```/* <![CDATA[ */
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 version of YUV to RGB upsampling functions.
11//
13
14#include "./dsp.h"
15
16#if defined(WEBP_USE_SSE2)
17
18#include <assert.h>
19#include <emmintrin.h>
20#include <string.h>
21#include "./yuv.h"
22
23#ifdef FANCY_UPSAMPLING
24
25// We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
26// u = (9*a + 3*b + 3*c + d + 8) / 16
27//   = (a + (a + 3*b + 3*c + d) / 8 + 1) / 2
28//   = (a + m + 1) / 2
29// where m = (a + 3*b + 3*c + d) / 8
30//         = ((a + b + c + d) / 2 + b + c) / 4
31//
32// Let's say  k = (a + b + c + d) / 4.
33// We can compute k as
34// k = (s + t + 1) / 2 - ((a^d) | (b^c) | (s^t)) & 1
35// where s = (a + d + 1) / 2 and t = (b + c + 1) / 2
36//
37// Then m can be written as
38// m = (k + t + 1) / 2 - (((b^c) & (s^t)) | (k^t)) & 1
39
40// Computes out = (k + in + 1) / 2 - ((ij & (s^t)) | (k^in)) & 1
41#define GET_M(ij, in, out) do {                                                \
42  const __m128i tmp0 = _mm_avg_epu8(k, (in));     /* (k + in + 1) / 2 */       \
43  const __m128i tmp1 = _mm_and_si128((ij), st);   /* (ij) & (s^t) */           \
44  const __m128i tmp2 = _mm_xor_si128(k, (in));    /* (k^in) */                 \
45  const __m128i tmp3 = _mm_or_si128(tmp1, tmp2);  /* ((ij) & (s^t)) | (k^in) */\
46  const __m128i tmp4 = _mm_and_si128(tmp3, one);  /* & 1 -> lsb_correction */  \
47  (out) = _mm_sub_epi8(tmp0, tmp4);    /* (k + in + 1) / 2 - lsb_correction */ \
48} while (0)
49
50// pack and store two alternating pixel rows
51#define PACK_AND_STORE(a, b, da, db, out) do {                                 \
52  const __m128i t_a = _mm_avg_epu8(a, da);  /* (9a + 3b + 3c +  d + 8) / 16 */ \
53  const __m128i t_b = _mm_avg_epu8(b, db);  /* (3a + 9b +  c + 3d + 8) / 16 */ \
54  const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b);                             \
55  const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b);                             \
56  _mm_store_si128(((__m128i*)(out)) + 0, t_1);                                 \
57  _mm_store_si128(((__m128i*)(out)) + 1, t_2);                                 \
58} while (0)
59
60// Loads 17 pixels each from rows r1 and r2 and generates 32 pixels.
61#define UPSAMPLE_32PIXELS(r1, r2, out) {                                       \
62  const __m128i one = _mm_set1_epi8(1);                                        \
63  const __m128i a = _mm_loadu_si128((__m128i*)&(r1)[0]);                       \
64  const __m128i b = _mm_loadu_si128((__m128i*)&(r1)[1]);                       \
65  const __m128i c = _mm_loadu_si128((__m128i*)&(r2)[0]);                       \
66  const __m128i d = _mm_loadu_si128((__m128i*)&(r2)[1]);                       \
67                                                                               \
68  const __m128i s = _mm_avg_epu8(a, d);        /* s = (a + d + 1) / 2 */       \
69  const __m128i t = _mm_avg_epu8(b, c);        /* t = (b + c + 1) / 2 */       \
70  const __m128i st = _mm_xor_si128(s, t);      /* st = s^t */                  \
71                                                                               \
72  const __m128i ad = _mm_xor_si128(a, d);      /* ad = a^d */                  \
73  const __m128i bc = _mm_xor_si128(b, c);      /* bc = b^c */                  \
74                                                                               \
75  const __m128i t1 = _mm_or_si128(ad, bc);     /* (a^d) | (b^c) */             \
76  const __m128i t2 = _mm_or_si128(t1, st);     /* (a^d) | (b^c) | (s^t) */     \
77  const __m128i t3 = _mm_and_si128(t2, one);   /* (a^d) | (b^c) | (s^t) & 1 */ \
78  const __m128i t4 = _mm_avg_epu8(s, t);                                       \
79  const __m128i k = _mm_sub_epi8(t4, t3);      /* k = (a + b + c + d) / 4 */   \
80  __m128i diag1, diag2;                                                        \
81                                                                               \
82  GET_M(bc, t, diag1);                  /* diag1 = (a + 3b + 3c + d) / 8 */    \
83  GET_M(ad, s, diag2);                  /* diag2 = (3a + b + c + 3d) / 8 */    \
84                                                                               \
85  /* pack the alternate pixels */                                              \
86  PACK_AND_STORE(a, b, diag1, diag2, out +      0);  /* store top */           \
87  PACK_AND_STORE(c, d, diag2, diag1, out + 2 * 32);  /* store bottom */        \
88}
89
90// Turn the macro into a function for reducing code-size when non-critical
91static void Upsample32Pixels(const uint8_t r1[], const uint8_t r2[],
92                             uint8_t* const out) {
93  UPSAMPLE_32PIXELS(r1, r2, out);
94}
95
96#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                         \
97  uint8_t r1[17], r2[17];                                                      \
98  memcpy(r1, (tb), (num_pixels));                                              \
99  memcpy(r2, (bb), (num_pixels));                                              \
100  /* replicate last byte */                                                    \
101  memset(r1 + (num_pixels), r1[(num_pixels) - 1], 17 - (num_pixels));          \
102  memset(r2 + (num_pixels), r2[(num_pixels) - 1], 17 - (num_pixels));          \
103  /* using the shared function instead of the macro saves ~3k code size */     \
104  Upsample32Pixels(r1, r2, out);                                               \
105}
106
107#define CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y,                              \
108                    top_dst, bottom_dst, cur_x, num_pixels) {                  \
109  int n;                                                                       \
110  for (n = 0; n < (num_pixels); ++n) {                                         \
111    FUNC(top_y[(cur_x) + n], r_u[n], r_v[n],                                   \
112         top_dst + ((cur_x) + n) * XSTEP);                                     \
113  }                                                                            \
114  if (bottom_y != NULL) {                                                      \
115    for (n = 0; n < (num_pixels); ++n) {                                       \
116      FUNC(bottom_y[(cur_x) + n], r_u[64 + n], r_v[64 + n],                    \
117           bottom_dst + ((cur_x) + n) * XSTEP);                                \
118    }                                                                          \
119  }                                                                            \
120}
121
122#define CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y,                           \
123                       top_dst, bottom_dst, cur_x) do {                        \
124  FUNC##32(top_y + (cur_x), r_u, r_v, top_dst + (cur_x) * XSTEP);              \
125  if (bottom_y != NULL) {                                                      \
126    FUNC##32(bottom_y + (cur_x), r_u + 64, r_v + 64,                           \
127             bottom_dst + (cur_x) * XSTEP);                                    \
128  }                                                                            \
129} while (0)
130
131#define SSE2_UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP)                             \
132static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
133                      const uint8_t* top_u, const uint8_t* top_v,              \
134                      const uint8_t* cur_u, const uint8_t* cur_v,              \
135                      uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
136  int uv_pos, pos;                                                             \
137  /* 16byte-aligned array to cache reconstructed u and v */                    \
138  uint8_t uv_buf[4 * 32 + 15];                                                 \
139  uint8_t* const r_u = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);             \
140  uint8_t* const r_v = r_u + 32;                                               \
141                                                                               \
142  assert(top_y != NULL);                                                       \
143  {   /* Treat the first pixel in regular way */                               \
144    const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                       \
145    const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                       \
146    const int u0_t = (top_u[0] + u_diag) >> 1;                                 \
147    const int v0_t = (top_v[0] + v_diag) >> 1;                                 \
148    FUNC(top_y[0], u0_t, v0_t, top_dst);                                       \
149    if (bottom_y != NULL) {                                                    \
150      const int u0_b = (cur_u[0] + u_diag) >> 1;                               \
151      const int v0_b = (cur_v[0] + v_diag) >> 1;                               \
152      FUNC(bottom_y[0], u0_b, v0_b, bottom_dst);                               \
153    }                                                                          \
154  }                                                                            \
155  /* For UPSAMPLE_32PIXELS, 17 u/v values must be read-able for each block */  \
156  for (pos = 1, uv_pos = 0; pos + 32 + 1 <= len; pos += 32, uv_pos += 16) {    \
157    UPSAMPLE_32PIXELS(top_u + uv_pos, cur_u + uv_pos, r_u);                    \
158    UPSAMPLE_32PIXELS(top_v + uv_pos, cur_v + uv_pos, r_v);                    \
159    CONVERT2RGB_32(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst, pos);    \
160  }                                                                            \
161  if (len > 1) {                                                               \
162    const int left_over = ((len + 1) >> 1) - (pos >> 1);                       \
163    assert(left_over > 0);                                                     \
164    UPSAMPLE_LAST_BLOCK(top_u + uv_pos, cur_u + uv_pos, left_over, r_u);       \
165    UPSAMPLE_LAST_BLOCK(top_v + uv_pos, cur_v + uv_pos, left_over, r_v);       \
166    CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, top_dst, bottom_dst,             \
167                pos, len - pos);                                               \
168  }                                                                            \
169}
170
171// SSE2 variants of the fancy upsampler.
172SSE2_UPSAMPLE_FUNC(UpsampleRgbLinePair,  VP8YuvToRgb,  3)
173SSE2_UPSAMPLE_FUNC(UpsampleBgrLinePair,  VP8YuvToBgr,  3)
174SSE2_UPSAMPLE_FUNC(UpsampleRgbaLinePair, VP8YuvToRgba, 4)
175SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePair, VP8YuvToBgra, 4)
176
177#undef GET_M
178#undef PACK_AND_STORE
179#undef UPSAMPLE_32PIXELS
180#undef UPSAMPLE_LAST_BLOCK
181#undef CONVERT2RGB
182#undef CONVERT2RGB_32
183#undef SSE2_UPSAMPLE_FUNC
184
185#endif  // FANCY_UPSAMPLING
186
187#endif   // WEBP_USE_SSE2
188
189//------------------------------------------------------------------------------
190
192
193#ifdef FANCY_UPSAMPLING
194
195extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
196
198#if defined(WEBP_USE_SSE2)
199  VP8YUVInitSSE2();
200  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePair;
201  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair;
202  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePair;
203  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair;
204  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair;
205  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair;
206#endif   // WEBP_USE_SSE2
207}
208
209#else
210
211// this empty function is to avoid an empty .o