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 /**
19  * @brief Functions for finding dominant direction of a set of colors.
20  */
21 #if !defined(ASTCENC_DECOMPRESS_ONLY)
22 
23 #include "astcenc_internal.h"
24 
25 #include <cassert>
26 
27 /**
28  * @brief Compute the average RGB color of each partition.
29  *
30  * The algorithm here uses a vectorized sequential scan and per-partition
31  * color accumulators, using select() to mask texel lanes in other partitions.
32  *
33  * We only accumulate sums for N-1 partitions during the scan; the value for
34  * the last partition can be computed given that we know the block-wide average
35  * already.
36  *
37  * Because of this we could reduce the loop iteration count so it "just" spans
38  * the max texel index needed for the N-1 partitions, which could need fewer
39  * iterations than the full block texel count. However, this makes the loop
40  * count erratic and causes more branch mispredictions so is a net loss.
41  *
42  * @param      pi         The partitioning to use.
43  * @param      blk        The block data to process.
44  * @param[out] averages   The output averages. Unused partition indices will
45  *                        not be initialized, and lane<3> will be zero.
46  */
compute_partition_averages_rgb(const partition_info & pi,const image_block & blk,vfloat4 averages[BLOCK_MAX_PARTITIONS])47 static void compute_partition_averages_rgb(
48 	const partition_info& pi,
49 	const image_block& blk,
50 	vfloat4 averages[BLOCK_MAX_PARTITIONS]
51 ) {
52 	unsigned int partition_count = pi.partition_count;
53 	unsigned int texel_count = blk.texel_count;
54 	promise(texel_count > 0);
55 
56 	// For 1 partition just use the precomputed mean
57 	if (partition_count == 1)
58 	{
59 		averages[0] = blk.data_mean.swz<0, 1, 2>();
60 	}
61 	// For 2 partitions scan results for partition 0, compute partition 1
62 	else if (partition_count == 2)
63 	{
64 		vfloatacc pp_avg_rgb[3] {};
65 
66 		vint lane_id = vint::lane_id();
67 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
68 		{
69 			vint texel_partition(pi.partition_of_texel + i);
70 
71 			vmask lane_mask = lane_id < vint(texel_count);
72 			lane_id += vint(ASTCENC_SIMD_WIDTH);
73 
74 			vmask p0_mask = lane_mask & (texel_partition == vint(0));
75 
76 			vfloat data_r = loada(blk.data_r + i);
77 			haccumulate(pp_avg_rgb[0], data_r, p0_mask);
78 
79 			vfloat data_g = loada(blk.data_g + i);
80 			haccumulate(pp_avg_rgb[1], data_g, p0_mask);
81 
82 			vfloat data_b = loada(blk.data_b + i);
83 			haccumulate(pp_avg_rgb[2], data_b, p0_mask);
84 		}
85 
86 		vfloat4 block_total = blk.data_mean.swz<0, 1, 2>() * static_cast<float>(blk.texel_count);
87 
88 		vfloat4 p0_total = vfloat3(hadd_s(pp_avg_rgb[0]),
89 		                           hadd_s(pp_avg_rgb[1]),
90 		                           hadd_s(pp_avg_rgb[2]));
91 
92 		vfloat4 p1_total = block_total - p0_total;
93 
94 		averages[0] = p0_total / static_cast<float>(pi.partition_texel_count[0]);
95 		averages[1] = p1_total / static_cast<float>(pi.partition_texel_count[1]);
96 	}
97 	// For 3 partitions scan results for partition 0/1, compute partition 2
98 	else if (partition_count == 3)
99 	{
100 		vfloatacc pp_avg_rgb[2][3] {};
101 
102 		vint lane_id = vint::lane_id();
103 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
104 		{
105 			vint texel_partition(pi.partition_of_texel + i);
106 
107 			vmask lane_mask = lane_id < vint(texel_count);
108 			lane_id += vint(ASTCENC_SIMD_WIDTH);
109 
110 			vmask p0_mask = lane_mask & (texel_partition == vint(0));
111 			vmask p1_mask = lane_mask & (texel_partition == vint(1));
112 
113 			vfloat data_r = loada(blk.data_r + i);
114 			haccumulate(pp_avg_rgb[0][0], data_r, p0_mask);
115 			haccumulate(pp_avg_rgb[1][0], data_r, p1_mask);
116 
117 			vfloat data_g = loada(blk.data_g + i);
118 			haccumulate(pp_avg_rgb[0][1], data_g, p0_mask);
119 			haccumulate(pp_avg_rgb[1][1], data_g, p1_mask);
120 
121 			vfloat data_b = loada(blk.data_b + i);
122 			haccumulate(pp_avg_rgb[0][2], data_b, p0_mask);
123 			haccumulate(pp_avg_rgb[1][2], data_b, p1_mask);
124 		}
125 
126 		vfloat4 block_total = blk.data_mean.swz<0, 1, 2>() * static_cast<float>(blk.texel_count);
127 
128 		vfloat4 p0_total = vfloat3(hadd_s(pp_avg_rgb[0][0]),
129 		                           hadd_s(pp_avg_rgb[0][1]),
130 		                           hadd_s(pp_avg_rgb[0][2]));
131 
132 		vfloat4 p1_total = vfloat3(hadd_s(pp_avg_rgb[1][0]),
133 		                           hadd_s(pp_avg_rgb[1][1]),
134 		                           hadd_s(pp_avg_rgb[1][2]));
135 
136 		vfloat4 p2_total = block_total - p0_total - p1_total;
137 
138 		averages[0] = p0_total / static_cast<float>(pi.partition_texel_count[0]);
139 		averages[1] = p1_total / static_cast<float>(pi.partition_texel_count[1]);
140 		averages[2] = p2_total / static_cast<float>(pi.partition_texel_count[2]);
141 	}
142 	else
143 	{
144 		// For 4 partitions scan results for partition 0/1/2, compute partition 3
145 		vfloatacc pp_avg_rgb[3][3] {};
146 
147 		vint lane_id = vint::lane_id();
148 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
149 		{
150 			vint texel_partition(pi.partition_of_texel + i);
151 
152 			vmask lane_mask = lane_id < vint(texel_count);
153 			lane_id += vint(ASTCENC_SIMD_WIDTH);
154 
155 			vmask p0_mask = lane_mask & (texel_partition == vint(0));
156 			vmask p1_mask = lane_mask & (texel_partition == vint(1));
157 			vmask p2_mask = lane_mask & (texel_partition == vint(2));
158 
159 			vfloat data_r = loada(blk.data_r + i);
160 			haccumulate(pp_avg_rgb[0][0], data_r, p0_mask);
161 			haccumulate(pp_avg_rgb[1][0], data_r, p1_mask);
162 			haccumulate(pp_avg_rgb[2][0], data_r, p2_mask);
163 
164 			vfloat data_g = loada(blk.data_g + i);
165 			haccumulate(pp_avg_rgb[0][1], data_g, p0_mask);
166 			haccumulate(pp_avg_rgb[1][1], data_g, p1_mask);
167 			haccumulate(pp_avg_rgb[2][1], data_g, p2_mask);
168 
169 			vfloat data_b = loada(blk.data_b + i);
170 			haccumulate(pp_avg_rgb[0][2], data_b, p0_mask);
171 			haccumulate(pp_avg_rgb[1][2], data_b, p1_mask);
172 			haccumulate(pp_avg_rgb[2][2], data_b, p2_mask);
173 		}
174 
175 		vfloat4 block_total = blk.data_mean.swz<0, 1, 2>() * static_cast<float>(blk.texel_count);
176 
177 		vfloat4 p0_total = vfloat3(hadd_s(pp_avg_rgb[0][0]),
178 		                           hadd_s(pp_avg_rgb[0][1]),
179 		                           hadd_s(pp_avg_rgb[0][2]));
180 
181 		vfloat4 p1_total = vfloat3(hadd_s(pp_avg_rgb[1][0]),
182 		                           hadd_s(pp_avg_rgb[1][1]),
183 		                           hadd_s(pp_avg_rgb[1][2]));
184 
185 		vfloat4 p2_total = vfloat3(hadd_s(pp_avg_rgb[2][0]),
186 		                           hadd_s(pp_avg_rgb[2][1]),
187 		                           hadd_s(pp_avg_rgb[2][2]));
188 
189 		vfloat4 p3_total = block_total - p0_total - p1_total- p2_total;
190 
191 		averages[0] = p0_total / static_cast<float>(pi.partition_texel_count[0]);
192 		averages[1] = p1_total / static_cast<float>(pi.partition_texel_count[1]);
193 		averages[2] = p2_total / static_cast<float>(pi.partition_texel_count[2]);
194 		averages[3] = p3_total / static_cast<float>(pi.partition_texel_count[3]);
195 	}
196 }
197 
198 /**
199  * @brief Compute the average RGBA color of each partition.
200  *
201  * The algorithm here uses a vectorized sequential scan and per-partition
202  * color accumulators, using select() to mask texel lanes in other partitions.
203  *
204  * We only accumulate sums for N-1 partitions during the scan; the value for
205  * the last partition can be computed given that we know the block-wide average
206  * already.
207  *
208  * Because of this we could reduce the loop iteration count so it "just" spans
209  * the max texel index needed for the N-1 partitions, which could need fewer
210  * iterations than the full block texel count. However, this makes the loop
211  * count erratic and causes more branch mispredictions so is a net loss.
212  *
213  * @param      pi         The partitioning to use.
214  * @param      blk        The block data to process.
215  * @param[out] averages   The output averages. Unused partition indices will
216  *                        not be initialized.
217  */
compute_partition_averages_rgba(const partition_info & pi,const image_block & blk,vfloat4 averages[BLOCK_MAX_PARTITIONS])218 static void compute_partition_averages_rgba(
219 	const partition_info& pi,
220 	const image_block& blk,
221 	vfloat4 averages[BLOCK_MAX_PARTITIONS]
222 ) {
223 	unsigned int partition_count = pi.partition_count;
224 	unsigned int texel_count = blk.texel_count;
225 	promise(texel_count > 0);
226 
227 	// For 1 partition just use the precomputed mean
228 	if (partition_count == 1)
229 	{
230 		averages[0] = blk.data_mean;
231 	}
232 	// For 2 partitions scan results for partition 0, compute partition 1
233 	else if (partition_count == 2)
234 	{
235 		vfloat4 pp_avg_rgba[4] {};
236 
237 		vint lane_id = vint::lane_id();
238 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
239 		{
240 			vint texel_partition(pi.partition_of_texel + i);
241 
242 			vmask lane_mask = lane_id < vint(texel_count);
243 			lane_id += vint(ASTCENC_SIMD_WIDTH);
244 
245 			vmask p0_mask = lane_mask & (texel_partition == vint(0));
246 
247 			vfloat data_r = loada(blk.data_r + i);
248 			haccumulate(pp_avg_rgba[0], data_r, p0_mask);
249 
250 			vfloat data_g = loada(blk.data_g + i);
251 			haccumulate(pp_avg_rgba[1], data_g, p0_mask);
252 
253 			vfloat data_b = loada(blk.data_b + i);
254 			haccumulate(pp_avg_rgba[2], data_b, p0_mask);
255 
256 			vfloat data_a = loada(blk.data_a + i);
257 			haccumulate(pp_avg_rgba[3], data_a, p0_mask);
258 		}
259 
260 		vfloat4 block_total = blk.data_mean * static_cast<float>(blk.texel_count);
261 
262 		vfloat4 p0_total = vfloat4(hadd_s(pp_avg_rgba[0]),
263 		                           hadd_s(pp_avg_rgba[1]),
264 		                           hadd_s(pp_avg_rgba[2]),
265 		                           hadd_s(pp_avg_rgba[3]));
266 
267 		vfloat4 p1_total = block_total - p0_total;
268 
269 		averages[0] = p0_total / static_cast<float>(pi.partition_texel_count[0]);
270 		averages[1] = p1_total / static_cast<float>(pi.partition_texel_count[1]);
271 	}
272 	// For 3 partitions scan results for partition 0/1, compute partition 2
273 	else if (partition_count == 3)
274 	{
275 		vfloat4 pp_avg_rgba[2][4] {};
276 
277 		vint lane_id = vint::lane_id();
278 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
279 		{
280 			vint texel_partition(pi.partition_of_texel + i);
281 
282 			vmask lane_mask = lane_id < vint(texel_count);
283 			lane_id += vint(ASTCENC_SIMD_WIDTH);
284 
285 			vmask p0_mask = lane_mask & (texel_partition == vint(0));
286 			vmask p1_mask = lane_mask & (texel_partition == vint(1));
287 
288 			vfloat data_r = loada(blk.data_r + i);
289 			haccumulate(pp_avg_rgba[0][0], data_r, p0_mask);
290 			haccumulate(pp_avg_rgba[1][0], data_r, p1_mask);
291 
292 			vfloat data_g = loada(blk.data_g + i);
293 			haccumulate(pp_avg_rgba[0][1], data_g, p0_mask);
294 			haccumulate(pp_avg_rgba[1][1], data_g, p1_mask);
295 
296 			vfloat data_b = loada(blk.data_b + i);
297 			haccumulate(pp_avg_rgba[0][2], data_b, p0_mask);
298 			haccumulate(pp_avg_rgba[1][2], data_b, p1_mask);
299 
300 			vfloat data_a = loada(blk.data_a + i);
301 			haccumulate(pp_avg_rgba[0][3], data_a, p0_mask);
302 			haccumulate(pp_avg_rgba[1][3], data_a, p1_mask);
303 		}
304 
305 		vfloat4 block_total = blk.data_mean * static_cast<float>(blk.texel_count);
306 
307 		vfloat4 p0_total = vfloat4(hadd_s(pp_avg_rgba[0][0]),
308 		                           hadd_s(pp_avg_rgba[0][1]),
309 		                           hadd_s(pp_avg_rgba[0][2]),
310 		                           hadd_s(pp_avg_rgba[0][3]));
311 
312 		vfloat4 p1_total = vfloat4(hadd_s(pp_avg_rgba[1][0]),
313 		                           hadd_s(pp_avg_rgba[1][1]),
314 		                           hadd_s(pp_avg_rgba[1][2]),
315 		                           hadd_s(pp_avg_rgba[1][3]));
316 
317 		vfloat4 p2_total = block_total - p0_total - p1_total;
318 
319 		averages[0] = p0_total / static_cast<float>(pi.partition_texel_count[0]);
320 		averages[1] = p1_total / static_cast<float>(pi.partition_texel_count[1]);
321 		averages[2] = p2_total / static_cast<float>(pi.partition_texel_count[2]);
322 	}
323 	else
324 	{
325 		// For 4 partitions scan results for partition 0/1/2, compute partition 3
326 		vfloat4 pp_avg_rgba[3][4] {};
327 
328 		vint lane_id = vint::lane_id();
329 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
330 		{
331 			vint texel_partition(pi.partition_of_texel + i);
332 
333 			vmask lane_mask = lane_id < vint(texel_count);
334 			lane_id += vint(ASTCENC_SIMD_WIDTH);
335 
336 			vmask p0_mask = lane_mask & (texel_partition == vint(0));
337 			vmask p1_mask = lane_mask & (texel_partition == vint(1));
338 			vmask p2_mask = lane_mask & (texel_partition == vint(2));
339 
340 			vfloat data_r = loada(blk.data_r + i);
341 			haccumulate(pp_avg_rgba[0][0], data_r, p0_mask);
342 			haccumulate(pp_avg_rgba[1][0], data_r, p1_mask);
343 			haccumulate(pp_avg_rgba[2][0], data_r, p2_mask);
344 
345 			vfloat data_g = loada(blk.data_g + i);
346 			haccumulate(pp_avg_rgba[0][1], data_g, p0_mask);
347 			haccumulate(pp_avg_rgba[1][1], data_g, p1_mask);
348 			haccumulate(pp_avg_rgba[2][1], data_g, p2_mask);
349 
350 			vfloat data_b = loada(blk.data_b + i);
351 			haccumulate(pp_avg_rgba[0][2], data_b, p0_mask);
352 			haccumulate(pp_avg_rgba[1][2], data_b, p1_mask);
353 			haccumulate(pp_avg_rgba[2][2], data_b, p2_mask);
354 
355 			vfloat data_a = loada(blk.data_a + i);
356 			haccumulate(pp_avg_rgba[0][3], data_a, p0_mask);
357 			haccumulate(pp_avg_rgba[1][3], data_a, p1_mask);
358 			haccumulate(pp_avg_rgba[2][3], data_a, p2_mask);
359 		}
360 
361 		vfloat4 block_total = blk.data_mean * static_cast<float>(blk.texel_count);
362 
363 		vfloat4 p0_total = vfloat4(hadd_s(pp_avg_rgba[0][0]),
364 		                           hadd_s(pp_avg_rgba[0][1]),
365 		                           hadd_s(pp_avg_rgba[0][2]),
366 		                           hadd_s(pp_avg_rgba[0][3]));
367 
368 		vfloat4 p1_total = vfloat4(hadd_s(pp_avg_rgba[1][0]),
369 		                           hadd_s(pp_avg_rgba[1][1]),
370 		                           hadd_s(pp_avg_rgba[1][2]),
371 		                           hadd_s(pp_avg_rgba[1][3]));
372 
373 		vfloat4 p2_total = vfloat4(hadd_s(pp_avg_rgba[2][0]),
374 		                           hadd_s(pp_avg_rgba[2][1]),
375 		                           hadd_s(pp_avg_rgba[2][2]),
376 		                           hadd_s(pp_avg_rgba[2][3]));
377 
378 		vfloat4 p3_total = block_total - p0_total - p1_total- p2_total;
379 
380 		averages[0] = p0_total / static_cast<float>(pi.partition_texel_count[0]);
381 		averages[1] = p1_total / static_cast<float>(pi.partition_texel_count[1]);
382 		averages[2] = p2_total / static_cast<float>(pi.partition_texel_count[2]);
383 		averages[3] = p3_total / static_cast<float>(pi.partition_texel_count[3]);
384 	}
385 }
386 
387 /* See header for documentation. */
compute_avgs_and_dirs_4_comp(const partition_info & pi,const image_block & blk,partition_metrics pm[BLOCK_MAX_PARTITIONS])388 void compute_avgs_and_dirs_4_comp(
389 	const partition_info& pi,
390 	const image_block& blk,
391 	partition_metrics pm[BLOCK_MAX_PARTITIONS]
392 ) {
393 	int partition_count = pi.partition_count;
394 	promise(partition_count > 0);
395 
396 	// Pre-compute partition_averages
397 	vfloat4 partition_averages[BLOCK_MAX_PARTITIONS];
398 	compute_partition_averages_rgba(pi, blk, partition_averages);
399 
400 	for (int partition = 0; partition < partition_count; partition++)
401 	{
402 		const uint8_t *texel_indexes = pi.texels_of_partition[partition];
403 		unsigned int texel_count = pi.partition_texel_count[partition];
404 		promise(texel_count > 0);
405 
406 		vfloat4 average = partition_averages[partition];
407 		pm[partition].avg = average;
408 
409 		vfloat4 sum_xp = vfloat4::zero();
410 		vfloat4 sum_yp = vfloat4::zero();
411 		vfloat4 sum_zp = vfloat4::zero();
412 		vfloat4 sum_wp = vfloat4::zero();
413 
414 		for (unsigned int i = 0; i < texel_count; i++)
415 		{
416 			unsigned int iwt = texel_indexes[i];
417 			vfloat4 texel_datum = blk.texel(iwt);
418 			texel_datum = texel_datum - average;
419 
420 			vfloat4 zero = vfloat4::zero();
421 
422 			vmask4 tdm0 = texel_datum.swz<0,0,0,0>() > zero;
423 			sum_xp += select(zero, texel_datum, tdm0);
424 
425 			vmask4 tdm1 = texel_datum.swz<1,1,1,1>() > zero;
426 			sum_yp += select(zero, texel_datum, tdm1);
427 
428 			vmask4 tdm2 = texel_datum.swz<2,2,2,2>() > zero;
429 			sum_zp += select(zero, texel_datum, tdm2);
430 
431 			vmask4 tdm3 = texel_datum.swz<3,3,3,3>() > zero;
432 			sum_wp += select(zero, texel_datum, tdm3);
433 		}
434 
435 		vfloat4 prod_xp = dot(sum_xp, sum_xp);
436 		vfloat4 prod_yp = dot(sum_yp, sum_yp);
437 		vfloat4 prod_zp = dot(sum_zp, sum_zp);
438 		vfloat4 prod_wp = dot(sum_wp, sum_wp);
439 
440 		vfloat4 best_vector = sum_xp;
441 		vfloat4 best_sum = prod_xp;
442 
443 		vmask4 mask = prod_yp > best_sum;
444 		best_vector = select(best_vector, sum_yp, mask);
445 		best_sum = select(best_sum, prod_yp, mask);
446 
447 		mask = prod_zp > best_sum;
448 		best_vector = select(best_vector, sum_zp, mask);
449 		best_sum = select(best_sum, prod_zp, mask);
450 
451 		mask = prod_wp > best_sum;
452 		best_vector = select(best_vector, sum_wp, mask);
453 
454 		pm[partition].dir = best_vector;
455 	}
456 }
457 
458 /* See header for documentation. */
compute_avgs_and_dirs_3_comp(const partition_info & pi,const image_block & blk,unsigned int omitted_component,partition_metrics pm[BLOCK_MAX_PARTITIONS])459 void compute_avgs_and_dirs_3_comp(
460 	const partition_info& pi,
461 	const image_block& blk,
462 	unsigned int omitted_component,
463 	partition_metrics pm[BLOCK_MAX_PARTITIONS]
464 ) {
465 	// Pre-compute partition_averages
466 	vfloat4 partition_averages[BLOCK_MAX_PARTITIONS];
467 	compute_partition_averages_rgba(pi, blk, partition_averages);
468 
469 	const float* data_vr = blk.data_r;
470 	const float* data_vg = blk.data_g;
471 	const float* data_vb = blk.data_b;
472 
473 	// TODO: Data-driven permute would be useful to avoid this ...
474 	if (omitted_component == 0)
475 	{
476 		partition_averages[0] = partition_averages[0].swz<1, 2, 3>();
477 		partition_averages[1] = partition_averages[1].swz<1, 2, 3>();
478 		partition_averages[2] = partition_averages[2].swz<1, 2, 3>();
479 		partition_averages[3] = partition_averages[3].swz<1, 2, 3>();
480 
481 		data_vr = blk.data_g;
482 		data_vg = blk.data_b;
483 		data_vb = blk.data_a;
484 	}
485 	else if (omitted_component == 1)
486 	{
487 		partition_averages[0] = partition_averages[0].swz<0, 2, 3>();
488 		partition_averages[1] = partition_averages[1].swz<0, 2, 3>();
489 		partition_averages[2] = partition_averages[2].swz<0, 2, 3>();
490 		partition_averages[3] = partition_averages[3].swz<0, 2, 3>();
491 
492 		data_vg = blk.data_b;
493 		data_vb = blk.data_a;
494 	}
495 	else if (omitted_component == 2)
496 	{
497 		partition_averages[0] = partition_averages[0].swz<0, 1, 3>();
498 		partition_averages[1] = partition_averages[1].swz<0, 1, 3>();
499 		partition_averages[2] = partition_averages[2].swz<0, 1, 3>();
500 		partition_averages[3] = partition_averages[3].swz<0, 1, 3>();
501 
502 		data_vb = blk.data_a;
503 	}
504 	else
505 	{
506 		partition_averages[0] = partition_averages[0].swz<0, 1, 2>();
507 		partition_averages[1] = partition_averages[1].swz<0, 1, 2>();
508 		partition_averages[2] = partition_averages[2].swz<0, 1, 2>();
509 		partition_averages[3] = partition_averages[3].swz<0, 1, 2>();
510 	}
511 
512 	unsigned int partition_count = pi.partition_count;
513 	promise(partition_count > 0);
514 
515 	for (unsigned int partition = 0; partition < partition_count; partition++)
516 	{
517 		const uint8_t *texel_indexes = pi.texels_of_partition[partition];
518 		unsigned int texel_count = pi.partition_texel_count[partition];
519 		promise(texel_count > 0);
520 
521 		vfloat4 average = partition_averages[partition];
522 		pm[partition].avg = average;
523 
524 		vfloat4 sum_xp = vfloat4::zero();
525 		vfloat4 sum_yp = vfloat4::zero();
526 		vfloat4 sum_zp = vfloat4::zero();
527 
528 		for (unsigned int i = 0; i < texel_count; i++)
529 		{
530 			unsigned int iwt = texel_indexes[i];
531 
532 			vfloat4 texel_datum = vfloat3(data_vr[iwt],
533 			                              data_vg[iwt],
534 			                              data_vb[iwt]);
535 			texel_datum = texel_datum - average;
536 
537 			vfloat4 zero = vfloat4::zero();
538 
539 			vmask4 tdm0 = texel_datum.swz<0,0,0,0>() > zero;
540 			sum_xp += select(zero, texel_datum, tdm0);
541 
542 			vmask4 tdm1 = texel_datum.swz<1,1,1,1>() > zero;
543 			sum_yp += select(zero, texel_datum, tdm1);
544 
545 			vmask4 tdm2 = texel_datum.swz<2,2,2,2>() > zero;
546 			sum_zp += select(zero, texel_datum, tdm2);
547 		}
548 
549 		vfloat4 prod_xp = dot(sum_xp, sum_xp);
550 		vfloat4 prod_yp = dot(sum_yp, sum_yp);
551 		vfloat4 prod_zp = dot(sum_zp, sum_zp);
552 
553 		vfloat4 best_vector = sum_xp;
554 		vfloat4 best_sum = prod_xp;
555 
556 		vmask4 mask = prod_yp > best_sum;
557 		best_vector = select(best_vector, sum_yp, mask);
558 		best_sum = select(best_sum, prod_yp, mask);
559 
560 		mask = prod_zp > best_sum;
561 		best_vector = select(best_vector, sum_zp, mask);
562 
563 		pm[partition].dir = best_vector;
564 	}
565 }
566 
567 /* See header for documentation. */
compute_avgs_and_dirs_3_comp_rgb(const partition_info & pi,const image_block & blk,partition_metrics pm[BLOCK_MAX_PARTITIONS])568 void compute_avgs_and_dirs_3_comp_rgb(
569 	const partition_info& pi,
570 	const image_block& blk,
571 	partition_metrics pm[BLOCK_MAX_PARTITIONS]
572 ) {
573 	unsigned int partition_count = pi.partition_count;
574 	promise(partition_count > 0);
575 
576 	// Pre-compute partition_averages
577 	vfloat4 partition_averages[BLOCK_MAX_PARTITIONS];
578 	compute_partition_averages_rgb(pi, blk, partition_averages);
579 
580 	for (unsigned int partition = 0; partition < partition_count; partition++)
581 	{
582 		const uint8_t *texel_indexes = pi.texels_of_partition[partition];
583 		unsigned int texel_count = pi.partition_texel_count[partition];
584 		promise(texel_count > 0);
585 
586 		vfloat4 average = partition_averages[partition];
587 		pm[partition].avg = average;
588 
589 		vfloat4 sum_xp = vfloat4::zero();
590 		vfloat4 sum_yp = vfloat4::zero();
591 		vfloat4 sum_zp = vfloat4::zero();
592 
593 		for (unsigned int i = 0; i < texel_count; i++)
594 		{
595 			unsigned int iwt = texel_indexes[i];
596 
597 			vfloat4 texel_datum = blk.texel3(iwt);
598 			texel_datum = texel_datum - average;
599 
600 			vfloat4 zero = vfloat4::zero();
601 
602 			vmask4 tdm0 = texel_datum.swz<0,0,0,0>() > zero;
603 			sum_xp += select(zero, texel_datum, tdm0);
604 
605 			vmask4 tdm1 = texel_datum.swz<1,1,1,1>() > zero;
606 			sum_yp += select(zero, texel_datum, tdm1);
607 
608 			vmask4 tdm2 = texel_datum.swz<2,2,2,2>() > zero;
609 			sum_zp += select(zero, texel_datum, tdm2);
610 		}
611 
612 		vfloat4 prod_xp = dot(sum_xp, sum_xp);
613 		vfloat4 prod_yp = dot(sum_yp, sum_yp);
614 		vfloat4 prod_zp = dot(sum_zp, sum_zp);
615 
616 		vfloat4 best_vector = sum_xp;
617 		vfloat4 best_sum = prod_xp;
618 
619 		vmask4 mask = prod_yp > best_sum;
620 		best_vector = select(best_vector, sum_yp, mask);
621 		best_sum = select(best_sum, prod_yp, mask);
622 
623 		mask = prod_zp > best_sum;
624 		best_vector = select(best_vector, sum_zp, mask);
625 
626 		pm[partition].dir = best_vector;
627 	}
628 }
629 
630 /* See header for documentation. */
compute_avgs_and_dirs_2_comp(const partition_info & pt,const image_block & blk,unsigned int component1,unsigned int component2,partition_metrics pm[BLOCK_MAX_PARTITIONS])631 void compute_avgs_and_dirs_2_comp(
632 	const partition_info& pt,
633 	const image_block& blk,
634 	unsigned int component1,
635 	unsigned int component2,
636 	partition_metrics pm[BLOCK_MAX_PARTITIONS]
637 ) {
638 	vfloat4 average;
639 
640 	const float* data_vr = nullptr;
641 	const float* data_vg = nullptr;
642 
643 	if (component1 == 0 && component2 == 1)
644 	{
645 		average = blk.data_mean.swz<0, 1>();
646 
647 		data_vr = blk.data_r;
648 		data_vg = blk.data_g;
649 	}
650 	else if (component1 == 0 && component2 == 2)
651 	{
652 		average = blk.data_mean.swz<0, 2>();
653 
654 		data_vr = blk.data_r;
655 		data_vg = blk.data_b;
656 	}
657 	else // (component1 == 1 && component2 == 2)
658 	{
659 		assert(component1 == 1 && component2 == 2);
660 
661 		average = blk.data_mean.swz<1, 2>();
662 
663 		data_vr = blk.data_g;
664 		data_vg = blk.data_b;
665 	}
666 
667 	unsigned int partition_count = pt.partition_count;
668 	promise(partition_count > 0);
669 
670 	for (unsigned int partition = 0; partition < partition_count; partition++)
671 	{
672 		const uint8_t *texel_indexes = pt.texels_of_partition[partition];
673 		unsigned int texel_count = pt.partition_texel_count[partition];
674 		promise(texel_count > 0);
675 
676 		// Only compute a partition mean if more than one partition
677 		if (partition_count > 1)
678 		{
679 			average = vfloat4::zero();
680 			for (unsigned int i = 0; i < texel_count; i++)
681 			{
682 				unsigned int iwt = texel_indexes[i];
683 				average += vfloat2(data_vr[iwt], data_vg[iwt]);
684 			}
685 
686 			average = average / static_cast<float>(texel_count);
687 		}
688 
689 		pm[partition].avg = average;
690 
691 		vfloat4 sum_xp = vfloat4::zero();
692 		vfloat4 sum_yp = vfloat4::zero();
693 
694 		for (unsigned int i = 0; i < texel_count; i++)
695 		{
696 			unsigned int iwt = texel_indexes[i];
697 			vfloat4 texel_datum = vfloat2(data_vr[iwt], data_vg[iwt]);
698 			texel_datum = texel_datum - average;
699 
700 			vfloat4 zero = vfloat4::zero();
701 
702 			vmask4 tdm0 = texel_datum.swz<0,0,0,0>() > zero;
703 			sum_xp += select(zero, texel_datum, tdm0);
704 
705 			vmask4 tdm1 = texel_datum.swz<1,1,1,1>() > zero;
706 			sum_yp += select(zero, texel_datum, tdm1);
707 		}
708 
709 		vfloat4 prod_xp = dot(sum_xp, sum_xp);
710 		vfloat4 prod_yp = dot(sum_yp, sum_yp);
711 
712 		vfloat4 best_vector = sum_xp;
713 		vfloat4 best_sum = prod_xp;
714 
715 		vmask4 mask = prod_yp > best_sum;
716 		best_vector = select(best_vector, sum_yp, mask);
717 
718 		pm[partition].dir = best_vector;
719 	}
720 }
721 
722 /* See header for documentation. */
compute_error_squared_rgba(const partition_info & pi,const image_block & blk,const processed_line4 uncor_plines[BLOCK_MAX_PARTITIONS],const processed_line4 samec_plines[BLOCK_MAX_PARTITIONS],float uncor_lengths[BLOCK_MAX_PARTITIONS],float samec_lengths[BLOCK_MAX_PARTITIONS],float & uncor_error,float & samec_error)723 void compute_error_squared_rgba(
724 	const partition_info& pi,
725 	const image_block& blk,
726 	const processed_line4 uncor_plines[BLOCK_MAX_PARTITIONS],
727 	const processed_line4 samec_plines[BLOCK_MAX_PARTITIONS],
728 	float uncor_lengths[BLOCK_MAX_PARTITIONS],
729 	float samec_lengths[BLOCK_MAX_PARTITIONS],
730 	float& uncor_error,
731 	float& samec_error
732 ) {
733 	unsigned int partition_count = pi.partition_count;
734 	promise(partition_count > 0);
735 
736 	vfloatacc uncor_errorsumv = vfloatacc::zero();
737 	vfloatacc samec_errorsumv = vfloatacc::zero();
738 
739 	for (unsigned int partition = 0; partition < partition_count; partition++)
740 	{
741 		const uint8_t *texel_indexes = pi.texels_of_partition[partition];
742 
743 		float uncor_loparam = 1e10f;
744 		float uncor_hiparam = -1e10f;
745 
746 		float samec_loparam = 1e10f;
747 		float samec_hiparam = -1e10f;
748 
749 		processed_line4 l_uncor = uncor_plines[partition];
750 		processed_line4 l_samec = samec_plines[partition];
751 
752 		unsigned int texel_count = pi.partition_texel_count[partition];
753 		promise(texel_count > 0);
754 
755 		// Vectorize some useful scalar inputs
756 		vfloat l_uncor_bs0(l_uncor.bs.lane<0>());
757 		vfloat l_uncor_bs1(l_uncor.bs.lane<1>());
758 		vfloat l_uncor_bs2(l_uncor.bs.lane<2>());
759 		vfloat l_uncor_bs3(l_uncor.bs.lane<3>());
760 
761 		vfloat l_uncor_amod0(l_uncor.amod.lane<0>());
762 		vfloat l_uncor_amod1(l_uncor.amod.lane<1>());
763 		vfloat l_uncor_amod2(l_uncor.amod.lane<2>());
764 		vfloat l_uncor_amod3(l_uncor.amod.lane<3>());
765 
766 		vfloat l_samec_bs0(l_samec.bs.lane<0>());
767 		vfloat l_samec_bs1(l_samec.bs.lane<1>());
768 		vfloat l_samec_bs2(l_samec.bs.lane<2>());
769 		vfloat l_samec_bs3(l_samec.bs.lane<3>());
770 
771 		assert(all(l_samec.amod == vfloat4(0.0f)));
772 
773 		vfloat uncor_loparamv(1e10f);
774 		vfloat uncor_hiparamv(-1e10f);
775 
776 		vfloat samec_loparamv(1e10f);
777 		vfloat samec_hiparamv(-1e10f);
778 
779 		vfloat ew_r(blk.channel_weight.lane<0>());
780 		vfloat ew_g(blk.channel_weight.lane<1>());
781 		vfloat ew_b(blk.channel_weight.lane<2>());
782 		vfloat ew_a(blk.channel_weight.lane<3>());
783 
784 		// This implementation over-shoots, but this is safe as we initialize the texel_indexes
785 		// array to extend the last value. This means min/max are not impacted, but we need to mask
786 		// out the dummy values when we compute the line weighting.
787 		vint lane_ids = vint::lane_id();
788 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
789 		{
790 			vmask mask = lane_ids < vint(texel_count);
791 			vint texel_idxs(texel_indexes + i);
792 
793 			vfloat data_r = gatherf(blk.data_r, texel_idxs);
794 			vfloat data_g = gatherf(blk.data_g, texel_idxs);
795 			vfloat data_b = gatherf(blk.data_b, texel_idxs);
796 			vfloat data_a = gatherf(blk.data_a, texel_idxs);
797 
798 			vfloat uncor_param = (data_r * l_uncor_bs0)
799 			                   + (data_g * l_uncor_bs1)
800 			                   + (data_b * l_uncor_bs2)
801 			                   + (data_a * l_uncor_bs3);
802 
803 			uncor_loparamv = min(uncor_param, uncor_loparamv);
804 			uncor_hiparamv = max(uncor_param, uncor_hiparamv);
805 
806 			vfloat uncor_dist0 = (l_uncor_amod0 - data_r)
807 			                   + (uncor_param * l_uncor_bs0);
808 			vfloat uncor_dist1 = (l_uncor_amod1 - data_g)
809 			                   + (uncor_param * l_uncor_bs1);
810 			vfloat uncor_dist2 = (l_uncor_amod2 - data_b)
811 			                   + (uncor_param * l_uncor_bs2);
812 			vfloat uncor_dist3 = (l_uncor_amod3 - data_a)
813 			                   + (uncor_param * l_uncor_bs3);
814 
815 			vfloat uncor_err = (ew_r * uncor_dist0 * uncor_dist0)
816 			                 + (ew_g * uncor_dist1 * uncor_dist1)
817 			                 + (ew_b * uncor_dist2 * uncor_dist2)
818 			                 + (ew_a * uncor_dist3 * uncor_dist3);
819 
820 			haccumulate(uncor_errorsumv, uncor_err, mask);
821 
822 			// Process samechroma data
823 			vfloat samec_param = (data_r * l_samec_bs0)
824 			                   + (data_g * l_samec_bs1)
825 			                   + (data_b * l_samec_bs2)
826 			                   + (data_a * l_samec_bs3);
827 
828 			samec_loparamv = min(samec_param, samec_loparamv);
829 			samec_hiparamv = max(samec_param, samec_hiparamv);
830 
831 			vfloat samec_dist0 = samec_param * l_samec_bs0 - data_r;
832 			vfloat samec_dist1 = samec_param * l_samec_bs1 - data_g;
833 			vfloat samec_dist2 = samec_param * l_samec_bs2 - data_b;
834 			vfloat samec_dist3 = samec_param * l_samec_bs3 - data_a;
835 
836 			vfloat samec_err = (ew_r * samec_dist0 * samec_dist0)
837 			                 + (ew_g * samec_dist1 * samec_dist1)
838 			                 + (ew_b * samec_dist2 * samec_dist2)
839 			                 + (ew_a * samec_dist3 * samec_dist3);
840 
841 			haccumulate(samec_errorsumv, samec_err, mask);
842 
843 			lane_ids += vint(ASTCENC_SIMD_WIDTH);
844 		}
845 
846 		uncor_loparam = hmin_s(uncor_loparamv);
847 		uncor_hiparam = hmax_s(uncor_hiparamv);
848 
849 		samec_loparam = hmin_s(samec_loparamv);
850 		samec_hiparam = hmax_s(samec_hiparamv);
851 
852 		float uncor_linelen = uncor_hiparam - uncor_loparam;
853 		float samec_linelen = samec_hiparam - samec_loparam;
854 
855 		// Turn very small numbers and NaNs into a small number
856 		uncor_lengths[partition] = astc::max(uncor_linelen, 1e-7f);
857 		samec_lengths[partition] = astc::max(samec_linelen, 1e-7f);
858 	}
859 
860 	uncor_error = hadd_s(uncor_errorsumv);
861 	samec_error = hadd_s(samec_errorsumv);
862 }
863 
864 /* See header for documentation. */
compute_error_squared_rgb(const partition_info & pi,const image_block & blk,partition_lines3 plines[BLOCK_MAX_PARTITIONS],float & uncor_error,float & samec_error)865 void compute_error_squared_rgb(
866 	const partition_info& pi,
867 	const image_block& blk,
868 	partition_lines3 plines[BLOCK_MAX_PARTITIONS],
869 	float& uncor_error,
870 	float& samec_error
871 ) {
872 	unsigned int partition_count = pi.partition_count;
873 	promise(partition_count > 0);
874 
875 	vfloatacc uncor_errorsumv = vfloatacc::zero();
876 	vfloatacc samec_errorsumv = vfloatacc::zero();
877 
878 	for (unsigned int partition = 0; partition < partition_count; partition++)
879 	{
880 		partition_lines3& pl = plines[partition];
881 		const uint8_t *texel_indexes = pi.texels_of_partition[partition];
882 		unsigned int texel_count = pi.partition_texel_count[partition];
883 		promise(texel_count > 0);
884 
885 		float uncor_loparam = 1e10f;
886 		float uncor_hiparam = -1e10f;
887 
888 		float samec_loparam = 1e10f;
889 		float samec_hiparam = -1e10f;
890 
891 		processed_line3 l_uncor = pl.uncor_pline;
892 		processed_line3 l_samec = pl.samec_pline;
893 
894 		// This implementation is an example vectorization of this function.
895 		// It works for - the codec is a 2-4% faster than not vectorizing - but
896 		// the benefit is limited by the use of gathers and register pressure
897 
898 		// Vectorize some useful scalar inputs
899 		vfloat l_uncor_bs0(l_uncor.bs.lane<0>());
900 		vfloat l_uncor_bs1(l_uncor.bs.lane<1>());
901 		vfloat l_uncor_bs2(l_uncor.bs.lane<2>());
902 
903 		vfloat l_uncor_amod0(l_uncor.amod.lane<0>());
904 		vfloat l_uncor_amod1(l_uncor.amod.lane<1>());
905 		vfloat l_uncor_amod2(l_uncor.amod.lane<2>());
906 
907 		vfloat l_samec_bs0(l_samec.bs.lane<0>());
908 		vfloat l_samec_bs1(l_samec.bs.lane<1>());
909 		vfloat l_samec_bs2(l_samec.bs.lane<2>());
910 
911 		assert(all(l_samec.amod == vfloat4(0.0f)));
912 
913 		vfloat uncor_loparamv(1e10f);
914 		vfloat uncor_hiparamv(-1e10f);
915 
916 		vfloat samec_loparamv(1e10f);
917 		vfloat samec_hiparamv(-1e10f);
918 
919 		vfloat ew_r(blk.channel_weight.lane<0>());
920 		vfloat ew_g(blk.channel_weight.lane<1>());
921 		vfloat ew_b(blk.channel_weight.lane<2>());
922 
923 		// This implementation over-shoots, but this is safe as we initialize the weights array
924 		// to extend the last value. This means min/max are not impacted, but we need to mask
925 		// out the dummy values when we compute the line weighting.
926 		vint lane_ids = vint::lane_id();
927 		for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
928 		{
929 			vmask mask = lane_ids < vint(texel_count);
930 			vint texel_idxs(texel_indexes + i);
931 
932 			vfloat data_r = gatherf(blk.data_r, texel_idxs);
933 			vfloat data_g = gatherf(blk.data_g, texel_idxs);
934 			vfloat data_b = gatherf(blk.data_b, texel_idxs);
935 
936 			vfloat uncor_param = (data_r * l_uncor_bs0)
937 			                   + (data_g * l_uncor_bs1)
938 			                   + (data_b * l_uncor_bs2);
939 
940 			uncor_loparamv = min(uncor_param, uncor_loparamv);
941 			uncor_hiparamv = max(uncor_param, uncor_hiparamv);
942 
943 			vfloat uncor_dist0 = (l_uncor_amod0 - data_r)
944 			                   + (uncor_param * l_uncor_bs0);
945 			vfloat uncor_dist1 = (l_uncor_amod1 - data_g)
946 			                   + (uncor_param * l_uncor_bs1);
947 			vfloat uncor_dist2 = (l_uncor_amod2 - data_b)
948 			                   + (uncor_param * l_uncor_bs2);
949 
950 			vfloat uncor_err = (ew_r * uncor_dist0 * uncor_dist0)
951 			                 + (ew_g * uncor_dist1 * uncor_dist1)
952 			                 + (ew_b * uncor_dist2 * uncor_dist2);
953 
954 			haccumulate(uncor_errorsumv, uncor_err, mask);
955 
956 			// Process samechroma data
957 			vfloat samec_param = (data_r * l_samec_bs0)
958 			                   + (data_g * l_samec_bs1)
959 			                   + (data_b * l_samec_bs2);
960 
961 			samec_loparamv = min(samec_param, samec_loparamv);
962 			samec_hiparamv = max(samec_param, samec_hiparamv);
963 
964 			vfloat samec_dist0 = samec_param * l_samec_bs0 - data_r;
965 			vfloat samec_dist1 = samec_param * l_samec_bs1 - data_g;
966 			vfloat samec_dist2 = samec_param * l_samec_bs2 - data_b;
967 
968 			vfloat samec_err = (ew_r * samec_dist0 * samec_dist0)
969 			                 + (ew_g * samec_dist1 * samec_dist1)
970 			                 + (ew_b * samec_dist2 * samec_dist2);
971 
972 			haccumulate(samec_errorsumv, samec_err, mask);
973 
974 			lane_ids += vint(ASTCENC_SIMD_WIDTH);
975 		}
976 
977 		uncor_loparam = hmin_s(uncor_loparamv);
978 		uncor_hiparam = hmax_s(uncor_hiparamv);
979 
980 		samec_loparam = hmin_s(samec_loparamv);
981 		samec_hiparam = hmax_s(samec_hiparamv);
982 
983 		float uncor_linelen = uncor_hiparam - uncor_loparam;
984 		float samec_linelen = samec_hiparam - samec_loparam;
985 
986 		// Turn very small numbers and NaNs into a small number
987 		pl.uncor_line_len = astc::max(uncor_linelen, 1e-7f);
988 		pl.samec_line_len = astc::max(samec_linelen, 1e-7f);
989 	}
990 
991 	uncor_error = hadd_s(uncor_errorsumv);
992 	samec_error = hadd_s(samec_errorsumv);
993 }
994 
995 #endif
996