1// Copyright 2011 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// Speed-critical encoding functions.
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include <stdlib.h>  // for abs()
15#include "./dsp.h"
16#include "../enc/vp8enci.h"
17
18#if defined(__cplusplus) || defined(c_plusplus)
19extern "C" {
20#endif
21
22static WEBP_INLINE uint8_t clip_8b(int v) {
23  return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
24}
25
26static WEBP_INLINE int clip_max(int v, int max) {
27  return (v > max) ? max : v;
28}
29
30//------------------------------------------------------------------------------
31// Compute susceptibility based on DCT-coeff histograms:
32// the higher, the "easier" the macroblock is to compress.
33
34const int VP8DspScan[16 + 4 + 4] = {
35  // Luma
36  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
37  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
38  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
39  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
40
41  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
42  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
43};
44
45static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
46                             int start_block, int end_block,
47                             VP8Histogram* const histo) {
48  int j;
49  for (j = start_block; j < end_block; ++j) {
50    int k;
51    int16_t out[16];
52
53    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
54
55    // Convert coefficients to bin.
56    for (k = 0; k < 16; ++k) {
57      const int v = abs(out[k]) >> 3;  // TODO(skal): add rounding?
58      const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
59      histo->distribution[clipped_value]++;
60    }
61  }
62}
63
64//------------------------------------------------------------------------------
65// run-time tables (~4k)
66
67static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
68
69// We declare this variable 'volatile' to prevent instruction reordering
70// and make sure it's set to true _last_ (so as to be thread-safe)
71static volatile int tables_ok = 0;
72
73static void InitTables(void) {
74  if (!tables_ok) {
75    int i;
76    for (i = -255; i <= 255 + 255; ++i) {
77      clip1[255 + i] = clip_8b(i);
78    }
79    tables_ok = 1;
80  }
81}
82
83
84//------------------------------------------------------------------------------
85// Transforms (Paragraph 14.4)
86
87#define STORE(x, y, v) \
88  dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3))
89
90static const int kC1 = 20091 + (1 << 16);
91static const int kC2 = 35468;
92#define MUL(a, b) (((a) * (b)) >> 16)
93
94static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
95                                      uint8_t* dst) {
96  int C[4 * 4], *tmp;
97  int i;
98  tmp = C;
99  for (i = 0; i < 4; ++i) {    // vertical pass
100    const int a = in[0] + in[8];
101    const int b = in[0] - in[8];
102    const int c = MUL(in[4], kC2) - MUL(in[12], kC1);
103    const int d = MUL(in[4], kC1) + MUL(in[12], kC2);
104    tmp[0] = a + d;
105    tmp[1] = b + c;
106    tmp[2] = b - c;
107    tmp[3] = a - d;
108    tmp += 4;
109    in++;
110  }
111
112  tmp = C;
113  for (i = 0; i < 4; ++i) {    // horizontal pass
114    const int dc = tmp[0] + 4;
115    const int a =  dc +  tmp[8];
116    const int b =  dc -  tmp[8];
117    const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1);
118    const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2);
119    STORE(0, i, a + d);
120    STORE(1, i, b + c);
121    STORE(2, i, b - c);
122    STORE(3, i, a - d);
123    tmp++;
124  }
125}
126
127static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
128                       int do_two) {
129  ITransformOne(ref, in, dst);
130  if (do_two) {
131    ITransformOne(ref + 4, in + 16, dst + 4);
132  }
133}
134
135static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
136  int i;
137  int tmp[16];
138  for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
139    const int d0 = src[0] - ref[0];   // 9bit dynamic range ([-255,255])
140    const int d1 = src[1] - ref[1];
141    const int d2 = src[2] - ref[2];
142    const int d3 = src[3] - ref[3];
143    const int a0 = (d0 + d3);         // 10b                      [-510,510]
144    const int a1 = (d1 + d2);
145    const int a2 = (d1 - d2);
146    const int a3 = (d0 - d3);
147    tmp[0 + i * 4] = (a0 + a1) * 8;   // 14b                      [-8160,8160]
148    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
149    tmp[2 + i * 4] = (a0 - a1) * 8;
150    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  937) >> 9;
151  }
152  for (i = 0; i < 4; ++i) {
153    const int a0 = (tmp[0 + i] + tmp[12 + i]);  // 15b
154    const int a1 = (tmp[4 + i] + tmp[ 8 + i]);
155    const int a2 = (tmp[4 + i] - tmp[ 8 + i]);
156    const int a3 = (tmp[0 + i] - tmp[12 + i]);
157    out[0 + i] = (a0 + a1 + 7) >> 4;            // 12b
158    out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0);
159    out[8 + i] = (a0 - a1 + 7) >> 4;
160    out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16);
161  }
162}
163
164static void ITransformWHT(const int16_t* in, int16_t* out) {
165  int tmp[16];
166  int i;
167  for (i = 0; i < 4; ++i) {
168    const int a0 = in[0 + i] + in[12 + i];
169    const int a1 = in[4 + i] + in[ 8 + i];
170    const int a2 = in[4 + i] - in[ 8 + i];
171    const int a3 = in[0 + i] - in[12 + i];
172    tmp[0  + i] = a0 + a1;
173    tmp[8  + i] = a0 - a1;
174    tmp[4  + i] = a3 + a2;
175    tmp[12 + i] = a3 - a2;
176  }
177  for (i = 0; i < 4; ++i) {
178    const int dc = tmp[0 + i * 4] + 3;    // w/ rounder
179    const int a0 = dc             + tmp[3 + i * 4];
180    const int a1 = tmp[1 + i * 4] + tmp[2 + i * 4];
181    const int a2 = tmp[1 + i * 4] - tmp[2 + i * 4];
182    const int a3 = dc             - tmp[3 + i * 4];
183    out[ 0] = (a0 + a1) >> 3;
184    out[16] = (a3 + a2) >> 3;
185    out[32] = (a0 - a1) >> 3;
186    out[48] = (a3 - a2) >> 3;
187    out += 64;
188  }
189}
190
191static void FTransformWHT(const int16_t* in, int16_t* out) {
192  // input is 12b signed
193  int16_t tmp[16];
194  int i;
195  for (i = 0; i < 4; ++i, in += 64) {
196    const int a0 = (in[0 * 16] + in[2 * 16]);  // 13b
197    const int a1 = (in[1 * 16] + in[3 * 16]);
198    const int a2 = (in[1 * 16] - in[3 * 16]);
199    const int a3 = (in[0 * 16] - in[2 * 16]);
200    tmp[0 + i * 4] = a0 + a1;   // 14b
201    tmp[1 + i * 4] = a3 + a2;
202    tmp[2 + i * 4] = a3 - a2;
203    tmp[3 + i * 4] = a0 - a1;
204  }
205  for (i = 0; i < 4; ++i) {
206    const int a0 = (tmp[0 + i] + tmp[8 + i]);  // 15b
207    const int a1 = (tmp[4 + i] + tmp[12+ i]);
208    const int a2 = (tmp[4 + i] - tmp[12+ i]);
209    const int a3 = (tmp[0 + i] - tmp[8 + i]);
210    const int b0 = a0 + a1;    // 16b
211    const int b1 = a3 + a2;
212    const int b2 = a3 - a2;
213    const int b3 = a0 - a1;
214    out[ 0 + i] = b0 >> 1;     // 15b
215    out[ 4 + i] = b1 >> 1;
216    out[ 8 + i] = b2 >> 1;
217    out[12 + i] = b3 >> 1;
218  }
219}
220
221#undef MUL
222#undef STORE
223
224//------------------------------------------------------------------------------
225// Intra predictions
226
227#define DST(x, y) dst[(x) + (y) * BPS]
228
229static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
230  int j;
231  for (j = 0; j < size; ++j) {
232    memset(dst + j * BPS, value, size);
233  }
234}
235
236static WEBP_INLINE void VerticalPred(uint8_t* dst,
237                                     const uint8_t* top, int size) {
238  int j;
239  if (top) {
240    for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
241  } else {
242    Fill(dst, 127, size);
243  }
244}
245
246static WEBP_INLINE void HorizontalPred(uint8_t* dst,
247                                       const uint8_t* left, int size) {
248  if (left) {
249    int j;
250    for (j = 0; j < size; ++j) {
251      memset(dst + j * BPS, left[j], size);
252    }
253  } else {
254    Fill(dst, 129, size);
255  }
256}
257
258static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
259                                   const uint8_t* top, int size) {
260  int y;
261  if (left) {
262    if (top) {
263      const uint8_t* const clip = clip1 + 255 - left[-1];
264      for (y = 0; y < size; ++y) {
265        const uint8_t* const clip_table = clip + left[y];
266        int x;
267        for (x = 0; x < size; ++x) {
268          dst[x] = clip_table[top[x]];
269        }
270        dst += BPS;
271      }
272    } else {
273      HorizontalPred(dst, left, size);
274    }
275  } else {
276    // true motion without left samples (hence: with default 129 value)
277    // is equivalent to VE prediction where you just copy the top samples.
278    // Note that if top samples are not available, the default value is
279    // then 129, and not 127 as in the VerticalPred case.
280    if (top) {
281      VerticalPred(dst, top, size);
282    } else {
283      Fill(dst, 129, size);
284    }
285  }
286}
287
288static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
289                               const uint8_t* top,
290                               int size, int round, int shift) {
291  int DC = 0;
292  int j;
293  if (top) {
294    for (j = 0; j < size; ++j) DC += top[j];
295    if (left) {   // top and left present
296      for (j = 0; j < size; ++j) DC += left[j];
297    } else {      // top, but no left
298      DC += DC;
299    }
300    DC = (DC + round) >> shift;
301  } else if (left) {   // left but no top
302    for (j = 0; j < size; ++j) DC += left[j];
303    DC += DC;
304    DC = (DC + round) >> shift;
305  } else {   // no top, no left, nothing.
306    DC = 0x80;
307  }
308  Fill(dst, DC, size);
309}
310
311//------------------------------------------------------------------------------
312// Chroma 8x8 prediction (paragraph 12.2)
313
314static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
315                             const uint8_t* top) {
316  // U block
317  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
318  VerticalPred(C8VE8 + dst, top, 8);
319  HorizontalPred(C8HE8 + dst, left, 8);
320  TrueMotion(C8TM8 + dst, left, top, 8);
321  // V block
322  dst += 8;
323  if (top) top += 8;
324  if (left) left += 16;
325  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
326  VerticalPred(C8VE8 + dst, top, 8);
327  HorizontalPred(C8HE8 + dst, left, 8);
328  TrueMotion(C8TM8 + dst, left, top, 8);
329}
330
331//------------------------------------------------------------------------------
332// luma 16x16 prediction (paragraph 12.3)
333
334static void Intra16Preds(uint8_t* dst,
335                         const uint8_t* left, const uint8_t* top) {
336  DCMode(I16DC16 + dst, left, top, 16, 16, 5);
337  VerticalPred(I16VE16 + dst, top, 16);
338  HorizontalPred(I16HE16 + dst, left, 16);
339  TrueMotion(I16TM16 + dst, left, top, 16);
340}
341
342//------------------------------------------------------------------------------
343// luma 4x4 prediction
344
345#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
346#define AVG2(a, b) (((a) + (b) + 1) >> 1)
347
348static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
349  const uint8_t vals[4] = {
350    AVG3(top[-1], top[0], top[1]),
351    AVG3(top[ 0], top[1], top[2]),
352    AVG3(top[ 1], top[2], top[3]),
353    AVG3(top[ 2], top[3], top[4])
354  };
355  int i;
356  for (i = 0; i < 4; ++i) {
357    memcpy(dst + i * BPS, vals, 4);
358  }
359}
360
361static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
362  const int X = top[-1];
363  const int I = top[-2];
364  const int J = top[-3];
365  const int K = top[-4];
366  const int L = top[-5];
367  *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(X, I, J);
368  *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(I, J, K);
369  *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(J, K, L);
370  *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(K, L, L);
371}
372
373static void DC4(uint8_t* dst, const uint8_t* top) {
374  uint32_t dc = 4;
375  int i;
376  for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
377  Fill(dst, dc >> 3, 4);
378}
379
380static void RD4(uint8_t* dst, const uint8_t* top) {
381  const int X = top[-1];
382  const int I = top[-2];
383  const int J = top[-3];
384  const int K = top[-4];
385  const int L = top[-5];
386  const int A = top[0];
387  const int B = top[1];
388  const int C = top[2];
389  const int D = top[3];
390  DST(0, 3)                                     = AVG3(J, K, L);
391  DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
392  DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
393  DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
394  DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
395  DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
396  DST(3, 0)                                     = AVG3(D, C, B);
397}
398
399static void LD4(uint8_t* dst, const uint8_t* top) {
400  const int A = top[0];
401  const int B = top[1];
402  const int C = top[2];
403  const int D = top[3];
404  const int E = top[4];
405  const int F = top[5];
406  const int G = top[6];
407  const int H = top[7];
408  DST(0, 0)                                     = AVG3(A, B, C);
409  DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
410  DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
411  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
412  DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
413  DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
414  DST(3, 3)                                     = AVG3(G, H, H);
415}
416
417static void VR4(uint8_t* dst, const uint8_t* top) {
418  const int X = top[-1];
419  const int I = top[-2];
420  const int J = top[-3];
421  const int K = top[-4];
422  const int A = top[0];
423  const int B = top[1];
424  const int C = top[2];
425  const int D = top[3];
426  DST(0, 0) = DST(1, 2) = AVG2(X, A);
427  DST(1, 0) = DST(2, 2) = AVG2(A, B);
428  DST(2, 0) = DST(3, 2) = AVG2(B, C);
429  DST(3, 0)             = AVG2(C, D);
430
431  DST(0, 3) =             AVG3(K, J, I);
432  DST(0, 2) =             AVG3(J, I, X);
433  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
434  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
435  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
436  DST(3, 1) =             AVG3(B, C, D);
437}
438
439static void VL4(uint8_t* dst, const uint8_t* top) {
440  const int A = top[0];
441  const int B = top[1];
442  const int C = top[2];
443  const int D = top[3];
444  const int E = top[4];
445  const int F = top[5];
446  const int G = top[6];
447  const int H = top[7];
448  DST(0, 0) =             AVG2(A, B);
449  DST(1, 0) = DST(0, 2) = AVG2(B, C);
450  DST(2, 0) = DST(1, 2) = AVG2(C, D);
451  DST(3, 0) = DST(2, 2) = AVG2(D, E);
452
453  DST(0, 1) =             AVG3(A, B, C);
454  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
455  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
456  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
457              DST(3, 2) = AVG3(E, F, G);
458              DST(3, 3) = AVG3(F, G, H);
459}
460
461static void HU4(uint8_t* dst, const uint8_t* top) {
462  const int I = top[-2];
463  const int J = top[-3];
464  const int K = top[-4];
465  const int L = top[-5];
466  DST(0, 0) =             AVG2(I, J);
467  DST(2, 0) = DST(0, 1) = AVG2(J, K);
468  DST(2, 1) = DST(0, 2) = AVG2(K, L);
469  DST(1, 0) =             AVG3(I, J, K);
470  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
471  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
472  DST(3, 2) = DST(2, 2) =
473  DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
474}
475
476static void HD4(uint8_t* dst, const uint8_t* top) {
477  const int X = top[-1];
478  const int I = top[-2];
479  const int J = top[-3];
480  const int K = top[-4];
481  const int L = top[-5];
482  const int A = top[0];
483  const int B = top[1];
484  const int C = top[2];
485
486  DST(0, 0) = DST(2, 1) = AVG2(I, X);
487  DST(0, 1) = DST(2, 2) = AVG2(J, I);
488  DST(0, 2) = DST(2, 3) = AVG2(K, J);
489  DST(0, 3)             = AVG2(L, K);
490
491  DST(3, 0)             = AVG3(A, B, C);
492  DST(2, 0)             = AVG3(X, A, B);
493  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
494  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
495  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
496  DST(1, 3)             = AVG3(L, K, J);
497}
498
499static void TM4(uint8_t* dst, const uint8_t* top) {
500  int x, y;
501  const uint8_t* const clip = clip1 + 255 - top[-1];
502  for (y = 0; y < 4; ++y) {
503    const uint8_t* const clip_table = clip + top[-2 - y];
504    for (x = 0; x < 4; ++x) {
505      dst[x] = clip_table[top[x]];
506    }
507    dst += BPS;
508  }
509}
510
511#undef DST
512#undef AVG3
513#undef AVG2
514
515// Left samples are top[-5 .. -2], top_left is top[-1], top are
516// located at top[0..3], and top right is top[4..7]
517static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
518  DC4(I4DC4 + dst, top);
519  TM4(I4TM4 + dst, top);
520  VE4(I4VE4 + dst, top);
521  HE4(I4HE4 + dst, top);
522  RD4(I4RD4 + dst, top);
523  VR4(I4VR4 + dst, top);
524  LD4(I4LD4 + dst, top);
525  VL4(I4VL4 + dst, top);
526  HD4(I4HD4 + dst, top);
527  HU4(I4HU4 + dst, top);
528}
529
530//------------------------------------------------------------------------------
531// Metric
532
533static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
534                              int w, int h) {
535  int count = 0;
536  int y, x;
537  for (y = 0; y < h; ++y) {
538    for (x = 0; x < w; ++x) {
539      const int diff = (int)a[x] - b[x];
540      count += diff * diff;
541    }
542    a += BPS;
543    b += BPS;
544  }
545  return count;
546}
547
548static int SSE16x16(const uint8_t* a, const uint8_t* b) {
549  return GetSSE(a, b, 16, 16);
550}
551static int SSE16x8(const uint8_t* a, const uint8_t* b) {
552  return GetSSE(a, b, 16, 8);
553}
554static int SSE8x8(const uint8_t* a, const uint8_t* b) {
555  return GetSSE(a, b, 8, 8);
556}
557static int SSE4x4(const uint8_t* a, const uint8_t* b) {
558  return GetSSE(a, b, 4, 4);
559}
560
561//------------------------------------------------------------------------------
562// Texture distortion
563//
564// We try to match the spectral content (weighted) between source and
565// reconstructed samples.
566
567// Hadamard transform
568// Returns the weighted sum of the absolute value of transformed coefficients.
569static int TTransform(const uint8_t* in, const uint16_t* w) {
570  int sum = 0;
571  int tmp[16];
572  int i;
573  // horizontal pass
574  for (i = 0; i < 4; ++i, in += BPS) {
575    const int a0 = in[0] + in[2];
576    const int a1 = in[1] + in[3];
577    const int a2 = in[1] - in[3];
578    const int a3 = in[0] - in[2];
579    tmp[0 + i * 4] = a0 + a1;
580    tmp[1 + i * 4] = a3 + a2;
581    tmp[2 + i * 4] = a3 - a2;
582    tmp[3 + i * 4] = a0 - a1;
583  }
584  // vertical pass
585  for (i = 0; i < 4; ++i, ++w) {
586    const int a0 = tmp[0 + i] + tmp[8 + i];
587    const int a1 = tmp[4 + i] + tmp[12+ i];
588    const int a2 = tmp[4 + i] - tmp[12+ i];
589    const int a3 = tmp[0 + i] - tmp[8 + i];
590    const int b0 = a0 + a1;
591    const int b1 = a3 + a2;
592    const int b2 = a3 - a2;
593    const int b3 = a0 - a1;
594
595    sum += w[ 0] * abs(b0);
596    sum += w[ 4] * abs(b1);
597    sum += w[ 8] * abs(b2);
598    sum += w[12] * abs(b3);
599  }
600  return sum;
601}
602
603static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
604                    const uint16_t* const w) {
605  const int sum1 = TTransform(a, w);
606  const int sum2 = TTransform(b, w);
607  return abs(sum2 - sum1) >> 5;
608}
609
610static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
611                      const uint16_t* const w) {
612  int D = 0;
613  int x, y;
614  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
615    for (x = 0; x < 16; x += 4) {
616      D += Disto4x4(a + x + y, b + x + y, w);
617    }
618  }
619  return D;
620}
621
622//------------------------------------------------------------------------------
623// Quantization
624//
625
626static const uint8_t kZigzag[16] = {
627  0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
628};
629
630// Simple quantization
631static int QuantizeBlock(int16_t in[16], int16_t out[16],
632                         int n, const VP8Matrix* const mtx) {
633  int last = -1;
634  for (; n < 16; ++n) {
635    const int j = kZigzag[n];
636    const int sign = (in[j] < 0);
637    const int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
638    if (coeff > mtx->zthresh_[j]) {
639      const int Q = mtx->q_[j];
640      const int iQ = mtx->iq_[j];
641      const int B = mtx->bias_[j];
642      out[n] = QUANTDIV(coeff, iQ, B);
643      if (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
644      if (sign) out[n] = -out[n];
645      in[j] = out[n] * Q;
646      if (out[n]) last = n;
647    } else {
648      out[n] = 0;
649      in[j] = 0;
650    }
651  }
652  return (last >= 0);
653}
654
655//------------------------------------------------------------------------------
656// Block copy
657
658static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) {
659  int y;
660  for (y = 0; y < size; ++y) {
661    memcpy(dst, src, size);
662    src += BPS;
663    dst += BPS;
664  }
665}
666
667static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
668
669//------------------------------------------------------------------------------
670// Initialization
671
672// Speed-critical function pointers. We have to initialize them to the default
673// implementations within VP8EncDspInit().
674VP8CHisto VP8CollectHistogram;
675VP8Idct VP8ITransform;
676VP8Fdct VP8FTransform;
677VP8WHT VP8ITransformWHT;
678VP8WHT VP8FTransformWHT;
679VP8Intra4Preds VP8EncPredLuma4;
680VP8IntraPreds VP8EncPredLuma16;
681VP8IntraPreds VP8EncPredChroma8;
682VP8Metric VP8SSE16x16;
683VP8Metric VP8SSE8x8;
684VP8Metric VP8SSE16x8;
685VP8Metric VP8SSE4x4;
686VP8WMetric VP8TDisto4x4;
687VP8WMetric VP8TDisto16x16;
688VP8QuantizeBlock VP8EncQuantizeBlock;
689VP8BlockCopy VP8Copy4x4;
690
691extern void VP8EncDspInitSSE2(void);
692extern void VP8EncDspInitNEON(void);
693
694void VP8EncDspInit(void) {
695  InitTables();
696
697  // default C implementations
698  VP8CollectHistogram = CollectHistogram;
699  VP8ITransform = ITransform;
700  VP8FTransform = FTransform;
701  VP8ITransformWHT = ITransformWHT;
702  VP8FTransformWHT = FTransformWHT;
703  VP8EncPredLuma4 = Intra4Preds;
704  VP8EncPredLuma16 = Intra16Preds;
705  VP8EncPredChroma8 = IntraChromaPreds;
706  VP8SSE16x16 = SSE16x16;
707  VP8SSE8x8 = SSE8x8;
708  VP8SSE16x8 = SSE16x8;
709  VP8SSE4x4 = SSE4x4;
710  VP8TDisto4x4 = Disto4x4;
711  VP8TDisto16x16 = Disto16x16;
712  VP8EncQuantizeBlock = QuantizeBlock;
713  VP8Copy4x4 = Copy4x4;
714
715  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
716  if (VP8GetCPUInfo) {
717#if defined(WEBP_USE_SSE2)
718    if (VP8GetCPUInfo(kSSE2)) {
719      VP8EncDspInitSSE2();
720    }
721#elif defined(WEBP_USE_NEON)
722    if (VP8GetCPUInfo(kNEON)) {
723      VP8EncDspInitNEON();
724    }
725#endif
726  }
727}
728
729#if defined(__cplusplus) || defined(c_plusplus)
730}    // extern "C"
731#endif
732