filters_neon.c (11048B)
1 // Copyright 2017 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 alpha filters 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 <assert.h> 19 #include "src/dsp/neon.h" 20 21 //------------------------------------------------------------------------------ 22 // Helpful macros. 23 24 #define DCHECK(in, out) \ 25 do { \ 26 assert((in) != NULL); \ 27 assert((out) != NULL); \ 28 assert((in) != (out)); \ 29 assert(width > 0); \ 30 assert(height > 0); \ 31 assert(stride >= width); \ 32 } while (0) 33 34 // load eight u8 and widen to s16 35 #define U8_TO_S16(A) vreinterpretq_s16_u16(vmovl_u8(A)) 36 #define LOAD_U8_TO_S16(A) U8_TO_S16(vld1_u8(A)) 37 38 // shift left or right by N byte, inserting zeros 39 #define SHIFT_RIGHT_N_Q(A, N) vextq_u8((A), zero, (N)) 40 #define SHIFT_LEFT_N_Q(A, N) vextq_u8(zero, (A), (16 - (N)) % 16) 41 42 // rotate left by N bytes 43 #define ROTATE_LEFT_N(A, N) vext_u8((A), (A), (N)) 44 // rotate right by N bytes 45 #define ROTATE_RIGHT_N(A, N) vext_u8((A), (A), (8 - (N)) % 8) 46 47 static void PredictLine_NEON(const uint8_t* src, const uint8_t* pred, 48 uint8_t* WEBP_RESTRICT dst, int length) { 49 int i; 50 assert(length >= 0); 51 for (i = 0; i + 16 <= length; i += 16) { 52 const uint8x16_t A = vld1q_u8(&src[i]); 53 const uint8x16_t B = vld1q_u8(&pred[i]); 54 const uint8x16_t C = vsubq_u8(A, B); 55 vst1q_u8(&dst[i], C); 56 } 57 for (; i < length; ++i) dst[i] = src[i] - pred[i]; 58 } 59 60 // Special case for left-based prediction (when preds==dst-1 or preds==src-1). 61 static void PredictLineLeft_NEON(const uint8_t* WEBP_RESTRICT src, 62 uint8_t* WEBP_RESTRICT dst, int length) { 63 PredictLine_NEON(src, src - 1, dst, length); 64 } 65 66 //------------------------------------------------------------------------------ 67 // Horizontal filter. 68 69 static WEBP_INLINE void DoHorizontalFilter_NEON( 70 const uint8_t* WEBP_RESTRICT in, int width, int height, int stride, 71 uint8_t* WEBP_RESTRICT out) { 72 int row; 73 DCHECK(in, out); 74 75 // Leftmost pixel is the same as input for topmost scanline. 76 out[0] = in[0]; 77 PredictLineLeft_NEON(in + 1, out + 1, width - 1); 78 in += stride; 79 out += stride; 80 81 // Filter line-by-line. 82 for (row = 1; row < height; ++row) { 83 // Leftmost pixel is predicted from above. 84 out[0] = in[0] - in[-stride]; 85 PredictLineLeft_NEON(in + 1, out + 1, width - 1); 86 in += stride; 87 out += stride; 88 } 89 } 90 91 static void HorizontalFilter_NEON(const uint8_t* WEBP_RESTRICT data, 92 int width, int height, int stride, 93 uint8_t* WEBP_RESTRICT filtered_data) { 94 DoHorizontalFilter_NEON(data, width, height, stride, filtered_data); 95 } 96 97 //------------------------------------------------------------------------------ 98 // Vertical filter. 99 100 static WEBP_INLINE void DoVerticalFilter_NEON(const uint8_t* WEBP_RESTRICT in, 101 int width, int height, int stride, 102 uint8_t* WEBP_RESTRICT out) { 103 int row; 104 DCHECK(in, out); 105 106 // Very first top-left pixel is copied. 107 out[0] = in[0]; 108 // Rest of top scan-line is left-predicted. 109 PredictLineLeft_NEON(in + 1, out + 1, width - 1); 110 in += stride; 111 out += stride; 112 113 // Filter line-by-line. 114 for (row = 1; row < height; ++row) { 115 PredictLine_NEON(in, in - stride, out, width); 116 in += stride; 117 out += stride; 118 } 119 } 120 121 static void VerticalFilter_NEON(const uint8_t* WEBP_RESTRICT data, 122 int width, int height, int stride, 123 uint8_t* WEBP_RESTRICT filtered_data) { 124 DoVerticalFilter_NEON(data, width, height, stride, filtered_data); 125 } 126 127 //------------------------------------------------------------------------------ 128 // Gradient filter. 129 130 static WEBP_INLINE int GradientPredictor_C(uint8_t a, uint8_t b, uint8_t c) { 131 const int g = a + b - c; 132 return ((g & ~0xff) == 0) ? g : (g < 0) ? 0 : 255; // clip to 8bit 133 } 134 135 static void GradientPredictDirect_NEON(const uint8_t* const row, 136 const uint8_t* const top, 137 uint8_t* WEBP_RESTRICT const out, 138 int length) { 139 int i; 140 for (i = 0; i + 8 <= length; i += 8) { 141 const uint8x8_t A = vld1_u8(&row[i - 1]); 142 const uint8x8_t B = vld1_u8(&top[i + 0]); 143 const int16x8_t C = vreinterpretq_s16_u16(vaddl_u8(A, B)); 144 const int16x8_t D = LOAD_U8_TO_S16(&top[i - 1]); 145 const uint8x8_t E = vqmovun_s16(vsubq_s16(C, D)); 146 const uint8x8_t F = vld1_u8(&row[i + 0]); 147 vst1_u8(&out[i], vsub_u8(F, E)); 148 } 149 for (; i < length; ++i) { 150 out[i] = row[i] - GradientPredictor_C(row[i - 1], top[i], top[i - 1]); 151 } 152 } 153 154 static WEBP_INLINE void DoGradientFilter_NEON(const uint8_t* WEBP_RESTRICT in, 155 int width, int height, int stride, 156 uint8_t* WEBP_RESTRICT out) { 157 int row; 158 DCHECK(in, out); 159 160 // left prediction for top scan-line 161 out[0] = in[0]; 162 PredictLineLeft_NEON(in + 1, out + 1, width - 1); 163 in += stride; 164 out += stride; 165 166 // Filter line-by-line. 167 for (row = 1; row < height; ++row) { 168 out[0] = in[0] - in[-stride]; 169 GradientPredictDirect_NEON(in + 1, in + 1 - stride, out + 1, width - 1); 170 in += stride; 171 out += stride; 172 } 173 } 174 175 static void GradientFilter_NEON(const uint8_t* WEBP_RESTRICT data, 176 int width, int height, int stride, 177 uint8_t* WEBP_RESTRICT filtered_data) { 178 DoGradientFilter_NEON(data, width, height, stride, filtered_data); 179 } 180 181 #undef DCHECK 182 183 //------------------------------------------------------------------------------ 184 // Inverse transforms 185 186 static void HorizontalUnfilter_NEON(const uint8_t* prev, const uint8_t* in, 187 uint8_t* out, int width) { 188 int i; 189 const uint8x16_t zero = vdupq_n_u8(0); 190 uint8x16_t last; 191 out[0] = in[0] + (prev == NULL ? 0 : prev[0]); 192 if (width <= 1) return; 193 last = vsetq_lane_u8(out[0], zero, 0); 194 for (i = 1; i + 16 <= width; i += 16) { 195 const uint8x16_t A0 = vld1q_u8(&in[i]); 196 const uint8x16_t A1 = vaddq_u8(A0, last); 197 const uint8x16_t A2 = SHIFT_LEFT_N_Q(A1, 1); 198 const uint8x16_t A3 = vaddq_u8(A1, A2); 199 const uint8x16_t A4 = SHIFT_LEFT_N_Q(A3, 2); 200 const uint8x16_t A5 = vaddq_u8(A3, A4); 201 const uint8x16_t A6 = SHIFT_LEFT_N_Q(A5, 4); 202 const uint8x16_t A7 = vaddq_u8(A5, A6); 203 const uint8x16_t A8 = SHIFT_LEFT_N_Q(A7, 8); 204 const uint8x16_t A9 = vaddq_u8(A7, A8); 205 vst1q_u8(&out[i], A9); 206 last = SHIFT_RIGHT_N_Q(A9, 15); 207 } 208 for (; i < width; ++i) out[i] = in[i] + out[i - 1]; 209 } 210 211 static void VerticalUnfilter_NEON(const uint8_t* prev, const uint8_t* in, 212 uint8_t* out, int width) { 213 if (prev == NULL) { 214 HorizontalUnfilter_NEON(NULL, in, out, width); 215 } else { 216 int i; 217 assert(width >= 0); 218 for (i = 0; i + 16 <= width; i += 16) { 219 const uint8x16_t A = vld1q_u8(&in[i]); 220 const uint8x16_t B = vld1q_u8(&prev[i]); 221 const uint8x16_t C = vaddq_u8(A, B); 222 vst1q_u8(&out[i], C); 223 } 224 for (; i < width; ++i) out[i] = in[i] + prev[i]; 225 } 226 } 227 228 // GradientUnfilter_NEON is correct but slower than the C-version, 229 // at least on ARM64. For armv7, it's a wash. 230 // So best is to disable it for now, but keep the idea around... 231 #if !defined(USE_GRADIENT_UNFILTER) 232 #define USE_GRADIENT_UNFILTER 0 // ALTERNATE_CODE 233 #endif 234 235 #if (USE_GRADIENT_UNFILTER == 1) 236 #define GRAD_PROCESS_LANE(L) do { \ 237 const uint8x8_t tmp1 = ROTATE_RIGHT_N(pred, 1); /* rotate predictor in */ \ 238 const int16x8_t tmp2 = vaddq_s16(BC, U8_TO_S16(tmp1)); \ 239 const uint8x8_t delta = vqmovun_s16(tmp2); \ 240 pred = vadd_u8(D, delta); \ 241 out = vext_u8(out, ROTATE_LEFT_N(pred, (L)), 1); \ 242 } while (0) 243 244 static void GradientPredictInverse_NEON(const uint8_t* const in, 245 const uint8_t* const top, 246 uint8_t* const row, int length) { 247 if (length > 0) { 248 int i; 249 uint8x8_t pred = vdup_n_u8(row[-1]); // left sample 250 uint8x8_t out = vdup_n_u8(0); 251 for (i = 0; i + 8 <= length; i += 8) { 252 const int16x8_t B = LOAD_U8_TO_S16(&top[i + 0]); 253 const int16x8_t C = LOAD_U8_TO_S16(&top[i - 1]); 254 const int16x8_t BC = vsubq_s16(B, C); // unclipped gradient basis B - C 255 const uint8x8_t D = vld1_u8(&in[i]); // base input 256 GRAD_PROCESS_LANE(0); 257 GRAD_PROCESS_LANE(1); 258 GRAD_PROCESS_LANE(2); 259 GRAD_PROCESS_LANE(3); 260 GRAD_PROCESS_LANE(4); 261 GRAD_PROCESS_LANE(5); 262 GRAD_PROCESS_LANE(6); 263 GRAD_PROCESS_LANE(7); 264 vst1_u8(&row[i], out); 265 } 266 for (; i < length; ++i) { 267 row[i] = in[i] + GradientPredictor_C(row[i - 1], top[i], top[i - 1]); 268 } 269 } 270 } 271 #undef GRAD_PROCESS_LANE 272 273 static void GradientUnfilter_NEON(const uint8_t* prev, const uint8_t* in, 274 uint8_t* out, int width) { 275 if (prev == NULL) { 276 HorizontalUnfilter_NEON(NULL, in, out, width); 277 } else { 278 out[0] = in[0] + prev[0]; // predict from above 279 GradientPredictInverse_NEON(in + 1, prev + 1, out + 1, width - 1); 280 } 281 } 282 283 #endif // USE_GRADIENT_UNFILTER 284 285 //------------------------------------------------------------------------------ 286 // Entry point 287 288 extern void VP8FiltersInitNEON(void); 289 290 WEBP_TSAN_IGNORE_FUNCTION void VP8FiltersInitNEON(void) { 291 WebPUnfilters[WEBP_FILTER_HORIZONTAL] = HorizontalUnfilter_NEON; 292 WebPUnfilters[WEBP_FILTER_VERTICAL] = VerticalUnfilter_NEON; 293 #if (USE_GRADIENT_UNFILTER == 1) 294 WebPUnfilters[WEBP_FILTER_GRADIENT] = GradientUnfilter_NEON; 295 #endif 296 297 WebPFilters[WEBP_FILTER_HORIZONTAL] = HorizontalFilter_NEON; 298 WebPFilters[WEBP_FILTER_VERTICAL] = VerticalFilter_NEON; 299 WebPFilters[WEBP_FILTER_GRADIENT] = GradientFilter_NEON; 300 } 301 302 #else // !WEBP_USE_NEON 303 304 WEBP_DSP_INIT_STUB(VP8FiltersInitNEON) 305 306 #endif // WEBP_USE_NEON