1// Copyright 2014 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// YUV->RGB conversion functions
11//
12// Author: Skal (pascal.massimino@gmail.com)
13
14#include "./yuv.h"
15
16#if defined(WEBP_USE_SSE2)
17
18#include <emmintrin.h>
19#include <string.h>   // for memcpy
20
21typedef union {   // handy struct for converting SSE2 registers
22  int32_t i32[4];
23  uint8_t u8[16];
24  __m128i m;
25} VP8kCstSSE2;
26
27#if defined(WEBP_YUV_USE_SSE2_TABLES)
28
29#include "./yuv_tables_sse2.h"
30
31void VP8YUVInitSSE2(void) {}
32
33#else
34
35static int done_sse2 = 0;
36static VP8kCstSSE2 VP8kUtoRGBA[256], VP8kVtoRGBA[256], VP8kYtoRGBA[256];
37
38void VP8YUVInitSSE2(void) {
39  if (!done_sse2) {
40    int i;
41    for (i = 0; i < 256; ++i) {
42      VP8kYtoRGBA[i].i32[0] =
43        VP8kYtoRGBA[i].i32[1] =
44        VP8kYtoRGBA[i].i32[2] = (i - 16) * kYScale + YUV_HALF2;
45      VP8kYtoRGBA[i].i32[3] = 0xff << YUV_FIX2;
46
47      VP8kUtoRGBA[i].i32[0] = 0;
48      VP8kUtoRGBA[i].i32[1] = -kUToG * (i - 128);
49      VP8kUtoRGBA[i].i32[2] =  kUToB * (i - 128);
50      VP8kUtoRGBA[i].i32[3] = 0;
51
52      VP8kVtoRGBA[i].i32[0] =  kVToR * (i - 128);
53      VP8kVtoRGBA[i].i32[1] = -kVToG * (i - 128);
54      VP8kVtoRGBA[i].i32[2] = 0;
55      VP8kVtoRGBA[i].i32[3] = 0;
56    }
57    done_sse2 = 1;
58
59#if 0   // code used to generate 'yuv_tables_sse2.h'
60    printf("static const VP8kCstSSE2 VP8kYtoRGBA[256] = {\n");
61    for (i = 0; i < 256; ++i) {
62      printf("  {{0x%.8x, 0x%.8x, 0x%.8x, 0x%.8x}},\n",
63             VP8kYtoRGBA[i].i32[0], VP8kYtoRGBA[i].i32[1],
64             VP8kYtoRGBA[i].i32[2], VP8kYtoRGBA[i].i32[3]);
65    }
66    printf("};\n\n");
67    printf("static const VP8kCstSSE2 VP8kUtoRGBA[256] = {\n");
68    for (i = 0; i < 256; ++i) {
69      printf("  {{0, 0x%.8x, 0x%.8x, 0}},\n",
70             VP8kUtoRGBA[i].i32[1], VP8kUtoRGBA[i].i32[2]);
71    }
72    printf("};\n\n");
73    printf("static VP8kCstSSE2 VP8kVtoRGBA[256] = {\n");
74    for (i = 0; i < 256; ++i) {
75      printf("  {{0x%.8x, 0x%.8x, 0, 0}},\n",
76             VP8kVtoRGBA[i].i32[0], VP8kVtoRGBA[i].i32[1]);
77    }
78    printf("};\n\n");
79#endif
80  }
81}
82
83#endif  // WEBP_YUV_USE_SSE2_TABLES
84
85//-----------------------------------------------------------------------------
86
87static WEBP_INLINE __m128i LoadUVPart(int u, int v) {
88  const __m128i u_part = _mm_loadu_si128(&VP8kUtoRGBA[u].m);
89  const __m128i v_part = _mm_loadu_si128(&VP8kVtoRGBA[v].m);
90  const __m128i uv_part = _mm_add_epi32(u_part, v_part);
91  return uv_part;
92}
93
94static WEBP_INLINE __m128i GetRGBA32bWithUV(int y, const __m128i uv_part) {
95  const __m128i y_part = _mm_loadu_si128(&VP8kYtoRGBA[y].m);
96  const __m128i rgba1 = _mm_add_epi32(y_part, uv_part);
97  const __m128i rgba2 = _mm_srai_epi32(rgba1, YUV_FIX2);
98  return rgba2;
99}
100
101static WEBP_INLINE __m128i GetRGBA32b(int y, int u, int v) {
102  const __m128i uv_part = LoadUVPart(u, v);
103  return GetRGBA32bWithUV(y, uv_part);
104}
105
106static WEBP_INLINE void YuvToRgbSSE2(uint8_t y, uint8_t u, uint8_t v,
107                                     uint8_t* const rgb) {
108  const __m128i tmp0 = GetRGBA32b(y, u, v);
109  const __m128i tmp1 = _mm_packs_epi32(tmp0, tmp0);
110  const __m128i tmp2 = _mm_packus_epi16(tmp1, tmp1);
111  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
112  _mm_storel_epi64((__m128i*)rgb, tmp2);
113}
114
115static WEBP_INLINE void YuvToBgrSSE2(uint8_t y, uint8_t u, uint8_t v,
116                                     uint8_t* const bgr) {
117  const __m128i tmp0 = GetRGBA32b(y, u, v);
118  const __m128i tmp1 = _mm_shuffle_epi32(tmp0, _MM_SHUFFLE(3, 0, 1, 2));
119  const __m128i tmp2 = _mm_packs_epi32(tmp1, tmp1);
120  const __m128i tmp3 = _mm_packus_epi16(tmp2, tmp2);
121  // Note: we store 8 bytes at a time, not 3 bytes! -> memory stomp
122  _mm_storel_epi64((__m128i*)bgr, tmp3);
123}
124
125//-----------------------------------------------------------------------------
126// Convert spans of 32 pixels to various RGB formats for the fancy upsampler.
127
128#ifdef FANCY_UPSAMPLING
129
130void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
131                    uint8_t* dst) {
132  int n;
133  for (n = 0; n < 32; n += 4) {
134    const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
135    const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
136    const __m128i tmp0_3 = GetRGBA32b(y[n + 2], u[n + 2], v[n + 2]);
137    const __m128i tmp0_4 = GetRGBA32b(y[n + 3], u[n + 3], v[n + 3]);
138    const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
139    const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
140    const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
141    _mm_storeu_si128((__m128i*)dst, tmp2);
142    dst += 4 * 4;
143  }
144}
145
146void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
147                    uint8_t* dst) {
148  int n;
149  for (n = 0; n < 32; n += 2) {
150    const __m128i tmp0_1 = GetRGBA32b(y[n + 0], u[n + 0], v[n + 0]);
151    const __m128i tmp0_2 = GetRGBA32b(y[n + 1], u[n + 1], v[n + 1]);
152    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
153    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
154    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
155    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
156    _mm_storel_epi64((__m128i*)dst, tmp3);
157    dst += 4 * 2;
158  }
159}
160
161void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
162                   uint8_t* dst) {
163  int n;
164  uint8_t tmp0[2 * 3 + 5 + 15];
165  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
166  for (n = 0; n < 30; ++n) {   // we directly stomp the *dst memory
167    YuvToRgbSSE2(y[n], u[n], v[n], dst + n * 3);
168  }
169  // Last two pixels are special: we write in a tmp buffer before sending
170  // to dst.
171  YuvToRgbSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
172  YuvToRgbSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
173  memcpy(dst + n * 3, tmp, 2 * 3);
174}
175
176void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v,
177                   uint8_t* dst) {
178  int n;
179  uint8_t tmp0[2 * 3 + 5 + 15];
180  uint8_t* const tmp = (uint8_t*)((uintptr_t)(tmp0 + 15) & ~15);  // align
181  for (n = 0; n < 30; ++n) {
182    YuvToBgrSSE2(y[n], u[n], v[n], dst + n * 3);
183  }
184  YuvToBgrSSE2(y[n + 0], u[n + 0], v[n + 0], tmp + 0);
185  YuvToBgrSSE2(y[n + 1], u[n + 1], v[n + 1], tmp + 3);
186  memcpy(dst + n * 3, tmp, 2 * 3);
187}
188
189#endif  // FANCY_UPSAMPLING
190
191//-----------------------------------------------------------------------------
192// Arbitrary-length row conversion functions
193
194static void YuvToRgbaRowSSE2(const uint8_t* y,
195                             const uint8_t* u, const uint8_t* v,
196                             uint8_t* dst, int len) {
197  int n;
198  for (n = 0; n + 4 <= len; n += 4) {
199    const __m128i uv_0 = LoadUVPart(u[0], v[0]);
200    const __m128i uv_1 = LoadUVPart(u[1], v[1]);
201    const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
202    const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
203    const __m128i tmp0_3 = GetRGBA32bWithUV(y[2], uv_1);
204    const __m128i tmp0_4 = GetRGBA32bWithUV(y[3], uv_1);
205    const __m128i tmp1_1 = _mm_packs_epi32(tmp0_1, tmp0_2);
206    const __m128i tmp1_2 = _mm_packs_epi32(tmp0_3, tmp0_4);
207    const __m128i tmp2 = _mm_packus_epi16(tmp1_1, tmp1_2);
208    _mm_storeu_si128((__m128i*)dst, tmp2);
209    dst += 4 * 4;
210    y += 4;
211    u += 2;
212    v += 2;
213  }
214  // Finish off
215  while (n < len) {
216    VP8YuvToRgba(y[0], u[0], v[0], dst);
217    dst += 4;
218    ++y;
219    u += (n & 1);
220    v += (n & 1);
221    ++n;
222  }
223}
224
225static void YuvToBgraRowSSE2(const uint8_t* y,
226                             const uint8_t* u, const uint8_t* v,
227                             uint8_t* dst, int len) {
228  int n;
229  for (n = 0; n + 2 <= len; n += 2) {
230    const __m128i uv_0 = LoadUVPart(u[0], v[0]);
231    const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
232    const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
233    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(3, 0, 1, 2));
234    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(3, 0, 1, 2));
235    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
236    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
237    _mm_storel_epi64((__m128i*)dst, tmp3);
238    dst += 4 * 2;
239    y += 2;
240    ++u;
241    ++v;
242  }
243  // Finish off
244  if (len & 1) {
245    VP8YuvToBgra(y[0], u[0], v[0], dst);
246  }
247}
248
249static void YuvToArgbRowSSE2(const uint8_t* y,
250                             const uint8_t* u, const uint8_t* v,
251                             uint8_t* dst, int len) {
252  int n;
253  for (n = 0; n + 2 <= len; n += 2) {
254    const __m128i uv_0 = LoadUVPart(u[0], v[0]);
255    const __m128i tmp0_1 = GetRGBA32bWithUV(y[0], uv_0);
256    const __m128i tmp0_2 = GetRGBA32bWithUV(y[1], uv_0);
257    const __m128i tmp1_1 = _mm_shuffle_epi32(tmp0_1, _MM_SHUFFLE(2, 1, 0, 3));
258    const __m128i tmp1_2 = _mm_shuffle_epi32(tmp0_2, _MM_SHUFFLE(2, 1, 0, 3));
259    const __m128i tmp2_1 = _mm_packs_epi32(tmp1_1, tmp1_2);
260    const __m128i tmp3 = _mm_packus_epi16(tmp2_1, tmp2_1);
261    _mm_storel_epi64((__m128i*)dst, tmp3);
262    dst += 4 * 2;
263    y += 2;
264    ++u;
265    ++v;
266  }
267  // Finish off
268  if (len & 1) {
269    VP8YuvToArgb(y[0], u[0], v[0], dst);
270  }
271}
272
273static void YuvToRgbRowSSE2(const uint8_t* y,
274                            const uint8_t* u, const uint8_t* v,
275                            uint8_t* dst, int len) {
276  int n;
277  for (n = 0; n + 2 < len; ++n) {   // we directly stomp the *dst memory
278    YuvToRgbSSE2(y[0], u[0], v[0], dst);  // stomps 8 bytes
279    dst += 3;
280    ++y;
281    u += (n & 1);
282    v += (n & 1);
283  }
284  VP8YuvToRgb(y[0], u[0], v[0], dst);
285  if (len > 1) {
286    VP8YuvToRgb(y[1], u[n & 1], v[n & 1], dst + 3);
287  }
288}
289
290static void YuvToBgrRowSSE2(const uint8_t* y,
291                            const uint8_t* u, const uint8_t* v,
292                            uint8_t* dst, int len) {
293  int n;
294  for (n = 0; n + 2 < len; ++n) {   // we directly stomp the *dst memory
295    YuvToBgrSSE2(y[0], u[0], v[0], dst);  // stomps 8 bytes
296    dst += 3;
297    ++y;
298    u += (n & 1);
299    v += (n & 1);
300  }
301  VP8YuvToBgr(y[0], u[0], v[0], dst + 0);
302  if (len > 1) {
303    VP8YuvToBgr(y[1], u[n & 1], v[n & 1], dst + 3);
304  }
305}
306
307#endif  // WEBP_USE_SSE2
308
309//------------------------------------------------------------------------------
310// Entry point
311
312extern void WebPInitSamplersSSE2(void);
313
314void WebPInitSamplersSSE2(void) {
315#if defined(WEBP_USE_SSE2)
316  WebPSamplers[MODE_RGB]  = YuvToRgbRowSSE2;
317  WebPSamplers[MODE_RGBA] = YuvToRgbaRowSSE2;
318  WebPSamplers[MODE_BGR]  = YuvToBgrRowSSE2;
319  WebPSamplers[MODE_BGRA] = YuvToBgraRowSSE2;
320  WebPSamplers[MODE_ARGB] = YuvToArgbRowSSE2;
321#endif  // WEBP_USE_SSE2
322}
323