CoCalc -- lossless

GitHub Repository: godotengine/godot
Path: blob/master/thirdparty/libwebp/src/dsp/lossless_sse2.c
⁹⁹¹⁸ views
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 ([email protected])
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

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((int)c0), zero);
31
  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c1), zero);
32
  const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)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
  return (uint32_t)_mm_cvtsi128_si32(b);
37
}
38

39
static WEBP_INLINE uint32_t ClampedAddSubtractHalf_SSE2(uint32_t c0,
40
                                                        uint32_t c1,
41
                                                        uint32_t c2) {
42
  const __m128i zero = _mm_setzero_si128();
43
  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c0), zero);
44
  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c1), zero);
45
  const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)c2), zero);
46
  const __m128i avg = _mm_add_epi16(C1, C0);
47
  const __m128i A0 = _mm_srli_epi16(avg, 1);
48
  const __m128i A1 = _mm_sub_epi16(A0, B0);
49
  const __m128i BgtA = _mm_cmpgt_epi16(B0, A0);
50
  const __m128i A2 = _mm_sub_epi16(A1, BgtA);
51
  const __m128i A3 = _mm_srai_epi16(A2, 1);
52
  const __m128i A4 = _mm_add_epi16(A0, A3);
53
  const __m128i A5 = _mm_packus_epi16(A4, A4);
54
  return (uint32_t)_mm_cvtsi128_si32(A5);
55
}
56

57
static WEBP_INLINE uint32_t Select_SSE2(uint32_t a, uint32_t b, uint32_t c) {
58
  int pa_minus_pb;
59
  const __m128i zero = _mm_setzero_si128();
60
  const __m128i A0 = _mm_cvtsi32_si128((int)a);
61
  const __m128i B0 = _mm_cvtsi32_si128((int)b);
62
  const __m128i C0 = _mm_cvtsi32_si128((int)c);
63
  const __m128i AC0 = _mm_subs_epu8(A0, C0);
64
  const __m128i CA0 = _mm_subs_epu8(C0, A0);
65
  const __m128i BC0 = _mm_subs_epu8(B0, C0);
66
  const __m128i CB0 = _mm_subs_epu8(C0, B0);
67
  const __m128i AC = _mm_or_si128(AC0, CA0);
68
  const __m128i BC = _mm_or_si128(BC0, CB0);
69
  const __m128i pa = _mm_unpacklo_epi8(AC, zero);  // |a - c|
70
  const __m128i pb = _mm_unpacklo_epi8(BC, zero);  // |b - c|
71
  const __m128i diff = _mm_sub_epi16(pb, pa);
72
  {
73
    int16_t out[8];
74
    _mm_storeu_si128((__m128i*)out, diff);
75
    pa_minus_pb = out[0] + out[1] + out[2] + out[3];
76
  }
77
  return (pa_minus_pb <= 0) ? a : b;
78
}
79

80
static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
81
                                       const __m128i* const a1,
82
                                       __m128i* const avg) {
83
  // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
84
  const __m128i ones = _mm_set1_epi8(1);
85
  const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
86
  const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
87
  *avg = _mm_sub_epi8(avg1, one);
88
}
89

90
static WEBP_INLINE void Average2_uint32_SSE2(const uint32_t a0,
91
                                             const uint32_t a1,
92
                                             __m128i* const avg) {
93
  // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
94
  const __m128i ones = _mm_set1_epi8(1);
95
  const __m128i A0 = _mm_cvtsi32_si128((int)a0);
96
  const __m128i A1 = _mm_cvtsi32_si128((int)a1);
97
  const __m128i avg1 = _mm_avg_epu8(A0, A1);
98
  const __m128i one = _mm_and_si128(_mm_xor_si128(A0, A1), ones);
99
  *avg = _mm_sub_epi8(avg1, one);
100
}
101

102
static WEBP_INLINE __m128i Average2_uint32_16_SSE2(uint32_t a0, uint32_t a1) {
103
  const __m128i zero = _mm_setzero_si128();
104
  const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a0), zero);
105
  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a1), zero);
106
  const __m128i sum = _mm_add_epi16(A1, A0);
107
  return _mm_srli_epi16(sum, 1);
108
}
109

110
static WEBP_INLINE uint32_t Average2_SSE2(uint32_t a0, uint32_t a1) {
111
  __m128i output;
112
  Average2_uint32_SSE2(a0, a1, &output);
113
  return (uint32_t)_mm_cvtsi128_si32(output);
114
}
115

116
static WEBP_INLINE uint32_t Average3_SSE2(uint32_t a0, uint32_t a1,
117
                                          uint32_t a2) {
118
  const __m128i zero = _mm_setzero_si128();
119
  const __m128i avg1 = Average2_uint32_16_SSE2(a0, a2);
120
  const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128((int)a1), zero);
121
  const __m128i sum = _mm_add_epi16(avg1, A1);
122
  const __m128i avg2 = _mm_srli_epi16(sum, 1);
123
  const __m128i A2 = _mm_packus_epi16(avg2, avg2);
124
  return (uint32_t)_mm_cvtsi128_si32(A2);
125
}
126

127
static WEBP_INLINE uint32_t Average4_SSE2(uint32_t a0, uint32_t a1,
128
                                          uint32_t a2, uint32_t a3) {
129
  const __m128i avg1 = Average2_uint32_16_SSE2(a0, a1);
130
  const __m128i avg2 = Average2_uint32_16_SSE2(a2, a3);
131
  const __m128i sum = _mm_add_epi16(avg2, avg1);
132
  const __m128i avg3 = _mm_srli_epi16(sum, 1);
133
  const __m128i A0 = _mm_packus_epi16(avg3, avg3);
134
  return (uint32_t)_mm_cvtsi128_si32(A0);
135
}
136

137
static uint32_t Predictor5_SSE2(const uint32_t* const left,
138
                                const uint32_t* const top) {
139
  const uint32_t pred = Average3_SSE2(*left, top[0], top[1]);
140
  return pred;
141
}
142
static uint32_t Predictor6_SSE2(const uint32_t* const left,
143
                                const uint32_t* const top) {
144
  const uint32_t pred = Average2_SSE2(*left, top[-1]);
145
  return pred;
146
}
147
static uint32_t Predictor7_SSE2(const uint32_t* const left,
148
                                const uint32_t* const top) {
149
  const uint32_t pred = Average2_SSE2(*left, top[0]);
150
  return pred;
151
}
152
static uint32_t Predictor8_SSE2(const uint32_t* const left,
153
                                const uint32_t* const top) {
154
  const uint32_t pred = Average2_SSE2(top[-1], top[0]);
155
  (void)left;
156
  return pred;
157
}
158
static uint32_t Predictor9_SSE2(const uint32_t* const left,
159
                                const uint32_t* const top) {
160
  const uint32_t pred = Average2_SSE2(top[0], top[1]);
161
  (void)left;
162
  return pred;
163
}
164
static uint32_t Predictor10_SSE2(const uint32_t* const left,
165
                                 const uint32_t* const top) {
166
  const uint32_t pred = Average4_SSE2(*left, top[-1], top[0], top[1]);
167
  return pred;
168
}
169
static uint32_t Predictor11_SSE2(const uint32_t* const left,
170
                                 const uint32_t* const top) {
171
  const uint32_t pred = Select_SSE2(top[0], *left, top[-1]);
172
  return pred;
173
}
174
static uint32_t Predictor12_SSE2(const uint32_t* const left,
175
                                 const uint32_t* const top) {
176
  const uint32_t pred = ClampedAddSubtractFull_SSE2(*left, top[0], top[-1]);
177
  return pred;
178
}
179
static uint32_t Predictor13_SSE2(const uint32_t* const left,
180
                                 const uint32_t* const top) {
181
  const uint32_t pred = ClampedAddSubtractHalf_SSE2(*left, top[0], top[-1]);
182
  return pred;
183
}
184

185
// Batch versions of those functions.
186

187
// Predictor0: ARGB_BLACK.
188
static void PredictorAdd0_SSE2(const uint32_t* in, const uint32_t* upper,
189
                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
190
  int i;
191
  const __m128i black = _mm_set1_epi32((int)ARGB_BLACK);
192
  for (i = 0; i + 4 <= num_pixels; i += 4) {
193
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
194
    const __m128i res = _mm_add_epi8(src, black);
195
    _mm_storeu_si128((__m128i*)&out[i], res);
196
  }
197
  if (i != num_pixels) {
198
    VP8LPredictorsAdd_C[0](in + i, NULL, num_pixels - i, out + i);
199
  }
200
  (void)upper;
201
}
202

203
// Predictor1: left.
204
static void PredictorAdd1_SSE2(const uint32_t* in, const uint32_t* upper,
205
                               int num_pixels, uint32_t* WEBP_RESTRICT out) {
206
  int i;
207
  __m128i prev = _mm_set1_epi32((int)out[-1]);
208
  for (i = 0; i + 4 <= num_pixels; i += 4) {
209
    // a | b | c | d
210
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
211
    // 0 | a | b | c
212
    const __m128i shift0 = _mm_slli_si128(src, 4);
213
    // a | a + b | b + c | c + d
214
    const __m128i sum0 = _mm_add_epi8(src, shift0);
215
    // 0 | 0 | a | a + b
216
    const __m128i shift1 = _mm_slli_si128(sum0, 8);
217
    // a | a + b | a + b + c | a + b + c + d
218
    const __m128i sum1 = _mm_add_epi8(sum0, shift1);
219
    const __m128i res = _mm_add_epi8(sum1, prev);
220
    _mm_storeu_si128((__m128i*)&out[i], res);
221
    // replicate prev output on the four lanes
222
    prev = _mm_shuffle_epi32(res, (3 << 0) | (3 << 2) | (3 << 4) | (3 << 6));
223
  }
224
  if (i != num_pixels) {
225
    VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i);
226
  }
227
}
228

229
// Macro that adds 32-bit integers from IN using mod 256 arithmetic
230
// per 8 bit channel.
231
#define GENERATE_PREDICTOR_1(X, IN)                                           \
232
static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
233
                                   int num_pixels,                            \
234
                                   uint32_t* WEBP_RESTRICT out) {             \
235
  int i;                                                                      \
236
  for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
237
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);              \
238
    const __m128i other = _mm_loadu_si128((const __m128i*)&(IN));             \
239
    const __m128i res = _mm_add_epi8(src, other);                             \
240
    _mm_storeu_si128((__m128i*)&out[i], res);                                 \
241
  }                                                                           \
242
  if (i != num_pixels) {                                                      \
243
    VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i);     \
244
  }                                                                           \
245
}
246

247
// Predictor2: Top.
248
GENERATE_PREDICTOR_1(2, upper[i])
249
// Predictor3: Top-right.
250
GENERATE_PREDICTOR_1(3, upper[i + 1])
251
// Predictor4: Top-left.
252
GENERATE_PREDICTOR_1(4, upper[i - 1])
253
#undef GENERATE_PREDICTOR_1
254

255
// Due to averages with integers, values cannot be accumulated in parallel for
256
// predictors 5 to 7.
257
GENERATE_PREDICTOR_ADD(Predictor5_SSE2, PredictorAdd5_SSE2)
258
GENERATE_PREDICTOR_ADD(Predictor6_SSE2, PredictorAdd6_SSE2)
259
GENERATE_PREDICTOR_ADD(Predictor7_SSE2, PredictorAdd7_SSE2)
260

261
#define GENERATE_PREDICTOR_2(X, IN)                                           \
262
static void PredictorAdd##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
263
                                   int num_pixels,                            \
264
                                   uint32_t* WEBP_RESTRICT out) {             \
265
  int i;                                                                      \
266
  for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
267
    const __m128i Tother = _mm_loadu_si128((const __m128i*)&(IN));            \
268
    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);             \
269
    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);              \
270
    __m128i avg, res;                                                         \
271
    Average2_m128i(&T, &Tother, &avg);                                        \
272
    res = _mm_add_epi8(avg, src);                                             \
273
    _mm_storeu_si128((__m128i*)&out[i], res);                                 \
274
  }                                                                           \
275
  if (i != num_pixels) {                                                      \
276
    VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i);     \
277
  }                                                                           \
278
}
279
// Predictor8: average TL T.
280
GENERATE_PREDICTOR_2(8, upper[i - 1])
281
// Predictor9: average T TR.
282
GENERATE_PREDICTOR_2(9, upper[i + 1])
283
#undef GENERATE_PREDICTOR_2
284

285
// Predictor10: average of (average of (L,TL), average of (T, TR)).
286
#define DO_PRED10(OUT) do {                         \
287
  __m128i avgLTL, avg;                              \
288
  Average2_m128i(&L, &TL, &avgLTL);                 \
289
  Average2_m128i(&avgTTR, &avgLTL, &avg);           \
290
  L = _mm_add_epi8(avg, src);                       \
291
  out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(L);  \
292
} while (0)
293

294
#define DO_PRED10_SHIFT do {                                  \
295
  /* Rotate the pre-computed values for the next iteration.*/ \
296
  avgTTR = _mm_srli_si128(avgTTR, 4);                         \
297
  TL = _mm_srli_si128(TL, 4);                                 \
298
  src = _mm_srli_si128(src, 4);                               \
299
} while (0)
300

301
static void PredictorAdd10_SSE2(const uint32_t* in, const uint32_t* upper,
302
                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
303
  int i;
304
  __m128i L = _mm_cvtsi32_si128((int)out[-1]);
305
  for (i = 0; i + 4 <= num_pixels; i += 4) {
306
    __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
307
    __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
308
    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
309
    const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
310
    __m128i avgTTR;
311
    Average2_m128i(&T, &TR, &avgTTR);
312
    DO_PRED10(0);
313
    DO_PRED10_SHIFT;
314
    DO_PRED10(1);
315
    DO_PRED10_SHIFT;
316
    DO_PRED10(2);
317
    DO_PRED10_SHIFT;
318
    DO_PRED10(3);
319
  }
320
  if (i != num_pixels) {
321
    VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
322
  }
323
}
324
#undef DO_PRED10
325
#undef DO_PRED10_SHIFT
326

327
// Predictor11: select.
328
#define DO_PRED11(OUT) do {                                            \
329
  const __m128i L_lo = _mm_unpacklo_epi32(L, T);                       \
330
  const __m128i TL_lo = _mm_unpacklo_epi32(TL, T);                     \
331
  const __m128i pb = _mm_sad_epu8(L_lo, TL_lo); /* pb = sum |L-TL|*/   \
332
  const __m128i mask = _mm_cmpgt_epi32(pb, pa);                        \
333
  const __m128i A = _mm_and_si128(mask, L);                            \
334
  const __m128i B = _mm_andnot_si128(mask, T);                         \
335
  const __m128i pred = _mm_or_si128(A, B); /* pred = (pa > b)? L : T*/ \
336
  L = _mm_add_epi8(src, pred);                                         \
337
  out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(L);                     \
338
} while (0)
339

340
#define DO_PRED11_SHIFT do {                                \
341
  /* Shift the pre-computed value for the next iteration.*/ \
342
  T = _mm_srli_si128(T, 4);                                 \
343
  TL = _mm_srli_si128(TL, 4);                               \
344
  src = _mm_srli_si128(src, 4);                             \
345
  pa = _mm_srli_si128(pa, 4);                               \
346
} while (0)
347

348
static void PredictorAdd11_SSE2(const uint32_t* in, const uint32_t* upper,
349
                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
350
  int i;
351
  __m128i pa;
352
  __m128i L = _mm_cvtsi32_si128((int)out[-1]);
353
  for (i = 0; i + 4 <= num_pixels; i += 4) {
354
    __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
355
    __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
356
    __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
357
    {
358
      // We can unpack with any value on the upper 32 bits, provided it's the
359
      // same on both operands (so that their sum of abs diff is zero). Here we
360
      // use T.
361
      const __m128i T_lo = _mm_unpacklo_epi32(T, T);
362
      const __m128i TL_lo = _mm_unpacklo_epi32(TL, T);
363
      const __m128i T_hi = _mm_unpackhi_epi32(T, T);
364
      const __m128i TL_hi = _mm_unpackhi_epi32(TL, T);
365
      const __m128i s_lo = _mm_sad_epu8(T_lo, TL_lo);
366
      const __m128i s_hi = _mm_sad_epu8(T_hi, TL_hi);
367
      pa = _mm_packs_epi32(s_lo, s_hi);  // pa = sum |T-TL|
368
    }
369
    DO_PRED11(0);
370
    DO_PRED11_SHIFT;
371
    DO_PRED11(1);
372
    DO_PRED11_SHIFT;
373
    DO_PRED11(2);
374
    DO_PRED11_SHIFT;
375
    DO_PRED11(3);
376
  }
377
  if (i != num_pixels) {
378
    VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
379
  }
380
}
381
#undef DO_PRED11
382
#undef DO_PRED11_SHIFT
383

384
// Predictor12: ClampedAddSubtractFull.
385
#define DO_PRED12(DIFF, LANE, OUT) do {              \
386
  const __m128i all = _mm_add_epi16(L, (DIFF));      \
387
  const __m128i alls = _mm_packus_epi16(all, all);   \
388
  const __m128i res = _mm_add_epi8(src, alls);       \
389
  out[i + (OUT)] = (uint32_t)_mm_cvtsi128_si32(res); \
390
  L = _mm_unpacklo_epi8(res, zero);                  \
391
} while (0)
392

393
#define DO_PRED12_SHIFT(DIFF, LANE) do {                    \
394
  /* Shift the pre-computed value for the next iteration.*/ \
395
  if ((LANE) == 0) (DIFF) = _mm_srli_si128((DIFF), 8);      \
396
  src = _mm_srli_si128(src, 4);                             \
397
} while (0)
398

399
static void PredictorAdd12_SSE2(const uint32_t* in, const uint32_t* upper,
400
                                int num_pixels, uint32_t* WEBP_RESTRICT out) {
401
  int i;
402
  const __m128i zero = _mm_setzero_si128();
403
  const __m128i L8 = _mm_cvtsi32_si128((int)out[-1]);
404
  __m128i L = _mm_unpacklo_epi8(L8, zero);
405
  for (i = 0; i + 4 <= num_pixels; i += 4) {
406
    // Load 4 pixels at a time.
407
    __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
408
    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
409
    const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
410
    const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
411
    const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
412
    const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
413
    const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
414
    __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
415
    __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
416
    DO_PRED12(diff_lo, 0, 0);
417
    DO_PRED12_SHIFT(diff_lo, 0);
418
    DO_PRED12(diff_lo, 1, 1);
419
    DO_PRED12_SHIFT(diff_lo, 1);
420
    DO_PRED12(diff_hi, 0, 2);
421
    DO_PRED12_SHIFT(diff_hi, 0);
422
    DO_PRED12(diff_hi, 1, 3);
423
  }
424
  if (i != num_pixels) {
425
    VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
426
  }
427
}
428
#undef DO_PRED12
429
#undef DO_PRED12_SHIFT
430

431
// Due to averages with integers, values cannot be accumulated in parallel for
432
// predictors 13.
433
GENERATE_PREDICTOR_ADD(Predictor13_SSE2, PredictorAdd13_SSE2)
434

435
//------------------------------------------------------------------------------
436
// Subtract-Green Transform
437

438
static void AddGreenToBlueAndRed_SSE2(const uint32_t* const src, int num_pixels,
439
                                      uint32_t* dst) {
440
  int i;
441
  for (i = 0; i + 4 <= num_pixels; i += 4) {
442
    const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
443
    const __m128i A = _mm_srli_epi16(in, 8);     // 0 a 0 g
444
    const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
445
    const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0));  // 0g0g
446
    const __m128i out = _mm_add_epi8(in, C);
447
    _mm_storeu_si128((__m128i*)&dst[i], out);
448
  }
449
  // fallthrough and finish off with plain-C
450
  if (i != num_pixels) {
451
    VP8LAddGreenToBlueAndRed_C(src + i, num_pixels - i, dst + i);
452
  }
453
}
454

455
//------------------------------------------------------------------------------
456
// Color Transform
457

458
static void TransformColorInverse_SSE2(const VP8LMultipliers* const m,
459
                                       const uint32_t* const src,
460
                                       int num_pixels, uint32_t* dst) {
461
// sign-extended multiplying constants, pre-shifted by 5.
462
#define CST(X)  (((int16_t)(m->X << 8)) >> 5)   // sign-extend
463
#define MK_CST_16(HI, LO) \
464
  _mm_set1_epi32((int)(((uint32_t)(HI) << 16) | ((LO) & 0xffff)))
465
  const __m128i mults_rb = MK_CST_16(CST(green_to_red_), CST(green_to_blue_));
466
  const __m128i mults_b2 = MK_CST_16(CST(red_to_blue_), 0);
467
#undef MK_CST_16
468
#undef CST
469
  const __m128i mask_ag = _mm_set1_epi32((int)0xff00ff00);  // alpha-green masks
470
  int i;
471
  for (i = 0; i + 4 <= num_pixels; i += 4) {
472
    const __m128i in = _mm_loadu_si128((const __m128i*)&src[i]); // argb
473
    const __m128i A = _mm_and_si128(in, mask_ag);     // a   0   g   0
474
    const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
475
    const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0));  // g0g0
476
    const __m128i D = _mm_mulhi_epi16(C, mults_rb);    // x dr  x db1
477
    const __m128i E = _mm_add_epi8(in, D);             // x r'  x   b'
478
    const __m128i F = _mm_slli_epi16(E, 8);            // r' 0   b' 0
479
    const __m128i G = _mm_mulhi_epi16(F, mults_b2);    // x db2  0  0
480
    const __m128i H = _mm_srli_epi32(G, 8);            // 0  x db2  0
481
    const __m128i I = _mm_add_epi8(H, F);              // r' x  b'' 0
482
    const __m128i J = _mm_srli_epi16(I, 8);            // 0  r'  0  b''
483
    const __m128i out = _mm_or_si128(J, A);
484
    _mm_storeu_si128((__m128i*)&dst[i], out);
485
  }
486
  // Fall-back to C-version for left-overs.
487
  if (i != num_pixels) {
488
    VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
489
  }
490
}
491

492
//------------------------------------------------------------------------------
493
// Color-space conversion functions
494

495
static void ConvertBGRAToRGB_SSE2(const uint32_t* WEBP_RESTRICT src,
496
                                  int num_pixels, uint8_t* WEBP_RESTRICT dst) {
497
  const __m128i* in = (const __m128i*)src;
498
  __m128i* out = (__m128i*)dst;
499

500
  while (num_pixels >= 32) {
501
    // Load the BGRA buffers.
502
    __m128i in0 = _mm_loadu_si128(in + 0);
503
    __m128i in1 = _mm_loadu_si128(in + 1);
504
    __m128i in2 = _mm_loadu_si128(in + 2);
505
    __m128i in3 = _mm_loadu_si128(in + 3);
506
    __m128i in4 = _mm_loadu_si128(in + 4);
507
    __m128i in5 = _mm_loadu_si128(in + 5);
508
    __m128i in6 = _mm_loadu_si128(in + 6);
509
    __m128i in7 = _mm_loadu_si128(in + 7);
510
    VP8L32bToPlanar_SSE2(&in0, &in1, &in2, &in3);
511
    VP8L32bToPlanar_SSE2(&in4, &in5, &in6, &in7);
512
    // At this points, in1/in5 contains red only, in2/in6 green only ...
513
    // Pack the colors in 24b RGB.
514
    VP8PlanarTo24b_SSE2(&in1, &in5, &in2, &in6, &in3, &in7);
515
    _mm_storeu_si128(out + 0, in1);
516
    _mm_storeu_si128(out + 1, in5);
517
    _mm_storeu_si128(out + 2, in2);
518
    _mm_storeu_si128(out + 3, in6);
519
    _mm_storeu_si128(out + 4, in3);
520
    _mm_storeu_si128(out + 5, in7);
521
    in += 8;
522
    out += 6;
523
    num_pixels -= 32;
524
  }
525
  // left-overs
526
  if (num_pixels > 0) {
527
    VP8LConvertBGRAToRGB_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
528
  }
529
}
530

531
static void ConvertBGRAToRGBA_SSE2(const uint32_t* WEBP_RESTRICT src,
532
                                   int num_pixels, uint8_t* WEBP_RESTRICT dst) {
533
  const __m128i red_blue_mask = _mm_set1_epi32(0x00ff00ff);
534
  const __m128i* in = (const __m128i*)src;
535
  __m128i* out = (__m128i*)dst;
536
  while (num_pixels >= 8) {
537
    const __m128i A1 = _mm_loadu_si128(in++);
538
    const __m128i A2 = _mm_loadu_si128(in++);
539
    const __m128i B1 = _mm_and_si128(A1, red_blue_mask);     // R 0 B 0
540
    const __m128i B2 = _mm_and_si128(A2, red_blue_mask);     // R 0 B 0
541
    const __m128i C1 = _mm_andnot_si128(red_blue_mask, A1);  // 0 G 0 A
542
    const __m128i C2 = _mm_andnot_si128(red_blue_mask, A2);  // 0 G 0 A
543
    const __m128i D1 = _mm_shufflelo_epi16(B1, _MM_SHUFFLE(2, 3, 0, 1));
544
    const __m128i D2 = _mm_shufflelo_epi16(B2, _MM_SHUFFLE(2, 3, 0, 1));
545
    const __m128i E1 = _mm_shufflehi_epi16(D1, _MM_SHUFFLE(2, 3, 0, 1));
546
    const __m128i E2 = _mm_shufflehi_epi16(D2, _MM_SHUFFLE(2, 3, 0, 1));
547
    const __m128i F1 = _mm_or_si128(E1, C1);
548
    const __m128i F2 = _mm_or_si128(E2, C2);
549
    _mm_storeu_si128(out++, F1);
550
    _mm_storeu_si128(out++, F2);
551
    num_pixels -= 8;
552
  }
553
  // left-overs
554
  if (num_pixels > 0) {
555
    VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
556
  }
557
}
558

559
static void ConvertBGRAToRGBA4444_SSE2(const uint32_t* WEBP_RESTRICT src,
560
                                       int num_pixels,
561
                                       uint8_t* WEBP_RESTRICT dst) {
562
  const __m128i mask_0x0f = _mm_set1_epi8(0x0f);
563
  const __m128i mask_0xf0 = _mm_set1_epi8((char)0xf0);
564
  const __m128i* in = (const __m128i*)src;
565
  __m128i* out = (__m128i*)dst;
566
  while (num_pixels >= 8) {
567
    const __m128i bgra0 = _mm_loadu_si128(in++);     // bgra0|bgra1|bgra2|bgra3
568
    const __m128i bgra4 = _mm_loadu_si128(in++);     // bgra4|bgra5|bgra6|bgra7
569
    const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4);  // b0b4g0g4r0r4a0a4...
570
    const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4);  // b2b6g2g6r2r6a2a6...
571
    const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h);    // b0b2b4b6g0g2g4g6...
572
    const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h);    // b1b3b5b7g1g3g5g7...
573
    const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h);    // b0...b7 | g0...g7
574
    const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h);    // r0...r7 | a0...a7
575
    const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h);   // g0...g7 | a0...a7
576
    const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l);   // r0...r7 | b0...b7
577
    const __m128i ga1 = _mm_srli_epi16(ga0, 4);         // g0-|g1-|...|a6-|a7-
578
    const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0);  // -r0|-r1|...|-b6|-a7
579
    const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f);  // g0-|g1-|...|a6-|a7-
580
    const __m128i rgba0 = _mm_or_si128(ga2, rb1);       // rg0..rg7 | ba0..ba7
581
    const __m128i rgba1 = _mm_srli_si128(rgba0, 8);     // ba0..ba7 | 0
582
#if (WEBP_SWAP_16BIT_CSP == 1)
583
    const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0);  // barg0...barg7
584
#else
585
    const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1);  // rgba0...rgba7
586
#endif
587
    _mm_storeu_si128(out++, rgba);
588
    num_pixels -= 8;
589
  }
590
  // left-overs
591
  if (num_pixels > 0) {
592
    VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out);
593
  }
594
}
595

596
static void ConvertBGRAToRGB565_SSE2(const uint32_t* WEBP_RESTRICT src,
597
                                     int num_pixels,
598
                                     uint8_t* WEBP_RESTRICT dst) {
599
  const __m128i mask_0xe0 = _mm_set1_epi8((char)0xe0);
600
  const __m128i mask_0xf8 = _mm_set1_epi8((char)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

638
static void ConvertBGRAToBGR_SSE2(const uint32_t* WEBP_RESTRICT src,
639
                                  int num_pixels, uint8_t* WEBP_RESTRICT 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

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

718
Product

Resources

Company
