• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2011 Google Inc. All Rights Reserved.
2 //
3 // This code is licensed under the same terms as WebM:
4 //  Software License Agreement:  http://www.webmproject.org/license/software/
5 //  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
6 // -----------------------------------------------------------------------------
7 //
8 // Speed-critical encoding functions.
9 //
10 // Author: Skal (pascal.massimino@gmail.com)
11 
12 #include <stdlib.h>  // for abs()
13 #include "./dsp.h"
14 #include "../enc/vp8enci.h"
15 
16 #if defined(__cplusplus) || defined(c_plusplus)
17 extern "C" {
18 #endif
19 
clip_8b(int v)20 static WEBP_INLINE uint8_t clip_8b(int v) {
21   return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
22 }
23 
clip_max(int v,int max)24 static 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 
32 const 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 
CollectHistogram(const uint8_t * ref,const uint8_t * pred,int start_block,int end_block,VP8Histogram * const histo)43 static 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 
65 static 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)
69 static volatile int tables_ok = 0;
70 
InitTables(void)71 static 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 
88 static const int kC1 = 20091 + (1 << 16);
89 static const int kC2 = 35468;
90 #define MUL(a, b) (((a) * (b)) >> 16)
91 
ITransformOne(const uint8_t * ref,const int16_t * in,uint8_t * dst)92 static 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 
ITransform(const uint8_t * ref,const int16_t * in,uint8_t * dst,int do_two)125 static 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 
FTransform(const uint8_t * src,const uint8_t * ref,int16_t * out)133 static 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) << 3;  // 14b                      [-8160,8160]
146     tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
147     tmp[2 + i * 4] = (a0 - a1) << 3;
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 
ITransformWHT(const int16_t * in,int16_t * out)162 static 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 
FTransformWHT(const int16_t * in,int16_t * out)189 static void FTransformWHT(const int16_t* in, int16_t* out) {
190   int tmp[16];
191   int i;
192   for (i = 0; i < 4; ++i, in += 64) {
193     const int a0 = (in[0 * 16] + in[2 * 16]) << 2;
194     const int a1 = (in[1 * 16] + in[3 * 16]) << 2;
195     const int a2 = (in[1 * 16] - in[3 * 16]) << 2;
196     const int a3 = (in[0 * 16] - in[2 * 16]) << 2;
197     tmp[0 + i * 4] = (a0 + a1) + (a0 != 0);
198     tmp[1 + i * 4] = a3 + a2;
199     tmp[2 + i * 4] = a3 - a2;
200     tmp[3 + i * 4] = a0 - a1;
201   }
202   for (i = 0; i < 4; ++i) {
203     const int a0 = (tmp[0 + i] + tmp[8 + i]);
204     const int a1 = (tmp[4 + i] + tmp[12+ i]);
205     const int a2 = (tmp[4 + i] - tmp[12+ i]);
206     const int a3 = (tmp[0 + i] - tmp[8 + i]);
207     const int b0 = a0 + a1;
208     const int b1 = a3 + a2;
209     const int b2 = a3 - a2;
210     const int b3 = a0 - a1;
211     out[ 0 + i] = (b0 + (b0 > 0) + 3) >> 3;
212     out[ 4 + i] = (b1 + (b1 > 0) + 3) >> 3;
213     out[ 8 + i] = (b2 + (b2 > 0) + 3) >> 3;
214     out[12 + i] = (b3 + (b3 > 0) + 3) >> 3;
215   }
216 }
217 
218 #undef MUL
219 #undef STORE
220 
221 //------------------------------------------------------------------------------
222 // Intra predictions
223 
224 #define DST(x, y) dst[(x) + (y) * BPS]
225 
Fill(uint8_t * dst,int value,int size)226 static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
227   int j;
228   for (j = 0; j < size; ++j) {
229     memset(dst + j * BPS, value, size);
230   }
231 }
232 
VerticalPred(uint8_t * dst,const uint8_t * top,int size)233 static WEBP_INLINE void VerticalPred(uint8_t* dst,
234                                      const uint8_t* top, int size) {
235   int j;
236   if (top) {
237     for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
238   } else {
239     Fill(dst, 127, size);
240   }
241 }
242 
HorizontalPred(uint8_t * dst,const uint8_t * left,int size)243 static WEBP_INLINE void HorizontalPred(uint8_t* dst,
244                                        const uint8_t* left, int size) {
245   if (left) {
246     int j;
247     for (j = 0; j < size; ++j) {
248       memset(dst + j * BPS, left[j], size);
249     }
250   } else {
251     Fill(dst, 129, size);
252   }
253 }
254 
TrueMotion(uint8_t * dst,const uint8_t * left,const uint8_t * top,int size)255 static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
256                                    const uint8_t* top, int size) {
257   int y;
258   if (left) {
259     if (top) {
260       const uint8_t* const clip = clip1 + 255 - left[-1];
261       for (y = 0; y < size; ++y) {
262         const uint8_t* const clip_table = clip + left[y];
263         int x;
264         for (x = 0; x < size; ++x) {
265           dst[x] = clip_table[top[x]];
266         }
267         dst += BPS;
268       }
269     } else {
270       HorizontalPred(dst, left, size);
271     }
272   } else {
273     // true motion without left samples (hence: with default 129 value)
274     // is equivalent to VE prediction where you just copy the top samples.
275     // Note that if top samples are not available, the default value is
276     // then 129, and not 127 as in the VerticalPred case.
277     if (top) {
278       VerticalPred(dst, top, size);
279     } else {
280       Fill(dst, 129, size);
281     }
282   }
283 }
284 
DCMode(uint8_t * dst,const uint8_t * left,const uint8_t * top,int size,int round,int shift)285 static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
286                                const uint8_t* top,
287                                int size, int round, int shift) {
288   int DC = 0;
289   int j;
290   if (top) {
291     for (j = 0; j < size; ++j) DC += top[j];
292     if (left) {   // top and left present
293       for (j = 0; j < size; ++j) DC += left[j];
294     } else {      // top, but no left
295       DC += DC;
296     }
297     DC = (DC + round) >> shift;
298   } else if (left) {   // left but no top
299     for (j = 0; j < size; ++j) DC += left[j];
300     DC += DC;
301     DC = (DC + round) >> shift;
302   } else {   // no top, no left, nothing.
303     DC = 0x80;
304   }
305   Fill(dst, DC, size);
306 }
307 
308 //------------------------------------------------------------------------------
309 // Chroma 8x8 prediction (paragraph 12.2)
310 
IntraChromaPreds(uint8_t * dst,const uint8_t * left,const uint8_t * top)311 static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
312                              const uint8_t* top) {
313   // U block
314   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
315   VerticalPred(C8VE8 + dst, top, 8);
316   HorizontalPred(C8HE8 + dst, left, 8);
317   TrueMotion(C8TM8 + dst, left, top, 8);
318   // V block
319   dst += 8;
320   if (top) top += 8;
321   if (left) left += 16;
322   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
323   VerticalPred(C8VE8 + dst, top, 8);
324   HorizontalPred(C8HE8 + dst, left, 8);
325   TrueMotion(C8TM8 + dst, left, top, 8);
326 }
327 
328 //------------------------------------------------------------------------------
329 // luma 16x16 prediction (paragraph 12.3)
330 
Intra16Preds(uint8_t * dst,const uint8_t * left,const uint8_t * top)331 static void Intra16Preds(uint8_t* dst,
332                          const uint8_t* left, const uint8_t* top) {
333   DCMode(I16DC16 + dst, left, top, 16, 16, 5);
334   VerticalPred(I16VE16 + dst, top, 16);
335   HorizontalPred(I16HE16 + dst, left, 16);
336   TrueMotion(I16TM16 + dst, left, top, 16);
337 }
338 
339 //------------------------------------------------------------------------------
340 // luma 4x4 prediction
341 
342 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
343 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
344 
VE4(uint8_t * dst,const uint8_t * top)345 static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
346   const uint8_t vals[4] = {
347     AVG3(top[-1], top[0], top[1]),
348     AVG3(top[ 0], top[1], top[2]),
349     AVG3(top[ 1], top[2], top[3]),
350     AVG3(top[ 2], top[3], top[4])
351   };
352   int i;
353   for (i = 0; i < 4; ++i) {
354     memcpy(dst + i * BPS, vals, 4);
355   }
356 }
357 
HE4(uint8_t * dst,const uint8_t * top)358 static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
359   const int X = top[-1];
360   const int I = top[-2];
361   const int J = top[-3];
362   const int K = top[-4];
363   const int L = top[-5];
364   *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(X, I, J);
365   *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(I, J, K);
366   *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(J, K, L);
367   *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(K, L, L);
368 }
369 
DC4(uint8_t * dst,const uint8_t * top)370 static void DC4(uint8_t* dst, const uint8_t* top) {
371   uint32_t dc = 4;
372   int i;
373   for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
374   Fill(dst, dc >> 3, 4);
375 }
376 
RD4(uint8_t * dst,const uint8_t * top)377 static void RD4(uint8_t* dst, const uint8_t* top) {
378   const int X = top[-1];
379   const int I = top[-2];
380   const int J = top[-3];
381   const int K = top[-4];
382   const int L = top[-5];
383   const int A = top[0];
384   const int B = top[1];
385   const int C = top[2];
386   const int D = top[3];
387   DST(0, 3)                                     = AVG3(J, K, L);
388   DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
389   DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
390   DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
391   DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
392   DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
393   DST(3, 0)                                     = AVG3(D, C, B);
394 }
395 
LD4(uint8_t * dst,const uint8_t * top)396 static void LD4(uint8_t* dst, const uint8_t* top) {
397   const int A = top[0];
398   const int B = top[1];
399   const int C = top[2];
400   const int D = top[3];
401   const int E = top[4];
402   const int F = top[5];
403   const int G = top[6];
404   const int H = top[7];
405   DST(0, 0)                                     = AVG3(A, B, C);
406   DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
407   DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
408   DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
409   DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
410   DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
411   DST(3, 3)                                     = AVG3(G, H, H);
412 }
413 
VR4(uint8_t * dst,const uint8_t * top)414 static void VR4(uint8_t* dst, const uint8_t* top) {
415   const int X = top[-1];
416   const int I = top[-2];
417   const int J = top[-3];
418   const int K = top[-4];
419   const int A = top[0];
420   const int B = top[1];
421   const int C = top[2];
422   const int D = top[3];
423   DST(0, 0) = DST(1, 2) = AVG2(X, A);
424   DST(1, 0) = DST(2, 2) = AVG2(A, B);
425   DST(2, 0) = DST(3, 2) = AVG2(B, C);
426   DST(3, 0)             = AVG2(C, D);
427 
428   DST(0, 3) =             AVG3(K, J, I);
429   DST(0, 2) =             AVG3(J, I, X);
430   DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
431   DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
432   DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
433   DST(3, 1) =             AVG3(B, C, D);
434 }
435 
VL4(uint8_t * dst,const uint8_t * top)436 static void VL4(uint8_t* dst, const uint8_t* top) {
437   const int A = top[0];
438   const int B = top[1];
439   const int C = top[2];
440   const int D = top[3];
441   const int E = top[4];
442   const int F = top[5];
443   const int G = top[6];
444   const int H = top[7];
445   DST(0, 0) =             AVG2(A, B);
446   DST(1, 0) = DST(0, 2) = AVG2(B, C);
447   DST(2, 0) = DST(1, 2) = AVG2(C, D);
448   DST(3, 0) = DST(2, 2) = AVG2(D, E);
449 
450   DST(0, 1) =             AVG3(A, B, C);
451   DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
452   DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
453   DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
454               DST(3, 2) = AVG3(E, F, G);
455               DST(3, 3) = AVG3(F, G, H);
456 }
457 
HU4(uint8_t * dst,const uint8_t * top)458 static void HU4(uint8_t* dst, const uint8_t* top) {
459   const int I = top[-2];
460   const int J = top[-3];
461   const int K = top[-4];
462   const int L = top[-5];
463   DST(0, 0) =             AVG2(I, J);
464   DST(2, 0) = DST(0, 1) = AVG2(J, K);
465   DST(2, 1) = DST(0, 2) = AVG2(K, L);
466   DST(1, 0) =             AVG3(I, J, K);
467   DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
468   DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
469   DST(3, 2) = DST(2, 2) =
470   DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
471 }
472 
HD4(uint8_t * dst,const uint8_t * top)473 static void HD4(uint8_t* dst, const uint8_t* top) {
474   const int X = top[-1];
475   const int I = top[-2];
476   const int J = top[-3];
477   const int K = top[-4];
478   const int L = top[-5];
479   const int A = top[0];
480   const int B = top[1];
481   const int C = top[2];
482 
483   DST(0, 0) = DST(2, 1) = AVG2(I, X);
484   DST(0, 1) = DST(2, 2) = AVG2(J, I);
485   DST(0, 2) = DST(2, 3) = AVG2(K, J);
486   DST(0, 3)             = AVG2(L, K);
487 
488   DST(3, 0)             = AVG3(A, B, C);
489   DST(2, 0)             = AVG3(X, A, B);
490   DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
491   DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
492   DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
493   DST(1, 3)             = AVG3(L, K, J);
494 }
495 
TM4(uint8_t * dst,const uint8_t * top)496 static void TM4(uint8_t* dst, const uint8_t* top) {
497   int x, y;
498   const uint8_t* const clip = clip1 + 255 - top[-1];
499   for (y = 0; y < 4; ++y) {
500     const uint8_t* const clip_table = clip + top[-2 - y];
501     for (x = 0; x < 4; ++x) {
502       dst[x] = clip_table[top[x]];
503     }
504     dst += BPS;
505   }
506 }
507 
508 #undef DST
509 #undef AVG3
510 #undef AVG2
511 
512 // Left samples are top[-5 .. -2], top_left is top[-1], top are
513 // located at top[0..3], and top right is top[4..7]
Intra4Preds(uint8_t * dst,const uint8_t * top)514 static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
515   DC4(I4DC4 + dst, top);
516   TM4(I4TM4 + dst, top);
517   VE4(I4VE4 + dst, top);
518   HE4(I4HE4 + dst, top);
519   RD4(I4RD4 + dst, top);
520   VR4(I4VR4 + dst, top);
521   LD4(I4LD4 + dst, top);
522   VL4(I4VL4 + dst, top);
523   HD4(I4HD4 + dst, top);
524   HU4(I4HU4 + dst, top);
525 }
526 
527 //------------------------------------------------------------------------------
528 // Metric
529 
GetSSE(const uint8_t * a,const uint8_t * b,int w,int h)530 static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
531                               int w, int h) {
532   int count = 0;
533   int y, x;
534   for (y = 0; y < h; ++y) {
535     for (x = 0; x < w; ++x) {
536       const int diff = (int)a[x] - b[x];
537       count += diff * diff;
538     }
539     a += BPS;
540     b += BPS;
541   }
542   return count;
543 }
544 
SSE16x16(const uint8_t * a,const uint8_t * b)545 static int SSE16x16(const uint8_t* a, const uint8_t* b) {
546   return GetSSE(a, b, 16, 16);
547 }
SSE16x8(const uint8_t * a,const uint8_t * b)548 static int SSE16x8(const uint8_t* a, const uint8_t* b) {
549   return GetSSE(a, b, 16, 8);
550 }
SSE8x8(const uint8_t * a,const uint8_t * b)551 static int SSE8x8(const uint8_t* a, const uint8_t* b) {
552   return GetSSE(a, b, 8, 8);
553 }
SSE4x4(const uint8_t * a,const uint8_t * b)554 static int SSE4x4(const uint8_t* a, const uint8_t* b) {
555   return GetSSE(a, b, 4, 4);
556 }
557 
558 //------------------------------------------------------------------------------
559 // Texture distortion
560 //
561 // We try to match the spectral content (weighted) between source and
562 // reconstructed samples.
563 
564 // Hadamard transform
565 // Returns the weighted sum of the absolute value of transformed coefficients.
TTransform(const uint8_t * in,const uint16_t * w)566 static int TTransform(const uint8_t* in, const uint16_t* w) {
567   int sum = 0;
568   int tmp[16];
569   int i;
570   // horizontal pass
571   for (i = 0; i < 4; ++i, in += BPS) {
572     const int a0 = in[0] + in[2];
573     const int a1 = in[1] + in[3];
574     const int a2 = in[1] - in[3];
575     const int a3 = in[0] - in[2];
576     tmp[0 + i * 4] = a0 + a1;
577     tmp[1 + i * 4] = a3 + a2;
578     tmp[2 + i * 4] = a3 - a2;
579     tmp[3 + i * 4] = a0 - a1;
580   }
581   // vertical pass
582   for (i = 0; i < 4; ++i, ++w) {
583     const int a0 = tmp[0 + i] + tmp[8 + i];
584     const int a1 = tmp[4 + i] + tmp[12+ i];
585     const int a2 = tmp[4 + i] - tmp[12+ i];
586     const int a3 = tmp[0 + i] - tmp[8 + i];
587     const int b0 = a0 + a1;
588     const int b1 = a3 + a2;
589     const int b2 = a3 - a2;
590     const int b3 = a0 - a1;
591 
592     sum += w[ 0] * abs(b0);
593     sum += w[ 4] * abs(b1);
594     sum += w[ 8] * abs(b2);
595     sum += w[12] * abs(b3);
596   }
597   return sum;
598 }
599 
Disto4x4(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)600 static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
601                     const uint16_t* const w) {
602   const int sum1 = TTransform(a, w);
603   const int sum2 = TTransform(b, w);
604   return abs(sum2 - sum1) >> 5;
605 }
606 
Disto16x16(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)607 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
608                       const uint16_t* const w) {
609   int D = 0;
610   int x, y;
611   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
612     for (x = 0; x < 16; x += 4) {
613       D += Disto4x4(a + x + y, b + x + y, w);
614     }
615   }
616   return D;
617 }
618 
619 //------------------------------------------------------------------------------
620 // Quantization
621 //
622 
623 static const uint8_t kZigzag[16] = {
624   0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
625 };
626 
627 // Simple quantization
QuantizeBlock(int16_t in[16],int16_t out[16],int n,const VP8Matrix * const mtx)628 static int QuantizeBlock(int16_t in[16], int16_t out[16],
629                          int n, const VP8Matrix* const mtx) {
630   int last = -1;
631   for (; n < 16; ++n) {
632     const int j = kZigzag[n];
633     const int sign = (in[j] < 0);
634     int coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
635     if (coeff > 2047) coeff = 2047;
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 (sign) out[n] = -out[n];
642       in[j] = out[n] * Q;
643       if (out[n]) last = n;
644     } else {
645       out[n] = 0;
646       in[j] = 0;
647     }
648   }
649   return (last >= 0);
650 }
651 
652 //------------------------------------------------------------------------------
653 // Block copy
654 
Copy(const uint8_t * src,uint8_t * dst,int size)655 static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) {
656   int y;
657   for (y = 0; y < size; ++y) {
658     memcpy(dst, src, size);
659     src += BPS;
660     dst += BPS;
661   }
662 }
663 
Copy4x4(const uint8_t * src,uint8_t * dst)664 static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
665 
666 //------------------------------------------------------------------------------
667 // Initialization
668 
669 // Speed-critical function pointers. We have to initialize them to the default
670 // implementations within VP8EncDspInit().
671 VP8CHisto VP8CollectHistogram;
672 VP8Idct VP8ITransform;
673 VP8Fdct VP8FTransform;
674 VP8WHT VP8ITransformWHT;
675 VP8WHT VP8FTransformWHT;
676 VP8Intra4Preds VP8EncPredLuma4;
677 VP8IntraPreds VP8EncPredLuma16;
678 VP8IntraPreds VP8EncPredChroma8;
679 VP8Metric VP8SSE16x16;
680 VP8Metric VP8SSE8x8;
681 VP8Metric VP8SSE16x8;
682 VP8Metric VP8SSE4x4;
683 VP8WMetric VP8TDisto4x4;
684 VP8WMetric VP8TDisto16x16;
685 VP8QuantizeBlock VP8EncQuantizeBlock;
686 VP8BlockCopy VP8Copy4x4;
687 
688 extern void VP8EncDspInitSSE2(void);
689 extern void VP8EncDspInitNEON(void);
690 
VP8EncDspInit(void)691 void VP8EncDspInit(void) {
692   InitTables();
693 
694   // default C implementations
695   VP8CollectHistogram = CollectHistogram;
696   VP8ITransform = ITransform;
697   VP8FTransform = FTransform;
698   VP8ITransformWHT = ITransformWHT;
699   VP8FTransformWHT = FTransformWHT;
700   VP8EncPredLuma4 = Intra4Preds;
701   VP8EncPredLuma16 = Intra16Preds;
702   VP8EncPredChroma8 = IntraChromaPreds;
703   VP8SSE16x16 = SSE16x16;
704   VP8SSE8x8 = SSE8x8;
705   VP8SSE16x8 = SSE16x8;
706   VP8SSE4x4 = SSE4x4;
707   VP8TDisto4x4 = Disto4x4;
708   VP8TDisto16x16 = Disto16x16;
709   VP8EncQuantizeBlock = QuantizeBlock;
710   VP8Copy4x4 = Copy4x4;
711 
712   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
713   if (VP8GetCPUInfo) {
714 #if defined(WEBP_USE_SSE2)
715     if (VP8GetCPUInfo(kSSE2)) {
716       VP8EncDspInitSSE2();
717     }
718 #elif defined(WEBP_USE_NEON)
719     if (VP8GetCPUInfo(kNEON)) {
720       VP8EncDspInitNEON();
721     }
722 #endif
723   }
724 }
725 
726 #if defined(__cplusplus) || defined(c_plusplus)
727 }    // extern "C"
728 #endif
729