• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/vp8i_enc.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 
43 // general-purpose util function
VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH+1],VP8Histogram * const histo)44 void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1],
45                          VP8Histogram* const histo) {
46   int max_value = 0, last_non_zero = 1;
47   int k;
48   for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
49     const int value = distribution[k];
50     if (value > 0) {
51       if (value > max_value) max_value = value;
52       last_non_zero = k;
53     }
54   }
55   histo->max_value = max_value;
56   histo->last_non_zero = last_non_zero;
57 }
58 
CollectHistogram(const uint8_t * ref,const uint8_t * pred,int start_block,int end_block,VP8Histogram * const histo)59 static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
60                              int start_block, int end_block,
61                              VP8Histogram* const histo) {
62   int j;
63   int distribution[MAX_COEFF_THRESH + 1] = { 0 };
64   for (j = start_block; j < end_block; ++j) {
65     int k;
66     int16_t out[16];
67 
68     VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
69 
70     // Convert coefficients to bin.
71     for (k = 0; k < 16; ++k) {
72       const int v = abs(out[k]) >> 3;
73       const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
74       ++distribution[clipped_value];
75     }
76   }
77   VP8SetHistogramData(distribution, histo);
78 }
79 
80 //------------------------------------------------------------------------------
81 // run-time tables (~4k)
82 
83 static uint8_t clip1[255 + 510 + 1];    // clips [-255,510] to [0,255]
84 
85 // We declare this variable 'volatile' to prevent instruction reordering
86 // and make sure it's set to true _last_ (so as to be thread-safe)
87 static volatile int tables_ok = 0;
88 
InitTables(void)89 static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) {
90   if (!tables_ok) {
91     int i;
92     for (i = -255; i <= 255 + 255; ++i) {
93       clip1[255 + i] = clip_8b(i);
94     }
95     tables_ok = 1;
96   }
97 }
98 
99 
100 //------------------------------------------------------------------------------
101 // Transforms (Paragraph 14.4)
102 
103 #define STORE(x, y, v) \
104   dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3))
105 
106 static const int kC1 = 20091 + (1 << 16);
107 static const int kC2 = 35468;
108 #define MUL(a, b) (((a) * (b)) >> 16)
109 
ITransformOne(const uint8_t * ref,const int16_t * in,uint8_t * dst)110 static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
111                                       uint8_t* dst) {
112   int C[4 * 4], *tmp;
113   int i;
114   tmp = C;
115   for (i = 0; i < 4; ++i) {    // vertical pass
116     const int a = in[0] + in[8];
117     const int b = in[0] - in[8];
118     const int c = MUL(in[4], kC2) - MUL(in[12], kC1);
119     const int d = MUL(in[4], kC1) + MUL(in[12], kC2);
120     tmp[0] = a + d;
121     tmp[1] = b + c;
122     tmp[2] = b - c;
123     tmp[3] = a - d;
124     tmp += 4;
125     in++;
126   }
127 
128   tmp = C;
129   for (i = 0; i < 4; ++i) {    // horizontal pass
130     const int dc = tmp[0] + 4;
131     const int a =  dc +  tmp[8];
132     const int b =  dc -  tmp[8];
133     const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1);
134     const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2);
135     STORE(0, i, a + d);
136     STORE(1, i, b + c);
137     STORE(2, i, b - c);
138     STORE(3, i, a - d);
139     tmp++;
140   }
141 }
142 
ITransform(const uint8_t * ref,const int16_t * in,uint8_t * dst,int do_two)143 static void ITransform(const uint8_t* ref, const int16_t* in, uint8_t* dst,
144                        int do_two) {
145   ITransformOne(ref, in, dst);
146   if (do_two) {
147     ITransformOne(ref + 4, in + 16, dst + 4);
148   }
149 }
150 
FTransform(const uint8_t * src,const uint8_t * ref,int16_t * out)151 static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
152   int i;
153   int tmp[16];
154   for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
155     const int d0 = src[0] - ref[0];   // 9bit dynamic range ([-255,255])
156     const int d1 = src[1] - ref[1];
157     const int d2 = src[2] - ref[2];
158     const int d3 = src[3] - ref[3];
159     const int a0 = (d0 + d3);         // 10b                      [-510,510]
160     const int a1 = (d1 + d2);
161     const int a2 = (d1 - d2);
162     const int a3 = (d0 - d3);
163     tmp[0 + i * 4] = (a0 + a1) * 8;   // 14b                      [-8160,8160]
164     tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
165     tmp[2 + i * 4] = (a0 - a1) * 8;
166     tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  937) >> 9;
167   }
168   for (i = 0; i < 4; ++i) {
169     const int a0 = (tmp[0 + i] + tmp[12 + i]);  // 15b
170     const int a1 = (tmp[4 + i] + tmp[ 8 + i]);
171     const int a2 = (tmp[4 + i] - tmp[ 8 + i]);
172     const int a3 = (tmp[0 + i] - tmp[12 + i]);
173     out[0 + i] = (a0 + a1 + 7) >> 4;            // 12b
174     out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0);
175     out[8 + i] = (a0 - a1 + 7) >> 4;
176     out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16);
177   }
178 }
179 
FTransform2(const uint8_t * src,const uint8_t * ref,int16_t * out)180 static void FTransform2(const uint8_t* src, const uint8_t* ref, int16_t* out) {
181   VP8FTransform(src, ref, out);
182   VP8FTransform(src + 4, ref + 4, out + 16);
183 }
184 
FTransformWHT(const int16_t * in,int16_t * out)185 static void FTransformWHT(const int16_t* in, int16_t* out) {
186   // input is 12b signed
187   int32_t tmp[16];
188   int i;
189   for (i = 0; i < 4; ++i, in += 64) {
190     const int a0 = (in[0 * 16] + in[2 * 16]);  // 13b
191     const int a1 = (in[1 * 16] + in[3 * 16]);
192     const int a2 = (in[1 * 16] - in[3 * 16]);
193     const int a3 = (in[0 * 16] - in[2 * 16]);
194     tmp[0 + i * 4] = a0 + a1;   // 14b
195     tmp[1 + i * 4] = a3 + a2;
196     tmp[2 + i * 4] = a3 - a2;
197     tmp[3 + i * 4] = a0 - a1;
198   }
199   for (i = 0; i < 4; ++i) {
200     const int a0 = (tmp[0 + i] + tmp[8 + i]);  // 15b
201     const int a1 = (tmp[4 + i] + tmp[12+ i]);
202     const int a2 = (tmp[4 + i] - tmp[12+ i]);
203     const int a3 = (tmp[0 + i] - tmp[8 + i]);
204     const int b0 = a0 + a1;    // 16b
205     const int b1 = a3 + a2;
206     const int b2 = a3 - a2;
207     const int b3 = a0 - a1;
208     out[ 0 + i] = b0 >> 1;     // 15b
209     out[ 4 + i] = b1 >> 1;
210     out[ 8 + i] = b2 >> 1;
211     out[12 + i] = b3 >> 1;
212   }
213 }
214 
215 #undef MUL
216 #undef STORE
217 
218 //------------------------------------------------------------------------------
219 // Intra predictions
220 
Fill(uint8_t * dst,int value,int size)221 static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) {
222   int j;
223   for (j = 0; j < size; ++j) {
224     memset(dst + j * BPS, value, size);
225   }
226 }
227 
VerticalPred(uint8_t * dst,const uint8_t * top,int size)228 static WEBP_INLINE void VerticalPred(uint8_t* dst,
229                                      const uint8_t* top, int size) {
230   int j;
231   if (top != NULL) {
232     for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size);
233   } else {
234     Fill(dst, 127, size);
235   }
236 }
237 
HorizontalPred(uint8_t * dst,const uint8_t * left,int size)238 static WEBP_INLINE void HorizontalPred(uint8_t* dst,
239                                        const uint8_t* left, int size) {
240   if (left != NULL) {
241     int j;
242     for (j = 0; j < size; ++j) {
243       memset(dst + j * BPS, left[j], size);
244     }
245   } else {
246     Fill(dst, 129, size);
247   }
248 }
249 
TrueMotion(uint8_t * dst,const uint8_t * left,const uint8_t * top,int size)250 static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left,
251                                    const uint8_t* top, int size) {
252   int y;
253   if (left != NULL) {
254     if (top != NULL) {
255       const uint8_t* const clip = clip1 + 255 - left[-1];
256       for (y = 0; y < size; ++y) {
257         const uint8_t* const clip_table = clip + left[y];
258         int x;
259         for (x = 0; x < size; ++x) {
260           dst[x] = clip_table[top[x]];
261         }
262         dst += BPS;
263       }
264     } else {
265       HorizontalPred(dst, left, size);
266     }
267   } else {
268     // true motion without left samples (hence: with default 129 value)
269     // is equivalent to VE prediction where you just copy the top samples.
270     // Note that if top samples are not available, the default value is
271     // then 129, and not 127 as in the VerticalPred case.
272     if (top != NULL) {
273       VerticalPred(dst, top, size);
274     } else {
275       Fill(dst, 129, size);
276     }
277   }
278 }
279 
DCMode(uint8_t * dst,const uint8_t * left,const uint8_t * top,int size,int round,int shift)280 static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left,
281                                const uint8_t* top,
282                                int size, int round, int shift) {
283   int DC = 0;
284   int j;
285   if (top != NULL) {
286     for (j = 0; j < size; ++j) DC += top[j];
287     if (left != NULL) {   // top and left present
288       for (j = 0; j < size; ++j) DC += left[j];
289     } else {      // top, but no left
290       DC += DC;
291     }
292     DC = (DC + round) >> shift;
293   } else if (left != NULL) {   // left but no top
294     for (j = 0; j < size; ++j) DC += left[j];
295     DC += DC;
296     DC = (DC + round) >> shift;
297   } else {   // no top, no left, nothing.
298     DC = 0x80;
299   }
300   Fill(dst, DC, size);
301 }
302 
303 //------------------------------------------------------------------------------
304 // Chroma 8x8 prediction (paragraph 12.2)
305 
IntraChromaPreds(uint8_t * dst,const uint8_t * left,const uint8_t * top)306 static void IntraChromaPreds(uint8_t* dst, const uint8_t* left,
307                              const uint8_t* top) {
308   // U block
309   DCMode(C8DC8 + dst, left, top, 8, 8, 4);
310   VerticalPred(C8VE8 + dst, top, 8);
311   HorizontalPred(C8HE8 + dst, left, 8);
312   TrueMotion(C8TM8 + dst, left, top, 8);
313   // V block
314   dst += 8;
315   if (top != NULL) top += 8;
316   if (left != NULL) left += 16;
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 }
322 
323 //------------------------------------------------------------------------------
324 // luma 16x16 prediction (paragraph 12.3)
325 
Intra16Preds(uint8_t * dst,const uint8_t * left,const uint8_t * top)326 static void Intra16Preds(uint8_t* dst,
327                          const uint8_t* left, const uint8_t* top) {
328   DCMode(I16DC16 + dst, left, top, 16, 16, 5);
329   VerticalPred(I16VE16 + dst, top, 16);
330   HorizontalPred(I16HE16 + dst, left, 16);
331   TrueMotion(I16TM16 + dst, left, top, 16);
332 }
333 
334 //------------------------------------------------------------------------------
335 // luma 4x4 prediction
336 
337 #define DST(x, y) dst[(x) + (y) * BPS]
338 #define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2))
339 #define AVG2(a, b) (((a) + (b) + 1) >> 1)
340 
VE4(uint8_t * dst,const uint8_t * top)341 static void VE4(uint8_t* dst, const uint8_t* top) {    // vertical
342   const uint8_t vals[4] = {
343     AVG3(top[-1], top[0], top[1]),
344     AVG3(top[ 0], top[1], top[2]),
345     AVG3(top[ 1], top[2], top[3]),
346     AVG3(top[ 2], top[3], top[4])
347   };
348   int i;
349   for (i = 0; i < 4; ++i) {
350     memcpy(dst + i * BPS, vals, 4);
351   }
352 }
353 
HE4(uint8_t * dst,const uint8_t * top)354 static void HE4(uint8_t* dst, const uint8_t* top) {    // horizontal
355   const int X = top[-1];
356   const int I = top[-2];
357   const int J = top[-3];
358   const int K = top[-4];
359   const int L = top[-5];
360   WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J));
361   WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K));
362   WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L));
363   WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L));
364 }
365 
DC4(uint8_t * dst,const uint8_t * top)366 static void DC4(uint8_t* dst, const uint8_t* top) {
367   uint32_t dc = 4;
368   int i;
369   for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i];
370   Fill(dst, dc >> 3, 4);
371 }
372 
RD4(uint8_t * dst,const uint8_t * top)373 static void RD4(uint8_t* dst, const uint8_t* top) {
374   const int X = top[-1];
375   const int I = top[-2];
376   const int J = top[-3];
377   const int K = top[-4];
378   const int L = top[-5];
379   const int A = top[0];
380   const int B = top[1];
381   const int C = top[2];
382   const int D = top[3];
383   DST(0, 3)                                     = AVG3(J, K, L);
384   DST(0, 2) = DST(1, 3)                         = AVG3(I, J, K);
385   DST(0, 1) = DST(1, 2) = DST(2, 3)             = AVG3(X, I, J);
386   DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I);
387   DST(1, 0) = DST(2, 1) = DST(3, 2)             = AVG3(B, A, X);
388   DST(2, 0) = DST(3, 1)                         = AVG3(C, B, A);
389   DST(3, 0)                                     = AVG3(D, C, B);
390 }
391 
LD4(uint8_t * dst,const uint8_t * top)392 static void LD4(uint8_t* dst, const uint8_t* top) {
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   const int E = top[4];
398   const int F = top[5];
399   const int G = top[6];
400   const int H = top[7];
401   DST(0, 0)                                     = AVG3(A, B, C);
402   DST(1, 0) = DST(0, 1)                         = AVG3(B, C, D);
403   DST(2, 0) = DST(1, 1) = DST(0, 2)             = AVG3(C, D, E);
404   DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
405   DST(3, 1) = DST(2, 2) = DST(1, 3)             = AVG3(E, F, G);
406   DST(3, 2) = DST(2, 3)                         = AVG3(F, G, H);
407   DST(3, 3)                                     = AVG3(G, H, H);
408 }
409 
VR4(uint8_t * dst,const uint8_t * top)410 static void VR4(uint8_t* dst, const uint8_t* top) {
411   const int X = top[-1];
412   const int I = top[-2];
413   const int J = top[-3];
414   const int K = top[-4];
415   const int A = top[0];
416   const int B = top[1];
417   const int C = top[2];
418   const int D = top[3];
419   DST(0, 0) = DST(1, 2) = AVG2(X, A);
420   DST(1, 0) = DST(2, 2) = AVG2(A, B);
421   DST(2, 0) = DST(3, 2) = AVG2(B, C);
422   DST(3, 0)             = AVG2(C, D);
423 
424   DST(0, 3) =             AVG3(K, J, I);
425   DST(0, 2) =             AVG3(J, I, X);
426   DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
427   DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
428   DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
429   DST(3, 1) =             AVG3(B, C, D);
430 }
431 
VL4(uint8_t * dst,const uint8_t * top)432 static void VL4(uint8_t* dst, const uint8_t* top) {
433   const int A = top[0];
434   const int B = top[1];
435   const int C = top[2];
436   const int D = top[3];
437   const int E = top[4];
438   const int F = top[5];
439   const int G = top[6];
440   const int H = top[7];
441   DST(0, 0) =             AVG2(A, B);
442   DST(1, 0) = DST(0, 2) = AVG2(B, C);
443   DST(2, 0) = DST(1, 2) = AVG2(C, D);
444   DST(3, 0) = DST(2, 2) = AVG2(D, E);
445 
446   DST(0, 1) =             AVG3(A, B, C);
447   DST(1, 1) = DST(0, 3) = AVG3(B, C, D);
448   DST(2, 1) = DST(1, 3) = AVG3(C, D, E);
449   DST(3, 1) = DST(2, 3) = AVG3(D, E, F);
450               DST(3, 2) = AVG3(E, F, G);
451               DST(3, 3) = AVG3(F, G, H);
452 }
453 
HU4(uint8_t * dst,const uint8_t * top)454 static void HU4(uint8_t* dst, const uint8_t* top) {
455   const int I = top[-2];
456   const int J = top[-3];
457   const int K = top[-4];
458   const int L = top[-5];
459   DST(0, 0) =             AVG2(I, J);
460   DST(2, 0) = DST(0, 1) = AVG2(J, K);
461   DST(2, 1) = DST(0, 2) = AVG2(K, L);
462   DST(1, 0) =             AVG3(I, J, K);
463   DST(3, 0) = DST(1, 1) = AVG3(J, K, L);
464   DST(3, 1) = DST(1, 2) = AVG3(K, L, L);
465   DST(3, 2) = DST(2, 2) =
466   DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L;
467 }
468 
HD4(uint8_t * dst,const uint8_t * top)469 static void HD4(uint8_t* dst, const uint8_t* top) {
470   const int X = top[-1];
471   const int I = top[-2];
472   const int J = top[-3];
473   const int K = top[-4];
474   const int L = top[-5];
475   const int A = top[0];
476   const int B = top[1];
477   const int C = top[2];
478 
479   DST(0, 0) = DST(2, 1) = AVG2(I, X);
480   DST(0, 1) = DST(2, 2) = AVG2(J, I);
481   DST(0, 2) = DST(2, 3) = AVG2(K, J);
482   DST(0, 3)             = AVG2(L, K);
483 
484   DST(3, 0)             = AVG3(A, B, C);
485   DST(2, 0)             = AVG3(X, A, B);
486   DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
487   DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
488   DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
489   DST(1, 3)             = AVG3(L, K, J);
490 }
491 
TM4(uint8_t * dst,const uint8_t * top)492 static void TM4(uint8_t* dst, const uint8_t* top) {
493   int x, y;
494   const uint8_t* const clip = clip1 + 255 - top[-1];
495   for (y = 0; y < 4; ++y) {
496     const uint8_t* const clip_table = clip + top[-2 - y];
497     for (x = 0; x < 4; ++x) {
498       dst[x] = clip_table[top[x]];
499     }
500     dst += BPS;
501   }
502 }
503 
504 #undef DST
505 #undef AVG3
506 #undef AVG2
507 
508 // Left samples are top[-5 .. -2], top_left is top[-1], top are
509 // located at top[0..3], and top right is top[4..7]
Intra4Preds(uint8_t * dst,const uint8_t * top)510 static void Intra4Preds(uint8_t* dst, const uint8_t* top) {
511   DC4(I4DC4 + dst, top);
512   TM4(I4TM4 + dst, top);
513   VE4(I4VE4 + dst, top);
514   HE4(I4HE4 + dst, top);
515   RD4(I4RD4 + dst, top);
516   VR4(I4VR4 + dst, top);
517   LD4(I4LD4 + dst, top);
518   VL4(I4VL4 + dst, top);
519   HD4(I4HD4 + dst, top);
520   HU4(I4HU4 + dst, top);
521 }
522 
523 //------------------------------------------------------------------------------
524 // Metric
525 
GetSSE(const uint8_t * a,const uint8_t * b,int w,int h)526 static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b,
527                               int w, int h) {
528   int count = 0;
529   int y, x;
530   for (y = 0; y < h; ++y) {
531     for (x = 0; x < w; ++x) {
532       const int diff = (int)a[x] - b[x];
533       count += diff * diff;
534     }
535     a += BPS;
536     b += BPS;
537   }
538   return count;
539 }
540 
SSE16x16(const uint8_t * a,const uint8_t * b)541 static int SSE16x16(const uint8_t* a, const uint8_t* b) {
542   return GetSSE(a, b, 16, 16);
543 }
SSE16x8(const uint8_t * a,const uint8_t * b)544 static int SSE16x8(const uint8_t* a, const uint8_t* b) {
545   return GetSSE(a, b, 16, 8);
546 }
SSE8x8(const uint8_t * a,const uint8_t * b)547 static int SSE8x8(const uint8_t* a, const uint8_t* b) {
548   return GetSSE(a, b, 8, 8);
549 }
SSE4x4(const uint8_t * a,const uint8_t * b)550 static int SSE4x4(const uint8_t* a, const uint8_t* b) {
551   return GetSSE(a, b, 4, 4);
552 }
553 
Mean16x4(const uint8_t * ref,uint32_t dc[4])554 static void Mean16x4(const uint8_t* ref, uint32_t dc[4]) {
555   int k, x, y;
556   for (k = 0; k < 4; ++k) {
557     uint32_t avg = 0;
558     for (y = 0; y < 4; ++y) {
559       for (x = 0; x < 4; ++x) {
560         avg += ref[x + y * BPS];
561       }
562     }
563     dc[k] = avg;
564     ref += 4;   // go to next 4x4 block.
565   }
566 }
567 
568 //------------------------------------------------------------------------------
569 // Texture distortion
570 //
571 // We try to match the spectral content (weighted) between source and
572 // reconstructed samples.
573 
574 // Hadamard transform
575 // Returns the weighted sum of the absolute value of transformed coefficients.
576 // w[] contains a row-major 4 by 4 symmetric matrix.
TTransform(const uint8_t * in,const uint16_t * w)577 static int TTransform(const uint8_t* in, const uint16_t* w) {
578   int sum = 0;
579   int tmp[16];
580   int i;
581   // horizontal pass
582   for (i = 0; i < 4; ++i, in += BPS) {
583     const int a0 = in[0] + in[2];
584     const int a1 = in[1] + in[3];
585     const int a2 = in[1] - in[3];
586     const int a3 = in[0] - in[2];
587     tmp[0 + i * 4] = a0 + a1;
588     tmp[1 + i * 4] = a3 + a2;
589     tmp[2 + i * 4] = a3 - a2;
590     tmp[3 + i * 4] = a0 - a1;
591   }
592   // vertical pass
593   for (i = 0; i < 4; ++i, ++w) {
594     const int a0 = tmp[0 + i] + tmp[8 + i];
595     const int a1 = tmp[4 + i] + tmp[12+ i];
596     const int a2 = tmp[4 + i] - tmp[12+ i];
597     const int a3 = tmp[0 + i] - tmp[8 + i];
598     const int b0 = a0 + a1;
599     const int b1 = a3 + a2;
600     const int b2 = a3 - a2;
601     const int b3 = a0 - a1;
602 
603     sum += w[ 0] * abs(b0);
604     sum += w[ 4] * abs(b1);
605     sum += w[ 8] * abs(b2);
606     sum += w[12] * abs(b3);
607   }
608   return sum;
609 }
610 
Disto4x4(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)611 static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
612                     const uint16_t* const w) {
613   const int sum1 = TTransform(a, w);
614   const int sum2 = TTransform(b, w);
615   return abs(sum2 - sum1) >> 5;
616 }
617 
Disto16x16(const uint8_t * const a,const uint8_t * const b,const uint16_t * const w)618 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
619                       const uint16_t* const w) {
620   int D = 0;
621   int x, y;
622   for (y = 0; y < 16 * BPS; y += 4 * BPS) {
623     for (x = 0; x < 16; x += 4) {
624       D += Disto4x4(a + x + y, b + x + y, w);
625     }
626   }
627   return D;
628 }
629 
630 //------------------------------------------------------------------------------
631 // Quantization
632 //
633 
634 static const uint8_t kZigzag[16] = {
635   0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15
636 };
637 
638 // Simple quantization
QuantizeBlock(int16_t in[16],int16_t out[16],const VP8Matrix * const mtx)639 static int QuantizeBlock(int16_t in[16], int16_t out[16],
640                          const VP8Matrix* const mtx) {
641   int last = -1;
642   int n;
643   for (n = 0; n < 16; ++n) {
644     const int j = kZigzag[n];
645     const int sign = (in[j] < 0);
646     const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j];
647     if (coeff > mtx->zthresh_[j]) {
648       const uint32_t Q = mtx->q_[j];
649       const uint32_t iQ = mtx->iq_[j];
650       const uint32_t B = mtx->bias_[j];
651       int level = QUANTDIV(coeff, iQ, B);
652       if (level > MAX_LEVEL) level = MAX_LEVEL;
653       if (sign) level = -level;
654       in[j] = level * (int)Q;
655       out[n] = level;
656       if (level) last = n;
657     } else {
658       out[n] = 0;
659       in[j] = 0;
660     }
661   }
662   return (last >= 0);
663 }
664 
Quantize2Blocks(int16_t in[32],int16_t out[32],const VP8Matrix * const mtx)665 static int Quantize2Blocks(int16_t in[32], int16_t out[32],
666                            const VP8Matrix* const mtx) {
667   int nz;
668   nz  = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0;
669   nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1;
670   return nz;
671 }
672 
673 //------------------------------------------------------------------------------
674 // Block copy
675 
Copy(const uint8_t * src,uint8_t * dst,int w,int h)676 static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) {
677   int y;
678   for (y = 0; y < h; ++y) {
679     memcpy(dst, src, w);
680     src += BPS;
681     dst += BPS;
682   }
683 }
684 
Copy4x4(const uint8_t * src,uint8_t * dst)685 static void Copy4x4(const uint8_t* src, uint8_t* dst) {
686   Copy(src, dst, 4, 4);
687 }
688 
Copy16x8(const uint8_t * src,uint8_t * dst)689 static void Copy16x8(const uint8_t* src, uint8_t* dst) {
690   Copy(src, dst, 16, 8);
691 }
692 
693 //------------------------------------------------------------------------------
694 // SSIM / PSNR
695 
696 // hat-shaped filter. Sum of coefficients is equal to 16.
697 static const uint32_t kWeight[2 * VP8_SSIM_KERNEL + 1] = {
698   1, 2, 3, 4, 3, 2, 1
699 };
700 static const uint32_t kWeightSum = 16 * 16;   // sum{kWeight}^2
701 
SSIMCalculation(const VP8DistoStats * const stats,uint32_t N)702 static WEBP_INLINE double SSIMCalculation(
703     const VP8DistoStats* const stats, uint32_t N  /*num samples*/) {
704   const uint32_t w2 =  N * N;
705   const uint32_t C1 = 20 * w2;
706   const uint32_t C2 = 60 * w2;
707   const uint32_t C3 = 8 * 8 * w2;   // 'dark' limit ~= 6
708   const uint64_t xmxm = (uint64_t)stats->xm * stats->xm;
709   const uint64_t ymym = (uint64_t)stats->ym * stats->ym;
710   if (xmxm + ymym >= C3) {
711     const int64_t xmym = (int64_t)stats->xm * stats->ym;
712     const int64_t sxy = (int64_t)stats->xym * N - xmym;    // can be negative
713     const uint64_t sxx = (uint64_t)stats->xxm * N - xmxm;
714     const uint64_t syy = (uint64_t)stats->yym * N - ymym;
715     // we descale by 8 to prevent overflow during the fnum/fden multiply.
716     const uint64_t num_S = (2 * (uint64_t)(sxy < 0 ? 0 : sxy) + C2) >> 8;
717     const uint64_t den_S = (sxx + syy + C2) >> 8;
718     const uint64_t fnum = (2 * xmym + C1) * num_S;
719     const uint64_t fden = (xmxm + ymym + C1) * den_S;
720     const double r = (double)fnum / fden;
721     assert(r >= 0. && r <= 1.0);
722     return r;
723   }
724   return 1.;   // area is too dark to contribute meaningfully
725 }
726 
VP8SSIMFromStats(const VP8DistoStats * const stats)727 double VP8SSIMFromStats(const VP8DistoStats* const stats) {
728   return SSIMCalculation(stats, kWeightSum);
729 }
730 
VP8SSIMFromStatsClipped(const VP8DistoStats * const stats)731 double VP8SSIMFromStatsClipped(const VP8DistoStats* const stats) {
732   return SSIMCalculation(stats, stats->w);
733 }
734 
SSIMGetClipped_C(const uint8_t * src1,int stride1,const uint8_t * src2,int stride2,int xo,int yo,int W,int H)735 static double SSIMGetClipped_C(const uint8_t* src1, int stride1,
736                                const uint8_t* src2, int stride2,
737                                int xo, int yo, int W, int H) {
738   VP8DistoStats stats = { 0, 0, 0, 0, 0, 0 };
739   const int ymin = (yo - VP8_SSIM_KERNEL < 0) ? 0 : yo - VP8_SSIM_KERNEL;
740   const int ymax = (yo + VP8_SSIM_KERNEL > H - 1) ? H - 1
741                                                   : yo + VP8_SSIM_KERNEL;
742   const int xmin = (xo - VP8_SSIM_KERNEL < 0) ? 0 : xo - VP8_SSIM_KERNEL;
743   const int xmax = (xo + VP8_SSIM_KERNEL > W - 1) ? W - 1
744                                                   : xo + VP8_SSIM_KERNEL;
745   int x, y;
746   src1 += ymin * stride1;
747   src2 += ymin * stride2;
748   for (y = ymin; y <= ymax; ++y, src1 += stride1, src2 += stride2) {
749     for (x = xmin; x <= xmax; ++x) {
750       const uint32_t w = kWeight[VP8_SSIM_KERNEL + x - xo]
751                        * kWeight[VP8_SSIM_KERNEL + y - yo];
752       const uint32_t s1 = src1[x];
753       const uint32_t s2 = src2[x];
754       stats.w   += w;
755       stats.xm  += w * s1;
756       stats.ym  += w * s2;
757       stats.xxm += w * s1 * s1;
758       stats.xym += w * s1 * s2;
759       stats.yym += w * s2 * s2;
760     }
761   }
762   return VP8SSIMFromStatsClipped(&stats);
763 }
764 
SSIMGet_C(const uint8_t * src1,int stride1,const uint8_t * src2,int stride2)765 static double SSIMGet_C(const uint8_t* src1, int stride1,
766                         const uint8_t* src2, int stride2) {
767   VP8DistoStats stats = { 0, 0, 0, 0, 0, 0 };
768   int x, y;
769   for (y = 0; y <= 2 * VP8_SSIM_KERNEL; ++y, src1 += stride1, src2 += stride2) {
770     for (x = 0; x <= 2 * VP8_SSIM_KERNEL; ++x) {
771       const uint32_t w = kWeight[x] * kWeight[y];
772       const uint32_t s1 = src1[x];
773       const uint32_t s2 = src2[x];
774       stats.xm  += w * s1;
775       stats.ym  += w * s2;
776       stats.xxm += w * s1 * s1;
777       stats.xym += w * s1 * s2;
778       stats.yym += w * s2 * s2;
779     }
780   }
781   return VP8SSIMFromStats(&stats);
782 }
783 
784 //------------------------------------------------------------------------------
785 
AccumulateSSE(const uint8_t * src1,const uint8_t * src2,int len)786 static uint32_t AccumulateSSE(const uint8_t* src1,
787                               const uint8_t* src2, int len) {
788   int i;
789   uint32_t sse2 = 0;
790   assert(len <= 65535);  // to ensure that accumulation fits within uint32_t
791   for (i = 0; i < len; ++i) {
792     const int32_t diff = src1[i] - src2[i];
793     sse2 += diff * diff;
794   }
795   return sse2;
796 }
797 
798 //------------------------------------------------------------------------------
799 
800 VP8SSIMGetFunc VP8SSIMGet;
801 VP8SSIMGetClippedFunc VP8SSIMGetClipped;
802 VP8AccumulateSSEFunc VP8AccumulateSSE;
803 
804 extern void VP8SSIMDspInitSSE2(void);
805 
806 static volatile VP8CPUInfo ssim_last_cpuinfo_used =
807     (VP8CPUInfo)&ssim_last_cpuinfo_used;
808 
VP8SSIMDspInit(void)809 WEBP_TSAN_IGNORE_FUNCTION void VP8SSIMDspInit(void) {
810   if (ssim_last_cpuinfo_used == VP8GetCPUInfo) return;
811 
812   VP8SSIMGetClipped = SSIMGetClipped_C;
813   VP8SSIMGet = SSIMGet_C;
814 
815   VP8AccumulateSSE = AccumulateSSE;
816   if (VP8GetCPUInfo != NULL) {
817 #if defined(WEBP_USE_SSE2)
818     if (VP8GetCPUInfo(kSSE2)) {
819       VP8SSIMDspInitSSE2();
820     }
821 #endif
822   }
823 
824   ssim_last_cpuinfo_used = VP8GetCPUInfo;
825 }
826 
827 //------------------------------------------------------------------------------
828 // Initialization
829 
830 // Speed-critical function pointers. We have to initialize them to the default
831 // implementations within VP8EncDspInit().
832 VP8CHisto VP8CollectHistogram;
833 VP8Idct VP8ITransform;
834 VP8Fdct VP8FTransform;
835 VP8Fdct VP8FTransform2;
836 VP8WHT VP8FTransformWHT;
837 VP8Intra4Preds VP8EncPredLuma4;
838 VP8IntraPreds VP8EncPredLuma16;
839 VP8IntraPreds VP8EncPredChroma8;
840 VP8Metric VP8SSE16x16;
841 VP8Metric VP8SSE8x8;
842 VP8Metric VP8SSE16x8;
843 VP8Metric VP8SSE4x4;
844 VP8WMetric VP8TDisto4x4;
845 VP8WMetric VP8TDisto16x16;
846 VP8MeanMetric VP8Mean16x4;
847 VP8QuantizeBlock VP8EncQuantizeBlock;
848 VP8Quantize2Blocks VP8EncQuantize2Blocks;
849 VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT;
850 VP8BlockCopy VP8Copy4x4;
851 VP8BlockCopy VP8Copy16x8;
852 
853 extern void VP8EncDspInitSSE2(void);
854 extern void VP8EncDspInitSSE41(void);
855 extern void VP8EncDspInitAVX2(void);
856 extern void VP8EncDspInitNEON(void);
857 extern void VP8EncDspInitMIPS32(void);
858 extern void VP8EncDspInitMIPSdspR2(void);
859 extern void VP8EncDspInitMSA(void);
860 
861 static volatile VP8CPUInfo enc_last_cpuinfo_used =
862     (VP8CPUInfo)&enc_last_cpuinfo_used;
863 
VP8EncDspInit(void)864 WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspInit(void) {
865   if (enc_last_cpuinfo_used == VP8GetCPUInfo) return;
866 
867   VP8DspInit();  // common inverse transforms
868   InitTables();
869 
870   // default C implementations
871   VP8CollectHistogram = CollectHistogram;
872   VP8ITransform = ITransform;
873   VP8FTransform = FTransform;
874   VP8FTransform2 = FTransform2;
875   VP8FTransformWHT = FTransformWHT;
876   VP8EncPredLuma4 = Intra4Preds;
877   VP8EncPredLuma16 = Intra16Preds;
878   VP8EncPredChroma8 = IntraChromaPreds;
879   VP8SSE16x16 = SSE16x16;
880   VP8SSE8x8 = SSE8x8;
881   VP8SSE16x8 = SSE16x8;
882   VP8SSE4x4 = SSE4x4;
883   VP8TDisto4x4 = Disto4x4;
884   VP8TDisto16x16 = Disto16x16;
885   VP8Mean16x4 = Mean16x4;
886   VP8EncQuantizeBlock = QuantizeBlock;
887   VP8EncQuantize2Blocks = Quantize2Blocks;
888   VP8EncQuantizeBlockWHT = QuantizeBlock;
889   VP8Copy4x4 = Copy4x4;
890   VP8Copy16x8 = Copy16x8;
891 
892   // If defined, use CPUInfo() to overwrite some pointers with faster versions.
893   if (VP8GetCPUInfo != NULL) {
894 #if defined(WEBP_USE_SSE2)
895     if (VP8GetCPUInfo(kSSE2)) {
896       VP8EncDspInitSSE2();
897 #if defined(WEBP_USE_SSE41)
898       if (VP8GetCPUInfo(kSSE4_1)) {
899         VP8EncDspInitSSE41();
900       }
901 #endif
902     }
903 #endif
904 #if defined(WEBP_USE_AVX2)
905     if (VP8GetCPUInfo(kAVX2)) {
906       VP8EncDspInitAVX2();
907     }
908 #endif
909 #if defined(WEBP_USE_NEON)
910     if (VP8GetCPUInfo(kNEON)) {
911       VP8EncDspInitNEON();
912     }
913 #endif
914 #if defined(WEBP_USE_MIPS32)
915     if (VP8GetCPUInfo(kMIPS32)) {
916       VP8EncDspInitMIPS32();
917     }
918 #endif
919 #if defined(WEBP_USE_MIPS_DSP_R2)
920     if (VP8GetCPUInfo(kMIPSdspR2)) {
921       VP8EncDspInitMIPSdspR2();
922     }
923 #endif
924 #if defined(WEBP_USE_MSA)
925     if (VP8GetCPUInfo(kMSA)) {
926       VP8EncDspInitMSA();
927     }
928 #endif
929   }
930   enc_last_cpuinfo_used = VP8GetCPUInfo;
931 }
932