OLD | NEW |
---|---|
(Empty) | |
1 // Copyright 2015 The Chromium Authors. All rights reserved. | |
2 // Use of this source code is governed by a BSD-style license that can be | |
3 // found in the LICENSE file. | |
4 | |
5 #include "cc/resources/texture_compressor_etc1_sse.h" | |
6 | |
7 #include <assert.h> | |
reveman
2015/05/07 16:26:29
is this needed?
radu.velea
2015/05/08 10:52:11
Done.
| |
8 #include <emmintrin.h> | |
9 #include <cmath> | |
10 #include <limits> | |
11 | |
12 #include "base/compiler_specific.h" | |
13 #include "base/logging.h" | |
14 // Using this header for common functions such as Color handling | |
15 // and codeword table. | |
16 #include "cc/resources/texture_compressor_etc1.h" | |
17 | |
18 namespace cc { | |
19 | |
20 namespace { | |
21 | |
22 inline uint32_t SetETC1MaxError(uint32_t avg_error) { | |
23 // ETC1 codeword table is sorted ascending. | |
24 // Our algorithm will try to identify the index that generates the minimum | |
25 // error. | |
26 // The min error calculated during ComputeLuminance main loop will converge | |
27 // towards that value. | |
28 // We use this threashold to determine when it doesn't make sense to iterate | |
29 // further through the array. | |
30 return avg_error + avg_error / 2 + 384; | |
31 } | |
32 | |
33 struct __sse_data { | |
34 // This is used to store raw data. | |
35 uint8_t* block; | |
36 // This is used to store 8 bit packed values. | |
37 __m128i* packed; | |
38 // This is used to store 32 bit zero extended values into 4x4 arrays. | |
39 __m128i* blue; | |
40 __m128i* green; | |
41 __m128i* red; | |
42 }; | |
43 | |
44 // Commonly used registers throughout the code. | |
45 static const __m128i __sse_zero = _mm_set1_epi32(0); | |
46 static const __m128i __sse_max_int = _mm_set1_epi32(0x7FFFFFFF); | |
47 | |
48 inline __m128i AddAndClamp(const __m128i x, const __m128i y) { | |
49 static const __m128i color_max = _mm_set1_epi32(0xFF); | |
50 return _mm_max_epi16(__sse_zero, | |
51 _mm_min_epi16(_mm_add_epi16(x, y), color_max)); | |
52 } | |
53 | |
54 inline __m128i GetColorErrorSSE(const __m128i x, const __m128i y) { | |
55 // Changed from _mm_mullo_epi32 (SSE4) to _mm_mullo_epi16 (SSE2). | |
56 __m128i ret = _mm_sub_epi16(x, y); | |
57 return _mm_mullo_epi16(ret, ret); | |
58 } | |
59 | |
60 inline __m128i AddChannelError(const __m128i x, | |
61 const __m128i y, | |
62 const __m128i z) { | |
63 return _mm_add_epi32(x, _mm_add_epi32(y, z)); | |
64 } | |
65 | |
66 inline uint32_t SumSSE(const __m128i x) { | |
67 __m128i sum = _mm_add_epi32(x, _mm_shuffle_epi32(x, 0x4E)); | |
68 sum = _mm_add_epi32(sum, _mm_shuffle_epi32(sum, 0xB1)); | |
69 | |
70 return _mm_cvtsi128_si32(sum); | |
71 } | |
72 | |
73 inline uint32_t GetVerticalError(const __sse_data* data, | |
74 const __m128i* blue_avg, | |
75 const __m128i* green_avg, | |
76 const __m128i* red_avg, | |
77 uint32_t* verror) { | |
78 __m128i error = __sse_zero; | |
79 | |
80 for (int i = 0; i < 4; i++) { | |
81 error = _mm_add_epi32(error, GetColorErrorSSE(data->blue[i], blue_avg[0])); | |
82 error = | |
83 _mm_add_epi32(error, GetColorErrorSSE(data->green[i], green_avg[0])); | |
84 error = _mm_add_epi32(error, GetColorErrorSSE(data->red[i], red_avg[0])); | |
85 } | |
86 | |
87 error = _mm_add_epi32(error, _mm_shuffle_epi32(error, 0x4E)); | |
88 | |
89 verror[0] = _mm_cvtsi128_si32(error); | |
90 verror[1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(error, 0xB1)); | |
91 | |
92 return verror[0] + verror[1]; | |
93 } | |
94 | |
95 inline uint32_t GetHorizontalError(const __sse_data* data, | |
96 const __m128i* blue_avg, | |
97 const __m128i* green_avg, | |
98 const __m128i* red_avg, | |
99 uint32_t* verror) { | |
100 __m128i error = __sse_zero; | |
101 int first_index, second_index; | |
102 | |
103 for (int i = 0; i < 2; i++) { | |
104 first_index = 2 * i; | |
105 second_index = first_index + 1; | |
106 | |
107 error = _mm_add_epi32( | |
108 error, GetColorErrorSSE(data->blue[first_index], blue_avg[i])); | |
109 error = _mm_add_epi32( | |
110 error, GetColorErrorSSE(data->blue[second_index], blue_avg[i])); | |
111 error = _mm_add_epi32( | |
112 error, GetColorErrorSSE(data->green[first_index], green_avg[i])); | |
113 error = _mm_add_epi32( | |
114 error, GetColorErrorSSE(data->green[second_index], green_avg[i])); | |
115 error = _mm_add_epi32(error, | |
116 GetColorErrorSSE(data->red[first_index], red_avg[i])); | |
117 error = _mm_add_epi32( | |
118 error, GetColorErrorSSE(data->red[second_index], red_avg[i])); | |
119 } | |
120 | |
121 error = _mm_add_epi32(error, _mm_shuffle_epi32(error, 0x4E)); | |
122 | |
123 verror[0] = _mm_cvtsi128_si32(error); | |
124 verror[1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(error, 0xB1)); | |
125 | |
126 return verror[0] + verror[1]; | |
127 } | |
128 | |
129 inline void GetAvgColors(const __sse_data* data, | |
130 float* output, | |
131 bool* __sse_use_diff) { | |
132 __m128i sum[2], tmp; | |
133 | |
134 // TODO(radu.velea): _mm_avg_epu8 on packed data maybe. | |
135 | |
136 // Compute avg red value. | |
137 // [S0 S0 S1 S1] | |
138 sum[0] = _mm_add_epi32(data->red[0], data->red[1]); | |
139 sum[0] = _mm_add_epi32(sum[0], _mm_shuffle_epi32(sum[0], 0xB1)); | |
140 | |
141 // [S2 S2 S3 S3] | |
142 sum[1] = _mm_add_epi32(data->red[2], data->red[3]); | |
143 sum[1] = _mm_add_epi32(sum[1], _mm_shuffle_epi32(sum[1], 0xB1)); | |
144 | |
145 float hred[2], vred[2]; | |
146 hred[0] = (_mm_cvtsi128_si32( | |
147 _mm_add_epi32(sum[0], _mm_shuffle_epi32(sum[0], 0x4E)))) / | |
148 8.0f; | |
149 hred[1] = (_mm_cvtsi128_si32( | |
150 _mm_add_epi32(sum[1], _mm_shuffle_epi32(sum[1], 0x4E)))) / | |
151 8.0f; | |
152 | |
153 tmp = _mm_add_epi32(sum[0], sum[1]); | |
154 vred[0] = (_mm_cvtsi128_si32(tmp)) / 8.0f; | |
155 vred[1] = (_mm_cvtsi128_si32(_mm_shuffle_epi32(tmp, 0x2))) / 8.0f; | |
156 | |
157 // Compute avg green value. | |
158 // [S0 S0 S1 S1] | |
159 sum[0] = _mm_add_epi32(data->green[0], data->green[1]); | |
160 sum[0] = _mm_add_epi32(sum[0], _mm_shuffle_epi32(sum[0], 0xB1)); | |
161 | |
162 // [S2 S2 S3 S3] | |
163 sum[1] = _mm_add_epi32(data->green[2], data->green[3]); | |
164 sum[1] = _mm_add_epi32(sum[1], _mm_shuffle_epi32(sum[1], 0xB1)); | |
165 | |
166 float hgreen[2], vgreen[2]; | |
167 hgreen[0] = (_mm_cvtsi128_si32( | |
168 _mm_add_epi32(sum[0], _mm_shuffle_epi32(sum[0], 0x4E)))) / | |
169 8.0f; | |
170 hgreen[1] = (_mm_cvtsi128_si32( | |
171 _mm_add_epi32(sum[1], _mm_shuffle_epi32(sum[1], 0x4E)))) / | |
172 8.0f; | |
173 | |
174 tmp = _mm_add_epi32(sum[0], sum[1]); | |
175 vgreen[0] = (_mm_cvtsi128_si32(tmp)) / 8.0f; | |
176 vgreen[1] = (_mm_cvtsi128_si32(_mm_shuffle_epi32(tmp, 0x2))) / 8.0f; | |
177 | |
178 // Compute avg blue value. | |
179 // [S0 S0 S1 S1] | |
180 sum[0] = _mm_add_epi32(data->blue[0], data->blue[1]); | |
181 sum[0] = _mm_add_epi32(sum[0], _mm_shuffle_epi32(sum[0], 0xB1)); | |
182 | |
183 // [S2 S2 S3 S3] | |
184 sum[1] = _mm_add_epi32(data->blue[2], data->blue[3]); | |
185 sum[1] = _mm_add_epi32(sum[1], _mm_shuffle_epi32(sum[1], 0xB1)); | |
186 | |
187 float hblue[2], vblue[2]; | |
188 hblue[0] = (_mm_cvtsi128_si32( | |
189 _mm_add_epi32(sum[0], _mm_shuffle_epi32(sum[0], 0x4E)))) / | |
190 8.0f; | |
191 hblue[1] = (_mm_cvtsi128_si32( | |
192 _mm_add_epi32(sum[1], _mm_shuffle_epi32(sum[1], 0x4E)))) / | |
193 8.0f; | |
194 | |
195 tmp = _mm_add_epi32(sum[0], sum[1]); | |
196 vblue[0] = (_mm_cvtsi128_si32(tmp)) / 8.0f; | |
197 vblue[1] = (_mm_cvtsi128_si32(_mm_shuffle_epi32(tmp, 0x2))) / 8.0f; | |
198 | |
199 // TODO(radu.velea): Return int's instead of floats, based on Quality. | |
200 output[0] = vblue[0]; | |
201 output[1] = vgreen[0]; | |
202 output[2] = vred[0]; | |
203 | |
204 output[3] = vblue[1]; | |
205 output[4] = vgreen[1]; | |
206 output[5] = vred[1]; | |
207 | |
208 output[6] = hblue[0]; | |
209 output[7] = hgreen[0]; | |
210 output[8] = hred[0]; | |
211 | |
212 output[9] = hblue[1]; | |
213 output[10] = hgreen[1]; | |
214 output[11] = hred[1]; | |
215 | |
216 __m128i threshold_upper = _mm_set1_epi32(3); | |
217 __m128i threshold_lower = _mm_set1_epi32(-4); | |
218 | |
219 __m128 factor_v = _mm_set1_ps(31.0f / 255.0f); | |
220 __m128 rounding_v = _mm_set1_ps(0.5f); | |
221 __m128 h_avg_0 = _mm_set_ps(hblue[0], hgreen[0], hred[0], 0); | |
222 __m128 h_avg_1 = _mm_set_ps(hblue[1], hgreen[1], hred[1], 0); | |
223 | |
224 __m128 v_avg_0 = _mm_set_ps(vblue[0], vgreen[0], vred[0], 0); | |
225 __m128 v_avg_1 = _mm_set_ps(vblue[1], vgreen[1], vred[1], 0); | |
226 | |
227 h_avg_0 = _mm_mul_ps(h_avg_0, factor_v); | |
228 h_avg_1 = _mm_mul_ps(h_avg_1, factor_v); | |
229 v_avg_0 = _mm_mul_ps(v_avg_0, factor_v); | |
230 v_avg_1 = _mm_mul_ps(v_avg_1, factor_v); | |
231 | |
232 h_avg_0 = _mm_add_ps(h_avg_0, rounding_v); | |
233 h_avg_1 = _mm_add_ps(h_avg_1, rounding_v); | |
234 v_avg_0 = _mm_add_ps(v_avg_0, rounding_v); | |
235 v_avg_1 = _mm_add_ps(v_avg_1, rounding_v); | |
236 | |
237 __m128i h_avg_0i = _mm_cvttps_epi32(h_avg_0); | |
238 __m128i h_avg_1i = _mm_cvttps_epi32(h_avg_1); | |
239 | |
240 __m128i v_avg_0i = _mm_cvttps_epi32(v_avg_0); | |
241 __m128i v_avg_1i = _mm_cvttps_epi32(v_avg_1); | |
242 | |
243 h_avg_0i = _mm_sub_epi32(h_avg_1i, h_avg_0i); | |
244 v_avg_0i = _mm_sub_epi32(v_avg_1i, v_avg_0i); | |
245 | |
246 __sse_use_diff[0] = | |
247 (0 == _mm_movemask_epi8(_mm_cmplt_epi32(v_avg_0i, threshold_lower))); | |
248 __sse_use_diff[0] &= | |
249 (0 == _mm_movemask_epi8(_mm_cmpgt_epi32(v_avg_0i, threshold_upper))); | |
250 | |
251 __sse_use_diff[1] = | |
252 (0 == _mm_movemask_epi8(_mm_cmplt_epi32(h_avg_0i, threshold_lower))); | |
253 __sse_use_diff[1] &= | |
254 (0 == _mm_movemask_epi8(_mm_cmpgt_epi32(h_avg_0i, threshold_upper))); | |
255 } | |
256 | |
257 void ComputeLuminance(uint8_t* block, | |
258 const Color& base, | |
259 const int sub_block_id, | |
260 const uint8_t* idx_to_num_tab, | |
261 const __sse_data* data, | |
262 const uint32_t expected_error) { | |
263 uint8_t best_tbl_idx = 0; | |
264 uint32_t best_error = 0x7FFFFFFF; | |
265 uint8_t best_mod_idx[8][8]; // [table][texel] | |
266 | |
267 const __m128i base_blue = _mm_set1_epi32(base.channels.b); | |
268 const __m128i base_green = _mm_set1_epi32(base.channels.g); | |
269 const __m128i base_red = _mm_set1_epi32(base.channels.r); | |
270 | |
271 __m128i test_red, test_blue, test_green, tmp, tmp_blue, tmp_green, tmp_red; | |
272 __m128i block_error, mask; | |
273 | |
274 // This will have the minimum errors for each 4 pixels. | |
275 __m128i first_half_min; | |
276 __m128i second_half_min; | |
277 | |
278 // This will have the matching table index combo for each 4 pixels. | |
279 __m128i first_half_pattern; | |
280 __m128i second_half_pattern; | |
281 | |
282 const __m128i first_blue_data_block = data->blue[2 * sub_block_id]; | |
283 const __m128i first_green_data_block = data->green[2 * sub_block_id]; | |
284 const __m128i first_red_data_block = data->red[2 * sub_block_id]; | |
285 | |
286 const __m128i second_blue_data_block = data->blue[2 * sub_block_id + 1]; | |
287 const __m128i second_green_data_block = data->green[2 * sub_block_id + 1]; | |
288 const __m128i second_red_data_block = data->red[2 * sub_block_id + 1]; | |
289 | |
290 uint32_t min; | |
291 // Fail early to increase speed. | |
292 long delta = INT32_MAX; | |
293 uint32_t last_min = INT32_MAX; | |
294 | |
295 const uint8_t shuffle_mask[] = { | |
296 0x1B, 0x4E, 0xB1, 0xE4}; // Important they are sorted ascending. | |
297 | |
298 for (unsigned int tbl_idx = 0; tbl_idx < 8; ++tbl_idx) { | |
299 tmp = _mm_set_epi32( | |
300 g_codeword_tables[tbl_idx][3], g_codeword_tables[tbl_idx][2], | |
301 g_codeword_tables[tbl_idx][1], g_codeword_tables[tbl_idx][0]); | |
302 | |
303 test_blue = AddAndClamp(tmp, base_blue); | |
304 test_green = AddAndClamp(tmp, base_green); | |
305 test_red = AddAndClamp(tmp, base_red); | |
306 | |
307 first_half_min = __sse_max_int; | |
308 second_half_min = __sse_max_int; | |
309 | |
310 first_half_pattern = __sse_zero; | |
311 second_half_pattern = __sse_zero; | |
312 | |
313 for (uint8_t imm8 : shuffle_mask) { | |
314 switch (imm8) { | |
315 case 0x1B: | |
316 tmp_blue = _mm_shuffle_epi32(test_blue, 0x1B); | |
317 tmp_green = _mm_shuffle_epi32(test_green, 0x1B); | |
318 tmp_red = _mm_shuffle_epi32(test_red, 0x1B); | |
319 break; | |
320 case 0x4E: | |
321 tmp_blue = _mm_shuffle_epi32(test_blue, 0x4E); | |
322 tmp_green = _mm_shuffle_epi32(test_green, 0x4E); | |
323 tmp_red = _mm_shuffle_epi32(test_red, 0x4E); | |
324 break; | |
325 case 0xB1: | |
326 tmp_blue = _mm_shuffle_epi32(test_blue, 0xB1); | |
327 tmp_green = _mm_shuffle_epi32(test_green, 0xB1); | |
328 tmp_red = _mm_shuffle_epi32(test_red, 0xB1); | |
329 break; | |
330 case 0xE4: | |
331 tmp_blue = _mm_shuffle_epi32(test_blue, 0xE4); | |
332 tmp_green = _mm_shuffle_epi32(test_green, 0xE4); | |
333 tmp_red = _mm_shuffle_epi32(test_red, 0xE4); | |
334 break; | |
335 default: | |
336 tmp_blue = test_blue; | |
337 tmp_green = test_green; | |
338 tmp_red = test_red; | |
339 } | |
340 | |
341 tmp = _mm_set1_epi32(imm8); | |
342 | |
343 block_error = | |
344 AddChannelError(GetColorErrorSSE(tmp_blue, first_blue_data_block), | |
345 GetColorErrorSSE(tmp_green, first_green_data_block), | |
346 GetColorErrorSSE(tmp_red, first_red_data_block)); | |
347 | |
348 // Save winning pattern. | |
349 first_half_pattern = _mm_max_epi16( | |
350 first_half_pattern, | |
351 _mm_and_si128(tmp, _mm_cmpgt_epi32(first_half_min, block_error))); | |
352 // Should use _mm_min_epi32(first_half_min, block_error); from SSE4 | |
353 // otherwise we have a small performance penalty. | |
354 mask = _mm_cmplt_epi32(block_error, first_half_min); | |
355 first_half_min = _mm_add_epi32(_mm_and_si128(mask, block_error), | |
356 _mm_andnot_si128(mask, first_half_min)); | |
357 | |
358 // Compute second part of the block. | |
359 block_error = | |
360 AddChannelError(GetColorErrorSSE(tmp_blue, second_blue_data_block), | |
361 GetColorErrorSSE(tmp_green, second_green_data_block), | |
362 GetColorErrorSSE(tmp_red, second_red_data_block)); | |
363 | |
364 // Save winning pattern. | |
365 second_half_pattern = _mm_max_epi16( | |
366 second_half_pattern, | |
367 _mm_and_si128(tmp, _mm_cmpgt_epi32(second_half_min, block_error))); | |
368 // Should use _mm_min_epi32(second_half_min, block_error); from SSE4 | |
369 // otherwise we have a small performance penalty. | |
370 mask = _mm_cmplt_epi32(block_error, second_half_min); | |
371 second_half_min = _mm_add_epi32(_mm_and_si128(mask, block_error), | |
372 _mm_andnot_si128(mask, second_half_min)); | |
373 } | |
374 | |
375 first_half_min = _mm_add_epi32(first_half_min, second_half_min); | |
376 first_half_min = | |
377 _mm_add_epi32(first_half_min, _mm_shuffle_epi32(first_half_min, 0x4E)); | |
378 first_half_min = | |
379 _mm_add_epi32(first_half_min, _mm_shuffle_epi32(first_half_min, 0xB1)); | |
380 | |
381 min = _mm_cvtsi128_si32(first_half_min); | |
382 | |
383 delta = min - last_min; | |
384 last_min = min; | |
385 | |
386 if (min < best_error) { | |
387 best_tbl_idx = tbl_idx; | |
388 best_error = min; | |
389 | |
390 best_mod_idx[tbl_idx][0] = | |
391 (_mm_cvtsi128_si32(first_half_pattern) >> (0)) & 3; | |
392 best_mod_idx[tbl_idx][4] = | |
393 (_mm_cvtsi128_si32(second_half_pattern) >> (0)) & 3; | |
394 | |
395 best_mod_idx[tbl_idx][1] = | |
396 (_mm_cvtsi128_si32(_mm_shuffle_epi32(first_half_pattern, 0x1)) >> | |
397 (2)) & | |
398 3; | |
399 best_mod_idx[tbl_idx][5] = | |
400 (_mm_cvtsi128_si32(_mm_shuffle_epi32(second_half_pattern, 0x1)) >> | |
401 (2)) & | |
402 3; | |
403 | |
404 best_mod_idx[tbl_idx][2] = | |
405 (_mm_cvtsi128_si32(_mm_shuffle_epi32(first_half_pattern, 0x2)) >> | |
406 (4)) & | |
407 3; | |
408 best_mod_idx[tbl_idx][6] = | |
409 (_mm_cvtsi128_si32(_mm_shuffle_epi32(second_half_pattern, 0x2)) >> | |
410 (4)) & | |
411 3; | |
412 | |
413 best_mod_idx[tbl_idx][3] = | |
414 (_mm_cvtsi128_si32(_mm_shuffle_epi32(first_half_pattern, 0x3)) >> | |
415 (6)) & | |
416 3; | |
417 best_mod_idx[tbl_idx][7] = | |
418 (_mm_cvtsi128_si32(_mm_shuffle_epi32(second_half_pattern, 0x3)) >> | |
419 (6)) & | |
420 3; | |
421 | |
422 if (best_error == 0) { | |
423 break; | |
424 } | |
425 } else if (delta > 0 && expected_error < min) { | |
426 // The error is growing and is well beyond expected threshold. | |
427 break; | |
428 } | |
429 } | |
430 | |
431 WriteCodewordTable(block, sub_block_id, best_tbl_idx); | |
432 | |
433 uint32_t pix_data = 0; | |
434 uint8_t mod_idx; | |
435 uint8_t pix_idx; | |
436 uint32_t lsb; | |
437 uint32_t msb; | |
438 int texel_num; | |
439 | |
440 for (unsigned int i = 0; i < 8; ++i) { | |
441 mod_idx = best_mod_idx[best_tbl_idx][i]; | |
442 pix_idx = g_mod_to_pix[mod_idx]; | |
443 | |
444 lsb = pix_idx & 0x1; | |
445 msb = pix_idx >> 1; | |
446 | |
447 // Obtain the texel number as specified in the standard. | |
448 texel_num = idx_to_num_tab[i]; | |
449 pix_data |= msb << (texel_num + 16); | |
450 pix_data |= lsb << (texel_num); | |
451 } | |
452 | |
453 WritePixelData(block, pix_data); | |
454 } | |
455 | |
456 void CompressBlock(uint8_t* dst, __sse_data* data) { | |
457 // First 3 values are for vertical 1, second 3 vertical 2, third 3 horizontal | |
458 // 1, last 3 | |
459 // horizontal 2. | |
460 float __sse_avg_colors[12] = { | |
461 0, | |
462 }; | |
463 bool use_differential[2] = {true, true}; | |
464 GetAvgColors(data, __sse_avg_colors, use_differential); | |
465 Color sub_block_avg[4]; | |
466 | |
467 // TODO(radu.velea): Remove floating point operations and use only int's + | |
468 // normal rounding and shifts for reduced Quality. | |
469 for (int i = 0, j = 1; i < 4; i += 2, j += 2) { | |
470 if (use_differential[i / 2] == false) { | |
471 sub_block_avg[i] = MakeColor444(&__sse_avg_colors[i * 3]); | |
472 sub_block_avg[j] = MakeColor444(&__sse_avg_colors[j * 3]); | |
473 } else { | |
474 sub_block_avg[i] = MakeColor555(&__sse_avg_colors[i * 3]); | |
475 sub_block_avg[j] = MakeColor555(&__sse_avg_colors[j * 3]); | |
476 } | |
477 } | |
478 | |
479 __m128i red_avg[2], green_avg[2], blue_avg[2]; | |
480 | |
481 // TODO(radu.velea): Perfect accuracy, maybe skip floating variables. | |
482 blue_avg[0] = _mm_set_epi32(static_cast<int>(__sse_avg_colors[3]), | |
483 static_cast<int>(__sse_avg_colors[3]), | |
484 static_cast<int>(__sse_avg_colors[0]), | |
485 static_cast<int>(__sse_avg_colors[0])); | |
486 | |
487 green_avg[0] = _mm_set_epi32(static_cast<int>(__sse_avg_colors[4]), | |
488 static_cast<int>(__sse_avg_colors[4]), | |
489 static_cast<int>(__sse_avg_colors[1]), | |
490 static_cast<int>(__sse_avg_colors[1])); | |
491 | |
492 red_avg[0] = _mm_set_epi32(static_cast<int>(__sse_avg_colors[5]), | |
493 static_cast<int>(__sse_avg_colors[5]), | |
494 static_cast<int>(__sse_avg_colors[2]), | |
495 static_cast<int>(__sse_avg_colors[2])); | |
496 | |
497 uint32_t vertical_error[2]; | |
498 GetVerticalError(data, blue_avg, green_avg, red_avg, vertical_error); | |
499 | |
500 // TODO(radu.velea): Perfect accuracy, maybe skip floating variables. | |
501 blue_avg[0] = _mm_set1_epi32(static_cast<int>(__sse_avg_colors[6])); | |
502 blue_avg[1] = _mm_set1_epi32(static_cast<int>(__sse_avg_colors[9])); | |
503 | |
504 green_avg[0] = _mm_set1_epi32(static_cast<int>(__sse_avg_colors[7])); | |
505 green_avg[1] = _mm_set1_epi32(static_cast<int>(__sse_avg_colors[10])); | |
506 | |
507 red_avg[0] = _mm_set1_epi32(static_cast<int>(__sse_avg_colors[8])); | |
508 red_avg[1] = _mm_set1_epi32(static_cast<int>(__sse_avg_colors[11])); | |
509 | |
510 uint32_t horizontal_error[2]; | |
511 GetHorizontalError(data, blue_avg, green_avg, red_avg, horizontal_error); | |
512 | |
513 bool flip = (horizontal_error[0] + horizontal_error[1]) < | |
514 (vertical_error[0] + vertical_error[1]); | |
515 uint32_t* expected_errors = flip ? horizontal_error : vertical_error; | |
516 | |
517 // Clear destination buffer so that we can "or" in the results. | |
518 memset(dst, 0, 8); | |
519 | |
520 WriteDiff(dst, use_differential[!!flip]); | |
521 WriteFlip(dst, flip); | |
522 | |
523 uint8_t sub_block_off_0 = flip ? 2 : 0; | |
524 uint8_t sub_block_off_1 = sub_block_off_0 + 1; | |
525 | |
526 if (use_differential[!!flip]) { | |
527 WriteColors555(dst, sub_block_avg[sub_block_off_0], | |
528 sub_block_avg[sub_block_off_1]); | |
529 } else { | |
530 WriteColors444(dst, sub_block_avg[sub_block_off_0], | |
531 sub_block_avg[sub_block_off_1]); | |
532 } | |
533 | |
534 if (!flip) { | |
535 // Transpose vertical data into horizontal lines. | |
536 __m128i tmp; | |
537 for (int i = 0; i < 4; i += 2) { | |
538 tmp = data->blue[i]; | |
539 data->blue[i] = _mm_add_epi32( | |
540 _mm_move_epi64(data->blue[i]), | |
541 _mm_shuffle_epi32(_mm_move_epi64(data->blue[i + 1]), 0x4E)); | |
542 data->blue[i + 1] = _mm_add_epi32( | |
543 _mm_move_epi64(_mm_shuffle_epi32(tmp, 0x4E)), | |
544 _mm_shuffle_epi32( | |
545 _mm_move_epi64(_mm_shuffle_epi32(data->blue[i + 1], 0x4E)), | |
546 0x4E)); | |
547 | |
548 tmp = data->green[i]; | |
549 data->green[i] = _mm_add_epi32( | |
550 _mm_move_epi64(data->green[i]), | |
551 _mm_shuffle_epi32(_mm_move_epi64(data->green[i + 1]), 0x4E)); | |
552 data->green[i + 1] = _mm_add_epi32( | |
553 _mm_move_epi64(_mm_shuffle_epi32(tmp, 0x4E)), | |
554 _mm_shuffle_epi32( | |
555 _mm_move_epi64(_mm_shuffle_epi32(data->green[i + 1], 0x4E)), | |
556 0x4E)); | |
557 | |
558 tmp = data->red[i]; | |
559 data->red[i] = _mm_add_epi32( | |
560 _mm_move_epi64(data->red[i]), | |
561 _mm_shuffle_epi32(_mm_move_epi64(data->red[i + 1]), 0x4E)); | |
562 data->red[i + 1] = _mm_add_epi32( | |
563 _mm_move_epi64(_mm_shuffle_epi32(tmp, 0x4E)), | |
564 _mm_shuffle_epi32( | |
565 _mm_move_epi64(_mm_shuffle_epi32(data->red[i + 1], 0x4E)), 0x4E)); | |
566 } | |
567 | |
568 tmp = data->blue[1]; | |
569 data->blue[1] = data->blue[2]; | |
570 data->blue[2] = tmp; | |
571 | |
572 tmp = data->green[1]; | |
573 data->green[1] = data->green[2]; | |
574 data->green[2] = tmp; | |
575 | |
576 tmp = data->red[1]; | |
577 data->red[1] = data->red[2]; | |
578 data->red[2] = tmp; | |
579 } | |
580 | |
581 // Compute luminance for the first sub block. | |
582 ComputeLuminance(dst, sub_block_avg[sub_block_off_0], 0, | |
583 g_idx_to_num[sub_block_off_0], data, | |
584 SetETC1MaxError(expected_errors[0])); | |
585 // Compute luminance for the second sub block. | |
586 ComputeLuminance(dst, sub_block_avg[sub_block_off_1], 1, | |
587 g_idx_to_num[sub_block_off_1], data, | |
588 SetETC1MaxError(expected_errors[1])); | |
589 } | |
590 | |
591 static void ExtractBlock(uint8_t* dst, const uint8_t* src, int width) { | |
592 for (int j = 0; j < 4; ++j) { | |
593 memcpy(&dst[j * 4 * 4], src, 4 * 4); | |
594 src += width * 4; | |
595 } | |
596 } | |
597 | |
598 inline bool TransposeBlock(uint8_t* block, __m128i* transposed) { | |
599 // This function transforms an incommig block of RGBA or GBRA pixels into 4 | |
600 // registers, each containing the data corresponding for a single channel. | |
601 // Ex: transposed[0] will have all the R values for a RGBA block, | |
602 // transposed[1] will have G, etc. | |
603 // The values are packed as 8 bit unsigned values in the SSE registers. | |
604 | |
605 // Before doing any work we check if the block is solid. | |
606 __m128i tmp3, tmp2, tmp1, tmp0; | |
607 __m128i test_solid = _mm_set1_epi32(*((uint32_t*)block)); | |
608 uint16_t mask = 0xFFFF; | |
609 | |
610 // a0,a1,a2,...a7, ...a15 | |
611 transposed[0] = _mm_loadu_si128((__m128i*)(block)); | |
612 // b0, b1,b2,...b7.... b15 | |
613 transposed[1] = _mm_loadu_si128((__m128i*)(block + 16)); | |
614 // c0, c1,c2,...c7....c15 | |
615 transposed[2] = _mm_loadu_si128((__m128i*)(block + 32)); | |
616 // d0,d1,d2,...d7....d15 | |
617 transposed[3] = _mm_loadu_si128((__m128i*)(block + 48)); | |
618 | |
619 for (int i = 0; i < 4; i++) { | |
620 mask &= _mm_movemask_epi8(_mm_cmpeq_epi8(transposed[i], test_solid)); | |
621 } | |
622 | |
623 if (mask == 0xFFFF) { | |
624 // Block is solid, no need to do any more work. | |
625 return false; | |
626 } | |
627 | |
628 // a0,b0, a1,b1, a2,b2, a3,b3,....a7,b7 | |
629 tmp0 = _mm_unpacklo_epi8(transposed[0], transposed[1]); | |
630 // c0,d0, c1,d1, c2,d2, c3,d3,... c7,d7 | |
631 tmp1 = _mm_unpacklo_epi8(transposed[2], transposed[3]); | |
632 // a8,b8, a9,b9, a10,b10, a11,b11,...a15,b15 | |
633 tmp2 = _mm_unpackhi_epi8(transposed[0], transposed[1]); | |
634 // c8,d8, c9,d9, c10,d10, c11,d11,...c15,d15 | |
635 tmp3 = _mm_unpackhi_epi8(transposed[2], transposed[3]); | |
636 | |
637 // a0,a8, b0,b8, a1,a9, b1,b9, ....a3,a11, b3,b11 | |
638 transposed[0] = _mm_unpacklo_epi8(tmp0, tmp2); | |
639 // a4,a12, b4,b12, a5,a13, b5,b13,....a7,a15,b7,b15 | |
640 transposed[1] = _mm_unpackhi_epi8(tmp0, tmp2); | |
641 // c0,c8, d0,d8, c1,c9, d1,d9.....d3,d11 | |
642 transposed[2] = _mm_unpacklo_epi8(tmp1, tmp3); | |
643 // c4,c12,d4,d12, c5,c13, d5,d13,....d7,d15 | |
644 transposed[3] = _mm_unpackhi_epi8(tmp1, tmp3); | |
645 | |
646 // a0,a8, b0,b8, c0,c8, d0,d8, a1,a9, b1,b9, c1,c9, d1,d9 | |
647 tmp0 = _mm_unpacklo_epi32(transposed[0], transposed[2]); | |
648 // a2,a10, b2,b10, c2,c10, d2,d10, a3,a11, b3,b11, c3,c11, d3,d11 | |
649 tmp1 = _mm_unpackhi_epi32(transposed[0], transposed[2]); | |
650 // a4,a12, b4,b12, c4,c12, d4,d12, a5,a13, b5,b13, c5,c13, d5,d13 | |
651 tmp2 = _mm_unpacklo_epi32(transposed[1], transposed[3]); | |
652 // a6,a14, b6,b14, c6,c14, d6,d14, a7,a15, b7,b15, c7,c15, d7,d15 | |
653 tmp3 = _mm_unpackhi_epi32(transposed[1], transposed[3]); | |
654 | |
655 // a0,a4, a8,a12, b0,b4, b8,b12, c0,c4, c8,c12, d0,d4, d8,d12 | |
656 transposed[0] = _mm_unpacklo_epi8(tmp0, tmp2); | |
657 // a1,a5, a9,a13, b1,b5, b9,b13, c1,c5, c9,c13, d1,d5, d9,d13 | |
658 transposed[1] = _mm_unpackhi_epi8(tmp0, tmp2); | |
659 // a2,a6, a10,a14, b2,b6, b10,b14, c2,c6, c10,c14, d2,d6, d10,d14 | |
660 transposed[2] = _mm_unpacklo_epi8(tmp1, tmp3); | |
661 // a3,a7, a11,a15, b3,b7, b11,b15, c3,c7, c11,c15, d3,d7, d11,d15 | |
662 transposed[3] = _mm_unpackhi_epi8(tmp1, tmp3); | |
663 | |
664 return true; | |
665 } | |
666 | |
667 inline void UnpackBlock(__m128i* packed, | |
668 __m128i* red, | |
669 __m128i* green, | |
670 __m128i* blue, | |
671 __m128i* alpha) { | |
672 const __m128i zero = _mm_set1_epi8(0); | |
673 __m128i tmp_low, tmp_high; | |
674 | |
675 // Unpack red. | |
676 tmp_low = _mm_unpacklo_epi8(packed[0], zero); | |
677 tmp_high = _mm_unpackhi_epi8(packed[0], zero); | |
678 | |
679 red[0] = _mm_unpacklo_epi16(tmp_low, zero); | |
680 red[1] = _mm_unpackhi_epi16(tmp_low, zero); | |
681 | |
682 red[2] = _mm_unpacklo_epi16(tmp_high, zero); | |
683 red[3] = _mm_unpackhi_epi16(tmp_high, zero); | |
684 | |
685 // Unpack green. | |
686 tmp_low = _mm_unpacklo_epi8(packed[1], zero); | |
687 tmp_high = _mm_unpackhi_epi8(packed[1], zero); | |
688 | |
689 green[0] = _mm_unpacklo_epi16(tmp_low, zero); | |
690 green[1] = _mm_unpackhi_epi16(tmp_low, zero); | |
691 | |
692 green[2] = _mm_unpacklo_epi16(tmp_high, zero); | |
693 green[3] = _mm_unpackhi_epi16(tmp_high, zero); | |
694 | |
695 // Unpack blue. | |
696 tmp_low = _mm_unpacklo_epi8(packed[2], zero); | |
697 tmp_high = _mm_unpackhi_epi8(packed[2], zero); | |
698 | |
699 blue[0] = _mm_unpacklo_epi16(tmp_low, zero); | |
700 blue[1] = _mm_unpackhi_epi16(tmp_low, zero); | |
701 | |
702 blue[2] = _mm_unpacklo_epi16(tmp_high, zero); | |
703 blue[3] = _mm_unpackhi_epi16(tmp_high, zero); | |
704 | |
705 // Unpack alpha - unused for ETC1. | |
706 tmp_low = _mm_unpacklo_epi8(packed[3], zero); | |
707 tmp_high = _mm_unpackhi_epi8(packed[3], zero); | |
708 | |
709 alpha[0] = _mm_unpacklo_epi16(tmp_low, zero); | |
710 alpha[1] = _mm_unpackhi_epi16(tmp_low, zero); | |
711 | |
712 alpha[2] = _mm_unpacklo_epi16(tmp_high, zero); | |
713 alpha[3] = _mm_unpackhi_epi16(tmp_high, zero); | |
714 } | |
715 | |
716 inline void CompressSolid(uint8_t* dst, uint8_t* block) { | |
717 // Clear destination buffer so that we can "or" in the results. | |
718 memset(dst, 0, 8); | |
719 | |
720 const float src_color_float[3] = {static_cast<float>(block[0]), | |
721 static_cast<float>(block[1]), | |
722 static_cast<float>(block[2])}; | |
723 const Color base = MakeColor555(src_color_float); | |
724 const __m128i base_v = | |
725 _mm_set_epi32(0, base.channels.r, base.channels.g, base.channels.b); | |
726 | |
727 const __m128i constant = _mm_set_epi32(0, block[2], block[1], block[0]); | |
728 __m128i lum; | |
729 __m128i colors[4]; | |
730 static const __m128i rgb = | |
731 _mm_set_epi32(0, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF); | |
732 | |
733 WriteDiff(dst, true); | |
734 WriteFlip(dst, false); | |
735 | |
736 WriteColors555(dst, base, base); | |
737 | |
738 uint8_t best_tbl_idx = 0; | |
739 uint8_t best_mod_idx = 0; | |
740 uint32_t best_mod_err = INT32_MAX; | |
741 | |
742 for (unsigned int tbl_idx = 0; tbl_idx < 8; ++tbl_idx) { | |
743 lum = _mm_set_epi32( | |
744 g_codeword_tables[tbl_idx][3], g_codeword_tables[tbl_idx][2], | |
745 g_codeword_tables[tbl_idx][1], g_codeword_tables[tbl_idx][0]); | |
746 colors[0] = AddAndClamp(base_v, _mm_shuffle_epi32(lum, 0x0)); | |
747 colors[1] = AddAndClamp(base_v, _mm_shuffle_epi32(lum, 0x55)); | |
748 colors[2] = AddAndClamp(base_v, _mm_shuffle_epi32(lum, 0xAA)); | |
749 colors[3] = AddAndClamp(base_v, _mm_shuffle_epi32(lum, 0xFF)); | |
750 | |
751 for (int i = 0; i < 4; i++) { | |
752 uint32_t mod_err = | |
753 SumSSE(GetColorErrorSSE(constant, _mm_and_si128(colors[i], rgb))); | |
754 colors[i] = _mm_and_si128(colors[i], rgb); | |
755 if (mod_err < best_mod_err) { | |
756 best_tbl_idx = tbl_idx; | |
757 best_mod_idx = i; | |
758 best_mod_err = mod_err; | |
759 | |
760 if (mod_err == 0) { | |
761 break; // We cannot do any better than this. | |
762 } | |
763 } | |
764 } | |
765 } | |
766 | |
767 WriteCodewordTable(dst, 0, best_tbl_idx); | |
768 WriteCodewordTable(dst, 1, best_tbl_idx); | |
769 | |
770 uint8_t pix_idx = g_mod_to_pix[best_mod_idx]; | |
771 uint32_t lsb = pix_idx & 0x1; | |
772 uint32_t msb = pix_idx >> 1; | |
773 | |
774 uint32_t pix_data = 0; | |
775 for (unsigned int i = 0; i < 2; ++i) { | |
776 for (unsigned int j = 0; j < 8; ++j) { | |
777 // Obtain the texel number as specified in the standard. | |
778 int texel_num = g_idx_to_num[i][j]; | |
779 pix_data |= msb << (texel_num + 16); | |
780 pix_data |= lsb << (texel_num); | |
781 } | |
782 } | |
783 | |
784 WritePixelData(dst, pix_data); | |
785 } | |
786 | |
787 } // namespace | |
788 | |
789 void TextureCompressorETC1SSE::Compress(const uint8_t* src, | |
790 uint8_t* dst, | |
791 int width, | |
792 int height, | |
793 Quality quality) { | |
794 DCHECK_GE(width, 4); | |
795 DCHECK_EQ((width & 3), 0); | |
796 DCHECK_GE(height, 4); | |
797 DCHECK_EQ((height & 3), 0); | |
798 | |
799 ALIGNAS(16) uint8_t block[64]; | |
800 __m128i packed[4]; | |
801 __m128i red[4], green[4], blue[4], alpha[4]; | |
802 __sse_data data; | |
803 | |
804 for (int y = 0; y < height; y += 4, src += width * 4 * 4) { | |
805 for (int x = 0; x < width; x += 4, dst += 8) { | |
806 ExtractBlock(block, src + x * 4, width); | |
807 if (TransposeBlock(block, packed) == false) { | |
808 CompressSolid(dst, block); | |
809 } else { | |
810 UnpackBlock(packed, blue, green, red, alpha); | |
811 | |
812 data.block = block; | |
813 data.packed = packed; | |
814 data.red = red; | |
815 data.blue = blue; | |
816 data.green = green; | |
817 | |
818 CompressBlock(dst, &data); | |
819 } | |
820 } | |
821 } | |
822 } | |
823 | |
824 } // namespace cc | |
OLD | NEW |