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 (at) 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 // Subtract-Green Transform 144 145 // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use 146 // non-standard versions there. 147 #if defined(__APPLE__) && defined(__aarch64__) && \ 148 defined(__apple_build_version__) && (__apple_build_version__< 6020037) 149 #define USE_VTBLQ 150 #endif 151 152 #ifdef USE_VTBLQ 153 // 255 = byte will be zeroed 154 static const uint8_t kGreenShuffle[16] = { 155 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13, 255 156 }; 157 158 static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb, 159 const uint8x16_t shuffle) { 160 return vcombine_u8(vtbl1q_u8(argb, vget_low_u8(shuffle)), 161 vtbl1q_u8(argb, vget_high_u8(shuffle))); 162 } 163 #else // !USE_VTBLQ 164 // 255 = byte will be zeroed 165 static const uint8_t kGreenShuffle[8] = { 1, 255, 1, 255, 5, 255, 5, 255 }; 166 167 static WEBP_INLINE uint8x16_t DoGreenShuffle(const uint8x16_t argb, 168 const uint8x8_t shuffle) { 169 return vcombine_u8(vtbl1_u8(vget_low_u8(argb), shuffle), 170 vtbl1_u8(vget_high_u8(argb), shuffle)); 171 } 172 #endif // USE_VTBLQ 173 174 static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) { 175 const uint32_t* const end = argb_data + (num_pixels & ~3); 176 #ifdef USE_VTBLQ 177 const uint8x16_t shuffle = vld1q_u8(kGreenShuffle); 178 #else 179 const uint8x8_t shuffle = vld1_u8(kGreenShuffle); 180 #endif 181 for (; argb_data < end; argb_data += 4) { 182 const uint8x16_t argb = vld1q_u8((uint8_t*)argb_data); 183 const uint8x16_t greens = DoGreenShuffle(argb, shuffle); 184 vst1q_u8((uint8_t*)argb_data, vaddq_u8(argb, greens)); 185 } 186 // fallthrough and finish off with plain-C 187 VP8LAddGreenToBlueAndRed_C(argb_data, num_pixels & 3); 188 } 189 190 //------------------------------------------------------------------------------ 191 // Color Transform 192 193 static void TransformColorInverse(const VP8LMultipliers* const m, 194 uint32_t* argb_data, int num_pixels) { 195 // sign-extended multiplying constants, pre-shifted by 6. 196 #define CST(X) (((int16_t)(m->X << 8)) >> 6) 197 const int16_t rb[8] = { 198 CST(green_to_blue_), CST(green_to_red_), 199 CST(green_to_blue_), CST(green_to_red_), 200 CST(green_to_blue_), CST(green_to_red_), 201 CST(green_to_blue_), CST(green_to_red_) 202 }; 203 const int16x8_t mults_rb = vld1q_s16(rb); 204 const int16_t b2[8] = { 205 0, CST(red_to_blue_), 0, CST(red_to_blue_), 206 0, CST(red_to_blue_), 0, CST(red_to_blue_), 207 }; 208 const int16x8_t mults_b2 = vld1q_s16(b2); 209 #undef CST 210 #ifdef USE_VTBLQ 211 static const uint8_t kg0g0[16] = { 212 255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13 213 }; 214 const uint8x16_t shuffle = vld1q_u8(kg0g0); 215 #else 216 static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 }; 217 const uint8x8_t shuffle = vld1_u8(k0g0g); 218 #endif 219 const uint32x4_t mask_ag = vdupq_n_u32(0xff00ff00u); 220 int i; 221 for (i = 0; i + 4 <= num_pixels; i += 4) { 222 const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i)); 223 const uint32x4_t a0g0 = vandq_u32(vreinterpretq_u32_u8(in), mask_ag); 224 // 0 g 0 g 225 const uint8x16_t greens = DoGreenShuffle(in, shuffle); 226 // x dr x db1 227 const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb); 228 // x r' x b' 229 const int8x16_t B = vaddq_s8(vreinterpretq_s8_u8(in), 230 vreinterpretq_s8_s16(A)); 231 // r' 0 b' 0 232 const int16x8_t C = vshlq_n_s16(vreinterpretq_s16_s8(B), 8); 233 // x db2 0 0 234 const int16x8_t D = vqdmulhq_s16(C, mults_b2); 235 // 0 x db2 0 236 const uint32x4_t E = vshrq_n_u32(vreinterpretq_u32_s16(D), 8); 237 // r' x b'' 0 238 const int8x16_t F = vaddq_s8(vreinterpretq_s8_u32(E), 239 vreinterpretq_s8_s16(C)); 240 // 0 r' 0 b'' 241 const uint16x8_t G = vshrq_n_u16(vreinterpretq_u16_s8(F), 8); 242 const uint32x4_t out = vorrq_u32(vreinterpretq_u32_u16(G), a0g0); 243 vst1q_u32(argb_data + i, out); 244 } 245 // Fall-back to C-version for left-overs. 246 VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i); 247 } 248 249 #undef USE_VTBLQ 250 251 //------------------------------------------------------------------------------ 252 // Entry point 253 254 extern void VP8LDspInitNEON(void); 255 256 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitNEON(void) { 257 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; 258 VP8LConvertBGRAToBGR = ConvertBGRAToBGR; 259 VP8LConvertBGRAToRGB = ConvertBGRAToRGB; 260 261 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; 262 VP8LTransformColorInverse = TransformColorInverse; 263 } 264 265 #else // !WEBP_USE_NEON 266 267 WEBP_DSP_INIT_STUB(VP8LDspInitNEON) 268 269 #endif // WEBP_USE_NEON 270
