1 | // SPDX-License-Identifier: Apache-2.0 |
2 | // ---------------------------------------------------------------------------- |
3 | // Copyright 2011-2023 Arm Limited |
4 | // |
5 | // Licensed under the Apache License, Version 2.0 (the "License"); you may not |
6 | // use this file except in compliance with the License. You may obtain a copy |
7 | // of the License at: |
8 | // |
9 | // http://www.apache.org/licenses/LICENSE-2.0 |
10 | // |
11 | // Unless required by applicable law or agreed to in writing, software |
12 | // distributed under the License is distributed on an "AS IS" BASIS, WITHOUT |
13 | // WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the |
14 | // License for the specific language governing permissions and limitations |
15 | // under the License. |
16 | // ---------------------------------------------------------------------------- |
17 | |
18 | #if !defined(ASTCENC_DECOMPRESS_ONLY) |
19 | |
20 | /** |
21 | * @brief Functions for color quantization. |
22 | * |
23 | * The design of the color quantization functionality requires the caller to use higher level error |
24 | * analysis to determine the base encoding that should be used. This earlier analysis will select |
25 | * the basic type of the endpoint that should be used: |
26 | * |
27 | * * Mode: LDR or HDR |
28 | * * Quantization level |
29 | * * Channel count: L, LA, RGB, or RGBA |
30 | * * Endpoint 2 type: Direct color endcode, or scaled from endpoint 1. |
31 | * |
32 | * However, this leaves a number of decisions about exactly how to pack the endpoints open. In |
33 | * particular we need to determine if blue contraction can be used, or/and if delta encoding can be |
34 | * used. If they can be applied these will allow us to maintain higher precision in the endpoints |
35 | * without needing additional storage. |
36 | */ |
37 | |
38 | #include <stdio.h> |
39 | #include <assert.h> |
40 | |
41 | #include "astcenc_internal.h" |
42 | |
43 | /** |
44 | * @brief Determine the quantized value given a quantization level. |
45 | * |
46 | * @param quant_level The quantization level to use. |
47 | * @param value The value to convert. This must be in the 0-255 range. |
48 | * |
49 | * @return The unpacked quantized value, returned in 0-255 range. |
50 | */ |
51 | static inline uint8_t quant_color( |
52 | quant_method quant_level, |
53 | int value |
54 | ) { |
55 | int index = value * 2 + 1; |
56 | return color_unquant_to_uquant_tables[quant_level - QUANT_6][index]; |
57 | } |
58 | |
59 | /** |
60 | * @brief Determine the quantized value given a quantization level and residual. |
61 | * |
62 | * @param quant_level The quantization level to use. |
63 | * @param value The value to convert. This must be in the 0-255 range. |
64 | * @param valuef The original value before rounding, used to compute a residual. |
65 | * |
66 | * @return The unpacked quantized value, returned in 0-255 range. |
67 | */ |
68 | static inline uint8_t quant_color( |
69 | quant_method quant_level, |
70 | int value, |
71 | float valuef |
72 | ) { |
73 | int index = value * 2; |
74 | |
75 | // Compute the residual to determine if we should round down or up ties. |
76 | // Test should be residual >= 0, but empirical testing shows small bias helps. |
77 | float residual = valuef - static_cast<float>(value); |
78 | if (residual >= -0.1f) |
79 | { |
80 | index++; |
81 | } |
82 | |
83 | return color_unquant_to_uquant_tables[quant_level - QUANT_6][index]; |
84 | } |
85 | |
86 | /** |
87 | * @brief Quantize an LDR RGB color. |
88 | * |
89 | * Since this is a fall-back encoding, we cannot actually fail but must produce a sensible result. |
90 | * For this encoding @c color0 cannot be larger than @c color1. If @c color0 is actually larger |
91 | * than @c color1, @c color0 is reduced and @c color1 is increased until the constraint is met. |
92 | * |
93 | * @param color0 The input unquantized color0 endpoint. |
94 | * @param color1 The input unquantized color1 endpoint. |
95 | * @param[out] output The output endpoints, returned as (r0, r1, g0, g1, b0, b1). |
96 | * @param quant_level The quantization level to use. |
97 | */ |
98 | static void quantize_rgb( |
99 | vfloat4 color0, |
100 | vfloat4 color1, |
101 | uint8_t output[6], |
102 | quant_method quant_level |
103 | ) { |
104 | float scale = 1.0f / 257.0f; |
105 | |
106 | float r0 = astc::clamp255f(color0.lane<0>() * scale); |
107 | float g0 = astc::clamp255f(color0.lane<1>() * scale); |
108 | float b0 = astc::clamp255f(color0.lane<2>() * scale); |
109 | |
110 | float r1 = astc::clamp255f(color1.lane<0>() * scale); |
111 | float g1 = astc::clamp255f(color1.lane<1>() * scale); |
112 | float b1 = astc::clamp255f(color1.lane<2>() * scale); |
113 | |
114 | int ri0, gi0, bi0, ri1, gi1, bi1; |
115 | float rgb0_addon = 0.0f; |
116 | float rgb1_addon = 0.0f; |
117 | do |
118 | { |
119 | ri0 = quant_color(quant_level, astc::max(astc::flt2int_rtn(r0 + rgb0_addon), 0), r0 + rgb0_addon); |
120 | gi0 = quant_color(quant_level, astc::max(astc::flt2int_rtn(g0 + rgb0_addon), 0), g0 + rgb0_addon); |
121 | bi0 = quant_color(quant_level, astc::max(astc::flt2int_rtn(b0 + rgb0_addon), 0), b0 + rgb0_addon); |
122 | ri1 = quant_color(quant_level, astc::min(astc::flt2int_rtn(r1 + rgb1_addon), 255), r1 + rgb1_addon); |
123 | gi1 = quant_color(quant_level, astc::min(astc::flt2int_rtn(g1 + rgb1_addon), 255), g1 + rgb1_addon); |
124 | bi1 = quant_color(quant_level, astc::min(astc::flt2int_rtn(b1 + rgb1_addon), 255), b1 + rgb1_addon); |
125 | |
126 | rgb0_addon -= 0.2f; |
127 | rgb1_addon += 0.2f; |
128 | } while (ri0 + gi0 + bi0 > ri1 + gi1 + bi1); |
129 | |
130 | output[0] = static_cast<uint8_t>(ri0); |
131 | output[1] = static_cast<uint8_t>(ri1); |
132 | output[2] = static_cast<uint8_t>(gi0); |
133 | output[3] = static_cast<uint8_t>(gi1); |
134 | output[4] = static_cast<uint8_t>(bi0); |
135 | output[5] = static_cast<uint8_t>(bi1); |
136 | } |
137 | |
138 | /** |
139 | * @brief Quantize an LDR RGBA color. |
140 | * |
141 | * Since this is a fall-back encoding, we cannot actually fail but must produce a sensible result. |
142 | * For this encoding @c color0.rgb cannot be larger than @c color1.rgb (this indicates blue |
143 | * contraction). If @c color0.rgb is actually larger than @c color1.rgb, @c color0.rgb is reduced |
144 | * and @c color1.rgb is increased until the constraint is met. |
145 | * |
146 | * @param color0 The input unquantized color0 endpoint. |
147 | * @param color1 The input unquantized color1 endpoint. |
148 | * @param[out] output The output endpoints, returned as (r0, r1, g0, g1, b0, b1, a0, a1). |
149 | * @param quant_level The quantization level to use. |
150 | */ |
151 | static void quantize_rgba( |
152 | vfloat4 color0, |
153 | vfloat4 color1, |
154 | uint8_t output[8], |
155 | quant_method quant_level |
156 | ) { |
157 | float scale = 1.0f / 257.0f; |
158 | |
159 | float a0 = astc::clamp255f(color0.lane<3>() * scale); |
160 | float a1 = astc::clamp255f(color1.lane<3>() * scale); |
161 | |
162 | output[6] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); |
163 | output[7] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); |
164 | |
165 | quantize_rgb(color0, color1, output, quant_level); |
166 | } |
167 | |
168 | /** |
169 | * @brief Try to quantize an LDR RGB color using blue-contraction. |
170 | * |
171 | * Blue-contraction is only usable if encoded color 1 is larger than color 0. |
172 | * |
173 | * @param color0 The input unquantized color0 endpoint. |
174 | * @param color1 The input unquantized color1 endpoint. |
175 | * @param[out] output The output endpoints, returned as (r1, r0, g1, g0, b1, b0). |
176 | * @param quant_level The quantization level to use. |
177 | * |
178 | * @return Returns @c false on failure, @c true on success. |
179 | */ |
180 | static bool try_quantize_rgb_blue_contract( |
181 | vfloat4 color0, |
182 | vfloat4 color1, |
183 | uint8_t output[6], |
184 | quant_method quant_level |
185 | ) { |
186 | float scale = 1.0f / 257.0f; |
187 | |
188 | float r0 = color0.lane<0>() * scale; |
189 | float g0 = color0.lane<1>() * scale; |
190 | float b0 = color0.lane<2>() * scale; |
191 | |
192 | float r1 = color1.lane<0>() * scale; |
193 | float g1 = color1.lane<1>() * scale; |
194 | float b1 = color1.lane<2>() * scale; |
195 | |
196 | // Apply inverse blue-contraction. This can produce an overflow; which means BC cannot be used. |
197 | r0 += (r0 - b0); |
198 | g0 += (g0 - b0); |
199 | r1 += (r1 - b1); |
200 | g1 += (g1 - b1); |
201 | |
202 | if (r0 < 0.0f || r0 > 255.0f || g0 < 0.0f || g0 > 255.0f || b0 < 0.0f || b0 > 255.0f || |
203 | r1 < 0.0f || r1 > 255.0f || g1 < 0.0f || g1 > 255.0f || b1 < 0.0f || b1 > 255.0f) |
204 | { |
205 | return false; |
206 | } |
207 | |
208 | // Quantize the inverse-blue-contracted color |
209 | int ri0 = quant_color(quant_level, astc::flt2int_rtn(r0), r0); |
210 | int gi0 = quant_color(quant_level, astc::flt2int_rtn(g0), g0); |
211 | int bi0 = quant_color(quant_level, astc::flt2int_rtn(b0), b0); |
212 | |
213 | int ri1 = quant_color(quant_level, astc::flt2int_rtn(r1), r1); |
214 | int gi1 = quant_color(quant_level, astc::flt2int_rtn(g1), g1); |
215 | int bi1 = quant_color(quant_level, astc::flt2int_rtn(b1), b1); |
216 | |
217 | // If color #1 is not larger than color #0 then blue-contraction cannot be used. Note that |
218 | // blue-contraction and quantization change this order, which is why we must test afterwards. |
219 | if (ri1 + gi1 + bi1 <= ri0 + gi0 + bi0) |
220 | { |
221 | return false; |
222 | } |
223 | |
224 | output[0] = static_cast<uint8_t>(ri1); |
225 | output[1] = static_cast<uint8_t>(ri0); |
226 | output[2] = static_cast<uint8_t>(gi1); |
227 | output[3] = static_cast<uint8_t>(gi0); |
228 | output[4] = static_cast<uint8_t>(bi1); |
229 | output[5] = static_cast<uint8_t>(bi0); |
230 | |
231 | return true; |
232 | } |
233 | |
234 | /** |
235 | * @brief Try to quantize an LDR RGBA color using blue-contraction. |
236 | * |
237 | * Blue-contraction is only usable if encoded color 1 RGB is larger than color 0 RGB. |
238 | * |
239 | * @param color0 The input unquantized color0 endpoint. |
240 | * @param color1 The input unquantized color1 endpoint. |
241 | * @param[out] output The output endpoints, returned as (r1, r0, g1, g0, b1, b0, a1, a0). |
242 | * @param quant_level The quantization level to use. |
243 | * |
244 | * @return Returns @c false on failure, @c true on success. |
245 | */ |
246 | static bool try_quantize_rgba_blue_contract( |
247 | vfloat4 color0, |
248 | vfloat4 color1, |
249 | uint8_t output[8], |
250 | quant_method quant_level |
251 | ) { |
252 | float scale = 1.0f / 257.0f; |
253 | |
254 | float a0 = astc::clamp255f(color0.lane<3>() * scale); |
255 | float a1 = astc::clamp255f(color1.lane<3>() * scale); |
256 | |
257 | output[6] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); |
258 | output[7] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); |
259 | |
260 | return try_quantize_rgb_blue_contract(color0, color1, output, quant_level); |
261 | } |
262 | |
263 | /** |
264 | * @brief Try to quantize an LDR RGB color using delta encoding. |
265 | * |
266 | * At decode time we move one bit from the offset to the base and seize another bit as a sign bit; |
267 | * we then unquantize both values as if they contain one extra bit. If the sum of the offsets is |
268 | * non-negative, then we encode a regular delta. |
269 | * |
270 | * @param color0 The input unquantized color0 endpoint. |
271 | * @param color1 The input unquantized color1 endpoint. |
272 | * @param[out] output The output endpoints, returned as (r0, r1, g0, g1, b0, b1). |
273 | * @param quant_level The quantization level to use. |
274 | * |
275 | * @return Returns @c false on failure, @c true on success. |
276 | */ |
277 | static bool try_quantize_rgb_delta( |
278 | vfloat4 color0, |
279 | vfloat4 color1, |
280 | uint8_t output[6], |
281 | quant_method quant_level |
282 | ) { |
283 | float scale = 1.0f / 257.0f; |
284 | |
285 | float r0 = astc::clamp255f(color0.lane<0>() * scale); |
286 | float g0 = astc::clamp255f(color0.lane<1>() * scale); |
287 | float b0 = astc::clamp255f(color0.lane<2>() * scale); |
288 | |
289 | float r1 = astc::clamp255f(color1.lane<0>() * scale); |
290 | float g1 = astc::clamp255f(color1.lane<1>() * scale); |
291 | float b1 = astc::clamp255f(color1.lane<2>() * scale); |
292 | |
293 | // Transform r0 to unorm9 |
294 | int r0a = astc::flt2int_rtn(r0); |
295 | int g0a = astc::flt2int_rtn(g0); |
296 | int b0a = astc::flt2int_rtn(b0); |
297 | |
298 | r0a <<= 1; |
299 | g0a <<= 1; |
300 | b0a <<= 1; |
301 | |
302 | // Mask off the top bit |
303 | int r0b = r0a & 0xFF; |
304 | int g0b = g0a & 0xFF; |
305 | int b0b = b0a & 0xFF; |
306 | |
307 | // Quantize then unquantize in order to get a value that we take differences against |
308 | int r0be = quant_color(quant_level, r0b); |
309 | int g0be = quant_color(quant_level, g0b); |
310 | int b0be = quant_color(quant_level, b0b); |
311 | |
312 | r0b = r0be | (r0a & 0x100); |
313 | g0b = g0be | (g0a & 0x100); |
314 | b0b = b0be | (b0a & 0x100); |
315 | |
316 | // Get hold of the second value |
317 | int r1d = astc::flt2int_rtn(r1); |
318 | int g1d = astc::flt2int_rtn(g1); |
319 | int b1d = astc::flt2int_rtn(b1); |
320 | |
321 | r1d <<= 1; |
322 | g1d <<= 1; |
323 | b1d <<= 1; |
324 | |
325 | // ... and take differences |
326 | r1d -= r0b; |
327 | g1d -= g0b; |
328 | b1d -= b0b; |
329 | |
330 | // Check if the difference is too large to be encodable |
331 | if (r1d > 63 || g1d > 63 || b1d > 63 || r1d < -64 || g1d < -64 || b1d < -64) |
332 | { |
333 | return false; |
334 | } |
335 | |
336 | // Insert top bit of the base into the offset |
337 | r1d &= 0x7F; |
338 | g1d &= 0x7F; |
339 | b1d &= 0x7F; |
340 | |
341 | r1d |= (r0b & 0x100) >> 1; |
342 | g1d |= (g0b & 0x100) >> 1; |
343 | b1d |= (b0b & 0x100) >> 1; |
344 | |
345 | // Then quantize and unquantize; if this causes either top two bits to flip, then encoding fails |
346 | // since we have then corrupted either the top bit of the base or the sign bit of the offset |
347 | int r1de = quant_color(quant_level, r1d); |
348 | int g1de = quant_color(quant_level, g1d); |
349 | int b1de = quant_color(quant_level, b1d); |
350 | |
351 | if (((r1d ^ r1de) | (g1d ^ g1de) | (b1d ^ b1de)) & 0xC0) |
352 | { |
353 | return false; |
354 | } |
355 | |
356 | // If the sum of offsets triggers blue-contraction then encoding fails |
357 | vint4 ep0(r0be, g0be, b0be, 0); |
358 | vint4 ep1(r1de, g1de, b1de, 0); |
359 | bit_transfer_signed(ep1, ep0); |
360 | if (hadd_rgb_s(ep1) < 0) |
361 | { |
362 | return false; |
363 | } |
364 | |
365 | // Check that the offsets produce legitimate sums as well |
366 | ep0 = ep0 + ep1; |
367 | if (any((ep0 < vint4(0)) | (ep0 > vint4(0xFF)))) |
368 | { |
369 | return false; |
370 | } |
371 | |
372 | output[0] = static_cast<uint8_t>(r0be); |
373 | output[1] = static_cast<uint8_t>(r1de); |
374 | output[2] = static_cast<uint8_t>(g0be); |
375 | output[3] = static_cast<uint8_t>(g1de); |
376 | output[4] = static_cast<uint8_t>(b0be); |
377 | output[5] = static_cast<uint8_t>(b1de); |
378 | |
379 | return true; |
380 | } |
381 | |
382 | static bool try_quantize_rgb_delta_blue_contract( |
383 | vfloat4 color0, |
384 | vfloat4 color1, |
385 | uint8_t output[6], |
386 | quant_method quant_level |
387 | ) { |
388 | // Note: Switch around endpoint colors already at start |
389 | float scale = 1.0f / 257.0f; |
390 | |
391 | float r1 = color0.lane<0>() * scale; |
392 | float g1 = color0.lane<1>() * scale; |
393 | float b1 = color0.lane<2>() * scale; |
394 | |
395 | float r0 = color1.lane<0>() * scale; |
396 | float g0 = color1.lane<1>() * scale; |
397 | float b0 = color1.lane<2>() * scale; |
398 | |
399 | // Apply inverse blue-contraction. This can produce an overflow; which means BC cannot be used. |
400 | r0 += (r0 - b0); |
401 | g0 += (g0 - b0); |
402 | r1 += (r1 - b1); |
403 | g1 += (g1 - b1); |
404 | |
405 | if (r0 < 0.0f || r0 > 255.0f || g0 < 0.0f || g0 > 255.0f || b0 < 0.0f || b0 > 255.0f || |
406 | r1 < 0.0f || r1 > 255.0f || g1 < 0.0f || g1 > 255.0f || b1 < 0.0f || b1 > 255.0f) |
407 | { |
408 | return false; |
409 | } |
410 | |
411 | // Transform r0 to unorm9 |
412 | int r0a = astc::flt2int_rtn(r0); |
413 | int g0a = astc::flt2int_rtn(g0); |
414 | int b0a = astc::flt2int_rtn(b0); |
415 | r0a <<= 1; |
416 | g0a <<= 1; |
417 | b0a <<= 1; |
418 | |
419 | // Mask off the top bit |
420 | int r0b = r0a & 0xFF; |
421 | int g0b = g0a & 0xFF; |
422 | int b0b = b0a & 0xFF; |
423 | |
424 | // Quantize, then unquantize in order to get a value that we take differences against. |
425 | int r0be = quant_color(quant_level, r0b); |
426 | int g0be = quant_color(quant_level, g0b); |
427 | int b0be = quant_color(quant_level, b0b); |
428 | |
429 | r0b = r0be | (r0a & 0x100); |
430 | g0b = g0be | (g0a & 0x100); |
431 | b0b = b0be | (b0a & 0x100); |
432 | |
433 | // Get hold of the second value |
434 | int r1d = astc::flt2int_rtn(r1); |
435 | int g1d = astc::flt2int_rtn(g1); |
436 | int b1d = astc::flt2int_rtn(b1); |
437 | |
438 | r1d <<= 1; |
439 | g1d <<= 1; |
440 | b1d <<= 1; |
441 | |
442 | // .. and take differences! |
443 | r1d -= r0b; |
444 | g1d -= g0b; |
445 | b1d -= b0b; |
446 | |
447 | // Check if the difference is too large to be encodable |
448 | if (r1d > 63 || g1d > 63 || b1d > 63 || r1d < -64 || g1d < -64 || b1d < -64) |
449 | { |
450 | return false; |
451 | } |
452 | |
453 | // Insert top bit of the base into the offset |
454 | r1d &= 0x7F; |
455 | g1d &= 0x7F; |
456 | b1d &= 0x7F; |
457 | |
458 | r1d |= (r0b & 0x100) >> 1; |
459 | g1d |= (g0b & 0x100) >> 1; |
460 | b1d |= (b0b & 0x100) >> 1; |
461 | |
462 | // Then quantize and unquantize; if this causes any of the top two bits to flip, |
463 | // then encoding fails, since we have then corrupted either the top bit of the base |
464 | // or the sign bit of the offset. |
465 | int r1de = quant_color(quant_level, r1d); |
466 | int g1de = quant_color(quant_level, g1d); |
467 | int b1de = quant_color(quant_level, b1d); |
468 | |
469 | if (((r1d ^ r1de) | (g1d ^ g1de) | (b1d ^ b1de)) & 0xC0) |
470 | { |
471 | return false; |
472 | } |
473 | |
474 | // If the sum of offsets does not trigger blue-contraction then encoding fails |
475 | vint4 ep0(r0be, g0be, b0be, 0); |
476 | vint4 ep1(r1de, g1de, b1de, 0); |
477 | bit_transfer_signed(ep1, ep0); |
478 | if (hadd_rgb_s(ep1) >= 0) |
479 | { |
480 | return false; |
481 | } |
482 | |
483 | // Check that the offsets produce legitimate sums as well |
484 | ep0 = ep0 + ep1; |
485 | if (any((ep0 < vint4(0)) | (ep0 > vint4(0xFF)))) |
486 | { |
487 | return false; |
488 | } |
489 | |
490 | output[0] = static_cast<uint8_t>(r0be); |
491 | output[1] = static_cast<uint8_t>(r1de); |
492 | output[2] = static_cast<uint8_t>(g0be); |
493 | output[3] = static_cast<uint8_t>(g1de); |
494 | output[4] = static_cast<uint8_t>(b0be); |
495 | output[5] = static_cast<uint8_t>(b1de); |
496 | |
497 | return true; |
498 | } |
499 | |
500 | /** |
501 | * @brief Try to quantize an LDR A color using delta encoding. |
502 | * |
503 | * At decode time we move one bit from the offset to the base and seize another bit as a sign bit; |
504 | * we then unquantize both values as if they contain one extra bit. If the sum of the offsets is |
505 | * non-negative, then we encode a regular delta. |
506 | * |
507 | * This function only compressed the alpha - the other elements in the output array are not touched. |
508 | * |
509 | * @param color0 The input unquantized color0 endpoint. |
510 | * @param color1 The input unquantized color1 endpoint. |
511 | * @param[out] output The output endpoints, returned as (x, x, x, x, x, x, a0, a1). |
512 | * @param quant_level The quantization level to use. |
513 | * |
514 | * @return Returns @c false on failure, @c true on success. |
515 | */ |
516 | static bool try_quantize_alpha_delta( |
517 | vfloat4 color0, |
518 | vfloat4 color1, |
519 | uint8_t output[8], |
520 | quant_method quant_level |
521 | ) { |
522 | float scale = 1.0f / 257.0f; |
523 | |
524 | float a0 = astc::clamp255f(color0.lane<3>() * scale); |
525 | float a1 = astc::clamp255f(color1.lane<3>() * scale); |
526 | |
527 | int a0a = astc::flt2int_rtn(a0); |
528 | a0a <<= 1; |
529 | int a0b = a0a & 0xFF; |
530 | int a0be = quant_color(quant_level, a0b); |
531 | a0b = a0be; |
532 | a0b |= a0a & 0x100; |
533 | int a1d = astc::flt2int_rtn(a1); |
534 | a1d <<= 1; |
535 | a1d -= a0b; |
536 | |
537 | if (a1d > 63 || a1d < -64) |
538 | { |
539 | return false; |
540 | } |
541 | |
542 | a1d &= 0x7F; |
543 | a1d |= (a0b & 0x100) >> 1; |
544 | |
545 | int a1de = quant_color(quant_level, a1d); |
546 | int a1du = a1de; |
547 | if ((a1d ^ a1du) & 0xC0) |
548 | { |
549 | return false; |
550 | } |
551 | |
552 | a1du &= 0x7F; |
553 | if (a1du & 0x40) |
554 | { |
555 | a1du -= 0x80; |
556 | } |
557 | |
558 | a1du += a0b; |
559 | if (a1du < 0 || a1du > 0x1FF) |
560 | { |
561 | return false; |
562 | } |
563 | |
564 | output[6] = static_cast<uint8_t>(a0be); |
565 | output[7] = static_cast<uint8_t>(a1de); |
566 | |
567 | return true; |
568 | } |
569 | |
570 | /** |
571 | * @brief Try to quantize an LDR LA color using delta encoding. |
572 | * |
573 | * At decode time we move one bit from the offset to the base and seize another bit as a sign bit; |
574 | * we then unquantize both values as if they contain one extra bit. If the sum of the offsets is |
575 | * non-negative, then we encode a regular delta. |
576 | * |
577 | * This function only compressed the alpha - the other elements in the output array are not touched. |
578 | * |
579 | * @param color0 The input unquantized color0 endpoint. |
580 | * @param color1 The input unquantized color1 endpoint. |
581 | * @param[out] output The output endpoints, returned as (l0, l1, a0, a1). |
582 | * @param quant_level The quantization level to use. |
583 | * |
584 | * @return Returns @c false on failure, @c true on success. |
585 | */ |
586 | static bool try_quantize_luminance_alpha_delta( |
587 | vfloat4 color0, |
588 | vfloat4 color1, |
589 | uint8_t output[4], |
590 | quant_method quant_level |
591 | ) { |
592 | float scale = 1.0f / 257.0f; |
593 | |
594 | float l0 = astc::clamp255f(hadd_rgb_s(color0) * ((1.0f / 3.0f) * scale)); |
595 | float l1 = astc::clamp255f(hadd_rgb_s(color1) * ((1.0f / 3.0f) * scale)); |
596 | |
597 | float a0 = astc::clamp255f(color0.lane<3>() * scale); |
598 | float a1 = astc::clamp255f(color1.lane<3>() * scale); |
599 | |
600 | int l0a = astc::flt2int_rtn(l0); |
601 | int a0a = astc::flt2int_rtn(a0); |
602 | l0a <<= 1; |
603 | a0a <<= 1; |
604 | |
605 | int l0b = l0a & 0xFF; |
606 | int a0b = a0a & 0xFF; |
607 | int l0be = quant_color(quant_level, l0b); |
608 | int a0be = quant_color(quant_level, a0b); |
609 | l0b = l0be; |
610 | a0b = a0be; |
611 | l0b |= l0a & 0x100; |
612 | a0b |= a0a & 0x100; |
613 | |
614 | int l1d = astc::flt2int_rtn(l1); |
615 | int a1d = astc::flt2int_rtn(a1); |
616 | l1d <<= 1; |
617 | a1d <<= 1; |
618 | l1d -= l0b; |
619 | a1d -= a0b; |
620 | |
621 | if (l1d > 63 || l1d < -64) |
622 | { |
623 | return false; |
624 | } |
625 | |
626 | if (a1d > 63 || a1d < -64) |
627 | { |
628 | return false; |
629 | } |
630 | |
631 | l1d &= 0x7F; |
632 | a1d &= 0x7F; |
633 | l1d |= (l0b & 0x100) >> 1; |
634 | a1d |= (a0b & 0x100) >> 1; |
635 | |
636 | int l1de = quant_color(quant_level, l1d); |
637 | int a1de = quant_color(quant_level, a1d); |
638 | int l1du = l1de; |
639 | int a1du = a1de; |
640 | |
641 | if ((l1d ^ l1du) & 0xC0) |
642 | { |
643 | return false; |
644 | } |
645 | |
646 | if ((a1d ^ a1du) & 0xC0) |
647 | { |
648 | return false; |
649 | } |
650 | |
651 | l1du &= 0x7F; |
652 | a1du &= 0x7F; |
653 | |
654 | if (l1du & 0x40) |
655 | { |
656 | l1du -= 0x80; |
657 | } |
658 | |
659 | if (a1du & 0x40) |
660 | { |
661 | a1du -= 0x80; |
662 | } |
663 | |
664 | l1du += l0b; |
665 | a1du += a0b; |
666 | |
667 | if (l1du < 0 || l1du > 0x1FF) |
668 | { |
669 | return false; |
670 | } |
671 | |
672 | if (a1du < 0 || a1du > 0x1FF) |
673 | { |
674 | return false; |
675 | } |
676 | |
677 | output[0] = static_cast<uint8_t>(l0be); |
678 | output[1] = static_cast<uint8_t>(l1de); |
679 | output[2] = static_cast<uint8_t>(a0be); |
680 | output[3] = static_cast<uint8_t>(a1de); |
681 | |
682 | return true; |
683 | } |
684 | |
685 | /** |
686 | * @brief Try to quantize an LDR RGBA color using delta encoding. |
687 | * |
688 | * At decode time we move one bit from the offset to the base and seize another bit as a sign bit; |
689 | * we then unquantize both values as if they contain one extra bit. If the sum of the offsets is |
690 | * non-negative, then we encode a regular delta. |
691 | * |
692 | * This function only compressed the alpha - the other elements in the output array are not touched. |
693 | * |
694 | * @param color0 The input unquantized color0 endpoint. |
695 | * @param color1 The input unquantized color1 endpoint. |
696 | * @param[out] output The output endpoints, returned as (r0, r1, b0, b1, g0, g1, a0, a1). |
697 | * @param quant_level The quantization level to use. |
698 | * |
699 | * @return Returns @c false on failure, @c true on success. |
700 | */ |
701 | static bool try_quantize_rgba_delta( |
702 | vfloat4 color0, |
703 | vfloat4 color1, |
704 | uint8_t output[8], |
705 | quant_method quant_level |
706 | ) { |
707 | return try_quantize_rgb_delta(color0, color1, output, quant_level) && |
708 | try_quantize_alpha_delta(color0, color1, output, quant_level); |
709 | } |
710 | |
711 | |
712 | /** |
713 | * @brief Try to quantize an LDR RGBA color using delta and blue contract encoding. |
714 | * |
715 | * At decode time we move one bit from the offset to the base and seize another bit as a sign bit; |
716 | * we then unquantize both values as if they contain one extra bit. If the sum of the offsets is |
717 | * non-negative, then we encode a regular delta. |
718 | * |
719 | * This function only compressed the alpha - the other elements in the output array are not touched. |
720 | * |
721 | * @param color0 The input unquantized color0 endpoint. |
722 | * @param color1 The input unquantized color1 endpoint. |
723 | * @param[out] output The output endpoints, returned as (r0, r1, b0, b1, g0, g1, a0, a1). |
724 | * @param quant_level The quantization level to use. |
725 | * |
726 | * @return Returns @c false on failure, @c true on success. |
727 | */ |
728 | static bool try_quantize_rgba_delta_blue_contract( |
729 | vfloat4 color0, |
730 | vfloat4 color1, |
731 | uint8_t output[8], |
732 | quant_method quant_level |
733 | ) { |
734 | // Note that we swap the color0 and color1 ordering for alpha to match RGB blue-contract |
735 | return try_quantize_rgb_delta_blue_contract(color0, color1, output, quant_level) && |
736 | try_quantize_alpha_delta(color1, color0, output, quant_level); |
737 | } |
738 | |
739 | /** |
740 | * @brief Quantize an LDR RGB color using scale encoding. |
741 | * |
742 | * @param color The input unquantized color endpoint and scale factor. |
743 | * @param[out] output The output endpoints, returned as (r0, g0, b0, s). |
744 | * @param quant_level The quantization level to use. |
745 | */ |
746 | static void quantize_rgbs( |
747 | vfloat4 color, |
748 | uint8_t output[4], |
749 | quant_method quant_level |
750 | ) { |
751 | float scale = 1.0f / 257.0f; |
752 | |
753 | float r = astc::clamp255f(color.lane<0>() * scale); |
754 | float g = astc::clamp255f(color.lane<1>() * scale); |
755 | float b = astc::clamp255f(color.lane<2>() * scale); |
756 | |
757 | int ri = quant_color(quant_level, astc::flt2int_rtn(r), r); |
758 | int gi = quant_color(quant_level, astc::flt2int_rtn(g), g); |
759 | int bi = quant_color(quant_level, astc::flt2int_rtn(b), b); |
760 | |
761 | float oldcolorsum = hadd_rgb_s(color) * scale; |
762 | float newcolorsum = static_cast<float>(ri + gi + bi); |
763 | |
764 | float scalea = astc::clamp1f(color.lane<3>() * (oldcolorsum + 1e-10f) / (newcolorsum + 1e-10f)); |
765 | int scale_idx = astc::flt2int_rtn(scalea * 256.0f); |
766 | scale_idx = astc::clamp(scale_idx, 0, 255); |
767 | |
768 | output[0] = static_cast<uint8_t>(ri); |
769 | output[1] = static_cast<uint8_t>(gi); |
770 | output[2] = static_cast<uint8_t>(bi); |
771 | output[3] = quant_color(quant_level, scale_idx); |
772 | } |
773 | |
774 | /** |
775 | * @brief Quantize an LDR RGBA color using scale encoding. |
776 | * |
777 | * @param color The input unquantized color endpoint and scale factor. |
778 | * @param[out] output The output endpoints, returned as (r0, g0, b0, s, a0, a1). |
779 | * @param quant_level The quantization level to use. |
780 | */ |
781 | static void quantize_rgbs_alpha( |
782 | vfloat4 color0, |
783 | vfloat4 color1, |
784 | vfloat4 color, |
785 | uint8_t output[6], |
786 | quant_method quant_level |
787 | ) { |
788 | float scale = 1.0f / 257.0f; |
789 | |
790 | float a0 = astc::clamp255f(color0.lane<3>() * scale); |
791 | float a1 = astc::clamp255f(color1.lane<3>() * scale); |
792 | |
793 | output[4] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); |
794 | output[5] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); |
795 | |
796 | quantize_rgbs(color, output, quant_level); |
797 | } |
798 | |
799 | /** |
800 | * @brief Quantize a LDR L color. |
801 | * |
802 | * @param color0 The input unquantized color0 endpoint. |
803 | * @param color1 The input unquantized color1 endpoint. |
804 | * @param[out] output The output endpoints, returned as (l0, l1). |
805 | * @param quant_level The quantization level to use. |
806 | */ |
807 | static void quantize_luminance( |
808 | vfloat4 color0, |
809 | vfloat4 color1, |
810 | uint8_t output[2], |
811 | quant_method quant_level |
812 | ) { |
813 | float scale = 1.0f / 257.0f; |
814 | |
815 | color0 = color0 * scale; |
816 | color1 = color1 * scale; |
817 | |
818 | float lum0 = astc::clamp255f(hadd_rgb_s(color0) * (1.0f / 3.0f)); |
819 | float lum1 = astc::clamp255f(hadd_rgb_s(color1) * (1.0f / 3.0f)); |
820 | |
821 | if (lum0 > lum1) |
822 | { |
823 | float avg = (lum0 + lum1) * 0.5f; |
824 | lum0 = avg; |
825 | lum1 = avg; |
826 | } |
827 | |
828 | output[0] = quant_color(quant_level, astc::flt2int_rtn(lum0), lum0); |
829 | output[1] = quant_color(quant_level, astc::flt2int_rtn(lum1), lum1); |
830 | } |
831 | |
832 | /** |
833 | * @brief Quantize a LDR LA color. |
834 | * |
835 | * @param color0 The input unquantized color0 endpoint. |
836 | * @param color1 The input unquantized color1 endpoint. |
837 | * @param[out] output The output endpoints, returned as (l0, l1, a0, a1). |
838 | * @param quant_level The quantization level to use. |
839 | */ |
840 | static void quantize_luminance_alpha( |
841 | vfloat4 color0, |
842 | vfloat4 color1, |
843 | uint8_t output[4], |
844 | quant_method quant_level |
845 | ) { |
846 | float scale = 1.0f / 257.0f; |
847 | |
848 | color0 = color0 * scale; |
849 | color1 = color1 * scale; |
850 | |
851 | float lum0 = astc::clamp255f(hadd_rgb_s(color0) * (1.0f / 3.0f)); |
852 | float lum1 = astc::clamp255f(hadd_rgb_s(color1) * (1.0f / 3.0f)); |
853 | |
854 | float a0 = astc::clamp255f(color0.lane<3>()); |
855 | float a1 = astc::clamp255f(color1.lane<3>()); |
856 | |
857 | output[0] = quant_color(quant_level, astc::flt2int_rtn(lum0), lum0); |
858 | output[1] = quant_color(quant_level, astc::flt2int_rtn(lum1), lum1); |
859 | output[2] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); |
860 | output[3] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); |
861 | } |
862 | |
863 | /** |
864 | * @brief Quantize and unquantize a value ensuring top two bits are the same. |
865 | * |
866 | * @param quant_level The quantization level to use. |
867 | * @param value The input unquantized value. |
868 | * @param[out] quant_value The quantized value. |
869 | */ |
870 | static inline void quantize_and_unquantize_retain_top_two_bits( |
871 | quant_method quant_level, |
872 | uint8_t value, |
873 | uint8_t& quant_value |
874 | ) { |
875 | int perform_loop; |
876 | uint8_t quantval; |
877 | |
878 | do |
879 | { |
880 | quantval = quant_color(quant_level, value); |
881 | |
882 | // Perform looping if the top two bits were modified by quant/unquant |
883 | perform_loop = (value & 0xC0) != (quantval & 0xC0); |
884 | |
885 | if ((quantval & 0xC0) > (value & 0xC0)) |
886 | { |
887 | // Quant/unquant rounded UP so that the top two bits changed; |
888 | // decrement the input in hopes that this will avoid rounding up. |
889 | value--; |
890 | } |
891 | else if ((quantval & 0xC0) < (value & 0xC0)) |
892 | { |
893 | // Quant/unquant rounded DOWN so that the top two bits changed; |
894 | // decrement the input in hopes that this will avoid rounding down. |
895 | value--; |
896 | } |
897 | } while (perform_loop); |
898 | |
899 | quant_value = quantval; |
900 | } |
901 | |
902 | /** |
903 | * @brief Quantize and unquantize a value ensuring top four bits are the same. |
904 | * |
905 | * @param quant_level The quantization level to use. |
906 | * @param value The input unquantized value. |
907 | * @param[out] quant_value The quantized value in 0-255 range. |
908 | */ |
909 | static inline void quantize_and_unquantize_retain_top_four_bits( |
910 | quant_method quant_level, |
911 | uint8_t value, |
912 | uint8_t& quant_value |
913 | ) { |
914 | uint8_t perform_loop; |
915 | uint8_t quantval; |
916 | |
917 | do |
918 | { |
919 | quantval = quant_color(quant_level, value); |
920 | // Perform looping if the top four bits were modified by quant/unquant |
921 | perform_loop = (value & 0xF0) != (quantval & 0xF0); |
922 | |
923 | if ((quantval & 0xF0) > (value & 0xF0)) |
924 | { |
925 | // Quant/unquant rounded UP so that the top four bits changed; |
926 | // decrement the input value in hopes that this will avoid rounding up. |
927 | value--; |
928 | } |
929 | else if ((quantval & 0xF0) < (value & 0xF0)) |
930 | { |
931 | // Quant/unquant rounded DOWN so that the top four bits changed; |
932 | // decrement the input value in hopes that this will avoid rounding down. |
933 | value--; |
934 | } |
935 | } while (perform_loop); |
936 | |
937 | quant_value = quantval; |
938 | } |
939 | |
940 | /** |
941 | * @brief Quantize a HDR RGB color using RGB + offset. |
942 | * |
943 | * @param color The input unquantized color endpoint and offset. |
944 | * @param[out] output The output endpoints, returned as packed RGBS with some mode bits. |
945 | * @param quant_level The quantization level to use. |
946 | */ |
947 | static void quantize_hdr_rgbo( |
948 | vfloat4 color, |
949 | uint8_t output[4], |
950 | quant_method quant_level |
951 | ) { |
952 | color.set_lane<0>(color.lane<0>() + color.lane<3>()); |
953 | color.set_lane<1>(color.lane<1>() + color.lane<3>()); |
954 | color.set_lane<2>(color.lane<2>() + color.lane<3>()); |
955 | |
956 | color = clamp(0.0f, 65535.0f, color); |
957 | |
958 | vfloat4 color_bak = color; |
959 | |
960 | int majcomp; |
961 | if (color.lane<0>() > color.lane<1>() && color.lane<0>() > color.lane<2>()) |
962 | { |
963 | majcomp = 0; // red is largest component |
964 | } |
965 | else if (color.lane<1>() > color.lane<2>()) |
966 | { |
967 | majcomp = 1; // green is largest component |
968 | } |
969 | else |
970 | { |
971 | majcomp = 2; // blue is largest component |
972 | } |
973 | |
974 | // swap around the red component and the largest component. |
975 | switch (majcomp) |
976 | { |
977 | case 1: |
978 | color = color.swz<1, 0, 2, 3>(); |
979 | break; |
980 | case 2: |
981 | color = color.swz<2, 1, 0, 3>(); |
982 | break; |
983 | default: |
984 | break; |
985 | } |
986 | |
987 | static const int mode_bits[5][3] { |
988 | {11, 5, 7}, |
989 | {11, 6, 5}, |
990 | {10, 5, 8}, |
991 | {9, 6, 7}, |
992 | {8, 7, 6} |
993 | }; |
994 | |
995 | static const float mode_cutoffs[5][2] { |
996 | {1024, 4096}, |
997 | {2048, 1024}, |
998 | {2048, 16384}, |
999 | {8192, 16384}, |
1000 | {32768, 16384} |
1001 | }; |
1002 | |
1003 | static const float mode_rscales[5] { |
1004 | 32.0f, |
1005 | 32.0f, |
1006 | 64.0f, |
1007 | 128.0f, |
1008 | 256.0f, |
1009 | }; |
1010 | |
1011 | static const float mode_scales[5] { |
1012 | 1.0f / 32.0f, |
1013 | 1.0f / 32.0f, |
1014 | 1.0f / 64.0f, |
1015 | 1.0f / 128.0f, |
1016 | 1.0f / 256.0f, |
1017 | }; |
1018 | |
1019 | float r_base = color.lane<0>(); |
1020 | float g_base = color.lane<0>() - color.lane<1>() ; |
1021 | float b_base = color.lane<0>() - color.lane<2>() ; |
1022 | float s_base = color.lane<3>() ; |
1023 | |
1024 | for (int mode = 0; mode < 5; mode++) |
1025 | { |
1026 | if (g_base > mode_cutoffs[mode][0] || b_base > mode_cutoffs[mode][0] || s_base > mode_cutoffs[mode][1]) |
1027 | { |
1028 | continue; |
1029 | } |
1030 | |
1031 | // Encode the mode into a 4-bit vector |
1032 | int mode_enc = mode < 4 ? (mode | (majcomp << 2)) : (majcomp | 0xC); |
1033 | |
1034 | float mode_scale = mode_scales[mode]; |
1035 | float mode_rscale = mode_rscales[mode]; |
1036 | |
1037 | int gb_intcutoff = 1 << mode_bits[mode][1]; |
1038 | int s_intcutoff = 1 << mode_bits[mode][2]; |
1039 | |
1040 | // Quantize and unquantize R |
1041 | int r_intval = astc::flt2int_rtn(r_base * mode_scale); |
1042 | |
1043 | int r_lowbits = r_intval & 0x3f; |
1044 | |
1045 | r_lowbits |= (mode_enc & 3) << 6; |
1046 | |
1047 | uint8_t r_quantval; |
1048 | quantize_and_unquantize_retain_top_two_bits( |
1049 | quant_level, static_cast<uint8_t>(r_lowbits), r_quantval); |
1050 | |
1051 | r_intval = (r_intval & ~0x3f) | (r_quantval & 0x3f); |
1052 | float r_fval = static_cast<float>(r_intval) * mode_rscale; |
1053 | |
1054 | // Recompute G and B, then quantize and unquantize them |
1055 | float g_fval = r_fval - color.lane<1>() ; |
1056 | float b_fval = r_fval - color.lane<2>() ; |
1057 | |
1058 | g_fval = astc::clamp(g_fval, 0.0f, 65535.0f); |
1059 | b_fval = astc::clamp(b_fval, 0.0f, 65535.0f); |
1060 | |
1061 | int g_intval = astc::flt2int_rtn(g_fval * mode_scale); |
1062 | int b_intval = astc::flt2int_rtn(b_fval * mode_scale); |
1063 | |
1064 | if (g_intval >= gb_intcutoff || b_intval >= gb_intcutoff) |
1065 | { |
1066 | continue; |
1067 | } |
1068 | |
1069 | int g_lowbits = g_intval & 0x1f; |
1070 | int b_lowbits = b_intval & 0x1f; |
1071 | |
1072 | int bit0 = 0; |
1073 | int bit1 = 0; |
1074 | int bit2 = 0; |
1075 | int bit3 = 0; |
1076 | |
1077 | switch (mode) |
1078 | { |
1079 | case 0: |
1080 | case 2: |
1081 | bit0 = (r_intval >> 9) & 1; |
1082 | break; |
1083 | case 1: |
1084 | case 3: |
1085 | bit0 = (r_intval >> 8) & 1; |
1086 | break; |
1087 | case 4: |
1088 | case 5: |
1089 | bit0 = (g_intval >> 6) & 1; |
1090 | break; |
1091 | } |
1092 | |
1093 | switch (mode) |
1094 | { |
1095 | case 0: |
1096 | case 1: |
1097 | case 2: |
1098 | case 3: |
1099 | bit2 = (r_intval >> 7) & 1; |
1100 | break; |
1101 | case 4: |
1102 | case 5: |
1103 | bit2 = (b_intval >> 6) & 1; |
1104 | break; |
1105 | } |
1106 | |
1107 | switch (mode) |
1108 | { |
1109 | case 0: |
1110 | case 2: |
1111 | bit1 = (r_intval >> 8) & 1; |
1112 | break; |
1113 | case 1: |
1114 | case 3: |
1115 | case 4: |
1116 | case 5: |
1117 | bit1 = (g_intval >> 5) & 1; |
1118 | break; |
1119 | } |
1120 | |
1121 | switch (mode) |
1122 | { |
1123 | case 0: |
1124 | bit3 = (r_intval >> 10) & 1; |
1125 | break; |
1126 | case 2: |
1127 | bit3 = (r_intval >> 6) & 1; |
1128 | break; |
1129 | case 1: |
1130 | case 3: |
1131 | case 4: |
1132 | case 5: |
1133 | bit3 = (b_intval >> 5) & 1; |
1134 | break; |
1135 | } |
1136 | |
1137 | g_lowbits |= (mode_enc & 0x4) << 5; |
1138 | b_lowbits |= (mode_enc & 0x8) << 4; |
1139 | |
1140 | g_lowbits |= bit0 << 6; |
1141 | g_lowbits |= bit1 << 5; |
1142 | b_lowbits |= bit2 << 6; |
1143 | b_lowbits |= bit3 << 5; |
1144 | |
1145 | uint8_t g_quantval; |
1146 | uint8_t b_quantval; |
1147 | |
1148 | quantize_and_unquantize_retain_top_four_bits( |
1149 | quant_level, static_cast<uint8_t>(g_lowbits), g_quantval); |
1150 | quantize_and_unquantize_retain_top_four_bits( |
1151 | quant_level, static_cast<uint8_t>(b_lowbits), b_quantval); |
1152 | |
1153 | g_intval = (g_intval & ~0x1f) | (g_quantval & 0x1f); |
1154 | b_intval = (b_intval & ~0x1f) | (b_quantval & 0x1f); |
1155 | |
1156 | g_fval = static_cast<float>(g_intval) * mode_rscale; |
1157 | b_fval = static_cast<float>(b_intval) * mode_rscale; |
1158 | |
1159 | // Recompute the scale value, based on the errors introduced to red, green and blue |
1160 | |
1161 | // If the error is positive, then the R,G,B errors combined have raised the color |
1162 | // value overall; as such, the scale value needs to be increased. |
1163 | float rgb_errorsum = (r_fval - color.lane<0>() ) + (r_fval - g_fval - color.lane<1>() ) + (r_fval - b_fval - color.lane<2>() ); |
1164 | |
1165 | float s_fval = s_base + rgb_errorsum * (1.0f / 3.0f); |
1166 | s_fval = astc::clamp(s_fval, 0.0f, 1e9f); |
1167 | |
1168 | int s_intval = astc::flt2int_rtn(s_fval * mode_scale); |
1169 | |
1170 | if (s_intval >= s_intcutoff) |
1171 | { |
1172 | continue; |
1173 | } |
1174 | |
1175 | int s_lowbits = s_intval & 0x1f; |
1176 | |
1177 | int bit4; |
1178 | int bit5; |
1179 | int bit6; |
1180 | switch (mode) |
1181 | { |
1182 | case 1: |
1183 | bit6 = (r_intval >> 9) & 1; |
1184 | break; |
1185 | default: |
1186 | bit6 = (s_intval >> 5) & 1; |
1187 | break; |
1188 | } |
1189 | |
1190 | switch (mode) |
1191 | { |
1192 | case 4: |
1193 | bit5 = (r_intval >> 7) & 1; |
1194 | break; |
1195 | case 1: |
1196 | bit5 = (r_intval >> 10) & 1; |
1197 | break; |
1198 | default: |
1199 | bit5 = (s_intval >> 6) & 1; |
1200 | break; |
1201 | } |
1202 | |
1203 | switch (mode) |
1204 | { |
1205 | case 2: |
1206 | bit4 = (s_intval >> 7) & 1; |
1207 | break; |
1208 | default: |
1209 | bit4 = (r_intval >> 6) & 1; |
1210 | break; |
1211 | } |
1212 | |
1213 | s_lowbits |= bit6 << 5; |
1214 | s_lowbits |= bit5 << 6; |
1215 | s_lowbits |= bit4 << 7; |
1216 | |
1217 | uint8_t s_quantval; |
1218 | |
1219 | quantize_and_unquantize_retain_top_four_bits( |
1220 | quant_level, static_cast<uint8_t>(s_lowbits), s_quantval); |
1221 | |
1222 | output[0] = r_quantval; |
1223 | output[1] = g_quantval; |
1224 | output[2] = b_quantval; |
1225 | output[3] = s_quantval; |
1226 | return; |
1227 | } |
1228 | |
1229 | // Failed to encode any of the modes above? In that case encode using mode #5 |
1230 | float vals[4]; |
1231 | vals[0] = color_bak.lane<0>(); |
1232 | vals[1] = color_bak.lane<1>(); |
1233 | vals[2] = color_bak.lane<2>(); |
1234 | vals[3] = color_bak.lane<3>(); |
1235 | |
1236 | int ivals[4]; |
1237 | float cvals[3]; |
1238 | |
1239 | for (int i = 0; i < 3; i++) |
1240 | { |
1241 | vals[i] = astc::clamp(vals[i], 0.0f, 65020.0f); |
1242 | ivals[i] = astc::flt2int_rtn(vals[i] * (1.0f / 512.0f)); |
1243 | cvals[i] = static_cast<float>(ivals[i]) * 512.0f; |
1244 | } |
1245 | |
1246 | float rgb_errorsum = (cvals[0] - vals[0]) + (cvals[1] - vals[1]) + (cvals[2] - vals[2]); |
1247 | vals[3] += rgb_errorsum * (1.0f / 3.0f); |
1248 | |
1249 | vals[3] = astc::clamp(vals[3], 0.0f, 65020.0f); |
1250 | ivals[3] = astc::flt2int_rtn(vals[3] * (1.0f / 512.0f)); |
1251 | |
1252 | int encvals[4]; |
1253 | encvals[0] = (ivals[0] & 0x3f) | 0xC0; |
1254 | encvals[1] = (ivals[1] & 0x7f) | 0x80; |
1255 | encvals[2] = (ivals[2] & 0x7f) | 0x80; |
1256 | encvals[3] = (ivals[3] & 0x7f) | ((ivals[0] & 0x40) << 1); |
1257 | |
1258 | for (uint8_t i = 0; i < 4; i++) |
1259 | { |
1260 | quantize_and_unquantize_retain_top_four_bits( |
1261 | quant_level, static_cast<uint8_t>(encvals[i]), output[i]); |
1262 | } |
1263 | |
1264 | return; |
1265 | } |
1266 | |
1267 | /** |
1268 | * @brief Quantize a HDR RGB color using direct RGB encoding. |
1269 | * |
1270 | * @param color0 The input unquantized color0 endpoint. |
1271 | * @param color1 The input unquantized color1 endpoint. |
1272 | * @param[out] output The output endpoints, returned as packed RGB+RGB pairs with mode bits. |
1273 | * @param quant_level The quantization level to use. |
1274 | */ |
1275 | static void quantize_hdr_rgb( |
1276 | vfloat4 color0, |
1277 | vfloat4 color1, |
1278 | uint8_t output[6], |
1279 | quant_method quant_level |
1280 | ) { |
1281 | // Note: color*.lane<3> is not used so we can ignore it |
1282 | color0 = clamp(0.0f, 65535.0f, color0); |
1283 | color1 = clamp(0.0f, 65535.0f, color1); |
1284 | |
1285 | vfloat4 color0_bak = color0; |
1286 | vfloat4 color1_bak = color1; |
1287 | |
1288 | int majcomp; |
1289 | if (color1.lane<0>() > color1.lane<1>() && color1.lane<0>() > color1.lane<2>()) |
1290 | { |
1291 | majcomp = 0; |
1292 | } |
1293 | else if (color1.lane<1>() > color1.lane<2>()) |
1294 | { |
1295 | majcomp = 1; |
1296 | } |
1297 | else |
1298 | { |
1299 | majcomp = 2; |
1300 | } |
1301 | |
1302 | // Swizzle the components |
1303 | switch (majcomp) |
1304 | { |
1305 | case 1: // red-green swap |
1306 | color0 = color0.swz<1, 0, 2, 3>(); |
1307 | color1 = color1.swz<1, 0, 2, 3>(); |
1308 | break; |
1309 | case 2: // red-blue swap |
1310 | color0 = color0.swz<2, 1, 0, 3>(); |
1311 | color1 = color1.swz<2, 1, 0, 3>(); |
1312 | break; |
1313 | default: |
1314 | break; |
1315 | } |
1316 | |
1317 | float a_base = color1.lane<0>(); |
1318 | a_base = astc::clamp(a_base, 0.0f, 65535.0f); |
1319 | |
1320 | float b0_base = a_base - color1.lane<1>(); |
1321 | float b1_base = a_base - color1.lane<2>(); |
1322 | float c_base = a_base - color0.lane<0>(); |
1323 | float d0_base = a_base - b0_base - c_base - color0.lane<1>(); |
1324 | float d1_base = a_base - b1_base - c_base - color0.lane<2>(); |
1325 | |
1326 | // Number of bits in the various fields in the various modes |
1327 | static const int mode_bits[8][4] { |
1328 | {9, 7, 6, 7}, |
1329 | {9, 8, 6, 6}, |
1330 | {10, 6, 7, 7}, |
1331 | {10, 7, 7, 6}, |
1332 | {11, 8, 6, 5}, |
1333 | {11, 6, 8, 6}, |
1334 | {12, 7, 7, 5}, |
1335 | {12, 6, 7, 6} |
1336 | }; |
1337 | |
1338 | // Cutoffs to use for the computed values of a,b,c,d, assuming the |
1339 | // range 0..65535 are LNS values corresponding to fp16. |
1340 | static const float mode_cutoffs[8][4] { |
1341 | {16384, 8192, 8192, 8}, // mode 0: 9,7,6,7 |
1342 | {32768, 8192, 4096, 8}, // mode 1: 9,8,6,6 |
1343 | {4096, 8192, 4096, 4}, // mode 2: 10,6,7,7 |
1344 | {8192, 8192, 2048, 4}, // mode 3: 10,7,7,6 |
1345 | {8192, 2048, 512, 2}, // mode 4: 11,8,6,5 |
1346 | {2048, 8192, 1024, 2}, // mode 5: 11,6,8,6 |
1347 | {2048, 2048, 256, 1}, // mode 6: 12,7,7,5 |
1348 | {1024, 2048, 512, 1}, // mode 7: 12,6,7,6 |
1349 | }; |
1350 | |
1351 | static const float mode_scales[8] { |
1352 | 1.0f / 128.0f, |
1353 | 1.0f / 128.0f, |
1354 | 1.0f / 64.0f, |
1355 | 1.0f / 64.0f, |
1356 | 1.0f / 32.0f, |
1357 | 1.0f / 32.0f, |
1358 | 1.0f / 16.0f, |
1359 | 1.0f / 16.0f, |
1360 | }; |
1361 | |
1362 | // Scaling factors when going from what was encoded in the mode to 16 bits. |
1363 | static const float mode_rscales[8] { |
1364 | 128.0f, |
1365 | 128.0f, |
1366 | 64.0f, |
1367 | 64.0f, |
1368 | 32.0f, |
1369 | 32.0f, |
1370 | 16.0f, |
1371 | 16.0f |
1372 | }; |
1373 | |
1374 | // Try modes one by one, with the highest-precision mode first. |
1375 | for (int mode = 7; mode >= 0; mode--) |
1376 | { |
1377 | // For each mode, test if we can in fact accommodate the computed b, c, and d values. |
1378 | // If we clearly can't, then we skip to the next mode. |
1379 | |
1380 | float b_cutoff = mode_cutoffs[mode][0]; |
1381 | float c_cutoff = mode_cutoffs[mode][1]; |
1382 | float d_cutoff = mode_cutoffs[mode][2]; |
1383 | |
1384 | if (b0_base > b_cutoff || b1_base > b_cutoff || c_base > c_cutoff || fabsf(d0_base) > d_cutoff || fabsf(d1_base) > d_cutoff) |
1385 | { |
1386 | continue; |
1387 | } |
1388 | |
1389 | float mode_scale = mode_scales[mode]; |
1390 | float mode_rscale = mode_rscales[mode]; |
1391 | |
1392 | int b_intcutoff = 1 << mode_bits[mode][1]; |
1393 | int c_intcutoff = 1 << mode_bits[mode][2]; |
1394 | int d_intcutoff = 1 << (mode_bits[mode][3] - 1); |
1395 | |
1396 | // Quantize and unquantize A, with the assumption that its high bits can be handled safely. |
1397 | int a_intval = astc::flt2int_rtn(a_base * mode_scale); |
1398 | int a_lowbits = a_intval & 0xFF; |
1399 | |
1400 | int a_quantval = quant_color(quant_level, a_lowbits); |
1401 | int a_uquantval = a_quantval; |
1402 | a_intval = (a_intval & ~0xFF) | a_uquantval; |
1403 | float a_fval = static_cast<float>(a_intval) * mode_rscale; |
1404 | |
1405 | // Recompute C, then quantize and unquantize it |
1406 | float c_fval = a_fval - color0.lane<0>(); |
1407 | c_fval = astc::clamp(c_fval, 0.0f, 65535.0f); |
1408 | |
1409 | int c_intval = astc::flt2int_rtn(c_fval * mode_scale); |
1410 | |
1411 | if (c_intval >= c_intcutoff) |
1412 | { |
1413 | continue; |
1414 | } |
1415 | |
1416 | int c_lowbits = c_intval & 0x3f; |
1417 | |
1418 | c_lowbits |= (mode & 1) << 7; |
1419 | c_lowbits |= (a_intval & 0x100) >> 2; |
1420 | |
1421 | uint8_t c_quantval; |
1422 | |
1423 | quantize_and_unquantize_retain_top_two_bits( |
1424 | quant_level, static_cast<uint8_t>(c_lowbits), c_quantval); |
1425 | |
1426 | c_intval = (c_intval & ~0x3F) | (c_quantval & 0x3F); |
1427 | c_fval = static_cast<float>(c_intval) * mode_rscale; |
1428 | |
1429 | // Recompute B0 and B1, then quantize and unquantize them |
1430 | float b0_fval = a_fval - color1.lane<1>(); |
1431 | float b1_fval = a_fval - color1.lane<2>(); |
1432 | |
1433 | b0_fval = astc::clamp(b0_fval, 0.0f, 65535.0f); |
1434 | b1_fval = astc::clamp(b1_fval, 0.0f, 65535.0f); |
1435 | int b0_intval = astc::flt2int_rtn(b0_fval * mode_scale); |
1436 | int b1_intval = astc::flt2int_rtn(b1_fval * mode_scale); |
1437 | |
1438 | if (b0_intval >= b_intcutoff || b1_intval >= b_intcutoff) |
1439 | { |
1440 | continue; |
1441 | } |
1442 | |
1443 | int b0_lowbits = b0_intval & 0x3f; |
1444 | int b1_lowbits = b1_intval & 0x3f; |
1445 | |
1446 | int bit0 = 0; |
1447 | int bit1 = 0; |
1448 | switch (mode) |
1449 | { |
1450 | case 0: |
1451 | case 1: |
1452 | case 3: |
1453 | case 4: |
1454 | case 6: |
1455 | bit0 = (b0_intval >> 6) & 1; |
1456 | break; |
1457 | case 2: |
1458 | case 5: |
1459 | case 7: |
1460 | bit0 = (a_intval >> 9) & 1; |
1461 | break; |
1462 | } |
1463 | |
1464 | switch (mode) |
1465 | { |
1466 | case 0: |
1467 | case 1: |
1468 | case 3: |
1469 | case 4: |
1470 | case 6: |
1471 | bit1 = (b1_intval >> 6) & 1; |
1472 | break; |
1473 | case 2: |
1474 | bit1 = (c_intval >> 6) & 1; |
1475 | break; |
1476 | case 5: |
1477 | case 7: |
1478 | bit1 = (a_intval >> 10) & 1; |
1479 | break; |
1480 | } |
1481 | |
1482 | b0_lowbits |= bit0 << 6; |
1483 | b1_lowbits |= bit1 << 6; |
1484 | |
1485 | b0_lowbits |= ((mode >> 1) & 1) << 7; |
1486 | b1_lowbits |= ((mode >> 2) & 1) << 7; |
1487 | |
1488 | uint8_t b0_quantval; |
1489 | uint8_t b1_quantval; |
1490 | |
1491 | quantize_and_unquantize_retain_top_two_bits( |
1492 | quant_level, static_cast<uint8_t>(b0_lowbits), b0_quantval); |
1493 | quantize_and_unquantize_retain_top_two_bits( |
1494 | quant_level, static_cast<uint8_t>(b1_lowbits), b1_quantval); |
1495 | |
1496 | b0_intval = (b0_intval & ~0x3f) | (b0_quantval & 0x3f); |
1497 | b1_intval = (b1_intval & ~0x3f) | (b1_quantval & 0x3f); |
1498 | b0_fval = static_cast<float>(b0_intval) * mode_rscale; |
1499 | b1_fval = static_cast<float>(b1_intval) * mode_rscale; |
1500 | |
1501 | // Recompute D0 and D1, then quantize and unquantize them |
1502 | float d0_fval = a_fval - b0_fval - c_fval - color0.lane<1>(); |
1503 | float d1_fval = a_fval - b1_fval - c_fval - color0.lane<2>(); |
1504 | |
1505 | d0_fval = astc::clamp(d0_fval, -65535.0f, 65535.0f); |
1506 | d1_fval = astc::clamp(d1_fval, -65535.0f, 65535.0f); |
1507 | |
1508 | int d0_intval = astc::flt2int_rtn(d0_fval * mode_scale); |
1509 | int d1_intval = astc::flt2int_rtn(d1_fval * mode_scale); |
1510 | |
1511 | if (abs(d0_intval) >= d_intcutoff || abs(d1_intval) >= d_intcutoff) |
1512 | { |
1513 | continue; |
1514 | } |
1515 | |
1516 | int d0_lowbits = d0_intval & 0x1f; |
1517 | int d1_lowbits = d1_intval & 0x1f; |
1518 | |
1519 | int bit2 = 0; |
1520 | int bit3 = 0; |
1521 | int bit4; |
1522 | int bit5; |
1523 | switch (mode) |
1524 | { |
1525 | case 0: |
1526 | case 2: |
1527 | bit2 = (d0_intval >> 6) & 1; |
1528 | break; |
1529 | case 1: |
1530 | case 4: |
1531 | bit2 = (b0_intval >> 7) & 1; |
1532 | break; |
1533 | case 3: |
1534 | bit2 = (a_intval >> 9) & 1; |
1535 | break; |
1536 | case 5: |
1537 | bit2 = (c_intval >> 7) & 1; |
1538 | break; |
1539 | case 6: |
1540 | case 7: |
1541 | bit2 = (a_intval >> 11) & 1; |
1542 | break; |
1543 | } |
1544 | switch (mode) |
1545 | { |
1546 | case 0: |
1547 | case 2: |
1548 | bit3 = (d1_intval >> 6) & 1; |
1549 | break; |
1550 | case 1: |
1551 | case 4: |
1552 | bit3 = (b1_intval >> 7) & 1; |
1553 | break; |
1554 | case 3: |
1555 | case 5: |
1556 | case 6: |
1557 | case 7: |
1558 | bit3 = (c_intval >> 6) & 1; |
1559 | break; |
1560 | } |
1561 | |
1562 | switch (mode) |
1563 | { |
1564 | case 4: |
1565 | case 6: |
1566 | bit4 = (a_intval >> 9) & 1; |
1567 | bit5 = (a_intval >> 10) & 1; |
1568 | break; |
1569 | default: |
1570 | bit4 = (d0_intval >> 5) & 1; |
1571 | bit5 = (d1_intval >> 5) & 1; |
1572 | break; |
1573 | } |
1574 | |
1575 | d0_lowbits |= bit2 << 6; |
1576 | d1_lowbits |= bit3 << 6; |
1577 | d0_lowbits |= bit4 << 5; |
1578 | d1_lowbits |= bit5 << 5; |
1579 | |
1580 | d0_lowbits |= (majcomp & 1) << 7; |
1581 | d1_lowbits |= ((majcomp >> 1) & 1) << 7; |
1582 | |
1583 | uint8_t d0_quantval; |
1584 | uint8_t d1_quantval; |
1585 | |
1586 | quantize_and_unquantize_retain_top_four_bits( |
1587 | quant_level, static_cast<uint8_t>(d0_lowbits), d0_quantval); |
1588 | quantize_and_unquantize_retain_top_four_bits( |
1589 | quant_level, static_cast<uint8_t>(d1_lowbits), d1_quantval); |
1590 | |
1591 | output[0] = static_cast<uint8_t>(a_quantval); |
1592 | output[1] = c_quantval; |
1593 | output[2] = b0_quantval; |
1594 | output[3] = b1_quantval; |
1595 | output[4] = d0_quantval; |
1596 | output[5] = d1_quantval; |
1597 | return; |
1598 | } |
1599 | |
1600 | // If neither of the modes fit we will use a flat representation for storing data, using 8 bits |
1601 | // for red and green, and 7 bits for blue. This gives color accuracy roughly similar to LDR |
1602 | // 4:4:3 which is not at all great but usable. This representation is used if the light color is |
1603 | // more than 4x the color value of the dark color. |
1604 | float vals[6]; |
1605 | vals[0] = color0_bak.lane<0>(); |
1606 | vals[1] = color1_bak.lane<0>(); |
1607 | vals[2] = color0_bak.lane<1>(); |
1608 | vals[3] = color1_bak.lane<1>(); |
1609 | vals[4] = color0_bak.lane<2>(); |
1610 | vals[5] = color1_bak.lane<2>(); |
1611 | |
1612 | for (int i = 0; i < 6; i++) |
1613 | { |
1614 | vals[i] = astc::clamp(vals[i], 0.0f, 65020.0f); |
1615 | } |
1616 | |
1617 | for (int i = 0; i < 4; i++) |
1618 | { |
1619 | int idx = astc::flt2int_rtn(vals[i] * 1.0f / 256.0f); |
1620 | output[i] = quant_color(quant_level, idx); |
1621 | } |
1622 | |
1623 | for (int i = 4; i < 6; i++) |
1624 | { |
1625 | int idx = astc::flt2int_rtn(vals[i] * 1.0f / 512.0f) + 128; |
1626 | quantize_and_unquantize_retain_top_two_bits( |
1627 | quant_level, static_cast<uint8_t>(idx), output[i]); |
1628 | } |
1629 | |
1630 | return; |
1631 | } |
1632 | |
1633 | /** |
1634 | * @brief Quantize a HDR RGB + LDR A color using direct RGBA encoding. |
1635 | * |
1636 | * @param color0 The input unquantized color0 endpoint. |
1637 | * @param color1 The input unquantized color1 endpoint. |
1638 | * @param[out] output The output endpoints, returned as packed RGBA+RGBA pairs with mode bits. |
1639 | * @param quant_level The quantization level to use. |
1640 | */ |
1641 | static void quantize_hdr_rgb_ldr_alpha( |
1642 | vfloat4 color0, |
1643 | vfloat4 color1, |
1644 | uint8_t output[8], |
1645 | quant_method quant_level |
1646 | ) { |
1647 | float scale = 1.0f / 257.0f; |
1648 | |
1649 | float a0 = astc::clamp255f(color0.lane<3>() * scale); |
1650 | float a1 = astc::clamp255f(color1.lane<3>() * scale); |
1651 | |
1652 | output[6] = quant_color(quant_level, astc::flt2int_rtn(a0), a0); |
1653 | output[7] = quant_color(quant_level, astc::flt2int_rtn(a1), a1); |
1654 | |
1655 | quantize_hdr_rgb(color0, color1, output, quant_level); |
1656 | } |
1657 | |
1658 | /** |
1659 | * @brief Quantize a HDR L color using the large range encoding. |
1660 | * |
1661 | * @param color0 The input unquantized color0 endpoint. |
1662 | * @param color1 The input unquantized color1 endpoint. |
1663 | * @param[out] output The output endpoints, returned as packed (l0, l1). |
1664 | * @param quant_level The quantization level to use. |
1665 | */ |
1666 | static void quantize_hdr_luminance_large_range( |
1667 | vfloat4 color0, |
1668 | vfloat4 color1, |
1669 | uint8_t output[2], |
1670 | quant_method quant_level |
1671 | ) { |
1672 | float lum0 = hadd_rgb_s(color0) * (1.0f / 3.0f); |
1673 | float lum1 = hadd_rgb_s(color1) * (1.0f / 3.0f); |
1674 | |
1675 | if (lum1 < lum0) |
1676 | { |
1677 | float avg = (lum0 + lum1) * 0.5f; |
1678 | lum0 = avg; |
1679 | lum1 = avg; |
1680 | } |
1681 | |
1682 | int ilum1 = astc::flt2int_rtn(lum1); |
1683 | int ilum0 = astc::flt2int_rtn(lum0); |
1684 | |
1685 | // Find the closest encodable point in the upper half of the code-point space |
1686 | int upper_v0 = (ilum0 + 128) >> 8; |
1687 | int upper_v1 = (ilum1 + 128) >> 8; |
1688 | |
1689 | upper_v0 = astc::clamp(upper_v0, 0, 255); |
1690 | upper_v1 = astc::clamp(upper_v1, 0, 255); |
1691 | |
1692 | // Find the closest encodable point in the lower half of the code-point space |
1693 | int lower_v0 = (ilum1 + 256) >> 8; |
1694 | int lower_v1 = ilum0 >> 8; |
1695 | |
1696 | lower_v0 = astc::clamp(lower_v0, 0, 255); |
1697 | lower_v1 = astc::clamp(lower_v1, 0, 255); |
1698 | |
1699 | // Determine the distance between the point in code-point space and the input value |
1700 | int upper0_dec = upper_v0 << 8; |
1701 | int upper1_dec = upper_v1 << 8; |
1702 | int lower0_dec = (lower_v1 << 8) + 128; |
1703 | int lower1_dec = (lower_v0 << 8) - 128; |
1704 | |
1705 | int upper0_diff = upper0_dec - ilum0; |
1706 | int upper1_diff = upper1_dec - ilum1; |
1707 | int lower0_diff = lower0_dec - ilum0; |
1708 | int lower1_diff = lower1_dec - ilum1; |
1709 | |
1710 | int upper_error = (upper0_diff * upper0_diff) + (upper1_diff * upper1_diff); |
1711 | int lower_error = (lower0_diff * lower0_diff) + (lower1_diff * lower1_diff); |
1712 | |
1713 | int v0, v1; |
1714 | if (upper_error < lower_error) |
1715 | { |
1716 | v0 = upper_v0; |
1717 | v1 = upper_v1; |
1718 | } |
1719 | else |
1720 | { |
1721 | v0 = lower_v0; |
1722 | v1 = lower_v1; |
1723 | } |
1724 | |
1725 | // OK; encode |
1726 | output[0] = quant_color(quant_level, v0); |
1727 | output[1] = quant_color(quant_level, v1); |
1728 | } |
1729 | |
1730 | /** |
1731 | * @brief Quantize a HDR L color using the small range encoding. |
1732 | * |
1733 | * @param color0 The input unquantized color0 endpoint. |
1734 | * @param color1 The input unquantized color1 endpoint. |
1735 | * @param[out] output The output endpoints, returned as packed (l0, l1) with mode bits. |
1736 | * @param quant_level The quantization level to use. |
1737 | * |
1738 | * @return Returns @c false on failure, @c true on success. |
1739 | */ |
1740 | static bool try_quantize_hdr_luminance_small_range( |
1741 | vfloat4 color0, |
1742 | vfloat4 color1, |
1743 | uint8_t output[2], |
1744 | quant_method quant_level |
1745 | ) { |
1746 | float lum0 = hadd_rgb_s(color0) * (1.0f / 3.0f); |
1747 | float lum1 = hadd_rgb_s(color1) * (1.0f / 3.0f); |
1748 | |
1749 | if (lum1 < lum0) |
1750 | { |
1751 | float avg = (lum0 + lum1) * 0.5f; |
1752 | lum0 = avg; |
1753 | lum1 = avg; |
1754 | } |
1755 | |
1756 | int ilum1 = astc::flt2int_rtn(lum1); |
1757 | int ilum0 = astc::flt2int_rtn(lum0); |
1758 | |
1759 | // Difference of more than a factor-of-2 results in immediate failure |
1760 | if (ilum1 - ilum0 > 2048) |
1761 | { |
1762 | return false; |
1763 | } |
1764 | |
1765 | int lowval, highval, diffval; |
1766 | int v0, v1; |
1767 | int v0e, v1e; |
1768 | int v0d, v1d; |
1769 | |
1770 | // Try to encode the high-precision submode |
1771 | lowval = (ilum0 + 16) >> 5; |
1772 | highval = (ilum1 + 16) >> 5; |
1773 | |
1774 | lowval = astc::clamp(lowval, 0, 2047); |
1775 | highval = astc::clamp(highval, 0, 2047); |
1776 | |
1777 | v0 = lowval & 0x7F; |
1778 | v0e = quant_color(quant_level, v0); |
1779 | v0d = v0e; |
1780 | |
1781 | if (v0d < 0x80) |
1782 | { |
1783 | lowval = (lowval & ~0x7F) | v0d; |
1784 | diffval = highval - lowval; |
1785 | if (diffval >= 0 && diffval <= 15) |
1786 | { |
1787 | v1 = ((lowval >> 3) & 0xF0) | diffval; |
1788 | v1e = quant_color(quant_level, v1); |
1789 | v1d = v1e; |
1790 | if ((v1d & 0xF0) == (v1 & 0xF0)) |
1791 | { |
1792 | output[0] = static_cast<uint8_t>(v0e); |
1793 | output[1] = static_cast<uint8_t>(v1e); |
1794 | return true; |
1795 | } |
1796 | } |
1797 | } |
1798 | |
1799 | // Try to encode the low-precision submode |
1800 | lowval = (ilum0 + 32) >> 6; |
1801 | highval = (ilum1 + 32) >> 6; |
1802 | |
1803 | lowval = astc::clamp(lowval, 0, 1023); |
1804 | highval = astc::clamp(highval, 0, 1023); |
1805 | |
1806 | v0 = (lowval & 0x7F) | 0x80; |
1807 | v0e = quant_color(quant_level, v0); |
1808 | v0d = v0e; |
1809 | if ((v0d & 0x80) == 0) |
1810 | { |
1811 | return false; |
1812 | } |
1813 | |
1814 | lowval = (lowval & ~0x7F) | (v0d & 0x7F); |
1815 | diffval = highval - lowval; |
1816 | if (diffval < 0 || diffval > 31) |
1817 | { |
1818 | return false; |
1819 | } |
1820 | |
1821 | v1 = ((lowval >> 2) & 0xE0) | diffval; |
1822 | v1e = quant_color(quant_level, v1); |
1823 | v1d = v1e; |
1824 | if ((v1d & 0xE0) != (v1 & 0xE0)) |
1825 | { |
1826 | return false; |
1827 | } |
1828 | |
1829 | output[0] = static_cast<uint8_t>(v0e); |
1830 | output[1] = static_cast<uint8_t>(v1e); |
1831 | return true; |
1832 | } |
1833 | |
1834 | /** |
1835 | * @brief Quantize a HDR A color using either delta or direct RGBA encoding. |
1836 | * |
1837 | * @param alpha0 The input unquantized color0 endpoint. |
1838 | * @param alpha1 The input unquantized color1 endpoint. |
1839 | * @param[out] output The output endpoints, returned as packed RGBA+RGBA pairs with mode bits. |
1840 | * @param quant_level The quantization level to use. |
1841 | */ |
1842 | static void quantize_hdr_alpha( |
1843 | float alpha0, |
1844 | float alpha1, |
1845 | uint8_t output[2], |
1846 | quant_method quant_level |
1847 | ) { |
1848 | alpha0 = astc::clamp(alpha0, 0.0f, 65280.0f); |
1849 | alpha1 = astc::clamp(alpha1, 0.0f, 65280.0f); |
1850 | |
1851 | int ialpha0 = astc::flt2int_rtn(alpha0); |
1852 | int ialpha1 = astc::flt2int_rtn(alpha1); |
1853 | |
1854 | int val0, val1, diffval; |
1855 | int v6, v7; |
1856 | int v6e, v7e; |
1857 | int v6d, v7d; |
1858 | |
1859 | // Try to encode one of the delta submodes, in decreasing-precision order |
1860 | for (int i = 2; i >= 0; i--) |
1861 | { |
1862 | val0 = (ialpha0 + (128 >> i)) >> (8 - i); |
1863 | val1 = (ialpha1 + (128 >> i)) >> (8 - i); |
1864 | |
1865 | v6 = (val0 & 0x7F) | ((i & 1) << 7); |
1866 | v6e = quant_color(quant_level, v6); |
1867 | v6d = v6e; |
1868 | |
1869 | if ((v6 ^ v6d) & 0x80) |
1870 | { |
1871 | continue; |
1872 | } |
1873 | |
1874 | val0 = (val0 & ~0x7f) | (v6d & 0x7f); |
1875 | diffval = val1 - val0; |
1876 | int cutoff = 32 >> i; |
1877 | int mask = 2 * cutoff - 1; |
1878 | |
1879 | if (diffval < -cutoff || diffval >= cutoff) |
1880 | { |
1881 | continue; |
1882 | } |
1883 | |
1884 | v7 = ((i & 2) << 6) | ((val0 >> 7) << (6 - i)) | (diffval & mask); |
1885 | v7e = quant_color(quant_level, v7); |
1886 | v7d = v7e; |
1887 | |
1888 | static const int testbits[3] { 0xE0, 0xF0, 0xF8 }; |
1889 | |
1890 | if ((v7 ^ v7d) & testbits[i]) |
1891 | { |
1892 | continue; |
1893 | } |
1894 | |
1895 | output[0] = static_cast<uint8_t>(v6e); |
1896 | output[1] = static_cast<uint8_t>(v7e); |
1897 | return; |
1898 | } |
1899 | |
1900 | // Could not encode any of the delta modes; instead encode a flat value |
1901 | val0 = (ialpha0 + 256) >> 9; |
1902 | val1 = (ialpha1 + 256) >> 9; |
1903 | v6 = val0 | 0x80; |
1904 | v7 = val1 | 0x80; |
1905 | |
1906 | output[0] = quant_color(quant_level, v6); |
1907 | output[1] = quant_color(quant_level, v7); |
1908 | |
1909 | return; |
1910 | } |
1911 | |
1912 | /** |
1913 | * @brief Quantize a HDR RGBA color using either delta or direct RGBA encoding. |
1914 | * |
1915 | * @param color0 The input unquantized color0 endpoint. |
1916 | * @param color1 The input unquantized color1 endpoint. |
1917 | * @param[out] output The output endpoints, returned as packed RGBA+RGBA pairs with mode bits. |
1918 | * @param quant_level The quantization level to use. |
1919 | */ |
1920 | static void quantize_hdr_rgb_alpha( |
1921 | vfloat4 color0, |
1922 | vfloat4 color1, |
1923 | uint8_t output[8], |
1924 | quant_method quant_level |
1925 | ) { |
1926 | quantize_hdr_rgb(color0, color1, output, quant_level); |
1927 | quantize_hdr_alpha(color0.lane<3>(), color1.lane<3>(), output + 6, quant_level); |
1928 | } |
1929 | |
1930 | /* See header for documentation. */ |
1931 | uint8_t pack_color_endpoints( |
1932 | vfloat4 color0, |
1933 | vfloat4 color1, |
1934 | vfloat4 rgbs_color, |
1935 | vfloat4 rgbo_color, |
1936 | int format, |
1937 | uint8_t* output, |
1938 | quant_method quant_level |
1939 | ) { |
1940 | assert(QUANT_6 <= quant_level && quant_level <= QUANT_256); |
1941 | |
1942 | // We do not support negative colors |
1943 | color0 = max(color0, 0.0f); |
1944 | color1 = max(color1, 0.0f); |
1945 | |
1946 | uint8_t retval = 0; |
1947 | |
1948 | switch (format) |
1949 | { |
1950 | case FMT_RGB: |
1951 | if (quant_level <= QUANT_160) |
1952 | { |
1953 | if (try_quantize_rgb_delta_blue_contract(color0, color1, output, quant_level)) |
1954 | { |
1955 | retval = FMT_RGB_DELTA; |
1956 | break; |
1957 | } |
1958 | if (try_quantize_rgb_delta(color0, color1, output, quant_level)) |
1959 | { |
1960 | retval = FMT_RGB_DELTA; |
1961 | break; |
1962 | } |
1963 | } |
1964 | if (quant_level < QUANT_256 && try_quantize_rgb_blue_contract(color0, color1, output, quant_level)) |
1965 | { |
1966 | retval = FMT_RGB; |
1967 | break; |
1968 | } |
1969 | quantize_rgb(color0, color1, output, quant_level); |
1970 | retval = FMT_RGB; |
1971 | break; |
1972 | |
1973 | case FMT_RGBA: |
1974 | if (quant_level <= QUANT_160) |
1975 | { |
1976 | if (try_quantize_rgba_delta_blue_contract(color0, color1, output, quant_level)) |
1977 | { |
1978 | retval = FMT_RGBA_DELTA; |
1979 | break; |
1980 | } |
1981 | if (try_quantize_rgba_delta(color0, color1, output, quant_level)) |
1982 | { |
1983 | retval = FMT_RGBA_DELTA; |
1984 | break; |
1985 | } |
1986 | } |
1987 | if (quant_level < QUANT_256 && try_quantize_rgba_blue_contract(color0, color1, output, quant_level)) |
1988 | { |
1989 | retval = FMT_RGBA; |
1990 | break; |
1991 | } |
1992 | quantize_rgba(color0, color1, output, quant_level); |
1993 | retval = FMT_RGBA; |
1994 | break; |
1995 | |
1996 | case FMT_RGB_SCALE: |
1997 | quantize_rgbs(rgbs_color, output, quant_level); |
1998 | retval = FMT_RGB_SCALE; |
1999 | break; |
2000 | |
2001 | case FMT_HDR_RGB_SCALE: |
2002 | quantize_hdr_rgbo(rgbo_color, output, quant_level); |
2003 | retval = FMT_HDR_RGB_SCALE; |
2004 | break; |
2005 | |
2006 | case FMT_HDR_RGB: |
2007 | quantize_hdr_rgb(color0, color1, output, quant_level); |
2008 | retval = FMT_HDR_RGB; |
2009 | break; |
2010 | |
2011 | case FMT_RGB_SCALE_ALPHA: |
2012 | quantize_rgbs_alpha(color0, color1, rgbs_color, output, quant_level); |
2013 | retval = FMT_RGB_SCALE_ALPHA; |
2014 | break; |
2015 | |
2016 | case FMT_HDR_LUMINANCE_SMALL_RANGE: |
2017 | case FMT_HDR_LUMINANCE_LARGE_RANGE: |
2018 | if (try_quantize_hdr_luminance_small_range(color0, color1, output, quant_level)) |
2019 | { |
2020 | retval = FMT_HDR_LUMINANCE_SMALL_RANGE; |
2021 | break; |
2022 | } |
2023 | quantize_hdr_luminance_large_range(color0, color1, output, quant_level); |
2024 | retval = FMT_HDR_LUMINANCE_LARGE_RANGE; |
2025 | break; |
2026 | |
2027 | case FMT_LUMINANCE: |
2028 | quantize_luminance(color0, color1, output, quant_level); |
2029 | retval = FMT_LUMINANCE; |
2030 | break; |
2031 | |
2032 | case FMT_LUMINANCE_ALPHA: |
2033 | if (quant_level <= 18) |
2034 | { |
2035 | if (try_quantize_luminance_alpha_delta(color0, color1, output, quant_level)) |
2036 | { |
2037 | retval = FMT_LUMINANCE_ALPHA_DELTA; |
2038 | break; |
2039 | } |
2040 | } |
2041 | quantize_luminance_alpha(color0, color1, output, quant_level); |
2042 | retval = FMT_LUMINANCE_ALPHA; |
2043 | break; |
2044 | |
2045 | case FMT_HDR_RGB_LDR_ALPHA: |
2046 | quantize_hdr_rgb_ldr_alpha(color0, color1, output, quant_level); |
2047 | retval = FMT_HDR_RGB_LDR_ALPHA; |
2048 | break; |
2049 | |
2050 | case FMT_HDR_RGBA: |
2051 | quantize_hdr_rgb_alpha(color0, color1, output, quant_level); |
2052 | retval = FMT_HDR_RGBA; |
2053 | break; |
2054 | } |
2055 | |
2056 | return retval; |
2057 | } |
2058 | |
2059 | #endif |
2060 | |