rfreeimage 0.1.0

data/vendor/FreeImage/Source/LibWebP/src/dsp/dsp.upsampling_mips_dsp_r2.c

@@ -0,0 +1,280 @@
+// Copyright 2014 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// YUV to RGB upsampling functions.
+//
+// Author(s): Branimir Vasic (branimir.vasic@imgtec.com)
+//            Djordje Pesut  (djordje.pesut@imgtec.com)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_MIPS_DSP_R2)
+
+#include "./yuv.h"
+#include <assert.h>
+
+#if !defined(WEBP_YUV_USE_TABLE)
+
+#define YUV_TO_RGB(Y, U, V, R, G, B) do {                                      \
+    const int t1 = kYScale * Y;                                                \
+    const int t2 = kVToG * V;                                                  \
+    R = kVToR * V;                                                             \
+    G = kUToG * U;                                                             \
+    B = kUToB * U;                                                             \
+    R = t1 + R;                                                                \
+    G = t1 - G;                                                                \
+    B = t1 + B;                                                                \
+    R = R + kRCst;                                                             \
+    G = G - t2 + kGCst;                                                        \
+    B = B + kBCst;                                                             \
+    __asm__ volatile (                                                         \
+      "shll_s.w         %["#R"],      %["#R"],        9              \n\t"     \
+      "shll_s.w         %["#G"],      %["#G"],        9              \n\t"     \
+      "shll_s.w         %["#B"],      %["#B"],        9              \n\t"     \
+      "precrqu_s.qb.ph  %["#R"],      %["#R"],        $zero          \n\t"     \
+      "precrqu_s.qb.ph  %["#G"],      %["#G"],        $zero          \n\t"     \
+      "precrqu_s.qb.ph  %["#B"],      %["#B"],        $zero          \n\t"     \
+      "srl              %["#R"],      %["#R"],        24             \n\t"     \
+      "srl              %["#G"],      %["#G"],        24             \n\t"     \
+      "srl              %["#B"],      %["#B"],        24             \n\t"     \
+      : [R]"+r"(R), [G]"+r"(G), [B]"+r"(B)                                     \
+      :                                                                        \
+    );                                                                         \
+  } while (0)
+
+static WEBP_INLINE void YuvToRgb(int y, int u, int v, uint8_t* const rgb) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  rgb[0] = r;
+  rgb[1] = g;
+  rgb[2] = b;
+}
+static WEBP_INLINE void YuvToBgr(int y, int u, int v, uint8_t* const bgr) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  bgr[0] = b;
+  bgr[1] = g;
+  bgr[2] = r;
+}
+static WEBP_INLINE void YuvToRgb565(int y, int u, int v, uint8_t* const rgb) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  {
+    const int rg = (r & 0xf8) | (g >> 5);
+    const int gb = ((g << 3) & 0xe0) | (b >> 3);
+#ifdef WEBP_SWAP_16BIT_CSP
+    rgb[0] = gb;
+    rgb[1] = rg;
+#else
+    rgb[0] = rg;
+    rgb[1] = gb;
+#endif
+  }
+}
+static WEBP_INLINE void YuvToRgba4444(int y, int u, int v,
+                                      uint8_t* const argb) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  {
+    const int rg = (r & 0xf0) | (g >> 4);
+    const int ba = (b & 0xf0) | 0x0f;     // overwrite the lower 4 bits
+#ifdef WEBP_SWAP_16BIT_CSP
+    argb[0] = ba;
+    argb[1] = rg;
+#else
+    argb[0] = rg;
+    argb[1] = ba;
+#endif
+   }
+}
+#endif  // WEBP_YUV_USE_TABLE
+
+//-----------------------------------------------------------------------------
+// Alpha handling variants
+
+static WEBP_INLINE void YuvToArgb(uint8_t y, uint8_t u, uint8_t v,
+                                  uint8_t* const argb) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  argb[0] = 0xff;
+  argb[1] = r;
+  argb[2] = g;
+  argb[3] = b;
+}
+static WEBP_INLINE void YuvToBgra(uint8_t y, uint8_t u, uint8_t v,
+                                  uint8_t* const bgra) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  bgra[0] = b;
+  bgra[1] = g;
+  bgra[2] = r;
+  bgra[3] = 0xff;
+}
+static WEBP_INLINE void YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
+                                  uint8_t* const rgba) {
+  int r, g, b;
+  YUV_TO_RGB(y, u, v, r, g, b);
+  rgba[0] = r;
+  rgba[1] = g;
+  rgba[2] = b;
+  rgba[3] = 0xff;
+}
+
+//------------------------------------------------------------------------------
+// Fancy upsampler
+
+#ifdef FANCY_UPSAMPLING
+
+// Given samples laid out in a square as:
+//  [a b]
+//  [c d]
+// we interpolate u/v as:
+//  ([9*a + 3*b + 3*c +   d    3*a + 9*b + 3*c +   d] + [8 8]) / 16
+//  ([3*a +   b + 9*c + 3*d      a + 3*b + 3*c + 9*d]   [8 8]) / 16
+
+// We process u and v together stashed into 32bit (16bit each).
+#define LOAD_UV(u, v) ((u) | ((v) << 16))
+
+#define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                                  \
+static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
+                      const uint8_t* top_u, const uint8_t* top_v,              \
+                      const uint8_t* cur_u, const uint8_t* cur_v,              \
+                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
+  int x;                                                                       \
+  const int last_pixel_pair = (len - 1) >> 1;                                  \
+  uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]);   /* top-left sample */        \
+  uint32_t l_uv  = LOAD_UV(cur_u[0], cur_v[0]);   /* left-sample */            \
+  assert(top_y != NULL);                                                       \
+  {                                                                            \
+    const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2;                \
+    FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst);                          \
+  }                                                                            \
+  if (bottom_y != NULL) {                                                      \
+    const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2;                \
+    FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst);                    \
+  }                                                                            \
+  for (x = 1; x <= last_pixel_pair; ++x) {                                     \
+    const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]);  /* top sample */       \
+    const uint32_t uv   = LOAD_UV(cur_u[x], cur_v[x]);  /* sample */           \
+    /* precompute invariant values associated with first and second diagonals*/\
+    const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u;               \
+    const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3;                   \
+    const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3;                    \
+    {                                                                          \
+      const uint32_t uv0 = (diag_12 + tl_uv) >> 1;                             \
+      const uint32_t uv1 = (diag_03 + t_uv) >> 1;                              \
+      FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16),                          \
+           top_dst + (2 * x - 1) * XSTEP);                                     \
+      FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16),                          \
+           top_dst + (2 * x - 0) * XSTEP);                                     \
+    }                                                                          \
+    if (bottom_y != NULL) {                                                    \
+      const uint32_t uv0 = (diag_03 + l_uv) >> 1;                              \
+      const uint32_t uv1 = (diag_12 + uv) >> 1;                                \
+      FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16),                       \
+           bottom_dst + (2 * x - 1) * XSTEP);                                  \
+      FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16),                       \
+           bottom_dst + (2 * x + 0) * XSTEP);                                  \
+    }                                                                          \
+    tl_uv = t_uv;                                                              \
+    l_uv = uv;                                                                 \
+  }                                                                            \
+  if (!(len & 1)) {                                                            \
+    {                                                                          \
+      const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2;              \
+      FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16),                            \
+           top_dst + (len - 1) * XSTEP);                                       \
+    }                                                                          \
+    if (bottom_y != NULL) {                                                    \
+      const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2;              \
+      FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16),                         \
+           bottom_dst + (len - 1) * XSTEP);                                    \
+    }                                                                          \
+  }                                                                            \
+}
+
+// All variants implemented.
+UPSAMPLE_FUNC(UpsampleRgbLinePair,      YuvToRgb,      3)
+UPSAMPLE_FUNC(UpsampleBgrLinePair,      YuvToBgr,      3)
+UPSAMPLE_FUNC(UpsampleRgbaLinePair,     YuvToRgba,     4)
+UPSAMPLE_FUNC(UpsampleBgraLinePair,     YuvToBgra,     4)
+UPSAMPLE_FUNC(UpsampleArgbLinePair,     YuvToArgb,     4)
+UPSAMPLE_FUNC(UpsampleRgba4444LinePair, YuvToRgba4444, 2)
+UPSAMPLE_FUNC(UpsampleRgb565LinePair,   YuvToRgb565,   2)
+
+#undef LOAD_UV
+#undef UPSAMPLE_FUNC
+
+#endif  // FANCY_UPSAMPLING
+
+#endif  // WEBP_USE_MIPS_DSP_R2
+
+extern void WebPInitUpsamplersMIPSdspR2(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersMIPSdspR2(void) {
+#if defined(WEBP_USE_MIPS_DSP_R2)
+#ifdef FANCY_UPSAMPLING
+  WebPUpsamplers[MODE_RGB]       = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_RGBA]      = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_BGR]       = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_BGRA]      = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_ARGB]      = UpsampleArgbLinePair;
+  WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair;
+  WebPUpsamplers[MODE_RGB_565]   = UpsampleRgb565LinePair;
+  WebPUpsamplers[MODE_rgbA]      = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA]      = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_Argb]      = UpsampleArgbLinePair;
+  WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair;
+#endif  // FANCY_UPSAMPLING
+#endif  // WEBP_USE_MIPS_DSP_R2
+}
+
+//------------------------------------------------------------------------------
+// YUV444 converter
+
+#if defined(WEBP_USE_MIPS_DSP_R2)
+
+#define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP)                                    \
+static void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v,    \
+                      uint8_t* dst, int len) {                                 \
+  int i;                                                                       \
+  for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * XSTEP]);           \
+}
+
+YUV444_FUNC(Yuv444ToRgb,      YuvToRgb,      3)
+YUV444_FUNC(Yuv444ToBgr,      YuvToBgr,      3)
+YUV444_FUNC(Yuv444ToRgba,     YuvToRgba,     4)
+YUV444_FUNC(Yuv444ToBgra,     YuvToBgra,     4)
+YUV444_FUNC(Yuv444ToArgb,     YuvToArgb,     4)
+YUV444_FUNC(Yuv444ToRgba4444, YuvToRgba4444, 2)
+YUV444_FUNC(Yuv444ToRgb565,   YuvToRgb565,   2)
+
+#undef YUV444_FUNC
+
+#endif  // WEBP_USE_MIPS_DSP_R2
+
+extern void WebPInitYUV444ConvertersMIPSdspR2(void);
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitYUV444ConvertersMIPSdspR2(void) {
+#if defined(WEBP_USE_MIPS_DSP_R2)
+  WebPYUV444Converters[MODE_RGB]       = Yuv444ToRgb;
+  WebPYUV444Converters[MODE_RGBA]      = Yuv444ToRgba;
+  WebPYUV444Converters[MODE_BGR]       = Yuv444ToBgr;
+  WebPYUV444Converters[MODE_BGRA]      = Yuv444ToBgra;
+  WebPYUV444Converters[MODE_ARGB]      = Yuv444ToArgb;
+  WebPYUV444Converters[MODE_RGBA_4444] = Yuv444ToRgba4444;
+  WebPYUV444Converters[MODE_RGB_565]   = Yuv444ToRgb565;
+  WebPYUV444Converters[MODE_rgbA]      = Yuv444ToRgba;
+  WebPYUV444Converters[MODE_bgrA]      = Yuv444ToBgra;
+  WebPYUV444Converters[MODE_Argb]      = Yuv444ToArgb;
+  WebPYUV444Converters[MODE_rgbA_4444] = Yuv444ToRgba4444;
+#endif  // WEBP_USE_MIPS_DSP_R2
+}
+
+//------------------------------------------------------------------------------

data/vendor/FreeImage/Source/LibWebP/src/dsp/dsp.upsampling_neon.c

@@ -0,0 +1,267 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the COPYING file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS. All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+// -----------------------------------------------------------------------------
+//
+// NEON version of YUV to RGB upsampling functions.
+//
+// Author: mans@mansr.com (Mans Rullgard)
+// Based on SSE code by: somnath@google.com (Somnath Banerjee)
+
+#include "./dsp.h"
+
+#if defined(WEBP_USE_NEON)
+
+#include <assert.h>
+#include <arm_neon.h>
+#include <string.h>
+#include "./neon.h"
+#include "./yuv.h"
+
+#ifdef FANCY_UPSAMPLING
+
+//-----------------------------------------------------------------------------
+// U/V upsampling
+
+// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
+#define UPSAMPLE_16PIXELS(r1, r2, out) {                                \
+  uint8x8_t a = vld1_u8(r1);                                            \
+  uint8x8_t b = vld1_u8(r1 + 1);                                        \
+  uint8x8_t c = vld1_u8(r2);                                            \
+  uint8x8_t d = vld1_u8(r2 + 1);                                        \
+                                                                        \
+  uint16x8_t al = vshll_n_u8(a, 1);                                     \
+  uint16x8_t bl = vshll_n_u8(b, 1);                                     \
+  uint16x8_t cl = vshll_n_u8(c, 1);                                     \
+  uint16x8_t dl = vshll_n_u8(d, 1);                                     \
+                                                                        \
+  uint8x8_t diag1, diag2;                                               \
+  uint16x8_t sl;                                                        \
+                                                                        \
+  /* a + b + c + d */                                                   \
+  sl = vaddl_u8(a,  b);                                                 \
+  sl = vaddw_u8(sl, c);                                                 \
+  sl = vaddw_u8(sl, d);                                                 \
+                                                                        \
+  al = vaddq_u16(sl, al); /* 3a +  b +  c +  d */                       \
+  bl = vaddq_u16(sl, bl); /*  a + 3b +  c +  d */                       \
+                                                                        \
+  al = vaddq_u16(al, dl); /* 3a +  b +  c + 3d */                       \
+  bl = vaddq_u16(bl, cl); /*  a + 3b + 3c +  d */                       \
+                                                                        \
+  diag2 = vshrn_n_u16(al, 3);                                           \
+  diag1 = vshrn_n_u16(bl, 3);                                           \
+                                                                        \
+  a = vrhadd_u8(a, diag1);                                              \
+  b = vrhadd_u8(b, diag2);                                              \
+  c = vrhadd_u8(c, diag2);                                              \
+  d = vrhadd_u8(d, diag1);                                              \
+                                                                        \
+  {                                                                     \
+    uint8x8x2_t a_b, c_d;                                               \
+    INIT_VECTOR2(a_b, a, b);                                            \
+    INIT_VECTOR2(c_d, c, d);                                            \
+    vst2_u8(out,      a_b);                                             \
+    vst2_u8(out + 32, c_d);                                             \
+  }                                                                     \
+}
+
+// Turn the macro into a function for reducing code-size when non-critical
+static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
+                             uint8_t *out) {
+  UPSAMPLE_16PIXELS(r1, r2, out);
+}
+
+#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                  \
+  uint8_t r1[9], r2[9];                                                 \
+  memcpy(r1, (tb), (num_pixels));                                       \
+  memcpy(r2, (bb), (num_pixels));                                       \
+  /* replicate last byte */                                             \
+  memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels));    \
+  memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels));    \
+  Upsample16Pixels(r1, r2, out);                                        \
+}
+
+//-----------------------------------------------------------------------------
+// YUV->RGB conversion
+
+static const int16_t kCoeffs[4] = { kYScale, kVToR, kUToG, kVToG };
+
+#define v255 vdup_n_u8(255)
+
+#define STORE_Rgb(out, r, g, b) do {                                    \
+  uint8x8x3_t r_g_b;                                                    \
+  INIT_VECTOR3(r_g_b, r, g, b);                                         \
+  vst3_u8(out, r_g_b);                                                  \
+} while (0)
+
+#define STORE_Bgr(out, r, g, b) do {                                    \
+  uint8x8x3_t b_g_r;                                                    \
+  INIT_VECTOR3(b_g_r, b, g, r);                                         \
+  vst3_u8(out, b_g_r);                                                  \
+} while (0)
+
+#define STORE_Rgba(out, r, g, b) do {                                   \
+  uint8x8x4_t r_g_b_v255;                                               \
+  INIT_VECTOR4(r_g_b_v255, r, g, b, v255);                              \
+  vst4_u8(out, r_g_b_v255);                                             \
+} while (0)
+
+#define STORE_Bgra(out, r, g, b) do {                                   \
+  uint8x8x4_t b_g_r_v255;                                               \
+  INIT_VECTOR4(b_g_r_v255, b, g, r, v255);                              \
+  vst4_u8(out, b_g_r_v255);                                             \
+} while (0)
+
+#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) {            \
+  int i;                                                                \
+  for (i = 0; i < N; i += 8) {                                          \
+    const int off = ((cur_x) + i) * XSTEP;                              \
+    uint8x8_t y  = vld1_u8((src_y) + (cur_x)  + i);                     \
+    uint8x8_t u  = vld1_u8((src_uv) + i);                               \
+    uint8x8_t v  = vld1_u8((src_uv) + i + 16);                          \
+    const int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16));       \
+    const int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128));      \
+    const int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128));      \
+    int32x4_t yl = vmull_lane_s16(vget_low_s16(yy),  cf16, 0);          \
+    int32x4_t yh = vmull_lane_s16(vget_high_s16(yy), cf16, 0);          \
+    const int32x4_t rl = vmlal_lane_s16(yl, vget_low_s16(vv),  cf16, 1);\
+    const int32x4_t rh = vmlal_lane_s16(yh, vget_high_s16(vv), cf16, 1);\
+    int32x4_t gl = vmlsl_lane_s16(yl, vget_low_s16(uu),  cf16, 2);      \
+    int32x4_t gh = vmlsl_lane_s16(yh, vget_high_s16(uu), cf16, 2);      \
+    const int32x4_t bl = vmovl_s16(vget_low_s16(uu));                   \
+    const int32x4_t bh = vmovl_s16(vget_high_s16(uu));                  \
+    gl = vmlsl_lane_s16(gl, vget_low_s16(vv),  cf16, 3);                \
+    gh = vmlsl_lane_s16(gh, vget_high_s16(vv), cf16, 3);                \
+    yl = vmlaq_lane_s32(yl, bl, cf32, 0);                               \
+    yh = vmlaq_lane_s32(yh, bh, cf32, 0);                               \
+    /* vrshrn_n_s32() already incorporates the rounding constant */     \
+    y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, YUV_FIX2),            \
+                                 vrshrn_n_s32(rh, YUV_FIX2)));          \
+    u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, YUV_FIX2),            \
+                                 vrshrn_n_s32(gh, YUV_FIX2)));          \
+    v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(yl, YUV_FIX2),            \
+                                 vrshrn_n_s32(yh, YUV_FIX2)));          \
+    STORE_ ## FMT(out + off, y, u, v);                                  \
+  }                                                                     \
+}
+
+#define CONVERT1(FUNC, XSTEP, N, src_y, src_uv, rgb, cur_x) {           \
+  int i;                                                                \
+  for (i = 0; i < N; i++) {                                             \
+    const int off = ((cur_x) + i) * XSTEP;                              \
+    const int y = src_y[(cur_x) + i];                                   \
+    const int u = (src_uv)[i];                                          \
+    const int v = (src_uv)[i + 16];                                     \
+    FUNC(y, u, v, rgb + off);                                           \
+  }                                                                     \
+}
+
+#define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv,                  \
+                      top_dst, bottom_dst, cur_x, len) {                \
+  CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x)                  \
+  if (bottom_y != NULL) {                                               \
+    CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x)   \
+  }                                                                     \
+}
+
+#define CONVERT2RGB_1(FUNC, XSTEP, top_y, bottom_y, uv,                 \
+                      top_dst, bottom_dst, cur_x, len) {                \
+  CONVERT1(FUNC, XSTEP, len, top_y, uv, top_dst, cur_x);                \
+  if (bottom_y != NULL) {                                               \
+    CONVERT1(FUNC, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x); \
+  }                                                                     \
+}
+
+#define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP)                       \
+static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y,    \
+                      const uint8_t *top_u, const uint8_t *top_v,       \
+                      const uint8_t *cur_u, const uint8_t *cur_v,       \
+                      uint8_t *top_dst, uint8_t *bottom_dst, int len) { \
+  int block;                                                            \
+  /* 16 byte aligned array to cache reconstructed u and v */            \
+  uint8_t uv_buf[2 * 32 + 15];                                          \
+  uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);     \
+  const int uv_len = (len + 1) >> 1;                                    \
+  /* 9 pixels must be read-able for each block */                       \
+  const int num_blocks = (uv_len - 1) >> 3;                             \
+  const int leftover = uv_len - num_blocks * 8;                         \
+  const int last_pos = 1 + 16 * num_blocks;                             \
+                                                                        \
+  const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                  \
+  const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                  \
+                                                                        \
+  const int16x4_t cf16 = vld1_s16(kCoeffs);                             \
+  const int32x2_t cf32 = vdup_n_s32(kUToB);                             \
+  const uint8x8_t u16  = vdup_n_u8(16);                                 \
+  const uint8x8_t u128 = vdup_n_u8(128);                                \
+                                                                        \
+  /* Treat the first pixel in regular way */                            \
+  assert(top_y != NULL);                                                \
+  {                                                                     \
+    const int u0 = (top_u[0] + u_diag) >> 1;                            \
+    const int v0 = (top_v[0] + v_diag) >> 1;                            \
+    VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst);                         \
+  }                                                                     \
+  if (bottom_y != NULL) {                                               \
+    const int u0 = (cur_u[0] + u_diag) >> 1;                            \
+    const int v0 = (cur_v[0] + v_diag) >> 1;                            \
+    VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst);                   \
+  }                                                                     \
+                                                                        \
+  for (block = 0; block < num_blocks; ++block) {                        \
+    UPSAMPLE_16PIXELS(top_u, cur_u, r_uv);                              \
+    UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16);                         \
+    CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv,                    \
+                  top_dst, bottom_dst, 16 * block + 1, 16);             \
+    top_u += 8;                                                         \
+    cur_u += 8;                                                         \
+    top_v += 8;                                                         \
+    cur_v += 8;                                                         \
+  }                                                                     \
+                                                                        \
+  UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv);                    \
+  UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16);               \
+  CONVERT2RGB_1(VP8YuvTo ## FMT, XSTEP, top_y, bottom_y, r_uv,          \
+                top_dst, bottom_dst, last_pos, len - last_pos);         \
+}
+
+// NEON variants of the fancy upsampler.
+NEON_UPSAMPLE_FUNC(UpsampleRgbLinePair,  Rgb,  3)
+NEON_UPSAMPLE_FUNC(UpsampleBgrLinePair,  Bgr,  3)
+NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePair, Rgba, 4)
+NEON_UPSAMPLE_FUNC(UpsampleBgraLinePair, Bgra, 4)
+
+#endif  // FANCY_UPSAMPLING
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+
+extern void WebPInitUpsamplersNEON(void);
+
+#ifdef FANCY_UPSAMPLING
+
+extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
+
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersNEON(void) {
+#if defined(WEBP_USE_NEON)
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
+#endif   // WEBP_USE_NEON
+}
+
+#else
+
+// this empty function is to avoid an empty .o
+WEBP_TSAN_IGNORE_FUNCTION void WebPInitUpsamplersNEON(void) {}
+
+#endif  // FANCY_UPSAMPLING