1 | // SPDX-License-Identifier: Apache-2.0 |
2 | // ---------------------------------------------------------------------------- |
3 | // Copyright 2011-2023 Arm Limited |
4 | // |
5 | // Licensed under the Apache License, Version 2.0 (the "License"); you may not |
6 | // use this file except in compliance with the License. You may obtain a copy |
7 | // of the License at: |
8 | // |
9 | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | // |
11 | // Unless required by applicable law or agreed to in writing, software |
12 | // distributed under the License is distributed on an "AS IS" BASIS, WITHOUT |
13 | // WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the |
14 | // License for the specific language governing permissions and limitations |
15 | // under the License. |
16 | // ---------------------------------------------------------------------------- |
17 | |
18 | #if !defined(ASTCENC_DECOMPRESS_ONLY) |
19 | |
20 | /** |
21 | * @brief Functions for angular-sum algorithm for weight alignment. |
22 | * |
23 | * This algorithm works as follows: |
24 | * - we compute a complex number P as (cos s*i, sin s*i) for each weight, |
25 | * where i is the input value and s is a scaling factor based on the spacing between the weights. |
26 | * - we then add together complex numbers for all the weights. |
27 | * - we then compute the length and angle of the resulting sum. |
28 | * |
29 | * This should produce the following results: |
30 | * - perfect alignment results in a vector whose length is equal to the sum of lengths of all inputs |
31 | * - even distribution results in a vector of length 0. |
32 | * - all samples identical results in perfect alignment for every scaling. |
33 | * |
34 | * For each scaling factor within a given set, we compute an alignment factor from 0 to 1. This |
35 | * should then result in some scalings standing out as having particularly good alignment factors; |
36 | * we can use this to produce a set of candidate scale/shift values for various quantization levels; |
37 | * we should then actually try them and see what happens. |
38 | */ |
39 | |
40 | #include "astcenc_internal.h" |
41 | #include "astcenc_vecmathlib.h" |
42 | |
43 | #include <stdio.h> |
44 | #include <cassert> |
45 | #include <cstring> |
46 | |
47 | static constexpr unsigned int ANGULAR_STEPS { 32 }; |
48 | |
49 | static_assert((ANGULAR_STEPS % ASTCENC_SIMD_WIDTH) == 0, |
50 | "ANGULAR_STEPS must be multiple of ASTCENC_SIMD_WIDTH" ); |
51 | |
52 | static_assert(ANGULAR_STEPS >= 32, |
53 | "ANGULAR_STEPS must be at least max(steps_for_quant_level)" ); |
54 | |
55 | // Store a reduced sin/cos table for 64 possible weight values; this causes |
56 | // slight quality loss compared to using sin() and cos() directly. Must be 2^N. |
57 | static constexpr unsigned int SINCOS_STEPS { 64 }; |
58 | |
59 | static const uint8_t steps_for_quant_level[12] { |
60 | 2, 3, 4, 5, 6, 8, 10, 12, 16, 20, 24, 32 |
61 | }; |
62 | |
63 | alignas(ASTCENC_VECALIGN) static float sin_table[SINCOS_STEPS][ANGULAR_STEPS]; |
64 | alignas(ASTCENC_VECALIGN) static float cos_table[SINCOS_STEPS][ANGULAR_STEPS]; |
65 | |
66 | #if defined(ASTCENC_DIAGNOSTICS) |
67 | static bool print_once { true }; |
68 | #endif |
69 | |
70 | /* See header for documentation. */ |
71 | void prepare_angular_tables() |
72 | { |
73 | for (unsigned int i = 0; i < ANGULAR_STEPS; i++) |
74 | { |
75 | float angle_step = static_cast<float>(i + 1); |
76 | |
77 | for (unsigned int j = 0; j < SINCOS_STEPS; j++) |
78 | { |
79 | sin_table[j][i] = static_cast<float>(sinf((2.0f * astc::PI / (SINCOS_STEPS - 1.0f)) * angle_step * static_cast<float>(j))); |
80 | cos_table[j][i] = static_cast<float>(cosf((2.0f * astc::PI / (SINCOS_STEPS - 1.0f)) * angle_step * static_cast<float>(j))); |
81 | } |
82 | } |
83 | } |
84 | |
85 | /** |
86 | * @brief Compute the angular alignment factors and offsets. |
87 | * |
88 | * @param weight_count The number of (decimated) weights. |
89 | * @param dec_weight_ideal_value The ideal decimated unquantized weight values. |
90 | * @param max_angular_steps The maximum number of steps to be tested. |
91 | * @param[out] offsets The output angular offsets array. |
92 | */ |
93 | static void compute_angular_offsets( |
94 | unsigned int weight_count, |
95 | const float* dec_weight_ideal_value, |
96 | unsigned int max_angular_steps, |
97 | float* offsets |
98 | ) { |
99 | promise(weight_count > 0); |
100 | promise(max_angular_steps > 0); |
101 | |
102 | alignas(ASTCENC_VECALIGN) int isamplev[BLOCK_MAX_WEIGHTS]; |
103 | |
104 | // Precompute isample; arrays are always allocated 64 elements long |
105 | for (unsigned int i = 0; i < weight_count; i += ASTCENC_SIMD_WIDTH) |
106 | { |
107 | // Add 2^23 and interpreting bits extracts round-to-nearest int |
108 | vfloat sample = loada(dec_weight_ideal_value + i) * (SINCOS_STEPS - 1.0f) + vfloat(12582912.0f); |
109 | vint isample = float_as_int(sample) & vint((SINCOS_STEPS - 1)); |
110 | storea(isample, isamplev + i); |
111 | } |
112 | |
113 | // Arrays are multiple of SIMD width (ANGULAR_STEPS), safe to overshoot max |
114 | vfloat mult = vfloat(1.0f / (2.0f * astc::PI)); |
115 | |
116 | for (unsigned int i = 0; i < max_angular_steps; i += ASTCENC_SIMD_WIDTH) |
117 | { |
118 | vfloat anglesum_x = vfloat::zero(); |
119 | vfloat anglesum_y = vfloat::zero(); |
120 | |
121 | for (unsigned int j = 0; j < weight_count; j++) |
122 | { |
123 | int isample = isamplev[j]; |
124 | anglesum_x += loada(cos_table[isample] + i); |
125 | anglesum_y += loada(sin_table[isample] + i); |
126 | } |
127 | |
128 | vfloat angle = atan2(anglesum_y, anglesum_x); |
129 | vfloat ofs = angle * mult; |
130 | storea(ofs, offsets + i); |
131 | } |
132 | } |
133 | |
134 | /** |
135 | * @brief For a given step size compute the lowest and highest weight. |
136 | * |
137 | * Compute the lowest and highest weight that results from quantizing using the given stepsize and |
138 | * offset, and then compute the resulting error. The cut errors indicate the error that results from |
139 | * forcing samples that should have had one weight value one step up or down. |
140 | * |
141 | * @param weight_count The number of (decimated) weights. |
142 | * @param dec_weight_ideal_value The ideal decimated unquantized weight values. |
143 | * @param max_angular_steps The maximum number of steps to be tested. |
144 | * @param max_quant_steps The maximum quantization level to be tested. |
145 | * @param offsets The angular offsets array. |
146 | * @param[out] lowest_weight Per angular step, the lowest weight. |
147 | * @param[out] weight_span Per angular step, the span between lowest and highest weight. |
148 | * @param[out] error Per angular step, the error. |
149 | * @param[out] cut_low_weight_error Per angular step, the low weight cut error. |
150 | * @param[out] cut_high_weight_error Per angular step, the high weight cut error. |
151 | */ |
152 | static void compute_lowest_and_highest_weight( |
153 | unsigned int weight_count, |
154 | const float* dec_weight_ideal_value, |
155 | unsigned int max_angular_steps, |
156 | unsigned int max_quant_steps, |
157 | const float* offsets, |
158 | float* lowest_weight, |
159 | int* weight_span, |
160 | float* error, |
161 | float* cut_low_weight_error, |
162 | float* cut_high_weight_error |
163 | ) { |
164 | promise(weight_count > 0); |
165 | promise(max_angular_steps > 0); |
166 | |
167 | vfloat rcp_stepsize = vfloat::lane_id() + vfloat(1.0f); |
168 | |
169 | // Arrays are ANGULAR_STEPS long, so always safe to run full vectors |
170 | for (unsigned int sp = 0; sp < max_angular_steps; sp += ASTCENC_SIMD_WIDTH) |
171 | { |
172 | vfloat minidx(128.0f); |
173 | vfloat maxidx(-128.0f); |
174 | vfloat errval = vfloat::zero(); |
175 | vfloat cut_low_weight_err = vfloat::zero(); |
176 | vfloat cut_high_weight_err = vfloat::zero(); |
177 | vfloat offset = loada(offsets + sp); |
178 | |
179 | for (unsigned int j = 0; j < weight_count; j++) |
180 | { |
181 | vfloat sval = load1(dec_weight_ideal_value + j) * rcp_stepsize - offset; |
182 | vfloat svalrte = round(sval); |
183 | vfloat diff = sval - svalrte; |
184 | errval += diff * diff; |
185 | |
186 | // Reset tracker on min hit |
187 | vmask mask = svalrte < minidx; |
188 | minidx = select(minidx, svalrte, mask); |
189 | cut_low_weight_err = select(cut_low_weight_err, vfloat::zero(), mask); |
190 | |
191 | // Accumulate on min hit |
192 | mask = svalrte == minidx; |
193 | vfloat accum = cut_low_weight_err + vfloat(1.0f) - vfloat(2.0f) * diff; |
194 | cut_low_weight_err = select(cut_low_weight_err, accum, mask); |
195 | |
196 | // Reset tracker on max hit |
197 | mask = svalrte > maxidx; |
198 | maxidx = select(maxidx, svalrte, mask); |
199 | cut_high_weight_err = select(cut_high_weight_err, vfloat::zero(), mask); |
200 | |
201 | // Accumulate on max hit |
202 | mask = svalrte == maxidx; |
203 | accum = cut_high_weight_err + vfloat(1.0f) + vfloat(2.0f) * diff; |
204 | cut_high_weight_err = select(cut_high_weight_err, accum, mask); |
205 | } |
206 | |
207 | // Write out min weight and weight span; clamp span to a usable range |
208 | vint span = float_to_int(maxidx - minidx + vfloat(1)); |
209 | span = min(span, vint(max_quant_steps + 3)); |
210 | span = max(span, vint(2)); |
211 | storea(minidx, lowest_weight + sp); |
212 | storea(span, weight_span + sp); |
213 | |
214 | // The cut_(lowest/highest)_weight_error indicate the error that results from forcing |
215 | // samples that should have had the weight value one step (up/down). |
216 | vfloat ssize = 1.0f / rcp_stepsize; |
217 | vfloat errscale = ssize * ssize; |
218 | storea(errval * errscale, error + sp); |
219 | storea(cut_low_weight_err * errscale, cut_low_weight_error + sp); |
220 | storea(cut_high_weight_err * errscale, cut_high_weight_error + sp); |
221 | |
222 | rcp_stepsize = rcp_stepsize + vfloat(ASTCENC_SIMD_WIDTH); |
223 | } |
224 | } |
225 | |
226 | /** |
227 | * @brief The main function for the angular algorithm. |
228 | * |
229 | * @param weight_count The number of (decimated) weights. |
230 | * @param dec_weight_ideal_value The ideal decimated unquantized weight values. |
231 | * @param max_quant_level The maximum quantization level to be tested. |
232 | * @param[out] low_value Per angular step, the lowest weight value. |
233 | * @param[out] high_value Per angular step, the highest weight value. |
234 | */ |
235 | static void compute_angular_endpoints_for_quant_levels( |
236 | unsigned int weight_count, |
237 | const float* dec_weight_ideal_value, |
238 | unsigned int max_quant_level, |
239 | float low_value[TUNE_MAX_ANGULAR_QUANT + 1], |
240 | float high_value[TUNE_MAX_ANGULAR_QUANT + 1] |
241 | ) { |
242 | unsigned int max_quant_steps = steps_for_quant_level[max_quant_level]; |
243 | unsigned int max_angular_steps = steps_for_quant_level[max_quant_level]; |
244 | |
245 | alignas(ASTCENC_VECALIGN) float angular_offsets[ANGULAR_STEPS]; |
246 | |
247 | compute_angular_offsets(weight_count, dec_weight_ideal_value, |
248 | max_angular_steps, angular_offsets); |
249 | |
250 | alignas(ASTCENC_VECALIGN) float lowest_weight[ANGULAR_STEPS]; |
251 | alignas(ASTCENC_VECALIGN) int32_t weight_span[ANGULAR_STEPS]; |
252 | alignas(ASTCENC_VECALIGN) float error[ANGULAR_STEPS]; |
253 | alignas(ASTCENC_VECALIGN) float cut_low_weight_error[ANGULAR_STEPS]; |
254 | alignas(ASTCENC_VECALIGN) float cut_high_weight_error[ANGULAR_STEPS]; |
255 | |
256 | compute_lowest_and_highest_weight(weight_count, dec_weight_ideal_value, |
257 | max_angular_steps, max_quant_steps, |
258 | angular_offsets, lowest_weight, weight_span, error, |
259 | cut_low_weight_error, cut_high_weight_error); |
260 | |
261 | // For each quantization level, find the best error terms. Use packed vectors so data-dependent |
262 | // branches can become selects. This involves some integer to float casts, but the values are |
263 | // small enough so they never round the wrong way. |
264 | vfloat4 best_results[36]; |
265 | |
266 | // Initialize the array to some safe defaults |
267 | promise(max_quant_steps > 0); |
268 | for (unsigned int i = 0; i < (max_quant_steps + 4); i++) |
269 | { |
270 | // Lane<0> = Best error |
271 | // Lane<1> = Best scale; -1 indicates no solution found |
272 | // Lane<2> = Cut low weight |
273 | best_results[i] = vfloat4(ERROR_CALC_DEFAULT, -1.0f, 0.0f, 0.0f); |
274 | } |
275 | |
276 | promise(max_angular_steps > 0); |
277 | for (unsigned int i = 0; i < max_angular_steps; i++) |
278 | { |
279 | float i_flt = static_cast<float>(i); |
280 | |
281 | int idx_span = weight_span[i]; |
282 | |
283 | float error_cut_low = error[i] + cut_low_weight_error[i]; |
284 | float error_cut_high = error[i] + cut_high_weight_error[i]; |
285 | float error_cut_low_high = error[i] + cut_low_weight_error[i] + cut_high_weight_error[i]; |
286 | |
287 | // Check best error against record N |
288 | vfloat4 best_result = best_results[idx_span]; |
289 | vfloat4 new_result = vfloat4(error[i], i_flt, 0.0f, 0.0f); |
290 | vmask4 mask = vfloat4(best_result.lane<0>()) > vfloat4(error[i]); |
291 | best_results[idx_span] = select(best_result, new_result, mask); |
292 | |
293 | // Check best error against record N-1 with either cut low or cut high |
294 | best_result = best_results[idx_span - 1]; |
295 | |
296 | new_result = vfloat4(error_cut_low, i_flt, 1.0f, 0.0f); |
297 | mask = vfloat4(best_result.lane<0>()) > vfloat4(error_cut_low); |
298 | best_result = select(best_result, new_result, mask); |
299 | |
300 | new_result = vfloat4(error_cut_high, i_flt, 0.0f, 0.0f); |
301 | mask = vfloat4(best_result.lane<0>()) > vfloat4(error_cut_high); |
302 | best_results[idx_span - 1] = select(best_result, new_result, mask); |
303 | |
304 | // Check best error against record N-2 with both cut low and high |
305 | best_result = best_results[idx_span - 2]; |
306 | new_result = vfloat4(error_cut_low_high, i_flt, 1.0f, 0.0f); |
307 | mask = vfloat4(best_result.lane<0>()) > vfloat4(error_cut_low_high); |
308 | best_results[idx_span - 2] = select(best_result, new_result, mask); |
309 | } |
310 | |
311 | for (unsigned int i = 0; i <= max_quant_level; i++) |
312 | { |
313 | unsigned int q = steps_for_quant_level[i]; |
314 | int bsi = static_cast<int>(best_results[q].lane<1>()); |
315 | |
316 | // Did we find anything? |
317 | #if defined(ASTCENC_DIAGNOSTICS) |
318 | if ((bsi < 0) && print_once) |
319 | { |
320 | print_once = false; |
321 | printf("INFO: Unable to find full encoding within search error limit.\n\n" ); |
322 | } |
323 | #endif |
324 | |
325 | bsi = astc::max(0, bsi); |
326 | |
327 | float lwi = lowest_weight[bsi] + best_results[q].lane<2>(); |
328 | float hwi = lwi + static_cast<float>(q) - 1.0f; |
329 | |
330 | float stepsize = 1.0f / (1.0f + static_cast<float>(bsi)); |
331 | low_value[i] = (angular_offsets[bsi] + lwi) * stepsize; |
332 | high_value[i] = (angular_offsets[bsi] + hwi) * stepsize; |
333 | } |
334 | } |
335 | |
336 | /* See header for documentation. */ |
337 | void compute_angular_endpoints_1plane( |
338 | bool only_always, |
339 | const block_size_descriptor& bsd, |
340 | const float* dec_weight_ideal_value, |
341 | unsigned int max_weight_quant, |
342 | compression_working_buffers& tmpbuf |
343 | ) { |
344 | float (&low_value)[WEIGHTS_MAX_BLOCK_MODES] = tmpbuf.weight_low_value1; |
345 | float (&high_value)[WEIGHTS_MAX_BLOCK_MODES] = tmpbuf.weight_high_value1; |
346 | |
347 | float (&low_values)[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1] = tmpbuf.weight_low_values1; |
348 | float (&high_values)[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1] = tmpbuf.weight_high_values1; |
349 | |
350 | unsigned int max_decimation_modes = only_always ? bsd.decimation_mode_count_always |
351 | : bsd.decimation_mode_count_selected; |
352 | promise(max_decimation_modes > 0); |
353 | for (unsigned int i = 0; i < max_decimation_modes; i++) |
354 | { |
355 | const decimation_mode& dm = bsd.decimation_modes[i]; |
356 | if (!dm.is_ref_1plane(static_cast<quant_method>(max_weight_quant))) |
357 | { |
358 | continue; |
359 | } |
360 | |
361 | unsigned int weight_count = bsd.get_decimation_info(i).weight_count; |
362 | |
363 | unsigned int max_precision = dm.maxprec_1plane; |
364 | if (max_precision > TUNE_MAX_ANGULAR_QUANT) |
365 | { |
366 | max_precision = TUNE_MAX_ANGULAR_QUANT; |
367 | } |
368 | |
369 | if (max_precision > max_weight_quant) |
370 | { |
371 | max_precision = max_weight_quant; |
372 | } |
373 | |
374 | compute_angular_endpoints_for_quant_levels( |
375 | weight_count, |
376 | dec_weight_ideal_value + i * BLOCK_MAX_WEIGHTS, |
377 | max_precision, low_values[i], high_values[i]); |
378 | } |
379 | |
380 | unsigned int max_block_modes = only_always ? bsd.block_mode_count_1plane_always |
381 | : bsd.block_mode_count_1plane_selected; |
382 | promise(max_block_modes > 0); |
383 | for (unsigned int i = 0; i < max_block_modes; i++) |
384 | { |
385 | const block_mode& bm = bsd.block_modes[i]; |
386 | assert(!bm.is_dual_plane); |
387 | |
388 | unsigned int quant_mode = bm.quant_mode; |
389 | unsigned int decim_mode = bm.decimation_mode; |
390 | |
391 | if (quant_mode <= TUNE_MAX_ANGULAR_QUANT) |
392 | { |
393 | low_value[i] = low_values[decim_mode][quant_mode]; |
394 | high_value[i] = high_values[decim_mode][quant_mode]; |
395 | } |
396 | else |
397 | { |
398 | low_value[i] = 0.0f; |
399 | high_value[i] = 1.0f; |
400 | } |
401 | } |
402 | } |
403 | |
404 | /* See header for documentation. */ |
405 | void compute_angular_endpoints_2planes( |
406 | const block_size_descriptor& bsd, |
407 | const float* dec_weight_ideal_value, |
408 | unsigned int max_weight_quant, |
409 | compression_working_buffers& tmpbuf |
410 | ) { |
411 | float (&low_value1)[WEIGHTS_MAX_BLOCK_MODES] = tmpbuf.weight_low_value1; |
412 | float (&high_value1)[WEIGHTS_MAX_BLOCK_MODES] = tmpbuf.weight_high_value1; |
413 | float (&low_value2)[WEIGHTS_MAX_BLOCK_MODES] = tmpbuf.weight_low_value2; |
414 | float (&high_value2)[WEIGHTS_MAX_BLOCK_MODES] = tmpbuf.weight_high_value2; |
415 | |
416 | float (&low_values1)[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1] = tmpbuf.weight_low_values1; |
417 | float (&high_values1)[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1] = tmpbuf.weight_high_values1; |
418 | float (&low_values2)[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1] = tmpbuf.weight_low_values2; |
419 | float (&high_values2)[WEIGHTS_MAX_DECIMATION_MODES][TUNE_MAX_ANGULAR_QUANT + 1] = tmpbuf.weight_high_values2; |
420 | |
421 | promise(bsd.decimation_mode_count_selected > 0); |
422 | for (unsigned int i = 0; i < bsd.decimation_mode_count_selected; i++) |
423 | { |
424 | const decimation_mode& dm = bsd.decimation_modes[i]; |
425 | if (!dm.is_ref_2plane(static_cast<quant_method>(max_weight_quant))) |
426 | { |
427 | continue; |
428 | } |
429 | |
430 | unsigned int weight_count = bsd.get_decimation_info(i).weight_count; |
431 | |
432 | unsigned int max_precision = dm.maxprec_2planes; |
433 | if (max_precision > TUNE_MAX_ANGULAR_QUANT) |
434 | { |
435 | max_precision = TUNE_MAX_ANGULAR_QUANT; |
436 | } |
437 | |
438 | if (max_precision > max_weight_quant) |
439 | { |
440 | max_precision = max_weight_quant; |
441 | } |
442 | |
443 | compute_angular_endpoints_for_quant_levels( |
444 | weight_count, |
445 | dec_weight_ideal_value + i * BLOCK_MAX_WEIGHTS, |
446 | max_precision, low_values1[i], high_values1[i]); |
447 | |
448 | compute_angular_endpoints_for_quant_levels( |
449 | weight_count, |
450 | dec_weight_ideal_value + i * BLOCK_MAX_WEIGHTS + WEIGHTS_PLANE2_OFFSET, |
451 | max_precision, low_values2[i], high_values2[i]); |
452 | } |
453 | |
454 | unsigned int start = bsd.block_mode_count_1plane_selected; |
455 | unsigned int end = bsd.block_mode_count_1plane_2plane_selected; |
456 | for (unsigned int i = start; i < end; i++) |
457 | { |
458 | const block_mode& bm = bsd.block_modes[i]; |
459 | unsigned int quant_mode = bm.quant_mode; |
460 | unsigned int decim_mode = bm.decimation_mode; |
461 | |
462 | if (quant_mode <= TUNE_MAX_ANGULAR_QUANT) |
463 | { |
464 | low_value1[i] = low_values1[decim_mode][quant_mode]; |
465 | high_value1[i] = high_values1[decim_mode][quant_mode]; |
466 | low_value2[i] = low_values2[decim_mode][quant_mode]; |
467 | high_value2[i] = high_values2[decim_mode][quant_mode]; |
468 | } |
469 | else |
470 | { |
471 | low_value1[i] = 0.0f; |
472 | high_value1[i] = 1.0f; |
473 | low_value2[i] = 0.0f; |
474 | high_value2[i] = 1.0f; |
475 | } |
476 | } |
477 | } |
478 | |
479 | #endif |
480 | |