/src/libwebp/src/dsp/upsampling.c
Line | Count | Source |
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 <assert.h> |
15 | | #include <stddef.h> |
16 | | |
17 | | #include "src/dsp/cpu.h" |
18 | | #include "src/dsp/dsp.h" |
19 | | #include "src/dsp/yuv.h" |
20 | | #include "src/webp/decode.h" |
21 | | #include "src/webp/types.h" |
22 | | |
23 | | //------------------------------------------------------------------------------ |
24 | | // Fancy upsampler |
25 | | |
26 | | #ifdef FANCY_UPSAMPLING |
27 | | |
28 | | // Fancy upsampling functions to convert YUV to RGB |
29 | | WebPUpsampleLinePairFunc WebPUpsamplers[MODE_LAST]; |
30 | | |
31 | | // Given samples laid out in a square as: |
32 | | // [a b] |
33 | | // [c d] |
34 | | // we interpolate u/v as: |
35 | | // ([9*a + 3*b + 3*c + d 3*a + 9*b + 3*c + d] + [8 8]) / 16 |
36 | | // ([3*a + b + 9*c + 3*d a + 3*b + 3*c + 9*d] [8 8]) / 16 |
37 | | |
38 | | // We process u and v together stashed into 32bit (16bit each). |
39 | 0 | #define LOAD_UV(u, v) ((u) | ((v) << 16)) |
40 | | |
41 | | #define UPSAMPLE_FUNC(FUNC_NAME, FUNC, XSTEP) \ |
42 | | static void FUNC_NAME( \ |
43 | | const uint8_t* WEBP_RESTRICT top_y, \ |
44 | | const uint8_t* WEBP_RESTRICT bottom_y, \ |
45 | | const uint8_t* WEBP_RESTRICT top_u, const uint8_t* WEBP_RESTRICT top_v, \ |
46 | | const uint8_t* WEBP_RESTRICT cur_u, const uint8_t* WEBP_RESTRICT cur_v, \ |
47 | | uint8_t* WEBP_RESTRICT top_dst, uint8_t* WEBP_RESTRICT bottom_dst, \ |
48 | 0 | int len) { \ |
49 | 0 | int x; \ |
50 | 0 | const int last_pixel_pair = (len - 1) >> 1; \ |
51 | 0 | uint32_t tl_uv = LOAD_UV(top_u[0], top_v[0]); /* top-left sample */ \ |
52 | 0 | uint32_t l_uv = LOAD_UV(cur_u[0], cur_v[0]); /* left-sample */ \ |
53 | 0 | assert(top_y != NULL); \ |
54 | 0 | { \ |
55 | 0 | const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ |
56 | 0 | FUNC(top_y[0], uv0 & 0xff, (uv0 >> 16), top_dst); \ |
57 | 0 | } \ |
58 | 0 | if (bottom_y != NULL) { \ |
59 | 0 | const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ |
60 | 0 | FUNC(bottom_y[0], uv0 & 0xff, (uv0 >> 16), bottom_dst); \ |
61 | 0 | } \ |
62 | 0 | for (x = 1; x <= last_pixel_pair; ++x) { \ |
63 | 0 | const uint32_t t_uv = LOAD_UV(top_u[x], top_v[x]); /* top sample */ \ |
64 | 0 | const uint32_t uv = LOAD_UV(cur_u[x], cur_v[x]); /* sample */ \ |
65 | 0 | /* precompute invariant values associated with first and second \ |
66 | 0 | * diagonals*/ \ |
67 | 0 | const uint32_t avg = tl_uv + t_uv + l_uv + uv + 0x00080008u; \ |
68 | 0 | const uint32_t diag_12 = (avg + 2 * (t_uv + l_uv)) >> 3; \ |
69 | 0 | const uint32_t diag_03 = (avg + 2 * (tl_uv + uv)) >> 3; \ |
70 | 0 | { \ |
71 | 0 | const uint32_t uv0 = (diag_12 + tl_uv) >> 1; \ |
72 | 0 | const uint32_t uv1 = (diag_03 + t_uv) >> 1; \ |
73 | 0 | FUNC(top_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ |
74 | 0 | top_dst + (2 * x - 1) * (XSTEP)); \ |
75 | 0 | FUNC(top_y[2 * x - 0], uv1 & 0xff, (uv1 >> 16), \ |
76 | 0 | top_dst + (2 * x - 0) * (XSTEP)); \ |
77 | 0 | } \ |
78 | 0 | if (bottom_y != NULL) { \ |
79 | 0 | const uint32_t uv0 = (diag_03 + l_uv) >> 1; \ |
80 | 0 | const uint32_t uv1 = (diag_12 + uv) >> 1; \ |
81 | 0 | FUNC(bottom_y[2 * x - 1], uv0 & 0xff, (uv0 >> 16), \ |
82 | 0 | bottom_dst + (2 * x - 1) * (XSTEP)); \ |
83 | 0 | FUNC(bottom_y[2 * x + 0], uv1 & 0xff, (uv1 >> 16), \ |
84 | 0 | bottom_dst + (2 * x + 0) * (XSTEP)); \ |
85 | 0 | } \ |
86 | 0 | tl_uv = t_uv; \ |
87 | 0 | l_uv = uv; \ |
88 | 0 | } \ |
89 | 0 | if (!(len & 1)) { \ |
90 | 0 | { \ |
91 | 0 | const uint32_t uv0 = (3 * tl_uv + l_uv + 0x00020002u) >> 2; \ |
92 | 0 | FUNC(top_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ |
93 | 0 | top_dst + (len - 1) * (XSTEP)); \ |
94 | 0 | } \ |
95 | 0 | if (bottom_y != NULL) { \ |
96 | 0 | const uint32_t uv0 = (3 * l_uv + tl_uv + 0x00020002u) >> 2; \ |
97 | 0 | FUNC(bottom_y[len - 1], uv0 & 0xff, (uv0 >> 16), \ |
98 | 0 | bottom_dst + (len - 1) * (XSTEP)); \ |
99 | 0 | } \ |
100 | 0 | } \ |
101 | 0 | } |
102 | | |
103 | | // All variants implemented. |
104 | | #if !WEBP_NEON_OMIT_C_CODE |
105 | 0 | UPSAMPLE_FUNC(UpsampleRgbaLinePair_C, VP8YuvToRgba, 4) |
106 | 0 | UPSAMPLE_FUNC(UpsampleBgraLinePair_C, VP8YuvToBgra, 4) |
107 | | #if !defined(WEBP_REDUCE_CSP) |
108 | 0 | UPSAMPLE_FUNC(UpsampleArgbLinePair_C, VP8YuvToArgb, 4) |
109 | 0 | UPSAMPLE_FUNC(UpsampleRgbLinePair_C, VP8YuvToRgb, 3) |
110 | 0 | UPSAMPLE_FUNC(UpsampleBgrLinePair_C, VP8YuvToBgr, 3) |
111 | 0 | UPSAMPLE_FUNC(UpsampleRgba4444LinePair_C, VP8YuvToRgba4444, 2) |
112 | 0 | UPSAMPLE_FUNC(UpsampleRgb565LinePair_C, VP8YuvToRgb565, 2) |
113 | | #else |
114 | | static void EmptyUpsampleFunc(const uint8_t* top_y, const uint8_t* bottom_y, |
115 | | const uint8_t* top_u, const uint8_t* top_v, |
116 | | const uint8_t* cur_u, const uint8_t* cur_v, |
117 | | uint8_t* top_dst, uint8_t* bottom_dst, int len) { |
118 | | (void)top_y; |
119 | | (void)bottom_y; |
120 | | (void)top_u; |
121 | | (void)top_v; |
122 | | (void)cur_u; |
123 | | (void)cur_v; |
124 | | (void)top_dst; |
125 | | (void)bottom_dst; |
126 | | (void)len; |
127 | | assert(0); // COLORSPACE SUPPORT NOT COMPILED |
128 | | } |
129 | | #define UpsampleArgbLinePair_C EmptyUpsampleFunc |
130 | | #define UpsampleRgbLinePair_C EmptyUpsampleFunc |
131 | | #define UpsampleBgrLinePair_C EmptyUpsampleFunc |
132 | | #define UpsampleRgba4444LinePair_C EmptyUpsampleFunc |
133 | | #define UpsampleRgb565LinePair_C EmptyUpsampleFunc |
134 | | #endif // WEBP_REDUCE_CSP |
135 | | |
136 | | #endif |
137 | | |
138 | | #undef LOAD_UV |
139 | | #undef UPSAMPLE_FUNC |
140 | | |
141 | | #endif // FANCY_UPSAMPLING |
142 | | |
143 | | //------------------------------------------------------------------------------ |
144 | | |
145 | | #if !defined(FANCY_UPSAMPLING) |
146 | | #define DUAL_SAMPLE_FUNC(FUNC_NAME, FUNC) \ |
147 | | static void FUNC_NAME( \ |
148 | | const uint8_t* WEBP_RESTRICT top_y, const uint8_t* WEBP_RESTRICT bot_y, \ |
149 | | const uint8_t* WEBP_RESTRICT top_u, const uint8_t* WEBP_RESTRICT top_v, \ |
150 | | const uint8_t* WEBP_RESTRICT bot_u, const uint8_t* WEBP_RESTRICT bot_v, \ |
151 | | uint8_t* WEBP_RESTRICT top_dst, uint8_t* WEBP_RESTRICT bot_dst, \ |
152 | | int len) { \ |
153 | | const int half_len = len >> 1; \ |
154 | | int x; \ |
155 | | assert(top_dst != NULL); \ |
156 | | { \ |
157 | | for (x = 0; x < half_len; ++x) { \ |
158 | | FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x + 0); \ |
159 | | FUNC(top_y[2 * x + 1], top_u[x], top_v[x], top_dst + 8 * x + 4); \ |
160 | | } \ |
161 | | if (len & 1) \ |
162 | | FUNC(top_y[2 * x + 0], top_u[x], top_v[x], top_dst + 8 * x); \ |
163 | | } \ |
164 | | if (bot_dst != NULL) { \ |
165 | | for (x = 0; x < half_len; ++x) { \ |
166 | | FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x + 0); \ |
167 | | FUNC(bot_y[2 * x + 1], bot_u[x], bot_v[x], bot_dst + 8 * x + 4); \ |
168 | | } \ |
169 | | if (len & 1) \ |
170 | | FUNC(bot_y[2 * x + 0], bot_u[x], bot_v[x], bot_dst + 8 * x); \ |
171 | | } \ |
172 | | } |
173 | | |
174 | | DUAL_SAMPLE_FUNC(DualLineSamplerBGRA, VP8YuvToBgra) |
175 | | DUAL_SAMPLE_FUNC(DualLineSamplerARGB, VP8YuvToArgb) |
176 | | #undef DUAL_SAMPLE_FUNC |
177 | | |
178 | | #endif // !FANCY_UPSAMPLING |
179 | | |
180 | 0 | WebPUpsampleLinePairFunc WebPGetLinePairConverter(int alpha_is_last) { |
181 | 0 | WebPInitUpsamplers(); |
182 | 0 | #ifdef FANCY_UPSAMPLING |
183 | 0 | return WebPUpsamplers[alpha_is_last ? MODE_BGRA : MODE_ARGB]; |
184 | | #else |
185 | | return (alpha_is_last ? DualLineSamplerBGRA : DualLineSamplerARGB); |
186 | | #endif |
187 | 0 | } |
188 | | |
189 | | //------------------------------------------------------------------------------ |
190 | | // YUV444 converter |
191 | | |
192 | | #define YUV444_FUNC(FUNC_NAME, FUNC, XSTEP) \ |
193 | | extern void FUNC_NAME( \ |
194 | | const uint8_t* WEBP_RESTRICT y, const uint8_t* WEBP_RESTRICT u, \ |
195 | | const uint8_t* WEBP_RESTRICT v, uint8_t* WEBP_RESTRICT dst, int len); \ |
196 | | void FUNC_NAME( \ |
197 | | const uint8_t* WEBP_RESTRICT y, const uint8_t* WEBP_RESTRICT u, \ |
198 | 0 | const uint8_t* WEBP_RESTRICT v, uint8_t* WEBP_RESTRICT dst, int len) { \ |
199 | 0 | int i; \ |
200 | 0 | for (i = 0; i < len; ++i) FUNC(y[i], u[i], v[i], &dst[i * (XSTEP)]); \ |
201 | 0 | } Unexecuted instantiation: WebPYuv444ToRgba_C Unexecuted instantiation: WebPYuv444ToBgra_C Unexecuted instantiation: WebPYuv444ToRgb_C Unexecuted instantiation: WebPYuv444ToBgr_C Unexecuted instantiation: WebPYuv444ToArgb_C Unexecuted instantiation: WebPYuv444ToRgba4444_C Unexecuted instantiation: WebPYuv444ToRgb565_C |
202 | | |
203 | | YUV444_FUNC(WebPYuv444ToRgba_C, VP8YuvToRgba, 4) |
204 | | YUV444_FUNC(WebPYuv444ToBgra_C, VP8YuvToBgra, 4) |
205 | | #if !defined(WEBP_REDUCE_CSP) |
206 | | YUV444_FUNC(WebPYuv444ToRgb_C, VP8YuvToRgb, 3) |
207 | | YUV444_FUNC(WebPYuv444ToBgr_C, VP8YuvToBgr, 3) |
208 | | YUV444_FUNC(WebPYuv444ToArgb_C, VP8YuvToArgb, 4) |
209 | | YUV444_FUNC(WebPYuv444ToRgba4444_C, VP8YuvToRgba4444, 2) |
210 | | YUV444_FUNC(WebPYuv444ToRgb565_C, VP8YuvToRgb565, 2) |
211 | | #else |
212 | | static void EmptyYuv444Func(const uint8_t* y, const uint8_t* u, |
213 | | const uint8_t* v, uint8_t* dst, int len) { |
214 | | (void)y; |
215 | | (void)u; |
216 | | (void)v; |
217 | | (void)dst; |
218 | | (void)len; |
219 | | } |
220 | | #define WebPYuv444ToRgb_C EmptyYuv444Func |
221 | | #define WebPYuv444ToBgr_C EmptyYuv444Func |
222 | | #define WebPYuv444ToArgb_C EmptyYuv444Func |
223 | | #define WebPYuv444ToRgba4444_C EmptyYuv444Func |
224 | | #define WebPYuv444ToRgb565_C EmptyYuv444Func |
225 | | #endif // WEBP_REDUCE_CSP |
226 | | |
227 | | #undef YUV444_FUNC |
228 | | |
229 | | WebPYUV444Converter WebPYUV444Converters[MODE_LAST]; |
230 | | |
231 | | extern VP8CPUInfo VP8GetCPUInfo; |
232 | | extern void WebPInitYUV444ConvertersMIPSdspR2(void); |
233 | | extern void WebPInitYUV444ConvertersSSE2(void); |
234 | | extern void WebPInitYUV444ConvertersSSE41(void); |
235 | | |
236 | 0 | WEBP_DSP_INIT_FUNC(WebPInitYUV444Converters) { |
237 | 0 | WebPYUV444Converters[MODE_RGBA] = WebPYuv444ToRgba_C; |
238 | 0 | WebPYUV444Converters[MODE_BGRA] = WebPYuv444ToBgra_C; |
239 | 0 | WebPYUV444Converters[MODE_RGB] = WebPYuv444ToRgb_C; |
240 | 0 | WebPYUV444Converters[MODE_BGR] = WebPYuv444ToBgr_C; |
241 | 0 | WebPYUV444Converters[MODE_ARGB] = WebPYuv444ToArgb_C; |
242 | 0 | WebPYUV444Converters[MODE_RGBA_4444] = WebPYuv444ToRgba4444_C; |
243 | 0 | WebPYUV444Converters[MODE_RGB_565] = WebPYuv444ToRgb565_C; |
244 | 0 | WebPYUV444Converters[MODE_rgbA] = WebPYuv444ToRgba_C; |
245 | 0 | WebPYUV444Converters[MODE_bgrA] = WebPYuv444ToBgra_C; |
246 | 0 | WebPYUV444Converters[MODE_Argb] = WebPYuv444ToArgb_C; |
247 | 0 | WebPYUV444Converters[MODE_rgbA_4444] = WebPYuv444ToRgba4444_C; |
248 | |
|
249 | 0 | if (VP8GetCPUInfo != NULL) { |
250 | 0 | #if defined(WEBP_HAVE_SSE2) |
251 | 0 | if (VP8GetCPUInfo(kSSE2)) { |
252 | 0 | WebPInitYUV444ConvertersSSE2(); |
253 | 0 | } |
254 | 0 | #endif |
255 | 0 | #if defined(WEBP_HAVE_SSE41) |
256 | 0 | if (VP8GetCPUInfo(kSSE4_1)) { |
257 | 0 | WebPInitYUV444ConvertersSSE41(); |
258 | 0 | } |
259 | 0 | #endif |
260 | | #if defined(WEBP_USE_MIPS_DSP_R2) |
261 | | if (VP8GetCPUInfo(kMIPSdspR2)) { |
262 | | WebPInitYUV444ConvertersMIPSdspR2(); |
263 | | } |
264 | | #endif |
265 | 0 | } |
266 | 0 | } |
267 | | |
268 | | //------------------------------------------------------------------------------ |
269 | | // Main calls |
270 | | |
271 | | extern void WebPInitUpsamplersSSE2(void); |
272 | | extern void WebPInitUpsamplersSSE41(void); |
273 | | extern void WebPInitUpsamplersNEON(void); |
274 | | extern void WebPInitUpsamplersMIPSdspR2(void); |
275 | | extern void WebPInitUpsamplersMSA(void); |
276 | | |
277 | 1 | WEBP_DSP_INIT_FUNC(WebPInitUpsamplers) { |
278 | 1 | #ifdef FANCY_UPSAMPLING |
279 | 1 | #if !WEBP_NEON_OMIT_C_CODE |
280 | 1 | WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePair_C; |
281 | 1 | WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePair_C; |
282 | 1 | WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePair_C; |
283 | 1 | WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePair_C; |
284 | 1 | WebPUpsamplers[MODE_RGB] = UpsampleRgbLinePair_C; |
285 | 1 | WebPUpsamplers[MODE_BGR] = UpsampleBgrLinePair_C; |
286 | 1 | WebPUpsamplers[MODE_ARGB] = UpsampleArgbLinePair_C; |
287 | 1 | WebPUpsamplers[MODE_RGBA_4444] = UpsampleRgba4444LinePair_C; |
288 | 1 | WebPUpsamplers[MODE_RGB_565] = UpsampleRgb565LinePair_C; |
289 | 1 | WebPUpsamplers[MODE_Argb] = UpsampleArgbLinePair_C; |
290 | 1 | WebPUpsamplers[MODE_rgbA_4444] = UpsampleRgba4444LinePair_C; |
291 | 1 | #endif |
292 | | |
293 | | // If defined, use CPUInfo() to overwrite some pointers with faster versions. |
294 | 1 | if (VP8GetCPUInfo != NULL) { |
295 | 1 | #if defined(WEBP_HAVE_SSE2) |
296 | 1 | if (VP8GetCPUInfo(kSSE2)) { |
297 | 1 | WebPInitUpsamplersSSE2(); |
298 | 1 | } |
299 | 1 | #endif |
300 | 1 | #if defined(WEBP_HAVE_SSE41) |
301 | 1 | if (VP8GetCPUInfo(kSSE4_1)) { |
302 | 1 | WebPInitUpsamplersSSE41(); |
303 | 1 | } |
304 | 1 | #endif |
305 | | #if defined(WEBP_USE_MIPS_DSP_R2) |
306 | | if (VP8GetCPUInfo(kMIPSdspR2)) { |
307 | | WebPInitUpsamplersMIPSdspR2(); |
308 | | } |
309 | | #endif |
310 | | #if defined(WEBP_USE_MSA) |
311 | | if (VP8GetCPUInfo(kMSA)) { |
312 | | WebPInitUpsamplersMSA(); |
313 | | } |
314 | | #endif |
315 | 1 | } |
316 | | |
317 | | #if defined(WEBP_HAVE_NEON) |
318 | | if (WEBP_NEON_OMIT_C_CODE || |
319 | | (VP8GetCPUInfo != NULL && VP8GetCPUInfo(kNEON))) { |
320 | | WebPInitUpsamplersNEON(); |
321 | | } |
322 | | #endif |
323 | | |
324 | 1 | assert(WebPUpsamplers[MODE_RGBA] != NULL); |
325 | 1 | assert(WebPUpsamplers[MODE_BGRA] != NULL); |
326 | 1 | assert(WebPUpsamplers[MODE_rgbA] != NULL); |
327 | 1 | assert(WebPUpsamplers[MODE_bgrA] != NULL); |
328 | 1 | #if !defined(WEBP_REDUCE_CSP) || !WEBP_NEON_OMIT_C_CODE |
329 | 1 | assert(WebPUpsamplers[MODE_RGB] != NULL); |
330 | 1 | assert(WebPUpsamplers[MODE_BGR] != NULL); |
331 | 1 | assert(WebPUpsamplers[MODE_ARGB] != NULL); |
332 | 1 | assert(WebPUpsamplers[MODE_RGBA_4444] != NULL); |
333 | 1 | assert(WebPUpsamplers[MODE_RGB_565] != NULL); |
334 | 1 | assert(WebPUpsamplers[MODE_Argb] != NULL); |
335 | 1 | assert(WebPUpsamplers[MODE_rgbA_4444] != NULL); |
336 | 1 | #endif |
337 | | |
338 | 1 | #endif // FANCY_UPSAMPLING |
339 | 1 | } |
340 | | |
341 | | //------------------------------------------------------------------------------ |