1// Copyright 2016 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// SSE2 code common to several files.
11//
12// Author: Vincent Rabaud (vrabaud@google.com)
13
14#ifndef WEBP_DSP_COMMON_SSE2_H_
15#define WEBP_DSP_COMMON_SSE2_H_
16
17#ifdef __cplusplus
18extern "C" {
19#endif
20
21#if defined(WEBP_USE_SSE2)
22
23#include <emmintrin.h>
24
25//------------------------------------------------------------------------------
26// Quite useful macro for debugging. Left here for convenience.
27
28#if 0
29#include <stdio.h>
30static WEBP_INLINE void PrintReg(const __m128i r, const char* const name,
31                                 int size) {
32  int n;
33  union {
34    __m128i r;
35    uint8_t i8[16];
36    uint16_t i16[8];
37    uint32_t i32[4];
38    uint64_t i64[2];
39  } tmp;
40  tmp.r = r;
41  fprintf(stderr, "%s\t: ", name);
42  if (size == 8) {
43    for (n = 0; n < 16; ++n) fprintf(stderr, "%.2x ", tmp.i8[n]);
44  } else if (size == 16) {
45    for (n = 0; n < 8; ++n) fprintf(stderr, "%.4x ", tmp.i16[n]);
46  } else if (size == 32) {
47    for (n = 0; n < 4; ++n) fprintf(stderr, "%.8x ", tmp.i32[n]);
48  } else {
49    for (n = 0; n < 2; ++n) fprintf(stderr, "%.16lx ", tmp.i64[n]);
50  }
51  fprintf(stderr, "\n");
52}
53#endif
54
55//------------------------------------------------------------------------------
56// Math functions.
57
58// Return the sum of all the 8b in the register.
59static WEBP_INLINE int VP8HorizontalAdd8b(const __m128i* const a) {
60  const __m128i zero = _mm_setzero_si128();
61  const __m128i sad8x2 = _mm_sad_epu8(*a, zero);
62  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
63  const __m128i sum = _mm_add_epi32(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
64  return _mm_cvtsi128_si32(sum);
65}
66
67// Transpose two 4x4 16b matrices horizontally stored in registers.
68static WEBP_INLINE void VP8Transpose_2_4x4_16b(
69    const __m128i* const in0, const __m128i* const in1,
70    const __m128i* const in2, const __m128i* const in3, __m128i* const out0,
71    __m128i* const out1, __m128i* const out2, __m128i* const out3) {
72  // Transpose the two 4x4.
73  // a00 a01 a02 a03   b00 b01 b02 b03
74  // a10 a11 a12 a13   b10 b11 b12 b13
75  // a20 a21 a22 a23   b20 b21 b22 b23
76  // a30 a31 a32 a33   b30 b31 b32 b33
77  const __m128i transpose0_0 = _mm_unpacklo_epi16(*in0, *in1);
78  const __m128i transpose0_1 = _mm_unpacklo_epi16(*in2, *in3);
79  const __m128i transpose0_2 = _mm_unpackhi_epi16(*in0, *in1);
80  const __m128i transpose0_3 = _mm_unpackhi_epi16(*in2, *in3);
81  // a00 a10 a01 a11   a02 a12 a03 a13
82  // a20 a30 a21 a31   a22 a32 a23 a33
83  // b00 b10 b01 b11   b02 b12 b03 b13
84  // b20 b30 b21 b31   b22 b32 b23 b33
85  const __m128i transpose1_0 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
86  const __m128i transpose1_1 = _mm_unpacklo_epi32(transpose0_2, transpose0_3);
87  const __m128i transpose1_2 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
88  const __m128i transpose1_3 = _mm_unpackhi_epi32(transpose0_2, transpose0_3);
89  // a00 a10 a20 a30 a01 a11 a21 a31
90  // b00 b10 b20 b30 b01 b11 b21 b31
91  // a02 a12 a22 a32 a03 a13 a23 a33
92  // b02 b12 a22 b32 b03 b13 b23 b33
93  *out0 = _mm_unpacklo_epi64(transpose1_0, transpose1_1);
94  *out1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1);
95  *out2 = _mm_unpacklo_epi64(transpose1_2, transpose1_3);
96  *out3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3);
97  // a00 a10 a20 a30   b00 b10 b20 b30
98  // a01 a11 a21 a31   b01 b11 b21 b31
99  // a02 a12 a22 a32   b02 b12 b22 b32
100  // a03 a13 a23 a33   b03 b13 b23 b33
101}
102
103//------------------------------------------------------------------------------
104// Channel mixing.
105
106// Function used several times in VP8PlanarTo24b.
107// It samples the in buffer as follows: one every two unsigned char is stored
108// at the beginning of the buffer, while the other half is stored at the end.
109#define VP8PlanarTo24bHelper(IN, OUT)                            \
110  do {                                                           \
111    const __m128i v_mask = _mm_set1_epi16(0x00ff);               \
112    /* Take one every two upper 8b values.*/                     \
113    (OUT##0) = _mm_packus_epi16(_mm_and_si128((IN##0), v_mask),  \
114                                _mm_and_si128((IN##1), v_mask)); \
115    (OUT##1) = _mm_packus_epi16(_mm_and_si128((IN##2), v_mask),  \
116                                _mm_and_si128((IN##3), v_mask)); \
117    (OUT##2) = _mm_packus_epi16(_mm_and_si128((IN##4), v_mask),  \
118                                _mm_and_si128((IN##5), v_mask)); \
119    /* Take one every two lower 8b values.*/                     \
120    (OUT##3) = _mm_packus_epi16(_mm_srli_epi16((IN##0), 8),      \
121                                _mm_srli_epi16((IN##1), 8));     \
122    (OUT##4) = _mm_packus_epi16(_mm_srli_epi16((IN##2), 8),      \
123                                _mm_srli_epi16((IN##3), 8));     \
124    (OUT##5) = _mm_packus_epi16(_mm_srli_epi16((IN##4), 8),      \
125                                _mm_srli_epi16((IN##5), 8));     \
126  } while (0)
127
128// Pack the planar buffers
129// rrrr... rrrr... gggg... gggg... bbbb... bbbb....
130// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ...
131static WEBP_INLINE void VP8PlanarTo24b(__m128i* const in0, __m128i* const in1,
132                                       __m128i* const in2, __m128i* const in3,
133                                       __m128i* const in4, __m128i* const in5) {
134  // The input is 6 registers of sixteen 8b but for the sake of explanation,
135  // let's take 6 registers of four 8b values.
136  // To pack, we will keep taking one every two 8b integer and move it
137  // around as follows:
138  // Input:
139  //   r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7
140  // Split the 6 registers in two sets of 3 registers: the first set as the even
141  // 8b bytes, the second the odd ones:
142  //   r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7
143  // Repeat the same permutations twice more:
144  //   r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7
145  //   r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7
146  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
147  VP8PlanarTo24bHelper(*in, tmp);
148  VP8PlanarTo24bHelper(tmp, *in);
149  VP8PlanarTo24bHelper(*in, tmp);
150  // We need to do it two more times than the example as we have sixteen bytes.
151  {
152    __m128i out0, out1, out2, out3, out4, out5;
153    VP8PlanarTo24bHelper(tmp, out);
154    VP8PlanarTo24bHelper(out, *in);
155  }
156}
157
158#undef VP8PlanarTo24bHelper
159
160// Convert four packed four-channel buffers like argbargbargbargb... into the
161// split channels aaaaa ... rrrr ... gggg .... bbbbb ......
162static WEBP_INLINE void VP8L32bToPlanar(__m128i* const in0,
163                                        __m128i* const in1,
164                                        __m128i* const in2,
165                                        __m128i* const in3) {
166  // Column-wise transpose.
167  const __m128i A0 = _mm_unpacklo_epi8(*in0, *in1);
168  const __m128i A1 = _mm_unpackhi_epi8(*in0, *in1);
169  const __m128i A2 = _mm_unpacklo_epi8(*in2, *in3);
170  const __m128i A3 = _mm_unpackhi_epi8(*in2, *in3);
171  const __m128i B0 = _mm_unpacklo_epi8(A0, A1);
172  const __m128i B1 = _mm_unpackhi_epi8(A0, A1);
173  const __m128i B2 = _mm_unpacklo_epi8(A2, A3);
174  const __m128i B3 = _mm_unpackhi_epi8(A2, A3);
175  // C0 = g7 g6 ... g1 g0 | b7 b6 ... b1 b0
176  // C1 = a7 a6 ... a1 a0 | r7 r6 ... r1 r0
177  const __m128i C0 = _mm_unpacklo_epi8(B0, B1);
178  const __m128i C1 = _mm_unpackhi_epi8(B0, B1);
179  const __m128i C2 = _mm_unpacklo_epi8(B2, B3);
180  const __m128i C3 = _mm_unpackhi_epi8(B2, B3);
181  // Gather the channels.
182  *in0 = _mm_unpackhi_epi64(C1, C3);
183  *in1 = _mm_unpacklo_epi64(C1, C3);
184  *in2 = _mm_unpackhi_epi64(C0, C2);
185  *in3 = _mm_unpacklo_epi64(C0, C2);
186}
187
188#endif  // WEBP_USE_SSE2
189
190#ifdef __cplusplus
191}  // extern "C"
192#endif
193
194#endif  // WEBP_DSP_COMMON_SSE2_H_
195