1 | // SPDX-License-Identifier: Apache-2.0 |
2 | // ---------------------------------------------------------------------------- |
3 | // Copyright 2011-2023 Arm Limited |
4 | // |
5 | // Licensed under the Apache License, Version 2.0 (the "License"); you may not |
6 | // use this file except in compliance with the License. You may obtain a copy |
7 | // of the License at: |
8 | // |
9 | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | // |
11 | // Unless required by applicable law or agreed to in writing, software |
12 | // distributed under the License is distributed on an "AS IS" BASIS, WITHOUT |
13 | // WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the |
14 | // License for the specific language governing permissions and limitations |
15 | // under the License. |
16 | // ---------------------------------------------------------------------------- |
17 | |
18 | #if !defined(ASTCENC_DECOMPRESS_ONLY) |
19 | |
20 | /** |
21 | * @brief Functions for finding best partition for a block. |
22 | * |
23 | * The partition search operates in two stages. The first pass uses kmeans clustering to group |
24 | * texels into an ideal partitioning for the requested partition count, and then compares that |
25 | * against the 1024 partitionings generated by the ASTC partition hash function. The generated |
26 | * partitions are then ranked by the number of texels in the wrong partition, compared to the ideal |
27 | * clustering. All 1024 partitions are tested for similarity and ranked, apart from duplicates and |
28 | * partitionings that actually generate fewer than the requested partition count, but only the top |
29 | * N candidates are actually put through a more detailed search. N is determined by the compressor |
30 | * quality preset. |
31 | * |
32 | * For the detailed search, each candidate is checked against two possible encoding methods: |
33 | * |
34 | * - The best partitioning assuming different chroma colors (RGB + RGB or RGB + delta endpoints). |
35 | * - The best partitioning assuming same chroma colors (RGB + scale endpoints). |
36 | * |
37 | * This is implemented by computing the compute mean color and dominant direction for each |
38 | * partition. This defines two lines, both of which go through the mean color value. |
39 | * |
40 | * - One line has a direction defined by the dominant direction; this is used to assess the error |
41 | * from using an uncorrelated color representation. |
42 | * - The other line goes through (0,0,0,1) and is used to assess the error from using a same chroma |
43 | * (RGB + scale) color representation. |
44 | * |
45 | * The best candidate is selected by computing the squared-errors that result from using these |
46 | * lines for endpoint selection. |
47 | */ |
48 | |
49 | #include <limits> |
50 | #include "astcenc_internal.h" |
51 | |
52 | /** |
53 | * @brief Pick some initial kmeans cluster centers. |
54 | * |
55 | * @param blk The image block color data to compress. |
56 | * @param texel_count The number of texels in the block. |
57 | * @param partition_count The number of partitions in the block. |
58 | * @param[out] cluster_centers The initial partition cluster center colors. |
59 | */ |
60 | static void kmeans_init( |
61 | const image_block& blk, |
62 | unsigned int texel_count, |
63 | unsigned int partition_count, |
64 | vfloat4 cluster_centers[BLOCK_MAX_PARTITIONS] |
65 | ) { |
66 | promise(texel_count > 0); |
67 | promise(partition_count > 0); |
68 | |
69 | unsigned int clusters_selected = 0; |
70 | float distances[BLOCK_MAX_TEXELS]; |
71 | |
72 | // Pick a random sample as first cluster center; 145897 from random.org |
73 | unsigned int sample = 145897 % texel_count; |
74 | vfloat4 center_color = blk.texel(sample); |
75 | cluster_centers[clusters_selected] = center_color; |
76 | clusters_selected++; |
77 | |
78 | // Compute the distance to the first cluster center |
79 | float distance_sum = 0.0f; |
80 | for (unsigned int i = 0; i < texel_count; i++) |
81 | { |
82 | vfloat4 color = blk.texel(i); |
83 | vfloat4 diff = color - center_color; |
84 | float distance = dot_s(diff * diff, blk.channel_weight); |
85 | distance_sum += distance; |
86 | distances[i] = distance; |
87 | } |
88 | |
89 | // More numbers from random.org for weighted-random center selection |
90 | const float cluster_cutoffs[9] { |
91 | 0.626220f, 0.932770f, 0.275454f, |
92 | 0.318558f, 0.240113f, 0.009190f, |
93 | 0.347661f, 0.731960f, 0.156391f |
94 | }; |
95 | |
96 | unsigned int cutoff = (clusters_selected - 1) + 3 * (partition_count - 2); |
97 | |
98 | // Pick the remaining samples as needed |
99 | while (true) |
100 | { |
101 | // Pick the next center in a weighted-random fashion. |
102 | float summa = 0.0f; |
103 | float distance_cutoff = distance_sum * cluster_cutoffs[cutoff++]; |
104 | for (sample = 0; sample < texel_count; sample++) |
105 | { |
106 | summa += distances[sample]; |
107 | if (summa >= distance_cutoff) |
108 | { |
109 | break; |
110 | } |
111 | } |
112 | |
113 | // Clamp to a valid range and store the selected cluster center |
114 | sample = astc::min(sample, texel_count - 1); |
115 | |
116 | center_color = blk.texel(sample); |
117 | cluster_centers[clusters_selected++] = center_color; |
118 | if (clusters_selected >= partition_count) |
119 | { |
120 | break; |
121 | } |
122 | |
123 | // Compute the distance to the new cluster center, keep the min dist |
124 | distance_sum = 0.0f; |
125 | for (unsigned int i = 0; i < texel_count; i++) |
126 | { |
127 | vfloat4 color = blk.texel(i); |
128 | vfloat4 diff = color - center_color; |
129 | float distance = dot_s(diff * diff, blk.channel_weight); |
130 | distance = astc::min(distance, distances[i]); |
131 | distance_sum += distance; |
132 | distances[i] = distance; |
133 | } |
134 | } |
135 | } |
136 | |
137 | /** |
138 | * @brief Assign texels to clusters, based on a set of chosen center points. |
139 | * |
140 | * @param blk The image block color data to compress. |
141 | * @param texel_count The number of texels in the block. |
142 | * @param partition_count The number of partitions in the block. |
143 | * @param cluster_centers The partition cluster center colors. |
144 | * @param[out] partition_of_texel The partition assigned for each texel. |
145 | */ |
146 | static void kmeans_assign( |
147 | const image_block& blk, |
148 | unsigned int texel_count, |
149 | unsigned int partition_count, |
150 | const vfloat4 cluster_centers[BLOCK_MAX_PARTITIONS], |
151 | uint8_t partition_of_texel[BLOCK_MAX_TEXELS] |
152 | ) { |
153 | promise(texel_count > 0); |
154 | promise(partition_count > 0); |
155 | |
156 | uint8_t partition_texel_count[BLOCK_MAX_PARTITIONS] { 0 }; |
157 | |
158 | // Find the best partition for every texel |
159 | for (unsigned int i = 0; i < texel_count; i++) |
160 | { |
161 | float best_distance = std::numeric_limits<float>::max(); |
162 | unsigned int best_partition = 0; |
163 | |
164 | vfloat4 color = blk.texel(i); |
165 | for (unsigned int j = 0; j < partition_count; j++) |
166 | { |
167 | vfloat4 diff = color - cluster_centers[j]; |
168 | float distance = dot_s(diff * diff, blk.channel_weight); |
169 | if (distance < best_distance) |
170 | { |
171 | best_distance = distance; |
172 | best_partition = j; |
173 | } |
174 | } |
175 | |
176 | partition_of_texel[i] = static_cast<uint8_t>(best_partition); |
177 | partition_texel_count[best_partition]++; |
178 | } |
179 | |
180 | // It is possible to get a situation where a partition ends up without any texels. In this case, |
181 | // assign texel N to partition N. This is silly, but ensures that every partition retains at |
182 | // least one texel. Reassigning a texel in this manner may cause another partition to go empty, |
183 | // so if we actually did a reassignment, run the whole loop over again. |
184 | bool problem_case; |
185 | do |
186 | { |
187 | problem_case = false; |
188 | for (unsigned int i = 0; i < partition_count; i++) |
189 | { |
190 | if (partition_texel_count[i] == 0) |
191 | { |
192 | partition_texel_count[partition_of_texel[i]]--; |
193 | partition_texel_count[i]++; |
194 | partition_of_texel[i] = static_cast<uint8_t>(i); |
195 | problem_case = true; |
196 | } |
197 | } |
198 | } while (problem_case); |
199 | } |
200 | |
201 | /** |
202 | * @brief Compute new cluster centers based on their center of gravity. |
203 | * |
204 | * @param blk The image block color data to compress. |
205 | * @param texel_count The number of texels in the block. |
206 | * @param partition_count The number of partitions in the block. |
207 | * @param[out] cluster_centers The new cluster center colors. |
208 | * @param partition_of_texel The partition assigned for each texel. |
209 | */ |
210 | static void kmeans_update( |
211 | const image_block& blk, |
212 | unsigned int texel_count, |
213 | unsigned int partition_count, |
214 | vfloat4 cluster_centers[BLOCK_MAX_PARTITIONS], |
215 | const uint8_t partition_of_texel[BLOCK_MAX_TEXELS] |
216 | ) { |
217 | promise(texel_count > 0); |
218 | promise(partition_count > 0); |
219 | |
220 | vfloat4 color_sum[BLOCK_MAX_PARTITIONS] { |
221 | vfloat4::zero(), |
222 | vfloat4::zero(), |
223 | vfloat4::zero(), |
224 | vfloat4::zero() |
225 | }; |
226 | |
227 | uint8_t partition_texel_count[BLOCK_MAX_PARTITIONS] { 0 }; |
228 | |
229 | // Find the center-of-gravity in each cluster |
230 | for (unsigned int i = 0; i < texel_count; i++) |
231 | { |
232 | uint8_t partition = partition_of_texel[i]; |
233 | color_sum[partition] += blk.texel(i); |
234 | partition_texel_count[partition]++; |
235 | } |
236 | |
237 | // Set the center of gravity to be the new cluster center |
238 | for (unsigned int i = 0; i < partition_count; i++) |
239 | { |
240 | float scale = 1.0f / static_cast<float>(partition_texel_count[i]); |
241 | cluster_centers[i] = color_sum[i] * scale; |
242 | } |
243 | } |
244 | |
245 | /** |
246 | * @brief Compute bit-mismatch for partitioning in 2-partition mode. |
247 | * |
248 | * @param a The texel assignment bitvector for the block. |
249 | * @param b The texel assignment bitvector for the partition table. |
250 | * |
251 | * @return The number of bit mismatches. |
252 | */ |
253 | static inline unsigned int partition_mismatch2( |
254 | const uint64_t a[2], |
255 | const uint64_t b[2] |
256 | ) { |
257 | int v1 = popcount(a[0] ^ b[0]) + popcount(a[1] ^ b[1]); |
258 | int v2 = popcount(a[0] ^ b[1]) + popcount(a[1] ^ b[0]); |
259 | return astc::min(v1, v2); |
260 | } |
261 | |
262 | /** |
263 | * @brief Compute bit-mismatch for partitioning in 3-partition mode. |
264 | * |
265 | * @param a The texel assignment bitvector for the block. |
266 | * @param b The texel assignment bitvector for the partition table. |
267 | * |
268 | * @return The number of bit mismatches. |
269 | */ |
270 | static inline unsigned int partition_mismatch3( |
271 | const uint64_t a[3], |
272 | const uint64_t b[3] |
273 | ) { |
274 | int p00 = popcount(a[0] ^ b[0]); |
275 | int p01 = popcount(a[0] ^ b[1]); |
276 | int p02 = popcount(a[0] ^ b[2]); |
277 | |
278 | int p10 = popcount(a[1] ^ b[0]); |
279 | int p11 = popcount(a[1] ^ b[1]); |
280 | int p12 = popcount(a[1] ^ b[2]); |
281 | |
282 | int p20 = popcount(a[2] ^ b[0]); |
283 | int p21 = popcount(a[2] ^ b[1]); |
284 | int p22 = popcount(a[2] ^ b[2]); |
285 | |
286 | int s0 = p11 + p22; |
287 | int s1 = p12 + p21; |
288 | int v0 = astc::min(s0, s1) + p00; |
289 | |
290 | int s2 = p10 + p22; |
291 | int s3 = p12 + p20; |
292 | int v1 = astc::min(s2, s3) + p01; |
293 | |
294 | int s4 = p10 + p21; |
295 | int s5 = p11 + p20; |
296 | int v2 = astc::min(s4, s5) + p02; |
297 | |
298 | return astc::min(v0, v1, v2); |
299 | } |
300 | |
301 | /** |
302 | * @brief Compute bit-mismatch for partitioning in 4-partition mode. |
303 | * |
304 | * @param a The texel assignment bitvector for the block. |
305 | * @param b The texel assignment bitvector for the partition table. |
306 | * |
307 | * @return The number of bit mismatches. |
308 | */ |
309 | static inline unsigned int partition_mismatch4( |
310 | const uint64_t a[4], |
311 | const uint64_t b[4] |
312 | ) { |
313 | int p00 = popcount(a[0] ^ b[0]); |
314 | int p01 = popcount(a[0] ^ b[1]); |
315 | int p02 = popcount(a[0] ^ b[2]); |
316 | int p03 = popcount(a[0] ^ b[3]); |
317 | |
318 | int p10 = popcount(a[1] ^ b[0]); |
319 | int p11 = popcount(a[1] ^ b[1]); |
320 | int p12 = popcount(a[1] ^ b[2]); |
321 | int p13 = popcount(a[1] ^ b[3]); |
322 | |
323 | int p20 = popcount(a[2] ^ b[0]); |
324 | int p21 = popcount(a[2] ^ b[1]); |
325 | int p22 = popcount(a[2] ^ b[2]); |
326 | int p23 = popcount(a[2] ^ b[3]); |
327 | |
328 | int p30 = popcount(a[3] ^ b[0]); |
329 | int p31 = popcount(a[3] ^ b[1]); |
330 | int p32 = popcount(a[3] ^ b[2]); |
331 | int p33 = popcount(a[3] ^ b[3]); |
332 | |
333 | int mx23 = astc::min(p22 + p33, p23 + p32); |
334 | int mx13 = astc::min(p21 + p33, p23 + p31); |
335 | int mx12 = astc::min(p21 + p32, p22 + p31); |
336 | int mx03 = astc::min(p20 + p33, p23 + p30); |
337 | int mx02 = astc::min(p20 + p32, p22 + p30); |
338 | int mx01 = astc::min(p21 + p30, p20 + p31); |
339 | |
340 | int v0 = p00 + astc::min(p11 + mx23, p12 + mx13, p13 + mx12); |
341 | int v1 = p01 + astc::min(p10 + mx23, p12 + mx03, p13 + mx02); |
342 | int v2 = p02 + astc::min(p11 + mx03, p10 + mx13, p13 + mx01); |
343 | int v3 = p03 + astc::min(p11 + mx02, p12 + mx01, p10 + mx12); |
344 | |
345 | return astc::min(v0, v1, v2, v3); |
346 | } |
347 | |
348 | using mismatch_dispatch = unsigned int (*)(const uint64_t*, const uint64_t*); |
349 | |
350 | /** |
351 | * @brief Count the partition table mismatches vs the data clustering. |
352 | * |
353 | * @param bsd The block size information. |
354 | * @param partition_count The number of partitions in the block. |
355 | * @param bitmaps The block texel partition assignment patterns. |
356 | * @param[out] mismatch_counts The array storing per partitioning mismatch counts. |
357 | */ |
358 | static void count_partition_mismatch_bits( |
359 | const block_size_descriptor& bsd, |
360 | unsigned int partition_count, |
361 | const uint64_t bitmaps[BLOCK_MAX_PARTITIONS], |
362 | unsigned int mismatch_counts[BLOCK_MAX_PARTITIONINGS] |
363 | ) { |
364 | unsigned int active_count = bsd.partitioning_count_selected[partition_count - 1]; |
365 | promise(active_count > 0); |
366 | |
367 | if (partition_count == 2) |
368 | { |
369 | for (unsigned int i = 0; i < active_count; i++) |
370 | { |
371 | mismatch_counts[i] = partition_mismatch2(bitmaps, bsd.coverage_bitmaps_2[i]); |
372 | } |
373 | } |
374 | else if (partition_count == 3) |
375 | { |
376 | for (unsigned int i = 0; i < active_count; i++) |
377 | { |
378 | mismatch_counts[i] = partition_mismatch3(bitmaps, bsd.coverage_bitmaps_3[i]); |
379 | } |
380 | } |
381 | else |
382 | { |
383 | for (unsigned int i = 0; i < active_count; i++) |
384 | { |
385 | mismatch_counts[i] = partition_mismatch4(bitmaps, bsd.coverage_bitmaps_4[i]); |
386 | } |
387 | } |
388 | } |
389 | |
390 | /** |
391 | * @brief Use counting sort on the mismatch array to sort partition candidates. |
392 | * |
393 | * @param partitioning_count The number of packed partitionings. |
394 | * @param mismatch_count Partitioning mismatch counts, in index order. |
395 | * @param[out] partition_ordering Partition index values, in mismatch order. |
396 | * |
397 | * @return The number of active partitions in this selection. |
398 | */ |
399 | static unsigned int get_partition_ordering_by_mismatch_bits( |
400 | unsigned int partitioning_count, |
401 | const unsigned int mismatch_count[BLOCK_MAX_PARTITIONINGS], |
402 | unsigned int partition_ordering[BLOCK_MAX_PARTITIONINGS] |
403 | ) { |
404 | promise(partitioning_count > 0); |
405 | unsigned int mscount[256] { 0 }; |
406 | |
407 | // Create the histogram of mismatch counts |
408 | for (unsigned int i = 0; i < partitioning_count; i++) |
409 | { |
410 | mscount[mismatch_count[i]]++; |
411 | } |
412 | |
413 | unsigned int active_count = partitioning_count - mscount[255]; |
414 | |
415 | // Create a running sum from the histogram array |
416 | // Cells store previous values only; i.e. exclude self after sum |
417 | unsigned int summa = 0; |
418 | for (unsigned int i = 0; i < 256; i++) |
419 | { |
420 | unsigned int cnt = mscount[i]; |
421 | mscount[i] = summa; |
422 | summa += cnt; |
423 | } |
424 | |
425 | // Use the running sum as the index, incrementing after read to allow |
426 | // sequential entries with the same count |
427 | for (unsigned int i = 0; i < partitioning_count; i++) |
428 | { |
429 | unsigned int idx = mscount[mismatch_count[i]]++; |
430 | partition_ordering[idx] = i; |
431 | } |
432 | |
433 | return active_count; |
434 | } |
435 | |
436 | /** |
437 | * @brief Use k-means clustering to compute a partition ordering for a block.. |
438 | * |
439 | * @param bsd The block size information. |
440 | * @param blk The image block color data to compress. |
441 | * @param partition_count The desired number of partitions in the block. |
442 | * @param[out] partition_ordering The list of recommended partition indices, in priority order. |
443 | * |
444 | * @return The number of active partitionings in this selection. |
445 | */ |
446 | static unsigned int compute_kmeans_partition_ordering( |
447 | const block_size_descriptor& bsd, |
448 | const image_block& blk, |
449 | unsigned int partition_count, |
450 | unsigned int partition_ordering[BLOCK_MAX_PARTITIONINGS] |
451 | ) { |
452 | vfloat4 cluster_centers[BLOCK_MAX_PARTITIONS]; |
453 | uint8_t texel_partitions[BLOCK_MAX_TEXELS]; |
454 | |
455 | // Use three passes of k-means clustering to partition the block data |
456 | for (unsigned int i = 0; i < 3; i++) |
457 | { |
458 | if (i == 0) |
459 | { |
460 | kmeans_init(blk, bsd.texel_count, partition_count, cluster_centers); |
461 | } |
462 | else |
463 | { |
464 | kmeans_update(blk, bsd.texel_count, partition_count, cluster_centers, texel_partitions); |
465 | } |
466 | |
467 | kmeans_assign(blk, bsd.texel_count, partition_count, cluster_centers, texel_partitions); |
468 | } |
469 | |
470 | // Construct the block bitmaps of texel assignments to each partition |
471 | uint64_t bitmaps[BLOCK_MAX_PARTITIONS] { 0 }; |
472 | unsigned int texels_to_process = astc::min(bsd.texel_count, BLOCK_MAX_KMEANS_TEXELS); |
473 | promise(texels_to_process > 0); |
474 | for (unsigned int i = 0; i < texels_to_process; i++) |
475 | { |
476 | unsigned int idx = bsd.kmeans_texels[i]; |
477 | bitmaps[texel_partitions[idx]] |= 1ULL << i; |
478 | } |
479 | |
480 | // Count the mismatch between the block and the format's partition tables |
481 | unsigned int mismatch_counts[BLOCK_MAX_PARTITIONINGS]; |
482 | count_partition_mismatch_bits(bsd, partition_count, bitmaps, mismatch_counts); |
483 | |
484 | // Sort the partitions based on the number of mismatched bits |
485 | return get_partition_ordering_by_mismatch_bits( |
486 | bsd.partitioning_count_selected[partition_count - 1], |
487 | mismatch_counts, partition_ordering); |
488 | } |
489 | |
490 | /** |
491 | * @brief Insert a partitioning into an order list of results, sorted by error. |
492 | * |
493 | * @param max_values The max number of entries in the best result arrays. |
494 | * @param this_error The error of the new entry. |
495 | * @param this_partition The partition ID of the new entry. |
496 | * @param[out] best_errors The array of best error values. |
497 | * @param[out] best_partitions The array of best partition values. |
498 | */ |
499 | static void insert_result( |
500 | unsigned int max_values, |
501 | float this_error, |
502 | unsigned int this_partition, |
503 | float* best_errors, |
504 | unsigned int* best_partitions) |
505 | { |
506 | promise(max_values > 0); |
507 | |
508 | // Don't bother searching if the current worst error beats the new error |
509 | if (this_error >= best_errors[max_values - 1]) |
510 | { |
511 | return; |
512 | } |
513 | |
514 | // Else insert into the list in error-order |
515 | for (unsigned int i = 0; i < max_values; i++) |
516 | { |
517 | // Existing result is better - move on ... |
518 | if (this_error > best_errors[i]) |
519 | { |
520 | continue; |
521 | } |
522 | |
523 | // Move existing results down one |
524 | for (unsigned int j = max_values - 1; j > i; j--) |
525 | { |
526 | best_errors[j] = best_errors[j - 1]; |
527 | best_partitions[j] = best_partitions[j - 1]; |
528 | } |
529 | |
530 | // Insert new result |
531 | best_errors[i] = this_error; |
532 | best_partitions[i] = this_partition; |
533 | break; |
534 | } |
535 | } |
536 | |
537 | /* See header for documentation. */ |
538 | unsigned int find_best_partition_candidates( |
539 | const block_size_descriptor& bsd, |
540 | const image_block& blk, |
541 | unsigned int partition_count, |
542 | unsigned int partition_search_limit, |
543 | unsigned int best_partitions[TUNE_MAX_PARTITIONING_CANDIDATES], |
544 | unsigned int requested_candidates |
545 | ) { |
546 | // Constant used to estimate quantization error for a given partitioning; the optimal value for |
547 | // this depends on bitrate. These values have been determined empirically. |
548 | unsigned int texels_per_block = bsd.texel_count; |
549 | float weight_imprecision_estim = 0.055f; |
550 | if (texels_per_block <= 20) |
551 | { |
552 | weight_imprecision_estim = 0.03f; |
553 | } |
554 | else if (texels_per_block <= 31) |
555 | { |
556 | weight_imprecision_estim = 0.04f; |
557 | } |
558 | else if (texels_per_block <= 41) |
559 | { |
560 | weight_imprecision_estim = 0.05f; |
561 | } |
562 | |
563 | promise(partition_count > 0); |
564 | promise(partition_search_limit > 0); |
565 | |
566 | weight_imprecision_estim = weight_imprecision_estim * weight_imprecision_estim; |
567 | |
568 | unsigned int partition_sequence[BLOCK_MAX_PARTITIONINGS]; |
569 | unsigned int sequence_len = compute_kmeans_partition_ordering(bsd, blk, partition_count, partition_sequence); |
570 | partition_search_limit = astc::min(partition_search_limit, sequence_len); |
571 | requested_candidates = astc::min(partition_search_limit, requested_candidates); |
572 | |
573 | bool uses_alpha = !blk.is_constant_channel(3); |
574 | |
575 | // Partitioning errors assuming uncorrelated-chrominance endpoints |
576 | float uncor_best_errors[TUNE_MAX_PARTITIONING_CANDIDATES]; |
577 | unsigned int uncor_best_partitions[TUNE_MAX_PARTITIONING_CANDIDATES]; |
578 | |
579 | // Partitioning errors assuming same-chrominance endpoints |
580 | float samec_best_errors[TUNE_MAX_PARTITIONING_CANDIDATES]; |
581 | unsigned int samec_best_partitions[TUNE_MAX_PARTITIONING_CANDIDATES]; |
582 | |
583 | for (unsigned int i = 0; i < requested_candidates; i++) |
584 | { |
585 | uncor_best_errors[i] = ERROR_CALC_DEFAULT; |
586 | samec_best_errors[i] = ERROR_CALC_DEFAULT; |
587 | } |
588 | |
589 | if (uses_alpha) |
590 | { |
591 | for (unsigned int i = 0; i < partition_search_limit; i++) |
592 | { |
593 | unsigned int partition = partition_sequence[i]; |
594 | const auto& pi = bsd.get_raw_partition_info(partition_count, partition); |
595 | |
596 | // Compute weighting to give to each component in each partition |
597 | partition_metrics pms[BLOCK_MAX_PARTITIONS]; |
598 | |
599 | compute_avgs_and_dirs_4_comp(pi, blk, pms); |
600 | |
601 | line4 uncor_lines[BLOCK_MAX_PARTITIONS]; |
602 | line4 samec_lines[BLOCK_MAX_PARTITIONS]; |
603 | |
604 | processed_line4 uncor_plines[BLOCK_MAX_PARTITIONS]; |
605 | processed_line4 samec_plines[BLOCK_MAX_PARTITIONS]; |
606 | |
607 | float line_lengths[BLOCK_MAX_PARTITIONS]; |
608 | |
609 | for (unsigned int j = 0; j < partition_count; j++) |
610 | { |
611 | partition_metrics& pm = pms[j]; |
612 | |
613 | uncor_lines[j].a = pm.avg; |
614 | uncor_lines[j].b = normalize_safe(pm.dir, unit4()); |
615 | |
616 | uncor_plines[j].amod = uncor_lines[j].a - uncor_lines[j].b * dot(uncor_lines[j].a, uncor_lines[j].b); |
617 | uncor_plines[j].bs = uncor_lines[j].b; |
618 | |
619 | samec_lines[j].a = vfloat4::zero(); |
620 | samec_lines[j].b = normalize_safe(pm.avg, unit4()); |
621 | |
622 | samec_plines[j].amod = vfloat4::zero(); |
623 | samec_plines[j].bs = samec_lines[j].b; |
624 | } |
625 | |
626 | float uncor_error = 0.0f; |
627 | float samec_error = 0.0f; |
628 | |
629 | compute_error_squared_rgba(pi, |
630 | blk, |
631 | uncor_plines, |
632 | samec_plines, |
633 | line_lengths, |
634 | uncor_error, |
635 | samec_error); |
636 | |
637 | // Compute an estimate of error introduced by weight quantization imprecision. |
638 | // This error is computed as follows, for each partition |
639 | // 1: compute the principal-axis vector (full length) in error-space |
640 | // 2: convert the principal-axis vector to regular RGB-space |
641 | // 3: scale the vector by a constant that estimates average quantization error |
642 | // 4: for each texel, square the vector, then do a dot-product with the texel's |
643 | // error weight; sum up the results across all texels. |
644 | // 4(optimized): square the vector once, then do a dot-product with the average |
645 | // texel error, then multiply by the number of texels. |
646 | |
647 | for (unsigned int j = 0; j < partition_count; j++) |
648 | { |
649 | float tpp = static_cast<float>(pi.partition_texel_count[j]); |
650 | vfloat4 error_weights(tpp * weight_imprecision_estim); |
651 | |
652 | vfloat4 uncor_vector = uncor_lines[j].b * line_lengths[j]; |
653 | vfloat4 samec_vector = samec_lines[j].b * line_lengths[j]; |
654 | |
655 | uncor_error += dot_s(uncor_vector * uncor_vector, error_weights); |
656 | samec_error += dot_s(samec_vector * samec_vector, error_weights); |
657 | } |
658 | |
659 | insert_result(requested_candidates, uncor_error, partition, uncor_best_errors, uncor_best_partitions); |
660 | insert_result(requested_candidates, samec_error, partition, samec_best_errors, samec_best_partitions); |
661 | } |
662 | } |
663 | else |
664 | { |
665 | for (unsigned int i = 0; i < partition_search_limit; i++) |
666 | { |
667 | unsigned int partition = partition_sequence[i]; |
668 | const auto& pi = bsd.get_raw_partition_info(partition_count, partition); |
669 | |
670 | // Compute weighting to give to each component in each partition |
671 | partition_metrics pms[BLOCK_MAX_PARTITIONS]; |
672 | compute_avgs_and_dirs_3_comp_rgb(pi, blk, pms); |
673 | |
674 | partition_lines3 plines[BLOCK_MAX_PARTITIONS]; |
675 | |
676 | for (unsigned int j = 0; j < partition_count; j++) |
677 | { |
678 | partition_metrics& pm = pms[j]; |
679 | partition_lines3& pl = plines[j]; |
680 | |
681 | pl.uncor_line.a = pm.avg; |
682 | pl.uncor_line.b = normalize_safe(pm.dir, unit3()); |
683 | |
684 | pl.samec_line.a = vfloat4::zero(); |
685 | pl.samec_line.b = normalize_safe(pm.avg, unit3()); |
686 | |
687 | pl.uncor_pline.amod = pl.uncor_line.a - pl.uncor_line.b * dot3(pl.uncor_line.a, pl.uncor_line.b); |
688 | pl.uncor_pline.bs = pl.uncor_line.b; |
689 | |
690 | pl.samec_pline.amod = vfloat4::zero(); |
691 | pl.samec_pline.bs = pl.samec_line.b; |
692 | } |
693 | |
694 | float uncor_error = 0.0f; |
695 | float samec_error = 0.0f; |
696 | |
697 | compute_error_squared_rgb(pi, |
698 | blk, |
699 | plines, |
700 | uncor_error, |
701 | samec_error); |
702 | |
703 | // Compute an estimate of error introduced by weight quantization imprecision. |
704 | // This error is computed as follows, for each partition |
705 | // 1: compute the principal-axis vector (full length) in error-space |
706 | // 2: convert the principal-axis vector to regular RGB-space |
707 | // 3: scale the vector by a constant that estimates average quantization error |
708 | // 4: for each texel, square the vector, then do a dot-product with the texel's |
709 | // error weight; sum up the results across all texels. |
710 | // 4(optimized): square the vector once, then do a dot-product with the average |
711 | // texel error, then multiply by the number of texels. |
712 | |
713 | for (unsigned int j = 0; j < partition_count; j++) |
714 | { |
715 | partition_lines3& pl = plines[j]; |
716 | |
717 | float tpp = static_cast<float>(pi.partition_texel_count[j]); |
718 | vfloat4 error_weights(tpp * weight_imprecision_estim); |
719 | |
720 | vfloat4 uncor_vector = pl.uncor_line.b * pl.line_length; |
721 | vfloat4 samec_vector = pl.samec_line.b * pl.line_length; |
722 | |
723 | uncor_error += dot3_s(uncor_vector * uncor_vector, error_weights); |
724 | samec_error += dot3_s(samec_vector * samec_vector, error_weights); |
725 | } |
726 | |
727 | insert_result(requested_candidates, uncor_error, partition, uncor_best_errors, uncor_best_partitions); |
728 | insert_result(requested_candidates, samec_error, partition, samec_best_errors, samec_best_partitions); |
729 | } |
730 | } |
731 | |
732 | unsigned int interleave[2 * TUNE_MAX_PARTITIONING_CANDIDATES]; |
733 | for (unsigned int i = 0; i < requested_candidates; i++) |
734 | { |
735 | interleave[2 * i] = bsd.get_raw_partition_info(partition_count, uncor_best_partitions[i]).partition_index; |
736 | interleave[2 * i + 1] = bsd.get_raw_partition_info(partition_count, samec_best_partitions[i]).partition_index; |
737 | } |
738 | |
739 | uint64_t bitmasks[1024/64] { 0 }; |
740 | unsigned int emitted = 0; |
741 | |
742 | // Deduplicate the first "requested" entries |
743 | for (unsigned int i = 0; i < requested_candidates * 2; i++) |
744 | { |
745 | unsigned int partition = interleave[i]; |
746 | |
747 | unsigned int word = partition / 64; |
748 | unsigned int bit = partition % 64; |
749 | |
750 | bool written = bitmasks[word] & (1ull << bit); |
751 | |
752 | if (!written) |
753 | { |
754 | best_partitions[emitted] = partition; |
755 | bitmasks[word] |= 1ull << bit; |
756 | emitted++; |
757 | |
758 | if (emitted == requested_candidates) |
759 | { |
760 | break; |
761 | } |
762 | } |
763 | } |
764 | |
765 | return emitted; |
766 | } |
767 | |
768 | #endif |
769 | |