1 // Copyright 2014 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 // NEON variant of methods for lossless decoder
11 //
12 // Author: Skal (pascal.massimino@gmail.com)
13
14 #include "src/dsp/dsp.h"
15
16 #if defined(WEBP_USE_NEON)
17
18 #include <arm_neon.h>
19
20 #include "src/dsp/lossless.h"
21 #include "src/dsp/neon.h"
22
23 //------------------------------------------------------------------------------
24 // Colorspace conversion functions
25
26 #if !defined(WORK_AROUND_GCC)
27 // gcc 4.6.0 had some trouble (NDK-r9) with this code. We only use it for
28 // gcc-4.8.x at least.
ConvertBGRAToRGBA_NEON(const uint32_t * src,int num_pixels,uint8_t * dst)29 static void ConvertBGRAToRGBA_NEON(const uint32_t* src,
30 int num_pixels, uint8_t* dst) {
31 const uint32_t* const end = src + (num_pixels & ~15);
32 for (; src < end; src += 16) {
33 uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
34 // swap B and R. (VSWP d0,d2 has no intrinsics equivalent!)
35 const uint8x16_t tmp = pixel.val[0];
36 pixel.val[0] = pixel.val[2];
37 pixel.val[2] = tmp;
38 vst4q_u8(dst, pixel);
39 dst += 64;
40 }
41 VP8LConvertBGRAToRGBA_C(src, num_pixels & 15, dst); // left-overs
42 }
43
ConvertBGRAToBGR_NEON(const uint32_t * src,int num_pixels,uint8_t * dst)44 static void ConvertBGRAToBGR_NEON(const uint32_t* src,
45 int num_pixels, uint8_t* dst) {
46 const uint32_t* const end = src + (num_pixels & ~15);
47 for (; src < end; src += 16) {
48 const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
49 const uint8x16x3_t tmp = { { pixel.val[0], pixel.val[1], pixel.val[2] } };
50 vst3q_u8(dst, tmp);
51 dst += 48;
52 }
53 VP8LConvertBGRAToBGR_C(src, num_pixels & 15, dst); // left-overs
54 }
55
ConvertBGRAToRGB_NEON(const uint32_t * src,int num_pixels,uint8_t * dst)56 static void ConvertBGRAToRGB_NEON(const uint32_t* src,
57 int num_pixels, uint8_t* dst) {
58 const uint32_t* const end = src + (num_pixels & ~15);
59 for (; src < end; src += 16) {
60 const uint8x16x4_t pixel = vld4q_u8((uint8_t*)src);
61 const uint8x16x3_t tmp = { { pixel.val[2], pixel.val[1], pixel.val[0] } };
62 vst3q_u8(dst, tmp);
63 dst += 48;
64 }
65 VP8LConvertBGRAToRGB_C(src, num_pixels & 15, dst); // left-overs
66 }
67
68 #else // WORK_AROUND_GCC
69
70 // gcc-4.6.0 fallback
71
72 static const uint8_t kRGBAShuffle[8] = { 2, 1, 0, 3, 6, 5, 4, 7 };
73
ConvertBGRAToRGBA_NEON(const uint32_t * src,int num_pixels,uint8_t * dst)74 static void ConvertBGRAToRGBA_NEON(const uint32_t* src,
75 int num_pixels, uint8_t* dst) {
76 const uint32_t* const end = src + (num_pixels & ~1);
77 const uint8x8_t shuffle = vld1_u8(kRGBAShuffle);
78 for (; src < end; src += 2) {
79 const uint8x8_t pixels = vld1_u8((uint8_t*)src);
80 vst1_u8(dst, vtbl1_u8(pixels, shuffle));
81 dst += 8;
82 }
83 VP8LConvertBGRAToRGBA_C(src, num_pixels & 1, dst); // left-overs
84 }
85
86 static const uint8_t kBGRShuffle[3][8] = {
87 { 0, 1, 2, 4, 5, 6, 8, 9 },
88 { 10, 12, 13, 14, 16, 17, 18, 20 },
89 { 21, 22, 24, 25, 26, 28, 29, 30 }
90 };
91
ConvertBGRAToBGR_NEON(const uint32_t * src,int num_pixels,uint8_t * dst)92 static void ConvertBGRAToBGR_NEON(const uint32_t* src,
93 int num_pixels, uint8_t* dst) {
94 const uint32_t* const end = src + (num_pixels & ~7);
95 const uint8x8_t shuffle0 = vld1_u8(kBGRShuffle[0]);
96 const uint8x8_t shuffle1 = vld1_u8(kBGRShuffle[1]);
97 const uint8x8_t shuffle2 = vld1_u8(kBGRShuffle[2]);
98 for (; src < end; src += 8) {
99 uint8x8x4_t pixels;
100 INIT_VECTOR4(pixels,
101 vld1_u8((const uint8_t*)(src + 0)),
102 vld1_u8((const uint8_t*)(src + 2)),
103 vld1_u8((const uint8_t*)(src + 4)),
104 vld1_u8((const uint8_t*)(src + 6)));
105 vst1_u8(dst + 0, vtbl4_u8(pixels, shuffle0));
106 vst1_u8(dst + 8, vtbl4_u8(pixels, shuffle1));
107 vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2));
108 dst += 8 * 3;
109 }
110 VP8LConvertBGRAToBGR_C(src, num_pixels & 7, dst); // left-overs
111 }
112
113 static const uint8_t kRGBShuffle[3][8] = {
114 { 2, 1, 0, 6, 5, 4, 10, 9 },
115 { 8, 14, 13, 12, 18, 17, 16, 22 },
116 { 21, 20, 26, 25, 24, 30, 29, 28 }
117 };
118
ConvertBGRAToRGB_NEON(const uint32_t * src,int num_pixels,uint8_t * dst)119 static void ConvertBGRAToRGB_NEON(const uint32_t* src,
120 int num_pixels, uint8_t* dst) {
121 const uint32_t* const end = src + (num_pixels & ~7);
122 const uint8x8_t shuffle0 = vld1_u8(kRGBShuffle[0]);
123 const uint8x8_t shuffle1 = vld1_u8(kRGBShuffle[1]);
124 const uint8x8_t shuffle2 = vld1_u8(kRGBShuffle[2]);
125 for (; src < end; src += 8) {
126 uint8x8x4_t pixels;
127 INIT_VECTOR4(pixels,
128 vld1_u8((const uint8_t*)(src + 0)),
129 vld1_u8((const uint8_t*)(src + 2)),
130 vld1_u8((const uint8_t*)(src + 4)),
131 vld1_u8((const uint8_t*)(src + 6)));
132 vst1_u8(dst + 0, vtbl4_u8(pixels, shuffle0));
133 vst1_u8(dst + 8, vtbl4_u8(pixels, shuffle1));
134 vst1_u8(dst + 16, vtbl4_u8(pixels, shuffle2));
135 dst += 8 * 3;
136 }
137 VP8LConvertBGRAToRGB_C(src, num_pixels & 7, dst); // left-overs
138 }
139
140 #endif // !WORK_AROUND_GCC
141
142 //------------------------------------------------------------------------------
143 // Predictor Transform
144
145 #define LOAD_U32_AS_U8(IN) vreinterpret_u8_u32(vdup_n_u32((IN)))
146 #define LOAD_U32P_AS_U8(IN) vreinterpret_u8_u32(vld1_u32((IN)))
147 #define LOADQ_U32_AS_U8(IN) vreinterpretq_u8_u32(vdupq_n_u32((IN)))
148 #define LOADQ_U32P_AS_U8(IN) vreinterpretq_u8_u32(vld1q_u32((IN)))
149 #define GET_U8_AS_U32(IN) vget_lane_u32(vreinterpret_u32_u8((IN)), 0);
150 #define GETQ_U8_AS_U32(IN) vgetq_lane_u32(vreinterpretq_u32_u8((IN)), 0);
151 #define STOREQ_U8_AS_U32P(OUT, IN) vst1q_u32((OUT), vreinterpretq_u32_u8((IN)));
152 #define ROTATE32_LEFT(L) vextq_u8((L), (L), 12) // D|C|B|A -> C|B|A|D
153
Average2_u8_NEON(uint32_t a0,uint32_t a1)154 static WEBP_INLINE uint8x8_t Average2_u8_NEON(uint32_t a0, uint32_t a1) {
155 const uint8x8_t A0 = LOAD_U32_AS_U8(a0);
156 const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
157 return vhadd_u8(A0, A1);
158 }
159
ClampedAddSubtractHalf_NEON(uint32_t c0,uint32_t c1,uint32_t c2)160 static WEBP_INLINE uint32_t ClampedAddSubtractHalf_NEON(uint32_t c0,
161 uint32_t c1,
162 uint32_t c2) {
163 const uint8x8_t avg = Average2_u8_NEON(c0, c1);
164 // Remove one to c2 when bigger than avg.
165 const uint8x8_t C2 = LOAD_U32_AS_U8(c2);
166 const uint8x8_t cmp = vcgt_u8(C2, avg);
167 const uint8x8_t C2_1 = vadd_u8(C2, cmp);
168 // Compute half of the difference between avg and c2.
169 const int8x8_t diff_avg = vreinterpret_s8_u8(vhsub_u8(avg, C2_1));
170 // Compute the sum with avg and saturate.
171 const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(avg));
172 const uint8x8_t res = vqmovun_s16(vaddw_s8(avg_16, diff_avg));
173 const uint32_t output = GET_U8_AS_U32(res);
174 return output;
175 }
176
Average2_NEON(uint32_t a0,uint32_t a1)177 static WEBP_INLINE uint32_t Average2_NEON(uint32_t a0, uint32_t a1) {
178 const uint8x8_t avg_u8x8 = Average2_u8_NEON(a0, a1);
179 const uint32_t avg = GET_U8_AS_U32(avg_u8x8);
180 return avg;
181 }
182
Average3_NEON(uint32_t a0,uint32_t a1,uint32_t a2)183 static WEBP_INLINE uint32_t Average3_NEON(uint32_t a0, uint32_t a1,
184 uint32_t a2) {
185 const uint8x8_t avg0 = Average2_u8_NEON(a0, a2);
186 const uint8x8_t A1 = LOAD_U32_AS_U8(a1);
187 const uint32_t avg = GET_U8_AS_U32(vhadd_u8(avg0, A1));
188 return avg;
189 }
190
Predictor5_NEON(const uint32_t * const left,const uint32_t * const top)191 static uint32_t Predictor5_NEON(const uint32_t* const left,
192 const uint32_t* const top) {
193 return Average3_NEON(*left, top[0], top[1]);
194 }
Predictor6_NEON(const uint32_t * const left,const uint32_t * const top)195 static uint32_t Predictor6_NEON(const uint32_t* const left,
196 const uint32_t* const top) {
197 return Average2_NEON(*left, top[-1]);
198 }
Predictor7_NEON(const uint32_t * const left,const uint32_t * const top)199 static uint32_t Predictor7_NEON(const uint32_t* const left,
200 const uint32_t* const top) {
201 return Average2_NEON(*left, top[0]);
202 }
Predictor13_NEON(const uint32_t * const left,const uint32_t * const top)203 static uint32_t Predictor13_NEON(const uint32_t* const left,
204 const uint32_t* const top) {
205 return ClampedAddSubtractHalf_NEON(*left, top[0], top[-1]);
206 }
207
208 // Batch versions of those functions.
209
210 // Predictor0: ARGB_BLACK.
PredictorAdd0_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)211 static void PredictorAdd0_NEON(const uint32_t* in, const uint32_t* upper,
212 int num_pixels, uint32_t* out) {
213 int i;
214 const uint8x16_t black = vreinterpretq_u8_u32(vdupq_n_u32(ARGB_BLACK));
215 for (i = 0; i + 4 <= num_pixels; i += 4) {
216 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
217 const uint8x16_t res = vaddq_u8(src, black);
218 STOREQ_U8_AS_U32P(&out[i], res);
219 }
220 VP8LPredictorsAdd_C[0](in + i, upper + i, num_pixels - i, out + i);
221 }
222
223 // Predictor1: left.
PredictorAdd1_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)224 static void PredictorAdd1_NEON(const uint32_t* in, const uint32_t* upper,
225 int num_pixels, uint32_t* out) {
226 int i;
227 const uint8x16_t zero = LOADQ_U32_AS_U8(0);
228 for (i = 0; i + 4 <= num_pixels; i += 4) {
229 // a | b | c | d
230 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
231 // 0 | a | b | c
232 const uint8x16_t shift0 = vextq_u8(zero, src, 12);
233 // a | a + b | b + c | c + d
234 const uint8x16_t sum0 = vaddq_u8(src, shift0);
235 // 0 | 0 | a | a + b
236 const uint8x16_t shift1 = vextq_u8(zero, sum0, 8);
237 // a | a + b | a + b + c | a + b + c + d
238 const uint8x16_t sum1 = vaddq_u8(sum0, shift1);
239 const uint8x16_t prev = LOADQ_U32_AS_U8(out[i - 1]);
240 const uint8x16_t res = vaddq_u8(sum1, prev);
241 STOREQ_U8_AS_U32P(&out[i], res);
242 }
243 VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i);
244 }
245
246 // Macro that adds 32-bit integers from IN using mod 256 arithmetic
247 // per 8 bit channel.
248 #define GENERATE_PREDICTOR_1(X, IN) \
249 static void PredictorAdd##X##_NEON(const uint32_t* in, \
250 const uint32_t* upper, int num_pixels, \
251 uint32_t* out) { \
252 int i; \
253 for (i = 0; i + 4 <= num_pixels; i += 4) { \
254 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]); \
255 const uint8x16_t other = LOADQ_U32P_AS_U8(&(IN)); \
256 const uint8x16_t res = vaddq_u8(src, other); \
257 STOREQ_U8_AS_U32P(&out[i], res); \
258 } \
259 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
260 }
261 // Predictor2: Top.
262 GENERATE_PREDICTOR_1(2, upper[i])
263 // Predictor3: Top-right.
264 GENERATE_PREDICTOR_1(3, upper[i + 1])
265 // Predictor4: Top-left.
266 GENERATE_PREDICTOR_1(4, upper[i - 1])
267 #undef GENERATE_PREDICTOR_1
268
269 // Predictor5: average(average(left, TR), T)
270 #define DO_PRED5(LANE) do { \
271 const uint8x16_t avgLTR = vhaddq_u8(L, TR); \
272 const uint8x16_t avg = vhaddq_u8(avgLTR, T); \
273 const uint8x16_t res = vaddq_u8(avg, src); \
274 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
275 L = ROTATE32_LEFT(res); \
276 } while (0)
277
PredictorAdd5_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)278 static void PredictorAdd5_NEON(const uint32_t* in, const uint32_t* upper,
279 int num_pixels, uint32_t* out) {
280 int i;
281 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
282 for (i = 0; i + 4 <= num_pixels; i += 4) {
283 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
284 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i + 0]);
285 const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + 1]);
286 DO_PRED5(0);
287 DO_PRED5(1);
288 DO_PRED5(2);
289 DO_PRED5(3);
290 }
291 VP8LPredictorsAdd_C[5](in + i, upper + i, num_pixels - i, out + i);
292 }
293 #undef DO_PRED5
294
295 #define DO_PRED67(LANE) do { \
296 const uint8x16_t avg = vhaddq_u8(L, top); \
297 const uint8x16_t res = vaddq_u8(avg, src); \
298 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
299 L = ROTATE32_LEFT(res); \
300 } while (0)
301
302 // Predictor6: average(left, TL)
PredictorAdd6_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)303 static void PredictorAdd6_NEON(const uint32_t* in, const uint32_t* upper,
304 int num_pixels, uint32_t* out) {
305 int i;
306 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
307 for (i = 0; i + 4 <= num_pixels; i += 4) {
308 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
309 const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i - 1]);
310 DO_PRED67(0);
311 DO_PRED67(1);
312 DO_PRED67(2);
313 DO_PRED67(3);
314 }
315 VP8LPredictorsAdd_C[6](in + i, upper + i, num_pixels - i, out + i);
316 }
317
318 // Predictor7: average(left, T)
PredictorAdd7_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)319 static void PredictorAdd7_NEON(const uint32_t* in, const uint32_t* upper,
320 int num_pixels, uint32_t* out) {
321 int i;
322 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
323 for (i = 0; i + 4 <= num_pixels; i += 4) {
324 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
325 const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i]);
326 DO_PRED67(0);
327 DO_PRED67(1);
328 DO_PRED67(2);
329 DO_PRED67(3);
330 }
331 VP8LPredictorsAdd_C[7](in + i, upper + i, num_pixels - i, out + i);
332 }
333 #undef DO_PRED67
334
335 #define GENERATE_PREDICTOR_2(X, IN) \
336 static void PredictorAdd##X##_NEON(const uint32_t* in, \
337 const uint32_t* upper, int num_pixels, \
338 uint32_t* out) { \
339 int i; \
340 for (i = 0; i + 4 <= num_pixels; i += 4) { \
341 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]); \
342 const uint8x16_t Tother = LOADQ_U32P_AS_U8(&(IN)); \
343 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]); \
344 const uint8x16_t avg = vhaddq_u8(T, Tother); \
345 const uint8x16_t res = vaddq_u8(avg, src); \
346 STOREQ_U8_AS_U32P(&out[i], res); \
347 } \
348 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
349 }
350 // Predictor8: average TL T.
351 GENERATE_PREDICTOR_2(8, upper[i - 1])
352 // Predictor9: average T TR.
353 GENERATE_PREDICTOR_2(9, upper[i + 1])
354 #undef GENERATE_PREDICTOR_2
355
356 // Predictor10: average of (average of (L,TL), average of (T, TR)).
357 #define DO_PRED10(LANE) do { \
358 const uint8x16_t avgLTL = vhaddq_u8(L, TL); \
359 const uint8x16_t avg = vhaddq_u8(avgTTR, avgLTL); \
360 const uint8x16_t res = vaddq_u8(avg, src); \
361 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
362 L = ROTATE32_LEFT(res); \
363 } while (0)
364
PredictorAdd10_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)365 static void PredictorAdd10_NEON(const uint32_t* in, const uint32_t* upper,
366 int num_pixels, uint32_t* out) {
367 int i;
368 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
369 for (i = 0; i + 4 <= num_pixels; i += 4) {
370 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
371 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
372 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
373 const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + 1]);
374 const uint8x16_t avgTTR = vhaddq_u8(T, TR);
375 DO_PRED10(0);
376 DO_PRED10(1);
377 DO_PRED10(2);
378 DO_PRED10(3);
379 }
380 VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
381 }
382 #undef DO_PRED10
383
384 // Predictor11: select.
385 #define DO_PRED11(LANE) do { \
386 const uint8x16_t sumLin = vaddq_u8(L, src); /* in + L */ \
387 const uint8x16_t pLTL = vabdq_u8(L, TL); /* |L - TL| */ \
388 const uint16x8_t sum_LTL = vpaddlq_u8(pLTL); \
389 const uint32x4_t pa = vpaddlq_u16(sum_LTL); \
390 const uint32x4_t mask = vcleq_u32(pa, pb); \
391 const uint8x16_t res = vbslq_u8(vreinterpretq_u8_u32(mask), sumTin, sumLin); \
392 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
393 L = ROTATE32_LEFT(res); \
394 } while (0)
395
PredictorAdd11_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)396 static void PredictorAdd11_NEON(const uint32_t* in, const uint32_t* upper,
397 int num_pixels, uint32_t* out) {
398 int i;
399 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
400 for (i = 0; i + 4 <= num_pixels; i += 4) {
401 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
402 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
403 const uint8x16_t pTTL = vabdq_u8(T, TL); // |T - TL|
404 const uint16x8_t sum_TTL = vpaddlq_u8(pTTL);
405 const uint32x4_t pb = vpaddlq_u16(sum_TTL);
406 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
407 const uint8x16_t sumTin = vaddq_u8(T, src); // in + T
408 DO_PRED11(0);
409 DO_PRED11(1);
410 DO_PRED11(2);
411 DO_PRED11(3);
412 }
413 VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
414 }
415 #undef DO_PRED11
416
417 // Predictor12: ClampedAddSubtractFull.
418 #define DO_PRED12(DIFF, LANE) do { \
419 const uint8x8_t pred = \
420 vqmovun_s16(vaddq_s16(vreinterpretq_s16_u16(L), (DIFF))); \
421 const uint8x8_t res = \
422 vadd_u8(pred, (LANE <= 1) ? vget_low_u8(src) : vget_high_u8(src)); \
423 const uint16x8_t res16 = vmovl_u8(res); \
424 vst1_lane_u32(&out[i + (LANE)], vreinterpret_u32_u8(res), (LANE) & 1); \
425 /* rotate in the left predictor for next iteration */ \
426 L = vextq_u16(res16, res16, 4); \
427 } while (0)
428
PredictorAdd12_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)429 static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* upper,
430 int num_pixels, uint32_t* out) {
431 int i;
432 uint16x8_t L = vmovl_u8(LOAD_U32_AS_U8(out[-1]));
433 for (i = 0; i + 4 <= num_pixels; i += 4) {
434 // load four pixels of source
435 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
436 // precompute the difference T - TL once for all, stored as s16
437 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
438 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
439 const int16x8_t diff_lo =
440 vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(T), vget_low_u8(TL)));
441 const int16x8_t diff_hi =
442 vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T), vget_high_u8(TL)));
443 // loop over the four reconstructed pixels
444 DO_PRED12(diff_lo, 0);
445 DO_PRED12(diff_lo, 1);
446 DO_PRED12(diff_hi, 2);
447 DO_PRED12(diff_hi, 3);
448 }
449 VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
450 }
451 #undef DO_PRED12
452
453 // Predictor13: ClampedAddSubtractHalf
454 #define DO_PRED13(LANE, LOW_OR_HI) do { \
455 const uint8x16_t avg = vhaddq_u8(L, T); \
456 const uint8x16_t cmp = vcgtq_u8(TL, avg); \
457 const uint8x16_t TL_1 = vaddq_u8(TL, cmp); \
458 /* Compute half of the difference between avg and TL'. */ \
459 const int8x8_t diff_avg = \
460 vreinterpret_s8_u8(LOW_OR_HI(vhsubq_u8(avg, TL_1))); \
461 /* Compute the sum with avg and saturate. */ \
462 const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(LOW_OR_HI(avg))); \
463 const uint8x8_t delta = vqmovun_s16(vaddw_s8(avg_16, diff_avg)); \
464 const uint8x8_t res = vadd_u8(LOW_OR_HI(src), delta); \
465 const uint8x16_t res2 = vcombine_u8(res, res); \
466 vst1_lane_u32(&out[i + (LANE)], vreinterpret_u32_u8(res), (LANE) & 1); \
467 L = ROTATE32_LEFT(res2); \
468 } while (0)
469
PredictorAdd13_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)470 static void PredictorAdd13_NEON(const uint32_t* in, const uint32_t* upper,
471 int num_pixels, uint32_t* out) {
472 int i;
473 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
474 for (i = 0; i + 4 <= num_pixels; i += 4) {
475 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
476 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
477 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
478 DO_PRED13(0, vget_low_u8);
479 DO_PRED13(1, vget_low_u8);
480 DO_PRED13(2, vget_high_u8);
481 DO_PRED13(3, vget_high_u8);
482 }
483 VP8LPredictorsAdd_C[13](in + i, upper + i, num_pixels - i, out + i);
484 }
485 #undef DO_PRED13
486
487 #undef LOAD_U32_AS_U8
488 #undef LOAD_U32P_AS_U8
489 #undef LOADQ_U32_AS_U8
490 #undef LOADQ_U32P_AS_U8
491 #undef GET_U8_AS_U32
492 #undef GETQ_U8_AS_U32
493 #undef STOREQ_U8_AS_U32P
494 #undef ROTATE32_LEFT
495
496 //------------------------------------------------------------------------------
497 // Subtract-Green Transform
498
499 // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
500 // non-standard versions there.
501 #if defined(__APPLE__) && defined(__aarch64__) && \
502 defined(__apple_build_version__) && (__apple_build_version__< 6020037)
503 #define USE_VTBLQ
504 #endif
505
506 #ifdef USE_VTBLQ
507 // 255 = byte will be zeroed
508 static const uint8_t kGreenShuffle[16] = {
509 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255
510 };
511
DoGreenShuffle_NEON(const uint8x16_t argb,const uint8x16_t shuffle)512 static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb,
513 const uint8x16_t shuffle) {
514 return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
515 vtbl1q_u8(argb, vget_high_u8(shuffle)));
516 }
517 #else // !USE_VTBLQ
518 // 255 = byte will be zeroed
519 static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 };
520
DoGreenShuffle_NEON(const uint8x16_t argb,const uint8x8_t shuffle)521 static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb,
522 const uint8x8_t shuffle) {
523 return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
524 vtbl1_u8(vget_high_u8(argb), shuffle));
525 }
526 #endif // USE_VTBLQ
527
AddGreenToBlueAndRed_NEON(const uint32_t * src,int num_pixels,uint32_t * dst)528 static void AddGreenToBlueAndRed_NEON(const uint32_t* src, int num_pixels,
529 uint32_t* dst) {
530 const uint32_t* const end = src + (num_pixels & ~3);
531 #ifdef USE_VTBLQ
532 const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
533 #else
534 const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
535 #endif
536 for (; src < end; src += 4, dst += 4) {
537 const uint8x16_t argb = vld1q_u8((const uint8_t*)src);
538 const uint8x16_t greens = DoGreenShuffle_NEON(argb, shuffle);
539 vst1q_u8((uint8_t*)dst, vaddq_u8(argb, greens));
540 }
541 // fallthrough and finish off with plain-C
542 VP8LAddGreenToBlueAndRed_C(src, num_pixels & 3, dst);
543 }
544
545 //------------------------------------------------------------------------------
546 // Color Transform
547
TransformColorInverse_NEON(const VP8LMultipliers * const m,const uint32_t * const src,int num_pixels,uint32_t * dst)548 static void TransformColorInverse_NEON(const VP8LMultipliers* const m,
549 const uint32_t* const src,
550 int num_pixels, uint32_t* dst) {
551 // sign-extended multiplying constants, pre-shifted by 6.
552 #define CST(X) (((int16_t)(m->X << 8)) >> 6)
553 const int16_t rb[8] = {
554 CST(green_to_blue_), CST(green_to_red_),
555 CST(green_to_blue_), CST(green_to_red_),
556 CST(green_to_blue_), CST(green_to_red_),
557 CST(green_to_blue_), CST(green_to_red_)
558 };
559 const int16x8_t mults_rb = vld1q_s16(rb);
560 const int16_t b2[8] = {
561 0, CST(red_to_blue_), 0, CST(red_to_blue_),
562 0, CST(red_to_blue_), 0, CST(red_to_blue_),
563 };
564 const int16x8_t mults_b2 = vld1q_s16(b2);
565 #undef CST
566 #ifdef USE_VTBLQ
567 static const uint8_t kg0g0[16] = {
568 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13
569 };
570 const uint8x16_t shuffle = vld1q_u8(kg0g0);
571 #else
572 static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 };
573 const uint8x8_t shuffle = vld1_u8(k0g0g);
574 #endif
575 const uint32x4_t mask_ag = vdupq_n_u32(0xff00ff00u);
576 int i;
577 for (i = 0; i + 4 <= num_pixels; i += 4) {
578 const uint8x16_t in = vld1q_u8((const uint8_t*)(src + i));
579 const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in), mask_ag);
580 // 0 g 0 g
581 const uint8x16_t greens = DoGreenShuffle_NEON(in, shuffle);
582 // x dr x db1
583 const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb);
584 // x r' x b'
585 const int8x16_t B = vaddq_s8(vreinterpretq_s8_u8(in),
586 vreinterpretq_s8_s16(A));
587 // r' 0 b' 0
588 const int16x8_t C = vshlq_n_s16(vreinterpretq_s16_s8(B), 8);
589 // x db2 0 0
590 const int16x8_t D = vqdmulhq_s16(C, mults_b2);
591 // 0 x db2 0
592 const uint32x4_t E = vshrq_n_u32(vreinterpretq_u32_s16(D), 8);
593 // r' x b'' 0
594 const int8x16_t F = vaddq_s8(vreinterpretq_s8_u32(E),
595 vreinterpretq_s8_s16(C));
596 // 0 r' 0 b''
597 const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F), 8);
598 const uint32x4_t out = vorrq_u32(vreinterpretq_u32_u16(G), a0g0);
599 vst1q_u32(dst + i, out);
600 }
601 // Fall-back to C-version for left-overs.
602 VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
603 }
604
605 #undef USE_VTBLQ
606
607 //------------------------------------------------------------------------------
608 // Entry point
609
610 extern void VP8LDspInitNEON(void);
611
VP8LDspInitNEON(void)612 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) {
613 VP8LPredictors[5] = Predictor5_NEON;
614 VP8LPredictors[6] = Predictor6_NEON;
615 VP8LPredictors[7] = Predictor7_NEON;
616 VP8LPredictors[13] = Predictor13_NEON;
617
618 VP8LPredictorsAdd[0] = PredictorAdd0_NEON;
619 VP8LPredictorsAdd[1] = PredictorAdd1_NEON;
620 VP8LPredictorsAdd[2] = PredictorAdd2_NEON;
621 VP8LPredictorsAdd[3] = PredictorAdd3_NEON;
622 VP8LPredictorsAdd[4] = PredictorAdd4_NEON;
623 VP8LPredictorsAdd[5] = PredictorAdd5_NEON;
624 VP8LPredictorsAdd[6] = PredictorAdd6_NEON;
625 VP8LPredictorsAdd[7] = PredictorAdd7_NEON;
626 VP8LPredictorsAdd[8] = PredictorAdd8_NEON;
627 VP8LPredictorsAdd[9] = PredictorAdd9_NEON;
628 VP8LPredictorsAdd[10] = PredictorAdd10_NEON;
629 VP8LPredictorsAdd[11] = PredictorAdd11_NEON;
630 VP8LPredictorsAdd[12] = PredictorAdd12_NEON;
631 VP8LPredictorsAdd[13] = PredictorAdd13_NEON;
632
633 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_NEON;
634 VP8LConvertBGRAToBGR = ConvertBGRAToBGR_NEON;
635 VP8LConvertBGRAToRGB = ConvertBGRAToRGB_NEON;
636
637 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_NEON;
638 VP8LTransformColorInverse = TransformColorInverse_NEON;
639 }
640
641 #else // !WEBP_USE_NEON
642
643 WEBP_DSP_INIT_STUB(VP8LDspInitNEON)
644
645 #endif // WEBP_USE_NEON
646