/src/libavif/ext/libyuv/source/row_common.cc
Line | Count | Source |
1 | | /* |
2 | | * Copyright 2011 The LibYuv Project Authors. All rights reserved. |
3 | | * |
4 | | * Use of this source code is governed by a BSD-style license |
5 | | * that can be found in the LICENSE file in the root of the source |
6 | | * tree. An additional intellectual property rights grant can be found |
7 | | * in the file PATENTS. All contributing project authors may |
8 | | * be found in the AUTHORS file in the root of the source tree. |
9 | | */ |
10 | | |
11 | | #include "libyuv/row.h" |
12 | | |
13 | | #include <assert.h> |
14 | | #include <string.h> // For memcpy and memset. |
15 | | |
16 | | #include "libyuv/basic_types.h" |
17 | | #include "libyuv/convert_argb.h" // For kYuvI601Constants |
18 | | |
19 | | #ifdef __cplusplus |
20 | | namespace libyuv { |
21 | | extern "C" { |
22 | | #endif |
23 | | |
24 | | #ifdef __cplusplus |
25 | 6.81G | #define STATIC_CAST(type, expr) static_cast<type>(expr) |
26 | | #else |
27 | | #define STATIC_CAST(type, expr) (type)(expr) |
28 | | #endif |
29 | | |
30 | | // This macro controls YUV to RGB using unsigned math to extend range of |
31 | | // YUV to RGB coefficients to 0 to 4 instead of 0 to 2 for more accuracy on B: |
32 | | // LIBYUV_UNLIMITED_DATA |
33 | | |
34 | | // Macros to enable unlimited data for each colorspace |
35 | | // LIBYUV_UNLIMITED_BT601 |
36 | | // LIBYUV_UNLIMITED_BT709 |
37 | | // LIBYUV_UNLIMITED_BT2020 |
38 | | |
39 | | #if defined(LIBYUV_BIT_EXACT) |
40 | | #define LIBYUV_UNATTENUATE_DUP 1 |
41 | | #endif |
42 | | |
43 | | // llvm x86 is poor at ternary operator, so use branchless min/max. |
44 | | |
45 | | #define USE_BRANCHLESS 1 |
46 | | #if defined(USE_BRANCHLESS) |
47 | 0 | static __inline int32_t clamp0(int32_t v) { |
48 | 0 | return -(v >= 0) & v; |
49 | 0 | } |
50 | | // TODO(fbarchard): make clamp255 preserve negative values. |
51 | 0 | static __inline int32_t clamp255(int32_t v) { |
52 | 0 | return (-(v >= 255) | v) & 255; |
53 | 0 | } |
54 | | |
55 | 0 | static __inline int32_t clamp1023(int32_t v) { |
56 | 0 | return (-(v >= 1023) | v) & 1023; |
57 | 0 | } |
58 | | |
59 | | // clamp to max |
60 | 0 | static __inline int32_t ClampMax(int32_t v, int32_t max) { |
61 | 0 | return (-(v >= max) | v) & max; |
62 | 0 | } |
63 | | |
64 | 0 | static __inline uint32_t Abs(int32_t v) { |
65 | 0 | int m = -(v < 0); |
66 | 0 | return (v + m) ^ m; |
67 | 0 | } |
68 | | #else // USE_BRANCHLESS |
69 | | static __inline int32_t clamp0(int32_t v) { |
70 | | return (v < 0) ? 0 : v; |
71 | | } |
72 | | |
73 | | static __inline int32_t clamp255(int32_t v) { |
74 | | return (v > 255) ? 255 : v; |
75 | | } |
76 | | |
77 | | static __inline int32_t clamp1023(int32_t v) { |
78 | | return (v > 1023) ? 1023 : v; |
79 | | } |
80 | | |
81 | | static __inline int32_t ClampMax(int32_t v, int32_t max) { |
82 | | return (v > max) ? max : v; |
83 | | } |
84 | | |
85 | | static __inline uint32_t Abs(int32_t v) { |
86 | | return (v < 0) ? -v : v; |
87 | | } |
88 | | #endif // USE_BRANCHLESS |
89 | 0 | static __inline uint32_t Clamp(int32_t val) { |
90 | 0 | int v = clamp0(val); |
91 | 0 | return (uint32_t)(clamp255(v)); |
92 | 0 | } |
93 | | |
94 | 0 | static __inline uint32_t Clamp10(int32_t val) { |
95 | 0 | int v = clamp0(val); |
96 | 0 | return (uint32_t)(clamp1023(v)); |
97 | 0 | } |
98 | | |
99 | | // Little Endian |
100 | | #if defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \ |
101 | | defined(_M_IX86) || defined(__arm__) || defined(_M_ARM) || \ |
102 | | (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) |
103 | 0 | #define WRITEWORD(p, v) *(uint32_t*)(p) = v |
104 | | #else |
105 | | static inline void WRITEWORD(uint8_t* p, uint32_t v) { |
106 | | p[0] = (uint8_t)(v & 255); |
107 | | p[1] = (uint8_t)((v >> 8) & 255); |
108 | | p[2] = (uint8_t)((v >> 16) & 255); |
109 | | p[3] = (uint8_t)((v >> 24) & 255); |
110 | | } |
111 | | #endif |
112 | | |
113 | 0 | void RGB24ToARGBRow_C(const uint8_t* src_rgb24, uint8_t* dst_argb, int width) { |
114 | 0 | int x; |
115 | 0 | for (x = 0; x < width; ++x) { |
116 | 0 | uint8_t b = src_rgb24[0]; |
117 | 0 | uint8_t g = src_rgb24[1]; |
118 | 0 | uint8_t r = src_rgb24[2]; |
119 | 0 | dst_argb[0] = b; |
120 | 0 | dst_argb[1] = g; |
121 | 0 | dst_argb[2] = r; |
122 | 0 | dst_argb[3] = 255u; |
123 | 0 | dst_argb += 4; |
124 | 0 | src_rgb24 += 3; |
125 | 0 | } |
126 | 0 | } |
127 | | |
128 | 0 | void RAWToARGBRow_C(const uint8_t* src_raw, uint8_t* dst_argb, int width) { |
129 | 0 | int x; |
130 | 0 | for (x = 0; x < width; ++x) { |
131 | 0 | uint8_t r = src_raw[0]; |
132 | 0 | uint8_t g = src_raw[1]; |
133 | 0 | uint8_t b = src_raw[2]; |
134 | 0 | dst_argb[0] = b; |
135 | 0 | dst_argb[1] = g; |
136 | 0 | dst_argb[2] = r; |
137 | 0 | dst_argb[3] = 255u; |
138 | 0 | dst_argb += 4; |
139 | 0 | src_raw += 3; |
140 | 0 | } |
141 | 0 | } |
142 | | |
143 | 0 | void RAWToRGBARow_C(const uint8_t* src_raw, uint8_t* dst_rgba, int width) { |
144 | 0 | int x; |
145 | 0 | for (x = 0; x < width; ++x) { |
146 | 0 | uint8_t r = src_raw[0]; |
147 | 0 | uint8_t g = src_raw[1]; |
148 | 0 | uint8_t b = src_raw[2]; |
149 | 0 | dst_rgba[0] = 255u; |
150 | 0 | dst_rgba[1] = b; |
151 | 0 | dst_rgba[2] = g; |
152 | 0 | dst_rgba[3] = r; |
153 | 0 | dst_rgba += 4; |
154 | 0 | src_raw += 3; |
155 | 0 | } |
156 | 0 | } |
157 | | |
158 | 0 | void RAWToRGB24Row_C(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) { |
159 | 0 | int x; |
160 | 0 | for (x = 0; x < width; ++x) { |
161 | 0 | uint8_t r = src_raw[0]; |
162 | 0 | uint8_t g = src_raw[1]; |
163 | 0 | uint8_t b = src_raw[2]; |
164 | 0 | dst_rgb24[0] = b; |
165 | 0 | dst_rgb24[1] = g; |
166 | 0 | dst_rgb24[2] = r; |
167 | 0 | dst_rgb24 += 3; |
168 | 0 | src_raw += 3; |
169 | 0 | } |
170 | 0 | } |
171 | | |
172 | | void RGB565ToARGBRow_C(const uint8_t* src_rgb565, |
173 | | uint8_t* dst_argb, |
174 | 0 | int width) { |
175 | 0 | int x; |
176 | 0 | for (x = 0; x < width; ++x) { |
177 | 0 | uint8_t b = STATIC_CAST(uint8_t, src_rgb565[0] & 0x1f); |
178 | 0 | uint8_t g = STATIC_CAST( |
179 | 0 | uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3)); |
180 | 0 | uint8_t r = STATIC_CAST(uint8_t, src_rgb565[1] >> 3); |
181 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, (b << 3) | (b >> 2)); |
182 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, (g << 2) | (g >> 4)); |
183 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, (r << 3) | (r >> 2)); |
184 | 0 | dst_argb[3] = 255u; |
185 | 0 | dst_argb += 4; |
186 | 0 | src_rgb565 += 2; |
187 | 0 | } |
188 | 0 | } |
189 | | |
190 | | void ARGB1555ToARGBRow_C(const uint8_t* src_argb1555, |
191 | | uint8_t* dst_argb, |
192 | 0 | int width) { |
193 | 0 | int x; |
194 | 0 | for (x = 0; x < width; ++x) { |
195 | 0 | uint8_t b = STATIC_CAST(uint8_t, src_argb1555[0] & 0x1f); |
196 | 0 | uint8_t g = STATIC_CAST( |
197 | 0 | uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3)); |
198 | 0 | uint8_t r = STATIC_CAST(uint8_t, (src_argb1555[1] & 0x7c) >> 2); |
199 | 0 | uint8_t a = STATIC_CAST(uint8_t, src_argb1555[1] >> 7); |
200 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, (b << 3) | (b >> 2)); |
201 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, (g << 3) | (g >> 2)); |
202 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, (r << 3) | (r >> 2)); |
203 | 0 | dst_argb[3] = -a; |
204 | 0 | dst_argb += 4; |
205 | 0 | src_argb1555 += 2; |
206 | 0 | } |
207 | 0 | } |
208 | | |
209 | | void ARGB4444ToARGBRow_C(const uint8_t* src_argb4444, |
210 | | uint8_t* dst_argb, |
211 | 0 | int width) { |
212 | 0 | int x; |
213 | 0 | for (x = 0; x < width; ++x) { |
214 | 0 | uint8_t b = STATIC_CAST(uint8_t, src_argb4444[0] & 0x0f); |
215 | 0 | uint8_t g = STATIC_CAST(uint8_t, src_argb4444[0] >> 4); |
216 | 0 | uint8_t r = STATIC_CAST(uint8_t, src_argb4444[1] & 0x0f); |
217 | 0 | uint8_t a = STATIC_CAST(uint8_t, src_argb4444[1] >> 4); |
218 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, (b << 4) | b); |
219 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, (g << 4) | g); |
220 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, (r << 4) | r); |
221 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, (a << 4) | a); |
222 | 0 | dst_argb += 4; |
223 | 0 | src_argb4444 += 2; |
224 | 0 | } |
225 | 0 | } |
226 | | |
227 | 0 | void AR30ToARGBRow_C(const uint8_t* src_ar30, uint8_t* dst_argb, int width) { |
228 | 0 | int x; |
229 | 0 | for (x = 0; x < width; ++x) { |
230 | 0 | uint32_t ar30; |
231 | 0 | memcpy(&ar30, src_ar30, sizeof ar30); |
232 | 0 | uint32_t b = (ar30 >> 2) & 0xff; |
233 | 0 | uint32_t g = (ar30 >> 12) & 0xff; |
234 | 0 | uint32_t r = (ar30 >> 22) & 0xff; |
235 | 0 | uint32_t a = (ar30 >> 30) * 0x55; // Replicate 2 bits to 8 bits. |
236 | 0 | *(uint32_t*)(dst_argb) = b | (g << 8) | (r << 16) | (a << 24); |
237 | 0 | dst_argb += 4; |
238 | 0 | src_ar30 += 4; |
239 | 0 | } |
240 | 0 | } |
241 | | |
242 | 0 | void AR30ToABGRRow_C(const uint8_t* src_ar30, uint8_t* dst_abgr, int width) { |
243 | 0 | int x; |
244 | 0 | for (x = 0; x < width; ++x) { |
245 | 0 | uint32_t ar30; |
246 | 0 | memcpy(&ar30, src_ar30, sizeof ar30); |
247 | 0 | uint32_t b = (ar30 >> 2) & 0xff; |
248 | 0 | uint32_t g = (ar30 >> 12) & 0xff; |
249 | 0 | uint32_t r = (ar30 >> 22) & 0xff; |
250 | 0 | uint32_t a = (ar30 >> 30) * 0x55; // Replicate 2 bits to 8 bits. |
251 | 0 | *(uint32_t*)(dst_abgr) = r | (g << 8) | (b << 16) | (a << 24); |
252 | 0 | dst_abgr += 4; |
253 | 0 | src_ar30 += 4; |
254 | 0 | } |
255 | 0 | } |
256 | | |
257 | 0 | void AR30ToAB30Row_C(const uint8_t* src_ar30, uint8_t* dst_ab30, int width) { |
258 | 0 | int x; |
259 | 0 | for (x = 0; x < width; ++x) { |
260 | 0 | uint32_t ar30; |
261 | 0 | memcpy(&ar30, src_ar30, sizeof ar30); |
262 | 0 | uint32_t b = ar30 & 0x3ff; |
263 | 0 | uint32_t ga = ar30 & 0xc00ffc00; |
264 | 0 | uint32_t r = (ar30 >> 20) & 0x3ff; |
265 | 0 | *(uint32_t*)(dst_ab30) = r | ga | (b << 20); |
266 | 0 | dst_ab30 += 4; |
267 | 0 | src_ar30 += 4; |
268 | 0 | } |
269 | 0 | } |
270 | | |
271 | 0 | void ARGBToABGRRow_C(const uint8_t* src_argb, uint8_t* dst_abgr, int width) { |
272 | 0 | int x; |
273 | 0 | for (x = 0; x < width; ++x) { |
274 | 0 | uint8_t b = src_argb[0]; |
275 | 0 | uint8_t g = src_argb[1]; |
276 | 0 | uint8_t r = src_argb[2]; |
277 | 0 | uint8_t a = src_argb[3]; |
278 | 0 | dst_abgr[0] = r; |
279 | 0 | dst_abgr[1] = g; |
280 | 0 | dst_abgr[2] = b; |
281 | 0 | dst_abgr[3] = a; |
282 | 0 | dst_abgr += 4; |
283 | 0 | src_argb += 4; |
284 | 0 | } |
285 | 0 | } |
286 | | |
287 | 0 | void ARGBToBGRARow_C(const uint8_t* src_argb, uint8_t* dst_bgra, int width) { |
288 | 0 | int x; |
289 | 0 | for (x = 0; x < width; ++x) { |
290 | 0 | uint8_t b = src_argb[0]; |
291 | 0 | uint8_t g = src_argb[1]; |
292 | 0 | uint8_t r = src_argb[2]; |
293 | 0 | uint8_t a = src_argb[3]; |
294 | 0 | dst_bgra[0] = a; |
295 | 0 | dst_bgra[1] = r; |
296 | 0 | dst_bgra[2] = g; |
297 | 0 | dst_bgra[3] = b; |
298 | 0 | dst_bgra += 4; |
299 | 0 | src_argb += 4; |
300 | 0 | } |
301 | 0 | } |
302 | | |
303 | 0 | void ARGBToRGBARow_C(const uint8_t* src_argb, uint8_t* dst_rgba, int width) { |
304 | 0 | int x; |
305 | 0 | for (x = 0; x < width; ++x) { |
306 | 0 | uint8_t b = src_argb[0]; |
307 | 0 | uint8_t g = src_argb[1]; |
308 | 0 | uint8_t r = src_argb[2]; |
309 | 0 | uint8_t a = src_argb[3]; |
310 | 0 | dst_rgba[0] = a; |
311 | 0 | dst_rgba[1] = b; |
312 | 0 | dst_rgba[2] = g; |
313 | 0 | dst_rgba[3] = r; |
314 | 0 | dst_rgba += 4; |
315 | 0 | src_argb += 4; |
316 | 0 | } |
317 | 0 | } |
318 | | |
319 | 0 | void ARGBToRGB24Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) { |
320 | 0 | int x; |
321 | 0 | for (x = 0; x < width; ++x) { |
322 | 0 | uint8_t b = src_argb[0]; |
323 | 0 | uint8_t g = src_argb[1]; |
324 | 0 | uint8_t r = src_argb[2]; |
325 | 0 | dst_rgb[0] = b; |
326 | 0 | dst_rgb[1] = g; |
327 | 0 | dst_rgb[2] = r; |
328 | 0 | dst_rgb += 3; |
329 | 0 | src_argb += 4; |
330 | 0 | } |
331 | 0 | } |
332 | | |
333 | 0 | void ARGBToRAWRow_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) { |
334 | 0 | int x; |
335 | 0 | for (x = 0; x < width; ++x) { |
336 | 0 | uint8_t b = src_argb[0]; |
337 | 0 | uint8_t g = src_argb[1]; |
338 | 0 | uint8_t r = src_argb[2]; |
339 | 0 | dst_rgb[0] = r; |
340 | 0 | dst_rgb[1] = g; |
341 | 0 | dst_rgb[2] = b; |
342 | 0 | dst_rgb += 3; |
343 | 0 | src_argb += 4; |
344 | 0 | } |
345 | 0 | } |
346 | | |
347 | 0 | void RGBAToARGBRow_C(const uint8_t* src_rgba, uint8_t* dst_argb, int width) { |
348 | 0 | int x; |
349 | 0 | for (x = 0; x < width; ++x) { |
350 | 0 | uint8_t a = src_rgba[0]; |
351 | 0 | uint8_t b = src_rgba[1]; |
352 | 0 | uint8_t g = src_rgba[2]; |
353 | 0 | uint8_t r = src_rgba[3]; |
354 | 0 | dst_argb[0] = b; |
355 | 0 | dst_argb[1] = g; |
356 | 0 | dst_argb[2] = r; |
357 | 0 | dst_argb[3] = a; |
358 | 0 | dst_argb += 4; |
359 | 0 | src_rgba += 4; |
360 | 0 | } |
361 | 0 | } |
362 | | |
363 | 0 | void ARGBToRGB565Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) { |
364 | 0 | int x; |
365 | 0 | for (x = 0; x < width - 1; x += 2) { |
366 | 0 | uint8_t b0 = src_argb[0] >> 3; |
367 | 0 | uint8_t g0 = src_argb[1] >> 2; |
368 | 0 | uint8_t r0 = src_argb[2] >> 3; |
369 | 0 | uint8_t b1 = src_argb[4] >> 3; |
370 | 0 | uint8_t g1 = src_argb[5] >> 2; |
371 | 0 | uint8_t r1 = src_argb[6] >> 3; |
372 | 0 | WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) | (b1 << 16) | (g1 << 21) | |
373 | 0 | (r1 << 27)); |
374 | 0 | dst_rgb += 4; |
375 | 0 | src_argb += 8; |
376 | 0 | } |
377 | 0 | if (width & 1) { |
378 | 0 | uint8_t b0 = src_argb[0] >> 3; |
379 | 0 | uint8_t g0 = src_argb[1] >> 2; |
380 | 0 | uint8_t r0 = src_argb[2] >> 3; |
381 | 0 | *(uint16_t*)(dst_rgb) = STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11)); |
382 | 0 | } |
383 | 0 | } |
384 | | |
385 | | // dither4 is a row of 4 values from 4x4 dither matrix. |
386 | | // The 4x4 matrix contains values to increase RGB. When converting to |
387 | | // fewer bits (565) this provides an ordered dither. |
388 | | // The order in the 4x4 matrix in first byte is upper left. |
389 | | // The 4 values are passed as an int, then referenced as an array, so |
390 | | // endian will not affect order of the original matrix. But the dither4 |
391 | | // will containing the first pixel in the lower byte for little endian |
392 | | // or the upper byte for big endian. |
393 | | void ARGBToRGB565DitherRow_C(const uint8_t* src_argb, |
394 | | uint8_t* dst_rgb, |
395 | | uint32_t dither4, |
396 | 0 | int width) { |
397 | 0 | int x; |
398 | 0 | for (x = 0; x < width - 1; x += 2) { |
399 | 0 | int dither0 = ((const unsigned char*)(&dither4))[x & 3]; |
400 | 0 | int dither1 = ((const unsigned char*)(&dither4))[(x + 1) & 3]; |
401 | 0 | uint8_t b0 = STATIC_CAST(uint8_t, clamp255(src_argb[0] + dither0) >> 3); |
402 | 0 | uint8_t g0 = STATIC_CAST(uint8_t, clamp255(src_argb[1] + dither0) >> 2); |
403 | 0 | uint8_t r0 = STATIC_CAST(uint8_t, clamp255(src_argb[2] + dither0) >> 3); |
404 | 0 | uint8_t b1 = STATIC_CAST(uint8_t, clamp255(src_argb[4] + dither1) >> 3); |
405 | 0 | uint8_t g1 = STATIC_CAST(uint8_t, clamp255(src_argb[5] + dither1) >> 2); |
406 | 0 | uint8_t r1 = STATIC_CAST(uint8_t, clamp255(src_argb[6] + dither1) >> 3); |
407 | 0 | *(uint16_t*)(dst_rgb + 0) = |
408 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11)); |
409 | 0 | *(uint16_t*)(dst_rgb + 2) = |
410 | 0 | STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 11)); |
411 | 0 | dst_rgb += 4; |
412 | 0 | src_argb += 8; |
413 | 0 | } |
414 | 0 | if (width & 1) { |
415 | 0 | int dither0 = ((const unsigned char*)(&dither4))[(width - 1) & 3]; |
416 | 0 | uint8_t b0 = STATIC_CAST(uint8_t, clamp255(src_argb[0] + dither0) >> 3); |
417 | 0 | uint8_t g0 = STATIC_CAST(uint8_t, clamp255(src_argb[1] + dither0) >> 2); |
418 | 0 | uint8_t r0 = STATIC_CAST(uint8_t, clamp255(src_argb[2] + dither0) >> 3); |
419 | 0 | *(uint16_t*)(dst_rgb) = STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11)); |
420 | 0 | } |
421 | 0 | } |
422 | | |
423 | 0 | void ARGBToARGB1555Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) { |
424 | 0 | int x; |
425 | 0 | for (x = 0; x < width - 1; x += 2) { |
426 | 0 | uint8_t b0 = src_argb[0] >> 3; |
427 | 0 | uint8_t g0 = src_argb[1] >> 3; |
428 | 0 | uint8_t r0 = src_argb[2] >> 3; |
429 | 0 | uint8_t a0 = src_argb[3] >> 7; |
430 | 0 | uint8_t b1 = src_argb[4] >> 3; |
431 | 0 | uint8_t g1 = src_argb[5] >> 3; |
432 | 0 | uint8_t r1 = src_argb[6] >> 3; |
433 | 0 | uint8_t a1 = src_argb[7] >> 7; |
434 | 0 | *(uint16_t*)(dst_rgb + 0) = |
435 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | (a0 << 15)); |
436 | 0 | *(uint16_t*)(dst_rgb + 2) = |
437 | 0 | STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 10) | (a1 << 15)); |
438 | 0 | dst_rgb += 4; |
439 | 0 | src_argb += 8; |
440 | 0 | } |
441 | 0 | if (width & 1) { |
442 | 0 | uint8_t b0 = src_argb[0] >> 3; |
443 | 0 | uint8_t g0 = src_argb[1] >> 3; |
444 | 0 | uint8_t r0 = src_argb[2] >> 3; |
445 | 0 | uint8_t a0 = src_argb[3] >> 7; |
446 | 0 | *(uint16_t*)(dst_rgb) = |
447 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | (a0 << 15)); |
448 | 0 | } |
449 | 0 | } |
450 | | |
451 | 0 | void ARGBToARGB4444Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) { |
452 | 0 | int x; |
453 | 0 | for (x = 0; x < width - 1; x += 2) { |
454 | 0 | uint8_t b0 = src_argb[0] >> 4; |
455 | 0 | uint8_t g0 = src_argb[1] >> 4; |
456 | 0 | uint8_t r0 = src_argb[2] >> 4; |
457 | 0 | uint8_t a0 = src_argb[3] >> 4; |
458 | 0 | uint8_t b1 = src_argb[4] >> 4; |
459 | 0 | uint8_t g1 = src_argb[5] >> 4; |
460 | 0 | uint8_t r1 = src_argb[6] >> 4; |
461 | 0 | uint8_t a1 = src_argb[7] >> 4; |
462 | 0 | *(uint16_t*)(dst_rgb + 0) = |
463 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | (a0 << 12)); |
464 | 0 | *(uint16_t*)(dst_rgb + 2) = |
465 | 0 | STATIC_CAST(uint16_t, b1 | (g1 << 4) | (r1 << 8) | (a1 << 12)); |
466 | 0 | dst_rgb += 4; |
467 | 0 | src_argb += 8; |
468 | 0 | } |
469 | 0 | if (width & 1) { |
470 | 0 | uint8_t b0 = src_argb[0] >> 4; |
471 | 0 | uint8_t g0 = src_argb[1] >> 4; |
472 | 0 | uint8_t r0 = src_argb[2] >> 4; |
473 | 0 | uint8_t a0 = src_argb[3] >> 4; |
474 | 0 | *(uint16_t*)(dst_rgb) = |
475 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | (a0 << 12)); |
476 | 0 | } |
477 | 0 | } |
478 | | |
479 | 0 | void ABGRToAR30Row_C(const uint8_t* src_abgr, uint8_t* dst_ar30, int width) { |
480 | 0 | int x; |
481 | 0 | for (x = 0; x < width; ++x) { |
482 | 0 | uint32_t r0 = (src_abgr[0] >> 6) | ((uint32_t)(src_abgr[0]) << 2); |
483 | 0 | uint32_t g0 = (src_abgr[1] >> 6) | ((uint32_t)(src_abgr[1]) << 2); |
484 | 0 | uint32_t b0 = (src_abgr[2] >> 6) | ((uint32_t)(src_abgr[2]) << 2); |
485 | 0 | uint32_t a0 = (src_abgr[3] >> 6); |
486 | 0 | *(uint32_t*)(dst_ar30) = |
487 | 0 | STATIC_CAST(uint32_t, b0 | (g0 << 10) | (r0 << 20) | (a0 << 30)); |
488 | 0 | dst_ar30 += 4; |
489 | 0 | src_abgr += 4; |
490 | 0 | } |
491 | 0 | } |
492 | | |
493 | 0 | void ARGBToAR30Row_C(const uint8_t* src_argb, uint8_t* dst_ar30, int width) { |
494 | 0 | int x; |
495 | 0 | for (x = 0; x < width; ++x) { |
496 | 0 | uint32_t b0 = (src_argb[0] >> 6) | ((uint32_t)(src_argb[0]) << 2); |
497 | 0 | uint32_t g0 = (src_argb[1] >> 6) | ((uint32_t)(src_argb[1]) << 2); |
498 | 0 | uint32_t r0 = (src_argb[2] >> 6) | ((uint32_t)(src_argb[2]) << 2); |
499 | 0 | uint32_t a0 = (src_argb[3] >> 6); |
500 | 0 | *(uint32_t*)(dst_ar30) = |
501 | 0 | STATIC_CAST(uint32_t, b0 | (g0 << 10) | (r0 << 20) | (a0 << 30)); |
502 | 0 | dst_ar30 += 4; |
503 | 0 | src_argb += 4; |
504 | 0 | } |
505 | 0 | } |
506 | | |
507 | 0 | void ARGBToAR64Row_C(const uint8_t* src_argb, uint16_t* dst_ar64, int width) { |
508 | 0 | int x; |
509 | 0 | for (x = 0; x < width; ++x) { |
510 | 0 | uint16_t b = src_argb[0] * 0x0101; |
511 | 0 | uint16_t g = src_argb[1] * 0x0101; |
512 | 0 | uint16_t r = src_argb[2] * 0x0101; |
513 | 0 | uint16_t a = src_argb[3] * 0x0101; |
514 | 0 | dst_ar64[0] = b; |
515 | 0 | dst_ar64[1] = g; |
516 | 0 | dst_ar64[2] = r; |
517 | 0 | dst_ar64[3] = a; |
518 | 0 | dst_ar64 += 4; |
519 | 0 | src_argb += 4; |
520 | 0 | } |
521 | 0 | } |
522 | | |
523 | 0 | void ARGBToAB64Row_C(const uint8_t* src_argb, uint16_t* dst_ab64, int width) { |
524 | 0 | int x; |
525 | 0 | for (x = 0; x < width; ++x) { |
526 | 0 | uint16_t b = src_argb[0] * 0x0101; |
527 | 0 | uint16_t g = src_argb[1] * 0x0101; |
528 | 0 | uint16_t r = src_argb[2] * 0x0101; |
529 | 0 | uint16_t a = src_argb[3] * 0x0101; |
530 | 0 | dst_ab64[0] = r; |
531 | 0 | dst_ab64[1] = g; |
532 | 0 | dst_ab64[2] = b; |
533 | 0 | dst_ab64[3] = a; |
534 | 0 | dst_ab64 += 4; |
535 | 0 | src_argb += 4; |
536 | 0 | } |
537 | 0 | } |
538 | | |
539 | 0 | void AR64ToARGBRow_C(const uint16_t* src_ar64, uint8_t* dst_argb, int width) { |
540 | 0 | int x; |
541 | 0 | for (x = 0; x < width; ++x) { |
542 | 0 | uint8_t b = src_ar64[0] >> 8; |
543 | 0 | uint8_t g = src_ar64[1] >> 8; |
544 | 0 | uint8_t r = src_ar64[2] >> 8; |
545 | 0 | uint8_t a = src_ar64[3] >> 8; |
546 | 0 | dst_argb[0] = b; |
547 | 0 | dst_argb[1] = g; |
548 | 0 | dst_argb[2] = r; |
549 | 0 | dst_argb[3] = a; |
550 | 0 | dst_argb += 4; |
551 | 0 | src_ar64 += 4; |
552 | 0 | } |
553 | 0 | } |
554 | | |
555 | 0 | void AB64ToARGBRow_C(const uint16_t* src_ab64, uint8_t* dst_argb, int width) { |
556 | 0 | int x; |
557 | 0 | for (x = 0; x < width; ++x) { |
558 | 0 | uint8_t r = src_ab64[0] >> 8; |
559 | 0 | uint8_t g = src_ab64[1] >> 8; |
560 | 0 | uint8_t b = src_ab64[2] >> 8; |
561 | 0 | uint8_t a = src_ab64[3] >> 8; |
562 | 0 | dst_argb[0] = b; |
563 | 0 | dst_argb[1] = g; |
564 | 0 | dst_argb[2] = r; |
565 | 0 | dst_argb[3] = a; |
566 | 0 | dst_argb += 4; |
567 | 0 | src_ab64 += 4; |
568 | 0 | } |
569 | 0 | } |
570 | | |
571 | 0 | void AR64ToAB64Row_C(const uint16_t* src_ar64, uint16_t* dst_ab64, int width) { |
572 | 0 | int x; |
573 | 0 | for (x = 0; x < width; ++x) { |
574 | 0 | uint16_t b = src_ar64[0]; |
575 | 0 | uint16_t g = src_ar64[1]; |
576 | 0 | uint16_t r = src_ar64[2]; |
577 | 0 | uint16_t a = src_ar64[3]; |
578 | 0 | dst_ab64[0] = r; |
579 | 0 | dst_ab64[1] = g; |
580 | 0 | dst_ab64[2] = b; |
581 | 0 | dst_ab64[3] = a; |
582 | 0 | dst_ab64 += 4; |
583 | 0 | src_ar64 += 4; |
584 | 0 | } |
585 | 0 | } |
586 | | |
587 | | // TODO(fbarchard): Make shuffle compatible with SIMD versions |
588 | | void AR64ShuffleRow_C(const uint8_t* src_ar64, |
589 | | uint8_t* dst_ar64, |
590 | | const uint8_t* shuffler, |
591 | 0 | int width) { |
592 | 0 | const uint16_t* src_ar64_16 = (const uint16_t*)src_ar64; |
593 | 0 | uint16_t* dst_ar64_16 = (uint16_t*)dst_ar64; |
594 | 0 | int index0 = shuffler[0] / 2; |
595 | 0 | int index1 = shuffler[2] / 2; |
596 | 0 | int index2 = shuffler[4] / 2; |
597 | 0 | int index3 = shuffler[6] / 2; |
598 | | // Shuffle a row of AR64. |
599 | 0 | int x; |
600 | 0 | for (x = 0; x < width / 2; ++x) { |
601 | | // To support in-place conversion. |
602 | 0 | uint16_t b = src_ar64_16[index0]; |
603 | 0 | uint16_t g = src_ar64_16[index1]; |
604 | 0 | uint16_t r = src_ar64_16[index2]; |
605 | 0 | uint16_t a = src_ar64_16[index3]; |
606 | 0 | dst_ar64_16[0] = b; |
607 | 0 | dst_ar64_16[1] = g; |
608 | 0 | dst_ar64_16[2] = r; |
609 | 0 | dst_ar64_16[3] = a; |
610 | 0 | src_ar64_16 += 4; |
611 | 0 | dst_ar64_16 += 4; |
612 | 0 | } |
613 | 0 | } |
614 | | // BT601 8 bit Y: |
615 | | // b 0.114 * 219 = 24.966 = 25 |
616 | | // g 0.587 * 219 = 128.553 = 129 |
617 | | // r 0.299 * 219 = 65.481 = 66 |
618 | | // BT601 8 bit U: |
619 | | // b 0.875 * 128 = 112.0 = 112 |
620 | | // g -0.5781 * 128 = −73.9968 = -74 |
621 | | // r -0.2969 * 128 = −38.0032 = -38 |
622 | | // BT601 8 bit V: |
623 | | // b -0.1406 * 128 = −17.9968 = -18 |
624 | | // g -0.7344 * 128 = −94.0032 = -94 |
625 | | // r 0.875 * 128 = 112.0 = 112 |
626 | 0 | static __inline uint8_t RGBToY(uint8_t r, uint8_t g, uint8_t b) { |
627 | 0 | return STATIC_CAST(uint8_t, (66 * r + 129 * g + 25 * b + 0x1080) >> 8); |
628 | 0 | } |
629 | 0 | static __inline uint8_t RGBToU(uint8_t r, uint8_t g, uint8_t b) { |
630 | 0 | return STATIC_CAST(uint8_t, (112 * b - 74 * g - 38 * r + 0x8000) >> 8); |
631 | 0 | } |
632 | 0 | static __inline uint8_t RGBToV(uint8_t r, uint8_t g, uint8_t b) { |
633 | 0 | return STATIC_CAST(uint8_t, (112 * r - 94 * g - 18 * b + 0x8000) >> 8); |
634 | 0 | } |
635 | 0 | #define AVGB(a, b) (((a) + (b) + 1) >> 1) |
636 | | |
637 | | #define MAKEROWY(NAME, R, G, B, BPP) \ |
638 | 0 | void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \ |
639 | 0 | int x; \ |
640 | 0 | for (x = 0; x < width; ++x) { \ |
641 | 0 | dst_y[0] = RGBToY(src_rgb[R], src_rgb[G], src_rgb[B]); \ |
642 | 0 | src_rgb += BPP; \ |
643 | 0 | dst_y += 1; \ |
644 | 0 | } \ |
645 | 0 | } \ Unexecuted instantiation: ARGBToYRow_C Unexecuted instantiation: BGRAToYRow_C Unexecuted instantiation: ABGRToYRow_C Unexecuted instantiation: RGBAToYRow_C Unexecuted instantiation: RGB24ToYRow_C Unexecuted instantiation: RAWToYRow_C |
646 | | void NAME##ToUVRow_C(const uint8_t* src_rgb, int src_stride_rgb, \ |
647 | 0 | uint8_t* dst_u, uint8_t* dst_v, int width) { \ |
648 | 0 | const uint8_t* src_rgb1 = src_rgb + src_stride_rgb; \ |
649 | 0 | int x; \ |
650 | 0 | for (x = 0; x < width - 1; x += 2) { \ |
651 | 0 | uint8_t ab = (src_rgb[B] + src_rgb[B + BPP] + src_rgb1[B] + \ |
652 | 0 | src_rgb1[B + BPP] + 2) >> \ |
653 | 0 | 2; \ |
654 | 0 | uint8_t ag = (src_rgb[G] + src_rgb[G + BPP] + src_rgb1[G] + \ |
655 | 0 | src_rgb1[G + BPP] + 2) >> \ |
656 | 0 | 2; \ |
657 | 0 | uint8_t ar = (src_rgb[R] + src_rgb[R + BPP] + src_rgb1[R] + \ |
658 | 0 | src_rgb1[R + BPP] + 2) >> \ |
659 | 0 | 2; \ |
660 | 0 | dst_u[0] = RGBToU(ar, ag, ab); \ |
661 | 0 | dst_v[0] = RGBToV(ar, ag, ab); \ |
662 | 0 | src_rgb += BPP * 2; \ |
663 | 0 | src_rgb1 += BPP * 2; \ |
664 | 0 | dst_u += 1; \ |
665 | 0 | dst_v += 1; \ |
666 | 0 | } \ |
667 | 0 | if (width & 1) { \ |
668 | 0 | uint8_t ab = (src_rgb[B] + src_rgb1[B] + 1) >> 1; \ |
669 | 0 | uint8_t ag = (src_rgb[G] + src_rgb1[G] + 1) >> 1; \ |
670 | 0 | uint8_t ar = (src_rgb[R] + src_rgb1[R] + 1) >> 1; \ |
671 | 0 | dst_u[0] = RGBToU(ar, ag, ab); \ |
672 | 0 | dst_v[0] = RGBToV(ar, ag, ab); \ |
673 | 0 | } \ |
674 | 0 | } Unexecuted instantiation: ARGBToUVRow_C Unexecuted instantiation: BGRAToUVRow_C Unexecuted instantiation: ABGRToUVRow_C Unexecuted instantiation: RGBAToUVRow_C Unexecuted instantiation: RGB24ToUVRow_C Unexecuted instantiation: RAWToUVRow_C |
675 | | |
676 | | MAKEROWY(ARGB, 2, 1, 0, 4) |
677 | | MAKEROWY(BGRA, 1, 2, 3, 4) |
678 | | MAKEROWY(ABGR, 0, 1, 2, 4) |
679 | | MAKEROWY(RGBA, 3, 2, 1, 4) |
680 | | MAKEROWY(RGB24, 2, 1, 0, 3) |
681 | | MAKEROWY(RAW, 0, 1, 2, 3) |
682 | | #undef MAKEROWY |
683 | | |
684 | | // JPeg uses BT.601-1 full range |
685 | | // y = 0.29900 * r + 0.58700 * g + 0.11400 * b |
686 | | // u = -0.16874 * r - 0.33126 * g + 0.50000 * b + center |
687 | | // v = 0.50000 * r - 0.41869 * g - 0.08131 * b + center |
688 | | // JPeg 8 bit Y: |
689 | | // b 0.11400 * 256 = 29.184 = 29 |
690 | | // g 0.58700 * 256 = 150.272 = 150 |
691 | | // r 0.29900 * 256 = 76.544 = 77 |
692 | | // JPeg 8 bit U: |
693 | | // b 0.50000 * 256 = 128.0 = 128 |
694 | | // g -0.33126 * 256 = −84.80256 = -85 |
695 | | // r -0.16874 * 256 = −43.19744 = -43 |
696 | | // JPeg 8 bit V: |
697 | | // b -0.08131 * 256 = −20.81536 = -21 |
698 | | // g -0.41869 * 256 = −107.18464 = -107 |
699 | | // r 0.50000 * 256 = 128.0 = 128 |
700 | | |
701 | | // 8 bit |
702 | 0 | static __inline uint8_t RGBToYJ(uint8_t r, uint8_t g, uint8_t b) { |
703 | 0 | return (77 * r + 150 * g + 29 * b + 128) >> 8; |
704 | 0 | } |
705 | 0 | static __inline uint8_t RGBToUJ(uint8_t r, uint8_t g, uint8_t b) { |
706 | 0 | return (128 * b - 85 * g - 43 * r + 0x8000) >> 8; |
707 | 0 | } |
708 | 0 | static __inline uint8_t RGBToVJ(uint8_t r, uint8_t g, uint8_t b) { |
709 | 0 | return (128 * r - 107 * g - 21 * b + 0x8000) >> 8; |
710 | 0 | } |
711 | | |
712 | | // ARGBToYJ_C and ARGBToUVJ_C |
713 | | #define MAKEROWYJ(NAME, R, G, B, BPP) \ |
714 | 0 | void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \ |
715 | 0 | int x; \ |
716 | 0 | for (x = 0; x < width; ++x) { \ |
717 | 0 | dst_y[0] = RGBToYJ(src_rgb[R], src_rgb[G], src_rgb[B]); \ |
718 | 0 | src_rgb += BPP; \ |
719 | 0 | dst_y += 1; \ |
720 | 0 | } \ |
721 | 0 | } \ Unexecuted instantiation: ARGBToYJRow_C Unexecuted instantiation: ABGRToYJRow_C Unexecuted instantiation: RGBAToYJRow_C Unexecuted instantiation: RGB24ToYJRow_C Unexecuted instantiation: RAWToYJRow_C |
722 | | void NAME##ToUVJRow_C(const uint8_t* src_rgb, int src_stride_rgb, \ |
723 | 0 | uint8_t* dst_u, uint8_t* dst_v, int width) { \ |
724 | 0 | const uint8_t* src_rgb1 = src_rgb + src_stride_rgb; \ |
725 | 0 | int x; \ |
726 | 0 | for (x = 0; x < width - 1; x += 2) { \ |
727 | 0 | uint8_t ab = (src_rgb[B] + src_rgb[B + BPP] + src_rgb1[B] + \ |
728 | 0 | src_rgb1[B + BPP] + 2) >> \ |
729 | 0 | 2; \ |
730 | 0 | uint8_t ag = (src_rgb[G] + src_rgb[G + BPP] + src_rgb1[G] + \ |
731 | 0 | src_rgb1[G + BPP] + 2) >> \ |
732 | 0 | 2; \ |
733 | 0 | uint8_t ar = (src_rgb[R] + src_rgb[R + BPP] + src_rgb1[R] + \ |
734 | 0 | src_rgb1[R + BPP] + 2) >> \ |
735 | 0 | 2; \ |
736 | 0 | dst_u[0] = RGBToUJ(ar, ag, ab); \ |
737 | 0 | dst_v[0] = RGBToVJ(ar, ag, ab); \ |
738 | 0 | src_rgb += BPP * 2; \ |
739 | 0 | src_rgb1 += BPP * 2; \ |
740 | 0 | dst_u += 1; \ |
741 | 0 | dst_v += 1; \ |
742 | 0 | } \ |
743 | 0 | if (width & 1) { \ |
744 | 0 | uint16_t ab = (src_rgb[B] + src_rgb1[B] + 1) >> 1; \ |
745 | 0 | uint16_t ag = (src_rgb[G] + src_rgb1[G] + 1) >> 1; \ |
746 | 0 | uint16_t ar = (src_rgb[R] + src_rgb1[R] + 1) >> 1; \ |
747 | 0 | dst_u[0] = RGBToUJ(ar, ag, ab); \ |
748 | 0 | dst_v[0] = RGBToVJ(ar, ag, ab); \ |
749 | 0 | } \ |
750 | 0 | } Unexecuted instantiation: ARGBToUVJRow_C Unexecuted instantiation: ABGRToUVJRow_C Unexecuted instantiation: RGBAToUVJRow_C Unexecuted instantiation: RGB24ToUVJRow_C Unexecuted instantiation: RAWToUVJRow_C |
751 | | |
752 | | MAKEROWYJ(ARGB, 2, 1, 0, 4) |
753 | | MAKEROWYJ(ABGR, 0, 1, 2, 4) |
754 | | MAKEROWYJ(RGBA, 3, 2, 1, 4) |
755 | | MAKEROWYJ(RGB24, 2, 1, 0, 3) |
756 | | MAKEROWYJ(RAW, 0, 1, 2, 3) |
757 | | #undef MAKEROWYJ |
758 | | |
759 | 0 | void RGB565ToYRow_C(const uint8_t* src_rgb565, uint8_t* dst_y, int width) { |
760 | 0 | int x; |
761 | 0 | for (x = 0; x < width; ++x) { |
762 | 0 | uint8_t b = src_rgb565[0] & 0x1f; |
763 | 0 | uint8_t g = STATIC_CAST( |
764 | 0 | uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3)); |
765 | 0 | uint8_t r = src_rgb565[1] >> 3; |
766 | 0 | b = STATIC_CAST(uint8_t, (b << 3) | (b >> 2)); |
767 | 0 | g = STATIC_CAST(uint8_t, (g << 2) | (g >> 4)); |
768 | 0 | r = STATIC_CAST(uint8_t, (r << 3) | (r >> 2)); |
769 | 0 | dst_y[0] = RGBToY(r, g, b); |
770 | 0 | src_rgb565 += 2; |
771 | 0 | dst_y += 1; |
772 | 0 | } |
773 | 0 | } |
774 | | |
775 | 0 | void ARGB1555ToYRow_C(const uint8_t* src_argb1555, uint8_t* dst_y, int width) { |
776 | 0 | int x; |
777 | 0 | for (x = 0; x < width; ++x) { |
778 | 0 | uint8_t b = src_argb1555[0] & 0x1f; |
779 | 0 | uint8_t g = STATIC_CAST( |
780 | 0 | uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3)); |
781 | 0 | uint8_t r = (src_argb1555[1] & 0x7c) >> 2; |
782 | 0 | b = STATIC_CAST(uint8_t, (b << 3) | (b >> 2)); |
783 | 0 | g = STATIC_CAST(uint8_t, (g << 3) | (g >> 2)); |
784 | 0 | r = STATIC_CAST(uint8_t, (r << 3) | (r >> 2)); |
785 | 0 | dst_y[0] = RGBToY(r, g, b); |
786 | 0 | src_argb1555 += 2; |
787 | 0 | dst_y += 1; |
788 | 0 | } |
789 | 0 | } |
790 | | |
791 | 0 | void ARGB4444ToYRow_C(const uint8_t* src_argb4444, uint8_t* dst_y, int width) { |
792 | 0 | int x; |
793 | 0 | for (x = 0; x < width; ++x) { |
794 | 0 | uint8_t b = src_argb4444[0] & 0x0f; |
795 | 0 | uint8_t g = src_argb4444[0] >> 4; |
796 | 0 | uint8_t r = src_argb4444[1] & 0x0f; |
797 | 0 | b = STATIC_CAST(uint8_t, (b << 4) | b); |
798 | 0 | g = STATIC_CAST(uint8_t, (g << 4) | g); |
799 | 0 | r = STATIC_CAST(uint8_t, (r << 4) | r); |
800 | 0 | dst_y[0] = RGBToY(r, g, b); |
801 | 0 | src_argb4444 += 2; |
802 | 0 | dst_y += 1; |
803 | 0 | } |
804 | 0 | } |
805 | | |
806 | | void RGB565ToUVRow_C(const uint8_t* src_rgb565, |
807 | | int src_stride_rgb565, |
808 | | uint8_t* dst_u, |
809 | | uint8_t* dst_v, |
810 | 0 | int width) { |
811 | 0 | const uint8_t* next_rgb565 = src_rgb565 + src_stride_rgb565; |
812 | 0 | int x; |
813 | 0 | for (x = 0; x < width - 1; x += 2) { |
814 | 0 | uint8_t b0 = STATIC_CAST(uint8_t, src_rgb565[0] & 0x1f); |
815 | 0 | uint8_t g0 = STATIC_CAST( |
816 | 0 | uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3)); |
817 | 0 | uint8_t r0 = STATIC_CAST(uint8_t, src_rgb565[1] >> 3); |
818 | 0 | uint8_t b1 = STATIC_CAST(uint8_t, src_rgb565[2] & 0x1f); |
819 | 0 | uint8_t g1 = STATIC_CAST( |
820 | 0 | uint8_t, (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3)); |
821 | 0 | uint8_t r1 = STATIC_CAST(uint8_t, src_rgb565[3] >> 3); |
822 | 0 | uint8_t b2 = STATIC_CAST(uint8_t, next_rgb565[0] & 0x1f); |
823 | 0 | uint8_t g2 = STATIC_CAST( |
824 | 0 | uint8_t, (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3)); |
825 | 0 | uint8_t r2 = STATIC_CAST(uint8_t, next_rgb565[1] >> 3); |
826 | 0 | uint8_t b3 = STATIC_CAST(uint8_t, next_rgb565[2] & 0x1f); |
827 | 0 | uint8_t g3 = STATIC_CAST( |
828 | 0 | uint8_t, (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3)); |
829 | 0 | uint8_t r3 = STATIC_CAST(uint8_t, next_rgb565[3] >> 3); |
830 | |
|
831 | 0 | b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2)); |
832 | 0 | g0 = STATIC_CAST(uint8_t, (g0 << 2) | (g0 >> 4)); |
833 | 0 | r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2)); |
834 | 0 | b1 = STATIC_CAST(uint8_t, (b1 << 3) | (b1 >> 2)); |
835 | 0 | g1 = STATIC_CAST(uint8_t, (g1 << 2) | (g1 >> 4)); |
836 | 0 | r1 = STATIC_CAST(uint8_t, (r1 << 3) | (r1 >> 2)); |
837 | 0 | b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2)); |
838 | 0 | g2 = STATIC_CAST(uint8_t, (g2 << 2) | (g2 >> 4)); |
839 | 0 | r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2)); |
840 | 0 | b3 = STATIC_CAST(uint8_t, (b3 << 3) | (b3 >> 2)); |
841 | 0 | g3 = STATIC_CAST(uint8_t, (g3 << 2) | (g3 >> 4)); |
842 | 0 | r3 = STATIC_CAST(uint8_t, (r3 << 3) | (r3 >> 2)); |
843 | |
|
844 | 0 | uint8_t b = (b0 + b1 + b2 + b3 + 2) >> 2; |
845 | 0 | uint8_t g = (g0 + g1 + g2 + g3 + 2) >> 2; |
846 | 0 | uint8_t r = (r0 + r1 + r2 + r3 + 2) >> 2; |
847 | 0 | dst_u[0] = RGBToU(r, g, b); |
848 | 0 | dst_v[0] = RGBToV(r, g, b); |
849 | |
|
850 | 0 | src_rgb565 += 4; |
851 | 0 | next_rgb565 += 4; |
852 | 0 | dst_u += 1; |
853 | 0 | dst_v += 1; |
854 | 0 | } |
855 | 0 | if (width & 1) { |
856 | 0 | uint8_t b0 = STATIC_CAST(uint8_t, src_rgb565[0] & 0x1f); |
857 | 0 | uint8_t g0 = STATIC_CAST( |
858 | 0 | uint8_t, (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3)); |
859 | 0 | uint8_t r0 = STATIC_CAST(uint8_t, src_rgb565[1] >> 3); |
860 | 0 | uint8_t b2 = STATIC_CAST(uint8_t, next_rgb565[0] & 0x1f); |
861 | 0 | uint8_t g2 = STATIC_CAST( |
862 | 0 | uint8_t, (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3)); |
863 | 0 | uint8_t r2 = STATIC_CAST(uint8_t, next_rgb565[1] >> 3); |
864 | 0 | b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2)); |
865 | 0 | g0 = STATIC_CAST(uint8_t, (g0 << 2) | (g0 >> 4)); |
866 | 0 | r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2)); |
867 | 0 | b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2)); |
868 | 0 | g2 = STATIC_CAST(uint8_t, (g2 << 2) | (g2 >> 4)); |
869 | 0 | r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2)); |
870 | |
|
871 | 0 | uint8_t ab = AVGB(b0, b2); |
872 | 0 | uint8_t ag = AVGB(g0, g2); |
873 | 0 | uint8_t ar = AVGB(r0, r2); |
874 | 0 | dst_u[0] = RGBToU(ar, ag, ab); |
875 | 0 | dst_v[0] = RGBToV(ar, ag, ab); |
876 | 0 | } |
877 | 0 | } |
878 | | |
879 | | void ARGB1555ToUVRow_C(const uint8_t* src_argb1555, |
880 | | int src_stride_argb1555, |
881 | | uint8_t* dst_u, |
882 | | uint8_t* dst_v, |
883 | 0 | int width) { |
884 | 0 | const uint8_t* next_argb1555 = src_argb1555 + src_stride_argb1555; |
885 | 0 | int x; |
886 | 0 | for (x = 0; x < width - 1; x += 2) { |
887 | 0 | uint8_t b0 = STATIC_CAST(uint8_t, src_argb1555[0] & 0x1f); |
888 | 0 | uint8_t g0 = STATIC_CAST( |
889 | 0 | uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3)); |
890 | 0 | uint8_t r0 = STATIC_CAST(uint8_t, (src_argb1555[1] & 0x7c) >> 2); |
891 | 0 | uint8_t b1 = STATIC_CAST(uint8_t, src_argb1555[2] & 0x1f); |
892 | 0 | uint8_t g1 = STATIC_CAST( |
893 | 0 | uint8_t, (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3)); |
894 | 0 | uint8_t r1 = STATIC_CAST(uint8_t, (src_argb1555[3] & 0x7c) >> 2); |
895 | 0 | uint8_t b2 = STATIC_CAST(uint8_t, next_argb1555[0] & 0x1f); |
896 | 0 | uint8_t g2 = STATIC_CAST( |
897 | 0 | uint8_t, (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3)); |
898 | 0 | uint8_t r2 = STATIC_CAST(uint8_t, (next_argb1555[1] & 0x7c) >> 2); |
899 | 0 | uint8_t b3 = STATIC_CAST(uint8_t, next_argb1555[2] & 0x1f); |
900 | 0 | uint8_t g3 = STATIC_CAST( |
901 | 0 | uint8_t, (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3)); |
902 | 0 | uint8_t r3 = STATIC_CAST(uint8_t, (next_argb1555[3] & 0x7c) >> 2); |
903 | |
|
904 | 0 | b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2)); |
905 | 0 | g0 = STATIC_CAST(uint8_t, (g0 << 3) | (g0 >> 2)); |
906 | 0 | r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2)); |
907 | 0 | b1 = STATIC_CAST(uint8_t, (b1 << 3) | (b1 >> 2)); |
908 | 0 | g1 = STATIC_CAST(uint8_t, (g1 << 3) | (g1 >> 2)); |
909 | 0 | r1 = STATIC_CAST(uint8_t, (r1 << 3) | (r1 >> 2)); |
910 | 0 | b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2)); |
911 | 0 | g2 = STATIC_CAST(uint8_t, (g2 << 3) | (g2 >> 2)); |
912 | 0 | r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2)); |
913 | 0 | b3 = STATIC_CAST(uint8_t, (b3 << 3) | (b3 >> 2)); |
914 | 0 | g3 = STATIC_CAST(uint8_t, (g3 << 3) | (g3 >> 2)); |
915 | 0 | r3 = STATIC_CAST(uint8_t, (r3 << 3) | (r3 >> 2)); |
916 | |
|
917 | 0 | uint8_t b = (b0 + b1 + b2 + b3 + 2) >> 2; |
918 | 0 | uint8_t g = (g0 + g1 + g2 + g3 + 2) >> 2; |
919 | 0 | uint8_t r = (r0 + r1 + r2 + r3 + 2) >> 2; |
920 | 0 | dst_u[0] = RGBToU(r, g, b); |
921 | 0 | dst_v[0] = RGBToV(r, g, b); |
922 | |
|
923 | 0 | src_argb1555 += 4; |
924 | 0 | next_argb1555 += 4; |
925 | 0 | dst_u += 1; |
926 | 0 | dst_v += 1; |
927 | 0 | } |
928 | 0 | if (width & 1) { |
929 | 0 | uint8_t b0 = STATIC_CAST(uint8_t, src_argb1555[0] & 0x1f); |
930 | 0 | uint8_t g0 = STATIC_CAST( |
931 | 0 | uint8_t, (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3)); |
932 | 0 | uint8_t r0 = STATIC_CAST(uint8_t, (src_argb1555[1] & 0x7c) >> 2); |
933 | 0 | uint8_t b2 = STATIC_CAST(uint8_t, next_argb1555[0] & 0x1f); |
934 | 0 | uint8_t g2 = STATIC_CAST( |
935 | 0 | uint8_t, (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3)); |
936 | 0 | uint8_t r2 = STATIC_CAST(uint8_t, (next_argb1555[1] & 0x7c) >> 2); |
937 | |
|
938 | 0 | b0 = STATIC_CAST(uint8_t, (b0 << 3) | (b0 >> 2)); |
939 | 0 | g0 = STATIC_CAST(uint8_t, (g0 << 3) | (g0 >> 2)); |
940 | 0 | r0 = STATIC_CAST(uint8_t, (r0 << 3) | (r0 >> 2)); |
941 | 0 | b2 = STATIC_CAST(uint8_t, (b2 << 3) | (b2 >> 2)); |
942 | 0 | g2 = STATIC_CAST(uint8_t, (g2 << 3) | (g2 >> 2)); |
943 | 0 | r2 = STATIC_CAST(uint8_t, (r2 << 3) | (r2 >> 2)); |
944 | |
|
945 | 0 | uint8_t ab = AVGB(b0, b2); |
946 | 0 | uint8_t ag = AVGB(g0, g2); |
947 | 0 | uint8_t ar = AVGB(r0, r2); |
948 | 0 | dst_u[0] = RGBToU(ar, ag, ab); |
949 | 0 | dst_v[0] = RGBToV(ar, ag, ab); |
950 | 0 | } |
951 | 0 | } |
952 | | |
953 | | void ARGB4444ToUVRow_C(const uint8_t* src_argb4444, |
954 | | int src_stride_argb4444, |
955 | | uint8_t* dst_u, |
956 | | uint8_t* dst_v, |
957 | 0 | int width) { |
958 | 0 | const uint8_t* next_argb4444 = src_argb4444 + src_stride_argb4444; |
959 | 0 | int x; |
960 | 0 | for (x = 0; x < width - 1; x += 2) { |
961 | 0 | uint8_t b0 = src_argb4444[0] & 0x0f; |
962 | 0 | uint8_t g0 = src_argb4444[0] >> 4; |
963 | 0 | uint8_t r0 = src_argb4444[1] & 0x0f; |
964 | 0 | uint8_t b1 = src_argb4444[2] & 0x0f; |
965 | 0 | uint8_t g1 = src_argb4444[2] >> 4; |
966 | 0 | uint8_t r1 = src_argb4444[3] & 0x0f; |
967 | 0 | uint8_t b2 = next_argb4444[0] & 0x0f; |
968 | 0 | uint8_t g2 = next_argb4444[0] >> 4; |
969 | 0 | uint8_t r2 = next_argb4444[1] & 0x0f; |
970 | 0 | uint8_t b3 = next_argb4444[2] & 0x0f; |
971 | 0 | uint8_t g3 = next_argb4444[2] >> 4; |
972 | 0 | uint8_t r3 = next_argb4444[3] & 0x0f; |
973 | |
|
974 | 0 | b0 = STATIC_CAST(uint8_t, (b0 << 4) | b0); |
975 | 0 | g0 = STATIC_CAST(uint8_t, (g0 << 4) | g0); |
976 | 0 | r0 = STATIC_CAST(uint8_t, (r0 << 4) | r0); |
977 | 0 | b1 = STATIC_CAST(uint8_t, (b1 << 4) | b1); |
978 | 0 | g1 = STATIC_CAST(uint8_t, (g1 << 4) | g1); |
979 | 0 | r1 = STATIC_CAST(uint8_t, (r1 << 4) | r1); |
980 | 0 | b2 = STATIC_CAST(uint8_t, (b2 << 4) | b2); |
981 | 0 | g2 = STATIC_CAST(uint8_t, (g2 << 4) | g2); |
982 | 0 | r2 = STATIC_CAST(uint8_t, (r2 << 4) | r2); |
983 | 0 | b3 = STATIC_CAST(uint8_t, (b3 << 4) | b3); |
984 | 0 | g3 = STATIC_CAST(uint8_t, (g3 << 4) | g3); |
985 | 0 | r3 = STATIC_CAST(uint8_t, (r3 << 4) | r3); |
986 | |
|
987 | 0 | uint8_t b = (b0 + b1 + b2 + b3 + 2) >> 2; |
988 | 0 | uint8_t g = (g0 + g1 + g2 + g3 + 2) >> 2; |
989 | 0 | uint8_t r = (r0 + r1 + r2 + r3 + 2) >> 2; |
990 | 0 | dst_u[0] = RGBToU(r, g, b); |
991 | 0 | dst_v[0] = RGBToV(r, g, b); |
992 | |
|
993 | 0 | src_argb4444 += 4; |
994 | 0 | next_argb4444 += 4; |
995 | 0 | dst_u += 1; |
996 | 0 | dst_v += 1; |
997 | 0 | } |
998 | 0 | if (width & 1) { |
999 | 0 | uint8_t b0 = src_argb4444[0] & 0x0f; |
1000 | 0 | uint8_t g0 = src_argb4444[0] >> 4; |
1001 | 0 | uint8_t r0 = src_argb4444[1] & 0x0f; |
1002 | 0 | uint8_t b2 = next_argb4444[0] & 0x0f; |
1003 | 0 | uint8_t g2 = next_argb4444[0] >> 4; |
1004 | 0 | uint8_t r2 = next_argb4444[1] & 0x0f; |
1005 | |
|
1006 | 0 | b0 = STATIC_CAST(uint8_t, (b0 << 4) | b0); |
1007 | 0 | g0 = STATIC_CAST(uint8_t, (g0 << 4) | g0); |
1008 | 0 | r0 = STATIC_CAST(uint8_t, (r0 << 4) | r0); |
1009 | 0 | b2 = STATIC_CAST(uint8_t, (b2 << 4) | b2); |
1010 | 0 | g2 = STATIC_CAST(uint8_t, (g2 << 4) | g2); |
1011 | 0 | r2 = STATIC_CAST(uint8_t, (r2 << 4) | r2); |
1012 | |
|
1013 | 0 | uint8_t ab = AVGB(b0, b2); |
1014 | 0 | uint8_t ag = AVGB(g0, g2); |
1015 | 0 | uint8_t ar = AVGB(r0, r2); |
1016 | 0 | dst_u[0] = RGBToU(ar, ag, ab); |
1017 | 0 | dst_v[0] = RGBToV(ar, ag, ab); |
1018 | 0 | } |
1019 | 0 | } |
1020 | | |
1021 | | void ARGBToUV444Row_C(const uint8_t* src_argb, |
1022 | | uint8_t* dst_u, |
1023 | | uint8_t* dst_v, |
1024 | 0 | int width) { |
1025 | 0 | int x; |
1026 | 0 | for (x = 0; x < width; ++x) { |
1027 | 0 | uint8_t ab = src_argb[0]; |
1028 | 0 | uint8_t ag = src_argb[1]; |
1029 | 0 | uint8_t ar = src_argb[2]; |
1030 | 0 | dst_u[0] = RGBToU(ar, ag, ab); |
1031 | 0 | dst_v[0] = RGBToV(ar, ag, ab); |
1032 | 0 | src_argb += 4; |
1033 | 0 | dst_u += 1; |
1034 | 0 | dst_v += 1; |
1035 | 0 | } |
1036 | 0 | } |
1037 | | |
1038 | | void ARGBToUVJ444Row_C(const uint8_t* src_argb, |
1039 | | uint8_t* dst_u, |
1040 | | uint8_t* dst_v, |
1041 | 0 | int width) { |
1042 | 0 | int x; |
1043 | 0 | for (x = 0; x < width; ++x) { |
1044 | 0 | uint8_t ab = src_argb[0]; |
1045 | 0 | uint8_t ag = src_argb[1]; |
1046 | 0 | uint8_t ar = src_argb[2]; |
1047 | 0 | dst_u[0] = RGBToUJ(ar, ag, ab); |
1048 | 0 | dst_v[0] = RGBToVJ(ar, ag, ab); |
1049 | 0 | src_argb += 4; |
1050 | 0 | dst_u += 1; |
1051 | 0 | dst_v += 1; |
1052 | 0 | } |
1053 | 0 | } |
1054 | | |
1055 | 0 | void ARGBGrayRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width) { |
1056 | 0 | int x; |
1057 | 0 | for (x = 0; x < width; ++x) { |
1058 | 0 | uint8_t y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]); |
1059 | 0 | dst_argb[2] = dst_argb[1] = dst_argb[0] = y; |
1060 | 0 | dst_argb[3] = src_argb[3]; |
1061 | 0 | dst_argb += 4; |
1062 | 0 | src_argb += 4; |
1063 | 0 | } |
1064 | 0 | } |
1065 | | |
1066 | | // Convert a row of image to Sepia tone. |
1067 | 0 | void ARGBSepiaRow_C(uint8_t* dst_argb, int width) { |
1068 | 0 | int x; |
1069 | 0 | for (x = 0; x < width; ++x) { |
1070 | 0 | int b = dst_argb[0]; |
1071 | 0 | int g = dst_argb[1]; |
1072 | 0 | int r = dst_argb[2]; |
1073 | 0 | int sb = (b * 17 + g * 68 + r * 35) >> 7; |
1074 | 0 | int sg = (b * 22 + g * 88 + r * 45) >> 7; |
1075 | 0 | int sr = (b * 24 + g * 98 + r * 50) >> 7; |
1076 | | // b does not over flow. a is preserved from original. |
1077 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, sb); |
1078 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, clamp255(sg)); |
1079 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, clamp255(sr)); |
1080 | 0 | dst_argb += 4; |
1081 | 0 | } |
1082 | 0 | } |
1083 | | |
1084 | | // Apply color matrix to a row of image. Matrix is signed. |
1085 | | // TODO(fbarchard): Consider adding rounding (+32). |
1086 | | void ARGBColorMatrixRow_C(const uint8_t* src_argb, |
1087 | | uint8_t* dst_argb, |
1088 | | const int8_t* matrix_argb, |
1089 | 0 | int width) { |
1090 | 0 | int x; |
1091 | 0 | for (x = 0; x < width; ++x) { |
1092 | 0 | int b = src_argb[0]; |
1093 | 0 | int g = src_argb[1]; |
1094 | 0 | int r = src_argb[2]; |
1095 | 0 | int a = src_argb[3]; |
1096 | 0 | int sb = (b * matrix_argb[0] + g * matrix_argb[1] + r * matrix_argb[2] + |
1097 | 0 | a * matrix_argb[3]) >> |
1098 | 0 | 6; |
1099 | 0 | int sg = (b * matrix_argb[4] + g * matrix_argb[5] + r * matrix_argb[6] + |
1100 | 0 | a * matrix_argb[7]) >> |
1101 | 0 | 6; |
1102 | 0 | int sr = (b * matrix_argb[8] + g * matrix_argb[9] + r * matrix_argb[10] + |
1103 | 0 | a * matrix_argb[11]) >> |
1104 | 0 | 6; |
1105 | 0 | int sa = (b * matrix_argb[12] + g * matrix_argb[13] + r * matrix_argb[14] + |
1106 | 0 | a * matrix_argb[15]) >> |
1107 | 0 | 6; |
1108 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, Clamp(sb)); |
1109 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, Clamp(sg)); |
1110 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, Clamp(sr)); |
1111 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, Clamp(sa)); |
1112 | 0 | src_argb += 4; |
1113 | 0 | dst_argb += 4; |
1114 | 0 | } |
1115 | 0 | } |
1116 | | |
1117 | | // Apply color table to a row of image. |
1118 | | void ARGBColorTableRow_C(uint8_t* dst_argb, |
1119 | | const uint8_t* table_argb, |
1120 | 0 | int width) { |
1121 | 0 | int x; |
1122 | 0 | for (x = 0; x < width; ++x) { |
1123 | 0 | int b = dst_argb[0]; |
1124 | 0 | int g = dst_argb[1]; |
1125 | 0 | int r = dst_argb[2]; |
1126 | 0 | int a = dst_argb[3]; |
1127 | 0 | dst_argb[0] = table_argb[b * 4 + 0]; |
1128 | 0 | dst_argb[1] = table_argb[g * 4 + 1]; |
1129 | 0 | dst_argb[2] = table_argb[r * 4 + 2]; |
1130 | 0 | dst_argb[3] = table_argb[a * 4 + 3]; |
1131 | 0 | dst_argb += 4; |
1132 | 0 | } |
1133 | 0 | } |
1134 | | |
1135 | | // Apply color table to a row of image. |
1136 | | void RGBColorTableRow_C(uint8_t* dst_argb, |
1137 | | const uint8_t* table_argb, |
1138 | 0 | int width) { |
1139 | 0 | int x; |
1140 | 0 | for (x = 0; x < width; ++x) { |
1141 | 0 | int b = dst_argb[0]; |
1142 | 0 | int g = dst_argb[1]; |
1143 | 0 | int r = dst_argb[2]; |
1144 | 0 | dst_argb[0] = table_argb[b * 4 + 0]; |
1145 | 0 | dst_argb[1] = table_argb[g * 4 + 1]; |
1146 | 0 | dst_argb[2] = table_argb[r * 4 + 2]; |
1147 | 0 | dst_argb += 4; |
1148 | 0 | } |
1149 | 0 | } |
1150 | | |
1151 | | void ARGBQuantizeRow_C(uint8_t* dst_argb, |
1152 | | int scale, |
1153 | | int interval_size, |
1154 | | int interval_offset, |
1155 | 0 | int width) { |
1156 | 0 | int x; |
1157 | 0 | for (x = 0; x < width; ++x) { |
1158 | 0 | int b = dst_argb[0]; |
1159 | 0 | int g = dst_argb[1]; |
1160 | 0 | int r = dst_argb[2]; |
1161 | 0 | dst_argb[0] = STATIC_CAST( |
1162 | 0 | uint8_t, (b * scale >> 16) * interval_size + interval_offset); |
1163 | 0 | dst_argb[1] = STATIC_CAST( |
1164 | 0 | uint8_t, (g * scale >> 16) * interval_size + interval_offset); |
1165 | 0 | dst_argb[2] = STATIC_CAST( |
1166 | 0 | uint8_t, (r * scale >> 16) * interval_size + interval_offset); |
1167 | 0 | dst_argb += 4; |
1168 | 0 | } |
1169 | 0 | } |
1170 | | |
1171 | 0 | #define REPEAT8(v) (v) | ((v) << 8) |
1172 | 0 | #define SHADE(f, v) v* f >> 24 |
1173 | | |
1174 | | void ARGBShadeRow_C(const uint8_t* src_argb, |
1175 | | uint8_t* dst_argb, |
1176 | | int width, |
1177 | 0 | uint32_t value) { |
1178 | 0 | const uint32_t b_scale = REPEAT8(value & 0xff); |
1179 | 0 | const uint32_t g_scale = REPEAT8((value >> 8) & 0xff); |
1180 | 0 | const uint32_t r_scale = REPEAT8((value >> 16) & 0xff); |
1181 | 0 | const uint32_t a_scale = REPEAT8(value >> 24); |
1182 | |
|
1183 | 0 | int i; |
1184 | 0 | for (i = 0; i < width; ++i) { |
1185 | 0 | const uint32_t b = REPEAT8(src_argb[0]); |
1186 | 0 | const uint32_t g = REPEAT8(src_argb[1]); |
1187 | 0 | const uint32_t r = REPEAT8(src_argb[2]); |
1188 | 0 | const uint32_t a = REPEAT8(src_argb[3]); |
1189 | 0 | dst_argb[0] = SHADE(b, b_scale); |
1190 | 0 | dst_argb[1] = SHADE(g, g_scale); |
1191 | 0 | dst_argb[2] = SHADE(r, r_scale); |
1192 | 0 | dst_argb[3] = SHADE(a, a_scale); |
1193 | 0 | src_argb += 4; |
1194 | 0 | dst_argb += 4; |
1195 | 0 | } |
1196 | 0 | } |
1197 | | #undef REPEAT8 |
1198 | | #undef SHADE |
1199 | | |
1200 | | void ARGBMultiplyRow_C(const uint8_t* src_argb, |
1201 | | const uint8_t* src_argb1, |
1202 | | uint8_t* dst_argb, |
1203 | 0 | int width) { |
1204 | 0 | int i; |
1205 | 0 | for (i = 0; i < width; ++i) { |
1206 | 0 | const uint32_t b = src_argb[0]; |
1207 | 0 | const uint32_t g = src_argb[1]; |
1208 | 0 | const uint32_t r = src_argb[2]; |
1209 | 0 | const uint32_t a = src_argb[3]; |
1210 | 0 | const uint32_t b_scale = src_argb1[0]; |
1211 | 0 | const uint32_t g_scale = src_argb1[1]; |
1212 | 0 | const uint32_t r_scale = src_argb1[2]; |
1213 | 0 | const uint32_t a_scale = src_argb1[3]; |
1214 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, (b * b_scale + 128) >> 8); |
1215 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, (g * g_scale + 128) >> 8); |
1216 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, (r * r_scale + 128) >> 8); |
1217 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, (a * a_scale + 128) >> 8); |
1218 | 0 | src_argb += 4; |
1219 | 0 | src_argb1 += 4; |
1220 | 0 | dst_argb += 4; |
1221 | 0 | } |
1222 | 0 | } |
1223 | | |
1224 | | #define SHADE(f, v) clamp255(v + f) |
1225 | | |
1226 | | void ARGBAddRow_C(const uint8_t* src_argb, |
1227 | | const uint8_t* src_argb1, |
1228 | | uint8_t* dst_argb, |
1229 | 0 | int width) { |
1230 | 0 | int i; |
1231 | 0 | for (i = 0; i < width; ++i) { |
1232 | 0 | const int b = src_argb[0]; |
1233 | 0 | const int g = src_argb[1]; |
1234 | 0 | const int r = src_argb[2]; |
1235 | 0 | const int a = src_argb[3]; |
1236 | 0 | const int b_add = src_argb1[0]; |
1237 | 0 | const int g_add = src_argb1[1]; |
1238 | 0 | const int r_add = src_argb1[2]; |
1239 | 0 | const int a_add = src_argb1[3]; |
1240 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, SHADE(b, b_add)); |
1241 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, SHADE(g, g_add)); |
1242 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, SHADE(r, r_add)); |
1243 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, SHADE(a, a_add)); |
1244 | 0 | src_argb += 4; |
1245 | 0 | src_argb1 += 4; |
1246 | 0 | dst_argb += 4; |
1247 | 0 | } |
1248 | 0 | } |
1249 | | #undef SHADE |
1250 | | |
1251 | | #define SHADE(f, v) clamp0(f - v) |
1252 | | |
1253 | | void ARGBSubtractRow_C(const uint8_t* src_argb, |
1254 | | const uint8_t* src_argb1, |
1255 | | uint8_t* dst_argb, |
1256 | 0 | int width) { |
1257 | 0 | int i; |
1258 | 0 | for (i = 0; i < width; ++i) { |
1259 | 0 | const int b = src_argb[0]; |
1260 | 0 | const int g = src_argb[1]; |
1261 | 0 | const int r = src_argb[2]; |
1262 | 0 | const int a = src_argb[3]; |
1263 | 0 | const int b_sub = src_argb1[0]; |
1264 | 0 | const int g_sub = src_argb1[1]; |
1265 | 0 | const int r_sub = src_argb1[2]; |
1266 | 0 | const int a_sub = src_argb1[3]; |
1267 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, SHADE(b, b_sub)); |
1268 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, SHADE(g, g_sub)); |
1269 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, SHADE(r, r_sub)); |
1270 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, SHADE(a, a_sub)); |
1271 | 0 | src_argb += 4; |
1272 | 0 | src_argb1 += 4; |
1273 | 0 | dst_argb += 4; |
1274 | 0 | } |
1275 | 0 | } |
1276 | | #undef SHADE |
1277 | | |
1278 | | // Sobel functions which mimics SSSE3. |
1279 | | void SobelXRow_C(const uint8_t* src_y0, |
1280 | | const uint8_t* src_y1, |
1281 | | const uint8_t* src_y2, |
1282 | | uint8_t* dst_sobelx, |
1283 | 0 | int width) { |
1284 | 0 | int i; |
1285 | 0 | for (i = 0; i < width; ++i) { |
1286 | 0 | int a = src_y0[i]; |
1287 | 0 | int b = src_y1[i]; |
1288 | 0 | int c = src_y2[i]; |
1289 | 0 | int a_sub = src_y0[i + 2]; |
1290 | 0 | int b_sub = src_y1[i + 2]; |
1291 | 0 | int c_sub = src_y2[i + 2]; |
1292 | 0 | int a_diff = a - a_sub; |
1293 | 0 | int b_diff = b - b_sub; |
1294 | 0 | int c_diff = c - c_sub; |
1295 | 0 | int sobel = Abs(a_diff + b_diff * 2 + c_diff); |
1296 | 0 | dst_sobelx[i] = (uint8_t)(clamp255(sobel)); |
1297 | 0 | } |
1298 | 0 | } |
1299 | | |
1300 | | void SobelYRow_C(const uint8_t* src_y0, |
1301 | | const uint8_t* src_y1, |
1302 | | uint8_t* dst_sobely, |
1303 | 0 | int width) { |
1304 | 0 | int i; |
1305 | 0 | for (i = 0; i < width; ++i) { |
1306 | 0 | int a = src_y0[i + 0]; |
1307 | 0 | int b = src_y0[i + 1]; |
1308 | 0 | int c = src_y0[i + 2]; |
1309 | 0 | int a_sub = src_y1[i + 0]; |
1310 | 0 | int b_sub = src_y1[i + 1]; |
1311 | 0 | int c_sub = src_y1[i + 2]; |
1312 | 0 | int a_diff = a - a_sub; |
1313 | 0 | int b_diff = b - b_sub; |
1314 | 0 | int c_diff = c - c_sub; |
1315 | 0 | int sobel = Abs(a_diff + b_diff * 2 + c_diff); |
1316 | 0 | dst_sobely[i] = (uint8_t)(clamp255(sobel)); |
1317 | 0 | } |
1318 | 0 | } |
1319 | | |
1320 | | void SobelRow_C(const uint8_t* src_sobelx, |
1321 | | const uint8_t* src_sobely, |
1322 | | uint8_t* dst_argb, |
1323 | 0 | int width) { |
1324 | 0 | int i; |
1325 | 0 | for (i = 0; i < width; ++i) { |
1326 | 0 | int r = src_sobelx[i]; |
1327 | 0 | int b = src_sobely[i]; |
1328 | 0 | int s = clamp255(r + b); |
1329 | 0 | dst_argb[0] = (uint8_t)(s); |
1330 | 0 | dst_argb[1] = (uint8_t)(s); |
1331 | 0 | dst_argb[2] = (uint8_t)(s); |
1332 | 0 | dst_argb[3] = (uint8_t)(255u); |
1333 | 0 | dst_argb += 4; |
1334 | 0 | } |
1335 | 0 | } |
1336 | | |
1337 | | void SobelToPlaneRow_C(const uint8_t* src_sobelx, |
1338 | | const uint8_t* src_sobely, |
1339 | | uint8_t* dst_y, |
1340 | 0 | int width) { |
1341 | 0 | int i; |
1342 | 0 | for (i = 0; i < width; ++i) { |
1343 | 0 | int r = src_sobelx[i]; |
1344 | 0 | int b = src_sobely[i]; |
1345 | 0 | int s = clamp255(r + b); |
1346 | 0 | dst_y[i] = (uint8_t)(s); |
1347 | 0 | } |
1348 | 0 | } |
1349 | | |
1350 | | void SobelXYRow_C(const uint8_t* src_sobelx, |
1351 | | const uint8_t* src_sobely, |
1352 | | uint8_t* dst_argb, |
1353 | 0 | int width) { |
1354 | 0 | int i; |
1355 | 0 | for (i = 0; i < width; ++i) { |
1356 | 0 | int r = src_sobelx[i]; |
1357 | 0 | int b = src_sobely[i]; |
1358 | 0 | int g = clamp255(r + b); |
1359 | 0 | dst_argb[0] = (uint8_t)(b); |
1360 | 0 | dst_argb[1] = (uint8_t)(g); |
1361 | 0 | dst_argb[2] = (uint8_t)(r); |
1362 | 0 | dst_argb[3] = (uint8_t)(255u); |
1363 | 0 | dst_argb += 4; |
1364 | 0 | } |
1365 | 0 | } |
1366 | | |
1367 | 0 | void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width) { |
1368 | | // Copy a Y to RGB. |
1369 | 0 | int x; |
1370 | 0 | for (x = 0; x < width; ++x) { |
1371 | 0 | uint8_t y = src_y[0]; |
1372 | 0 | dst_argb[2] = dst_argb[1] = dst_argb[0] = y; |
1373 | 0 | dst_argb[3] = 255u; |
1374 | 0 | dst_argb += 4; |
1375 | 0 | ++src_y; |
1376 | 0 | } |
1377 | 0 | } |
1378 | | |
1379 | | // Macros to create SIMD specific yuv to rgb conversion constants. |
1380 | | |
1381 | | // clang-format off |
1382 | | |
1383 | | #if defined(__aarch64__) || defined(__arm__) || defined(__riscv) |
1384 | | // Bias values include subtract 128 from U and V, bias from Y and rounding. |
1385 | | // For B and R bias is negative. For G bias is positive. |
1386 | | #define YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR) \ |
1387 | | {{UB, VR, UG, VG, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}, \ |
1388 | | {YG, (UB * 128 - YB), (UG * 128 + VG * 128 + YB), (VR * 128 - YB), YB, 0, \ |
1389 | | 0, 0}} |
1390 | | #else |
1391 | | #define YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR) \ |
1392 | | {{UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, \ |
1393 | | UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0}, \ |
1394 | | {UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, \ |
1395 | | UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG}, \ |
1396 | | {0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, \ |
1397 | | 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR}, \ |
1398 | | {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}, \ |
1399 | | {YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB}} |
1400 | | #endif |
1401 | | |
1402 | | // clang-format on |
1403 | | |
1404 | | #define MAKEYUVCONSTANTS(name, YG, YB, UB, UG, VG, VR) \ |
1405 | | const struct YuvConstants SIMD_ALIGNED(kYuv##name##Constants) = \ |
1406 | | YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR); \ |
1407 | | const struct YuvConstants SIMD_ALIGNED(kYvu##name##Constants) = \ |
1408 | | YUVCONSTANTSBODY(YG, YB, VR, VG, UG, UB); |
1409 | | |
1410 | | // TODO(fbarchard): Generate SIMD structures from float matrix. |
1411 | | |
1412 | | // BT.601 limited range YUV to RGB reference |
1413 | | // R = (Y - 16) * 1.164 + V * 1.596 |
1414 | | // G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813 |
1415 | | // B = (Y - 16) * 1.164 + U * 2.018 |
1416 | | // KR = 0.299; KB = 0.114 |
1417 | | |
1418 | | // U and V contributions to R,G,B. |
1419 | | #if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT601) |
1420 | | #define UB 129 /* round(2.018 * 64) */ |
1421 | | #else |
1422 | | #define UB 128 /* max(128, round(2.018 * 64)) */ |
1423 | | #endif |
1424 | | #define UG 25 /* round(0.391 * 64) */ |
1425 | | #define VG 52 /* round(0.813 * 64) */ |
1426 | | #define VR 102 /* round(1.596 * 64) */ |
1427 | | |
1428 | | // Y contribution to R,G,B. Scale and bias. |
1429 | | #define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ |
1430 | | #define YB -1160 /* 1.164 * 64 * -16 + 64 / 2 */ |
1431 | | |
1432 | | MAKEYUVCONSTANTS(I601, YG, YB, UB, UG, VG, VR) |
1433 | | |
1434 | | #undef YG |
1435 | | #undef YB |
1436 | | #undef UB |
1437 | | #undef UG |
1438 | | #undef VG |
1439 | | #undef VR |
1440 | | |
1441 | | // BT.601 full range YUV to RGB reference (aka JPEG) |
1442 | | // * R = Y + V * 1.40200 |
1443 | | // * G = Y - U * 0.34414 - V * 0.71414 |
1444 | | // * B = Y + U * 1.77200 |
1445 | | // KR = 0.299; KB = 0.114 |
1446 | | |
1447 | | // U and V contributions to R,G,B. |
1448 | | #define UB 113 /* round(1.77200 * 64) */ |
1449 | | #define UG 22 /* round(0.34414 * 64) */ |
1450 | | #define VG 46 /* round(0.71414 * 64) */ |
1451 | | #define VR 90 /* round(1.40200 * 64) */ |
1452 | | |
1453 | | // Y contribution to R,G,B. Scale and bias. |
1454 | | #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */ |
1455 | | #define YB 32 /* 64 / 2 */ |
1456 | | |
1457 | | MAKEYUVCONSTANTS(JPEG, YG, YB, UB, UG, VG, VR) |
1458 | | |
1459 | | #undef YG |
1460 | | #undef YB |
1461 | | #undef UB |
1462 | | #undef UG |
1463 | | #undef VG |
1464 | | #undef VR |
1465 | | |
1466 | | // BT.709 limited range YUV to RGB reference |
1467 | | // R = (Y - 16) * 1.164 + V * 1.793 |
1468 | | // G = (Y - 16) * 1.164 - U * 0.213 - V * 0.533 |
1469 | | // B = (Y - 16) * 1.164 + U * 2.112 |
1470 | | // KR = 0.2126, KB = 0.0722 |
1471 | | |
1472 | | // U and V contributions to R,G,B. |
1473 | | #if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT709) |
1474 | | #define UB 135 /* round(2.112 * 64) */ |
1475 | | #else |
1476 | | #define UB 128 /* max(128, round(2.112 * 64)) */ |
1477 | | #endif |
1478 | | #define UG 14 /* round(0.213 * 64) */ |
1479 | | #define VG 34 /* round(0.533 * 64) */ |
1480 | | #define VR 115 /* round(1.793 * 64) */ |
1481 | | |
1482 | | // Y contribution to R,G,B. Scale and bias. |
1483 | | #define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */ |
1484 | | #define YB -1160 /* 1.164 * 64 * -16 + 64 / 2 */ |
1485 | | |
1486 | | MAKEYUVCONSTANTS(H709, YG, YB, UB, UG, VG, VR) |
1487 | | |
1488 | | #undef YG |
1489 | | #undef YB |
1490 | | #undef UB |
1491 | | #undef UG |
1492 | | #undef VG |
1493 | | #undef VR |
1494 | | |
1495 | | // BT.709 full range YUV to RGB reference |
1496 | | // R = Y + V * 1.5748 |
1497 | | // G = Y - U * 0.18732 - V * 0.46812 |
1498 | | // B = Y + U * 1.8556 |
1499 | | // KR = 0.2126, KB = 0.0722 |
1500 | | |
1501 | | // U and V contributions to R,G,B. |
1502 | | #define UB 119 /* round(1.8556 * 64) */ |
1503 | | #define UG 12 /* round(0.18732 * 64) */ |
1504 | | #define VG 30 /* round(0.46812 * 64) */ |
1505 | | #define VR 101 /* round(1.5748 * 64) */ |
1506 | | |
1507 | | // Y contribution to R,G,B. Scale and bias. (same as jpeg) |
1508 | | #define YG 16320 /* round(1 * 64 * 256 * 256 / 257) */ |
1509 | | #define YB 32 /* 64 / 2 */ |
1510 | | |
1511 | | MAKEYUVCONSTANTS(F709, YG, YB, UB, UG, VG, VR) |
1512 | | |
1513 | | #undef YG |
1514 | | #undef YB |
1515 | | #undef UB |
1516 | | #undef UG |
1517 | | #undef VG |
1518 | | #undef VR |
1519 | | |
1520 | | // BT.2020 limited range YUV to RGB reference |
1521 | | // R = (Y - 16) * 1.164384 + V * 1.67867 |
1522 | | // G = (Y - 16) * 1.164384 - U * 0.187326 - V * 0.65042 |
1523 | | // B = (Y - 16) * 1.164384 + U * 2.14177 |
1524 | | // KR = 0.2627; KB = 0.0593 |
1525 | | |
1526 | | // U and V contributions to R,G,B. |
1527 | | #if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT2020) |
1528 | | #define UB 137 /* round(2.142 * 64) */ |
1529 | | #else |
1530 | | #define UB 128 /* max(128, round(2.142 * 64)) */ |
1531 | | #endif |
1532 | | #define UG 12 /* round(0.187326 * 64) */ |
1533 | | #define VG 42 /* round(0.65042 * 64) */ |
1534 | | #define VR 107 /* round(1.67867 * 64) */ |
1535 | | |
1536 | | // Y contribution to R,G,B. Scale and bias. |
1537 | | #define YG 19003 /* round(1.164384 * 64 * 256 * 256 / 257) */ |
1538 | | #define YB -1160 /* 1.164384 * 64 * -16 + 64 / 2 */ |
1539 | | |
1540 | | MAKEYUVCONSTANTS(2020, YG, YB, UB, UG, VG, VR) |
1541 | | |
1542 | | #undef YG |
1543 | | #undef YB |
1544 | | #undef UB |
1545 | | #undef UG |
1546 | | #undef VG |
1547 | | #undef VR |
1548 | | |
1549 | | // BT.2020 full range YUV to RGB reference |
1550 | | // R = Y + V * 1.474600 |
1551 | | // G = Y - U * 0.164553 - V * 0.571353 |
1552 | | // B = Y + U * 1.881400 |
1553 | | // KR = 0.2627; KB = 0.0593 |
1554 | | |
1555 | | #define UB 120 /* round(1.881400 * 64) */ |
1556 | | #define UG 11 /* round(0.164553 * 64) */ |
1557 | | #define VG 37 /* round(0.571353 * 64) */ |
1558 | | #define VR 94 /* round(1.474600 * 64) */ |
1559 | | |
1560 | | // Y contribution to R,G,B. Scale and bias. (same as jpeg) |
1561 | | #define YG 16320 /* round(1 * 64 * 256 * 256 / 257) */ |
1562 | | #define YB 32 /* 64 / 2 */ |
1563 | | |
1564 | | MAKEYUVCONSTANTS(V2020, YG, YB, UB, UG, VG, VR) |
1565 | | |
1566 | | #undef YG |
1567 | | #undef YB |
1568 | | #undef UB |
1569 | | #undef UG |
1570 | | #undef VG |
1571 | | #undef VR |
1572 | | |
1573 | | #undef BB |
1574 | | #undef BG |
1575 | | #undef BR |
1576 | | |
1577 | | #undef MAKEYUVCONSTANTS |
1578 | | |
1579 | | #if defined(__aarch64__) || defined(__arm__) || defined(__riscv) |
1580 | | #define LOAD_YUV_CONSTANTS \ |
1581 | | int ub = yuvconstants->kUVCoeff[0]; \ |
1582 | | int vr = yuvconstants->kUVCoeff[1]; \ |
1583 | | int ug = yuvconstants->kUVCoeff[2]; \ |
1584 | | int vg = yuvconstants->kUVCoeff[3]; \ |
1585 | | int yg = yuvconstants->kRGBCoeffBias[0]; \ |
1586 | | int bb = yuvconstants->kRGBCoeffBias[1]; \ |
1587 | | int bg = yuvconstants->kRGBCoeffBias[2]; \ |
1588 | | int br = yuvconstants->kRGBCoeffBias[3] |
1589 | | |
1590 | | #define CALC_RGB16 \ |
1591 | | int32_t y1 = (uint32_t)(y32 * yg) >> 16; \ |
1592 | | int b16 = y1 + (u * ub) - bb; \ |
1593 | | int g16 = y1 + bg - (u * ug + v * vg); \ |
1594 | | int r16 = y1 + (v * vr) - br |
1595 | | #else |
1596 | | #define LOAD_YUV_CONSTANTS \ |
1597 | 0 | int ub = yuvconstants->kUVToB[0]; \ |
1598 | 0 | int ug = yuvconstants->kUVToG[0]; \ |
1599 | 0 | int vg = yuvconstants->kUVToG[1]; \ |
1600 | 0 | int vr = yuvconstants->kUVToR[1]; \ |
1601 | 0 | int yg = yuvconstants->kYToRgb[0]; \ |
1602 | 0 | int yb = yuvconstants->kYBiasToRgb[0] |
1603 | | |
1604 | | #define CALC_RGB16 \ |
1605 | 0 | int32_t y1 = ((uint32_t)(y32 * yg) >> 16) + yb; \ |
1606 | 0 | int8_t ui = (int8_t)u; \ |
1607 | 0 | int8_t vi = (int8_t)v; \ |
1608 | 0 | ui -= 0x80; \ |
1609 | 0 | vi -= 0x80; \ |
1610 | 0 | int b16 = y1 + (ui * ub); \ |
1611 | 0 | int g16 = y1 - (ui * ug + vi * vg); \ |
1612 | 0 | int r16 = y1 + (vi * vr) |
1613 | | #endif |
1614 | | |
1615 | | // C reference code that mimics the YUV assembly. |
1616 | | // Reads 8 bit YUV and leaves result as 16 bit. |
1617 | | static __inline void YuvPixel(uint8_t y, |
1618 | | uint8_t u, |
1619 | | uint8_t v, |
1620 | | uint8_t* b, |
1621 | | uint8_t* g, |
1622 | | uint8_t* r, |
1623 | 0 | const struct YuvConstants* yuvconstants) { |
1624 | 0 | LOAD_YUV_CONSTANTS; |
1625 | 0 | uint32_t y32 = y * 0x0101; |
1626 | 0 | CALC_RGB16; |
1627 | 0 | *b = STATIC_CAST(uint8_t, Clamp((int32_t)(b16) >> 6)); |
1628 | 0 | *g = STATIC_CAST(uint8_t, Clamp((int32_t)(g16) >> 6)); |
1629 | 0 | *r = STATIC_CAST(uint8_t, Clamp((int32_t)(r16) >> 6)); |
1630 | 0 | } |
1631 | | |
1632 | | // Reads 8 bit YUV and leaves result as 16 bit. |
1633 | | static __inline void YuvPixel8_16(uint8_t y, |
1634 | | uint8_t u, |
1635 | | uint8_t v, |
1636 | | int* b, |
1637 | | int* g, |
1638 | | int* r, |
1639 | 0 | const struct YuvConstants* yuvconstants) { |
1640 | 0 | LOAD_YUV_CONSTANTS; |
1641 | 0 | uint32_t y32 = y * 0x0101; |
1642 | 0 | CALC_RGB16; |
1643 | 0 | *b = b16; |
1644 | 0 | *g = g16; |
1645 | 0 | *r = r16; |
1646 | 0 | } |
1647 | | |
1648 | | // C reference code that mimics the YUV 16 bit assembly. |
1649 | | // Reads 10 bit YUV and leaves result as 16 bit. |
1650 | | static __inline void YuvPixel10_16(uint16_t y, |
1651 | | uint16_t u, |
1652 | | uint16_t v, |
1653 | | int* b, |
1654 | | int* g, |
1655 | | int* r, |
1656 | 0 | const struct YuvConstants* yuvconstants) { |
1657 | 0 | LOAD_YUV_CONSTANTS; |
1658 | 0 | uint32_t y32 = (y << 6) | (y >> 4); |
1659 | 0 | u = STATIC_CAST(uint8_t, clamp255(u >> 2)); |
1660 | 0 | v = STATIC_CAST(uint8_t, clamp255(v >> 2)); |
1661 | 0 | CALC_RGB16; |
1662 | 0 | *b = b16; |
1663 | 0 | *g = g16; |
1664 | 0 | *r = r16; |
1665 | 0 | } |
1666 | | |
1667 | | // C reference code that mimics the YUV 16 bit assembly. |
1668 | | // Reads 12 bit YUV and leaves result as 16 bit. |
1669 | | static __inline void YuvPixel12_16(int16_t y, |
1670 | | int16_t u, |
1671 | | int16_t v, |
1672 | | int* b, |
1673 | | int* g, |
1674 | | int* r, |
1675 | 0 | const struct YuvConstants* yuvconstants) { |
1676 | 0 | LOAD_YUV_CONSTANTS; |
1677 | 0 | uint32_t y32 = (y << 4) | (y >> 8); |
1678 | 0 | u = STATIC_CAST(uint8_t, clamp255(u >> 4)); |
1679 | 0 | v = STATIC_CAST(uint8_t, clamp255(v >> 4)); |
1680 | 0 | CALC_RGB16; |
1681 | 0 | *b = b16; |
1682 | 0 | *g = g16; |
1683 | 0 | *r = r16; |
1684 | 0 | } |
1685 | | |
1686 | | // C reference code that mimics the YUV 10 bit assembly. |
1687 | | // Reads 10 bit YUV and clamps down to 8 bit RGB. |
1688 | | static __inline void YuvPixel10(uint16_t y, |
1689 | | uint16_t u, |
1690 | | uint16_t v, |
1691 | | uint8_t* b, |
1692 | | uint8_t* g, |
1693 | | uint8_t* r, |
1694 | 0 | const struct YuvConstants* yuvconstants) { |
1695 | 0 | int b16; |
1696 | 0 | int g16; |
1697 | 0 | int r16; |
1698 | 0 | YuvPixel10_16(y, u, v, &b16, &g16, &r16, yuvconstants); |
1699 | 0 | *b = STATIC_CAST(uint8_t, Clamp(b16 >> 6)); |
1700 | 0 | *g = STATIC_CAST(uint8_t, Clamp(g16 >> 6)); |
1701 | 0 | *r = STATIC_CAST(uint8_t, Clamp(r16 >> 6)); |
1702 | 0 | } |
1703 | | |
1704 | | // C reference code that mimics the YUV 12 bit assembly. |
1705 | | // Reads 12 bit YUV and clamps down to 8 bit RGB. |
1706 | | static __inline void YuvPixel12(uint16_t y, |
1707 | | uint16_t u, |
1708 | | uint16_t v, |
1709 | | uint8_t* b, |
1710 | | uint8_t* g, |
1711 | | uint8_t* r, |
1712 | 0 | const struct YuvConstants* yuvconstants) { |
1713 | 0 | int b16; |
1714 | 0 | int g16; |
1715 | 0 | int r16; |
1716 | 0 | YuvPixel12_16(y, u, v, &b16, &g16, &r16, yuvconstants); |
1717 | 0 | *b = STATIC_CAST(uint8_t, Clamp(b16 >> 6)); |
1718 | 0 | *g = STATIC_CAST(uint8_t, Clamp(g16 >> 6)); |
1719 | 0 | *r = STATIC_CAST(uint8_t, Clamp(r16 >> 6)); |
1720 | 0 | } |
1721 | | |
1722 | | // C reference code that mimics the YUV 16 bit assembly. |
1723 | | // Reads 16 bit YUV and leaves result as 8 bit. |
1724 | | static __inline void YuvPixel16_8(uint16_t y, |
1725 | | uint16_t u, |
1726 | | uint16_t v, |
1727 | | uint8_t* b, |
1728 | | uint8_t* g, |
1729 | | uint8_t* r, |
1730 | 0 | const struct YuvConstants* yuvconstants) { |
1731 | 0 | LOAD_YUV_CONSTANTS; |
1732 | 0 | uint32_t y32 = y; |
1733 | 0 | u = STATIC_CAST(uint16_t, clamp255(u >> 8)); |
1734 | 0 | v = STATIC_CAST(uint16_t, clamp255(v >> 8)); |
1735 | 0 | CALC_RGB16; |
1736 | 0 | *b = STATIC_CAST(uint8_t, Clamp((int32_t)(b16) >> 6)); |
1737 | 0 | *g = STATIC_CAST(uint8_t, Clamp((int32_t)(g16) >> 6)); |
1738 | 0 | *r = STATIC_CAST(uint8_t, Clamp((int32_t)(r16) >> 6)); |
1739 | 0 | } |
1740 | | |
1741 | | // C reference code that mimics the YUV 16 bit assembly. |
1742 | | // Reads 16 bit YUV and leaves result as 16 bit. |
1743 | | static __inline void YuvPixel16_16(uint16_t y, |
1744 | | uint16_t u, |
1745 | | uint16_t v, |
1746 | | int* b, |
1747 | | int* g, |
1748 | | int* r, |
1749 | 0 | const struct YuvConstants* yuvconstants) { |
1750 | 0 | LOAD_YUV_CONSTANTS; |
1751 | 0 | uint32_t y32 = y; |
1752 | 0 | u = STATIC_CAST(uint16_t, clamp255(u >> 8)); |
1753 | 0 | v = STATIC_CAST(uint16_t, clamp255(v >> 8)); |
1754 | 0 | CALC_RGB16; |
1755 | 0 | *b = b16; |
1756 | 0 | *g = g16; |
1757 | 0 | *r = r16; |
1758 | 0 | } |
1759 | | |
1760 | | // C reference code that mimics the YUV assembly. |
1761 | | // Reads 8 bit YUV and leaves result as 8 bit. |
1762 | | static __inline void YPixel(uint8_t y, |
1763 | | uint8_t* b, |
1764 | | uint8_t* g, |
1765 | | uint8_t* r, |
1766 | 0 | const struct YuvConstants* yuvconstants) { |
1767 | | #if defined(__aarch64__) || defined(__arm__) || defined(__riscv) |
1768 | | int yg = yuvconstants->kRGBCoeffBias[0]; |
1769 | | int ygb = yuvconstants->kRGBCoeffBias[4]; |
1770 | | #else |
1771 | 0 | int ygb = yuvconstants->kYBiasToRgb[0]; |
1772 | 0 | int yg = yuvconstants->kYToRgb[0]; |
1773 | 0 | #endif |
1774 | 0 | uint32_t y1 = (uint32_t)(y * 0x0101 * yg) >> 16; |
1775 | 0 | uint8_t b8 = STATIC_CAST(uint8_t, Clamp(((int32_t)(y1) + ygb) >> 6)); |
1776 | 0 | *b = b8; |
1777 | 0 | *g = b8; |
1778 | 0 | *r = b8; |
1779 | 0 | } |
1780 | | |
1781 | | void I444ToARGBRow_C(const uint8_t* src_y, |
1782 | | const uint8_t* src_u, |
1783 | | const uint8_t* src_v, |
1784 | | uint8_t* rgb_buf, |
1785 | | const struct YuvConstants* yuvconstants, |
1786 | 0 | int width) { |
1787 | 0 | int x; |
1788 | 0 | for (x = 0; x < width; ++x) { |
1789 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1790 | 0 | rgb_buf + 2, yuvconstants); |
1791 | 0 | rgb_buf[3] = 255; |
1792 | 0 | src_y += 1; |
1793 | 0 | src_u += 1; |
1794 | 0 | src_v += 1; |
1795 | 0 | rgb_buf += 4; // Advance 1 pixel. |
1796 | 0 | } |
1797 | 0 | } |
1798 | | |
1799 | | void I444ToRGB24Row_C(const uint8_t* src_y, |
1800 | | const uint8_t* src_u, |
1801 | | const uint8_t* src_v, |
1802 | | uint8_t* rgb_buf, |
1803 | | const struct YuvConstants* yuvconstants, |
1804 | 0 | int width) { |
1805 | 0 | int x; |
1806 | 0 | for (x = 0; x < width; ++x) { |
1807 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1808 | 0 | rgb_buf + 2, yuvconstants); |
1809 | 0 | src_y += 1; |
1810 | 0 | src_u += 1; |
1811 | 0 | src_v += 1; |
1812 | 0 | rgb_buf += 3; // Advance 1 pixel. |
1813 | 0 | } |
1814 | 0 | } |
1815 | | |
1816 | | // Also used for 420 |
1817 | | void I422ToARGBRow_C(const uint8_t* src_y, |
1818 | | const uint8_t* src_u, |
1819 | | const uint8_t* src_v, |
1820 | | uint8_t* rgb_buf, |
1821 | | const struct YuvConstants* yuvconstants, |
1822 | 0 | int width) { |
1823 | 0 | int x; |
1824 | 0 | for (x = 0; x < width - 1; x += 2) { |
1825 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1826 | 0 | rgb_buf + 2, yuvconstants); |
1827 | 0 | rgb_buf[3] = 255; |
1828 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5, |
1829 | 0 | rgb_buf + 6, yuvconstants); |
1830 | 0 | rgb_buf[7] = 255; |
1831 | 0 | src_y += 2; |
1832 | 0 | src_u += 1; |
1833 | 0 | src_v += 1; |
1834 | 0 | rgb_buf += 8; // Advance 2 pixels. |
1835 | 0 | } |
1836 | 0 | if (width & 1) { |
1837 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1838 | 0 | rgb_buf + 2, yuvconstants); |
1839 | 0 | rgb_buf[3] = 255; |
1840 | 0 | } |
1841 | 0 | } |
1842 | | |
1843 | | // 10 bit YUV to ARGB |
1844 | | void I210ToARGBRow_C(const uint16_t* src_y, |
1845 | | const uint16_t* src_u, |
1846 | | const uint16_t* src_v, |
1847 | | uint8_t* rgb_buf, |
1848 | | const struct YuvConstants* yuvconstants, |
1849 | 0 | int width) { |
1850 | 0 | int x; |
1851 | 0 | for (x = 0; x < width - 1; x += 2) { |
1852 | 0 | YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1853 | 0 | rgb_buf + 2, yuvconstants); |
1854 | 0 | rgb_buf[3] = 255; |
1855 | 0 | YuvPixel10(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5, |
1856 | 0 | rgb_buf + 6, yuvconstants); |
1857 | 0 | rgb_buf[7] = 255; |
1858 | 0 | src_y += 2; |
1859 | 0 | src_u += 1; |
1860 | 0 | src_v += 1; |
1861 | 0 | rgb_buf += 8; // Advance 2 pixels. |
1862 | 0 | } |
1863 | 0 | if (width & 1) { |
1864 | 0 | YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1865 | 0 | rgb_buf + 2, yuvconstants); |
1866 | 0 | rgb_buf[3] = 255; |
1867 | 0 | } |
1868 | 0 | } |
1869 | | |
1870 | | void I410ToARGBRow_C(const uint16_t* src_y, |
1871 | | const uint16_t* src_u, |
1872 | | const uint16_t* src_v, |
1873 | | uint8_t* rgb_buf, |
1874 | | const struct YuvConstants* yuvconstants, |
1875 | 0 | int width) { |
1876 | 0 | int x; |
1877 | 0 | for (x = 0; x < width; ++x) { |
1878 | 0 | YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1879 | 0 | rgb_buf + 2, yuvconstants); |
1880 | 0 | rgb_buf[3] = 255; |
1881 | 0 | src_y += 1; |
1882 | 0 | src_u += 1; |
1883 | 0 | src_v += 1; |
1884 | 0 | rgb_buf += 4; // Advance 1 pixels. |
1885 | 0 | } |
1886 | 0 | } |
1887 | | |
1888 | | void I210AlphaToARGBRow_C(const uint16_t* src_y, |
1889 | | const uint16_t* src_u, |
1890 | | const uint16_t* src_v, |
1891 | | const uint16_t* src_a, |
1892 | | uint8_t* rgb_buf, |
1893 | | const struct YuvConstants* yuvconstants, |
1894 | 0 | int width) { |
1895 | 0 | int x; |
1896 | 0 | for (x = 0; x < width - 1; x += 2) { |
1897 | 0 | YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1898 | 0 | rgb_buf + 2, yuvconstants); |
1899 | 0 | rgb_buf[3] = STATIC_CAST(uint8_t, clamp255(src_a[0] >> 2)); |
1900 | 0 | YuvPixel10(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5, |
1901 | 0 | rgb_buf + 6, yuvconstants); |
1902 | 0 | rgb_buf[7] = STATIC_CAST(uint8_t, clamp255(src_a[1] >> 2)); |
1903 | 0 | src_y += 2; |
1904 | 0 | src_u += 1; |
1905 | 0 | src_v += 1; |
1906 | 0 | src_a += 2; |
1907 | 0 | rgb_buf += 8; // Advance 2 pixels. |
1908 | 0 | } |
1909 | 0 | if (width & 1) { |
1910 | 0 | YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1911 | 0 | rgb_buf + 2, yuvconstants); |
1912 | 0 | rgb_buf[3] = STATIC_CAST(uint8_t, clamp255(src_a[0] >> 2)); |
1913 | 0 | } |
1914 | 0 | } |
1915 | | |
1916 | | void I410AlphaToARGBRow_C(const uint16_t* src_y, |
1917 | | const uint16_t* src_u, |
1918 | | const uint16_t* src_v, |
1919 | | const uint16_t* src_a, |
1920 | | uint8_t* rgb_buf, |
1921 | | const struct YuvConstants* yuvconstants, |
1922 | 0 | int width) { |
1923 | 0 | int x; |
1924 | 0 | for (x = 0; x < width; ++x) { |
1925 | 0 | YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1926 | 0 | rgb_buf + 2, yuvconstants); |
1927 | 0 | rgb_buf[3] = STATIC_CAST(uint8_t, clamp255(src_a[0] >> 2)); |
1928 | 0 | src_y += 1; |
1929 | 0 | src_u += 1; |
1930 | 0 | src_v += 1; |
1931 | 0 | src_a += 1; |
1932 | 0 | rgb_buf += 4; // Advance 1 pixels. |
1933 | 0 | } |
1934 | 0 | } |
1935 | | |
1936 | | // 12 bit YUV to ARGB |
1937 | | void I212ToARGBRow_C(const uint16_t* src_y, |
1938 | | const uint16_t* src_u, |
1939 | | const uint16_t* src_v, |
1940 | | uint8_t* rgb_buf, |
1941 | | const struct YuvConstants* yuvconstants, |
1942 | 0 | int width) { |
1943 | 0 | int x; |
1944 | 0 | for (x = 0; x < width - 1; x += 2) { |
1945 | 0 | YuvPixel12(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1946 | 0 | rgb_buf + 2, yuvconstants); |
1947 | 0 | rgb_buf[3] = 255; |
1948 | 0 | YuvPixel12(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5, |
1949 | 0 | rgb_buf + 6, yuvconstants); |
1950 | 0 | rgb_buf[7] = 255; |
1951 | 0 | src_y += 2; |
1952 | 0 | src_u += 1; |
1953 | 0 | src_v += 1; |
1954 | 0 | rgb_buf += 8; // Advance 2 pixels. |
1955 | 0 | } |
1956 | 0 | if (width & 1) { |
1957 | 0 | YuvPixel12(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
1958 | 0 | rgb_buf + 2, yuvconstants); |
1959 | 0 | rgb_buf[3] = 255; |
1960 | 0 | } |
1961 | 0 | } |
1962 | | |
1963 | 0 | static void StoreAR30(uint8_t* rgb_buf, int b, int g, int r) { |
1964 | 0 | uint32_t ar30; |
1965 | 0 | b = b >> 4; // convert 8 bit 10.6 to 10 bit. |
1966 | 0 | g = g >> 4; |
1967 | 0 | r = r >> 4; |
1968 | 0 | b = Clamp10(b); |
1969 | 0 | g = Clamp10(g); |
1970 | 0 | r = Clamp10(r); |
1971 | 0 | ar30 = b | ((uint32_t)g << 10) | ((uint32_t)r << 20) | 0xc0000000; |
1972 | 0 | (*(uint32_t*)rgb_buf) = ar30; |
1973 | 0 | } |
1974 | | |
1975 | | // 10 bit YUV to 10 bit AR30 |
1976 | | void I210ToAR30Row_C(const uint16_t* src_y, |
1977 | | const uint16_t* src_u, |
1978 | | const uint16_t* src_v, |
1979 | | uint8_t* rgb_buf, |
1980 | | const struct YuvConstants* yuvconstants, |
1981 | 0 | int width) { |
1982 | 0 | int x; |
1983 | 0 | int b; |
1984 | 0 | int g; |
1985 | 0 | int r; |
1986 | 0 | for (x = 0; x < width - 1; x += 2) { |
1987 | 0 | YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
1988 | 0 | StoreAR30(rgb_buf, b, g, r); |
1989 | 0 | YuvPixel10_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
1990 | 0 | StoreAR30(rgb_buf + 4, b, g, r); |
1991 | 0 | src_y += 2; |
1992 | 0 | src_u += 1; |
1993 | 0 | src_v += 1; |
1994 | 0 | rgb_buf += 8; // Advance 2 pixels. |
1995 | 0 | } |
1996 | 0 | if (width & 1) { |
1997 | 0 | YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
1998 | 0 | StoreAR30(rgb_buf, b, g, r); |
1999 | 0 | } |
2000 | 0 | } |
2001 | | |
2002 | | // 12 bit YUV to 10 bit AR30 |
2003 | | void I212ToAR30Row_C(const uint16_t* src_y, |
2004 | | const uint16_t* src_u, |
2005 | | const uint16_t* src_v, |
2006 | | uint8_t* rgb_buf, |
2007 | | const struct YuvConstants* yuvconstants, |
2008 | 0 | int width) { |
2009 | 0 | int x; |
2010 | 0 | int b; |
2011 | 0 | int g; |
2012 | 0 | int r; |
2013 | 0 | for (x = 0; x < width - 1; x += 2) { |
2014 | 0 | YuvPixel12_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2015 | 0 | StoreAR30(rgb_buf, b, g, r); |
2016 | 0 | YuvPixel12_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2017 | 0 | StoreAR30(rgb_buf + 4, b, g, r); |
2018 | 0 | src_y += 2; |
2019 | 0 | src_u += 1; |
2020 | 0 | src_v += 1; |
2021 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2022 | 0 | } |
2023 | 0 | if (width & 1) { |
2024 | 0 | YuvPixel12_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2025 | 0 | StoreAR30(rgb_buf, b, g, r); |
2026 | 0 | } |
2027 | 0 | } |
2028 | | |
2029 | | void I410ToAR30Row_C(const uint16_t* src_y, |
2030 | | const uint16_t* src_u, |
2031 | | const uint16_t* src_v, |
2032 | | uint8_t* rgb_buf, |
2033 | | const struct YuvConstants* yuvconstants, |
2034 | 0 | int width) { |
2035 | 0 | int x; |
2036 | 0 | int b; |
2037 | 0 | int g; |
2038 | 0 | int r; |
2039 | 0 | for (x = 0; x < width; ++x) { |
2040 | 0 | YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2041 | 0 | StoreAR30(rgb_buf, b, g, r); |
2042 | 0 | src_y += 1; |
2043 | 0 | src_u += 1; |
2044 | 0 | src_v += 1; |
2045 | 0 | rgb_buf += 4; // Advance 1 pixel. |
2046 | 0 | } |
2047 | 0 | } |
2048 | | |
2049 | | // P210 has 10 bits in msb of 16 bit NV12 style layout. |
2050 | | void P210ToARGBRow_C(const uint16_t* src_y, |
2051 | | const uint16_t* src_uv, |
2052 | | uint8_t* dst_argb, |
2053 | | const struct YuvConstants* yuvconstants, |
2054 | 0 | int width) { |
2055 | 0 | int x; |
2056 | 0 | for (x = 0; x < width - 1; x += 2) { |
2057 | 0 | YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1, |
2058 | 0 | dst_argb + 2, yuvconstants); |
2059 | 0 | dst_argb[3] = 255; |
2060 | 0 | YuvPixel16_8(src_y[1], src_uv[0], src_uv[1], dst_argb + 4, dst_argb + 5, |
2061 | 0 | dst_argb + 6, yuvconstants); |
2062 | 0 | dst_argb[7] = 255; |
2063 | 0 | src_y += 2; |
2064 | 0 | src_uv += 2; |
2065 | 0 | dst_argb += 8; // Advance 2 pixels. |
2066 | 0 | } |
2067 | 0 | if (width & 1) { |
2068 | 0 | YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1, |
2069 | 0 | dst_argb + 2, yuvconstants); |
2070 | 0 | dst_argb[3] = 255; |
2071 | 0 | } |
2072 | 0 | } |
2073 | | |
2074 | | void P410ToARGBRow_C(const uint16_t* src_y, |
2075 | | const uint16_t* src_uv, |
2076 | | uint8_t* dst_argb, |
2077 | | const struct YuvConstants* yuvconstants, |
2078 | 0 | int width) { |
2079 | 0 | int x; |
2080 | 0 | for (x = 0; x < width; ++x) { |
2081 | 0 | YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1, |
2082 | 0 | dst_argb + 2, yuvconstants); |
2083 | 0 | dst_argb[3] = 255; |
2084 | 0 | src_y += 1; |
2085 | 0 | src_uv += 2; |
2086 | 0 | dst_argb += 4; // Advance 1 pixels. |
2087 | 0 | } |
2088 | 0 | } |
2089 | | |
2090 | | void P210ToAR30Row_C(const uint16_t* src_y, |
2091 | | const uint16_t* src_uv, |
2092 | | uint8_t* dst_ar30, |
2093 | | const struct YuvConstants* yuvconstants, |
2094 | 0 | int width) { |
2095 | 0 | int x; |
2096 | 0 | int b; |
2097 | 0 | int g; |
2098 | 0 | int r; |
2099 | 0 | for (x = 0; x < width - 1; x += 2) { |
2100 | 0 | YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants); |
2101 | 0 | StoreAR30(dst_ar30, b, g, r); |
2102 | 0 | YuvPixel16_16(src_y[1], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants); |
2103 | 0 | StoreAR30(dst_ar30 + 4, b, g, r); |
2104 | 0 | src_y += 2; |
2105 | 0 | src_uv += 2; |
2106 | 0 | dst_ar30 += 8; // Advance 2 pixels. |
2107 | 0 | } |
2108 | 0 | if (width & 1) { |
2109 | 0 | YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants); |
2110 | 0 | StoreAR30(dst_ar30, b, g, r); |
2111 | 0 | } |
2112 | 0 | } |
2113 | | |
2114 | | void P410ToAR30Row_C(const uint16_t* src_y, |
2115 | | const uint16_t* src_uv, |
2116 | | uint8_t* dst_ar30, |
2117 | | const struct YuvConstants* yuvconstants, |
2118 | 0 | int width) { |
2119 | 0 | int x; |
2120 | 0 | int b; |
2121 | 0 | int g; |
2122 | 0 | int r; |
2123 | 0 | for (x = 0; x < width; ++x) { |
2124 | 0 | YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants); |
2125 | 0 | StoreAR30(dst_ar30, b, g, r); |
2126 | 0 | src_y += 1; |
2127 | 0 | src_uv += 2; |
2128 | 0 | dst_ar30 += 4; // Advance 1 pixel. |
2129 | 0 | } |
2130 | 0 | } |
2131 | | |
2132 | | // 8 bit YUV to 10 bit AR30 |
2133 | | // Uses same code as 10 bit YUV bit shifts the 8 bit values up to 10 bits. |
2134 | | void I422ToAR30Row_C(const uint8_t* src_y, |
2135 | | const uint8_t* src_u, |
2136 | | const uint8_t* src_v, |
2137 | | uint8_t* rgb_buf, |
2138 | | const struct YuvConstants* yuvconstants, |
2139 | 0 | int width) { |
2140 | 0 | int x; |
2141 | 0 | int b; |
2142 | 0 | int g; |
2143 | 0 | int r; |
2144 | 0 | for (x = 0; x < width - 1; x += 2) { |
2145 | 0 | YuvPixel8_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2146 | 0 | StoreAR30(rgb_buf, b, g, r); |
2147 | 0 | YuvPixel8_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2148 | 0 | StoreAR30(rgb_buf + 4, b, g, r); |
2149 | 0 | src_y += 2; |
2150 | 0 | src_u += 1; |
2151 | 0 | src_v += 1; |
2152 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2153 | 0 | } |
2154 | 0 | if (width & 1) { |
2155 | 0 | YuvPixel8_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants); |
2156 | 0 | StoreAR30(rgb_buf, b, g, r); |
2157 | 0 | } |
2158 | 0 | } |
2159 | | |
2160 | | void I444AlphaToARGBRow_C(const uint8_t* src_y, |
2161 | | const uint8_t* src_u, |
2162 | | const uint8_t* src_v, |
2163 | | const uint8_t* src_a, |
2164 | | uint8_t* rgb_buf, |
2165 | | const struct YuvConstants* yuvconstants, |
2166 | 0 | int width) { |
2167 | 0 | int x; |
2168 | 0 | for (x = 0; x < width; ++x) { |
2169 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
2170 | 0 | rgb_buf + 2, yuvconstants); |
2171 | 0 | rgb_buf[3] = src_a[0]; |
2172 | 0 | src_y += 1; |
2173 | 0 | src_u += 1; |
2174 | 0 | src_v += 1; |
2175 | 0 | src_a += 1; |
2176 | 0 | rgb_buf += 4; // Advance 1 pixel. |
2177 | 0 | } |
2178 | 0 | } |
2179 | | |
2180 | | void I422AlphaToARGBRow_C(const uint8_t* src_y, |
2181 | | const uint8_t* src_u, |
2182 | | const uint8_t* src_v, |
2183 | | const uint8_t* src_a, |
2184 | | uint8_t* rgb_buf, |
2185 | | const struct YuvConstants* yuvconstants, |
2186 | 0 | int width) { |
2187 | 0 | int x; |
2188 | 0 | for (x = 0; x < width - 1; x += 2) { |
2189 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
2190 | 0 | rgb_buf + 2, yuvconstants); |
2191 | 0 | rgb_buf[3] = src_a[0]; |
2192 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5, |
2193 | 0 | rgb_buf + 6, yuvconstants); |
2194 | 0 | rgb_buf[7] = src_a[1]; |
2195 | 0 | src_y += 2; |
2196 | 0 | src_u += 1; |
2197 | 0 | src_v += 1; |
2198 | 0 | src_a += 2; |
2199 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2200 | 0 | } |
2201 | 0 | if (width & 1) { |
2202 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
2203 | 0 | rgb_buf + 2, yuvconstants); |
2204 | 0 | rgb_buf[3] = src_a[0]; |
2205 | 0 | } |
2206 | 0 | } |
2207 | | |
2208 | | void I422ToRGB24Row_C(const uint8_t* src_y, |
2209 | | const uint8_t* src_u, |
2210 | | const uint8_t* src_v, |
2211 | | uint8_t* rgb_buf, |
2212 | | const struct YuvConstants* yuvconstants, |
2213 | 0 | int width) { |
2214 | 0 | int x; |
2215 | 0 | for (x = 0; x < width - 1; x += 2) { |
2216 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
2217 | 0 | rgb_buf + 2, yuvconstants); |
2218 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 3, rgb_buf + 4, |
2219 | 0 | rgb_buf + 5, yuvconstants); |
2220 | 0 | src_y += 2; |
2221 | 0 | src_u += 1; |
2222 | 0 | src_v += 1; |
2223 | 0 | rgb_buf += 6; // Advance 2 pixels. |
2224 | 0 | } |
2225 | 0 | if (width & 1) { |
2226 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1, |
2227 | 0 | rgb_buf + 2, yuvconstants); |
2228 | 0 | } |
2229 | 0 | } |
2230 | | |
2231 | | void I422ToARGB4444Row_C(const uint8_t* src_y, |
2232 | | const uint8_t* src_u, |
2233 | | const uint8_t* src_v, |
2234 | | uint8_t* dst_argb4444, |
2235 | | const struct YuvConstants* yuvconstants, |
2236 | 0 | int width) { |
2237 | 0 | uint8_t b0; |
2238 | 0 | uint8_t g0; |
2239 | 0 | uint8_t r0; |
2240 | 0 | uint8_t b1; |
2241 | 0 | uint8_t g1; |
2242 | 0 | uint8_t r1; |
2243 | 0 | int x; |
2244 | 0 | for (x = 0; x < width - 1; x += 2) { |
2245 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants); |
2246 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1, yuvconstants); |
2247 | 0 | b0 = b0 >> 4; |
2248 | 0 | g0 = g0 >> 4; |
2249 | 0 | r0 = r0 >> 4; |
2250 | 0 | b1 = b1 >> 4; |
2251 | 0 | g1 = g1 >> 4; |
2252 | 0 | r1 = r1 >> 4; |
2253 | 0 | *(uint16_t*)(dst_argb4444 + 0) = |
2254 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | 0xf000); |
2255 | 0 | *(uint16_t*)(dst_argb4444 + 2) = |
2256 | 0 | STATIC_CAST(uint16_t, b1 | (g1 << 4) | (r1 << 8) | 0xf000); |
2257 | 0 | src_y += 2; |
2258 | 0 | src_u += 1; |
2259 | 0 | src_v += 1; |
2260 | 0 | dst_argb4444 += 4; // Advance 2 pixels. |
2261 | 0 | } |
2262 | 0 | if (width & 1) { |
2263 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants); |
2264 | 0 | b0 = b0 >> 4; |
2265 | 0 | g0 = g0 >> 4; |
2266 | 0 | r0 = r0 >> 4; |
2267 | 0 | *(uint16_t*)(dst_argb4444) = |
2268 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 4) | (r0 << 8) | 0xf000); |
2269 | 0 | } |
2270 | 0 | } |
2271 | | |
2272 | | void I422ToARGB1555Row_C(const uint8_t* src_y, |
2273 | | const uint8_t* src_u, |
2274 | | const uint8_t* src_v, |
2275 | | uint8_t* dst_argb1555, |
2276 | | const struct YuvConstants* yuvconstants, |
2277 | 0 | int width) { |
2278 | 0 | uint8_t b0; |
2279 | 0 | uint8_t g0; |
2280 | 0 | uint8_t r0; |
2281 | 0 | uint8_t b1; |
2282 | 0 | uint8_t g1; |
2283 | 0 | uint8_t r1; |
2284 | 0 | int x; |
2285 | 0 | for (x = 0; x < width - 1; x += 2) { |
2286 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants); |
2287 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1, yuvconstants); |
2288 | 0 | b0 = b0 >> 3; |
2289 | 0 | g0 = g0 >> 3; |
2290 | 0 | r0 = r0 >> 3; |
2291 | 0 | b1 = b1 >> 3; |
2292 | 0 | g1 = g1 >> 3; |
2293 | 0 | r1 = r1 >> 3; |
2294 | 0 | *(uint16_t*)(dst_argb1555 + 0) = |
2295 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | 0x8000); |
2296 | 0 | *(uint16_t*)(dst_argb1555 + 2) = |
2297 | 0 | STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 10) | 0x8000); |
2298 | 0 | src_y += 2; |
2299 | 0 | src_u += 1; |
2300 | 0 | src_v += 1; |
2301 | 0 | dst_argb1555 += 4; // Advance 2 pixels. |
2302 | 0 | } |
2303 | 0 | if (width & 1) { |
2304 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants); |
2305 | 0 | b0 = b0 >> 3; |
2306 | 0 | g0 = g0 >> 3; |
2307 | 0 | r0 = r0 >> 3; |
2308 | 0 | *(uint16_t*)(dst_argb1555) = |
2309 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 10) | 0x8000); |
2310 | 0 | } |
2311 | 0 | } |
2312 | | |
2313 | | void I422ToRGB565Row_C(const uint8_t* src_y, |
2314 | | const uint8_t* src_u, |
2315 | | const uint8_t* src_v, |
2316 | | uint8_t* dst_rgb565, |
2317 | | const struct YuvConstants* yuvconstants, |
2318 | 0 | int width) { |
2319 | 0 | uint8_t b0; |
2320 | 0 | uint8_t g0; |
2321 | 0 | uint8_t r0; |
2322 | 0 | uint8_t b1; |
2323 | 0 | uint8_t g1; |
2324 | 0 | uint8_t r1; |
2325 | 0 | int x; |
2326 | 0 | for (x = 0; x < width - 1; x += 2) { |
2327 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants); |
2328 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1, yuvconstants); |
2329 | 0 | b0 = b0 >> 3; |
2330 | 0 | g0 = g0 >> 2; |
2331 | 0 | r0 = r0 >> 3; |
2332 | 0 | b1 = b1 >> 3; |
2333 | 0 | g1 = g1 >> 2; |
2334 | 0 | r1 = r1 >> 3; |
2335 | 0 | *(uint16_t*)(dst_rgb565 + 0) = |
2336 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11)); |
2337 | 0 | *(uint16_t*)(dst_rgb565 + 2) = |
2338 | 0 | STATIC_CAST(uint16_t, b1 | (g1 << 5) | (r1 << 11)); |
2339 | 0 | src_y += 2; |
2340 | 0 | src_u += 1; |
2341 | 0 | src_v += 1; |
2342 | 0 | dst_rgb565 += 4; // Advance 2 pixels. |
2343 | 0 | } |
2344 | 0 | if (width & 1) { |
2345 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0, yuvconstants); |
2346 | 0 | b0 = b0 >> 3; |
2347 | 0 | g0 = g0 >> 2; |
2348 | 0 | r0 = r0 >> 3; |
2349 | 0 | *(uint16_t*)(dst_rgb565 + 0) = |
2350 | 0 | STATIC_CAST(uint16_t, b0 | (g0 << 5) | (r0 << 11)); |
2351 | 0 | } |
2352 | 0 | } |
2353 | | |
2354 | | void NV12ToARGBRow_C(const uint8_t* src_y, |
2355 | | const uint8_t* src_uv, |
2356 | | uint8_t* rgb_buf, |
2357 | | const struct YuvConstants* yuvconstants, |
2358 | 0 | int width) { |
2359 | 0 | int x; |
2360 | 0 | for (x = 0; x < width - 1; x += 2) { |
2361 | 0 | YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1, |
2362 | 0 | rgb_buf + 2, yuvconstants); |
2363 | 0 | rgb_buf[3] = 255; |
2364 | 0 | YuvPixel(src_y[1], src_uv[0], src_uv[1], rgb_buf + 4, rgb_buf + 5, |
2365 | 0 | rgb_buf + 6, yuvconstants); |
2366 | 0 | rgb_buf[7] = 255; |
2367 | 0 | src_y += 2; |
2368 | 0 | src_uv += 2; |
2369 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2370 | 0 | } |
2371 | 0 | if (width & 1) { |
2372 | 0 | YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1, |
2373 | 0 | rgb_buf + 2, yuvconstants); |
2374 | 0 | rgb_buf[3] = 255; |
2375 | 0 | } |
2376 | 0 | } |
2377 | | |
2378 | | void NV21ToARGBRow_C(const uint8_t* src_y, |
2379 | | const uint8_t* src_vu, |
2380 | | uint8_t* rgb_buf, |
2381 | | const struct YuvConstants* yuvconstants, |
2382 | 0 | int width) { |
2383 | 0 | int x; |
2384 | 0 | for (x = 0; x < width - 1; x += 2) { |
2385 | 0 | YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1, |
2386 | 0 | rgb_buf + 2, yuvconstants); |
2387 | 0 | rgb_buf[3] = 255; |
2388 | 0 | YuvPixel(src_y[1], src_vu[1], src_vu[0], rgb_buf + 4, rgb_buf + 5, |
2389 | 0 | rgb_buf + 6, yuvconstants); |
2390 | 0 | rgb_buf[7] = 255; |
2391 | 0 | src_y += 2; |
2392 | 0 | src_vu += 2; |
2393 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2394 | 0 | } |
2395 | 0 | if (width & 1) { |
2396 | 0 | YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1, |
2397 | 0 | rgb_buf + 2, yuvconstants); |
2398 | 0 | rgb_buf[3] = 255; |
2399 | 0 | } |
2400 | 0 | } |
2401 | | |
2402 | | void NV12ToRGB24Row_C(const uint8_t* src_y, |
2403 | | const uint8_t* src_uv, |
2404 | | uint8_t* rgb_buf, |
2405 | | const struct YuvConstants* yuvconstants, |
2406 | 0 | int width) { |
2407 | 0 | int x; |
2408 | 0 | for (x = 0; x < width - 1; x += 2) { |
2409 | 0 | YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1, |
2410 | 0 | rgb_buf + 2, yuvconstants); |
2411 | 0 | YuvPixel(src_y[1], src_uv[0], src_uv[1], rgb_buf + 3, rgb_buf + 4, |
2412 | 0 | rgb_buf + 5, yuvconstants); |
2413 | 0 | src_y += 2; |
2414 | 0 | src_uv += 2; |
2415 | 0 | rgb_buf += 6; // Advance 2 pixels. |
2416 | 0 | } |
2417 | 0 | if (width & 1) { |
2418 | 0 | YuvPixel(src_y[0], src_uv[0], src_uv[1], rgb_buf + 0, rgb_buf + 1, |
2419 | 0 | rgb_buf + 2, yuvconstants); |
2420 | 0 | } |
2421 | 0 | } |
2422 | | |
2423 | | void NV21ToRGB24Row_C(const uint8_t* src_y, |
2424 | | const uint8_t* src_vu, |
2425 | | uint8_t* rgb_buf, |
2426 | | const struct YuvConstants* yuvconstants, |
2427 | 0 | int width) { |
2428 | 0 | int x; |
2429 | 0 | for (x = 0; x < width - 1; x += 2) { |
2430 | 0 | YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1, |
2431 | 0 | rgb_buf + 2, yuvconstants); |
2432 | 0 | YuvPixel(src_y[1], src_vu[1], src_vu[0], rgb_buf + 3, rgb_buf + 4, |
2433 | 0 | rgb_buf + 5, yuvconstants); |
2434 | 0 | src_y += 2; |
2435 | 0 | src_vu += 2; |
2436 | 0 | rgb_buf += 6; // Advance 2 pixels. |
2437 | 0 | } |
2438 | 0 | if (width & 1) { |
2439 | 0 | YuvPixel(src_y[0], src_vu[1], src_vu[0], rgb_buf + 0, rgb_buf + 1, |
2440 | 0 | rgb_buf + 2, yuvconstants); |
2441 | 0 | } |
2442 | 0 | } |
2443 | | |
2444 | | void NV12ToRGB565Row_C(const uint8_t* src_y, |
2445 | | const uint8_t* src_uv, |
2446 | | uint8_t* dst_rgb565, |
2447 | | const struct YuvConstants* yuvconstants, |
2448 | 0 | int width) { |
2449 | 0 | uint8_t b0; |
2450 | 0 | uint8_t g0; |
2451 | 0 | uint8_t r0; |
2452 | 0 | uint8_t b1; |
2453 | 0 | uint8_t g1; |
2454 | 0 | uint8_t r1; |
2455 | 0 | int x; |
2456 | 0 | for (x = 0; x < width - 1; x += 2) { |
2457 | 0 | YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0, yuvconstants); |
2458 | 0 | YuvPixel(src_y[1], src_uv[0], src_uv[1], &b1, &g1, &r1, yuvconstants); |
2459 | 0 | b0 = b0 >> 3; |
2460 | 0 | g0 = g0 >> 2; |
2461 | 0 | r0 = r0 >> 3; |
2462 | 0 | b1 = b1 >> 3; |
2463 | 0 | g1 = g1 >> 2; |
2464 | 0 | r1 = r1 >> 3; |
2465 | 0 | *(uint16_t*)(dst_rgb565 + 0) = STATIC_CAST(uint16_t, b0) | |
2466 | 0 | STATIC_CAST(uint16_t, g0 << 5) | |
2467 | 0 | STATIC_CAST(uint16_t, r0 << 11); |
2468 | 0 | *(uint16_t*)(dst_rgb565 + 2) = STATIC_CAST(uint16_t, b1) | |
2469 | 0 | STATIC_CAST(uint16_t, g1 << 5) | |
2470 | 0 | STATIC_CAST(uint16_t, r1 << 11); |
2471 | 0 | src_y += 2; |
2472 | 0 | src_uv += 2; |
2473 | 0 | dst_rgb565 += 4; // Advance 2 pixels. |
2474 | 0 | } |
2475 | 0 | if (width & 1) { |
2476 | 0 | YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0, yuvconstants); |
2477 | 0 | b0 = b0 >> 3; |
2478 | 0 | g0 = g0 >> 2; |
2479 | 0 | r0 = r0 >> 3; |
2480 | 0 | *(uint16_t*)(dst_rgb565) = STATIC_CAST(uint16_t, b0) | |
2481 | 0 | STATIC_CAST(uint16_t, g0 << 5) | |
2482 | 0 | STATIC_CAST(uint16_t, r0 << 11); |
2483 | 0 | } |
2484 | 0 | } |
2485 | | |
2486 | | void YUY2ToARGBRow_C(const uint8_t* src_yuy2, |
2487 | | uint8_t* rgb_buf, |
2488 | | const struct YuvConstants* yuvconstants, |
2489 | 0 | int width) { |
2490 | 0 | int x; |
2491 | 0 | for (x = 0; x < width - 1; x += 2) { |
2492 | 0 | YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3], rgb_buf + 0, rgb_buf + 1, |
2493 | 0 | rgb_buf + 2, yuvconstants); |
2494 | 0 | rgb_buf[3] = 255; |
2495 | 0 | YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3], rgb_buf + 4, rgb_buf + 5, |
2496 | 0 | rgb_buf + 6, yuvconstants); |
2497 | 0 | rgb_buf[7] = 255; |
2498 | 0 | src_yuy2 += 4; |
2499 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2500 | 0 | } |
2501 | 0 | if (width & 1) { |
2502 | 0 | YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3], rgb_buf + 0, rgb_buf + 1, |
2503 | 0 | rgb_buf + 2, yuvconstants); |
2504 | 0 | rgb_buf[3] = 255; |
2505 | 0 | } |
2506 | 0 | } |
2507 | | |
2508 | | void UYVYToARGBRow_C(const uint8_t* src_uyvy, |
2509 | | uint8_t* rgb_buf, |
2510 | | const struct YuvConstants* yuvconstants, |
2511 | 0 | int width) { |
2512 | 0 | int x; |
2513 | 0 | for (x = 0; x < width - 1; x += 2) { |
2514 | 0 | YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2], rgb_buf + 0, rgb_buf + 1, |
2515 | 0 | rgb_buf + 2, yuvconstants); |
2516 | 0 | rgb_buf[3] = 255; |
2517 | 0 | YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2], rgb_buf + 4, rgb_buf + 5, |
2518 | 0 | rgb_buf + 6, yuvconstants); |
2519 | 0 | rgb_buf[7] = 255; |
2520 | 0 | src_uyvy += 4; |
2521 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2522 | 0 | } |
2523 | 0 | if (width & 1) { |
2524 | 0 | YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2], rgb_buf + 0, rgb_buf + 1, |
2525 | 0 | rgb_buf + 2, yuvconstants); |
2526 | 0 | rgb_buf[3] = 255; |
2527 | 0 | } |
2528 | 0 | } |
2529 | | |
2530 | | void I422ToRGBARow_C(const uint8_t* src_y, |
2531 | | const uint8_t* src_u, |
2532 | | const uint8_t* src_v, |
2533 | | uint8_t* rgb_buf, |
2534 | | const struct YuvConstants* yuvconstants, |
2535 | 0 | int width) { |
2536 | 0 | int x; |
2537 | 0 | for (x = 0; x < width - 1; x += 2) { |
2538 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 1, rgb_buf + 2, |
2539 | 0 | rgb_buf + 3, yuvconstants); |
2540 | 0 | rgb_buf[0] = 255; |
2541 | 0 | YuvPixel(src_y[1], src_u[0], src_v[0], rgb_buf + 5, rgb_buf + 6, |
2542 | 0 | rgb_buf + 7, yuvconstants); |
2543 | 0 | rgb_buf[4] = 255; |
2544 | 0 | src_y += 2; |
2545 | 0 | src_u += 1; |
2546 | 0 | src_v += 1; |
2547 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2548 | 0 | } |
2549 | 0 | if (width & 1) { |
2550 | 0 | YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 1, rgb_buf + 2, |
2551 | 0 | rgb_buf + 3, yuvconstants); |
2552 | 0 | rgb_buf[0] = 255; |
2553 | 0 | } |
2554 | 0 | } |
2555 | | |
2556 | | void I400ToARGBRow_C(const uint8_t* src_y, |
2557 | | uint8_t* rgb_buf, |
2558 | | const struct YuvConstants* yuvconstants, |
2559 | 0 | int width) { |
2560 | 0 | int x; |
2561 | 0 | for (x = 0; x < width - 1; x += 2) { |
2562 | 0 | YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants); |
2563 | 0 | rgb_buf[3] = 255; |
2564 | 0 | YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants); |
2565 | 0 | rgb_buf[7] = 255; |
2566 | 0 | src_y += 2; |
2567 | 0 | rgb_buf += 8; // Advance 2 pixels. |
2568 | 0 | } |
2569 | 0 | if (width & 1) { |
2570 | 0 | YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants); |
2571 | 0 | rgb_buf[3] = 255; |
2572 | 0 | } |
2573 | 0 | } |
2574 | | |
2575 | 0 | void MirrorRow_C(const uint8_t* src, uint8_t* dst, int width) { |
2576 | 0 | int x; |
2577 | 0 | src += width - 1; |
2578 | 0 | for (x = 0; x < width - 1; x += 2) { |
2579 | 0 | dst[x] = src[0]; |
2580 | 0 | dst[x + 1] = src[-1]; |
2581 | 0 | src -= 2; |
2582 | 0 | } |
2583 | 0 | if (width & 1) { |
2584 | 0 | dst[width - 1] = src[0]; |
2585 | 0 | } |
2586 | 0 | } |
2587 | | |
2588 | 0 | void MirrorRow_16_C(const uint16_t* src, uint16_t* dst, int width) { |
2589 | 0 | int x; |
2590 | 0 | src += width - 1; |
2591 | 0 | for (x = 0; x < width - 1; x += 2) { |
2592 | 0 | dst[x] = src[0]; |
2593 | 0 | dst[x + 1] = src[-1]; |
2594 | 0 | src -= 2; |
2595 | 0 | } |
2596 | 0 | if (width & 1) { |
2597 | 0 | dst[width - 1] = src[0]; |
2598 | 0 | } |
2599 | 0 | } |
2600 | | |
2601 | 0 | void MirrorUVRow_C(const uint8_t* src_uv, uint8_t* dst_uv, int width) { |
2602 | 0 | int x; |
2603 | 0 | src_uv += (width - 1) << 1; |
2604 | 0 | for (x = 0; x < width; ++x) { |
2605 | 0 | dst_uv[0] = src_uv[0]; |
2606 | 0 | dst_uv[1] = src_uv[1]; |
2607 | 0 | src_uv -= 2; |
2608 | 0 | dst_uv += 2; |
2609 | 0 | } |
2610 | 0 | } |
2611 | | |
2612 | | void MirrorSplitUVRow_C(const uint8_t* src_uv, |
2613 | | uint8_t* dst_u, |
2614 | | uint8_t* dst_v, |
2615 | 0 | int width) { |
2616 | 0 | int x; |
2617 | 0 | src_uv += (width - 1) << 1; |
2618 | 0 | for (x = 0; x < width - 1; x += 2) { |
2619 | 0 | dst_u[x] = src_uv[0]; |
2620 | 0 | dst_u[x + 1] = src_uv[-2]; |
2621 | 0 | dst_v[x] = src_uv[1]; |
2622 | 0 | dst_v[x + 1] = src_uv[-2 + 1]; |
2623 | 0 | src_uv -= 4; |
2624 | 0 | } |
2625 | 0 | if (width & 1) { |
2626 | 0 | dst_u[width - 1] = src_uv[0]; |
2627 | 0 | dst_v[width - 1] = src_uv[1]; |
2628 | 0 | } |
2629 | 0 | } |
2630 | | |
2631 | 0 | void ARGBMirrorRow_C(const uint8_t* src, uint8_t* dst, int width) { |
2632 | 0 | int x; |
2633 | 0 | const uint32_t* src32 = (const uint32_t*)(src); |
2634 | 0 | uint32_t* dst32 = (uint32_t*)(dst); |
2635 | 0 | src32 += width - 1; |
2636 | 0 | for (x = 0; x < width - 1; x += 2) { |
2637 | 0 | dst32[x] = src32[0]; |
2638 | 0 | dst32[x + 1] = src32[-1]; |
2639 | 0 | src32 -= 2; |
2640 | 0 | } |
2641 | 0 | if (width & 1) { |
2642 | 0 | dst32[width - 1] = src32[0]; |
2643 | 0 | } |
2644 | 0 | } |
2645 | | |
2646 | 0 | void RGB24MirrorRow_C(const uint8_t* src_rgb24, uint8_t* dst_rgb24, int width) { |
2647 | 0 | int x; |
2648 | 0 | src_rgb24 += width * 3 - 3; |
2649 | 0 | for (x = 0; x < width; ++x) { |
2650 | 0 | uint8_t b = src_rgb24[0]; |
2651 | 0 | uint8_t g = src_rgb24[1]; |
2652 | 0 | uint8_t r = src_rgb24[2]; |
2653 | 0 | dst_rgb24[0] = b; |
2654 | 0 | dst_rgb24[1] = g; |
2655 | 0 | dst_rgb24[2] = r; |
2656 | 0 | src_rgb24 -= 3; |
2657 | 0 | dst_rgb24 += 3; |
2658 | 0 | } |
2659 | 0 | } |
2660 | | |
2661 | | void SplitUVRow_C(const uint8_t* src_uv, |
2662 | | uint8_t* dst_u, |
2663 | | uint8_t* dst_v, |
2664 | 0 | int width) { |
2665 | 0 | int x; |
2666 | 0 | for (x = 0; x < width - 1; x += 2) { |
2667 | 0 | dst_u[x] = src_uv[0]; |
2668 | 0 | dst_u[x + 1] = src_uv[2]; |
2669 | 0 | dst_v[x] = src_uv[1]; |
2670 | 0 | dst_v[x + 1] = src_uv[3]; |
2671 | 0 | src_uv += 4; |
2672 | 0 | } |
2673 | 0 | if (width & 1) { |
2674 | 0 | dst_u[width - 1] = src_uv[0]; |
2675 | 0 | dst_v[width - 1] = src_uv[1]; |
2676 | 0 | } |
2677 | 0 | } |
2678 | | |
2679 | | void MergeUVRow_C(const uint8_t* src_u, |
2680 | | const uint8_t* src_v, |
2681 | | uint8_t* dst_uv, |
2682 | 0 | int width) { |
2683 | 0 | int x; |
2684 | 0 | for (x = 0; x < width - 1; x += 2) { |
2685 | 0 | dst_uv[0] = src_u[x]; |
2686 | 0 | dst_uv[1] = src_v[x]; |
2687 | 0 | dst_uv[2] = src_u[x + 1]; |
2688 | 0 | dst_uv[3] = src_v[x + 1]; |
2689 | 0 | dst_uv += 4; |
2690 | 0 | } |
2691 | 0 | if (width & 1) { |
2692 | 0 | dst_uv[0] = src_u[width - 1]; |
2693 | 0 | dst_uv[1] = src_v[width - 1]; |
2694 | 0 | } |
2695 | 0 | } |
2696 | | |
2697 | | void DetileRow_C(const uint8_t* src, |
2698 | | ptrdiff_t src_tile_stride, |
2699 | | uint8_t* dst, |
2700 | 0 | int width) { |
2701 | 0 | int x; |
2702 | 0 | for (x = 0; x < width - 15; x += 16) { |
2703 | 0 | memcpy(dst, src, 16); |
2704 | 0 | dst += 16; |
2705 | 0 | src += src_tile_stride; |
2706 | 0 | } |
2707 | 0 | if (width & 15) { |
2708 | 0 | memcpy(dst, src, width & 15); |
2709 | 0 | } |
2710 | 0 | } |
2711 | | |
2712 | | void DetileRow_16_C(const uint16_t* src, |
2713 | | ptrdiff_t src_tile_stride, |
2714 | | uint16_t* dst, |
2715 | 0 | int width) { |
2716 | 0 | int x; |
2717 | 0 | for (x = 0; x < width - 15; x += 16) { |
2718 | 0 | memcpy(dst, src, 16 * sizeof(uint16_t)); |
2719 | 0 | dst += 16; |
2720 | 0 | src += src_tile_stride; |
2721 | 0 | } |
2722 | 0 | if (width & 15) { |
2723 | 0 | memcpy(dst, src, (width & 15) * sizeof(uint16_t)); |
2724 | 0 | } |
2725 | 0 | } |
2726 | | |
2727 | | void DetileSplitUVRow_C(const uint8_t* src_uv, |
2728 | | ptrdiff_t src_tile_stride, |
2729 | | uint8_t* dst_u, |
2730 | | uint8_t* dst_v, |
2731 | 0 | int width) { |
2732 | 0 | int x; |
2733 | 0 | for (x = 0; x < width - 15; x += 16) { |
2734 | 0 | SplitUVRow_C(src_uv, dst_u, dst_v, 8); |
2735 | 0 | dst_u += 8; |
2736 | 0 | dst_v += 8; |
2737 | 0 | src_uv += src_tile_stride; |
2738 | 0 | } |
2739 | 0 | if (width & 15) { |
2740 | 0 | SplitUVRow_C(src_uv, dst_u, dst_v, ((width & 15) + 1) / 2); |
2741 | 0 | } |
2742 | 0 | } |
2743 | | |
2744 | | void DetileToYUY2_C(const uint8_t* src_y, |
2745 | | ptrdiff_t src_y_tile_stride, |
2746 | | const uint8_t* src_uv, |
2747 | | ptrdiff_t src_uv_tile_stride, |
2748 | | uint8_t* dst_yuy2, |
2749 | 0 | int width) { |
2750 | 0 | for (int x = 0; x < width - 15; x += 16) { |
2751 | 0 | for (int i = 0; i < 8; i++) { |
2752 | 0 | dst_yuy2[0] = src_y[0]; |
2753 | 0 | dst_yuy2[1] = src_uv[0]; |
2754 | 0 | dst_yuy2[2] = src_y[1]; |
2755 | 0 | dst_yuy2[3] = src_uv[1]; |
2756 | 0 | dst_yuy2 += 4; |
2757 | 0 | src_y += 2; |
2758 | 0 | src_uv += 2; |
2759 | 0 | } |
2760 | 0 | src_y += src_y_tile_stride - 16; |
2761 | 0 | src_uv += src_uv_tile_stride - 16; |
2762 | 0 | } |
2763 | 0 | } |
2764 | | |
2765 | | // Unpack MT2T into tiled P010 64 pixels at a time. MT2T's bitstream is encoded |
2766 | | // in 80 byte blocks representing 64 pixels each. The first 16 bytes of the |
2767 | | // block contain all of the lower 2 bits of each pixel packed together, and the |
2768 | | // next 64 bytes represent all the upper 8 bits of the pixel. The lower bits are |
2769 | | // packed into 1x4 blocks, whereas the upper bits are packed in normal raster |
2770 | | // order. |
2771 | 0 | void UnpackMT2T_C(const uint8_t* src, uint16_t* dst, size_t size) { |
2772 | 0 | for (size_t i = 0; i < size; i += 80) { |
2773 | 0 | const uint8_t* src_lower_bits = src; |
2774 | 0 | const uint8_t* src_upper_bits = src + 16; |
2775 | |
|
2776 | 0 | for (int j = 0; j < 4; j++) { |
2777 | 0 | for (int k = 0; k < 16; k++) { |
2778 | 0 | *dst++ = ((src_lower_bits[k] >> (j * 2)) & 0x3) << 6 | |
2779 | 0 | (uint16_t)*src_upper_bits << 8 | |
2780 | 0 | (uint16_t)*src_upper_bits >> 2; |
2781 | 0 | src_upper_bits++; |
2782 | 0 | } |
2783 | 0 | } |
2784 | |
|
2785 | 0 | src += 80; |
2786 | 0 | } |
2787 | 0 | } |
2788 | | |
2789 | | void SplitRGBRow_C(const uint8_t* src_rgb, |
2790 | | uint8_t* dst_r, |
2791 | | uint8_t* dst_g, |
2792 | | uint8_t* dst_b, |
2793 | 0 | int width) { |
2794 | 0 | int x; |
2795 | 0 | for (x = 0; x < width; ++x) { |
2796 | 0 | dst_r[x] = src_rgb[0]; |
2797 | 0 | dst_g[x] = src_rgb[1]; |
2798 | 0 | dst_b[x] = src_rgb[2]; |
2799 | 0 | src_rgb += 3; |
2800 | 0 | } |
2801 | 0 | } |
2802 | | |
2803 | | void MergeRGBRow_C(const uint8_t* src_r, |
2804 | | const uint8_t* src_g, |
2805 | | const uint8_t* src_b, |
2806 | | uint8_t* dst_rgb, |
2807 | 0 | int width) { |
2808 | 0 | int x; |
2809 | 0 | for (x = 0; x < width; ++x) { |
2810 | 0 | dst_rgb[0] = src_r[x]; |
2811 | 0 | dst_rgb[1] = src_g[x]; |
2812 | 0 | dst_rgb[2] = src_b[x]; |
2813 | 0 | dst_rgb += 3; |
2814 | 0 | } |
2815 | 0 | } |
2816 | | |
2817 | | void SplitARGBRow_C(const uint8_t* src_argb, |
2818 | | uint8_t* dst_r, |
2819 | | uint8_t* dst_g, |
2820 | | uint8_t* dst_b, |
2821 | | uint8_t* dst_a, |
2822 | 0 | int width) { |
2823 | 0 | int x; |
2824 | 0 | for (x = 0; x < width; ++x) { |
2825 | 0 | dst_b[x] = src_argb[0]; |
2826 | 0 | dst_g[x] = src_argb[1]; |
2827 | 0 | dst_r[x] = src_argb[2]; |
2828 | 0 | dst_a[x] = src_argb[3]; |
2829 | 0 | src_argb += 4; |
2830 | 0 | } |
2831 | 0 | } |
2832 | | |
2833 | | void MergeARGBRow_C(const uint8_t* src_r, |
2834 | | const uint8_t* src_g, |
2835 | | const uint8_t* src_b, |
2836 | | const uint8_t* src_a, |
2837 | | uint8_t* dst_argb, |
2838 | 0 | int width) { |
2839 | 0 | int x; |
2840 | 0 | for (x = 0; x < width; ++x) { |
2841 | 0 | dst_argb[0] = src_b[x]; |
2842 | 0 | dst_argb[1] = src_g[x]; |
2843 | 0 | dst_argb[2] = src_r[x]; |
2844 | 0 | dst_argb[3] = src_a[x]; |
2845 | 0 | dst_argb += 4; |
2846 | 0 | } |
2847 | 0 | } |
2848 | | |
2849 | | void MergeXR30Row_C(const uint16_t* src_r, |
2850 | | const uint16_t* src_g, |
2851 | | const uint16_t* src_b, |
2852 | | uint8_t* dst_ar30, |
2853 | | int depth, |
2854 | 0 | int width) { |
2855 | 0 | assert(depth >= 10); |
2856 | 0 | assert(depth <= 16); |
2857 | 0 | int x; |
2858 | 0 | int shift = depth - 10; |
2859 | 0 | uint32_t* dst_ar30_32 = (uint32_t*)dst_ar30; |
2860 | 0 | for (x = 0; x < width; ++x) { |
2861 | 0 | uint32_t r = clamp1023(src_r[x] >> shift); |
2862 | 0 | uint32_t g = clamp1023(src_g[x] >> shift); |
2863 | 0 | uint32_t b = clamp1023(src_b[x] >> shift); |
2864 | 0 | dst_ar30_32[x] = b | (g << 10) | (r << 20) | 0xc0000000; |
2865 | 0 | } |
2866 | 0 | } |
2867 | | |
2868 | | void MergeAR64Row_C(const uint16_t* src_r, |
2869 | | const uint16_t* src_g, |
2870 | | const uint16_t* src_b, |
2871 | | const uint16_t* src_a, |
2872 | | uint16_t* dst_ar64, |
2873 | | int depth, |
2874 | 0 | int width) { |
2875 | 0 | assert(depth >= 1); |
2876 | 0 | assert(depth <= 16); |
2877 | 0 | int x; |
2878 | 0 | int shift = 16 - depth; |
2879 | 0 | int max = (1 << depth) - 1; |
2880 | 0 | for (x = 0; x < width; ++x) { |
2881 | 0 | dst_ar64[0] = STATIC_CAST(uint16_t, ClampMax(src_b[x], max) << shift); |
2882 | 0 | dst_ar64[1] = STATIC_CAST(uint16_t, ClampMax(src_g[x], max) << shift); |
2883 | 0 | dst_ar64[2] = STATIC_CAST(uint16_t, ClampMax(src_r[x], max) << shift); |
2884 | 0 | dst_ar64[3] = STATIC_CAST(uint16_t, ClampMax(src_a[x], max) << shift); |
2885 | 0 | dst_ar64 += 4; |
2886 | 0 | } |
2887 | 0 | } |
2888 | | |
2889 | | void MergeARGB16To8Row_C(const uint16_t* src_r, |
2890 | | const uint16_t* src_g, |
2891 | | const uint16_t* src_b, |
2892 | | const uint16_t* src_a, |
2893 | | uint8_t* dst_argb, |
2894 | | int depth, |
2895 | 0 | int width) { |
2896 | 0 | assert(depth >= 8); |
2897 | 0 | assert(depth <= 16); |
2898 | 0 | int x; |
2899 | 0 | int shift = depth - 8; |
2900 | 0 | for (x = 0; x < width; ++x) { |
2901 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, clamp255(src_b[x] >> shift)); |
2902 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, clamp255(src_g[x] >> shift)); |
2903 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, clamp255(src_r[x] >> shift)); |
2904 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, clamp255(src_a[x] >> shift)); |
2905 | 0 | dst_argb += 4; |
2906 | 0 | } |
2907 | 0 | } |
2908 | | |
2909 | | void MergeXR64Row_C(const uint16_t* src_r, |
2910 | | const uint16_t* src_g, |
2911 | | const uint16_t* src_b, |
2912 | | uint16_t* dst_ar64, |
2913 | | int depth, |
2914 | 0 | int width) { |
2915 | 0 | assert(depth >= 1); |
2916 | 0 | assert(depth <= 16); |
2917 | 0 | int x; |
2918 | 0 | int shift = 16 - depth; |
2919 | 0 | int max = (1 << depth) - 1; |
2920 | 0 | for (x = 0; x < width; ++x) { |
2921 | 0 | dst_ar64[0] = STATIC_CAST(uint16_t, ClampMax(src_b[x], max) << shift); |
2922 | 0 | dst_ar64[1] = STATIC_CAST(uint16_t, ClampMax(src_g[x], max) << shift); |
2923 | 0 | dst_ar64[2] = STATIC_CAST(uint16_t, ClampMax(src_r[x], max) << shift); |
2924 | 0 | dst_ar64[3] = 0xffff; |
2925 | 0 | dst_ar64 += 4; |
2926 | 0 | } |
2927 | 0 | } |
2928 | | |
2929 | | void MergeXRGB16To8Row_C(const uint16_t* src_r, |
2930 | | const uint16_t* src_g, |
2931 | | const uint16_t* src_b, |
2932 | | uint8_t* dst_argb, |
2933 | | int depth, |
2934 | 0 | int width) { |
2935 | 0 | assert(depth >= 8); |
2936 | 0 | assert(depth <= 16); |
2937 | 0 | int x; |
2938 | 0 | int shift = depth - 8; |
2939 | 0 | for (x = 0; x < width; ++x) { |
2940 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, clamp255(src_b[x] >> shift)); |
2941 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, clamp255(src_g[x] >> shift)); |
2942 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, clamp255(src_r[x] >> shift)); |
2943 | 0 | dst_argb[3] = 0xff; |
2944 | 0 | dst_argb += 4; |
2945 | 0 | } |
2946 | 0 | } |
2947 | | |
2948 | | void SplitXRGBRow_C(const uint8_t* src_argb, |
2949 | | uint8_t* dst_r, |
2950 | | uint8_t* dst_g, |
2951 | | uint8_t* dst_b, |
2952 | 0 | int width) { |
2953 | 0 | int x; |
2954 | 0 | for (x = 0; x < width; ++x) { |
2955 | 0 | dst_b[x] = src_argb[0]; |
2956 | 0 | dst_g[x] = src_argb[1]; |
2957 | 0 | dst_r[x] = src_argb[2]; |
2958 | 0 | src_argb += 4; |
2959 | 0 | } |
2960 | 0 | } |
2961 | | |
2962 | | void MergeXRGBRow_C(const uint8_t* src_r, |
2963 | | const uint8_t* src_g, |
2964 | | const uint8_t* src_b, |
2965 | | uint8_t* dst_argb, |
2966 | 0 | int width) { |
2967 | 0 | int x; |
2968 | 0 | for (x = 0; x < width; ++x) { |
2969 | 0 | dst_argb[0] = src_b[x]; |
2970 | 0 | dst_argb[1] = src_g[x]; |
2971 | 0 | dst_argb[2] = src_r[x]; |
2972 | 0 | dst_argb[3] = 255; |
2973 | 0 | dst_argb += 4; |
2974 | 0 | } |
2975 | 0 | } |
2976 | | |
2977 | | // Convert lsb formats to msb, depending on sample depth. |
2978 | | void MergeUVRow_16_C(const uint16_t* src_u, |
2979 | | const uint16_t* src_v, |
2980 | | uint16_t* dst_uv, |
2981 | | int depth, |
2982 | 0 | int width) { |
2983 | 0 | int shift = 16 - depth; |
2984 | 0 | assert(depth >= 8); |
2985 | 0 | assert(depth <= 16); |
2986 | 0 | int x; |
2987 | 0 | for (x = 0; x < width; ++x) { |
2988 | 0 | dst_uv[0] = STATIC_CAST(uint16_t, src_u[x] << shift); |
2989 | 0 | dst_uv[1] = STATIC_CAST(uint16_t, src_v[x] << shift); |
2990 | 0 | dst_uv += 2; |
2991 | 0 | } |
2992 | 0 | } |
2993 | | |
2994 | | // Convert msb formats to lsb, depending on sample depth. |
2995 | | void SplitUVRow_16_C(const uint16_t* src_uv, |
2996 | | uint16_t* dst_u, |
2997 | | uint16_t* dst_v, |
2998 | | int depth, |
2999 | 0 | int width) { |
3000 | 0 | int shift = 16 - depth; |
3001 | 0 | int x; |
3002 | 0 | assert(depth >= 8); |
3003 | 0 | assert(depth <= 16); |
3004 | 0 | for (x = 0; x < width; ++x) { |
3005 | 0 | dst_u[x] = src_uv[0] >> shift; |
3006 | 0 | dst_v[x] = src_uv[1] >> shift; |
3007 | 0 | src_uv += 2; |
3008 | 0 | } |
3009 | 0 | } |
3010 | | |
3011 | | void MultiplyRow_16_C(const uint16_t* src_y, |
3012 | | uint16_t* dst_y, |
3013 | | int scale, |
3014 | 0 | int width) { |
3015 | 0 | int x; |
3016 | 0 | for (x = 0; x < width; ++x) { |
3017 | 0 | dst_y[x] = STATIC_CAST(uint16_t, src_y[x] * scale); |
3018 | 0 | } |
3019 | 0 | } |
3020 | | |
3021 | | void DivideRow_16_C(const uint16_t* src_y, |
3022 | | uint16_t* dst_y, |
3023 | | int scale, |
3024 | 0 | int width) { |
3025 | 0 | int x; |
3026 | 0 | for (x = 0; x < width; ++x) { |
3027 | 0 | dst_y[x] = (src_y[x] * scale) >> 16; |
3028 | 0 | } |
3029 | 0 | } |
3030 | | |
3031 | | // Use scale to convert lsb formats to msb, depending how many bits there are: |
3032 | | // 32768 = 9 bits |
3033 | | // 16384 = 10 bits |
3034 | | // 4096 = 12 bits |
3035 | | // 256 = 16 bits |
3036 | | // TODO(fbarchard): change scale to bits |
3037 | | #define C16TO8(v, scale) clamp255(((v) * (scale)) >> 16) |
3038 | | |
3039 | | void Convert16To8Row_C(const uint16_t* src_y, |
3040 | | uint8_t* dst_y, |
3041 | | int scale, |
3042 | 0 | int width) { |
3043 | 0 | int x; |
3044 | 0 | assert(scale >= 256); |
3045 | 0 | assert(scale <= 32768); |
3046 | |
|
3047 | 0 | for (x = 0; x < width; ++x) { |
3048 | 0 | dst_y[x] = STATIC_CAST(uint8_t, C16TO8(src_y[x], scale)); |
3049 | 0 | } |
3050 | 0 | } |
3051 | | |
3052 | | // Use scale to convert lsb formats to msb, depending how many bits there are: |
3053 | | // 1024 = 10 bits |
3054 | | void Convert8To16Row_C(const uint8_t* src_y, |
3055 | | uint16_t* dst_y, |
3056 | | int scale, |
3057 | 0 | int width) { |
3058 | 0 | int x; |
3059 | 0 | scale *= 0x0101; // replicates the byte. |
3060 | 0 | for (x = 0; x < width; ++x) { |
3061 | 0 | dst_y[x] = (src_y[x] * scale) >> 16; |
3062 | 0 | } |
3063 | 0 | } |
3064 | | |
3065 | | // Use scale to convert J420 to I420 |
3066 | | // scale parameter is 8.8 fixed point but limited to 0 to 255 |
3067 | | // Function is based on DivideRow, but adds a bias |
3068 | | // Does not clamp |
3069 | | void Convert8To8Row_C(const uint8_t* src_y, |
3070 | | uint8_t* dst_y, |
3071 | | int scale, |
3072 | | int bias, |
3073 | 0 | int width) { |
3074 | 0 | int x; |
3075 | 0 | assert(scale >= 0); |
3076 | 0 | assert(scale <= 255); |
3077 | |
|
3078 | 0 | for (x = 0; x < width; ++x) { |
3079 | 0 | dst_y[x] = ((src_y[x] * scale) >> 8) + bias; |
3080 | 0 | } |
3081 | 0 | } |
3082 | | |
3083 | 0 | void CopyRow_C(const uint8_t* src, uint8_t* dst, int count) { |
3084 | 0 | memcpy(dst, src, count); |
3085 | 0 | } |
3086 | | |
3087 | 0 | void CopyRow_16_C(const uint16_t* src, uint16_t* dst, int count) { |
3088 | 0 | memcpy(dst, src, count * 2); |
3089 | 0 | } |
3090 | | |
3091 | 0 | void SetRow_C(uint8_t* dst, uint8_t v8, int width) { |
3092 | 0 | memset(dst, v8, width); |
3093 | 0 | } |
3094 | | |
3095 | 0 | void ARGBSetRow_C(uint8_t* dst_argb, uint32_t v32, int width) { |
3096 | 0 | int x; |
3097 | 0 | for (x = 0; x < width; ++x) { |
3098 | 0 | memcpy(dst_argb + x * sizeof v32, &v32, sizeof v32); |
3099 | 0 | } |
3100 | 0 | } |
3101 | | |
3102 | | // Filter 2 rows of YUY2 UV's (422) into U and V (420). |
3103 | | void YUY2ToUVRow_C(const uint8_t* src_yuy2, |
3104 | | int src_stride_yuy2, |
3105 | | uint8_t* dst_u, |
3106 | | uint8_t* dst_v, |
3107 | 0 | int width) { |
3108 | | // Output a row of UV values, filtering 2 rows of YUY2. |
3109 | 0 | int x; |
3110 | 0 | for (x = 0; x < width; x += 2) { |
3111 | 0 | dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1; |
3112 | 0 | dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1; |
3113 | 0 | src_yuy2 += 4; |
3114 | 0 | dst_u += 1; |
3115 | 0 | dst_v += 1; |
3116 | 0 | } |
3117 | 0 | } |
3118 | | |
3119 | | // Filter 2 rows of YUY2 UV's (422) into UV (NV12). |
3120 | | void YUY2ToNVUVRow_C(const uint8_t* src_yuy2, |
3121 | | int src_stride_yuy2, |
3122 | | uint8_t* dst_uv, |
3123 | 0 | int width) { |
3124 | | // Output a row of UV values, filtering 2 rows of YUY2. |
3125 | 0 | int x; |
3126 | 0 | for (x = 0; x < width; x += 2) { |
3127 | 0 | dst_uv[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1; |
3128 | 0 | dst_uv[1] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1; |
3129 | 0 | src_yuy2 += 4; |
3130 | 0 | dst_uv += 2; |
3131 | 0 | } |
3132 | 0 | } |
3133 | | |
3134 | | // Copy row of YUY2 UV's (422) into U and V (422). |
3135 | | void YUY2ToUV422Row_C(const uint8_t* src_yuy2, |
3136 | | uint8_t* dst_u, |
3137 | | uint8_t* dst_v, |
3138 | 0 | int width) { |
3139 | | // Output a row of UV values. |
3140 | 0 | int x; |
3141 | 0 | for (x = 0; x < width; x += 2) { |
3142 | 0 | dst_u[0] = src_yuy2[1]; |
3143 | 0 | dst_v[0] = src_yuy2[3]; |
3144 | 0 | src_yuy2 += 4; |
3145 | 0 | dst_u += 1; |
3146 | 0 | dst_v += 1; |
3147 | 0 | } |
3148 | 0 | } |
3149 | | |
3150 | | // Copy row of YUY2 Y's (422) into Y (420/422). |
3151 | 0 | void YUY2ToYRow_C(const uint8_t* src_yuy2, uint8_t* dst_y, int width) { |
3152 | | // Output a row of Y values. |
3153 | 0 | int x; |
3154 | 0 | for (x = 0; x < width - 1; x += 2) { |
3155 | 0 | dst_y[x] = src_yuy2[0]; |
3156 | 0 | dst_y[x + 1] = src_yuy2[2]; |
3157 | 0 | src_yuy2 += 4; |
3158 | 0 | } |
3159 | 0 | if (width & 1) { |
3160 | 0 | dst_y[width - 1] = src_yuy2[0]; |
3161 | 0 | } |
3162 | 0 | } |
3163 | | |
3164 | | // Filter 2 rows of UYVY UV's (422) into U and V (420). |
3165 | | void UYVYToUVRow_C(const uint8_t* src_uyvy, |
3166 | | int src_stride_uyvy, |
3167 | | uint8_t* dst_u, |
3168 | | uint8_t* dst_v, |
3169 | 0 | int width) { |
3170 | | // Output a row of UV values. |
3171 | 0 | int x; |
3172 | 0 | for (x = 0; x < width; x += 2) { |
3173 | 0 | dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1; |
3174 | 0 | dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1; |
3175 | 0 | src_uyvy += 4; |
3176 | 0 | dst_u += 1; |
3177 | 0 | dst_v += 1; |
3178 | 0 | } |
3179 | 0 | } |
3180 | | |
3181 | | // Copy row of UYVY UV's (422) into U and V (422). |
3182 | | void UYVYToUV422Row_C(const uint8_t* src_uyvy, |
3183 | | uint8_t* dst_u, |
3184 | | uint8_t* dst_v, |
3185 | 0 | int width) { |
3186 | | // Output a row of UV values. |
3187 | 0 | int x; |
3188 | 0 | for (x = 0; x < width; x += 2) { |
3189 | 0 | dst_u[0] = src_uyvy[0]; |
3190 | 0 | dst_v[0] = src_uyvy[2]; |
3191 | 0 | src_uyvy += 4; |
3192 | 0 | dst_u += 1; |
3193 | 0 | dst_v += 1; |
3194 | 0 | } |
3195 | 0 | } |
3196 | | |
3197 | | // Copy row of UYVY Y's (422) into Y (420/422). |
3198 | 0 | void UYVYToYRow_C(const uint8_t* src_uyvy, uint8_t* dst_y, int width) { |
3199 | | // Output a row of Y values. |
3200 | 0 | int x; |
3201 | 0 | for (x = 0; x < width - 1; x += 2) { |
3202 | 0 | dst_y[x] = src_uyvy[1]; |
3203 | 0 | dst_y[x + 1] = src_uyvy[3]; |
3204 | 0 | src_uyvy += 4; |
3205 | 0 | } |
3206 | 0 | if (width & 1) { |
3207 | 0 | dst_y[width - 1] = src_uyvy[1]; |
3208 | 0 | } |
3209 | 0 | } |
3210 | | |
3211 | | #define BLEND(f, b, a) clamp255((((256 - a) * b) >> 8) + f) |
3212 | | |
3213 | | // Blend src_argb over src_argb1 and store to dst_argb. |
3214 | | // dst_argb may be src_argb or src_argb1. |
3215 | | // This code mimics the SSSE3 version for better testability. |
3216 | | void ARGBBlendRow_C(const uint8_t* src_argb, |
3217 | | const uint8_t* src_argb1, |
3218 | | uint8_t* dst_argb, |
3219 | 0 | int width) { |
3220 | 0 | int x; |
3221 | 0 | for (x = 0; x < width - 1; x += 2) { |
3222 | 0 | uint32_t fb = src_argb[0]; |
3223 | 0 | uint32_t fg = src_argb[1]; |
3224 | 0 | uint32_t fr = src_argb[2]; |
3225 | 0 | uint32_t a = src_argb[3]; |
3226 | 0 | uint32_t bb = src_argb1[0]; |
3227 | 0 | uint32_t bg = src_argb1[1]; |
3228 | 0 | uint32_t br = src_argb1[2]; |
3229 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, BLEND(fb, bb, a)); |
3230 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, BLEND(fg, bg, a)); |
3231 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, BLEND(fr, br, a)); |
3232 | 0 | dst_argb[3] = 255u; |
3233 | |
|
3234 | 0 | fb = src_argb[4 + 0]; |
3235 | 0 | fg = src_argb[4 + 1]; |
3236 | 0 | fr = src_argb[4 + 2]; |
3237 | 0 | a = src_argb[4 + 3]; |
3238 | 0 | bb = src_argb1[4 + 0]; |
3239 | 0 | bg = src_argb1[4 + 1]; |
3240 | 0 | br = src_argb1[4 + 2]; |
3241 | 0 | dst_argb[4 + 0] = STATIC_CAST(uint8_t, BLEND(fb, bb, a)); |
3242 | 0 | dst_argb[4 + 1] = STATIC_CAST(uint8_t, BLEND(fg, bg, a)); |
3243 | 0 | dst_argb[4 + 2] = STATIC_CAST(uint8_t, BLEND(fr, br, a)); |
3244 | 0 | dst_argb[4 + 3] = 255u; |
3245 | 0 | src_argb += 8; |
3246 | 0 | src_argb1 += 8; |
3247 | 0 | dst_argb += 8; |
3248 | 0 | } |
3249 | |
|
3250 | 0 | if (width & 1) { |
3251 | 0 | uint32_t fb = src_argb[0]; |
3252 | 0 | uint32_t fg = src_argb[1]; |
3253 | 0 | uint32_t fr = src_argb[2]; |
3254 | 0 | uint32_t a = src_argb[3]; |
3255 | 0 | uint32_t bb = src_argb1[0]; |
3256 | 0 | uint32_t bg = src_argb1[1]; |
3257 | 0 | uint32_t br = src_argb1[2]; |
3258 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, BLEND(fb, bb, a)); |
3259 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, BLEND(fg, bg, a)); |
3260 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, BLEND(fr, br, a)); |
3261 | 0 | dst_argb[3] = 255u; |
3262 | 0 | } |
3263 | 0 | } |
3264 | | #undef BLEND |
3265 | | |
3266 | 0 | #define UBLEND(f, b, a) (((a)*f) + ((255 - a) * b) + 255) >> 8 |
3267 | | void BlendPlaneRow_C(const uint8_t* src0, |
3268 | | const uint8_t* src1, |
3269 | | const uint8_t* alpha, |
3270 | | uint8_t* dst, |
3271 | 0 | int width) { |
3272 | 0 | int x; |
3273 | 0 | for (x = 0; x < width - 1; x += 2) { |
3274 | 0 | dst[0] = UBLEND(src0[0], src1[0], alpha[0]); |
3275 | 0 | dst[1] = UBLEND(src0[1], src1[1], alpha[1]); |
3276 | 0 | src0 += 2; |
3277 | 0 | src1 += 2; |
3278 | 0 | alpha += 2; |
3279 | 0 | dst += 2; |
3280 | 0 | } |
3281 | 0 | if (width & 1) { |
3282 | 0 | dst[0] = UBLEND(src0[0], src1[0], alpha[0]); |
3283 | 0 | } |
3284 | 0 | } |
3285 | | #undef UBLEND |
3286 | | |
3287 | 0 | #define ATTENUATE(f, a) (f * a + 255) >> 8 |
3288 | | |
3289 | | // Multiply source RGB by alpha and store to destination. |
3290 | 0 | void ARGBAttenuateRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width) { |
3291 | 0 | int i; |
3292 | 0 | for (i = 0; i < width - 1; i += 2) { |
3293 | 0 | uint32_t b = src_argb[0]; |
3294 | 0 | uint32_t g = src_argb[1]; |
3295 | 0 | uint32_t r = src_argb[2]; |
3296 | 0 | uint32_t a = src_argb[3]; |
3297 | 0 | dst_argb[0] = ATTENUATE(b, a); |
3298 | 0 | dst_argb[1] = ATTENUATE(g, a); |
3299 | 0 | dst_argb[2] = ATTENUATE(r, a); |
3300 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, a); |
3301 | 0 | b = src_argb[4]; |
3302 | 0 | g = src_argb[5]; |
3303 | 0 | r = src_argb[6]; |
3304 | 0 | a = src_argb[7]; |
3305 | 0 | dst_argb[4] = ATTENUATE(b, a); |
3306 | 0 | dst_argb[5] = ATTENUATE(g, a); |
3307 | 0 | dst_argb[6] = ATTENUATE(r, a); |
3308 | 0 | dst_argb[7] = STATIC_CAST(uint8_t, a); |
3309 | 0 | src_argb += 8; |
3310 | 0 | dst_argb += 8; |
3311 | 0 | } |
3312 | |
|
3313 | 0 | if (width & 1) { |
3314 | 0 | const uint32_t b = src_argb[0]; |
3315 | 0 | const uint32_t g = src_argb[1]; |
3316 | 0 | const uint32_t r = src_argb[2]; |
3317 | 0 | const uint32_t a = src_argb[3]; |
3318 | 0 | dst_argb[0] = ATTENUATE(b, a); |
3319 | 0 | dst_argb[1] = ATTENUATE(g, a); |
3320 | 0 | dst_argb[2] = ATTENUATE(r, a); |
3321 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, a); |
3322 | 0 | } |
3323 | 0 | } |
3324 | | #undef ATTENUATE |
3325 | | |
3326 | | // Divide source RGB by alpha and store to destination. |
3327 | | // b = (b * 255 + (a / 2)) / a; |
3328 | | // g = (g * 255 + (a / 2)) / a; |
3329 | | // r = (r * 255 + (a / 2)) / a; |
3330 | | // Reciprocal method is off by 1 on some values. ie 125 |
3331 | | // 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower. |
3332 | | #define T(a) 0x01000000 + (0x10000 / a) |
3333 | | const uint32_t fixed_invtbl8[256] = { |
3334 | | 0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), |
3335 | | T(0x07), T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), |
3336 | | T(0x0e), T(0x0f), T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), |
3337 | | T(0x15), T(0x16), T(0x17), T(0x18), T(0x19), T(0x1a), T(0x1b), |
3338 | | T(0x1c), T(0x1d), T(0x1e), T(0x1f), T(0x20), T(0x21), T(0x22), |
3339 | | T(0x23), T(0x24), T(0x25), T(0x26), T(0x27), T(0x28), T(0x29), |
3340 | | T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f), T(0x30), |
3341 | | T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37), |
3342 | | T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), |
3343 | | T(0x3f), T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), |
3344 | | T(0x46), T(0x47), T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), |
3345 | | T(0x4d), T(0x4e), T(0x4f), T(0x50), T(0x51), T(0x52), T(0x53), |
3346 | | T(0x54), T(0x55), T(0x56), T(0x57), T(0x58), T(0x59), T(0x5a), |
3347 | | T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f), T(0x60), T(0x61), |
3348 | | T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67), T(0x68), |
3349 | | T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f), |
3350 | | T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), |
3351 | | T(0x77), T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), |
3352 | | T(0x7e), T(0x7f), T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), |
3353 | | T(0x85), T(0x86), T(0x87), T(0x88), T(0x89), T(0x8a), T(0x8b), |
3354 | | T(0x8c), T(0x8d), T(0x8e), T(0x8f), T(0x90), T(0x91), T(0x92), |
3355 | | T(0x93), T(0x94), T(0x95), T(0x96), T(0x97), T(0x98), T(0x99), |
3356 | | T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f), T(0xa0), |
3357 | | T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7), |
3358 | | T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), |
3359 | | T(0xaf), T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), |
3360 | | T(0xb6), T(0xb7), T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), |
3361 | | T(0xbd), T(0xbe), T(0xbf), T(0xc0), T(0xc1), T(0xc2), T(0xc3), |
3362 | | T(0xc4), T(0xc5), T(0xc6), T(0xc7), T(0xc8), T(0xc9), T(0xca), |
3363 | | T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf), T(0xd0), T(0xd1), |
3364 | | T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7), T(0xd8), |
3365 | | T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf), |
3366 | | T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), |
3367 | | T(0xe7), T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), |
3368 | | T(0xee), T(0xef), T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), |
3369 | | T(0xf5), T(0xf6), T(0xf7), T(0xf8), T(0xf9), T(0xfa), T(0xfb), |
3370 | | T(0xfc), T(0xfd), T(0xfe), 0x01000100}; |
3371 | | #undef T |
3372 | | |
3373 | | #if defined(LIBYUV_UNATTENUATE_DUP) |
3374 | | // This code mimics the Intel SIMD version for better testability. |
3375 | | #define UNATTENUATE(f, ia) clamp255(((f | (f << 8)) * ia) >> 16) |
3376 | | #else |
3377 | | #define UNATTENUATE(f, ia) clamp255((f * ia) >> 8) |
3378 | | #endif |
3379 | | |
3380 | | // mimics the Intel SIMD code for exactness. |
3381 | | void ARGBUnattenuateRow_C(const uint8_t* src_argb, |
3382 | | uint8_t* dst_argb, |
3383 | 0 | int width) { |
3384 | 0 | int i; |
3385 | 0 | for (i = 0; i < width; ++i) { |
3386 | 0 | uint32_t b = src_argb[0]; |
3387 | 0 | uint32_t g = src_argb[1]; |
3388 | 0 | uint32_t r = src_argb[2]; |
3389 | 0 | const uint32_t a = src_argb[3]; |
3390 | 0 | const uint32_t ia = fixed_invtbl8[a] & 0xffff; // 8.8 fixed point |
3391 | | |
3392 | | // Clamping should not be necessary but is free in assembly. |
3393 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, UNATTENUATE(b, ia)); |
3394 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, UNATTENUATE(g, ia)); |
3395 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, UNATTENUATE(r, ia)); |
3396 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, a); |
3397 | 0 | src_argb += 4; |
3398 | 0 | dst_argb += 4; |
3399 | 0 | } |
3400 | 0 | } |
3401 | | |
3402 | | void ComputeCumulativeSumRow_C(const uint8_t* row, |
3403 | | int32_t* cumsum, |
3404 | | const int32_t* previous_cumsum, |
3405 | 0 | int width) { |
3406 | 0 | int32_t row_sum[4] = {0, 0, 0, 0}; |
3407 | 0 | int x; |
3408 | 0 | for (x = 0; x < width; ++x) { |
3409 | 0 | row_sum[0] += row[x * 4 + 0]; |
3410 | 0 | row_sum[1] += row[x * 4 + 1]; |
3411 | 0 | row_sum[2] += row[x * 4 + 2]; |
3412 | 0 | row_sum[3] += row[x * 4 + 3]; |
3413 | 0 | cumsum[x * 4 + 0] = row_sum[0] + previous_cumsum[x * 4 + 0]; |
3414 | 0 | cumsum[x * 4 + 1] = row_sum[1] + previous_cumsum[x * 4 + 1]; |
3415 | 0 | cumsum[x * 4 + 2] = row_sum[2] + previous_cumsum[x * 4 + 2]; |
3416 | 0 | cumsum[x * 4 + 3] = row_sum[3] + previous_cumsum[x * 4 + 3]; |
3417 | 0 | } |
3418 | 0 | } |
3419 | | |
3420 | | void CumulativeSumToAverageRow_C(const int32_t* tl, |
3421 | | const int32_t* bl, |
3422 | | int w, |
3423 | | int area, |
3424 | | uint8_t* dst, |
3425 | 0 | int count) { |
3426 | 0 | float ooa; |
3427 | 0 | int i; |
3428 | 0 | assert(area != 0); |
3429 | |
|
3430 | 0 | ooa = 1.0f / STATIC_CAST(float, area); |
3431 | 0 | for (i = 0; i < count; ++i) { |
3432 | 0 | dst[0] = |
3433 | 0 | (uint8_t)(STATIC_CAST(float, bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * |
3434 | 0 | ooa); |
3435 | 0 | dst[1] = |
3436 | 0 | (uint8_t)(STATIC_CAST(float, bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * |
3437 | 0 | ooa); |
3438 | 0 | dst[2] = |
3439 | 0 | (uint8_t)(STATIC_CAST(float, bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * |
3440 | 0 | ooa); |
3441 | 0 | dst[3] = |
3442 | 0 | (uint8_t)(STATIC_CAST(float, bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * |
3443 | 0 | ooa); |
3444 | 0 | dst += 4; |
3445 | 0 | tl += 4; |
3446 | 0 | bl += 4; |
3447 | 0 | } |
3448 | 0 | } |
3449 | | |
3450 | | // Copy pixels from rotated source to destination row with a slope. |
3451 | | LIBYUV_API |
3452 | | void ARGBAffineRow_C(const uint8_t* src_argb, |
3453 | | int src_argb_stride, |
3454 | | uint8_t* dst_argb, |
3455 | | const float* uv_dudv, |
3456 | 0 | int width) { |
3457 | 0 | int i; |
3458 | | // Render a row of pixels from source into a buffer. |
3459 | 0 | float uv[2]; |
3460 | 0 | uv[0] = uv_dudv[0]; |
3461 | 0 | uv[1] = uv_dudv[1]; |
3462 | 0 | for (i = 0; i < width; ++i) { |
3463 | 0 | int x = (int)(uv[0]); |
3464 | 0 | int y = (int)(uv[1]); |
3465 | 0 | *(uint32_t*)(dst_argb) = |
3466 | 0 | *(const uint32_t*)(src_argb + y * src_argb_stride + x * 4); |
3467 | 0 | dst_argb += 4; |
3468 | 0 | uv[0] += uv_dudv[2]; |
3469 | 0 | uv[1] += uv_dudv[3]; |
3470 | 0 | } |
3471 | 0 | } |
3472 | | |
3473 | | // Blend 2 rows into 1. |
3474 | | static void HalfRow_C(const uint8_t* src_uv, |
3475 | | ptrdiff_t src_uv_stride, |
3476 | | uint8_t* dst_uv, |
3477 | 0 | int width) { |
3478 | 0 | int x; |
3479 | 0 | for (x = 0; x < width; ++x) { |
3480 | 0 | dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1; |
3481 | 0 | } |
3482 | 0 | } |
3483 | | |
3484 | | static void HalfRow_16_C(const uint16_t* src_uv, |
3485 | | ptrdiff_t src_uv_stride, |
3486 | | uint16_t* dst_uv, |
3487 | 22.0k | int width) { |
3488 | 22.0k | int x; |
3489 | 27.3M | for (x = 0; x < width; ++x) { |
3490 | 27.2M | dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1; |
3491 | 27.2M | } |
3492 | 22.0k | } |
3493 | | |
3494 | | static void HalfRow_16To8_C(const uint16_t* src_uv, |
3495 | | ptrdiff_t src_uv_stride, |
3496 | | uint8_t* dst_uv, |
3497 | | int scale, |
3498 | 0 | int width) { |
3499 | 0 | int x; |
3500 | 0 | for (x = 0; x < width; ++x) { |
3501 | 0 | dst_uv[x] = STATIC_CAST( |
3502 | 0 | uint8_t, |
3503 | 0 | C16TO8((src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1, scale)); |
3504 | 0 | } |
3505 | 0 | } |
3506 | | |
3507 | | // C version 2x2 -> 2x1. |
3508 | | void InterpolateRow_C(uint8_t* dst_ptr, |
3509 | | const uint8_t* src_ptr, |
3510 | | ptrdiff_t src_stride, |
3511 | | int width, |
3512 | 0 | int source_y_fraction) { |
3513 | 0 | int y1_fraction = source_y_fraction; |
3514 | 0 | int y0_fraction = 256 - y1_fraction; |
3515 | 0 | const uint8_t* src_ptr1 = src_ptr + src_stride; |
3516 | 0 | int x; |
3517 | 0 | assert(source_y_fraction >= 0); |
3518 | 0 | assert(source_y_fraction < 256); |
3519 | |
|
3520 | 0 | if (y1_fraction == 0) { |
3521 | 0 | memcpy(dst_ptr, src_ptr, width); |
3522 | 0 | return; |
3523 | 0 | } |
3524 | 0 | if (y1_fraction == 128) { |
3525 | 0 | HalfRow_C(src_ptr, src_stride, dst_ptr, width); |
3526 | 0 | return; |
3527 | 0 | } |
3528 | 0 | for (x = 0; x < width; ++x) { |
3529 | 0 | dst_ptr[0] = STATIC_CAST( |
3530 | 0 | uint8_t, |
3531 | 0 | (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8); |
3532 | 0 | ++src_ptr; |
3533 | 0 | ++src_ptr1; |
3534 | 0 | ++dst_ptr; |
3535 | 0 | } |
3536 | 0 | } |
3537 | | |
3538 | | // C version 2x2 -> 2x1. |
3539 | | void InterpolateRow_16_C(uint16_t* dst_ptr, |
3540 | | const uint16_t* src_ptr, |
3541 | | ptrdiff_t src_stride, |
3542 | | int width, |
3543 | 6.09M | int source_y_fraction) { |
3544 | 6.09M | int y1_fraction = source_y_fraction; |
3545 | 6.09M | int y0_fraction = 256 - y1_fraction; |
3546 | 6.09M | const uint16_t* src_ptr1 = src_ptr + src_stride; |
3547 | 6.09M | int x; |
3548 | 6.09M | assert(source_y_fraction >= 0); |
3549 | 6.09M | assert(source_y_fraction < 256); |
3550 | | |
3551 | 6.09M | if (y1_fraction == 0) { |
3552 | 880k | memcpy(dst_ptr, src_ptr, width * 2); |
3553 | 880k | return; |
3554 | 880k | } |
3555 | 5.21M | if (y1_fraction == 128) { |
3556 | 22.0k | HalfRow_16_C(src_ptr, src_stride, dst_ptr, width); |
3557 | 22.0k | return; |
3558 | 22.0k | } |
3559 | 6.82G | for (x = 0; x < width; ++x) { |
3560 | 6.81G | dst_ptr[0] = STATIC_CAST( |
3561 | 6.81G | uint16_t, |
3562 | 6.81G | (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8); |
3563 | 6.81G | ++src_ptr; |
3564 | 6.81G | ++src_ptr1; |
3565 | 6.81G | ++dst_ptr; |
3566 | 6.81G | } |
3567 | 5.19M | } |
3568 | | |
3569 | | // C version 2x2 16 bit-> 2x1 8 bit. |
3570 | | // Use scale to convert lsb formats to msb, depending how many bits there are: |
3571 | | // 32768 = 9 bits |
3572 | | // 16384 = 10 bits |
3573 | | // 4096 = 12 bits |
3574 | | // 256 = 16 bits |
3575 | | // TODO(fbarchard): change scale to bits |
3576 | | |
3577 | | void InterpolateRow_16To8_C(uint8_t* dst_ptr, |
3578 | | const uint16_t* src_ptr, |
3579 | | ptrdiff_t src_stride, |
3580 | | int scale, |
3581 | | int width, |
3582 | 0 | int source_y_fraction) { |
3583 | 0 | int y1_fraction = source_y_fraction; |
3584 | 0 | int y0_fraction = 256 - y1_fraction; |
3585 | 0 | const uint16_t* src_ptr1 = src_ptr + src_stride; |
3586 | 0 | int x; |
3587 | 0 | assert(source_y_fraction >= 0); |
3588 | 0 | assert(source_y_fraction < 256); |
3589 | |
|
3590 | 0 | if (source_y_fraction == 0) { |
3591 | 0 | Convert16To8Row_C(src_ptr, dst_ptr, scale, width); |
3592 | 0 | return; |
3593 | 0 | } |
3594 | 0 | if (source_y_fraction == 128) { |
3595 | 0 | HalfRow_16To8_C(src_ptr, src_stride, dst_ptr, scale, width); |
3596 | 0 | return; |
3597 | 0 | } |
3598 | 0 | for (x = 0; x < width; ++x) { |
3599 | 0 | dst_ptr[0] = STATIC_CAST( |
3600 | 0 | uint8_t, |
3601 | 0 | C16TO8( |
3602 | 0 | (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8, |
3603 | 0 | scale)); |
3604 | 0 | src_ptr += 1; |
3605 | 0 | src_ptr1 += 1; |
3606 | 0 | dst_ptr += 1; |
3607 | 0 | } |
3608 | 0 | } |
3609 | | |
3610 | | // Use first 4 shuffler values to reorder ARGB channels. |
3611 | | void ARGBShuffleRow_C(const uint8_t* src_argb, |
3612 | | uint8_t* dst_argb, |
3613 | | const uint8_t* shuffler, |
3614 | 0 | int width) { |
3615 | 0 | int index0 = shuffler[0]; |
3616 | 0 | int index1 = shuffler[1]; |
3617 | 0 | int index2 = shuffler[2]; |
3618 | 0 | int index3 = shuffler[3]; |
3619 | | // Shuffle a row of ARGB. |
3620 | 0 | int x; |
3621 | 0 | for (x = 0; x < width; ++x) { |
3622 | | // To support in-place conversion. |
3623 | 0 | uint8_t b = src_argb[index0]; |
3624 | 0 | uint8_t g = src_argb[index1]; |
3625 | 0 | uint8_t r = src_argb[index2]; |
3626 | 0 | uint8_t a = src_argb[index3]; |
3627 | 0 | dst_argb[0] = b; |
3628 | 0 | dst_argb[1] = g; |
3629 | 0 | dst_argb[2] = r; |
3630 | 0 | dst_argb[3] = a; |
3631 | 0 | src_argb += 4; |
3632 | 0 | dst_argb += 4; |
3633 | 0 | } |
3634 | 0 | } |
3635 | | |
3636 | | void I422ToYUY2Row_C(const uint8_t* src_y, |
3637 | | const uint8_t* src_u, |
3638 | | const uint8_t* src_v, |
3639 | | uint8_t* dst_frame, |
3640 | 0 | int width) { |
3641 | 0 | int x; |
3642 | 0 | for (x = 0; x < width - 1; x += 2) { |
3643 | 0 | dst_frame[0] = src_y[0]; |
3644 | 0 | dst_frame[1] = src_u[0]; |
3645 | 0 | dst_frame[2] = src_y[1]; |
3646 | 0 | dst_frame[3] = src_v[0]; |
3647 | 0 | dst_frame += 4; |
3648 | 0 | src_y += 2; |
3649 | 0 | src_u += 1; |
3650 | 0 | src_v += 1; |
3651 | 0 | } |
3652 | 0 | if (width & 1) { |
3653 | 0 | dst_frame[0] = src_y[0]; |
3654 | 0 | dst_frame[1] = src_u[0]; |
3655 | 0 | dst_frame[2] = 0; |
3656 | 0 | dst_frame[3] = src_v[0]; |
3657 | 0 | } |
3658 | 0 | } |
3659 | | |
3660 | | void I422ToUYVYRow_C(const uint8_t* src_y, |
3661 | | const uint8_t* src_u, |
3662 | | const uint8_t* src_v, |
3663 | | uint8_t* dst_frame, |
3664 | 0 | int width) { |
3665 | 0 | int x; |
3666 | 0 | for (x = 0; x < width - 1; x += 2) { |
3667 | 0 | dst_frame[0] = src_u[0]; |
3668 | 0 | dst_frame[1] = src_y[0]; |
3669 | 0 | dst_frame[2] = src_v[0]; |
3670 | 0 | dst_frame[3] = src_y[1]; |
3671 | 0 | dst_frame += 4; |
3672 | 0 | src_y += 2; |
3673 | 0 | src_u += 1; |
3674 | 0 | src_v += 1; |
3675 | 0 | } |
3676 | 0 | if (width & 1) { |
3677 | 0 | dst_frame[0] = src_u[0]; |
3678 | 0 | dst_frame[1] = src_y[0]; |
3679 | 0 | dst_frame[2] = src_v[0]; |
3680 | 0 | dst_frame[3] = 0; |
3681 | 0 | } |
3682 | 0 | } |
3683 | | |
3684 | | void ARGBPolynomialRow_C(const uint8_t* src_argb, |
3685 | | uint8_t* dst_argb, |
3686 | | const float* poly, |
3687 | 0 | int width) { |
3688 | 0 | int i; |
3689 | 0 | for (i = 0; i < width; ++i) { |
3690 | 0 | float b = (float)(src_argb[0]); |
3691 | 0 | float g = (float)(src_argb[1]); |
3692 | 0 | float r = (float)(src_argb[2]); |
3693 | 0 | float a = (float)(src_argb[3]); |
3694 | 0 | float b2 = b * b; |
3695 | 0 | float g2 = g * g; |
3696 | 0 | float r2 = r * r; |
3697 | 0 | float a2 = a * a; |
3698 | 0 | float db = poly[0] + poly[4] * b; |
3699 | 0 | float dg = poly[1] + poly[5] * g; |
3700 | 0 | float dr = poly[2] + poly[6] * r; |
3701 | 0 | float da = poly[3] + poly[7] * a; |
3702 | 0 | float b3 = b2 * b; |
3703 | 0 | float g3 = g2 * g; |
3704 | 0 | float r3 = r2 * r; |
3705 | 0 | float a3 = a2 * a; |
3706 | 0 | db += poly[8] * b2; |
3707 | 0 | dg += poly[9] * g2; |
3708 | 0 | dr += poly[10] * r2; |
3709 | 0 | da += poly[11] * a2; |
3710 | 0 | db += poly[12] * b3; |
3711 | 0 | dg += poly[13] * g3; |
3712 | 0 | dr += poly[14] * r3; |
3713 | 0 | da += poly[15] * a3; |
3714 | |
|
3715 | 0 | dst_argb[0] = STATIC_CAST(uint8_t, Clamp((int32_t)(db))); |
3716 | 0 | dst_argb[1] = STATIC_CAST(uint8_t, Clamp((int32_t)(dg))); |
3717 | 0 | dst_argb[2] = STATIC_CAST(uint8_t, Clamp((int32_t)(dr))); |
3718 | 0 | dst_argb[3] = STATIC_CAST(uint8_t, Clamp((int32_t)(da))); |
3719 | 0 | src_argb += 4; |
3720 | 0 | dst_argb += 4; |
3721 | 0 | } |
3722 | 0 | } |
3723 | | |
3724 | | // Samples assumed to be unsigned in low 9, 10 or 12 bits. Scale factor |
3725 | | // adjust the source integer range to the half float range desired. |
3726 | | |
3727 | | // This magic constant is 2^-112. Multiplying by this |
3728 | | // is the same as subtracting 112 from the exponent, which |
3729 | | // is the difference in exponent bias between 32-bit and |
3730 | | // 16-bit floats. Once we've done this subtraction, we can |
3731 | | // simply extract the low bits of the exponent and the high |
3732 | | // bits of the mantissa from our float and we're done. |
3733 | | |
3734 | | // Work around GCC 7 punning warning -Wstrict-aliasing |
3735 | | #if defined(__GNUC__) |
3736 | | typedef uint32_t __attribute__((__may_alias__)) uint32_alias_t; |
3737 | | #else |
3738 | | typedef uint32_t uint32_alias_t; |
3739 | | #endif |
3740 | | |
3741 | | void HalfFloatRow_C(const uint16_t* src, |
3742 | | uint16_t* dst, |
3743 | | float scale, |
3744 | 0 | int width) { |
3745 | 0 | int i; |
3746 | 0 | float mult = 1.9259299444e-34f * scale; |
3747 | 0 | for (i = 0; i < width; ++i) { |
3748 | 0 | float value = src[i] * mult; |
3749 | 0 | dst[i] = (uint16_t)((*(const uint32_alias_t*)&value) >> 13); |
3750 | 0 | } |
3751 | 0 | } |
3752 | | |
3753 | 0 | void ByteToFloatRow_C(const uint8_t* src, float* dst, float scale, int width) { |
3754 | 0 | int i; |
3755 | 0 | for (i = 0; i < width; ++i) { |
3756 | 0 | float value = src[i] * scale; |
3757 | 0 | dst[i] = value; |
3758 | 0 | } |
3759 | 0 | } |
3760 | | |
3761 | | void ARGBLumaColorTableRow_C(const uint8_t* src_argb, |
3762 | | uint8_t* dst_argb, |
3763 | | int width, |
3764 | | const uint8_t* luma, |
3765 | 0 | uint32_t lumacoeff) { |
3766 | 0 | uint32_t bc = lumacoeff & 0xff; |
3767 | 0 | uint32_t gc = (lumacoeff >> 8) & 0xff; |
3768 | 0 | uint32_t rc = (lumacoeff >> 16) & 0xff; |
3769 | |
|
3770 | 0 | int i; |
3771 | 0 | for (i = 0; i < width - 1; i += 2) { |
3772 | | // Luminance in rows, color values in columns. |
3773 | 0 | const uint8_t* luma0 = |
3774 | 0 | ((src_argb[0] * bc + src_argb[1] * gc + src_argb[2] * rc) & 0x7F00u) + |
3775 | 0 | luma; |
3776 | 0 | const uint8_t* luma1; |
3777 | 0 | dst_argb[0] = luma0[src_argb[0]]; |
3778 | 0 | dst_argb[1] = luma0[src_argb[1]]; |
3779 | 0 | dst_argb[2] = luma0[src_argb[2]]; |
3780 | 0 | dst_argb[3] = src_argb[3]; |
3781 | 0 | luma1 = |
3782 | 0 | ((src_argb[4] * bc + src_argb[5] * gc + src_argb[6] * rc) & 0x7F00u) + |
3783 | 0 | luma; |
3784 | 0 | dst_argb[4] = luma1[src_argb[4]]; |
3785 | 0 | dst_argb[5] = luma1[src_argb[5]]; |
3786 | 0 | dst_argb[6] = luma1[src_argb[6]]; |
3787 | 0 | dst_argb[7] = src_argb[7]; |
3788 | 0 | src_argb += 8; |
3789 | 0 | dst_argb += 8; |
3790 | 0 | } |
3791 | 0 | if (width & 1) { |
3792 | | // Luminance in rows, color values in columns. |
3793 | 0 | const uint8_t* luma0 = |
3794 | 0 | ((src_argb[0] * bc + src_argb[1] * gc + src_argb[2] * rc) & 0x7F00u) + |
3795 | 0 | luma; |
3796 | 0 | dst_argb[0] = luma0[src_argb[0]]; |
3797 | 0 | dst_argb[1] = luma0[src_argb[1]]; |
3798 | 0 | dst_argb[2] = luma0[src_argb[2]]; |
3799 | 0 | dst_argb[3] = src_argb[3]; |
3800 | 0 | } |
3801 | 0 | } |
3802 | | |
3803 | 0 | void ARGBCopyAlphaRow_C(const uint8_t* src, uint8_t* dst, int width) { |
3804 | 0 | int i; |
3805 | 0 | for (i = 0; i < width - 1; i += 2) { |
3806 | 0 | dst[3] = src[3]; |
3807 | 0 | dst[7] = src[7]; |
3808 | 0 | dst += 8; |
3809 | 0 | src += 8; |
3810 | 0 | } |
3811 | 0 | if (width & 1) { |
3812 | 0 | dst[3] = src[3]; |
3813 | 0 | } |
3814 | 0 | } |
3815 | | |
3816 | 0 | void ARGBExtractAlphaRow_C(const uint8_t* src_argb, uint8_t* dst_a, int width) { |
3817 | 0 | int i; |
3818 | 0 | for (i = 0; i < width - 1; i += 2) { |
3819 | 0 | dst_a[0] = src_argb[3]; |
3820 | 0 | dst_a[1] = src_argb[7]; |
3821 | 0 | dst_a += 2; |
3822 | 0 | src_argb += 8; |
3823 | 0 | } |
3824 | 0 | if (width & 1) { |
3825 | 0 | dst_a[0] = src_argb[3]; |
3826 | 0 | } |
3827 | 0 | } |
3828 | | |
3829 | 0 | void ARGBCopyYToAlphaRow_C(const uint8_t* src, uint8_t* dst, int width) { |
3830 | 0 | int i; |
3831 | 0 | for (i = 0; i < width - 1; i += 2) { |
3832 | 0 | dst[3] = src[0]; |
3833 | 0 | dst[7] = src[1]; |
3834 | 0 | dst += 8; |
3835 | 0 | src += 2; |
3836 | 0 | } |
3837 | 0 | if (width & 1) { |
3838 | 0 | dst[3] = src[0]; |
3839 | 0 | } |
3840 | 0 | } |
3841 | | |
3842 | | // Maximum temporary width for wrappers to process at a time, in pixels. |
3843 | 223k | #define MAXTWIDTH 2048 |
3844 | | |
3845 | | #if !(defined(_MSC_VER) && !defined(__clang__) && defined(_M_IX86)) && \ |
3846 | | defined(HAS_I422TORGB565ROW_SSSE3) && !defined(LIBYUV_ENABLE_ROWWIN) |
3847 | | // row_win.cc has asm version, but GCC uses 2 step wrapper. |
3848 | | void I422ToRGB565Row_SSSE3(const uint8_t* src_y, |
3849 | | const uint8_t* src_u, |
3850 | | const uint8_t* src_v, |
3851 | | uint8_t* dst_rgb565, |
3852 | | const struct YuvConstants* yuvconstants, |
3853 | 0 | int width) { |
3854 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3855 | 0 | while (width > 0) { |
3856 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3857 | 0 | I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, yuvconstants, twidth); |
3858 | 0 | ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); |
3859 | 0 | src_y += twidth; |
3860 | 0 | src_u += twidth / 2; |
3861 | 0 | src_v += twidth / 2; |
3862 | 0 | dst_rgb565 += twidth * 2; |
3863 | 0 | width -= twidth; |
3864 | 0 | } |
3865 | 0 | } |
3866 | | #endif |
3867 | | |
3868 | | #if defined(HAS_I422TOARGB1555ROW_SSSE3) |
3869 | | void I422ToARGB1555Row_SSSE3(const uint8_t* src_y, |
3870 | | const uint8_t* src_u, |
3871 | | const uint8_t* src_v, |
3872 | | uint8_t* dst_argb1555, |
3873 | | const struct YuvConstants* yuvconstants, |
3874 | 0 | int width) { |
3875 | | // Row buffer for intermediate ARGB pixels. |
3876 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3877 | 0 | while (width > 0) { |
3878 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3879 | 0 | I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, yuvconstants, twidth); |
3880 | 0 | ARGBToARGB1555Row_SSE2(row, dst_argb1555, twidth); |
3881 | 0 | src_y += twidth; |
3882 | 0 | src_u += twidth / 2; |
3883 | 0 | src_v += twidth / 2; |
3884 | 0 | dst_argb1555 += twidth * 2; |
3885 | 0 | width -= twidth; |
3886 | 0 | } |
3887 | 0 | } |
3888 | | #endif |
3889 | | |
3890 | | #if defined(HAS_I422TOARGB4444ROW_SSSE3) |
3891 | | void I422ToARGB4444Row_SSSE3(const uint8_t* src_y, |
3892 | | const uint8_t* src_u, |
3893 | | const uint8_t* src_v, |
3894 | | uint8_t* dst_argb4444, |
3895 | | const struct YuvConstants* yuvconstants, |
3896 | 0 | int width) { |
3897 | | // Row buffer for intermediate ARGB pixels. |
3898 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3899 | 0 | while (width > 0) { |
3900 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3901 | 0 | I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, yuvconstants, twidth); |
3902 | 0 | ARGBToARGB4444Row_SSE2(row, dst_argb4444, twidth); |
3903 | 0 | src_y += twidth; |
3904 | 0 | src_u += twidth / 2; |
3905 | 0 | src_v += twidth / 2; |
3906 | 0 | dst_argb4444 += twidth * 2; |
3907 | 0 | width -= twidth; |
3908 | 0 | } |
3909 | 0 | } |
3910 | | #endif |
3911 | | |
3912 | | #if defined(HAS_NV12TORGB565ROW_SSSE3) |
3913 | | void NV12ToRGB565Row_SSSE3(const uint8_t* src_y, |
3914 | | const uint8_t* src_uv, |
3915 | | uint8_t* dst_rgb565, |
3916 | | const struct YuvConstants* yuvconstants, |
3917 | 0 | int width) { |
3918 | | // Row buffer for intermediate ARGB pixels. |
3919 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3920 | 0 | while (width > 0) { |
3921 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3922 | 0 | NV12ToARGBRow_SSSE3(src_y, src_uv, row, yuvconstants, twidth); |
3923 | 0 | ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); |
3924 | 0 | src_y += twidth; |
3925 | 0 | src_uv += twidth; |
3926 | 0 | dst_rgb565 += twidth * 2; |
3927 | 0 | width -= twidth; |
3928 | 0 | } |
3929 | 0 | } |
3930 | | #endif |
3931 | | |
3932 | | #if defined(HAS_NV12TORGB24ROW_SSSE3) |
3933 | | void NV12ToRGB24Row_SSSE3(const uint8_t* src_y, |
3934 | | const uint8_t* src_uv, |
3935 | | uint8_t* dst_rgb24, |
3936 | | const struct YuvConstants* yuvconstants, |
3937 | 0 | int width) { |
3938 | | // Row buffer for intermediate ARGB pixels. |
3939 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3940 | 0 | while (width > 0) { |
3941 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3942 | 0 | NV12ToARGBRow_SSSE3(src_y, src_uv, row, yuvconstants, twidth); |
3943 | 0 | ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); |
3944 | 0 | src_y += twidth; |
3945 | 0 | src_uv += twidth; |
3946 | 0 | dst_rgb24 += twidth * 3; |
3947 | 0 | width -= twidth; |
3948 | 0 | } |
3949 | 0 | } |
3950 | | #endif |
3951 | | |
3952 | | #if defined(HAS_NV21TORGB24ROW_SSSE3) |
3953 | | void NV21ToRGB24Row_SSSE3(const uint8_t* src_y, |
3954 | | const uint8_t* src_vu, |
3955 | | uint8_t* dst_rgb24, |
3956 | | const struct YuvConstants* yuvconstants, |
3957 | 0 | int width) { |
3958 | | // Row buffer for intermediate ARGB pixels. |
3959 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3960 | 0 | while (width > 0) { |
3961 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3962 | 0 | NV21ToARGBRow_SSSE3(src_y, src_vu, row, yuvconstants, twidth); |
3963 | 0 | ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); |
3964 | 0 | src_y += twidth; |
3965 | 0 | src_vu += twidth; |
3966 | 0 | dst_rgb24 += twidth * 3; |
3967 | 0 | width -= twidth; |
3968 | 0 | } |
3969 | 0 | } |
3970 | | #endif |
3971 | | |
3972 | | #if defined(HAS_NV12TORGB24ROW_AVX2) |
3973 | | void NV12ToRGB24Row_AVX2(const uint8_t* src_y, |
3974 | | const uint8_t* src_uv, |
3975 | | uint8_t* dst_rgb24, |
3976 | | const struct YuvConstants* yuvconstants, |
3977 | 0 | int width) { |
3978 | | // Row buffer for intermediate ARGB pixels. |
3979 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
3980 | 0 | while (width > 0) { |
3981 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
3982 | 0 | NV12ToARGBRow_AVX2(src_y, src_uv, row, yuvconstants, twidth); |
3983 | 0 | #if defined(HAS_ARGBTORGB24ROW_AVX2) |
3984 | 0 | ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth); |
3985 | | #else |
3986 | | ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); |
3987 | | #endif |
3988 | 0 | src_y += twidth; |
3989 | 0 | src_uv += twidth; |
3990 | 0 | dst_rgb24 += twidth * 3; |
3991 | 0 | width -= twidth; |
3992 | 0 | } |
3993 | 0 | } |
3994 | | #endif |
3995 | | |
3996 | | #if defined(HAS_NV21TORGB24ROW_AVX2) |
3997 | | void NV21ToRGB24Row_AVX2(const uint8_t* src_y, |
3998 | | const uint8_t* src_vu, |
3999 | | uint8_t* dst_rgb24, |
4000 | | const struct YuvConstants* yuvconstants, |
4001 | 0 | int width) { |
4002 | | // Row buffer for intermediate ARGB pixels. |
4003 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4004 | 0 | while (width > 0) { |
4005 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4006 | 0 | NV21ToARGBRow_AVX2(src_y, src_vu, row, yuvconstants, twidth); |
4007 | 0 | #if defined(HAS_ARGBTORGB24ROW_AVX2) |
4008 | 0 | ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth); |
4009 | | #else |
4010 | | ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); |
4011 | | #endif |
4012 | 0 | src_y += twidth; |
4013 | 0 | src_vu += twidth; |
4014 | 0 | dst_rgb24 += twidth * 3; |
4015 | 0 | width -= twidth; |
4016 | 0 | } |
4017 | 0 | } |
4018 | | #endif |
4019 | | |
4020 | | #if defined(HAS_I422TORGB565ROW_AVX2) |
4021 | | void I422ToRGB565Row_AVX2(const uint8_t* src_y, |
4022 | | const uint8_t* src_u, |
4023 | | const uint8_t* src_v, |
4024 | | uint8_t* dst_rgb565, |
4025 | | const struct YuvConstants* yuvconstants, |
4026 | 2.42k | int width) { |
4027 | 2.42k | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4028 | 4.85k | while (width > 0) { |
4029 | 2.42k | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4030 | 2.42k | I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth); |
4031 | | #if defined(HAS_ARGBTORGB565ROW_AVX2) |
4032 | | ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth); |
4033 | | #else |
4034 | 2.42k | ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); |
4035 | 2.42k | #endif |
4036 | 2.42k | src_y += twidth; |
4037 | 2.42k | src_u += twidth / 2; |
4038 | 2.42k | src_v += twidth / 2; |
4039 | 2.42k | dst_rgb565 += twidth * 2; |
4040 | 2.42k | width -= twidth; |
4041 | 2.42k | } |
4042 | 2.42k | } |
4043 | | #endif |
4044 | | |
4045 | | #if defined(HAS_I422TOARGB1555ROW_AVX2) |
4046 | | void I422ToARGB1555Row_AVX2(const uint8_t* src_y, |
4047 | | const uint8_t* src_u, |
4048 | | const uint8_t* src_v, |
4049 | | uint8_t* dst_argb1555, |
4050 | | const struct YuvConstants* yuvconstants, |
4051 | 0 | int width) { |
4052 | | // Row buffer for intermediate ARGB pixels. |
4053 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4054 | 0 | while (width > 0) { |
4055 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4056 | 0 | I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth); |
4057 | | #if defined(HAS_ARGBTOARGB1555ROW_AVX2) |
4058 | | ARGBToARGB1555Row_AVX2(row, dst_argb1555, twidth); |
4059 | | #else |
4060 | 0 | ARGBToARGB1555Row_SSE2(row, dst_argb1555, twidth); |
4061 | 0 | #endif |
4062 | 0 | src_y += twidth; |
4063 | 0 | src_u += twidth / 2; |
4064 | 0 | src_v += twidth / 2; |
4065 | 0 | dst_argb1555 += twidth * 2; |
4066 | 0 | width -= twidth; |
4067 | 0 | } |
4068 | 0 | } |
4069 | | #endif |
4070 | | |
4071 | | #if defined(HAS_I422TOARGB4444ROW_AVX2) |
4072 | | void I422ToARGB4444Row_AVX2(const uint8_t* src_y, |
4073 | | const uint8_t* src_u, |
4074 | | const uint8_t* src_v, |
4075 | | uint8_t* dst_argb4444, |
4076 | | const struct YuvConstants* yuvconstants, |
4077 | 0 | int width) { |
4078 | | // Row buffer for intermediate ARGB pixels. |
4079 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4080 | 0 | while (width > 0) { |
4081 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4082 | 0 | I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth); |
4083 | | #if defined(HAS_ARGBTOARGB4444ROW_AVX2) |
4084 | | ARGBToARGB4444Row_AVX2(row, dst_argb4444, twidth); |
4085 | | #else |
4086 | 0 | ARGBToARGB4444Row_SSE2(row, dst_argb4444, twidth); |
4087 | 0 | #endif |
4088 | 0 | src_y += twidth; |
4089 | 0 | src_u += twidth / 2; |
4090 | 0 | src_v += twidth / 2; |
4091 | 0 | dst_argb4444 += twidth * 2; |
4092 | 0 | width -= twidth; |
4093 | 0 | } |
4094 | 0 | } |
4095 | | #endif |
4096 | | |
4097 | | #if defined(HAS_I422TORGB24ROW_AVX2) |
4098 | | void I422ToRGB24Row_AVX2(const uint8_t* src_y, |
4099 | | const uint8_t* src_u, |
4100 | | const uint8_t* src_v, |
4101 | | uint8_t* dst_rgb24, |
4102 | | const struct YuvConstants* yuvconstants, |
4103 | 6.09k | int width) { |
4104 | | // Row buffer for intermediate ARGB pixels. |
4105 | 6.09k | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4106 | 12.1k | while (width > 0) { |
4107 | 6.09k | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4108 | 6.09k | I422ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth); |
4109 | 6.09k | #if defined(HAS_ARGBTORGB24ROW_AVX2) |
4110 | 6.09k | ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth); |
4111 | | #else |
4112 | | ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); |
4113 | | #endif |
4114 | 6.09k | src_y += twidth; |
4115 | 6.09k | src_u += twidth / 2; |
4116 | 6.09k | src_v += twidth / 2; |
4117 | 6.09k | dst_rgb24 += twidth * 3; |
4118 | 6.09k | width -= twidth; |
4119 | 6.09k | } |
4120 | 6.09k | } |
4121 | | #endif |
4122 | | |
4123 | | #if defined(HAS_I444TORGB24ROW_AVX2) |
4124 | | void I444ToRGB24Row_AVX2(const uint8_t* src_y, |
4125 | | const uint8_t* src_u, |
4126 | | const uint8_t* src_v, |
4127 | | uint8_t* dst_rgb24, |
4128 | | const struct YuvConstants* yuvconstants, |
4129 | 6.57k | int width) { |
4130 | | // Row buffer for intermediate ARGB pixels. |
4131 | 6.57k | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4132 | 13.3k | while (width > 0) { |
4133 | 6.82k | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4134 | 6.82k | I444ToARGBRow_AVX2(src_y, src_u, src_v, row, yuvconstants, twidth); |
4135 | 6.82k | #if defined(HAS_ARGBTORGB24ROW_AVX2) |
4136 | 6.82k | ARGBToRGB24Row_AVX2(row, dst_rgb24, twidth); |
4137 | | #else |
4138 | | ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth); |
4139 | | #endif |
4140 | 6.82k | src_y += twidth; |
4141 | 6.82k | src_u += twidth; |
4142 | 6.82k | src_v += twidth; |
4143 | 6.82k | dst_rgb24 += twidth * 3; |
4144 | 6.82k | width -= twidth; |
4145 | 6.82k | } |
4146 | 6.57k | } |
4147 | | #endif |
4148 | | |
4149 | | #if defined(HAS_NV12TORGB565ROW_AVX2) |
4150 | | void NV12ToRGB565Row_AVX2(const uint8_t* src_y, |
4151 | | const uint8_t* src_uv, |
4152 | | uint8_t* dst_rgb565, |
4153 | | const struct YuvConstants* yuvconstants, |
4154 | 0 | int width) { |
4155 | | // Row buffer for intermediate ARGB pixels. |
4156 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4157 | 0 | while (width > 0) { |
4158 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4159 | 0 | NV12ToARGBRow_AVX2(src_y, src_uv, row, yuvconstants, twidth); |
4160 | | #if defined(HAS_ARGBTORGB565ROW_AVX2) |
4161 | | ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth); |
4162 | | #else |
4163 | 0 | ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth); |
4164 | 0 | #endif |
4165 | 0 | src_y += twidth; |
4166 | 0 | src_uv += twidth; |
4167 | 0 | dst_rgb565 += twidth * 2; |
4168 | 0 | width -= twidth; |
4169 | 0 | } |
4170 | 0 | } |
4171 | | #endif |
4172 | | |
4173 | | #ifdef HAS_RGB24TOYJROW_AVX2 |
4174 | | // Convert 16 RGB24 pixels (64 bytes) to 16 YJ values. |
4175 | 0 | void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) { |
4176 | | // Row buffer for intermediate ARGB pixels. |
4177 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4178 | 0 | while (width > 0) { |
4179 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4180 | 0 | RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth); |
4181 | 0 | ARGBToYJRow_AVX2(row, dst_yj, twidth); |
4182 | 0 | src_rgb24 += twidth * 3; |
4183 | 0 | dst_yj += twidth; |
4184 | 0 | width -= twidth; |
4185 | 0 | } |
4186 | 0 | } |
4187 | | #endif // HAS_RGB24TOYJROW_AVX2 |
4188 | | |
4189 | | #ifdef HAS_RAWTOYJROW_AVX2 |
4190 | | // Convert 32 RAW pixels (128 bytes) to 32 YJ values. |
4191 | 40 | void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_yj, int width) { |
4192 | | // Row buffer for intermediate ARGB pixels. |
4193 | 40 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4194 | 103k | while (width > 0) { |
4195 | 103k | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4196 | 103k | #ifdef HAS_RAWTOARGBROW_AVX2 |
4197 | 103k | RAWToARGBRow_AVX2(src_raw, row, twidth); |
4198 | | #else |
4199 | | RAWToARGBRow_SSSE3(src_raw, row, twidth); |
4200 | | #endif |
4201 | 103k | ARGBToYJRow_AVX2(row, dst_yj, twidth); |
4202 | 103k | src_raw += twidth * 3; |
4203 | 103k | dst_yj += twidth; |
4204 | 103k | width -= twidth; |
4205 | 103k | } |
4206 | 40 | } |
4207 | | #endif // HAS_RAWTOYJROW_AVX2 |
4208 | | |
4209 | | #ifdef HAS_RGB24TOYJROW_SSSE3 |
4210 | | // Convert 16 RGB24 pixels (64 bytes) to 16 YJ values. |
4211 | 0 | void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) { |
4212 | | // Row buffer for intermediate ARGB pixels. |
4213 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4214 | 0 | while (width > 0) { |
4215 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4216 | 0 | RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth); |
4217 | 0 | ARGBToYJRow_SSSE3(row, dst_yj, twidth); |
4218 | 0 | src_rgb24 += twidth * 3; |
4219 | 0 | dst_yj += twidth; |
4220 | 0 | width -= twidth; |
4221 | 0 | } |
4222 | 0 | } |
4223 | | #endif // HAS_RGB24TOYJROW_SSSE3 |
4224 | | |
4225 | | #ifdef HAS_RAWTOYJROW_SSSE3 |
4226 | | // Convert 16 RAW pixels (64 bytes) to 16 YJ values. |
4227 | 0 | void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_yj, int width) { |
4228 | | // Row buffer for intermediate ARGB pixels. |
4229 | 0 | SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]); |
4230 | 0 | while (width > 0) { |
4231 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4232 | 0 | RAWToARGBRow_SSSE3(src_raw, row, twidth); |
4233 | 0 | ARGBToYJRow_SSSE3(row, dst_yj, twidth); |
4234 | 0 | src_raw += twidth * 3; |
4235 | 0 | dst_yj += twidth; |
4236 | 0 | width -= twidth; |
4237 | 0 | } |
4238 | 0 | } |
4239 | | #endif // HAS_RAWTOYJROW_SSSE3 |
4240 | | |
4241 | | #ifdef HAS_INTERPOLATEROW_16TO8_AVX2 |
4242 | | void InterpolateRow_16To8_AVX2(uint8_t* dst_ptr, |
4243 | | const uint16_t* src_ptr, |
4244 | | ptrdiff_t src_stride, |
4245 | | int scale, |
4246 | | int width, |
4247 | 0 | int source_y_fraction) { |
4248 | | // Row buffer for intermediate 16 bit pixels. |
4249 | 0 | SIMD_ALIGNED(uint16_t row[MAXTWIDTH]); |
4250 | 0 | while (width > 0) { |
4251 | 0 | int twidth = width > MAXTWIDTH ? MAXTWIDTH : width; |
4252 | 0 | InterpolateRow_16_C(row, src_ptr, src_stride, twidth, source_y_fraction); |
4253 | 0 | Convert16To8Row_AVX2(row, dst_ptr, scale, twidth); |
4254 | 0 | src_ptr += twidth; |
4255 | 0 | dst_ptr += twidth; |
4256 | 0 | width -= twidth; |
4257 | 0 | } |
4258 | 0 | } |
4259 | | #endif // HAS_INTERPOLATEROW_16TO8_AVX2 |
4260 | | |
4261 | 0 | float ScaleSumSamples_C(const float* src, float* dst, float scale, int width) { |
4262 | 0 | float fsum = 0.f; |
4263 | 0 | int i; |
4264 | 0 | for (i = 0; i < width; ++i) { |
4265 | 0 | float v = *src++; |
4266 | 0 | fsum += v * v; |
4267 | 0 | *dst++ = v * scale; |
4268 | 0 | } |
4269 | 0 | return fsum; |
4270 | 0 | } |
4271 | | |
4272 | 0 | float ScaleMaxSamples_C(const float* src, float* dst, float scale, int width) { |
4273 | 0 | float fmax = 0.f; |
4274 | 0 | int i; |
4275 | 0 | for (i = 0; i < width; ++i) { |
4276 | 0 | float v = *src++; |
4277 | 0 | float vs = v * scale; |
4278 | 0 | fmax = (v > fmax) ? v : fmax; |
4279 | 0 | *dst++ = vs; |
4280 | 0 | } |
4281 | 0 | return fmax; |
4282 | 0 | } |
4283 | | |
4284 | 0 | void ScaleSamples_C(const float* src, float* dst, float scale, int width) { |
4285 | 0 | int i; |
4286 | 0 | for (i = 0; i < width; ++i) { |
4287 | 0 | *dst++ = *src++ * scale; |
4288 | 0 | } |
4289 | 0 | } |
4290 | | |
4291 | 0 | void GaussRow_C(const uint32_t* src, uint16_t* dst, int width) { |
4292 | 0 | int i; |
4293 | 0 | for (i = 0; i < width; ++i) { |
4294 | 0 | *dst++ = STATIC_CAST( |
4295 | 0 | uint16_t, |
4296 | 0 | (src[0] + src[1] * 4 + src[2] * 6 + src[3] * 4 + src[4] + 128) >> 8); |
4297 | 0 | ++src; |
4298 | 0 | } |
4299 | 0 | } |
4300 | | |
4301 | | // filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row. |
4302 | | void GaussCol_C(const uint16_t* src0, |
4303 | | const uint16_t* src1, |
4304 | | const uint16_t* src2, |
4305 | | const uint16_t* src3, |
4306 | | const uint16_t* src4, |
4307 | | uint32_t* dst, |
4308 | 0 | int width) { |
4309 | 0 | int i; |
4310 | 0 | for (i = 0; i < width; ++i) { |
4311 | 0 | *dst++ = *src0++ + *src1++ * 4 + *src2++ * 6 + *src3++ * 4 + *src4++; |
4312 | 0 | } |
4313 | 0 | } |
4314 | | |
4315 | 0 | void GaussRow_F32_C(const float* src, float* dst, int width) { |
4316 | 0 | int i; |
4317 | 0 | for (i = 0; i < width; ++i) { |
4318 | 0 | *dst++ = (src[0] + src[1] * 4 + src[2] * 6 + src[3] * 4 + src[4]) * |
4319 | 0 | (1.0f / 256.0f); |
4320 | 0 | ++src; |
4321 | 0 | } |
4322 | 0 | } |
4323 | | |
4324 | | // filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row. |
4325 | | void GaussCol_F32_C(const float* src0, |
4326 | | const float* src1, |
4327 | | const float* src2, |
4328 | | const float* src3, |
4329 | | const float* src4, |
4330 | | float* dst, |
4331 | 0 | int width) { |
4332 | 0 | int i; |
4333 | 0 | for (i = 0; i < width; ++i) { |
4334 | 0 | *dst++ = *src0++ + *src1++ * 4 + *src2++ * 6 + *src3++ * 4 + *src4++; |
4335 | 0 | } |
4336 | 0 | } |
4337 | | |
4338 | | // Convert biplanar NV21 to packed YUV24 |
4339 | | void NV21ToYUV24Row_C(const uint8_t* src_y, |
4340 | | const uint8_t* src_vu, |
4341 | | uint8_t* dst_yuv24, |
4342 | 0 | int width) { |
4343 | 0 | int x; |
4344 | 0 | for (x = 0; x < width - 1; x += 2) { |
4345 | 0 | dst_yuv24[0] = src_vu[0]; // V |
4346 | 0 | dst_yuv24[1] = src_vu[1]; // U |
4347 | 0 | dst_yuv24[2] = src_y[0]; // Y0 |
4348 | 0 | dst_yuv24[3] = src_vu[0]; // V |
4349 | 0 | dst_yuv24[4] = src_vu[1]; // U |
4350 | 0 | dst_yuv24[5] = src_y[1]; // Y1 |
4351 | 0 | src_y += 2; |
4352 | 0 | src_vu += 2; |
4353 | 0 | dst_yuv24 += 6; // Advance 2 pixels. |
4354 | 0 | } |
4355 | 0 | if (width & 1) { |
4356 | 0 | dst_yuv24[0] = src_vu[0]; // V |
4357 | 0 | dst_yuv24[1] = src_vu[1]; // U |
4358 | 0 | dst_yuv24[2] = src_y[0]; // Y0 |
4359 | 0 | } |
4360 | 0 | } |
4361 | | |
4362 | | // Filter 2 rows of AYUV UV's (444) into UV (420). |
4363 | | // AYUV is VUYA in memory. UV for NV12 is UV order in memory. |
4364 | | void AYUVToUVRow_C(const uint8_t* src_ayuv, |
4365 | | int src_stride_ayuv, |
4366 | | uint8_t* dst_uv, |
4367 | 0 | int width) { |
4368 | | // Output a row of UV values, filtering 2x2 rows of AYUV. |
4369 | 0 | int x; |
4370 | 0 | for (x = 0; x < width - 1; x += 2) { |
4371 | 0 | dst_uv[0] = (src_ayuv[1] + src_ayuv[5] + src_ayuv[src_stride_ayuv + 1] + |
4372 | 0 | src_ayuv[src_stride_ayuv + 5] + 2) >> |
4373 | 0 | 2; |
4374 | 0 | dst_uv[1] = (src_ayuv[0] + src_ayuv[4] + src_ayuv[src_stride_ayuv + 0] + |
4375 | 0 | src_ayuv[src_stride_ayuv + 4] + 2) >> |
4376 | 0 | 2; |
4377 | 0 | src_ayuv += 8; |
4378 | 0 | dst_uv += 2; |
4379 | 0 | } |
4380 | 0 | if (width & 1) { |
4381 | 0 | dst_uv[0] = (src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] + 1) >> 1; |
4382 | 0 | dst_uv[1] = (src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] + 1) >> 1; |
4383 | 0 | } |
4384 | 0 | } |
4385 | | |
4386 | | // Filter 2 rows of AYUV UV's (444) into VU (420). |
4387 | | void AYUVToVURow_C(const uint8_t* src_ayuv, |
4388 | | int src_stride_ayuv, |
4389 | | uint8_t* dst_vu, |
4390 | 0 | int width) { |
4391 | | // Output a row of VU values, filtering 2x2 rows of AYUV. |
4392 | 0 | int x; |
4393 | 0 | for (x = 0; x < width - 1; x += 2) { |
4394 | 0 | dst_vu[0] = (src_ayuv[0] + src_ayuv[4] + src_ayuv[src_stride_ayuv + 0] + |
4395 | 0 | src_ayuv[src_stride_ayuv + 4] + 2) >> |
4396 | 0 | 2; |
4397 | 0 | dst_vu[1] = (src_ayuv[1] + src_ayuv[5] + src_ayuv[src_stride_ayuv + 1] + |
4398 | 0 | src_ayuv[src_stride_ayuv + 5] + 2) >> |
4399 | 0 | 2; |
4400 | 0 | src_ayuv += 8; |
4401 | 0 | dst_vu += 2; |
4402 | 0 | } |
4403 | 0 | if (width & 1) { |
4404 | 0 | dst_vu[0] = (src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] + 1) >> 1; |
4405 | 0 | dst_vu[1] = (src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] + 1) >> 1; |
4406 | 0 | } |
4407 | 0 | } |
4408 | | |
4409 | | // Copy row of AYUV Y's into Y |
4410 | 0 | void AYUVToYRow_C(const uint8_t* src_ayuv, uint8_t* dst_y, int width) { |
4411 | | // Output a row of Y values. |
4412 | 0 | int x; |
4413 | 0 | for (x = 0; x < width; ++x) { |
4414 | 0 | dst_y[x] = src_ayuv[2]; // v,u,y,a |
4415 | 0 | src_ayuv += 4; |
4416 | 0 | } |
4417 | 0 | } |
4418 | | |
4419 | | // Convert UV plane of NV12 to VU of NV21. |
4420 | 0 | void SwapUVRow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width) { |
4421 | 0 | int x; |
4422 | 0 | for (x = 0; x < width; ++x) { |
4423 | 0 | uint8_t u = src_uv[0]; |
4424 | 0 | uint8_t v = src_uv[1]; |
4425 | 0 | dst_vu[0] = v; |
4426 | 0 | dst_vu[1] = u; |
4427 | 0 | src_uv += 2; |
4428 | 0 | dst_vu += 2; |
4429 | 0 | } |
4430 | 0 | } |
4431 | | |
4432 | | void HalfMergeUVRow_C(const uint8_t* src_u, |
4433 | | int src_stride_u, |
4434 | | const uint8_t* src_v, |
4435 | | int src_stride_v, |
4436 | | uint8_t* dst_uv, |
4437 | 0 | int width) { |
4438 | 0 | int x; |
4439 | 0 | for (x = 0; x < width - 1; x += 2) { |
4440 | 0 | dst_uv[0] = (src_u[0] + src_u[1] + src_u[src_stride_u] + |
4441 | 0 | src_u[src_stride_u + 1] + 2) >> |
4442 | 0 | 2; |
4443 | 0 | dst_uv[1] = (src_v[0] + src_v[1] + src_v[src_stride_v] + |
4444 | 0 | src_v[src_stride_v + 1] + 2) >> |
4445 | 0 | 2; |
4446 | 0 | src_u += 2; |
4447 | 0 | src_v += 2; |
4448 | 0 | dst_uv += 2; |
4449 | 0 | } |
4450 | 0 | if (width & 1) { |
4451 | 0 | dst_uv[0] = (src_u[0] + src_u[src_stride_u] + 1) >> 1; |
4452 | 0 | dst_uv[1] = (src_v[0] + src_v[src_stride_v] + 1) >> 1; |
4453 | 0 | } |
4454 | 0 | } |
4455 | | |
4456 | | #undef STATIC_CAST |
4457 | | |
4458 | | #ifdef __cplusplus |
4459 | | } // extern "C" |
4460 | | } // namespace libyuv |
4461 | | #endif |