rfreeimage 0.1.0

data/vendor/FreeImage/Source/LibWebP/src/dsp/mips_macro.h

@@ -0,0 +1,200 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+// MIPS common macros
+
+#ifndef WEBP_DSP_MIPS_MACRO_H_
+#define WEBP_DSP_MIPS_MACRO_H_
+
+#if defined(__GNUC__) && defined(__ANDROID__) && LOCAL_GCC_VERSION == 0x409
+#define WORK_AROUND_GCC
+#endif
+
+#define STR(s) #s
+#define XSTR(s) STR(s)
+
+// O0[31..16 | 15..0] = I0[31..16 | 15..0] + I1[31..16 | 15..0]
+// O1[31..16 | 15..0] = I0[31..16 | 15..0] - I1[31..16 | 15..0]
+// O - output
+// I - input (macro doesn't change it)
+#define ADD_SUB_HALVES(O0, O1,                                                 \
+                       I0, I1)                                                 \
+  "addq.ph          %["#O0"],   %["#I0"],  %["#I1"]           \n\t"            \
+  "subq.ph          %["#O1"],   %["#I0"],  %["#I1"]           \n\t"
+
+// O - output
+// I - input (macro doesn't change it)
+// I[0/1] - offset in bytes
+#define LOAD_IN_X2(O0, O1,                                                     \
+                   I0, I1)                                                     \
+  "lh               %["#O0"],   "#I0"(%[in])                  \n\t"            \
+  "lh               %["#O1"],   "#I1"(%[in])                  \n\t"
+
+// I0 - location
+// I1..I9 - offsets in bytes
+#define LOAD_WITH_OFFSET_X4(O0, O1, O2, O3,                                    \
+                            I0, I1, I2, I3, I4, I5, I6, I7, I8, I9)            \
+  "ulw    %["#O0"],    "#I1"+"XSTR(I9)"*"#I5"(%["#I0"])       \n\t"            \
+  "ulw    %["#O1"],    "#I2"+"XSTR(I9)"*"#I6"(%["#I0"])       \n\t"            \
+  "ulw    %["#O2"],    "#I3"+"XSTR(I9)"*"#I7"(%["#I0"])       \n\t"            \
+  "ulw    %["#O3"],    "#I4"+"XSTR(I9)"*"#I8"(%["#I0"])       \n\t"
+
+// O - output
+// IO - input/output
+// I - input (macro doesn't change it)
+#define MUL_SHIFT_SUM(O0, O1, O2, O3, O4, O5, O6, O7,                          \
+                      IO0, IO1, IO2, IO3,                                      \
+                      I0, I1, I2, I3, I4, I5, I6, I7)                          \
+  "mul              %["#O0"],   %["#I0"],   %[kC2]            \n\t"            \
+  "mul              %["#O1"],   %["#I0"],   %[kC1]            \n\t"            \
+  "mul              %["#O2"],   %["#I1"],   %[kC2]            \n\t"            \
+  "mul              %["#O3"],   %["#I1"],   %[kC1]            \n\t"            \
+  "mul              %["#O4"],   %["#I2"],   %[kC2]            \n\t"            \
+  "mul              %["#O5"],   %["#I2"],   %[kC1]            \n\t"            \
+  "mul              %["#O6"],   %["#I3"],   %[kC2]            \n\t"            \
+  "mul              %["#O7"],   %["#I3"],   %[kC1]            \n\t"            \
+  "sra              %["#O0"],   %["#O0"],   16                \n\t"            \
+  "sra              %["#O1"],   %["#O1"],   16                \n\t"            \
+  "sra              %["#O2"],   %["#O2"],   16                \n\t"            \
+  "sra              %["#O3"],   %["#O3"],   16                \n\t"            \
+  "sra              %["#O4"],   %["#O4"],   16                \n\t"            \
+  "sra              %["#O5"],   %["#O5"],   16                \n\t"            \
+  "sra              %["#O6"],   %["#O6"],   16                \n\t"            \
+  "sra              %["#O7"],   %["#O7"],   16                \n\t"            \
+  "addu             %["#IO0"],  %["#IO0"],  %["#I4"]          \n\t"            \
+  "addu             %["#IO1"],  %["#IO1"],  %["#I5"]          \n\t"            \
+  "subu             %["#IO2"],  %["#IO2"],  %["#I6"]          \n\t"            \
+  "subu             %["#IO3"],  %["#IO3"],  %["#I7"]          \n\t"
+
+// O - output
+// I - input (macro doesn't change it)
+#define INSERT_HALF_X2(O0, O1,                                                 \
+                       I0, I1)                                                 \
+  "ins              %["#O0"],   %["#I0"], 16,    16           \n\t"            \
+  "ins              %["#O1"],   %["#I1"], 16,    16           \n\t"
+
+// O - output
+// I - input (macro doesn't change it)
+#define SRA_16(O0, O1, O2, O3,                                                 \
+               I0, I1, I2, I3)                                                 \
+  "sra              %["#O0"],  %["#I0"],  16                  \n\t"            \
+  "sra              %["#O1"],  %["#I1"],  16                  \n\t"            \
+  "sra              %["#O2"],  %["#I2"],  16                  \n\t"            \
+  "sra              %["#O3"],  %["#I3"],  16                  \n\t"
+
+// temp0[31..16 | 15..0] = temp8[31..16 | 15..0] + temp12[31..16 | 15..0]
+// temp1[31..16 | 15..0] = temp8[31..16 | 15..0] - temp12[31..16 | 15..0]
+// temp0[31..16 | 15..0] = temp0[31..16 >> 3 | 15..0 >> 3]
+// temp1[31..16 | 15..0] = temp1[31..16 >> 3 | 15..0 >> 3]
+// O - output
+// I - input (macro doesn't change it)
+#define SHIFT_R_SUM_X2(O0, O1, O2, O3, O4, O5, O6, O7,                         \
+                       I0, I1, I2, I3, I4, I5, I6, I7)                         \
+  "addq.ph          %["#O0"],   %["#I0"],   %["#I4"]          \n\t"            \
+  "subq.ph          %["#O1"],   %["#I0"],   %["#I4"]          \n\t"            \
+  "addq.ph          %["#O2"],   %["#I1"],   %["#I5"]          \n\t"            \
+  "subq.ph          %["#O3"],   %["#I1"],   %["#I5"]          \n\t"            \
+  "addq.ph          %["#O4"],   %["#I2"],   %["#I6"]          \n\t"            \
+  "subq.ph          %["#O5"],   %["#I2"],   %["#I6"]          \n\t"            \
+  "addq.ph          %["#O6"],   %["#I3"],   %["#I7"]          \n\t"            \
+  "subq.ph          %["#O7"],   %["#I3"],   %["#I7"]          \n\t"            \
+  "shra.ph          %["#O0"],   %["#O0"],   3                 \n\t"            \
+  "shra.ph          %["#O1"],   %["#O1"],   3                 \n\t"            \
+  "shra.ph          %["#O2"],   %["#O2"],   3                 \n\t"            \
+  "shra.ph          %["#O3"],   %["#O3"],   3                 \n\t"            \
+  "shra.ph          %["#O4"],   %["#O4"],   3                 \n\t"            \
+  "shra.ph          %["#O5"],   %["#O5"],   3                 \n\t"            \
+  "shra.ph          %["#O6"],   %["#O6"],   3                 \n\t"            \
+  "shra.ph          %["#O7"],   %["#O7"],   3                 \n\t"
+
+// precrq.ph.w temp0, temp8, temp2
+//   temp0 = temp8[31..16] | temp2[31..16]
+// ins temp2, temp8, 16, 16
+//   temp2 = temp8[31..16] | temp2[15..0]
+// O - output
+// IO - input/output
+// I - input (macro doesn't change it)
+#define PACK_2_HALVES_TO_WORD(O0, O1, O2, O3,                                  \
+                              IO0, IO1, IO2, IO3,                              \
+                              I0, I1, I2, I3)                                  \
+  "precrq.ph.w      %["#O0"],    %["#I0"],  %["#IO0"]         \n\t"            \
+  "precrq.ph.w      %["#O1"],    %["#I1"],  %["#IO1"]         \n\t"            \
+  "ins              %["#IO0"],   %["#I0"],  16,    16         \n\t"            \
+  "ins              %["#IO1"],   %["#I1"],  16,    16         \n\t"            \
+  "precrq.ph.w      %["#O2"],    %["#I2"],  %["#IO2"]         \n\t"            \
+  "precrq.ph.w      %["#O3"],    %["#I3"],  %["#IO3"]         \n\t"            \
+  "ins              %["#IO2"],   %["#I2"],  16,    16         \n\t"            \
+  "ins              %["#IO3"],   %["#I3"],  16,    16         \n\t"
+
+// preceu.ph.qbr temp0, temp8
+//   temp0 = 0 | 0 | temp8[23..16] | temp8[7..0]
+// preceu.ph.qbl temp1, temp8
+//   temp1 = temp8[23..16] | temp8[7..0] | 0 | 0
+// O - output
+// I - input (macro doesn't change it)
+#define CONVERT_2_BYTES_TO_HALF(O0, O1, O2, O3, O4, O5, O6, O7,                \
+                                I0, I1, I2, I3)                                \
+  "preceu.ph.qbr    %["#O0"],   %["#I0"]                      \n\t"            \
+  "preceu.ph.qbl    %["#O1"],   %["#I0"]                      \n\t"            \
+  "preceu.ph.qbr    %["#O2"],   %["#I1"]                      \n\t"            \
+  "preceu.ph.qbl    %["#O3"],   %["#I1"]                      \n\t"            \
+  "preceu.ph.qbr    %["#O4"],   %["#I2"]                      \n\t"            \
+  "preceu.ph.qbl    %["#O5"],   %["#I2"]                      \n\t"            \
+  "preceu.ph.qbr    %["#O6"],   %["#I3"]                      \n\t"            \
+  "preceu.ph.qbl    %["#O7"],   %["#I3"]                      \n\t"
+
+// temp0[31..16 | 15..0] = temp0[31..16 | 15..0] + temp8[31..16 | 15..0]
+// temp0[31..16 | 15..0] = temp0[31..16 <<(s) 7 | 15..0 <<(s) 7]
+// temp1..temp7 same as temp0
+// precrqu_s.qb.ph temp0, temp1, temp0:
+//   temp0 = temp1[31..24] | temp1[15..8] | temp0[31..24] | temp0[15..8]
+// store temp0 to dst
+// IO - input/output
+// I - input (macro doesn't change it)
+#define STORE_SAT_SUM_X2(IO0, IO1, IO2, IO3, IO4, IO5, IO6, IO7,               \
+                         I0, I1, I2, I3, I4, I5, I6, I7,                       \
+                         I8, I9, I10, I11, I12, I13)                           \
+  "addq.ph          %["#IO0"],  %["#IO0"],  %["#I0"]          \n\t"            \
+  "addq.ph          %["#IO1"],  %["#IO1"],  %["#I1"]          \n\t"            \
+  "addq.ph          %["#IO2"],  %["#IO2"],  %["#I2"]          \n\t"            \
+  "addq.ph          %["#IO3"],  %["#IO3"],  %["#I3"]          \n\t"            \
+  "addq.ph          %["#IO4"],  %["#IO4"],  %["#I4"]          \n\t"            \
+  "addq.ph          %["#IO5"],  %["#IO5"],  %["#I5"]          \n\t"            \
+  "addq.ph          %["#IO6"],  %["#IO6"],  %["#I6"]          \n\t"            \
+  "addq.ph          %["#IO7"],  %["#IO7"],  %["#I7"]          \n\t"            \
+  "shll_s.ph        %["#IO0"],  %["#IO0"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO1"],  %["#IO1"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO2"],  %["#IO2"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO3"],  %["#IO3"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO4"],  %["#IO4"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO5"],  %["#IO5"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO6"],  %["#IO6"],  7                 \n\t"            \
+  "shll_s.ph        %["#IO7"],  %["#IO7"],  7                 \n\t"            \
+  "precrqu_s.qb.ph  %["#IO0"],  %["#IO1"],  %["#IO0"]         \n\t"            \
+  "precrqu_s.qb.ph  %["#IO2"],  %["#IO3"],  %["#IO2"]         \n\t"            \
+  "precrqu_s.qb.ph  %["#IO4"],  %["#IO5"],  %["#IO4"]         \n\t"            \
+  "precrqu_s.qb.ph  %["#IO6"],  %["#IO7"],  %["#IO6"]         \n\t"            \
+  "usw              %["#IO0"],  "XSTR(I13)"*"#I9"(%["#I8"])   \n\t"            \
+  "usw              %["#IO2"],  "XSTR(I13)"*"#I10"(%["#I8"])  \n\t"            \
+  "usw              %["#IO4"],  "XSTR(I13)"*"#I11"(%["#I8"])  \n\t"            \
+  "usw              %["#IO6"],  "XSTR(I13)"*"#I12"(%["#I8"])  \n\t"
+
+#define OUTPUT_EARLY_CLOBBER_REGS_10()                                         \
+  : [temp1]"=&r"(temp1), [temp2]"=&r"(temp2), [temp3]"=&r"(temp3),             \
+    [temp4]"=&r"(temp4), [temp5]"=&r"(temp5), [temp6]"=&r"(temp6),             \
+    [temp7]"=&r"(temp7), [temp8]"=&r"(temp8), [temp9]"=&r"(temp9),             \
+    [temp10]"=&r"(temp10)
+
+#define OUTPUT_EARLY_CLOBBER_REGS_18()                                         \
+  OUTPUT_EARLY_CLOBBER_REGS_10(),                                              \
+  [temp11]"=&r"(temp11), [temp12]"=&r"(temp12), [temp13]"=&r"(temp13),         \
+  [temp14]"=&r"(temp14), [temp15]"=&r"(temp15), [temp16]"=&r"(temp16),         \
+  [temp17]"=&r"(temp17), [temp18]"=&r"(temp18)
+
+#endif  // WEBP_DSP_MIPS_MACRO_H_

data/vendor/FreeImage/Source/LibWebP/src/dsp/neon.h

@@ -0,0 +1,82 @@
+// 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.
+// -----------------------------------------------------------------------------
+//
+//  NEON common code.
+
+#ifndef WEBP_DSP_NEON_H_
+#define WEBP_DSP_NEON_H_
+
+#include <arm_neon.h>
+
+#include "./dsp.h"
+
+// Right now, some intrinsics functions seem slower, so we disable them
+// everywhere except aarch64 where the inline assembly is incompatible.
+#if defined(__aarch64__)
+#define WEBP_USE_INTRINSICS   // use intrinsics when possible
+#endif
+
+#define INIT_VECTOR2(v, a, b) do {  \
+  v.val[0] = a;                     \
+  v.val[1] = b;                     \
+} while (0)
+
+#define INIT_VECTOR3(v, a, b, c) do {  \
+  v.val[0] = a;                        \
+  v.val[1] = b;                        \
+  v.val[2] = c;                        \
+} while (0)
+
+#define INIT_VECTOR4(v, a, b, c, d) do {  \
+  v.val[0] = a;                           \
+  v.val[1] = b;                           \
+  v.val[2] = c;                           \
+  v.val[3] = d;                           \
+} while (0)
+
+// if using intrinsics, this flag avoids some functions that make gcc-4.6.3
+// crash ("internal compiler error: in immed_double_const, at emit-rtl.").
+// (probably similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=48183)
+#if !(LOCAL_GCC_PREREQ(4,8) || defined(__aarch64__))
+#define WORK_AROUND_GCC
+#endif
+
+static WEBP_INLINE int32x4x4_t Transpose4x4(const int32x4x4_t rows) {
+  uint64x2x2_t row01, row23;
+
+  row01.val[0] = vreinterpretq_u64_s32(rows.val[0]);
+  row01.val[1] = vreinterpretq_u64_s32(rows.val[1]);
+  row23.val[0] = vreinterpretq_u64_s32(rows.val[2]);
+  row23.val[1] = vreinterpretq_u64_s32(rows.val[3]);
+  // Transpose 64-bit values (there's no vswp equivalent)
+  {
+    const uint64x1_t row0h = vget_high_u64(row01.val[0]);
+    const uint64x1_t row2l = vget_low_u64(row23.val[0]);
+    const uint64x1_t row1h = vget_high_u64(row01.val[1]);
+    const uint64x1_t row3l = vget_low_u64(row23.val[1]);
+    row01.val[0] = vcombine_u64(vget_low_u64(row01.val[0]), row2l);
+    row23.val[0] = vcombine_u64(row0h, vget_high_u64(row23.val[0]));
+    row01.val[1] = vcombine_u64(vget_low_u64(row01.val[1]), row3l);
+    row23.val[1] = vcombine_u64(row1h, vget_high_u64(row23.val[1]));
+  }
+  {
+    const int32x4x2_t out01 = vtrnq_s32(vreinterpretq_s32_u64(row01.val[0]),
+                                        vreinterpretq_s32_u64(row01.val[1]));
+    const int32x4x2_t out23 = vtrnq_s32(vreinterpretq_s32_u64(row23.val[0]),
+                                        vreinterpretq_s32_u64(row23.val[1]));
+    int32x4x4_t out;
+    out.val[0] = out01.val[0];
+    out.val[1] = out01.val[1];
+    out.val[2] = out23.val[0];
+    out.val[3] = out23.val[1];
+    return out;
+  }
+}
+
+#endif  // WEBP_DSP_NEON_H_

data/vendor/FreeImage/Source/LibWebP/src/dsp/yuv.h

@@ -0,0 +1,321 @@
+// Copyright 2010 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.
+// -----------------------------------------------------------------------------
+//
+// inline YUV<->RGB conversion function
+//
+// The exact naming is Y'CbCr, following the ITU-R BT.601 standard.
+// More information at: http://en.wikipedia.org/wiki/YCbCr
+// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16
+// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128
+// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128
+// We use 16bit fixed point operations for RGB->YUV conversion (YUV_FIX).
+//
+// For the Y'CbCr to RGB conversion, the BT.601 specification reads:
+//   R = 1.164 * (Y-16) + 1.596 * (V-128)
+//   G = 1.164 * (Y-16) - 0.813 * (V-128) - 0.391 * (U-128)
+//   B = 1.164 * (Y-16)                   + 2.018 * (U-128)
+// where Y is in the [16,235] range, and U/V in the [16,240] range.
+// In the table-lookup version (WEBP_YUV_USE_TABLE), the common factor
+// "1.164 * (Y-16)" can be handled as an offset in the VP8kClip[] table.
+// So in this case the formulae should read:
+//   R = 1.164 * [Y + 1.371 * (V-128)                  ] - 18.624
+//   G = 1.164 * [Y - 0.698 * (V-128) - 0.336 * (U-128)] - 18.624
+//   B = 1.164 * [Y                   + 1.733 * (U-128)] - 18.624
+// once factorized.
+// For YUV->RGB conversion, only 14bit fixed precision is used (YUV_FIX2).
+// That's the maximum possible for a convenient ARM implementation.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#ifndef WEBP_DSP_YUV_H_
+#define WEBP_DSP_YUV_H_
+
+#include "./dsp.h"
+#include "../dec/decode_vp8.h"
+
+// Define the following to use the LUT-based code:
+// #define WEBP_YUV_USE_TABLE
+
+#if defined(WEBP_EXPERIMENTAL_FEATURES)
+// Do NOT activate this feature for real compression. This is only experimental!
+// This flag is for comparison purpose against JPEG's "YUVj" natural colorspace.
+// This colorspace is close to Rec.601's Y'CbCr model with the notable
+// difference of allowing larger range for luma/chroma.
+// See http://en.wikipedia.org/wiki/YCbCr#JPEG_conversion paragraph, and its
+// difference with http://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion
+// #define USE_YUVj
+#endif
+
+//------------------------------------------------------------------------------
+// YUV -> RGB conversion
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  YUV_FIX = 16,                    // fixed-point precision for RGB->YUV
+  YUV_HALF = 1 << (YUV_FIX - 1),
+  YUV_MASK = (256 << YUV_FIX) - 1,
+  YUV_RANGE_MIN = -227,            // min value of r/g/b output
+  YUV_RANGE_MAX = 256 + 226,       // max value of r/g/b output
+
+  YUV_FIX2 = 14,                   // fixed-point precision for YUV->RGB
+  YUV_HALF2 = 1 << (YUV_FIX2 - 1),
+  YUV_MASK2 = (256 << YUV_FIX2) - 1
+};
+
+// These constants are 14b fixed-point version of ITU-R BT.601 constants.
+#define kYScale 19077    // 1.164 = 255 / 219
+#define kVToR   26149    // 1.596 = 255 / 112 * 0.701
+#define kUToG   6419     // 0.391 = 255 / 112 * 0.886 * 0.114 / 0.587
+#define kVToG   13320    // 0.813 = 255 / 112 * 0.701 * 0.299 / 0.587
+#define kUToB   33050    // 2.018 = 255 / 112 * 0.886
+#define kRCst (-kYScale * 16 - kVToR * 128 + YUV_HALF2)
+#define kGCst (-kYScale * 16 + kUToG * 128 + kVToG * 128 + YUV_HALF2)
+#define kBCst (-kYScale * 16 - kUToB * 128 + YUV_HALF2)
+
+//------------------------------------------------------------------------------
+
+#if !defined(WEBP_YUV_USE_TABLE)
+
+// slower on x86 by ~7-8%, but bit-exact with the SSE2 version
+
+static WEBP_INLINE int VP8Clip8(int v) {
+  return ((v & ~YUV_MASK2) == 0) ? (v >> YUV_FIX2) : (v < 0) ? 0 : 255;
+}
+
+static WEBP_INLINE int VP8YUVToR(int y, int v) {
+  return VP8Clip8(kYScale * y + kVToR * v + kRCst);
+}
+
+static WEBP_INLINE int VP8YUVToG(int y, int u, int v) {
+  return VP8Clip8(kYScale * y - kUToG * u - kVToG * v + kGCst);
+}
+
+static WEBP_INLINE int VP8YUVToB(int y, int u) {
+  return VP8Clip8(kYScale * y + kUToB * u + kBCst);
+}
+
+static WEBP_INLINE void VP8YuvToRgb(int y, int u, int v,
+                                    uint8_t* const rgb) {
+  rgb[0] = VP8YUVToR(y, v);
+  rgb[1] = VP8YUVToG(y, u, v);
+  rgb[2] = VP8YUVToB(y, u);
+}
+
+static WEBP_INLINE void VP8YuvToBgr(int y, int u, int v,
+                                    uint8_t* const bgr) {
+  bgr[0] = VP8YUVToB(y, u);
+  bgr[1] = VP8YUVToG(y, u, v);
+  bgr[2] = VP8YUVToR(y, v);
+}
+
+static WEBP_INLINE void VP8YuvToRgb565(int y, int u, int v,
+                                       uint8_t* const rgb) {
+  const int r = VP8YUVToR(y, v);      // 5 usable bits
+  const int g = VP8YUVToG(y, u, v);   // 6 usable bits
+  const int b = VP8YUVToB(y, u);      // 5 usable bits
+  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 VP8YuvToRgba4444(int y, int u, int v,
+                                         uint8_t* const argb) {
+  const int r = VP8YUVToR(y, v);        // 4 usable bits
+  const int g = VP8YUVToG(y, u, v);     // 4 usable bits
+  const int b = VP8YUVToB(y, u);        // 4 usable bits
+  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
+}
+
+#else
+
+// Table-based version, not totally equivalent to the SSE2 version.
+// Rounding diff is only +/-1 though.
+
+extern int16_t VP8kVToR[256], VP8kUToB[256];
+extern int32_t VP8kVToG[256], VP8kUToG[256];
+extern uint8_t VP8kClip[YUV_RANGE_MAX - YUV_RANGE_MIN];
+extern uint8_t VP8kClip4Bits[YUV_RANGE_MAX - YUV_RANGE_MIN];
+
+static WEBP_INLINE void VP8YuvToRgb(int y, int u, int v,
+                                    uint8_t* const rgb) {
+  const int r_off = VP8kVToR[v];
+  const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
+  const int b_off = VP8kUToB[u];
+  rgb[0] = VP8kClip[y + r_off - YUV_RANGE_MIN];
+  rgb[1] = VP8kClip[y + g_off - YUV_RANGE_MIN];
+  rgb[2] = VP8kClip[y + b_off - YUV_RANGE_MIN];
+}
+
+static WEBP_INLINE void VP8YuvToBgr(int y, int u, int v,
+                                    uint8_t* const bgr) {
+  const int r_off = VP8kVToR[v];
+  const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
+  const int b_off = VP8kUToB[u];
+  bgr[0] = VP8kClip[y + b_off - YUV_RANGE_MIN];
+  bgr[1] = VP8kClip[y + g_off - YUV_RANGE_MIN];
+  bgr[2] = VP8kClip[y + r_off - YUV_RANGE_MIN];
+}
+
+static WEBP_INLINE void VP8YuvToRgb565(int y, int u, int v,
+                                       uint8_t* const rgb) {
+  const int r_off = VP8kVToR[v];
+  const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
+  const int b_off = VP8kUToB[u];
+  const int rg = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) |
+                  (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5));
+  const int gb = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) |
+                   (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3));
+#ifdef WEBP_SWAP_16BIT_CSP
+  rgb[0] = gb;
+  rgb[1] = rg;
+#else
+  rgb[0] = rg;
+  rgb[1] = gb;
+#endif
+}
+
+static WEBP_INLINE void VP8YuvToRgba4444(int y, int u, int v,
+                                         uint8_t* const argb) {
+  const int r_off = VP8kVToR[v];
+  const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
+  const int b_off = VP8kUToB[u];
+  const int rg = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) |
+                   VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]);
+  const int ba = (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4) | 0x0f;
+#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 VP8YuvToArgb(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const argb) {
+  argb[0] = 0xff;
+  VP8YuvToRgb(y, u, v, argb + 1);
+}
+
+static WEBP_INLINE void VP8YuvToBgra(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const bgra) {
+  VP8YuvToBgr(y, u, v, bgra);
+  bgra[3] = 0xff;
+}
+
+static WEBP_INLINE void VP8YuvToRgba(uint8_t y, uint8_t u, uint8_t v,
+                                     uint8_t* const rgba) {
+  VP8YuvToRgb(y, u, v, rgba);
+  rgba[3] = 0xff;
+}
+
+// Must be called before everything, to initialize the tables.
+void VP8YUVInit(void);
+
+//-----------------------------------------------------------------------------
+// SSE2 extra functions (mostly for upsampling_sse2.c)
+
+#if defined(WEBP_USE_SSE2)
+
+// When the following is defined, tables are initialized statically, adding ~12k
+// to the binary size. Otherwise, they are initialized at run-time (small cost).
+#define WEBP_YUV_USE_SSE2_TABLES
+
+#if defined(FANCY_UPSAMPLING)
+// Process 32 pixels and store the result (24b or 32b per pixel) in *dst.
+void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst);
+void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst);
+void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                    uint8_t* dst);
+void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
+                   uint8_t* dst);
+#endif  // FANCY_UPSAMPLING
+
+// Must be called to initialize tables before using the functions.
+void VP8YUVInitSSE2(void);
+
+#endif    // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// RGB -> YUV conversion
+
+// Stub functions that can be called with various rounding values:
+static WEBP_INLINE int VP8ClipUV(int uv, int rounding) {
+  uv = (uv + rounding + (128 << (YUV_FIX + 2))) >> (YUV_FIX + 2);
+  return ((uv & ~0xff) == 0) ? uv : (uv < 0) ? 0 : 255;
+}
+
+#ifndef USE_YUVj
+
+static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) {
+  const int luma = 16839 * r + 33059 * g + 6420 * b;
+  return (luma + rounding + (16 << YUV_FIX)) >> YUV_FIX;  // no need to clip
+}
+
+static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) {
+  const int u = -9719 * r - 19081 * g + 28800 * b;
+  return VP8ClipUV(u, rounding);
+}
+
+static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) {
+  const int v = +28800 * r - 24116 * g - 4684 * b;
+  return VP8ClipUV(v, rounding);
+}
+
+#else
+
+// This JPEG-YUV colorspace, only for comparison!
+// These are also 16bit precision coefficients from Rec.601, but with full
+// [0..255] output range.
+static WEBP_INLINE int VP8RGBToY(int r, int g, int b, int rounding) {
+  const int luma = 19595 * r + 38470 * g + 7471 * b;
+  return (luma + rounding) >> YUV_FIX;  // no need to clip
+}
+
+static WEBP_INLINE int VP8RGBToU(int r, int g, int b, int rounding) {
+  const int u = -11058 * r - 21710 * g + 32768 * b;
+  return VP8ClipUV(u, rounding);
+}
+
+static WEBP_INLINE int VP8RGBToV(int r, int g, int b, int rounding) {
+  const int v = 32768 * r - 27439 * g - 5329 * b;
+  return VP8ClipUV(v, rounding);
+}
+
+#endif    // USE_YUVj
+
+#ifdef __cplusplus
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_DSP_YUV_H_ */