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 #if !defined(ASTCENC_DECOMPRESS_ONLY)
19 
20 /**
21  * @brief Functions to compress a symbolic block.
22  */
23 
24 #include "astcenc_internal.h"
25 #include "astcenc_diagnostic_trace.h"
26 
27 #include <cassert>
28 #ifdef ASTC_CUSTOMIZED_ENABLE
29 AstcCustomizedSoManager g_astcCustomizedSoManager;
30 #endif
31 
32 /**
33  * @brief Merge two planes of endpoints into a single vector.
34  *
35  * @param      ep_plane1          The endpoints for plane 1.
36  * @param      ep_plane2          The endpoints for plane 2.
37  * @param      component_plane2   The color component for plane 2.
38  * @param[out] result             The merged output.
39  */
merge_endpoints( const endpoints& ep_plane1, const endpoints& ep_plane2, unsigned int component_plane2, endpoints& result )40 static void merge_endpoints(
41 	const endpoints& ep_plane1,
42 	const endpoints& ep_plane2,
43 	unsigned int component_plane2,
44 	endpoints& result
45 ) {
46 	unsigned int partition_count = ep_plane1.partition_count;
47 	assert(partition_count == 1);
48 
49 	vmask4 sep_mask = vint4::lane_id() == vint4(component_plane2);
50 
51 	result.partition_count = partition_count;
52 	result.endpt0[0] = select(ep_plane1.endpt0[0], ep_plane2.endpt0[0], sep_mask);
53 	result.endpt1[0] = select(ep_plane1.endpt1[0], ep_plane2.endpt1[0], sep_mask);
54 }
55 
56 /**
57  * @brief Attempt to improve weights given a chosen configuration.
58  *
59  * Given a fixed weight grid decimation and weight value quantization, iterate over all weights (per
60  * partition and per plane) and attempt to improve image quality by moving each weight up by one or
61  * down by one quantization step.
62  *
63  * This is a specialized function which only supports operating on undecimated weight grids,
64  * therefore primarily improving the performance of 4x4 and 5x5 blocks where grid decimation
65  * is needed less often.
66  *
67  * @param      decode_mode   The decode mode (LDR, HDR).
68  * @param      bsd           The block size information.
69  * @param      blk           The image block color data to compress.
70  * @param[out] scb           The symbolic compressed block output.
71  */
72 #if ASTCENC_NEON != 0
realign_weights_undecimated( astcenc_profile decode_mode, const block_size_descriptor& bsd, const image_block& blk, symbolic_compressed_block& scb )73 static bool realign_weights_undecimated(
74 	astcenc_profile decode_mode,
75 	const block_size_descriptor& bsd,
76 	const image_block& blk,
77 	symbolic_compressed_block& scb
78 ) {
79 	// Get the partition descriptor
80 	unsigned int partition_count = scb.partition_count;
81 	const auto& pi = bsd.get_partition_info(partition_count, scb.partition_index);
82 
83 	// Get the quantization table
84 	const block_mode& bm = bsd.get_block_mode(scb.block_mode);
85 	unsigned int weight_quant_level = bm.quant_mode;
86 	const quant_and_transfer_table& qat = quant_and_xfer_tables[weight_quant_level];
87 
88 	unsigned int max_plane = bm.is_dual_plane;
89 	int plane2_component = scb.plane2_component;
90 	vmask4 plane_mask = vint4::lane_id() == vint4(plane2_component);
91 
92 	// Decode the color endpoints
93 	bool rgb_hdr;
94 	bool alpha_hdr;
95 	vint4 endpnt0[BLOCK_MAX_PARTITIONS];
96 	vint4 endpnt1[BLOCK_MAX_PARTITIONS];
97 	vfloat4 endpnt0f[BLOCK_MAX_PARTITIONS];
98 	vfloat4 offset[BLOCK_MAX_PARTITIONS];
99 
100 	promise(partition_count > 0);
101 
102 	for (unsigned int pa_idx = 0; pa_idx < partition_count; pa_idx++)
103 	{
104 		unpack_color_endpoints(decode_mode,
105 		                       scb.color_formats[pa_idx],
106 		                       scb.color_values[pa_idx],
107 		                       rgb_hdr, alpha_hdr,
108 		                       endpnt0[pa_idx],
109 		                       endpnt1[pa_idx]);
110 	}
111 
112 	uint8_t* dec_weights_uquant = scb.weights;
113 	bool adjustments = false;
114 
115 	// For each plane and partition ...
116 	for (unsigned int pl_idx = 0; pl_idx <= max_plane; pl_idx++)
117 	{
118 		for (unsigned int pa_idx = 0; pa_idx < partition_count; pa_idx++)
119 		{
120 			// Compute the endpoint delta for all components in current plane
121 			vint4 epd = endpnt1[pa_idx] - endpnt0[pa_idx];
122 			epd = select(epd, vint4::zero(), plane_mask);
123 
124 			endpnt0f[pa_idx] = int_to_float(endpnt0[pa_idx]);
125 			offset[pa_idx] = int_to_float(epd) * (1.0f / 64.0f);
126 		}
127 
128 		// For each weight compute previous, current, and next errors
129 		promise(bsd.texel_count > 0);
130 
131 		unsigned int texel = 0;
132 		for (; texel + ASTCENC_SIMD_WIDTH <= bsd.texel_count; texel += ASTCENC_SIMD_WIDTH)
133 		{
134 			int uqw0 = dec_weights_uquant[texel];
135 			int uqw1 = dec_weights_uquant[texel + 1];
136 			int uqw2 = dec_weights_uquant[texel + 2];
137 			int uqw3 = dec_weights_uquant[texel + 3];
138 
139 			vint4 uqw_vec = vint4(uqw0, uqw1, uqw2, uqw3);
140 			vint4 prev_and_next_vec = vint4(qat.prev_next_values[uqw0], qat.prev_next_values[uqw1],
141 							qat.prev_next_values[uqw2], qat.prev_next_values[uqw3]);
142 
143 			vint4 mask = vint4(0xFF, 0xFF, 0xFF, 0xFF);
144 			vint4 uqw_down_vec = prev_and_next_vec & mask;
145 			vint4 uqw_up_vec = vint4(vshrq_n_s32(prev_and_next_vec.m, 8)) & mask;
146 
147 			vfloat4 weight_base_vec = int_to_float(uqw_vec);
148 			vfloat4 weight_down_vec = int_to_float(uqw_down_vec) - weight_base_vec;
149 			vfloat4 weight_up_vec = int_to_float(uqw_up_vec) - weight_base_vec;
150 
151 			unsigned int partition0 = pi.partition_of_texel[texel];
152 			unsigned int partition1 = pi.partition_of_texel[texel + 1];
153 			unsigned int partition2 = pi.partition_of_texel[texel + 2];
154 			unsigned int partition3 = pi.partition_of_texel[texel + 3];
155 
156 			vfloat4 color_offset0 = offset[partition0];
157 			vfloat4 color_offset1 = offset[partition1];
158 			vfloat4 color_offset2 = offset[partition2];
159 			vfloat4 color_offset3 = offset[partition3];
160 
161 			vfloat4 color_base0 = endpnt0f[partition0];
162 			vfloat4 color_base1 = endpnt0f[partition1];
163 			vfloat4 color_base2 = endpnt0f[partition2];
164 			vfloat4 color_base3 = endpnt0f[partition3];
165 
166 			vfloat4 color0 = color_base0 + color_offset0 * weight_base_vec.lane<0>();
167 			vfloat4 color1 = color_base1 + color_offset1 * weight_base_vec.lane<1>();
168 			vfloat4 color2 = color_base2 + color_offset2 * weight_base_vec.lane<2>();
169 			vfloat4 color3 = color_base3 + color_offset3 * weight_base_vec.lane<3>();
170 
171 			vfloat4 orig_color0 = blk.texel(texel);
172 			vfloat4 orig_color1 = blk.texel(texel + 1);
173 			vfloat4 orig_color2 = blk.texel(texel + 2);
174 			vfloat4 orig_color3 = blk.texel(texel + 3);
175 
176 			vfloat4 error_weight = blk.channel_weight;
177 
178 			vfloat4 color_diff0 = color0 - orig_color0;
179 			vfloat4 color_diff1 = color1 - orig_color1;
180 			vfloat4 color_diff2 = color2 - orig_color2;
181 			vfloat4 color_diff3 = color3 - orig_color3;
182 
183 			vfloat4 color_diff_down0 = color_diff0 + color_offset0 * weight_down_vec.lane<0>();
184 			vfloat4 color_diff_down1 = color_diff1 + color_offset1 * weight_down_vec.lane<1>();
185 			vfloat4 color_diff_down2 = color_diff2 + color_offset2 * weight_down_vec.lane<2>();
186 			vfloat4 color_diff_down3 = color_diff3 + color_offset3 * weight_down_vec.lane<3>();
187 
188 			vfloat4 color_diff_up0 = color_diff0 + color_offset0 * weight_up_vec.lane<0>();
189 			vfloat4 color_diff_up1 = color_diff1 + color_offset1 * weight_up_vec.lane<1>();
190 			vfloat4 color_diff_up2 = color_diff2 + color_offset2 * weight_up_vec.lane<2>();
191 			vfloat4 color_diff_up3 = color_diff3 + color_offset3 * weight_up_vec.lane<3>();
192 
193 			float error_base0 = dot_s(color_diff0 * color_diff0, error_weight);
194 			float error_base1 = dot_s(color_diff1 * color_diff1, error_weight);
195 			float error_base2 = dot_s(color_diff2 * color_diff2, error_weight);
196 			float error_base3 = dot_s(color_diff3 * color_diff3, error_weight);
197 
198 			float error_down0 = dot_s(color_diff_down0 * color_diff_down0, error_weight);
199 			float error_down1 = dot_s(color_diff_down1 * color_diff_down1, error_weight);
200 			float error_down2 = dot_s(color_diff_down2 * color_diff_down2, error_weight);
201 			float error_down3 = dot_s(color_diff_down3 * color_diff_down3, error_weight);
202 
203 			float error_up0 = dot_s(color_diff_up0 * color_diff_up0, error_weight);
204 			float error_up1 = dot_s(color_diff_up1 * color_diff_up1, error_weight);
205 			float error_up2 = dot_s(color_diff_up2 * color_diff_up2, error_weight);
206 			float error_up3 = dot_s(color_diff_up3 * color_diff_up3, error_weight);
207 
208 			vfloat4 error_base_vec = vfloat4(error_base0, error_base1, error_base2, error_base3);
209 			vfloat4 error_down_vec = vfloat4(error_down0, error_down1, error_down2, error_down3);
210 			vfloat4 error_up_vec = vfloat4(error_up0, error_up1, error_up2, error_up3);
211 
212 			vmask4 check_result_up = (error_up_vec < error_base_vec) &
213 			        (error_up_vec < error_down_vec) & (uqw_vec < vint4(64));
214 
215 			vmask4 check_result_down = (error_down_vec < error_base_vec) & (uqw_vec > vint4::zero());
216 			check_result_down = check_result_down & (~check_result_up);
217 
218 			if (popcount(check_result_up | check_result_down) != 0)
219 			{
220 				uqw_vec = select(uqw_vec, uqw_up_vec, check_result_up);
221 				uqw_vec = select(uqw_vec, uqw_down_vec, check_result_down);
222 
223 				dec_weights_uquant[texel] = uqw_vec.lane<0>();
224 				dec_weights_uquant[texel + 1] = uqw_vec.lane<1>();
225 				dec_weights_uquant[texel + 2] = uqw_vec.lane<2>();    // channel 2
226 				dec_weights_uquant[texel + 3] = uqw_vec.lane<3>();    // channel 3
227 				adjustments = true;
228 			}
229 		};
230 
231 		for (; texel < bsd.texel_count; texel++)
232 		{
233 			int uqw = dec_weights_uquant[texel];
234 
235 			uint32_t prev_and_next = qat.prev_next_values[uqw];
236 			int uqw_down = prev_and_next & 0xFF;
237 			int uqw_up = (prev_and_next >> 8) & 0xFF;
238 
239 			// Interpolate the colors to create the diffs
240 			float weight_base = static_cast<float>(uqw);
241 			float weight_down = static_cast<float>(uqw_down - uqw);
242 			float weight_up = static_cast<float>(uqw_up - uqw);
243 
244 			unsigned int partition = pi.partition_of_texel[texel];
245 			vfloat4 color_offset = offset[partition];
246 			vfloat4 color_base   = endpnt0f[partition];
247 
248 			vfloat4 color = color_base + color_offset * weight_base;
249 			vfloat4 orig_color   = blk.texel(texel);
250 			vfloat4 error_weight = blk.channel_weight;
251 
252 			vfloat4 color_diff      = color - orig_color;
253 			vfloat4 color_diff_down = color_diff + color_offset * weight_down;
254 			vfloat4 color_diff_up   = color_diff + color_offset * weight_up;
255 
256 			float error_base = dot_s(color_diff      * color_diff,      error_weight);
257 			float error_down = dot_s(color_diff_down * color_diff_down, error_weight);
258 			float error_up   = dot_s(color_diff_up   * color_diff_up,   error_weight);
259 
260 			// Check if the prev or next error is better, and if so use it
261 			if ((error_up < error_base) && (error_up < error_down) && (uqw < 64))
262 			{
263 				dec_weights_uquant[texel] = static_cast<uint8_t>(uqw_up);
264 				adjustments = true;
265 			}
266 			else if ((error_down < error_base) && (uqw > 0))
267 			{
268 				dec_weights_uquant[texel] = static_cast<uint8_t>(uqw_down);
269 				adjustments = true;
270 			}
271 		}
272 
273 		// Prepare iteration for plane 2
274 		dec_weights_uquant += WEIGHTS_PLANE2_OFFSET;
275 		plane_mask = ~plane_mask;
276 	}
277 	return adjustments;
278 }
279 #else
realign_weights_undecimated( astcenc_profile decode_mode, const block_size_descriptor& bsd, const image_block& blk, symbolic_compressed_block& scb )280 static bool realign_weights_undecimated(
281 	astcenc_profile decode_mode,
282 	const block_size_descriptor& bsd,
283 	const image_block& blk,
284 	symbolic_compressed_block& scb
285 ) {
286 	// Get the partition descriptor
287 	unsigned int partition_count = scb.partition_count;
288 	const auto& pi = bsd.get_partition_info(partition_count, scb.partition_index);
289 
290 	// Get the quantization table
291 	const block_mode& bm = bsd.get_block_mode(scb.block_mode);
292 	unsigned int weight_quant_level = bm.quant_mode;
293 	const quant_and_transfer_table& qat = quant_and_xfer_tables[weight_quant_level];
294 
295 	unsigned int max_plane = bm.is_dual_plane;
296 	int plane2_component = scb.plane2_component;
297 	vmask4 plane_mask = vint4::lane_id() == vint4(plane2_component);
298 
299 	// Decode the color endpoints
300 	bool rgb_hdr;
301 	bool alpha_hdr;
302 	vint4 endpnt0[BLOCK_MAX_PARTITIONS];
303 	vint4 endpnt1[BLOCK_MAX_PARTITIONS];
304 	vfloat4 endpnt0f[BLOCK_MAX_PARTITIONS];
305 	vfloat4 offset[BLOCK_MAX_PARTITIONS];
306 
307 	promise(partition_count > 0);
308 
309 	for (unsigned int pa_idx = 0; pa_idx < partition_count; pa_idx++)
310 	{
311 		unpack_color_endpoints(decode_mode,
312 		                       scb.color_formats[pa_idx],
313 		                       scb.color_values[pa_idx],
314 		                       rgb_hdr, alpha_hdr,
315 		                       endpnt0[pa_idx],
316 		                       endpnt1[pa_idx]);
317 	}
318 
319 	uint8_t* dec_weights_uquant = scb.weights;
320 	bool adjustments = false;
321 
322 	// For each plane and partition ...
323 	for (unsigned int pl_idx = 0; pl_idx <= max_plane; pl_idx++)
324 	{
325 		for (unsigned int pa_idx = 0; pa_idx < partition_count; pa_idx++)
326 		{
327 			// Compute the endpoint delta for all components in current plane
328 			vint4 epd = endpnt1[pa_idx] - endpnt0[pa_idx];
329 			epd = select(epd, vint4::zero(), plane_mask);
330 
331 			endpnt0f[pa_idx] = int_to_float(endpnt0[pa_idx]);
332 			offset[pa_idx] = int_to_float(epd) * (1.0f / 64.0f);
333 		}
334 
335 		// For each weight compute previous, current, and next errors
336 		promise(bsd.texel_count > 0);
337 		for (unsigned int texel = 0; texel < bsd.texel_count; texel++)
338 		{
339 			int uqw = dec_weights_uquant[texel];
340 
341 			uint32_t prev_and_next = qat.prev_next_values[uqw];
342 			int uqw_down = prev_and_next & 0xFF;
343 			int uqw_up = (prev_and_next >> 8) & 0xFF;
344 
345 			// Interpolate the colors to create the diffs
346 			float weight_base = static_cast<float>(uqw);
347 			float weight_down = static_cast<float>(uqw_down - uqw);
348 			float weight_up = static_cast<float>(uqw_up - uqw);
349 
350 			unsigned int partition = pi.partition_of_texel[texel];
351 			vfloat4 color_offset = offset[partition];
352 			vfloat4 color_base   = endpnt0f[partition];
353 
354 			vfloat4 color = color_base + color_offset * weight_base;
355 			vfloat4 orig_color   = blk.texel(texel);
356 			vfloat4 error_weight = blk.channel_weight;
357 
358 			vfloat4 color_diff      = color - orig_color;
359 			vfloat4 color_diff_down = color_diff + color_offset * weight_down;
360 			vfloat4 color_diff_up   = color_diff + color_offset * weight_up;
361 
362 			float error_base = dot_s(color_diff      * color_diff,      error_weight);
363 			float error_down = dot_s(color_diff_down * color_diff_down, error_weight);
364 			float error_up   = dot_s(color_diff_up   * color_diff_up,   error_weight);
365 
366 			// Check if the prev or next error is better, and if so use it
367 			if ((error_up < error_base) && (error_up < error_down) && (uqw < 64))
368 			{
369 				dec_weights_uquant[texel] = static_cast<uint8_t>(uqw_up);
370 				adjustments = true;
371 			}
372 			else if ((error_down < error_base) && (uqw > 0))
373 			{
374 				dec_weights_uquant[texel] = static_cast<uint8_t>(uqw_down);
375 				adjustments = true;
376 			}
377 		}
378 
379 		// Prepare iteration for plane 2
380 		dec_weights_uquant += WEIGHTS_PLANE2_OFFSET;
381 		plane_mask = ~plane_mask;
382 	}
383 
384 	return adjustments;
385 }
386 #endif
387 
388 /**
389  * @brief Attempt to improve weights given a chosen configuration.
390  *
391  * Given a fixed weight grid decimation and weight value quantization, iterate over all weights (per
392  * partition and per plane) and attempt to improve image quality by moving each weight up by one or
393  * down by one quantization step.
394  *
395  * @param      decode_mode   The decode mode (LDR, HDR).
396  * @param      bsd           The block size information.
397  * @param      blk           The image block color data to compress.
398  * @param[out] scb           The symbolic compressed block output.
399  */
realign_weights_decimated( astcenc_profile decode_mode, const block_size_descriptor& bsd, const image_block& blk, symbolic_compressed_block& scb )400 static bool realign_weights_decimated(
401 	astcenc_profile decode_mode,
402 	const block_size_descriptor& bsd,
403 	const image_block& blk,
404 	symbolic_compressed_block& scb
405 ) {
406 	// Get the partition descriptor
407 	unsigned int partition_count = scb.partition_count;
408 	const auto& pi = bsd.get_partition_info(partition_count, scb.partition_index);
409 
410 	// Get the quantization table
411 	const block_mode& bm = bsd.get_block_mode(scb.block_mode);
412 	unsigned int weight_quant_level = bm.quant_mode;
413 	const quant_and_transfer_table& qat = quant_and_xfer_tables[weight_quant_level];
414 
415 	// Get the decimation table
416 	const decimation_info& di = bsd.get_decimation_info(bm.decimation_mode);
417 	unsigned int weight_count = di.weight_count;
418 	assert(weight_count != bsd.texel_count);
419 
420 	unsigned int max_plane = bm.is_dual_plane;
421 	int plane2_component = scb.plane2_component;
422 	vmask4 plane_mask = vint4::lane_id() == vint4(plane2_component);
423 
424 	// Decode the color endpoints
425 	bool rgb_hdr;
426 	bool alpha_hdr;
427 	vint4 endpnt0[BLOCK_MAX_PARTITIONS];
428 	vint4 endpnt1[BLOCK_MAX_PARTITIONS];
429 	vfloat4 endpnt0f[BLOCK_MAX_PARTITIONS];
430 	vfloat4 offset[BLOCK_MAX_PARTITIONS];
431 
432 	promise(partition_count > 0);
433 	promise(weight_count > 0);
434 
435 	for (unsigned int pa_idx = 0; pa_idx < partition_count; pa_idx++)
436 	{
437 		unpack_color_endpoints(decode_mode,
438 		                       scb.color_formats[pa_idx],
439 		                       scb.color_values[pa_idx],
440 		                       rgb_hdr, alpha_hdr,
441 		                       endpnt0[pa_idx],
442 		                       endpnt1[pa_idx]);
443 	}
444 
445 	uint8_t* dec_weights_uquant = scb.weights;
446 	bool adjustments = false;
447 
448 	// For each plane and partition ...
449 	for (unsigned int pl_idx = 0; pl_idx <= max_plane; pl_idx++)
450 	{
451 		for (unsigned int pa_idx = 0; pa_idx < partition_count; pa_idx++)
452 		{
453 			// Compute the endpoint delta for all components in current plane
454 			vint4 epd = endpnt1[pa_idx] - endpnt0[pa_idx];
455 			epd = select(epd, vint4::zero(), plane_mask);
456 
457 			endpnt0f[pa_idx] = int_to_float(endpnt0[pa_idx]);
458 			offset[pa_idx] = int_to_float(epd) * (1.0f / 64.0f);
459 		}
460 
461 		// Create an unquantized weight grid for this decimation level
462 		ASTCENC_ALIGNAS float uq_weightsf[BLOCK_MAX_WEIGHTS];
463 		for (unsigned int we_idx = 0; we_idx < weight_count; we_idx += ASTCENC_SIMD_WIDTH)
464 		{
465 			vint unquant_value(dec_weights_uquant + we_idx);
466 			vfloat unquant_valuef = int_to_float(unquant_value);
467 			storea(unquant_valuef, uq_weightsf + we_idx);
468 		}
469 
470 		// For each weight compute previous, current, and next errors
471 		for (unsigned int we_idx = 0; we_idx < weight_count; we_idx++)
472 		{
473 			int uqw = dec_weights_uquant[we_idx];
474 			uint32_t prev_and_next = qat.prev_next_values[uqw];
475 
476 			float uqw_base = uq_weightsf[we_idx];
477 			float uqw_down = static_cast<float>(prev_and_next & 0xFF);
478 			float uqw_up = static_cast<float>((prev_and_next >> 8) & 0xFF);
479 
480 			float uqw_diff_down = uqw_down - uqw_base;
481 			float uqw_diff_up = uqw_up - uqw_base;
482 
483 			vfloat4 error_basev = vfloat4::zero();
484 			vfloat4 error_downv = vfloat4::zero();
485 			vfloat4 error_upv = vfloat4::zero();
486 
487 			// Interpolate the colors to create the diffs
488 			unsigned int texels_to_evaluate = di.weight_texel_count[we_idx];
489 			promise(texels_to_evaluate > 0);
490 			for (unsigned int te_idx = 0; te_idx < texels_to_evaluate; te_idx++)
491 			{
492 				unsigned int texel = di.weight_texels_tr[te_idx][we_idx];
493 
494 				float tw_base = di.texel_contrib_for_weight[te_idx][we_idx];
495 
496 				float weight_base = (uq_weightsf[di.texel_weights_tr[0][texel]] * di.texel_weight_contribs_float_tr[0][texel]
497 				                   + uq_weightsf[di.texel_weights_tr[1][texel]] * di.texel_weight_contribs_float_tr[1][texel])
498 					              + (uq_weightsf[di.texel_weights_tr[2][texel]] * di.texel_weight_contribs_float_tr[2][texel]
499 				                   + uq_weightsf[di.texel_weights_tr[3][texel]] * di.texel_weight_contribs_float_tr[3][texel]);
500 
501 				// Ideally this is integer rounded, but IQ gain it isn't worth the overhead
502 				// float weight = astc::flt_rd(weight_base + 0.5f);
503 				// float weight_down = astc::flt_rd(weight_base + 0.5f + uqw_diff_down * tw_base) - weight;
504 				// float weight_up = astc::flt_rd(weight_base + 0.5f + uqw_diff_up * tw_base) - weight;
505 				float weight_down = weight_base + uqw_diff_down * tw_base - weight_base;
506 				float weight_up = weight_base + uqw_diff_up * tw_base - weight_base;
507 
508 				unsigned int partition = pi.partition_of_texel[texel];
509 				vfloat4 color_offset = offset[partition];
510 				vfloat4 color_base   = endpnt0f[partition];
511 
512 				vfloat4 color = color_base + color_offset * weight_base;
513 				vfloat4 orig_color = blk.texel(texel);
514 
515 				vfloat4 color_diff      = color - orig_color;
516 				vfloat4 color_down_diff = color_diff + color_offset * weight_down;
517 				vfloat4 color_up_diff   = color_diff + color_offset * weight_up;
518 
519 				error_basev += color_diff * color_diff;
520 				error_downv += color_down_diff * color_down_diff;
521 				error_upv   += color_up_diff * color_up_diff;
522 			}
523 
524 			vfloat4 error_weight = blk.channel_weight;
525 			float error_base = hadd_s(error_basev * error_weight);
526 			float error_down = hadd_s(error_downv * error_weight);
527 			float error_up   = hadd_s(error_upv   * error_weight);
528 
529 			// Check if the prev or next error is better, and if so use it
530 			if ((error_up < error_base) && (error_up < error_down) && (uqw < 64))
531 			{
532 				uq_weightsf[we_idx] = uqw_up;
533 				dec_weights_uquant[we_idx] = static_cast<uint8_t>(uqw_up);
534 				adjustments = true;
535 			}
536 			else if ((error_down < error_base) && (uqw > 0))
537 			{
538 				uq_weightsf[we_idx] = uqw_down;
539 				dec_weights_uquant[we_idx] = static_cast<uint8_t>(uqw_down);
540 				adjustments = true;
541 			}
542 		}
543 
544 		// Prepare iteration for plane 2
545 		dec_weights_uquant += WEIGHTS_PLANE2_OFFSET;
546 		plane_mask = ~plane_mask;
547 	}
548 
549 	return adjustments;
550 }
551 
552 /**
553  * @brief Compress a block using a chosen partitioning and 1 plane of weights.
554  *
555  * @param      config                    The compressor configuration.
556  * @param      bsd                       The block size information.
557  * @param      blk                       The image block color data to compress.
558  * @param      only_always               True if we only use "always" percentile block modes.
559  * @param      tune_errorval_threshold   The error value threshold.
560  * @param      partition_count           The partition count.
561  * @param      partition_index           The partition index if @c partition_count is 2-4.
562  * @param[out] scb                       The symbolic compressed block output.
563  * @param[out] tmpbuf                    The quantized weights for plane 1.
564  */
compress_symbolic_block_for_partition_1plane( QualityProfile privateProfile, const astcenc_config& config, const block_size_descriptor& bsd, const image_block& blk, bool only_always, float tune_errorval_threshold, unsigned int partition_count, unsigned int partition_index, symbolic_compressed_block& scb, compression_working_buffers& tmpbuf, int quant_limit )565 static float compress_symbolic_block_for_partition_1plane(
566 	QualityProfile privateProfile,
567 	const astcenc_config& config,
568 	const block_size_descriptor& bsd,
569 	const image_block& blk,
570 	bool only_always,
571 	float tune_errorval_threshold,
572 	unsigned int partition_count,
573 	unsigned int partition_index,
574 	symbolic_compressed_block& scb,
575 	compression_working_buffers& tmpbuf,
576 	int quant_limit
577 ) {
578 	promise(partition_count > 0);
579 	promise(config.tune_candidate_limit > 0);
580 	promise(config.tune_refinement_limit > 0);
581 
582 	int max_weight_quant = astc::min(static_cast<int>(QUANT_32), quant_limit);
583 
584 	auto compute_difference = &compute_symbolic_block_difference_1plane;
585 	if ((partition_count == 1) && !(config.flags & ASTCENC_FLG_MAP_RGBM))
586 	{
587 		compute_difference = &compute_symbolic_block_difference_1plane_1partition;
588 	}
589 
590 	const auto& pi = bsd.get_partition_info(partition_count, partition_index);
591 
592 	// Compute ideal weights and endpoint colors, with no quantization or decimation
593 	endpoints_and_weights& ei = tmpbuf.ei1;
594 	compute_ideal_colors_and_weights_1plane(blk, pi, ei);
595 
596 	// Compute ideal weights and endpoint colors for every decimation
597 	float* dec_weights_ideal = tmpbuf.dec_weights_ideal;
598 	uint8_t* dec_weights_uquant = tmpbuf.dec_weights_uquant;
599 
600 	// For each decimation mode, compute an ideal set of weights with no quantization
601 	unsigned int max_decimation_modes = only_always ? bsd.decimation_mode_count_always
602 	                                                : bsd.decimation_mode_count_selected;
603 	promise(max_decimation_modes > 0);
604 	for (unsigned int i = 0; i < max_decimation_modes; i++)
605 	{
606 		const auto& dm = bsd.get_decimation_mode(i);
607 		if (!dm.is_ref_1plane(static_cast<quant_method>(max_weight_quant)))
608 		{
609 			continue;
610 		}
611 
612 		const auto& di = bsd.get_decimation_info(i);
613 
614 		compute_ideal_weights_for_decimation(
615 		    ei,
616 		    di,
617 		    dec_weights_ideal + i * BLOCK_MAX_WEIGHTS);
618 	}
619 
620 	// Compute maximum colors for the endpoints and ideal weights, then for each endpoint and ideal
621 	// weight pair, compute the smallest weight that will result in a color value greater than 1
622 	vfloat4 min_ep(10.0f);
623 	for (unsigned int i = 0; i < partition_count; i++)
624 	{
625 		vfloat4 ep = (vfloat4(1.0f) - ei.ep.endpt0[i]) / (ei.ep.endpt1[i] - ei.ep.endpt0[i]);
626 
627 		vmask4 use_ep = (ep > vfloat4(0.5f)) & (ep < min_ep);
628 		min_ep = select(min_ep, ep, use_ep);
629 	}
630 
631 	float min_wt_cutoff = hmin_s(min_ep);
632 
633 	// For each mode, use the angular method to compute a shift
634 	compute_angular_endpoints_1plane(
635 	    privateProfile, only_always, bsd, dec_weights_ideal, max_weight_quant, tmpbuf);
636 
637 	float* weight_low_value = tmpbuf.weight_low_value1;
638 	float* weight_high_value = tmpbuf.weight_high_value1;
639 	int8_t* qwt_bitcounts = tmpbuf.qwt_bitcounts;
640 	float* qwt_errors = tmpbuf.qwt_errors;
641 
642 	// For each mode (which specifies a decimation and a quantization):
643 	//     * Compute number of bits needed for the quantized weights
644 	//     * Generate an optimized set of quantized weights
645 	//     * Compute quantization errors for the mode
646 
647 
648 	static const int8_t free_bits_for_partition_count[4] {
649 		115 - 4, 111 - 4 - PARTITION_INDEX_BITS, 108 - 4 - PARTITION_INDEX_BITS, 105 - 4 - PARTITION_INDEX_BITS
650 	};
651 
652 	unsigned int max_block_modes = only_always ? bsd.block_mode_count_1plane_always
653 	                                           : bsd.block_mode_count_1plane_selected;
654 	promise(max_block_modes > 0);
655 	for (unsigned int i = 0; i < max_block_modes; i++)
656 	{
657 		const block_mode& bm = bsd.block_modes[i];
658 
659 		if (bm.quant_mode > max_weight_quant)
660 		{
661 			qwt_errors[i] = 1e38f;
662 			continue;
663 		}
664 
665 		assert(!bm.is_dual_plane);
666 		int bitcount = free_bits_for_partition_count[partition_count - 1] - bm.weight_bits;
667 		if (bitcount <= 0)
668 		{
669 			qwt_errors[i] = 1e38f;
670 			continue;
671 		}
672 
673 		if (weight_high_value[i] > 1.02f * min_wt_cutoff)
674 		{
675 			weight_high_value[i] = 1.0f;
676 		}
677 
678 		int decimation_mode = bm.decimation_mode;
679 		const auto& di = bsd.get_decimation_info(decimation_mode);
680 
681 		qwt_bitcounts[i] = static_cast<int8_t>(bitcount);
682 
683 		ASTCENC_ALIGNAS float dec_weights_uquantf[BLOCK_MAX_WEIGHTS];
684 
685 		// Generate the optimized set of weights for the weight mode
686 		compute_quantized_weights_for_decimation(
687 		    di,
688 		    weight_low_value[i], weight_high_value[i],
689 		    dec_weights_ideal + BLOCK_MAX_WEIGHTS * decimation_mode,
690 		    dec_weights_uquantf,
691 		    dec_weights_uquant + BLOCK_MAX_WEIGHTS * i,
692 		    bm.get_weight_quant_mode());
693 
694 		// Compute weight quantization errors for the block mode
695 		qwt_errors[i] = compute_error_of_weight_set_1plane(
696 		    ei,
697 		    di,
698 		    dec_weights_uquantf);
699 	}
700 
701 	// Decide the optimal combination of color endpoint encodings and weight encodings
702 	uint8_t partition_format_specifiers[TUNE_MAX_TRIAL_CANDIDATES][BLOCK_MAX_PARTITIONS];
703 	int block_mode_index[TUNE_MAX_TRIAL_CANDIDATES];
704 
705 	quant_method color_quant_level[TUNE_MAX_TRIAL_CANDIDATES];
706 	quant_method color_quant_level_mod[TUNE_MAX_TRIAL_CANDIDATES];
707 
708 	unsigned int candidate_count = compute_ideal_endpoint_formats(
709 	    privateProfile,
710 	    pi, blk, ei.ep, qwt_bitcounts, qwt_errors,
711 	    config.tune_candidate_limit, 0, max_block_modes,
712 	    partition_format_specifiers, block_mode_index,
713 	    color_quant_level, color_quant_level_mod, tmpbuf);
714 
715 	// Iterate over the N believed-to-be-best modes to find out which one is actually best
716 	float best_errorval_in_mode = ERROR_CALC_DEFAULT;
717 	float best_errorval_in_scb = scb.errorval;
718 
719 	for (unsigned int i = 0; i < candidate_count; i++)
720 	{
721 		TRACE_NODE(node0, "candidate");
722 
723 		const int bm_packed_index = block_mode_index[i];
724 		assert(bm_packed_index >= 0 && bm_packed_index < static_cast<int>(bsd.block_mode_count_1plane_selected));
725 		const block_mode& qw_bm = bsd.block_modes[bm_packed_index];
726 
727 		int decimation_mode = qw_bm.decimation_mode;
728 		const auto& di = bsd.get_decimation_info(decimation_mode);
729 		promise(di.weight_count > 0);
730 
731 		trace_add_data("weight_x", di.weight_x);
732 		trace_add_data("weight_y", di.weight_y);
733 		trace_add_data("weight_z", di.weight_z);
734 		trace_add_data("weight_quant", qw_bm.quant_mode);
735 
736 		// Recompute the ideal color endpoints before storing them
737 		vfloat4 rgbs_colors[BLOCK_MAX_PARTITIONS];
738 		vfloat4 rgbo_colors[BLOCK_MAX_PARTITIONS];
739 
740 		symbolic_compressed_block workscb;
741 		endpoints workep = ei.ep;
742 
743 		uint8_t* u8_weight_src = dec_weights_uquant + BLOCK_MAX_WEIGHTS * bm_packed_index;
744 
745 		for (unsigned int j = 0; j < di.weight_count; j++)
746 		{
747 			workscb.weights[j] = u8_weight_src[j];
748 		}
749 
750 		for (unsigned int l = 0; l < config.tune_refinement_limit; l++)
751 		{
752 			recompute_ideal_colors_1plane(
753 			    blk, pi, di, workscb.weights,
754 			    workep, rgbs_colors, rgbo_colors);
755 
756 			// Quantize the chosen color, tracking if worth trying the mod value
757 			bool all_same = color_quant_level[i] != color_quant_level_mod[i];
758 			for (unsigned int j = 0; j < partition_count; j++)
759 			{
760 				workscb.color_formats[j] = pack_color_endpoints(
761 				    privateProfile,
762 				    workep.endpt0[j],
763 				    workep.endpt1[j],
764 				    rgbs_colors[j],
765 				    rgbo_colors[j],
766 				    partition_format_specifiers[i][j],
767 				    workscb.color_values[j],
768 				    color_quant_level[i]);
769 
770 				all_same = all_same && workscb.color_formats[j] == workscb.color_formats[0];
771 			}
772 
773 			// If all the color endpoint modes are the same, we get a few more bits to store colors;
774 			// let's see if we can take advantage of this: requantize all the colors and see if the
775 			// endpoint modes remain the same.
776 			workscb.color_formats_matched = 0;
777 			if (partition_count >= 2 && all_same)
778 			{
779 				uint8_t colorvals[BLOCK_MAX_PARTITIONS][8];
780 				uint8_t color_formats_mod[BLOCK_MAX_PARTITIONS] { 0 };
781 				bool all_same_mod = true;
782 				for (unsigned int j = 0; j < partition_count; j++)
783 				{
784 					color_formats_mod[j] = pack_color_endpoints(
785 					    privateProfile,
786 					    workep.endpt0[j],
787 					    workep.endpt1[j],
788 					    rgbs_colors[j],
789 					    rgbo_colors[j],
790 					    partition_format_specifiers[i][j],
791 					    colorvals[j],
792 					    color_quant_level_mod[i]);
793 
794 					// Early out as soon as it's no longer possible to use mod
795 					if (color_formats_mod[j] != color_formats_mod[0])
796 					{
797 						all_same_mod = false;
798 						break;
799 					}
800 				}
801 
802 				if (all_same_mod)
803 				{
804 					workscb.color_formats_matched = 1;
805 					for (unsigned int j = 0; j < BLOCK_MAX_PARTITIONS; j++)
806 					{
807 						for (unsigned int k = 0; k < 8; k++)
808 						{
809 							workscb.color_values[j][k] = colorvals[j][k];
810 						}
811 
812 						workscb.color_formats[j] = color_formats_mod[j];
813 					}
814 				}
815 			}
816 
817 			// Store header fields
818 			workscb.partition_count = static_cast<uint8_t>(partition_count);
819 			workscb.partition_index = static_cast<uint16_t>(partition_index);
820 			workscb.plane2_component = -1;
821 			workscb.quant_mode = workscb.color_formats_matched ? color_quant_level_mod[i] : color_quant_level[i];
822 			workscb.block_mode = qw_bm.mode_index;
823 			workscb.block_type = SYM_BTYPE_NONCONST;
824 			if (privateProfile == HIGH_SPEED_PROFILE)
825 			{
826 				workscb.errorval = 0;
827 				scb = workscb;
828 				break;
829 			}
830 			// Pre-realign test
831 			if (l == 0)
832 			{
833 				float errorval = compute_difference(config, bsd, workscb, blk);
834 				if (errorval == -ERROR_CALC_DEFAULT)
835 				{
836 					errorval = -errorval;
837 					workscb.block_type = SYM_BTYPE_ERROR;
838 				}
839 
840 				trace_add_data("error_prerealign", errorval);
841 				best_errorval_in_mode = astc::min(errorval, best_errorval_in_mode);
842 
843 				// Average refinement improvement is 3.5% per iteration (allow 4.5%), but the first
844 				// iteration can help more so we give it a extra 8% leeway. Use this knowledge to
845 				// drive a heuristic to skip blocks that are unlikely to catch up with the best
846 				// block we have already.
847 				unsigned int iters_remaining = config.tune_refinement_limit - l;
848 				float threshold = (0.045f * static_cast<float>(iters_remaining)) + 1.08f;
849 				if (errorval > (threshold * best_errorval_in_scb))
850 				{
851 					break;
852 				}
853 
854 				if (errorval < best_errorval_in_scb)
855 				{
856 					best_errorval_in_scb = errorval;
857 					workscb.errorval = errorval;
858 					scb = workscb;
859 
860 					if (errorval < tune_errorval_threshold)
861 					{
862 						// Skip remaining candidates - this is "good enough"
863 						i = candidate_count;
864 						break;
865 					}
866 				}
867 			}
868 
869 			bool adjustments;
870 			if (di.weight_count != bsd.texel_count)
871 			{
872 				adjustments = realign_weights_decimated(
873 					config.profile, bsd, blk, workscb);
874 			}
875 			else
876 			{
877 				adjustments = realign_weights_undecimated(
878 					config.profile, bsd, blk, workscb);
879 			}
880 
881 			// Post-realign test
882 			float errorval = compute_difference(config, bsd, workscb, blk);
883 			if (errorval == -ERROR_CALC_DEFAULT)
884 			{
885 				errorval = -errorval;
886 				workscb.block_type = SYM_BTYPE_ERROR;
887 			}
888 
889 			trace_add_data("error_postrealign", errorval);
890 			best_errorval_in_mode = astc::min(errorval, best_errorval_in_mode);
891 
892 			// Average refinement improvement is 3.5% per iteration, so skip blocks that are
893 			// unlikely to catch up with the best block we have already. Assume a 4.5% per step to
894 			// give benefit of the doubt ...
895 			unsigned int iters_remaining = config.tune_refinement_limit - 1 - l;
896 			float threshold = (0.045f * static_cast<float>(iters_remaining)) + 1.0f;
897 			if (errorval > (threshold * best_errorval_in_scb))
898 			{
899 				break;
900 			}
901 
902 			if (errorval < best_errorval_in_scb)
903 			{
904 				best_errorval_in_scb = errorval;
905 				workscb.errorval = errorval;
906 				scb = workscb;
907 
908 				if (errorval < tune_errorval_threshold)
909 				{
910 					// Skip remaining candidates - this is "good enough"
911 					i = candidate_count;
912 					break;
913 				}
914 			}
915 
916 			if (!adjustments)
917 			{
918 				break;
919 			}
920 		}
921 	}
922 
923 	return best_errorval_in_mode;
924 }
925 
926 /**
927  * @brief Compress a block using a chosen partitioning and 2 planes of weights.
928  *
929  * @param      config                    The compressor configuration.
930  * @param      bsd                       The block size information.
931  * @param      blk                       The image block color data to compress.
932  * @param      tune_errorval_threshold   The error value threshold.
933  * @param      plane2_component          The component index for the second plane of weights.
934  * @param[out] scb                       The symbolic compressed block output.
935  * @param[out] tmpbuf                    The quantized weights for plane 1.
936  */
compress_symbolic_block_for_partition_2planes( QualityProfile privateProfile, const astcenc_config& config, const block_size_descriptor& bsd, const image_block& blk, float tune_errorval_threshold, unsigned int plane2_component, symbolic_compressed_block& scb, compression_working_buffers& tmpbuf, int quant_limit )937 static float compress_symbolic_block_for_partition_2planes(
938 	QualityProfile privateProfile,
939 	const astcenc_config& config,
940 	const block_size_descriptor& bsd,
941 	const image_block& blk,
942 	float tune_errorval_threshold,
943 	unsigned int plane2_component,
944 	symbolic_compressed_block& scb,
945 	compression_working_buffers& tmpbuf,
946 	int quant_limit
947 ) {
948 	promise(config.tune_candidate_limit > 0);
949 	promise(config.tune_refinement_limit > 0);
950 	promise(bsd.decimation_mode_count_selected > 0);
951 
952 	int max_weight_quant = astc::min(static_cast<int>(QUANT_32), quant_limit);
953 
954 	// Compute ideal weights and endpoint colors, with no quantization or decimation
955 	endpoints_and_weights& ei1 = tmpbuf.ei1;
956 	endpoints_and_weights& ei2 = tmpbuf.ei2;
957 
958 	compute_ideal_colors_and_weights_2planes(bsd, blk, plane2_component, ei1, ei2);
959 
960 	// Compute ideal weights and endpoint colors for every decimation
961 	float* dec_weights_ideal = tmpbuf.dec_weights_ideal;
962 	uint8_t* dec_weights_uquant = tmpbuf.dec_weights_uquant;
963 
964 	// For each decimation mode, compute an ideal set of weights with no quantization
965 	for (unsigned int i = 0; i < bsd.decimation_mode_count_selected; i++)
966 	{
967 		const auto& dm = bsd.get_decimation_mode(i);
968 		if (!dm.is_ref_2plane(static_cast<quant_method>(max_weight_quant)))
969 		{
970 			continue;
971 		}
972 
973 		const auto& di = bsd.get_decimation_info(i);
974 
975 		compute_ideal_weights_for_decimation(
976 		    ei1,
977 		    di,
978 		    dec_weights_ideal + i * BLOCK_MAX_WEIGHTS);
979 
980 		compute_ideal_weights_for_decimation(
981 		    ei2,
982 		    di,
983 		    dec_weights_ideal + i * BLOCK_MAX_WEIGHTS + WEIGHTS_PLANE2_OFFSET);
984 	}
985 
986 	// Compute maximum colors for the endpoints and ideal weights, then for each endpoint and ideal
987 	// weight pair, compute the smallest weight that will result in a color value greater than 1
988 	vfloat4 min_ep1(10.0f);
989 	vfloat4 min_ep2(10.0f);
990 
991 	vfloat4 ep1 = (vfloat4(1.0f) - ei1.ep.endpt0[0]) / (ei1.ep.endpt1[0] - ei1.ep.endpt0[0]);
992 	vmask4 use_ep1 = (ep1 > vfloat4(0.5f)) & (ep1 < min_ep1);
993 	min_ep1 = select(min_ep1, ep1, use_ep1);
994 
995 	vfloat4 ep2 = (vfloat4(1.0f) - ei2.ep.endpt0[0]) / (ei2.ep.endpt1[0] - ei2.ep.endpt0[0]);
996 	vmask4 use_ep2 = (ep2 > vfloat4(0.5f)) & (ep2 < min_ep2);
997 	min_ep2 = select(min_ep2, ep2, use_ep2);
998 
999 	vfloat4 err_max(ERROR_CALC_DEFAULT);
1000 	vmask4 err_mask = vint4::lane_id() == vint4(plane2_component);
1001 
1002 	// Set the plane2 component to max error in ep1
1003 	min_ep1 = select(min_ep1, err_max, err_mask);
1004 
1005 	float min_wt_cutoff1 = hmin_s(min_ep1);
1006 
1007 	// Set the minwt2 to the plane2 component min in ep2
1008 	float min_wt_cutoff2 = hmin_s(select(err_max, min_ep2, err_mask));
1009 
1010 	compute_angular_endpoints_2planes(
1011 	    privateProfile, bsd, dec_weights_ideal, max_weight_quant, tmpbuf);
1012 
1013 	// For each mode (which specifies a decimation and a quantization):
1014 	//     * Compute number of bits needed for the quantized weights
1015 	//     * Generate an optimized set of quantized weights
1016 	//     * Compute quantization errors for the mode
1017 
1018 	float* weight_low_value1 = tmpbuf.weight_low_value1;
1019 	float* weight_high_value1 = tmpbuf.weight_high_value1;
1020 	float* weight_low_value2 = tmpbuf.weight_low_value2;
1021 	float* weight_high_value2 = tmpbuf.weight_high_value2;
1022 
1023 	int8_t* qwt_bitcounts = tmpbuf.qwt_bitcounts;
1024 	float* qwt_errors = tmpbuf.qwt_errors;
1025 
1026 	unsigned int start_2plane = bsd.block_mode_count_1plane_selected;
1027 	unsigned int end_2plane = bsd.block_mode_count_1plane_2plane_selected;
1028 
1029 	for (unsigned int i = start_2plane; i < end_2plane; i++)
1030 	{
1031 		const block_mode& bm = bsd.block_modes[i];
1032 		assert(bm.is_dual_plane);
1033 
1034 		if (bm.quant_mode > max_weight_quant)
1035 		{
1036 			qwt_errors[i] = 1e38f;
1037 			continue;
1038 		}
1039 
1040 		qwt_bitcounts[i] = static_cast<int8_t>(109 - bm.weight_bits);
1041 
1042 		if (weight_high_value1[i] > 1.02f * min_wt_cutoff1)
1043 		{
1044 			weight_high_value1[i] = 1.0f;
1045 		}
1046 
1047 		if (weight_high_value2[i] > 1.02f * min_wt_cutoff2)
1048 		{
1049 			weight_high_value2[i] = 1.0f;
1050 		}
1051 
1052 		unsigned int decimation_mode = bm.decimation_mode;
1053 		const auto& di = bsd.get_decimation_info(decimation_mode);
1054 
1055 		ASTCENC_ALIGNAS float dec_weights_uquantf[BLOCK_MAX_WEIGHTS];
1056 
1057 		// Generate the optimized set of weights for the mode
1058 		compute_quantized_weights_for_decimation(
1059 		    di,
1060 		    weight_low_value1[i],
1061 		    weight_high_value1[i],
1062 		    dec_weights_ideal + BLOCK_MAX_WEIGHTS * decimation_mode,
1063 		    dec_weights_uquantf,
1064 		    dec_weights_uquant + BLOCK_MAX_WEIGHTS * i,
1065 		    bm.get_weight_quant_mode());
1066 
1067 		compute_quantized_weights_for_decimation(
1068 		    di,
1069 		    weight_low_value2[i],
1070 		    weight_high_value2[i],
1071 		    dec_weights_ideal + BLOCK_MAX_WEIGHTS * decimation_mode + WEIGHTS_PLANE2_OFFSET,
1072 		    dec_weights_uquantf + WEIGHTS_PLANE2_OFFSET,
1073 		    dec_weights_uquant + BLOCK_MAX_WEIGHTS * i + WEIGHTS_PLANE2_OFFSET,
1074 		    bm.get_weight_quant_mode());
1075 
1076 		// Compute weight quantization errors for the block mode
1077 		qwt_errors[i] = compute_error_of_weight_set_2planes(
1078 		    ei1,
1079 		    ei2,
1080 		    di,
1081 		    dec_weights_uquantf,
1082 		    dec_weights_uquantf + WEIGHTS_PLANE2_OFFSET);
1083 	}
1084 
1085 	// Decide the optimal combination of color endpoint encodings and weight encodings
1086 	uint8_t partition_format_specifiers[TUNE_MAX_TRIAL_CANDIDATES][BLOCK_MAX_PARTITIONS];
1087 	int block_mode_index[TUNE_MAX_TRIAL_CANDIDATES];
1088 
1089 	quant_method color_quant_level[TUNE_MAX_TRIAL_CANDIDATES];
1090 	quant_method color_quant_level_mod[TUNE_MAX_TRIAL_CANDIDATES];
1091 
1092 	endpoints epm;
1093 	merge_endpoints(ei1.ep, ei2.ep, plane2_component, epm);
1094 
1095 	const auto& pi = bsd.get_partition_info(1, 0);
1096 	unsigned int candidate_count = compute_ideal_endpoint_formats(
1097 	    config.privateProfile,
1098 	    pi, blk, epm, qwt_bitcounts, qwt_errors,
1099 	    config.tune_candidate_limit,
1100 		bsd.block_mode_count_1plane_selected, bsd.block_mode_count_1plane_2plane_selected,
1101 	    partition_format_specifiers, block_mode_index,
1102 	    color_quant_level, color_quant_level_mod, tmpbuf);
1103 
1104 	// Iterate over the N believed-to-be-best modes to find out which one is actually best
1105 	float best_errorval_in_mode = ERROR_CALC_DEFAULT;
1106 	float best_errorval_in_scb = scb.errorval;
1107 
1108 	for (unsigned int i = 0; i < candidate_count; i++)
1109 	{
1110 		TRACE_NODE(node0, "candidate");
1111 
1112 		const int bm_packed_index = block_mode_index[i];
1113 		assert(bm_packed_index >= static_cast<int>(bsd.block_mode_count_1plane_selected) &&
1114 		       bm_packed_index < static_cast<int>(bsd.block_mode_count_1plane_2plane_selected));
1115 		const block_mode& qw_bm = bsd.block_modes[bm_packed_index];
1116 
1117 		int decimation_mode = qw_bm.decimation_mode;
1118 		const auto& di = bsd.get_decimation_info(decimation_mode);
1119 		promise(di.weight_count > 0);
1120 
1121 		trace_add_data("weight_x", di.weight_x);
1122 		trace_add_data("weight_y", di.weight_y);
1123 		trace_add_data("weight_z", di.weight_z);
1124 		trace_add_data("weight_quant", qw_bm.quant_mode);
1125 
1126 		vfloat4 rgbs_color;
1127 		vfloat4 rgbo_color;
1128 
1129 		symbolic_compressed_block workscb;
1130 		endpoints workep = epm;
1131 
1132 		uint8_t* u8_weight1_src = dec_weights_uquant + BLOCK_MAX_WEIGHTS * bm_packed_index;
1133 		uint8_t* u8_weight2_src = dec_weights_uquant + BLOCK_MAX_WEIGHTS * bm_packed_index + WEIGHTS_PLANE2_OFFSET;
1134 
1135 		for (int j = 0; j < di.weight_count; j++)
1136 		{
1137 			workscb.weights[j] = u8_weight1_src[j];
1138 			workscb.weights[j + WEIGHTS_PLANE2_OFFSET] = u8_weight2_src[j];
1139 		}
1140 
1141 		for (unsigned int l = 0; l < config.tune_refinement_limit; l++)
1142 		{
1143 			recompute_ideal_colors_2planes(
1144 			    blk, bsd, di,
1145 			    workscb.weights, workscb.weights + WEIGHTS_PLANE2_OFFSET,
1146 			    workep, rgbs_color, rgbo_color, plane2_component);
1147 
1148 			// Quantize the chosen color
1149 			workscb.color_formats[0] = pack_color_endpoints(
1150 			                               privateProfile,
1151 			                               workep.endpt0[0],
1152 			                               workep.endpt1[0],
1153 			                               rgbs_color, rgbo_color,
1154 			                               partition_format_specifiers[i][0],
1155 			                               workscb.color_values[0],
1156 			                               color_quant_level[i]);
1157 
1158 			// Store header fields
1159 			workscb.partition_count = 1;
1160 			workscb.partition_index = 0;
1161 			workscb.quant_mode = color_quant_level[i];
1162 			workscb.color_formats_matched = 0;
1163 			workscb.block_mode = qw_bm.mode_index;
1164 			workscb.plane2_component = static_cast<int8_t>(plane2_component);
1165 			workscb.block_type = SYM_BTYPE_NONCONST;
1166 
1167 			// Pre-realign test
1168 			if (l == 0)
1169 			{
1170 				float errorval = compute_symbolic_block_difference_2plane(config, bsd, workscb, blk);
1171 				if (errorval == -ERROR_CALC_DEFAULT)
1172 				{
1173 					errorval = -errorval;
1174 					workscb.block_type = SYM_BTYPE_ERROR;
1175 				}
1176 
1177 				trace_add_data("error_prerealign", errorval);
1178 				best_errorval_in_mode = astc::min(errorval, best_errorval_in_mode);
1179 
1180 				// Average refinement improvement is 3.5% per iteration (allow 4.5%), but the first
1181 				// iteration can help more so we give it a extra 8% leeway. Use this knowledge to
1182 				// drive a heuristic to skip blocks that are unlikely to catch up with the best
1183 				// block we have already.
1184 				unsigned int iters_remaining = config.tune_refinement_limit - l;
1185 				float threshold = (0.045f * static_cast<float>(iters_remaining)) + 1.08f;
1186 				if (errorval > (threshold * best_errorval_in_scb))
1187 				{
1188 					break;
1189 				}
1190 
1191 				if (errorval < best_errorval_in_scb)
1192 				{
1193 					best_errorval_in_scb = errorval;
1194 					workscb.errorval = errorval;
1195 					scb = workscb;
1196 
1197 					if (errorval < tune_errorval_threshold)
1198 					{
1199 						// Skip remaining candidates - this is "good enough"
1200 						i = candidate_count;
1201 						break;
1202 					}
1203 				}
1204 			}
1205 
1206 			// Perform a final pass over the weights to try to improve them.
1207 			bool adjustments;
1208 			if (di.weight_count != bsd.texel_count)
1209 			{
1210 				adjustments = realign_weights_decimated(
1211 					config.profile, bsd, blk, workscb);
1212 			}
1213 			else
1214 			{
1215 				adjustments = realign_weights_undecimated(
1216 					config.profile, bsd, blk, workscb);
1217 			}
1218 
1219 			// Post-realign test
1220 			float errorval = compute_symbolic_block_difference_2plane(config, bsd, workscb, blk);
1221 			if (errorval == -ERROR_CALC_DEFAULT)
1222 			{
1223 				errorval = -errorval;
1224 				workscb.block_type = SYM_BTYPE_ERROR;
1225 			}
1226 
1227 			trace_add_data("error_postrealign", errorval);
1228 			best_errorval_in_mode = astc::min(errorval, best_errorval_in_mode);
1229 
1230 			// Average refinement improvement is 3.5% per iteration, so skip blocks that are
1231 			// unlikely to catch up with the best block we have already. Assume a 4.5% per step to
1232 			// give benefit of the doubt ...
1233 			unsigned int iters_remaining = config.tune_refinement_limit - 1 - l;
1234 			float threshold = (0.045f * static_cast<float>(iters_remaining)) + 1.0f;
1235 			if (errorval > (threshold * best_errorval_in_scb))
1236 			{
1237 				break;
1238 			}
1239 
1240 			if (errorval < best_errorval_in_scb)
1241 			{
1242 				best_errorval_in_scb = errorval;
1243 				workscb.errorval = errorval;
1244 				scb = workscb;
1245 
1246 				if (errorval < tune_errorval_threshold)
1247 				{
1248 					// Skip remaining candidates - this is "good enough"
1249 					i = candidate_count;
1250 					break;
1251 				}
1252 			}
1253 
1254 			if (!adjustments)
1255 			{
1256 				break;
1257 			}
1258 		}
1259 	}
1260 
1261 	return best_errorval_in_mode;
1262 }
1263 
1264 /**
1265  * @brief Determine the lowest cross-channel correlation factor.
1266  *
1267  * @param texels_per_block   The number of texels in a block.
1268  * @param blk                The image block color data to compress.
1269  *
1270  * @return Return the lowest correlation factor.
1271  */
prepare_block_statistics( int texels_per_block, const image_block& blk )1272 static float prepare_block_statistics(
1273 	int texels_per_block,
1274 	const image_block& blk
1275 ) {
1276 	// Compute covariance matrix, as a collection of 10 scalars that form the upper-triangular row
1277 	// of the matrix. The matrix is symmetric, so this is all we need for this use case.
1278 	float rs = 0.0f;
1279 	float gs = 0.0f;
1280 	float bs = 0.0f;
1281 	float as = 0.0f;
1282 	float rr_var = 0.0f;
1283 	float gg_var = 0.0f;
1284 	float bb_var = 0.0f;
1285 	float aa_var = 0.0f;
1286 	float rg_cov = 0.0f;
1287 	float rb_cov = 0.0f;
1288 	float ra_cov = 0.0f;
1289 	float gb_cov = 0.0f;
1290 	float ga_cov = 0.0f;
1291 	float ba_cov = 0.0f;
1292 
1293 	float weight_sum = 0.0f;
1294 
1295 	promise(texels_per_block > 0);
1296 	for (int i = 0; i < texels_per_block; i++)
1297 	{
1298 		float weight = hadd_s(blk.channel_weight) / 4.0f;
1299 		assert(weight >= 0.0f);
1300 		weight_sum += weight;
1301 
1302 		float r = blk.data_r[i];
1303 		float g = blk.data_g[i];
1304 		float b = blk.data_b[i];
1305 		float a = blk.data_a[i];
1306 
1307 		float rw = r * weight;
1308 		rs += rw;
1309 		rr_var += r * rw;
1310 		rg_cov += g * rw;
1311 		rb_cov += b * rw;
1312 		ra_cov += a * rw;
1313 
1314 		float gw = g * weight;
1315 		gs += gw;
1316 		gg_var += g * gw;
1317 		gb_cov += b * gw;
1318 		ga_cov += a * gw;
1319 
1320 		float bw = b * weight;
1321 		bs += bw;
1322 		bb_var += b * bw;
1323 		ba_cov += a * bw;
1324 
1325 		float aw = a * weight;
1326 		as += aw;
1327 		aa_var += a * aw;
1328 	}
1329 
1330 	float rpt = 1.0f / astc::max(weight_sum, 1e-7f);
1331 
1332 	rr_var -= rs * (rs * rpt);
1333 	rg_cov -= gs * (rs * rpt);
1334 	rb_cov -= bs * (rs * rpt);
1335 	ra_cov -= as * (rs * rpt);
1336 
1337 	gg_var -= gs * (gs * rpt);
1338 	gb_cov -= bs * (gs * rpt);
1339 	ga_cov -= as * (gs * rpt);
1340 
1341 	bb_var -= bs * (bs * rpt);
1342 	ba_cov -= as * (bs * rpt);
1343 
1344 	aa_var -= as * (as * rpt);
1345 
1346 	// These will give a NaN if a channel is constant - these are fixed up in the next step
1347 	rg_cov *= astc::rsqrt(rr_var * gg_var);
1348 	rb_cov *= astc::rsqrt(rr_var * bb_var);
1349 	ra_cov *= astc::rsqrt(rr_var * aa_var);
1350 	gb_cov *= astc::rsqrt(gg_var * bb_var);
1351 	ga_cov *= astc::rsqrt(gg_var * aa_var);
1352 	ba_cov *= astc::rsqrt(bb_var * aa_var);
1353 
1354 	if (astc::isnan(rg_cov)) rg_cov = 1.0f;
1355 	if (astc::isnan(rb_cov)) rb_cov = 1.0f;
1356 	if (astc::isnan(ra_cov)) ra_cov = 1.0f;
1357 	if (astc::isnan(gb_cov)) gb_cov = 1.0f;
1358 	if (astc::isnan(ga_cov)) ga_cov = 1.0f;
1359 	if (astc::isnan(ba_cov)) ba_cov = 1.0f;
1360 
1361 	float lowest_correlation = astc::min(fabsf(rg_cov),      fabsf(rb_cov));
1362 	lowest_correlation       = astc::min(lowest_correlation, fabsf(ra_cov));
1363 	lowest_correlation       = astc::min(lowest_correlation, fabsf(gb_cov));
1364 	lowest_correlation       = astc::min(lowest_correlation, fabsf(ga_cov));
1365 	lowest_correlation       = astc::min(lowest_correlation, fabsf(ba_cov));
1366 
1367 	// Diagnostic trace points
1368 	trace_add_data("min_r", blk.data_min.lane<0>());
1369 	trace_add_data("max_r", blk.data_max.lane<0>());
1370 	trace_add_data("min_g", blk.data_min.lane<1>());
1371 	trace_add_data("max_g", blk.data_max.lane<1>());
1372 	trace_add_data("min_b", blk.data_min.lane<2>());
1373 	trace_add_data("max_b", blk.data_max.lane<2>());
1374 	trace_add_data("min_a", blk.data_min.lane<3>());
1375 	trace_add_data("max_a", blk.data_max.lane<3>());
1376 	trace_add_data("cov_rg", fabsf(rg_cov));
1377 	trace_add_data("cov_rb", fabsf(rb_cov));
1378 	trace_add_data("cov_ra", fabsf(ra_cov));
1379 	trace_add_data("cov_gb", fabsf(gb_cov));
1380 	trace_add_data("cov_ga", fabsf(ga_cov));
1381 	trace_add_data("cov_ba", fabsf(ba_cov));
1382 
1383 	return lowest_correlation;
1384 }
1385 
1386 /* See header for documentation. */
compress_block( const astcenc_contexti& ctx, const image_block& blk, uint8_t pcb[16], compression_working_buffers& tmpbuf, bool calQualityEnable, int32_t *mseBlock[RGBA_COM] )1387 void compress_block(
1388 	const astcenc_contexti& ctx,
1389 	const image_block& blk,
1390 	uint8_t pcb[16],
1391 #if QUALITY_CONTROL
1392 	compression_working_buffers& tmpbuf,
1393 	bool calQualityEnable,
1394 	int32_t *mseBlock[RGBA_COM]
1395 #else
1396 	compression_working_buffers& tmpbuf
1397 #endif
1398 	)
1399 {
1400 	astcenc_profile decode_mode = ctx.config.profile;
1401 	symbolic_compressed_block scb;
1402 	const block_size_descriptor& bsd = *ctx.bsd;
1403 	float lowest_correl;
1404 
1405 	TRACE_NODE(node0, "block");
1406 	trace_add_data("pos_x", blk.xpos);
1407 	trace_add_data("pos_y", blk.ypos);
1408 	trace_add_data("pos_z", blk.zpos);
1409 
1410 	// Set stricter block targets for luminance data as we have more bits to play with
1411 	bool block_is_l = blk.is_luminance();
1412 	float block_is_l_scale = block_is_l ? 1.0f / 1.5f : 1.0f;
1413 
1414 	// Set slightly stricter block targets for lumalpha data as we have more bits to play with
1415 	bool block_is_la = blk.is_luminancealpha();
1416 	float block_is_la_scale = block_is_la ? 1.0f / 1.05f : 1.0f;
1417 
1418 	bool block_skip_two_plane = false;
1419 	int max_partitions;
1420 	if (ctx.config.privateProfile == HIGH_SPEED_PROFILE)
1421 	{
1422 		max_partitions = 1;
1423 	}
1424 #ifdef ASTC_CUSTOMIZED_ENABLE
1425 	else if (ctx.config.privateProfile == CUSTOMIZED_PROFILE)
1426 	{
1427 		if (!g_astcCustomizedSoManager.LoadSutCustomizedSo() ||
1428 			g_astcCustomizedSoManager.customizedMaxPartitionsFunc_ == nullptr)
1429 		{
1430 			printf("astcenc customized so dlopen failed or customizedMaxPartitionsFunc_ is nullptr!\n");
1431 			return;
1432 		}
1433 		max_partitions = g_astcCustomizedSoManager.customizedMaxPartitionsFunc_();
1434 	}
1435 #endif
1436 	else
1437 	{
1438 		max_partitions = ctx.config.tune_partition_count_limit;
1439 	}
1440 
1441 	unsigned int requested_partition_indices[3] {
1442 		ctx.config.tune_2partition_index_limit,
1443 		ctx.config.tune_3partition_index_limit,
1444 		ctx.config.tune_4partition_index_limit
1445 	};
1446 
1447 	unsigned int requested_partition_trials[3] {
1448 		ctx.config.tune_2partitioning_candidate_limit,
1449 		ctx.config.tune_3partitioning_candidate_limit,
1450 		ctx.config.tune_4partitioning_candidate_limit
1451 	};
1452 
1453 #if defined(ASTCENC_DIAGNOSTICS)
1454 	// Do this early in diagnostic builds so we can dump uniform metrics
1455 	// for every block. Do it later in release builds to avoid redundant work!
1456 	float error_weight_sum = hadd_s(blk.channel_weight) * bsd.texel_count;
1457 	float error_threshold = ctx.config.tune_db_limit
1458 	                      * error_weight_sum
1459 	                      * block_is_l_scale
1460 	                      * block_is_la_scale;
1461 
1462 	lowest_correl = prepare_block_statistics(bsd.texel_count, blk);
1463 	trace_add_data("lowest_correl", lowest_correl);
1464 	trace_add_data("tune_error_threshold", error_threshold);
1465 #endif
1466 
1467 	// Detected a constant-color block
1468 	if (all(blk.data_min == blk.data_max))
1469 	{
1470 		TRACE_NODE(node1, "pass");
1471 		trace_add_data("partition_count", 0);
1472 		trace_add_data("plane_count", 1);
1473 
1474 		scb.partition_count = 0;
1475 
1476 		// Encode as FP16 if using HDR
1477 		if ((decode_mode == ASTCENC_PRF_HDR) ||
1478 		    (decode_mode == ASTCENC_PRF_HDR_RGB_LDR_A))
1479 		{
1480 			scb.block_type = SYM_BTYPE_CONST_F16;
1481 			vint4 color_f16 = float_to_float16(blk.origin_texel);
1482 			store(color_f16, scb.constant_color);
1483 		}
1484 		// Encode as UNORM16 if NOT using HDR
1485 		else
1486 		{
1487 			scb.block_type = SYM_BTYPE_CONST_U16;
1488 			vfloat4 color_f32 = clamp(0.0f, 1.0f, blk.origin_texel) * 65535.0f;
1489 			vint4 color_u16 = float_to_int_rtn(color_f32);
1490 			store(color_u16, scb.constant_color);
1491 		}
1492 
1493 		trace_add_data("exit", "quality hit");
1494 		if (ctx.config.privateProfile != HIGH_QUALITY_PROFILE)
1495 		{
1496 			scb.block_type = SYM_BTYPE_NONCONST;
1497 			scb.partition_count = 1;
1498 			scb.color_formats_matched = 0;
1499 			scb.plane2_component = -1;
1500 			if (ctx.config.privateProfile == HIGH_SPEED_PROFILE)
1501 			{
1502 				scb.block_mode = HIGH_SPEED_PROFILE_BLOCK_MODE;
1503 			}
1504 #ifdef ASTC_CUSTOMIZED_ENABLE
1505 			else if (ctx.config.privateProfile == CUSTOMIZED_PROFILE)
1506 			{
1507 				if (!g_astcCustomizedSoManager.LoadSutCustomizedSo() ||
1508 					g_astcCustomizedSoManager.customizedBlockModeFunc_ == nullptr)
1509 				{
1510 					printf("astcenc customized so dlopen failed or customizedBlockModeFunc_ is nullptr!\n");
1511 					return;
1512 				}
1513 				scb.block_mode = g_astcCustomizedSoManager.customizedBlockModeFunc_();
1514 			}
1515 #endif
1516 			scb.partition_index = 0;
1517 			scb.quant_mode = QUANT_256;
1518 			scb.color_formats[0] = 12; // color format is 12 when block mode is HIGH_SPEED_PROFILE_BLOCK_MODE
1519 			for (int w = 0; w < 16; w++) { // weights num is 16 when block mode is HIGH_SPEED_PROFILE_BLOCK_MODE
1520 				scb.weights[w] = 0;
1521 			}
1522 			for (unsigned int pixel = 0; pixel < BLOCK_MAX_COMPONENTS; pixel++) { // scb.constant_color[pixel] is 16 bit
1523 				scb.color_values[0][pixel << 1] = scb.constant_color[pixel] & BYTE_MASK; // low byte
1524 				scb.color_values[0][(pixel << 1) + 1] = (scb.constant_color[pixel] >> 8) & BYTE_MASK; // high byte
1525 			}
1526 		}
1527 		scb.privateProfile = ctx.config.privateProfile;
1528 		symbolic_to_physical(bsd, scb, pcb);
1529 #if QUALITY_CONTROL
1530 	if (calQualityEnable) {
1531 		*mseBlock[R_COM] = *mseBlock[G_COM] = *mseBlock[B_COM] = *mseBlock[A_COM] = 0;
1532 	}
1533 #endif
1534 		return;
1535 	}
1536 
1537 #if !defined(ASTCENC_DIAGNOSTICS)
1538 	float error_weight_sum = hadd_s(blk.channel_weight) * bsd.texel_count;
1539 	float error_threshold = ctx.config.tune_db_limit
1540 	                      * error_weight_sum
1541 	                      * block_is_l_scale
1542 	                      * block_is_la_scale;
1543 #endif
1544 
1545 	// Set SCB and mode errors to a very high error value
1546 	scb.errorval = ERROR_CALC_DEFAULT;
1547 	scb.block_type = SYM_BTYPE_ERROR;
1548 
1549 	float best_errorvals_for_pcount[BLOCK_MAX_PARTITIONS] {
1550 		ERROR_CALC_DEFAULT, ERROR_CALC_DEFAULT, ERROR_CALC_DEFAULT, ERROR_CALC_DEFAULT
1551 	};
1552 
1553 	float exit_thresholds_for_pcount[BLOCK_MAX_PARTITIONS] {
1554 		0.0f,
1555 		ctx.config.tune_2partition_early_out_limit_factor,
1556 		ctx.config.tune_3partition_early_out_limit_factor,
1557 		0.0f
1558 	};
1559 
1560 	// Trial using 1 plane of weights and 1 partition.
1561 
1562 	// Most of the time we test it twice, first with a mode cutoff of 0 and then with the specified
1563 	// mode cutoff. This causes an early-out that speeds up encoding of easy blocks. However, this
1564 	// optimization is disabled for 4x4 and 5x4 blocks where it nearly always slows down the
1565 	// compression and slightly reduces image quality.
1566 
1567 	float errorval_mult[2] {
1568 		1.0f / ctx.config.tune_mse_overshoot,
1569 		1.0f
1570 	};
1571 
1572 	static const float errorval_overshoot = 1.0f / ctx.config.tune_mse_overshoot;
1573 
1574 	// Only enable MODE0 fast path if enabled
1575 	// Never enable for 3D blocks as no "always" block modes are available
1576 	int start_trial = 1;
1577  	if ((ctx.config.tune_search_mode0_enable >= TUNE_MIN_SEARCH_MODE0) && (bsd.zdim == 1))
1578 	{
1579 		start_trial = 0;
1580 	}
1581 
1582 	int quant_limit = QUANT_32;
1583 	for (int i = start_trial; i < 2; i++)
1584 	{
1585 		TRACE_NODE(node1, "pass");
1586 		trace_add_data("partition_count", 1);
1587 		trace_add_data("plane_count", 1);
1588 		trace_add_data("search_mode", i);
1589 
1590 		float errorval = compress_symbolic_block_for_partition_1plane(
1591 		    ctx.config.privateProfile,
1592 		    ctx.config, bsd, blk, i == 0,
1593 		    error_threshold * errorval_mult[i] * errorval_overshoot,
1594 		    1, 0,  scb, tmpbuf, QUANT_32);
1595 
1596 		// Record the quant level so we can use the filter later searches
1597 		const auto& bm = bsd.get_block_mode(scb.block_mode);
1598 		quant_limit = bm.get_weight_quant_mode();
1599 
1600 		best_errorvals_for_pcount[0] = astc::min(best_errorvals_for_pcount[0], errorval);
1601 		if ((ctx.config.privateProfile == HIGH_SPEED_PROFILE) || (errorval < (error_threshold * errorval_mult[i])))
1602 		{
1603 			trace_add_data("exit", "quality hit");
1604 			goto END_OF_TESTS;
1605 		}
1606 	}
1607 
1608 #if !defined(ASTCENC_DIAGNOSTICS)
1609 	lowest_correl = prepare_block_statistics(bsd.texel_count, blk);
1610 #endif
1611 
1612 	block_skip_two_plane = lowest_correl > ctx.config.tune_2plane_early_out_limit_correlation;
1613 
1614 	// Test the four possible 1-partition, 2-planes modes. Do this in reverse, as
1615 	// alpha is the most likely to be non-correlated if it is present in the data.
1616 	for (int i = BLOCK_MAX_COMPONENTS - 1; i >= 0; i--)
1617 	{
1618 		if (ctx.config.privateProfile != HIGH_QUALITY_PROFILE)
1619 		{
1620 			break;
1621 		}
1622 		TRACE_NODE(node1, "pass");
1623 		trace_add_data("partition_count", 1);
1624 		trace_add_data("plane_count", 2);
1625 		trace_add_data("plane_component", i);
1626 
1627 		if (block_skip_two_plane)
1628 		{
1629 			trace_add_data("skip", "tune_2plane_early_out_limit_correlation");
1630 			continue;
1631 		}
1632 
1633 		if (blk.grayscale && i != 3)
1634 		{
1635 			trace_add_data("skip", "grayscale block");
1636 			continue;
1637 		}
1638 
1639 		if (blk.is_constant_channel(i))
1640 		{
1641 			trace_add_data("skip", "constant component");
1642 			continue;
1643 		}
1644 
1645 		float errorval = compress_symbolic_block_for_partition_2planes(
1646 		    ctx.config.privateProfile,
1647 		    ctx.config, bsd, blk, error_threshold * errorval_overshoot,
1648 		    i, scb, tmpbuf, quant_limit);
1649 
1650 		// If attempting two planes is much worse than the best one plane result
1651 		// then further two plane searches are unlikely to help so move on ...
1652 		if (errorval > (best_errorvals_for_pcount[0] * 1.85f))
1653 		{
1654 			break;
1655 		}
1656 
1657 		if (errorval < error_threshold)
1658 		{
1659 			trace_add_data("exit", "quality hit");
1660 			goto END_OF_TESTS;
1661 		}
1662 	}
1663 
1664 	// Find best blocks for 2, 3 and 4 partitions
1665 	for (int partition_count = 2; partition_count <= max_partitions; partition_count++)
1666 	{
1667 		unsigned int partition_indices[TUNE_MAX_PARTITIONING_CANDIDATES];
1668 
1669 		unsigned int requested_indices = requested_partition_indices[partition_count - 2];
1670 
1671 		unsigned int requested_trials = requested_partition_trials[partition_count - 2];
1672 		requested_trials = astc::min(requested_trials, requested_indices);
1673 
1674 		unsigned int actual_trials = find_best_partition_candidates(
1675 		    bsd, blk, partition_count, requested_indices, partition_indices, requested_trials);
1676 
1677 		float best_error_in_prev = best_errorvals_for_pcount[partition_count - 2];
1678 
1679 		for (unsigned int i = 0; i < actual_trials; i++)
1680 		{
1681 			TRACE_NODE(node1, "pass");
1682 			trace_add_data("partition_count", partition_count);
1683 			trace_add_data("partition_index", partition_indices[i]);
1684 			trace_add_data("plane_count", 1);
1685 			trace_add_data("search_mode", i);
1686 
1687 			float errorval = compress_symbolic_block_for_partition_1plane(
1688 			    ctx.config.privateProfile,
1689 			    ctx.config, bsd, blk, false,
1690 			    error_threshold * errorval_overshoot,
1691 			    partition_count, partition_indices[i],
1692 			    scb, tmpbuf, quant_limit);
1693 
1694 			best_errorvals_for_pcount[partition_count - 1] = astc::min(best_errorvals_for_pcount[partition_count - 1], errorval);
1695 
1696 			// If using N partitions doesn't improve much over using N-1 partitions then skip trying
1697 			// N+1. Error can dramatically improve if the data is correlated or non-correlated and
1698 			// aligns with a partitioning that suits that encoding, so for this inner loop check add
1699 			// a large error scale because the "other" trial could be a lot better.
1700 			float best_error = best_errorvals_for_pcount[partition_count - 1];
1701 			float best_error_scale = exit_thresholds_for_pcount[partition_count - 1] * 1.85f;
1702 			if (best_error > (best_error_in_prev * best_error_scale))
1703 			{
1704 				trace_add_data("skip", "tune_partition_early_out_limit_factor");
1705 				goto END_OF_TESTS;
1706 			}
1707 
1708 			if (errorval < error_threshold)
1709 			{
1710 				trace_add_data("exit", "quality hit");
1711 				goto END_OF_TESTS;
1712 			}
1713 		}
1714 
1715 		// If using N partitions doesn't improve much over using N-1 partitions then skip trying N+1
1716 		float best_error = best_errorvals_for_pcount[partition_count - 1];
1717 		float best_error_scale = exit_thresholds_for_pcount[partition_count - 1];
1718 		if (best_error > (best_error_in_prev * best_error_scale))
1719 		{
1720 			trace_add_data("skip", "tune_partition_early_out_limit_factor");
1721 			goto END_OF_TESTS;
1722 		}
1723 	}
1724 
1725 	trace_add_data("exit", "quality not hit");
1726 
1727 END_OF_TESTS:
1728 	// If we still have an error block then convert to something we can encode
1729 	// TODO: Do something more sensible here, such as average color block
1730 	if (scb.block_type == SYM_BTYPE_ERROR)
1731 	{
1732 #if defined(ASTCENC_DIAGNOSTICS)
1733 		static bool printed_once = false;
1734 		if (!printed_once)
1735 		{
1736 			printed_once = true;
1737 			printf("WARN: At least one block failed to find a valid encoding.\n"
1738 			       "      Try increasing compression quality settings.\n\n");
1739 		}
1740 #endif
1741 
1742 		scb.block_type = SYM_BTYPE_CONST_U16;
1743 		vfloat4 color_f32 = clamp(0.0f, 1.0f, blk.origin_texel) * 65535.0f;
1744 		vint4 color_u16 = float_to_int_rtn(color_f32);
1745 		store(color_u16, scb.constant_color);
1746 	}
1747 
1748 	// Compress to a physical block
1749 	scb.privateProfile = ctx.config.privateProfile;
1750 	symbolic_to_physical(bsd, scb, pcb);
1751 #if QUALITY_CONTROL
1752 	if (calQualityEnable) {
1753 		image_block decBlk = blk;
1754 		decompress_symbolic_block(ctx.config.profile, bsd, blk.xpos, blk.ypos, blk.zpos, scb, decBlk);
1755 		vint4 colorSumDiff = vint4::zero();
1756 		for (size_t ii = 0; ii < bsd.texel_count; ii++) {
1757 			vint4 colorRef = float_to_int_rtn(blk.texel(ii) * 255.0f / 65535.0f);
1758 			vint4 colorTest = float_to_int_rtn(min(decBlk.texel(ii), 1.0f) * 255.0f);
1759 			vint4 colorDiff = colorRef - colorTest;
1760 			colorSumDiff += colorDiff * colorDiff;
1761 		}
1762 		*mseBlock[R_COM] = colorSumDiff.lane<0>();
1763 		*mseBlock[G_COM] = colorSumDiff.lane<1>();
1764 		*mseBlock[B_COM] = colorSumDiff.lane<2>();
1765 		*mseBlock[A_COM] = colorSumDiff.lane<3>();
1766     }
1767 #endif
1768 }
1769 
1770 #endif
1771