1 | // Copyright 2011 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 to RGB upsampling functions. |
11 | // |
12 | // Author: somnath@google.com (Somnath Banerjee) |
13 | |
14 | #include "src/dsp/dsp.h" |
15 | #include "src/dsp/yuv.h" |
16 | |
17 | #include <assert.h> |
18 | |
19 | //------------------------------------------------------------------------------ |
20 | // Fancy upsampler |
21 | |
22 | #ifdef FANCY_UPSAMPLING |
23 | |
24 | // Fancy upsampling functions to convert YUV to RGB |
25 | WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST]; |
26 | |
27 | // Given samples laid out in a square as: |
28 | // [a b] |
29 | // [c d] |
30 | // we interpolate u/v as: |
31 | // ([9*a + 3*b + 3*c + d 3*a + 9*b + 3*c + d] + [8 8]) / 16 |
32 | // ([3*a + b + 9*c + 3*d a + 3*b + 3*c + 9*d] [8 8]) / 16 |
33 | |
34 | // We process u and v together stashed into 32bit (16bit each). |
35 | #define LOAD_UV(u, v) ((u) | ((v) << 16)) |
36 | |
37 | #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ |
38 | static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y, \ |
39 | const uint8_t* top_u, const uint8_t* top_v, \ |
40 | const uint8_t* cur_u, const uint8_t* cur_v, \ |
41 | uint8_t* top_dst, uint8_t* bottom_dst, int len) { \ |
42 | int x; \ |
43 | const int last_pixel_pair = (len - 1) >> 1; \ |
44 | uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]); /* top-left sample */ \ |
45 | uint32_t l_uv = LOAD_UV(cur_u[0], cur_v[0]); /* left-sample */ \ |
46 | assert(top_y != NULL); \ |
47 | { \ |
48 | const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ |
49 | FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \ |
50 | } \ |
51 | if (bottom_y != NULL) { \ |
52 | const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ |
53 | FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst); \ |
54 | } \ |
55 | for (x = 1; x <= last_pixel_pair; ++x) { \ |
56 | const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]); /* top sample */ \ |
57 | const uint32_t uv = LOAD_UV(cur_u[x], cur_v[x]); /* sample */ \ |
58 | /* precompute invariant values associated with first and second diagonals*/\ |
59 | const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u; \ |
60 | const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3; \ |
61 | const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3; \ |
62 | { \ |
63 | const uint32_t uv0 = (diag_12 + tl_uv) >> 1; \ |
64 | const uint32_t uv1 = (diag_03 + t_uv) >> 1; \ |
65 | FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ |
66 | top_dst + (2 * x - 1) * (XSTEP)); \ |
67 | FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16), \ |
68 | top_dst + (2 * x - 0) * (XSTEP)); \ |
69 | } \ |
70 | if (bottom_y != NULL) { \ |
71 | const uint32_t uv0 = (diag_03 + l_uv) >> 1; \ |
72 | const uint32_t uv1 = (diag_12 + uv) >> 1; \ |
73 | FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ |
74 | bottom_dst + (2 * x - 1) * (XSTEP)); \ |
75 | FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16), \ |
76 | bottom_dst + (2 * x + 0) * (XSTEP)); \ |
77 | } \ |
78 | tl_uv = t_uv; \ |
79 | l_uv = uv; \ |
80 | } \ |
81 | if (!(len & 1)) { \ |
82 | { \ |
83 | const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ |
84 | FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ |
85 | top_dst + (len - 1) * (XSTEP)); \ |
86 | } \ |
87 | if (bottom_y != NULL) { \ |
88 | const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ |
89 | FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ |
90 | bottom_dst + (len - 1) * (XSTEP)); \ |
91 | } \ |
92 | } \ |
93 | } |
94 | |
95 | // All variants implemented. |
96 | #if !WEBP_NEON_OMIT_C_CODE |
97 | UPSAMPLE_FUNC(UpsampleRgbaLinePair_C, VP8YuvToRgba, 4) |
98 | UPSAMPLE_FUNC(UpsampleBgraLinePair_C, VP8YuvToBgra, 4) |
99 | #if !defined(WEBP_REDUCE_CSP) |
100 | UPSAMPLE_FUNC(UpsampleArgbLinePair_C, VP8YuvToArgb, 4) |
101 | UPSAMPLE_FUNC(UpsampleRgbLinePair_C, VP8YuvToRgb, 3) |
102 | UPSAMPLE_FUNC(UpsampleBgrLinePair_C, VP8YuvToBgr, 3) |
103 | UPSAMPLE_FUNC(UpsampleRgba4444LinePair_C, VP8YuvToRgba4444, 2) |
104 | UPSAMPLE_FUNC(UpsampleRgb565LinePair_C, VP8YuvToRgb565, 2) |
105 | #else |
106 | static void EmptyUpsampleFunc(const uint8_t* top_y, const uint8_t* bottom_y, |
107 | const uint8_t* top_u, const uint8_t* top_v, |
108 | const uint8_t* cur_u, const uint8_t* cur_v, |
109 | uint8_t* top_dst, uint8_t* bottom_dst, int len) { |
110 | (void)top_y; |
111 | (void)bottom_y; |
112 | (void)top_u; |
113 | (void)top_v; |
114 | (void)cur_u; |
115 | (void)cur_v; |
116 | (void)top_dst; |
117 | (void)bottom_dst; |
118 | (void)len; |
119 | assert(0); // COLORSPACE SUPPORT NOT COMPILED |
120 | } |
121 | #define UpsampleArgbLinePair_C EmptyUpsampleFunc |
122 | #define UpsampleRgbLinePair_C EmptyUpsampleFunc |
123 | #define UpsampleBgrLinePair_C EmptyUpsampleFunc |
124 | #define UpsampleRgba4444LinePair_C EmptyUpsampleFunc |
125 | #define UpsampleRgb565LinePair_C EmptyUpsampleFunc |
126 | #endif // WEBP_REDUCE_CSP |
127 | |
128 | #endif |
129 | |
130 | #undef LOAD_UV |
131 | #undef UPSAMPLE_FUNC |
132 | |
133 | #endif // FANCY_UPSAMPLING |
134 | |
135 | //------------------------------------------------------------------------------ |
136 | |
137 | #if !defined(FANCY_UPSAMPLING) |
138 | #define DUAL_SAMPLE_FUNC(FUNC_NAME, FUNC) \ |
139 | static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bot_y, \ |
140 | const uint8_t* top_u, const uint8_t* top_v, \ |
141 | const uint8_t* bot_u, const uint8_t* bot_v, \ |
142 | uint8_t* top_dst, uint8_t* bot_dst, int len) { \ |
143 | const int half_len = len >> 1; \ |
144 | int x; \ |
145 | assert(top_dst != NULL); \ |
146 | { \ |
147 | for (x = 0; x < half_len; ++x) { \ |
148 | FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x + 0); \ |
149 | FUNC(top_y[2 * x + 1], top_u[x], top_v[x], top_dst + 8 * x + 4); \ |
150 | } \ |
151 | if (len & 1) FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x); \ |
152 | } \ |
153 | if (bot_dst != NULL) { \ |
154 | for (x = 0; x < half_len; ++x) { \ |
155 | FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x + 0); \ |
156 | FUNC(bot_y[2 * x + 1], bot_u[x], bot_v[x], bot_dst + 8 * x + 4); \ |
157 | } \ |
158 | if (len & 1) FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x); \ |
159 | } \ |
160 | } |
161 | |
162 | DUAL_SAMPLE_FUNC(DualLineSamplerBGRA, VP8YuvToBgra) |
163 | DUAL_SAMPLE_FUNC(DualLineSamplerARGB, VP8YuvToArgb) |
164 | #undef DUAL_SAMPLE_FUNC |
165 | |
166 | #endif // !FANCY_UPSAMPLING |
167 | |
168 | WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) { |
169 | WebPInitUpsamplers(); |
170 | #ifdef FANCY_UPSAMPLING |
171 | return WebPUpsamplers[alpha_is_last ? MODE_BGRA : MODE_ARGB]; |
172 | #else |
173 | return (alpha_is_last ? DualLineSamplerBGRA : DualLineSamplerARGB); |
174 | #endif |
175 | } |
176 | |
177 | //------------------------------------------------------------------------------ |
178 | // YUV444 converter |
179 | |
180 | #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP) \ |
181 | extern void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ |
182 | uint8_t* dst, int len); \ |
183 | void FUNC_NAME(const uint8_t* y, const uint8_t* u, const uint8_t* v, \ |
184 | uint8_t* dst, int len) { \ |
185 | int i; \ |
186 | for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * (XSTEP)]); \ |
187 | } |
188 | |
189 | YUV444_FUNC(WebPYuv444ToRgba_C, VP8YuvToRgba, 4) |
190 | YUV444_FUNC(WebPYuv444ToBgra_C, VP8YuvToBgra, 4) |
191 | #if !defined(WEBP_REDUCE_CSP) |
192 | YUV444_FUNC(WebPYuv444ToRgb_C, VP8YuvToRgb, 3) |
193 | YUV444_FUNC(WebPYuv444ToBgr_C, VP8YuvToBgr, 3) |
194 | YUV444_FUNC(WebPYuv444ToArgb_C, VP8YuvToArgb, 4) |
195 | YUV444_FUNC(WebPYuv444ToRgba4444_C, VP8YuvToRgba4444, 2) |
196 | YUV444_FUNC(WebPYuv444ToRgb565_C, VP8YuvToRgb565, 2) |
197 | #else |
198 | static void EmptyYuv444Func(const uint8_t* y, |
199 | const uint8_t* u, const uint8_t* v, |
200 | uint8_t* dst, int len) { |
201 | (void)y; |
202 | (void)u; |
203 | (void)v; |
204 | (void)dst; |
205 | (void)len; |
206 | } |
207 | #define WebPYuv444ToRgb_C EmptyYuv444Func |
208 | #define WebPYuv444ToBgr_C EmptyYuv444Func |
209 | #define WebPYuv444ToArgb_C EmptyYuv444Func |
210 | #define WebPYuv444ToRgba4444_C EmptyYuv444Func |
211 | #define WebPYuv444ToRgb565_C EmptyYuv444Func |
212 | #endif // WEBP_REDUCE_CSP |
213 | |
214 | #undef YUV444_FUNC |
215 | |
216 | WebPYUV444Converter WebPYUV444Converters[MODE_LAST]; |
217 | |
218 | extern VP8CPUInfo VP8GetCPUInfo; |
219 | extern void WebPInitYUV444ConvertersMIPSdspR2(void); |
220 | extern void (void); |
221 | extern void (void); |
222 | |
223 | WEBP_DSP_INIT_FUNC(WebPInitYUV444Converters) { |
224 | WebPYUV444Converters[MODE_RGBA] = WebPYuv444ToRgba_C; |
225 | WebPYUV444Converters[MODE_BGRA] = WebPYuv444ToBgra_C; |
226 | WebPYUV444Converters[MODE_RGB] = WebPYuv444ToRgb_C; |
227 | WebPYUV444Converters[MODE_BGR] = WebPYuv444ToBgr_C; |
228 | WebPYUV444Converters[MODE_ARGB] = WebPYuv444ToArgb_C; |
229 | WebPYUV444Converters[MODE_RGBA_4444] = WebPYuv444ToRgba4444_C; |
230 | WebPYUV444Converters[MODE_RGB_565] = WebPYuv444ToRgb565_C; |
231 | WebPYUV444Converters[MODE_rgbA] = WebPYuv444ToRgba_C; |
232 | WebPYUV444Converters[MODE_bgrA] = WebPYuv444ToBgra_C; |
233 | WebPYUV444Converters[MODE_Argb] = WebPYuv444ToArgb_C; |
234 | WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444_C; |
235 | |
236 | if (VP8GetCPUInfo != NULL) { |
237 | #if defined(WEBP_HAVE_SSE2) |
238 | if (VP8GetCPUInfo(kSSE2)) { |
239 | WebPInitYUV444ConvertersSSE2(); |
240 | } |
241 | #endif |
242 | #if defined(WEBP_HAVE_SSE41) |
243 | if (VP8GetCPUInfo(kSSE4_1)) { |
244 | WebPInitYUV444ConvertersSSE41(); |
245 | } |
246 | #endif |
247 | #if defined(WEBP_USE_MIPS_DSP_R2) |
248 | if (VP8GetCPUInfo(kMIPSdspR2)) { |
249 | WebPInitYUV444ConvertersMIPSdspR2(); |
250 | } |
251 | #endif |
252 | } |
253 | } |
254 | |
255 | //------------------------------------------------------------------------------ |
256 | // Main calls |
257 | |
258 | extern void (void); |
259 | extern void (void); |
260 | extern void WebPInitUpsamplersNEON(void); |
261 | extern void WebPInitUpsamplersMIPSdspR2(void); |
262 | extern void WebPInitUpsamplersMSA(void); |
263 | |
264 | WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { |
265 | #ifdef FANCY_UPSAMPLING |
266 | #if !WEBP_NEON_OMIT_C_CODE |
267 | WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_C; |
268 | WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_C; |
269 | WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_C; |
270 | WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_C; |
271 | WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair_C; |
272 | WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair_C; |
273 | WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_C; |
274 | WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_C; |
275 | WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_C; |
276 | WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_C; |
277 | WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_C; |
278 | #endif |
279 | |
280 | // If defined, use CPUInfo() to overwrite some pointers with faster versions. |
281 | if (VP8GetCPUInfo != NULL) { |
282 | #if defined(WEBP_HAVE_SSE2) |
283 | if (VP8GetCPUInfo(kSSE2)) { |
284 | WebPInitUpsamplersSSE2(); |
285 | } |
286 | #endif |
287 | #if defined(WEBP_HAVE_SSE41) |
288 | if (VP8GetCPUInfo(kSSE4_1)) { |
289 | WebPInitUpsamplersSSE41(); |
290 | } |
291 | #endif |
292 | #if defined(WEBP_USE_MIPS_DSP_R2) |
293 | if (VP8GetCPUInfo(kMIPSdspR2)) { |
294 | WebPInitUpsamplersMIPSdspR2(); |
295 | } |
296 | #endif |
297 | #if defined(WEBP_USE_MSA) |
298 | if (VP8GetCPUInfo(kMSA)) { |
299 | WebPInitUpsamplersMSA(); |
300 | } |
301 | #endif |
302 | } |
303 | |
304 | #if defined(WEBP_HAVE_NEON) |
305 | if (WEBP_NEON_OMIT_C_CODE || |
306 | (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { |
307 | WebPInitUpsamplersNEON(); |
308 | } |
309 | #endif |
310 | |
311 | assert(WebPUpsamplers[MODE_RGBA] != NULL); |
312 | assert(WebPUpsamplers[MODE_BGRA] != NULL); |
313 | assert(WebPUpsamplers[MODE_rgbA] != NULL); |
314 | assert(WebPUpsamplers[MODE_bgrA] != NULL); |
315 | #if !defined(WEBP_REDUCE_CSP) || !WEBP_NEON_OMIT_C_CODE |
316 | assert(WebPUpsamplers[MODE_RGB] != NULL); |
317 | assert(WebPUpsamplers[MODE_BGR] != NULL); |
318 | assert(WebPUpsamplers[MODE_ARGB] != NULL); |
319 | assert(WebPUpsamplers[MODE_RGBA_4444] != NULL); |
320 | assert(WebPUpsamplers[MODE_RGB_565] != NULL); |
321 | assert(WebPUpsamplers[MODE_Argb] != NULL); |
322 | assert(WebPUpsamplers[MODE_rgbA_4444] != NULL); |
323 | #endif |
324 | |
325 | #endif // FANCY_UPSAMPLING |
326 | } |
327 | |
328 | //------------------------------------------------------------------------------ |
329 | |