third_party/libwebp/dsp/lossless_neon.c

   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 "./dsp.h"
  15
  16 #if defined(WEBP_USE_NEON)
  17
  18 #include <arm_neon.h>
  19
  20 #include "./lossless.h"
  21 #include "./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.
  29 static void ConvertBGRAToRGBA(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
  44 static void ConvertBGRAToBGR(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
  56 static void ConvertBGRAToRGB(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
  74 static void ConvertBGRAToRGBA(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
  92 static void ConvertBGRAToBGR(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
 119 static void ConvertBGRAToRGB(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
 144 #ifdef USE_INTRINSICS
 145
 146 static WEBP_INLINE uint32_t Average2(const uint32_t* const a,
 147                                      const uint32_t* const b) {
 148   const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
 149   const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
 150   const uint8x8_t avg = vhadd_u8(a0, b0);
 151   return vget_lane_u32(vreinterpret_u32_u8(avg), 0);
 152 }
 153
 154 static WEBP_INLINE uint32_t Average3(const uint32_t* const a,
 155                                      const uint32_t* const b,
 156                                      const uint32_t* const c) {
 157   const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
 158   const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
 159   const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
 160   const uint8x8_t avg1 = vhadd_u8(a0, c0);
 161   const uint8x8_t avg2 = vhadd_u8(avg1, b0);
 162   return vget_lane_u32(vreinterpret_u32_u8(avg2), 0);
 163 }
 164
 165 static WEBP_INLINE uint32_t Average4(const uint32_t* const a,
 166                                      const uint32_t* const b,
 167                                      const uint32_t* const c,
 168                                      const uint32_t* const d) {
 169   const uint8x8_t a0 = vreinterpret_u8_u64(vcreate_u64(*a));
 170   const uint8x8_t b0 = vreinterpret_u8_u64(vcreate_u64(*b));
 171   const uint8x8_t c0 = vreinterpret_u8_u64(vcreate_u64(*c));
 172   const uint8x8_t d0 = vreinterpret_u8_u64(vcreate_u64(*d));
 173   const uint8x8_t avg1 = vhadd_u8(a0, b0);
 174   const uint8x8_t avg2 = vhadd_u8(c0, d0);
 175   const uint8x8_t avg3 = vhadd_u8(avg1, avg2);
 176   return vget_lane_u32(vreinterpret_u32_u8(avg3), 0);
 177 }
 178
 179 static uint32_t Predictor5(uint32_t left, const uint32_t* const top) {
 180   return Average3(&left, top + 0, top + 1);
 181 }
 182
 183 static uint32_t Predictor6(uint32_t left, const uint32_t* const top) {
 184   return Average2(&left, top - 1);
 185 }
 186
 187 static uint32_t Predictor7(uint32_t left, const uint32_t* const top) {
 188   return Average2(&left, top + 0);
 189 }
 190
 191 static uint32_t Predictor8(uint32_t left, const uint32_t* const top) {
 192   (void)left;
 193   return Average2(top - 1, top + 0);
 194 }
 195
 196 static uint32_t Predictor9(uint32_t left, const uint32_t* const top) {
 197   (void)left;
 198   return Average2(top + 0, top + 1);
 199 }
 200
 201 static uint32_t Predictor10(uint32_t left, const uint32_t* const top) {
 202   return Average4(&left, top - 1, top + 0, top + 1);
 203 }
 204
 205 //------------------------------------------------------------------------------
 206
 207 static WEBP_INLINE uint32_t Select(const uint32_t* const c0,
 208                                    const uint32_t* const c1,
 209                                    const uint32_t* const c2) {
 210   const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
 211   const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
 212   const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
 213   const uint8x8_t bc = vabd_u8(p1, p2);   // |b-c|
 214   const uint8x8_t ac = vabd_u8(p0, p2);   // |a-c|
 215   const int16x4_t sum_bc = vreinterpret_s16_u16(vpaddl_u8(bc));
 216   const int16x4_t sum_ac = vreinterpret_s16_u16(vpaddl_u8(ac));
 217   const int32x2_t diff = vpaddl_s16(vsub_s16(sum_bc, sum_ac));
 218   const int32_t pa_minus_pb = vget_lane_s32(diff, 0);
 219   return (pa_minus_pb <= 0) ? *c0 : *c1;
 220 }
 221
 222 static uint32_t Predictor11(uint32_t left, const uint32_t* const top) {
 223   return Select(top + 0, &left, top - 1);
 224 }
 225
 226 static WEBP_INLINE uint32_t ClampedAddSubtractFull(const uint32_t* const c0,
 227                                                    const uint32_t* const c1,
 228                                                    const uint32_t* const c2) {
 229   const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
 230   const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
 231   const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
 232   const uint16x8_t sum0 = vaddl_u8(p0, p1);                // add and widen
 233   const uint16x8_t sum1 = vqsubq_u16(sum0, vmovl_u8(p2));  // widen and subtract
 234   const uint8x8_t out = vqmovn_u16(sum1);                  // narrow and clamp
 235   return vget_lane_u32(vreinterpret_u32_u8(out), 0);
 236 }
 237
 238 static uint32_t Predictor12(uint32_t left, const uint32_t* const top) {
 239   return ClampedAddSubtractFull(&left, top + 0, top - 1);
 240 }
 241
 242 static WEBP_INLINE uint32_t ClampedAddSubtractHalf(const uint32_t* const c0,
 243                                                    const uint32_t* const c1,
 244                                                    const uint32_t* const c2) {
 245   const uint8x8_t p0 = vreinterpret_u8_u64(vcreate_u64(*c0));
 246   const uint8x8_t p1 = vreinterpret_u8_u64(vcreate_u64(*c1));
 247   const uint8x8_t p2 = vreinterpret_u8_u64(vcreate_u64(*c2));
 248   const uint8x8_t avg = vhadd_u8(p0, p1);                  // Average(c0,c1)
 249   const uint8x8_t ab = vshr_n_u8(vqsub_u8(avg, p2), 1);    // (a-b)>>1 saturated
 250   const uint8x8_t ba = vshr_n_u8(vqsub_u8(p2, avg), 1);    // (b-a)>>1 saturated
 251   const uint8x8_t out = vqsub_u8(vqadd_u8(avg, ab), ba);
 252   return vget_lane_u32(vreinterpret_u32_u8(out), 0);
 253 }
 254
 255 static uint32_t Predictor13(uint32_t left, const uint32_t* const top) {
 256   return ClampedAddSubtractHalf(&left, top + 0, top - 1);
 257 }
 258
 259 //------------------------------------------------------------------------------
 260 // Subtract-Green Transform
 261
 262 // vtbl? are unavailable in iOS/arm64 builds.
 263 #if !defined(__aarch64__)
 264
 265 // 255 = byte will be zero'd
 266 static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255  };
 267
 268 static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
 269   const uint32_t* const end = argb_data + (num_pixels & ~3);
 270   const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
 271   for (; argb_data < end; argb_data += 4) {
 272     const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
 273     const uint8x16_t greens =
 274         vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
 275                     vtbl1_u8(vget_high_u8(argb), shuffle));
 276     vst1q_u8((uint8_t*)argb_data, vsubq_u8(argb, greens));
 277   }
 278   // fallthrough and finish off with plain-C
 279   VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3);
 280 }
 281
 282 static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) {
 283   const uint32_t* const end = argb_data + (num_pixels & ~3);
 284   const uint8x8_t shuffle = vld1_u8(kGreenShuffle);
 285   for (; argb_data < end; argb_data += 4) {
 286     const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data);
 287     const uint8x16_t greens =
 288         vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle),
 289                     vtbl1_u8(vget_high_u8(argb), shuffle));
 290     vst1q_u8((uint8_t*)argb_data, vaddq_u8(argb, greens));
 291   }
 292   // fallthrough and finish off with plain-C
 293   VP8LAddGreenToBlueAndRed_C(argb_data, num_pixels & 3);
 294 }
 295
 296 #endif   // !__aarch64__
 297
 298 #endif   // USE_INTRINSICS
 299
 300 #endif   // WEBP_USE_NEON
 301
 302 //------------------------------------------------------------------------------
 303
 304 extern void VP8LDspInitNEON(void);
 305
 306 void VP8LDspInitNEON(void) {
 307 #if defined(WEBP_USE_NEON)
 308   VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA;
 309   VP8LConvertBGRAToBGR = ConvertBGRAToBGR;
 310   VP8LConvertBGRAToRGB = ConvertBGRAToRGB;
 311
 312 #ifdef USE_INTRINSICS
 313   VP8LPredictors[5] = Predictor5;
 314   VP8LPredictors[6] = Predictor6;
 315   VP8LPredictors[7] = Predictor7;
 316   VP8LPredictors[8] = Predictor8;
 317   VP8LPredictors[9] = Predictor9;
 318   VP8LPredictors[10] = Predictor10;
 319   VP8LPredictors[11] = Predictor11;
 320   VP8LPredictors[12] = Predictor12;
 321   VP8LPredictors[13] = Predictor13;
 322
 323 #if !defined(__aarch64__)
 324   VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
 325   VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed;
 326 #endif
 327 #endif
 328
 329 #endif   // WEBP_USE_NEON
 330 }
 331
 332 //------------------------------------------------------------------------------