enc.c revision 8b720228d581a84fd173b6dcb2fa295b59db489a 
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 <assert.h>
15#include <stdlib.h>  // for abs()
16
17#include "./dsp.h"
18#include "../enc/vp8enci.h"
19
20static WEBP_INLINE uint8_t clip_8b(int v) {
21  return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
22}
23
24static WEBP_INLINE int clip_max(int v, int max) {
25  return (v > max) ? max : v;
26}
27
28//------------------------------------------------------------------------------
29// Compute susceptibility based on DCT-coeff histograms:
30// the higher, the "easier" the macroblock is to compress.
31
32const int VP8DspScan[16 + 4 + 4] = {
33  // Luma
34  0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
35  0 +  4 * BPS,  4 +  4 * BPS, 8 +  4 * BPS, 12 +  4 * BPS,
36  0 +  8 * BPS,  4 +  8 * BPS, 8 +  8 * BPS, 12 +  8 * BPS,
37  0 + 12 * BPS,  4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS,
38
39  0 + 0 * BPS,   4 + 0 * BPS, 0 + 4 * BPS,  4 + 4 * BPS,    // U
40  8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
41};
42
43static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
44                             int start_block, int end_block,
45                             VP8Histogram* const histo) {
46  int j;
47  for (j = start_block; j < end_block; ++j) {
48    int k;
49    int16_t out[16];
50
51    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
52
53    // Convert coefficients to bin.
54    for (k = 0; k < 16; ++k) {
55      const int v = abs(out[k]) >> 3;  // TODO(skal): add rounding?
56      const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
57      histo->distribution[clipped_value]++;
58    }
59  }
60}
61
62//------------------------------------------------------------------------------
63// run-time tables (~4k)
64
65static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
66
67// We declare this variable 'volatile' to prevent instruction reordering
68// and make sure it's set to true _last_ (so as to be thread-safe)
69static volatile int tables_ok = 0;
70
71static void InitTables(void) {
72  if (!tables_ok) {
73    int i;
74    for (i = -255; i <= 255 + 255; ++i) {
75      clip1[255 + i] = clip_8b(i);
76    }
77    tables_ok = 1;
78  }
79}
80
81
82//------------------------------------------------------------------------------
83// Transforms (Paragraph 14.4)
84
85#define STORE(x, y, v) \
86  dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3))
87
88static const int kC1 = 20091 + (1 << 16);
89static const int kC2 = 35468;
90#define MUL(a, b) (((a) * (b)) >> 16)
91
92static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
93                                      uint8_t* dst) {
94  int C[4 * 4], *tmp;
95  int i;
96  tmp = C;
97  for (i = 0; i < 4; ++i) {    // vertical pass
98    const int a = in[0] + in[8];
99    const int b = in[0] - in[8];
100    const int c = MUL(in[4], kC2) - MUL(in[12], kC1);
101    const int d = MUL(in[4], kC1) + MUL(in[12], kC2);
102    tmp[0] = a + d;
103    tmp[1] = b + c;
104    tmp[2] = b - c;
105    tmp[3] = a - d;
106    tmp += 4;
107    in++;
108  }
109
110  tmp = C;
111  for (i = 0; i < 4; ++i) {    // horizontal pass
112    const int dc = tmp[0] + 4;
113    const int a =  dc +  tmp[8];
114    const int b =  dc -  tmp[8];
115    const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1);
116    const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2);
117    STORE(0, i, a + d);
118    STORE(1, i, b + c);
119    STORE(2, i, b - c);
120    STORE(3, i, a - d);
121    tmp++;
122  }
123}
124
125static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
126                       int do_two) {
127  ITransformOne(ref, in, dst);
128  if (do_two) {
129    ITransformOne(ref + 4, in + 16, dst + 4);
130  }
131}
132
133static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
134  int i;
135  int tmp[16];
136  for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
137    const int d0 = src[0] - ref[0];   // 9bit dynamic range ([-255,255])
138    const int d1 = src[1] - ref[1];
139    const int d2 = src[2] - ref[2];
140    const int d3 = src[3] - ref[3];
141    const int a0 = (d0 + d3);         // 10b                      [-510,510]
142    const int a1 = (d1 + d2);
143    const int a2 = (d1 - d2);
144    const int a3 = (d0 - d3);
145    tmp[0 + i * 4] = (a0 + a1) * 8;   // 14b                      [-8160,8160]
146    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
147    tmp[2 + i * 4] = (a0 - a1) * 8;
148    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  937) >> 9;
149  }
150  for (i = 0; i < 4; ++i) {
151    const int a0 = (tmp[0 + i] + tmp[12 + i]);  // 15b
152    const int a1 = (tmp[4 + i] + tmp[ 8 + i]);
153    const int a2 = (tmp[4 + i] - tmp[ 8 + i]);
154    const int a3 = (tmp[0 + i] - tmp[12 + i]);
155    out[0 + i] = (a0 + a1 + 7) >> 4;            // 12b
156    out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0);
157    out[8 + i] = (a0 - a1 + 7) >> 4;
158    out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16);
159  }
160}
161
162static void ITransformWHT(const int16_t* in, int16_t* out) {
163  int tmp[16];
164  int i;
165  for (i = 0; i < 4; ++i) {
166    const int a0 = in[0 + i] + in[12 + i];
167    const int a1 = in[4 + i] + in[ 8 + i];
168    const int a2 = in[4 + i] - in[ 8 + i];
169    const int a3 = in[0 + i] - in[12 + i];
170    tmp[0  + i] = a0 + a1;
171    tmp[8  + i] = a0 - a1;
172    tmp[4  + i] = a3 + a2;
173    tmp[12 + i] = a3 - a2;
174  }
175  for (i = 0; i < 4; ++i) {
176    const int dc = tmp[0 + i * 4] + 3;    // w/ rounder
177    const int a0 = dc             + tmp[3 + i * 4];
178    const int a1 = tmp[1 + i * 4] + tmp[2 + i * 4];
179    const int a2 = tmp[1 + i * 4] - tmp[2 + i * 4];
180    const int a3 = dc             - tmp[3 + i * 4];
181    out[ 0] = (a0 + a1) >> 3;
182    out[16] = (a3 + a2) >> 3;
183    out[32] = (a0 - a1) >> 3;
184    out[48] = (a3 - a2) >> 3;
185    out += 64;
186  }
187}
188
189static void FTransformWHT(const int16_t* in, int16_t* out) {
190  // input is 12b signed
191  int32_t tmp[16];
192  int i;
193  for (i = 0; i < 4; ++i, in += 64) {
194    const int a0 = (in[0 * 16] + in[2 * 16]);  // 13b
195    const int a1 = (in[1 * 16] + in[3 * 16]);
196    const int a2 = (in[1 * 16] - in[3 * 16]);
197    const int a3 = (in[0 * 16] - in[2 * 16]);
198    tmp[0 + i * 4] = a0 + a1;   // 14b
199    tmp[1 + i * 4] = a3 + a2;
200    tmp[2 + i * 4] = a3 - a2;
201    tmp[3 + i * 4] = a0 - a1;
202  }
203  for (i = 0; i < 4; ++i) {
204    const int a0 = (tmp[0 + i] + tmp[8 + i]);  // 15b
205    const int a1 = (tmp[4 + i] + tmp[12+ i]);
206    const int a2 = (tmp[4 + i] - tmp[12+ i]);
207    const int a3 = (tmp[0 + i] - tmp[8 + i]);
208    const int b0 = a0 + a1;    // 16b
209    const int b1 = a3 + a2;
210    const int b2 = a3 - a2;
211    const int b3 = a0 - a1;
212    out[ 0 + i] = b0 >> 1;     // 15b
213    out[ 4 + i] = b1 >> 1;
214    out[ 8 + i] = b2 >> 1;
215    out[12 + i] = b3 >> 1;
216  }
217}
218
219#undef MUL
220#undef STORE
221
222//------------------------------------------------------------------------------
223// Intra predictions
224
225#define DST(x, y) dst[(x) + (y) * BPS]
226
227static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
228  int j;
229  for (j = 0; j < size; ++j) {
230    memset(dst + j * BPS, value, size);
231  }
232}
233
234static WEBP_INLINE void VerticalPred(uint8_t* dst,
235                                     const uint8_t* top, int size) {
236  int j;
237  if (top) {
238    for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
239  } else {
240    Fill(dst, 127, size);
241  }
242}
243
244static WEBP_INLINE void HorizontalPred(uint8_t* dst,
245                                       const uint8_t* left, int size) {
246  if (left) {
247    int j;
248    for (j = 0; j < size; ++j) {
249      memset(dst + j * BPS, left[j], size);
250    }
251  } else {
252    Fill(dst, 129, size);
253  }
254}
255
256static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
257                                   const uint8_t* top, int size) {
258  int y;
259  if (left) {
260    if (top) {
261      const uint8_t* const clip = clip1 + 255 - left[-1];
262      for (y = 0; y < size; ++y) {
263        const uint8_t* const clip_table = clip + left[y];
264        int x;
265        for (x = 0; x < size; ++x) {
266          dst[x] = clip_table[top[x]];
267        }
268        dst += BPS;
269      }
270    } else {
271      HorizontalPred(dst, left, size);
272    }
273  } else {
274    // true motion without left samples (hence: with default 129 value)
275    // is equivalent to VE prediction where you just copy the top samples.
276    // Note that if top samples are not available, the default value is
277    // then 129, and not 127 as in the VerticalPred case.
278    if (top) {
279      VerticalPred(dst, top, size);
280    } else {
281      Fill(dst, 129, size);
282    }
283  }
284}
285
286static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
287                               const uint8_t* top,
288                               int size, int round, int shift) {
289  int DC = 0;
290  int j;
291  if (top) {
292    for (j = 0; j < size; ++j) DC += top[j];
293    if (left) {   // top and left present
294      for (j = 0; j < size; ++j) DC += left[j];
295    } else {      // top, but no left
296      DC += DC;
297    }
298    DC = (DC + round) >> shift;
299  } else if (left) {   // left but no top
300    for (j = 0; j < size; ++j) DC += left[j];
301    DC += DC;
302    DC = (DC + round) >> shift;
303  } else {   // no top, no left, nothing.
304    DC = 0x80;
305  }
306  Fill(dst, DC, size);
307}
308
309//------------------------------------------------------------------------------
310// Chroma 8x8 prediction (paragraph 12.2)
311
312static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
313                             const uint8_t* top) {
314  // U block
315  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
316  VerticalPred(C8VE8 + dst, top, 8);
317  HorizontalPred(C8HE8 + dst, left, 8);
318  TrueMotion(C8TM8 + dst, left, top, 8);
319  // V block
320  dst += 8;
321  if (top) top += 8;
322  if (left) left += 16;
323  DCMode(C8DC8 + dst, left, top, 8, 8, 4);
324  VerticalPred(C8VE8 + dst, top, 8);
325  HorizontalPred(C8HE8 + dst, left, 8);
326  TrueMotion(C8TM8 + dst, left, top, 8);
327}
328
329//------------------------------------------------------------------------------
330// luma 16x16 prediction (paragraph 12.3)
331
332static void Intra16Preds(uint8_t* dst,
333                         const uint8_t* left, const uint8_t* top) {
334  DCMode(I16DC16 + dst, left, top, 16, 16, 5);
335  VerticalPred(I16VE16 + dst, top, 16);
336  HorizontalPred(I16HE16 + dst, left, 16);
337  TrueMotion(I16TM16 + dst, left, top, 16);
338}
339
340//------------------------------------------------------------------------------
341// luma 4x4 prediction
342
343#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
344#define AVG2(a, b) (((a) + (b) + 1) >> 1)
345
346static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
347  const uint8_t vals[4] = {
348    AVG3(top[-1], top[0], top[1]),
349    AVG3(top[ 0], top[1], top[2]),
350    AVG3(top[ 1], top[2], top[3]),
351    AVG3(top[ 2], top[3], top[4])
352  };
353  int i;
354  for (i = 0; i < 4; ++i) {
355    memcpy(dst + i * BPS, vals, 4);
356  }
357}
358
359static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
360  const int X = top[-1];
361  const int I = top[-2];
362  const int J = top[-3];
363  const int K = top[-4];
364  const int L = top[-5];
365  *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(X, I, J);
366  *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(I, J, K);
367  *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(J, K, L);
368  *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(K, L, L);
369}
370
371static void DC4(uint8_t* dst, const uint8_t* top) {
372  uint32_t dc = 4;
373  int i;
374  for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
375  Fill(dst, dc >> 3, 4);
376}
377
378static void RD4(uint8_t* dst, const uint8_t* top) {
379  const int X = top[-1];
380  const int I = top[-2];
381  const int J = top[-3];
382  const int K = top[-4];
383  const int L = top[-5];
384  const int A = top[0];
385  const int B = top[1];
386  const int C = top[2];
387  const int D = top[3];
388  DST(0, 3)                                     = AVG3(J, K, L);
389  DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
390  DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
391  DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
392  DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
393  DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
394  DST(3, 0)                                     = AVG3(D, C, B);
395}
396
397static void LD4(uint8_t* dst, const uint8_t* top) {
398  const int A = top[0];
399  const int B = top[1];
400  const int C = top[2];
401  const int D = top[3];
402  const int E = top[4];
403  const int F = top[5];
404  const int G = top[6];
405  const int H = top[7];
406  DST(0, 0)                                     = AVG3(A, B, C);
407  DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
408  DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
409  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
410  DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
411  DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
412  DST(3, 3)                                     = AVG3(G, H, H);
413}
414
415static void VR4(uint8_t* dst, const uint8_t* top) {
416  const int X = top[-1];
417  const int I = top[-2];
418  const int J = top[-3];
419  const int K = top[-4];
420  const int A = top[0];
421  const int B = top[1];
422  const int C = top[2];
423  const int D = top[3];
424  DST(0, 0) = DST(1, 2) = AVG2(X, A);
425  DST(1, 0) = DST(2, 2) = AVG2(A, B);
426  DST(2, 0) = DST(3, 2) = AVG2(B, C);
427  DST(3, 0)             = AVG2(C, D);
428
429  DST(0, 3) =             AVG3(K, J, I);
430  DST(0, 2) =             AVG3(J, I, X);
431  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
432  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
433  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
434  DST(3, 1) =             AVG3(B, C, D);
435}
436
437static void VL4(uint8_t* dst, const uint8_t* top) {
438  const int A = top[0];
439  const int B = top[1];
440  const int C = top[2];
441  const int D = top[3];
442  const int E = top[4];
443  const int F = top[5];
444  const int G = top[6];
445  const int H = top[7];
446  DST(0, 0) =             AVG2(A, B);
447  DST(1, 0) = DST(0, 2) = AVG2(B, C);
448  DST(2, 0) = DST(1, 2) = AVG2(C, D);
449  DST(3, 0) = DST(2, 2) = AVG2(D, E);
450
451  DST(0, 1) =             AVG3(A, B, C);
452  DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
453  DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
454  DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
455              DST(3, 2) = AVG3(E, F, G);
456              DST(3, 3) = AVG3(F, G, H);
457}
458
459static void HU4(uint8_t* dst, const uint8_t* top) {
460  const int I = top[-2];
461  const int J = top[-3];
462  const int K = top[-4];
463  const int L = top[-5];
464  DST(0, 0) =             AVG2(I, J);
465  DST(2, 0) = DST(0, 1) = AVG2(J, K);
466  DST(2, 1) = DST(0, 2) = AVG2(K, L);
467  DST(1, 0) =             AVG3(I, J, K);
468  DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
469  DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
470  DST(3, 2) = DST(2, 2) =
471  DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
472}
473
474static void HD4(uint8_t* dst, const uint8_t* top) {
475  const int X = top[-1];
476  const int I = top[-2];
477  const int J = top[-3];
478  const int K = top[-4];
479  const int L = top[-5];
480  const int A = top[0];
481  const int B = top[1];
482  const int C = top[2];
483
484  DST(0, 0) = DST(2, 1) = AVG2(I, X);
485  DST(0, 1) = DST(2, 2) = AVG2(J, I);
486  DST(0, 2) = DST(2, 3) = AVG2(K, J);
487  DST(0, 3)             = AVG2(L, K);
488
489  DST(3, 0)             = AVG3(A, B, C);
490  DST(2, 0)             = AVG3(X, A, B);
491  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
492  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
493  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
494  DST(1, 3)             = AVG3(L, K, J);
495}
496
497static void TM4(uint8_t* dst, const uint8_t* top) {
498  int x, y;
499  const uint8_t* const clip = clip1 + 255 - top[-1];
500  for (y = 0; y < 4; ++y) {
501    const uint8_t* const clip_table = clip + top[-2 - y];
502    for (x = 0; x < 4; ++x) {
503      dst[x] = clip_table[top[x]];
504    }
505    dst += BPS;
506  }
507}
508
509#undef DST
510#undef AVG3
511#undef AVG2
512
513// Left samples are top[-5 .. -2], top_left is top[-1], top are
514// located at top[0..3], and top right is top[4..7]
515static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
516  DC4(I4DC4 + dst, top);
517  TM4(I4TM4 + dst, top);
518  VE4(I4VE4 + dst, top);
519  HE4(I4HE4 + dst, top);
520  RD4(I4RD4 + dst, top);
521  VR4(I4VR4 + dst, top);
522  LD4(I4LD4 + dst, top);
523  VL4(I4VL4 + dst, top);
524  HD4(I4HD4 + dst, top);
525  HU4(I4HU4 + dst, top);
526}
527
528//------------------------------------------------------------------------------
529// Metric
530
531static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
532                              int w, int h) {
533  int count = 0;
534  int y, x;
535  for (y = 0; y < h; ++y) {
536    for (x = 0; x < w; ++x) {
537      const int diff = (int)a[x] - b[x];
538      count += diff * diff;
539    }
540    a += BPS;
541    b += BPS;
542  }
543  return count;
544}
545
546static int SSE16x16(const uint8_t* a, const uint8_t* b) {
547  return GetSSE(a, b, 16, 16);
548}
549static int SSE16x8(const uint8_t* a, const uint8_t* b) {
550  return GetSSE(a, b, 16, 8);
551}
552static int SSE8x8(const uint8_t* a, const uint8_t* b) {
553  return GetSSE(a, b, 8, 8);
554}
555static int SSE4x4(const uint8_t* a, const uint8_t* b) {
556  return GetSSE(a, b, 4, 4);
557}
558
559//------------------------------------------------------------------------------
560// Texture distortion
561//
562// We try to match the spectral content (weighted) between source and
563// reconstructed samples.
564
565// Hadamard transform
566// Returns the weighted sum of the absolute value of transformed coefficients.
567static int TTransform(const uint8_t* in, const uint16_t* w) {
568  int sum = 0;
569  int tmp[16];
570  int i;
571  // horizontal pass
572  for (i = 0; i < 4; ++i, in += BPS) {
573    const int a0 = in[0] + in[2];
574    const int a1 = in[1] + in[3];
575    const int a2 = in[1] - in[3];
576    const int a3 = in[0] - in[2];
577    tmp[0 + i * 4] = a0 + a1;
578    tmp[1 + i * 4] = a3 + a2;
579    tmp[2 + i * 4] = a3 - a2;
580    tmp[3 + i * 4] = a0 - a1;
581  }
582  // vertical pass
583  for (i = 0; i < 4; ++i, ++w) {
584    const int a0 = tmp[0 + i] + tmp[8 + i];
585    const int a1 = tmp[4 + i] + tmp[12+ i];
586    const int a2 = tmp[4 + i] - tmp[12+ i];
587    const int a3 = tmp[0 + i] - tmp[8 + i];
588    const int b0 = a0 + a1;
589    const int b1 = a3 + a2;
590    const int b2 = a3 - a2;
591    const int b3 = a0 - a1;
592
593    sum += w[ 0] * abs(b0);
594    sum += w[ 4] * abs(b1);
595    sum += w[ 8] * abs(b2);
596    sum += w[12] * abs(b3);
597  }
598  return sum;
599}
600
601static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
602                    const uint16_t* const w) {
603  const int sum1 = TTransform(a, w);
604  const int sum2 = TTransform(b, w);
605  return abs(sum2 - sum1) >> 5;
606}
607
608static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
609                      const uint16_t* const w) {
610  int D = 0;
611  int x, y;
612  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
613    for (x = 0; x < 16; x += 4) {
614      D += Disto4x4(a + x + y, b + x + y, w);
615    }
616  }
617  return D;
618}
619
620//------------------------------------------------------------------------------
621// Quantization
622//
623
624static const uint8_t kZigzag[16] = {
625  0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
626};
627
628// Simple quantization
629static int QuantizeBlock(int16_t in[16], int16_t out[16],
630                         int n, const VP8Matrix* const mtx) {
631  int last = -1;
632  for (; n < 16; ++n) {
633    const int j = kZigzag[n];
634    const int sign = (in[j] < 0);
635    const int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
636    if (coeff > mtx->zthresh_[j]) {
637      const int Q = mtx->q_[j];
638      const int iQ = mtx->iq_[j];
639      const int B = mtx->bias_[j];
640      out[n] = QUANTDIV(coeff, iQ, B);
641      if (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
642      if (sign) out[n] = -out[n];
643      in[j] = out[n] * Q;
644      if (out[n]) last = n;
645    } else {
646      out[n] = 0;
647      in[j] = 0;
648    }
649  }
650  return (last >= 0);
651}
652
653static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
654                            const VP8Matrix* const mtx) {
655  int n, last = -1;
656  for (n = 0; n < 16; ++n) {
657    const int j = kZigzag[n];
658    const int sign = (in[j] < 0);
659    const int coeff = sign ? -in[j] : in[j];
660    assert(mtx->sharpen_[j] == 0);
661    if (coeff > mtx->zthresh_[j]) {
662      const int Q = mtx->q_[j];
663      const int iQ = mtx->iq_[j];
664      const int B = mtx->bias_[j];
665      out[n] = QUANTDIV(coeff, iQ, B);
666      if (out[n] > MAX_LEVEL) out[n] = MAX_LEVEL;
667      if (sign) out[n] = -out[n];
668      in[j] = out[n] * Q;
669      if (out[n]) last = n;
670    } else {
671      out[n] = 0;
672      in[j] = 0;
673    }
674  }
675  return (last >= 0);
676}
677
678//------------------------------------------------------------------------------
679// Block copy
680
681static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) {
682  int y;
683  for (y = 0; y < size; ++y) {
684    memcpy(dst, src, size);
685    src += BPS;
686    dst += BPS;
687  }
688}
689
690static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
691
692//------------------------------------------------------------------------------
693// Initialization
694
695// Speed-critical function pointers. We have to initialize them to the default
696// implementations within VP8EncDspInit().
697VP8CHisto VP8CollectHistogram;
698VP8Idct VP8ITransform;
699VP8Fdct VP8FTransform;
700VP8WHT VP8ITransformWHT;
701VP8WHT VP8FTransformWHT;
702VP8Intra4Preds VP8EncPredLuma4;
703VP8IntraPreds VP8EncPredLuma16;
704VP8IntraPreds VP8EncPredChroma8;
705VP8Metric VP8SSE16x16;
706VP8Metric VP8SSE8x8;
707VP8Metric VP8SSE16x8;
708VP8Metric VP8SSE4x4;
709VP8WMetric VP8TDisto4x4;
710VP8WMetric VP8TDisto16x16;
711VP8QuantizeBlock VP8EncQuantizeBlock;
712VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
713VP8BlockCopy VP8Copy4x4;
714
715extern void VP8EncDspInitSSE2(void);
716extern void VP8EncDspInitNEON(void);
717
718void VP8EncDspInit(void) {
719  InitTables();
720
721  // default C implementations
722  VP8CollectHistogram = CollectHistogram;
723  VP8ITransform = ITransform;
724  VP8FTransform = FTransform;
725  VP8ITransformWHT = ITransformWHT;
726  VP8FTransformWHT = FTransformWHT;
727  VP8EncPredLuma4 = Intra4Preds;
728  VP8EncPredLuma16 = Intra16Preds;
729  VP8EncPredChroma8 = IntraChromaPreds;
730  VP8SSE16x16 = SSE16x16;
731  VP8SSE8x8 = SSE8x8;
732  VP8SSE16x8 = SSE16x8;
733  VP8SSE4x4 = SSE4x4;
734  VP8TDisto4x4 = Disto4x4;
735  VP8TDisto16x16 = Disto16x16;
736  VP8EncQuantizeBlock = QuantizeBlock;
737  VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
738  VP8Copy4x4 = Copy4x4;
739
740  // If defined, use CPUInfo() to overwrite some pointers with faster versions.
741  if (VP8GetCPUInfo) {
742#if defined(WEBP_USE_SSE2)
743    if (VP8GetCPUInfo(kSSE2)) {
744      VP8EncDspInitSSE2();
745    }
746#elif defined(WEBP_USE_NEON)
747    if (VP8GetCPUInfo(kNEON)) {
748      VP8EncDspInitNEON();
749    }
750#endif
751  }
752}
753
754