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 "src/dsp/dsp.h" |
18 | #include "src/enc/vp8i_enc.h" |
19 | |
20 | static WEBP_INLINE uint8_t clip_8b(int v) { |
21 | return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255; |
22 | } |
23 | |
24 | #if !WEBP_NEON_OMIT_C_CODE |
25 | static WEBP_INLINE int clip_max(int v, int max) { |
26 | return (v > max) ? max : v; |
27 | } |
28 | #endif // !WEBP_NEON_OMIT_C_CODE |
29 | |
30 | //------------------------------------------------------------------------------ |
31 | // Compute susceptibility based on DCT-coeff histograms: |
32 | // the higher, the "easier" the macroblock is to compress. |
33 | |
34 | const int VP8DspScan[16 + 4 + 4] = { |
35 | // Luma |
36 | 0 + 0 * BPS, 4 + 0 * BPS, 8 + 0 * BPS, 12 + 0 * BPS, |
37 | 0 + 4 * BPS, 4 + 4 * BPS, 8 + 4 * BPS, 12 + 4 * BPS, |
38 | 0 + 8 * BPS, 4 + 8 * BPS, 8 + 8 * BPS, 12 + 8 * BPS, |
39 | 0 + 12 * BPS, 4 + 12 * BPS, 8 + 12 * BPS, 12 + 12 * BPS, |
40 | |
41 | 0 + 0 * BPS, 4 + 0 * BPS, 0 + 4 * BPS, 4 + 4 * BPS, // U |
42 | 8 + 0 * BPS, 12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS // V |
43 | }; |
44 | |
45 | // general-purpose util function |
46 | void VP8SetHistogramData(const int distribution[MAX_COEFF_THRESH + 1], |
47 | VP8Histogram* const histo) { |
48 | int max_value = 0, last_non_zero = 1; |
49 | int k; |
50 | for (k = 0; k <= MAX_COEFF_THRESH; ++k) { |
51 | const int value = distribution[k]; |
52 | if (value > 0) { |
53 | if (value > max_value) max_value = value; |
54 | last_non_zero = k; |
55 | } |
56 | } |
57 | histo->max_value = max_value; |
58 | histo->last_non_zero = last_non_zero; |
59 | } |
60 | |
61 | #if !WEBP_NEON_OMIT_C_CODE |
62 | static void CollectHistogram_C(const uint8_t* ref, const uint8_t* pred, |
63 | int start_block, int end_block, |
64 | VP8Histogram* const histo) { |
65 | int j; |
66 | int distribution[MAX_COEFF_THRESH + 1] = { 0 }; |
67 | for (j = start_block; j < end_block; ++j) { |
68 | int k; |
69 | int16_t out[16]; |
70 | |
71 | VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out); |
72 | |
73 | // Convert coefficients to bin. |
74 | for (k = 0; k < 16; ++k) { |
75 | const int v = abs(out[k]) >> 3; |
76 | const int clipped_value = clip_max(v, MAX_COEFF_THRESH); |
77 | ++distribution[clipped_value]; |
78 | } |
79 | } |
80 | VP8SetHistogramData(distribution, histo); |
81 | } |
82 | #endif // !WEBP_NEON_OMIT_C_CODE |
83 | |
84 | //------------------------------------------------------------------------------ |
85 | // run-time tables (~4k) |
86 | |
87 | static uint8_t clip1[255 + 510 + 1]; // clips [-255,510] to [0,255] |
88 | |
89 | // We declare this variable 'volatile' to prevent instruction reordering |
90 | // and make sure it's set to true _last_ (so as to be thread-safe) |
91 | static volatile int tables_ok = 0; |
92 | |
93 | static WEBP_TSAN_IGNORE_FUNCTION void InitTables(void) { |
94 | if (!tables_ok) { |
95 | int i; |
96 | for (i = -255; i <= 255 + 255; ++i) { |
97 | clip1[255 + i] = clip_8b(i); |
98 | } |
99 | tables_ok = 1; |
100 | } |
101 | } |
102 | |
103 | |
104 | //------------------------------------------------------------------------------ |
105 | // Transforms (Paragraph 14.4) |
106 | |
107 | #if !WEBP_NEON_OMIT_C_CODE |
108 | |
109 | #define STORE(x, y, v) \ |
110 | dst[(x) + (y) * BPS] = clip_8b(ref[(x) + (y) * BPS] + ((v) >> 3)) |
111 | |
112 | static const int kC1 = 20091 + (1 << 16); |
113 | static const int kC2 = 35468; |
114 | #define MUL(a, b) (((a) * (b)) >> 16) |
115 | |
116 | static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in, |
117 | uint8_t* dst) { |
118 | int C[4 * 4], *tmp; |
119 | int i; |
120 | tmp = C; |
121 | for (i = 0; i < 4; ++i) { // vertical pass |
122 | const int a = in[0] + in[8]; |
123 | const int b = in[0] - in[8]; |
124 | const int c = MUL(in[4], kC2) - MUL(in[12], kC1); |
125 | const int d = MUL(in[4], kC1) + MUL(in[12], kC2); |
126 | tmp[0] = a + d; |
127 | tmp[1] = b + c; |
128 | tmp[2] = b - c; |
129 | tmp[3] = a - d; |
130 | tmp += 4; |
131 | in++; |
132 | } |
133 | |
134 | tmp = C; |
135 | for (i = 0; i < 4; ++i) { // horizontal pass |
136 | const int dc = tmp[0] + 4; |
137 | const int a = dc + tmp[8]; |
138 | const int b = dc - tmp[8]; |
139 | const int c = MUL(tmp[4], kC2) - MUL(tmp[12], kC1); |
140 | const int d = MUL(tmp[4], kC1) + MUL(tmp[12], kC2); |
141 | STORE(0, i, a + d); |
142 | STORE(1, i, b + c); |
143 | STORE(2, i, b - c); |
144 | STORE(3, i, a - d); |
145 | tmp++; |
146 | } |
147 | } |
148 | |
149 | static void ITransform_C(const uint8_t* ref, const int16_t* in, uint8_t* dst, |
150 | int do_two) { |
151 | ITransformOne(ref, in, dst); |
152 | if (do_two) { |
153 | ITransformOne(ref + 4, in + 16, dst + 4); |
154 | } |
155 | } |
156 | |
157 | static void FTransform_C(const uint8_t* src, const uint8_t* ref, int16_t* out) { |
158 | int i; |
159 | int tmp[16]; |
160 | for (i = 0; i < 4; ++i, src += BPS, ref += BPS) { |
161 | const int d0 = src[0] - ref[0]; // 9bit dynamic range ([-255,255]) |
162 | const int d1 = src[1] - ref[1]; |
163 | const int d2 = src[2] - ref[2]; |
164 | const int d3 = src[3] - ref[3]; |
165 | const int a0 = (d0 + d3); // 10b [-510,510] |
166 | const int a1 = (d1 + d2); |
167 | const int a2 = (d1 - d2); |
168 | const int a3 = (d0 - d3); |
169 | tmp[0 + i * 4] = (a0 + a1) * 8; // 14b [-8160,8160] |
170 | tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9; // [-7536,7542] |
171 | tmp[2 + i * 4] = (a0 - a1) * 8; |
172 | tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 + 937) >> 9; |
173 | } |
174 | for (i = 0; i < 4; ++i) { |
175 | const int a0 = (tmp[0 + i] + tmp[12 + i]); // 15b |
176 | const int a1 = (tmp[4 + i] + tmp[ 8 + i]); |
177 | const int a2 = (tmp[4 + i] - tmp[ 8 + i]); |
178 | const int a3 = (tmp[0 + i] - tmp[12 + i]); |
179 | out[0 + i] = (a0 + a1 + 7) >> 4; // 12b |
180 | out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0); |
181 | out[8 + i] = (a0 - a1 + 7) >> 4; |
182 | out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16); |
183 | } |
184 | } |
185 | #endif // !WEBP_NEON_OMIT_C_CODE |
186 | |
187 | static void FTransform2_C(const uint8_t* src, const uint8_t* ref, |
188 | int16_t* out) { |
189 | VP8FTransform(src, ref, out); |
190 | VP8FTransform(src + 4, ref + 4, out + 16); |
191 | } |
192 | |
193 | #if !WEBP_NEON_OMIT_C_CODE |
194 | static void FTransformWHT_C(const int16_t* in, int16_t* out) { |
195 | // input is 12b signed |
196 | int32_t tmp[16]; |
197 | int i; |
198 | for (i = 0; i < 4; ++i, in += 64) { |
199 | const int a0 = (in[0 * 16] + in[2 * 16]); // 13b |
200 | const int a1 = (in[1 * 16] + in[3 * 16]); |
201 | const int a2 = (in[1 * 16] - in[3 * 16]); |
202 | const int a3 = (in[0 * 16] - in[2 * 16]); |
203 | tmp[0 + i * 4] = a0 + a1; // 14b |
204 | tmp[1 + i * 4] = a3 + a2; |
205 | tmp[2 + i * 4] = a3 - a2; |
206 | tmp[3 + i * 4] = a0 - a1; |
207 | } |
208 | for (i = 0; i < 4; ++i) { |
209 | const int a0 = (tmp[0 + i] + tmp[8 + i]); // 15b |
210 | const int a1 = (tmp[4 + i] + tmp[12+ i]); |
211 | const int a2 = (tmp[4 + i] - tmp[12+ i]); |
212 | const int a3 = (tmp[0 + i] - tmp[8 + i]); |
213 | const int b0 = a0 + a1; // 16b |
214 | const int b1 = a3 + a2; |
215 | const int b2 = a3 - a2; |
216 | const int b3 = a0 - a1; |
217 | out[ 0 + i] = b0 >> 1; // 15b |
218 | out[ 4 + i] = b1 >> 1; |
219 | out[ 8 + i] = b2 >> 1; |
220 | out[12 + i] = b3 >> 1; |
221 | } |
222 | } |
223 | #endif // !WEBP_NEON_OMIT_C_CODE |
224 | |
225 | #undef MUL |
226 | #undef STORE |
227 | |
228 | //------------------------------------------------------------------------------ |
229 | // Intra predictions |
230 | |
231 | static WEBP_INLINE void Fill(uint8_t* dst, int value, int size) { |
232 | int j; |
233 | for (j = 0; j < size; ++j) { |
234 | memset(dst + j * BPS, value, size); |
235 | } |
236 | } |
237 | |
238 | static WEBP_INLINE void VerticalPred(uint8_t* dst, |
239 | const uint8_t* top, int size) { |
240 | int j; |
241 | if (top != NULL) { |
242 | for (j = 0; j < size; ++j) memcpy(dst + j * BPS, top, size); |
243 | } else { |
244 | Fill(dst, 127, size); |
245 | } |
246 | } |
247 | |
248 | static WEBP_INLINE void HorizontalPred(uint8_t* dst, |
249 | const uint8_t* left, int size) { |
250 | if (left != NULL) { |
251 | int j; |
252 | for (j = 0; j < size; ++j) { |
253 | memset(dst + j * BPS, left[j], size); |
254 | } |
255 | } else { |
256 | Fill(dst, 129, size); |
257 | } |
258 | } |
259 | |
260 | static WEBP_INLINE void TrueMotion(uint8_t* dst, const uint8_t* left, |
261 | const uint8_t* top, int size) { |
262 | int y; |
263 | if (left != NULL) { |
264 | if (top != NULL) { |
265 | const uint8_t* const clip = clip1 + 255 - left[-1]; |
266 | for (y = 0; y < size; ++y) { |
267 | const uint8_t* const clip_table = clip + left[y]; |
268 | int x; |
269 | for (x = 0; x < size; ++x) { |
270 | dst[x] = clip_table[top[x]]; |
271 | } |
272 | dst += BPS; |
273 | } |
274 | } else { |
275 | HorizontalPred(dst, left, size); |
276 | } |
277 | } else { |
278 | // true motion without left samples (hence: with default 129 value) |
279 | // is equivalent to VE prediction where you just copy the top samples. |
280 | // Note that if top samples are not available, the default value is |
281 | // then 129, and not 127 as in the VerticalPred case. |
282 | if (top != NULL) { |
283 | VerticalPred(dst, top, size); |
284 | } else { |
285 | Fill(dst, 129, size); |
286 | } |
287 | } |
288 | } |
289 | |
290 | static WEBP_INLINE void DCMode(uint8_t* dst, const uint8_t* left, |
291 | const uint8_t* top, |
292 | int size, int round, int shift) { |
293 | int DC = 0; |
294 | int j; |
295 | if (top != NULL) { |
296 | for (j = 0; j < size; ++j) DC += top[j]; |
297 | if (left != NULL) { // top and left present |
298 | for (j = 0; j < size; ++j) DC += left[j]; |
299 | } else { // top, but no left |
300 | DC += DC; |
301 | } |
302 | DC = (DC + round) >> shift; |
303 | } else if (left != NULL) { // left but no top |
304 | for (j = 0; j < size; ++j) DC += left[j]; |
305 | DC += DC; |
306 | DC = (DC + round) >> shift; |
307 | } else { // no top, no left, nothing. |
308 | DC = 0x80; |
309 | } |
310 | Fill(dst, DC, size); |
311 | } |
312 | |
313 | //------------------------------------------------------------------------------ |
314 | // Chroma 8x8 prediction (paragraph 12.2) |
315 | |
316 | static void IntraChromaPreds_C(uint8_t* dst, const uint8_t* left, |
317 | const uint8_t* top) { |
318 | // U block |
319 | DCMode(C8DC8 + dst, left, top, 8, 8, 4); |
320 | VerticalPred(C8VE8 + dst, top, 8); |
321 | HorizontalPred(C8HE8 + dst, left, 8); |
322 | TrueMotion(C8TM8 + dst, left, top, 8); |
323 | // V block |
324 | dst += 8; |
325 | if (top != NULL) top += 8; |
326 | if (left != NULL) left += 16; |
327 | DCMode(C8DC8 + dst, left, top, 8, 8, 4); |
328 | VerticalPred(C8VE8 + dst, top, 8); |
329 | HorizontalPred(C8HE8 + dst, left, 8); |
330 | TrueMotion(C8TM8 + dst, left, top, 8); |
331 | } |
332 | |
333 | //------------------------------------------------------------------------------ |
334 | // luma 16x16 prediction (paragraph 12.3) |
335 | |
336 | static void Intra16Preds_C(uint8_t* dst, |
337 | const uint8_t* left, const uint8_t* top) { |
338 | DCMode(I16DC16 + dst, left, top, 16, 16, 5); |
339 | VerticalPred(I16VE16 + dst, top, 16); |
340 | HorizontalPred(I16HE16 + dst, left, 16); |
341 | TrueMotion(I16TM16 + dst, left, top, 16); |
342 | } |
343 | |
344 | //------------------------------------------------------------------------------ |
345 | // luma 4x4 prediction |
346 | |
347 | #define DST(x, y) dst[(x) + (y) * BPS] |
348 | #define AVG3(a, b, c) ((uint8_t)(((a) + 2 * (b) + (c) + 2) >> 2)) |
349 | #define AVG2(a, b) (((a) + (b) + 1) >> 1) |
350 | |
351 | static void VE4(uint8_t* dst, const uint8_t* top) { // vertical |
352 | const uint8_t vals[4] = { |
353 | AVG3(top[-1], top[0], top[1]), |
354 | AVG3(top[ 0], top[1], top[2]), |
355 | AVG3(top[ 1], top[2], top[3]), |
356 | AVG3(top[ 2], top[3], top[4]) |
357 | }; |
358 | int i; |
359 | for (i = 0; i < 4; ++i) { |
360 | memcpy(dst + i * BPS, vals, 4); |
361 | } |
362 | } |
363 | |
364 | static void HE4(uint8_t* dst, const uint8_t* top) { // horizontal |
365 | const int X = top[-1]; |
366 | const int I = top[-2]; |
367 | const int J = top[-3]; |
368 | const int K = top[-4]; |
369 | const int L = top[-5]; |
370 | WebPUint32ToMem(dst + 0 * BPS, 0x01010101U * AVG3(X, I, J)); |
371 | WebPUint32ToMem(dst + 1 * BPS, 0x01010101U * AVG3(I, J, K)); |
372 | WebPUint32ToMem(dst + 2 * BPS, 0x01010101U * AVG3(J, K, L)); |
373 | WebPUint32ToMem(dst + 3 * BPS, 0x01010101U * AVG3(K, L, L)); |
374 | } |
375 | |
376 | static void DC4(uint8_t* dst, const uint8_t* top) { |
377 | uint32_t dc = 4; |
378 | int i; |
379 | for (i = 0; i < 4; ++i) dc += top[i] + top[-5 + i]; |
380 | Fill(dst, dc >> 3, 4); |
381 | } |
382 | |
383 | static void RD4(uint8_t* dst, const uint8_t* top) { |
384 | const int X = top[-1]; |
385 | const int I = top[-2]; |
386 | const int J = top[-3]; |
387 | const int K = top[-4]; |
388 | const int L = top[-5]; |
389 | const int A = top[0]; |
390 | const int B = top[1]; |
391 | const int C = top[2]; |
392 | const int D = top[3]; |
393 | DST(0, 3) = AVG3(J, K, L); |
394 | DST(0, 2) = DST(1, 3) = AVG3(I, J, K); |
395 | DST(0, 1) = DST(1, 2) = DST(2, 3) = AVG3(X, I, J); |
396 | DST(0, 0) = DST(1, 1) = DST(2, 2) = DST(3, 3) = AVG3(A, X, I); |
397 | DST(1, 0) = DST(2, 1) = DST(3, 2) = AVG3(B, A, X); |
398 | DST(2, 0) = DST(3, 1) = AVG3(C, B, A); |
399 | DST(3, 0) = AVG3(D, C, B); |
400 | } |
401 | |
402 | static void LD4(uint8_t* dst, const uint8_t* top) { |
403 | const int A = top[0]; |
404 | const int B = top[1]; |
405 | const int C = top[2]; |
406 | const int D = top[3]; |
407 | const int E = top[4]; |
408 | const int F = top[5]; |
409 | const int G = top[6]; |
410 | const int H = top[7]; |
411 | DST(0, 0) = AVG3(A, B, C); |
412 | DST(1, 0) = DST(0, 1) = AVG3(B, C, D); |
413 | DST(2, 0) = DST(1, 1) = DST(0, 2) = AVG3(C, D, E); |
414 | DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F); |
415 | DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G); |
416 | DST(3, 2) = DST(2, 3) = AVG3(F, G, H); |
417 | DST(3, 3) = AVG3(G, H, H); |
418 | } |
419 | |
420 | static void VR4(uint8_t* dst, const uint8_t* top) { |
421 | const int X = top[-1]; |
422 | const int I = top[-2]; |
423 | const int J = top[-3]; |
424 | const int K = top[-4]; |
425 | const int A = top[0]; |
426 | const int B = top[1]; |
427 | const int C = top[2]; |
428 | const int D = top[3]; |
429 | DST(0, 0) = DST(1, 2) = AVG2(X, A); |
430 | DST(1, 0) = DST(2, 2) = AVG2(A, B); |
431 | DST(2, 0) = DST(3, 2) = AVG2(B, C); |
432 | DST(3, 0) = AVG2(C, D); |
433 | |
434 | DST(0, 3) = AVG3(K, J, I); |
435 | DST(0, 2) = AVG3(J, I, X); |
436 | DST(0, 1) = DST(1, 3) = AVG3(I, X, A); |
437 | DST(1, 1) = DST(2, 3) = AVG3(X, A, B); |
438 | DST(2, 1) = DST(3, 3) = AVG3(A, B, C); |
439 | DST(3, 1) = AVG3(B, C, D); |
440 | } |
441 | |
442 | static void VL4(uint8_t* dst, const uint8_t* top) { |
443 | const int A = top[0]; |
444 | const int B = top[1]; |
445 | const int C = top[2]; |
446 | const int D = top[3]; |
447 | const int E = top[4]; |
448 | const int F = top[5]; |
449 | const int G = top[6]; |
450 | const int H = top[7]; |
451 | DST(0, 0) = AVG2(A, B); |
452 | DST(1, 0) = DST(0, 2) = AVG2(B, C); |
453 | DST(2, 0) = DST(1, 2) = AVG2(C, D); |
454 | DST(3, 0) = DST(2, 2) = AVG2(D, E); |
455 | |
456 | DST(0, 1) = AVG3(A, B, C); |
457 | DST(1, 1) = DST(0, 3) = AVG3(B, C, D); |
458 | DST(2, 1) = DST(1, 3) = AVG3(C, D, E); |
459 | DST(3, 1) = DST(2, 3) = AVG3(D, E, F); |
460 | DST(3, 2) = AVG3(E, F, G); |
461 | DST(3, 3) = AVG3(F, G, H); |
462 | } |
463 | |
464 | static void HU4(uint8_t* dst, const uint8_t* top) { |
465 | const int I = top[-2]; |
466 | const int J = top[-3]; |
467 | const int K = top[-4]; |
468 | const int L = top[-5]; |
469 | DST(0, 0) = AVG2(I, J); |
470 | DST(2, 0) = DST(0, 1) = AVG2(J, K); |
471 | DST(2, 1) = DST(0, 2) = AVG2(K, L); |
472 | DST(1, 0) = AVG3(I, J, K); |
473 | DST(3, 0) = DST(1, 1) = AVG3(J, K, L); |
474 | DST(3, 1) = DST(1, 2) = AVG3(K, L, L); |
475 | DST(3, 2) = DST(2, 2) = |
476 | DST(0, 3) = DST(1, 3) = DST(2, 3) = DST(3, 3) = L; |
477 | } |
478 | |
479 | static void HD4(uint8_t* dst, const uint8_t* top) { |
480 | const int X = top[-1]; |
481 | const int I = top[-2]; |
482 | const int J = top[-3]; |
483 | const int K = top[-4]; |
484 | const int L = top[-5]; |
485 | const int A = top[0]; |
486 | const int B = top[1]; |
487 | const int C = top[2]; |
488 | |
489 | DST(0, 0) = DST(2, 1) = AVG2(I, X); |
490 | DST(0, 1) = DST(2, 2) = AVG2(J, I); |
491 | DST(0, 2) = DST(2, 3) = AVG2(K, J); |
492 | DST(0, 3) = AVG2(L, K); |
493 | |
494 | DST(3, 0) = AVG3(A, B, C); |
495 | DST(2, 0) = AVG3(X, A, B); |
496 | DST(1, 0) = DST(3, 1) = AVG3(I, X, A); |
497 | DST(1, 1) = DST(3, 2) = AVG3(J, I, X); |
498 | DST(1, 2) = DST(3, 3) = AVG3(K, J, I); |
499 | DST(1, 3) = AVG3(L, K, J); |
500 | } |
501 | |
502 | static void TM4(uint8_t* dst, const uint8_t* top) { |
503 | int x, y; |
504 | const uint8_t* const clip = clip1 + 255 - top[-1]; |
505 | for (y = 0; y < 4; ++y) { |
506 | const uint8_t* const clip_table = clip + top[-2 - y]; |
507 | for (x = 0; x < 4; ++x) { |
508 | dst[x] = clip_table[top[x]]; |
509 | } |
510 | dst += BPS; |
511 | } |
512 | } |
513 | |
514 | #undef DST |
515 | #undef AVG3 |
516 | #undef AVG2 |
517 | |
518 | // Left samples are top[-5 .. -2], top_left is top[-1], top are |
519 | // located at top[0..3], and top right is top[4..7] |
520 | static void Intra4Preds_C(uint8_t* dst, const uint8_t* top) { |
521 | DC4(I4DC4 + dst, top); |
522 | TM4(I4TM4 + dst, top); |
523 | VE4(I4VE4 + dst, top); |
524 | HE4(I4HE4 + dst, top); |
525 | RD4(I4RD4 + dst, top); |
526 | VR4(I4VR4 + dst, top); |
527 | LD4(I4LD4 + dst, top); |
528 | VL4(I4VL4 + dst, top); |
529 | HD4(I4HD4 + dst, top); |
530 | HU4(I4HU4 + dst, top); |
531 | } |
532 | |
533 | //------------------------------------------------------------------------------ |
534 | // Metric |
535 | |
536 | #if !WEBP_NEON_OMIT_C_CODE |
537 | static WEBP_INLINE int GetSSE(const uint8_t* a, const uint8_t* b, |
538 | int w, int h) { |
539 | int count = 0; |
540 | int y, x; |
541 | for (y = 0; y < h; ++y) { |
542 | for (x = 0; x < w; ++x) { |
543 | const int diff = (int)a[x] - b[x]; |
544 | count += diff * diff; |
545 | } |
546 | a += BPS; |
547 | b += BPS; |
548 | } |
549 | return count; |
550 | } |
551 | |
552 | static int SSE16x16_C(const uint8_t* a, const uint8_t* b) { |
553 | return GetSSE(a, b, 16, 16); |
554 | } |
555 | static int SSE16x8_C(const uint8_t* a, const uint8_t* b) { |
556 | return GetSSE(a, b, 16, 8); |
557 | } |
558 | static int SSE8x8_C(const uint8_t* a, const uint8_t* b) { |
559 | return GetSSE(a, b, 8, 8); |
560 | } |
561 | static int SSE4x4_C(const uint8_t* a, const uint8_t* b) { |
562 | return GetSSE(a, b, 4, 4); |
563 | } |
564 | #endif // !WEBP_NEON_OMIT_C_CODE |
565 | |
566 | static void Mean16x4_C(const uint8_t* ref, uint32_t dc[4]) { |
567 | int k, x, y; |
568 | for (k = 0; k < 4; ++k) { |
569 | uint32_t avg = 0; |
570 | for (y = 0; y < 4; ++y) { |
571 | for (x = 0; x < 4; ++x) { |
572 | avg += ref[x + y * BPS]; |
573 | } |
574 | } |
575 | dc[k] = avg; |
576 | ref += 4; // go to next 4x4 block. |
577 | } |
578 | } |
579 | |
580 | //------------------------------------------------------------------------------ |
581 | // Texture distortion |
582 | // |
583 | // We try to match the spectral content (weighted) between source and |
584 | // reconstructed samples. |
585 | |
586 | #if !WEBP_NEON_OMIT_C_CODE |
587 | // Hadamard transform |
588 | // Returns the weighted sum of the absolute value of transformed coefficients. |
589 | // w[] contains a row-major 4 by 4 symmetric matrix. |
590 | static int TTransform(const uint8_t* in, const uint16_t* w) { |
591 | int sum = 0; |
592 | int tmp[16]; |
593 | int i; |
594 | // horizontal pass |
595 | for (i = 0; i < 4; ++i, in += BPS) { |
596 | const int a0 = in[0] + in[2]; |
597 | const int a1 = in[1] + in[3]; |
598 | const int a2 = in[1] - in[3]; |
599 | const int a3 = in[0] - in[2]; |
600 | tmp[0 + i * 4] = a0 + a1; |
601 | tmp[1 + i * 4] = a3 + a2; |
602 | tmp[2 + i * 4] = a3 - a2; |
603 | tmp[3 + i * 4] = a0 - a1; |
604 | } |
605 | // vertical pass |
606 | for (i = 0; i < 4; ++i, ++w) { |
607 | const int a0 = tmp[0 + i] + tmp[8 + i]; |
608 | const int a1 = tmp[4 + i] + tmp[12+ i]; |
609 | const int a2 = tmp[4 + i] - tmp[12+ i]; |
610 | const int a3 = tmp[0 + i] - tmp[8 + i]; |
611 | const int b0 = a0 + a1; |
612 | const int b1 = a3 + a2; |
613 | const int b2 = a3 - a2; |
614 | const int b3 = a0 - a1; |
615 | |
616 | sum += w[ 0] * abs(b0); |
617 | sum += w[ 4] * abs(b1); |
618 | sum += w[ 8] * abs(b2); |
619 | sum += w[12] * abs(b3); |
620 | } |
621 | return sum; |
622 | } |
623 | |
624 | static int Disto4x4_C(const uint8_t* const a, const uint8_t* const b, |
625 | const uint16_t* const w) { |
626 | const int sum1 = TTransform(a, w); |
627 | const int sum2 = TTransform(b, w); |
628 | return abs(sum2 - sum1) >> 5; |
629 | } |
630 | |
631 | static int Disto16x16_C(const uint8_t* const a, const uint8_t* const b, |
632 | const uint16_t* const w) { |
633 | int D = 0; |
634 | int x, y; |
635 | for (y = 0; y < 16 * BPS; y += 4 * BPS) { |
636 | for (x = 0; x < 16; x += 4) { |
637 | D += Disto4x4_C(a + x + y, b + x + y, w); |
638 | } |
639 | } |
640 | return D; |
641 | } |
642 | #endif // !WEBP_NEON_OMIT_C_CODE |
643 | |
644 | //------------------------------------------------------------------------------ |
645 | // Quantization |
646 | // |
647 | |
648 | static const uint8_t kZigzag[16] = { |
649 | 0, 1, 4, 8, 5, 2, 3, 6, 9, 12, 13, 10, 7, 11, 14, 15 |
650 | }; |
651 | |
652 | // Simple quantization |
653 | static int QuantizeBlock_C(int16_t in[16], int16_t out[16], |
654 | const VP8Matrix* const mtx) { |
655 | int last = -1; |
656 | int n; |
657 | for (n = 0; n < 16; ++n) { |
658 | const int j = kZigzag[n]; |
659 | const int sign = (in[j] < 0); |
660 | const uint32_t coeff = (sign ? -in[j] : in[j]) + mtx->sharpen_[j]; |
661 | if (coeff > mtx->zthresh_[j]) { |
662 | const uint32_t Q = mtx->q_[j]; |
663 | const uint32_t iQ = mtx->iq_[j]; |
664 | const uint32_t B = mtx->bias_[j]; |
665 | int level = QUANTDIV(coeff, iQ, B); |
666 | if (level > MAX_LEVEL) level = MAX_LEVEL; |
667 | if (sign) level = -level; |
668 | in[j] = level * (int)Q; |
669 | out[n] = level; |
670 | if (level) last = n; |
671 | } else { |
672 | out[n] = 0; |
673 | in[j] = 0; |
674 | } |
675 | } |
676 | return (last >= 0); |
677 | } |
678 | |
679 | #if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC |
680 | static int Quantize2Blocks_C(int16_t in[32], int16_t out[32], |
681 | const VP8Matrix* const mtx) { |
682 | int nz; |
683 | nz = VP8EncQuantizeBlock(in + 0 * 16, out + 0 * 16, mtx) << 0; |
684 | nz |= VP8EncQuantizeBlock(in + 1 * 16, out + 1 * 16, mtx) << 1; |
685 | return nz; |
686 | } |
687 | #endif // !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC |
688 | |
689 | //------------------------------------------------------------------------------ |
690 | // Block copy |
691 | |
692 | static WEBP_INLINE void Copy(const uint8_t* src, uint8_t* dst, int w, int h) { |
693 | int y; |
694 | for (y = 0; y < h; ++y) { |
695 | memcpy(dst, src, w); |
696 | src += BPS; |
697 | dst += BPS; |
698 | } |
699 | } |
700 | |
701 | static void Copy4x4_C(const uint8_t* src, uint8_t* dst) { |
702 | Copy(src, dst, 4, 4); |
703 | } |
704 | |
705 | static void Copy16x8_C(const uint8_t* src, uint8_t* dst) { |
706 | Copy(src, dst, 16, 8); |
707 | } |
708 | |
709 | //------------------------------------------------------------------------------ |
710 | // Initialization |
711 | |
712 | // Speed-critical function pointers. We have to initialize them to the default |
713 | // implementations within VP8EncDspInit(). |
714 | VP8CHisto VP8CollectHistogram; |
715 | VP8Idct VP8ITransform; |
716 | VP8Fdct VP8FTransform; |
717 | VP8Fdct VP8FTransform2; |
718 | VP8WHT VP8FTransformWHT; |
719 | VP8Intra4Preds VP8EncPredLuma4; |
720 | VP8IntraPreds VP8EncPredLuma16; |
721 | VP8IntraPreds VP8EncPredChroma8; |
722 | VP8Metric VP8SSE16x16; |
723 | VP8Metric VP8SSE8x8; |
724 | VP8Metric VP8SSE16x8; |
725 | VP8Metric VP8SSE4x4; |
726 | VP8WMetric VP8TDisto4x4; |
727 | VP8WMetric VP8TDisto16x16; |
728 | VP8MeanMetric VP8Mean16x4; |
729 | VP8QuantizeBlock VP8EncQuantizeBlock; |
730 | VP8Quantize2Blocks VP8EncQuantize2Blocks; |
731 | VP8QuantizeBlockWHT VP8EncQuantizeBlockWHT; |
732 | VP8BlockCopy VP8Copy4x4; |
733 | VP8BlockCopy VP8Copy16x8; |
734 | |
735 | extern void VP8EncDspInitSSE2(void); |
736 | extern void VP8EncDspInitSSE41(void); |
737 | extern void VP8EncDspInitNEON(void); |
738 | extern void VP8EncDspInitMIPS32(void); |
739 | extern void VP8EncDspInitMIPSdspR2(void); |
740 | extern void VP8EncDspInitMSA(void); |
741 | |
742 | WEBP_DSP_INIT_FUNC(VP8EncDspInit) { |
743 | VP8DspInit(); // common inverse transforms |
744 | InitTables(); |
745 | |
746 | // default C implementations |
747 | #if !WEBP_NEON_OMIT_C_CODE |
748 | VP8ITransform = ITransform_C; |
749 | VP8FTransform = FTransform_C; |
750 | VP8FTransformWHT = FTransformWHT_C; |
751 | VP8TDisto4x4 = Disto4x4_C; |
752 | VP8TDisto16x16 = Disto16x16_C; |
753 | VP8CollectHistogram = CollectHistogram_C; |
754 | VP8SSE16x16 = SSE16x16_C; |
755 | VP8SSE16x8 = SSE16x8_C; |
756 | VP8SSE8x8 = SSE8x8_C; |
757 | VP8SSE4x4 = SSE4x4_C; |
758 | #endif |
759 | |
760 | #if !WEBP_NEON_OMIT_C_CODE || WEBP_NEON_WORK_AROUND_GCC |
761 | VP8EncQuantizeBlock = QuantizeBlock_C; |
762 | VP8EncQuantize2Blocks = Quantize2Blocks_C; |
763 | #endif |
764 | |
765 | VP8FTransform2 = FTransform2_C; |
766 | VP8EncPredLuma4 = Intra4Preds_C; |
767 | VP8EncPredLuma16 = Intra16Preds_C; |
768 | VP8EncPredChroma8 = IntraChromaPreds_C; |
769 | VP8Mean16x4 = Mean16x4_C; |
770 | VP8EncQuantizeBlockWHT = QuantizeBlock_C; |
771 | VP8Copy4x4 = Copy4x4_C; |
772 | VP8Copy16x8 = Copy16x8_C; |
773 | |
774 | // If defined, use CPUInfo() to overwrite some pointers with faster versions. |
775 | if (VP8GetCPUInfo != NULL) { |
776 | #if defined(WEBP_USE_SSE2) |
777 | if (VP8GetCPUInfo(kSSE2)) { |
778 | VP8EncDspInitSSE2(); |
779 | #if defined(WEBP_USE_SSE41) |
780 | if (VP8GetCPUInfo(kSSE4_1)) { |
781 | VP8EncDspInitSSE41(); |
782 | } |
783 | #endif |
784 | } |
785 | #endif |
786 | #if defined(WEBP_USE_MIPS32) |
787 | if (VP8GetCPUInfo(kMIPS32)) { |
788 | VP8EncDspInitMIPS32(); |
789 | } |
790 | #endif |
791 | #if defined(WEBP_USE_MIPS_DSP_R2) |
792 | if (VP8GetCPUInfo(kMIPSdspR2)) { |
793 | VP8EncDspInitMIPSdspR2(); |
794 | } |
795 | #endif |
796 | #if defined(WEBP_USE_MSA) |
797 | if (VP8GetCPUInfo(kMSA)) { |
798 | VP8EncDspInitMSA(); |
799 | } |
800 | #endif |
801 | } |
802 | |
803 | #if defined(WEBP_USE_NEON) |
804 | if (WEBP_NEON_OMIT_C_CODE || |
805 | (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { |
806 | VP8EncDspInitNEON(); |
807 | } |
808 | #endif |
809 | |
810 | assert(VP8ITransform != NULL); |
811 | assert(VP8FTransform != NULL); |
812 | assert(VP8FTransformWHT != NULL); |
813 | assert(VP8TDisto4x4 != NULL); |
814 | assert(VP8TDisto16x16 != NULL); |
815 | assert(VP8CollectHistogram != NULL); |
816 | assert(VP8SSE16x16 != NULL); |
817 | assert(VP8SSE16x8 != NULL); |
818 | assert(VP8SSE8x8 != NULL); |
819 | assert(VP8SSE4x4 != NULL); |
820 | assert(VP8EncQuantizeBlock != NULL); |
821 | assert(VP8EncQuantize2Blocks != NULL); |
822 | assert(VP8FTransform2 != NULL); |
823 | assert(VP8EncPredLuma4 != NULL); |
824 | assert(VP8EncPredLuma16 != NULL); |
825 | assert(VP8EncPredChroma8 != NULL); |
826 | assert(VP8Mean16x4 != NULL); |
827 | assert(VP8EncQuantizeBlockWHT != NULL); |
828 | assert(VP8Copy4x4 != NULL); |
829 | assert(VP8Copy16x8 != NULL); |
830 | } |
831 | |