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(uint32_t left,const uint32_t * const top)191 static uint32_t Predictor5_NEON(uint32_t left, const uint32_t* const top) {
192 return Average3_NEON(left, top[0], top[1]);
193 }
Predictor6_NEON(uint32_t left,const uint32_t * const top)194 static uint32_t Predictor6_NEON(uint32_t left, const uint32_t* const top) {
195 return Average2_NEON(left, top[-1]);
196 }
Predictor7_NEON(uint32_t left,const uint32_t * const top)197 static uint32_t Predictor7_NEON(uint32_t left, const uint32_t* const top) {
198 return Average2_NEON(left, top[0]);
199 }
Predictor13_NEON(uint32_t left,const uint32_t * const top)200 static uint32_t Predictor13_NEON(uint32_t left, const uint32_t* const top) {
201 return ClampedAddSubtractHalf_NEON(left, top[0], top[-1]);
202 }
203
204 // Batch versions of those functions.
205
206 // Predictor0: ARGB_BLACK.
PredictorAdd0_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)207 static void PredictorAdd0_NEON(const uint32_t* in, const uint32_t* upper,
208 int num_pixels, uint32_t* out) {
209 int i;
210 const uint8x16_t black = vreinterpretq_u8_u32(vdupq_n_u32(ARGB_BLACK));
211 for (i = 0; i + 4 <= num_pixels; i += 4) {
212 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
213 const uint8x16_t res = vaddq_u8(src, black);
214 STOREQ_U8_AS_U32P(&out[i], res);
215 }
216 VP8LPredictorsAdd_C[0](in + i, upper + i, num_pixels - i, out + i);
217 }
218
219 // Predictor1: left.
PredictorAdd1_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)220 static void PredictorAdd1_NEON(const uint32_t* in, const uint32_t* upper,
221 int num_pixels, uint32_t* out) {
222 int i;
223 const uint8x16_t zero = LOADQ_U32_AS_U8(0);
224 for (i = 0; i + 4 <= num_pixels; i += 4) {
225 // a | b | c | d
226 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
227 // 0 | a | b | c
228 const uint8x16_t shift0 = vextq_u8(zero, src, 12);
229 // a | a + b | b + c | c + d
230 const uint8x16_t sum0 = vaddq_u8(src, shift0);
231 // 0 | 0 | a | a + b
232 const uint8x16_t shift1 = vextq_u8(zero, sum0, 8);
233 // a | a + b | a + b + c | a + b + c + d
234 const uint8x16_t sum1 = vaddq_u8(sum0, shift1);
235 const uint8x16_t prev = LOADQ_U32_AS_U8(out[i - 1]);
236 const uint8x16_t res = vaddq_u8(sum1, prev);
237 STOREQ_U8_AS_U32P(&out[i], res);
238 }
239 VP8LPredictorsAdd_C[1](in + i, upper + i, num_pixels - i, out + i);
240 }
241
242 // Macro that adds 32-bit integers from IN using mod 256 arithmetic
243 // per 8 bit channel.
244 #define GENERATE_PREDICTOR_1(X, IN) \
245 static void PredictorAdd##X##_NEON(const uint32_t* in, \
246 const uint32_t* upper, int num_pixels, \
247 uint32_t* out) { \
248 int i; \
249 for (i = 0; i + 4 <= num_pixels; i += 4) { \
250 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]); \
251 const uint8x16_t other = LOADQ_U32P_AS_U8(&(IN)); \
252 const uint8x16_t res = vaddq_u8(src, other); \
253 STOREQ_U8_AS_U32P(&out[i], res); \
254 } \
255 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
256 }
257 // Predictor2: Top.
258 GENERATE_PREDICTOR_1(2, upper[i])
259 // Predictor3: Top-right.
260 GENERATE_PREDICTOR_1(3, upper[i + 1])
261 // Predictor4: Top-left.
262 GENERATE_PREDICTOR_1(4, upper[i - 1])
263 #undef GENERATE_PREDICTOR_1
264
265 // Predictor5: average(average(left, TR), T)
266 #define DO_PRED5(LANE) do { \
267 const uint8x16_t avgLTR = vhaddq_u8(L, TR); \
268 const uint8x16_t avg = vhaddq_u8(avgLTR, T); \
269 const uint8x16_t res = vaddq_u8(avg, src); \
270 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
271 L = ROTATE32_LEFT(res); \
272 } while (0)
273
PredictorAdd5_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)274 static void PredictorAdd5_NEON(const uint32_t* in, const uint32_t* upper,
275 int num_pixels, uint32_t* out) {
276 int i;
277 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
278 for (i = 0; i + 4 <= num_pixels; i += 4) {
279 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
280 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i + 0]);
281 const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + 1]);
282 DO_PRED5(0);
283 DO_PRED5(1);
284 DO_PRED5(2);
285 DO_PRED5(3);
286 }
287 VP8LPredictorsAdd_C[5](in + i, upper + i, num_pixels - i, out + i);
288 }
289 #undef DO_PRED5
290
291 #define DO_PRED67(LANE) do { \
292 const uint8x16_t avg = vhaddq_u8(L, top); \
293 const uint8x16_t res = vaddq_u8(avg, src); \
294 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
295 L = ROTATE32_LEFT(res); \
296 } while (0)
297
298 // Predictor6: average(left, TL)
PredictorAdd6_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)299 static void PredictorAdd6_NEON(const uint32_t* in, const uint32_t* upper,
300 int num_pixels, uint32_t* out) {
301 int i;
302 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
303 for (i = 0; i + 4 <= num_pixels; i += 4) {
304 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
305 const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i - 1]);
306 DO_PRED67(0);
307 DO_PRED67(1);
308 DO_PRED67(2);
309 DO_PRED67(3);
310 }
311 VP8LPredictorsAdd_C[6](in + i, upper + i, num_pixels - i, out + i);
312 }
313
314 // Predictor7: average(left, T)
PredictorAdd7_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)315 static void PredictorAdd7_NEON(const uint32_t* in, const uint32_t* upper,
316 int num_pixels, uint32_t* out) {
317 int i;
318 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
319 for (i = 0; i + 4 <= num_pixels; i += 4) {
320 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
321 const uint8x16_t top = LOADQ_U32P_AS_U8(&upper[i]);
322 DO_PRED67(0);
323 DO_PRED67(1);
324 DO_PRED67(2);
325 DO_PRED67(3);
326 }
327 VP8LPredictorsAdd_C[7](in + i, upper + i, num_pixels - i, out + i);
328 }
329 #undef DO_PRED67
330
331 #define GENERATE_PREDICTOR_2(X, IN) \
332 static void PredictorAdd##X##_NEON(const uint32_t* in, \
333 const uint32_t* upper, int num_pixels, \
334 uint32_t* out) { \
335 int i; \
336 for (i = 0; i + 4 <= num_pixels; i += 4) { \
337 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]); \
338 const uint8x16_t Tother = LOADQ_U32P_AS_U8(&(IN)); \
339 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]); \
340 const uint8x16_t avg = vhaddq_u8(T, Tother); \
341 const uint8x16_t res = vaddq_u8(avg, src); \
342 STOREQ_U8_AS_U32P(&out[i], res); \
343 } \
344 VP8LPredictorsAdd_C[(X)](in + i, upper + i, num_pixels - i, out + i); \
345 }
346 // Predictor8: average TL T.
347 GENERATE_PREDICTOR_2(8, upper[i - 1])
348 // Predictor9: average T TR.
349 GENERATE_PREDICTOR_2(9, upper[i + 1])
350 #undef GENERATE_PREDICTOR_2
351
352 // Predictor10: average of (average of (L,TL), average of (T, TR)).
353 #define DO_PRED10(LANE) do { \
354 const uint8x16_t avgLTL = vhaddq_u8(L, TL); \
355 const uint8x16_t avg = vhaddq_u8(avgTTR, avgLTL); \
356 const uint8x16_t res = vaddq_u8(avg, src); \
357 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
358 L = ROTATE32_LEFT(res); \
359 } while (0)
360
PredictorAdd10_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)361 static void PredictorAdd10_NEON(const uint32_t* in, const uint32_t* upper,
362 int num_pixels, uint32_t* out) {
363 int i;
364 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
365 for (i = 0; i + 4 <= num_pixels; i += 4) {
366 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
367 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
368 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
369 const uint8x16_t TR = LOADQ_U32P_AS_U8(&upper[i + 1]);
370 const uint8x16_t avgTTR = vhaddq_u8(T, TR);
371 DO_PRED10(0);
372 DO_PRED10(1);
373 DO_PRED10(2);
374 DO_PRED10(3);
375 }
376 VP8LPredictorsAdd_C[10](in + i, upper + i, num_pixels - i, out + i);
377 }
378 #undef DO_PRED10
379
380 // Predictor11: select.
381 #define DO_PRED11(LANE) do { \
382 const uint8x16_t sumLin = vaddq_u8(L, src); /* in + L */ \
383 const uint8x16_t pLTL = vabdq_u8(L, TL); /* |L - TL| */ \
384 const uint16x8_t sum_LTL = vpaddlq_u8(pLTL); \
385 const uint32x4_t pa = vpaddlq_u16(sum_LTL); \
386 const uint32x4_t mask = vcleq_u32(pa, pb); \
387 const uint8x16_t res = vbslq_u8(vreinterpretq_u8_u32(mask), sumTin, sumLin); \
388 vst1q_lane_u32(&out[i + (LANE)], vreinterpretq_u32_u8(res), (LANE)); \
389 L = ROTATE32_LEFT(res); \
390 } while (0)
391
PredictorAdd11_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)392 static void PredictorAdd11_NEON(const uint32_t* in, const uint32_t* upper,
393 int num_pixels, uint32_t* out) {
394 int i;
395 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
396 for (i = 0; i + 4 <= num_pixels; i += 4) {
397 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
398 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
399 const uint8x16_t pTTL = vabdq_u8(T, TL); // |T - TL|
400 const uint16x8_t sum_TTL = vpaddlq_u8(pTTL);
401 const uint32x4_t pb = vpaddlq_u16(sum_TTL);
402 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
403 const uint8x16_t sumTin = vaddq_u8(T, src); // in + T
404 DO_PRED11(0);
405 DO_PRED11(1);
406 DO_PRED11(2);
407 DO_PRED11(3);
408 }
409 VP8LPredictorsAdd_C[11](in + i, upper + i, num_pixels - i, out + i);
410 }
411 #undef DO_PRED11
412
413 // Predictor12: ClampedAddSubtractFull.
414 #define DO_PRED12(DIFF, LANE) do { \
415 const uint8x8_t pred = \
416 vqmovun_s16(vaddq_s16(vreinterpretq_s16_u16(L), (DIFF))); \
417 const uint8x8_t res = \
418 vadd_u8(pred, (LANE <= 1) ? vget_low_u8(src) : vget_high_u8(src)); \
419 const uint16x8_t res16 = vmovl_u8(res); \
420 vst1_lane_u32(&out[i + (LANE)], vreinterpret_u32_u8(res), (LANE) & 1); \
421 /* rotate in the left predictor for next iteration */ \
422 L = vextq_u16(res16, res16, 4); \
423 } while (0)
424
PredictorAdd12_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)425 static void PredictorAdd12_NEON(const uint32_t* in, const uint32_t* upper,
426 int num_pixels, uint32_t* out) {
427 int i;
428 uint16x8_t L = vmovl_u8(LOAD_U32_AS_U8(out[-1]));
429 for (i = 0; i + 4 <= num_pixels; i += 4) {
430 // load four pixels of source
431 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
432 // precompute the difference T - TL once for all, stored as s16
433 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
434 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
435 const int16x8_t diff_lo =
436 vreinterpretq_s16_u16(vsubl_u8(vget_low_u8(T), vget_low_u8(TL)));
437 const int16x8_t diff_hi =
438 vreinterpretq_s16_u16(vsubl_u8(vget_high_u8(T), vget_high_u8(TL)));
439 // loop over the four reconstructed pixels
440 DO_PRED12(diff_lo, 0);
441 DO_PRED12(diff_lo, 1);
442 DO_PRED12(diff_hi, 2);
443 DO_PRED12(diff_hi, 3);
444 }
445 VP8LPredictorsAdd_C[12](in + i, upper + i, num_pixels - i, out + i);
446 }
447 #undef DO_PRED12
448
449 // Predictor13: ClampedAddSubtractHalf
450 #define DO_PRED13(LANE, LOW_OR_HI) do { \
451 const uint8x16_t avg = vhaddq_u8(L, T); \
452 const uint8x16_t cmp = vcgtq_u8(TL, avg); \
453 const uint8x16_t TL_1 = vaddq_u8(TL, cmp); \
454 /* Compute half of the difference between avg and TL'. */ \
455 const int8x8_t diff_avg = \
456 vreinterpret_s8_u8(LOW_OR_HI(vhsubq_u8(avg, TL_1))); \
457 /* Compute the sum with avg and saturate. */ \
458 const int16x8_t avg_16 = vreinterpretq_s16_u16(vmovl_u8(LOW_OR_HI(avg))); \
459 const uint8x8_t delta = vqmovun_s16(vaddw_s8(avg_16, diff_avg)); \
460 const uint8x8_t res = vadd_u8(LOW_OR_HI(src), delta); \
461 const uint8x16_t res2 = vcombine_u8(res, res); \
462 vst1_lane_u32(&out[i + (LANE)], vreinterpret_u32_u8(res), (LANE) & 1); \
463 L = ROTATE32_LEFT(res2); \
464 } while (0)
465
PredictorAdd13_NEON(const uint32_t * in,const uint32_t * upper,int num_pixels,uint32_t * out)466 static void PredictorAdd13_NEON(const uint32_t* in, const uint32_t* upper,
467 int num_pixels, uint32_t* out) {
468 int i;
469 uint8x16_t L = LOADQ_U32_AS_U8(out[-1]);
470 for (i = 0; i + 4 <= num_pixels; i += 4) {
471 const uint8x16_t src = LOADQ_U32P_AS_U8(&in[i]);
472 const uint8x16_t T = LOADQ_U32P_AS_U8(&upper[i]);
473 const uint8x16_t TL = LOADQ_U32P_AS_U8(&upper[i - 1]);
474 DO_PRED13(0, vget_low_u8);
475 DO_PRED13(1, vget_low_u8);
476 DO_PRED13(2, vget_high_u8);
477 DO_PRED13(3, vget_high_u8);
478 }
479 VP8LPredictorsAdd_C[13](in + i, upper + i, num_pixels - i, out + i);
480 }
481 #undef DO_PRED13
482
483 #undef LOAD_U32_AS_U8
484 #undef LOAD_U32P_AS_U8
485 #undef LOADQ_U32_AS_U8
486 #undef LOADQ_U32P_AS_U8
487 #undef GET_U8_AS_U32
488 #undef GETQ_U8_AS_U32
489 #undef STOREQ_U8_AS_U32P
490 #undef ROTATE32_LEFT
491
492 //------------------------------------------------------------------------------
493 // Subtract-Green Transform
494
495 // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
496 // non-standard versions there.
497 #if defined(__APPLE__) && defined(__aarch64__) && \
498 defined(__apple_build_version__) && (__apple_build_version__< 6020037)
499 #define USE_VTBLQ
500 #endif
501
502 #ifdef USE_VTBLQ
503 // 255 = byte will be zeroed
504 static const uint8_t kGreenShuffle[16] = {
505 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255
506 };
507
DoGreenShuffle_NEON(const uint8x16_t argb,const uint8x16_t shuffle)508 static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb,
509 const uint8x16_t shuffle) {
510 return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)),
511 vtbl1q_u8(argb, vget_high_u8(shuffle)));
512 }
513 #else // !USE_VTBLQ
514 // 255 = byte will be zeroed
515 static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 };
516
DoGreenShuffle_NEON(const uint8x16_t argb,const uint8x8_t shuffle)517 static WEBP_INLINE uint8x16_t DoGreenShuffle_NEON(const uint8x16_t argb,
518 const uint8x8_t shuffle) {
519 return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
520 vtbl1_u8(vget_high_u8(argb), shuffle));
521 }
522 #endif // USE_VTBLQ
523
AddGreenToBlueAndRed_NEON(const uint32_t * src,int num_pixels,uint32_t * dst)524 static void AddGreenToBlueAndRed_NEON(const uint32_t* src, int num_pixels,
525 uint32_t* dst) {
526 const uint32_t* const end = src + (num_pixels & ~3);
527 #ifdef USE_VTBLQ
528 const uint8x16_t shuffle = vld1q_u8(kGreenShuffle);
529 #else
530 const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
531 #endif
532 for (; src < end; src += 4, dst += 4) {
533 const uint8x16_t argb = vld1q_u8((const uint8_t*)src);
534 const uint8x16_t greens = DoGreenShuffle_NEON(argb, shuffle);
535 vst1q_u8((uint8_t*)dst, vaddq_u8(argb, greens));
536 }
537 // fallthrough and finish off with plain-C
538 VP8LAddGreenToBlueAndRed_C(src, num_pixels & 3, dst);
539 }
540
541 //------------------------------------------------------------------------------
542 // Color Transform
543
TransformColorInverse_NEON(const VP8LMultipliers * const m,const uint32_t * const src,int num_pixels,uint32_t * dst)544 static void TransformColorInverse_NEON(const VP8LMultipliers* const m,
545 const uint32_t* const src,
546 int num_pixels, uint32_t* dst) {
547 // sign-extended multiplying constants, pre-shifted by 6.
548 #define CST(X) (((int16_t)(m->X << 8)) >> 6)
549 const int16_t rb[8] = {
550 CST(green_to_blue_), CST(green_to_red_),
551 CST(green_to_blue_), CST(green_to_red_),
552 CST(green_to_blue_), CST(green_to_red_),
553 CST(green_to_blue_), CST(green_to_red_)
554 };
555 const int16x8_t mults_rb = vld1q_s16(rb);
556 const int16_t b2[8] = {
557 0, CST(red_to_blue_), 0, CST(red_to_blue_),
558 0, CST(red_to_blue_), 0, CST(red_to_blue_),
559 };
560 const int16x8_t mults_b2 = vld1q_s16(b2);
561 #undef CST
562 #ifdef USE_VTBLQ
563 static const uint8_t kg0g0[16] = {
564 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13
565 };
566 const uint8x16_t shuffle = vld1q_u8(kg0g0);
567 #else
568 static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 };
569 const uint8x8_t shuffle = vld1_u8(k0g0g);
570 #endif
571 const uint32x4_t mask_ag = vdupq_n_u32(0xff00ff00u);
572 int i;
573 for (i = 0; i + 4 <= num_pixels; i += 4) {
574 const uint8x16_t in = vld1q_u8((const uint8_t*)(src + i));
575 const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in), mask_ag);
576 // 0 g 0 g
577 const uint8x16_t greens = DoGreenShuffle_NEON(in, shuffle);
578 // x dr x db1
579 const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb);
580 // x r' x b'
581 const int8x16_t B = vaddq_s8(vreinterpretq_s8_u8(in),
582 vreinterpretq_s8_s16(A));
583 // r' 0 b' 0
584 const int16x8_t C = vshlq_n_s16(vreinterpretq_s16_s8(B), 8);
585 // x db2 0 0
586 const int16x8_t D = vqdmulhq_s16(C, mults_b2);
587 // 0 x db2 0
588 const uint32x4_t E = vshrq_n_u32(vreinterpretq_u32_s16(D), 8);
589 // r' x b'' 0
590 const int8x16_t F = vaddq_s8(vreinterpretq_s8_u32(E),
591 vreinterpretq_s8_s16(C));
592 // 0 r' 0 b''
593 const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F), 8);
594 const uint32x4_t out = vorrq_u32(vreinterpretq_u32_u16(G), a0g0);
595 vst1q_u32(dst + i, out);
596 }
597 // Fall-back to C-version for left-overs.
598 VP8LTransformColorInverse_C(m, src + i, num_pixels - i, dst + i);
599 }
600
601 #undef USE_VTBLQ
602
603 //------------------------------------------------------------------------------
604 // Entry point
605
606 extern void VP8LDspInitNEON(void);
607
VP8LDspInitNEON(void)608 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) {
609 VP8LPredictors[5] = Predictor5_NEON;
610 VP8LPredictors[6] = Predictor6_NEON;
611 VP8LPredictors[7] = Predictor7_NEON;
612 VP8LPredictors[13] = Predictor13_NEON;
613
614 VP8LPredictorsAdd[0] = PredictorAdd0_NEON;
615 VP8LPredictorsAdd[1] = PredictorAdd1_NEON;
616 VP8LPredictorsAdd[2] = PredictorAdd2_NEON;
617 VP8LPredictorsAdd[3] = PredictorAdd3_NEON;
618 VP8LPredictorsAdd[4] = PredictorAdd4_NEON;
619 VP8LPredictorsAdd[5] = PredictorAdd5_NEON;
620 VP8LPredictorsAdd[6] = PredictorAdd6_NEON;
621 VP8LPredictorsAdd[7] = PredictorAdd7_NEON;
622 VP8LPredictorsAdd[8] = PredictorAdd8_NEON;
623 VP8LPredictorsAdd[9] = PredictorAdd9_NEON;
624 VP8LPredictorsAdd[10] = PredictorAdd10_NEON;
625 VP8LPredictorsAdd[11] = PredictorAdd11_NEON;
626 VP8LPredictorsAdd[12] = PredictorAdd12_NEON;
627 VP8LPredictorsAdd[13] = PredictorAdd13_NEON;
628
629 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA_NEON;
630 VP8LConvertBGRAToBGR = ConvertBGRAToBGR_NEON;
631 VP8LConvertBGRAToRGB = ConvertBGRAToRGB_NEON;
632
633 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed_NEON;
634 VP8LTransformColorInverse = TransformColorInverse_NEON;
635 }
636
637 #else // !WEBP_USE_NEON
638
639 WEBP_DSP_INIT_STUB(VP8LDspInitNEON)
640
641 #endif // WEBP_USE_NEON
642