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 
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) * 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 
FTransformWHT(const int16_t * in,int16_t * out)162 static void FTransformWHT(const int16_t* in, int16_t* out) {
163   // input is 12b signed
164   int32_t tmp[16];
165   int i;
166   for (i = 0; i < 4; ++i, in += 64) {
167     const int a0 = (in[0 * 16] + in[2 * 16]);  // 13b
168     const int a1 = (in[1 * 16] + in[3 * 16]);
169     const int a2 = (in[1 * 16] - in[3 * 16]);
170     const int a3 = (in[0 * 16] - in[2 * 16]);
171     tmp[0 + i * 4] = a0 + a1;   // 14b
172     tmp[1 + i * 4] = a3 + a2;
173     tmp[2 + i * 4] = a3 - a2;
174     tmp[3 + i * 4] = a0 - a1;
175   }
176   for (i = 0; i < 4; ++i) {
177     const int a0 = (tmp[0 + i] + tmp[8 + i]);  // 15b
178     const int a1 = (tmp[4 + i] + tmp[12+ i]);
179     const int a2 = (tmp[4 + i] - tmp[12+ i]);
180     const int a3 = (tmp[0 + i] - tmp[8 + i]);
181     const int b0 = a0 + a1;    // 16b
182     const int b1 = a3 + a2;
183     const int b2 = a3 - a2;
184     const int b3 = a0 - a1;
185     out[ 0 + i] = b0 >> 1;     // 15b
186     out[ 4 + i] = b1 >> 1;
187     out[ 8 + i] = b2 >> 1;
188     out[12 + i] = b3 >> 1;
189   }
190 }
191 
192 #undef MUL
193 #undef STORE
194 
195 //------------------------------------------------------------------------------
196 // Intra predictions
197 
198 #define DST(x, y) dst[(x) + (y) * BPS]
199 
Fill(uint8_t * dst,int value,int size)200 static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
201   int j;
202   for (j = 0; j < size; ++j) {
203     memset(dst + j * BPS, value, size);
204   }
205 }
206 
VerticalPred(uint8_t * dst,const uint8_t * top,int size)207 static WEBP_INLINE void VerticalPred(uint8_t* dst,
208                                      const uint8_t* top, int size) {
209   int j;
210   if (top) {
211     for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
212   } else {
213     Fill(dst, 127, size);
214   }
215 }
216 
HorizontalPred(uint8_t * dst,const uint8_t * left,int size)217 static WEBP_INLINE void HorizontalPred(uint8_t* dst,
218                                        const uint8_t* left, int size) {
219   if (left) {
220     int j;
221     for (j = 0; j < size; ++j) {
222       memset(dst + j * BPS, left[j], size);
223     }
224   } else {
225     Fill(dst, 129, size);
226   }
227 }
228 
TrueMotion(uint8_t * dst,const uint8_t * left,const uint8_t * top,int size)229 static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
230                                    const uint8_t* top, int size) {
231   int y;
232   if (left) {
233     if (top) {
234       const uint8_t* const clip = clip1 + 255 - left[-1];
235       for (y = 0; y < size; ++y) {
236         const uint8_t* const clip_table = clip + left[y];
237         int x;
238         for (x = 0; x < size; ++x) {
239           dst[x] = clip_table[top[x]];
240         }
241         dst += BPS;
242       }
243     } else {
244       HorizontalPred(dst, left, size);
245     }
246   } else {
247     // true motion without left samples (hence: with default 129 value)
248     // is equivalent to VE prediction where you just copy the top samples.
249     // Note that if top samples are not available, the default value is
250     // then 129, and not 127 as in the VerticalPred case.
251     if (top) {
252       VerticalPred(dst, top, size);
253     } else {
254       Fill(dst, 129, size);
255     }
256   }
257 }
258 
DCMode(uint8_t * dst,const uint8_t * left,const uint8_t * top,int size,int round,int shift)259 static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
260                                const uint8_t* top,
261                                int size, int round, int shift) {
262   int DC = 0;
263   int j;
264   if (top) {
265     for (j = 0; j < size; ++j) DC += top[j];
266     if (left) {   // top and left present
267       for (j = 0; j < size; ++j) DC += left[j];
268     } else {      // top, but no left
269       DC += DC;
270     }
271     DC = (DC + round) >> shift;
272   } else if (left) {   // left but no top
273     for (j = 0; j < size; ++j) DC += left[j];
274     DC += DC;
275     DC = (DC + round) >> shift;
276   } else {   // no top, no left, nothing.
277     DC = 0x80;
278   }
279   Fill(dst, DC, size);
280 }
281 
282 //------------------------------------------------------------------------------
283 // Chroma 8x8 prediction (paragraph 12.2)
284 
IntraChromaPreds(uint8_t * dst,const uint8_t * left,const uint8_t * top)285 static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
286                              const uint8_t* top) {
287   // U block
288   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
289   VerticalPred(C8VE8 + dst, top, 8);
290   HorizontalPred(C8HE8 + dst, left, 8);
291   TrueMotion(C8TM8 + dst, left, top, 8);
292   // V block
293   dst += 8;
294   if (top) top += 8;
295   if (left) left += 16;
296   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
297   VerticalPred(C8VE8 + dst, top, 8);
298   HorizontalPred(C8HE8 + dst, left, 8);
299   TrueMotion(C8TM8 + dst, left, top, 8);
300 }
301 
302 //------------------------------------------------------------------------------
303 // luma 16x16 prediction (paragraph 12.3)
304 
Intra16Preds(uint8_t * dst,const uint8_t * left,const uint8_t * top)305 static void Intra16Preds(uint8_t* dst,
306                          const uint8_t* left, const uint8_t* top) {
307   DCMode(I16DC16 + dst, left, top, 16, 16, 5);
308   VerticalPred(I16VE16 + dst, top, 16);
309   HorizontalPred(I16HE16 + dst, left, 16);
310   TrueMotion(I16TM16 + dst, left, top, 16);
311 }
312 
313 //------------------------------------------------------------------------------
314 // luma 4x4 prediction
315 
316 #define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
317 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
318 
VE4(uint8_t * dst,const uint8_t * top)319 static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
320   const uint8_t vals[4] = {
321     AVG3(top[-1], top[0], top[1]),
322     AVG3(top[ 0], top[1], top[2]),
323     AVG3(top[ 1], top[2], top[3]),
324     AVG3(top[ 2], top[3], top[4])
325   };
326   int i;
327   for (i = 0; i < 4; ++i) {
328     memcpy(dst + i * BPS, vals, 4);
329   }
330 }
331 
HE4(uint8_t * dst,const uint8_t * top)332 static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
333   const int X = top[-1];
334   const int I = top[-2];
335   const int J = top[-3];
336   const int K = top[-4];
337   const int L = top[-5];
338   *(uint32_t*)(dst + 0 * BPS) = 0x01010101U * AVG3(X, I, J);
339   *(uint32_t*)(dst + 1 * BPS) = 0x01010101U * AVG3(I, J, K);
340   *(uint32_t*)(dst + 2 * BPS) = 0x01010101U * AVG3(J, K, L);
341   *(uint32_t*)(dst + 3 * BPS) = 0x01010101U * AVG3(K, L, L);
342 }
343 
DC4(uint8_t * dst,const uint8_t * top)344 static void DC4(uint8_t* dst, const uint8_t* top) {
345   uint32_t dc = 4;
346   int i;
347   for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
348   Fill(dst, dc >> 3, 4);
349 }
350 
RD4(uint8_t * dst,const uint8_t * top)351 static void RD4(uint8_t* dst, const uint8_t* top) {
352   const int X = top[-1];
353   const int I = top[-2];
354   const int J = top[-3];
355   const int K = top[-4];
356   const int L = top[-5];
357   const int A = top[0];
358   const int B = top[1];
359   const int C = top[2];
360   const int D = top[3];
361   DST(0, 3)                                     = AVG3(J, K, L);
362   DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
363   DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
364   DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
365   DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
366   DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
367   DST(3, 0)                                     = AVG3(D, C, B);
368 }
369 
LD4(uint8_t * dst,const uint8_t * top)370 static void LD4(uint8_t* dst, const uint8_t* top) {
371   const int A = top[0];
372   const int B = top[1];
373   const int C = top[2];
374   const int D = top[3];
375   const int E = top[4];
376   const int F = top[5];
377   const int G = top[6];
378   const int H = top[7];
379   DST(0, 0)                                     = AVG3(A, B, C);
380   DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
381   DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
382   DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
383   DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
384   DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
385   DST(3, 3)                                     = AVG3(G, H, H);
386 }
387 
VR4(uint8_t * dst,const uint8_t * top)388 static void VR4(uint8_t* dst, const uint8_t* top) {
389   const int X = top[-1];
390   const int I = top[-2];
391   const int J = top[-3];
392   const int K = top[-4];
393   const int A = top[0];
394   const int B = top[1];
395   const int C = top[2];
396   const int D = top[3];
397   DST(0, 0) = DST(1, 2) = AVG2(X, A);
398   DST(1, 0) = DST(2, 2) = AVG2(A, B);
399   DST(2, 0) = DST(3, 2) = AVG2(B, C);
400   DST(3, 0)             = AVG2(C, D);
401 
402   DST(0, 3) =             AVG3(K, J, I);
403   DST(0, 2) =             AVG3(J, I, X);
404   DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
405   DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
406   DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
407   DST(3, 1) =             AVG3(B, C, D);
408 }
409 
VL4(uint8_t * dst,const uint8_t * top)410 static void VL4(uint8_t* dst, const uint8_t* top) {
411   const int A = top[0];
412   const int B = top[1];
413   const int C = top[2];
414   const int D = top[3];
415   const int E = top[4];
416   const int F = top[5];
417   const int G = top[6];
418   const int H = top[7];
419   DST(0, 0) =             AVG2(A, B);
420   DST(1, 0) = DST(0, 2) = AVG2(B, C);
421   DST(2, 0) = DST(1, 2) = AVG2(C, D);
422   DST(3, 0) = DST(2, 2) = AVG2(D, E);
423 
424   DST(0, 1) =             AVG3(A, B, C);
425   DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
426   DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
427   DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
428               DST(3, 2) = AVG3(E, F, G);
429               DST(3, 3) = AVG3(F, G, H);
430 }
431 
HU4(uint8_t * dst,const uint8_t * top)432 static void HU4(uint8_t* dst, const uint8_t* top) {
433   const int I = top[-2];
434   const int J = top[-3];
435   const int K = top[-4];
436   const int L = top[-5];
437   DST(0, 0) =             AVG2(I, J);
438   DST(2, 0) = DST(0, 1) = AVG2(J, K);
439   DST(2, 1) = DST(0, 2) = AVG2(K, L);
440   DST(1, 0) =             AVG3(I, J, K);
441   DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
442   DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
443   DST(3, 2) = DST(2, 2) =
444   DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
445 }
446 
HD4(uint8_t * dst,const uint8_t * top)447 static void HD4(uint8_t* dst, const uint8_t* top) {
448   const int X = top[-1];
449   const int I = top[-2];
450   const int J = top[-3];
451   const int K = top[-4];
452   const int L = top[-5];
453   const int A = top[0];
454   const int B = top[1];
455   const int C = top[2];
456 
457   DST(0, 0) = DST(2, 1) = AVG2(I, X);
458   DST(0, 1) = DST(2, 2) = AVG2(J, I);
459   DST(0, 2) = DST(2, 3) = AVG2(K, J);
460   DST(0, 3)             = AVG2(L, K);
461 
462   DST(3, 0)             = AVG3(A, B, C);
463   DST(2, 0)             = AVG3(X, A, B);
464   DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
465   DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
466   DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
467   DST(1, 3)             = AVG3(L, K, J);
468 }
469 
TM4(uint8_t * dst,const uint8_t * top)470 static void TM4(uint8_t* dst, const uint8_t* top) {
471   int x, y;
472   const uint8_t* const clip = clip1 + 255 - top[-1];
473   for (y = 0; y < 4; ++y) {
474     const uint8_t* const clip_table = clip + top[-2 - y];
475     for (x = 0; x < 4; ++x) {
476       dst[x] = clip_table[top[x]];
477     }
478     dst += BPS;
479   }
480 }
481 
482 #undef DST
483 #undef AVG3
484 #undef AVG2
485 
486 // Left samples are top[-5 .. -2], top_left is top[-1], top are
487 // located at top[0..3], and top right is top[4..7]
Intra4Preds(uint8_t * dst,const uint8_t * top)488 static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
489   DC4(I4DC4 + dst, top);
490   TM4(I4TM4 + dst, top);
491   VE4(I4VE4 + dst, top);
492   HE4(I4HE4 + dst, top);
493   RD4(I4RD4 + dst, top);
494   VR4(I4VR4 + dst, top);
495   LD4(I4LD4 + dst, top);
496   VL4(I4VL4 + dst, top);
497   HD4(I4HD4 + dst, top);
498   HU4(I4HU4 + dst, top);
499 }
500 
501 //------------------------------------------------------------------------------
502 // Metric
503 
GetSSE(const uint8_t * a,const uint8_t * b,int w,int h)504 static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
505                               int w, int h) {
506   int count = 0;
507   int y, x;
508   for (y = 0; y < h; ++y) {
509     for (x = 0; x < w; ++x) {
510       const int diff = (int)a[x] - b[x];
511       count += diff * diff;
512     }
513     a += BPS;
514     b += BPS;
515   }
516   return count;
517 }
518 
SSE16x16(const uint8_t * a,const uint8_t * b)519 static int SSE16x16(const uint8_t* a, const uint8_t* b) {
520   return GetSSE(a, b, 16, 16);
521 }
SSE16x8(const uint8_t * a,const uint8_t * b)522 static int SSE16x8(const uint8_t* a, const uint8_t* b) {
523   return GetSSE(a, b, 16, 8);
524 }
SSE8x8(const uint8_t * a,const uint8_t * b)525 static int SSE8x8(const uint8_t* a, const uint8_t* b) {
526   return GetSSE(a, b, 8, 8);
527 }
SSE4x4(const uint8_t * a,const uint8_t * b)528 static int SSE4x4(const uint8_t* a, const uint8_t* b) {
529   return GetSSE(a, b, 4, 4);
530 }
531 
532 //------------------------------------------------------------------------------
533 // Texture distortion
534 //
535 // We try to match the spectral content (weighted) between source and
536 // reconstructed samples.
537 
538 // Hadamard transform
539 // Returns the weighted sum of the absolute value of transformed coefficients.
TTransform(const uint8_t * in,const uint16_t * w)540 static int TTransform(const uint8_t* in, const uint16_t* w) {
541   int sum = 0;
542   int tmp[16];
543   int i;
544   // horizontal pass
545   for (i = 0; i < 4; ++i, in += BPS) {
546     const int a0 = in[0] + in[2];
547     const int a1 = in[1] + in[3];
548     const int a2 = in[1] - in[3];
549     const int a3 = in[0] - in[2];
550     tmp[0 + i * 4] = a0 + a1;
551     tmp[1 + i * 4] = a3 + a2;
552     tmp[2 + i * 4] = a3 - a2;
553     tmp[3 + i * 4] = a0 - a1;
554   }
555   // vertical pass
556   for (i = 0; i < 4; ++i, ++w) {
557     const int a0 = tmp[0 + i] + tmp[8 + i];
558     const int a1 = tmp[4 + i] + tmp[12+ i];
559     const int a2 = tmp[4 + i] - tmp[12+ i];
560     const int a3 = tmp[0 + i] - tmp[8 + i];
561     const int b0 = a0 + a1;
562     const int b1 = a3 + a2;
563     const int b2 = a3 - a2;
564     const int b3 = a0 - a1;
565 
566     sum += w[ 0] * abs(b0);
567     sum += w[ 4] * abs(b1);
568     sum += w[ 8] * abs(b2);
569     sum += w[12] * abs(b3);
570   }
571   return sum;
572 }
573 
Disto4x4(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)574 static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
575                     const uint16_t* const w) {
576   const int sum1 = TTransform(a, w);
577   const int sum2 = TTransform(b, w);
578   return abs(sum2 - sum1) >> 5;
579 }
580 
Disto16x16(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)581 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
582                       const uint16_t* const w) {
583   int D = 0;
584   int x, y;
585   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
586     for (x = 0; x < 16; x += 4) {
587       D += Disto4x4(a + x + y, b + x + y, w);
588     }
589   }
590   return D;
591 }
592 
593 //------------------------------------------------------------------------------
594 // Quantization
595 //
596 
597 static const uint8_t kZigzag[16] = {
598   0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
599 };
600 
601 // Simple quantization
QuantizeBlock(int16_t in[16],int16_t out[16],const VP8Matrix * const mtx)602 static int QuantizeBlock(int16_t in[16], int16_t out[16],
603                          const VP8Matrix* const mtx) {
604   int last = -1;
605   int n;
606   for (n = 0; n < 16; ++n) {
607     const int j = kZigzag[n];
608     const int sign = (in[j] < 0);
609     const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
610     if (coeff > mtx->zthresh_[j]) {
611       const uint32_t Q = mtx->q_[j];
612       const uint32_t iQ = mtx->iq_[j];
613       const uint32_t B = mtx->bias_[j];
614       int level = QUANTDIV(coeff, iQ, B);
615       if (level > MAX_LEVEL) level = MAX_LEVEL;
616       if (sign) level = -level;
617       in[j] = level * Q;
618       out[n] = level;
619       if (level) last = n;
620     } else {
621       out[n] = 0;
622       in[j] = 0;
623     }
624   }
625   return (last >= 0);
626 }
627 
QuantizeBlockWHT(int16_t in[16],int16_t out[16],const VP8Matrix * const mtx)628 static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
629                             const VP8Matrix* const mtx) {
630   int n, last = -1;
631   for (n = 0; n < 16; ++n) {
632     const int j = kZigzag[n];
633     const int sign = (in[j] < 0);
634     const uint32_t coeff = sign ? -in[j] : in[j];
635     assert(mtx->sharpen_[j] == 0);
636     if (coeff > mtx->zthresh_[j]) {
637       const uint32_t Q = mtx->q_[j];
638       const uint32_t iQ = mtx->iq_[j];
639       const uint32_t B = mtx->bias_[j];
640       int level = QUANTDIV(coeff, iQ, B);
641       if (level > MAX_LEVEL) level = MAX_LEVEL;
642       if (sign) level = -level;
643       in[j] = level * Q;
644       out[n] = level;
645       if (level) last = n;
646     } else {
647       out[n] = 0;
648       in[j] = 0;
649     }
650   }
651   return (last >= 0);
652 }
653 
654 //------------------------------------------------------------------------------
655 // Block copy
656 
Copy(const uint8_t * src,uint8_t * dst,int size)657 static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int size) {
658   int y;
659   for (y = 0; y < size; ++y) {
660     memcpy(dst, src, size);
661     src += BPS;
662     dst += BPS;
663   }
664 }
665 
Copy4x4(const uint8_t * src,uint8_t * dst)666 static void Copy4x4(const uint8_t* src, uint8_t* dst) { Copy(src, dst, 4); }
667 
668 //------------------------------------------------------------------------------
669 // Initialization
670 
671 // Speed-critical function pointers. We have to initialize them to the default
672 // implementations within VP8EncDspInit().
673 VP8CHisto VP8CollectHistogram;
674 VP8Idct VP8ITransform;
675 VP8Fdct VP8FTransform;
676 VP8WHT VP8FTransformWHT;
677 VP8Intra4Preds VP8EncPredLuma4;
678 VP8IntraPreds VP8EncPredLuma16;
679 VP8IntraPreds VP8EncPredChroma8;
680 VP8Metric VP8SSE16x16;
681 VP8Metric VP8SSE8x8;
682 VP8Metric VP8SSE16x8;
683 VP8Metric VP8SSE4x4;
684 VP8WMetric VP8TDisto4x4;
685 VP8WMetric VP8TDisto16x16;
686 VP8QuantizeBlock VP8EncQuantizeBlock;
687 VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
688 VP8BlockCopy VP8Copy4x4;
689 
690 extern void VP8EncDspInitSSE2(void);
691 extern void VP8EncDspInitAVX2(void);
692 extern void VP8EncDspInitNEON(void);
693 extern void VP8EncDspInitMIPS32(void);
694 
695 static volatile VP8CPUInfo enc_last_cpuinfo_used =
696     (VP8CPUInfo)&enc_last_cpuinfo_used;
697 
VP8EncDspInit(void)698 void VP8EncDspInit(void) {
699   if (enc_last_cpuinfo_used == VP8GetCPUInfo) return;
700 
701   VP8DspInit();  // common inverse transforms
702   InitTables();
703 
704   // default C implementations
705   VP8CollectHistogram = CollectHistogram;
706   VP8ITransform = ITransform;
707   VP8FTransform = FTransform;
708   VP8FTransformWHT = FTransformWHT;
709   VP8EncPredLuma4 = Intra4Preds;
710   VP8EncPredLuma16 = Intra16Preds;
711   VP8EncPredChroma8 = IntraChromaPreds;
712   VP8SSE16x16 = SSE16x16;
713   VP8SSE8x8 = SSE8x8;
714   VP8SSE16x8 = SSE16x8;
715   VP8SSE4x4 = SSE4x4;
716   VP8TDisto4x4 = Disto4x4;
717   VP8TDisto16x16 = Disto16x16;
718   VP8EncQuantizeBlock = QuantizeBlock;
719   VP8EncQuantizeBlockWHT = QuantizeBlockWHT;
720   VP8Copy4x4 = Copy4x4;
721 
722   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
723   if (VP8GetCPUInfo != NULL) {
724 #if defined(WEBP_USE_SSE2)
725     if (VP8GetCPUInfo(kSSE2)) {
726       VP8EncDspInitSSE2();
727     }
728 #endif
729 #if defined(WEBP_USE_AVX2)
730     if (VP8GetCPUInfo(kAVX2)) {
731       VP8EncDspInitAVX2();
732     }
733 #endif
734 #if defined(WEBP_USE_NEON)
735     if (VP8GetCPUInfo(kNEON)) {
736       VP8EncDspInitNEON();
737     }
738 #endif
739 #if defined(WEBP_USE_MIPS32)
740     if (VP8GetCPUInfo(kMIPS32)) {
741       VP8EncDspInitMIPS32();
742     }
743 #endif
744   }
745   enc_last_cpuinfo_used = VP8GetCPUInfo;
746 }
747 
748