1 // SPDX-License-Identifier: Apache-2.0
2 // ----------------------------------------------------------------------------
3 // Copyright 2011-2022 Arm Limited
4 //
5 // Licensed under the Apache License, Version 2.0 (the "License"); you may not
6 // use this file except in compliance with the License. You may obtain a copy
7 // of the License at:
8 //
9 //     http://www.apache.org/licenses/LICENSE-2.0
10 //
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
13 // WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
14 // License for the specific language governing permissions and limitations
15 // under the License.
16 // ----------------------------------------------------------------------------
17 
18 #if !defined(ASTCENC_DECOMPRESS_ONLY)
19 
20 /**
21  * @brief Functions for finding best endpoint format.
22  *
23  * We assume there are two independent sources of error in any given partition:
24  *
25  *   - Encoding choice errors
26  *   - Quantization errors
27  *
28  * Encoding choice errors are caused by encoder decisions. For example:
29  *
30  *   - Using luminance instead of separate RGB components.
31  *   - Using a constant 1.0 alpha instead of storing an alpha component.
32  *   - Using RGB+scale instead of storing two full RGB endpoints.
33  *
34  * Quantization errors occur due to the limited precision we use for storage. These errors generally
35  * scale with quantization level, but are not actually independent of color encoding. In particular:
36  *
37  *   - If we can use offset encoding then quantization error is halved.
38  *   - If we can use blue-contraction then quantization error for RG is halved.
39  *   - If we use HDR endpoints the quantization error is higher.
40  *
41  * Apart from these effects, we assume the error is proportional to the quantization step size.
42  */
43 
44 
45 #include "astcenc_internal.h"
46 #include "astcenc_vecmathlib.h"
47 
48 #include <assert.h>
49 
50 /**
51  * @brief Compute the errors of the endpoint line options for one partition.
52  *
53  * Uncorrelated data assumes storing completely independent RGBA channels for each endpoint. Same
54  * chroma data assumes storing RGBA endpoints which pass though the origin (LDR only). RGBL data
55  * assumes storing RGB + lumashift (HDR only). Luminance error assumes storing RGB channels as a
56  * single value.
57  *
58  *
59  * @param      pi                The partition info data.
60  * @param      partition_index   The partition index to compule the error for.
61  * @param      blk               The image block.
62  * @param      uncor_pline       The endpoint line assuming uncorrelated endpoints.
63  * @param[out] uncor_err         The computed error for the uncorrelated endpoint line.
64  * @param      samec_pline       The endpoint line assuming the same chroma for both endpoints.
65  * @param[out] samec_err         The computed error for the uncorrelated endpoint line.
66  * @param      rgbl_pline        The endpoint line assuming RGB + lumashift data.
67  * @param[out] rgbl_err          The computed error for the RGB + lumashift endpoint line.
68  * @param      l_pline           The endpoint line assuming luminance data.
69  * @param[out] l_err             The computed error for the luminance endpoint line.
70  * @param[out] a_drop_err        The computed error for dropping the alpha component.
71  */
compute_error_squared_rgb_single_partition( const partition_info& pi, int partition_index, const image_block& blk, const processed_line3& uncor_pline, float& uncor_err, const processed_line3& samec_pline, float& samec_err, const processed_line3& rgbl_pline, float& rgbl_err, const processed_line3& l_pline, float& l_err, float& a_drop_err )72 static void compute_error_squared_rgb_single_partition(
73 	const partition_info& pi,
74 	int partition_index,
75 	const image_block& blk,
76 	const processed_line3& uncor_pline,
77 	float& uncor_err,
78 	const processed_line3& samec_pline,
79 	float& samec_err,
80 	const processed_line3& rgbl_pline,
81 	float& rgbl_err,
82 	const processed_line3& l_pline,
83 	float& l_err,
84 	float& a_drop_err
85 ) {
86 	vfloat4 ews = blk.channel_weight;
87 
88 	unsigned int texel_count = pi.partition_texel_count[partition_index];
89 	const uint8_t* texel_indexes = pi.texels_of_partition[partition_index];
90 	promise(texel_count > 0);
91 
92 	vfloatacc a_drop_errv = vfloatacc::zero();
93 	vfloat default_a(blk.get_default_alpha());
94 
95 	vfloatacc uncor_errv = vfloatacc::zero();
96 	vfloat uncor_bs0(uncor_pline.bs.lane<0>());
97 	vfloat uncor_bs1(uncor_pline.bs.lane<1>());
98 	vfloat uncor_bs2(uncor_pline.bs.lane<2>());
99 
100 	vfloat uncor_amod0(uncor_pline.amod.lane<0>());
101 	vfloat uncor_amod1(uncor_pline.amod.lane<1>());
102 	vfloat uncor_amod2(uncor_pline.amod.lane<2>());
103 
104 	vfloatacc samec_errv = vfloatacc::zero();
105 	vfloat samec_bs0(samec_pline.bs.lane<0>());
106 	vfloat samec_bs1(samec_pline.bs.lane<1>());
107 	vfloat samec_bs2(samec_pline.bs.lane<2>());
108 
109 	vfloatacc rgbl_errv = vfloatacc::zero();
110 	vfloat rgbl_bs0(rgbl_pline.bs.lane<0>());
111 	vfloat rgbl_bs1(rgbl_pline.bs.lane<1>());
112 	vfloat rgbl_bs2(rgbl_pline.bs.lane<2>());
113 
114 	vfloat rgbl_amod0(rgbl_pline.amod.lane<0>());
115 	vfloat rgbl_amod1(rgbl_pline.amod.lane<1>());
116 	vfloat rgbl_amod2(rgbl_pline.amod.lane<2>());
117 
118 	vfloatacc l_errv = vfloatacc::zero();
119 	vfloat l_bs0(l_pline.bs.lane<0>());
120 	vfloat l_bs1(l_pline.bs.lane<1>());
121 	vfloat l_bs2(l_pline.bs.lane<2>());
122 
123 	vfloat one_third(1/3.0f, 1/3.0f, 1/3.0f, 1/3.0f);
124 	vfloat uncor_errv0 = vfloat::zero();
125 	vfloat uncor_errv1 = vfloat::zero();
126 	vfloat uncor_errv2 = vfloat::zero();
127 	vfloat samec_errv0 = vfloat::zero();
128 	vfloat samec_errv1 = vfloat::zero();
129 	vfloat samec_errv2 = vfloat::zero();
130 	vfloat rgbl_errv0 = vfloat::zero();
131 	vfloat rgbl_errv1 = vfloat::zero();
132 	vfloat rgbl_errv2 = vfloat::zero();
133 	vfloat l_errv0 = vfloat::zero();
134 	vfloat l_errv1 = vfloat::zero();
135 	vfloat l_errv2 = vfloat::zero();
136 
137 	unsigned int i = 0;
138 	for (; i + ASTCENC_SIMD_WIDTH <= texel_count; i += ASTCENC_SIMD_WIDTH)
139 	{
140 #ifdef ASTCENC_USE_COMMON_GATHERF
141 		const uint8_t* tix = texel_indexes + i;
142 #else
143 		vint tix(texel_indexes + i);
144 #endif
145 
146 		// Compute the error that arises from just ditching alpha
147 		vfloat data_a = gatherf(blk.data_a, tix);
148 		vfloat alpha_diff = data_a - default_a;
149 		alpha_diff = alpha_diff * alpha_diff;
150 
151 		haccumulate(a_drop_errv, alpha_diff);
152 
153 		vfloat data_r = gatherf(blk.data_r, tix);
154 		vfloat data_g = gatherf(blk.data_g, tix);
155 		vfloat data_b = gatherf(blk.data_b, tix);
156 
157 		vfloat data_rgb_avg = (data_r + data_g + data_b) * one_third;
158 		vfloat data_rgb_0 = data_rgb_avg - data_r;
159 		vfloat data_rgb_1 = data_rgb_avg - data_g;
160 		vfloat data_rgb_2 = data_rgb_avg - data_b;
161 
162 		// Compute uncorrelated error
163 		vfloat param = data_r * uncor_bs0
164 		             + data_g * uncor_bs1
165 		             + data_b * uncor_bs2;
166 
167 		vfloat dist0 = (uncor_amod0 + param * uncor_bs0) - data_r;
168 		vfloat dist1 = (uncor_amod1 + param * uncor_bs1) - data_g;
169 		vfloat dist2 = (uncor_amod2 + param * uncor_bs2) - data_b;
170 
171 		haccumulate(uncor_errv0, dist0 * dist0);
172 		haccumulate(uncor_errv1, dist1 * dist1);
173 		haccumulate(uncor_errv2, dist2 * dist2);
174 
175 		// Compute same chroma error - no "amod", its always zero
176 		param = data_r * samec_bs0
177 		      + data_g * samec_bs1
178 		      + data_b * samec_bs2;
179 
180 		dist0 = (param * samec_bs0) - data_r;
181 		dist1 = (param * samec_bs1) - data_g;
182 		dist2 = (param * samec_bs2) - data_b;
183 
184 		haccumulate(uncor_errv0, dist0 * dist0);
185 		haccumulate(uncor_errv1, dist1 * dist1);
186 		haccumulate(uncor_errv2, dist2 * dist2);
187 
188 		// Compute rgbl error
189 		dist0 = rgbl_amod0 + data_rgb_0;
190 		dist1 = rgbl_amod1 + data_rgb_1;
191 		dist2 = rgbl_amod2 + data_rgb_2;
192 
193 		haccumulate(rgbl_errv0, dist0 * dist0);
194 		haccumulate(rgbl_errv1, dist1 * dist1);
195 		haccumulate(rgbl_errv2, dist2 * dist2);
196 
197 		// Compute luma error - no "amod", its always zero
198 		dist0 = data_rgb_0;
199 		dist1 = data_rgb_1;
200 		dist2 = data_rgb_2;
201 
202 		haccumulate(l_errv0, dist0 * dist0);
203 		haccumulate(l_errv1, dist1 * dist1);
204 		haccumulate(l_errv2, dist2 * dist2);
205 	}
206 
207 	uncor_errv = uncor_errv0 * ews.lane<0>() + uncor_errv1 * ews.lane<1>() + uncor_errv2 * ews.lane<2>(); // channel 0,1,2
208 	samec_errv = samec_errv0 * ews.lane<0>() + samec_errv1 * ews.lane<1>() + samec_errv2 * ews.lane<2>(); // channel 0,1,2
209 	rgbl_errv = rgbl_errv0 * ews.lane<0>() + rgbl_errv1 * ews.lane<1>() + rgbl_errv2 * ews.lane<2>(); // channel 0,1,2
210 	l_errv = l_errv0 * ews.lane<0>() + l_errv1 * ews.lane<1>() + l_errv2 * ews.lane<2>(); // channel 0,1,2
211 
212 	if (i < texel_count)
213 	{
214 		vint lane_ids = vint::lane_id() + i;
215 		vint tix(texel_indexes + i);
216 
217 		vmask mask = lane_ids < vint(texel_count);
218 		lane_ids += vint(ASTCENC_SIMD_WIDTH);
219 
220 		// Compute the error that arises from just ditching alpha
221 		vfloat data_a = gatherf(blk.data_a, tix);
222 		vfloat alpha_diff = data_a - default_a;
223 		alpha_diff = alpha_diff * alpha_diff;
224 
225 		haccumulate(a_drop_errv, alpha_diff, mask);
226 
227 		vfloat data_r = gatherf(blk.data_r, tix);
228 		vfloat data_g = gatherf(blk.data_g, tix);
229 		vfloat data_b = gatherf(blk.data_b, tix);
230 
231 		vfloat data_rgb_avg = (data_r + data_g + data_b) * one_third;
232 		vfloat data_rgb_0 = data_rgb_avg - data_r;
233 		vfloat data_rgb_1 = data_rgb_avg - data_g;
234 		vfloat data_rgb_2 = data_rgb_avg - data_b;
235 
236 		// Compute uncorrelated error
237 		vfloat param = data_r * uncor_bs0
238 		             + data_g * uncor_bs1
239 		             + data_b * uncor_bs2;
240 
241 		vfloat dist0 = (uncor_amod0 + param * uncor_bs0) - data_r;
242 		vfloat dist1 = (uncor_amod1 + param * uncor_bs1) - data_g;
243 		vfloat dist2 = (uncor_amod2 + param * uncor_bs2) - data_b;
244 
245 		vfloat error = dist0 * dist0 * ews.lane<0>()
246 		             + dist1 * dist1 * ews.lane<1>()
247 		             + dist2 * dist2 * ews.lane<2>();
248 
249 		haccumulate(uncor_errv, error, mask);
250 
251 		// Compute same chroma error - no "amod", its always zero
252 		param = data_r * samec_bs0
253 		      + data_g * samec_bs1
254 		      + data_b * samec_bs2;
255 
256 		dist0 = (param * samec_bs0) - data_r;
257 		dist1 = (param * samec_bs1) - data_g;
258 		dist2 = (param * samec_bs2) - data_b;
259 
260 		error = dist0 * dist0 * ews.lane<0>()
261 		      + dist1 * dist1 * ews.lane<1>()
262 		      + dist2 * dist2 * ews.lane<2>();
263 
264 		haccumulate(samec_errv, error, mask);
265 
266 		// Compute rgbl error
267 		dist0 = rgbl_amod0 + data_rgb_0;
268 		dist1 = rgbl_amod1 + data_rgb_1;
269 		dist2 = rgbl_amod2 + data_rgb_2;
270 
271 		error = dist0 * dist0 * ews.lane<0>()
272 		      + dist1 * dist1 * ews.lane<1>()
273 		      + dist2 * dist2 * ews.lane<2>();
274 
275 		haccumulate(rgbl_errv, error, mask);
276 
277 		// Compute luma error - no "amod", its always zero
278 		dist0 = data_rgb_0;
279 		dist1 = data_rgb_1;
280 		dist2 = data_rgb_2;
281 
282 		error = dist0 * dist0 * ews.lane<0>()
283 		      + dist1 * dist1 * ews.lane<1>()
284 		      + dist2 * dist2 * ews.lane<2>();
285 
286 		haccumulate(l_errv, error, mask);
287 	}
288 
289 	a_drop_err = hadd_s(a_drop_errv) * ews.lane<3>();
290 	uncor_err = hadd_s(uncor_errv);
291 	samec_err = hadd_s(samec_errv);
292 	rgbl_err = hadd_s(rgbl_errv);
293 	l_err = hadd_s(l_errv);
294 }
295 
296 /**
297  * @brief For a given set of input colors and partitioning determine endpoint encode errors.
298  *
299  * This function determines the color error that results from RGB-scale encoding (LDR only),
300  * RGB-lumashift encoding (HDR only), luminance-encoding, and alpha drop. Also determines whether
301  * the endpoints are eligible for offset encoding or blue-contraction
302  *
303  * @param      blk   The image block.
304  * @param      pi    The partition info data.
305  * @param      ep    The idealized endpoints.
306  * @param[out] eci   The resulting encoding choice error metrics.
307   */
308 static void compute_encoding_choice_errors(
309 	QualityProfile privateProfile,
310 	const image_block& blk,
311 	const partition_info& pi,
312 	const endpoints& ep,
313 	encoding_choice_errors eci[BLOCK_MAX_PARTITIONS])
314 {
315 	int partition_count = pi.partition_count;
316 	promise(partition_count > 0);
317 
318 	partition_metrics *pms = reinterpret_cast<partition_metrics *>(&blk.pms[0]);
319 
320 	if (!blk.is_constant_channel(3) || (partition_count != 1 && privateProfile == HIGH_QUALITY_PROFILE))
321 	{
322 		compute_avgs_and_dirs_3_comp_rgb(pi, blk, pms);
323 	}
324 
325 	for (int i = 0; i < partition_count; i++)
326 	{
327 		partition_metrics& pm = pms[i];
328 
329 		line3 uncor_rgb_lines;
330 		line3 samec_rgb_lines;  // for LDR-RGB-scale
331 		line3 rgb_luma_lines;   // for HDR-RGB-scale
332 
333 		processed_line3 uncor_rgb_plines;
334 		processed_line3 samec_rgb_plines;
335 		processed_line3 rgb_luma_plines;
336 		processed_line3 luminance_plines;
337 
338 		float uncorr_rgb_error;
339 		float samechroma_rgb_error;
340 		float rgb_luma_error;
341 		float luminance_rgb_error;
342 		float alpha_drop_error;
343 
344 		uncor_rgb_lines.a = pm.avg;
345 		uncor_rgb_lines.b = normalize_safe(pm.dir, unit3());
346 
347 		samec_rgb_lines.a = vfloat4::zero();
348 		samec_rgb_lines.b = normalize_safe(pm.avg, unit3());
349 
350 		rgb_luma_lines.a = pm.avg;
351 		rgb_luma_lines.b = unit3();
352 
353 		uncor_rgb_plines.amod = uncor_rgb_lines.a - uncor_rgb_lines.b * dot3(uncor_rgb_lines.a, uncor_rgb_lines.b);
354 		uncor_rgb_plines.bs   = uncor_rgb_lines.b;
355 
356 		// Same chroma always goes though zero, so this is simpler than the others
357 		samec_rgb_plines.amod = vfloat4::zero();
358 		samec_rgb_plines.bs   = samec_rgb_lines.b;
359 
360 		rgb_luma_plines.amod = rgb_luma_lines.a - rgb_luma_lines.b * dot3(rgb_luma_lines.a, rgb_luma_lines.b);
361 		rgb_luma_plines.bs   = rgb_luma_lines.b;
362 
363 		// Luminance always goes though zero, so this is simpler than the others
364 		luminance_plines.amod = vfloat4::zero();
365 		luminance_plines.bs   = unit3();
366 
367 		compute_error_squared_rgb_single_partition(
368 		    pi, i, blk,
369 		    uncor_rgb_plines, uncorr_rgb_error,
370 		    samec_rgb_plines, samechroma_rgb_error,
371 		    rgb_luma_plines,  rgb_luma_error,
372 		    luminance_plines, luminance_rgb_error,
373 		                      alpha_drop_error);
374 
375 		// Determine if we can offset encode RGB lanes
376 		vfloat4 endpt0 = ep.endpt0[i];
377 		vfloat4 endpt1 = ep.endpt1[i];
378 		vfloat4 endpt_diff = abs(endpt1 - endpt0);
379 		vmask4 endpt_can_offset = endpt_diff < vfloat4(0.12f * 65535.0f);
380 		bool can_offset_encode = (mask(endpt_can_offset) & 0x7) == 0x7;
381 
382 		// Store out the settings
383 		eci[i].rgb_scale_error = (samechroma_rgb_error - uncorr_rgb_error) * 0.7f;  // empirical
384 		eci[i].rgb_luma_error  = (rgb_luma_error - uncorr_rgb_error) * 1.5f;        // wild guess
385 		eci[i].luminance_error = (luminance_rgb_error - uncorr_rgb_error) * 3.0f;   // empirical
386 		eci[i].alpha_drop_error = alpha_drop_error * 3.0f;
387 		eci[i].can_offset_encode = can_offset_encode;
388 		eci[i].can_blue_contract = !blk.is_luminance();
389 	}
390 }
391 
392 /**
393  * @brief For a given partition compute the error for every endpoint integer count and quant level.
394  *
395  * @param      encode_hdr_rgb     @c true if using HDR for RGB, @c false for LDR.
396  * @param      encode_hdr_alpha   @c true if using HDR for alpha, @c false for LDR.
397  * @param      partition_index    The partition index.
398  * @param      pi                 The partition info.
399  * @param      eci                The encoding choice error metrics.
400  * @param      ep                 The idealized endpoints.
401  * @param      error_weight       The resulting encoding choice error metrics.
402  * @param[out] best_error         The best error for each integer count and quant level.
403  * @param[out] format_of_choice   The preferred endpoint format for each integer count and quant level.
404  */
405 static void compute_color_error_for_every_integer_count_and_quant_level(
406 	bool encode_hdr_rgb,
407 	bool encode_hdr_alpha,
408 	int partition_index,
409 	const partition_info& pi,
410 	const encoding_choice_errors& eci,
411 	const endpoints& ep,
412 	vfloat4 error_weight,
413 	float best_error[21][4],
414 	uint8_t format_of_choice[21][4]
415 ) {
416 	int partition_size = pi.partition_texel_count[partition_index];
417 
418 	static const float baseline_quant_error[21 - QUANT_6] {
419 		(65536.0f * 65536.0f / 18.0f) / (5 * 5),
420 		(65536.0f * 65536.0f / 18.0f) / (7 * 7),
421 		(65536.0f * 65536.0f / 18.0f) / (9 * 9),
422 		(65536.0f * 65536.0f / 18.0f) / (11 * 11),
423 		(65536.0f * 65536.0f / 18.0f) / (15 * 15),
424 		(65536.0f * 65536.0f / 18.0f) / (19 * 19),
425 		(65536.0f * 65536.0f / 18.0f) / (23 * 23),
426 		(65536.0f * 65536.0f / 18.0f) / (31 * 31),
427 		(65536.0f * 65536.0f / 18.0f) / (39 * 39),
428 		(65536.0f * 65536.0f / 18.0f) / (47 * 47),
429 		(65536.0f * 65536.0f / 18.0f) / (63 * 63),
430 		(65536.0f * 65536.0f / 18.0f) / (79 * 79),
431 		(65536.0f * 65536.0f / 18.0f) / (95 * 95),
432 		(65536.0f * 65536.0f / 18.0f) / (127 * 127),
433 		(65536.0f * 65536.0f / 18.0f) / (159 * 159),
434 		(65536.0f * 65536.0f / 18.0f) / (191 * 191),
435 		(65536.0f * 65536.0f / 18.0f) / (255 * 255)
436 	};
437 
438 	vfloat4 ep0 = ep.endpt0[partition_index];
439 	vfloat4 ep1 = ep.endpt1[partition_index];
440 
441 	float ep1_min = hmin_rgb_s(ep1);
442 	ep1_min = astc::max(ep1_min, 0.0f);
443 
444 	float error_weight_rgbsum = hadd_rgb_s(error_weight);
445 
446 	float range_upper_limit_rgb = encode_hdr_rgb ? 61440.0f : 65535.0f;
447 	float range_upper_limit_alpha = encode_hdr_alpha ? 61440.0f : 65535.0f;
448 
449 	// It is possible to get endpoint colors significantly outside [0,upper-limit] even if the
450 	// input data are safely contained in [0,upper-limit]; we need to add an error term for this
451 	vfloat4 offset(range_upper_limit_rgb, range_upper_limit_rgb, range_upper_limit_rgb, range_upper_limit_alpha);
452 	vfloat4 ep0_range_error_high = max(ep0 - offset, 0.0f);
453 	vfloat4 ep1_range_error_high = max(ep1 - offset, 0.0f);
454 
455 	vfloat4 ep0_range_error_low = min(ep0, 0.0f);
456 	vfloat4 ep1_range_error_low = min(ep1, 0.0f);
457 
458 	vfloat4 sum_range_error =
459 		(ep0_range_error_low * ep0_range_error_low) +
460 		(ep1_range_error_low * ep1_range_error_low) +
461 		(ep0_range_error_high * ep0_range_error_high) +
462 		(ep1_range_error_high * ep1_range_error_high);
463 
464 	float rgb_range_error = dot3_s(sum_range_error, error_weight)
465 	                      * 0.5f * static_cast<float>(partition_size);
466 	float alpha_range_error = sum_range_error.lane<3>() * error_weight.lane<3>()
467 	                        * 0.5f * static_cast<float>(partition_size);
468 
469 	if (encode_hdr_rgb)
470 	{
471 
472 		// Collect some statistics
473 		float af, cf;
474 		if (ep1.lane<0>() > ep1.lane<1>() && ep1.lane<0>() > ep1.lane<2>())
475 		{
476 			af = ep1.lane<0>();
477 			cf = ep1.lane<0>() - ep0.lane<0>();
478 		}
479 		else if (ep1.lane<1>() > ep1.lane<2>())
480 		{
481 			af = ep1.lane<1>();
482 			cf = ep1.lane<1>() - ep0.lane<1>();
483 		}
484 		else
485 		{
486 			af = ep1.lane<2>();
487 			cf = ep1.lane<2>() - ep0.lane<2>();
488 		}
489 
490 		// Estimate of color-component spread in high endpoint color
491 		float bf = af - ep1_min;
492 		vfloat4 prd = (ep1 - vfloat4(cf)).swz<0, 1, 2>();
493 		vfloat4 pdif = prd - ep0.swz<0, 1, 2>();
494 		// Estimate of color-component spread in low endpoint color
495 		float df = hmax_s(abs(pdif));
496 
497 		int b = static_cast<int>(bf);
498 		int c = static_cast<int>(cf);
499 		int d = static_cast<int>(df);
500 
501 		// Determine which one of the 6 submodes is likely to be used in case of an RGBO-mode
502 		int rgbo_mode = 5;		// 7 bits per component
503 		// mode 4: 8 7 6
504 		if (b < 32768 && c < 16384)
505 		{
506 			rgbo_mode = 4;
507 		}
508 
509 		// mode 3: 9 6 7
510 		if (b < 8192 && c < 16384)
511 		{
512 			rgbo_mode = 3;
513 		}
514 
515 		// mode 2: 10 5 8
516 		if (b < 2048 && c < 16384)
517 		{
518 			rgbo_mode = 2;
519 		}
520 
521 		// mode 1: 11 6 5
522 		if (b < 2048 && c < 1024)
523 		{
524 			rgbo_mode = 1;
525 		}
526 
527 		// mode 0: 11 5 7
528 		if (b < 1024 && c < 4096)
529 		{
530 			rgbo_mode = 0;
531 		}
532 
533 		// Determine which one of the 9 submodes is likely to be used in case of an RGB-mode.
534 		int rgb_mode = 8;		// 8 bits per component, except 7 bits for blue
535 
536 		// mode 0: 9 7 6 7
537 		if (b < 16384 && c < 8192 && d < 8192)
538 		{
539 			rgb_mode = 0;
540 		}
541 
542 		// mode 1: 9 8 6 6
543 		if (b < 32768 && c < 8192 && d < 4096)
544 		{
545 			rgb_mode = 1;
546 		}
547 
548 		// mode 2: 10 6 7 7
549 		if (b < 4096 && c < 8192 && d < 4096)
550 		{
551 			rgb_mode = 2;
552 		}
553 
554 		// mode 3: 10 7 7 6
555 		if (b < 8192 && c < 8192 && d < 2048)
556 		{
557 			rgb_mode = 3;
558 		}
559 
560 		// mode 4: 11 8 6 5
561 		if (b < 8192 && c < 2048 && d < 512)
562 		{
563 			rgb_mode = 4;
564 		}
565 
566 		// mode 5: 11 6 8 6
567 		if (b < 2048 && c < 8192 && d < 1024)
568 		{
569 			rgb_mode = 5;
570 		}
571 
572 		// mode 6: 12 7 7 5
573 		if (b < 2048 && c < 2048 && d < 256)
574 		{
575 			rgb_mode = 6;
576 		}
577 
578 		// mode 7: 12 6 7 6
579 		if (b < 1024 && c < 2048 && d < 512)
580 		{
581 			rgb_mode = 7;
582 		}
583 
584 		static const float rgbo_error_scales[6] { 4.0f, 4.0f, 16.0f, 64.0f, 256.0f, 1024.0f };
585 		static const float rgb_error_scales[9] { 64.0f, 64.0f, 16.0f, 16.0f, 4.0f, 4.0f, 1.0f, 1.0f, 384.0f };
586 
587 		float mode7mult = rgbo_error_scales[rgbo_mode] * 0.0015f;  // Empirically determined ....
588 		float mode11mult = rgb_error_scales[rgb_mode] * 0.010f;    // Empirically determined ....
589 
590 
591 		float lum_high = hadd_rgb_s(ep1) * (1.0f / 3.0f);
592 		float lum_low = hadd_rgb_s(ep0) * (1.0f / 3.0f);
593 		float lumdif = lum_high - lum_low;
594 		float mode23mult = lumdif < 960 ? 4.0f : lumdif < 3968 ? 16.0f : 128.0f;
595 
596 		mode23mult *= 0.0005f;  // Empirically determined ....
597 
598 		// Pick among the available HDR endpoint modes
599 		for (int i = QUANT_2; i < QUANT_16; i++)
600 		{
601 			best_error[i][3] = ERROR_CALC_DEFAULT;
602 			best_error[i][2] = ERROR_CALC_DEFAULT;
603 			best_error[i][1] = ERROR_CALC_DEFAULT;
604 			best_error[i][0] = ERROR_CALC_DEFAULT;
605 
606 			format_of_choice[i][3] = static_cast<uint8_t>(encode_hdr_alpha ? FMT_HDR_RGBA : FMT_HDR_RGB_LDR_ALPHA);
607 			format_of_choice[i][2] = FMT_HDR_RGB;
608 			format_of_choice[i][1] = FMT_HDR_RGB_SCALE;
609 			format_of_choice[i][0] = FMT_HDR_LUMINANCE_LARGE_RANGE;
610 		}
611 
612 		for (int i = QUANT_16; i <= QUANT_256; i++)
613 		{
614 			// The base_quant_error should depend on the scale-factor that would be used during
615 			// actual encode of the color value
616 
617 			float base_quant_error = baseline_quant_error[i - QUANT_6] * static_cast<float>(partition_size);
618 			float rgb_quantization_error = error_weight_rgbsum * base_quant_error * 2.0f;
619 			float alpha_quantization_error = error_weight.lane<3>() * base_quant_error * 2.0f;
620 			float rgba_quantization_error = rgb_quantization_error + alpha_quantization_error;
621 
622 			// For 8 integers, we have two encodings: one with HDR A and another one with LDR A
623 
624 			float full_hdr_rgba_error = rgba_quantization_error + rgb_range_error + alpha_range_error;
625 			best_error[i][3] = full_hdr_rgba_error;
626 			format_of_choice[i][3] = static_cast<uint8_t>(encode_hdr_alpha ? FMT_HDR_RGBA : FMT_HDR_RGB_LDR_ALPHA);
627 
628 			// For 6 integers, we have one HDR-RGB encoding
629 			float full_hdr_rgb_error = (rgb_quantization_error * mode11mult) + rgb_range_error + eci.alpha_drop_error;
630 			best_error[i][2] = full_hdr_rgb_error;
631 			format_of_choice[i][2] = FMT_HDR_RGB;
632 
633 			// For 4 integers, we have one HDR-RGB-Scale encoding
634 			float hdr_rgb_scale_error = (rgb_quantization_error * mode7mult) + rgb_range_error + eci.alpha_drop_error + eci.rgb_luma_error;
635 
636 			best_error[i][1] = hdr_rgb_scale_error;
637 			format_of_choice[i][1] = FMT_HDR_RGB_SCALE;
638 
639 			// For 2 integers, we assume luminance-with-large-range
640 			float hdr_luminance_error = (rgb_quantization_error * mode23mult) + rgb_range_error + eci.alpha_drop_error + eci.luminance_error;
641 			best_error[i][0] = hdr_luminance_error;
642 			format_of_choice[i][0] = FMT_HDR_LUMINANCE_LARGE_RANGE;
643 		}
644 	}
645 	else
646 	{
647 		for (int i = QUANT_2; i < QUANT_6; i++)
648 		{
649 			best_error[i][3] = ERROR_CALC_DEFAULT;
650 			best_error[i][2] = ERROR_CALC_DEFAULT;
651 			best_error[i][1] = ERROR_CALC_DEFAULT;
652 			best_error[i][0] = ERROR_CALC_DEFAULT;
653 
654 			format_of_choice[i][3] = FMT_RGBA;
655 			format_of_choice[i][2] = FMT_RGB;
656 			format_of_choice[i][1] = FMT_RGB_SCALE;
657 			format_of_choice[i][0] = FMT_LUMINANCE;
658 		}
659 
660 		float base_quant_error_rgb = error_weight_rgbsum * static_cast<float>(partition_size);
661 		float base_quant_error_a = error_weight.lane<3>() * static_cast<float>(partition_size);
662 		float base_quant_error_rgba = base_quant_error_rgb + base_quant_error_a;
663 
664 		float error_scale_bc_rgba = eci.can_blue_contract ? 0.625f : 1.0f;
665 		float error_scale_oe_rgba = eci.can_offset_encode ? 0.5f : 1.0f;
666 
667 		float error_scale_bc_rgb = eci.can_blue_contract ? 0.5f : 1.0f;
668 		float error_scale_oe_rgb = eci.can_offset_encode ? 0.25f : 1.0f;
669 
670 		// Pick among the available LDR endpoint modes
671 		for (int i = QUANT_6; i <= QUANT_256; i++)
672 		{
673 			// Offset encoding not possible at higher quant levels
674 			if (i >= QUANT_192)
675 			{
676 				error_scale_oe_rgba = 1.0f;
677 				error_scale_oe_rgb = 1.0f;
678 			}
679 
680 			float base_quant_error = baseline_quant_error[i - QUANT_6];
681 			float quant_error_rgb  = base_quant_error_rgb * base_quant_error;
682 			float quant_error_rgba = base_quant_error_rgba * base_quant_error;
683 
684 			// 8 integers can encode as RGBA+RGBA
685 			float full_ldr_rgba_error = quant_error_rgba
686 			                          * error_scale_bc_rgba
687 			                          * error_scale_oe_rgba
688 			                          + rgb_range_error
689 			                          + alpha_range_error;
690 
691 			best_error[i][3] = full_ldr_rgba_error;
692 			format_of_choice[i][3] = FMT_RGBA;
693 
694 			// 6 integers can encode as RGB+RGB or RGBS+AA
695 			float full_ldr_rgb_error = quant_error_rgb
696 			                         * error_scale_bc_rgb
697 			                         * error_scale_oe_rgb
698 			                         + rgb_range_error
699 			                         + eci.alpha_drop_error;
700 
701 			float rgbs_alpha_error = quant_error_rgba
702 			                       + eci.rgb_scale_error
703 			                       + rgb_range_error
704 			                       + alpha_range_error;
705 
706 			if (rgbs_alpha_error < full_ldr_rgb_error)
707 			{
708 				best_error[i][2] = rgbs_alpha_error;
709 				format_of_choice[i][2] = FMT_RGB_SCALE_ALPHA;
710 			}
711 			else
712 			{
713 				best_error[i][2] = full_ldr_rgb_error;
714 				format_of_choice[i][2] = FMT_RGB;
715 			}
716 
717 			// 4 integers can encode as RGBS or LA+LA
718 			float ldr_rgbs_error = quant_error_rgb
719 			                     + rgb_range_error
720 			                     + eci.alpha_drop_error
721 			                     + eci.rgb_scale_error;
722 
723 			float lum_alpha_error = quant_error_rgba
724 			                      + rgb_range_error
725 			                      + alpha_range_error
726 			                      + eci.luminance_error;
727 
728 			if (ldr_rgbs_error < lum_alpha_error)
729 			{
730 				best_error[i][1] = ldr_rgbs_error;
731 				format_of_choice[i][1] = FMT_RGB_SCALE;
732 			}
733 			else
734 			{
735 				best_error[i][1] = lum_alpha_error;
736 				format_of_choice[i][1] = FMT_LUMINANCE_ALPHA;
737 			}
738 
739 			// 2 integers can encode as L+L
740 			float luminance_error = quant_error_rgb
741 			                      + rgb_range_error
742 			                      + eci.alpha_drop_error
743 			                      + eci.luminance_error;
744 
745 			best_error[i][0] = luminance_error;
746 			format_of_choice[i][0] = FMT_LUMINANCE;
747 		}
748 	}
749 }
750 
751 /**
752  * @brief For one partition compute the best format and quantization for a given bit count.
753  *
754  * @param      best_combined_error    The best error for each quant level and integer count.
755  * @param      best_combined_format   The best format for each quant level and integer count.
756  * @param      bits_available         The number of bits available for encoding.
757  * @param[out] best_quant_level       The output best color quant level.
758  * @param[out] best_format            The output best color format.
759  *
760  * @return The output error for the best pairing.
761  */
762 static float one_partition_find_best_combination_for_bitcount(
763 	QualityProfile privateProfile,
764 	const float best_combined_error[21][4],
765 	const uint8_t best_combined_format[21][4],
766 	int bits_available,
767 	uint8_t& best_quant_level,
768 	uint8_t& best_format
769 ) {
770 	int best_integer_count = 0;
771 	float best_integer_count_error = ERROR_CALC_DEFAULT;
772 
773 	for (int integer_count = 1; integer_count <= 4;  integer_count++)
774 	{
775 		if (privateProfile != HIGH_QUALITY_PROFILE)
776 		{
777 			integer_count = 4; // constant 4 bit count for HIGH_SPEED_PROFILE mode
778 		}
779 		// Compute the quantization level for a given number of integers and a given number of bits
780 		int quant_level = quant_mode_table[integer_count][bits_available];
781 
782 		// Don't have enough bits to represent a given endpoint format at all!
783 		if (quant_level < QUANT_6)
784 		{
785 			continue;
786 		}
787 
788 		float integer_count_error = best_combined_error[quant_level][integer_count - 1];
789 		if (integer_count_error < best_integer_count_error)
790 		{
791 			best_integer_count_error = integer_count_error;
792 			best_integer_count = integer_count - 1;
793 		}
794 	}
795 
796 	int ql = quant_mode_table[best_integer_count + 1][bits_available];
797 
798 	best_quant_level = static_cast<uint8_t>(ql);
799 	if (privateProfile != HIGH_QUALITY_PROFILE) // keep openSource code style
800 	{
801 		best_format = FMT_RGBA;
802 	}
803 	else
804 	{
805 		best_format = FMT_LUMINANCE;
806 
807 		if (ql >= QUANT_6)
808 		{
809 			best_format = best_combined_format[ql][best_integer_count];
810 		}
811 	}
812 
813 	return best_integer_count_error;
814 }
815 
816 /**
817  * @brief For 2 partitions compute the best format combinations for every pair of quant mode and integer count.
818  *
819  * @param      best_error             The best error for a single endpoint quant level and integer count.
820  * @param      best_format            The best format for a single endpoint quant level and integer count.
821  * @param[out] best_combined_error    The best combined error pairings for the 2 partitions.
822  * @param[out] best_combined_format   The best combined format pairings for the 2 partitions.
823  */
824 static void two_partitions_find_best_combination_for_every_quantization_and_integer_count(
825 	const float best_error[2][21][4],	// indexed by (partition, quant-level, integer-pair-count-minus-1)
826 	const uint8_t best_format[2][21][4],
827 	float best_combined_error[21][7],	// indexed by (quant-level, integer-pair-count-minus-2)
828 	uint8_t best_combined_format[21][7][2]
829 ) {
830 	for (int i = QUANT_2; i <= QUANT_256; i++)
831 	{
832 		for (int j = 0; j < 7; j++)
833 		{
834 			best_combined_error[i][j] = ERROR_CALC_DEFAULT;
835 		}
836 	}
837 
838 	for (int quant = QUANT_6; quant <= QUANT_256; quant++)
839 	{
840 		for (int i = 0; i < 4; i++)	// integer-count for first endpoint-pair
841 		{
842 			for (int j = 0; j < 4; j++)	// integer-count for second endpoint-pair
843 			{
844 				int low2 = astc::min(i, j);
845 				int high2 = astc::max(i, j);
846 				if ((high2 - low2) > 1)
847 				{
848 					continue;
849 				}
850 
851 				int intcnt = i + j;
852 				float errorterm = astc::min(best_error[0][quant][i] + best_error[1][quant][j], 1e10f);
853 				if (errorterm <= best_combined_error[quant][intcnt])
854 				{
855 					best_combined_error[quant][intcnt] = errorterm;
856 					best_combined_format[quant][intcnt][0] = best_format[0][quant][i];
857 					best_combined_format[quant][intcnt][1] = best_format[1][quant][j];
858 				}
859 			}
860 		}
861 	}
862 }
863 
864 /**
865  * @brief For 2 partitions compute the best format and quantization for a given bit count.
866  *
867  * @param      best_combined_error    The best error for each quant level and integer count.
868  * @param      best_combined_format   The best format for each quant level and integer count.
869  * @param      bits_available         The number of bits available for encoding.
870  * @param[out] best_quant_level       The output best color quant level.
871  * @param[out] best_quant_level_mod   The output best color quant level assuming two more bits are available.
872  * @param[out] best_formats           The output best color formats.
873  *
874  * @return The output error for the best pairing.
875  */
876 static float two_partitions_find_best_combination_for_bitcount(
877 	unsigned int privateProfile,
878 	float best_combined_error[21][7],
879 	uint8_t best_combined_format[21][7][2],
880 	int bits_available,
881 	uint8_t& best_quant_level,
882 	uint8_t& best_quant_level_mod,
883 	uint8_t* best_formats
884 ) {
885 	int best_integer_count = 0;
886 	float best_integer_count_error = ERROR_CALC_DEFAULT;
887 	int integer_count = 2;
888 	if (privateProfile != HIGH_QUALITY_PROFILE)
889 	{
890 		integer_count = 8;  // constant 8 bit count
891 	}
892 
893 	for (; integer_count <= 8; integer_count++)
894 	{
895 		// Compute the quantization level for a given number of integers and a given number of bits
896 		int quant_level = quant_mode_table[integer_count][bits_available];
897 
898 		// Don't have enough bits to represent a given endpoint format at all!
899 		if (quant_level < QUANT_6)
900 		{
901 			break;
902 		}
903 
904 		float integer_count_error = best_combined_error[quant_level][integer_count - 2];
905 		if (integer_count_error < best_integer_count_error)
906 		{
907 			best_integer_count_error = integer_count_error;
908 			best_integer_count = integer_count;
909 		}
910 	}
911 
912 	int ql = quant_mode_table[best_integer_count][bits_available];
913 	int ql_mod = quant_mode_table[best_integer_count][bits_available + 2];
914 
915 	best_quant_level = static_cast<uint8_t>(ql);
916 	best_quant_level_mod = static_cast<uint8_t>(ql_mod);
917 
918 	if (ql >= QUANT_6)
919 	{
920 		for (int i = 0; i < 2; i++)
921 		{
922 			best_formats[i] = best_combined_format[ql][best_integer_count - 2][i];
923 		}
924 	}
925 	else
926 	{
927 		for (int i = 0; i < 2; i++)
928 		{
929 			best_formats[i] = FMT_LUMINANCE;
930 		}
931 	}
932 
933 	return best_integer_count_error;
934 }
935 
936 /**
937  * @brief For 3 partitions compute the best format combinations for every pair of quant mode and integer count.
938  *
939  * @param      best_error             The best error for a single endpoint quant level and integer count.
940  * @param      best_format            The best format for a single endpoint quant level and integer count.
941  * @param[out] best_combined_error    The best combined error pairings for the 3 partitions.
942  * @param[out] best_combined_format   The best combined format pairings for the 3 partitions.
943  */
944 static void three_partitions_find_best_combination_for_every_quantization_and_integer_count(
945 	const float best_error[3][21][4],	// indexed by (partition, quant-level, integer-count)
946 	const uint8_t best_format[3][21][4],
947 	float best_combined_error[21][10],
948 	uint8_t best_combined_format[21][10][3]
949 ) {
950 	for (int i = QUANT_2; i <= QUANT_256; i++)
951 	{
952 		for (int j = 0; j < 10; j++)
953 		{
954 			best_combined_error[i][j] = ERROR_CALC_DEFAULT;
955 		}
956 	}
957 
958 	for (int quant = QUANT_6; quant <= QUANT_256; quant++)
959 	{
960 		for (int i = 0; i < 4; i++)	// integer-count for first endpoint-pair
961 		{
962 			for (int j = 0; j < 4; j++)	// integer-count for second endpoint-pair
963 			{
964 				int low2 = astc::min(i, j);
965 				int high2 = astc::max(i, j);
966 				if ((high2 - low2) > 1)
967 				{
968 					continue;
969 				}
970 
971 				for (int k = 0; k < 4; k++)	// integer-count for third endpoint-pair
972 				{
973 					int low3 = astc::min(k, low2);
974 					int high3 = astc::max(k, high2);
975 					if ((high3 - low3) > 1)
976 					{
977 						continue;
978 					}
979 
980 					int intcnt = i + j + k;
981 					float errorterm = astc::min(best_error[0][quant][i] + best_error[1][quant][j] + best_error[2][quant][k], 1e10f);
982 					if (errorterm <= best_combined_error[quant][intcnt])
983 					{
984 						best_combined_error[quant][intcnt] = errorterm;
985 						best_combined_format[quant][intcnt][0] = best_format[0][quant][i];
986 						best_combined_format[quant][intcnt][1] = best_format[1][quant][j];
987 						best_combined_format[quant][intcnt][2] = best_format[2][quant][k];
988 					}
989 				}
990 			}
991 		}
992 	}
993 }
994 
995 /**
996  * @brief For 3 partitions compute the best format and quantization for a given bit count.
997  *
998  * @param      best_combined_error    The best error for each quant level and integer count.
999  * @param      best_combined_format   The best format for each quant level and integer count.
1000  * @param      bits_available         The number of bits available for encoding.
1001  * @param[out] best_quant_level       The output best color quant level.
1002  * @param[out] best_quant_level_mod   The output best color quant level assuming two more bits are available.
1003  * @param[out] best_formats           The output best color formats.
1004  *
1005  * @return The output error for the best pairing.
1006  */
1007 static float three_partitions_find_best_combination_for_bitcount(
1008 	const float best_combined_error[21][10],
1009 	const uint8_t best_combined_format[21][10][3],
1010 	int bits_available,
1011 	uint8_t& best_quant_level,
1012 	uint8_t& best_quant_level_mod,
1013 	uint8_t* best_formats
1014 ) {
1015 	int best_integer_count = 0;
1016 	float best_integer_count_error = ERROR_CALC_DEFAULT;
1017 
1018 	for (int integer_count = 3; integer_count <= 9; integer_count++)
1019 	{
1020 		// Compute the quantization level for a given number of integers and a given number of bits
1021 		int quant_level = quant_mode_table[integer_count][bits_available];
1022 
1023 		// Don't have enough bits to represent a given endpoint format at all!
1024 		if (quant_level < QUANT_6)
1025 		{
1026 			break;
1027 		}
1028 
1029 		float integer_count_error = best_combined_error[quant_level][integer_count - 3];
1030 		if (integer_count_error < best_integer_count_error)
1031 		{
1032 			best_integer_count_error = integer_count_error;
1033 			best_integer_count = integer_count;
1034 		}
1035 	}
1036 
1037 	int ql = quant_mode_table[best_integer_count][bits_available];
1038 	int ql_mod = quant_mode_table[best_integer_count][bits_available + 5];
1039 
1040 	best_quant_level = static_cast<uint8_t>(ql);
1041 	best_quant_level_mod = static_cast<uint8_t>(ql_mod);
1042 
1043 	if (ql >= QUANT_6)
1044 	{
1045 		for (int i = 0; i < 3; i++)
1046 		{
1047 			best_formats[i] = best_combined_format[ql][best_integer_count - 3][i];
1048 		}
1049 	}
1050 	else
1051 	{
1052 		for (int i = 0; i < 3; i++)
1053 		{
1054 			best_formats[i] = FMT_LUMINANCE;
1055 		}
1056 	}
1057 
1058 	return best_integer_count_error;
1059 }
1060 
1061 /**
1062  * @brief For 4 partitions compute the best format combinations for every pair of quant mode and integer count.
1063  *
1064  * @param      best_error             The best error for a single endpoint quant level and integer count.
1065  * @param      best_format            The best format for a single endpoint quant level and integer count.
1066  * @param[out] best_combined_error    The best combined error pairings for the 4 partitions.
1067  * @param[out] best_combined_format   The best combined format pairings for the 4 partitions.
1068  */
1069 static void four_partitions_find_best_combination_for_every_quantization_and_integer_count(
1070 	const float best_error[4][21][4],	// indexed by (partition, quant-level, integer-count)
1071 	const uint8_t best_format[4][21][4],
1072 	float best_combined_error[21][13],
1073 	uint8_t best_combined_format[21][13][4]
1074 ) {
1075 	for (int i = QUANT_2; i <= QUANT_256; i++)
1076 	{
1077 		for (int j = 0; j < 13; j++)
1078 		{
1079 			best_combined_error[i][j] = ERROR_CALC_DEFAULT;
1080 		}
1081 	}
1082 
1083 	for (int quant = QUANT_6; quant <= QUANT_256; quant++)
1084 	{
1085 		for (int i = 0; i < 4; i++)	// integer-count for first endpoint-pair
1086 		{
1087 			for (int j = 0; j < 4; j++)	// integer-count for second endpoint-pair
1088 			{
1089 				int low2 = astc::min(i, j);
1090 				int high2 = astc::max(i, j);
1091 				if ((high2 - low2) > 1)
1092 				{
1093 					continue;
1094 				}
1095 
1096 				for (int k = 0; k < 4; k++)	// integer-count for third endpoint-pair
1097 				{
1098 					int low3 = astc::min(k, low2);
1099 					int high3 = astc::max(k, high2);
1100 					if ((high3 - low3) > 1)
1101 					{
1102 						continue;
1103 					}
1104 
1105 					for (int l = 0; l < 4; l++)	// integer-count for fourth endpoint-pair
1106 					{
1107 						int low4 = astc::min(l, low3);
1108 						int high4 = astc::max(l, high3);
1109 						if ((high4 - low4) > 1)
1110 						{
1111 							continue;
1112 						}
1113 
1114 						int intcnt = i + j + k + l;
1115 						float errorterm = astc::min(best_error[0][quant][i] + best_error[1][quant][j] + best_error[2][quant][k] + best_error[3][quant][l], 1e10f);
1116 						if (errorterm <= best_combined_error[quant][intcnt])
1117 						{
1118 							best_combined_error[quant][intcnt] = errorterm;
1119 							best_combined_format[quant][intcnt][0] = best_format[0][quant][i];
1120 							best_combined_format[quant][intcnt][1] = best_format[1][quant][j];
1121 							best_combined_format[quant][intcnt][2] = best_format[2][quant][k];
1122 							best_combined_format[quant][intcnt][3] = best_format[3][quant][l];
1123 						}
1124 					}
1125 				}
1126 			}
1127 		}
1128 	}
1129 }
1130 
1131 /**
1132  * @brief For 4 partitions compute the best format and quantization for a given bit count.
1133  *
1134  * @param      best_combined_error    The best error for each quant level and integer count.
1135  * @param      best_combined_format   The best format for each quant level and integer count.
1136  * @param      bits_available         The number of bits available for encoding.
1137  * @param[out] best_quant_level       The output best color quant level.
1138  * @param[out] best_quant_level_mod   The output best color quant level assuming two more bits are available.
1139  * @param[out] best_formats           The output best color formats.
1140  *
1141  * @return best_error The output error for the best pairing.
1142  */
1143 static float four_partitions_find_best_combination_for_bitcount(
1144 	const float best_combined_error[21][13],
1145 	const uint8_t best_combined_format[21][13][4],
1146 	int bits_available,
1147 	uint8_t& best_quant_level,
1148 	uint8_t& best_quant_level_mod,
1149 	uint8_t* best_formats
1150 ) {
1151 	int best_integer_count = 0;
1152 	float best_integer_count_error = ERROR_CALC_DEFAULT;
1153 
1154 	for (int integer_count = 4; integer_count <= 9; integer_count++)
1155 	{
1156 		// Compute the quantization level for a given number of integers and a given number of bits
1157 		int quant_level = quant_mode_table[integer_count][bits_available];
1158 
1159 		// Don't have enough bits to represent a given endpoint format at all!
1160 		if (quant_level < QUANT_6)
1161 		{
1162 			break;
1163 		}
1164 
1165 		float integer_count_error = best_combined_error[quant_level][integer_count - 4];
1166 		if (integer_count_error < best_integer_count_error)
1167 		{
1168 			best_integer_count_error = integer_count_error;
1169 			best_integer_count = integer_count;
1170 		}
1171 	}
1172 
1173 	int ql = quant_mode_table[best_integer_count][bits_available];
1174 	int ql_mod = quant_mode_table[best_integer_count][bits_available + 8];
1175 
1176 	best_quant_level = static_cast<uint8_t>(ql);
1177 	best_quant_level_mod = static_cast<uint8_t>(ql_mod);
1178 
1179 	if (ql >= QUANT_6)
1180 	{
1181 		for (int i = 0; i < 4; i++)
1182 		{
1183 			best_formats[i] = best_combined_format[ql][best_integer_count - 4][i];
1184 		}
1185 	}
1186 	else
1187 	{
1188 		for (int i = 0; i < 4; i++)
1189 		{
1190 			best_formats[i] = FMT_LUMINANCE;
1191 		}
1192 	}
1193 
1194 	return best_integer_count_error;
1195 }
1196 
1197 /* See header for documentation. */
1198 unsigned int compute_ideal_endpoint_formats(
1199 	QualityProfile privateProfile,
1200 	const partition_info& pi,
1201 	const image_block& blk,
1202 	const endpoints& ep,
1203 	 // bitcounts and errors computed for the various quantization methods
1204 	const int8_t* qwt_bitcounts,
1205 	const float* qwt_errors,
1206 	unsigned int tune_candidate_limit,
1207 	unsigned int start_block_mode,
1208 	unsigned int end_block_mode,
1209 	// output data
1210 	uint8_t partition_format_specifiers[TUNE_MAX_TRIAL_CANDIDATES][BLOCK_MAX_PARTITIONS],
1211 	int block_mode[TUNE_MAX_TRIAL_CANDIDATES],
1212 	quant_method quant_level[TUNE_MAX_TRIAL_CANDIDATES],
1213 	quant_method quant_level_mod[TUNE_MAX_TRIAL_CANDIDATES],
1214 	compression_working_buffers& tmpbuf
1215 ) {
1216 	int partition_count = pi.partition_count;
1217 
1218 	promise(partition_count > 0);
1219 
1220 	bool encode_hdr_rgb = static_cast<bool>(blk.rgb_lns[0]);
1221 	bool encode_hdr_alpha = static_cast<bool>(blk.alpha_lns[0]);
1222 
1223 	// Compute the errors that result from various encoding choices (such as using luminance instead
1224 	// of RGB, discarding Alpha, using RGB-scale in place of two separate RGB endpoints and so on)
1225 	encoding_choice_errors eci[BLOCK_MAX_PARTITIONS];
1226 	compute_encoding_choice_errors(privateProfile, blk, pi, ep, eci);
1227 
1228 	float best_error[BLOCK_MAX_PARTITIONS][21][4];
1229 	uint8_t format_of_choice[BLOCK_MAX_PARTITIONS][21][4];
1230 	for (int i = 0; i < partition_count; i++)
1231 	{
1232 		compute_color_error_for_every_integer_count_and_quant_level(
1233 		    encode_hdr_rgb, encode_hdr_alpha, i,
1234 		    pi, eci[i], ep, blk.channel_weight, best_error[i],
1235 		    format_of_choice[i]);
1236 	}
1237 
1238 	float* errors_of_best_combination = tmpbuf.errors_of_best_combination;
1239 	uint8_t* best_quant_levels = tmpbuf.best_quant_levels;
1240 	uint8_t* best_quant_levels_mod = tmpbuf.best_quant_levels_mod;
1241 	uint8_t (&best_ep_formats)[WEIGHTS_MAX_BLOCK_MODES][BLOCK_MAX_PARTITIONS] = tmpbuf.best_ep_formats;
1242 
1243 	// Ensure that the first iteration understep contains data that will never be picked
1244 	vfloat clear_error(ERROR_CALC_DEFAULT);
1245 	vint clear_quant(0);
1246 
1247 	unsigned int packed_start_block_mode = round_down_to_simd_multiple_vla(start_block_mode);
1248 	storea(clear_error, errors_of_best_combination + packed_start_block_mode);
1249 	store_nbytes(clear_quant, best_quant_levels + packed_start_block_mode);
1250 	store_nbytes(clear_quant, best_quant_levels_mod + packed_start_block_mode);
1251 
1252 	// Ensure that last iteration overstep contains data that will never be picked
1253 	unsigned int packed_end_block_mode = round_down_to_simd_multiple_vla(end_block_mode - 1);
1254 	storea(clear_error, errors_of_best_combination + packed_end_block_mode);
1255 	store_nbytes(clear_quant, best_quant_levels + packed_end_block_mode);
1256 	store_nbytes(clear_quant, best_quant_levels_mod + packed_end_block_mode);
1257 
1258 	// Track a scalar best to avoid expensive search at least once ...
1259 	float error_of_best_combination = ERROR_CALC_DEFAULT;
1260 	int index_of_best_combination = -1;
1261 
1262 	// The block contains 1 partition
1263 	if (partition_count == 1)
1264 	{
1265 		for (unsigned int i = start_block_mode; i < end_block_mode; i++)
1266 		{
1267 			if (qwt_errors[i] >= ERROR_CALC_DEFAULT)
1268 			{
1269 				errors_of_best_combination[i] = ERROR_CALC_DEFAULT;
1270 				continue;
1271 			}
1272 
1273 			float error_of_best = one_partition_find_best_combination_for_bitcount(
1274 			    privateProfile,
1275 			    best_error[0], format_of_choice[0], qwt_bitcounts[i],
1276 			    best_quant_levels[i], best_ep_formats[i][0]);
1277 
1278 			float total_error = error_of_best + qwt_errors[i];
1279 			errors_of_best_combination[i] = total_error;
1280 			best_quant_levels_mod[i] = best_quant_levels[i];
1281 
1282 			if (total_error < error_of_best_combination)
1283 			{
1284 				error_of_best_combination = total_error;
1285 				index_of_best_combination = i;
1286 			}
1287 		}
1288 	}
1289 	// The block contains 2 partitions
1290 	else if (partition_count == 2)
1291 	{
1292 		float combined_best_error[21][7];
1293 		uint8_t formats_of_choice[21][7][2];
1294 
1295 		two_partitions_find_best_combination_for_every_quantization_and_integer_count(
1296 		    best_error, format_of_choice, combined_best_error, formats_of_choice);
1297 
1298 		assert(start_block_mode == 0);
1299 		for (unsigned int i = 0; i < end_block_mode; i++)
1300 		{
1301 			if (qwt_errors[i] >= ERROR_CALC_DEFAULT)
1302 			{
1303 				errors_of_best_combination[i] = ERROR_CALC_DEFAULT;
1304 				continue;
1305 			}
1306 
1307 			float error_of_best = two_partitions_find_best_combination_for_bitcount(
1308 				privateProfile,
1309 				combined_best_error, formats_of_choice, qwt_bitcounts[i],
1310 				best_quant_levels[i], best_quant_levels_mod[i],
1311 				best_ep_formats[i]);
1312 
1313 			float total_error = error_of_best + qwt_errors[i];
1314 			errors_of_best_combination[i] = total_error;
1315 
1316 			if (total_error < error_of_best_combination)
1317 			{
1318 				error_of_best_combination = total_error;
1319 				index_of_best_combination = i;
1320 			}
1321 		}
1322 	}
1323 	// The block contains 3 partitions
1324 	else if (partition_count == 3)
1325 	{
1326 		float combined_best_error[21][10];
1327 		uint8_t formats_of_choice[21][10][3];
1328 
1329 		three_partitions_find_best_combination_for_every_quantization_and_integer_count(
1330 		    best_error, format_of_choice, combined_best_error, formats_of_choice);
1331 
1332 		assert(start_block_mode == 0);
1333 		for (unsigned int i = 0; i < end_block_mode; i++)
1334 		{
1335 			if (qwt_errors[i] >= ERROR_CALC_DEFAULT)
1336 			{
1337 				errors_of_best_combination[i] = ERROR_CALC_DEFAULT;
1338 				continue;
1339 			}
1340 
1341 			float error_of_best = three_partitions_find_best_combination_for_bitcount(
1342 			    combined_best_error, formats_of_choice, qwt_bitcounts[i],
1343 			    best_quant_levels[i], best_quant_levels_mod[i],
1344 			    best_ep_formats[i]);
1345 
1346 			float total_error = error_of_best + qwt_errors[i];
1347 			errors_of_best_combination[i] = total_error;
1348 
1349 			if (total_error < error_of_best_combination)
1350 			{
1351 				error_of_best_combination = total_error;
1352 				index_of_best_combination = i;
1353 			}
1354 		}
1355 	}
1356 	// The block contains 4 partitions
1357 	else // if (partition_count == 4)
1358 	{
1359 		assert(partition_count == 4);
1360 		float combined_best_error[21][13];
1361 		uint8_t formats_of_choice[21][13][4];
1362 
1363 		four_partitions_find_best_combination_for_every_quantization_and_integer_count(
1364 		    best_error, format_of_choice, combined_best_error, formats_of_choice);
1365 
1366 		assert(start_block_mode == 0);
1367 		for (unsigned int i = 0; i < end_block_mode; i++)
1368 		{
1369 			if (qwt_errors[i] >= ERROR_CALC_DEFAULT)
1370 			{
1371 				errors_of_best_combination[i] = ERROR_CALC_DEFAULT;
1372 				continue;
1373 			}
1374 
1375 			float error_of_best = four_partitions_find_best_combination_for_bitcount(
1376 			    combined_best_error, formats_of_choice, qwt_bitcounts[i],
1377 			    best_quant_levels[i], best_quant_levels_mod[i],
1378 			    best_ep_formats[i]);
1379 
1380 			float total_error = error_of_best + qwt_errors[i];
1381 			errors_of_best_combination[i] = total_error;
1382 
1383 			if (total_error < error_of_best_combination)
1384 			{
1385 				error_of_best_combination = total_error;
1386 				index_of_best_combination = i;
1387 			}
1388 		}
1389 	}
1390 
1391 	int best_error_weights[TUNE_MAX_TRIAL_CANDIDATES];
1392 
1393 	// Fast path the first result and avoid the list search for trial 0
1394 	best_error_weights[0] = index_of_best_combination;
1395 	if (index_of_best_combination >= 0)
1396 	{
1397 		errors_of_best_combination[index_of_best_combination] = ERROR_CALC_DEFAULT;
1398 	}
1399 
1400 	// Search the remaining results and pick the best candidate modes for trial 1+
1401 	for (unsigned int i = 1; i < tune_candidate_limit; i++)
1402 	{
1403 		vint vbest_error_index(-1);
1404 		vfloat vbest_ep_error(ERROR_CALC_DEFAULT);
1405 
1406 		start_block_mode = round_down_to_simd_multiple_vla(start_block_mode);
1407 		vint lane_ids = vint::lane_id() + vint(start_block_mode);
1408 		for (unsigned int j = start_block_mode; j < end_block_mode; j += ASTCENC_SIMD_WIDTH)
1409 		{
1410 			vfloat err = vfloat(errors_of_best_combination + j);
1411 			vmask mask = err < vbest_ep_error;
1412 			vbest_ep_error = select(vbest_ep_error, err, mask);
1413 			vbest_error_index = select(vbest_error_index, lane_ids, mask);
1414 			lane_ids += vint(ASTCENC_SIMD_WIDTH);
1415 		}
1416 
1417 		// Pick best mode from the SIMD result, using lowest matching index to ensure invariance
1418 		vmask lanes_min_error = vbest_ep_error == hmin(vbest_ep_error);
1419 		vbest_error_index = select(vint(0x7FFFFFFF), vbest_error_index, lanes_min_error);
1420 		vbest_error_index = hmin(vbest_error_index);
1421 		int best_error_index = vbest_error_index.lane<0>();
1422 
1423 		best_error_weights[i] = best_error_index;
1424 
1425 		// Max the error for this candidate so we don't pick it again
1426 		if (best_error_index >= 0)
1427 		{
1428 			errors_of_best_combination[best_error_index] = ERROR_CALC_DEFAULT;
1429 		}
1430 		// Early-out if no more candidates are valid
1431 		else
1432 		{
1433 			break;
1434 		}
1435 	}
1436 
1437 	for (unsigned int i = 0; i < tune_candidate_limit; i++)
1438 	{
1439 		if (best_error_weights[i] < 0)
1440 		{
1441 			return i;
1442 		}
1443 
1444 		block_mode[i] = best_error_weights[i];
1445 
1446 		quant_level[i] = static_cast<quant_method>(best_quant_levels[best_error_weights[i]]);
1447 		quant_level_mod[i] = static_cast<quant_method>(best_quant_levels_mod[best_error_weights[i]]);
1448 
1449 		assert(quant_level[i] >= QUANT_6 && quant_level[i] <= QUANT_256);
1450 		assert(quant_level_mod[i] >= QUANT_6 && quant_level_mod[i] <= QUANT_256);
1451 
1452 		for (int j = 0; j < partition_count; j++)
1453 		{
1454 			partition_format_specifiers[i][j] = best_ep_formats[best_error_weights[i]][j];
1455 		}
1456 	}
1457 
1458 	return tune_candidate_limit;
1459 }
1460 
1461 #endif
1462