src/third_party/libwebp/dsp/upsampling_neon.c

   1 // Copyright 2011 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 version of YUV to RGB upsampling functions.
  11 //
  12 // Author: mans@mansr.com (Mans Rullgard)
  13 // Based on SSE code by: somnath@google.com (Somnath Banerjee)
  14
  15 #include "./dsp.h"
  16
  17 #if defined(WEBP_USE_NEON)
  18
  19 #include <assert.h>
  20 #include <arm_neon.h>
  21 #include <string.h>
  22 #include "./neon.h"
  23 #include "./yuv.h"
  24
  25 #ifdef FANCY_UPSAMPLING
  26
  27 //-----------------------------------------------------------------------------
  28 // U/V upsampling
  29
  30 // Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
  31 #define UPSAMPLE_16PIXELS(r1, r2, out) {                                \
  32   uint8x8_t a = vld1_u8(r1);                                            \
  33   uint8x8_t b = vld1_u8(r1 + 1);                                        \
  34   uint8x8_t c = vld1_u8(r2);                                            \
  35   uint8x8_t d = vld1_u8(r2 + 1);                                        \
  36                                                                         \
  37   uint16x8_t al = vshll_n_u8(a, 1);                                     \
  38   uint16x8_t bl = vshll_n_u8(b, 1);                                     \
  39   uint16x8_t cl = vshll_n_u8(c, 1);                                     \
  40   uint16x8_t dl = vshll_n_u8(d, 1);                                     \
  41                                                                         \
  42   uint8x8_t diag1, diag2;                                               \
  43   uint16x8_t sl;                                                        \
  44                                                                         \
  45   /* a + b + c + d */                                                   \
  46   sl = vaddl_u8(a,  b);                                                 \
  47   sl = vaddw_u8(sl, c);                                                 \
  48   sl = vaddw_u8(sl, d);                                                 \
  49                                                                         \
  50   al = vaddq_u16(sl, al); /* 3a +  b +  c +  d */                       \
  51   bl = vaddq_u16(sl, bl); /*  a + 3b +  c +  d */                       \
  52                                                                         \
  53   al = vaddq_u16(al, dl); /* 3a +  b +  c + 3d */                       \
  54   bl = vaddq_u16(bl, cl); /*  a + 3b + 3c +  d */                       \
  55                                                                         \
  56   diag2 = vshrn_n_u16(al, 3);                                           \
  57   diag1 = vshrn_n_u16(bl, 3);                                           \
  58                                                                         \
  59   a = vrhadd_u8(a, diag1);                                              \
  60   b = vrhadd_u8(b, diag2);                                              \
  61   c = vrhadd_u8(c, diag2);                                              \
  62   d = vrhadd_u8(d, diag1);                                              \
  63                                                                         \
  64   {                                                                     \
  65     uint8x8x2_t a_b, c_d;                                               \
  66     INIT_VECTOR2(a_b, a, b);                                            \
  67     INIT_VECTOR2(c_d, c, d);                                            \
  68     vst2_u8(out,      a_b);                                             \
  69     vst2_u8(out + 32, c_d);                                             \
  70   }                                                                     \
  71 }
  72
  73 // Turn the macro into a function for reducing code-size when non-critical
  74 static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
  75                              uint8_t *out) {
  76   UPSAMPLE_16PIXELS(r1, r2, out);
  77 }
  78
  79 #define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                  \
  80   uint8_t r1[9], r2[9];                                                 \
  81   memcpy(r1, (tb), (num_pixels));                                       \
  82   memcpy(r2, (bb), (num_pixels));                                       \
  83   /* replicate last byte */                                             \
  84   memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels));    \
  85   memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels));    \
  86   Upsample16Pixels(r1, r2, out);                                        \
  87 }
  88
  89 //-----------------------------------------------------------------------------
  90 // YUV->RGB conversion
  91
  92 static const int16_t kCoeffs[4] = { kYScale, kVToR, kUToG, kVToG };
  93
  94 #define v255 vdup_n_u8(255)
  95
  96 #define STORE_Rgb(out, r, g, b) do {                                    \
  97   uint8x8x3_t r_g_b;                                                    \
  98   INIT_VECTOR3(r_g_b, r, g, b);                                         \
  99   vst3_u8(out, r_g_b);                                                  \
 100 } while (0)
 101
 102 #define STORE_Bgr(out, r, g, b) do {                                    \
 103   uint8x8x3_t b_g_r;                                                    \
 104   INIT_VECTOR3(b_g_r, b, g, r);                                         \
 105   vst3_u8(out, b_g_r);                                                  \
 106 } while (0)
 107
 108 #define STORE_Rgba(out, r, g, b) do {                                   \
 109   uint8x8x4_t r_g_b_v255;                                               \
 110   INIT_VECTOR4(r_g_b_v255, r, g, b, v255);                              \
 111   vst4_u8(out, r_g_b_v255);                                             \
 112 } while (0)
 113
 114 #define STORE_Bgra(out, r, g, b) do {                                   \
 115   uint8x8x4_t b_g_r_v255;                                               \
 116   INIT_VECTOR4(b_g_r_v255, b, g, r, v255);                              \
 117   vst4_u8(out, b_g_r_v255);                                             \
 118 } while (0)
 119
 120 #define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) {            \
 121   int i;                                                                \
 122   for (i = 0; i < N; i += 8) {                                          \
 123     const int off = ((cur_x) + i) * XSTEP;                              \
 124     uint8x8_t y  = vld1_u8((src_y) + (cur_x)  + i);                     \
 125     uint8x8_t u  = vld1_u8((src_uv) + i);                               \
 126     uint8x8_t v  = vld1_u8((src_uv) + i + 16);                          \
 127     const int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16));       \
 128     const int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128));      \
 129     const int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128));      \
 130     int32x4_t yl = vmull_lane_s16(vget_low_s16(yy),  cf16, 0);          \
 131     int32x4_t yh = vmull_lane_s16(vget_high_s16(yy), cf16, 0);          \
 132     const int32x4_t rl = vmlal_lane_s16(yl, vget_low_s16(vv),  cf16, 1);\
 133     const int32x4_t rh = vmlal_lane_s16(yh, vget_high_s16(vv), cf16, 1);\
 134     int32x4_t gl = vmlsl_lane_s16(yl, vget_low_s16(uu),  cf16, 2);      \
 135     int32x4_t gh = vmlsl_lane_s16(yh, vget_high_s16(uu), cf16, 2);      \
 136     const int32x4_t bl = vmovl_s16(vget_low_s16(uu));                   \
 137     const int32x4_t bh = vmovl_s16(vget_high_s16(uu));                  \
 138     gl = vmlsl_lane_s16(gl, vget_low_s16(vv),  cf16, 3);                \
 139     gh = vmlsl_lane_s16(gh, vget_high_s16(vv), cf16, 3);                \
 140     yl = vmlaq_lane_s32(yl, bl, cf32, 0);                               \
 141     yh = vmlaq_lane_s32(yh, bh, cf32, 0);                               \
 142     /* vrshrn_n_s32() already incorporates the rounding constant */     \
 143     y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, YUV_FIX2),            \
 144                                  vrshrn_n_s32(rh, YUV_FIX2)));          \
 145     u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, YUV_FIX2),            \
 146                                  vrshrn_n_s32(gh, YUV_FIX2)));          \
 147     v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(yl, YUV_FIX2),            \
 148                                  vrshrn_n_s32(yh, YUV_FIX2)));          \
 149     STORE_ ## FMT(out + off, y, u, v);                                  \
 150   }                                                                     \
 151 }
 152
 153 #define CONVERT1(FUNC, XSTEP, N, src_y, src_uv, rgb, cur_x) {           \
 154   int i;                                                                \
 155   for (i = 0; i < N; i++) {                                             \
 156     const int off = ((cur_x) + i) * XSTEP;                              \
 157     const int y = src_y[(cur_x) + i];                                   \
 158     const int u = (src_uv)[i];                                          \
 159     const int v = (src_uv)[i + 16];                                     \
 160     FUNC(y, u, v, rgb + off);                                           \
 161   }                                                                     \
 162 }
 163
 164 #define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv,                  \
 165                       top_dst, bottom_dst, cur_x, len) {                \
 166   CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x)                  \
 167   if (bottom_y != NULL) {                                               \
 168     CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x)   \
 169   }                                                                     \
 170 }
 171
 172 #define CONVERT2RGB_1(FUNC, XSTEP, top_y, bottom_y, uv,                 \
 173                       top_dst, bottom_dst, cur_x, len) {                \
 174   CONVERT1(FUNC, XSTEP, len, top_y, uv, top_dst, cur_x);                \
 175   if (bottom_y != NULL) {                                               \
 176     CONVERT1(FUNC, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \
 177   }                                                                     \
 178 }
 179
 180 #define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP)                       \
 181 static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y,    \
 182                       const uint8_t *top_u, const uint8_t *top_v,       \
 183                       const uint8_t *cur_u, const uint8_t *cur_v,       \
 184                       uint8_t *top_dst, uint8_t *bottom_dst, int len) { \
 185   int block;                                                            \
 186   /* 16 byte aligned array to cache reconstructed u and v */            \
 187   uint8_t uv_buf[2 * 32 + 15];                                          \
 188   uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);     \
 189   const int uv_len = (len + 1) >> 1;                                    \
 190   /* 9 pixels must be read-able for each block */                       \
 191   const int num_blocks = (uv_len - 1) >> 3;                             \
 192   const int leftover = uv_len - num_blocks * 8;                         \
 193   const int last_pos = 1 + 16 * num_blocks;                             \
 194                                                                         \
 195   const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                  \
 196   const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                  \
 197                                                                         \
 198   const int16x4_t cf16 = vld1_s16(kCoeffs);                             \
 199   const int32x2_t cf32 = vdup_n_s32(kUToB);                             \
 200   const uint8x8_t u16  = vdup_n_u8(16);                                 \
 201   const uint8x8_t u128 = vdup_n_u8(128);                                \
 202                                                                         \
 203   /* Treat the first pixel in regular way */                            \
 204   assert(top_y != NULL);                                                \
 205   {                                                                     \
 206     const int u0 = (top_u[0] + u_diag) >> 1;                            \
 207     const int v0 = (top_v[0] + v_diag) >> 1;                            \
 208     VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst);                         \
 209   }                                                                     \
 210   if (bottom_y != NULL) {                                               \
 211     const int u0 = (cur_u[0] + u_diag) >> 1;                            \
 212     const int v0 = (cur_v[0] + v_diag) >> 1;                            \
 213     VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst);                   \
 214   }                                                                     \
 215                                                                         \
 216   for (block = 0; block < num_blocks; ++block) {                        \
 217     UPSAMPLE_16PIXELS(top_u, cur_u, r_uv);                              \
 218     UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16);                         \
 219     CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv,                    \
 220                   top_dst, bottom_dst, 16 * block + 1, 16);             \
 221     top_u += 8;                                                         \
 222     cur_u += 8;                                                         \
 223     top_v += 8;                                                         \
 224     cur_v += 8;                                                         \
 225   }                                                                     \
 226                                                                         \
 227   UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv);                    \
 228   UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16);               \
 229   CONVERT2RGB_1(VP8YuvTo ## FMT, XSTEP, top_y, bottom_y, r_uv,          \
 230                 top_dst, bottom_dst, last_pos, len - last_pos);         \
 231 }
 232
 233 // NEON variants of the fancy upsampler.
 234 NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair,  Rgb,  3)
 235 NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair,  Bgr,  3)
 236 NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePair, Rgba, 4)
 237 NEON_UPSAMPLE_FUNC(UpsampleBgraLinePair, Bgra, 4)
 238
 239 #endif  // FANCY_UPSAMPLING
 240
 241 #endif   // WEBP_USE_NEON
 242
 243 //------------------------------------------------------------------------------
 244
 245 extern void WebPInitUpsamplersNEON(void);
 246
 247 #ifdef FANCY_UPSAMPLING
 248
 249 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 250
 251 void WebPInitUpsamplersNEON(void) {
 252 #if defined(WEBP_USE_NEON)
 253   WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair;
 254   WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
 255   WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair;
 256   WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
 257   WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
 258   WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
 259 #endif   // WEBP_USE_NEON
 260 }
 261
 262 #else
 263
 264 // this empty function is to avoid an empty .o
 265 void WebPInitUpsamplersNEON(void) {}
 266
 267 #endif  // FANCY_UPSAMPLING