1 // Copyright 2014 Google Inc. All Rights Reserved.
2 //
3 // Use of this source code is governed by a BSD-style license
4 // that can be found in the COPYING file in the root of the source
5 // tree. An additional intellectual property rights grant can be found
6 // in the file PATENTS. All contributing project authors may
7 // be found in the AUTHORS file in the root of the source tree.
8 // -----------------------------------------------------------------------------
9 //
10 // SSE2 variant of methods for lossless decoder
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13
14 #include "src/dsp/dsp.h"
15
16 #if defined(WEBP_USE_SSE2)
17
18 #include "src/dsp/common_sse2.h"
19 #include "src/dsp/lossless.h"
20 #include "src/dsp/lossless_common.h"
21 #include <emmintrin.h>
22
23 //------------------------------------------------------------------------------
24 // Predictor Transform
25
ClampedAddSubtractFull_SSE2(uint32_t c0,uint32_t c1,uint32_t c2)26 static WEBP_INLINE uint32_t ClampedAddSubtractFull_SSE2(uint32_t c0,
27 uint32_t c1,
28 uint32_t c2) {
29 const __m128i zero = _mm_setzero_si128();
30 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
31 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
32 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
33 const __m128i V1 = _mm_add_epi16(C0, C1);
34 const __m128i V2 = _mm_sub_epi16(V1, C2);
35 const __m128i b = _mm_packus_epi16(V2, V2);
36 const uint32_t output = _mm_cvtsi128_si32(b);
37 return output;
38 }
39
ClampedAddSubtractHalf_SSE2(uint32_t c0,uint32_t c1,uint32_t c2)40 static WEBP_INLINE uint32_t ClampedAddSubtractHalf_SSE2(uint32_t c0,
41 uint32_t c1,
42 uint32_t c2) {
43 const __m128i zero = _mm_setzero_si128();
44 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
45 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
46 const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
47 const __m128i avg = _mm_add_epi16(C1, C0);
48 const __m128i A0 = _mm_srli_epi16(avg, 1);
49 const __m128i A1 = _mm_sub_epi16(A0, B0);
50 const __m128i BgtA = _mm_cmpgt_epi16(B0, A0);
51 const __m128i A2 = _mm_sub_epi16(A1, BgtA);
52 const __m128i A3 = _mm_srai_epi16(A2, 1);
53 const __m128i A4 = _mm_add_epi16(A0, A3);
54 const __m128i A5 = _mm_packus_epi16(A4, A4);
55 const uint32_t output = _mm_cvtsi128_si32(A5);
56 return output;
57 }
58
Select_SSE2(uint32_t a,uint32_t b,uint32_t c)59 static WEBP_INLINE uint32_t Select_SSE2(uint32_t a, uint32_t b, uint32_t c) {
60 int pa_minus_pb;
61 const __m128i zero = _mm_setzero_si128();
62 const __m128i A0 = _mm_cvtsi32_si128(a);
63 const __m128i B0 = _mm_cvtsi32_si128(b);
64 const __m128i C0 = _mm_cvtsi32_si128(c);
65 const __m128i AC0 = _mm_subs_epu8(A0, C0);
66 const __m128i CA0 = _mm_subs_epu8(C0, A0);
67 const __m128i BC0 = _mm_subs_epu8(B0, C0);
68 const __m128i CB0 = _mm_subs_epu8(C0, B0);
69 const __m128i AC = _mm_or_si128(AC0, CA0);
70 const __m128i BC = _mm_or_si128(BC0, CB0);
71 const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c|
72 const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c|
73 const __m128i diff = _mm_sub_epi16(pb, pa);
74 {
75 int16_t out[8];
76 _mm_storeu_si128((__m128i*)out, diff);
77 pa_minus_pb = out[0] + out[1] + out[2] + out[3];
78 }
79 return (pa_minus_pb <= 0) ? a : b;
80 }
81
Average2_m128i(const __m128i * const a0,const __m128i * const a1,__m128i * const avg)82 static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
83 const __m128i* const a1,
84 __m128i* const avg) {
85 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
86 const __m128i ones = _mm_set1_epi8(1);
87 const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
88 const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
89 *avg = _mm_sub_epi8(avg1, one);
90 }
91
Average2_uint32_SSE2(const uint32_t a0,const uint32_t a1,__m128i * const avg)92 static WEBP_INLINE void Average2_uint32_SSE2(const uint32_t a0,
93 const uint32_t a1,
94 __m128i* const avg) {
95 // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
96 const __m128i ones = _mm_set1_epi8(1);
97 const __m128i A0 = _mm_cvtsi32_si128(a0);
98 const __m128i A1 = _mm_cvtsi32_si128(a1);
99 const __m128i avg1 = _mm_avg_epu8(A0, A1);
100 const __m128i one = _mm_and_si128(_mm_xor_si128(A0, A1), ones);
101 *avg = _mm_sub_epi8(avg1, one);
102 }
103
Average2_uint32_16_SSE2(uint32_t a0,uint32_t a1)104 static WEBP_INLINE __m128i Average2_uint32_16_SSE2(uint32_t a0, uint32_t a1) {
105 const __m128i zero = _mm_setzero_si128();
106 const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero);
107 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
108 const __m128i sum = _mm_add_epi16(A1, A0);
109 return _mm_srli_epi16(sum, 1);
110 }
111
Average2_SSE2(uint32_t a0,uint32_t a1)112 static WEBP_INLINE uint32_t Average2_SSE2(uint32_t a0, uint32_t a1) {
113 __m128i output;
114 Average2_uint32_SSE2(a0, a1, &output);
115 return _mm_cvtsi128_si32(output);
116 }
117
Average3_SSE2(uint32_t a0,uint32_t a1,uint32_t a2)118 static WEBP_INLINE uint32_t Average3_SSE2(uint32_t a0, uint32_t a1,
119 uint32_t a2) {
120 const __m128i zero = _mm_setzero_si128();
121 const __m128i avg1 = Average2_uint32_16_SSE2(a0, a2);
122 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero);
123 const __m128i sum = _mm_add_epi16(avg1, A1);
124 const __m128i avg2 = _mm_srli_epi16(sum, 1);
125 const __m128i A2 = _mm_packus_epi16(avg2, avg2);
126 const uint32_t output = _mm_cvtsi128_si32(A2);
127 return output;
128 }
129
Average4_SSE2(uint32_t a0,uint32_t a1,uint32_t a2,uint32_t a3)130 static WEBP_INLINE uint32_t Average4_SSE2(uint32_t a0, uint32_t a1,
131 uint32_t a2, uint32_t a3) {
132 const __m128i avg1 = Average2_uint32_16_SSE2(a0, a1);
133 const __m128i avg2 = Average2_uint32_16_SSE2(a2, a3);
134 const __m128i sum = _mm_add_epi16(avg2, avg1);
135 const __m128i avg3 = _mm_srli_epi16(sum, 1);
136 const __m128i A0 = _mm_packus_epi16(avg3, avg3);
137 const uint32_t output = _mm_cvtsi128_si32(A0);
138 return output;
139 }
140
Predictor5_SSE2(const uint32_t * const left,const uint32_t * const top)141 static uint32_t Predictor5_SSE2(const uint32_t* const left,
142 const uint32_t* const top) {
143 const uint32_t pred = Average3_SSE2(*left, top[0], top[1]);
144 return pred;
145 }
Predictor6_SSE2(const uint32_t * const left,const uint32_t * const top)146 static uint32_t Predictor6_SSE2(const uint32_t* const left,
147 const uint32_t* const top) {
148 const uint32_t pred = Average2_SSE2(*left, top[-1]);
149 return pred;
150 }
Predictor7_SSE2(const uint32_t * const left,const uint32_t * const top)151 static uint32_t Predictor7_SSE2(const uint32_t* const left,
152 const uint32_t* const top) {
153 const uint32_t pred = Average2_SSE2(*left, top[0]);
154 return pred;
155 }
Predictor8_SSE2(const uint32_t * const left,const uint32_t * const top)156 static uint32_t Predictor8_SSE2(const uint32_t* const left,
157 const uint32_t* const top) {
158 const uint32_t pred = Average2_SSE2(top[-1], top[0]);
159 (void)left;
160 return pred;
161 }
Predictor9_SSE2(const uint32_t * const left,const uint32_t * const top)162 static uint32_t Predictor9_SSE2(const uint32_t* const left,
163 const uint32_t* const top) {
164 const uint32_t pred = Average2_SSE2(top[0], top[1]);
165 (void)left;
166 return pred;
167 }
Predictor10_SSE2(const uint32_t * const left,const uint32_t * const top)168 static uint32_t Predictor10_SSE2(const uint32_t* const left,
169 const uint32_t* const top) {
170 const uint32_t pred = Average4_SSE2(*left, top[-1], top[0], top[1]);
171 return pred;
172 }
Predictor11_SSE2(const uint32_t * const left,const uint32_t * const top)173 static uint32_t Predictor11_SSE2(const uint32_t* const left,
174 const uint32_t* const top) {
175 const uint32_t pred = Select_SSE2(top[0], *left, top[-1]);
176 return pred;
177 }
Predictor12_SSE2(const uint32_t * const left,const uint32_t * const top)178 static uint32_t Predictor12_SSE2(const uint32_t* const left,
179 const uint32_t* const top) {
180 const uint32_t pred = ClampedAddSubtractFull_SSE2(*left, top[0], top[-1]);
181 return pred;
182 }
Predictor13_SSE2(const uint32_t * const left,const uint32_t * const top)183 static uint32_t Predictor13_SSE2(const uint32_t* const left,
184 const uint32_t* const top) {
185 const uint32_t pred = ClampedAddSubtractHalf_SSE2(*left, top[0], top[-1]);
186 return pred;
187 }
188
189 // Batch versions of those functions.
190
191 // Predictor0: ARGB_BLACK.
PredictorAdd0_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)192 static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
193 int num_pixels, uint32_t* out) {
194 int i;
195 const __m128i black = _mm_set1_epi32(ARGB_BLACK);
196 for (i = 0; i + 4 <= num_pixels; i += 4) {
197 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
198 const __m128i res = _mm_add_epi8(src, black);
199 _mm_storeu_si128((__m128i*)&out[i], res);
200 }
201 if (i != num_pixels) {
202 VP8LPredictorsAdd_C[0](in + i, NULL, num_pixels - i, out + i);
203 }
204 (void)upper;
205 }
206
207 // Predictor1: left.
PredictorAdd1_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)208 static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper,
209 int num_pixels, uint32_t* out) {
210 int i;
211 __m128i prev = _mm_set1_epi32(out[-1]);
212 for (i = 0; i + 4 <= num_pixels; i += 4) {
213 // a | b | c | d
214 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
215 // 0 | a | b | c
216 const __m128i shift0 = _mm_slli_si128(src, 4);
217 // a | a + b | b + c | c + d
218 const __m128i sum0 = _mm_add_epi8(src, shift0);
219 // 0 | 0 | a | a + b
220 const __m128i shift1 = _mm_slli_si128(sum0, 8);
221 // a | a + b | a + b + c | a + b + c + d
222 const __m128i sum1 = _mm_add_epi8(sum0, shift1);
223 const __m128i res = _mm_add_epi8(sum1, prev);
224 _mm_storeu_si128((__m128i*)&out[i], res);
225 // replicate prev output on the four lanes
226 prev = _mm_shuffle_epi32(res, (3 << 0) | (3 << 2) | (3 << 4) | (3 << 6));
227 }
228 if (i != num_pixels) {
229 VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i);
230 }
231 }
232
233 // Macro that adds 32-bit integers from IN using mod 256 arithmetic
234 // per 8 bit channel.
235 #define GENERATE_PREDICTOR_1(X, IN) \
236 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
237 int num_pixels, uint32_t* out) { \
238 int i; \
239 for (i = 0; i + 4 <= num_pixels; i += 4) { \
240 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
241 const __m128i other = _mm_loadu_si128((const __m128i*)&(IN)); \
242 const __m128i res = _mm_add_epi8(src, other); \
243 _mm_storeu_si128((__m128i*)&out[i], res); \
244 } \
245 if (i != num_pixels) { \
246 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
247 } \
248 }
249
250 // Predictor2: Top.
251 GENERATE_PREDICTOR_1(2, upper[i])
252 // Predictor3: Top-right.
253 GENERATE_PREDICTOR_1(3, upper[i + 1])
254 // Predictor4: Top-left.
255 GENERATE_PREDICTOR_1(4, upper[i - 1])
256 #undef GENERATE_PREDICTOR_1
257
258 // Due to averages with integers, values cannot be accumulated in parallel for
259 // predictors 5 to 7.
GENERATE_PREDICTOR_ADD(Predictor5_SSE2,PredictorAdd5_SSE2)260 GENERATE_PREDICTOR_ADD(Predictor5_SSE2, PredictorAdd5_SSE2)
261 GENERATE_PREDICTOR_ADD(Predictor6_SSE2, PredictorAdd6_SSE2)
262 GENERATE_PREDICTOR_ADD(Predictor7_SSE2, PredictorAdd7_SSE2)
263
264 #define GENERATE_PREDICTOR_2(X, IN) \
265 static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
266 int num_pixels, uint32_t* out) { \
267 int i; \
268 for (i = 0; i + 4 <= num_pixels; i += 4) { \
269 const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN)); \
270 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]); \
271 const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]); \
272 __m128i avg, res; \
273 Average2_m128i(&T, &Tother, &avg); \
274 res = _mm_add_epi8(avg, src); \
275 _mm_storeu_si128((__m128i*)&out[i], res); \
276 } \
277 if (i != num_pixels) { \
278 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
279 } \
280 }
281 // Predictor8: average TL T.
282 GENERATE_PREDICTOR_2(8, upper[i - 1])
283 // Predictor9: average T TR.
284 GENERATE_PREDICTOR_2(9, upper[i + 1])
285 #undef GENERATE_PREDICTOR_2
286
287 // Predictor10: average of (average of (L,TL), average of (T, TR)).
288 #define DO_PRED10(OUT) do { \
289 __m128i avgLTL, avg; \
290 Average2_m128i(&L, &TL, &avgLTL); \
291 Average2_m128i(&avgTTR, &avgLTL, &avg); \
292 L = _mm_add_epi8(avg, src); \
293 out[i + (OUT)] = _mm_cvtsi128_si32(L); \
294 } while (0)
295
296 #define DO_PRED10_SHIFT do { \
297 /* Rotate the pre-computed values for the next iteration.*/ \
298 avgTTR = _mm_srli_si128(avgTTR, 4); \
299 TL = _mm_srli_si128(TL, 4); \
300 src = _mm_srli_si128(src, 4); \
301 } while (0)
302
303 static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
304 int num_pixels, uint32_t* out) {
305 int i;
306 __m128i L = _mm_cvtsi32_si128(out[-1]);
307 for (i = 0; i + 4 <= num_pixels; i += 4) {
308 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
309 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
310 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
311 const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
312 __m128i avgTTR;
313 Average2_m128i(&T, &TR, &avgTTR);
314 DO_PRED10(0);
315 DO_PRED10_SHIFT;
316 DO_PRED10(1);
317 DO_PRED10_SHIFT;
318 DO_PRED10(2);
319 DO_PRED10_SHIFT;
320 DO_PRED10(3);
321 }
322 if (i != num_pixels) {
323 VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
324 }
325 }
326 #undef DO_PRED10
327 #undef DO_PRED10_SHIFT
328
329 // Predictor11: select.
330 #define DO_PRED11(OUT) do { \
331 const __m128i L_lo = _mm_unpacklo_epi32(L, T); \
332 const __m128i TL_lo = _mm_unpacklo_epi32(TL, T); \
333 const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); /* pb = sum |L-TL|*/ \
334 const __m128i mask = _mm_cmpgt_epi32(pb, pa); \
335 const __m128i A = _mm_and_si128(mask, L); \
336 const __m128i B = _mm_andnot_si128(mask, T); \
337 const __m128i pred = _mm_or_si128(A, B); /* pred = (pa > b)? L : T*/ \
338 L = _mm_add_epi8(src, pred); \
339 out[i + (OUT)] = _mm_cvtsi128_si32(L); \
340 } while (0)
341
342 #define DO_PRED11_SHIFT do { \
343 /* Shift the pre-computed value for the next iteration.*/ \
344 T = _mm_srli_si128(T, 4); \
345 TL = _mm_srli_si128(TL, 4); \
346 src = _mm_srli_si128(src, 4); \
347 pa = _mm_srli_si128(pa, 4); \
348 } while (0)
349
PredictorAdd11_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)350 static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
351 int num_pixels, uint32_t* out) {
352 int i;
353 __m128i pa;
354 __m128i L = _mm_cvtsi32_si128(out[-1]);
355 for (i = 0; i + 4 <= num_pixels; i += 4) {
356 __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
357 __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
358 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
359 {
360 // We can unpack with any value on the upper 32 bits, provided it's the
361 // same on both operands (so that their sum of abs diff is zero). Here we
362 // use T.
363 const __m128i T_lo = _mm_unpacklo_epi32(T, T);
364 const __m128i TL_lo = _mm_unpacklo_epi32(TL, T);
365 const __m128i T_hi = _mm_unpackhi_epi32(T, T);
366 const __m128i TL_hi = _mm_unpackhi_epi32(TL, T);
367 const __m128i s_lo = _mm_sad_epu8(T_lo, TL_lo);
368 const __m128i s_hi = _mm_sad_epu8(T_hi, TL_hi);
369 pa = _mm_packs_epi32(s_lo, s_hi); // pa = sum |T-TL|
370 }
371 DO_PRED11(0);
372 DO_PRED11_SHIFT;
373 DO_PRED11(1);
374 DO_PRED11_SHIFT;
375 DO_PRED11(2);
376 DO_PRED11_SHIFT;
377 DO_PRED11(3);
378 }
379 if (i != num_pixels) {
380 VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
381 }
382 }
383 #undef DO_PRED11
384 #undef DO_PRED11_SHIFT
385
386 // Predictor12: ClampedAddSubtractFull.
387 #define DO_PRED12(DIFF, LANE, OUT) do { \
388 const __m128i all = _mm_add_epi16(L, (DIFF)); \
389 const __m128i alls = _mm_packus_epi16(all, all); \
390 const __m128i res = _mm_add_epi8(src, alls); \
391 out[i + (OUT)] = _mm_cvtsi128_si32(res); \
392 L = _mm_unpacklo_epi8(res, zero); \
393 } while (0)
394
395 #define DO_PRED12_SHIFT(DIFF, LANE) do { \
396 /* Shift the pre-computed value for the next iteration.*/ \
397 if ((LANE) == 0) (DIFF) = _mm_srli_si128((DIFF), 8); \
398 src = _mm_srli_si128(src, 4); \
399 } while (0)
400
PredictorAdd12_SSE2(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)401 static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
402 int num_pixels, uint32_t* out) {
403 int i;
404 const __m128i zero = _mm_setzero_si128();
405 const __m128i L8 = _mm_cvtsi32_si128(out[-1]);
406 __m128i L = _mm_unpacklo_epi8(L8, zero);
407 for (i = 0; i + 4 <= num_pixels; i += 4) {
408 // Load 4 pixels at a time.
409 __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
410 const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
411 const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
412 const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
413 const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
414 const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
415 const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
416 __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
417 __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
418 DO_PRED12(diff_lo, 0, 0);
419 DO_PRED12_SHIFT(diff_lo, 0);
420 DO_PRED12(diff_lo, 1, 1);
421 DO_PRED12_SHIFT(diff_lo, 1);
422 DO_PRED12(diff_hi, 0, 2);
423 DO_PRED12_SHIFT(diff_hi, 0);
424 DO_PRED12(diff_hi, 1, 3);
425 }
426 if (i != num_pixels) {
427 VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
428 }
429 }
430 #undef DO_PRED12
431 #undef DO_PRED12_SHIFT
432
433 // Due to averages with integers, values cannot be accumulated in parallel for
434 // predictors 13.
GENERATE_PREDICTOR_ADD(Predictor13_SSE2,PredictorAdd13_SSE2)435 GENERATE_PREDICTOR_ADD(Predictor13_SSE2, PredictorAdd13_SSE2)
436
437 //------------------------------------------------------------------------------
438 // Subtract-Green Transform
439
440 static void AddGreenToBlueAndRed_SSE2(const uint32_t* const src, int num_pixels,
441 uint32_t* dst) {
442 int i;
443 for (i = 0; i + 4 <= num_pixels; i += 4) {
444 const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
445 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g
446 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
447 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g
448 const __m128i out = _mm_add_epi8(in, C);
449 _mm_storeu_si128((__m128i*)&dst[i], out);
450 }
451 // fallthrough and finish off with plain-C
452 if (i != num_pixels) {
453 VP8LAddGreenToBlueAndRed_C(src + i, num_pixels - i, dst + i);
454 }
455 }
456
457 //------------------------------------------------------------------------------
458 // Color Transform
459
TransformColorInverse_SSE2(const VP8LMultipliers * const m,const uint32_t * const src,int num_pixels,uint32_t * dst)460 static void TransformColorInverse_SSE2(const VP8LMultipliers* const m,
461 const uint32_t* const src,
462 int num_pixels, uint32_t* dst) {
463 // sign-extended multiplying constants, pre-shifted by 5.
464 #define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend
465 #define MK_CST_16(HI, LO) \
466 _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
467 const __m128i mults_rb = MK_CST_16(CST(green_to_red_), CST(green_to_blue_));
468 const __m128i mults_b2 = MK_CST_16(CST(red_to_blue_), 0);
469 #undef MK_CST_16
470 #undef CST
471 const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks
472 int i;
473 for (i = 0; i + 4 <= num_pixels; i += 4) {
474 const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
475 const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0
476 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
477 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0
478 const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1
479 const __m128i E = _mm_add_epi8(in, D); // x r' x b'
480 const __m128i F = _mm_slli_epi16(E, 8); // r' 0 b' 0
481 const __m128i G = _mm_mulhi_epi16(F, mults_b2); // x db2 0 0
482 const __m128i H = _mm_srli_epi32(G, 8); // 0 x db2 0
483 const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0
484 const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b''
485 const __m128i out = _mm_or_si128(J, A);
486 _mm_storeu_si128((__m128i*)&dst[i], out);
487 }
488 // Fall-back to C-version for left-overs.
489 if (i != num_pixels) {
490 VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
491 }
492 }
493
494 //------------------------------------------------------------------------------
495 // Color-space conversion functions
496
ConvertBGRAToRGB_SSE2(const uint32_t * src,int num_pixels,uint8_t * dst)497 static void ConvertBGRAToRGB_SSE2(const uint32_t* src, int num_pixels,
498 uint8_t* dst) {
499 const __m128i* in = (const __m128i*)src;
500 __m128i* out = (__m128i*)dst;
501
502 while (num_pixels >= 32) {
503 // Load the BGRA buffers.
504 __m128i in0 = _mm_loadu_si128(in + 0);
505 __m128i in1 = _mm_loadu_si128(in + 1);
506 __m128i in2 = _mm_loadu_si128(in + 2);
507 __m128i in3 = _mm_loadu_si128(in + 3);
508 __m128i in4 = _mm_loadu_si128(in + 4);
509 __m128i in5 = _mm_loadu_si128(in + 5);
510 __m128i in6 = _mm_loadu_si128(in + 6);
511 __m128i in7 = _mm_loadu_si128(in + 7);
512 VP8L32bToPlanar_SSE2(&in0, &in1, &in2, &in3);
513 VP8L32bToPlanar_SSE2(&in4, &in5, &in6, &in7);
514 // At this points, in1/in5 contains red only, in2/in6 green only ...
515 // Pack the colors in 24b RGB.
516 VP8PlanarTo24b_SSE2(&in1, &in5, &in2, &in6, &in3, &in7);
517 _mm_storeu_si128(out + 0, in1);
518 _mm_storeu_si128(out + 1, in5);
519 _mm_storeu_si128(out + 2, in2);
520 _mm_storeu_si128(out + 3, in6);
521 _mm_storeu_si128(out + 4, in3);
522 _mm_storeu_si128(out + 5, in7);
523 in += 8;
524 out += 6;
525 num_pixels -= 32;
526 }
527 // left-overs
528 if (num_pixels > 0) {
529 VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
530 }
531 }
532
ConvertBGRAToRGBA_SSE2(const uint32_t * src,int num_pixels,uint8_t * dst)533 static void ConvertBGRAToRGBA_SSE2(const uint32_t* src,
534 int num_pixels, uint8_t* dst) {
535 const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ffu);
536 const __m128i* in = (const __m128i*)src;
537 __m128i* out = (__m128i*)dst;
538 while (num_pixels >= 8) {
539 const __m128i A1 = _mm_loadu_si128(in++);
540 const __m128i A2 = _mm_loadu_si128(in++);
541 const __m128i B1 = _mm_and_si128(A1, red_blue_mask); // R 0 B 0
542 const __m128i B2 = _mm_and_si128(A2, red_blue_mask); // R 0 B 0
543 const __m128i C1 = _mm_andnot_si128(red_blue_mask, A1); // 0 G 0 A
544 const __m128i C2 = _mm_andnot_si128(red_blue_mask, A2); // 0 G 0 A
545 const __m128i D1 = _mm_shufflelo_epi16(B1, _MM_SHUFFLE(2, 3, 0, 1));
546 const __m128i D2 = _mm_shufflelo_epi16(B2, _MM_SHUFFLE(2, 3, 0, 1));
547 const __m128i E1 = _mm_shufflehi_epi16(D1, _MM_SHUFFLE(2, 3, 0, 1));
548 const __m128i E2 = _mm_shufflehi_epi16(D2, _MM_SHUFFLE(2, 3, 0, 1));
549 const __m128i F1 = _mm_or_si128(E1, C1);
550 const __m128i F2 = _mm_or_si128(E2, C2);
551 _mm_storeu_si128(out++, F1);
552 _mm_storeu_si128(out++, F2);
553 num_pixels -= 8;
554 }
555 // left-overs
556 if (num_pixels > 0) {
557 VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
558 }
559 }
560
ConvertBGRAToRGBA4444_SSE2(const uint32_t * src,int num_pixels,uint8_t * dst)561 static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* src,
562 int num_pixels, uint8_t* dst) {
563 const __m128i mask_0x0f = _mm_set1_epi8(0x0f);
564 const __m128i mask_0xf0 = _mm_set1_epi8(0xf0);
565 const __m128i* in = (const __m128i*)src;
566 __m128i* out = (__m128i*)dst;
567 while (num_pixels >= 8) {
568 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3
569 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7
570 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4...
571 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6...
572 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6...
573 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7...
574 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7
575 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7
576 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7
577 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7
578 const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7-
579 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7
580 const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7-
581 const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7
582 const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0
583 #if (WEBP_SWAP_16BIT_CSP == 1)
584 const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7
585 #else
586 const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7
587 #endif
588 _mm_storeu_si128(out++, rgba);
589 num_pixels -= 8;
590 }
591 // left-overs
592 if (num_pixels > 0) {
593 VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
594 }
595 }
596
ConvertBGRAToRGB565_SSE2(const uint32_t * src,int num_pixels,uint8_t * dst)597 static void ConvertBGRAToRGB565_SSE2(const uint32_t* src,
598 int num_pixels, uint8_t* dst) {
599 const __m128i mask_0xe0 = _mm_set1_epi8(0xe0);
600 const __m128i mask_0xf8 = _mm_set1_epi8(0xf8);
601 const __m128i mask_0x07 = _mm_set1_epi8(0x07);
602 const __m128i* in = (const __m128i*)src;
603 __m128i* out = (__m128i*)dst;
604 while (num_pixels >= 8) {
605 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3
606 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7
607 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4...
608 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6...
609 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6...
610 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7...
611 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7
612 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7
613 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7
614 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7
615 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7
616 const __m128i g_lo1 = _mm_srli_epi16(ga0, 5);
617 const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b)
618 const __m128i g_hi1 = _mm_slli_epi16(ga0, 3);
619 const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b)
620 const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0
621 const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx
622 const __m128i b1 = _mm_srli_epi16(b0, 3);
623 const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx
624 #if (WEBP_SWAP_16BIT_CSP == 1)
625 const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7
626 #else
627 const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7
628 #endif
629 _mm_storeu_si128(out++, rgba);
630 num_pixels -= 8;
631 }
632 // left-overs
633 if (num_pixels > 0) {
634 VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
635 }
636 }
637
ConvertBGRAToBGR_SSE2(const uint32_t * src,int num_pixels,uint8_t * dst)638 static void ConvertBGRAToBGR_SSE2(const uint32_t* src,
639 int num_pixels, uint8_t* dst) {
640 const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff);
641 const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0);
642 const __m128i* in = (const __m128i*)src;
643 const uint8_t* const end = dst + num_pixels * 3;
644 // the last storel_epi64 below writes 8 bytes starting at offset 18
645 while (dst + 26 <= end) {
646 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3
647 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7
648 const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0
649 const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0
650 const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0
651 const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0
652 const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00
653 const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00
654 const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00
655 const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00
656 const __m128i c2 = _mm_srli_si128(c0, 8);
657 const __m128i c6 = _mm_srli_si128(c4, 8);
658 _mm_storel_epi64((__m128i*)(dst + 0), c0);
659 _mm_storel_epi64((__m128i*)(dst + 6), c2);
660 _mm_storel_epi64((__m128i*)(dst + 12), c4);
661 _mm_storel_epi64((__m128i*)(dst + 18), c6);
662 dst += 24;
663 num_pixels -= 8;
664 }
665 // left-overs
666 if (num_pixels > 0) {
667 VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst);
668 }
669 }
670
671 //------------------------------------------------------------------------------
672 // Entry point
673
674 extern void VP8LDspInitSSE2(void);
675
VP8LDspInitSSE2(void)676 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) {
677 VP8LPredictors[5] = Predictor5_SSE2;
678 VP8LPredictors[6] = Predictor6_SSE2;
679 VP8LPredictors[7] = Predictor7_SSE2;
680 VP8LPredictors[8] = Predictor8_SSE2;
681 VP8LPredictors[9] = Predictor9_SSE2;
682 VP8LPredictors[10] = Predictor10_SSE2;
683 VP8LPredictors[11] = Predictor11_SSE2;
684 VP8LPredictors[12] = Predictor12_SSE2;
685 VP8LPredictors[13] = Predictor13_SSE2;
686
687 VP8LPredictorsAdd[0] = PredictorAdd0_SSE2;
688 VP8LPredictorsAdd[1] = PredictorAdd1_SSE2;
689 VP8LPredictorsAdd[2] = PredictorAdd2_SSE2;
690 VP8LPredictorsAdd[3] = PredictorAdd3_SSE2;
691 VP8LPredictorsAdd[4] = PredictorAdd4_SSE2;
692 VP8LPredictorsAdd[5] = PredictorAdd5_SSE2;
693 VP8LPredictorsAdd[6] = PredictorAdd6_SSE2;
694 VP8LPredictorsAdd[7] = PredictorAdd7_SSE2;
695 VP8LPredictorsAdd[8] = PredictorAdd8_SSE2;
696 VP8LPredictorsAdd[9] = PredictorAdd9_SSE2;
697 VP8LPredictorsAdd[10] = PredictorAdd10_SSE2;
698 VP8LPredictorsAdd[11] = PredictorAdd11_SSE2;
699 VP8LPredictorsAdd[12] = PredictorAdd12_SSE2;
700 VP8LPredictorsAdd[13] = PredictorAdd13_SSE2;
701
702 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_SSE2;
703 VP8LTransformColorInverse = TransformColorInverse_SSE2;
704
705 VP8LConvertBGRAToRGB = ConvertBGRAToRGB_SSE2;
706 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_SSE2;
707 VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444_SSE2;
708 VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565_SSE2;
709 VP8LConvertBGRAToBGR = ConvertBGRAToBGR_SSE2;
710 }
711
712 #else // !WEBP_USE_SSE2
713
714 WEBP_DSP_INIT_STUB(VP8LDspInitSSE2)
715
716 #endif // WEBP_USE_SSE2
717