1 // SPDX-License-Identifier: Apache-2.0
2 // ----------------------------------------------------------------------------
3 // Copyright 2011-2024 Arm Limited
4 //
5 // Licensed under the Apache License, Version 2.0 (the "License"); you may not
6 // use this file except in compliance with the License. You may obtain a copy
7 // of the License at:
8 //
9 // http://www.apache.org/licenses/LICENSE-2.0
10 //
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
13 // WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
14 // License for the specific language governing permissions and limitations
15 // under the License.
16 // ----------------------------------------------------------------------------
17
18 /**
19 * @brief Functions to decompress a symbolic block.
20 */
21
22 #include "astcenc_internal.h"
23
24 #include <stdio.h>
25 #include <assert.h>
26
27 /**
28 * @brief Compute the integer linear interpolation of two color endpoints.
29 *
30 * @param u8_mask The mask for lanes using decode_unorm8 rather than decode_f16.
31 * @param color0 The endpoint0 color.
32 * @param color1 The endpoint1 color.
33 * @param weights The interpolation weight (between 0 and 64).
34 *
35 * @return The interpolated color.
36 */
lerp_color_int( vmask4 u8_mask, vint4 color0, vint4 color1, vint4 weights )37 static vint4 lerp_color_int(
38 vmask4 u8_mask,
39 vint4 color0,
40 vint4 color1,
41 vint4 weights
42 ) {
43 vint4 weight1 = weights;
44 vint4 weight0 = vint4(64) - weight1;
45
46 vint4 color = (color0 * weight0) + (color1 * weight1) + vint4(32);
47 color = asr<6>(color);
48
49 // For decode_unorm8 values force the codec to bit replicate. This allows the
50 // rest of the codec to assume the full 0xFFFF range for everything and ignore
51 // the decode_mode setting
52 vint4 color_u8 = asr<8>(color) * vint4(257);
53 color = select(color, color_u8, u8_mask);
54
55 return color;
56 }
57
58 /**
59 * @brief Convert integer color value into a float value for the decoder.
60 *
61 * @param data The integer color value post-interpolation.
62 * @param lns_mask If set treat lane as HDR (LNS) else LDR (unorm16).
63 *
64 * @return The float color value.
65 */
decode_texel( vint4 data, vmask4 lns_mask )66 static inline vfloat4 decode_texel(
67 vint4 data,
68 vmask4 lns_mask
69 ) {
70 vint4 color_lns = vint4::zero();
71 vint4 color_unorm = vint4::zero();
72
73 if (any(lns_mask))
74 {
75 color_lns = lns_to_sf16(data);
76 }
77
78 if (!all(lns_mask))
79 {
80 color_unorm = unorm16_to_sf16(data);
81 }
82
83 // Pick components and then convert to FP16
84 vint4 datai = select(color_unorm, color_lns, lns_mask);
85 return float16_to_float(datai);
86 }
87
88 /* See header for documentation. */
unpack_weights( const block_size_descriptor& bsd, const symbolic_compressed_block& scb, const decimation_info& di, bool is_dual_plane, int weights_plane1[BLOCK_MAX_TEXELS], int weights_plane2[BLOCK_MAX_TEXELS] )89 void unpack_weights(
90 const block_size_descriptor& bsd,
91 const symbolic_compressed_block& scb,
92 const decimation_info& di,
93 bool is_dual_plane,
94 int weights_plane1[BLOCK_MAX_TEXELS],
95 int weights_plane2[BLOCK_MAX_TEXELS]
96 ) {
97 // Safe to overshoot as all arrays are allocated to full size
98 if (!is_dual_plane)
99 {
100 // Build full 64-entry weight lookup table
101 vint4 tab0 = vint4::load(scb.weights + 0);
102 vint4 tab1 = vint4::load(scb.weights + 16);
103 vint4 tab2 = vint4::load(scb.weights + 32);
104 vint4 tab3 = vint4::load(scb.weights + 48);
105
106 vint tab0p, tab1p, tab2p, tab3p;
107 vtable_prepare(tab0, tab1, tab2, tab3, tab0p, tab1p, tab2p, tab3p);
108
109 for (unsigned int i = 0; i < bsd.texel_count; i += ASTCENC_SIMD_WIDTH)
110 {
111 vint summed_value(8);
112 vint weight_count(di.texel_weight_count + i);
113 int max_weight_count = hmax(weight_count).lane<0>();
114
115 promise(max_weight_count > 0);
116 for (int j = 0; j < max_weight_count; j++)
117 {
118 vint texel_weights(di.texel_weights_tr[j] + i);
119 vint texel_weights_int(di.texel_weight_contribs_int_tr[j] + i);
120
121 summed_value += vtable_8bt_32bi(tab0p, tab1p, tab2p, tab3p, texel_weights) * texel_weights_int;
122 }
123
124 store(lsr<4>(summed_value), weights_plane1 + i);
125 }
126 }
127 else
128 {
129 // Build a 32-entry weight lookup table per plane
130 // Plane 1
131 vint4 tab0_plane1 = vint4::load(scb.weights + 0);
132 vint4 tab1_plane1 = vint4::load(scb.weights + 16);
133 vint tab0_plane1p, tab1_plane1p;
134 vtable_prepare(tab0_plane1, tab1_plane1, tab0_plane1p, tab1_plane1p);
135
136 // Plane 2
137 vint4 tab0_plane2 = vint4::load(scb.weights + 32);
138 vint4 tab1_plane2 = vint4::load(scb.weights + 48);
139 vint tab0_plane2p, tab1_plane2p;
140 vtable_prepare(tab0_plane2, tab1_plane2, tab0_plane2p, tab1_plane2p);
141
142 for (unsigned int i = 0; i < bsd.texel_count; i += ASTCENC_SIMD_WIDTH)
143 {
144 vint sum_plane1(8);
145 vint sum_plane2(8);
146
147 vint weight_count(di.texel_weight_count + i);
148 int max_weight_count = hmax(weight_count).lane<0>();
149
150 promise(max_weight_count > 0);
151 for (int j = 0; j < max_weight_count; j++)
152 {
153 vint texel_weights(di.texel_weights_tr[j] + i);
154 vint texel_weights_int(di.texel_weight_contribs_int_tr[j] + i);
155
156 sum_plane1 += vtable_8bt_32bi(tab0_plane1p, tab1_plane1p, texel_weights) * texel_weights_int;
157 sum_plane2 += vtable_8bt_32bi(tab0_plane2p, tab1_plane2p, texel_weights) * texel_weights_int;
158 }
159
160 store(lsr<4>(sum_plane1), weights_plane1 + i);
161 store(lsr<4>(sum_plane2), weights_plane2 + i);
162 }
163 }
164 }
165
166 /**
167 * @brief Return an FP32 NaN value for use in error colors.
168 *
169 * This NaN encoding will turn into 0xFFFF when converted to an FP16 NaN.
170 *
171 * @return The float color value.
172 */
error_color_nan()173 static float error_color_nan()
174 {
175 if32 v;
176 v.u = 0xFFFFE000U;
177 return v.f;
178 }
179
180 /* See header for documentation. */
decompress_symbolic_block( astcenc_profile decode_mode, const block_size_descriptor& bsd, int xpos, int ypos, int zpos, const symbolic_compressed_block& scb, image_block& blk )181 void decompress_symbolic_block(
182 astcenc_profile decode_mode,
183 const block_size_descriptor& bsd,
184 int xpos,
185 int ypos,
186 int zpos,
187 const symbolic_compressed_block& scb,
188 image_block& blk
189 ) {
190 blk.xpos = xpos;
191 blk.ypos = ypos;
192 blk.zpos = zpos;
193
194 blk.data_min = vfloat4::zero();
195 blk.data_mean = vfloat4::zero();
196 blk.data_max = vfloat4::zero();
197 blk.grayscale = false;
198
199 // If we detected an error-block, blow up immediately.
200 if (scb.block_type == SYM_BTYPE_ERROR)
201 {
202 for (unsigned int i = 0; i < bsd.texel_count; i++)
203 {
204 blk.data_r[i] = error_color_nan();
205 blk.data_g[i] = error_color_nan();
206 blk.data_b[i] = error_color_nan();
207 blk.data_a[i] = error_color_nan();
208 blk.rgb_lns[i] = 0;
209 blk.alpha_lns[i] = 0;
210 }
211
212 return;
213 }
214
215 if ((scb.block_type == SYM_BTYPE_CONST_F16) ||
216 (scb.block_type == SYM_BTYPE_CONST_U16))
217 {
218 vfloat4 color;
219 uint8_t use_lns = 0;
220
221 // UNORM16 constant color block
222 if (scb.block_type == SYM_BTYPE_CONST_U16)
223 {
224 vint4 colori(scb.constant_color);
225
226 // Determine the UNORM8 rounding on the decode
227 vmask4 u8_mask = get_u8_component_mask(decode_mode, blk);
228
229 // The real decoder would just use the top 8 bits, but we rescale
230 // in to a 16-bit value that rounds correctly.
231 vint4 colori_u8 = asr<8>(colori) * 257;
232 colori = select(colori, colori_u8, u8_mask);
233
234 vint4 colorf16 = unorm16_to_sf16(colori);
235 color = float16_to_float(colorf16);
236 }
237 // FLOAT16 constant color block
238 else
239 {
240 switch (decode_mode)
241 {
242 case ASTCENC_PRF_LDR_SRGB:
243 case ASTCENC_PRF_LDR:
244 color = vfloat4(error_color_nan());
245 break;
246 case ASTCENC_PRF_HDR_RGB_LDR_A:
247 case ASTCENC_PRF_HDR:
248 // Constant-color block; unpack from FP16 to FP32.
249 color = float16_to_float(vint4(scb.constant_color));
250 use_lns = 1;
251 break;
252 }
253 }
254
255 for (unsigned int i = 0; i < bsd.texel_count; i++)
256 {
257 blk.data_r[i] = color.lane<0>();
258 blk.data_g[i] = color.lane<1>();
259 blk.data_b[i] = color.lane<2>();
260 blk.data_a[i] = color.lane<3>();
261 blk.rgb_lns[i] = use_lns;
262 blk.alpha_lns[i] = use_lns;
263 }
264
265 return;
266 }
267
268 // Get the appropriate partition-table entry
269 int partition_count = scb.partition_count;
270 const auto& pi = bsd.get_partition_info(partition_count, scb.partition_index);
271
272 // Get the appropriate block descriptors
273 const auto& bm = bsd.get_block_mode(scb.block_mode);
274 const auto& di = bsd.get_decimation_info(bm.decimation_mode);
275
276 bool is_dual_plane = static_cast<bool>(bm.is_dual_plane);
277
278 // Unquantize and undecimate the weights
279 int plane1_weights[BLOCK_MAX_TEXELS];
280 int plane2_weights[BLOCK_MAX_TEXELS];
281 unpack_weights(bsd, scb, di, is_dual_plane, plane1_weights, plane2_weights);
282
283 // Now that we have endpoint colors and weights, we can unpack texel colors
284 int plane2_component = scb.plane2_component;
285 vmask4 plane2_mask = vint4::lane_id() == vint4(plane2_component);
286
287 vmask4 u8_mask = get_u8_component_mask(decode_mode, blk);
288
289 for (int i = 0; i < partition_count; i++)
290 {
291 // Decode the color endpoints for this partition
292 vint4 ep0;
293 vint4 ep1;
294 bool rgb_lns;
295 bool a_lns;
296
297 unpack_color_endpoints(decode_mode,
298 scb.color_formats[i],
299 scb.color_values[i],
300 rgb_lns, a_lns,
301 ep0, ep1);
302
303 vmask4 lns_mask(rgb_lns, rgb_lns, rgb_lns, a_lns);
304
305 int texel_count = pi.partition_texel_count[i];
306 for (int j = 0; j < texel_count; j++)
307 {
308 int tix = pi.texels_of_partition[i][j];
309 vint4 weight = select(vint4(plane1_weights[tix]), vint4(plane2_weights[tix]), plane2_mask);
310 vint4 color = lerp_color_int(u8_mask, ep0, ep1, weight);
311 vfloat4 colorf = decode_texel(color, lns_mask);
312
313 blk.data_r[tix] = colorf.lane<0>();
314 blk.data_g[tix] = colorf.lane<1>();
315 blk.data_b[tix] = colorf.lane<2>();
316 blk.data_a[tix] = colorf.lane<3>();
317 }
318 }
319 }
320
321 #if !defined(ASTCENC_DECOMPRESS_ONLY)
322
323 /* See header for documentation. */
compute_symbolic_block_difference_2plane( const astcenc_config& config, const block_size_descriptor& bsd, const symbolic_compressed_block& scb, const image_block& blk )324 float compute_symbolic_block_difference_2plane(
325 const astcenc_config& config,
326 const block_size_descriptor& bsd,
327 const symbolic_compressed_block& scb,
328 const image_block& blk
329 ) {
330 // If we detected an error-block, blow up immediately.
331 if (scb.block_type == SYM_BTYPE_ERROR)
332 {
333 return ERROR_CALC_DEFAULT;
334 }
335
336 assert(scb.block_mode >= 0);
337 assert(scb.partition_count == 1);
338 assert(bsd.get_block_mode(scb.block_mode).is_dual_plane == 1);
339
340 // Get the appropriate block descriptor
341 const block_mode& bm = bsd.get_block_mode(scb.block_mode);
342 const decimation_info& di = bsd.get_decimation_info(bm.decimation_mode);
343
344 // Unquantize and undecimate the weights
345 int plane1_weights[BLOCK_MAX_TEXELS];
346 int plane2_weights[BLOCK_MAX_TEXELS];
347 unpack_weights(bsd, scb, di, true, plane1_weights, plane2_weights);
348
349 vmask4 plane2_mask = vint4::lane_id() == vint4(scb.plane2_component);
350
351 vfloat4 summa = vfloat4::zero();
352
353 // Decode the color endpoints for this partition
354 vint4 ep0;
355 vint4 ep1;
356 bool rgb_lns;
357 bool a_lns;
358
359 unpack_color_endpoints(config.profile,
360 scb.color_formats[0],
361 scb.color_values[0],
362 rgb_lns, a_lns,
363 ep0, ep1);
364
365 vmask4 u8_mask = get_u8_component_mask(config.profile, blk);
366
367 // Unpack and compute error for each texel in the partition
368 unsigned int texel_count = bsd.texel_count;
369 for (unsigned int i = 0; i < texel_count; i++)
370 {
371 vint4 weight = select(vint4(plane1_weights[i]), vint4(plane2_weights[i]), plane2_mask);
372 vint4 colori = lerp_color_int(u8_mask, ep0, ep1, weight);
373
374 vfloat4 color = int_to_float(colori);
375 vfloat4 oldColor = blk.texel(i);
376
377 // Compare error using a perceptual decode metric for RGBM textures
378 if (config.flags & ASTCENC_FLG_MAP_RGBM)
379 {
380 // Fail encodings that result in zero weight M pixels. Note that this can cause
381 // "interesting" artifacts if we reject all useful encodings - we typically get max
382 // brightness encodings instead which look just as bad. We recommend users apply a
383 // bias to their stored M value, limiting the lower value to 16 or 32 to avoid
384 // getting small M values post-quantization, but we can't prove it would never
385 // happen, especially at low bit rates ...
386 if (color.lane<3>() == 0.0f)
387 {
388 return -ERROR_CALC_DEFAULT;
389 }
390
391 // Compute error based on decoded RGBM color
392 color = vfloat4(
393 color.lane<0>() * color.lane<3>() * config.rgbm_m_scale,
394 color.lane<1>() * color.lane<3>() * config.rgbm_m_scale,
395 color.lane<2>() * color.lane<3>() * config.rgbm_m_scale,
396 1.0f
397 );
398
399 oldColor = vfloat4(
400 oldColor.lane<0>() * oldColor.lane<3>() * config.rgbm_m_scale,
401 oldColor.lane<1>() * oldColor.lane<3>() * config.rgbm_m_scale,
402 oldColor.lane<2>() * oldColor.lane<3>() * config.rgbm_m_scale,
403 1.0f
404 );
405 }
406
407 vfloat4 error = oldColor - color;
408 error = min(abs(error), 1e15f);
409 error = error * error;
410
411 summa += min(dot(error, blk.channel_weight), ERROR_CALC_DEFAULT);
412 }
413
414 return summa.lane<0>();
415 }
416
417 /* See header for documentation. */
compute_symbolic_block_difference_1plane( const astcenc_config& config, const block_size_descriptor& bsd, const symbolic_compressed_block& scb, const image_block& blk )418 float compute_symbolic_block_difference_1plane(
419 const astcenc_config& config,
420 const block_size_descriptor& bsd,
421 const symbolic_compressed_block& scb,
422 const image_block& blk
423 ) {
424 assert(bsd.get_block_mode(scb.block_mode).is_dual_plane == 0);
425
426 // If we detected an error-block, blow up immediately.
427 if (scb.block_type == SYM_BTYPE_ERROR)
428 {
429 return ERROR_CALC_DEFAULT;
430 }
431
432 assert(scb.block_mode >= 0);
433
434 // Get the appropriate partition-table entry
435 unsigned int partition_count = scb.partition_count;
436 const auto& pi = bsd.get_partition_info(partition_count, scb.partition_index);
437
438 // Get the appropriate block descriptor
439 const block_mode& bm = bsd.get_block_mode(scb.block_mode);
440 const decimation_info& di = bsd.get_decimation_info(bm.decimation_mode);
441
442 // Unquantize and undecimate the weights
443 int plane1_weights[BLOCK_MAX_TEXELS];
444 unpack_weights(bsd, scb, di, false, plane1_weights, nullptr);
445
446 vmask4 u8_mask = get_u8_component_mask(config.profile, blk);
447
448 vfloat4 summa = vfloat4::zero();
449 for (unsigned int i = 0; i < partition_count; i++)
450 {
451 // Decode the color endpoints for this partition
452 vint4 ep0;
453 vint4 ep1;
454 bool rgb_lns;
455 bool a_lns;
456
457 unpack_color_endpoints(config.profile,
458 scb.color_formats[i],
459 scb.color_values[i],
460 rgb_lns, a_lns,
461 ep0, ep1);
462
463 // Unpack and compute error for each texel in the partition
464 unsigned int texel_count = pi.partition_texel_count[i];
465 for (unsigned int j = 0; j < texel_count; j++)
466 {
467 unsigned int tix = pi.texels_of_partition[i][j];
468 vint4 colori = lerp_color_int(u8_mask, ep0, ep1,
469 vint4(plane1_weights[tix]));
470
471 vfloat4 color = int_to_float(colori);
472 vfloat4 oldColor = blk.texel(tix);
473
474 // Compare error using a perceptual decode metric for RGBM textures
475 if (config.flags & ASTCENC_FLG_MAP_RGBM)
476 {
477 // Fail encodings that result in zero weight M pixels. Note that this can cause
478 // "interesting" artifacts if we reject all useful encodings - we typically get max
479 // brightness encodings instead which look just as bad. We recommend users apply a
480 // bias to their stored M value, limiting the lower value to 16 or 32 to avoid
481 // getting small M values post-quantization, but we can't prove it would never
482 // happen, especially at low bit rates ...
483 if (color.lane<3>() == 0.0f)
484 {
485 return -ERROR_CALC_DEFAULT;
486 }
487
488 // Compute error based on decoded RGBM color
489 color = vfloat4(
490 color.lane<0>() * color.lane<3>() * config.rgbm_m_scale,
491 color.lane<1>() * color.lane<3>() * config.rgbm_m_scale,
492 color.lane<2>() * color.lane<3>() * config.rgbm_m_scale,
493 1.0f
494 );
495
496 oldColor = vfloat4(
497 oldColor.lane<0>() * oldColor.lane<3>() * config.rgbm_m_scale,
498 oldColor.lane<1>() * oldColor.lane<3>() * config.rgbm_m_scale,
499 oldColor.lane<2>() * oldColor.lane<3>() * config.rgbm_m_scale,
500 1.0f
501 );
502 }
503
504 vfloat4 error = oldColor - color;
505 error = min(abs(error), 1e15f);
506 error = error * error;
507
508 summa += min(dot(error, blk.channel_weight), ERROR_CALC_DEFAULT);
509 }
510 }
511
512 return summa.lane<0>();
513 }
514
515 /* See header for documentation. */
compute_symbolic_block_difference_1plane_1partition( const astcenc_config& config, const block_size_descriptor& bsd, const symbolic_compressed_block& scb, const image_block& blk )516 float compute_symbolic_block_difference_1plane_1partition(
517 const astcenc_config& config,
518 const block_size_descriptor& bsd,
519 const symbolic_compressed_block& scb,
520 const image_block& blk
521 ) {
522 // If we detected an error-block, blow up immediately.
523 if (scb.block_type == SYM_BTYPE_ERROR)
524 {
525 return ERROR_CALC_DEFAULT;
526 }
527
528 assert(scb.block_mode >= 0);
529 assert(bsd.get_partition_info(scb.partition_count, scb.partition_index).partition_count == 1);
530
531 // Get the appropriate block descriptor
532 const block_mode& bm = bsd.get_block_mode(scb.block_mode);
533 const decimation_info& di = bsd.get_decimation_info(bm.decimation_mode);
534
535 // Unquantize and undecimate the weights
536 ASTCENC_ALIGNAS int plane1_weights[BLOCK_MAX_TEXELS];
537 unpack_weights(bsd, scb, di, false, plane1_weights, nullptr);
538
539 // Decode the color endpoints for this partition
540 vint4 ep0;
541 vint4 ep1;
542 bool rgb_lns;
543 bool a_lns;
544
545 unpack_color_endpoints(config.profile,
546 scb.color_formats[0],
547 scb.color_values[0],
548 rgb_lns, a_lns,
549 ep0, ep1);
550
551 vmask4 u8_mask = get_u8_component_mask(config.profile, blk);
552
553 // Unpack and compute error for each texel in the partition
554 vfloatacc summav = vfloatacc::zero();
555
556 vint lane_id = vint::lane_id();
557
558 unsigned int texel_count = bsd.texel_count;
559 for (unsigned int i = 0; i < texel_count; i += ASTCENC_SIMD_WIDTH)
560 {
561 // Compute EP1 contribution
562 vint weight1 = vint::loada(plane1_weights + i);
563 vint ep1_r = vint(ep1.lane<0>()) * weight1;
564 vint ep1_g = vint(ep1.lane<1>()) * weight1;
565 vint ep1_b = vint(ep1.lane<2>()) * weight1;
566 vint ep1_a = vint(ep1.lane<3>()) * weight1;
567
568 // Compute EP0 contribution
569 vint weight0 = vint(64) - weight1;
570 vint ep0_r = vint(ep0.lane<0>()) * weight0;
571 vint ep0_g = vint(ep0.lane<1>()) * weight0;
572 vint ep0_b = vint(ep0.lane<2>()) * weight0;
573 vint ep0_a = vint(ep0.lane<3>()) * weight0;
574
575 // Combine contributions
576 vint colori_r = asr<6>(ep0_r + ep1_r + vint(32));
577 vint colori_g = asr<6>(ep0_g + ep1_g + vint(32));
578 vint colori_b = asr<6>(ep0_b + ep1_b + vint(32));
579 vint colori_a = asr<6>(ep0_a + ep1_a + vint(32));
580
581 // If using a U8 decode mode bit replicate top 8 bits
582 // so rest of codec can assume 0xFFFF max range everywhere
583 vint colori_r8 = asr<8>(colori_r) * vint(257);
584 colori_r = select(colori_r, colori_r8, vmask(u8_mask.lane<0>()));
585
586 vint colori_g8 = asr<8>(colori_g) * vint(257);
587 colori_g = select(colori_g, colori_g8, vmask(u8_mask.lane<1>()));
588
589 vint colori_b8 = asr<8>(colori_b) * vint(257);
590 colori_b = select(colori_b, colori_b8, vmask(u8_mask.lane<2>()));
591
592 vint colori_a8 = asr<8>(colori_a) * vint(257);
593 colori_a = select(colori_a, colori_a8, vmask(u8_mask.lane<3>()));
594
595 // Compute color diff
596 vfloat color_r = int_to_float(colori_r);
597 vfloat color_g = int_to_float(colori_g);
598 vfloat color_b = int_to_float(colori_b);
599 vfloat color_a = int_to_float(colori_a);
600
601 vfloat color_orig_r = loada(blk.data_r + i);
602 vfloat color_orig_g = loada(blk.data_g + i);
603 vfloat color_orig_b = loada(blk.data_b + i);
604 vfloat color_orig_a = loada(blk.data_a + i);
605
606 vfloat color_error_r = min(abs(color_orig_r - color_r), vfloat(1e15f));
607 vfloat color_error_g = min(abs(color_orig_g - color_g), vfloat(1e15f));
608 vfloat color_error_b = min(abs(color_orig_b - color_b), vfloat(1e15f));
609 vfloat color_error_a = min(abs(color_orig_a - color_a), vfloat(1e15f));
610
611 // Compute squared error metric
612 color_error_r = color_error_r * color_error_r;
613 color_error_g = color_error_g * color_error_g;
614 color_error_b = color_error_b * color_error_b;
615 color_error_a = color_error_a * color_error_a;
616
617 vfloat metric = color_error_r * blk.channel_weight.lane<0>()
618 + color_error_g * blk.channel_weight.lane<1>()
619 + color_error_b * blk.channel_weight.lane<2>()
620 + color_error_a * blk.channel_weight.lane<3>();
621
622 // Mask off bad lanes
623 vmask mask = lane_id < vint(texel_count);
624 lane_id += vint(ASTCENC_SIMD_WIDTH);
625 haccumulate(summav, metric, mask);
626 }
627
628 return hadd_s(summav);
629 }
630
631 #endif
632