1 // SPDX-License-Identifier: Apache-2.0
2 // ----------------------------------------------------------------------------
3 // Copyright 2011-2021 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 for converting between symbolic and physical encodings.
20 */
21
22 #include "astcenc_internal.h"
23
24 #include <cassert>
25
26 /**
27 * @brief Write up to 8 bits at an arbitrary bit offset.
28 *
29 * The stored value is at most 8 bits, but can be stored at an offset of between 0 and 7 bits so
30 * may span two separate bytes in memory.
31 *
32 * @param value The value to write.
33 * @param bitcount The number of bits to write, starting from LSB.
34 * @param bitoffset The bit offset to store at, between 0 and 7.
35 * @param[in,out] ptr The data pointer to write to.
36 */
write_bits(int value,int bitcount,int bitoffset,uint8_t * ptr)37 static inline void write_bits(
38 int value,
39 int bitcount,
40 int bitoffset,
41 uint8_t* ptr
42 ) {
43 int mask = (1 << bitcount) - 1;
44 value &= mask;
45 ptr += bitoffset >> 3;
46 bitoffset &= 7;
47 value <<= bitoffset;
48 mask <<= bitoffset;
49 mask = ~mask;
50
51 ptr[0] &= mask;
52 ptr[0] |= value;
53 ptr[1] &= mask >> 8;
54 ptr[1] |= value >> 8;
55 }
56
57 /**
58 * @brief Read up to 8 bits at an arbitrary bit offset.
59 *
60 * The stored value is at most 8 bits, but can be stored at an offset of between 0 and 7 bits so may
61 * span two separate bytes in memory.
62 *
63 * @param bitcount The number of bits to read.
64 * @param bitoffset The bit offset to read from, between 0 and 7.
65 * @param[in,out] ptr The data pointer to read from.
66 *
67 * @return The read value.
68 */
read_bits(int bitcount,int bitoffset,const uint8_t * ptr)69 static inline int read_bits(
70 int bitcount,
71 int bitoffset,
72 const uint8_t* ptr
73 ) {
74 int mask = (1 << bitcount) - 1;
75 ptr += bitoffset >> 3;
76 bitoffset &= 7;
77 int value = ptr[0] | (ptr[1] << 8);
78 value >>= bitoffset;
79 value &= mask;
80 return value;
81 }
82
83 /**
84 * @brief Reverse bits in a byte.
85 *
86 * @param p The value to reverse.
87 *
88 * @return The reversed result.
89 */
bitrev8(int p)90 static inline int bitrev8(int p)
91 {
92 p = ((p & 0x0F) << 4) | ((p >> 4) & 0x0F);
93 p = ((p & 0x33) << 2) | ((p >> 2) & 0x33);
94 p = ((p & 0x55) << 1) | ((p >> 1) & 0x55);
95 return p;
96 }
97
98 /* See header for documentation. */
symbolic_to_physical(const block_size_descriptor & bsd,const symbolic_compressed_block & scb,physical_compressed_block & pcb)99 void symbolic_to_physical(
100 const block_size_descriptor& bsd,
101 const symbolic_compressed_block& scb,
102 physical_compressed_block& pcb
103 ) {
104 assert(scb.block_type != SYM_BTYPE_ERROR);
105
106 // Constant color block using UNORM16 colors
107 if (scb.block_type == SYM_BTYPE_CONST_U16)
108 {
109 // There is currently no attempt to coalesce larger void-extents
110 static const uint8_t cbytes[8] { 0xFC, 0xFD, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF };
111 for (unsigned int i = 0; i < 8; i++)
112 {
113 pcb.data[i] = cbytes[i];
114 }
115
116 for (unsigned int i = 0; i < BLOCK_MAX_COMPONENTS; i++)
117 {
118 pcb.data[2 * i + 8] = scb.constant_color[i] & 0xFF;
119 pcb.data[2 * i + 9] = (scb.constant_color[i] >> 8) & 0xFF;
120 }
121
122 return;
123 }
124
125 // Constant color block using FP16 colors
126 if (scb.block_type == SYM_BTYPE_CONST_F16)
127 {
128 // There is currently no attempt to coalesce larger void-extents
129 static const uint8_t cbytes[8] { 0xFC, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF };
130 for (unsigned int i = 0; i < 8; i++)
131 {
132 pcb.data[i] = cbytes[i];
133 }
134
135 for (unsigned int i = 0; i < BLOCK_MAX_COMPONENTS; i++)
136 {
137 pcb.data[2 * i + 8] = scb.constant_color[i] & 0xFF;
138 pcb.data[2 * i + 9] = (scb.constant_color[i] >> 8) & 0xFF;
139 }
140
141 return;
142 }
143
144 unsigned int partition_count = scb.partition_count;
145
146 // Compress the weights.
147 // They are encoded as an ordinary integer-sequence, then bit-reversed
148 uint8_t weightbuf[16] { 0 };
149
150 const auto& bm = bsd.get_block_mode(scb.block_mode);
151 const auto& di = bsd.get_decimation_info(bm.decimation_mode);
152 int weight_count = di.weight_count;
153 quant_method weight_quant_method = bm.get_weight_quant_mode();
154 float weight_quant_levels = static_cast<float>(get_quant_level(weight_quant_method));
155 int is_dual_plane = bm.is_dual_plane;
156
157 const auto& qat = quant_and_xfer_tables[weight_quant_method];
158
159 int real_weight_count = is_dual_plane ? 2 * weight_count : weight_count;
160
161 int bits_for_weights = get_ise_sequence_bitcount(real_weight_count, weight_quant_method);
162
163 uint8_t weights[64];
164 if (is_dual_plane)
165 {
166 for (int i = 0; i < weight_count; i++)
167 {
168 float uqw = static_cast<float>(scb.weights[i]);
169 float qw = (uqw / 64.0f) * (weight_quant_levels - 1.0f);
170 int qwi = static_cast<int>(qw + 0.5f);
171 weights[2 * i] = qat.scramble_map[qwi];
172
173 uqw = static_cast<float>(scb.weights[i + WEIGHTS_PLANE2_OFFSET]);
174 qw = (uqw / 64.0f) * (weight_quant_levels - 1.0f);
175 qwi = static_cast<int>(qw + 0.5f);
176 weights[2 * i + 1] = qat.scramble_map[qwi];
177 }
178 }
179 else
180 {
181 for (int i = 0; i < weight_count; i++)
182 {
183 float uqw = static_cast<float>(scb.weights[i]);
184 float qw = (uqw / 64.0f) * (weight_quant_levels - 1.0f);
185 int qwi = static_cast<int>(qw + 0.5f);
186 weights[i] = qat.scramble_map[qwi];
187 }
188 }
189
190 encode_ise(weight_quant_method, real_weight_count, weights, weightbuf, 0);
191
192 for (int i = 0; i < 16; i++)
193 {
194 pcb.data[i] = static_cast<uint8_t>(bitrev8(weightbuf[15 - i]));
195 }
196
197 write_bits(scb.block_mode, 11, 0, pcb.data);
198 write_bits(partition_count - 1, 2, 11, pcb.data);
199
200 int below_weights_pos = 128 - bits_for_weights;
201
202 // Encode partition index and color endpoint types for blocks with 2+ partitions
203 if (partition_count > 1)
204 {
205 write_bits(scb.partition_index, 6, 13, pcb.data);
206 write_bits(scb.partition_index >> 6, PARTITION_INDEX_BITS - 6, 19, pcb.data);
207
208 if (scb.color_formats_matched)
209 {
210 write_bits(scb.color_formats[0] << 2, 6, 13 + PARTITION_INDEX_BITS, pcb.data);
211 }
212 else
213 {
214 // Check endpoint types for each partition to determine the lowest class present
215 int low_class = 4;
216
217 for (unsigned int i = 0; i < partition_count; i++)
218 {
219 int class_of_format = scb.color_formats[i] >> 2;
220 low_class = astc::min(class_of_format, low_class);
221 }
222
223 if (low_class == 3)
224 {
225 low_class = 2;
226 }
227
228 int encoded_type = low_class + 1;
229 int bitpos = 2;
230
231 for (unsigned int i = 0; i < partition_count; i++)
232 {
233 int classbit_of_format = (scb.color_formats[i] >> 2) - low_class;
234 encoded_type |= classbit_of_format << bitpos;
235 bitpos++;
236 }
237
238 for (unsigned int i = 0; i < partition_count; i++)
239 {
240 int lowbits_of_format = scb.color_formats[i] & 3;
241 encoded_type |= lowbits_of_format << bitpos;
242 bitpos += 2;
243 }
244
245 int encoded_type_lowpart = encoded_type & 0x3F;
246 int encoded_type_highpart = encoded_type >> 6;
247 int encoded_type_highpart_size = (3 * partition_count) - 4;
248 int encoded_type_highpart_pos = 128 - bits_for_weights - encoded_type_highpart_size;
249 write_bits(encoded_type_lowpart, 6, 13 + PARTITION_INDEX_BITS, pcb.data);
250 write_bits(encoded_type_highpart, encoded_type_highpart_size, encoded_type_highpart_pos, pcb.data);
251 below_weights_pos -= encoded_type_highpart_size;
252 }
253 }
254 else
255 {
256 write_bits(scb.color_formats[0], 4, 13, pcb.data);
257 }
258
259 // In dual-plane mode, encode the color component of the second plane of weights
260 if (is_dual_plane)
261 {
262 write_bits(scb.plane2_component, 2, below_weights_pos - 2, pcb.data);
263 }
264
265 // Encode the color components
266 uint8_t values_to_encode[32];
267 int valuecount_to_encode = 0;
268 for (unsigned int i = 0; i < scb.partition_count; i++)
269 {
270 int vals = 2 * (scb.color_formats[i] >> 2) + 2;
271 assert(vals <= 8);
272 for (int j = 0; j < vals; j++)
273 {
274 values_to_encode[j + valuecount_to_encode] = scb.color_values[i][j];
275 }
276 valuecount_to_encode += vals;
277 }
278
279 encode_ise(scb.get_color_quant_mode(), valuecount_to_encode, values_to_encode, pcb.data,
280 scb.partition_count == 1 ? 17 : 19 + PARTITION_INDEX_BITS);
281 }
282
283 /* See header for documentation. */
physical_to_symbolic(const block_size_descriptor & bsd,const physical_compressed_block & pcb,symbolic_compressed_block & scb)284 void physical_to_symbolic(
285 const block_size_descriptor& bsd,
286 const physical_compressed_block& pcb,
287 symbolic_compressed_block& scb
288 ) {
289 uint8_t bswapped[16];
290
291 scb.block_type = SYM_BTYPE_NONCONST;
292
293 // Extract header fields
294 int block_mode = read_bits(11, 0, pcb.data);
295 if ((block_mode & 0x1FF) == 0x1FC)
296 {
297 // Constant color block
298
299 // Check what format the data has
300 if (block_mode & 0x200)
301 {
302 scb.block_type = SYM_BTYPE_CONST_F16;
303 }
304 else
305 {
306 scb.block_type = SYM_BTYPE_CONST_U16;
307 }
308
309 scb.partition_count = 0;
310 for (int i = 0; i < 4; i++)
311 {
312 scb.constant_color[i] = pcb.data[2 * i + 8] | (pcb.data[2 * i + 9] << 8);
313 }
314
315 // Additionally, check that the void-extent
316 if (bsd.zdim == 1)
317 {
318 // 2D void-extent
319 int rsvbits = read_bits(2, 10, pcb.data);
320 if (rsvbits != 3)
321 {
322 scb.block_type = SYM_BTYPE_ERROR;
323 return;
324 }
325
326 int vx_low_s = read_bits(8, 12, pcb.data) | (read_bits(5, 12 + 8, pcb.data) << 8);
327 int vx_high_s = read_bits(8, 25, pcb.data) | (read_bits(5, 25 + 8, pcb.data) << 8);
328 int vx_low_t = read_bits(8, 38, pcb.data) | (read_bits(5, 38 + 8, pcb.data) << 8);
329 int vx_high_t = read_bits(8, 51, pcb.data) | (read_bits(5, 51 + 8, pcb.data) << 8);
330
331 int all_ones = vx_low_s == 0x1FFF && vx_high_s == 0x1FFF && vx_low_t == 0x1FFF && vx_high_t == 0x1FFF;
332
333 if ((vx_low_s >= vx_high_s || vx_low_t >= vx_high_t) && !all_ones)
334 {
335 scb.block_type = SYM_BTYPE_ERROR;
336 return;
337 }
338 }
339 else
340 {
341 // 3D void-extent
342 int vx_low_s = read_bits(9, 10, pcb.data);
343 int vx_high_s = read_bits(9, 19, pcb.data);
344 int vx_low_t = read_bits(9, 28, pcb.data);
345 int vx_high_t = read_bits(9, 37, pcb.data);
346 int vx_low_p = read_bits(9, 46, pcb.data);
347 int vx_high_p = read_bits(9, 55, pcb.data);
348
349 int all_ones = vx_low_s == 0x1FF && vx_high_s == 0x1FF && vx_low_t == 0x1FF && vx_high_t == 0x1FF && vx_low_p == 0x1FF && vx_high_p == 0x1FF;
350
351 if ((vx_low_s >= vx_high_s || vx_low_t >= vx_high_t || vx_low_p >= vx_high_p) && !all_ones)
352 {
353 scb.block_type = SYM_BTYPE_ERROR;
354 return;
355 }
356 }
357
358 return;
359 }
360
361 unsigned int packed_index = bsd.block_mode_packed_index[block_mode];
362 if (packed_index == BLOCK_BAD_BLOCK_MODE)
363 {
364 scb.block_type = SYM_BTYPE_ERROR;
365 return;
366 }
367
368 const auto& bm = bsd.get_block_mode(block_mode);
369 const auto& di = bsd.get_decimation_info(bm.decimation_mode);
370
371 int weight_count = di.weight_count;
372 quant_method weight_quant_method = static_cast<quant_method>(bm.quant_mode);
373 int is_dual_plane = bm.is_dual_plane;
374
375 int real_weight_count = is_dual_plane ? 2 * weight_count : weight_count;
376
377 int partition_count = read_bits(2, 11, pcb.data) + 1;
378
379 scb.block_mode = static_cast<uint16_t>(block_mode);
380 scb.partition_count = static_cast<uint8_t>(partition_count);
381
382 for (int i = 0; i < 16; i++)
383 {
384 bswapped[i] = static_cast<uint8_t>(bitrev8(pcb.data[15 - i]));
385 }
386
387 int bits_for_weights = get_ise_sequence_bitcount(real_weight_count, weight_quant_method);
388
389 int below_weights_pos = 128 - bits_for_weights;
390
391 uint8_t indices[64];
392 const auto& qat = quant_and_xfer_tables[weight_quant_method];
393
394 decode_ise(weight_quant_method, real_weight_count, bswapped, indices, 0);
395
396 if (is_dual_plane)
397 {
398 for (int i = 0; i < weight_count; i++)
399 {
400 scb.weights[i] = qat.unscramble_and_unquant_map[indices[2 * i]];
401 scb.weights[i + WEIGHTS_PLANE2_OFFSET] = qat.unscramble_and_unquant_map[indices[2 * i + 1]];
402 }
403 }
404 else
405 {
406 for (int i = 0; i < weight_count; i++)
407 {
408 scb.weights[i] = qat.unscramble_and_unquant_map[indices[i]];
409 }
410 }
411
412 if (is_dual_plane && partition_count == 4)
413 {
414 scb.block_type = SYM_BTYPE_ERROR;
415 return;
416 }
417
418 scb.color_formats_matched = 0;
419
420 // Determine the format of each endpoint pair
421 int color_formats[BLOCK_MAX_PARTITIONS];
422 int encoded_type_highpart_size = 0;
423 if (partition_count == 1)
424 {
425 color_formats[0] = read_bits(4, 13, pcb.data);
426 scb.partition_index = 0;
427 }
428 else
429 {
430 encoded_type_highpart_size = (3 * partition_count) - 4;
431 below_weights_pos -= encoded_type_highpart_size;
432 int encoded_type = read_bits(6, 13 + PARTITION_INDEX_BITS, pcb.data) | (read_bits(encoded_type_highpart_size, below_weights_pos, pcb.data) << 6);
433 int baseclass = encoded_type & 0x3;
434 if (baseclass == 0)
435 {
436 for (int i = 0; i < partition_count; i++)
437 {
438 color_formats[i] = (encoded_type >> 2) & 0xF;
439 }
440
441 below_weights_pos += encoded_type_highpart_size;
442 scb.color_formats_matched = 1;
443 encoded_type_highpart_size = 0;
444 }
445 else
446 {
447 int bitpos = 2;
448 baseclass--;
449
450 for (int i = 0; i < partition_count; i++)
451 {
452 color_formats[i] = (((encoded_type >> bitpos) & 1) + baseclass) << 2;
453 bitpos++;
454 }
455
456 for (int i = 0; i < partition_count; i++)
457 {
458 color_formats[i] |= (encoded_type >> bitpos) & 3;
459 bitpos += 2;
460 }
461 }
462 scb.partition_index = static_cast<uint16_t>(read_bits(6, 13, pcb.data) | (read_bits(PARTITION_INDEX_BITS - 6, 19, pcb.data) << 6));
463 }
464
465 for (int i = 0; i < partition_count; i++)
466 {
467 scb.color_formats[i] = static_cast<uint8_t>(color_formats[i]);
468 }
469
470 // Determine number of color endpoint integers
471 int color_integer_count = 0;
472 for (int i = 0; i < partition_count; i++)
473 {
474 int endpoint_class = color_formats[i] >> 2;
475 color_integer_count += (endpoint_class + 1) * 2;
476 }
477
478 if (color_integer_count > 18)
479 {
480 scb.block_type = SYM_BTYPE_ERROR;
481 return;
482 }
483
484 // Determine the color endpoint format to use
485 static const int color_bits_arr[5] { -1, 115 - 4, 113 - 4 - PARTITION_INDEX_BITS, 113 - 4 - PARTITION_INDEX_BITS, 113 - 4 - PARTITION_INDEX_BITS };
486 int color_bits = color_bits_arr[partition_count] - bits_for_weights - encoded_type_highpart_size;
487 if (is_dual_plane)
488 {
489 color_bits -= 2;
490 }
491
492 if (color_bits < 0)
493 {
494 color_bits = 0;
495 }
496
497 int color_quant_level = quant_mode_table[color_integer_count >> 1][color_bits];
498 if (color_quant_level < QUANT_6)
499 {
500 scb.block_type = SYM_BTYPE_ERROR;
501 return;
502 }
503
504 // Unpack the integer color values and assign to endpoints
505 scb.quant_mode = static_cast<quant_method>(color_quant_level);
506 uint8_t values_to_decode[32];
507 decode_ise(static_cast<quant_method>(color_quant_level), color_integer_count, pcb.data,
508 values_to_decode, (partition_count == 1 ? 17 : 19 + PARTITION_INDEX_BITS));
509
510 int valuecount_to_decode = 0;
511 for (int i = 0; i < partition_count; i++)
512 {
513 int vals = 2 * (color_formats[i] >> 2) + 2;
514 for (int j = 0; j < vals; j++)
515 {
516 scb.color_values[i][j] = values_to_decode[j + valuecount_to_decode];
517 }
518 valuecount_to_decode += vals;
519 }
520
521 // Fetch component for second-plane in the case of dual plane of weights.
522 if (is_dual_plane)
523 {
524 scb.plane2_component = static_cast<int8_t>(read_bits(2, below_weights_pos - 2, pcb.data));
525 }
526 }
527