) |
| 2975 | { |
| 2976 | for (uint32_t x = 0; x < 2; x++) |
| 2977 | { |
| 2978 | const ycbcr& c = decoded_uastc_block_ycbcr[y][subset * 2 + x]; |
| 2979 | const uint64_t best_index_err = minimum(color_diff(color_table_ycbcr[0], c) << 2, (color_diff(color_table_ycbcr[1], c) << 2) + 1, (color_diff(color_table_ycbcr[2], c) << 2) + 2, (color_diff(color_table_ycbcr[3], c) << 2) + 3); |
| 2980 | |
| 2981 | const uint32_t best_index = (uint32_t)best_index_err & 3; |
| 2982 | err += color_diff(block_ycbcr[y][subset * 2 + x], color_table_ycbcr[best_index]); |
| 2983 | } |
| 2984 | if (err >= best_err) |
| 2985 | break; |
| 2986 | } |
| 2987 | } |
| 2988 | |
| 2989 | } // subset |
| 2990 | |
| 2991 | if (err < best_err) |
| 2992 | { |
| 2993 | best_err = err; |
| 2994 | |
| 2995 | best_etc1_blk = trial_block; |
| 2996 | best_etc1_bias = bias; |
| 2997 | } |
| 2998 | |
| 2999 | } // bias_iter |
| 3000 | |
| 3001 | } // individ |
| 3002 | |
| 3003 | } // flip |
| 3004 | } |
| 3005 | |
| 3006 | struct uastc_pack_eac_a8_results |
| 3007 | { |
| 3008 | uint32_t m_base; |
| 3009 | uint32_t m_table; |
| 3010 | uint32_t m_multiplier; |
| 3011 | }; |
| 3012 | |
| 3013 | static uint64_t uastc_pack_eac_a8(uastc_pack_eac_a8_results& results, const uint8_t* pPixels, uint32_t num_pixels, uint32_t base_search_rad, uint32_t mul_search_rad, uint32_t table_mask) |
| 3014 | { |
| 3015 | assert(num_pixels <= 16); |
| 3016 | |
| 3017 | uint32_t min_alpha = 255, max_alpha = 0; |
| 3018 | for (uint32_t i = 0; i < num_pixels; i++) |
| 3019 | { |
| 3020 | const uint32_t a = pPixels[i]; |
| 3021 | if (a < min_alpha) min_alpha = a; |
| 3022 | if (a > max_alpha) max_alpha = a; |
| 3023 | } |
| 3024 | |
| 3025 | if (min_alpha == max_alpha) |
| 3026 | { |
| 3027 | results.m_base = min_alpha; |
| 3028 | results.m_table = 13; |
| 3029 | results.m_multiplier = 1; |
| 3030 | return 0; |
| 3031 | } |
| 3032 | |
| 3033 | const uint32_t alpha_range = max_alpha - min_alpha; |
| 3034 | |
| 3035 | uint64_t best_err = UINT64_MAX; |
| 3036 | |
| 3037 | for (uint32_t table = 0; table < 16; table++) |
| 3038 | { |
| 3039 | if ((table_mask & (1U << table)) == 0) |
| 3040 | continue; |
| 3041 | |
| 3042 | const float range = (float)(g_etc2_eac_tables[table][ETC2_EAC_MAX_VALUE_SELECTOR] - g_etc2_eac_tables[table][ETC2_EAC_MIN_VALUE_SELECTOR]); |
| 3043 | const int center = (int)roundf(lerp((float)min_alpha, (float)max_alpha, (float)(0 - g_etc2_eac_tables[table][ETC2_EAC_MIN_VALUE_SELECTOR]) / range)); |
| 3044 | |
| 3045 | const int base_min = clamp255(center - base_search_rad); |
| 3046 | const int base_max = clamp255(center + base_search_rad); |
| 3047 | |
| 3048 | const int mul = (int)roundf(alpha_range / range); |
| 3049 | const int mul_low = clamp<int>(mul - mul_search_rad, 1, 15); |
| 3050 | const int mul_high = clamp<int>(mul + mul_search_rad, 1, 15); |
| 3051 | |
| 3052 | for (int base = base_min; base <= base_max; base++) |
| 3053 | { |
| 3054 | for (int multiplier = mul_low; multiplier <= mul_high; multiplier++) |
| 3055 | { |
| 3056 | uint64_t total_err = 0; |
| 3057 | |
| 3058 | for (uint32_t i = 0; i < num_pixels; i++) |
| 3059 | { |
| 3060 | const int a = pPixels[i]; |
| 3061 | |
| 3062 | uint32_t best_s_err = UINT32_MAX; |
| 3063 | //uint32_t best_s = 0; |
| 3064 | for (uint32_t s = 0; s < 8; s++) |
| 3065 | { |
| 3066 | const int v = clamp255((int)multiplier * g_etc2_eac_tables[table][s] + (int)base); |
| 3067 | |
| 3068 | uint32_t err = iabs(a - v); |
| 3069 | if (err < best_s_err) |
| 3070 | { |
| 3071 | best_s_err = err; |
| 3072 | //best_s = s; |
| 3073 | } |
| 3074 | } |
| 3075 | |
| 3076 | total_err += best_s_err * best_s_err; |
| 3077 | if (total_err >= best_err) |
| 3078 | break; |
| 3079 | } |
| 3080 | |
| 3081 | if (total_err < best_err) |
| 3082 | { |
| 3083 | best_err = total_err; |
| 3084 | results.m_base = base; |
| 3085 | results.m_multiplier = multiplier; |
| 3086 | results.m_table = table; |
| 3087 | if (!best_err) |
| 3088 | return best_err; |
| 3089 | } |
| 3090 | |
| 3091 | } // table |
| 3092 | |
| 3093 | } // multiplier |
| 3094 | |
| 3095 | } // base |
| 3096 | |
| 3097 | return best_err; |
| 3098 | } |
| 3099 | |
| 3100 | const int32_t DEFAULT_BC7_ERROR_WEIGHT = 50; |
| 3101 | const float UASTC_ERROR_THRESH = 1.3f; |
| 3102 | |
| 3103 | // TODO: This is a quick hack to favor certain modes when we know we'll be followed up with an RDO postprocess. |
| 3104 | static inline float get_uastc_mode_weight(uint32_t mode) |
| 3105 | { |
| 3106 | const float FAVORED_MODE_WEIGHT = .8f; |
| 3107 | |
| 3108 | switch (mode) |
| 3109 | { |
| 3110 | case 0: |
| 3111 | case 10: |
| 3112 | return FAVORED_MODE_WEIGHT; |
| 3113 | default: |
| 3114 | break; |
| 3115 | } |
| 3116 | |
| 3117 | return 1.0f; |
| 3118 | } |
| 3119 | |
| 3120 | void encode_uastc(const uint8_t* pRGBAPixels, uastc_block& output_block, uint32_t flags) |
| 3121 | { |
| 3122 | // printf("encode_uastc: \n"); |
| 3123 | // for (int i = 0; i < 16; i++) |
| 3124 | // printf("[%u %u %u %u] ", pRGBAPixels[i * 4 + 0], pRGBAPixels[i * 4 + 1], pRGBAPixels[i * 4 + 2], pRGBAPixels[i * 4 + 3]); |
| 3125 | // printf("\n"); |
| 3126 | |
| 3127 | const color_rgba(*block)[4] = reinterpret_cast<const color_rgba(*)[4]>(pRGBAPixels); |
| 3128 | |
| 3129 | bool solid_color = true, has_alpha = false, is_la = true; |
| 3130 | |
| 3131 | const color_rgba first_color(block[0][0]); |
| 3132 | for (uint32_t y = 0; y < 4; y++) |
| 3133 | { |
| 3134 | for (uint32_t x = 0; x < 4; x++) |
| 3135 | { |
| 3136 | if (block[y][x].a < 255) |
| 3137 | has_alpha = true; |
| 3138 | |
| 3139 | if (block[y][x] != first_color) |
| 3140 | solid_color = false; |
| 3141 | |
| 3142 | if ((block[y][x].r != block[y][x].g) || (block[y][x].r != block[y][x].b)) |
| 3143 | is_la = false; |
| 3144 | } |
| 3145 | } |
| 3146 | |
| 3147 | if (solid_color) |
| 3148 | { |
| 3149 | // Solid color blocks are so common that we handle them specially and as quickly as we can. |
| 3150 | uastc_encode_results solid_results; |
| 3151 | solid_results.m_uastc_mode = UASTC_MODE_INDEX_SOLID_COLOR; |
| 3152 | solid_results.m_astc_err = 0; |
| 3153 | solid_results.m_common_pattern = 0; |
| 3154 | solid_results.m_solid_color = first_color; |
| 3155 | memset(&solid_results.m_astc, 0, sizeof(solid_results.m_astc)); |
| 3156 | |
| 3157 | etc_block etc1_blk; |
| 3158 | uint32_t etc1_bias = 0; |
| 3159 | |
| 3160 | pack_etc1_block_solid_color(etc1_blk, &first_color.m_comps[0]); |
| 3161 | |
| 3162 | eac_a8_block eac_a8_blk; |
| 3163 | eac_a8_blk.m_table = 0; |
| 3164 | eac_a8_blk.m_multiplier = 1; |
| 3165 | |
| 3166 | pack_uastc(output_block, solid_results, etc1_blk, etc1_bias, eac_a8_blk, false, false); |
| 3167 | |
| 3168 | // printf(" Solid\n"); |
| 3169 | |
| 3170 | return; |
| 3171 | } |
| 3172 | |
| 3173 | int level = flags & 7; |
| 3174 | const bool favor_uastc_error = (flags & cPackUASTCFavorUASTCError) != 0; |
| 3175 | const bool favor_bc7_error = !favor_uastc_error && ((flags & cPackUASTCFavorBC7Error) != 0); |
| 3176 | //const bool etc1_perceptual = true; |
| 3177 | |
| 3178 | uastc_encode_results results[MAX_ENCODE_RESULTS]; |
| 3179 | |
| 3180 | level = clampi(level, cPackUASTCLevelFastest, cPackUASTCLevelVerySlow); |
| 3181 | |
| 3182 | // Set all options to slowest, then configure from there depending on the selected level. |
| 3183 | uint32_t mode_mask = UINT32_MAX; |
| 3184 | uint32_t uber_level = 6; |
| 3185 | bool estimate_partition = false; |
| 3186 | bool always_try_alpha_modes = true; |
| 3187 | uint32_t eac_a8_mul_search_rad = 3; |
| 3188 | uint32_t eac_a8_table_mask = UINT32_MAX; |
| 3189 | uint32_t least_squares_passes = 2; |
| 3190 | bool bc1_hints = true; |
| 3191 | bool only_use_la_on_transparent_blocks = false; |
| 3192 | |
| 3193 | switch (level) |
| 3194 | { |
| 3195 | case cPackUASTCLevelFastest: |
| 3196 | { |
| 3197 | mode_mask = (1 << 0) | (1 << 8) | |
| 3198 | (1 << 11) | (1 << 12) | |
| 3199 | (1 << 15); |
| 3200 | always_try_alpha_modes = false; |
| 3201 | eac_a8_mul_search_rad = 0; |
| 3202 | eac_a8_table_mask = (1 << 2) | (1 << 8) | (1 << 11) | (1 << 13); |
| 3203 | uber_level = 0; |
| 3204 | least_squares_passes = 1; |
| 3205 | bc1_hints = false; |
| 3206 | estimate_partition = true; |
| 3207 | only_use_la_on_transparent_blocks = true; |
| 3208 | break; |
| 3209 | } |
| 3210 | case cPackUASTCLevelFaster: |
| 3211 | { |
| 3212 | mode_mask = (1 << 0) | (1 << 4) | (1 << 6) | (1 << 8) | |
| 3213 | (1 << 9) | (1 << 11) | (1 << 12) | |
| 3214 | (1 << 15) | (1 << 17); |
| 3215 | always_try_alpha_modes = false; |
| 3216 | eac_a8_mul_search_rad = 0; |
| 3217 | eac_a8_table_mask = (1 << 2) | (1 << 8) | (1 << 11) | (1 << 13); |
| 3218 | uber_level = 0; |
| 3219 | least_squares_passes = 1; |
| 3220 | estimate_partition = true; |
| 3221 | break; |
| 3222 | } |
| 3223 | case cPackUASTCLevelDefault: |
| 3224 | { |
| 3225 | mode_mask = (1 << 0) | (1 << 1) | (1 << 4) | (1 << 5) | (1 << 6) | (1 << 8) | |
| 3226 | (1 << 9) | (1 << 10) | (1 << 11) | (1 << 12) | (1 << 13) | |
| 3227 | (1 << 15) | (1 << 16) | (1 << 17); |
| 3228 | always_try_alpha_modes = false; |
| 3229 | eac_a8_mul_search_rad = 1; |
| 3230 | eac_a8_table_mask = (1 << 0) | (1 << 2) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 10) | (1 << 11) | (1 << 13); |
| 3231 | uber_level = 1; |
| 3232 | least_squares_passes = 1; |
| 3233 | estimate_partition = true; |
| 3234 | break; |
| 3235 | } |
| 3236 | case cPackUASTCLevelSlower: |
| 3237 | { |
| 3238 | always_try_alpha_modes = false; |
| 3239 | eac_a8_mul_search_rad = 2; |
| 3240 | uber_level = 3; |
| 3241 | estimate_partition = true; |
| 3242 | break; |
| 3243 | } |
| 3244 | case cPackUASTCLevelVerySlow: |
| 3245 | { |
| 3246 | break; |
| 3247 | } |
| 3248 | } |
| 3249 | |
| 3250 | #if BASISU_SUPPORT_FORCE_MODE |
| 3251 | static int force_mode = -1; |
| 3252 | force_mode = (force_mode + 1) % TOTAL_UASTC_MODES; |
| 3253 | mode_mask = UINT32_MAX; |
| 3254 | always_try_alpha_modes = true; |
| 3255 | only_use_la_on_transparent_blocks = false; |
| 3256 | #endif |
| 3257 | |
| 3258 | // HACK HACK |
| 3259 | //mode_mask &= ~(1 << 18); |
| 3260 | //mode_mask = (1 << 18)| (1 << 10); |
| 3261 | |
| 3262 | uint32_t total_results = 0; |
| 3263 | |
| 3264 | if (only_use_la_on_transparent_blocks) |
| 3265 | { |
| 3266 | if ((is_la) && (!has_alpha)) |
| 3267 | is_la = false; |
| 3268 | } |
| 3269 | |
| 3270 | const bool try_alpha_modes = has_alpha || always_try_alpha_modes; |
| 3271 | |
| 3272 | bc7enc_compress_block_params comp_params; |
| 3273 | memset(&comp_params, 0, sizeof(comp_params)); |
| 3274 | comp_params.m_max_partitions_mode1 = 64; |
| 3275 | comp_params.m_least_squares_passes = least_squares_passes; |
| 3276 | comp_params.m_weights[0] = 1; |
| 3277 | comp_params.m_weights[1] = 1; |
| 3278 | comp_params.m_weights[2] = 1; |
| 3279 | comp_params.m_weights[3] = 1; |
| 3280 | comp_params.m_uber_level = uber_level; |
| 3281 | |
| 3282 | if (is_la) |
| 3283 | { |
| 3284 | if (mode_mask & (1U << 15)) |
| 3285 | astc_mode15(block, results, total_results, comp_params); |
| 3286 | |
| 3287 | if (mode_mask & (1U << 16)) |
| 3288 | astc_mode9_or_16(16, block, results, total_results, comp_params, estimate_partition ? 4 : 0); |
| 3289 | |
| 3290 | if (mode_mask & (1U << 17)) |
| 3291 | astc_mode11_or_17(17, block, results, total_results, comp_params); |
| 3292 | } |
| 3293 | |
| 3294 | if (!has_alpha) |
| 3295 | { |
| 3296 | if (mode_mask & (1U << 0)) |
| 3297 | astc_mode0_or_18(0, block, results, total_results, comp_params); |
| 3298 | |
| 3299 | if (mode_mask & (1U << 1)) |
| 3300 | astc_mode1(block, results, total_results, comp_params); |
| 3301 | |
| 3302 | if (mode_mask & (1U << 2)) |
| 3303 | astc_mode2(block, results, total_results, comp_params, estimate_partition); |
| 3304 | |
| 3305 | if (mode_mask & (1U << 3)) |
| 3306 | astc_mode3(block, results, total_results, comp_params, estimate_partition); |
| 3307 | |
| 3308 | if (mode_mask & (1U << 4)) |
| 3309 | astc_mode4(block, results, total_results, comp_params, estimate_partition); |
| 3310 | |
| 3311 | if (mode_mask & (1U << 5)) |
| 3312 | astc_mode5(block, results, total_results, comp_params); |
| 3313 | |
| 3314 | if (mode_mask & (1U << 6)) |
| 3315 | astc_mode6(block, results, total_results, comp_params); |
| 3316 | |
| 3317 | if (mode_mask & (1U << 7)) |
| 3318 | astc_mode7(block, results, total_results, comp_params, estimate_partition); |
| 3319 | |
| 3320 | if (mode_mask & (1U << 18)) |
| 3321 | astc_mode0_or_18(18, block, results, total_results, comp_params); |
| 3322 | } |
| 3323 | |
| 3324 | if (try_alpha_modes) |
| 3325 | { |
| 3326 | if (mode_mask & (1U << 9)) |
| 3327 | astc_mode9_or_16(9, block, results, total_results, comp_params, estimate_partition ? 4 : 0); |
| 3328 | |
| 3329 | if (mode_mask & (1U << 10)) |
| 3330 | astc_mode10(block, results, total_results, comp_params); |
| 3331 | |
| 3332 | if (mode_mask & (1U << 11)) |
| 3333 | astc_mode11_or_17(11, block, results, total_results, comp_params); |
| 3334 | |
| 3335 | if (mode_mask & (1U << 12)) |
| 3336 | astc_mode12(block, results, total_results, comp_params); |
| 3337 | |
| 3338 | if (mode_mask & (1U << 13)) |
| 3339 | astc_mode13(block, results, total_results, comp_params); |
| 3340 | |
| 3341 | if (mode_mask & (1U << 14)) |
| 3342 | astc_mode14(block, results, total_results, comp_params); |
| 3343 | } |
| 3344 | |
| 3345 | assert(total_results); |
| 3346 | |
| 3347 | // Fix up the errors so we consistently have LA, RGB, or RGBA error. |
| 3348 | for (uint32_t i = 0; i < total_results; i++) |
| 3349 | { |
| 3350 | uastc_encode_results& r = results[i]; |
| 3351 | if (!is_la) |
| 3352 | { |
| 3353 | if (g_uastc_mode_is_la[r.m_uastc_mode]) |
| 3354 | { |
| 3355 | color_rgba unpacked_block[16]; |
| 3356 | unpack_uastc(r.m_uastc_mode, r.m_common_pattern, r.m_solid_color.get_color32(), r.m_astc, (basist::color32 *)unpacked_block, false); |
| 3357 | |
| 3358 | uint64_t total_err = 0; |
| 3359 | for (uint32_t j = 0; j < 16; j++) |
| 3360 | total_err += color_distance(unpacked_block[j], ((const color_rgba*)block)[j], true); |
| 3361 | |
| 3362 | r.m_astc_err = total_err; |
| 3363 | } |
| 3364 | } |
| 3365 | else |
| 3366 | { |
| 3367 | if (!g_uastc_mode_is_la[r.m_uastc_mode]) |
| 3368 | { |
| 3369 | color_rgba unpacked_block[16]; |
| 3370 | unpack_uastc(r.m_uastc_mode, r.m_common_pattern, r.m_solid_color.get_color32(), r.m_astc, (basist::color32 *)unpacked_block, false); |
| 3371 | |
| 3372 | uint64_t total_err = 0; |
| 3373 | for (uint32_t j = 0; j < 16; j++) |
| 3374 | total_err += color_distance_la(unpacked_block[j], ((const color_rgba*)block)[j]); |
| 3375 | |
| 3376 | r.m_astc_err = total_err; |
| 3377 | } |
| 3378 | } |
| 3379 | } |
| 3380 | |
| 3381 | unpacked_uastc_block unpacked_ublock; |
| 3382 | memset(&unpacked_ublock, 0, sizeof(unpacked_ublock)); |
| 3383 | |
| 3384 | uint64_t total_overall_err[MAX_ENCODE_RESULTS]; |
| 3385 | float uastc_err_f[MAX_ENCODE_RESULTS]; |
| 3386 | double best_uastc_err_f = 1e+20f; |
| 3387 | |
| 3388 | int best_index = -1; |
| 3389 | |
| 3390 | if (total_results == 1) |
| 3391 | { |
| 3392 | best_index = 0; |
| 3393 | } |
| 3394 | else |
| 3395 | { |
| 3396 | const uint32_t bc7_err_weight = favor_bc7_error ? 100 : ((favor_uastc_error ? 0 : DEFAULT_BC7_ERROR_WEIGHT)); |
| 3397 | const uint32_t uastc_err_weight = favor_bc7_error ? 0 : 100; |
| 3398 | |
| 3399 | // Find best overall results, balancing UASTC and UASTC->BC7 error. |
| 3400 | // We purposely allow UASTC error to increase a little, if doing so lowers the BC7 error. |
| 3401 | for (uint32_t i = 0; i < total_results; i++) |
| 3402 | { |
| 3403 | #if BASISU_SUPPORT_FORCE_MODE |
| 3404 | if (results[i].m_uastc_mode == force_mode) |
| 3405 | { |
| 3406 | best_index = i; |
| 3407 | break; |
| 3408 | } |
| 3409 | #endif |
| 3410 | |
| 3411 | unpacked_ublock.m_mode = results[i].m_uastc_mode; |
| 3412 | unpacked_ublock.m_astc = results[i].m_astc; |
| 3413 | unpacked_ublock.m_common_pattern = results[i].m_common_pattern; |
| 3414 | unpacked_ublock.m_solid_color = results[i].m_solid_color.get_color32(); |
| 3415 | |
| 3416 | color_rgba decoded_uastc_block[4][4]; |
| 3417 | bool success = unpack_uastc(results[i].m_uastc_mode, results[i].m_common_pattern, results[i].m_solid_color.get_color32(), results[i].m_astc, (basist::color32 *)&decoded_uastc_block[0][0], false); |
| 3418 | (void)success; |
| 3419 | VALIDATE(success); |
| 3420 | |
| 3421 | uint64_t total_uastc_rgb_err, total_uastc_rgba_err, total_uastc_la_err; |
| 3422 | compute_block_error(block, decoded_uastc_block, total_uastc_rgb_err, total_uastc_rgba_err, total_uastc_la_err); |
| 3423 | |
| 3424 | // Validate the computed error, or we're go mad if it's inaccurate. |
| 3425 | if (results[i].m_uastc_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 3426 | { |
| 3427 | VALIDATE(total_uastc_rgba_err == 0); |
| 3428 | } |
| 3429 | else if (is_la) |
| 3430 | { |
| 3431 | VALIDATE(total_uastc_la_err == results[i].m_astc_err); |
| 3432 | } |
| 3433 | else if (g_uastc_mode_has_alpha[results[i].m_uastc_mode]) |
| 3434 | { |
| 3435 | VALIDATE(total_uastc_rgba_err == results[i].m_astc_err); |
| 3436 | } |
| 3437 | else |
| 3438 | { |
| 3439 | VALIDATE(total_uastc_rgb_err == results[i].m_astc_err); |
| 3440 | } |
| 3441 | |
| 3442 | // Transcode to BC7 |
| 3443 | bc7_optimization_results bc7_results; |
| 3444 | transcode_uastc_to_bc7(unpacked_ublock, bc7_results); |
| 3445 | |
| 3446 | bc7_block bc7_data; |
| 3447 | encode_bc7_block(&bc7_data, &bc7_results); |
| 3448 | |
| 3449 | color_rgba decoded_bc7_block[4][4]; |
| 3450 | unpack_block(texture_format::cBC7, &bc7_data, &decoded_bc7_block[0][0]); |
| 3451 | |
| 3452 | // Compute BC7 error |
| 3453 | uint64_t total_bc7_la_err, total_bc7_rgb_err, total_bc7_rgba_err; |
| 3454 | compute_block_error(block, decoded_bc7_block, total_bc7_rgb_err, total_bc7_rgba_err, total_bc7_la_err); |
| 3455 | |
| 3456 | if (results[i].m_uastc_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 3457 | { |
| 3458 | VALIDATE(total_bc7_rgba_err == 0); |
| 3459 | |
| 3460 | best_index = i; |
| 3461 | break; |
| 3462 | } |
| 3463 | |
| 3464 | uint64_t total_uastc_err = 0, total_bc7_err = 0; |
| 3465 | if (is_la) |
| 3466 | { |
| 3467 | total_bc7_err = total_bc7_la_err; |
| 3468 | total_uastc_err = total_uastc_la_err; |
| 3469 | } |
| 3470 | else if (has_alpha) |
| 3471 | { |
| 3472 | total_bc7_err = total_bc7_rgba_err; |
| 3473 | total_uastc_err = total_uastc_rgba_err; |
| 3474 | } |
| 3475 | else |
| 3476 | { |
| 3477 | total_bc7_err = total_bc7_rgb_err; |
| 3478 | total_uastc_err = total_uastc_rgb_err; |
| 3479 | } |
| 3480 | |
| 3481 | total_overall_err[i] = ((total_bc7_err * bc7_err_weight) / 100) + ((total_uastc_err * uastc_err_weight) / 100); |
| 3482 | if (!total_overall_err[i]) |
| 3483 | { |
| 3484 | best_index = i; |
| 3485 | break; |
| 3486 | } |
| 3487 | |
| 3488 | uastc_err_f[i] = sqrtf((float)total_uastc_err); |
| 3489 | |
| 3490 | if (uastc_err_f[i] < best_uastc_err_f) |
| 3491 | { |
| 3492 | best_uastc_err_f = uastc_err_f[i]; |
| 3493 | } |
| 3494 | |
| 3495 | } // total_results |
| 3496 | |
| 3497 | if (best_index < 0) |
| 3498 | { |
| 3499 | uint64_t best_err = UINT64_MAX; |
| 3500 | |
| 3501 | if ((best_uastc_err_f == 0.0f) || (favor_bc7_error)) |
| 3502 | { |
| 3503 | for (uint32_t i = 0; i < total_results; i++) |
| 3504 | { |
| 3505 | // TODO: This is a quick hack to favor modes 0 or 10 for better RDO compression. |
| 3506 | const float err_weight = (flags & cPackUASTCFavorSimplerModes) ? get_uastc_mode_weight(results[i].m_uastc_mode) : 1.0f; |
| 3507 | |
| 3508 | const uint64_t w = (uint64_t)(total_overall_err[i] * err_weight); |
| 3509 | if (w < best_err) |
| 3510 | { |
| 3511 | best_err = w; |
| 3512 | best_index = i; |
| 3513 | if (!best_err) |
| 3514 | break; |
| 3515 | } |
| 3516 | } // i |
| 3517 | } |
| 3518 | else |
| 3519 | { |
| 3520 | // Scan the UASTC results, and consider all results within a window that has the best UASTC+BC7 error. |
| 3521 | for (uint32_t i = 0; i < total_results; i++) |
| 3522 | { |
| 3523 | double err_delta = uastc_err_f[i] / best_uastc_err_f; |
| 3524 | |
| 3525 | if (err_delta <= UASTC_ERROR_THRESH) |
| 3526 | { |
| 3527 | // TODO: This is a quick hack to favor modes 0 or 10 for better RDO compression. |
| 3528 | const float err_weight = (flags & cPackUASTCFavorSimplerModes) ? get_uastc_mode_weight(results[i].m_uastc_mode) : 1.0f; |
| 3529 | |
| 3530 | const uint64_t w = (uint64_t)(total_overall_err[i] * err_weight); |
| 3531 | if (w < best_err) |
| 3532 | { |
| 3533 | best_err = w; |
| 3534 | best_index = i; |
| 3535 | if (!best_err) |
| 3536 | break; |
| 3537 | } |
| 3538 | } |
| 3539 | } // i |
| 3540 | } |
| 3541 | } |
| 3542 | } |
| 3543 | |
| 3544 | const uastc_encode_results& best_results = results[best_index]; |
| 3545 | const uint32_t best_mode = best_results.m_uastc_mode; |
| 3546 | const astc_block_desc& best_astc_results = best_results.m_astc; |
| 3547 | |
| 3548 | color_rgba decoded_uastc_block[4][4]; |
| 3549 | bool success = unpack_uastc(best_mode, best_results.m_common_pattern, best_results.m_solid_color.get_color32(), best_astc_results, (basist::color32 *)&decoded_uastc_block[0][0], false); |
| 3550 | (void)success; |
| 3551 | VALIDATE(success); |
| 3552 | |
| 3553 | #if BASISU_VALIDATE_UASTC_ENC |
| 3554 | // Make sure that the UASTC block unpacks to the same exact pixels as the ASTC block does, using two different decoders. |
| 3555 | { |
| 3556 | // Round trip to packed UASTC and back, then decode to pixels. |
| 3557 | etc_block etc1_blk; |
| 3558 | memset(&etc1_blk, 0, sizeof(etc1_blk)); |
| 3559 | eac_a8_block etc_eac_a8_blk; |
| 3560 | memset(&etc_eac_a8_blk, 0, sizeof(etc_eac_a8_blk)); |
| 3561 | etc_eac_a8_blk.m_multiplier = 1; |
| 3562 | |
| 3563 | basist::uastc_block temp_block; |
| 3564 | pack_uastc(temp_block, best_results, etc1_blk, 0, etc_eac_a8_blk, false, false); |
| 3565 | |
| 3566 | basist::color32 temp_block_unpacked[4][4]; |
| 3567 | success = basist::unpack_uastc(temp_block, (basist::color32 *)temp_block_unpacked, false); |
| 3568 | VALIDATE(success); |
| 3569 | |
| 3570 | #if BASISU_USE_ASTC_DECOMPRESS |
| 3571 | // Now round trip to packed ASTC and back, then decode to pixels. |
| 3572 | uint32_t astc_data[4]; |
| 3573 | |
| 3574 | if (best_results.m_uastc_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 3575 | pack_astc_solid_block(astc_data, (color32 &)best_results.m_solid_color); |
| 3576 | else |
| 3577 | { |
| 3578 | success = pack_astc_block(astc_data, &best_astc_results, best_results.m_uastc_mode); |
| 3579 | VALIDATE(success); |
| 3580 | } |
| 3581 | |
| 3582 | color_rgba decoded_astc_block[4][4]; |
| 3583 | success = basisu_astc::astc::decompress((uint8_t*)decoded_astc_block, (uint8_t*)&astc_data, false, 4, 4); |
| 3584 | VALIDATE(success); |
| 3585 | |
| 3586 | for (uint32_t y = 0; y < 4; y++) |
| 3587 | { |
| 3588 | for (uint32_t x = 0; x < 4; x++) |
| 3589 | { |
| 3590 | VALIDATE(decoded_astc_block[y][x] == decoded_uastc_block[y][x]); |
| 3591 | |
| 3592 | VALIDATE(temp_block_unpacked[y][x].c[0] == decoded_uastc_block[y][x].r); |
| 3593 | VALIDATE(temp_block_unpacked[y][x].c[1] == decoded_uastc_block[y][x].g); |
| 3594 | VALIDATE(temp_block_unpacked[y][x].c[2] == decoded_uastc_block[y][x].b); |
| 3595 | VALIDATE(temp_block_unpacked[y][x].c[3] == decoded_uastc_block[y][x].a); |
| 3596 | } |
| 3597 | } |
| 3598 | #endif |
| 3599 | } |
| 3600 | #endif |
| 3601 | |
| 3602 | // Compute BC1 hints |
| 3603 | bool bc1_hint0 = false, bc1_hint1 = false; |
| 3604 | if (bc1_hints) |
| 3605 | compute_bc1_hints(bc1_hint0, bc1_hint1, best_results, block, decoded_uastc_block); |
| 3606 | |
| 3607 | eac_a8_block eac_a8_blk; |
| 3608 | if ((g_uastc_mode_has_alpha[best_mode]) && (best_mode != UASTC_MODE_INDEX_SOLID_COLOR)) |
| 3609 | { |
| 3610 | // Compute ETC2 hints |
| 3611 | uint8_t decoded_uastc_block_alpha[16]; |
| 3612 | for (uint32_t i = 0; i < 16; i++) |
| 3613 | decoded_uastc_block_alpha[i] = decoded_uastc_block[i >> 2][i & 3].a; |
| 3614 | |
| 3615 | uastc_pack_eac_a8_results eac8_a8_results; |
| 3616 | memset(&eac8_a8_results, 0, sizeof(eac8_a8_results)); |
| 3617 | uastc_pack_eac_a8(eac8_a8_results, decoded_uastc_block_alpha, 16, 0, eac_a8_mul_search_rad, eac_a8_table_mask); |
| 3618 | |
| 3619 | // All we care about for hinting is the table and multiplier. |
| 3620 | eac_a8_blk.m_table = eac8_a8_results.m_table; |
| 3621 | eac_a8_blk.m_multiplier = eac8_a8_results.m_multiplier; |
| 3622 | } |
| 3623 | else |
| 3624 | { |
| 3625 | memset(&eac_a8_blk, 0, sizeof(eac_a8_blk)); |
| 3626 | } |
| 3627 | |
| 3628 | // Compute ETC1 hints |
| 3629 | etc_block etc1_blk; |
| 3630 | uint32_t etc1_bias = 0; |
| 3631 | compute_etc1_hints(etc1_blk, etc1_bias, best_results, block, decoded_uastc_block, level, flags); |
| 3632 | |
| 3633 | // Finally, pack the UASTC block with its hints and we're done. |
| 3634 | pack_uastc(output_block, best_results, etc1_blk, etc1_bias, eac_a8_blk, bc1_hint0, bc1_hint1); |
| 3635 | |
| 3636 | // printf(" Packed: "); |
| 3637 | // for (int i = 0; i < 16; i++) |
| 3638 | // printf("%X ", output_block.m_bytes[i]); |
| 3639 | // printf("\n"); |
| 3640 | } |
| 3641 | |
| 3642 | static bool uastc_recompute_hints(basist::uastc_block* pBlock, const color_rgba* pBlock_pixels, uint32_t flags, const unpacked_uastc_block *pUnpacked_blk) |
| 3643 | { |
| 3644 | unpacked_uastc_block unpacked_blk; |
| 3645 | |
| 3646 | if (pUnpacked_blk) |
| 3647 | unpacked_blk = *pUnpacked_blk; |
| 3648 | else |
| 3649 | { |
| 3650 | if (!unpack_uastc(*pBlock, unpacked_blk, false, true)) |
| 3651 | return false; |
| 3652 | } |
| 3653 | color_rgba decoded_uastc_block[4][4]; |
| 3654 | if (!unpack_uastc(unpacked_blk, (basist::color32 *)decoded_uastc_block, false)) |
| 3655 | return false; |
| 3656 | uastc_encode_results results; |
| 3657 | results.m_uastc_mode = unpacked_blk.m_mode; |
| 3658 | results.m_common_pattern = unpacked_blk.m_common_pattern; |
| 3659 | results.m_astc = unpacked_blk.m_astc; |
| 3660 | results.m_solid_color = unpacked_blk.m_solid_color; |
| 3661 | results.m_astc_err = 0; |
| 3662 | bool bc1_hints = true; |
| 3663 | uint32_t eac_a8_mul_search_rad = 3; |
| 3664 | uint32_t eac_a8_table_mask = UINT32_MAX; |
| 3665 | const uint32_t level = flags & cPackUASTCLevelMask; |
| 3666 | switch (level) |
| 3667 | { |
| 3668 | case cPackUASTCLevelFastest: |
| 3669 | { |
| 3670 | eac_a8_mul_search_rad = 0; |
| 3671 | eac_a8_table_mask = (1 << 2) | (1 << 8) | (1 << 11) | (1 << 13); |
| 3672 | bc1_hints = false; |
| 3673 | break; |
| 3674 | } |
| 3675 | case cPackUASTCLevelFaster: |
| 3676 | { |
| 3677 | eac_a8_mul_search_rad = 0; |
| 3678 | eac_a8_table_mask = (1 << 2) | (1 << 8) | (1 << 11) | (1 << 13); |
| 3679 | break; |
| 3680 | } |
| 3681 | case cPackUASTCLevelDefault: |
| 3682 | { |
| 3683 | eac_a8_mul_search_rad = 1; |
| 3684 | eac_a8_table_mask = (1 << 0) | (1 << 2) | (1 << 6) | (1 << 7) | (1 << 8) | (1 << 10) | (1 << 11) | (1 << 13); |
| 3685 | break; |
| 3686 | } |
| 3687 | case cPackUASTCLevelSlower: |
| 3688 | { |
| 3689 | eac_a8_mul_search_rad = 2; |
| 3690 | break; |
| 3691 | } |
| 3692 | case cPackUASTCLevelVerySlow: |
| 3693 | { |
| 3694 | break; |
| 3695 | } |
| 3696 | } |
| 3697 | bool bc1_hint0 = false, bc1_hint1 = false; |
| 3698 | if (bc1_hints) |
| 3699 | compute_bc1_hints(bc1_hint0, bc1_hint1, results, (color_rgba (*)[4])pBlock_pixels, decoded_uastc_block); |
| 3700 | const uint32_t best_mode = unpacked_blk.m_mode; |
| 3701 | eac_a8_block eac_a8_blk; |
| 3702 | if ((g_uastc_mode_has_alpha[best_mode]) && (best_mode != UASTC_MODE_INDEX_SOLID_COLOR)) |
| 3703 | { |
| 3704 | uint8_t decoded_uastc_block_alpha[16]; |
| 3705 | for (uint32_t i = 0; i < 16; i++) |
| 3706 | decoded_uastc_block_alpha[i] = decoded_uastc_block[i >> 2][i & 3].a; |
| 3707 | uastc_pack_eac_a8_results eac8_a8_results; |
| 3708 | memset(&eac8_a8_results, 0, sizeof(eac8_a8_results)); |
| 3709 | uastc_pack_eac_a8(eac8_a8_results, decoded_uastc_block_alpha, 16, 0, eac_a8_mul_search_rad, eac_a8_table_mask); |
| 3710 | eac_a8_blk.m_table = eac8_a8_results.m_table; |
| 3711 | eac_a8_blk.m_multiplier = eac8_a8_results.m_multiplier; |
| 3712 | } |
| 3713 | else |
| 3714 | { |
| 3715 | memset(&eac_a8_blk, 0, sizeof(eac_a8_blk)); |
| 3716 | } |
| 3717 | etc_block etc1_blk; |
| 3718 | uint32_t etc1_bias = 0; |
| 3719 | compute_etc1_hints(etc1_blk, etc1_bias, results, (color_rgba (*)[4])pBlock_pixels, decoded_uastc_block, level, flags); |
| 3720 | pack_uastc(*pBlock, results, etc1_blk, etc1_bias, eac_a8_blk, bc1_hint0, bc1_hint1); |
| 3721 | return true; |
| 3722 | } |
| 3723 | |
| 3724 | static const uint8_t g_uastc_mode_selector_bits[TOTAL_UASTC_MODES][2] = |
| 3725 | { |
| 3726 | { 65, 63 }, { 69, 31 }, { 73, 46 }, { 89, 29 }, |
| 3727 | { 89, 30 }, { 68, 47 }, { 66, 62 }, { 89, 30 }, |
| 3728 | { 0, 0 }, { 97, 30 }, { 65, 63 }, { 66, 62 }, |
| 3729 | { 81, 47 }, { 94, 30 }, { 92, 31 }, { 62, 63 }, |
| 3730 | { 98, 30 }, { 61, 62 }, { 49, 79 } |
| 3731 | }; |
| 3732 | |
| 3733 | static inline uint32_t set_block_bits(uint8_t* pBytes, uint64_t val, uint32_t num_bits, uint32_t cur_ofs) |
| 3734 | { |
| 3735 | assert(num_bits <= 64); |
| 3736 | assert((num_bits == 64) || (val < (1ULL << num_bits))); |
| 3737 | uint64_t mask = (num_bits == 64) ? UINT64_MAX : ((1ULL << num_bits) - 1); |
| 3738 | while (num_bits) |
| 3739 | { |
| 3740 | const uint32_t n = basisu::minimum<uint32_t>(8U - (cur_ofs & 7U), num_bits); |
| 3741 | pBytes[cur_ofs >> 3] &= ~static_cast<uint8_t>(mask << (cur_ofs & 7U)); |
| 3742 | pBytes[cur_ofs >> 3] |= static_cast<uint8_t>(val << (cur_ofs & 7U)); |
| 3743 | val >>= n; |
| 3744 | mask >>= n; |
| 3745 | num_bits -= n; |
| 3746 | cur_ofs += n; |
| 3747 | } |
| 3748 | return cur_ofs; |
| 3749 | } |
| 3750 | |
| 3751 | static const uint8_t g_tdefl_small_dist_extra[512] = |
| 3752 | { |
| 3753 | 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, |
| 3754 | 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, |
| 3755 | 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, |
| 3756 | 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, |
| 3757 | 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, |
| 3758 | 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, |
| 3759 | 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, |
| 3760 | 7, 7, 7, 7, 7, 7, 7, 7 |
| 3761 | }; |
| 3762 | |
| 3763 | static const uint8_t g_tdefl_large_dist_extra[128] = |
| 3764 | { |
| 3765 | 0, 0, 8, 8, 9, 9, 9, 9, 10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, |
| 3766 | 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, |
| 3767 | 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13 |
| 3768 | }; |
| 3769 | |
| 3770 | static inline uint32_t compute_match_cost_estimate(uint32_t dist) |
| 3771 | { |
| 3772 | uint32_t len_cost = 7; |
| 3773 | uint32_t dist_cost = 5; |
| 3774 | if (dist < 512) |
| 3775 | dist_cost += g_tdefl_small_dist_extra[dist & 511]; |
| 3776 | else |
| 3777 | { |
| 3778 | dist_cost += g_tdefl_large_dist_extra[basisu::minimum<uint32_t>(dist, 32767) >> 8]; |
| 3779 | while (dist >= 32768) |
| 3780 | { |
| 3781 | dist_cost++; |
| 3782 | dist >>= 1; |
| 3783 | } |
| 3784 | } |
| 3785 | return len_cost + dist_cost; |
| 3786 | } |
| 3787 | |
| 3788 | struct selector_bitsequence |
| 3789 | { |
| 3790 | uint64_t m_sel; |
| 3791 | uint32_t m_ofs; |
| 3792 | selector_bitsequence() { } |
| 3793 | selector_bitsequence(uint32_t bit_ofs, uint64_t sel) : m_sel(sel), m_ofs(bit_ofs) { } |
| 3794 | bool operator== (const selector_bitsequence& other) const |
| 3795 | { |
| 3796 | return (m_ofs == other.m_ofs) && (m_sel == other.m_sel); |
| 3797 | } |
| 3798 | |
| 3799 | bool operator< (const selector_bitsequence& other) const |
| 3800 | { |
| 3801 | if (m_ofs < other.m_ofs) |
| 3802 | return true; |
| 3803 | else if (m_ofs == other.m_ofs) |
| 3804 | return m_sel < other.m_sel; |
| 3805 | |
| 3806 | return false; |
| 3807 | } |
| 3808 | }; |
| 3809 | |
| 3810 | struct selector_bitsequence_hash |
| 3811 | { |
| 3812 | std::size_t operator()(selector_bitsequence const& s) const noexcept |
| 3813 | { |
| 3814 | return static_cast<std::size_t>(hash_hsieh((uint8_t *)&s, sizeof(s)) ^ s.m_sel); |
| 3815 | } |
| 3816 | }; |
| 3817 | |
| 3818 | class tracked_stat |
| 3819 | { |
| 3820 | public: |
| 3821 | tracked_stat() { clear(); } |
| 3822 | |
| 3823 | void clear() { m_num = 0; m_total = 0; m_total2 = 0; } |
| 3824 | |
| 3825 | void update(uint32_t val) { m_num++; m_total += val; m_total2 += val * val; } |
| 3826 | |
| 3827 | tracked_stat& operator += (uint32_t val) { update(val); return *this; } |
| 3828 | |
| 3829 | uint32_t get_number_of_values() { return m_num; } |
| 3830 | uint64_t get_total() const { return m_total; } |
| 3831 | uint64_t get_total2() const { return m_total2; } |
| 3832 | |
| 3833 | float get_average() const { return m_num ? (float)m_total / m_num : 0.0f; }; |
| 3834 | float get_std_dev() const { return m_num ? sqrtf((float)(m_num * m_total2 - m_total * m_total)) / m_num : 0.0f; } |
| 3835 | float get_variance() const { float s = get_std_dev(); return s * s; } |
| 3836 | |
| 3837 | private: |
| 3838 | uint32_t m_num; |
| 3839 | uint64_t m_total; |
| 3840 | uint64_t m_total2; |
| 3841 | }; |
| 3842 | |
| 3843 | static bool uastc_rdo_blocks(uint32_t first_index, uint32_t last_index, basist::uastc_block* pBlocks, const color_rgba* pBlock_pixels, const uastc_rdo_params& params, uint32_t flags, |
| 3844 | uint32_t &total_skipped, uint32_t &total_refined, uint32_t &total_modified, uint32_t &total_smooth) |
| 3845 | { |
| 3846 | debug_printf("uastc_rdo_blocks: Processing blocks %u to %u\n", first_index, last_index); |
| 3847 | |
| 3848 | const int total_blocks_to_check = basisu::maximum<uint32_t>(1U, params.m_lz_dict_size / sizeof(basist::uastc_block)); |
| 3849 | const bool perceptual = false; |
| 3850 | |
| 3851 | std::unordered_map<selector_bitsequence, uint32_t, selector_bitsequence_hash> selector_history; |
| 3852 | |
| 3853 | for (uint32_t block_index = first_index; block_index < last_index; block_index++) |
| 3854 | { |
| 3855 | const basist::uastc_block& blk = pBlocks[block_index]; |
| 3856 | const color_rgba* pPixels = &pBlock_pixels[16 * block_index]; |
| 3857 | |
| 3858 | unpacked_uastc_block unpacked_blk; |
| 3859 | if (!unpack_uastc(blk, unpacked_blk, false, true)) |
| 3860 | return false; |
| 3861 | |
| 3862 | const uint32_t block_mode = unpacked_blk.m_mode; |
| 3863 | if (block_mode == UASTC_MODE_INDEX_SOLID_COLOR) |
| 3864 | continue; |
| 3865 | |
| 3866 | tracked_stat r_stats, g_stats, b_stats, a_stats; |
| 3867 | |
| 3868 | for (uint32_t i = 0; i < 16; i++) |
| 3869 | { |
| 3870 | r_stats.update(pPixels[i].r); |
| 3871 | g_stats.update(pPixels[i].g); |
| 3872 | b_stats.update(pPixels[i].b); |
| 3873 | a_stats.update(pPixels[i].a); |
| 3874 | } |
| 3875 | |
| 3876 | const float max_std_dev = basisu::maximum<float>(basisu::maximum<float>(basisu::maximum(r_stats.get_std_dev(), g_stats.get_std_dev()), b_stats.get_std_dev()), a_stats.get_std_dev()); |
| 3877 | |
| 3878 | float yl = clamp<float>(max_std_dev / params.m_max_smooth_block_std_dev, 0.0f, 1.0f); |
| 3879 | yl = yl * yl; |
| 3880 | const float smooth_block_error_scale = lerp<float>(params.m_smooth_block_max_error_scale, 1.0f, yl); |
| 3881 | if (smooth_block_error_scale > 1.0f) |
| 3882 | total_smooth++; |
| 3883 | |
| 3884 | color_rgba decoded_uastc_block[4][4]; |
| 3885 | if (!unpack_uastc(unpacked_blk, (basist::color32*)decoded_uastc_block, false)) |
| 3886 | return false; |
| 3887 | |
| 3888 | uint64_t uastc_err = 0; |
| 3889 | for (uint32_t i = 0; i < 16; i++) |
| 3890 | uastc_err += color_distance(perceptual, pPixels[i], ((color_rgba*)decoded_uastc_block)[i], true); |
| 3891 | |
| 3892 | // Transcode to BC7 |
| 3893 | bc7_optimization_results b7_results; |
| 3894 | if (!transcode_uastc_to_bc7(unpacked_blk, b7_results)) |
| 3895 | return false; |
| 3896 | |
| 3897 | basist::bc7_block b7_block; |
| 3898 | basist::encode_bc7_block(&b7_block, &b7_results); |
| 3899 | |
| 3900 | color_rgba decoded_b7_blk[4][4]; |
| 3901 | unpack_block(texture_format::cBC7, &b7_block, &decoded_b7_blk[0][0]); |
| 3902 | |
| 3903 | uint64_t bc7_err = 0; |
| 3904 | for (uint32_t i = 0; i < 16; i++) |
| 3905 | bc7_err += color_distance(perceptual, pPixels[i], ((color_rgba*)decoded_b7_blk)[i], true); |
| 3906 | |
| 3907 | uint64_t cur_err = (uastc_err + bc7_err) / 2; |
| 3908 | |
| 3909 | // Divide by 16*4 to compute RMS error |
| 3910 | const float cur_ms_err = (float)cur_err * (1.0f / 64.0f); |
| 3911 | const float cur_rms_err = sqrt(cur_ms_err); |
| 3912 | |
| 3913 | const uint32_t first_sel_bit = g_uastc_mode_selector_bits[block_mode][0]; |
| 3914 | const uint32_t total_sel_bits = g_uastc_mode_selector_bits[block_mode][1]; |
| 3915 | assert(first_sel_bit + total_sel_bits <= 128); |
| 3916 | assert(total_sel_bits > 0); |
| 3917 | |
| 3918 | uint32_t cur_bit_offset = first_sel_bit; |
| 3919 | uint64_t cur_sel_bits = read_bits((const uint8_t*)&blk, cur_bit_offset, basisu::minimum(64U, total_sel_bits)); |
| 3920 | |
| 3921 | if (cur_rms_err >= params.m_skip_block_rms_thresh) |
| 3922 | { |
| 3923 | auto cur_search_res = selector_history.insert(std::make_pair(selector_bitsequence(first_sel_bit, cur_sel_bits), block_index)); |
| 3924 | |
| 3925 | // Block already has too much error, so don't mess with it. |
| 3926 | if (!cur_search_res.second) |
| 3927 | (*cur_search_res.first).second = block_index; |
| 3928 | |
| 3929 | total_skipped++; |
| 3930 | continue; |
| 3931 | } |
| 3932 | |
| 3933 | int cur_bits; |
| 3934 | auto cur_find_res = selector_history.find(selector_bitsequence(first_sel_bit, cur_sel_bits)); |
| 3935 | if (cur_find_res == selector_history.end()) |
| 3936 | { |
| 3937 | // Wasn't found - wildly estimate literal cost |
| 3938 | //cur_bits = (total_sel_bits * 5) / 4; |
| 3939 | cur_bits = (total_sel_bits * params.m_lz_literal_cost) / 100; |
| 3940 | } |
| 3941 | else |
| 3942 | { |
| 3943 | // Was found - wildly estimate match cost |
| 3944 | uint32_t match_block_index = cur_find_res->second; |
| 3945 | const int block_dist_in_bytes = (block_index - match_block_index) * 16; |
| 3946 | cur_bits = compute_match_cost_estimate(block_dist_in_bytes); |
| 3947 | } |
| 3948 | |
| 3949 | int first_block_to_check = basisu::maximum<int>(first_index, block_index - total_blocks_to_check); |
| 3950 | int last_block_to_check = block_index - 1; |
| 3951 | |
| 3952 | basist::uastc_block best_block(blk); |
| 3953 | uint32_t best_block_index = block_index; |
| 3954 | |
| 3955 | float best_t = cur_ms_err * smooth_block_error_scale + cur_bits * params.m_lambda; |
| 3956 | |
| 3957 | // Now scan through previous blocks, insert their selector bit patterns into the current block, and find |
| 3958 | // selector bit patterns which don't increase the overall block error too much. |
| 3959 | for (int prev_block_index = last_block_to_check; prev_block_index >= first_block_to_check; --prev_block_index) |
| 3960 | { |
| 3961 | const basist::uastc_block& prev_blk = pBlocks[prev_block_index]; |
| 3962 | |
| 3963 | uint32_t bit_offset = first_sel_bit; |
| 3964 | uint64_t sel_bits = read_bits((const uint8_t*)&prev_blk, bit_offset, basisu::minimum(64U, total_sel_bits)); |
| 3965 | |
| 3966 | int match_block_index = prev_block_index; |
| 3967 | auto res = selector_history.find(selector_bitsequence(first_sel_bit, sel_bits)); |
| 3968 | if (res != selector_history.end()) |
| 3969 | match_block_index = res->second; |
| 3970 | // Have we already checked this bit pattern? If so then skip this block. |
| 3971 | if (match_block_index > prev_block_index) |
| 3972 | continue; |
| 3973 | |
| 3974 | unpacked_uastc_block unpacked_prev_blk; |
| 3975 | if (!unpack_uastc(prev_blk, unpacked_prev_blk, false, true)) |
| 3976 | return false; |
| 3977 | |
| 3978 | basist::uastc_block trial_blk(blk); |
| 3979 | |
| 3980 | set_block_bits((uint8_t*)&trial_blk, sel_bits, basisu::minimum(64U, total_sel_bits), first_sel_bit); |
| 3981 | |
| 3982 | if (total_sel_bits > 64) |
| 3983 | { |
| 3984 | sel_bits = read_bits((const uint8_t*)&prev_blk, bit_offset, total_sel_bits - 64U); |
| 3985 | |
| 3986 | set_block_bits((uint8_t*)&trial_blk, sel_bits, total_sel_bits - 64U, first_sel_bit + basisu::minimum(64U, total_sel_bits)); |
| 3987 | } |
| 3988 | |
| 3989 | unpacked_uastc_block unpacked_trial_blk; |
| 3990 | if (!unpack_uastc(trial_blk, unpacked_trial_blk, false, true)) |
| 3991 | continue; |
| 3992 | |
| 3993 | color_rgba decoded_trial_uastc_block[4][4]; |
| 3994 | if (!unpack_uastc(unpacked_trial_blk, (basist::color32*)decoded_trial_uastc_block, false)) |
| 3995 | continue; |
| 3996 | |
| 3997 | uint64_t trial_uastc_err = 0; |
| 3998 | for (uint32_t i = 0; i < 16; i++) |
| 3999 | trial_uastc_err += color_distance(perceptual, pPixels[i], ((color_rgba*)decoded_trial_uastc_block)[i], true); |
| 4000 | |
| 4001 | // Transcode trial to BC7, compute error |
| 4002 | bc7_optimization_results trial_b7_results; |
| 4003 | if (!transcode_uastc_to_bc7(unpacked_trial_blk, trial_b7_results)) |
| 4004 | return false; |
| 4005 | |
| 4006 | basist::bc7_block trial_b7_block; |
| 4007 | basist::encode_bc7_block(&trial_b7_block, &trial_b7_results); |
| 4008 | |
| 4009 | color_rgba decoded_trial_b7_blk[4][4]; |
| 4010 | unpack_block(texture_format::cBC7, &trial_b7_block, &decoded_trial_b7_blk[0][0]); |
| 4011 | |
| 4012 | uint64_t trial_bc7_err = 0; |
| 4013 | for (uint32_t i = 0; i < 16; i++) |
| 4014 | trial_bc7_err += color_distance(perceptual, pPixels[i], ((color_rgba*)decoded_trial_b7_blk)[i], true); |
| 4015 | |
| 4016 | uint64_t trial_err = (trial_uastc_err + trial_bc7_err) / 2; |
| 4017 | |
| 4018 | const float trial_ms_err = (float)trial_err * (1.0f / 64.0f); |
| 4019 | const float trial_rms_err = sqrtf(trial_ms_err); |
| 4020 | |
| 4021 | if (trial_rms_err > cur_rms_err * params.m_max_allowed_rms_increase_ratio) |
| 4022 | continue; |
| 4023 | |
| 4024 | const int block_dist_in_bytes = (block_index - match_block_index) * 16; |
| 4025 | const int match_bits = compute_match_cost_estimate(block_dist_in_bytes); |
| 4026 | |
| 4027 | float t = trial_ms_err * smooth_block_error_scale + match_bits * params.m_lambda; |
| 4028 | if (t < best_t) |
| 4029 | { |
| 4030 | best_t = t; |
| 4031 | best_block_index = prev_block_index; |
| 4032 | |
| 4033 | best_block = trial_blk; |
| 4034 | } |
| 4035 | |
| 4036 | } // prev_block_index |
| 4037 | |
| 4038 | if (best_block_index != block_index) |
| 4039 | { |
| 4040 | total_modified++; |
| 4041 | |
| 4042 | unpacked_uastc_block unpacked_best_blk; |
| 4043 | if (!unpack_uastc(best_block, unpacked_best_blk, false, false)) |
| 4044 | return false; |
| 4045 | |
| 4046 | if ((params.m_endpoint_refinement) && (block_mode == 0)) |
| 4047 | { |
| 4048 | // Attempt to refine mode 0 block's endpoints, using the new selectors. This doesn't help much, but it does help. |
| 4049 | // TODO: We could do this with the other modes too. |
| 4050 | color_rgba decoded_best_uastc_block[4][4]; |
| 4051 | if (!unpack_uastc(unpacked_best_blk, (basist::color32*)decoded_best_uastc_block, false)) |
| 4052 | return false; |
| 4053 | |
| 4054 | // Compute the block's current error (with the modified selectors). |
| 4055 | uint64_t best_uastc_err = 0; |
| 4056 | for (uint32_t i = 0; i < 16; i++) |
| 4057 | best_uastc_err += color_distance(perceptual, pPixels[i], ((color_rgba*)decoded_best_uastc_block)[i], true); |
| 4058 | |
| 4059 | bc7enc_compress_block_params comp_params; |
| 4060 | memset(&comp_params, 0, sizeof(comp_params)); |
| 4061 | comp_params.m_max_partitions_mode1 = 64; |
| 4062 | comp_params.m_least_squares_passes = 1; |
| 4063 | comp_params.m_weights[0] = 1; |
| 4064 | comp_params.m_weights[1] = 1; |
| 4065 | comp_params.m_weights[2] = 1; |
| 4066 | comp_params.m_weights[3] = 1; |
| 4067 | comp_params.m_uber_level = 0; |
| 4068 | |
| 4069 | uastc_encode_results results; |
| 4070 | uint32_t total_results = 0; |
| 4071 | astc_mode0_or_18(0, (color_rgba(*)[4])pPixels, &results, total_results, comp_params, unpacked_best_blk.m_astc.m_weights); |
| 4072 | assert(total_results == 1); |
| 4073 | |
| 4074 | // See if the overall error has actually gone done. |
| 4075 | |
| 4076 | color_rgba decoded_trial_uastc_block[4][4]; |
| 4077 | bool success = unpack_uastc(results.m_uastc_mode, results.m_common_pattern, results.m_solid_color.get_color32(), results.m_astc, (basist::color32*) & decoded_trial_uastc_block[0][0], false); |
| 4078 | assert(success); |
| 4079 | |
| 4080 | BASISU_NOTE_UNUSED(success); |
| 4081 | |
| 4082 | uint64_t trial_uastc_err = 0; |
| 4083 | for (uint32_t i = 0; i < 16; i++) |
| 4084 | trial_uastc_err += color_distance(perceptual, pPixels[i], ((color_rgba*)decoded_trial_uastc_block)[i], true); |
| 4085 | |
| 4086 | if (trial_uastc_err < best_uastc_err) |
| 4087 | { |
| 4088 | // The error went down, so accept the new endpoints. |
| 4089 | |
| 4090 | // Ensure the selectors haven't changed, otherwise we'll invalidate the LZ matches. |
| 4091 | for (uint32_t i = 0; i < 16; i++) |
| 4092 | assert(unpacked_best_blk.m_astc.m_weights[i] == results.m_astc.m_weights[i]); |
| 4093 | |
| 4094 | unpacked_best_blk.m_astc = results.m_astc; |
| 4095 | |
| 4096 | total_refined++; |
| 4097 | } |
| 4098 | } // if ((params.m_endpoint_refinement) && (block_mode == 0)) |
| 4099 | |
| 4100 | // The selectors have changed, so go recompute the block hints. |
| 4101 | if (!uastc_recompute_hints(&best_block, pPixels, flags, &unpacked_best_blk)) |
| 4102 | return false; |
| 4103 | |
| 4104 | // Write the modified block |
| 4105 | pBlocks[block_index] = best_block; |
| 4106 | |
| 4107 | } // if (best_block_index != block_index) |
| 4108 | |
| 4109 | { |
| 4110 | uint32_t bit_offset = first_sel_bit; |
| 4111 | uint64_t sel_bits = read_bits((const uint8_t*)&best_block, bit_offset, basisu::minimum(64U, total_sel_bits)); |
| 4112 | |
| 4113 | auto res = selector_history.insert(std::make_pair(selector_bitsequence(first_sel_bit, sel_bits), block_index)); |
| 4114 | if (!res.second) |
| 4115 | (*res.first).second = block_index; |
| 4116 | } |
| 4117 | |
| 4118 | } // block_index |
| 4119 | |
| 4120 | return true; |
| 4121 | } |
| 4122 | |
| 4123 | // This function implements a basic form of rate distortion optimization (RDO) for UASTC. |
| 4124 | // It only changes selectors and then updates the hints. It uses very approximate LZ bitprice estimation. |
| 4125 | // There's A LOT that can be done better in here, but it's a start. |
| 4126 | // One nice advantage of the method used here is that it works for any input, no matter which or how many modes it uses. |
| 4127 | bool uastc_rdo(uint32_t num_blocks, basist::uastc_block* pBlocks, const color_rgba* pBlock_pixels, const uastc_rdo_params& params, uint32_t flags, job_pool* pJob_pool, uint32_t total_jobs) |
| 4128 | { |
| 4129 | assert(params.m_max_allowed_rms_increase_ratio > 1.0f); |
| 4130 | assert(params.m_lz_dict_size > 0); |
| 4131 | assert(params.m_lambda > 0.0f); |
| 4132 | |
| 4133 | uint32_t total_skipped = 0, total_modified = 0, total_refined = 0, total_smooth = 0; |
| 4134 | |
| 4135 | uint32_t blocks_per_job = total_jobs ? (num_blocks / total_jobs) : 0; |
| 4136 | |
| 4137 | std::mutex stat_mutex; |
| 4138 | |
| 4139 | bool status = false; |
| 4140 | |
| 4141 | if ((!pJob_pool) || (total_jobs <= 1) || (blocks_per_job <= 8)) |
| 4142 | { |
| 4143 | status = uastc_rdo_blocks(0, num_blocks, pBlocks, pBlock_pixels, params, flags, total_skipped, total_refined, total_modified, total_smooth); |
| 4144 | } |
| 4145 | else |
| 4146 | { |
| 4147 | bool all_succeeded = true; |
| 4148 | |
| 4149 | for (uint32_t block_index_iter = 0; block_index_iter < num_blocks; block_index_iter += blocks_per_job) |
| 4150 | { |
| 4151 | const uint32_t first_index = block_index_iter; |
| 4152 | const uint32_t last_index = minimum<uint32_t>(num_blocks, block_index_iter + blocks_per_job); |
| 4153 | |
| 4154 | #ifndef __EMSCRIPTEN__ |
| 4155 | pJob_pool->add_job([first_index, last_index, pBlocks, pBlock_pixels, ¶ms, flags, &total_skipped, &total_modified, &total_refined, &total_smooth, &all_succeeded, &stat_mutex] { |
| 4156 | #endif |
| 4157 | |
| 4158 | uint32_t job_skipped = 0, job_modified = 0, job_refined = 0, job_smooth = 0; |
| 4159 | |
| 4160 | bool status = uastc_rdo_blocks(first_index, last_index, pBlocks, pBlock_pixels, params, flags, job_skipped, job_refined, job_modified, job_smooth); |
| 4161 | |
| 4162 | { |
| 4163 | std::lock_guard<std::mutex> lck(stat_mutex); |
| 4164 | |
| 4165 | all_succeeded = all_succeeded && status; |
| 4166 | total_skipped += job_skipped; |
| 4167 | total_modified += job_modified; |
| 4168 | total_refined += job_refined; |
| 4169 | total_smooth += job_smooth; |
| 4170 | } |
| 4171 | |
| 4172 | #ifndef __EMSCRIPTEN__ |
| 4173 | } |
| 4174 | ); |
| 4175 | #endif |
| 4176 | |
| 4177 | } // block_index_iter |
| 4178 | |
| 4179 | #ifndef __EMSCRIPTEN__ |
| 4180 | pJob_pool->wait_for_all(); |
| 4181 | #endif |
| 4182 | |
| 4183 | status = all_succeeded; |
| 4184 | } |
| 4185 | |
| 4186 | debug_printf("uastc_rdo: Total modified: %3.2f%%, total skipped: %3.2f%%, total refined: %3.2f%%, total smooth: %3.2f%%\n", total_modified * 100.0f / num_blocks, total_skipped * 100.0f / num_blocks, total_refined * 100.0f / num_blocks, total_smooth * 100.0f / num_blocks); |
| 4187 | |
| 4188 | return status; |
| 4189 | } |
| 4190 | } // namespace basisu |
| 4191 | |
| 4192 | |
| 4193 | |
| 4194 | |
| 4195 | |
| 4196 |
Generated on 2023-Sep-17 from project Godot revision 4.2
Powered by 
Code Browser 2.1
Generator usage only permitted with license.