2 * Copyright 2012 The LibYuv Project Authors. All rights reserved.
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.
11 #include "libyuv/row.h"
13 #include "libyuv/basic_types.h"
20 // TODO(fbarchard): Consider 'any' functions handling any quantity of pixels.
21 // TODO(fbarchard): Consider 'any' functions handling odd alignment.
22 // YUV to RGB does multiple of 8 with SIMD and remainder with C.
23 #define YANY(NAMEANY, I420TORGB_SIMD, I420TORGB_C, UV_SHIFT, BPP, MASK) \
24 void NAMEANY(const uint8* y_buf, \
29 int n = width & ~MASK; \
30 I420TORGB_SIMD(y_buf, u_buf, v_buf, rgb_buf, n); \
31 I420TORGB_C(y_buf + n, \
32 u_buf + (n >> UV_SHIFT), \
33 v_buf + (n >> UV_SHIFT), \
34 rgb_buf + n * BPP, width & MASK); \
37 #ifdef HAS_I422TOARGBROW_SSSE3
38 YANY(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, I422ToARGBRow_C,
40 #endif // HAS_I422TOARGBROW_SSSE3
41 #ifdef HAS_I444TOARGBROW_SSSE3
42 YANY(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, I444ToARGBRow_C,
44 YANY(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, I411ToARGBRow_C,
46 YANY(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_SSSE3, I422ToBGRARow_C,
48 YANY(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_SSSE3, I422ToABGRRow_C,
50 YANY(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, I422ToRGBARow_C,
52 // I422ToRGB565Row_SSSE3 is unaligned.
53 YANY(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, I422ToARGB4444Row_C,
55 YANY(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, I422ToARGB1555Row_C,
57 YANY(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, I422ToRGB565Row_C,
59 // I422ToRGB24Row_SSSE3 is unaligned.
60 YANY(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, I422ToRGB24Row_C, 1, 3, 7)
61 YANY(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, I422ToRAWRow_C, 1, 3, 7)
62 YANY(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, I422ToYUY2Row_C, 1, 2, 15)
63 YANY(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, I422ToUYVYRow_C, 1, 2, 15)
64 #endif // HAS_I444TOARGBROW_SSSE3
65 #ifdef HAS_I422TOARGBROW_AVX2
66 YANY(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, I422ToARGBRow_C, 1, 4, 15)
67 #endif // HAS_I422TOARGBROW_AVX2
68 #ifdef HAS_I422TOBGRAROW_AVX2
69 YANY(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, I422ToBGRARow_C, 1, 4, 15)
70 #endif // HAS_I422TOBGRAROW_AVX2
71 #ifdef HAS_I422TOARGBROW_NEON
72 YANY(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, I444ToARGBRow_C, 0, 4, 7)
73 YANY(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, I422ToARGBRow_C, 1, 4, 7)
74 YANY(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, I411ToARGBRow_C, 2, 4, 7)
75 YANY(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, I422ToBGRARow_C, 1, 4, 7)
76 YANY(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, I422ToABGRRow_C, 1, 4, 7)
77 YANY(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, I422ToRGBARow_C, 1, 4, 7)
78 YANY(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, I422ToRGB24Row_C, 1, 3, 7)
79 YANY(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, I422ToRAWRow_C, 1, 3, 7)
80 YANY(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, I422ToARGB4444Row_C,
82 YANY(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, I422ToARGB1555Row_C,
84 YANY(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, I422ToRGB565Row_C, 1, 2, 7)
85 #endif // HAS_I422TOARGBROW_NEON
86 #ifdef HAS_I422TOYUY2ROW_NEON
87 YANY(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, I422ToYUY2Row_C, 1, 2, 15)
88 #endif // HAS_I422TOYUY2ROW_NEON
89 #ifdef HAS_I422TOUYVYROW_NEON
90 YANY(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, I422ToUYVYRow_C, 1, 2, 15)
91 #endif // HAS_I422TOUYVYROW_NEON
94 // Wrappers to handle odd width
95 #define NV2NY(NAMEANY, NV12TORGB_SIMD, NV12TORGB_C, UV_SHIFT, BPP) \
96 void NAMEANY(const uint8* y_buf, \
97 const uint8* uv_buf, \
100 int n = width & ~7; \
101 NV12TORGB_SIMD(y_buf, uv_buf, rgb_buf, n); \
102 NV12TORGB_C(y_buf + n, \
103 uv_buf + (n >> UV_SHIFT), \
104 rgb_buf + n * BPP, width & 7); \
107 #ifdef HAS_NV12TOARGBROW_SSSE3
108 NV2NY(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_SSSE3, NV12ToARGBRow_C,
110 NV2NY(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, NV21ToARGBRow_C,
112 #endif // HAS_NV12TOARGBROW_SSSE3
113 #ifdef HAS_NV12TOARGBROW_NEON
114 NV2NY(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, NV12ToARGBRow_C, 0, 4)
115 NV2NY(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, NV21ToARGBRow_C, 0, 4)
116 #endif // HAS_NV12TOARGBROW_NEON
117 #ifdef HAS_NV12TORGB565ROW_SSSE3
118 NV2NY(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, NV12ToRGB565Row_C,
120 NV2NY(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, NV21ToRGB565Row_C,
122 #endif // HAS_NV12TORGB565ROW_SSSE3
123 #ifdef HAS_NV12TORGB565ROW_NEON
124 NV2NY(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, NV12ToRGB565Row_C, 0, 2)
125 NV2NY(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, NV21ToRGB565Row_C, 0, 2)
126 #endif // HAS_NV12TORGB565ROW_NEON
129 #define RGBANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP) \
130 void NAMEANY(const uint8* src, \
133 int n = width & ~MASK; \
134 ARGBTORGB_SIMD(src, dst, n); \
135 ARGBTORGB_C(src + n * SBPP, dst + n * BPP, width & MASK); \
138 #if defined(HAS_ARGBTORGB24ROW_SSSE3)
139 RGBANY(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, ARGBToRGB24Row_C,
141 RGBANY(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, ARGBToRAWRow_C,
143 RGBANY(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, ARGBToRGB565Row_C,
145 RGBANY(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, ARGBToARGB1555Row_C,
147 RGBANY(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, ARGBToARGB4444Row_C,
150 #if defined(HAS_I400TOARGBROW_SSE2)
151 RGBANY(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, I400ToARGBRow_C,
154 #if defined(HAS_YTOARGBROW_SSE2)
155 RGBANY(YToARGBRow_Any_SSE2, YToARGBRow_SSE2, YToARGBRow_C,
157 RGBANY(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, YUY2ToARGBRow_C,
159 RGBANY(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, UYVYToARGBRow_C,
161 // These require alignment on ARGB, so C is used for remainder.
162 RGBANY(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, RGB24ToARGBRow_C,
164 RGBANY(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, RAWToARGBRow_C,
166 RGBANY(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, RGB565ToARGBRow_C,
168 RGBANY(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, ARGB1555ToARGBRow_C,
170 RGBANY(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, ARGB4444ToARGBRow_C,
173 #if defined(HAS_ARGBTORGB24ROW_NEON)
174 RGBANY(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, ARGBToRGB24Row_C, 7, 4, 3)
175 RGBANY(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, ARGBToRAWRow_C, 7, 4, 3)
176 RGBANY(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, ARGBToRGB565Row_C,
178 RGBANY(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, ARGBToARGB1555Row_C,
180 RGBANY(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, ARGBToARGB4444Row_C,
182 RGBANY(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, I400ToARGBRow_C,
184 RGBANY(YToARGBRow_Any_NEON, YToARGBRow_NEON, YToARGBRow_C,
186 RGBANY(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, YUY2ToARGBRow_C,
188 RGBANY(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, UYVYToARGBRow_C,
193 // ARGB to Bayer does multiple of 4 pixels, SSSE3 aligned src, unaligned dst.
194 #define BAYERANY(NAMEANY, ARGBTORGB_SIMD, ARGBTORGB_C, MASK, SBPP, BPP) \
195 void NAMEANY(const uint8* src, \
196 uint8* dst, uint32 selector, \
198 int n = width & ~MASK; \
199 ARGBTORGB_SIMD(src, dst, selector, n); \
200 ARGBTORGB_C(src + n * SBPP, dst + n * BPP, selector, width & MASK); \
203 #if defined(HAS_ARGBTOBAYERROW_SSSE3)
204 BAYERANY(ARGBToBayerRow_Any_SSSE3, ARGBToBayerRow_SSSE3, ARGBToBayerRow_C,
207 #if defined(HAS_ARGBTOBAYERROW_NEON)
208 BAYERANY(ARGBToBayerRow_Any_NEON, ARGBToBayerRow_NEON, ARGBToBayerRow_C,
211 #if defined(HAS_ARGBTOBAYERGGROW_SSE2)
212 BAYERANY(ARGBToBayerGGRow_Any_SSE2, ARGBToBayerGGRow_SSE2, ARGBToBayerGGRow_C,
215 #if defined(HAS_ARGBTOBAYERGGROW_NEON)
216 BAYERANY(ARGBToBayerGGRow_Any_NEON, ARGBToBayerGGRow_NEON, ARGBToBayerGGRow_C,
222 // RGB/YUV to Y does multiple of 16 with SIMD and last 16 with SIMD.
223 #define YANY(NAMEANY, ARGBTOY_SIMD, SBPP, BPP, NUM) \
224 void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) { \
225 ARGBTOY_SIMD(src_argb, dst_y, width - NUM); \
226 ARGBTOY_SIMD(src_argb + (width - NUM) * SBPP, \
227 dst_y + (width - NUM) * BPP, NUM); \
230 #ifdef HAS_ARGBTOYROW_AVX2
231 YANY(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 4, 1, 32)
232 YANY(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 4, 1, 32)
233 YANY(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 2, 1, 32)
234 YANY(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 2, 1, 32)
236 #ifdef HAS_ARGBTOYROW_SSSE3
237 YANY(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 4, 1, 16)
239 #ifdef HAS_BGRATOYROW_SSSE3
240 YANY(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 4, 1, 16)
241 YANY(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 4, 1, 16)
242 YANY(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 4, 1, 16)
243 YANY(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 2, 1, 16)
244 YANY(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 2, 1, 16)
246 #ifdef HAS_ARGBTOYJROW_SSSE3
247 YANY(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 4, 1, 16)
249 #ifdef HAS_ARGBTOYROW_NEON
250 YANY(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 4, 1, 8)
252 #ifdef HAS_ARGBTOYJROW_NEON
253 YANY(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 4, 1, 8)
255 #ifdef HAS_BGRATOYROW_NEON
256 YANY(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 4, 1, 8)
258 #ifdef HAS_ABGRTOYROW_NEON
259 YANY(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 4, 1, 8)
261 #ifdef HAS_RGBATOYROW_NEON
262 YANY(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 4, 1, 8)
264 #ifdef HAS_RGB24TOYROW_NEON
265 YANY(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 3, 1, 8)
267 #ifdef HAS_RAWTOYROW_NEON
268 YANY(RAWToYRow_Any_NEON, RAWToYRow_NEON, 3, 1, 8)
270 #ifdef HAS_RGB565TOYROW_NEON
271 YANY(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 2, 1, 8)
273 #ifdef HAS_ARGB1555TOYROW_NEON
274 YANY(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 2, 1, 8)
276 #ifdef HAS_ARGB4444TOYROW_NEON
277 YANY(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 2, 1, 8)
279 #ifdef HAS_YUY2TOYROW_NEON
280 YANY(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 2, 1, 16)
282 #ifdef HAS_UYVYTOYROW_NEON
283 YANY(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 2, 1, 16)
285 #ifdef HAS_RGB24TOARGBROW_NEON
286 YANY(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 3, 4, 8)
288 #ifdef HAS_RAWTOARGBROW_NEON
289 YANY(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 3, 4, 8)
291 #ifdef HAS_RGB565TOARGBROW_NEON
292 YANY(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 2, 4, 8)
294 #ifdef HAS_ARGB1555TOARGBROW_NEON
295 YANY(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 2, 4, 8)
297 #ifdef HAS_ARGB4444TOARGBROW_NEON
298 YANY(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 2, 4, 8)
302 #define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK) \
303 void NAMEANY(const uint8* src_argb, uint8* dst_y, int width) { \
304 int n = width & ~MASK; \
305 ARGBTOY_SIMD(src_argb, dst_y, n); \
306 ARGBTOY_C(src_argb + n * SBPP, \
307 dst_y + n * BPP, width & MASK); \
310 // Attenuate is destructive so last16 method can not be used due to overlap.
311 #ifdef HAS_ARGBATTENUATEROW_SSSE3
312 YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
315 #ifdef HAS_ARGBATTENUATEROW_SSE2
316 YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
319 #ifdef HAS_ARGBUNATTENUATEROW_SSE2
320 YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
323 #ifdef HAS_ARGBATTENUATEROW_AVX2
324 YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
327 #ifdef HAS_ARGBUNATTENUATEROW_AVX2
328 YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
331 #ifdef HAS_ARGBATTENUATEROW_NEON
332 YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
337 // RGB/YUV to UV does multiple of 16 with SIMD and remainder with C.
338 #define UVANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK) \
339 void NAMEANY(const uint8* src_argb, int src_stride_argb, \
340 uint8* dst_u, uint8* dst_v, int width) { \
341 int n = width & ~MASK; \
342 ANYTOUV_SIMD(src_argb, src_stride_argb, dst_u, dst_v, n); \
343 ANYTOUV_C(src_argb + n * BPP, src_stride_argb, \
349 #ifdef HAS_ARGBTOUVROW_AVX2
350 UVANY(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, ARGBToUVRow_C, 4, 31)
351 UVANY(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, YUY2ToUVRow_C, 2, 31)
352 UVANY(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, UYVYToUVRow_C, 2, 31)
354 #ifdef HAS_ARGBTOUVROW_SSSE3
355 UVANY(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, ARGBToUVRow_C, 4, 15)
356 UVANY(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, ARGBToUVJRow_C, 4, 15)
357 UVANY(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, BGRAToUVRow_C, 4, 15)
358 UVANY(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, ABGRToUVRow_C, 4, 15)
359 UVANY(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, RGBAToUVRow_C, 4, 15)
361 #ifdef HAS_YUY2TOUVROW_SSE2
362 UVANY(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, YUY2ToUVRow_C, 2, 15)
363 UVANY(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, UYVYToUVRow_C, 2, 15)
365 #ifdef HAS_ARGBTOUVROW_NEON
366 UVANY(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, ARGBToUVRow_C, 4, 15)
368 #ifdef HAS_ARGBTOUVJROW_NEON
369 UVANY(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, ARGBToUVJRow_C, 4, 15)
371 #ifdef HAS_BGRATOUVROW_NEON
372 UVANY(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, BGRAToUVRow_C, 4, 15)
374 #ifdef HAS_ABGRTOUVROW_NEON
375 UVANY(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, ABGRToUVRow_C, 4, 15)
377 #ifdef HAS_RGBATOUVROW_NEON
378 UVANY(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, RGBAToUVRow_C, 4, 15)
380 #ifdef HAS_RGB24TOUVROW_NEON
381 UVANY(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, RGB24ToUVRow_C, 3, 15)
383 #ifdef HAS_RAWTOUVROW_NEON
384 UVANY(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, RAWToUVRow_C, 3, 15)
386 #ifdef HAS_RGB565TOUVROW_NEON
387 UVANY(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, RGB565ToUVRow_C, 2, 15)
389 #ifdef HAS_ARGB1555TOUVROW_NEON
390 UVANY(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, ARGB1555ToUVRow_C, 2, 15)
392 #ifdef HAS_ARGB4444TOUVROW_NEON
393 UVANY(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, ARGB4444ToUVRow_C, 2, 15)
395 #ifdef HAS_YUY2TOUVROW_NEON
396 UVANY(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, YUY2ToUVRow_C, 2, 15)
398 #ifdef HAS_UYVYTOUVROW_NEON
399 UVANY(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, UYVYToUVRow_C, 2, 15)
403 #define UV422ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, BPP, MASK, SHIFT) \
404 void NAMEANY(const uint8* src_uv, \
405 uint8* dst_u, uint8* dst_v, int width) { \
406 int n = width & ~MASK; \
407 ANYTOUV_SIMD(src_uv, dst_u, dst_v, n); \
408 ANYTOUV_C(src_uv + n * BPP, \
409 dst_u + (n >> SHIFT), \
410 dst_v + (n >> SHIFT), \
414 #ifdef HAS_ARGBTOUV444ROW_SSSE3
415 UV422ANY(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3,
416 ARGBToUV444Row_C, 4, 15, 0)
418 #ifdef HAS_YUY2TOUV422ROW_AVX2
419 UV422ANY(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2,
420 YUY2ToUV422Row_C, 2, 31, 1)
421 UV422ANY(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2,
422 UYVYToUV422Row_C, 2, 31, 1)
424 #ifdef HAS_ARGBTOUV422ROW_SSSE3
425 UV422ANY(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3,
426 ARGBToUV422Row_C, 4, 15, 1)
428 #ifdef HAS_YUY2TOUV422ROW_SSE2
429 UV422ANY(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2,
430 YUY2ToUV422Row_C, 2, 15, 1)
431 UV422ANY(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2,
432 UYVYToUV422Row_C, 2, 15, 1)
434 #ifdef HAS_YUY2TOUV422ROW_NEON
435 UV422ANY(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON,
436 ARGBToUV444Row_C, 4, 7, 0)
437 UV422ANY(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON,
438 ARGBToUV422Row_C, 4, 15, 1)
439 UV422ANY(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON,
440 ARGBToUV411Row_C, 4, 31, 2)
441 UV422ANY(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON,
442 YUY2ToUV422Row_C, 2, 15, 1)
443 UV422ANY(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON,
444 UYVYToUV422Row_C, 2, 15, 1)
448 #define SPLITUVROWANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK) \
449 void NAMEANY(const uint8* src_uv, \
450 uint8* dst_u, uint8* dst_v, int width) { \
451 int n = width & ~MASK; \
452 ANYTOUV_SIMD(src_uv, dst_u, dst_v, n); \
453 ANYTOUV_C(src_uv + n * 2, \
459 #ifdef HAS_SPLITUVROW_SSE2
460 SPLITUVROWANY(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, SplitUVRow_C, 15)
462 #ifdef HAS_SPLITUVROW_AVX2
463 SPLITUVROWANY(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, SplitUVRow_C, 31)
465 #ifdef HAS_SPLITUVROW_NEON
466 SPLITUVROWANY(SplitUVRow_Any_NEON, SplitUVRow_NEON, SplitUVRow_C, 15)
468 #ifdef HAS_SPLITUVROW_MIPS_DSPR2
469 SPLITUVROWANY(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_MIPS_DSPR2,
474 #define MERGEUVROW_ANY(NAMEANY, ANYTOUV_SIMD, ANYTOUV_C, MASK) \
475 void NAMEANY(const uint8* src_u, const uint8* src_v, \
476 uint8* dst_uv, int width) { \
477 int n = width & ~MASK; \
478 ANYTOUV_SIMD(src_u, src_v, dst_uv, n); \
479 ANYTOUV_C(src_u + n, \
485 #ifdef HAS_MERGEUVROW_SSE2
486 MERGEUVROW_ANY(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, MergeUVRow_C, 15)
488 #ifdef HAS_MERGEUVROW_AVX2
489 MERGEUVROW_ANY(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, MergeUVRow_C, 31)
491 #ifdef HAS_MERGEUVROW_NEON
492 MERGEUVROW_ANY(MergeUVRow_Any_NEON, MergeUVRow_NEON, MergeUVRow_C, 15)
494 #undef MERGEUVROW_ANY
496 #define MATHROW_ANY(NAMEANY, ARGBMATH_SIMD, ARGBMATH_C, MASK) \
497 void NAMEANY(const uint8* src_argb0, const uint8* src_argb1, \
498 uint8* dst_argb, int width) { \
499 int n = width & ~MASK; \
500 ARGBMATH_SIMD(src_argb0, src_argb1, dst_argb, n); \
501 ARGBMATH_C(src_argb0 + n * 4, \
507 #ifdef HAS_ARGBMULTIPLYROW_SSE2
508 MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
511 #ifdef HAS_ARGBADDROW_SSE2
512 MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
514 #ifdef HAS_ARGBSUBTRACTROW_SSE2
515 MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
518 #ifdef HAS_ARGBMULTIPLYROW_AVX2
519 MATHROW_ANY(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, ARGBMultiplyRow_C,
522 #ifdef HAS_ARGBADDROW_AVX2
523 MATHROW_ANY(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, ARGBAddRow_C, 7)
525 #ifdef HAS_ARGBSUBTRACTROW_AVX2
526 MATHROW_ANY(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, ARGBSubtractRow_C,
529 #ifdef HAS_ARGBMULTIPLYROW_NEON
530 MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
533 #ifdef HAS_ARGBADDROW_NEON
534 MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
536 #ifdef HAS_ARGBSUBTRACTROW_NEON
537 MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
542 // Shuffle may want to work in place, so last16 method can not be used.
543 #define YANY(NAMEANY, ARGBTOY_SIMD, ARGBTOY_C, SBPP, BPP, MASK) \
544 void NAMEANY(const uint8* src_argb, uint8* dst_argb, \
545 const uint8* shuffler, int width) { \
546 int n = width & ~MASK; \
547 ARGBTOY_SIMD(src_argb, dst_argb, shuffler, n); \
548 ARGBTOY_C(src_argb + n * SBPP, \
549 dst_argb + n * BPP, shuffler, width & MASK); \
552 #ifdef HAS_ARGBSHUFFLEROW_SSE2
553 YANY(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2,
554 ARGBShuffleRow_C, 4, 4, 3)
556 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
557 YANY(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3,
558 ARGBShuffleRow_C, 4, 4, 7)
560 #ifdef HAS_ARGBSHUFFLEROW_AVX2
561 YANY(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2,
562 ARGBShuffleRow_C, 4, 4, 15)
564 #ifdef HAS_ARGBSHUFFLEROW_NEON
565 YANY(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON,
566 ARGBShuffleRow_C, 4, 4, 3)
570 // Interpolate may want to work in place, so last16 method can not be used.
571 #define NANY(NAMEANY, TERP_SIMD, TERP_C, SBPP, BPP, MASK) \
572 void NAMEANY(uint8* dst_ptr, const uint8* src_ptr, \
573 ptrdiff_t src_stride_ptr, int width, \
574 int source_y_fraction) { \
575 int n = width & ~MASK; \
576 TERP_SIMD(dst_ptr, src_ptr, src_stride_ptr, \
577 n, source_y_fraction); \
578 TERP_C(dst_ptr + n * BPP, \
579 src_ptr + n * SBPP, src_stride_ptr, \
580 width & MASK, source_y_fraction); \
583 #ifdef HAS_INTERPOLATEROW_AVX2
584 NANY(InterpolateRow_Any_AVX2, InterpolateRow_AVX2,
585 InterpolateRow_C, 1, 1, 32)
587 #ifdef HAS_INTERPOLATEROW_SSSE3
588 NANY(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3,
589 InterpolateRow_C, 1, 1, 15)
591 #ifdef HAS_INTERPOLATEROW_SSE2
592 NANY(InterpolateRow_Any_SSE2, InterpolateRow_SSE2,
593 InterpolateRow_C, 1, 1, 15)
595 #ifdef HAS_INTERPOLATEROW_NEON
596 NANY(InterpolateRow_Any_NEON, InterpolateRow_NEON,
597 InterpolateRow_C, 1, 1, 15)
599 #ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
600 NANY(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2,
601 InterpolateRow_C, 1, 1, 3)
607 } // namespace libyuv