1// SPDX-License-Identifier: Apache-2.0
2// ----------------------------------------------------------------------------
3// Copyright 2011-2024 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/**
19 * @brief Functions and data declarations.
20 */
21
22#ifndef ASTCENC_INTERNAL_INCLUDED
23#define ASTCENC_INTERNAL_INCLUDED
24
25#include <algorithm>
26#include <cstddef>
27#include <cstdint>
28#if defined(ASTCENC_DIAGNOSTICS)
29	#include <cstdio>
30#endif
31#include <cstdlib>
32#include <limits>
33
34#ifdef ASTC_CUSTOMIZED_ENABLE
35#include <unistd.h>
36#include <string>
37#if defined(_WIN32) && !defined(__CYGWIN__)
38#define NOMINMAX
39#include <windows.h>
40#include <io.h>
41#else
42#include <dlfcn.h>
43#endif
44#endif
45
46#include "astcenc.h"
47#include "astcenc_mathlib.h"
48#include "astcenc_vecmathlib.h"
49
50/**
51 * @brief Make a promise to the compiler's optimizer.
52 *
53 * A promise is an expression that the optimizer is can assume is true for to help it generate
54 * faster code. Common use cases for this are to promise that a for loop will iterate more than
55 * once, or that the loop iteration count is a multiple of a vector length, which avoids pre-loop
56 * checks and can avoid loop tails if loops are unrolled by the auto-vectorizer.
57 */
58#if defined(NDEBUG)
59	#if !defined(__clang__) && defined(_MSC_VER)
60		#define promise(cond) __assume(cond)
61	#elif defined(__clang__)
62		#if __has_builtin(__builtin_assume)
63			#define promise(cond) __builtin_assume(cond)
64		#elif __has_builtin(__builtin_unreachable)
65			#define promise(cond) if (!(cond)) { __builtin_unreachable(); }
66		#else
67			#define promise(cond)
68		#endif
69	#else // Assume GCC
70		#define promise(cond) if (!(cond)) { __builtin_unreachable(); }
71	#endif
72#else
73	#define promise(cond) assert(cond)
74#endif
75
76/* ============================================================================
77  Constants
78============================================================================ */
79#if !defined(ASTCENC_BLOCK_MAX_TEXELS)
80	#define ASTCENC_BLOCK_MAX_TEXELS 216 // A 3D 6x6x6 block
81#endif
82
83/** @brief The maximum number of texels a block can support (6x6x6 block). */
84static constexpr unsigned int BLOCK_MAX_TEXELS { ASTCENC_BLOCK_MAX_TEXELS };
85
86/** @brief The maximum number of components a block can support. */
87static constexpr unsigned int BLOCK_MAX_COMPONENTS { 4 };
88
89/** @brief The maximum number of partitions a block can support. */
90static constexpr unsigned int BLOCK_MAX_PARTITIONS { 4 };
91
92/** @brief The number of partitionings, per partition count, suported by the ASTC format. */
93static constexpr unsigned int BLOCK_MAX_PARTITIONINGS { 1024 };
94
95/** @brief The maximum number of texels used during partition selection for texel clustering. */
96static constexpr uint8_t BLOCK_MAX_KMEANS_TEXELS { 64 };
97
98/** @brief The maximum number of weights a block can support. */
99static constexpr unsigned int BLOCK_MAX_WEIGHTS { 64 };
100
101/** @brief The maximum number of weights a block can support per plane in 2 plane mode. */
102static constexpr unsigned int BLOCK_MAX_WEIGHTS_2PLANE { BLOCK_MAX_WEIGHTS / 2 };
103
104/** @brief The minimum number of weight bits a candidate encoding must encode. */
105static constexpr unsigned int BLOCK_MIN_WEIGHT_BITS { 24 };
106
107/** @brief The maximum number of weight bits a candidate encoding can encode. */
108static constexpr unsigned int BLOCK_MAX_WEIGHT_BITS { 96 };
109
110/** @brief The index indicating a bad (unused) block mode in the remap array. */
111static constexpr uint16_t BLOCK_BAD_BLOCK_MODE { 0xFFFFu };
112
113/** @brief The index indicating a bad (unused) partitioning in the remap array. */
114static constexpr uint16_t BLOCK_BAD_PARTITIONING { 0xFFFFu };
115
116/** @brief The number of partition index bits supported by the ASTC format . */
117static constexpr unsigned int PARTITION_INDEX_BITS { 10 };
118
119/** @brief The offset of the plane 2 weights in shared weight arrays. */
120static constexpr unsigned int WEIGHTS_PLANE2_OFFSET { BLOCK_MAX_WEIGHTS_2PLANE };
121
122/** @brief The sum of quantized weights for one texel. */
123static constexpr float WEIGHTS_TEXEL_SUM { 16.0f };
124
125/** @brief The number of block modes supported by the ASTC format. */
126static constexpr unsigned int WEIGHTS_MAX_BLOCK_MODES { 2048 };
127
128/** @brief The number of weight grid decimation modes supported by the ASTC format. */
129static constexpr unsigned int WEIGHTS_MAX_DECIMATION_MODES { 87 };
130
131/** @brief The high default error used to initialize error trackers. */
132static constexpr float ERROR_CALC_DEFAULT { 1e30f };
133
134/**
135 * @brief The minimum tuning setting threshold for the one partition fast path.
136 */
137static constexpr float TUNE_MIN_SEARCH_MODE0 { 0.85f };
138
139/**
140 * @brief The maximum number of candidate encodings tested for each encoding mode.
141 *
142 * This can be dynamically reduced by the compression quality preset.
143 */
144static constexpr unsigned int TUNE_MAX_TRIAL_CANDIDATES { 8 };
145
146/**
147 * @brief The maximum number of candidate partitionings tested for each encoding mode.
148 *
149 * This can be dynamically reduced by the compression quality preset.
150 */
151static constexpr unsigned int TUNE_MAX_PARTITIONING_CANDIDATES { 8 };
152
153/**
154 * @brief The maximum quant level using full angular endpoint search method.
155 *
156 * The angular endpoint search is used to find the min/max weight that should
157 * be used for a given quantization level. It is effective but expensive, so
158 * we only use it where it has the most value - low quant levels with wide
159 * spacing. It is used below TUNE_MAX_ANGULAR_QUANT (inclusive). Above this we
160 * assume the min weight is 0.0f, and the max weight is 1.0f.
161 *
162 * Note the angular algorithm is vectorized, and using QUANT_12 exactly fills
163 * one 8-wide vector. Decreasing by one doesn't buy much performance, and
164 * increasing by one is disproportionately expensive.
165 */
166static constexpr unsigned int TUNE_MAX_ANGULAR_QUANT { 7 }; /* QUANT_12 */
167
168static_assert((BLOCK_MAX_TEXELS % ASTCENC_SIMD_WIDTH) == 0,
169              "BLOCK_MAX_TEXELS must be multiple of ASTCENC_SIMD_WIDTH");
170
171static_assert(BLOCK_MAX_TEXELS <= 216,
172              "BLOCK_MAX_TEXELS must not be greater than 216");
173
174static_assert((BLOCK_MAX_WEIGHTS % ASTCENC_SIMD_WIDTH) == 0,
175              "BLOCK_MAX_WEIGHTS must be multiple of ASTCENC_SIMD_WIDTH");
176
177static_assert((WEIGHTS_MAX_BLOCK_MODES % ASTCENC_SIMD_WIDTH) == 0,
178              "WEIGHTS_MAX_BLOCK_MODES must be multiple of ASTCENC_SIMD_WIDTH");
179
180
181/* ============================================================================
182  Commonly used data structures
183============================================================================ */
184
185/**
186 * @brief The ASTC endpoint formats.
187 *
188 * Note, the values here are used directly in the encoding in the format so do not rearrange.
189 */
190enum endpoint_formats
191{
192	FMT_LUMINANCE = 0,
193	FMT_LUMINANCE_DELTA = 1,
194	FMT_HDR_LUMINANCE_LARGE_RANGE = 2,
195	FMT_HDR_LUMINANCE_SMALL_RANGE = 3,
196	FMT_LUMINANCE_ALPHA = 4,
197	FMT_LUMINANCE_ALPHA_DELTA = 5,
198	FMT_RGB_SCALE = 6,
199	FMT_HDR_RGB_SCALE = 7,
200	FMT_RGB = 8,
201	FMT_RGB_DELTA = 9,
202	FMT_RGB_SCALE_ALPHA = 10,
203	FMT_HDR_RGB = 11,
204	FMT_RGBA = 12,
205	FMT_RGBA_DELTA = 13,
206	FMT_HDR_RGB_LDR_ALPHA = 14,
207	FMT_HDR_RGBA = 15
208};
209
210/**
211 * @brief The ASTC quantization methods.
212 *
213 * Note, the values here are used directly in the encoding in the format so do not rearrange.
214 */
215enum quant_method
216{
217	QUANT_2 = 0,
218	QUANT_3 = 1,
219	QUANT_4 = 2,
220	QUANT_5 = 3,
221	QUANT_6 = 4,
222	QUANT_8 = 5,
223	QUANT_10 = 6,
224	QUANT_12 = 7,
225	QUANT_16 = 8,
226	QUANT_20 = 9,
227	QUANT_24 = 10,
228	QUANT_32 = 11,
229	QUANT_40 = 12,
230	QUANT_48 = 13,
231	QUANT_64 = 14,
232	QUANT_80 = 15,
233	QUANT_96 = 16,
234	QUANT_128 = 17,
235	QUANT_160 = 18,
236	QUANT_192 = 19,
237	QUANT_256 = 20
238};
239
240/**
241 * @brief The number of levels use by an ASTC quantization method.
242 *
243 * @param method   The quantization method
244 *
245 * @return   The number of levels used by @c method.
246 */
247static inline unsigned int get_quant_level(quant_method method)
248{
249	switch (method)
250	{
251	case QUANT_2:   return   2;
252	case QUANT_3:   return   3;
253	case QUANT_4:   return   4;
254	case QUANT_5:   return   5;
255	case QUANT_6:   return   6;
256	case QUANT_8:   return   8;
257	case QUANT_10:  return  10;
258	case QUANT_12:  return  12;
259	case QUANT_16:  return  16;
260	case QUANT_20:  return  20;
261	case QUANT_24:  return  24;
262	case QUANT_32:  return  32;
263	case QUANT_40:  return  40;
264	case QUANT_48:  return  48;
265	case QUANT_64:  return  64;
266	case QUANT_80:  return  80;
267	case QUANT_96:  return  96;
268	case QUANT_128: return 128;
269	case QUANT_160: return 160;
270	case QUANT_192: return 192;
271	case QUANT_256: return 256;
272	}
273
274	// Unreachable - the enum is fully described
275	return 0;
276}
277
278/**
279 * @brief Computed metrics about a partition in a block.
280 */
281struct partition_metrics
282{
283	/** @brief The error-weighted average color in the partition. */
284	vfloat4 avg;
285
286	/** @brief The dominant error-weighted direction in the partition. */
287	vfloat4 dir;
288};
289
290/**
291 * @brief Computed lines for a a three component analysis.
292 */
293struct partition_lines3
294{
295	/** @brief Line for uncorrelated chroma. */
296	line3 uncor_line;
297
298	/** @brief Line for correlated chroma, passing though the origin. */
299	line3 samec_line;
300
301	/** @brief Post-processed line for uncorrelated chroma. */
302	processed_line3 uncor_pline;
303
304	/** @brief Post-processed line for correlated chroma, passing though the origin. */
305	processed_line3 samec_pline;
306
307	/**
308	 * @brief The length of the line for uncorrelated chroma.
309	 *
310	 * This is used for both the uncorrelated and same chroma lines - they are normally very similar
311	 * and only used for the relative ranking of partitionings against one another.
312	 */
313	float line_length;
314};
315
316/**
317 * @brief The partition information for a single partition.
318 *
319 * ASTC has a total of 1024 candidate partitions for each of 2/3/4 partition counts, although this
320 * 1024 includes seeds that generate duplicates of other seeds and seeds that generate completely
321 * empty partitions. These are both valid encodings, but astcenc will skip both during compression
322 * as they are not useful.
323 */
324struct partition_info
325{
326	/** @brief The number of partitions in this partitioning. */
327	uint16_t partition_count;
328
329	/** @brief The index (seed) of this partitioning. */
330	uint16_t partition_index;
331
332	/**
333	 * @brief The number of texels in each partition.
334	 *
335	 * Note that some seeds result in zero texels assigned to a partition. These are valid, but are
336	 * skipped by this compressor as there is no point spending bits encoding an unused endpoints.
337	 */
338	uint8_t partition_texel_count[BLOCK_MAX_PARTITIONS];
339
340	/** @brief The partition of each texel in the block. */
341	uint8_t partition_of_texel[BLOCK_MAX_TEXELS];
342
343	/** @brief The list of texels in each partition. */
344	uint8_t texels_of_partition[BLOCK_MAX_PARTITIONS][BLOCK_MAX_TEXELS];
345};
346
347/**
348 * @brief The weight grid information for a single decimation pattern.
349 *
350 * ASTC can store one weight per texel, but is also capable of storing lower resolution weight grids
351 * that are interpolated during decompression to assign a with to a texel. Storing fewer weights
352 * can free up a substantial amount of bits that we can then spend on more useful things, such as
353 * more accurate endpoints and weights, or additional partitions.
354 *
355 * This data structure is used to store information about a single weight grid decimation pattern,
356 * for a single block size.
357 */
358struct decimation_info
359{
360	/** @brief The total number of texels in the block. */
361	uint8_t texel_count;
362
363	/** @brief The maximum number of stored weights that contribute to each texel, between 1 and 4. */
364	uint8_t max_texel_weight_count;
365
366	/** @brief The total number of weights stored. */
367	uint8_t weight_count;
368
369	/** @brief The number of stored weights in the X dimension. */
370	uint8_t weight_x;
371
372	/** @brief The number of stored weights in the Y dimension. */
373	uint8_t weight_y;
374
375	/** @brief The number of stored weights in the Z dimension. */
376	uint8_t weight_z;
377
378	/**
379	 * @brief The number of weights that contribute to each texel.
380	 * Value is between 1 and 4.
381	 */
382	uint8_t texel_weight_count[BLOCK_MAX_TEXELS];
383
384	/**
385	 * @brief The weight index of the N weights that are interpolated for each texel.
386	 * Stored transposed to improve vectorization.
387	 */
388	uint8_t texel_weights_tr[4][BLOCK_MAX_TEXELS];
389
390	/**
391	 * @brief The bilinear contribution of the N weights that are interpolated for each texel.
392	 * Value is between 0 and 16, stored transposed to improve vectorization.
393	 */
394	uint8_t texel_weight_contribs_int_tr[4][BLOCK_MAX_TEXELS];
395
396	/**
397	 * @brief The bilinear contribution of the N weights that are interpolated for each texel.
398	 * Value is between 0 and 1, stored transposed to improve vectorization.
399	 */
400	ASTCENC_ALIGNAS float texel_weight_contribs_float_tr[4][BLOCK_MAX_TEXELS];
401
402	/** @brief The number of texels that each stored weight contributes to. */
403	uint8_t weight_texel_count[BLOCK_MAX_WEIGHTS];
404
405	/**
406	 * @brief The list of texels that use a specific weight index.
407	 * Stored transposed to improve vectorization.
408	 */
409	uint8_t weight_texels_tr[BLOCK_MAX_TEXELS][BLOCK_MAX_WEIGHTS];
410
411	/**
412	 * @brief The bilinear contribution to the N texels that use each weight.
413	 * Value is between 0 and 1, stored transposed to improve vectorization.
414	 */
415	ASTCENC_ALIGNAS float weights_texel_contribs_tr[BLOCK_MAX_TEXELS][BLOCK_MAX_WEIGHTS];
416
417	/**
418	 * @brief The bilinear contribution to the Nth texel that uses each weight.
419	 * Value is between 0 and 1, stored transposed to improve vectorization.
420	 */
421	float texel_contrib_for_weight[BLOCK_MAX_TEXELS][BLOCK_MAX_WEIGHTS];
422};
423
424/**
425 * @brief Metadata for single block mode for a specific block size.
426 */
427struct block_mode
428{
429	/** @brief The block mode index in the ASTC encoded form. */
430	uint16_t mode_index;
431
432	/** @brief The decimation mode index in the compressor reindexed list. */
433	uint8_t decimation_mode;
434
435	/** @brief The weight quantization used by this block mode. */
436	uint8_t quant_mode;
437
438	/** @brief The weight quantization used by this block mode. */
439	uint8_t weight_bits;
440
441	/** @brief Is a dual weight plane used by this block mode? */
442	uint8_t is_dual_plane : 1;
443
444	/**
445	 * @brief Get the weight quantization used by this block mode.
446	 *
447	 * @return The quantization level.
448	 */
449	inline quant_method get_weight_quant_mode() const
450	{
451		return static_cast<quant_method>(this->quant_mode);
452	}
453};
454
455/**
456 * @brief Metadata for single decimation mode for a specific block size.
457 */
458struct decimation_mode
459{
460	/** @brief The max weight precision for 1 plane, or -1 if not supported. */
461	int8_t maxprec_1plane;
462
463	/** @brief The max weight precision for 2 planes, or -1 if not supported. */
464	int8_t maxprec_2planes;
465
466	/**
467	 * @brief Bitvector indicating weight quant modes used by active 1 plane block modes.
468	 *
469	 * Bit 0 = QUANT_2, Bit 1 = QUANT_3, etc.
470	 */
471	uint16_t refprec_1plane;
472
473	/**
474	 * @brief Bitvector indicating weight quant methods used by active 2 plane block modes.
475	 *
476	 * Bit 0 = QUANT_2, Bit 1 = QUANT_3, etc.
477	 */
478	uint16_t refprec_2planes;
479
480	/**
481	 * @brief Set a 1 plane weight quant as active.
482	 *
483	 * @param weight_quant   The quant method to set.
484	 */
485	void set_ref_1plane(quant_method weight_quant)
486	{
487		refprec_1plane |= (1 << weight_quant);
488	}
489
490	/**
491	 * @brief Test if this mode is active below a given 1 plane weight quant (inclusive).
492	 *
493	 * @param max_weight_quant   The max quant method to test.
494	 */
495	bool is_ref_1plane(quant_method max_weight_quant) const
496	{
497		uint16_t mask = static_cast<uint16_t>((1 << (max_weight_quant + 1)) - 1);
498		return (refprec_1plane & mask) != 0;
499	}
500
501	/**
502	 * @brief Set a 2 plane weight quant as active.
503	 *
504	 * @param weight_quant   The quant method to set.
505	 */
506	void set_ref_2plane(quant_method weight_quant)
507	{
508		refprec_2planes |= static_cast<uint16_t>(1 << weight_quant);
509	}
510
511	/**
512	 * @brief Test if this mode is active below a given 2 plane weight quant (inclusive).
513	 *
514	 * @param max_weight_quant   The max quant method to test.
515	 */
516	bool is_ref_2plane(quant_method max_weight_quant) const
517	{
518		uint16_t mask = static_cast<uint16_t>((1 << (max_weight_quant + 1)) - 1);
519		return (refprec_2planes & mask) != 0;
520	}
521};
522
523/**
524 * @brief Data tables for a single block size.
525 *
526 * The decimation tables store the information to apply weight grid dimension reductions. We only
527 * store the decimation modes that are actually needed by the current context; many of the possible
528 * modes will be unused (too many weights for the current block size or disabled by heuristics). The
529 * actual number of weights stored is @c decimation_mode_count, and the @c decimation_modes and
530 * @c decimation_tables arrays store the active modes contiguously at the start of the array. These
531 * entries are not stored in any particular order.
532 *
533 * The block mode tables store the unpacked block mode settings. Block modes are stored in the
534 * compressed block as an 11 bit field, but for any given block size and set of compressor
535 * heuristics, only a subset of the block modes will be used. The actual number of block modes
536 * stored is indicated in @c block_mode_count, and the @c block_modes array store the active modes
537 * contiguously at the start of the array. These entries are stored in incrementing "packed" value
538 * order, which doesn't mean much once unpacked. To allow decompressors to reference the packed data
539 * efficiently the @c block_mode_packed_index array stores the mapping between physical ID and the
540 * actual remapped array index.
541 */
542struct block_size_descriptor
543{
544	/** @brief The block X dimension, in texels. */
545	uint8_t xdim;
546
547	/** @brief The block Y dimension, in texels. */
548	uint8_t ydim;
549
550	/** @brief The block Z dimension, in texels. */
551	uint8_t zdim;
552
553	/** @brief The block total texel count. */
554	uint8_t texel_count;
555
556	/**
557	 * @brief The number of stored decimation modes which are "always" modes.
558	 *
559	 * Always modes are stored at the start of the decimation_modes list.
560	 */
561	unsigned int decimation_mode_count_always;
562
563	/** @brief The number of stored decimation modes for selected encodings. */
564	unsigned int decimation_mode_count_selected;
565
566	/** @brief The number of stored decimation modes for any encoding. */
567	unsigned int decimation_mode_count_all;
568
569	/**
570	 * @brief The number of stored block modes which are "always" modes.
571	 *
572	 * Always modes are stored at the start of the block_modes list.
573	 */
574	unsigned int block_mode_count_1plane_always;
575
576	/** @brief The number of stored block modes for active 1 plane encodings. */
577	unsigned int block_mode_count_1plane_selected;
578
579	/** @brief The number of stored block modes for active 1 and 2 plane encodings. */
580	unsigned int block_mode_count_1plane_2plane_selected;
581
582	/** @brief The number of stored block modes for any encoding. */
583	unsigned int block_mode_count_all;
584
585	/** @brief The number of selected partitionings for 1/2/3/4 partitionings. */
586	unsigned int partitioning_count_selected[BLOCK_MAX_PARTITIONS];
587
588	/** @brief The number of partitionings for 1/2/3/4 partitionings. */
589	unsigned int partitioning_count_all[BLOCK_MAX_PARTITIONS];
590
591	/** @brief The active decimation modes, stored in low indices. */
592	decimation_mode decimation_modes[WEIGHTS_MAX_DECIMATION_MODES];
593
594	/** @brief The active decimation tables, stored in low indices. */
595	ASTCENC_ALIGNAS decimation_info decimation_tables[WEIGHTS_MAX_DECIMATION_MODES];
596
597	/** @brief The packed block mode array index, or @c BLOCK_BAD_BLOCK_MODE if not active. */
598	uint16_t block_mode_packed_index[WEIGHTS_MAX_BLOCK_MODES];
599
600	/** @brief The active block modes, stored in low indices. */
601	block_mode block_modes[WEIGHTS_MAX_BLOCK_MODES];
602
603	/** @brief The active partition tables, stored in low indices per-count. */
604	partition_info partitionings[(3 * BLOCK_MAX_PARTITIONINGS) + 1];
605
606	/**
607	 * @brief The packed partition table array index, or @c BLOCK_BAD_PARTITIONING if not active.
608	 *
609	 * Indexed by partition_count - 2, containing 2, 3 and 4 partitions.
610	 */
611	uint16_t partitioning_packed_index[3][BLOCK_MAX_PARTITIONINGS];
612
613	/** @brief The active texels for k-means partition selection. */
614	uint8_t kmeans_texels[BLOCK_MAX_KMEANS_TEXELS];
615
616	/**
617	 * @brief The canonical 2-partition coverage pattern used during block partition search.
618	 *
619	 * Indexed by remapped index, not physical index.
620	 */
621	uint64_t coverage_bitmaps_2[BLOCK_MAX_PARTITIONINGS][2];
622
623	/**
624	 * @brief The canonical 3-partition coverage pattern used during block partition search.
625	 *
626	 * Indexed by remapped index, not physical index.
627	 */
628	uint64_t coverage_bitmaps_3[BLOCK_MAX_PARTITIONINGS][3];
629
630	/**
631	 * @brief The canonical 4-partition coverage pattern used during block partition search.
632	 *
633	 * Indexed by remapped index, not physical index.
634	 */
635	uint64_t coverage_bitmaps_4[BLOCK_MAX_PARTITIONINGS][4];
636
637	/**
638	 * @brief Get the block mode structure for index @c block_mode.
639	 *
640	 * This function can only return block modes that are enabled by the current compressor config.
641	 * Decompression from an arbitrary source should not use this without first checking that the
642	 * packed block mode index is not @c BLOCK_BAD_BLOCK_MODE.
643	 *
644	 * @param block_mode   The packed block mode index.
645	 *
646	 * @return The block mode structure.
647	 */
648	const block_mode& get_block_mode(unsigned int block_mode) const
649	{
650		unsigned int packed_index = this->block_mode_packed_index[block_mode];
651		assert(packed_index != BLOCK_BAD_BLOCK_MODE && packed_index < this->block_mode_count_all);
652		return this->block_modes[packed_index];
653	}
654
655	/**
656	 * @brief Get the decimation mode structure for index @c decimation_mode.
657	 *
658	 * This function can only return decimation modes that are enabled by the current compressor
659	 * config. The mode array is stored packed, but this is only ever indexed by the packed index
660	 * stored in the @c block_mode and never exists in an unpacked form.
661	 *
662	 * @param decimation_mode   The packed decimation mode index.
663	 *
664	 * @return The decimation mode structure.
665	 */
666	const decimation_mode& get_decimation_mode(unsigned int decimation_mode) const
667	{
668		return this->decimation_modes[decimation_mode];
669	}
670
671	/**
672	 * @brief Get the decimation info structure for index @c decimation_mode.
673	 *
674	 * This function can only return decimation modes that are enabled by the current compressor
675	 * config. The mode array is stored packed, but this is only ever indexed by the packed index
676	 * stored in the @c block_mode and never exists in an unpacked form.
677	 *
678	 * @param decimation_mode   The packed decimation mode index.
679	 *
680	 * @return The decimation info structure.
681	 */
682	const decimation_info& get_decimation_info(unsigned int decimation_mode) const
683	{
684		return this->decimation_tables[decimation_mode];
685	}
686
687	/**
688	 * @brief Get the partition info table for a given partition count.
689	 *
690	 * @param partition_count   The number of partitions we want the table for.
691	 *
692	 * @return The pointer to the table of 1024 entries (for 2/3/4 parts) or 1 entry (for 1 part).
693	 */
694	const partition_info* get_partition_table(unsigned int partition_count) const
695	{
696		if (partition_count == 1)
697		{
698			partition_count = 5;
699		}
700		unsigned int index = (partition_count - 2) * BLOCK_MAX_PARTITIONINGS;
701		return this->partitionings + index;
702	}
703
704	/**
705	 * @brief Get the partition info structure for a given partition count and seed.
706	 *
707	 * @param partition_count   The number of partitions we want the info for.
708	 * @param index             The partition seed (between 0 and 1023).
709	 *
710	 * @return The partition info structure.
711	 */
712	const partition_info& get_partition_info(unsigned int partition_count, unsigned int index) const
713	{
714		unsigned int packed_index = 0;
715		if (partition_count >= 2)
716		{
717			packed_index = this->partitioning_packed_index[partition_count - 2][index];
718		}
719
720		assert(packed_index != BLOCK_BAD_PARTITIONING && packed_index < this->partitioning_count_all[partition_count - 1]);
721		auto& result = get_partition_table(partition_count)[packed_index];
722		assert(index == result.partition_index);
723		return result;
724	}
725
726	/**
727	 * @brief Get the partition info structure for a given partition count and seed.
728	 *
729	 * @param partition_count   The number of partitions we want the info for.
730	 * @param packed_index      The raw array offset.
731	 *
732	 * @return The partition info structure.
733	 */
734	const partition_info& get_raw_partition_info(unsigned int partition_count, unsigned int packed_index) const
735	{
736		assert(packed_index != BLOCK_BAD_PARTITIONING && packed_index < this->partitioning_count_all[partition_count - 1]);
737		auto& result = get_partition_table(partition_count)[packed_index];
738		return result;
739	}
740};
741
742/**
743 * @brief The image data for a single block.
744 *
745 * The @c data_[rgba] fields store the image data in an encoded SoA float form designed for easy
746 * vectorization. Input data is converted to float and stored as values between 0 and 65535. LDR
747 * data is stored as direct UNORM data, HDR data is stored as LNS data.
748 *
749 * The @c rgb_lns and @c alpha_lns fields that assigned a per-texel use of HDR are only used during
750 * decompression. The current compressor will always use HDR endpoint formats when in HDR mode.
751 */
752struct image_block
753{
754	/** @brief The input (compress) or output (decompress) data for the red color component. */
755	ASTCENC_ALIGNAS float data_r[BLOCK_MAX_TEXELS];
756
757	/** @brief The input (compress) or output (decompress) data for the green color component. */
758	ASTCENC_ALIGNAS float data_g[BLOCK_MAX_TEXELS];
759
760	/** @brief The input (compress) or output (decompress) data for the blue color component. */
761	ASTCENC_ALIGNAS float data_b[BLOCK_MAX_TEXELS];
762
763	/** @brief The input (compress) or output (decompress) data for the alpha color component. */
764	ASTCENC_ALIGNAS float data_a[BLOCK_MAX_TEXELS];
765
766	mutable partition_metrics pms[BLOCK_MAX_PARTITIONS];
767
768	/** @brief The number of texels in the block. */
769	uint8_t texel_count;
770
771	/** @brief The original data for texel 0 for constant color block encoding. */
772	vfloat4 origin_texel;
773
774	/** @brief The min component value of all texels in the block. */
775	vfloat4 data_min;
776
777	/** @brief The mean component value of all texels in the block. */
778	vfloat4 data_mean;
779
780	/** @brief The max component value of all texels in the block. */
781	vfloat4 data_max;
782
783	/** @brief The relative error significance of the color channels. */
784	vfloat4 channel_weight;
785
786	/** @brief Is this grayscale block where R == G == B for all texels? */
787	bool grayscale;
788
789	/** @brief Is the eventual decode using decode_unorm8 rounding? */
790	bool decode_unorm8;
791
792	/** @brief Set to 1 if a texel is using HDR RGB endpoints (decompression only). */
793	uint8_t rgb_lns[BLOCK_MAX_TEXELS];
794
795	/** @brief Set to 1 if a texel is using HDR alpha endpoints (decompression only). */
796	uint8_t alpha_lns[BLOCK_MAX_TEXELS];
797
798	/** @brief The X position of this block in the input or output image. */
799	unsigned int xpos;
800
801	/** @brief The Y position of this block in the input or output image. */
802	unsigned int ypos;
803
804	/** @brief The Z position of this block in the input or output image. */
805	unsigned int zpos;
806
807	/**
808	 * @brief Get an RGBA texel value from the data.
809	 *
810	 * @param index   The texel index.
811	 *
812	 * @return The texel in RGBA component ordering.
813	 */
814	inline vfloat4 texel(unsigned int index) const
815	{
816		return vfloat4(data_r[index],
817		               data_g[index],
818		               data_b[index],
819		               data_a[index]);
820	}
821
822	/**
823	 * @brief Get an RGB texel value from the data.
824	 *
825	 * @param index   The texel index.
826	 *
827	 * @return The texel in RGB0 component ordering.
828	 */
829	inline vfloat4 texel3(unsigned int index) const
830	{
831		return vfloat3(data_r[index],
832		               data_g[index],
833		               data_b[index]);
834	}
835
836	/**
837	 * @brief Get the default alpha value for endpoints that don't store it.
838	 *
839	 * The default depends on whether the alpha endpoint is LDR or HDR.
840	 *
841	 * @return The alpha value in the scaled range used by the compressor.
842	 */
843	inline float get_default_alpha() const
844	{
845		return this->alpha_lns[0] ? static_cast<float>(0x7800) : static_cast<float>(0xFFFF);
846	}
847
848	/**
849	 * @brief Test if a single color channel is constant across the block.
850	 *
851	 * Constant color channels are easier to compress as interpolating between two identical colors
852	 * always returns the same value, irrespective of the weight used. They therefore can be ignored
853	 * for the purposes of weight selection and use of a second weight plane.
854	 *
855	 * @return @c true if the channel is constant across the block, @c false otherwise.
856	 */
857	inline bool is_constant_channel(int channel) const
858	{
859		vmask4 lane_mask = vint4::lane_id() == vint4(channel);
860		vmask4 color_mask = this->data_min == this->data_max;
861		return any(lane_mask & color_mask);
862	}
863
864	/**
865	 * @brief Test if this block is a luminance block with constant 1.0 alpha.
866	 *
867	 * @return @c true if the block is a luminance block , @c false otherwise.
868	 */
869	inline bool is_luminance() const
870	{
871		float default_alpha = this->get_default_alpha();
872		bool alpha1 = (this->data_min.lane<3>() == default_alpha) &&
873		              (this->data_max.lane<3>() == default_alpha);
874		return this->grayscale && alpha1;
875	}
876
877	/**
878	 * @brief Test if this block is a luminance block with variable alpha.
879	 *
880	 * @return @c true if the block is a luminance + alpha block , @c false otherwise.
881	 */
882	inline bool is_luminancealpha() const
883	{
884		float default_alpha = this->get_default_alpha();
885		bool alpha1 = (this->data_min.lane<3>() == default_alpha) &&
886		              (this->data_max.lane<3>() == default_alpha);
887		return this->grayscale && !alpha1;
888	}
889};
890
891/**
892 * @brief Data structure storing the color endpoints for a block.
893 */
894struct endpoints
895{
896	/** @brief The number of partition endpoints stored. */
897	unsigned int partition_count;
898
899	/** @brief The colors for endpoint 0. */
900	vfloat4 endpt0[BLOCK_MAX_PARTITIONS];
901
902	/** @brief The colors for endpoint 1. */
903	vfloat4 endpt1[BLOCK_MAX_PARTITIONS];
904};
905
906/**
907 * @brief Data structure storing the color endpoints and weights.
908 */
909struct endpoints_and_weights
910{
911	/** @brief True if all active values in weight_error_scale are the same. */
912	bool is_constant_weight_error_scale;
913
914	/** @brief The color endpoints. */
915	endpoints ep;
916
917	/** @brief The ideal weight for each texel; may be undecimated or decimated. */
918	ASTCENC_ALIGNAS float weights[BLOCK_MAX_TEXELS];
919
920	/** @brief The ideal weight error scaling for each texel; may be undecimated or decimated. */
921	ASTCENC_ALIGNAS float weight_error_scale[BLOCK_MAX_TEXELS];
922};
923
924/**
925 * @brief Utility storing estimated errors from choosing particular endpoint encodings.
926 */
927struct encoding_choice_errors
928{
929	/** @brief Error of using LDR RGB-scale instead of complete endpoints. */
930	float rgb_scale_error;
931
932	/** @brief Error of using HDR RGB-scale instead of complete endpoints. */
933	float rgb_luma_error;
934
935	/** @brief Error of using luminance instead of RGB. */
936	float luminance_error;
937
938	/** @brief Error of discarding alpha and using a constant 1.0 alpha. */
939	float alpha_drop_error;
940
941	/** @brief Can we use delta offset encoding? */
942	bool can_offset_encode;
943
944	/** @brief Can we use blue contraction encoding? */
945	bool can_blue_contract;
946};
947
948/**
949 * @brief Preallocated working buffers, allocated per thread during context creation.
950 */
951struct ASTCENC_ALIGNAS compression_working_buffers
952{
953	/** @brief Ideal endpoints and weights for plane 1. */
954	endpoints_and_weights ei1;
955
956	/** @brief Ideal endpoints and weights for plane 2. */
957	endpoints_and_weights ei2;
958
959	/**
960	 * @brief Decimated ideal weight values in the ~0-1 range.
961	 *
962	 * Note that values can be slightly below zero or higher than one due to
963	 * endpoint extents being inside the ideal color representation.
964	 *
965	 * For two planes, second plane starts at @c WEIGHTS_PLANE2_OFFSET offsets.
966	 */
967	ASTCENC_ALIGNAS float dec_weights_ideal[WEIGHTS_MAX_DECIMATION_MODES * BLOCK_MAX_WEIGHTS];
968
969	/**
970	 * @brief Decimated quantized weight values in the unquantized 0-64 range.
971	 *
972	 * For two planes, second plane starts at @c WEIGHTS_PLANE2_OFFSET offsets.
973	 */
974	uint8_t dec_weights_uquant[WEIGHTS_MAX_BLOCK_MODES * BLOCK_MAX_WEIGHTS];
975
976	/** @brief Error of the best encoding combination for each block mode. */
977	ASTCENC_ALIGNAS float errors_of_best_combination[WEIGHTS_MAX_BLOCK_MODES];
978
979	/** @brief The best color quant for each block mode. */
980	uint8_t best_quant_levels[WEIGHTS_MAX_BLOCK_MODES];
981
982	/** @brief The best color quant for each block mode if modes are the same and we have spare bits. */
983	uint8_t best_quant_levels_mod[WEIGHTS_MAX_BLOCK_MODES];
984
985	/** @brief The best endpoint format for each partition. */
986	uint8_t best_ep_formats[WEIGHTS_MAX_BLOCK_MODES][BLOCK_MAX_PARTITIONS];
987
988	/** @brief The total bit storage needed for quantized weights for each block mode. */
989	int8_t qwt_bitcounts[WEIGHTS_MAX_BLOCK_MODES];
990
991	/** @brief The cumulative error for quantized weights for each block mode. */
992	float qwt_errors[WEIGHTS_MAX_BLOCK_MODES];
993
994	/** @brief The low weight value in plane 1 for each block mode. */
995	float weight_low_value1[WEIGHTS_MAX_BLOCK_MODES];
996
997	/** @brief The high weight value in plane 1 for each block mode. */
998	float weight_high_value1[WEIGHTS_MAX_BLOCK_MODES];
999
1000	/** @brief The low weight value in plane 1 for each quant level and decimation mode. */
1001	float weight_low_values1[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1];
1002
1003	/** @brief The high weight value in plane 1 for each quant level and decimation mode. */
1004	float weight_high_values1[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1];
1005
1006	/** @brief The low weight value in plane 2 for each block mode. */
1007	float weight_low_value2[WEIGHTS_MAX_BLOCK_MODES];
1008
1009	/** @brief The high weight value in plane 2 for each block mode. */
1010	float weight_high_value2[WEIGHTS_MAX_BLOCK_MODES];
1011
1012	/** @brief The low weight value in plane 2 for each quant level and decimation mode. */
1013	float weight_low_values2[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1];
1014
1015	/** @brief The high weight value in plane 2 for each quant level and decimation mode. */
1016	float weight_high_values2[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1];
1017};
1018
1019struct dt_init_working_buffers
1020{
1021	uint8_t weight_count_of_texel[BLOCK_MAX_TEXELS];
1022	uint8_t grid_weights_of_texel[BLOCK_MAX_TEXELS][4];
1023	uint8_t weights_of_texel[BLOCK_MAX_TEXELS][4];
1024
1025	uint8_t texel_count_of_weight[BLOCK_MAX_WEIGHTS];
1026	uint8_t texels_of_weight[BLOCK_MAX_WEIGHTS][BLOCK_MAX_TEXELS];
1027	uint8_t texel_weights_of_weight[BLOCK_MAX_WEIGHTS][BLOCK_MAX_TEXELS];
1028};
1029
1030/**
1031 * @brief Weight quantization transfer table.
1032 *
1033 * ASTC can store texel weights at many quantization levels, so for performance we store essential
1034 * information about each level as a precomputed data structure. Unquantized weights are integers
1035 * or floats in the range [0, 64].
1036 *
1037 * This structure provides a table, used to estimate the closest quantized weight for a given
1038 * floating-point weight. For each quantized weight, the corresponding unquantized values. For each
1039 * quantized weight, a previous-value and a next-value.
1040*/
1041struct quant_and_transfer_table
1042{
1043	/** @brief The unscrambled unquantized value. */
1044	uint8_t quant_to_unquant[32];
1045
1046	/** @brief The scrambling order: scrambled_quant = map[unscrambled_quant]. */
1047	uint8_t scramble_map[32];
1048
1049	/** @brief The unscrambling order: unscrambled_unquant = map[scrambled_quant]. */
1050	uint8_t unscramble_and_unquant_map[32];
1051
1052	/**
1053	 * @brief A table of previous-and-next weights, indexed by the current unquantized value.
1054	 *  * bits 7:0 = previous-index, unquantized
1055	 *  * bits 15:8 = next-index, unquantized
1056	 */
1057	uint16_t prev_next_values[65];
1058};
1059
1060/** @brief The precomputed quant and transfer table. */
1061extern const quant_and_transfer_table quant_and_xfer_tables[12];
1062
1063/** @brief The block is an error block, and will return error color or NaN. */
1064static constexpr uint8_t SYM_BTYPE_ERROR { 0 };
1065
1066/** @brief The block is a constant color block using FP16 colors. */
1067static constexpr uint8_t SYM_BTYPE_CONST_F16 { 1 };
1068
1069/** @brief The block is a constant color block using UNORM16 colors. */
1070static constexpr uint8_t SYM_BTYPE_CONST_U16 { 2 };
1071
1072/** @brief The block is a normal non-constant color block. */
1073static constexpr uint8_t SYM_BTYPE_NONCONST { 3 };
1074
1075/**
1076 * @brief A symbolic representation of a compressed block.
1077 *
1078 * The symbolic representation stores the unpacked content of a single
1079 * physical compressed block, in a form which is much easier to access for
1080 * the rest of the compressor code.
1081 */
1082struct symbolic_compressed_block
1083{
1084	/** @brief The block type, one of the @c SYM_BTYPE_* constants. */
1085	uint8_t block_type;
1086
1087	/** @brief The number of partitions; valid for @c NONCONST blocks. */
1088	uint8_t partition_count;
1089
1090	/** @brief Non-zero if the color formats matched; valid for @c NONCONST blocks. */
1091	uint8_t color_formats_matched;
1092
1093	/** @brief The plane 2 color component, or -1 if single plane; valid for @c NONCONST blocks. */
1094	int8_t plane2_component;
1095
1096	/** @brief The block mode; valid for @c NONCONST blocks. */
1097	uint16_t block_mode;
1098
1099	/** @brief The partition index; valid for @c NONCONST blocks if 2 or more partitions. */
1100	uint16_t partition_index;
1101
1102	/** @brief The endpoint color formats for each partition; valid for @c NONCONST blocks. */
1103	uint8_t color_formats[BLOCK_MAX_PARTITIONS];
1104
1105	/** @brief The endpoint color quant mode; valid for @c NONCONST blocks. */
1106	quant_method quant_mode;
1107
1108	/** @brief The error of the current encoding; valid for @c NONCONST blocks. */
1109	float errorval;
1110
1111	// We can't have both of these at the same time
1112	union {
1113		/** @brief The constant color; valid for @c CONST blocks. */
1114		int constant_color[BLOCK_MAX_COMPONENTS];
1115
1116		/** @brief The quantized endpoint color pairs; valid for @c NONCONST blocks. */
1117		uint8_t color_values[BLOCK_MAX_PARTITIONS][8];
1118	};
1119
1120	/** @brief The quantized and decimated weights.
1121	 *
1122	 * Weights are stored in the 0-64 unpacked range allowing them to be used
1123	 * directly in encoding passes without per-use unpacking. Packing happens
1124	 * when converting to/from the physical bitstream encoding.
1125	 *
1126	 * If dual plane, the second plane starts at @c weights[WEIGHTS_PLANE2_OFFSET].
1127	 */
1128	uint8_t weights[BLOCK_MAX_WEIGHTS];
1129
1130	/**
1131	 * @brief Get the weight quantization used by this block mode.
1132	 *
1133	 * @return The quantization level.
1134	 */
1135	inline quant_method get_color_quant_mode() const
1136	{
1137		return this->quant_mode;
1138	}
1139	QualityProfile privateProfile;
1140};
1141
1142/**
1143 * @brief Parameter structure for @c compute_pixel_region_variance().
1144 *
1145 * This function takes a structure to avoid spilling arguments to the stack on every function
1146 * invocation, as there are a lot of parameters.
1147 */
1148struct pixel_region_args
1149{
1150	/** @brief The image to analyze. */
1151	const astcenc_image* img;
1152
1153	/** @brief The component swizzle pattern. */
1154	astcenc_swizzle swz;
1155
1156	/** @brief Should the algorithm bother with Z axis processing? */
1157	bool have_z;
1158
1159	/** @brief The kernel radius for alpha processing. */
1160	unsigned int alpha_kernel_radius;
1161
1162	/** @brief The X dimension of the working data to process. */
1163	unsigned int size_x;
1164
1165	/** @brief The Y dimension of the working data to process. */
1166	unsigned int size_y;
1167
1168	/** @brief The Z dimension of the working data to process. */
1169	unsigned int size_z;
1170
1171	/** @brief The X position of first src and dst data in the data set. */
1172	unsigned int offset_x;
1173
1174	/** @brief The Y position of first src and dst data in the data set. */
1175	unsigned int offset_y;
1176
1177	/** @brief The Z position of first src and dst data in the data set. */
1178	unsigned int offset_z;
1179
1180	/** @brief The working memory buffer. */
1181	vfloat4 *work_memory;
1182};
1183
1184/**
1185 * @brief Parameter structure for @c compute_averages_proc().
1186 */
1187struct avg_args
1188{
1189	/** @brief The arguments for the nested variance computation. */
1190	pixel_region_args arg;
1191
1192	/** @brief The image Stride dimensions. */
1193	unsigned int img_size_stride;
1194
1195	/** @brief The image X dimensions. */
1196	unsigned int img_size_x;
1197
1198	/** @brief The image Y dimensions. */
1199	unsigned int img_size_y;
1200
1201	/** @brief The image Z dimensions. */
1202	unsigned int img_size_z;
1203
1204	/** @brief The maximum working block dimensions in X and Y dimensions. */
1205	unsigned int blk_size_xy;
1206
1207	/** @brief The maximum working block dimensions in Z dimensions. */
1208	unsigned int blk_size_z;
1209
1210	/** @brief The working block memory size. */
1211	unsigned int work_memory_size;
1212};
1213
1214#if defined(ASTCENC_DIAGNOSTICS)
1215/* See astcenc_diagnostic_trace header for details. */
1216class TraceLog;
1217#endif
1218
1219/**
1220 * @brief The astcenc compression context.
1221 */
1222struct astcenc_contexti
1223{
1224	/** @brief The configuration this context was created with. */
1225	astcenc_config config;
1226
1227	/** @brief The thread count supported by this context. */
1228	unsigned int thread_count;
1229
1230	/** @brief The block size descriptor this context was created with. */
1231	block_size_descriptor* bsd;
1232
1233	/*
1234	 * Fields below here are not needed in a decompress-only build, but some remain as they are
1235	 * small and it avoids littering the code with #ifdefs. The most significant contributors to
1236	 * large structure size are omitted.
1237	 */
1238
1239	/** @brief The input image alpha channel averages table, may be @c nullptr if not needed. */
1240	float* input_alpha_averages;
1241
1242	/** @brief The scratch working buffers, one per thread (see @c thread_count). */
1243	compression_working_buffers* working_buffers;
1244
1245#if !defined(ASTCENC_DECOMPRESS_ONLY)
1246	/** @brief The pixel region and variance worker arguments. */
1247	avg_args avg_preprocess_args;
1248#endif
1249
1250#if defined(ASTCENC_DIAGNOSTICS)
1251	/**
1252	 * @brief The diagnostic trace logger.
1253	 *
1254	 * Note that this is a singleton, so can only be used in single threaded mode. It only exists
1255	 * here so we have a reference to close the file at the end of the capture.
1256	 */
1257	TraceLog* trace_log;
1258#endif
1259};
1260
1261/* ============================================================================
1262  Functionality for managing block sizes and partition tables.
1263============================================================================ */
1264
1265/**
1266 * @brief Populate the block size descriptor for the target block size.
1267 *
1268 * This will also initialize the partition table metadata, which is stored as part of the BSD
1269 * structure.
1270 *
1271 * @param      x_texels                 The number of texels in the block X dimension.
1272 * @param      y_texels                 The number of texels in the block Y dimension.
1273 * @param      z_texels                 The number of texels in the block Z dimension.
1274 * @param      can_omit_modes           Can we discard modes and partitionings that astcenc won't use?
1275 * @param      partition_count_cutoff   The partition count cutoff to use, if we can omit partitionings.
1276 * @param      mode_cutoff              The block mode percentile cutoff [0-1].
1277 * @param[out] bsd                      The descriptor to initialize.
1278 */
1279#ifdef ASTC_CUSTOMIZED_ENABLE
1280bool init_block_size_descriptor(
1281#else
1282void init_block_size_descriptor(
1283#endif
1284	QualityProfile privateProfile,
1285	unsigned int x_texels,
1286	unsigned int y_texels,
1287	unsigned int z_texels,
1288	bool can_omit_modes,
1289	unsigned int partition_count_cutoff,
1290	float mode_cutoff,
1291	block_size_descriptor& bsd);
1292
1293/**
1294 * @brief Populate the partition tables for the target block size.
1295 *
1296 * Note the @c bsd descriptor must be initialized by calling @c init_block_size_descriptor() before
1297 * calling this function.
1298 *
1299 * @param[out] bsd                      The block size information structure to populate.
1300 * @param      can_omit_partitionings   True if we can we drop partitionings that astcenc won't use.
1301 * @param      partition_count_cutoff   The partition count cutoff to use, if we can omit partitionings.
1302 */
1303void init_partition_tables(
1304	block_size_descriptor& bsd,
1305	bool can_omit_partitionings,
1306	unsigned int partition_count_cutoff);
1307
1308/**
1309 * @brief Get the percentile table for 2D block modes.
1310 *
1311 * This is an empirically determined prioritization of which block modes to use in the search in
1312 * terms of their centile (lower centiles = more useful).
1313 *
1314 * Returns a dynamically allocated array; caller must free with delete[].
1315 *
1316 * @param xdim The block x size.
1317 * @param ydim The block y size.
1318 *
1319 * @return The unpacked table.
1320 */
1321const float* get_2d_percentile_table(
1322	unsigned int xdim,
1323	unsigned int ydim);
1324
1325/**
1326 * @brief Query if a 2D block size is legal.
1327 *
1328 * @return True if legal, false otherwise.
1329 */
1330bool is_legal_2d_block_size(
1331	unsigned int xdim,
1332	unsigned int ydim);
1333
1334/**
1335 * @brief Query if a 3D block size is legal.
1336 *
1337 * @return True if legal, false otherwise.
1338 */
1339bool is_legal_3d_block_size(
1340	unsigned int xdim,
1341	unsigned int ydim,
1342	unsigned int zdim);
1343
1344/* ============================================================================
1345  Functionality for managing BISE quantization and unquantization.
1346============================================================================ */
1347
1348/**
1349 * @brief The precomputed table for quantizing color values.
1350 *
1351 * Converts unquant value in 0-255 range into quant value in 0-255 range.
1352 * No BISE scrambling is applied at this stage.
1353 *
1354 * The BISE encoding results in ties where available quant<256> values are
1355 * equidistant the available quant<BISE> values. This table stores two values
1356 * for each input - one for use with a negative residual, and one for use with
1357 * a positive residual.
1358 *
1359 * Indexed by [quant_mode - 4][data_value * 2 + residual].
1360 */
1361extern const uint8_t color_unquant_to_uquant_tables[17][512];
1362
1363/**
1364 * @brief The precomputed table for packing quantized color values.
1365 *
1366 * Converts quant value in 0-255 range into packed quant value in 0-N range,
1367 * with BISE scrambling applied.
1368 *
1369 * Indexed by [quant_mode - 4][data_value].
1370 */
1371extern const uint8_t color_uquant_to_scrambled_pquant_tables[17][256];
1372
1373/**
1374 * @brief The precomputed table for unpacking color values.
1375 *
1376 * Converts quant value in 0-N range into unpacked value in 0-255 range,
1377 * with BISE unscrambling applied.
1378 *
1379 * Indexed by [quant_mode - 4][data_value].
1380 */
1381extern const uint8_t* color_scrambled_pquant_to_uquant_tables[17];
1382
1383/**
1384 * @brief The precomputed quant mode storage table.
1385 *
1386 * Indexing by [integer_count/2][bits] gives us the quantization level for a given integer count and
1387 * number of compressed storage bits. Returns -1 for cases where the requested integer count cannot
1388 * ever fit in the supplied storage size.
1389 */
1390extern const int8_t quant_mode_table[10][128];
1391
1392/**
1393 * @brief Encode a packed string using BISE.
1394 *
1395 * Note that BISE can return strings that are not a whole number of bytes in length, and ASTC can
1396 * start storing strings in a block at arbitrary bit offsets in the encoded data.
1397 *
1398 * @param         quant_level       The BISE alphabet size.
1399 * @param         character_count   The number of characters in the string.
1400 * @param         input_data        The unpacked string, one byte per character.
1401 * @param[in,out] output_data       The output packed string.
1402 * @param         bit_offset        The starting offset in the output storage.
1403 */
1404void encode_ise(
1405	quant_method quant_level,
1406	unsigned int character_count,
1407	const uint8_t* input_data,
1408	uint8_t* output_data,
1409	unsigned int bit_offset);
1410
1411/**
1412 * @brief Decode a packed string using BISE.
1413 *
1414 * Note that BISE input strings are not a whole number of bytes in length, and ASTC can start
1415 * strings at arbitrary bit offsets in the encoded data.
1416 *
1417 * @param         quant_level       The BISE alphabet size.
1418 * @param         character_count   The number of characters in the string.
1419 * @param         input_data        The packed string.
1420 * @param[in,out] output_data       The output storage, one byte per character.
1421 * @param         bit_offset        The starting offset in the output storage.
1422 */
1423void decode_ise(
1424	quant_method quant_level,
1425	unsigned int character_count,
1426	const uint8_t* input_data,
1427	uint8_t* output_data,
1428	unsigned int bit_offset);
1429
1430/**
1431 * @brief Return the number of bits needed to encode an ISE sequence.
1432 *
1433 * This implementation assumes that the @c quant level is untrusted, given it may come from random
1434 * data being decompressed, so we return an arbitrary unencodable size if that is the case.
1435 *
1436 * @param character_count   The number of items in the sequence.
1437 * @param quant_level       The desired quantization level.
1438 *
1439 * @return The number of bits needed to encode the BISE string.
1440 */
1441unsigned int get_ise_sequence_bitcount(
1442	unsigned int character_count,
1443	quant_method quant_level);
1444
1445/* ============================================================================
1446  Functionality for managing color partitioning.
1447============================================================================ */
1448
1449/**
1450 * @brief Compute averages and dominant directions for each partition in a 2 component texture.
1451 *
1452 * @param      pi           The partition info for the current trial.
1453 * @param      blk          The image block color data to be compressed.
1454 * @param      component1   The first component included in the analysis.
1455 * @param      component2   The second component included in the analysis.
1456 * @param[out] pm           The output partition metrics.
1457 *                          - Only pi.partition_count array entries actually get initialized.
1458 *                          - Direction vectors @c pm.dir are not normalized.
1459 */
1460void compute_avgs_and_dirs_2_comp(
1461	const partition_info& pi,
1462	const image_block& blk,
1463	unsigned int component1,
1464	unsigned int component2,
1465	partition_metrics pm[BLOCK_MAX_PARTITIONS]);
1466
1467/**
1468 * @brief Compute averages and dominant directions for each partition in a 3 component texture.
1469 *
1470 * @param      pi                  The partition info for the current trial.
1471 * @param      blk                 The image block color data to be compressed.
1472 * @param      omitted_component   The component excluded from the analysis.
1473 * @param[out] pm                  The output partition metrics.
1474 *                                 - Only pi.partition_count array entries actually get initialized.
1475 *                                 - Direction vectors @c pm.dir are not normalized.
1476 */
1477void compute_avgs_and_dirs_3_comp(
1478	const partition_info& pi,
1479	const image_block& blk,
1480	unsigned int omitted_component,
1481	partition_metrics pm[BLOCK_MAX_PARTITIONS]);
1482
1483/**
1484 * @brief Compute averages and dominant directions for each partition in a 3 component texture.
1485 *
1486 * This is a specialization of @c compute_avgs_and_dirs_3_comp where the omitted component is
1487 * always alpha, a common case during partition search.
1488 *
1489 * @param      pi    The partition info for the current trial.
1490 * @param      blk   The image block color data to be compressed.
1491 * @param[out] pm    The output partition metrics.
1492 *                   - Only pi.partition_count array entries actually get initialized.
1493 *                   - Direction vectors @c pm.dir are not normalized.
1494 */
1495void compute_avgs_and_dirs_3_comp_rgb(
1496	const partition_info& pi,
1497	const image_block& blk,
1498	partition_metrics pm[BLOCK_MAX_PARTITIONS]);
1499
1500/**
1501 * @brief Compute averages and dominant directions for each partition in a 4 component texture.
1502 *
1503 * @param      pi    The partition info for the current trial.
1504 * @param      blk   The image block color data to be compressed.
1505 * @param[out] pm    The output partition metrics.
1506 *                   - Only pi.partition_count array entries actually get initialized.
1507 *                   - Direction vectors @c pm.dir are not normalized.
1508 */
1509void compute_avgs_and_dirs_4_comp(
1510	const partition_info& pi,
1511	const image_block& blk,
1512	partition_metrics pm[BLOCK_MAX_PARTITIONS]);
1513
1514/**
1515 * @brief Compute the RGB error for uncorrelated and same chroma projections.
1516 *
1517 * The output of compute averages and dirs is post processed to define two lines, both of which go
1518 * through the mean-color-value.  One line has a direction defined by the dominant direction; this
1519 * is used to assess the error from using an uncorrelated color representation. The other line goes
1520 * through (0,0,0) and is used to assess the error from using an RGBS color representation.
1521 *
1522 * This function computes the squared error when using these two representations.
1523 *
1524 * @param         pi            The partition info for the current trial.
1525 * @param         blk           The image block color data to be compressed.
1526 * @param[in,out] plines        Processed line inputs, and line length outputs.
1527 * @param[out]    uncor_error   The cumulative error for using the uncorrelated line.
1528 * @param[out]    samec_error   The cumulative error for using the same chroma line.
1529 */
1530void compute_error_squared_rgb(
1531	const partition_info& pi,
1532	const image_block& blk,
1533	partition_lines3 plines[BLOCK_MAX_PARTITIONS],
1534	float& uncor_error,
1535	float& samec_error);
1536
1537/**
1538 * @brief Compute the RGBA error for uncorrelated and same chroma projections.
1539 *
1540 * The output of compute averages and dirs is post processed to define two lines, both of which go
1541 * through the mean-color-value.  One line has a direction defined by the dominant direction; this
1542 * is used to assess the error from using an uncorrelated color representation. The other line goes
1543 * through (0,0,0,1) and is used to assess the error from using an RGBS color representation.
1544 *
1545 * This function computes the squared error when using these two representations.
1546 *
1547 * @param      pi              The partition info for the current trial.
1548 * @param      blk             The image block color data to be compressed.
1549 * @param      uncor_plines    Processed uncorrelated partition lines for each partition.
1550 * @param      samec_plines    Processed same chroma partition lines for each partition.
1551 * @param[out] line_lengths    The length of each components deviation from the line.
1552 * @param[out] uncor_error     The cumulative error for using the uncorrelated line.
1553 * @param[out] samec_error     The cumulative error for using the same chroma line.
1554 */
1555void compute_error_squared_rgba(
1556	const partition_info& pi,
1557	const image_block& blk,
1558	const processed_line4 uncor_plines[BLOCK_MAX_PARTITIONS],
1559	const processed_line4 samec_plines[BLOCK_MAX_PARTITIONS],
1560	float line_lengths[BLOCK_MAX_PARTITIONS],
1561	float& uncor_error,
1562	float& samec_error);
1563
1564/**
1565 * @brief Find the best set of partitions to trial for a given block.
1566 *
1567 * On return the @c best_partitions list will contain the two best partition
1568 * candidates; one assuming data has uncorrelated chroma and one assuming the
1569 * data has correlated chroma. The best candidate is returned first in the list.
1570 *
1571 * @param      bsd                      The block size information.
1572 * @param      blk                      The image block color data to compress.
1573 * @param      partition_count          The number of partitions in the block.
1574 * @param      partition_search_limit   The number of candidate partition encodings to trial.
1575 * @param[out] best_partitions          The best partition candidates.
1576 * @param      requested_candidates     The number of requested partitionings. May return fewer if
1577 *                                      candidates are not available.
1578 *
1579 * @return The actual number of candidates returned.
1580 */
1581unsigned int find_best_partition_candidates(
1582	const block_size_descriptor& bsd,
1583	const image_block& blk,
1584	unsigned int partition_count,
1585	unsigned int partition_search_limit,
1586	unsigned int best_partitions[TUNE_MAX_PARTITIONING_CANDIDATES],
1587	unsigned int requested_candidates);
1588
1589/* ============================================================================
1590  Functionality for managing images and image related data.
1591============================================================================ */
1592
1593/**
1594 * @brief Get a vector mask indicating lanes decompressing into a UNORM8 value.
1595 *
1596 * @param decode_mode   The color profile for LDR_SRGB settings.
1597 * @param blk           The image block for output image bitness settings.
1598 *
1599 * @return The component mask vector.
1600 */
1601static inline vmask4 get_u8_component_mask(
1602	astcenc_profile decode_mode,
1603	const image_block& blk
1604) {
1605	vmask4 u8_mask(false);
1606	// Decode mode writing to a unorm8 output value
1607	if (blk.decode_unorm8)
1608	{
1609		u8_mask = vmask4(true);
1610	}
1611	// SRGB writing to a unorm8 RGB value
1612	else if (decode_mode == ASTCENC_PRF_LDR_SRGB)
1613	{
1614		u8_mask = vmask4(true, true, true, false);
1615	}
1616
1617	return u8_mask;
1618}
1619
1620/**
1621 * @brief Setup computation of regional averages in an image.
1622 *
1623 * This must be done by only a single thread per image, before any thread calls
1624 * @c compute_averages().
1625 *
1626 * Results are written back into @c img->input_alpha_averages.
1627 *
1628 * @param      img                   The input image data, also holds output data.
1629 * @param      alpha_kernel_radius   The kernel radius (in pixels) for alpha mods.
1630 * @param      swz                   Input data component swizzle.
1631 * @param[out] ag                    The average variance arguments to init.
1632 *
1633 * @return The number of tasks in the processing stage.
1634 */
1635unsigned int init_compute_averages(
1636	const astcenc_image& img,
1637	unsigned int alpha_kernel_radius,
1638	const astcenc_swizzle& swz,
1639	avg_args& ag);
1640
1641/**
1642 * @brief Compute averages for a pixel region.
1643 *
1644 * The routine computes both in a single pass, using a summed-area table to decouple the running
1645 * time from the averaging/variance kernel size.
1646 *
1647 * @param[out] ctx   The compressor context storing the output data.
1648 * @param      arg   The input parameter structure.
1649 */
1650void compute_pixel_region_variance(
1651	astcenc_contexti& ctx,
1652	const pixel_region_args& arg);
1653/**
1654 * @brief Load a single image block from the input image.
1655 *
1656 * @param      decode_mode   The compression color profile.
1657 * @param      img           The input image data.
1658 * @param[out] blk           The image block to populate.
1659 * @param      bsd           The block size information.
1660 * @param      xpos          The block X coordinate in the input image.
1661 * @param      ypos          The block Y coordinate in the input image.
1662 * @param      zpos          The block Z coordinate in the input image.
1663 * @param      swz           The swizzle to apply on load.
1664 */
1665void load_image_block(
1666	astcenc_profile decode_mode,
1667	const astcenc_image& img,
1668	image_block& blk,
1669	const block_size_descriptor& bsd,
1670	unsigned int xpos,
1671	unsigned int ypos,
1672	unsigned int zpos,
1673	const astcenc_swizzle& swz);
1674
1675/**
1676 * @brief Load a single image block from the input image.
1677 *
1678 * This specialized variant can be used only if the block is 2D LDR U8 data,
1679 * with no swizzle.
1680 *
1681 * @param      decode_mode   The compression color profile.
1682 * @param      img           The input image data.
1683 * @param[out] blk           The image block to populate.
1684 * @param      bsd           The block size information.
1685 * @param      xpos          The block X coordinate in the input image.
1686 * @param      ypos          The block Y coordinate in the input image.
1687 * @param      zpos          The block Z coordinate in the input image.
1688 * @param      swz           The swizzle to apply on load.
1689 */
1690void load_image_block_fast_ldr(
1691	astcenc_profile decode_mode,
1692	const astcenc_image& img,
1693	image_block& blk,
1694	const block_size_descriptor& bsd,
1695	unsigned int xpos,
1696	unsigned int ypos,
1697	unsigned int zpos,
1698	const astcenc_swizzle& swz);
1699
1700/**
1701 * @brief Store a single image block to the output image.
1702 *
1703 * @param[out] img    The output image data.
1704 * @param      blk    The image block to export.
1705 * @param      bsd    The block size information.
1706 * @param      xpos   The block X coordinate in the input image.
1707 * @param      ypos   The block Y coordinate in the input image.
1708 * @param      zpos   The block Z coordinate in the input image.
1709 * @param      swz    The swizzle to apply on store.
1710 */
1711void store_image_block(
1712	astcenc_image& img,
1713	const image_block& blk,
1714	const block_size_descriptor& bsd,
1715	unsigned int xpos,
1716	unsigned int ypos,
1717	unsigned int zpos,
1718	const astcenc_swizzle& swz);
1719
1720/* ============================================================================
1721  Functionality for computing endpoint colors and weights for a block.
1722============================================================================ */
1723
1724/**
1725 * @brief Compute ideal endpoint colors and weights for 1 plane of weights.
1726 *
1727 * The ideal endpoints define a color line for the partition. For each texel the ideal weight
1728 * defines an exact position on the partition color line. We can then use these to assess the error
1729 * introduced by removing and quantizing the weight grid.
1730 *
1731 * @param      blk   The image block color data to compress.
1732 * @param      pi    The partition info for the current trial.
1733 * @param[out] ei    The endpoint and weight values.
1734 */
1735void compute_ideal_colors_and_weights_1plane(
1736	const image_block& blk,
1737	const partition_info& pi,
1738	endpoints_and_weights& ei);
1739
1740/**
1741 * @brief Compute ideal endpoint colors and weights for 2 planes of weights.
1742 *
1743 * The ideal endpoints define a color line for the partition. For each texel the ideal weight
1744 * defines an exact position on the partition color line. We can then use these to assess the error
1745 * introduced by removing and quantizing the weight grid.
1746 *
1747 * @param      bsd                The block size information.
1748 * @param      blk                The image block color data to compress.
1749 * @param      plane2_component   The component assigned to plane 2.
1750 * @param[out] ei1                The endpoint and weight values for plane 1.
1751 * @param[out] ei2                The endpoint and weight values for plane 2.
1752 */
1753void compute_ideal_colors_and_weights_2planes(
1754	const block_size_descriptor& bsd,
1755	const image_block& blk,
1756	unsigned int plane2_component,
1757	endpoints_and_weights& ei1,
1758	endpoints_and_weights& ei2);
1759
1760/**
1761 * @brief Compute the optimal unquantized weights for a decimation table.
1762 *
1763 * After computing ideal weights for the case for a complete weight grid, we we want to compute the
1764 * ideal weights for the case where weights exist only for some texels. We do this with a
1765 * steepest-descent grid solver which works as follows:
1766 *
1767 * First, for each actual weight, perform a weighted averaging of the texels affected by the weight.
1768 * Then, set step size to <some initial value> and attempt one step towards the original ideal
1769 * weight if it helps to reduce error.
1770 *
1771 * @param      ei                       The non-decimated endpoints and weights.
1772 * @param      di                       The selected weight decimation.
1773 * @param[out] dec_weight_ideal_value   The ideal values for the decimated weight set.
1774 */
1775void compute_ideal_weights_for_decimation(
1776	const endpoints_and_weights& ei,
1777	const decimation_info& di,
1778	float* dec_weight_ideal_value);
1779
1780/**
1781 * @brief Compute the optimal quantized weights for a decimation table.
1782 *
1783 * We test the two closest weight indices in the allowed quantization range and keep the weight that
1784 * is the closest match.
1785 *
1786 * @param      di                        The selected weight decimation.
1787 * @param      low_bound                 The lowest weight allowed.
1788 * @param      high_bound                The highest weight allowed.
1789 * @param      dec_weight_ideal_value    The ideal weight set.
1790 * @param[out] dec_weight_quant_uvalue   The output quantized weight as a float.
1791 * @param[out] dec_weight_uquant         The output quantized weight as encoded int.
1792 * @param      quant_level               The desired weight quant level.
1793 */
1794void compute_quantized_weights_for_decimation(
1795	const decimation_info& di,
1796	float low_bound,
1797	float high_bound,
1798	const float* dec_weight_ideal_value,
1799	float* dec_weight_quant_uvalue,
1800	uint8_t* dec_weight_uquant,
1801	quant_method quant_level);
1802
1803/**
1804 * @brief Compute the error of a decimated weight set for 1 plane.
1805 *
1806 * After computing ideal weights for the case with one weight per texel, we want to compute the
1807 * error for decimated weight grids where weights are stored at a lower resolution. This function
1808 * computes the error of the reduced grid, compared to the full grid.
1809 *
1810 * @param eai                       The ideal weights for the full grid.
1811 * @param di                        The selected weight decimation.
1812 * @param dec_weight_quant_uvalue   The quantized weights for the decimated grid.
1813 *
1814 * @return The accumulated error.
1815 */
1816float compute_error_of_weight_set_1plane(
1817	const endpoints_and_weights& eai,
1818	const decimation_info& di,
1819	const float* dec_weight_quant_uvalue);
1820
1821/**
1822 * @brief Compute the error of a decimated weight set for 2 planes.
1823 *
1824 * After computing ideal weights for the case with one weight per texel, we want to compute the
1825 * error for decimated weight grids where weights are stored at a lower resolution. This function
1826 * computes the error of the reduced grid, compared to the full grid.
1827 *
1828 * @param eai1                             The ideal weights for the full grid and plane 1.
1829 * @param eai2                             The ideal weights for the full grid and plane 2.
1830 * @param di                               The selected weight decimation.
1831 * @param dec_weight_quant_uvalue_plane1   The quantized weights for the decimated grid plane 1.
1832 * @param dec_weight_quant_uvalue_plane2   The quantized weights for the decimated grid plane 2.
1833 *
1834 * @return The accumulated error.
1835 */
1836float compute_error_of_weight_set_2planes(
1837	const endpoints_and_weights& eai1,
1838	const endpoints_and_weights& eai2,
1839	const decimation_info& di,
1840	const float* dec_weight_quant_uvalue_plane1,
1841	const float* dec_weight_quant_uvalue_plane2);
1842
1843/**
1844 * @brief Pack a single pair of color endpoints as effectively as possible.
1845 *
1846 * The user requests a base color endpoint mode in @c format, but the quantizer may choose a
1847 * delta-based representation. It will report back the format variant it actually used.
1848 *
1849 * @param      color0        The input unquantized color0 endpoint for absolute endpoint pairs.
1850 * @param      color1        The input unquantized color1 endpoint for absolute endpoint pairs.
1851 * @param      rgbs_color    The input unquantized RGBS variant endpoint for same chroma endpoints.
1852 * @param      rgbo_color    The input unquantized RGBS variant endpoint for HDR endpoints.
1853 * @param      format        The desired base format.
1854 * @param[out] output        The output storage for the quantized colors/
1855 * @param      quant_level   The quantization level requested.
1856 *
1857 * @return The actual endpoint mode used.
1858 */
1859uint8_t pack_color_endpoints(
1860	QualityProfile privateProfile,
1861	vfloat4 color0,
1862	vfloat4 color1,
1863	vfloat4 rgbs_color,
1864	vfloat4 rgbo_color,
1865	int format,
1866	uint8_t* output,
1867	quant_method quant_level);
1868
1869/**
1870 * @brief Unpack a single pair of encoded endpoints.
1871 *
1872 * Endpoints must be unscrambled and converted into the 0-255 range before calling this functions.
1873 *
1874 * @param      decode_mode   The decode mode (LDR, HDR, etc).
1875 * @param      format        The color endpoint mode used.
1876 * @param      input         The raw array of encoded input integers. The length of this array
1877 *                           depends on @c format; it can be safely assumed to be large enough.
1878 * @param[out] rgb_hdr       Is the endpoint using HDR for the RGB channels?
1879 * @param[out] alpha_hdr     Is the endpoint using HDR for the A channel?
1880 * @param[out] output0       The output color for endpoint 0.
1881 * @param[out] output1       The output color for endpoint 1.
1882 */
1883void unpack_color_endpoints(
1884	astcenc_profile decode_mode,
1885	int format,
1886	const uint8_t* input,
1887	bool& rgb_hdr,
1888	bool& alpha_hdr,
1889	vint4& output0,
1890	vint4& output1);
1891
1892/**
1893 * @brief Unpack an LDR RGBA color that uses delta encoding.
1894 *
1895 * @param      input0    The packed endpoint 0 color.
1896 * @param      input1    The packed endpoint 1 color deltas.
1897 * @param[out] output0   The unpacked endpoint 0 color.
1898 * @param[out] output1   The unpacked endpoint 1 color.
1899 */
1900void rgba_delta_unpack(
1901	vint4 input0,
1902	vint4 input1,
1903	vint4& output0,
1904	vint4& output1);
1905
1906/**
1907 * @brief Unpack an LDR RGBA color that uses direct encoding.
1908 *
1909 * @param      input0    The packed endpoint 0 color.
1910 * @param      input1    The packed endpoint 1 color.
1911 * @param[out] output0   The unpacked endpoint 0 color.
1912 * @param[out] output1   The unpacked endpoint 1 color.
1913 */
1914void rgba_unpack(
1915	vint4 input0,
1916	vint4 input1,
1917	vint4& output0,
1918	vint4& output1);
1919
1920/**
1921 * @brief Unpack a set of quantized and decimated weights.
1922 *
1923 * TODO: Can we skip this for non-decimated weights now that the @c scb is
1924 * already storing unquantized weights?
1925 *
1926 * @param      bsd              The block size information.
1927 * @param      scb              The symbolic compressed encoding.
1928 * @param      di               The weight grid decimation table.
1929 * @param      is_dual_plane    @c true if this is a dual plane block, @c false otherwise.
1930 * @param[out] weights_plane1   The output array for storing the plane 1 weights.
1931 * @param[out] weights_plane2   The output array for storing the plane 2 weights.
1932 */
1933void unpack_weights(
1934	const block_size_descriptor& bsd,
1935	const symbolic_compressed_block& scb,
1936	const decimation_info& di,
1937	bool is_dual_plane,
1938	int weights_plane1[BLOCK_MAX_TEXELS],
1939	int weights_plane2[BLOCK_MAX_TEXELS]);
1940
1941/**
1942 * @brief Identify, for each mode, which set of color endpoint produces the best result.
1943 *
1944 * Returns the best @c tune_candidate_limit best looking modes, along with the ideal color encoding
1945 * combination for each. The modified quantization level can be used when all formats are the same,
1946 * as this frees up two additional bits of storage.
1947 *
1948 * @param      pi                            The partition info for the current trial.
1949 * @param      blk                           The image block color data to compress.
1950 * @param      ep                            The ideal endpoints.
1951 * @param      qwt_bitcounts                 Bit counts for different quantization methods.
1952 * @param      qwt_errors                    Errors for different quantization methods.
1953 * @param      tune_candidate_limit          The max number of candidates to return, may be less.
1954 * @param      start_block_mode              The first block mode to inspect.
1955 * @param      end_block_mode                The last block mode to inspect.
1956 * @param[out] partition_format_specifiers   The best formats per partition.
1957 * @param[out] block_mode                    The best packed block mode indexes.
1958 * @param[out] quant_level                   The best color quant level.
1959 * @param[out] quant_level_mod               The best color quant level if endpoints are the same.
1960 * @param[out] tmpbuf                        Preallocated scratch buffers for the compressor.
1961 *
1962 * @return The actual number of candidate matches returned.
1963 */
1964unsigned int compute_ideal_endpoint_formats(
1965	QualityProfile privateProfile,
1966	const partition_info& pi,
1967	const image_block& blk,
1968	const endpoints& ep,
1969	const int8_t* qwt_bitcounts,
1970	const float* qwt_errors,
1971	unsigned int tune_candidate_limit,
1972	unsigned int start_block_mode,
1973	unsigned int end_block_mode,
1974	uint8_t partition_format_specifiers[TUNE_MAX_TRIAL_CANDIDATES][BLOCK_MAX_PARTITIONS],
1975	int block_mode[TUNE_MAX_TRIAL_CANDIDATES],
1976	quant_method quant_level[TUNE_MAX_TRIAL_CANDIDATES],
1977	quant_method quant_level_mod[TUNE_MAX_TRIAL_CANDIDATES],
1978	compression_working_buffers& tmpbuf);
1979
1980/**
1981 * @brief For a given 1 plane weight set recompute the endpoint colors.
1982 *
1983 * As we quantize and decimate weights the optimal endpoint colors may change slightly, so we must
1984 * recompute the ideal colors for a specific weight set.
1985 *
1986 * @param         blk                  The image block color data to compress.
1987 * @param         pi                   The partition info for the current trial.
1988 * @param         di                   The weight grid decimation table.
1989 * @param         dec_weights_uquant   The quantized weight set.
1990 * @param[in,out] ep                   The color endpoints (modifed in place).
1991 * @param[out]    rgbs_vectors         The RGB+scale vectors for LDR blocks.
1992 * @param[out]    rgbo_vectors         The RGB+offset vectors for HDR blocks.
1993 */
1994void recompute_ideal_colors_1plane(
1995	const image_block& blk,
1996	const partition_info& pi,
1997	const decimation_info& di,
1998	const uint8_t* dec_weights_uquant,
1999	endpoints& ep,
2000	vfloat4 rgbs_vectors[BLOCK_MAX_PARTITIONS],
2001	vfloat4 rgbo_vectors[BLOCK_MAX_PARTITIONS]);
2002
2003/**
2004 * @brief For a given 2 plane weight set recompute the endpoint colors.
2005 *
2006 * As we quantize and decimate weights the optimal endpoint colors may change slightly, so we must
2007 * recompute the ideal colors for a specific weight set.
2008 *
2009 * @param         blk                         The image block color data to compress.
2010 * @param         bsd                         The block_size descriptor.
2011 * @param         di                          The weight grid decimation table.
2012 * @param         dec_weights_uquant_plane1   The quantized weight set for plane 1.
2013 * @param         dec_weights_uquant_plane2   The quantized weight set for plane 2.
2014 * @param[in,out] ep                          The color endpoints (modifed in place).
2015 * @param[out]    rgbs_vector                 The RGB+scale color for LDR blocks.
2016 * @param[out]    rgbo_vector                 The RGB+offset color for HDR blocks.
2017 * @param         plane2_component            The component assigned to plane 2.
2018 */
2019void recompute_ideal_colors_2planes(
2020	const image_block& blk,
2021	const block_size_descriptor& bsd,
2022	const decimation_info& di,
2023	const uint8_t* dec_weights_uquant_plane1,
2024	const uint8_t* dec_weights_uquant_plane2,
2025	endpoints& ep,
2026	vfloat4& rgbs_vector,
2027	vfloat4& rgbo_vector,
2028	int plane2_component);
2029
2030/**
2031 * @brief Expand the angular tables needed for the alternative to PCA that we use.
2032 */
2033void prepare_angular_tables();
2034
2035/**
2036 * @brief Compute the angular endpoints for one plane for each block mode.
2037 *
2038 * @param      only_always              Only consider block modes that are always enabled.
2039 * @param      bsd                      The block size descriptor for the current trial.
2040 * @param      dec_weight_ideal_value   The ideal decimated unquantized weight values.
2041 * @param      max_weight_quant         The maximum block mode weight quantization allowed.
2042 * @param[out] tmpbuf                   Preallocated scratch buffers for the compressor.
2043 */
2044void compute_angular_endpoints_1plane(
2045	QualityProfile privateProfile,
2046	bool only_always,
2047	const block_size_descriptor& bsd,
2048	const float* dec_weight_ideal_value,
2049	unsigned int max_weight_quant,
2050	compression_working_buffers& tmpbuf);
2051
2052/**
2053 * @brief Compute the angular endpoints for two planes for each block mode.
2054 *
2055 * @param      bsd                      The block size descriptor for the current trial.
2056 * @param      dec_weight_ideal_value   The ideal decimated unquantized weight values.
2057 * @param      max_weight_quant         The maximum block mode weight quantization allowed.
2058 * @param[out] tmpbuf                   Preallocated scratch buffers for the compressor.
2059 */
2060void compute_angular_endpoints_2planes(
2061	QualityProfile privateProfile,
2062	const block_size_descriptor& bsd,
2063	const float* dec_weight_ideal_value,
2064	unsigned int max_weight_quant,
2065	compression_working_buffers& tmpbuf);
2066
2067/* ============================================================================
2068  Functionality for high level compression and decompression access.
2069============================================================================ */
2070
2071/**
2072 * @brief Compress an image block into a physical block.
2073 *
2074 * @param      ctx      The compressor context and configuration.
2075 * @param      blk      The image block color data to compress.
2076 * @param[out] pcb      The physical compressed block output.
2077 * @param[out] tmpbuf   Preallocated scratch buffers for the compressor.
2078 */
2079void compress_block(
2080	const astcenc_contexti& ctx,
2081	const image_block& blk,
2082	uint8_t pcb[16],
2083#if QUALITY_CONTROL
2084	compression_working_buffers& tmpbuf,
2085	bool calQualityEnable,
2086	int32_t *mseBlock[RGBA_COM]
2087#else
2088    compression_working_buffers& tmpbuf
2089#endif
2090	);
2091
2092/**
2093 * @brief Decompress a symbolic block in to an image block.
2094 *
2095 * @param      decode_mode   The decode mode (LDR, HDR, etc).
2096 * @param      bsd           The block size information.
2097 * @param      xpos          The X coordinate of the block in the overall image.
2098 * @param      ypos          The Y coordinate of the block in the overall image.
2099 * @param      zpos          The Z coordinate of the block in the overall image.
2100 * @param[out] blk           The decompressed image block color data.
2101 */
2102void decompress_symbolic_block(
2103	astcenc_profile decode_mode,
2104	const block_size_descriptor& bsd,
2105	int xpos,
2106	int ypos,
2107	int zpos,
2108	const symbolic_compressed_block& scb,
2109	image_block& blk);
2110
2111/**
2112 * @brief Compute the error between a symbolic block and the original input data.
2113 *
2114 * This function is specialized for 2 plane and 1 partition search.
2115 *
2116 * In RGBM mode this will reject blocks that attempt to encode a zero M value.
2117 *
2118 * @param config   The compressor config.
2119 * @param bsd      The block size information.
2120 * @param scb      The symbolic compressed encoding.
2121 * @param blk      The original image block color data.
2122 *
2123 * @return Returns the computed error, or a negative value if the encoding
2124 *         should be rejected for any reason.
2125 */
2126float compute_symbolic_block_difference_2plane(
2127	const astcenc_config& config,
2128	const block_size_descriptor& bsd,
2129	const symbolic_compressed_block& scb,
2130	const image_block& blk);
2131
2132/**
2133 * @brief Compute the error between a symbolic block and the original input data.
2134 *
2135 * This function is specialized for 1 plane and N partition search.
2136 *
2137 * In RGBM mode this will reject blocks that attempt to encode a zero M value.
2138 *
2139 * @param config   The compressor config.
2140 * @param bsd      The block size information.
2141 * @param scb      The symbolic compressed encoding.
2142 * @param blk      The original image block color data.
2143 *
2144 * @return Returns the computed error, or a negative value if the encoding
2145 *         should be rejected for any reason.
2146 */
2147float compute_symbolic_block_difference_1plane(
2148	const astcenc_config& config,
2149	const block_size_descriptor& bsd,
2150	const symbolic_compressed_block& scb,
2151	const image_block& blk);
2152
2153/**
2154 * @brief Compute the error between a symbolic block and the original input data.
2155 *
2156 * This function is specialized for 1 plane and 1 partition search.
2157 *
2158 * In RGBM mode this will reject blocks that attempt to encode a zero M value.
2159 *
2160 * @param config   The compressor config.
2161 * @param bsd      The block size information.
2162 * @param scb      The symbolic compressed encoding.
2163 * @param blk      The original image block color data.
2164 *
2165 * @return Returns the computed error, or a negative value if the encoding
2166 *         should be rejected for any reason.
2167 */
2168float compute_symbolic_block_difference_1plane_1partition(
2169	const astcenc_config& config,
2170	const block_size_descriptor& bsd,
2171	const symbolic_compressed_block& scb,
2172	const image_block& blk);
2173
2174/**
2175 * @brief Convert a symbolic representation into a binary physical encoding.
2176 *
2177 * It is assumed that the symbolic encoding is valid and encodable, or
2178 * previously flagged as an error block if an error color it to be encoded.
2179 *
2180 * @param      bsd   The block size information.
2181 * @param      scb   The symbolic representation.
2182 * @param[out] pcb   The physical compressed block output.
2183 */
2184void symbolic_to_physical(
2185	const block_size_descriptor& bsd,
2186	const symbolic_compressed_block& scb,
2187	uint8_t pcb[16]);
2188
2189/**
2190 * @brief Convert a binary physical encoding into a symbolic representation.
2191 *
2192 * This function can cope with arbitrary input data; output blocks will be
2193 * flagged as an error block if the encoding is invalid.
2194 *
2195 * @param      bsd   The block size information.
2196 * @param      pcb   The physical compresesd block input.
2197 * @param[out] scb   The output symbolic representation.
2198 */
2199void physical_to_symbolic(
2200	const block_size_descriptor& bsd,
2201	const uint8_t pcb[16],
2202	symbolic_compressed_block& scb);
2203
2204/* ============================================================================
2205Platform-specific functions.
2206============================================================================ */
2207/**
2208 * @brief Allocate an aligned memory buffer.
2209 *
2210 * Allocated memory must be freed by aligned_free.
2211 *
2212 * @param size    The desired buffer size.
2213 * @param align   The desired buffer alignment; must be 2^N, may be increased
2214 *                by the implementation to a minimum allowable alignment.
2215 *
2216 * @return The memory buffer pointer or nullptr on allocation failure.
2217 */
2218template<typename T>
2219T* aligned_malloc(size_t size, size_t align)
2220{
2221	void* ptr;
2222	int error = 0;
2223
2224	// Don't allow this to under-align a type
2225	size_t min_align = astc::max(alignof(T), sizeof(void*));
2226	size_t real_align = astc::max(min_align, align);
2227
2228#if defined(_WIN32)
2229	ptr = _aligned_malloc(size, real_align);
2230#else
2231	error = posix_memalign(&ptr, real_align, size);
2232#endif
2233
2234	if (error || (!ptr))
2235	{
2236		return nullptr;
2237	}
2238
2239	return static_cast<T*>(ptr);
2240}
2241
2242/**
2243 * @brief Free an aligned memory buffer.
2244 *
2245 * @param ptr   The buffer to free.
2246 */
2247template<typename T>
2248void aligned_free(T* ptr)
2249{
2250#if defined(_WIN32)
2251	_aligned_free(ptr);
2252#else
2253	free(ptr);
2254#endif
2255}
2256
2257#ifdef ASTC_CUSTOMIZED_ENABLE
2258#ifdef BUILD_HMOS_SDK
2259#if defined(_WIN32) && !defined(__CYGWIN__)
2260const LPCSTR g_astcCustomizedSo = "../../hms/toolchains/lib/libastcCustomizedEncode.dll";
2261#elif defined(__APPLE__)
2262const std::string g_astcCustomizedSo = "../../hms/toolchains/lib/libastcCustomizedEncode.dylib";
2263#else
2264const std::string g_astcCustomizedSo = "../../hms/toolchains/lib/libastcCustomizedEncode.so";
2265#endif
2266#else
2267const std::string g_astcCustomizedSo = "/system/lib64/module/hms/graphic/libastcCustomizedEncode.z.so";
2268#endif
2269using IsCustomizedBlockMode = bool (*)(const int);
2270using CustomizedMaxPartitions = int (*)();
2271using CustomizedBlockMode = int (*)();
2272
2273class AstcCustomizedSoManager
2274{
2275public:
2276	AstcCustomizedSoManager()
2277	{
2278		astcCustomizedSoOpened_ = false;
2279		astcCustomizedSoHandle_ = nullptr;
2280		isCustomizedBlockModeFunc_ = nullptr;
2281		customizedMaxPartitionsFunc_ = nullptr;
2282		customizedBlockModeFunc_ = nullptr;
2283	}
2284	~AstcCustomizedSoManager()
2285	{
2286		if (!astcCustomizedSoOpened_ || astcCustomizedSoHandle_ == nullptr)
2287		{
2288			printf("astcenc customized so is not be opened when dlclose!\n");
2289			return;
2290		}
2291#if defined(_WIN32) && !defined(__CYGWIN__)
2292		if (!FreeLibrary(astcCustomizedSoHandle_))
2293		{
2294			printf("astc dll FreeLibrary failed: %s\n", g_astcCustomizedSo);
2295		}
2296#else
2297		if (dlclose(astcCustomizedSoHandle_) != 0)
2298		{
2299			printf("astcenc so dlclose failed: %s\n", g_astcCustomizedSo.c_str());
2300		}
2301#endif
2302	}
2303	IsCustomizedBlockMode isCustomizedBlockModeFunc_;
2304	CustomizedMaxPartitions customizedMaxPartitionsFunc_;
2305	CustomizedBlockMode customizedBlockModeFunc_;
2306	bool LoadSutCustomizedSo()
2307	{
2308		if (!astcCustomizedSoOpened_)
2309		{
2310#if defined(_WIN32) && !defined(__CYGWIN__)
2311			if ((_access(g_astcCustomizedSo, 0) == -1))
2312			{
2313				printf("astc customized dll(%s) is not found!\n", g_astcCustomizedSo);
2314				return false;
2315			}
2316			astcCustomizedSoHandle_ = LoadLibrary(g_astcCustomizedSo);
2317			if (astcCustomizedSoHandle_ == nullptr)
2318			{
2319				printf("astc libAstcCustomizedEnc LoadLibrary failed!\n");
2320				return false;
2321			}
2322			isCustomizedBlockModeFunc_ =
2323				reinterpret_cast<IsCustomizedBlockMode>(GetProcAddress(astcCustomizedSoHandle_,
2324				"IsCustomizedBlockMode"));
2325			if (isCustomizedBlockModeFunc_ == nullptr)
2326			{
2327				printf("astc isCustomizedBlockModeFunc_ GetProcAddress failed!\n");
2328				if (!FreeLibrary(astcCustomizedSoHandle_))
2329				{
2330					printf("astc isCustomizedBlockModeFunc_ FreeLibrary failed!\n");
2331				}
2332				return false;
2333			}
2334			customizedMaxPartitionsFunc_ =
2335				reinterpret_cast<CustomizedMaxPartitions>(GetProcAddress(astcCustomizedSoHandle_,
2336				"CustomizedMaxPartitions"));
2337			if (customizedMaxPartitionsFunc_ == nullptr)
2338			{
2339				printf("astc customizedMaxPartitionsFunc_ GetProcAddress failed!\n");
2340				if (!FreeLibrary(astcCustomizedSoHandle_))
2341				{
2342					printf("astc customizedMaxPartitionsFunc_ FreeLibrary failed!\n");
2343				}
2344				return false;
2345			}
2346			customizedBlockModeFunc_ =
2347				reinterpret_cast<CustomizedBlockMode>(GetProcAddress(astcCustomizedSoHandle_,
2348				"CustomizedBlockMode"));
2349			if (customizedBlockModeFunc_ == nullptr)
2350			{
2351				printf("astc customizedBlockModeFunc_ GetProcAddress failed!\n");
2352				if (!FreeLibrary(astcCustomizedSoHandle_))
2353				{
2354					printf("astc customizedBlockModeFunc_ FreeLibrary failed!\n");
2355				}
2356				return false;
2357			}
2358			printf("astcenc customized dll load success: %s!\n", g_astcCustomizedSo);
2359#else
2360			if (access(g_astcCustomizedSo.c_str(), F_OK) == -1)
2361			{
2362				printf("astc customized so(%s) is not found!\n", g_astcCustomizedSo.c_str());
2363				return false;
2364			}
2365			astcCustomizedSoHandle_ = dlopen(g_astcCustomizedSo.c_str(), 1);
2366			if (astcCustomizedSoHandle_ == nullptr)
2367			{
2368				printf("astc libAstcCustomizedEnc dlopen failed!\n");
2369				return false;
2370			}
2371			isCustomizedBlockModeFunc_ =
2372				reinterpret_cast<IsCustomizedBlockMode>(dlsym(astcCustomizedSoHandle_,
2373				"IsCustomizedBlockMode"));
2374			if (isCustomizedBlockModeFunc_ == nullptr)
2375			{
2376				printf("astc isCustomizedBlockModeFunc_ dlsym failed!\n");
2377				dlclose(astcCustomizedSoHandle_);
2378				astcCustomizedSoHandle_ = nullptr;
2379				return false;
2380			}
2381			customizedMaxPartitionsFunc_ =
2382				reinterpret_cast<CustomizedMaxPartitions>(dlsym(astcCustomizedSoHandle_,
2383				"CustomizedMaxPartitions"));
2384			if (customizedMaxPartitionsFunc_ == nullptr)
2385			{
2386				printf("astc customizedMaxPartitionsFunc_ dlsym failed!\n");
2387				dlclose(astcCustomizedSoHandle_);
2388				astcCustomizedSoHandle_ = nullptr;
2389				return false;
2390			}
2391			customizedBlockModeFunc_ =
2392				reinterpret_cast<CustomizedBlockMode>(dlsym(astcCustomizedSoHandle_,
2393				"CustomizedBlockMode"));
2394			if (customizedBlockModeFunc_ == nullptr)
2395			{
2396				printf("astc customizedBlockModeFunc_ dlsym failed!\n");
2397				dlclose(astcCustomizedSoHandle_);
2398				astcCustomizedSoHandle_ = nullptr;
2399				return false;
2400			}
2401			printf("astcenc customized so dlopen success: %s\n", g_astcCustomizedSo.c_str());
2402#endif
2403			astcCustomizedSoOpened_ = true;
2404		}
2405		return true;
2406	}
2407private:
2408	bool astcCustomizedSoOpened_;
2409#if defined(_WIN32) && !defined(__CYGWIN__)
2410	HINSTANCE astcCustomizedSoHandle_;
2411#else
2412	void *astcCustomizedSoHandle_;
2413#endif
2414};
2415extern AstcCustomizedSoManager g_astcCustomizedSoManager;
2416#endif
2417
2418#endif
2419