1// Copyright 2015 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 encoder
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "./dsp.h"
15
16#if defined(WEBP_USE_SSE2)
17#include <assert.h>
18#include <emmintrin.h>
19#include "./lossless.h"
20#include "./common_sse2.h"
21#include "./lossless_common.h"
22
23// For sign-extended multiplying constants, pre-shifted by 5:
24#define CST_5b(X)  (((int16_t)((uint16_t)X << 8)) >> 5)
25
26//------------------------------------------------------------------------------
27// Subtract-Green Transform
28
29static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
30  int i;
31  for (i = 0; i + 4 <= num_pixels; i += 4) {
32    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
33    const __m128i A = _mm_srli_epi16(in, 8);     // 0 a 0 g
34    const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
35    const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0));  // 0g0g
36    const __m128i out = _mm_sub_epi8(in, C);
37    _mm_storeu_si128((__m128i*)&argb_data[i], out);
38  }
39  // fallthrough and finish off with plain-C
40  if (i != num_pixels) {
41    VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i);
42  }
43}
44
45//------------------------------------------------------------------------------
46// Color Transform
47
48static void TransformColor(const VP8LMultipliers* const m,
49                           uint32_t* argb_data, int num_pixels) {
50  const __m128i mults_rb = _mm_set_epi16(
51      CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_),
52      CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_),
53      CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_),
54      CST_5b(m->green_to_red_), CST_5b(m->green_to_blue_));
55  const __m128i mults_b2 = _mm_set_epi16(
56      CST_5b(m->red_to_blue_), 0, CST_5b(m->red_to_blue_), 0,
57      CST_5b(m->red_to_blue_), 0, CST_5b(m->red_to_blue_), 0);
58  const __m128i mask_ag = _mm_set1_epi32(0xff00ff00);  // alpha-green masks
59  const __m128i mask_rb = _mm_set1_epi32(0x00ff00ff);  // red-blue masks
60  int i;
61  for (i = 0; i + 4 <= num_pixels; i += 4) {
62    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb
63    const __m128i A = _mm_and_si128(in, mask_ag);     // a   0   g   0
64    const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0));
65    const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0));  // g0g0
66    const __m128i D = _mm_mulhi_epi16(C, mults_rb);    // x dr  x db1
67    const __m128i E = _mm_slli_epi16(in, 8);           // r 0   b   0
68    const __m128i F = _mm_mulhi_epi16(E, mults_b2);    // x db2 0   0
69    const __m128i G = _mm_srli_epi32(F, 16);           // 0 0   x db2
70    const __m128i H = _mm_add_epi8(G, D);              // x dr  x  db
71    const __m128i I = _mm_and_si128(H, mask_rb);       // 0 dr  0  db
72    const __m128i out = _mm_sub_epi8(in, I);
73    _mm_storeu_si128((__m128i*)&argb_data[i], out);
74  }
75  // fallthrough and finish off with plain-C
76  if (i != num_pixels) {
77    VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
78  }
79}
80
81//------------------------------------------------------------------------------
82#define SPAN 8
83static void CollectColorBlueTransforms(const uint32_t* argb, int stride,
84                                       int tile_width, int tile_height,
85                                       int green_to_blue, int red_to_blue,
86                                       int histo[]) {
87  const __m128i mults_r = _mm_set_epi16(
88      CST_5b(red_to_blue), 0, CST_5b(red_to_blue), 0,
89      CST_5b(red_to_blue), 0, CST_5b(red_to_blue), 0);
90  const __m128i mults_g = _mm_set_epi16(
91      0, CST_5b(green_to_blue), 0, CST_5b(green_to_blue),
92      0, CST_5b(green_to_blue), 0, CST_5b(green_to_blue));
93  const __m128i mask_g = _mm_set1_epi32(0x00ff00);  // green mask
94  const __m128i mask_b = _mm_set1_epi32(0x0000ff);  // blue mask
95  int y;
96  for (y = 0; y < tile_height; ++y) {
97    const uint32_t* const src = argb + y * stride;
98    int i, x;
99    for (x = 0; x + SPAN <= tile_width; x += SPAN) {
100      uint16_t values[SPAN];
101      const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x +        0]);
102      const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
103      const __m128i A0 = _mm_slli_epi16(in0, 8);        // r 0  | b 0
104      const __m128i A1 = _mm_slli_epi16(in1, 8);
105      const __m128i B0 = _mm_and_si128(in0, mask_g);    // 0 0  | g 0
106      const __m128i B1 = _mm_and_si128(in1, mask_g);
107      const __m128i C0 = _mm_mulhi_epi16(A0, mults_r);  // x db | 0 0
108      const __m128i C1 = _mm_mulhi_epi16(A1, mults_r);
109      const __m128i D0 = _mm_mulhi_epi16(B0, mults_g);  // 0 0  | x db
110      const __m128i D1 = _mm_mulhi_epi16(B1, mults_g);
111      const __m128i E0 = _mm_sub_epi8(in0, D0);         // x x  | x b'
112      const __m128i E1 = _mm_sub_epi8(in1, D1);
113      const __m128i F0 = _mm_srli_epi32(C0, 16);        // 0 0  | x db
114      const __m128i F1 = _mm_srli_epi32(C1, 16);
115      const __m128i G0 = _mm_sub_epi8(E0, F0);          // 0 0  | x b'
116      const __m128i G1 = _mm_sub_epi8(E1, F1);
117      const __m128i H0 = _mm_and_si128(G0, mask_b);     // 0 0  | 0 b
118      const __m128i H1 = _mm_and_si128(G1, mask_b);
119      const __m128i I = _mm_packs_epi32(H0, H1);        // 0 b' | 0 b'
120      _mm_storeu_si128((__m128i*)values, I);
121      for (i = 0; i < SPAN; ++i) ++histo[values[i]];
122    }
123  }
124  {
125    const int left_over = tile_width & (SPAN - 1);
126    if (left_over > 0) {
127      VP8LCollectColorBlueTransforms_C(argb + tile_width - left_over, stride,
128                                       left_over, tile_height,
129                                       green_to_blue, red_to_blue, histo);
130    }
131  }
132}
133
134static void CollectColorRedTransforms(const uint32_t* argb, int stride,
135                                      int tile_width, int tile_height,
136                                      int green_to_red, int histo[]) {
137  const __m128i mults_g = _mm_set_epi16(
138      0, CST_5b(green_to_red), 0, CST_5b(green_to_red),
139      0, CST_5b(green_to_red), 0, CST_5b(green_to_red));
140  const __m128i mask_g = _mm_set1_epi32(0x00ff00);  // green mask
141  const __m128i mask = _mm_set1_epi32(0xff);
142
143  int y;
144  for (y = 0; y < tile_height; ++y) {
145    const uint32_t* const src = argb + y * stride;
146    int i, x;
147    for (x = 0; x + SPAN <= tile_width; x += SPAN) {
148      uint16_t values[SPAN];
149      const __m128i in0 = _mm_loadu_si128((__m128i*)&src[x +        0]);
150      const __m128i in1 = _mm_loadu_si128((__m128i*)&src[x + SPAN / 2]);
151      const __m128i A0 = _mm_and_si128(in0, mask_g);    // 0 0  | g 0
152      const __m128i A1 = _mm_and_si128(in1, mask_g);
153      const __m128i B0 = _mm_srli_epi32(in0, 16);       // 0 0  | x r
154      const __m128i B1 = _mm_srli_epi32(in1, 16);
155      const __m128i C0 = _mm_mulhi_epi16(A0, mults_g);  // 0 0  | x dr
156      const __m128i C1 = _mm_mulhi_epi16(A1, mults_g);
157      const __m128i E0 = _mm_sub_epi8(B0, C0);          // x x  | x r'
158      const __m128i E1 = _mm_sub_epi8(B1, C1);
159      const __m128i F0 = _mm_and_si128(E0, mask);       // 0 0  | 0 r'
160      const __m128i F1 = _mm_and_si128(E1, mask);
161      const __m128i I = _mm_packs_epi32(F0, F1);
162      _mm_storeu_si128((__m128i*)values, I);
163      for (i = 0; i < SPAN; ++i) ++histo[values[i]];
164    }
165  }
166  {
167    const int left_over = tile_width & (SPAN - 1);
168    if (left_over > 0) {
169      VP8LCollectColorRedTransforms_C(argb + tile_width - left_over, stride,
170                                      left_over, tile_height,
171                                      green_to_red, histo);
172    }
173  }
174}
175#undef SPAN
176
177//------------------------------------------------------------------------------
178
179#define LINE_SIZE 16    // 8 or 16
180static void AddVector(const uint32_t* a, const uint32_t* b, uint32_t* out,
181                      int size) {
182  int i;
183  assert(size % LINE_SIZE == 0);
184  for (i = 0; i < size; i += LINE_SIZE) {
185    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
186    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i +  4]);
187#if (LINE_SIZE == 16)
188    const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i +  8]);
189    const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
190#endif
191    const __m128i b0 = _mm_loadu_si128((const __m128i*)&b[i +  0]);
192    const __m128i b1 = _mm_loadu_si128((const __m128i*)&b[i +  4]);
193#if (LINE_SIZE == 16)
194    const __m128i b2 = _mm_loadu_si128((const __m128i*)&b[i +  8]);
195    const __m128i b3 = _mm_loadu_si128((const __m128i*)&b[i + 12]);
196#endif
197    _mm_storeu_si128((__m128i*)&out[i +  0], _mm_add_epi32(a0, b0));
198    _mm_storeu_si128((__m128i*)&out[i +  4], _mm_add_epi32(a1, b1));
199#if (LINE_SIZE == 16)
200    _mm_storeu_si128((__m128i*)&out[i +  8], _mm_add_epi32(a2, b2));
201    _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
202#endif
203  }
204}
205
206static void AddVectorEq(const uint32_t* a, uint32_t* out, int size) {
207  int i;
208  assert(size % LINE_SIZE == 0);
209  for (i = 0; i < size; i += LINE_SIZE) {
210    const __m128i a0 = _mm_loadu_si128((const __m128i*)&a[i +  0]);
211    const __m128i a1 = _mm_loadu_si128((const __m128i*)&a[i +  4]);
212#if (LINE_SIZE == 16)
213    const __m128i a2 = _mm_loadu_si128((const __m128i*)&a[i +  8]);
214    const __m128i a3 = _mm_loadu_si128((const __m128i*)&a[i + 12]);
215#endif
216    const __m128i b0 = _mm_loadu_si128((const __m128i*)&out[i +  0]);
217    const __m128i b1 = _mm_loadu_si128((const __m128i*)&out[i +  4]);
218#if (LINE_SIZE == 16)
219    const __m128i b2 = _mm_loadu_si128((const __m128i*)&out[i +  8]);
220    const __m128i b3 = _mm_loadu_si128((const __m128i*)&out[i + 12]);
221#endif
222    _mm_storeu_si128((__m128i*)&out[i +  0], _mm_add_epi32(a0, b0));
223    _mm_storeu_si128((__m128i*)&out[i +  4], _mm_add_epi32(a1, b1));
224#if (LINE_SIZE == 16)
225    _mm_storeu_si128((__m128i*)&out[i +  8], _mm_add_epi32(a2, b2));
226    _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3));
227#endif
228  }
229}
230#undef LINE_SIZE
231
232// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But
233// that's ok since the histogram values are less than 1<<28 (max picture size).
234static void HistogramAdd(const VP8LHistogram* const a,
235                         const VP8LHistogram* const b,
236                         VP8LHistogram* const out) {
237  int i;
238  const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_);
239  assert(a->palette_code_bits_ == b->palette_code_bits_);
240  if (b != out) {
241    AddVector(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES);
242    AddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES);
243    AddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES);
244    AddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES);
245  } else {
246    AddVectorEq(a->literal_, out->literal_, NUM_LITERAL_CODES);
247    AddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES);
248    AddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES);
249    AddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES);
250  }
251  for (i = NUM_LITERAL_CODES; i < literal_size; ++i) {
252    out->literal_[i] = a->literal_[i] + b->literal_[i];
253  }
254  for (i = 0; i < NUM_DISTANCE_CODES; ++i) {
255    out->distance_[i] = a->distance_[i] + b->distance_[i];
256  }
257}
258
259//------------------------------------------------------------------------------
260// Entropy
261
262// Checks whether the X or Y contribution is worth computing and adding.
263// Used in loop unrolling.
264#define ANALYZE_X_OR_Y(x_or_y, j)                                   \
265  do {                                                              \
266    if (x_or_y[i + j] != 0) retval -= VP8LFastSLog2(x_or_y[i + j]); \
267  } while (0)
268
269// Checks whether the X + Y contribution is worth computing and adding.
270// Used in loop unrolling.
271#define ANALYZE_XY(j)                  \
272  do {                                 \
273    if (tmp[j] != 0) {                 \
274      retval -= VP8LFastSLog2(tmp[j]); \
275      ANALYZE_X_OR_Y(X, j);            \
276    }                                  \
277  } while (0)
278
279static float CombinedShannonEntropy(const int X[256], const int Y[256]) {
280  int i;
281  double retval = 0.;
282  int sumX, sumXY;
283  int32_t tmp[4];
284  __m128i zero = _mm_setzero_si128();
285  // Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY).
286  __m128i sumXY_128 = zero;
287  __m128i sumX_128 = zero;
288
289  for (i = 0; i < 256; i += 4) {
290    const __m128i x = _mm_loadu_si128((const __m128i*)(X + i));
291    const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i));
292
293    // Check if any X is non-zero: this actually provides a speedup as X is
294    // usually sparse.
295    if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) {
296      const __m128i xy_128 = _mm_add_epi32(x, y);
297      sumXY_128 = _mm_add_epi32(sumXY_128, xy_128);
298
299      sumX_128 = _mm_add_epi32(sumX_128, x);
300
301      // Analyze the different X + Y.
302      _mm_storeu_si128((__m128i*)tmp, xy_128);
303
304      ANALYZE_XY(0);
305      ANALYZE_XY(1);
306      ANALYZE_XY(2);
307      ANALYZE_XY(3);
308    } else {
309      // X is fully 0, so only deal with Y.
310      sumXY_128 = _mm_add_epi32(sumXY_128, y);
311
312      ANALYZE_X_OR_Y(Y, 0);
313      ANALYZE_X_OR_Y(Y, 1);
314      ANALYZE_X_OR_Y(Y, 2);
315      ANALYZE_X_OR_Y(Y, 3);
316    }
317  }
318
319  // Sum up sumX_128 to get sumX.
320  _mm_storeu_si128((__m128i*)tmp, sumX_128);
321  sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0];
322
323  // Sum up sumXY_128 to get sumXY.
324  _mm_storeu_si128((__m128i*)tmp, sumXY_128);
325  sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0];
326
327  retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
328  return (float)retval;
329}
330#undef ANALYZE_X_OR_Y
331#undef ANALYZE_XY
332
333//------------------------------------------------------------------------------
334
335static int VectorMismatch(const uint32_t* const array1,
336                          const uint32_t* const array2, int length) {
337  int match_len;
338
339  if (length >= 12) {
340    __m128i A0 = _mm_loadu_si128((const __m128i*)&array1[0]);
341    __m128i A1 = _mm_loadu_si128((const __m128i*)&array2[0]);
342    match_len = 0;
343    do {
344      // Loop unrolling and early load both provide a speedup of 10% for the
345      // current function. Also, max_limit can be MAX_LENGTH=4096 at most.
346      const __m128i cmpA = _mm_cmpeq_epi32(A0, A1);
347      const __m128i B0 =
348          _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
349      const __m128i B1 =
350          _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
351      if (_mm_movemask_epi8(cmpA) != 0xffff) break;
352      match_len += 4;
353
354      {
355        const __m128i cmpB = _mm_cmpeq_epi32(B0, B1);
356        A0 = _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
357        A1 = _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
358        if (_mm_movemask_epi8(cmpB) != 0xffff) break;
359        match_len += 4;
360      }
361    } while (match_len + 12 < length);
362  } else {
363    match_len = 0;
364    // Unroll the potential first two loops.
365    if (length >= 4 &&
366        _mm_movemask_epi8(_mm_cmpeq_epi32(
367            _mm_loadu_si128((const __m128i*)&array1[0]),
368            _mm_loadu_si128((const __m128i*)&array2[0]))) == 0xffff) {
369      match_len = 4;
370      if (length >= 8 &&
371          _mm_movemask_epi8(_mm_cmpeq_epi32(
372              _mm_loadu_si128((const __m128i*)&array1[4]),
373              _mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff) {
374        match_len = 8;
375      }
376    }
377  }
378
379  while (match_len < length && array1[match_len] == array2[match_len]) {
380    ++match_len;
381  }
382  return match_len;
383}
384
385// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
386static void BundleColorMap_SSE2(const uint8_t* const row, int width, int xbits,
387                                uint32_t* dst) {
388  int x;
389  assert(xbits >= 0);
390  assert(xbits <= 3);
391  switch (xbits) {
392    case 0: {
393      const __m128i ff = _mm_set1_epi16(0xff00);
394      const __m128i zero = _mm_setzero_si128();
395      // Store 0xff000000 | (row[x] << 8).
396      for (x = 0; x + 16 <= width; x += 16, dst += 16) {
397        const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
398        const __m128i in_lo = _mm_unpacklo_epi8(zero, in);
399        const __m128i dst0 = _mm_unpacklo_epi16(in_lo, ff);
400        const __m128i dst1 = _mm_unpackhi_epi16(in_lo, ff);
401        const __m128i in_hi = _mm_unpackhi_epi8(zero, in);
402        const __m128i dst2 = _mm_unpacklo_epi16(in_hi, ff);
403        const __m128i dst3 = _mm_unpackhi_epi16(in_hi, ff);
404        _mm_storeu_si128((__m128i*)&dst[0], dst0);
405        _mm_storeu_si128((__m128i*)&dst[4], dst1);
406        _mm_storeu_si128((__m128i*)&dst[8], dst2);
407        _mm_storeu_si128((__m128i*)&dst[12], dst3);
408      }
409      break;
410    }
411    case 1: {
412      const __m128i ff = _mm_set1_epi16(0xff00);
413      const __m128i mul = _mm_set1_epi16(0x110);
414      for (x = 0; x + 16 <= width; x += 16, dst += 8) {
415        // 0a0b | (where a/b are 4 bits).
416        const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
417        const __m128i tmp = _mm_mullo_epi16(in, mul);  // aba0
418        const __m128i pack = _mm_and_si128(tmp, ff);   // ab00
419        const __m128i dst0 = _mm_unpacklo_epi16(pack, ff);
420        const __m128i dst1 = _mm_unpackhi_epi16(pack, ff);
421        _mm_storeu_si128((__m128i*)&dst[0], dst0);
422        _mm_storeu_si128((__m128i*)&dst[4], dst1);
423      }
424      break;
425    }
426    case 2: {
427      const __m128i mask_or = _mm_set1_epi32(0xff000000);
428      const __m128i mul_cst = _mm_set1_epi16(0x0104);
429      const __m128i mask_mul = _mm_set1_epi16(0x0f00);
430      for (x = 0; x + 16 <= width; x += 16, dst += 4) {
431        // 000a000b000c000d | (where a/b/c/d are 2 bits).
432        const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
433        const __m128i mul = _mm_mullo_epi16(in, mul_cst);  // 00ab00b000cd00d0
434        const __m128i tmp = _mm_and_si128(mul, mask_mul);  // 00ab000000cd0000
435        const __m128i shift = _mm_srli_epi32(tmp, 12);     // 00000000ab000000
436        const __m128i pack = _mm_or_si128(shift, tmp);     // 00000000abcd0000
437        // Convert to 0xff00**00.
438        const __m128i res = _mm_or_si128(pack, mask_or);
439        _mm_storeu_si128((__m128i*)dst, res);
440      }
441      break;
442    }
443    default: {
444      assert(xbits == 3);
445      for (x = 0; x + 16 <= width; x += 16, dst += 2) {
446        // 0000000a00000000b... | (where a/b are 1 bit).
447        const __m128i in = _mm_loadu_si128((const __m128i*)&row[x]);
448        const __m128i shift = _mm_slli_epi64(in, 7);
449        const uint32_t move = _mm_movemask_epi8(shift);
450        dst[0] = 0xff000000 | ((move & 0xff) << 8);
451        dst[1] = 0xff000000 | (move & 0xff00);
452      }
453      break;
454    }
455  }
456  if (x != width) {
457    VP8LBundleColorMap_C(row + x, width - x, xbits, dst);
458  }
459}
460
461//------------------------------------------------------------------------------
462// Batch version of Predictor Transform subtraction
463
464static WEBP_INLINE void Average2_m128i(const __m128i* const a0,
465                                       const __m128i* const a1,
466                                       __m128i* const avg) {
467  // (a + b) >> 1 = ((a + b + 1) >> 1) - ((a ^ b) & 1)
468  const __m128i ones = _mm_set1_epi8(1);
469  const __m128i avg1 = _mm_avg_epu8(*a0, *a1);
470  const __m128i one = _mm_and_si128(_mm_xor_si128(*a0, *a1), ones);
471  *avg = _mm_sub_epi8(avg1, one);
472}
473
474// Predictor0: ARGB_BLACK.
475static void PredictorSub0_SSE2(const uint32_t* in, const uint32_t* upper,
476                               int num_pixels, uint32_t* out) {
477  int i;
478  const __m128i black = _mm_set1_epi32(ARGB_BLACK);
479  for (i = 0; i + 4 <= num_pixels; i += 4) {
480    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
481    const __m128i res = _mm_sub_epi8(src, black);
482    _mm_storeu_si128((__m128i*)&out[i], res);
483  }
484  if (i != num_pixels) {
485    VP8LPredictorsSub_C[0](in + i, upper + i, num_pixels - i, out + i);
486  }
487}
488
489#define GENERATE_PREDICTOR_1(X, IN)                                           \
490static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
491                                   int num_pixels, uint32_t* out) {           \
492  int i;                                                                      \
493  for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
494    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);              \
495    const __m128i pred = _mm_loadu_si128((const __m128i*)&(IN));              \
496    const __m128i res = _mm_sub_epi8(src, pred);                              \
497    _mm_storeu_si128((__m128i*)&out[i], res);                                 \
498  }                                                                           \
499  if (i != num_pixels) {                                                      \
500    VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i);     \
501  }                                                                           \
502}
503
504GENERATE_PREDICTOR_1(1, in[i - 1])       // Predictor1: L
505GENERATE_PREDICTOR_1(2, upper[i])        // Predictor2: T
506GENERATE_PREDICTOR_1(3, upper[i + 1])    // Predictor3: TR
507GENERATE_PREDICTOR_1(4, upper[i - 1])    // Predictor4: TL
508#undef GENERATE_PREDICTOR_1
509
510// Predictor5: avg2(avg2(L, TR), T)
511static void PredictorSub5_SSE2(const uint32_t* in, const uint32_t* upper,
512                               int num_pixels, uint32_t* out) {
513  int i;
514  for (i = 0; i + 4 <= num_pixels; i += 4) {
515    const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
516    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
517    const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
518    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
519    __m128i avg, pred, res;
520    Average2_m128i(&L, &TR, &avg);
521    Average2_m128i(&avg, &T, &pred);
522    res = _mm_sub_epi8(src, pred);
523    _mm_storeu_si128((__m128i*)&out[i], res);
524  }
525  if (i != num_pixels) {
526    VP8LPredictorsSub_C[5](in + i, upper + i, num_pixels - i, out + i);
527  }
528}
529
530#define GENERATE_PREDICTOR_2(X, A, B)                                         \
531static void PredictorSub##X##_SSE2(const uint32_t* in, const uint32_t* upper, \
532                                   int num_pixels, uint32_t* out) {           \
533  int i;                                                                      \
534  for (i = 0; i + 4 <= num_pixels; i += 4) {                                  \
535    const __m128i tA = _mm_loadu_si128((const __m128i*)&(A));                 \
536    const __m128i tB = _mm_loadu_si128((const __m128i*)&(B));                 \
537    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);              \
538    __m128i pred, res;                                                        \
539    Average2_m128i(&tA, &tB, &pred);                                          \
540    res = _mm_sub_epi8(src, pred);                                            \
541    _mm_storeu_si128((__m128i*)&out[i], res);                                 \
542  }                                                                           \
543  if (i != num_pixels) {                                                      \
544    VP8LPredictorsSub_C[(X)](in + i, upper + i, num_pixels - i, out + i);     \
545  }                                                                           \
546}
547
548GENERATE_PREDICTOR_2(6, in[i - 1], upper[i - 1])   // Predictor6: avg(L, TL)
549GENERATE_PREDICTOR_2(7, in[i - 1], upper[i])       // Predictor7: avg(L, T)
550GENERATE_PREDICTOR_2(8, upper[i - 1], upper[i])    // Predictor8: avg(TL, T)
551GENERATE_PREDICTOR_2(9, upper[i], upper[i + 1])    // Predictor9: average(T, TR)
552#undef GENERATE_PREDICTOR_2
553
554// Predictor10: avg(avg(L,TL), avg(T, TR)).
555static void PredictorSub10_SSE2(const uint32_t* in, const uint32_t* upper,
556                                int num_pixels, uint32_t* out) {
557  int i;
558  for (i = 0; i + 4 <= num_pixels; i += 4) {
559    const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
560    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
561    const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
562    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
563    const __m128i TR = _mm_loadu_si128((const __m128i*)&upper[i + 1]);
564    __m128i avgTTR, avgLTL, avg, res;
565    Average2_m128i(&T, &TR, &avgTTR);
566    Average2_m128i(&L, &TL, &avgLTL);
567    Average2_m128i(&avgTTR, &avgLTL, &avg);
568    res = _mm_sub_epi8(src, avg);
569    _mm_storeu_si128((__m128i*)&out[i], res);
570  }
571  if (i != num_pixels) {
572    VP8LPredictorsSub_C[10](in + i, upper + i, num_pixels - i, out + i);
573  }
574}
575
576// Predictor11: select.
577static void GetSumAbsDiff32(const __m128i* const A, const __m128i* const B,
578                            __m128i* const out) {
579  // We can unpack with any value on the upper 32 bits, provided it's the same
580  // on both operands (to that their sum of abs diff is zero). Here we use *A.
581  const __m128i A_lo = _mm_unpacklo_epi32(*A, *A);
582  const __m128i B_lo = _mm_unpacklo_epi32(*B, *A);
583  const __m128i A_hi = _mm_unpackhi_epi32(*A, *A);
584  const __m128i B_hi = _mm_unpackhi_epi32(*B, *A);
585  const __m128i s_lo = _mm_sad_epu8(A_lo, B_lo);
586  const __m128i s_hi = _mm_sad_epu8(A_hi, B_hi);
587  *out = _mm_packs_epi32(s_lo, s_hi);
588}
589
590static void PredictorSub11_SSE2(const uint32_t* in, const uint32_t* upper,
591                                int num_pixels, uint32_t* out) {
592  int i;
593  for (i = 0; i + 4 <= num_pixels; i += 4) {
594    const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
595    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
596    const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
597    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
598    __m128i pa, pb;
599    GetSumAbsDiff32(&T, &TL, &pa);   // pa = sum |T-TL|
600    GetSumAbsDiff32(&L, &TL, &pb);   // pb = sum |L-TL|
601    {
602      const __m128i mask = _mm_cmpgt_epi32(pb, pa);
603      const __m128i A = _mm_and_si128(mask, L);
604      const __m128i B = _mm_andnot_si128(mask, T);
605      const __m128i pred = _mm_or_si128(A, B);    // pred = (L > T)? L : T
606      const __m128i res = _mm_sub_epi8(src, pred);
607      _mm_storeu_si128((__m128i*)&out[i], res);
608    }
609  }
610  if (i != num_pixels) {
611    VP8LPredictorsSub_C[11](in + i, upper + i, num_pixels - i, out + i);
612  }
613}
614
615// Predictor12: ClampedSubSubtractFull.
616static void PredictorSub12_SSE2(const uint32_t* in, const uint32_t* upper,
617                                int num_pixels, uint32_t* out) {
618  int i;
619  const __m128i zero = _mm_setzero_si128();
620  for (i = 0; i + 4 <= num_pixels; i += 4) {
621    const __m128i src = _mm_loadu_si128((const __m128i*)&in[i]);
622    const __m128i L = _mm_loadu_si128((const __m128i*)&in[i - 1]);
623    const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
624    const __m128i L_hi = _mm_unpackhi_epi8(L, zero);
625    const __m128i T = _mm_loadu_si128((const __m128i*)&upper[i]);
626    const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
627    const __m128i T_hi = _mm_unpackhi_epi8(T, zero);
628    const __m128i TL = _mm_loadu_si128((const __m128i*)&upper[i - 1]);
629    const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
630    const __m128i TL_hi = _mm_unpackhi_epi8(TL, zero);
631    const __m128i diff_lo = _mm_sub_epi16(T_lo, TL_lo);
632    const __m128i diff_hi = _mm_sub_epi16(T_hi, TL_hi);
633    const __m128i pred_lo = _mm_add_epi16(L_lo, diff_lo);
634    const __m128i pred_hi = _mm_add_epi16(L_hi, diff_hi);
635    const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi);
636    const __m128i res = _mm_sub_epi8(src, pred);
637    _mm_storeu_si128((__m128i*)&out[i], res);
638  }
639  if (i != num_pixels) {
640    VP8LPredictorsSub_C[12](in + i, upper + i, num_pixels - i, out + i);
641  }
642}
643
644// Predictors13: ClampedAddSubtractHalf
645static void PredictorSub13_SSE2(const uint32_t* in, const uint32_t* upper,
646                                int num_pixels, uint32_t* out) {
647  int i;
648  const __m128i zero = _mm_setzero_si128();
649  for (i = 0; i + 2 <= num_pixels; i += 2) {
650    // we can only process two pixels at a time
651    const __m128i L = _mm_loadl_epi64((const __m128i*)&in[i - 1]);
652    const __m128i src = _mm_loadl_epi64((const __m128i*)&in[i]);
653    const __m128i T = _mm_loadl_epi64((const __m128i*)&upper[i]);
654    const __m128i TL = _mm_loadl_epi64((const __m128i*)&upper[i - 1]);
655    const __m128i L_lo = _mm_unpacklo_epi8(L, zero);
656    const __m128i T_lo = _mm_unpacklo_epi8(T, zero);
657    const __m128i TL_lo = _mm_unpacklo_epi8(TL, zero);
658    const __m128i sum = _mm_add_epi16(T_lo, L_lo);
659    const __m128i avg = _mm_srli_epi16(sum, 1);
660    const __m128i A1 = _mm_sub_epi16(avg, TL_lo);
661    const __m128i bit_fix = _mm_cmpgt_epi16(TL_lo, avg);
662    const __m128i A2 = _mm_sub_epi16(A1, bit_fix);
663    const __m128i A3 = _mm_srai_epi16(A2, 1);
664    const __m128i A4 = _mm_add_epi16(avg, A3);
665    const __m128i pred = _mm_packus_epi16(A4, A4);
666    const __m128i res = _mm_sub_epi8(src, pred);
667    _mm_storel_epi64((__m128i*)&out[i], res);
668  }
669  if (i != num_pixels) {
670    VP8LPredictorsSub_C[13](in + i, upper + i, num_pixels - i, out + i);
671  }
672}
673
674//------------------------------------------------------------------------------
675// Entry point
676
677extern void VP8LEncDspInitSSE2(void);
678
679WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
680  VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
681  VP8LTransformColor = TransformColor;
682  VP8LCollectColorBlueTransforms = CollectColorBlueTransforms;
683  VP8LCollectColorRedTransforms = CollectColorRedTransforms;
684  VP8LHistogramAdd = HistogramAdd;
685  VP8LCombinedShannonEntropy = CombinedShannonEntropy;
686  VP8LVectorMismatch = VectorMismatch;
687  VP8LBundleColorMap = BundleColorMap_SSE2;
688
689  VP8LPredictorsSub[0] = PredictorSub0_SSE2;
690  VP8LPredictorsSub[1] = PredictorSub1_SSE2;
691  VP8LPredictorsSub[2] = PredictorSub2_SSE2;
692  VP8LPredictorsSub[3] = PredictorSub3_SSE2;
693  VP8LPredictorsSub[4] = PredictorSub4_SSE2;
694  VP8LPredictorsSub[5] = PredictorSub5_SSE2;
695  VP8LPredictorsSub[6] = PredictorSub6_SSE2;
696  VP8LPredictorsSub[7] = PredictorSub7_SSE2;
697  VP8LPredictorsSub[8] = PredictorSub8_SSE2;
698  VP8LPredictorsSub[9] = PredictorSub9_SSE2;
699  VP8LPredictorsSub[10] = PredictorSub10_SSE2;
700  VP8LPredictorsSub[11] = PredictorSub11_SSE2;
701  VP8LPredictorsSub[12] = PredictorSub12_SSE2;
702  VP8LPredictorsSub[13] = PredictorSub13_SSE2;
703  VP8LPredictorsSub[14] = PredictorSub0_SSE2;  // <- padding security sentinels
704  VP8LPredictorsSub[15] = PredictorSub0_SSE2;
705}
706
707#else  // !WEBP_USE_SSE2
708
709WEBP_DSP_INIT_STUB(VP8LEncDspInitSSE2)
710
711#endif  // WEBP_USE_SSE2
712